diff --git a/NeurIPS/2025/$_boldsymbol{_lambda}$-Orthogonality Regularization for Compatible Representation Learning/c36592bf-d5fb-4809-b5a5-c95787696dbf_content_list.json b/NeurIPS/2025/$_boldsymbol{_lambda}$-Orthogonality Regularization for Compatible Representation Learning/c36592bf-d5fb-4809-b5a5-c95787696dbf_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..9a2d2e76fc63c08c0ecf8da0ce0d14403bf810e7 --- /dev/null +++ b/NeurIPS/2025/$_boldsymbol{_lambda}$-Orthogonality Regularization for Compatible Representation Learning/c36592bf-d5fb-4809-b5a5-c95787696dbf_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:19cef4a0ea991f419012ce9cc650d7a31e890ad2ac6f430582aed5c9de187baa +size 174916 diff --git a/NeurIPS/2025/$_boldsymbol{_lambda}$-Orthogonality Regularization for Compatible Representation Learning/c36592bf-d5fb-4809-b5a5-c95787696dbf_model.json b/NeurIPS/2025/$_boldsymbol{_lambda}$-Orthogonality Regularization for Compatible Representation Learning/c36592bf-d5fb-4809-b5a5-c95787696dbf_model.json new file mode 100644 index 0000000000000000000000000000000000000000..6e5a583b67ec6078d64bcdbb08bcf8b281ccfc2e --- /dev/null +++ b/NeurIPS/2025/$_boldsymbol{_lambda}$-Orthogonality Regularization for Compatible Representation Learning/c36592bf-d5fb-4809-b5a5-c95787696dbf_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4715a381048ebf4709c5d484ff10c165be84f96498b3c358d4bfeb5b04d2a6b7 +size 222595 diff --git a/NeurIPS/2025/$_boldsymbol{_lambda}$-Orthogonality Regularization for Compatible Representation Learning/c36592bf-d5fb-4809-b5a5-c95787696dbf_origin.pdf b/NeurIPS/2025/$_boldsymbol{_lambda}$-Orthogonality Regularization for Compatible Representation Learning/c36592bf-d5fb-4809-b5a5-c95787696dbf_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..31751999c118fdcdb9950740334baf82e390bc2b --- /dev/null +++ b/NeurIPS/2025/$_boldsymbol{_lambda}$-Orthogonality Regularization for Compatible Representation Learning/c36592bf-d5fb-4809-b5a5-c95787696dbf_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:377397f1af6409e83afb2fddf9d752a809e60a71b6eb887d34d9ade64186dcc6 +size 2771191 diff --git a/NeurIPS/2025/$_boldsymbol{_lambda}$-Orthogonality Regularization for Compatible Representation Learning/full.md b/NeurIPS/2025/$_boldsymbol{_lambda}$-Orthogonality Regularization for Compatible Representation Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e677f666cb430afa18e8588f2e71fc5fecc1485d --- /dev/null +++ b/NeurIPS/2025/$_boldsymbol{_lambda}$-Orthogonality Regularization for Compatible Representation Learning/full.md @@ -0,0 +1,714 @@ +# $\lambda$ -Orthogonality Regularization for Compatible Representation Learning + +Simone Ricci $^{1,2*}$ Niccolò Biondi $^{1,2}$ Federico Pernici $^{1,2}$ + +Ioannis Patras3 Alberto Del Bimbo1,2 + +$^{1}$ DINFO (Department of Information Engineering), University of Florence, Italy + $^{2}$ MICC (Media Integration and Communication Center) + $^{3}$ Queen Mary University of London, UK + +# Abstract + +Retrieval systems rely on representations learned by increasingly powerful models. However, due to the high training cost and inconsistencies in learned representations, there is significant interest in facilitating communication between representations and ensuring compatibility across independently trained neural networks. In the literature, two primary approaches are commonly used to adapt different learned representations: affine transformations, which adapt well to specific distributions but can significantly alter the original representation, and orthogonal transformations, which preserve the original structure with strict geometric constraints but limit adaptability. A key challenge is adapting the latent spaces of updated models to align with those of previous models on downstream distributions while preserving the newly learned representation spaces. In this paper, we impose a relaxed orthogonality constraint, namely $\lambda$ -Orthogonality regularization, while learning an affine transformation, to obtain distribution-specific adaptation while retaining the original learned representations. Extensive experiments across various architectures and datasets validate our approach, demonstrating that it preserves the model's zero-shot performance and ensures compatibility across model updates. Code available at: https://github.com/miccunifi/lambda_orthogonality. + +# 1 Introduction + +Retrieval tasks are increasingly relevant in real-world applications such as face recognition [1, 2, 3], image localization [4, 5, 6], and object identification [7, 8, 9]. In image retrieval, a gallery of labeled images is matched to query images to identify related ones, ideally of the same class. Instead of high-dimensional images, retrieval uses low-dimensional feature vectors obtained from embedding models. Enhancing retrieval performance often involves updating embedding models [10, 11] to leverage more expressive network architectures [12], new training techniques (e.g., loss functions) or training paradigms [13, 14, 15]. However, neural networks rarely produce compatible features, even when trained on the same data with identical methods and architectures [16]. Consequently, matching the features of new queries with those of older galleries can degrade retrieval performance due to incompatibility [15]. To address this, replacing the gallery features generated by the old model with those produced by the new model—a computationally expensive process known as backfilling—is required. The challenge of updating a base model while ensuring its backward compatibility and avoiding backfilling has been extensively investigated [17, 15, 18, 19, 20, 21]. Furthermore, the + +![](images/ec1b6055a6504637945c21f924ca7be6f18f4b3d7389e44d2c8ee0c558bc5929.jpg) +Figure 1: Overview of the proposed approach for achieving representation compatibility during retrieval system updates. A newly independently trained model is aligned to the old representation space via an orthogonal transformation $B_{\perp}$ , which preserves geometric structure. A forward transformation $F$ maps the old representations to the backward-aligned space of the new model. Only the transformation parameters are optimized during training, while model parameters remain fixed. + +![](images/a6e943f78190b66d4cfff4e6f3623d274f1d1944ff410ee6278170042d333a0a.jpg) + +![](images/e894996c0fd9ce8e0e81a3fedbd32b390809b0966c120eb8bd8bc719cb411677.jpg) + +![](images/4fcfa2544a03119abbe01c4f7f4b85ba9ba1bec7735d5495e3212bfa71728ff7.jpg) + +![](images/27ae4c27de964514455a64223f10701374d8bb853fbbfac41d9abdf2d310e5c9.jpg) + +optimal strategy for gallery updates—known as partial backfilling—has recently begun to receive attention [22]. + +Architectural changes and additional losses to ensure compatibility can reduce the performance of the updated model [23, 24]. To address this issue, research has focused on aligning the representation of a base model with that of an improved independently trained model using parameter-efficient adapters [22, 25]. On the other hand, the manifold hypothesis [26, 27] suggests that neural networks typically produce latent space representations of identical data distributions that differ primarily by a transformation. Consequently, mapping one representation to another requires only a few parameters, as functionally equivalent models approximate the same latent manifold [28, 29, 27]. Thus, a simple transformation aligning the new representation space to the previous one can provide the backward compatibility of the updated model. + +Recent studies have focused on affine and orthogonal mappings to adapt the latent space of a base model (source space) to that of another model (target space), using specific data points as reference [30, 28, 31]. Within the plasticity-stability paradigm [32], affine mappings offer high adaptability (plasticity) but may alter the source space's configuration [33, 34]. Conversely, orthogonal mappings maintain the source space's geometric structure (stability), though they offer no adaptability to a different distribution. To preserve the geometric structure of the source space, particularly when it is more informative than the target space [28, 35], while enabling adaptability, we propose a novel regularization term. Different from previous work [36], our term constrains a transformation to remain within a specified proximity to the orthogonality condition, controlled by a hyperparameter $\lambda$ . + +In this paper, we address the challenge of ensuring compatibility between independently trained models by learning different transformations across representation spaces, as illustrated in Fig. 1. Our contributions are summarized as follows: + +- We propose $\lambda$ -Orthogonality regularization, a relaxed orthogonality constraint that retains the original representation space's global structure while enabling slight local adaptations for downstream tasks. +- We enhance representation compatibility by employing a supervised contrastive loss, which promotes intra-class clustering and inter-model alignment of feature representations, while remaining agnostic to model architecture. +- We conduct extensive experiments across diverse architectures and datasets, demonstrating that our method not only ensures compatibility between models but also promotes the preservation of the base model's latent space geometry, resulting in improved accuracy on downstream tasks. + +- We propose a novel architecture-agnostic backfilling strategy that improves retrieval performance while optimizing the gallery update process. + +# 2 Related Works + +As demonstrated by [16], feature representations from two models—even if trained on the same data—do not generally coincide, creating costly backfilling in retrieval systems. To avoid this, [15] introduced Backward Compatible Training (BCT), which keeps the old classifier fixed as a reference so new embeddings align with prior class prototypes. Additionally, they provided a formal definition of compatibility between model representations. Subsequent research has expanded on this foundation, incorporating additional regularization techniques to better align new representations with previous ones [21, 37, 20, 38, 39] and implementing specific architecture design [18, 13, 40]. However, the performance of the updated backward-compatible models frequently falls to reach that of models trained independently [23], a consequence of the regularization imposed to achieve compatibility. To avoid this, [23] and [24] suggested expanding the representation space to include new classes while ensuring that the representations of old classes remain aligned during updates. To ensure compatibility between models trained independently, mapping-based strategies have been developed [41, 42, 43]. Forward Compatible Training (FCT), as detailed by [25], introduces a function that aligns embeddings from an older model to those of a newer model's space, incorporating additional side information for each data point. As noted by [25], the computational overhead of these transformations is minimal compared to the demands of processing images through the embedding model. FastFill [22] improves forward transformation learning by using a new model classifier and proposes a Bayesian strategy to optimize the gallery backfilling process leveraging the new model. In contrast, we propose a set of transformation functions to ensure not only forward but also backward compatibility during model updates, with a particular focus on the orthogonality property in backward mappings. Additionally, we propose a supervised contrastive loss that promotes intra-class clustering and inter-modality alignment, thereby enhancing adaptation. Finally, we propose a novel gallery backfilling strategy based on a distance metric that directly operates on pre-extracted gallery representations, making it agnostic to the underlying architecture. + +# 3 Method + +To achieve compatible representations between independently trained models, we introduce a theoretically grounded pipeline composed of multiple transformations. First, in Sec. 3.1 we report the definition of compatibility introduced by [15]. In Sec. 3.2 and 3.3, we introduce a novel backward-optimability method, which aligns the new model's representations to those of the previous model using either a strict orthogonal transformation or when adapting to a downstream task a transformation regularized by our proposed $\lambda$ -Orthogonality constraint. Next, in Sec. 3.4, we present forward transformation learning, which aligns the previous model's representations to those of the newly adapted model via an affine or more complex transformation, enabling effective gallery set updates. We also apply a supervised contrastive loss (Sec. 3.5) during transformation training to improve alignment between model representations and enhance intra-class cluster compactness, thereby satisfying the compatibility criterion defined in Def. 3.1. Finally, in Sec. 3.6, we propose a novel ordering strategy for backfilling the gallery with improved representations in an optimized sequence. Throughout our methodology, all models serve as fixed feature extractors with frozen parameters, while only the transformation layers are trained. + +# 3.1 Backward-Compatible Representations Definition + +The formulation of Backward-Compatibility between representations, introduced by [15], is closely related to the concept of latent space communication between different models [30]. The formal definition of backward-compatible representations specifies: + +Definition 3.1 (Backward-Compatibility). The representation of a model learned at step $k$ is compatible with the representation of a distinct model learned at a subsequent step $t$ , where $k < t$ , if the following condition is satisfied: + +$$ +\forall i, j: \left(y _ {i} = y _ {j} \Rightarrow d \left(\mathbf {h} _ {i} ^ {t}, \mathbf {h} _ {j} ^ {k}\right) \leq d \left(\mathbf {h} _ {i} ^ {k}, \mathbf {h} _ {j} ^ {k}\right)\right) \wedge \left(y _ {i} \neq y _ {j} \Rightarrow d \left(\mathbf {h} _ {i} ^ {t}, \mathbf {h} _ {j} ^ {k}\right) \geq d \left(\mathbf {h} _ {i} ^ {k}, \mathbf {h} _ {j} ^ {k}\right)\right) \tag {1} +$$ + +![](images/e81a3a93ef4154799d1caa7356c10fccdb1ad6cdc5571f5a006736966033c690.jpg) +(a) Value of Eq. 6 at different $\lambda$ . + +![](images/8a3da53ab4e51f16a3395b76df48c3137d70c173bcad7fbf0843bb9e08ce7799.jpg) +(b) Effect of different $\alpha$ at $\lambda = 6$ + +![](images/720f7d8c68a7d7a5976cb4b13c80095f06e92a0d9830071343c8eb467f1cb481.jpg) +(c)KDE of $W$ angles. +Figure 2: Impact of $\lambda$ -Orthogonality regularization on affine transformations. Fig. 2a shows the variation of Eq. 6 for different values of $\lambda$ , demonstrating the influence of the threshold in the regularization. Fig. 2b illustrates the effect of varying $\alpha$ while keeping $\lambda = 6$ , highlighting its behavior in the sigmoid function. Fig. 2c presents the kernel density estimation (KDE) of angles between the columns of matrix $W$ for different values of $\lambda$ , showcasing the impact of regularization on orthogonality preservation. + +where $d(\cdot, \cdot)$ is a distance function and $y_i$ and $y_j$ are the class labels associated with the extracted representation vectors $\mathbf{h}_i$ and $\mathbf{h}_j$ , respectively. The inequalities in Def. 3.1 indicate that the new model's representation, when compared against the old representation, should perform at least as well as the previous model's in clustering images from the same class and separating them from those of different classes. + +# 3.2 Backward Transformation + +One of the contributions of relative encoding [30] is the observation that representation spaces, in practice, often differ only by an angle-preserving transformation when they share the same or similar data semantics. Furthermore, [28] demonstrates that when there is a difference in learned semantics, a transformation that preserves both angles and distances—learned with Procrustes analysis [44]—yields superior performance in cross-architecture and cross-modality classification tasks than only angle-preserving mappings. A transformation $T$ is defined as an isometry if it preserves angles and distances between any two points $a$ and $b$ in the space. Formally, a mapping $T: \mathbb{R}^n \to \mathbb{R}^n$ is an isometry if the following condition holds: $\| T(a) - T(b) \|_2 = \| a - b \|_2$ , $\forall a, b \in \mathbb{R}^n$ , where $\| \cdot \|_2$ denotes the Euclidean norm, or equivalently, a general distance metric in other spaces. We leverage this property to achieve backward-compatible representations, aligning the updated model's space with the base model's using an orthogonal transformation. This maintains a unified representation space across updates, preserving the geometric properties and performance of the updated model due to the isometric nature of the transformation. + +Given a base model $\phi^k$ and its updated version $\phi^t$ , with $k < t$ , and their corresponding representation vectors $\mathbf{h}^k \in \mathbb{R}^d$ and $\mathbf{h}^t \in \mathbb{R}^n$ , we learn an orthogonal transformation $B_\perp: \mathbb{R}^n \to \mathbb{R}^n$ that maps the embedding space of the updated model into the space of the base model. To enforce strict orthogonality, a generic transformation $B$ is parameterized as the matrix exponential of a skew-symmetric matrix $P$ , such that $B = e^P$ , where the upper triangular entries of $P$ are learnable parameters [45]. To enforce alignment between the updated and base representation spaces, we optimize the transformation $B_\perp$ by minimizing the Mean Squared Error loss between $\mathbf{h}^k$ and the transformed $\mathbf{h}^t$ : + +$$ +\mathcal {L} _ {B} = \left| \left| B _ {\perp} \left(\mathbf {h} ^ {t}\right) - \mathbf {h} ^ {k} \right| \right| _ {2} ^ {2} \tag {2} +$$ + +As the transformation $B_{\perp}$ is a square matrix, if the dimensionalities of the two representation spaces differ, the higher-dimensional feature vector is truncated to match the dimensionality of the smaller representation space. + +# 3.3 $\lambda$ -Orthogonality Regularization + +A strict orthogonal constraint (high stability) on a transformation $B$ might not be ideal when model distributions vary from those on which the adapter is trained—the case of private models providing only their extracted embeddings to the user. Imposing such a constraint can limit the integration of + +![](images/a4d7c2863ba0b9aca756601c3faa747a105f0b3d8fddfd4d754674bd2cdaa575.jpg) +(a) Source space + +![](images/2f51f40cb316847d058b7d23a91ccd0fd06d222cebb192d68ef207dfaa38394a.jpg) +Figure 3: Effects of affine (Fig. 3c), strictly orthogonal (Fig. 3d), and $\lambda$ -orthogonality (with $\lambda = 1$ ) regularized (Fig. 3e) transformations trained to align a source representation space (Fig. 3a) learned with a LeNet model (embedding dimension $= 2$ ) on the complete MNIST dataset, with a target representation space (Fig. 3b) learned on the first five classes of MNIST using the same architecture. + +![](images/bbb0c6dccdadfda39421b1849d6b2d4f793fd751e7b55ef4898c85b0cec961e1.jpg) +(b) Target space +(c) Affine + +![](images/e1291b1d7c0cd3e50dabc13bca69630714bc1cca4ba97afbec0db64db25347bd.jpg) +(d) Orthogonal + +![](images/039bfb48690065df2844d7deccb383ce2b0e7259db632450edf7abdc8db00831.jpg) +(e) $\lambda$ -orthogonality + +new, relevant information for downstream tasks. Conversely, an affine transformation (high plasticity) without geometric regularization may disrupt the updated model's representations [46, 47]. As described in [36], a soft orthogonality constraint can be applied to the transformation $B: \mathbb{R}^n \to \mathbb{R}^n$ , consisting of a weight matrix $W \in \mathbb{R}^{n \times n}$ and a bias term $b \in \mathbb{R}^n$ . Previous works [48, 49, 50] have proposed constraining the Gram matrix of the weight matrix to be close to the identity matrix by minimizing a loss function defined as: + +$$ +\mathcal {L} _ {\text {o r t h}} = \left\| W ^ {T} W - I \right\| _ {F} \tag {3} +$$ + +where $||\cdot ||_F$ denotes the Frobenius norm and $W$ is the weight of the transformation $B$ . This can be interpreted as a weight decay term that restricts the set of parameters to lie close to a Stiefel manifold [50]. However, this approach does not provide control over the specific level of orthogonality that can be imposed on the transformation. + +To this end, we introduce a threshold $\lambda$ , which specifies the desired proximity of the Gram matrix of the weight matrix to the identity matrix. A naive solution would be to stop the optimization of the $\mathcal{L}_{orth}$ loss once the Gram matrix of the weight matrix reaches the threshold $\lambda$ , as the loss directly influences the Gram matrix: + +$$ +\min _ {W} \left\| W ^ {T} W - I \right\| _ {F} \quad \text {s . t .} \quad \left\| W ^ {T} W - I \right\| _ {F} \geq \lambda \tag {4} +$$ + +This objective can be achieved directly through the use of a Heaviside step function [51, 52], shifted by the parameter $\lambda$ : $H(x - \lambda) = \mathbf{1}_{\{x \geq \lambda\}}$ . This function $H$ offers an efficient mechanism to control the degree of orthogonality during the minimization process, that effectively deactivating the regularization term in Eq. 3 when the Frobenius norm exceeds the threshold $\lambda$ : + +$$ +\mathcal {L} _ {\lambda} = H \left(\| W W ^ {T} - I \| _ {F} - \lambda\right) \cdot \| W W ^ {T} - I \| _ {F}. \tag {5} +$$ + +However, this approach introduces a discontinuity in the loss function, as highlighted by [53]. In particular, their work focuses on evaluating the closeness of these sigmoid functions to the Heaviside step function, providing precise upper and lower bounds for the Hausdorff distance. Building on their theoretical and empirical analysis, we propose a smooth modulating function that ensures the effect of the constraint is gradually adjusted, with the penalty becoming more or less significant depending on the distance from the threshold $\lambda$ . Specifically, we formulate a novel $\lambda$ -Orthogonality Regularization term by optimizing a loss function defined as: + +$$ +\mathcal {L} _ {\lambda} = \sigma (\alpha (\| W W ^ {T} - I \| _ {F} - \lambda)) \cdot \| W W ^ {T} - I \| _ {F} \tag {6} +$$ + +where $\sigma (\cdot)$ is the sigmoid function, and $\alpha$ is a scaling factor. The sigmoid function acts as a continuous switch that gradually turns the regularization term on and off near the value of $\lambda$ , as shown in Fig. 2a. Instead, the scaling factor $\alpha$ controls the steepness of the sigmoid function, which in turn determines how sharply the regularization is activated or deactivated as the value of $\| WW^T - I \|_F$ approaches the threshold $\lambda$ . In Fig. 2b, we illustrate different levels of steepness applied to the regularization loss. As $\alpha$ increases, its behavior converges more closely to the Heaviside step function. + +To further analyze the behavior of the $\lambda$ -orthogonality regularization, we optimize Eq. 6, applied to a transformation $B$ with a randomly initialized weight matrix $W$ . As shown in Fig. 2c, the kernel density estimation (KDE) of these angles changes based on the value of $\lambda$ used in the regularization. Smaller values of $\lambda$ result in column vectors that become increasingly orthogonal, specifically when + +$\lambda = 0$ our regularization is equal to Eq.3. Fig. 3 illustrates the effects of an affine (Fig. 3c), strictly orthogonal (Fig. 3d), and $\lambda$ -orthogonality regularized (Fig. 3e) transformations trained to align a source representation space (Fig. 3a) learned on the full MNIST dataset with a target representation space (Fig. 3b) learned on the first five classes of MNIST. The toy experiment shows that the $\lambda$ -orthogonal constraint improves alignment by relaxing strict orthogonality while encouraging preservation of the source feature space structure, in contrast to an unconstrained transformation. + +# 3.4 Forward Transformation + +In addition to a backward transformation that maps the representations of the new model to those of the previous model, it is possible to formulate a forward transformation $F: \mathbb{R}^d \to \mathbb{R}^n$ . This transformation maps the representation vector $\mathbf{h}^k \in \mathbb{R}^d$ of the previous model to $\mathbf{h}^t \in \mathbb{R}^n$ , the representation of the new model. Since the representation of the new model is superior to that of the previous model, the transformation $F$ should be affine (high plasticity) or multiple projection layers to better adapt to the improved representation. The transformation $F$ is learned by minimizing the Mean Squared Error between the two representations, as $||F(\mathbf{h}^k) - \mathbf{h}^t||_2^2$ , following the approach described in [25]. This concept is closely related to latent space communication [30, 31], where $d\big(\mathbf{h}_i^k,\mathbf{h}_j^k\big) = d\big(\mathcal{T}\mathbf{h}_i^t,\mathcal{T}\mathbf{h}_j^t\big)$ with $\mathcal{T}$ is a generic transformation. In previous approaches [25, 22] the old representations $\mathbf{h}^k$ are aligned directly with the new $\mathbf{h}^t$ through transformation $F$ , but incompatibility arises between $\mathbf{h}^k$ and $F(\mathbf{h}^k)$ . As mentioned in Sec. 3.2, a backward orthogonal transformation $B_{\perp}$ realigns new representations with the old ones. Instead of adapting old features directly to new representations $\mathbf{h}^t$ , we adapt them to $B_{\perp}(\mathbf{h}^t)$ , ensuring a unified alignment across model updates. Furthermore, the transformations $F$ and $B_{\perp}$ can be trained jointly, as they utilize the same training data. Consequently, the forward alignment loss in our methodology is defined as: + +$$ +\mathcal {L} _ {F} = \left| \left| F \left(\mathbf {h} ^ {k}\right) - B _ {\perp} \left(\mathbf {h} ^ {t}\right) \right| \right| _ {2} ^ {2} \tag {7} +$$ + +If the extracted representations are derived from a dataset different from the training sets of the two models, as discussed in Sec. 3.3, a $\lambda$ -orthogonal regularized transformation $B_{\lambda}$ can be employed in place of the strictly orthogonal $B_{\perp}$ . + +# 3.5 Intra-class Clustering and Inter-Model Alignment + +As discussed in Sec. 3.1, the compatibility inequalities defined in Def. 3.1 require not only alignment but also a higher concentration of clusters to achieve compatibility. To this end, [22] introduces an additional training loss, $\mathcal{L}_{disc}$ , that, unlike the influence loss in [15], relies directly on the new model's classifier rather than the old one. However, $\mathcal{L}_{disc}$ depends on access to the new model's classifier and training loss, limiting its applicability, especially when the new model's architecture is unknown (e.g., embedding vectors from private or online models). To overcome this, we propose the use of a supervised contrastive loss, applied directly to representation vectors. This loss requires no classifier or architectural knowledge, as it directly leverages representation vectors for alignment and clustering. The supervised contrastive loss [54] minimizes the cross-entropy loss between $\mathbf{q}_i$ and $\mathbf{p}_i$ : + +$$ +\mathcal {L} _ {\mathrm {c o n t r}} = - \sum_ {i = 1} ^ {K} \mathbf {p} _ {i} \log \mathbf {q} _ {i} \tag {8} +$$ + +where $\mathbf{q}_i$ denotes the probability assigned to sample $i$ by applying a temperature-scaled softmax over the dot-product similarities between the L2-normalized feature $\mathbf{h}$ and each other candidate, and $\mathbf{p}_i$ is the normalized ground-truth indicator distribution that places equal mass on all semantically matching (same-class) candidates and zero on all others. Specifically, we utilized a combination of this loss function, where the objective $\mathcal{L}_{\mathrm{C}}$ is defined as: + +$$ +\mathcal {L} _ {\mathrm {C}} = \mathcal {L} _ {\text {c o n t r}} \left(F \left(\mathbf {h} ^ {k}\right), B _ {\perp} \left(\mathbf {h} ^ {t}\right)\right) + \mathcal {L} _ {\text {c o n t r}} \left(F \left(\mathbf {h} ^ {k}\right), \mathbf {h} ^ {k}\right) \tag {9} +$$ + +This loss encourages clustering among the adapted representations while also aligning them with those of the previous model, thereby promoting intra-class clustering and inter-model alignment of feature representations. + +The overall loss function of our framework is defined as a weighted sum of four components: the forward alignment loss $\mathcal{L}_F$ , the backward alignment loss $\mathcal{L}_B$ , the contrastive loss $\mathcal{L}_{\mathrm{C}}$ , and the $\lambda$ -Orthogonality regularization term $\mathcal{L}_{\lambda}$ . Formally, the total loss is expressed as: + +$$ +\mathcal {L} = w _ {1} \cdot \mathcal {L} _ {F} + w _ {2} \cdot \mathcal {L} _ {B} + w _ {3} \cdot \mathcal {L} _ {C} + \mathcal {L} _ {\lambda} \tag {10} +$$ + +Table 1: Compatibility evaluation on ImageNet1K under two scenarios: (a) Extending classes setting, (b) Architecture update setting. For each case (highlighted with different colors), we report CMC-Top1 and mAP metrics. + +(a) Extending classes setting. Two models trained independently: $\phi_{\mathrm{old}}$ on first 500 classes, and $\phi_{\mathrm{new}}$ on full ImageNet1K. Both use ResNet-34 with an embedding dimension of 128. + +
MethodQuery/GalleryCMC-Top1mAP
Ind. Train.φold/φold43.5625.18
φnew/φold0.100.15
φnew/φnew61.6135.69
FCT [25]F(φold)/φold0.100.15
F(φold)/F(φold)50.1330.93
φnew/F(φold)57.2133.00
FastFill [22]F(φold)/φold0.100.15
F(φold)/F(φold)50.6331.48
φnew/F(φold)57.2133.19
OursF(φold)/φold44.5926.70
F(φold)/F(φold)51.4633.75
B⊥(φnew)/F(φold)57.4134.53
B⊥(φnew)/φold43.9425.75
B⊥(φnew)/B⊥(φnew)61.6135.69
+ +(b) Independently Pretrained Models setting: Two models trained independently on the full ImageNet1K dataset. The first model, $\phi_{\mathrm{old}}$ is a ResNet-18, whereas the second, $\phi_{\mathrm{new}}$ is a ViT-L-16. + +
MethodQuery/GalleryCMC-Top1mAP
Ind. Train.φold/φold55.6226.91
φnew/φold0.040.17
φnew/φnew76.6256.84
FCT [25]F(φold)/φold0.040.17
F(φold)/F(φold)59.3942.65
φnew/F(φold)72.5449.85
FastFill [22]F(φold)/φold0.040.17
F(φold)/F(φold)61.1746.28
φnew/F(φold)73.3352.83
OursF(φold)/φold60.8340.69
F(φold)/F(φold)61.1045.91
B⊥(φnew)/F(φold)73.5352.06
B⊥(φnew)/φold65.5438.55
B⊥(φnew)/B⊥(φnew)76.6256.84
+ +where $w_{1}, w_{2}$ , and $w_{3}$ denote scalar weights used to balance the contributions of each term. + +# 3.6 Partial Backfilling Strategy + +Determining an effective ordering for backfilling samples in the forward-adapted gallery set, where $F(\mathbf{h}^k)$ from the old model is replaced by $B_{\perp}(\mathbf{h}^t)$ , is critical for achieving the performance of the new independently trained model as efficiently as possible. However, identifying the optimal ordering of backfilling represents a computationally intractable combinatorial problem [22]. To address this challenge, FastFill [22] introduces an ordering inspired by Bayesian Deep Learning. This approach models the alignment error as a multivariate Gaussian distribution and minimizes the negative log-likelihood of this distribution during the training of the mapping function $F$ . However, from a retrieval perspective, the most representative instances—those that significantly enhance the separation between distinct classes—are identified as the embeddings closest to their respective class means [55, 56]. Accordingly, prioritizing the backfilling of the least informative embeddings will increase the system's performance by reinforcing class distinctions. To this end, we propose a novel method for estimating a backfill ordering based directly on the already extracted representation vector $F(\mathbf{h}^k)$ . First, we calculate the mean representation vector $\mu_c$ for each class $c$ in the forward-adapted gallery set. Then, we compute a distance metric $d$ of each embedding vector $F(\mathbf{h}^k)$ from its corresponding class mean $\mu_c$ . For instance, $d$ can be the Mean Squared Error, $d = \| F(\mathbf{h}^k) - \mu_c\|_2$ . Gallery embedding exhibiting the largest distance $d$ from $\mu$ are prioritized for backfilling, thereby facilitating the matching with queries generated by the new backward-adapted independently trained model $B_{\perp}(\mathbf{h}^t)$ . + +# 4 Experiments + +# 4.1 Image Retrieval Compatibility + +Backward compatibility is crucial in retrieval tasks involving a gallery set $\mathcal{G} = \{(\mathbf{x}_i, y_i)\}_{i=1}^{N_g}$ and a query set $\mathcal{Q} = \{(\mathbf{x}_i, y_i)\}_{i=1}^{N_q}$ , each containing $N_g$ and $N_q$ images respectively, with associated class labels. A base model indexes the gallery by extracting feature vectors from the images, which are then used to match with vectors from the query set in retrieval tasks. The compatibility definition presented in Def. 3.1 involves computing pairwise distances between all datapoints in the dataset. This process becomes increasingly computationally demanding as the dataset size grows. Then, a model updated at step $t$ is considered backward-compatible with the base model trained at step $k$ if + +Table 2: Compatibility results for two models pretrained on ImageNet1K and adapted to downstream tasks: $\phi_{\mathrm{old}}$ , a ResNet-18, and $\phi_{\mathrm{new}}$ , a ViT-L-16, using as backward adapter $B_{\lambda}$ with $\lambda = 12$ . The ZS column indicates the CMC-Top1 performance increase on ImageNet1K, with values in parentheses indicating the increment compared to the newly independently trained model. Each Query/Gallery case is highlighted with a different color to facilitate comparison of results. + +
MethodQuery/GalleryDataset
CUBCIFAR100
CMC-Top1ZSCMC-Top1ZS
Ind. Train.φold/φold44.8251.13
φnew/φold0.40.8
φnew/φnew71.7874.08
FCT [25]F(φold)/φold0.040.8
F(φold)/F(φold)51.1057.35
φnew/F(φold)62.1369.80
FastFill [22]F(φold)/φold0.40.8
F(φold)/F(φold)54.5066.17
φnew/F(φold)61.4967.23
OursF(φold)/φold51.1267.29
F(φold)/F(φold)59.9267.72
Bλ(φnew)/F(φold)70.7272.08
Bλ(φnew)/φold60.6471.85
Bλ(φnew)/Bλ(φnew)75.44 (+3.66)+0.02578.23 (+4.15)+0.112
+ +the Empirical Compatibility Criterion [15] is satisfied: + +$$ +M \left(\Phi_ {t} ^ {\mathcal {Q}}, \Phi_ {k} ^ {\mathcal {G}}\right) > M \left(\Phi_ {k} ^ {\mathcal {Q}}, \Phi_ {k} ^ {\mathcal {G}}\right), \quad \text {w i t h} k < t \tag {11} +$$ + +where $M$ denote a performance metric, $\Phi^{\mathcal{G}}$ and $\Phi^{\mathcal{Q}}$ represent the extracted gallery and query sets, respectively. Specifically, $M\big(\Phi_t^\mathcal{Q},\Phi_k^\mathcal{G}\big)$ assesses cross-model retrieval with gallery features from the updated model at step $t$ and query features from step $k$ . In contrast, $M\big(\Phi_k^\mathcal{Q},\Phi_k^\mathcal{G}\big)$ refers to same-model retrieval, where both gallery and query features originate from the same model at step $k$ . + +Partial Backfilling. Given an ordering $\pi$ of the images in the gallery set $\Phi^{\mathcal{G}}$ , denoted as $\mathbf{x}_{\pi_1}, \mathbf{x}_{\pi_2}, \ldots, \mathbf{x}_{\pi_n}$ , and a backfilling fraction $\beta \in [0,1]$ , we define the partially backfilled gallery set $\Phi_{\pi,\beta}^{\mathcal{G}}$ as follows. The first $N_{g,\beta} = \lfloor \beta N_g \rfloor$ images in the ordering are processed using the updated model, while the remaining images are processed using the old model. Here, $N_g$ denotes the total number of images in the gallery. To evaluate different backfilling strategies, we employ the backfilling metric $\widetilde{M}$ , introduced in [22], which is defined as: $\widetilde{M}(\Phi^{\mathcal{G}}, \Phi^{\mathcal{Q}}, \pi) = \mathbb{E}_{\beta \sim [0,1]} M(\Phi_{\pi,\beta}^{\mathcal{G}}, \Phi^{\mathcal{Q}})$ . This metric is the area under the backfilling curve when evaluating performance using $M$ . + +# 4.2 Evaluation Metrics and Datasets + +Following prior work on model compatibility [15, 25], we evaluate performance using two metrics. The Cumulative Matching Characteristics (CMC), which measures top- $k$ retrieval accuracy by computing distances between query and gallery features, considers retrieval successful if at least one of the $k$ closest gallery images shares the query's label. The mean Average Precision (mAP) measures the area under the precision-recall curve across the full recall range [0, 1]. + +To validate our approach, we utilize the following datasets: ImageNet1K [57], CIFAR100 [58], and CUB200 [59]. Each dataset's validation/test set serves as both the query and gallery, with each query image removed from the gallery to avoid trivial matches during search. The notation 'Query/Gallery' indicates the models used for extracting embeddings in all tables, respectively. CUB200 and CIFAR100 are employed as downstream tasks. + +# 4.3 Extending Classes Setting + +In this setting, we update a base model by extending the number of classes. We train two models independently: $\phi_{\mathrm{old}}$ on the first 500 classes and $\phi_{\mathrm{new}}$ on all 1000 classes of ImageNet1K, both using a ResNet-34 architecture with an embedding dimension of 128, following PyTorch's standard training + +![](images/e3918417b22d4e646eb5232306e9295e42288c168ae3b684cefa7502beab9702.jpg) +Figure 4: Partial backfilling results for the Extending Classes setting (top Figures) of Tab. 1a, and Independently Pretrained Models setting (bottom Figures) of Tab. 1b. We use features from the new model $\phi_{\mathrm{new}}$ for the query set (otherwise $B_{\perp}(\phi_{\mathrm{new}})$ if trained). For the gallery set, we begin with forward-adapted old features $F(\phi_{\mathrm{old}})$ and incrementally replace them with new features. + +![](images/b5cf9dff1fe71c3f390aed3a8c28312e0635c7ebdbbfc987bb3313bfe186bbe2.jpg) + +Table 3: Extended Classes setting + +
MethodM
CMC-Top1mAP
FCT [25]58.7233.57
FastFill [22]60.4935.59
Ours61.2036.46
+ +Table 4: Independently Pre-trained Models setting + +
MethodM
CMC-Top1mAP
FCT [25]73.8652.02
FastFill [22]75.0655.34
Ours76.5957.72
+ +recipe $^2$ . After training the two models independently, adapters are optimized using Adam with a learning rate of 0.001, while keeping the model layers frozen. We compare our method against FCT [25] and FastFill [22], two mapping methods used to achieve compatible representations. In Tab. 1a, the performance of each method is summarized following the metrics of Sec. 4.2. The results indicate that the new model, $\phi_{\mathrm{new}}$ , is not directly compatible with the old one $\phi_{\mathrm{old}}$ . Additionally, the two mapping methods, FCT and FastFill, enhance performance across both metrics for the adapted representations of the gallery and query sets. However, these methods achieve backward compatibility with the newly trained model but not with the original one. In contrast, our method aligns the new model with the old one through the orthogonal transformation $B_{\perp}$ . This ensures compatibility between the new and old representations and also enhances the performance provided by the forward adapter $F$ . Appendix A provides additional results on Places365 [60] dataset. + +# 4.4 Independently Pretrained Models adapted on Downstream Task + +Due to escalating training costs, pretrained models are increasingly used, especially for adapting to downstream tasks with local datasets. In this context, we employ two models—available in the PyTorch hub—pretrained on the ImageNet1K dataset: a ResNet-18 with an embedding size of 512, and a more advanced Vision Transformer (ViT-L-16) [61] with an embedding size of 1024. The ViT model is considered an update over the ResNet-18 due to its enhanced architecture. Tab. 1b shows adapter training results using the same dataset as the two pretrained models, revealing a trend similar to Tab. 1a and demonstrating our method's comparable performance to other baselines, but with compatibility between the updated model and the previous one. Unlike FastFill, our approach does not require the new model's classifier, relying directly on the extracted embedding vectors. In Appendix B, to further validate our method, we apply our approach to different architectures used as pretrained models. Instead, in Appendix C, we investigate update scenarios involving distribution or objective shifts using CLIP-like [62] models and self-supervised architectures such as DINOv2 [63]. + +Results for compatibility on downstream tasks are reported in Tab. 2, where adapters are trained on representations from local datasets (CUB200 or CIFAR100) different from the training dataset. Employing a transformation $B_{\lambda}$ with $\lambda$ -Orthogonality regularization, our method enhances local task performance and model compatibility, outperforming the baselines. Results on additional downstream datasets (Flower102 [64] and Places365) are reported in Appendix D. From Tab. 1a and Tab. 1b, we observe that a strict orthogonal transformation, $B_{\perp}$ , does not result in performance improvements relative to the independently trained model $\phi_{\mathrm{new}}$ . Conversely, $B_{\lambda}$ , which provides more plasticity with respect to $B_{\perp}$ , enables the new model to enhance performance in the downstream task. An ablation study on the hyperparameter $\lambda$ is presented in Appendix E, and a component-wise ablation of the loss terms in Eq. 3.5 is detailed in Appendix F. + +# 4.5 Backfilling Results + +In this section, we evaluate our novel backfill strategy discussed in Sec. 3.6, considering the experimental setting detailed in Tab. 1a and Tab. 1b. Given that FCT lacks a specific backfilling strategy, we employ a random ordering as in [22]. The results, depicted in Fig. 4, Tab. 3, and Tab. 4, demonstrate that our backfilling strategy outperforms the other baselines by a certain margin. Notably, Fig. 4 illustrates that with less than $50\%$ of the gallery backfilled, we achieve the same performance as the newly independently trained model. In Appendix G, we provide an ablation study using an alternative distance metric to the Mean Squared Error employed in the main experiments. + +# 5 Conclusion + +Model compatibility is a critical challenge in many large-scale retrieval systems and can hinder system updates when not achieved. In this paper, we introduce mapping transformations that align independently learned representations under a unified space, also providing a more feature clustering through supervised contrastive loss. We also propose a relaxation of the orthogonality constraint to aid adaptation to downstream tasks without compromising the integrity of newly trained independent models. Additionally, we propose a novel backfill ordering strategy that enables efficient partial backfilling of the gallery set, achieving the performance of a newly independently trained model with less than half of the gallery backfilled. Our approach demonstrates superior performance compared to previous methods, across the same and different data distributions on which the models are trained. To contextualize these results, the limitations of the approach are examined in detail in Appendix I. Furthermore, to evaluate its practical utility, we analyze its methodological complexity and broader applicability in Appendix H. + +# Acknowledgments + +This paper was partially funded by the project "Collaborative Explainable neuro-symbolic AI for Decision Support Assistant", CAI4DSA, CUP B13C23005640006. + +# References + +[1] Florian Schroff, Dmitry Kalenichenko, and James Philbin. Facenet: A unified embedding for face recognition and clustering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 815-823, 2015. +[2] Weiyang Liu, Yandong Wen, Zhiding Yu, Ming Li, Bhiksha Raj, and Le Song. Spherface: Deep hypersphere embedding for face recognition. In 2017 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2017, Honolulu, HI, USA, July 21-26, 2017, pages 6738-6746. IEEE Computer Society, 2017. +[3] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4690-4699, 2019. +[4] Relja Arandjelovic, Petr Gronat, Akihiko Torii, Tomas Pajdla, and Josef Sivic. Netvlad: Cnn architecture for weakly supervised place recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5297-5307, 2016. +[5] Bingyi Cao, Andre Araujo, and Jack Sim. Unifying deep local and global features for image search. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XX 16, pages 726-743. Springer, 2020. +[6] Stephen Hausler, Sourav Garg, Ming Xu, Michael Milford, and Tobias Fischer. Patch-netvlad: Multi-scale fusion of locally-global descriptors for place recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14141-14152, 2021. +[7] Hyeonwoo Noh, Andre Araujo, Jack Sim, Tobias Weyand, and Bohyung Han. Large-scale image retrieval with attentive deep local features. In Proceedings of the IEEE international conference on computer vision, pages 3456-3465, 2017. + +[8] Fuwen Tan, Jiangbo Yuan, and Vicente Ordonez. Instance-level image retrieval using reranking transformers. In proceedings of the IEEE/CVF international conference on computer vision, pages 12105-12115, 2021. +[9] Bin Yan, Yi Jiang, Jiannan Wu, Dong Wang, Ping Luo, Zehuan Yuan, and Huchuan Lu. Universal instance perception as object discovery and retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15325-15336, 2023. +[10] Colin Raffel. Building machine learning models like open source software. Commun. ACM, 66(2):38-40, jan 2023. +[11] Prateek Yadav, Colin Raffel, Mohammed Muqeeth, Lucas Caccia, Haokun Liu, Tianlong Chen, Mohit Bansal, Leshem Choshen, and Alessandro Sordoni. A survey on model moering: Recycling and routing among specialized experts for collaborative learning. Trans. Mach. Learn. Res., 2025. +[12] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. +[13] Niccolò Biondi, Federico Pernici, Simone Ricci, and Alberto Del Bimbo. Stationary representations: Optimally approximating compatibility and implications for improved model replacements. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. +[14] Jessica Maria Echterhoff, Fartash Faghri, Raviteja Vemulapalli, Ting-Yao Hu, Chun-Liang Li, Oncel Tuzel, and Hadi Pouransari. MUSCLE: A model update strategy for compatible LLM evolution. In EMNLP (Findings), pages 7320-7332. Association for Computational Linguistics, 2024. +[15] Yantao Shen, Yuanjun Xiong, Wei Xia, and Stefano Soatto. Towards backward-compatible representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6368-6377, 2020. +[16] Yixuan Li, Jason Yosinski, Jeff Clune, Hod Lipson, and John Hopcroft. Convergent learning: Do different neural networks learn the same representations? In Yoshua Bengio and Yann LeCun, editors, Feature Extraction: Modern Questions and Challenges, pages 196-212. PMLR, 2015. +[17] Sijie Yan, Yuanjun Xiong, Kaustav Kundu, Shuo Yang, Siqi Deng, Meng Wang, Wei Xia, and Stefano Soatto. Positive-congruent training: Towards regression-free model updates. In CVPR, pages 14299-14308. Computer Vision Foundation / IEEE, 2021. +[18] Niccolo Biondi, Federico Pernici, Matteo Bruni, and Alberto Del Bimbo. Cores: Compatible representations via stationarity. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1-16, 2023. +[19] Mitchell Wortsman, Gabriel Ilharco, Samir Ya Gadre, Rebecca Roelofs, Raphael Gontijo-Lopes, Ari S Morcos, Hongseok Namkoong, Ali Farhadi, Yair Carmon, Simon Kornblith, et al. Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time. In International conference on machine learning, pages 23965-23998. PMLR, 2022. +[20] Binjie Zhang, Yixiao Ge, Yantao Shen, Shupeng Su, Fanzi Wu, Chun Yuan, Xuyuan Xu, Yexin Wang, and Ying Shan. Towards universal backward-compatible representation learning. In *IJCAI*, pages 1615–1621. ijcai.org, 2022. +[21] Qiang Meng, Chixiang Zhang, Xiaogiang Xu, and Feng Zhou. Learning compatible embeddings. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 9939-9948, October 2021. +[22] Florian Jaeckle, Fartash Faghri, Ali Farhadi, Oncel Tuzel, and Hadi Pouransari. Fastfill: Efficient compatible model update. In International Conference on Learning Representations, 2023. +[23] Yifei Zhou, Zilu Li, Abhinav Shrivastava, Hengshuang Zhao, Antonio Torralba, Taipeng Tian, and Ser-Nam Lim. Bt^2: Backward-compatible training with basis transformation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11229-11238, 2023. +[24] Simone Ricci, Niccolò Biondi, Federico Pernici, and Alberto Del Bimbo. Backward-compatible aligned representations via an orthogonal transformation layer. In ECCV Workshops (17), volume 15639 of Lecture Notes in Computer Science, pages 451-464. Springer, 2024. + +[25] Vivek Ramanujan, Pavan Kumar Anasosalu Vasu, Ali Farhadi, Oncel Tuzel, and Hadi Pouransari. Forward compatible training for large-scale embedding retrieval systems. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19386-19395, 2022. +[26] Charles Fefferman, Sanjoy Mitter, and Hariharan Narayanan. Testing the manifold hypothesis. Journal of the American Mathematical Society, 29(4):983-1049, 2016. +[27] Minyoung Huh, Brian Cheung, Tongzhou Wang, and Phillip Isola. Position: The platonic representation hypothesis. In ICML. OpenReview.net, 2024. +[28] Valentino Maiorca, Luca Moschella, Antonio Norelli, Marco Fumero, Francesco Locatello, and Emanuele Rodolà. Latent space translation via semantic alignment. Advances in Neural Information Processing Systems, 36, 2024. +[29] Marco Fumero, Marco Pegoraro, Valentino Maiorca, Francesco Locatello, and Emanuele Rodola. Latent functional maps: a spectral framework for representation alignment. In NeurIPS, 2024. +[30] Luca Moschella, Valentino Maiorca, Marco Fumero, Antonio Norelli, Francesco Locatello, and Emanuele Rodolà. Relative representations enable zero-shot latent space communication. In International Conference on Learning Representations, 2023. +[31] Valentino Maiorca, Luca Moschella, Marco Fumero, Francesco Locatello, and Emanuele Rodolà. Latent space translation via inverse relative projection. arXiv preprint arXiv:2406.15057, 2024. +[32] Martial Mermillod, Aurélia Bugaiska, and Patrick Bonin. The stability-plasticity dilemma: Investigating the continuum from catastrophic forgetting to age-limited learning effects, 2013. +[33] Guoliang Lin, Hanlu Chu, and Hanjiang Lai. Towards better plasticity-stability trade-off in incremental learning: A simple linear connector. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 89-98, 2022. +[34] Dongwan Kim and Bohyung Han. On the stability-plasticity dilemma of class-incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20196-20204, 2023. +[35] Lirong Wu, Zicheng Liu, Jun Xia, Zelin Zang, Siyuan Li, and Stan Z Li. Generalized clustering and multi-manifold learning with geometric structure preservation. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 139-147, 2022. +[36] Nitin Bansal, Xiaohan Chen, and Zhangyang Wang. Can we gain more from orthogonality regularizations in training deep networks? Advances in Neural Information Processing Systems, 31, 2018. +[37] Binjie Zhang, Yixiao Ge, Yantao Shen, Yu Li, Chun Yuan, XUYUAN XU, Yexin Wang, and Ying Shan. Hot-refresh model upgrades with regression-free compatible training in image retrieval. In International Conference on Learning Representations, 2021. +[38] Tan Pan, Furong Xu, Xudong Yang, Sifeng He, Chen Jiang, Qingpei Guo, Feng Qian, Xiaobo Zhang, Yuan Cheng, Lei Yang, et al. Boundary-aware backward-compatible representation via adversarial learning in image retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15201-15210, 2023. +[39] Mateusz Budnik and Yannis Avrithis. Asymmetric metric learning for knowledge transfer. In CVPR, pages 8228-8238. Computer Vision Foundation / IEEE, 2021. +[40] Niccolo Biondi, Federico Pernici, Matteo Bruni, Daniele Mugnai, and Alberto Del Bimbo. Cl2r: Compatible lifelong learning representations. ACM Transactions on Multimedia Computing, Communications and Applications, 18(2s):1-22, 2023. +[41] Ahmet Iscen, Jeffrey Zhang, Svetlana Lazebnik, and Cordelia Schmid. Memory-efficient incremental learning through feature adaptation. In European Conference on Computer Vision, pages 699-715. Springer, 2020. +[42] Chien-Yi Wang, Ya-Liang Chang, Shang-Ta Yang, Dong Chen, and Shang-Hong Lai. Unified representation learning for cross model compatibility. In 31st British Machine Vision Conference 2020, BMVC 2020. BMVA Press, 2020. +[43] Shupeng Su, Binjie Zhang, Yixiao Ge, Xuyuan Xu, Yexin Wang, Chun Yuan, and Ying Shan. Privacy-preserving model upgrades with bidirectional compatible training in image retrieval. arXiv preprint arXiv:2204.13919, 2022. + +[44] Chang Wang and Sridhar Mahadevan. Manifold alignment using procrastes analysis. In Proceedings of the 25th international conference on Machine learning, pages 1120-1127, 2008. +[45] Mario Lezcano-Casado and David Martinez-Rubio. Cheap orthogonal constraints in neural networks: A simple parametrization of the orthogonal and unitary group. In International Conference on Machine Learning, pages 3794-3803. PMLR, 2019. +[46] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, et al. Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences, 114(13):3521-3526, 2017. +[47] Ronald Kemker, Marc McClure, Angelina Abitino, Tyler Hayes, and Christopher Kanan. Measuring catastrophic forgetting in neural networks. In Proceedings of the AAAI conference on artificial intelligence, volume 32, 2018. +[48] Mehrtash Harandi and Basura Fernando. Generalized backpropagation, etude de cas: Orthogonality. arXiv preprint arXiv:1611.05927, 2016. +[49] Mete Ozay and Takayuki Okatani. Optimization on submanifolds of convolution kernels in cnns. arXiv preprint arXiv:1610.07008, 2016. +[50] Lei Huang, Xianglong Liu, Bo Lang, Adams Yu, Yongliang Wang, and Bo Li. Orthogonal weight normalization: Solution to optimization over multiple dependent stiefel manifolds in deep neural networks. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 32, 2018. +[51] Milton Abramowitz and Irene A Stegun. Handbook of mathematical functions with formulas, graphs, and mathematical tables, volume 55. US Government printing office, 1968. +[52] Sagar Sharma, Simone Sharma, and Anidhya Athaiya. Activation functions in neural networks. Towards Data Sci, 6(12):310-316, 2017. +[53] A Iliev, Nikolay Kyurkchiev, and Svetoslav Markov. On the approximation of the step function by some sigmoid functions. Mathematics and Computers in Simulation, 133:223-234, 2017. +[54] Yonglong Tian, Lijie Fan, Phillip Isola, Huiwen Chang, and Dilip Krishnan. Stablerep: Synthetic images from text-to-image models make strong visual representation learners. Advances in Neural Information Processing Systems, 36, 2024. +[55] Björn Barz and Joachim Denzler. Hierarchy-based image embeddings for semantic image retrieval. In 2019 IEEE winter conference on applications of computer vision (WACV), pages 638-647. IEEE, 2019. +[56] Mikolaj Wieczorek, Barbara Rychalska, and Jacek Dabrowski. On the unreasonable effectiveness of centroids in image retrieval. In Neural Information Processing: 28th International Conference, ICONIP 2021, Sanur, Bali, Indonesia, December 8–12, 2021, Proceedings, Part IV 28, pages 212–223. Springer, 2021. +[57] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115(3):211-252, 2015. +[58] A. Krizhevsky. Learning Multiple Layers of Features from Tiny Images. Technical report, Univ. Toronto, 2009. +[59] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. +[60] Bolei Zhou, Agata Lapedriza, Aditya Khosla, Aude Oliva, and Antonio Torralba. Places: A 10 million image database for scene recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2017. +[61] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021. OpenReview.net, 2021. +[62] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PmLR, 2021. + +[63] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy V Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel HAZIZA, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. Transactions on Machine Learning Research. +[64] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In 2008 Sixth Indian conference on computer vision, graphics & image processing, pages 722-729. IEEE, 2008. +[65] Soravit Changpinyo, Piyush Sharma, Nan Ding, and Radu Soricut. Conceptual 12m: Pushing web-scale image-text pre-training to recognize long-tail visual concepts. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3558-3568, 2021. +[66] Marco Mistretta, Alberto Baldrati, Lorenzo Agnolucci, Marco Bertini, and Andrew D. Bagdanov. Cross the gap: Exposing the intra-modal misalignment in CLIP via modality inversion. In *The Thirteenth International Conference on Learning Representations*, ICLR 2025, Singapore, April 24-28, 2025. Open-Review.net, 2025. +[67] Wenzhuo Liu, Fei Zhu, Longhui Wei, and Qi Tian. C-clip: Multimodal continual learning for vision-language model. In The Thirteenth International Conference on Learning Representations, 2025. +[68] Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. Scaling laws for neural language models. arXiv preprint arXiv:2001.08361, 2020. +[69] Preetum Nakkiran, Gal Kaplun, Yamini Bansal, Tristan Yang, Boaz Barak, and Ilya Sutskever. Deep double descent: Where bigger models and more data hurt. Journal of Statistical Mechanics: Theory and Experiment, 2021(12):124003, 2021. +[70] Gabriele Prato, Simon Guiroy, Ethan Caballero, Irina Rish, and Sarath Chandar. Scaling laws for the out-of-distribution generalization of image classifiers. ICML 2021 Workshop on Uncertainty and Robustness in Deep Learning., 2021. +[71] Ethan Caballero, Kshitij Gupta, Irina Rish, and David Krueger. Broken neural scaling laws. In *The Eleventh International Conference on Learning Representations*, 2023. + +# NeurIPS Paper Checklist + +# 1. Claims + +Question: Do the main claims made in the abstract and introduction accurately reflect the paper's contributions and scope? + +Answer: [Yes] + +Justification: We support our claims on compatibility adaptation through experimental validations presented in Sec. 4 and a theoretically grounded explanation in Sec. 3. + +Guidelines: + +- The answer NA means that the abstract and introduction do not include the claims made in the paper. +- The abstract and/or introduction should clearly state the claims made, including the contributions made in the paper and important assumptions and limitations. A No or NA answer to this question will not be perceived well by the reviewers. +- The claims made should match theoretical and experimental results, and reflect how much the results can be expected to generalize to other settings. +- It is fine to include aspirational goals as motivation as long as it is clear that these goals are not attained by the paper. + +# 2. Limitations + +Question: Does the paper discuss the limitations of the work performed by the authors? + +Answer: [Yes] + +Justification: We discuss the limitations of our work in the Appendix I. + +Guidelines: + +- The answer NA means that the paper has no limitation while the answer No means that the paper has limitations, but those are not discussed in the paper. +- The authors are encouraged to create a separate "Limitations" section in their paper. +- The paper should point out any strong assumptions and how robust the results are to violations of these assumptions (e.g., independence assumptions, noiseless settings, model well-specification, asymptotic approximations only holding locally). The authors should reflect on how these assumptions might be violated in practice and what the implications would be. +- The authors should reflect on the scope of the claims made, e.g., if the approach was only tested on a few datasets or with a few runs. In general, empirical results often depend on implicit assumptions, which should be articulated. +- The authors should reflect on the factors that influence the performance of the approach. For example, a facial recognition algorithm may perform poorly when image resolution is low or images are taken in low lighting. Or a speech-to-text system might not be used reliably to provide closed captions for online lectures because it fails to handle technical jargon. +- The authors should discuss the computational efficiency of the proposed algorithms and how they scale with dataset size. +- If applicable, the authors should discuss possible limitations of their approach to address problems of privacy and fairness. +- While the authors might fear that complete honesty about limitations might be used by reviewers as grounds for rejection, a worse outcome might be that reviewers discover limitations that aren't acknowledged in the paper. The authors should use their best judgment and recognize that individual actions in favor of transparency play an important role in developing norms that preserve the integrity of the community. Reviewers will be specifically instructed to not penalize honesty concerning limitations. + +# 3. Theory assumptions and proofs + +Question: For each theoretical result, does the paper provide the full set of assumptions and a complete (and correct) proof? + +Answer: [NA] + +Justification: Our manuscript does not provide any theoretical result. + +# Guidelines: + +- The answer NA means that the paper does not include theoretical results. +- All the theorems, formulas, and proofs in the paper should be numbered and cross-referenced. +- All assumptions should be clearly stated or referenced in the statement of any theorems. +- The proofs can either appear in the main paper or the supplemental material, but if they appear in the supplemental material, the authors are encouraged to provide a short proof sketch to provide intuition. +- Inversely, any informal proof provided in the core of the paper should be complemented by formal proofs provided in appendix or supplemental material. +- Theorems and Lemmas that the proof relies upon should be properly referenced. + +# 4. Experimental result reproducibility + +Question: Does the paper fully disclose all the information needed to reproduce the main experimental results of the paper to the extent that it affects the main claims and/or conclusions of the paper (regardless of whether the code and data are provided or not)? + +# Answer: [Yes] + +Justification: The details of our method are described in Sec. 3, while Sec. 4 provides the hyperparameter settings used to produce the results reported in all tables. Additional ablation studies on the hyperparameters are presented in Appendix E and Appendix F. + +# Guidelines: + +- The answer NA means that the paper does not include experiments. +- If the paper includes experiments, a No answer to this question will not be perceived well by the reviewers: Making the paper reproducible is important, regardless of whether the code and data are provided or not. +- If the contribution is a dataset and/or model, the authors should describe the steps taken to make their results reproducible or verifiable. +- Depending on the contribution, reproducibility can be accomplished in various ways. For example, if the contribution is a novel architecture, describing the architecture fully might suffice, or if the contribution is a specific model and empirical evaluation, it may be necessary to either make it possible for others to replicate the model with the same dataset, or provide access to the model. In general, releasing code and data is often one good way to accomplish this, but reproducibility can also be provided via detailed instructions for how to replicate the results, access to a hosted model (e.g., in the case of a large language model), releasing of a model checkpoint, or other means that are appropriate to the research performed. +- While NeurIPS does not require releasing code, the conference does require all submissions to provide some reasonable avenue for reproducibility, which may depend on the nature of the contribution. For example +(a) If the contribution is primarily a new algorithm, the paper should make it clear how to reproduce that algorithm. +(b) If the contribution is primarily a new model architecture, the paper should describe the architecture clearly and fully. +(c) If the contribution is a new model (e.g., a large language model), then there should either be a way to access this model for reproducing the results or a way to reproduce the model (e.g., with an open-source dataset or instructions for how to construct the dataset). +(d) We recognize that reproducibility may be tricky in some cases, in which case authors are welcome to describe the particular way they provide for reproducibility. In the case of closed-source models, it may be that access to the model is limited in some way (e.g., to registered users), but it should be possible for other researchers to have some path to reproducing or verifying the results. + +# 5. Open access to data and code + +Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material? + +Answer: [No] + +Justification: The code will be released upon acceptance. + +# Guidelines: + +- The answer NA means that paper does not include experiments requiring code. +- Please see the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details. +- While we encourage the release of code and data, we understand that this might not be possible, so "No" is an acceptable answer. Papers cannot be rejected simply for not including code, unless this is central to the contribution (e.g., for a new open-source benchmark). +- The instructions should contain the exact command and environment needed to run to reproduce the results. See the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details. +- The authors should provide instructions on data access and preparation, including how to access the raw data, preprocessed data, intermediate data, and generated data, etc. +- The authors should provide scripts to reproduce all experimental results for the new proposed method and baselines. If only a subset of experiments are reproducible, they should state which ones are omitted from the script and why. +- At submission time, to preserve anonymity, the authors should release anonymized versions (if applicable). +- Providing as much information as possible in supplemental material (appended to the paper) is recommended, but including URLs to data and code is permitted. + +# 6. Experimental setting/details + +Question: Does the paper specify all the training and test details (e.g., data splits, hyperparameters, how they were chosen, type of optimizer, etc.) necessary to understand the results? + +Answer: [Yes] + +Justification: Sec. 4 provides the hyperparameter settings used to produce the results reported in all tables. Additional ablation studies on the hyperparameters are presented in Appendix E and Appendix F. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The experimental setting should be presented in the core of the paper to a level of detail that is necessary to appreciate the results and make sense of them. +- The full details can be provided either with the code, in appendix, or as supplemental material. + +# 7. Experiment statistical significance + +Question: Does the paper report error bars suitably and correctly defined or other appropriate information about the statistical significance of the experiments? + +Answer: [No] + +Justification: Unfortunately, due to limited GPU credits, we were unable to validate our experiments with different seed values. However, we used the same random initialization values (i.e., seed) for all methods, datasets, and network architectures in our experiments. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The authors should answer "Yes" if the results are accompanied by error bars, confidence intervals, or statistical significance tests, at least for the experiments that support the main claims of the paper. + +- The factors of variability that the error bars are capturing should be clearly stated (for example, train/test split, initialization, random drawing of some parameter, or overall run with given experimental conditions). +- The method for calculating the error bars should be explained (closed form formula, call to a library function, bootstrap, etc.) +- The assumptions made should be given (e.g., Normally distributed errors). +- It should be clear whether the error bar is the standard deviation or the standard error of the mean. +- It is OK to report 1-sigma error bars, but one should state it. The authors should preferably report a 2-sigma error bar than state that they have a $96\%$ CI, if the hypothesis of Normality of errors is not verified. +- For asymmetric distributions, the authors should be careful not to show in tables or figures symmetric error bars that would yield results that are out of range (e.g. negative error rates). +- If error bars are reported in tables or plots, The authors should explain in the text how they were calculated and reference the corresponding figures or tables in the text. + +# 8. Experiments compute resources + +Question: For each experiment, does the paper provide sufficient information on the computer resources (type of compute workers, memory, time of execution) needed to reproduce the experiments? + +Answer: [No] + +Justification: Our approach directly leverages pre-extracted features, requiring minimal GPU usage during training and thereby enabling reproducibility on any contemporary GPU. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The paper should indicate the type of compute workers CPU or GPU, internal cluster, or cloud provider, including relevant memory and storage. +- The paper should provide the amount of compute required for each of the individual experimental runs as well as estimate the total compute. +- The paper should disclose whether the full research project required more compute than the experiments reported in the paper (e.g., preliminary or failed experiments that didn't make it into the paper). + +# 9. Code of ethics + +Question: Does the research conducted in the paper conform, in every respect, with the NeurIPS Code of Ethics https://neurips.cc/public/EthicsGuidelines? + +Answer: [Yes] + +Justification: Yes, the research conducted in the paper conforms to the NeurIPS Code of Ethics. + +Guidelines: + +- The answer NA means that the authors have not reviewed the NeurIPS Code of Ethics. +- If the authors answer No, they should explain the special circumstances that require a deviation from the Code of Ethics. +- The authors should make sure to preserve anonymity (e.g., if there is a special consideration due to laws or regulations in their jurisdiction). + +# 10. Broader impacts + +Question: Does the paper discuss both potential positive societal impacts and negative societal impacts of the work performed? + +Answer: [NA] + +Justification: Our work seems not to have societal impact to date. Anyway, we discussed the positive impact that it may have in real-world retrieval systems in the introduction section (Sec. 1). + +# Guidelines: + +- The answer NA means that there is no societal impact of the work performed. +- If the authors answer NA or No, they should explain why their work has no societal impact or why the paper does not address societal impact. +- Examples of negative societal impacts include potential malicious or unintended uses (e.g., disinformation, generating fake profiles, surveillance), fairness considerations (e.g., deployment of technologies that could make decisions that unfairly impact specific groups), privacy considerations, and security considerations. +- The conference expects that many papers will be foundational research and not tied to particular applications, let alone deployments. However, if there is a direct path to any negative applications, the authors should point it out. For example, it is legitimate to point out that an improvement in the quality of generative models could be used to generate deepfakes for disinformation. On the other hand, it is not needed to point out that a generic algorithm for optimizing neural networks could enable people to train models that generate Deepfakes faster. +- The authors should consider possible harms that could arise when the technology is being used as intended and functioning correctly, harms that could arise when the technology is being used as intended but gives incorrect results, and harms following from (intentional or unintentional) misuse of the technology. +- If there are negative societal impacts, the authors could also discuss possible mitigation strategies (e.g., gated release of models, providing defenses in addition to attacks, mechanisms for monitoring misuse, mechanisms to monitor how a system learns from feedback over time, improving the efficiency and accessibility of ML). + +# 11. Safeguards + +Question: Does the paper describe safeguards that have been put in place for responsible release of data or models that have a high risk for misuse (e.g., pretrained language models, image generators, or scraped datasets)? + +Answer: [NA] + +Justification: The paper does not to pose such risks. + +# Guidelines: + +- The answer NA means that the paper poses no such risks. +- Released models that have a high risk for misuse or dual-use should be released with necessary safeguards to allow for controlled use of the model, for example by requiring that users adhere to usage guidelines or restrictions to access the model or implementing safety filters. +- Datasets that have been scraped from the Internet could pose safety risks. The authors should describe how they avoided releasing unsafe images. +- We recognize that providing effective safeguards is challenging, and many papers do not require this, but we encourage authors to take this into account and make a best faith effort. + +# 12. Licenses for existing assets + +Question: Are the creators or original owners of assets (e.g., code, data, models), used in the paper, properly credited and are the license and terms of use explicitly mentioned and properly respected? + +Answer: [Yes] + +Justification: All the code, data, and models that we used has been properly credited and their license and terms of use respected. + +# Guidelines: + +- The answer NA means that the paper does not use existing assets. +- The authors should cite the original paper that produced the code package or dataset. +- The authors should state which version of the asset is used and, if possible, include a URL. +- The name of the license (e.g., CC-BY 4.0) should be included for each asset. + +- For scraped data from a particular source (e.g., website), the copyright and terms of service of that source should be provided. +- If assets are released, the license, copyright information, and terms of use in the package should be provided. For popular datasets, paperswithcode.com/datasets has curated licenses for some datasets. Their licensing guide can help determine the license of a dataset. +- For existing datasets that are re-packaged, both the original license and the license of the derived asset (if it has changed) should be provided. +- If this information is not available online, the authors are encouraged to reach out to the asset's creators. + +# 13. New assets + +Question: Are new assets introduced in the paper well documented and is the documentation provided alongside the assets? + +Answer: [NA] + +Justification: The present work does not release any new asset. + +Guidelines: + +- The answer NA means that the paper does not release new assets. +- Researchers should communicate the details of the dataset/code/model as part of their submissions via structured templates. This includes details about training, license, limitations, etc. +- The paper should discuss whether and how consent was obtained from people whose asset is used. +- At submission time, remember to anonymize your assets (if applicable). You can either create an anonymized URL or include an anonymized zip file. + +# 14. Crowdsourcing and research with human subjects + +Question: For crowdsourcing experiments and research with human subjects, does the paper include the full text of instructions given to participants and screenshots, if applicable, as well as details about compensation (if any)? + +Answer: [NA] + +Justification: The paper does not involve crowdsourcing nor research with human subjects. + +Guidelines: + +- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. +- Including this information in the supplemental material is fine, but if the main contribution of the paper involves human subjects, then as much detail as possible should be included in the main paper. +- According to the NeurIPS Code of Ethics, workers involved in data collection, curation, or other labor should be paid at least the minimum wage in the country of the data collector. + +# 15. Institutional review board (IRB) approvals or equivalent for research with human subjects + +Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals (or an equivalent approval/review based on the requirements of your country or institution) were obtained? + +Answer: [NA] + +Justification: The paper does not involve crowdsourcing nor research with human subjects. + +Guidelines: + +- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. + +- We recognize that the procedures for this may vary significantly between institutions and locations, and we expect authors to adhere to the NeurIPS Code of Ethics and the guidelines for their institution. +- For initial submissions, do not include any information that would break anonymity (if applicable), such as the institution conducting the review. + +# 16. Declaration of LLM usage + +Question: Does the paper describe the usage of LLMs if it is an important, original, or non-standard component of the core methods in this research? Note that if the LLM is used only for writing, editing, or formatting purposes and does not impact the core methodology, scientific rigorousness, or originality of the research, declaration is not required. + +Answer: [NA] + +Justification: LLMs were used only for the purpose of writing, editing, or formatting purposes and does not impact the core methodology. + +Guidelines: + +- The answer NA means that the core method development in this research does not involve LLMs as any important, original, or non-standard components. +- Please refer to our LLM policy (https://neurips.cc/Conferences/2025/LLM) for what should or should not be described. + +Table 5: Compatibility evaluation on Places365 under the Extending Classes setting. We use two independently trained ResNet-50 models: $\phi_{\mathrm{old}}$ trained on the first 205 classes, and $\phi_{\mathrm{new}}$ trained on all classes of Places365. + +
MethodQuery/GalleryCMC-Top1mAP
Ind. Train.φold/φold33.8615.76
φnew/φold0.210.33
φnew/φnew37.3719.11
FCT [25]F(φold)/φold0.210.33
F(φold)/F(φold)36.4319.02
φnew/F(φold)37.0418.99
FastFill [22]F(φold)/φold0.210.33
F(φold)/F(φold)39.7123.98
φnew/F(φold)38.4219.94
OursF(φold)/φold38.6521.88
F(φold)/F(φold)39.9626.19
B⊥(φnew)/F(φold)38.5021.77
B⊥(φnew)/φold35.4717.98
B⊥(φnew)/B⊥(φnew)37.3719.11
+ +# A Extending Classes Setting on Places365 + +To validate our approach further, we evaluate it using a model trained on a dataset different from ImageNet1K. Specifically, we use a ResNet-50 pretrained on Places205 (from ViSSL) as the old model, and a ResNet-50 pretrained on Places365 (from CSAILVision) as the new model. Tab. 5 summarizes the performance of each method using the evaluation metrics defined in Sec.4.2. The results demonstrate that the new model $\phi_{\mathrm{new}}$ is not inherently compatible with the old model, $\phi_{\mathrm{old}}$ . Moreover, the adaptation $F(\phi_{\mathrm{old}})$ provided by FCT underperforms when compared to the new model alone. In contrast, methods that promote better clustering, such as FastFill and our proposed approach, achieve even higher performance than the standalone new model. This improvement arises from leveraging information from both the old and new models, effectively implementing a form of knowledge distillation during the learning of the forward adapter. Unlike the baselines, our method aligns all adapted representations within a unified representation space, thereby consistently maintaining compatibility with the old model. + +# B Additional Architecture for Independently Pretrained Models Setting + +We conduct additional experiments using a DenseNet-121 as the old model, $\phi_{\mathrm{old}}$ , and an EfficientNet-B3 as the new model, $\phi_{\mathrm{new}}$ , both pretrained on ImageNet1K and obtained from the PyTorch Hub. The results of these experiments on the ImageNet1K dataset are presented in Tab. 6. Our approach achieves the best performance across all metrics, outperforming the baselines in both cross-model and same-model retrieval scenarios. + +# C Additional Experiments with DINOV2 and CLIP as Independently Pretrained Models + +To investigate update scenarios involving data distribution or objective shifts, we conduct additional experiments using a ResNet-18 pretrained on ImageNet1K as the old model, and both a CLIP [62] pretrained on CC12M [65] dataset and a DINOv2 [63] (vit_small_batch14_dinov2) as the new models. To train both the forward and backward transformations, the ImageNet1K dataset and the same hyperparameters of Tab. 1b are used. This setup represents a considerable shift in both data distribution and model objective relative to the new models. Notably, FastFill cannot be applied in this context, as both CLIP and DINOv2 lack classifiers. In Tab. 7a, we report the results obtained using DINOv2 as the new, independently trained model. Our approach achieves better results than FCT, further validating its practical applicability to real-world problems. + +Table 6: Compatibility results on ImageNet1K under the Independently Pretrained Models setting. We use two independently trained models: DenseNet-121 as the old model, $\phi_{\mathrm{old}}$ , and an EfficientNet-B3 as the new model, $\phi_{\mathrm{new}}$ , both pretrained on ImageNet1K and obtained from the PyTorch Hub. + +
MethodQuery/GalleryCMC-Top1mAP
Ind. Train.φold/φold62.0232.95
φnew/φold0.110.16
φnew/φnew71.6054.90
FCT [25]F(φold)/φold0.110.16
F(φold)/F(φold)68.1653.22
φnew/F(φold)70.6454.63
FastFill [22]F(φold)/φold0.110.16
F(φold)/F(φold)67.7657.22
φnew/F(φold)69.4757.43
OursF(φold)/φold69.2550.20
F(φold)/F(φold)69.2957.36
B⊥(φnew)/F(φold)71.3357.50
B⊥(φnew)/φold67.2344.34
B⊥(φnew)/B⊥(φnew)71.6054.90
+ +Table 7: Compatibility evaluation involving data distribution or objective shifts: (a) ResNet-18 as old model and DINOv2 [63] (vit_small_batch14_dinov2) as the new models; (b) ResNet-18 as old model and CLIP [62] pretrained on CC12M [65] as the new model. For each case, we report CMC-Top1 and mAP metrics. + +(a) DINOv2 [63]. A shift in the objective function is present between the old and new models. + +
MethodQuery/GalleryCMC-Top1mAP
Ind. Train.φold/φold55.6226.91
φnew/φold0.040.17
φnew/φnew71.9244.07
FCT [25]F(φold)/φold0.040.17
F(φold)/F(φold)59.3337.53
φnew/F(φold)67.9741.07
OursF(φold)/φold54.8232.14
F(φold)/F(φold)61.3041.95
B⊥(φnew)/F(φold)68.7443.78
B⊥(φnew)/φold58.7331.50
B⊥(φnew)/B⊥(φnew)71.9244.07
+ +(b) CLIP [62]. A shift in the objective function and the data distribution is present between the old and new models. + +
MethodQuery/GalleryCMC-Top1mAP
Ind. Train.φold/φold55.6226.91
φnew/φold0.040.17
φnew/φnew44.2916.15
FCT [25]F(φold)/φold0.040.17
F(φold)/F(φold)42.5816.93
φnew/F(φold)42.9616.88
OursF(φold)/φold61.1341.22
F(φold)/F(φold)57.6941.08
B⊥(φnew)/F(φold)44.9329.26
B⊥(φnew)/φold30.0216.68
B⊥(φnew)/B⊥(φnew)44.2916.15
+ +Instead, in Tab. 7b, we report the results obtained using CLIP pretrained on CC12M as the new, independently trained model. In this scenario, the pretrained CLIP model exhibits lower retrieval performance on ImageNet1K compared to ResNet-18. This is a well-known limitation of multimodal training, where intra-modal misalignment can negatively impact the quality of single-modality representations [66]. Specifically, CLIP models are optimized for cross-modal retrieval rather than single-modality retrieval tasks, in contrast to DINOv2 or ResNet-18, which are trained exclusively on a single modality. This reduction in performance of the new model relative to the old one causes FCT to degrade the overall retrieval capacity of the system, failing to achieve compatibility, as it attempts to transform the higher-quality representations of the old model into the lower-performing representations of the new model. In contrast, our method introduces an additional loss that encourages both intra-class clustering and inter-model alignment of feature representations on the specific training dataset. As a result, the forward transformation, due to its greater flexibility, improves the performance of the old model's representations. Even in this challenging scenario, our approach outperforms FCT, further validating the robustness of our method. + +Table 8: Compatibility results on Places365 and Flowers102 for two models pretrained on ImageNet1K and adapted to downstream tasks: $\phi_{\mathrm{old}}$ , a ResNet-18, and $\phi_{\mathrm{new}}$ , a ViT-L-16, using a backward adapter $B_{\lambda}$ with $\lambda = 12$ . The ZS column indicates the CMC-Top1 performance increase on ImageNet1K, with values in parentheses showing the increment compared to the newly independently trained model. + +
MethodQuery/GalleryPlaces365Flowers102
CMC-Top1ZSCMC-Top1ZS
Ind. Train.φold/φold22.4184.35
φnew/φold0.201.20
φnew/φnew35.1599.39
FCT [25]F(φold)/φold0.201.20
F(φold)/F(φold)28.1786.71
φnew/F(φold)32.1299.07
FastFill [22]F(φold)/φold0.201.20
F(φold)/F(φold)26.3853.78
φnew/F(φold)33.0411.12
OursF(φold)/φold28.8483.36
F(φold)/F(φold)29.8089.90
Bλ(φnew)/F(φold)33.2799.41
Bλ(φnew)/φold29.9498.17
Bλ(φnew)/Bλ(φnew)36.38 (+1.23)+0.3899.54 (+0.15)+0.01
+ +# D Additional Datasets for Independently Pretrained Models adapted on Downstream Task setting + +We further extend our analysis of the Independently Pretrained Models Adapted on Downstream Task setting by including two additional datasets: the larger Places365 and the fine-grained Flowers102. These additions allow us to evaluate our method's effectiveness in more challenging scenarios. The results are reported in Tab. 8. In these experiments, the old model is a ResNet-18 and the new model is a ViT-L-16, both pretrained on ImageNet-1K. We employ an affine adapter with $\lambda = 12$ . On both additional datasets, our approach consistently outperforms the baseline methods. The proposed $\lambda$ -Orthogonality regularization not only improves retrieval performance on the downstream tasks but also encourages the adapted new model representation, $B_{\lambda}(\phi_{\mathrm{new}})$ , to remain consistent with its original form. As a result, retrieval performance on ImageNet1K is preserved. + +# E Ablation on the hyperparameter $\lambda$ + +In our experiments, we select $\lambda$ to maximize adaptability to downstream tasks while preserving the pre-trained model's performance on its original training dataset, ImageNet1K. To illustrate the impact of our approach, Tab. 9 reports the CMC-Top1 scores obtained by applying our proposed $\lambda$ -orthogonal regularizer to the new pre-trained model. The results, also reported in Fig. 5, indicate that increasing $\lambda$ enhances the performance of the new model's representations on the downstream task. + +However, this improvement comes at the expense of reduced performance on the original dataset, as evidenced by a decrease in zero-shot (ZS) scores, particularly pronounced in the absence of regularization $(\lambda = \infty)$ . Empirically, we find that setting $\lambda = 12$ yields the best trade-off across all metrics. [36] optimize a soft orthogonality constraint, equal to case where + +![](images/db53df6f98b06506208e24cba2dba8dc7d2c33864059b4353e8ec4a247cc6870.jpg) +Figure 5: Ablation on our $\lambda$ -orthogonal regularization on CUB dataset. Displayed are the compatibility metrics on CUB and the zero-shot (ZS) improvement on ImageNet1K at different value of $\lambda$ . Results correspond to those in Tab. 9. + +$\lambda = 0$ . However, this formulation does not lead to performance improvements and is outperformed by + +Table 9: Ablation over orthogonal regularization strength $\lambda$ on CUB dataset. Compatibility metrics on the target task and zero-shot (ZS) CMC-Top1 gain on ImageNet1K. Parentheses show the increment in CMC-Top1 over the independently trained new model on CUB dataset. + +
λF(φold)/F(φold)Bλ(φnew)/F(φold)Bλ(φnew)/Bλ(φnew)ZS
⊥ (strict orth.)57.5266.8971.78 (+0.000)+0.000
057.4966.7971.54 (-0.241)-0.001
357.5266.7272.00 (+0.224)+0.008
658.0968.7773.07 (+1.294)+0.028
1259.9270.7275.44 (+3.659)+0.028
1659.6870.2176.40 (+4.625)+0.062
2260.2069.5077.89 (+6.109)-0.318
3659.3263.3478.77 (+6.990)-3.008
∞ (no reg.)59.2662.8478.89 (+7.110)-3.526
+ +Table 10: Comparison of orthogonal regularization methods with different weight scales $w$ . Compatibility metrics on the downstream task CUB200 and zero-shot (ZS) CMC-Top1 gain on ImageNet1K. Parentheses show the increment in CMC-Top1 over the independently trained new model. The last column reports the final value of $\| W^{\top}W - I\|_{F}$ . + +
wMethodF(φold)/F(φold)Bλ(φnew)/F(φold)Bλ(φnew)/Bλ(φnew)ZS||W^T W - I ||_F
1SO57.4866.7971.54 (-0.241)-0.0010.09
1SRIP57.3866.5771.66 (-0.120)-0.0010.08
1Ours (λ = 12)59.9270.7275.44 (+3.659)+0.02812.05
10^-1SO59.1169.5674.88 (+3.106)+0.0229.50
10^-1SRIP58.8863.5878.77 (+6.990)-1.46729.55
10^-1Ours (λ = 12)59.9370.7075.20 (+3.419)+0.07612.12
10^-2SO59.0663.5479.06 (+7.283)-1.34429.27
10^-2SRIP59.2363.4278.73 (+6.955)-3.07735.42
10^-2Ours (λ = 12)59.0663.5479.06 (+7.283)-1.34429.27
10^-3SO58.7162.9178.78 (+7.007)-3.16235.54
10^-3SRIP58.8363.1878.92 (+7.145)-3.45738.63
10^-3Ours (λ = 12)58.7162.9178.78 (+7.007)-3.16235.54
+ +the use of a strictly orthogonal transformation. As discussed in Sec. 3.3, imposing strict orthogonality may hinder the model's ability to incorporate task-specific information. In contrast, our approach relaxes this constraint by introducing a tunable hyperparameter $\lambda$ that controls the deviation of the Gram matrix from the identity, allowing greater flexibility while preserving representational consistency. + +To further validate our approach we also study the effect of a scalar weight $w$ to the loss contributions of our $\lambda$ -orthogonal regularization compared with two different orthogonal regularizations: Soft Orthogonality (SO)[36]—witch correspond to the spacial case of $\lambda = 0$ in our approach—and Spectral Restricted Isometry Property (SRIP)[36]. We test the regularizers across different values of scalar weight: $w = 1$ , $w = 10^{-1}$ , $w = 10^{-2}$ , and $w = 10^{-3}$ . Additionally, we include a column reporting the exact value of $\| W^{\top}W - I\|_{F}$ reached by the backward transformation $B_{\lambda}$ at the end of training, to indicate the deviation from strict orthogonality. + +As shown in the Tab. 10, for both SRIP and SO, the final value of $\| W^{\top}W - I\|_{F}$ is governed by the optimization process and the chosen scalar weight $w$ . Unlike our $\lambda$ -orthogonal regularization, these approaches do not provide direct control over $\| W^{\top}W - I\|_{F}$ ; a smaller contribution of the regularizer to the total loss results in a diminished regularization effect on the backward transformation $B_{\lambda}$ . When the scalar weight $w$ of the regularizer is reduced, the optimization process is unable to fully minimize the regularization term, particularly because competing loss components (such as MSE and the contrastive loss $L_{C}$ ) may favor a non-orthogonal transformation. For instance, when $w = 10^{-3}$ and $w = 10^{-2}$ , the results obtained with SO, SRIP, and our $\lambda$ -orthogonal regularization are comparable to those observed in the case of $\lambda = \infty$ (see Tab. 9), where the orthogonality constraint is entirely ignored. This occurs because, at such a small value of $w$ , the contribution of the regularizer becomes negligible during optimization. To avoid this issue, in our method we set $w = 1$ for the $\lambda$ -orthogonal regularization, thereby ensuring that the regularization term is effectively + +incorporated into the optimization process during the training of the backward transformation. This ensures that the regularization term achieves the target threshold $\lambda$ , enabling precise control over the stability-plasticity trade-off in the backward transformation and leads to higher representation compatibility on the downstream task. As highlighted by the bold entries in the Tab. 10, our method produces stable results (minor fluctuations are attributable to stochastic optimization) for $w = 1$ and $w = 10^{-1}$ in contrast to SO and SRIP. Conversely, when $w$ is very low ( $10^{-2}$ or $10^{-3}$ ), the regularizer cannot be fully optimized, and our method behaves similarly to SO regularization, as our introduced constrains ( $\|W^{\top}W - I\|_{F} \geq \lambda$ ) influences the minimum of the objective, which is never reached in practice. In contrast, due to its approximate formulation and greater complexity relative to SO, SRIP exhibits an even weaker regularization effect when $w$ is low. + +# F Detailed Analysis of Loss Term Contributions + +In this section, we analyze the contribution of each term to the final loss (Eq. 3.5) optimized during training. Tab. 11 presents the results obtained when the adaptation dataset matches the dataset used to train the models from which the features were extracted, namely ImageNet1K. In this scenario, a strict orthogonal transformation $B_{\perp}$ is employed for backward-compatibility. We observe that when used independently, $\mathcal{L}_F$ ensures compatibility with the representations of the new model but significantly fails to achieve backward compatibility. This behavior highlights a pronounced forward bias inherent to $\mathcal{L}_F$ . The backward alignment loss $\mathcal{L}_B$ alone promotes backward compatibility but degrades forward-adapted representation performance. The contrastive loss $\mathcal{L}_C$ alone significantly improves inter-model alignment and intra-class clustering, supporting both backward and forward compatibility. The combination $\mathcal{L}_F + \mathcal{L}_B + \mathcal{L}_C$ achieves the highest overall performance across compatibility scenarios, underscoring the importance of each loss component in maintaining balance between forward and backward transformation learning. + +Tab. 12 shows the impact of these loss terms in a downstream task setting (CUB dataset), where $\phi_{old}$ is ResNet-18 and $\phi_{new}$ is ViT-L-16, using $\lambda$ -Orthogonality with $\lambda = 12$ . Similar to Tab. 11, excluding the backward loss $\mathcal{L}_B$ still yields good forward compatibility but significantly reduces backward compatibility performance. Excluding the contrastive loss $\mathcal{L}_C$ substantially decreases the adaptation to the downstream task leading to lower $B_{\lambda}(\phi_{\mathrm{new}}) / B_{\lambda}(\phi_{\mathrm{new}})$ values. Using all loss terms $\mathcal{L}_F + \mathcal{L}_B + \mathcal{L}_C$ consistently achieves the best or near-best results in forward and backward compatibility, demonstrating the complementary nature of these terms. + +These analyses underline that each loss term contributes uniquely and significantly to achieving comprehensive and model compatibility across various tasks. + +Table 11: CMC-Top1 (\%) on ImageNet1K for different loss combinations ( $\checkmark =$ included, $\times =$ excluded). The setting is the same of Tab. 1b, where the first model, $\phi_{\mathrm{old}}$ , is a ResNet-18, whereas the second, $\phi_{\mathrm{new}}$ , is a ViT-L-16. + +
LossesQuery/Gallery (CMC-Top1 %)
\( \mathcal{L}_F \)\( \mathcal{L}_B \)\( \mathcal{L}_C \)\( F(\phi_{old})/\phi_{old} \)\( F(\phi_{old})/F(\phi_{old}) \)\( B_{\perp}(\phi_{new})/\phi_{old} \)\( B_{\perp}(\phi_{new})/F(\phi_{old}) \)\( B_{\perp}(\phi_{new})/B_{\perp}(\phi_{new}) \)
✓××0.0459.090.0472.2776.63
×✓×0.0449.3462.750.0476.63
××✓61.2458.6364.9760.8376.63
✓✓×54.1859.2962.7772.4676.63
✓×✓61.2560.4365.1373.4476.63
×✓✓60.8559.0965.4257.9076.63
✓✓✓60.8361.1065.5473.5376.63
+ +# G Distance metric for Partial Backfilling Ordering + +Our proposed partial backfilling strategy is guided by a distance metric $d$ , which measures the dissimilarity between each embedding vector $F(\mathbf{h}^k)$ and its corresponding class mean $\mu_c$ . This section investigates the impact of different distance metrics on determining an effective ordering for backfilling images in the gallery set. We compare two distance metrics—Mean Squared Error (MSE) and Cosine Distance—for ranking images during partial backfilling. The performance of each metric is evaluated under two distinct experimental conditions: the Extending Classes setting (Tab. 13) and + +Table 12: CMC-Top1 (\%) on CUB for different loss combinations ( $\checkmark =$ included, $\times =$ excluded). The setting is the same of Tab. 2, where the first model, $\phi_{\mathrm{old}}$ , is a ResNet-18, whereas the second, $\phi_{\mathrm{new}}$ , is a ViT-L-16. A backward adapter $B_{\lambda}$ with $\lambda = 12$ is used to adapt the improved model on the downstream task. + +
LossesQuery/Gallery (CMC-Top1 %)
LFLBLCF(φold)/φoldF(φold)/F(φold)Bλ(φnew)/φoldBλ(φnew)/F(φold)Bλ(φnew)/Bλ(φnew)
✓××0.051.720.063.8272.14
×✓×0.035.2745.800.071.91
✓✓×37.1547.5646.5660.7069.76
××✓52.7959.1458.3866.4673.36
✓×✓50.4359.8858.5770.1374.86
×✓✓53.2758.6660.4559.4473.12
✓✓✓51.1259.9260.6470.7275.44
+ +![](images/690a9c77e0686c7843f7a016b24002a7c5cba88debc66baa54194c64439b7793.jpg) + +![](images/191b4fccb80afb8da77da8de22993bc4216d188b2e2d244ba95bbff02b3600d0.jpg) + +Table 13: Extended Classes setting + +
MethodM
CMC-Top1mAP
MSE61.2036.46
Cosine Distance61.6837.10
+ +![](images/039865476ca0a1502057839b3ecd0ecc1dcf201b1cb3a1b4877404d9159f588c.jpg) +Figure 6: Different distance metric ablation for our partial backfilling strategy. Results for the Extending Classes setting (top Figures) of Tab. 1a, and Independently Pretrained Models setting (bottom Figures) of Tab. 1b. We use features from the new backward-adapted model $B_{\perp}(\phi_{\mathrm{new}})$ for the query set. For the gallery set, we begin with forward-adapted old features $F(\phi_{\mathrm{old}})$ and incrementally replace them with new features. + +![](images/9b5d854d7c37bc929cbea5c2ea418c1aa6ca0a54f45e4ffa0e9676352659e81a.jpg) + +Table 14: Independently Pre-trained Models setting + +
MethodM
CMC-Top1mAP
MSE76.5957.72
Cosine Distance76.4958.18
+ +the Independently Pretrained Models setting (Tab. 14). MSE computes the Euclidean distance between feature vectors, capturing both angular and magnitude discrepancies. As shown in Tab. 13 and Tab. 14, MSE generally yields robust performance, particularly in terms of CMC-Top1. In contrast, Cosine Distance measures the angular distance between normalized feature vectors, emphasizing directional similarity while ignoring magnitude. The results indicate that Cosine Distance achieves slightly better performance in terms of mAP and provides comparable CMC-Top1 scores relative to MSE. + +# H Method Complexity and Broader Applicability + +Method Complexity. Our approach requires training only two matrices, resulting in a small number of parameters to optimize. Because our method operates solely on the extracted embeddings, it does not require any knowledge of the underlying models and is therefore applicable across different objectives (see Appendix C), architectures, and types of learned representations. + +In contrast to previous methods, which either focus solely on alignment loss without any representation clustering loss (e.g., FCT [25]), or require specific architectural components of the pretrained models (e.g., FastFill [22], which requires access to the classifier of the new model), our approach addresses these limitations. Additionally, while existing baselines provide only forward adaptation, our method is designed to achieve both forward and backward compatibility, thereby addressing practical needs that prior works do not meet. For instance: + +- $B_{\perp}(\phi_{\mathrm{new}}) / F(\phi_{\mathrm{old}})$ yields higher retrieval values compared to the baselines. +- $B_{\perp}(\phi_{\mathrm{new}}) / \phi_{\mathrm{old}}$ can be achieved exclusively by our method. From a practical standpoint, this allows compatibility to be established even before all gallery items are forward-adapted using $F$ . +- Since our approach provides a unified representation space, even when the gallery is in a hybrid form (i.e., with some elements already adapted by $F$ and others not), using $B_{\perp}(\phi_{\mathrm{new}})$ still ensures compatibility. This can not be achieved neither by FCT [25] nor FastFill [22]. + +The contrastive loss defined in Eq. 8 relies on the availability of class labels to encourage embeddings from the same class to cluster together while pushing apart embeddings from different classes. In scenarios where class labels are not available, Eq. 8 naturally reduces to an unsupervised contrastive loss, similar to the objective used for training CLIP models [62]. In this unsupervised setting, we contrast pairs of representations originating from different models, and clustering—since it cannot be enforced directly—becomes a byproduct resulting from embedding similarity. Consequently, our approach is flexible and can be applied in both supervised and unsupervised training scenarios, depending on the availability of labels for the downstream task. + +Broader Applicability. As it is demonstrated in [36], soft orthogonalization has been applied to regularize all the weights of a CNN during training, and could benefit from the increased plasticity offered by our proposed $\lambda$ -orthogonal regularization. While retrieval is the standard scenario for evaluating compatibility [15], our approach is broadly applicable to any task that requires representation adaptation, as it focuses on model alignment and clustering of learned representations. As demonstrated in our downstream task adaptation experiments (see Sec. 4.4), our regularization approach yields improved performance compared to a strict orthogonal constraint, making it a valuable approach in domain adaptation scenarios as well. Furthermore, enforcing geometrical consistency while allowing adaptability has recently been investigated in the context of continual learning for multimodal training [67]. However, the authors of [67] promote this property indirectly through a knowledge consolidation loss, rather than by directly applying a regularization constraint. This highlights both possible future research and the potential applicability of our $\lambda$ -orthogonal regularization across various areas of representation learning. + +# I Limitations + +Our approach relies on the assumption that the new model's embedding space is more expressive (e.g., higher retrieval accuracy, stronger clustering) than that of the old model. If the updated model is not comparable or lower quality, due, for instance, to domain mismatch, insufficient training data, or architectural regressions, then both the forward and backward adapters may fail to improve performance or could even degrade compatibility. In many practical systems, this assumption is justified by scaling laws [68, 69, 70, 71] (i.e., larger models and more data generally yield better feature representations). For downstream tasks adaptation, while our $\lambda$ -orthogonal regularized adapter shows strong performance and compatibility across various retrieval tasks, a manual tuning of the orthogonality threshold ( $\lambda$ ) is needed. The trade-off between preserving the original model's geometry and allowing sufficient plasticity to adapt to new data hinges critically on the choice of $\lambda$ . In practice, this hyperparameter could be selected via cross-validation or a small hyperparameter search on a held-out portion of the downstream dataset. Although we found that $\lambda = 12$ provides a good balance in our experiments (Appendix E), different downstream domains (e.g., fine-grained vs. coarse categories) and adapted representations may require different tuning of $\lambda$ to achieve optimal performance. Automating or self-tuning this parameter remains an open challenge. \ No newline at end of file diff --git a/NeurIPS/2025/$_boldsymbol{_lambda}$-Orthogonality Regularization for Compatible Representation Learning/images.zip b/NeurIPS/2025/$_boldsymbol{_lambda}$-Orthogonality Regularization for Compatible Representation Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..44bb2ee3303d69a19961574719f44c5121a7a975 --- /dev/null +++ b/NeurIPS/2025/$_boldsymbol{_lambda}$-Orthogonality Regularization for Compatible Representation Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3b58fc1c19561fabd2f762982c965096450c4e6a4ecde227372f8ba797c24b64 +size 824468 diff --git a/NeurIPS/2025/$_boldsymbol{_lambda}$-Orthogonality Regularization for Compatible Representation Learning/layout.json b/NeurIPS/2025/$_boldsymbol{_lambda}$-Orthogonality Regularization for Compatible Representation Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b4eb21872fa40205bc533d06a7a3ca19f4558f00 --- /dev/null +++ b/NeurIPS/2025/$_boldsymbol{_lambda}$-Orthogonality Regularization for Compatible Representation Learning/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8f1ce93aa69462dd0b482e08b7616cec6354fe6bd0e0b50f00b005ada3d226a5 +size 976727 diff --git a/NeurIPS/2025/$_epsilon$-Seg_ Sparsely Supervised Semantic Segmentation of Microscopy Data/8833d9e0-1854-4f4d-bdfe-51f3284cf8e9_content_list.json b/NeurIPS/2025/$_epsilon$-Seg_ Sparsely Supervised Semantic Segmentation of Microscopy Data/8833d9e0-1854-4f4d-bdfe-51f3284cf8e9_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..02219d5aa53a1cbc0265c6cf240efb69964a7524 --- /dev/null +++ b/NeurIPS/2025/$_epsilon$-Seg_ Sparsely Supervised Semantic Segmentation of Microscopy Data/8833d9e0-1854-4f4d-bdfe-51f3284cf8e9_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:54fe8cc91d589f4a9b5863fd5046277cd0d0635dc2342cc37aa8afe64f8930ec +size 145745 diff --git a/NeurIPS/2025/$_epsilon$-Seg_ Sparsely Supervised Semantic Segmentation of Microscopy Data/8833d9e0-1854-4f4d-bdfe-51f3284cf8e9_model.json b/NeurIPS/2025/$_epsilon$-Seg_ Sparsely Supervised Semantic Segmentation of Microscopy Data/8833d9e0-1854-4f4d-bdfe-51f3284cf8e9_model.json new file mode 100644 index 0000000000000000000000000000000000000000..1301b2c4149ddaf85c4f9b5d60c002a746401a5e --- /dev/null +++ b/NeurIPS/2025/$_epsilon$-Seg_ Sparsely Supervised Semantic Segmentation of Microscopy Data/8833d9e0-1854-4f4d-bdfe-51f3284cf8e9_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c39a43aa11050c425dda0e12d9e8feadc8ec7ed620097d300c531fffaf340a99 +size 192630 diff --git a/NeurIPS/2025/$_epsilon$-Seg_ Sparsely Supervised Semantic Segmentation of Microscopy Data/8833d9e0-1854-4f4d-bdfe-51f3284cf8e9_origin.pdf b/NeurIPS/2025/$_epsilon$-Seg_ Sparsely Supervised Semantic Segmentation of Microscopy Data/8833d9e0-1854-4f4d-bdfe-51f3284cf8e9_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..2aaf667617196d52a2e312f4aeece8b668ecbf3b --- /dev/null +++ b/NeurIPS/2025/$_epsilon$-Seg_ Sparsely Supervised Semantic Segmentation of Microscopy Data/8833d9e0-1854-4f4d-bdfe-51f3284cf8e9_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:548e5e023cea0b6b29cc2aaa42fd170b157dd35fb4506d74de4edd0f88fc6fe0 +size 6366888 diff --git a/NeurIPS/2025/$_epsilon$-Seg_ Sparsely Supervised Semantic Segmentation of Microscopy Data/full.md b/NeurIPS/2025/$_epsilon$-Seg_ Sparsely Supervised Semantic Segmentation of Microscopy Data/full.md new file mode 100644 index 0000000000000000000000000000000000000000..2b6a05bc8951fa0f4b4719f7231e13d7af82bd56 --- /dev/null +++ b/NeurIPS/2025/$_epsilon$-Seg_ Sparsely Supervised Semantic Segmentation of Microscopy Data/full.md @@ -0,0 +1,719 @@ +# $\epsilon$ -Seg: Sparsely Supervised Semantic Segmentation of Microscopy Data + +Sheida Rahnamai Kordasiabi $^{1,2}$ , Damian Dalle Nogare $^{1}$ , Florian Jug $^{1}$ + +$^{1}$ Human Technopole, Milan, Italy + +$^{2}$ Technical University of Dresden, Germany + +# Abstract + +Semantic segmentation of electron microscopy (EM) images of biological samples remains a challenge in the life sciences. EM data captures details of biological structures, sometimes with such complexity that even human observers can find it overwhelming. We introduce $\epsilon$ -Seg, a method based on hierarchical variational autoencoders (HVAEs), employing center-region masking, sparse label contrastive learning (CL), a Gaussian mixture model (GMM) prior, and clustering-free label prediction. Center-region masking and the inpainting loss encourage the model to learn robust and representative embeddings to distinguish the desired classes, even if training labels are sparse (0.05% of the total image data or less). For optimal performance, we employ CL and a GMM prior to shape the latent space of the HVAE such that encoded input patches tend to cluster w.r.t. the semantic classes we wish to distinguish. Finally, instead of clustering latent embeddings for semantic segmentation, we propose a MLP semantic segmentation head to directly predict class labels from latent embeddings. We show empirical results of $\epsilon$ -Seg and baseline methods on 2 dense EM datasets of biological tissues and demonstrate the applicability of our method also on fluorescence microscopy data. Our results show that $\epsilon$ -Seg is capable of achieving competitive sparsely-supervised segmentation results on complex biological image data, even if only limited amounts of training labels are available. Code available at https://github.com/juglab/eps-Seg. + +# 1 Introduction + +Electron Microscopy (EM) comes in multiple flavors and is without doubt the tool of choice for high-resolution investigations of biological samples [12]. Today, microscopists can capture fine cellular structures at nanometer resolution [22, 3]. Although this opens unprecedented possibilities for studying the very fabric of life, it also means that such microscopes are producing an unfathomable amount of raw image data that then are available to be analyzed [36]. + +A key module of nearly every analysis pipeline is the segmentation step, where specific structures of interest must be found in the entire body of captured image data. Performing this step manually, is typically not feasible as it takes an impossibly long time [16, 36, 22]. Unfortunately, even semantic segmentation of EM data of biological samples remains a challenge [3, 31]. + +Ideally, methods for segmenting EM data should $(i)$ lead to sufficiently good segmentation results for the downstream analysis tasks at hand with as few training labels as possible, $(ii)$ generalize well to different imaging conditions and image tissue types and/or be able to fine-tune on moderate amounts of new training data [9], $(iii)$ be able to benefit from sparse labeled data via supervised contrastive learning approaches, and if possible $(iv)$ operate on a hierarchy of spatial scales to distinguish objects not only by either detailed textures or larger scale shapes, but both. + +With this in mind, we introduce $\epsilon$ -Seg, a novel and sparsely supervised semantic segmentation framework for EM images that reduces the 'hunger' for labeled data by using a powerful hierarchical VAE (HVAE) [28, 21] with a GMM prior instead of a regular Gaussian one. Furthermore, our method uses center-region inpainting and contrastive learning to enhance feature consistency and segmentation robustness, even when training data is scarce. Hence, $\epsilon$ -Seg learns structured latent space representations with effective feature separation for the semantic classes of interest. Once such features are learned, they can be clustered to obtain meaningful semantic segmentations. However, since this process is computationally intensive, we integrate a dedicated semantic segmentation head that directly produces segmentation labels, improving both accuracy and runtime. + +# 2 Related Work + +Sparse Supervision. Deep learning has transformed microscopy image segmentation. The U-Net [26] has long been a standard architecture, achieving strong results when trained in a fully supervised setting. However, such approaches rely on dense annotations, which are costly and time-consuming to obtain. At the other extreme, self-supervised methods such as MAESTER [34] learn directly from raw data without labels, offering excellent scalability but typically at the cost of reduced segmentation accuracy compared to fully supervised approaches. Between these extremes lies a growing body of work on sparse or weak supervision, which seeks to achieve label efficiency while maintaining good performance. We aim to surpass self-supervised methods in accuracy while requiring only a fraction of the annotations needed by fully supervised methods. Comprehensive reviews on segmentation methods in large-scale EM with deep learning are available [3], with representative examples including slice-wise pseudo-label propagation for neuronal membranes (4S) [30], or domain adaptation variants of U-Net designed for limited-annotation settings [4]. + +Hierarchical Variational Autoencoders. Hierarchical architectures, like HVAEs [28, 21, 32, 7, 24], appear to be an interesting choice for segmenting biological microscopy data. Based on variational autoencoders [20], these powerful models learn a full approximate posterior, but are limited by the typically used Gaussian prior, making us wonder if a Gaussian mixture would not be a more suitable choice for the semantic segmentation task at hand. While the above-mentioned methods pursue label efficiency through different strategies, they do not explicitly enforce semantically disentangled latent representations. In contrast, we explicitly enforce semantically disentangled latent representations by combining a GMM prior with contrastive learning, ensuring that each latent component aligns with a distinct object class. This motivates our focus on HVAEs, which progressively encode features from fine to coarse across network layers. As higher-level semantic structure emerges in deeper layers, the latent space can be disentangled and aligned with semantic classes, enabling efficient segmentation and downstream biological analysis. + +Gaussian Mixture Models (GMMs). GMMs have been extensively used to model multimodal distributions and are a key component for many clustering methods [8, 27, 5, 10]. Many approaches integrate GMMs within autoencoder-based architectures, either explicitly as a clustering module [5] or by enforcing multimodal latent structure through a GMM prior [8, 10]. In VAEs, GMM priors enable structured latent spaces where each mixture component represents a distinct cluster or class [10, 8]. Some methods employ direct optimization of GMM objectives alongside autoencoders [5], while others leverage categorical latent variables within GMVAE frameworks, using discrete reparameterization techniques such as the Gumbel-Softmax [19] relaxation to improve scalability [8]. These techniques effectively combine deep generative models with Gaussian mixture priors, enhancing unsupervised representation learning and clustering performance in high-dimensional data spaces. + +Contrastive Learning (CL). CL has gained attention for its ability to refine feature representations by maximizing similarities between related samples and minimizing them between unrelated ones. Methods like SimCLR [6] and MoCo [15] demonstrated their effectiveness in many applications. In the context of EM segmentation, CL enables better alignment of latent representations with subcellular structures. We will use CL to ensure that each GMM component corresponds to a distinct semantic class, not just in the highest level of the hierarchy we learn. + +Next, we present our proposed method, which integrates hierarchical variational autoencoders with GMM-based priors and contrastive learning to achieve accurate and label-efficient EM segmentation. + +![](images/3d699add13f898d834320a520deac9daea279692777e2b5c83b936ac3e031a2f.jpg) +Figure 1: The overall pipeline of $\epsilon$ -Seg which is trained on an inpainting task (of center-region masked inputs). $\phi$ and $\theta$ are encoder and decoder of the network, respectively. Dotted arrows show sampling from a distribution (Gumbel-Softmax (Categorical-like distribution) for segmentation head and Normal distribution for conditional posterior). $h$ is an intermediate feature embedding of input $x$ coming from the encoder $\phi$ . $f(h)$ is a logit vector and $|f(h)| = C$ with $C$ being the number of different classes/GMM prior components (equal to 4 for "BetaSeg" [22]). $\beta$ and $\gamma$ are feature-wise linear modulation (FiLM's [23]) parameters (shifting and scaling factors) of features $h$ . $h'$ are the posterior distribution's parameters and are divided into two chunks shown as $\mu_L(x)$ and $\sigma_L(x)$ by c being the corresponding label of the masked center region of each input patch $x$ in the batch. $z_L$ is a sample from $\mathcal{N}(\mu_L(x), \sigma_L^2(x))$ . $\pmb{y}'$ is a differentiable sample from a Gumbel-Softmax [19] distribution. Green arrow shows positive pair of patches having similar labels, and red arrows show negative pairs of patches having dissimilar labels. $\mathcal{L}_{CL}$ is then computed on $\pmb{\mu}s$ (further explanation can be found in section 3). For $\mathcal{L}_I$ inpainting loss, $\mathcal{L}_{CL}$ contrastive loss, $\mathcal{L}_{CE}$ cross-entropy loss and $\mathcal{L}_{KL}$ refer to Equations 1, 16, 14 and 15 respectively. + +# 3 Methods + +The method we propose is based on a Hierarchical VAE (HVAE) backbone similar to the ones described in [28, 24]. We modify the standard HVAE setup by $(i)$ using a Gaussian mixture model (GMM) instead of the default Gaussian, so every semantic class we want to distinguish has its own predetermined Gaussian region, and by $(ii)$ adding a contrastive loss (CL), we further ensure that latent encodings are grouped by their semantic similarity through all hierarchy levels. + +As the basis for our work, we used the openly available HVAE backbone of Hierarchical DivNoising (HDN) [24]. HVAEs, as introduced elsewhere [28, 32, 21, 24], consist of a bottom-up path (encoder) and a top-down path (decoder) with trainable parameters $\phi$ and $\theta$ , respectively. The encoder extracts features from a given input $x$ at progressively coarser scales, creating a hierarchical latent encoding $z$ that splits into sub-spaces $z_{i}, i = 1 \dots L$ , with $L$ being the number of hierarchy levels, or latent layers, in the HVAE. The decoder network in regular HVAEs reconstructs $x$ , starting from the topmost latent variables $z_{L}$ . Here, we first switch from reconstructing $x$ to inpainting a masked central region in $x$ , as described next. + +Autoencoding vs. Inpainting. In contrast to regular VAEs and HVAEs that use a reconstruction loss on full input patches $\mathbf{x}$ , we are using masked autoencoding instead [18]. Since our aim is to learn semantic features that can be used for pixel-level semantic segmentation, the zero-masking we employed asks the network to only reconstruct the masked region, effectively learning features that best represent the masked semantic class. We conducted experiments with masked regions of various sizes and have always ensured that all masked pixels were from the same semantic class, see Table 8. + +The model is trained to reconstruct the masked center pixel(s) using an MSE-based inpainting loss on $X$ , a training batch of inputs, of size $B$ , as + +$$ +\mathcal {L} _ {\mathrm {I}} = \frac {1}{B} \sum_ {\boldsymbol {x} \in X} \left(\boldsymbol {x} ^ {\text {m a s k}} - \hat {\boldsymbol {x}} ^ {\text {m a s k}}\right) ^ {2}, \tag {1} +$$ + +where $\hat{x}^{\mathrm{mask}}$ is the inpainted masked region the decoder predicted, and $x^{\mathrm{mask}}$ is the mask region of the respective input patch prior to zero-masking. + +HVAEs with Gaussian Priors. The Gaussian prior of regular VAEs only applies to the topmost hierarchy level in HVAEs, where it remains $\mathcal{N}(0,I)$ as depicted in Figure S1. + +The latent variables $\mathbf{z}$ of a HVAE are split into $L$ layers $\mathbf{z}_i, i \in [1, \dots, L]$ so that + +$$ +p _ {\theta} (\boldsymbol {z}) = p _ {\theta} \left(\boldsymbol {z} _ {L}\right) \prod_ {i = 1} ^ {L - 1} p _ {\theta} \left(\boldsymbol {z} _ {i} \mid \boldsymbol {z} _ {i + 1}\right), \tag {2} +$$ + +$$ +p _ {\theta} \left(\boldsymbol {z} _ {L}\right) = \mathcal {N} \left(\boldsymbol {z} _ {L} \mid \boldsymbol {0}, \boldsymbol {I}\right), \tag {3} +$$ + +$$ +p _ {\theta} \left(\boldsymbol {z} _ {i} \mid \boldsymbol {z} _ {i + 1}\right) = \mathcal {N} \left(\boldsymbol {z} _ {i} \mid \mu_ {p, i} \left(\boldsymbol {z} _ {i + 1}\right), \sigma_ {p, i} ^ {2} \left(\boldsymbol {z} _ {i + 1}\right)\right) \text {a n d} \tag {4} +$$ + +$$ +p _ {\theta} (\boldsymbol {x} \mid \boldsymbol {z} _ {1}) = \mathcal {N} (\boldsymbol {x} \mid \mu_ {p, 0} (\boldsymbol {z} _ {1}), \sigma_ {p, 0} ^ {2} (\boldsymbol {z} _ {1})), \tag {5} +$$ + +where $\mu_{\theta}(\pmb{z}_i)$ and $\sigma_{\theta}^{2}(\pmb{z}_{i})$ represent the mean and the variance of the latent encoding, parameterized by $\theta$ . + +For each layer $i$ , the approximate posterior $q_{\phi}(z_i|\boldsymbol{x}, \boldsymbol{z}_{ _ {i} | \boldsymbol {x}\right)} \left[ \mathrm {K L} \left(q _ {\phi} \left(\boldsymbol {z} _ {i} | \boldsymbol {x}, \boldsymbol {z} _ {< i}\right) \| p _ {\theta} \left(\boldsymbol {z} _ {i} | \boldsymbol {z} _ {i + 1}\right)\right) \right], \tag {7} +$$ + +where $z_{>i}$ are all $z_{j}$ for $j > i$ . + +HVAEs with a GMM Prior. When replacing the topmost prior $p_{\theta}(\boldsymbol{z}_L)$ in an HVAE with a Gaussian mixture model (GMM), the prior becomes a weighted sum of Gaussians + +$$ +p _ {\theta} \left(\boldsymbol {z} _ {L}\right) = \sum_ {c = 1} ^ {C} \pi_ {c} \mathcal {N} \left(\boldsymbol {z} _ {L}; \mu_ {c}, \sigma_ {c} ^ {2}\right), \tag {8} +$$ + +where $C$ is the total number of Gaussian components and also the number of semantic classes we want to distinguish, $\pi_c$ are the mixing coefficients of the GMM with $\sum_{c=1}^{C} \pi_c = 1$ , and $\mathcal{N}(z_L; \mu_c, \sigma_c^2)$ is a Gaussian component with mean $\mu_c$ and standard deviation $\sigma_c$ . + +Note that there is a one-to-one correspondence between Gaussian components of the GMM and the semantic classes $\epsilon$ -Seg is supposed to distinguish. This would ensure that the latent variable follows a categorical distribution over the semantic classes; we ideally want the mixture assignment $\pi = (\pi_1, \dots, \pi_C)$ to act as a one-hot vector, i.e. one $\pi_c$ should be 1, and the rest should be 0. + +However, in practice, learning a fully discrete $\pi$ is challenging because the standard VAE framework with a GMM prior typically results in soft assignments [10]. To encourage hard assignments, one could $(i)$ use a Gumbel-Softmax [19] trick to approximate categorical sampling while maintaining differentiability [8], $(ii)$ introduce an entropy loss to encourage $\pi_c$ values to be closer to either 0 or 1. In our experiments, we used the Gumbel-Softmax during training, while reverting to the standard softmax at inference time. We also introduced an entropy loss term as a form of self-supervision, which yielded moderate improvements in the Gumbel-Softmax-based results (see Supplementary Material), but did not lead to significant gains w.r.t. the best-performing softmax configuration. We therefore report the softmax-based results as our main findings, without the additional training phase using the entropy loss. In future work, we plan to investigate alternative self-supervision strategies to + +further enhance the segmentation performance, leveraging the vast amount of available unlabeled data, within the proposed framework. + +The approximate posterior for the topmost latent $z_{L}$ , can now be expressed as + +$$ +q _ {\phi} \left(\boldsymbol {z} _ {L} \mid \boldsymbol {x}\right) = \sum_ {l = 1} ^ {C} q _ {\phi} (c = l \mid \boldsymbol {x}) q _ {\phi} \left(\boldsymbol {z} _ {L} \mid \boldsymbol {x}, c = l\right), \tag {9} +$$ + +where $q_{\phi}(c|\pmb{x})$ is the approximate posterior probability of the GMM component $c$ set to label $l$ given input $\pmb{x}$ and $q_{\phi}(\pmb{z}_L|\pmb{x},c)$ is the topmost approximate posterior conditioned on $\pmb{x}$ and component $c$ . We model $q_{\phi}(\pmb{z}_L|\pmb{x},c)$ over all possible labels itself with a Gaussian + +$$ +q _ {\phi} \left(\boldsymbol {z} _ {L} \mid \boldsymbol {x}, c\right) = \mathcal {N} \left(\boldsymbol {z} _ {L}; \mu_ {L} (\boldsymbol {x}), \sigma_ {L} (\boldsymbol {x})\right), \tag {10} +$$ + +by predicting $\mu_{L}(\pmb{x})$ and $\sigma_{L}(\pmb{x})$ (see boxes labeled with "posterior" in Figure 1). In practice, the parameters $\mu_{L}(\pmb{x})$ and $\sigma_{L}(\pmb{x})$ are computed once from the FiLM-conditioned encoder output and are shared across all components $l$ . As a result, the mixture in Equation 9 reduces to + +$$ +q _ {\phi} \left(\boldsymbol {z} _ {L} | \boldsymbol {x}\right) = \mathcal {N} \left(\boldsymbol {z} _ {L}; \mu_ {L} (\boldsymbol {x}), \sigma_ {L} (\boldsymbol {x})\right), \tag {11} +$$ + +as depicted in Figure 1. In order to predict $\mu_{L}(\pmb {x})$ and $\sigma_L(x)$ , we must compute the conditional posterior. + +Computing the Conditional Posterior. In this section, we describe the main backbone of our method leading from a given input patch $\pmb{x} \in \pmb{X}$ to the computed posteriors $q_{\phi} = \mathcal{N}(\pmb{\mu}(\pmb{x}), \pmb{\sigma}^2(\pmb{x}))$ . Figure 1 illustrates the overall pipeline of $\epsilon$ -Seg. + +The encoder, parametrized by $\phi$ , processes $x$ , leading to intermediate features $h$ in the topmost hierarchy level $L$ . These features are then passed through an MLP classifier (rouge box in Figure 1), producing a vector of logits $f(h)$ with dimensionality $C$ , coinciding with the number of classes $\epsilon$ -Seg is tasked to distinguish. + +Instead of directly using $h$ as our posterior distribution parameters, as done in our Vanilla HVAE baseline, we are using $f(h)$ , fed through two additional MLPs, $g_{\gamma}$ and $g_{\beta}$ (see violet boxes in Figure 1), to compute parameters, $\gamma$ and $\beta$ such that $\gamma = g_{\gamma}(f(h))$ and $\beta = g_{\beta}(f(h))$ . + +Those MLPs are mapping logits $f(h)$ into feature-wise scaling and shifting factors. In this way, the encoded features $h$ are modulated via these FiLM [23] parameters $\gamma$ and $\beta$ into $h'$ via computing $h' = \gamma \odot h + \beta$ , where $\odot$ denotes the Hadamard product (element-wise multiplication). The modulated feature representation $h'$ is then chunked into two parts, $\pmb{\mu}_L(\pmb{x})$ and $\pmb{\sigma}_L(\pmb{x})$ , and used to parameterize the conditional Gaussian posterior in Equation 11. + +The Latent Semantic Segmentation Head. To avoid computationally costly downstream latent space clustering to perform the semantic segmentation task (as done in Xie et al. [34] and Han et al. [14] using K-Means clustering), we are introducing a segmentation head tasked to perform the semantic pixel classification tasks directly from the computed logits $f(h)$ . + +To compute $q_{\phi}(c|\pmb{x})$ of Equation 9, we use a categorical reparameterization trick via Gumbel-Softmax [19]. + +The standard Gumbel-Softmax formula using the class probabilities $\pi_{i}$ is + +$$ +y _ {i} ^ {\prime} = \frac {\exp \left(\left(\log \pi_ {i} + g _ {i}\right) / \tau\right)}{\sum_ {j = 1} ^ {C} \exp \left(\left(\log \pi_ {j} + g _ {j}\right) / \tau\right)}, \tag {12} +$$ + +where $g_{i} \sim \mathrm{Gumbel}(0,1)$ are Gumbel noise samples. Instead of probabilities $\pi_{i}$ , we work with logits $f(h)$ (raw scores before softmax). The equivalent formula becomes + +$$ +y _ {i} ^ {\prime} = \frac {\exp \left(\left(f _ {i} (h) + g _ {i}\right) / \tau\right)}{\sum_ {j = 1} ^ {C} \exp \left(\left(f _ {j} (h) + g _ {j}\right) / \tau\right)}. \tag {13} +$$ + +The temperature parameter $\tau$ in the Gumbel-Softmax distribution plays a crucial role in controlling the degree of discreteness in the sampled values. During training, $\tau$ is often annealed from a higher value to a lower one, gradually transitioning from a smooth approximation to a discrete categorical distribution. + +In $\epsilon$ -Seg, we use a typical annealing schedule $\tau = \max(\tau_{\min}, \exp(-rt))$ , where $r = 0.999$ is the decay rate, $\tau_{\min} = 0.5$ , and $t$ is the training step. Therefore, Gumbel enables the differentiable sampling of categorical variables, improving gradient estimation, and semi-supervised classification [19]. + +Next, we draw a vector $\mathbf{y}'$ , representing the class assignment (segmentation prediction) for an input patch $\mathbf{x}^{(i)}$ in the batch $\mathbf{X}$ , by sampling from the Gumbel-Softmax distribution parameterized by logits $f(h)$ with temperature $\tau$ . + +For input patches $\pmb{x}^{(i)}\in \pmb{X}$ for which we know the class label $l_{i}$ , we want to ensure that $y_l^{\prime (i)}\in \pmb{y}^{\prime (i)}$ is the largest entry. We do so using the cross-entropy loss + +$$ +\mathcal {L} _ {C E} = - \sum_ {\boldsymbol {x} ^ {(i)} \in \boldsymbol {X}} \log y _ {l} ^ {\prime (i)}. \tag {14} +$$ + +Computing the Kullback Leibler Divergence. As it is commonly done in VAEs [20], the KL-divergence term is regularizing the parameters of our encoder, $\phi$ , such that the approximate posterior will be close to our prior $p_{\theta}(z)$ . In HVAEs, KL is computed at each hierarchy level. Changing from a standard Gaussian prior at the highest hierarchy level $L$ to using a GMM prior, as described earlier in this section, requires us to define a strategy to compute the KL-divergence appropriately. + +Hershey and Olsen [17] address the challenge of efficiently approximating the KL divergence between two GMMs, and Durrieu et al. [13] propose lower and upper bounds to estimate this divergence. While these approaches can be needed in practical setups [10, 8], we only need to compute the KL divergence between the posterior $q_{\phi}(z_L|\pmb{x})$ (Equation 11) and the $l$ -th GMM component, where $l$ is either the known class label for an input patch $\pmb{x}^{(i)}$ , or $l = \arg \max_{y^{\prime (j)}\in y^{\prime (j)}}y^{\prime (j)}$ for a patch $\pmb{x}^{(j)}$ for which we do not have a ground truth class label. + +Hence, Equation 8 becomes $p_{\theta,c}(z_L) = \mathcal{N}(z_L; \mu_l, \sigma_l^2)$ , and $\mathcal{L}_{KL}$ is therefore still computed as the divergence between two normal distributions. The KL loss over all hierarchy levels is therefore + +$$ +\mathcal {L} _ {K L} = - \left(\mathrm {K L} \left(q _ {\phi} \left(z _ {1} \mid \boldsymbol {x}\right) \| p _ {\theta} \left(z _ {1} \mid z _ {2}\right)\right) + \sum_ {i = 2} ^ {L - 1} \mathrm {K L} \left(q _ {\phi} \left(z _ {i} \mid z _ {i - 1}\right) \| p _ {\theta} \left(z _ {i} \mid z _ {i + 1}\right)\right) + \mathrm {K L} \left(q _ {\phi} \left(z _ {L} \mid z _ {L - 1}, c\right) \| p _ {\theta , c} \left(z _ {L}\right)\right)\right). \tag {15} +$$ + +Contrastive Loss. The contrastive loss consists of two terms, positive pair loss $\mathcal{L}_{+}$ , which encourages proximity between samples belonging to the same class, and negative pair loss $\mathcal{L}_{-}$ , that penalizes proximity between samples of different classes, ensuring inter-class separation. We define boolean matrices $P$ and $N$ for positive pairs and negative pairs, respectively, as $P_{ij} = \left\{ \begin{array}{ll}1 & \text{if } l_i = l_j \text{ and } i \neq j, \\ 0 & \text{otherwise} \end{array} \right.$ and $N_{ij} = \left\{ \begin{array}{ll}1 & \text{if } l_i \neq l_j, \\ 0 & \text{otherwise,} \end{array} \right.$ with $l_i$ and $l_j$ being the labels of patches $i$ and $j$ , respectively. These loss terms then become $\mathcal{L}_{+} = \frac{1}{\sum_{i,j} P_{ij}} \sum_{i,j} P_{ij} \cdot \mathcal{D}(\boldsymbol{\mu}^{(i)}, \boldsymbol{\mu}^{(j)})$ and $\mathcal{L}_{-} = \sum_{i,j} Nij \cdot \ell_{-}(\mathcal{D}(\boldsymbol{\mu}^{(i)}, \boldsymbol{\mu}^{(j)}))$ , with $\boldsymbol{\mu}^{(i)}$ being the predicted means of the posterior distribution over all hierarchy levels for a patch $i$ in batch $\mathbf{X}$ , and $\mathcal{D}(\boldsymbol{\mu}^{(i)}, \boldsymbol{\mu}^{(j)})$ a distance function. In our experiments, we used the Euclidean distance. Note that for $\mathcal{L}_{-}$ we define the penalty function $\ell_{-}(d) = \left\{ \begin{array}{ll}0 & \text{if } d \geq m, \\ (m - d)^2 & \text{otherwise}, \end{array} \right.$ with $m$ being the so-called margin, a hyperparameter that must be set appropriately, e.g. using grid-search. + +The full contrastive loss term is finally defined as + +$$ +\mathcal {L} _ {C L} = \lambda \mathcal {L} _ {+} + (1 - \lambda) \mathcal {L} _ {-}, \tag {16} +$$ + +with $\lambda$ being a hyperparameter that balances the positive and negative pair loss with each other. + +Readers might wonder why a contrastive loss is useful when a GMM prior is used, where for each structure to be classified (i.e. for each label) we have defined a Gaussian component in its own right. The main reason is that the GMM prior only takes effect at the uppermost hierarchy level $L$ . At all levels $i < L$ , $\mathcal{L}_{CL}$ is taking care of the desired label-wise segregation of latent encodings. + +The Overall Loss of $\epsilon$ -Seg. Taken all together, the overall loss of $\epsilon$ -Seg is + +$$ +\mathcal {L} = \mathcal {L} _ {I} + \alpha_ {1} \mathcal {L} _ {C E} + \alpha_ {2} \mathcal {L} _ {K L} + \alpha_ {3} \mathcal {L} _ {C L}, \tag {17} +$$ + +
Learning ParadigmModelUNGMAvg DSC
Self-SupervisedVanilla HVAE* [24]0.440.550.340.130.37
Han et al.* [14]----0.66
MAESTER* [34]0.840.950.560.790.79
Sparsely SupervisedLabkit [2]0.850.440.680.610.65
U-Net0.900.960.780.660.83
ε-Seg (ours)0.910.960.820.860.89
Fully SupervisedVanilla ViT [11]0.910.980.770.870.88
Segmenter [29]0.910.990.860.900.92
U-Net [26]0.940.990.900.870.93
+ +Table 1: Dice similarity coefficient per class and average across all classes on the "BetaSeg" dataset [22]. Methods marked with an asterisk use K-Means clustering on latent features to conduct semantic segmentation (see Section 3). U: Unrecognized, N: Nucleus, G: Granules, M: Mitochondria. + +![](images/5232133b04ad03391a4961cee3766cb5f89d6331df314aef3fadf1168ff2aec9.jpg) +Image + +![](images/1f0ed72da65651040afa5ad57eaa286754b846b6de4f15c6c030fb471fc14964.jpg) +MAESTER + +![](images/76021068a94dc3489052a82931f383cf5a3d12a2d3908f35d8c7fac08e13bc5e.jpg) +Labkit + +![](images/b9c95aeaae8e9781dd5b6ed1e704291e29a9a5bbd1f61fa18b859380f53fd132.jpg) +GT Labels + +![](images/e230907205e441c749bbda542b8657bc13b8987be76103237a6e61a443fa16d5.jpg) +$\epsilon$ -Seg (ours) +Figure 2: Qualitative segmentation result on part of the test image stack (here we show section 627 of high_c4 of the "BetaSeg" dataset [22]). + +U-Net +![](images/0e39160a3805be343a6d816795955e204e9bf2d63b31a5c8f61a32a59ff85bc7.jpg) +nucleus granules +mitochondria +unrecognized + +where $\alpha_{i}$ 's are hyperparameters to adjust the contribution of each loss with each other. We tuned those hyperparameters using grid search and by manual tuning. + +Next, we show empirical results we obtained using $\epsilon$ -Seg and comparisons to several baseline methods on two dense EM datasets and one fluorescence microscopy dataset. + +# 4 Experiments and Results + +Datasets. One of the datasets used in this study is the "BetaSeg" [22] dataset from OpenOrganelle [16], a public repository of high-resolution cellular imaging data. Acquired via Focused Ion Beam Scanning Electron Microscopy (FIB-SEM), the dataset focuses on primary mouse pancreatic islet $\beta$ cells from a high-glucose-dosage group, chosen for comparison with prior works. It underwent preprocessing, including rescaling each stack to form $4\times 4\times 4$ nm isotropic voxels, which can be viewed in any arbitrary orientations, and generating reference segmentations through human annotation or manually corrected deep learning models. The final dataset consists of four cell volumes with binary segmentation masks for seven subcellular structures, centrioles, nucleus, plasma membrane, microtubules, golgi body, granules, and mitochondria, along with an eighth "unrecognized" category. Notably, the nucleus, granules, mitochondria, and unrecognized regions dominate the dataset. For evaluation, cells 1, 2, and 3 were used for training, while cell 4 served as an independent test set. + +Next, We used "liver FIBSEM" dataset that samples were fresh needle biopsies fixed with $4\%$ PFA and $2\%$ GA in phosphate buffer. High contrast staining was performed with reduced osmium and Waltons lead aspartate stain [33] and embedded in Epon. Sample preparation and imaging was done on a ZEISS GeminiSEM according to prior reports [35]. The final dataset consists of one cell volume with 11 crops that have been extracted from a cell volume, annotated manually and used for training, validation and testing. The segmentation masks consist of six subcellular structures, mitochondria, + +![](images/d0e8f4829cb9470fbe194d444ede758b9f07da8991f6389b7ee0c17c6626fb6c.jpg) +(a) + +![](images/10c17108a0e4427c3b6392aeeec3d955bdf5afded9c511d81d50fd8419e37635.jpg) +Image + +![](images/97772fb136b5277bdbfe0e2301cd31572fefa651524629c44325cb8ea4f8790f.jpg) +GT Labels + +![](images/f149662bc7c8f4d4fdcc5d0304f7915980048488d1d3a870197a0b1644ba0f8c.jpg) +$\epsilon$ -Seg (ours) + +![](images/70bd60dca3139bdc03e06882d3607ba605247c7cdc2bca08218bfa8a08d4e41b.jpg) +U-Net +Figure 3: Qualitative segmentation result on two crops of the whole 3D volume. (a) and (b) are section 80 and 26 of crop00 and crop10 in "liver FIBSEM" dataset respectively. The U-Net is sparsely-supervised (for the fully-supervised U-Net result, see Figure S4). + +
ModelBMPLBMOBCCBCAvg DSC
U-net [26]-Fully Supervised0.970.950.850.790.520.870.900.84
U-net-Sparsely Supervised0.940.810.680.810.490.390.000.59
ε-Seg-Sparsely Supervised0.910.820.630.810.390.700.460.67
+ +peroxisomes, lipofuscin, basolateral membrane, open bile canaliculus and closed bile canaliculus, along with a seventh "background" category. + +Table 4: Dice similarity coefficient per class and average across all classes comparing our model with baselines for "liver FIBSEM" dataset. B: Background, M: Mitochondria, P: Peroxisomes, L: Lipofuscin, BM: Basolateral Membrane, OBC: Open Bile Canaliculus, CBC: Closed Bile Canaliculus. + +
Per-Class Dice CoefficientAvg DSC
BackgroundCytoplasmNuclei
0.940.860.900.90
+ +Table 2: Dice similarity coefficient per class and average for "Aitslab-bioimaging" datasets. + +
RLFPer-Class Dice CoefficientAvg DSC
UNGM
200.890.980.810.830.88
150.880.980.810.780.86
100.860.980.800.750.85
50.850.960.770.760.84
10.790.950.690.690.78
+ +Table 5: DSC per class and average across all classes. The "RLF" column (Relative Labeling Factor) specify a scaling factor where 20 corresponds to $0.05\%$ and 1 as small as $0.0025\%$ of the total labeles available. U: Unrecognized, N:Nucleus, G:Granules, M:Mitochondria. + +
Trained onPer-Class Dice CoefficientAvg DSC
UNGM
high_c10.850.380.680.610.63
high_c20.800.330.580.560.57
high_c30.820.440.630.420.58
+ +Table 3: Labkit results. Due to different image sizes, Labkit was trained on individual volumes. U: Unrecognized, N: Nucleus, G: Granules, M: Mitochondria. + +
Entropy LossPer-Class Dice CoefficientAvg DSC
UNGM
X0.810.970.740.710.81
✓0.860.980.800.750.85
+ +Table 6: Effect of entropy loss: The best checkpoint of a sparsely supervised model was further trained using batches with $50\%$ unlabeled data. U: Unrecognized, N:Nucleus, G:Granules, M:Mitochondria. + +While it is true that FIB-SEM datasets like "BetaSeg" [22] offer isotropic resolution suitable for 3D processing, this is not always the case in EM imaging, where data often comes in 2D slices (especially in higher-throughput screens). + +![](images/8d50b85d1933e80ff1a785c533cbdb51bde362ebc82470a928e8947bc5a879e6.jpg) +Cytoplasm + +![](images/5ddec6dd6c88f3284af2072fa4a5f6abcd691f77b9a1149fe7fac9fb7bdda3f8.jpg) +Nuclei + +![](images/3ed07d152fa9841a2b682fd8e85e501eb392496b0d4b8819aa6cfa1e311bed60.jpg) +GT Labels +Figure 4: Qualitative results on a representative 2-channel image from the overlapping subset of the "Aitslab-bioimaging1" and "Aitslab-bioimaging2" datasets. The first two panels show the fluorescence microscopy channels: EGFP-Galectin-3-labeled cytoplasm (left) and Hoechst 33342-stained nuclei (center-left). The center-right panel (GT) displays the ground truth semantic segmentation with nuclei (cyan) and cytoplasm (magenta). The rightmost panel ( $\epsilon$ -Seg) shows the prediction from our method. + +![](images/8601e794dfa96e52a38fef9d406da20182c25899632b125190b09caaf7247465.jpg) +$\epsilon$ -Seg + +Furthermore, we conducted an experiment on overlapping subset of two datasets Aitslab-bioimaging1 [1] and Aitslab-bioimaging2 [25]. The Aitslab-bioimaging1 dataset is a benchmarking fluorescence microscopy dataset containing 50 images of Hoechst 33342-stained U2OS osteosarcoma cell nuclei, including annotations for nuclei, nuclear fragments, and micronuclei, designed for training and evaluating neural networks for instance and semantic segmentation and the Aitslab-bioimaging2 dataset is a fluorescence microscopy dataset containing 60 images of EGFP-Galectin-3 labeled U2OS osteosarcoma cells with hand-annotated cell outlines, designed for training and benchmarking neural networks for instance and semantic segmentation, with over 2200 annotated cell objects and compatibility with object detection tasks. The overlapping subset of them contains 30, 2-channel images for training and 10 for testing. + +Evaluation Metrics. We used Dice Similarity Coefficient (DSC) to evaluate the segmentation performance. DSC is a widely used metric in image segmentation and measures the similarity between the predicted and actual segmentation masks. + +Let $A$ and $B$ be two sets representing the binary segmentation masks of the ground truth and the predicted segmentation. The Dice coefficient is defined as $\text{Dice}(A, B) = \frac{2|A \cap B|}{|A| + |B|}$ , where $|A \cap B|$ is the number of overlapping pixels between the predicted and ground truth masks, $|A|$ , the number of pixels in the ground truth mask, and $|B|$ , the number of pixels in the predicted mask. + +Experiments. We use an architecture similar to the one used in the HDN work [24]. For all hyperparameters we have introduced, we used grid searches to find a good balance between performance and stability. We first evaluate our method on "BetaSeg" dataset [22] and compare its performance against baseline methods shown in Table S1. They demonstrate that our approach outperforms existing baselines in terms of DSC (F1-score). For the Labkit baseline we trained per cell and show the results in Table 3 and report the best class-wise performance in Table S1. Quantitative segmentation result are shown in Figure 2 (complete Figure S3). + +To further validate the robustness of our method, we conduct experiments on the "liver FIBSEM" dataset, comparing it with U-Net baselines (fully and sparsely-supervised). Quantitative and qualitative results are shown in Table 4 and Figure 3, respectively (complete Figure S4). Additionally, we show $\epsilon$ -Seg results also on a fluorescent microscopy dataset (see Table 2 and Figure 4). + +Model Ablations. We strip our model down to a vanilla HVAE and then re-introduce one component at a time, showing how each of the modules we have introduced above contributes to the overall performance we report. These results on the "BetaSeg" dataset are shown in Table 7. + +Additionally, we evaluate how the quality of the results depends on the amount of available training labels. To this end, we are starting from $0.05\%$ of the total image data available in the "BetaSeg" dataset and gradually decreasing the used training labels down to $0.0025\%$ . The results of these experiments can be found in Table 5. As discussed in Section 3, $\mathcal{L}_H$ helps us to gain additional performance also from the unlabeled data, which we measure and report in Table 6. Finally, we measured the effect of differently sized masking regions in Table 8. + +
LossPrior DistributionPer-Class Dice CoefficientAvg DSC
KLCLCEUNGM
✓✗✗N0.440.550.340.130.37
✓✓✗N0.830.950.690.760.81
✓✗✓N0.810.970.800.750.83
✓✓✓N0.810.970.730.720.81
✓✗✓GMM0.820.970.720.750.82
✓✓✓GMM0.860.980.800.750.85
+ +Table 7: Loss components and prior distribution ablation on "BetaSeg" dataset. U: Unrecognized, N: Nucleus, G: Granules, M: Mitochondria. + +
Mask SizePer-Class Dice CoefficientAvg DSC
UnrecognizedNucleusGranulesMitochondria
9x90.830.950.650.730.79
7x70.840.970.720.750.82
5x50.870.940.780.800.85
3x30.880.970.810.800.87
1x10.860.980.800.750.85
+ +Table 8: Label consistency ablation on "BetaSeg" dataset. The "Mask Size" column indicates the size of the center-region mask, within which the pixel-wise ground truth labels are consistent. + +Limitations. While $\epsilon$ -Seg achieves competitive segmentation results using only sparse supervision, several limitations do remain. First, all experiments we present are conducted on 2D images. Extending the presented framework to operate in full 3D is an important next step, especially for volume EM data analysis. Second, we noticed that the effectiveness of our entropy-based loss must be improved, e.g. by replacing it with a more adaptive or data-driven strategy. Finally, in the presented form, hyperparameters such as the contrastive loss margin still require manual tuning, which is not ideal for the ease of use by biological experts. + +# 5 Discussion + +Here we presented $\epsilon$ -Seg, a novel semantic segmentation approach that leverages the variational latent representation of hierarchical variational autoencoders (HVAEs) trained on a limited amount of pixel-labels in an inpainting setup. We used a GMM prior instead of the traditionally employed Gaussian prior and introduced a novel segmentation head that incorporates both a cross-entropy loss and an entropy loss to leverage available data for which no ground truth (GT) class-labels are available. The integration of contrastive loss, combined with the structural advantages of the GMM prior, provides a means to effectively distinguish biological structures directly from the latent space encoding. + +Transformer-based architectures, as used in MAESTER [34], usually have a rather large number of trainable parameters (i.e. 328, 452, 352 trainable parameters in MAESTER). This makes such approaches less applicable to life-scientists since they require rather powerful compute setups. Even our biggest network, in contrast, only employs 3,800,869 trainable parameters (see Tables S2 and S3), making it fast to train and easy to use. Our experiments also highlight an interesting fact, namely that smaller mask sizes with consistent labels emerged as the best strategy. This stands in contrast to Transformer-based approaches where a relatively large fraction of the input images is masked during training [34]. + +By combining hierarchical representations with advanced regularization techniques such as contrastive learning, we have shown that we can achieve competitive segmentation performance on complex microscopy data, even with relatively small models and limited training data. The proposed approach tackles the challenge of label scarcity, enhances latent space representations tailored to structured biological data, and lays the groundwork for future exploration of semi-supervised learning techniques and adaptive latent priors. + +Overall, this work bridges the gap between fully supervised and unsupervised methods by offering a scalable approach for large-scale biomedical semantic image data segmentation. + +# References + +[1] Malou Arvidsson, Salma Kazemi Rashed, and Sonja Aits. An annotated high-content fluorescence microscopy dataset with Hoechst 33342-stained nuclei and manually labelled outlines. Data Brief, 46: 108769, 2023. +[2] Matthias Arzt, Joran Deschamps, Christopher Schmied, Tobias Pietzsch, Deborah Schmidt, Pavel Tomancak, Robert Haase, and Florian Jug. Labkit: Labeling and segmentation toolkit for big image data. Frontiers in Computer Science, 4, 2022. +[3] Abhinav Aswath, Abdulrahman Alsahaf, Ben N. G. Giepmans, and George Azzopardi. Segmentation in large-scale cellular electron microscopy with deep learning: A literature survey. Medical Image Analysis, 89:102920, 2023. + +[4] Róger Bermúdez-Chacón, Okan Altingövde, Carlos Becker, Mathieu Salzmann, and Pascal Fua. Visual correspondences for unsupervised domain adaptation on electron microscopy images. IEEE Trans. Med. Imaging, 39(4):1256-1267, 2020. +[5] Ahcène Boubekki, Michael Kampffmeyer, Robert Jenssen, and Ulf Brefeld. Joint optimization of an autoencoder for clustering and embedding. Machine Learning, 110(6):1901-1937, 2021. +[6] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In Proceedings of the 37th International Conference on Machine Learning (ICML), pages 1597-1607. PMLR, 2020. +[7] Rewon Child. Very deep VAEs generalize autoregressive models and can outperform them on images. In International Conference on Learning Representations (ICLR), 2021. +[8] Mark Collier and Hector Urdiales. Scalable deep unsupervised clustering with concrete GMVAEs. In 1st Workshop on Deep Continuous-Discrete Machine Learning, ECML, 2019. +[9] Ryan Conrad and Kedar Narayan. CEM500K, a large-scale heterogeneous unlabeled cellular electron microscopy image dataset for deep learning. eLife, 10:e65894, 2021. +[10] Nat Dilokthanakul, Pedro A. M. Mediano, Marta Garnelo, Matthew C. H. Lee, Hugh Salimbeni, Kai Arulkumaran, and Murray Shanahan. Deep unsupervised clustering with gaussian mixture variational autoencoders. arXiv preprint arXiv:1611.02648, 2016. Under review at ICLR 2017. +[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations (ICLR), 2021. +[12] Damjana Drobne. 3D imaging of cells and tissues by focused ion beam/scanning electron microscopy (FIB/SEM). Methods in Molecular Biology, 950:275-292, 2013. +[13] Jean-Louis Durrieu, Jean-Philippe Thiran, and Francis Kelly. Lower and upper bounds for approximation of the kullback-leibler divergence between gaussian mixture models. In Proceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 4833–4836, 2012. +[14] Hongqing Han, Mariia Dmitrieva, Alexander Sauer, Ka Chun Tam, and Jens Rittscher. Self-supervised voxel-level representation rediscovers subcellular structures in volume electron microscopy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 2276-2285, 2022. +[15] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9726-9735, 2020. +[16] Lars Heinrich, Daniel Bennett, David Ackerman, William Park, John Bogovic, Nico Eckstein, Alexander Petruncio, Joe Clements, Sharmistha Pang, Chao-Shun Xu, Jan Funke, Walter Korff, Harald F. Hess, Jennifer Lippincott-Schwartz, Stephan Saalfeld, Andrew V. Weigel, and COSEM Project Team. Whole-cell organelle segmentation in volume electron microscopy. Nature, 599(7883):141-146, 2021. +[17] John R. Hershey and Peder A. Olsen. Approximating the kullback–leibler divergence between gaussian mixture models. In Proceedings of the IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP), pages IV-317-IV-320, 2007. +[18] Zhicheng Huang, Xiaojie Jin, Chengze Lu, Qibin Hou, Ming-Ming Cheng, Dongmei Fu, Xiaohui Shen, and Jiashi Feng. Contrastive masked autoencoders are stronger vision learners. IEEE Trans. Pattern Anal. Mach. Intell., 2023. +[19] Eric Jang, Shixiang Gu, and Ben Poole. Categorical reparameterization with gumbel-softmax. In International Conference on Learning Representations (ICLR), 2017. +[20] Durk P. Kingma, Shakir Mohamed, Danilo Jimenez Rezende, and Max Welling. Semi-supervised learning with deep generative models. In Conference on Neural Information Processing Systems (NeurIPS), 2014. +[21] Lars Maaløe, Marco Fraccaro, Valentin Lievin, and Ole Winther. BIVA: a very deep hierarchy of latent variables for generative modeling. In Advances in Neural Information Processing Systems (NeurIPS), pages 6548-6559. Curran Associates, Inc., 2019. + +[22] Andreas Müller, Daniel Schmidt, Chao-Shun Xu, Sharmistha Pang, Justin V. D'Costa, Stefan Kretschmar, Christian Munster, Thorsten Kurth, Florian Jug, Martin Weigert, Harald F. Hess, and Michele Solimena. 3D FIB-SEM reconstruction of microtubule-organelle interaction in whole primary mouse $\beta$ cells. Journal of Cell Biology, 220(2):e202010039, 2021. +[23] Ethan Perez, Florian Strub, Harm de Vries, Vincent Dumoulin, and Aaron C. Courville. Film: Visual reasoning with a general conditioning layer. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), pages 3942-3951, 2018. +[24] Mangal Prakash, Maurizio Delbrazio, Peyman Milanfar, and Florian Jug. Interpretable unsupervised diversity denoising and artefact removal. In International Conference on Learning Representations (ICLR), 2022. +[25] Salma Kazemi Rashed, Malou Arvidsson, Rafsan Ahmed, and Sonja Aits. An annotated high-content fluorescence microscopy dataset with EGFP-galectin-3-stained cells and manually labelled outlines. Data Brief, 58:111148, 2025. +[26] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention (MICCAI), pages 234-241. Springer, 2015. +[27] Marek Śmieja, Maciej Wolczyk, Jacek Tabor, and Bernhard C. Geiger. SeGMA: Semi-supervised gaussian mixture autoencoder. IEEE Trans. Neural Netw. Learn. Syst., 32(9):3930-3941, 2021. +[28] Casper Kaae Sønderby, Tapani Raiko, Lars Maaløe, Søren Kaae Sønderby, and Ole Winther. Ladder variational autoencoders. In Advances in Neural Information Processing Systems (NeurIPS), pages 3745-3753. Curran Associates, Inc., 2016. +[29] Robin Strudel, Ricardo Garcia, Ivan Laptev, and Cordelia Schmid. Segmenter: Transformer for semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 7262-7272, 2021. +[30] Eichi Takaya, Yusuke Takeichi, Mamiko Ozaki, and Satoshi Kurihara. Sequential semi-supervised segmentation for serial electron microscopy image with small number of labels. J. Neurosci. Methods, 351: 109066, 2021. +[31] Kai Philipp Treder, Chenyang Huang, Jinseok S. Kim, and Angus I. Kirkland. Applications of deep learning in electron microscopy. Microscopy (Oxford), 71(Supplement_1):i100-i115, 2022. +[32] Arash Vahdat and Jan Kautz. NVAE: A deep hierarchical variational autoencoder. In Advances in Neural Information Processing Systems (NeurIPS), pages 19667-19679, 2020. +[33] J. Walton. Lead aspartate, an en bloc contrast stain particularly useful for ultrastructural enzymology. J. Histochem. Cytochem., 27(10):1337-1342, 1979. +[34] Ronald Xie, Kuan Pang, Gary D. Bader, and Bo Wang. MAESTER: Masked autoencoder guided segmentation at pixel resolution for accurate, self-supervised subcellular structure recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17521-17531, 2023. +[35] C. Shan Xu, Kenneth J. Hayworth, Zhiyuan Lu, Peter Grob, Ana M. Hassan, José G. García-Cerdán, Krishna K. Niyogi, Eva Nogales, Richard J. Weinberg, and Harald F. Hess. Enhanced FIB-SEM systems for large-volume 3D imaging. eLife, 6:e25916, 2017. +[36] Chao-Shun Xu, Sharmistha Pang, Gleb Shtengel, Andreas Müller, Anna T. Ritter, Heather K. Hoffman, Shin-Ya Takemura, Zhipeng Lu, Helene A. Pasolli, Nikhil Iyer, Jihoon Chung, Daniel Bennett, Andrew V. Weigel, Michael Freeman, Sean B. van Engelenburg, Tobias C. Walther, Robert V. Farese Jr, Jennifer Lippincott-Schwartz, Ira Mellman, Michele Solimena, and Harald F. Hess. An open-access volume electron microscopy atlas of whole cells and tissues. Nature, 599(7883):147-151, 2021. Erratum in: Nature, vol. 599, no. 7885, p. E5, 2021, doi:10.1038/s41586-021-04132-8. + +# $\epsilon$ -Seg: Sparsely Supervised Semantic Segmentation of Microscopy Data + +![](images/361d0272d55368c13a79d57d1e8b3b9182debfca89e0918b40dacc4bdb7ea04b.jpg) +Supplementary Material +Figure S1: The overall pipeline of Vanilla HVAE in Table S1 (first row in Table 7), which is trained on an inpainting task (of the center-region masked inputs). $\phi$ and $\theta$ are encoder and decoder of the network, respectively. Dotted arrows show sampling from a distribution. $h$ is an intermediate feature embedding of input $\pmb{x}$ coming from the encoder $\phi$ and it is posterior distribution's parameters which is divided into two chunks shown as $\mu_{L}$ and $\sigma_{L}$ . $z_{L}$ is a sample from $\mathcal{N}(\pmb{\mu}_{L}(\pmb{x}),\pmb{\sigma}_{L}^{2}(\pmb{x}))$ . For $\mathcal{L}_I$ inpainting loss and $\mathcal{L}_{KL}$ refer to Equations 1 and 7 respectively. + +![](images/f85d38b7d43b978180063b15c35551970ef333b9ad5f872522ccc8249017726c.jpg) +Figure S2: The overall pipeline of Vanilla HVAE with only CL added in the pipeline in the second row in Table 7, which is trained on an inpainting task (of the center-region masked inputs). Green and red arrows are showing positive and negative pair respectively, in a batch. $\phi$ and $\theta$ are encoder and decoder of the network, respectively. Dotted lines show sampling from a distribution. $h$ is an intermediate feature embedding of input $\pmb{x}$ coming from the encoder $\phi$ and it is posterior distribution's parameters which is divided into two chunks shown as $\mu_{L}$ and $\sigma_{L}$ . $\pmb{z}_{L}$ is a sample from $\mathcal{N}(\pmb{\mu}_L(\pmb{x}),\pmb{\sigma}_L^2 (\pmb{x}))$ . For $\mathcal{L}_I$ inpainting loss, $\mathcal{L}_{CL}$ contrastive loss and $\mathcal{L}_{KL}$ refer to Equations 1, 16 and 7 respectively. + +
ModelLearning ParadigmUNGMAvg DSC
Vanilla HVAE* [24]Self-Supervised0.440.550.340.130.37
Labkit [2]Sparsely Supervised0.850.440.680.610.65
U-net [26]Fully Supervised0.940.990.900.870.93
U-netSparsely Supervised0.900.960.780.660.83
Vanilla ViT [11]Fully Supervised0.910.980.770.870.88
Segmenter [29]Fully Supervised0.910.990.860.900.92
MAESTER* [34]Self-Supervised0.840.950.560.790.79
Han et al* [14]Self-Supervised----0.66
ε-Seg (+LH)Sparsely Supervised0.890.980.810.830.88
+ +Table S1: Dice similarity coefficient per class and average across all classes comparing our model with baselines on the "BetaSeg" dataset [22]. Methods marked with an asterisk use K-Means clustering on latent features to conduct semantic segmentation (more explanation can be found in Section 3). U: Unrecognized, N:Nucleus, G:Granules, M:Mitochondria. + +
# res. +blocksPer-Class Dice CoefficientAvg +DSC
UNGM
50.860.980.800.750.85
40.850.970.800.740.84
30.880.960.810.800.86
20.870.970.810.770.86
10.850.970.800.720.84
+ +Entropy-based Loss. When the sample $y'$ of the Gumbel-Softmax distribution is uniform, the network is maximally unsure about which class to predict for the current input patch. We noticed that this is commonly the case early during training, where the network has not yet seen a lot of patches for which ground truth labels are available. + +To encourage the network not to predict a uniform $\pmb{y}^{\prime}$ , we introduced an entropy loss for all patches $\pmb{x}^{(j)}\in \pmb{X}$ for which we do not have a ground truth class label. + +$$ +\mathcal {L} _ {H} = - \sum_ {\boldsymbol {x} ^ {(j)} \in \boldsymbol {X}} \boldsymbol {y} ^ {\prime (j)} \log \left(\boldsymbol {y} ^ {\prime (j)}\right). \tag {18} +$$ + +Table S2: Residual blocks ablation (3 latent variables). U: Unrecognized, N: Nucleus, G: Granules, M: Mitochondria. + +
#latentPer-Class Dice CoefficientAvg DSC
UNGM
20.870.980.810.760.86
30.860.980.800.750.85
+ +Table S3: Latent variables ablation (5 res. blocks/layer). U: Unrecognized, N:Nucleus, G:Granules, M:Mitochondria. + +![](images/8e8eb18026e3e5d65a0ece7e0044f96a010da4cef17b5dfd697f67ea7e106b0d.jpg) +Image + +![](images/93dc66719bf002ecb3b9336f1d96f22cf30002ab277eb944f5725817484a92e7.jpg) +MAESTER + +![](images/6821904a8efd03421caecd00dc92869955abdd72204f171911ab44a63213eec1.jpg) +GT Labels + +![](images/92098fed9d1af3c28bfd43a23c762bc2b1219e1b745a1338b07d7450f616dab9.jpg) +$\epsilon$ -Seg $(+\mathcal{L}_H)$ + +![](images/3b311e1cf32ac4706a3828cae8a52419a6aac0018095e4e5609c82556d0e6c0d.jpg) +Labkit + +![](images/9e93f5b50dfee566aa7dd17a92a87ffea10f40ff63fda810b7c681aae2f0cd90.jpg) +U-Net + +Vanilla HVAE +![](images/c5b73cb3b4a13c337649790d2615d030d0a7bb212d9734c6360442c94c8c62a0.jpg) +nucleus +mitochondria +unrecognized + +Han et al. +Figure S3: Qualitative segmentation result on part of the test image stack (section 627 of high_c4 in "BetaSeg" dataset). +![](images/633115e2a7611fc836d3b610e05404e63fb51cc3d1dfb1d7ca866a4d2842fbf2.jpg) +granules + +![](images/f92ece2eaf545e6c94b39bd870779b8ebd0aac06dfa0c0240285f8db89fc4189.jpg) +(a) +Image + +mitochondria +peroxisomes +lipofuscin +open bile canaliculus +closed bile canaliculus +basolateral membrane +$\bigcirc$ background + +![](images/6dc07f037a3cdb61a00032fa78a8bbb34d62f9bd4609825939ca19a322d6f577.jpg) + +![](images/d9c7f097ca0f0a9b9a1f5619f3afb7f856b165c6fa80d345324f5245ad45a7b9.jpg) +GT Labels +U-Net (sparse) + +![](images/ede399dd3ad416f1cb974a397ebb9785014d73a25fad892d61a527c506ba480f.jpg) +Image + +mitochondria +peroxisomes +lipofuscin +open bile canaliculus +closed bile canaliculus +basolateral membrane +$\bigcirc$ background + +![](images/ffce00d4c4cc095898b9797b7dacd9428dae309d592081bae6cfc34856add15b.jpg) +U-Net (sparse) +U-Net (full) +$\epsilon$ -Seg $(+\mathcal{L}_H)$ +U-Net (full) +Figure S4: Qualitative segmentation result on two crops of the whole 3D volume. (a) and (b) are section 80 and 26 of crop00 and crop10 in "liver FIBSEM" dataset respectively. U-Net (sparse) and (full) is sparsely-supervised and fully-supervised respectively. + +
RLFModelUNGMAvg DSC
20U-net0.630.750.510.120.50
ε-Seg0.890.980.810.830.88
15U-net0.530.640.410.140.43
ε-Seg0.880.980.810.780.86
10U-net0.300.200.420.340.31
ε-Seg0.860.980.800.750.85
5U-net0.710.000.000.030.18
ε-Seg0.850.960.770.760.84
1U-net0.170.000.370.020.14
ε-Seg0.790.950.690.690.78
+ +Table S4: Comparison between U-Net and $\epsilon$ -Seg on the "BetaSeg" dataset under varying label sparsity levels. "RLF" (Relative Labeling Factor) specifies the fraction of available labels, where 20 corresponds to $0.05\%$ and 1 to $0.0025\%$ of total labels. U: Unrecognized, N: Nucleus, G: Granules, M: Mitochondria. Although both models were trained with balanced supervision, using patches selected to include all classes, the U-Net still fails to segment the nucleus at very low labeling levels (RLF 1 and 5). This illustrates a key limitation of discriminative models such as U-Net, under extreme supervision sparsity, even balanced examples may not suffice to generalize fine-grained or context-sensitive structures like the nucleus. In contrast, $\epsilon$ -Seg benefits from its class-aware latent modeling via the GMM prior, which enables it to extract meaningful representations for different structures and distinguish them semantically. We note that the sparse U-Net reported earlier was trained on slice numbers 800, 600, and 500 of the "high_c1", "high_c2", and "high_c3" volumes of the "BetaSeg" dataset. For selecting the same amount of data used in $\epsilon$ -Seg, to train the 2D U-Net on, as reported in the table above, we extracted 64x64 patches where except background, different classes are approximately well balanced. + +# NeurIPS Paper Checklist + +# 1. Claims + +Question: Do the main claims made in the abstract and introduction accurately reflect the paper's contributions and scope? + +Answer: [Yes] + +Justification: The abstract and introduction clearly state the paper's contributions, including the design of $\epsilon$ -Seg, an HVAE-based segmentation framework with a GMM prior, center-region inpainting, contrastive learning, and a dedicated semantic segmentation head. These claims are appropriately scoped and supported by the methodology and experiments presented in the rest of the paper. The text also specifies that the method works with extremely limited supervision and addresses common practical challenges in EM segmentation, which are demonstrated through empirical results. + +# Guidelines: + +- The answer NA means that the abstract and introduction do not include the claims made in the paper. +- The abstract and/or introduction should clearly state the claims made, including the contributions made in the paper and important assumptions and limitations. A No or NA answer to this question will not be perceived well by the reviewers. +- The claims made should match theoretical and experimental results, and reflect how much the results can be expected to generalize to other settings. +- It is fine to include aspirational goals as motivation as long as it is clear that these goals are not attained by the paper. + +# 2. Limitations + +Question: Does the paper discuss the limitations of the work performed by the authors? + +Answer: [Yes] + +Justification: We included a dedicated Limitations mini-headline at the end of the Experiments and Result (Section 4). There, we discuss that the current method is restricted to 2D data and would likely benefit from a 3D extension. We also note that the entropy-based loss could be further optimized, and that dataset-specific tuning is required for some hyperparameters, such as the contrastive loss margin. + +# Guidelines: + +- The answer NA means that the paper has no limitation while the answer No means that the paper has limitations, but those are not discussed in the paper. +- The authors are encouraged to create a separate "Limitations" section in their paper. +- The paper should point out any strong assumptions and how robust the results are to violations of these assumptions (e.g., independence assumptions, noiseless settings, model well-specification, asymptotic approximations only holding locally). The authors should reflect on how these assumptions might be violated in practice and what the implications would be. +- The authors should reflect on the scope of the claims made, e.g., if the approach was only tested on a few datasets or with a few runs. In general, empirical results often depend on implicit assumptions, which should be articulated. +- The authors should reflect on the factors that influence the performance of the approach. For example, a facial recognition algorithm may perform poorly when image resolution is low or images are taken in low lighting. Or a speech-to-text system might not be used reliably to provide closed captions for online lectures because it fails to handle technical jargon. +- The authors should discuss the computational efficiency of the proposed algorithms and how they scale with dataset size. +- If applicable, the authors should discuss possible limitations of their approach to address problems of privacy and fairness. + +- While the authors might fear that complete honesty about limitations might be used by reviewers as grounds for rejection, a worse outcome might be that reviewers discover limitations that aren't acknowledged in the paper. The authors should use their best judgment and recognize that individual actions in favor of transparency play an important role in developing norms that preserve the integrity of the community. Reviewers will be specifically instructed to not penalize honesty concerning limitations. + +# 3. Theory assumptions and proofs + +Question: For each theoretical result, does the paper provide the full set of assumptions and a complete (and correct) proof? + +Answer: [NA] + +Justification: The paper does not include formal theoretical results or proofs (e.g., theorems or lemmas). However, it provides detailed derivations and explanations of the model components and loss functions (see Section 3), including the use of a GMM prior in the HVAE framework and KL divergence formulation. + +# Guidelines: + +- The answer NA means that the paper does not include theoretical results. +- All the theorems, formulas, and proofs in the paper should be numbered and cross-referenced. +- All assumptions should be clearly stated or referenced in the statement of any theorems. +- The proofs can either appear in the main paper or the supplemental material, but if they appear in the supplemental material, the authors are encouraged to provide a short proof sketch to provide intuition. +- Inversely, any informal proof provided in the core of the paper should be complemented by formal proofs provided in appendix or supplemental material. +- Theorems and Lemmas that the proof relies upon should be properly referenced. + +# 4. Experimental result reproducibility + +Question: Does the paper fully disclose all the information needed to reproduce the main experimental results of the paper to the extent that it affects the main claims and/or conclusions of the paper (regardless of whether the code and data are provided or not)? + +Answer: [Yes] + +Justification: The paper provides all necessary implementation details, including model architecture (Figure 1), training settings, dataset descriptions and evaluation metrics (Section 4). Loss terms, and component configurations are also disclosed to allow faithful reproduction of the reported results. + +# Guidelines: + +- The answer NA means that the paper does not include experiments. +- If the paper includes experiments, a No answer to this question will not be perceived well by the reviewers: Making the paper reproducible is important, regardless of whether the code and data are provided or not. +- If the contribution is a dataset and/or model, the authors should describe the steps taken to make their results reproducible or verifiable. +- Depending on the contribution, reproducibility can be accomplished in various ways. For example, if the contribution is a novel architecture, describing the architecture fully might suffice, or if the contribution is a specific model and empirical evaluation, it may be necessary to either make it possible for others to replicate the model with the same dataset, or provide access to the model. In general, releasing code and data is often one good way to accomplish this, but reproducibility can also be provided via detailed instructions for how to replicate the results, access to a hosted model (e.g., in the case of a large language model), releasing of a model checkpoint, or other means that are appropriate to the research performed. +- While NeurIPS does not require releasing code, the conference does require all submissions to provide some reasonable avenue for reproducibility, which may depend on the nature of the contribution. For example + +(a) If the contribution is primarily a new algorithm, the paper should make it clear how to reproduce that algorithm. +(b) If the contribution is primarily a new model architecture, the paper should describe the architecture clearly and fully. +(c) If the contribution is a new model (e.g., a large language model), then there should either be a way to access this model for reproducing the results or a way to reproduce the model (e.g., with an open-source dataset or instructions for how to construct the dataset). +(d) We recognize that reproducibility may be tricky in some cases, in which case authors are welcome to describe the particular way they provide for reproducibility. In the case of closed-source models, it may be that access to the model is limited in some way (e.g., to registered users), but it should be possible for other researchers to have some path to reproducing or verifying the results. + +# 5. Open access to data and code + +Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material? + +# Answer: [Yes] + +Justification: Two of the datasets used in our experiments are publicly available and referenced in the paper. The third dataset is private and cannot be shared due to data access restrictions. We will publicly release the code on GitHub along with detailed instructions to reproduce all experiments based on the public datasets. + +# Guidelines: + +- The answer NA means that paper does not include experiments requiring code. +- Please see the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details. +- While we encourage the release of code and data, we understand that this might not be possible, so "No" is an acceptable answer. Papers cannot be rejected simply for not including code, unless this is central to the contribution (e.g., for a new open-source benchmark). +- The instructions should contain the exact command and environment needed to run to reproduce the results. See the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details. +- The authors should provide instructions on data access and preparation, including how to access the raw data, preprocessed data, intermediate data, and generated data, etc. +- The authors should provide scripts to reproduce all experimental results for the new proposed method and baselines. If only a subset of experiments are reproducible, they should state which ones are omitted from the script and why. +- At submission time, to preserve anonymity, the authors should release anonymized versions (if applicable). +- Providing as much information as possible in supplemental material (appended to the paper) is recommended, but including URLs to data and code is permitted. + +# 6. Experimental setting/details + +Question: Does the paper specify all the training and test details (e.g., data splits, hyperparameters, how they were chosen, type of optimizer, etc.) necessary to understand the results? + +# Answer: [Yes] + +Justification: We provide all relevant training and evaluation details, including data splits, optimizer type, learning rate, batch size, and other key hyperparameters. Where appropriate, we explain how hyperparameters were chosen, either based on prior work or grid search. + +# Guidelines: + +- The answer NA means that the paper does not include experiments. +- The experimental setting should be presented in the core of the paper to a level of detail that is necessary to appreciate the results and make sense of them. + +- The full details can be provided either with the code, in appendix, or as supplemental material. + +# 7. Experiment statistical significance + +Question: Does the paper report error bars suitably and correctly defined or other appropriate information about the statistical significance of the experiments? + +Answer: [No] + +Justification: While our main experiment (Table S1) includes 5-fold cross-validation to mitigate variability due to data splits, we did not report error bars or perform statistical significance tests. Given the limited size of our dataset and the exploratory nature of our work, our focus was on assessing the feasibility of the proposed method rather than establishing statistically significant performance differences. + +# Guidelines: + +- The answer NA means that the paper does not include experiments. +- The authors should answer "Yes" if the results are accompanied by error bars, confidence intervals, or statistical significance tests, at least for the experiments that support the main claims of the paper. +- The factors of variability that the error bars are capturing should be clearly stated (for example, train/test split, initialization, random drawing of some parameter, or overall run with given experimental conditions). +- The method for calculating the error bars should be explained (closed form formula, call to a library function, bootstrap, etc.) +- The assumptions made should be given (e.g., Normally distributed errors). +- It should be clear whether the error bar is the standard deviation or the standard error of the mean. +- It is OK to report 1-sigma error bars, but one should state it. The authors should preferably report a 2-sigma error bar than state that they have a $96\%$ CI, if the hypothesis of Normality of errors is not verified. +- For asymmetric distributions, the authors should be careful not to show in tables or figures symmetric error bars that would yield results that are out of range (e.g. negative error rates). +- If error bars are reported in tables or plots, The authors should explain in the text how they were calculated and reference the corresponding figures or tables in the text. + +# 8. Experiments compute resources + +Question: For each experiment, does the paper provide sufficient information on the computer resources (type of compute workers, memory, time of execution) needed to reproduce the experiments? + +Answer: [Yes] + +Justification: In this work, we mentioned the number of parameters in our largest model and the efficiency of our approach. Our method improves upon previous techniques by eliminating the need for K-Means clustering, allowing the model to directly generate segmentation labels from the segmentation head. This change significantly accelerates the inference process, resulting in faster segmentation without sacrificing accuracy. + +# Guidelines: + +- The answer NA means that the paper does not include experiments. +- The paper should indicate the type of compute workers CPU or GPU, internal cluster, or cloud provider, including relevant memory and storage. +- The paper should provide the amount of compute required for each of the individual experimental runs as well as estimate the total compute. +- The paper should disclose whether the full research project required more compute than the experiments reported in the paper (e.g., preliminary or failed experiments that didn't make it into the paper). + +# 9. Code of ethics + +Question: Does the research conducted in the paper conform, in every respect, with the NeurIPS Code of Ethics https://neurips.cc/public/EthicsGuidelines? + +Answer: [Yes] + +Justification: Yes, the research conducted in this paper conforms with the NeurIPS Code of Ethics. We have adhered to all relevant ethical guidelines, ensuring transparency, fairness, and respect for privacy in our work. + +Guidelines: + +- The answer NA means that the authors have not reviewed the NeurIPS Code of Ethics. +- If the authors answer No, they should explain the special circumstances that require a deviation from the Code of Ethics. +- The authors should make sure to preserve anonymity (e.g., if there is a special consideration due to laws or regulations in their jurisdiction). + +# 10. Broader impacts + +Question: Does the paper discuss both potential positive societal impacts and negative societal impacts of the work performed? + +Answer: [NA] + +Justification: The focus of this paper is primarily on technical advancements in segmentation, and while it does not explicitly address societal impacts, the method may have positive implications in fields like medical imaging. However, the societal implications are, if at all, only indirect and we believe the answer 'NA' is most appropriate. + +Guidelines: + +- The answer NA means that there is no societal impact of the work performed. +- If the authors answer NA or No, they should explain why their work has no societal impact or why the paper does not address societal impact. +- Examples of negative societal impacts include potential malicious or unintended uses (e.g., disinformation, generating fake profiles, surveillance), fairness considerations (e.g., deployment of technologies that could make decisions that unfairly impact specific groups), privacy considerations, and security considerations. +- The conference expects that many papers will be foundational research and not tied to particular applications, let alone deployments. However, if there is a direct path to any negative applications, the authors should point it out. For example, it is legitimate to point out that an improvement in the quality of generative models could be used to generate deepfakes for disinformation. On the other hand, it is not needed to point out that a generic algorithm for optimizing neural networks could enable people to train models that generate Deepfakes faster. +- The authors should consider possible harms that could arise when the technology is being used as intended and functioning correctly, harms that could arise when the technology is being used as intended but gives incorrect results, and harms following from (intentional or unintentional) misuse of the technology. +- If there are negative societal impacts, the authors could also discuss possible mitigation strategies (e.g., gated release of models, providing defenses in addition to attacks, mechanisms for monitoring misuse, mechanisms to monitor how a system learns from feedback over time, improving the efficiency and accessibility of ML). + +# 11. Safeguards + +Question: Does the paper describe safeguards that have been put in place for responsible release of data or models that have a high risk for misuse (e.g., pretrained language models, image generators, or scraped datasets)? + +Answer: [NA] + +Justification: This paper does not involve models or data with a high risk for misuse, and thus does not describe any specific safeguards related to their release. + +Guidelines: + +- The answer NA means that the paper poses no such risks. + +- Released models that have a high risk for misuse or dual-use should be released with necessary safeguards to allow for controlled use of the model, for example by requiring that users adhere to usage guidelines or restrictions to access the model or implementing safety filters. +- Datasets that have been scraped from the Internet could pose safety risks. The authors should describe how they avoided releasing unsafe images. +- We recognize that providing effective safeguards is challenging, and many papers do not require this, but we encourage authors to take this into account and make a best faith effort. + +# 12. Licenses for existing assets + +Question: Are the creators or original owners of assets (e.g., code, data, models), used in the paper, properly credited and are the license and terms of use explicitly mentioned and properly respected? + +Answer: [Yes] + +Justification: Yes, all creators and original owners of assets used in this paper, including datasets, code, and models, have been properly credited. Additionally, the licenses and terms of use associated with these assets have been explicitly mentioned and respected. + +# Guidelines: + +- The answer NA means that the paper does not use existing assets. +- The authors should cite the original paper that produced the code package or dataset. +- The authors should state which version of the asset is used and, if possible, include a URI. +- The name of the license (e.g., CC-BY 4.0) should be included for each asset. +- For scraped data from a particular source (e.g., website), the copyright and terms of service of that source should be provided. +- If assets are released, the license, copyright information, and terms of use in the package should be provided. For popular datasets, paperswithcode.com/datasets has curated licenses for some datasets. Their licensing guide can help determine the license of a dataset. +- For existing datasets that are re-packaged, both the original license and the license of the derived asset (if it has changed) should be provided. +- If this information is not available online, the authors are encouraged to reach out to the asset's creators. + +# 13. New assets + +Question: Are new assets introduced in the paper well documented and is the documentation provided alongside the assets? + +Answer: [Yes] + +Justification: Yes, the private dataset used in this paper is well documented, including details on its structure, size, and usage. However, due to privacy and confidentiality constraints, the dataset is not publicly available. Access to the dataset is restricted, but interested parties can contact the authors to be connected to the dataset owners. + +# Guidelines: + +- The answer NA means that the paper does not release new assets. +- Researchers should communicate the details of the dataset/code/model as part of their submissions via structured templates. This includes details about training, license, limitations, etc. +- The paper should discuss whether and how consent was obtained from people whose asset is used. +- At submission time, remember to anonymize your assets (if applicable). You can either create an anonymized URL or include an anonymized zip file. + +# 14. Crowdsourcing and research with human subjects + +Question: For crowdsourcing experiments and research with human subjects, does the paper include the full text of instructions given to participants and screenshots, if applicable, as well as details about compensation (if any)? + +Answer: [NA] + +Justification: This paper does not involve crowdsourcing or research with human subjects. + +Guidelines: + +- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. +- Including this information in the supplemental material is fine, but if the main contribution of the paper involves human subjects, then as much detail as possible should be included in the main paper. +- According to the NeurIPS Code of Ethics, workers involved in data collection, curation, or other labor should be paid at least the minimum wage in the country of the data collector. + +# 15. Institutional review board (IRB) approvals or equivalent for research with human subjects + +Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals (or an equivalent approval/review based on the requirements of your country or institution) were obtained? + +Answer: [NA] + +Justification: This paper does not involve research with human subjects, and therefore, no IRB or equivalent approvals were required. + +Guidelines: + +- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. +- Depending on the country in which research is conducted, IRB approval (or equivalent) may be required for any human subjects research. If you obtained IRB approval, you should clearly state this in the paper. +- We recognize that the procedures for this may vary significantly between institutions and locations, and we expect authors to adhere to the NeurIPS Code of Ethics and the guidelines for their institution. +- For initial submissions, do not include any information that would break anonymity (if applicable), such as the institution conducting the review. + +# 16. Declaration of LLM usage + +Question: Does the paper describe the usage of LLMs if it is an important, original, or non-standard component of the core methods in this research? Note that if the LLM is used only for writing, editing, or formatting purposes and does not impact the core methodology, scientific rigorousness, or originality of the research, declaration is not required. + +Answer: [NA] + +Justification: No large language models (LLMs) were used as part of the core methods in this research. + +Guidelines: + +- The answer NA means that the core method development in this research does not involve LLMs as any important, original, or non-standard components. +- Please refer to our LLM policy (https://neurips.cc/Conferences/2025/LLM) for what should or should not be described. \ No newline at end of file diff --git a/NeurIPS/2025/$_epsilon$-Seg_ Sparsely Supervised Semantic Segmentation of Microscopy Data/images.zip b/NeurIPS/2025/$_epsilon$-Seg_ Sparsely Supervised Semantic Segmentation of Microscopy Data/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..395c32e5676547342fb0d5ae6af0c4fdd8244385 --- /dev/null +++ b/NeurIPS/2025/$_epsilon$-Seg_ Sparsely Supervised Semantic Segmentation of Microscopy Data/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f04c378ac4818b10635d5250a2398b06128595c09ec26e19b105633050b45a25 +size 1160048 diff --git a/NeurIPS/2025/$_epsilon$-Seg_ Sparsely Supervised Semantic Segmentation of Microscopy Data/layout.json b/NeurIPS/2025/$_epsilon$-Seg_ Sparsely Supervised Semantic Segmentation of Microscopy Data/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..e176156f489714b7a8ceef349d24a66711ebfc9c --- /dev/null +++ b/NeurIPS/2025/$_epsilon$-Seg_ Sparsely Supervised Semantic Segmentation of Microscopy Data/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5cbe1d7ae49549c77687bff1cea6a48240c91cfd7805da2a5b3549e5231a27e1 +size 911037 diff --git a/NeurIPS/2025/$_mathcal{X}^2$-DFD_ A framework for e$_mathcal{X}$plainable and e$_mathcal{X}$tendable Deepfake Detection/bcd9cdba-c978-4228-b2cb-d5da71ce3406_content_list.json b/NeurIPS/2025/$_mathcal{X}^2$-DFD_ A framework for e$_mathcal{X}$plainable and e$_mathcal{X}$tendable Deepfake Detection/bcd9cdba-c978-4228-b2cb-d5da71ce3406_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..383f4d4d2dde30ca655054596947a1c01f8a4a66 --- /dev/null +++ b/NeurIPS/2025/$_mathcal{X}^2$-DFD_ A framework for e$_mathcal{X}$plainable and e$_mathcal{X}$tendable Deepfake Detection/bcd9cdba-c978-4228-b2cb-d5da71ce3406_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0957a8904e5eaac94fcaac2b93fd1d59f9c2b17a311aa188593f0cdc761acfe6 +size 270220 diff --git a/NeurIPS/2025/$_mathcal{X}^2$-DFD_ A framework for e$_mathcal{X}$plainable and e$_mathcal{X}$tendable Deepfake Detection/bcd9cdba-c978-4228-b2cb-d5da71ce3406_model.json b/NeurIPS/2025/$_mathcal{X}^2$-DFD_ A framework for e$_mathcal{X}$plainable and e$_mathcal{X}$tendable Deepfake Detection/bcd9cdba-c978-4228-b2cb-d5da71ce3406_model.json new file mode 100644 index 0000000000000000000000000000000000000000..91d5e16ed2b23602767c4d31036143a67d697d4b --- /dev/null +++ b/NeurIPS/2025/$_mathcal{X}^2$-DFD_ A framework for e$_mathcal{X}$plainable and e$_mathcal{X}$tendable Deepfake Detection/bcd9cdba-c978-4228-b2cb-d5da71ce3406_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b2ddc2f614aa3ea9c6d86ca46449ed15354b419ae5f387cb2267e0b516fcf128 +size 327773 diff --git a/NeurIPS/2025/$_mathcal{X}^2$-DFD_ A framework for e$_mathcal{X}$plainable and e$_mathcal{X}$tendable Deepfake Detection/bcd9cdba-c978-4228-b2cb-d5da71ce3406_origin.pdf b/NeurIPS/2025/$_mathcal{X}^2$-DFD_ A framework for e$_mathcal{X}$plainable and e$_mathcal{X}$tendable Deepfake Detection/bcd9cdba-c978-4228-b2cb-d5da71ce3406_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..0f01345e0da2248eb91ea8b03608c162e13f7552 --- /dev/null +++ b/NeurIPS/2025/$_mathcal{X}^2$-DFD_ A framework for e$_mathcal{X}$plainable and e$_mathcal{X}$tendable Deepfake Detection/bcd9cdba-c978-4228-b2cb-d5da71ce3406_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:abd9426f0db8c0f4b97f0e95d3b199b5aaafe3f25ede4e486468abe9a9a90959 +size 14647175 diff --git a/NeurIPS/2025/$_mathcal{X}^2$-DFD_ A framework for e$_mathcal{X}$plainable and e$_mathcal{X}$tendable Deepfake Detection/full.md b/NeurIPS/2025/$_mathcal{X}^2$-DFD_ A framework for e$_mathcal{X}$plainable and e$_mathcal{X}$tendable Deepfake Detection/full.md new file mode 100644 index 0000000000000000000000000000000000000000..ea4e32211bbe23a2d599bca0d8ad8d79b4ab30ce --- /dev/null +++ b/NeurIPS/2025/$_mathcal{X}^2$-DFD_ A framework for e$_mathcal{X}$plainable and e$_mathcal{X}$tendable Deepfake Detection/full.md @@ -0,0 +1,1136 @@ +# $\mathcal{X}^2$ -DFD: A framework for $\mathsf{e}\mathcal{X}$ plainable and $\mathsf{e}\mathcal{X}$ tenable Deepfake Detection + +Yize Chen $^{1,2*}$ , Zhiyuan Yan $^{3*}$ , Guangliang Cheng $^{4}$ , Kangran Zhao $^{1,2}$ , Siwei Lyu $^{5}$ , Baoyuan Wu $^{1,2\dagger}$ + +1 The Chinese University of Hong Kong, Shenzhen + +2 Shenzhen Loop Area Institute + +$^{3}$ School of Electronic and Computer Engineering, Peking University, P.R. China + +$^{4}$ Department of Computer Science, University of Liverpool, Liverpool, L69 7ZX, UK + +$^{5}$ Department of Computer Science and Engineering, + +University at Buffalo, State University of New York, Buffalo, NY, USA + +# Abstract + +This paper proposes $\mathcal{X}^2$ -DFD, an eXplainable and eXtendable framework based on multimodal large-language models (MLLMs) for deepfake detection, consisting of three key stages (see Figure 1). The first stage, Model Feature Assessment, systematically evaluates the detectability of forgery-related features for the MLLM, generating a prioritized ranking of features based on their intrinsic importance to the model. The second stage, Explainable Dataset Construction, consists of two key modules: Strong Feature Strengthening, which is designed to enhance the model's existing detection and explanation capabilities by reinforcing its well-learned features, and Weak Feature Supplementing, which addresses gaps by integrating specific feature detectors (e.g., low-level artifact analyzers) to compensate for the MLLM's limitations. The third stage, Fine-tuning and Inference, involves fine-tuning the MLLM on the constructed dataset and deploying it for final detection and explanation. By integrating these three stages, our approach enhances the MLLM's strengths while supplementing its weaknesses, ultimately improving both the detectability and explainability. Extensive experiments and ablations, followed by a comprehensive human study, validate the improved performance of our approach compared to the original MLLMs. More encouragingly, our framework is designed to be plug-and-play, allowing it to seamlessly integrate with future more advanced MLLMs and specific feature detectors, leading to continual improvement and extension to face the challenges of rapidly evolving deepfakes. Code can be found on https://github.com/chenyize111/X2DFD. + +# 1 Introduction + +Current generative AI technologies have enabled easy manipulation of facial identities, with many applications such as filmmaking and entertainment [53]. However, these technologies can also be misused to create deepfakes2 for malicious purposes, including violating personal privacy, spreading misinformation, and eroding trust in digital media. Therefore, there is a pressing need to establish a reliable and robust system for detecting deepfakes. In recent years, numerous deepfake detection methods have been proposed [34, 45, 88, 32, 6, 61, 78, 74], with the majority focusing on addressing the generalization issue when the manipulation methods between training and testing vary. However, + +![](images/8ebd4c19c0c224ca050f1f8fe85bb1792917406fb0f27913e0f6669762f7ac42.jpg) +Figure 1: High-level overview of our framework, consisting of three key stages: (1) Model Feature Assessment (MFA) evaluates and ranks the forgery-related features (e.g., blending artifacts) to generate a feature set, (2) Strong Feature Strengthening (SFS) enhances the model's strong features for improved detection and explanation, while Weak Feature Supplementing (WFS) leverages Specific Feature Detector (SFD) to compensate the model's weak features, and eventually resulting in an explainable dataset, and (3) The MLLM is fine-tuned using the dataset and then used for inference. + +these methods typically only output a probability indicating whether a given input is AI-generated [8, 83], without providing intuitive and convincing explanations behind the prediction. + +Multimodal Large Language Models (MLLMs) have shown remarkable potential in many vision tasks [71, 77, 53]. Given their strong vision-language reasoning capabilities, MLLMs offer a promising avenue for addressing the explainability gap in visual forgery detection. Recent studies [28, 60, 31, 84] have explored this direction by prompting human annotators or LLMs to describe forgery cues from multiple dimensions, which the MLLMs are then trained to detect. However, these approaches often overlook a key challenge: the reliability of the generated explanations. Due to MLLMs' well-documented tendency to hallucinate, especially under uncertain conditions [4], it is crucial to ensure that the models rely on their "familiar" forgery cues with strong discrimination for detection. Intuitively, not all forgery features are equally useful—some can be effectively leveraged for detection, while others are weakly utilized or ignored altogether. + +To investigate this, we conduct a comprehensive analysis of how well pre-trained MLLMs can utilize various forgery-related cues. As shown in Figure 2, certain cues exhibit strong detection performance (e.g., facial structures and skin tone), whereas others offer limited discriminative value (e.g., blending artifacts and lighting inconsistencies). When a cue is unfamiliar or ineffective for the model, explanations based on it become unreliable. In contrast, cues that align well with the model's capabilities produce more robust and trustworthy explanations. Therefore, to ensure reliable explanations, it is essential to + +![](images/f183158963476e4e967332d5c3ca4ca92c8a5a451837ddfc2aa4ef8d517bba32.jpg) +Figure 2: The diagram shows that pretrained models (e.g., LLaVa) effectively distinguish real from fake content using semantic features (e.g., Skin tone, Contour), but perform poorly with signal features (e.g., Blending, Lighting). + +explicitly identify and promote cues that the MLLM can reliably understand and leverage. + +Inspired by the above investigations, we propose $\mathcal{X}^2$ -DFD, a novel framework that utilizes MLLMs for deepfake detection. The key idea of our approach is to enhance the strengths and supplement the weaknesses of the original MLLMs. Our framework operates through three core stages. First, the Model Feature Assessment (MFA) assesses the intrinsic capability of the original MLLMs in deepfake detection. This stage quantifies the discriminative capability of each forgery-related feature, producing a prioritized ranking based on its importance to the model. Second, the Strong Feature Strengthening (SFS) and Weak Feature Supplementing (WFS) reinforce strong features and compensate for weak ones, resulting in a more explainable dataset. Third, we use the created dataset from the second stage to fine-tune the MLLM and then use it for improved detection and explanation. This integration enables us to leverage the strengths of both MLLMs and Specific Feature Detectors (SFDs) effectively and fuse the large and small models adaptively. Encouragingly, the modular-based design of the proposed $\mathcal{X}^2$ -DFD framework enables seamless integration with future MLLMs and SFDs as their capabilities evolve. + +Our main contributions are threefold. + +- We systematically assess the intrinsic capabilities of MLLMs for deepfake detection: To our knowledge, we are the first work to provide an in-depth analysis of MLLMs' inherent ability in + +deepfake detection. Our findings reveal that MLLMs exhibit varying discrimination capabilities across different forgery features. + +- We enhance MLLMs' explainability by reinforcing their strong features: Building on their strengths, we fine-tune MLLMs to generate explanations based on their most "familiar" forgery features, improving both detection accuracy and explainability. +- We further integrate Specific Feature Detectors to supplement the model's weakness, For forgery features where MLLMs struggle, we incorporate SFDs to complement their limitations, creating a more robust detection system. + +# 2 Related Work + +Conventional Deepfake Detection Early detection methods typically focus on performing feature engineering to mine a manual feature such as eye blinking frequency [36], warping artifacts [34], headpose [82], and etc. Recent conventional deepfake detectors mainly focus on dealing with the issue of generalization [79], where the distribution of training and testing data varies. Until now, they have developed novel solutions from different directions: constructing pseudo-fake samples to capture the blending clues [34, 32, 62, 89], learning spatial-frequency anomalies [23, 45, 48, 54], focusing on the ID inconsistency clues between fake and corresponding real [17], performing disentanglement learning to learn the forgery-related features [78, 81, 22], performing reconstruction learning to learn the general forgery clues [5, 67], locating the spatial-temporal inconsistency [24, 69, 90, 80, 86], and etc. However, these methods can only provide real or fake predictions without providing detailed explanations. The lack of convincing and human-comprehensible explanations might confuse users about why the predictions are deemed fake. + +Deepfake Detection via Multimodal Large Language Model Vision and language are the two important signals for human perception, and visual-language multimodal learning has thus drawn a lot of attention in the AI community. Recently, the LLaVA series [44, 43, 42] have explored a simple and effective approach for visual-language multimodal modeling. In the field of deepfake detection, [28, 60] have investigated the potential of prompt engineering in face forgery analysis and proposed that existing MLLMs show better explainability than previous conventional deepfake detectors. In addition, [35, 21, 31] probed different MLLMs for explainable fake image detection and [35, 70] by presenting a labeled multimodal database for fine-tuning. In parallel, VIPGuard [40] explores explainable deepfake detection by leveraging identity information through an MLLM. More recently, [87] proposed using pairs of human-generated visual questions answering (VQA) to construct the fine-tuning dataset, but manually creating detailed annotations can be very costly. Addressing this limitation, [27] recently introduced an automated pipeline using GPT-4o [1] to generate VQA pairs for dataset construction and MLLM training. However, a new critical question was then raised: Can MLLMs (e.g., LLaVa) fully comprehend the fake clues identified by GPT-4o? We argue that there could remain a "capability gap" between different MLLMs, particularly between "annotation generators" (GPT-4o) and "consumer models" (LLaVA). This gap exposes two unresolved challenges: (1) systematically analyzing the limitations of MLLM-based detectors in understanding all synthetic forgery clues (e.g., identifying specific detection capabilities they lack) and (2) developing methods to enhance their existing strengths (e.g., semantic consistency analysis) while compensating for weaknesses (e.g., fine-grained artifact recognition). To our knowledge, most existing works fail to adequately address the two key challenges, leaving a critical void in building more robust and explainable deepfake detection systems. + +# 3 Method + +In this work, we propose a general explainable and extendable multimodal framework for deepfake detection, which consists of three key stages: (1) Model Feature Assessment (MFA) evaluates and ranks the forgery-related features, (2) Strong Feature Strengthening (SFS) enhances the model's strong features and Weak Feature Supplementing (WFS) leverages Specific Feature Detector (SFD) to compensate the model's weak features, and eventually resulting in an explainable dataset, and (3) The MLLM is fine-tuned on this dataset and then used for inference for enhanced deepfake detection and explanations. In the following content, we will introduce the technical details of these stages. + +![](images/6b3773486b6b8629ba72949409c4a012a6761cbf9e7369f1c63077a187036d50.jpg) +Figure 3: A comprehensive breakdown of the three-stage methodology for $\mathcal{X}^2$ -DFD. In Stage 1, an automated procedure for forger-related feature generation, evaluation, and ranking is implemented within the MFA (Model Feature Assessment) module. Stage 2 incorporates the SFS (Strong Feature Strengthening) module, which automates the generation of explanatory annotations for a fine-tuning dataset consisting of real and fake images, leveraging strong features, alongside the WFS (Weak Feature Supplementing) module, which employs a specific feature detector to produce explanations for weak features. Stage 3 entails model fine-tuning and inference, empowering the model to excel in detection performance and provide precise explanations, utilizing both its proficient strong features and less proficient weak features for improved detection and explanation. + +# 3.1 Model Feature Assessment (MFA) + +As depicted in the top row of Figure 3, the MFA module consists of three sequential stages: feature-related question generation, assessment, and ranking. Each stage plays a crucial role in identifying and prioritizing forgery-related features. + +Step 1: Feature-related Question Generation. For each candidate forgery-related feature, a corresponding question is formulated to assess its presence in an image. Given that these features are not predefined, a Large Language Model (LLM), such as GPT-40, is leveraged to generate a diverse set of $N_{f}$ questions, denoted as $F_{i}$ . These questions are designed to probe key forgery indicators, including facial inconsistencies, unnatural color, and texture mismatches elements critical for deepfake detection. + +Step 2: Model Feature Assessment. Each generated question is paired with an image from the assessment dataset, forming structured prompts for model inference. The Multi-Modal Large Language Model (MLLM) then responds with binary outputs (yes or no), which are aggregated into a confusion matrix to quantify the reliability of each feature. Specifically, for an image $x_{j}$ and question $F_{i}$ , where $i$ represents the index of forgery feature-related question, and $j$ represents index of an image, the MLLM produces: + +$$ +R _ {i, j} = \mathcal {M} _ {\text {b a s e}} \left(F _ {i}, x _ {j}\right), \tag {1} +$$ + +where $R_{i,j} \in \{\text{yes, no}\}$ , representing the model's response. This step ensures that the generated questions effectively capture forgery-related discrepancies. + +Step 3: Feature Ranking. To prioritize the most discriminative features, questions are ranked based on their Balanced Accuracy (BA): + +$$ +\mathrm {B A} _ {i} = \frac {1}{2} \left(\frac {\mathrm {T P} _ {i}}{\mathrm {T P} _ {i} + \mathrm {F N} _ {i}} + \frac {\mathrm {T N} _ {i}}{\mathrm {T N} _ {i} + \mathrm {F P} _ {i}}\right), \tag {2} +$$ + +where $\mathrm{TP}_i, \mathrm{TN}_i, \mathrm{FP}_i, \mathrm{FN}_i$ denote True Positives, True Negatives, False Positives, and False Negatives, respectively. Due to potential class imbalance between real and fake samples in the dataset, we use the simple and widely adopted Balanced Accuracy (BA) metric. This fairly evaluates both classes, aiding effective feature ranking. The ranking identifies the most reliable forgery-related features for discrimination. + +Following the automated ranking, human verification is further conducted to ensure the reliability of the identified fake features. This step mitigates potential biases or misinterpretations by the LLM, refining the final selection of discriminative features. Additionally, irrelevant or non-discriminative features are filtered out, with minimal instances of erroneous or unrelated outputs. + +# 3.2 Strong Feature Strengthening (SFS) + +The SFS module constructs datasets by leveraging the strong feature capabilities identified as high-performing by the MFA module. This process comprises two key steps. + +Step 1: Real/Fake Prompts Generation. Leveraging the strong features from the MFA module, we generate specialized prompts to guide the MLLM's focus during the fine-tuning phase. Specifically, we first utilize GPT-4o to summarize these strong features and construct two distinct prompts: one tailored for real images $(\mathbf{P}_{\mathrm{real}})$ and another for fake images $(\mathbf{P}_{\mathrm{fake}})$ . These prompts are formulated as: $\mathbf{P}_{\mathrm{real}} = f(\mathbf{F}_{\mathrm{real}}), \quad \mathbf{P}_{\mathrm{fake}} = f(\mathbf{F}_{\mathrm{fake}})$ , where $\mathbf{F}_{\mathrm{real}}$ and $\mathbf{F}_{\mathrm{fake}}$ denote the sets of strong features relevant to real and fake images, respectively. Also, $f$ represents any LLMs. Here, we employ GPT-4o for implementation. + +Step 2: Fine-tuning Dataset Construction. A fine-tuning dataset $D_{ft}$ comprising VQA-style (visual question answering) pairs, which is constructed by pairing each image with the corresponding (real or fake) prompt. Each image is annotated with the specific features it exhibits, and the standardized prompt $\mathbf{P}_{\mathrm{fixed}}$ is defined as: $\mathbf{P}_{\mathrm{fixed}} =$ "Is this image real or fake?" The model's response is structured to begin with a definitive statement—"This image is real/fake"—followed by an explanation based on the identified features. Formally, the final answer is represented as: $\mathbf{A}_{\mathrm{final}} =$ "This image is real/fake" + $\mathbf{A}_{\mathrm{real/fake}}$ . Consequently, each VQA-style pair of the fine-tuning dataset $D_{ft}$ is formalized as: $\mathbf{VQ}\mathbf{A} = (\mathbf{I}\mathbf{m}\mathbf{a}\mathbf{g}\mathbf{e},\mathbf{P}_{\mathrm{fixed}},\mathbf{A}_{\mathrm{final}})$ . + +# 3.3 Weak Feature Supplementing (WFS) + +The WFS module construct datasets by integrating specific feature detectors, which are specialized in detecting features where the MLLM shows weakness. This module follows two steps: + +Step 1: Specific Feature Detector Invocation. For features that the MLLM identifies as weak, we deploy an specific deepfake detector (e.g., a blending-based detector [41]). This specific feature detector processes the input image and generates a prediction. Note that we also employ other SFDs for implementation, and we provide an in-depth analysis for this in the Appendix G. Specifically, when utilizing a blending detector as an instance of SFD, a blending score $s$ is produced: $s = \sigma(\text{BlendDetector}(x))$ , where $x$ denotes the input image, and $\sigma$ denotes the sigmoid function that transforms the logits output of the BlendDetector into the 0-1 range. + +Step 2: Integration of Specific Feature Detection Results into the Fine-tuning Dataset. + +The blending score $s$ obtained from the specific detector is incorporated into the fine-tuning dataset by appending it to the existing prompts. This is done by adding a statement such as: "And the blending feature score of content is: $\underline{s}$ " Additionally, based on the score, a corresponding response aligned with the probability is included, specifically in the Fine-tuning Dataset Construction section of the SFS. This integration ensures that the MLLM benefits from both its intrinsic detection capabilities and the specialized insights provided by the SFD. + +# 3.4 Model Finetune and Inference + +After obtaining the constructed dataset, the following steps involve fine-tuning and inference. The stage following two steps: + +Step 1: MLLM Fine-tuning. The initial MLLM is fine-tuned using the dataset $D_{ft}$ . This process adjusts the projector to accurately link image artifacts with corresponding fake labels. Additionally, + +Low-Rank Adaptation (LoRA) [25] is applied to selectively update a subset of the model's parameters, enhancing its focus on deepfake-specific features while preserving overall model integrity. This fine-tuning can be expressed as: + +$$ +\mathcal {M} _ {\text {b a s e}} \xrightarrow {D _ {f t}} \mathcal {M} _ {\text {f i n e - t u n e d}}, +$$ + +where $\mathcal{M}_{\mathrm{fine - tuned}}$ represents the enhanced MLLM with superior deepfake detection capabilities. + +Step 2: Integration of Specific Feature Detection into Inference Prompts. Generally, during the inference, the SFD detector's blending score $s$ is incorporated into the MLLM's prompt-based reasoning process. The final output of the model is structured, to begin with a definitive statement: "This image is real/fake", followed by reasoning based on identified visual features. Based on the blending score $s$ , the model appends a descriptive statement: $\mathbf{A}_{\mathrm{final}} =$ "This image is real/fake" (binary results) + $\mathbf{A}_{\mathrm{real/fake}}$ (explanations) + "And this image contains obvious/minimal blending artifacts" (clues from SFD). The model acquires this response pattern through training. This approach ensures that the MLLM effectively leverages SFDs to enhance its detection performance, particularly for features where it initially demonstrated weakness. + +# 4 Experiment + +# 4.1 Experimental Setup + +Datasets. We evaluate our proposed method on a diverse set of widely-used deepfake detection datasets, including the Deepfake Detection Challenge (DFDC) [16], its preview version (DFDCP) [15], DeepfakeDetection (DFD) [13], Celeb-DF-v2 (CDF-v2) [37], FaceForensics++ (FF++) [58] (c23 version for training), DFo [29], WildDeepfake (WDF) [95], FFIW [92], and the newly released DF40 dataset [76], which incorporates state-of-the-art forgery techniques such as Facedancer [57], FSGAN [51], inSwap [59], e4s [33], Simswap [7], and Uniface [91]. In line with the standard deepfake benchmark [79], we use the c23 version of FF++ for training and other datasets for testing in the main table. Additionally, we evaluated a broader range of facial forgery types using the DiFF [9] dataset, a comprehensive collection of diffusion-generated facial images, which allowed us to test our method on a wider spectrum of forgery techniques. + +Evaluation Metrics. We assess the performance of our model in terms of both detection performance and explanation quality. For detection, we adopt the Area Under the Curve (AUC) as the primary metric to evaluate the model's ability to distinguish real from fake content across entire datasets, reporting both frame-level and video-level AUC scores. Additional metrics, including Accuracy (Acc.), Equal Error Rate (EER), and Average Precision (AP), are also provided for a comprehensive analysis. For explanation, we follow [87] by using human-annotated data to measure text similarity between model-generated explanations and human-labeled ground truth, employing standard metrics such as BLEU [52], CIDEr [66], ROUGE-L [38], METEOR [14], and SPICE [2]. Beyond text similarity, we engage human evaluators and GPT-4o to assess the quality of explanations regarding forgery content, following prior studies [73, 21]. Evaluators rate the explanations on a scale from 0 (very poor) to 5 (excellent), ensuring a robust qualitative evaluation. + +Implementation Details. We initialize our model with the LLaVA-base weights and fine-tune the LLaVA model [44] using its official implementation codebase. For the specific feature detectors (SFD), we adopt a blending-based approach as proposed in [41]. Training is performed on a single NVIDIA 4090 GPU for 3 epochs, with a learning rate of $2 \times 10^{-5}$ in two layer mlp projector and $2 \times 10^{-4}$ in others, a rank of 16, and an alpha value set conventionally to twice the rank at 32. We use a batch size of 4, a gradient accumulation step of 1, and a warmup ratio of 0.03 to stabilize training. + +# 4.2 Generalizability Evaluation + +Following the common settings of previous works [74, 10], we first compare our method with 33 SOTA detectors (including both conventional and multimodal-based detectors) via cross-dataset evaluations (see Table 1). The results of other compared baselines are mainly cited from their original papers. Our approach excels across both frame-level and video-level evaluations, maintaining superior results when compared to other methods. The table clearly highlights our method's capability to generalize and consistently achieve higher detection performance at the frame level and video level, respectively. + +Table 1: Cross-dataset evaluations with 33 existing detectors. The top two results are highlighted, with the best in bold and the second-best underlined. \* indicates our reproductions based on the pre-trained checkpoints released by the authors, and $\ddagger$ refers to the MLLM-based detectors, which can also output explanations. $\ddagger$ : The LAA-Net is trained on the high-quality FF++ (raw) data, whereas our method is trained on the compressed (c23) version. + +
Frame-Level AUCVideo-Level AUC
MethodCDFDFDCPDFDCDFDAvg.MethodCDFDFDCPDFDCDFDAvg.
Xception [12]73.773.770.881.675.0Xception [12]81.674.273.289.679.7
FWA [34]66.863.761.374.066.5PCL+I2G [89]90.074.467.5--
Efficient-b4 [65]74.972.869.681.574.7LipForensics [24]82.4-73.5--
Face X-ray [32]67.969.463.376.769.3FTCN [90]86.974.071.094.481.6
F3-Net [54]77.077.272.882.377.3ViT-B (CLIP) [19]88.482.576.190.084.3
SPSL [46]76.574.170.181.275.5CORE [50]80.972.072.188.278.3
SRM [48]75.574.170.081.275.2SBI* [61]90.687.775.288.285.4
ViT-B (IN21k) [55]75.075.673.486.477.6UIA-ViT [94]82.475.8-94.7-
ViT-B (CLIP) [19]81.780.273.586.680.5SLADD* [6]79.7-77.2--
RECCE [5]73.274.271.381.875.1DCL [64]88.276.975.092.183.1
IID [26]83.881.2---SeeABLE [30]87.386.375.9--
ICT [18]85.7--84.1-CFM [47]89.780.270.695.283.9
LSDA [74]83.081.573.688.081.5UCF [78]83.774.277.086.780.4
VLFFD† [63]83.283.2-94.8-NACO [86]89.5-76.7--
FFAA† [27]--74.092.0-AltFreeing [69]89.5----
RepDFD† [39]80.090.677.3--TALL-Swin [72]90.8-76.8--
MFCLIP † [39]83.586.1---StyleDFD [11]89.0--96.1-
KFD-VLM † [85]89.986.7-92.3-LAA-Net ‡ [49]95.486.9-98.4-
X²-DFD (7B)90.487.383.792.388.4X²-DFD (7B)95.489.386.095.891.6
X²-DFD (13B)91.390.383.492.589.4X²-DFD (13B)95.791.085.796.192.1
+ +# 4.3 Explainability Evaluation + +Annotated Explainability Evaluation. We assess the performance of our model using the DD-VQA [87] test dataset, which incorporates human-annotated data from FF++ [58]. The evaluation employs a suite of metrics, including BLEU [52], CIDEr [66], ROUGE-L [38], METEOR [14], and SPICE [2], to quantify the alignment between our model's responses and human-annotated ground truth. The MLLMs assessed for explanation quality include LLaVA [43], Llama3.2V [20], Qwen2VL [68] and GPT4o [1]. The models use the same prompt to generate explainable outputs. To ensure a fair comparison, particularly given GPT-4o's tendency to refuse responses with the same prompt, we adopt a prompting strategy from [28]. This leads GPT-4o to generate shorter responses, resulting in lower scores. The evaluation results are summarized in Table 2 (Annotated Explainability Evaluation). Due to the DD-VQA only annotating the artifact in some specific parts (e.g., nose, eyes), GPT-4o and human experts are required to evaluate both annotated and unannotated scenarios, ensuring a thorough assessment of model explainability across different scenarios. + +Table 2: Explainability evaluation across annotated and unannotated settings, comparing five scores for annotated explainability, alongside human and GPT-4o evaluations (scored 0-5) for unannotated explainability. The best result per metric is highlighted in bold. + +
ModelAnnotated ExplainabilityUnannoted Explainability
BLEUCIDErROUGE-LMETEORSPICEAvg.Human-EvalGPT4-EvalAvg.
LLaVA-7B [44]0.1830.0210.1390.1100.0850.1082.3681.5421.955
Llama3.2-Vision [20]0.1310.0090.1310.0810.1160.0932.2651.6671.966
Qwen2.5-VL-7B [68]0.1400.0120.1430.0810.1500.1052.0341.3831.709
GPT-4o [1]0.1230.0110.0820.0510.0720.0682.5592.0552.307
Ours0.2030.0270.1550.1480.1550.1383.5722.6683.120
+ +Unannnotated Explainability Evaluation. To evaluate unannotated explainability, we build on insights from prior work [73, 21, 93] and utilize both human evaluators and GPT-4o to assess model performance across three key dimensions: (1) detection ability, (2) reasonableness of explanations, and (3) level of detail. Each dimension is scored on a scale from 0 to 5. The evaluation results are + +summarized in Table 2 (Unannotated Explainability Evaluation). The results demonstrate that our model achieves strong performance in both human-eval and GPT-eval. Additional details include the experiment of the setting of human study, Graphical User Interface (GUI) of human study, and the evaluation prompt of GPT4o can be found in Appendix. + +# 4.4 Computational Complexity Evaluation + +We performed a simple evaluation of the time complexity. Compared to other pretrained MLLMs, the model does not require much additional inference time in the same inference framework, as seen in Table 3, making the time acceptable. However, when compared to conventional Non-MLLM methods, the model, despite gaining interpretability and a richer feature space, requires more inference time. Nevertheless, we believe that as MLLM inference frameworks advance, along with pruning and quantization lightweight methods, MLLMs will benefit from these improvements when applied to Deepfake detection. + +Table 3: Comparison of average inference time per image for different models. + +
ModelSeconds
Non-MLLMs - Xception~0.03
Non-MLLMs - CDFA~0.05
Llama-3.2-11B~3.3
Qwen2.5 VL-Instruct 7B~1.6
LLaVa-7B~1.2
Ours (X2DFD w/ two SFDs)~1.3
+ +# 5 Ablation Study and Analysis + +Here, we address several key research questions through ablation studies and in-depth analysis. + +Question 1: Why is fine-tuning MLLMs with their strong features more effective than using all? + +To enhance the reasoning and detection capabilities of MLLMs, we introduce the Strong Feature Strengthening (SFS) module. In this module, we focus on selectively amplifying the most discriminative forgery-related features—referred to as strong features. These strong features are identified through Model Feature Attribution (MFA), which ranks features based on their importance scores across different modalities and samples. Fine-tuning on these "strong features" is more effective because it focuses the learning process on highly discriminative, reliable forgery-related cues while avoiding the disruptive noise introduced by weak features. + +This design leads us to a critical question: Are these strong features truly more effective for improving model performance than using the full set of features? To answer this, we compare two strategies: (1) enhancing all features in the LLM-generated feature list and (2) enhancing only the strong features selected via MFA. As shown in Table 4, the latter consistently outperforms the former across all datasets. + +Table 4: Comparison of model detection performance and interpretation between using all features and strong features, both enhanced by the Strong Feature Strengthening (SFS) module. The evaluation metrics include AUC and GPT-4o evaluation. Additionally, results for the Top-K feature selection strategy and the No Feature Explanation are provided for comparison. + +
Model ConfigurationCDFUnifaceHPS-DiffAvgGPT4o Eval
No Feature Explanation80.381.984.782.3-
X2DFD (Top-K=25)83.084.387.184.82.91
X2DFD (Top-K=50)83.284.588.785.53.02
X2DFD (Top-K=75)81.983.284.683.22.77
X2DFD (Top-K=100)79.082.383.681.62.63
+ +These results confirm that selectively strengthening the most discriminative features not only improves the model's performance but also yields more reliable model explanations. + +Question 2: How to ensure SFS module works when it should, and stays silent when it shouldn't? + +To further improve model generalization and interpretability, we extend our framework with two new components: the Specific Feature Detection (SFD) module and the Weak Feature Supplementing (WFS) module. While the earlier SFS module focuses on enhancing strong, highly discriminative features, it may overlook subtle patterns critical for certain forgery types. To address this, WFS is designed to teach the model how to lever + +Table 5: Effect of excluding supplementary features during training (WFS) and including them at inference (SFD infer) on model performance. + +
VariantCDFDFDDFDCDFDCPAvg.
WFS × SFD infer ×83.291.479.282.084.0
WFS × SFD infer ✓81.790.679.181.383.2
WFS ✓ SFD infer ✓90.492.383.787.388.4
+ +age weak features provided by SFD—features + +that are otherwise hard for MLLMs to interpret directly. + +To investigate whether this combination yields synergistic benefits, we compare three variants: (1) baseline without WFS and SFD at inference, (2) enabling SFD only during inference (without WFS), and (3) enabling both WFS and SFD. As shown in Table 5, the model achieves the best performance when WFS is present, demonstrating that SFD's weak signals become more useful once the model has learned how to utilize them through WFS. Without WFS, simply adding SFD at inference may not help—and can even lead to degradation—indicating that "1+1" only becomes greater than 2 when weak features are integrated structurally during training. + +Beyond synergy, it is crucial to ensure that the introduction of SFD does not interfere in scenarios where its cues are irrelevant. For example, blending-based detectors may provide limited value on datasets like SRI, + +Table 6: Comparison of AUC performance for models trained on FF++ alone versus FF++ with SRI, evaluated on other datasets. + +
Train DataCDFDFDCPDFDCDFDUnifaceFsganInswapSimswapAvg.
FF++ √ SRI ×90.487.383.792.385.591.181.285.187.1
FF++ √ SRI √91.589.383.992.787.489.981.086.187.7
+ +which contain no blending traces. We test this by introducing SRI as a training set and comparing performance. As shown in Table 6, the model not only maintains its effectiveness but even improves, suggesting that when SFD signals are weak or absent, the model naturally downplays them. This demonstrates that our design is adaptive—SFD helps when it can, and steps aside when it should. + +Overall, our framework achieves both synergistic improvement and non-intrusive integration: WFS enables the model to benefit from weak features without forcing reliance, and SFD contributes only when its signals are relevant. + +Question 3: How can we generate the most suitable set of $N$ forgery-related questions in MFA? + +The Model Feature Assessment (MFA) module evaluates the model's discriminative ability by asking it to answer a curated set of N forgery-related questions. A key challenge here is: how to generate the most suitable questions that effectively probe the model's understanding of diverse forgery cues. To explore this, we compare different question-generation strategies: (1) human-written features based on expert knowledge [87], (2) features automatically generated by large language models (LLMs), including Claude3.5-Sonnet [3] and GPT-4o [1]. + +As shown in Table 7, questions generated by LLMs slightly outperform those crafted by humans in terms of detection performance across multiple datasets. This suggests that LLMs can capture a broader and potentially more nuanced range of forgery + +Table 7: Comparison between LLMs and human annotators for generating N forgery-related questions for MFA. + +
VariantCDFDFDCPDFDCDFDUnifaceFsganSimswapAvg
Human Writing [87]89.189.783.692.582.389.187.087.6
Claude3.5-Sonnet [3]90.188.583.593.084.990.085.687.9
GPT4o [1]90.389.783.592.585.289.984.987.8
+ +related features, possibly including cues overlooked by human experts. However, LLM-generated questions are not always ideal—they may sometimes be generic, redundant, or irrelevant (e.g., mistakenly treating "Photoshop traces" as core deepfake features). On the other hand, although human-designed features may be narrower in scope, they offer higher precision and domain relevance, leading to robust results. To balance these strengths and weaknesses, we adopt a hybrid strategy: (1) Use an LLM to generate a diverse pool of forgery-related questions. (2) Rank them by relevance scores. (3) Apply human verification to filter out irrelevant or low-quality questions. + +Question 4: Can framework benefit from extending multiple SFDs? + +With the continuous development of generative technologies, diffusion-based generative methods and techniques have been emerging rapidly. To test the scalability of our framework, we extended it to another SFD, AlignedForensics [56], and incorporated a new training dataset, consisting of fake faces generated through diffusion techniques from 1000 external images, which form a subset of the face-swapping data in DiFF [9]. We then tested not only the face-swapping scope, which was previously the focus of this work (e.g., CDF and Simswap), but also expanded the evaluation to include face editing, image-to-image editing, and text-to-image generation using the DiFF dataset, which also focuses on faces. + +As shown in Table 8, we found that the X2DFD framework not only effectively extended the CDFA's blending-based SFD but also enhanced the performance when extending both the CDFA and AlignedForensics (diffusion-based SFD). This extension allowed the framework to achieve strong results across various forgery types and methods. + +Table 8: AUC Performance Evaluation for SFD Integration X2DFD-Sig (CDFA only) vs. X2DFD-Mul (CDFA & AlignedForensics) + +
ModelCDFSimswapDiff-FEDiff-I2IDiff-T2I
CDFA87.976.174.681.787.3
X2DFD-Sig90.485.182.181.788.3
X2DFD-Mul90.388.592.188.692.2
+ +# 6 Conclusion + +In this paper, we propose $\mathcal{X}^2$ -DFD, a unified multimodal framework for explainable and extendable deepfake detection. For the first time, we systematically evaluate the intrinsic capabilities of the pre-trained MLLMs, revealing their varying effectiveness across different forgery-related features. Inspired by this, we implement a targeted fine-tuning strategy, which has largely improved the explainability of the MLLMs, specifically capitalizing on their strengths. Furthermore, by integrating specific feature detectors (SFD), we design an adaptive fusion module to combine the complementary advantages of both MLLMs and conventional detectors for improved detection. + +Limitations and Future Work. While our framework demonstrates strong performance in detecting identity-specific facial forgeries, it has certain limitations. First, multimodal large language models (MLLMs) operate in a much larger parameter space, which leads to a richer visual feature space related to Deepfake detection. However, this comes at the cost of slower inference speeds. Future advancements in the field could help accelerate inference, benefiting from the rapid development of the domain. Second, our current implementation focuses solely on static image detection. However, real-world applications increasingly involve multimodal forgeries across video and audio streams. Extending our method to handle videos and audio-visual deepfakes is a critical next step for building a comprehensive and practical detection system. Third, As forgery technologies advance and realistic deepfakes exhibit fewer detectable artifacts that can be captured by natural language descriptions, our MLLM-based method may become less effective in interpreting semantic artifacts, thereby requiring the SFD module to play a more crucial role in the framework. + +Broader Impacts. This research advances machine learning with a new framework to detect and explain Deepfake images, effectively identifying deepfakes and reducing misuse of generative models for significant societal benefit. However, it risks being used to improve deepfake realism. To counter this, following previous works [75, 76] implement access controls. We will urge researchers to minimize harms while maximizing the positive impact of this work. + +Ethics & Reproducibility statements. All facial images used are from publicly available datasets with proper citations, ensuring no violation of personal privacy. And the human study has received the IRB approval. + +Content Structure of the Appendix. Due to page constraints, we include additional analyses and experiments in the Appendix. Specifically, the Appendix contains the following sections: Overview of Appendix, Experiment Setting Details, Additional Experimental Results, Human Study and GPT4 Evaluation, Additional Analysis of Model Feature Assessment, Additional Analysis of Strong Feature Strengthening, Additional Analysis of Weak Feature Supplementing, Additional Analysis of Ablation Study, Sample Showing. For further details, please refer to the Appendix. + +# Acknowledgments + +Baoyuan Wu was supported by Guangdong Basic and Applied Basic Research Foundation (No. 2024B1515020095), Guangdong Provincial Program (No. 2023TQ07A352), Shenzhen Science and Technology Program (No. JCYJ20240813113608011), and Longgang District Key Laboratory of Intelligent Digital Economy Security. + +# References + +[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. +[2] Peter Anderson, Basura Fernando, Mark Johnson, and Stephen Gould. Spice: Semantic propositional image caption evaluation. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part V 14, pages 382-398. Springer, 2016. +[3] Anthrop. Introducing the next generation of claude. https://www.anthropic.com/news/claude-3-family, 2024. 2024.03. +[4] Zechen Bai, Pichao Wang, Tianjun Xiao, Tong He, Zongbo Han, Zheng Zhang, and Mike Zheng Shou. Hallucination of multimodal large language models: A survey. arXiv preprint arXiv:2404.18930, 2024. +[5] Junyi Cao, Chao Ma, Taiping Yao, Shen Chen, Shouhong Ding, and Xiaokang Yang. End-to-end reconstruction-classification learning for face forgery detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4113-4122, 2022. +[6] Liang Chen, Yong Zhang, Yibing Song, Lingqiao Liu, and Jue Wang. Self-supervised learning of adversarial example: Towards good generalizations for deepfake detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18710-18719, 2022. +[7] Renwang Chen, Xuanhong Chen, Bingbing Ni, and Yanhao Ge. Simswap: An efficient framework for high fidelity face swapping. In Proceedings of the 28th ACM international conference on multimedia, pages 2003–2011, 2020. +[8] Ruoxin Chen, Junwei Xi, Zhiyuan Yan, Ke-Yue Zhang, Shuang Wu, Jingyi Xie, Xu Chen, Lei Xu, Isabel Guan, Taiping Yao, et al. Dual data alignment makes ai-generated image detector easier generalizable. arXiv preprint arXiv:2505.14359, 2025. +[9] Harry Cheng, Yangyang Guo, Tianyi Wang, Liqiang Nie, and Mohan Kankanhalli. Diffusion facial forgery detection. In Proceedings of the 32nd ACM international conference on multimedia, pages 5939-5948, 2024. +[10] Jikang Cheng, Zhiyuan Yan, Ying Zhang, Yuhao Luo, Zhongyuan Wang, and Chen Li. Can we leave deepfake data behind in training deepfake detector? arXiv preprint arXiv:2408.17052, 2024. +[11] Jongwook Choi, Taehoon Kim, Yonghyun Jeong, Seungryul Baek, and Jongwon Choi. Exploiting style latent flows for generalizing deepfake video detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1133-1143, 2024. +[12] François Chollet. Xception: Deep learning with depthwise separable convolutions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1251-1258, 2017. +[13] Deepfakedetection., 2021. https://ai.googleblog.com/2019/09/contributing-data-to-deepfakedetection.html Accessed 2021-11-13. +[14] Michael Denkowski and Alon Lavie. Meteor universal: Language specific translation evaluation for any target language. In Proceedings of the ninth workshop on statistical machine translation, pages 376-380, 2014. +[15] B Dolhansky. The dee pfake detection challenge (dfdc) pre view dataset. arXiv preprint arXiv:1910.08854, 2019. +[16] Brian Dolhansky, Joanna Bitton, Ben Pflaum, Jikuo Lu, Russ Howes, Menglin Wang, and Cristian Canton Ferrer. The deepfake detection challenge (dfdc) dataset. arXiv preprint arXiv:2006.07397, 2020. +[17] Shichao Dong, Jin Wang, Renhe Ji, Jiajun Liang, Haoqiang Fan, and Zheng Ge. Implicit identity leakage: The stumbling block to improving deepfake detection generalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3994-4004, 2023. +[18] Xiaoyi Dong, Jianmin Bao, Dongdong Chen, Ting Zhang, Weiming Zhang, Nenghai Yu, Dong Chen, Fang Wen, and Baining Guo. Protecting celebrities from deepfake with identity consistency transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9468-9478, 2022. +[19] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. + +[20] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. +[21] Niki M Foteinopoulou, Enjie Ghorbel, and Djamila Aouada. A hitchhiker's guide to fine-grained face forgery detection using common sense reasoning. Advances in Neural Information Processing Systems, 37:2943-2976, 2025. +[22] Xinghe Fu, Zhiyuan Yan, Taiping Yao, Shen Chen, and Xi Li. Exploring unbiased deepfake detection via token-level shuffling and mixing. arXiv preprint arXiv:2501.04376, 2025. +[23] Qiqi Gu, Shen Chen, Taiping Yao, Yang Chen, Shouhong Ding, and Ran Yi. Exploiting fine-grained face forgery clues via progressive enhancement learning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 735-743, 2022. +[24] Alexandros Haliassos, Konstantinos Vougioukas, Stavros Petridis, and Maja Pantic. Lips don't lie: A generalisable and robust approach to face forgery detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5039-5049, 2021. +[25] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. +[26] Baojin Huang, Zhongyuan Wang, Jifan Yang, Jiaxin Ai, Qin Zou, Qian Wang, and Dengpan Ye. Implicit identity driven deepfake face swapping detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4490-4499, 2023. +[27] Zhengchao Huang, Bin Xia, Zicheng Lin, Zhun Mou, and Wenming Yang. Ffaa: Multimodal large language model based explainable open-world face forgery analysis assistant. arXiv preprint arXiv:2408.10072, 2024. +[28] Shan Jia, Reilin Lyu, Kangran Zhao, Yize Chen, Zhiyuan Yan, Yan Ju, Chuanbo Hu, Xin Li, Baoyuan Wu, and Siwei Lyu. Can chatgpt detect deepfakes? a study of using multimodal large language models for media forensics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4324-4333, 2024. +[29] Liming Jiang, Ren Li, Wayne Wu, Chen Qian, and Chen Change Loy. Deeperforensics-1.0: A large-scale dataset for real-world face forgery detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020. +[30] Nicolas Larue, Ngoc-Son Vu, Vitomir Struc, Peter Peer, and Vassilis Christophides. Seeable: Soft discrepancies and bounded contrastive learning for exposing deepfakes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21011-21021, 2023. +[31] Jiawei Li, Fanrui Zhang, Jiaying Zhu, Esther Sun, Qiang Zhang, and Zheng-Jun Zha. Forgerygpt: Multimodal large language model for explainable image forgery detection and localization. arXiv preprint arXiv:2410.10238, 2024. +[32] Lingzhi Li, Jianmin Bao, Ting Zhang, Hao Yang, Dong Chen, Fang Wen, and Baining Guo. Face x-ray for more general face forgery detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020. +[33] Maomao Li, Ge Yuan, Cairong Wang, Zhian Liu, Yong Zhang, Yongwei Nie, Jue Wang, and Dong Xu. E4s: Fine-grained face swapping via editing with regional gan inversion. arXiv preprint arXiv:2310.15081, 2023. +[34] Y Li. Exposing deepfake videos by detecting face warping artifacts. arXiv preprint arXiv:1811.00656, 2018. +[35] Yixuan Li, Xuelin Liu, Xiaoyang Wang, Shiqi Wang, and Weisi Lin. Fakebench: Uncover the achilles' heels of fake images with large multimodal models. arXiv preprint arXiv:2404.13306, 2024. +[36] Yuezun Li, Ming-Ching Chang, and Siwei Lyu. In ictu oculi: Exposing ai created fake videos by detecting eye blinking. In 2018 IEEE International workshop on information forensics and security (WIFS), pages 1-7. Ieee, 2018. +[37] Yuezun Li, Xin Yang, Pu Sun, Honggang Qi, and Siwei Lyu. Celeb-df: A new dataset for deepfake forensics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020. + +[38] Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, pages 74-81, 2004. +[39] Kaiqing Lin, Yuzhen Lin, Weixiang Li, Taiping Yao, and Bin Li. Standing on the shoulders of giants: Reprogramming visual-language model for general deepfake detection. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, pages 5262-5270, 2025. +[40] Kaiqing Lin, Zhiyuan Yan, Ke-Yue Zhang, Li Hao, Yue Zhou, Yuzhen Lin, Weixiang Li, Taiping Yao, Shouhong Ding, and Bin Li. Guard me if you know me: Protecting specific face-identity from deepfakes. arXiv preprint arXiv:2505.19582, 2025. +[41] Yuzhen Lin, Wentang Song, Bin Li, Yuezun Li, Jiangqun Ni, Han Chen, and Qiushi Li. Fake it till you make it: Curricular dynamic forgery augmentations towards general deepfake detection. In European Conference on Computer Vision, pages 104-122. Springer, 2024. +[42] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26296-26306, 2024. +[43] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava-last: Improved reasoning,OCR, and world knowledge, 2024. +[44] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. +[45] Honggu Liu, Xiaodan Li, Wenbo Zhou, Yuefeng Chen, Yuan He, Hui Xue, Weiming Zhang, and Nenghai Yu. Spatial-phase shallow learning: rethinking face forgery detection in frequency domain. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 772-781, 2021. +[46] Honggu Liu, Xiaodan Li, Wenbo Zhou, Yuefeng Chen, Yuan He, Hui Xue, Weiming Zhang, and Nenghai Yu. Spatial-phase shallow learning: rethinking face forgery detection in frequency domain. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021. +[47] Anwei Luo, Chenqi Kong, Jiwu Huang, Yongjian Hu, Xiangui Kang, and Alex C Kot. Beyond the prior forgery knowledge: Mining critical clues for general face forgery detection. IEEE Transactions on Information Forensics and Security, 19:1168-1182, 2023. +[48] Yuchen Luo, Yong Zhang, Junchi Yan, and Wei Liu. Generalizing face forgery detection with high-frequency features. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16317-16326, 2021. +[49] Dat Nguyen, Nesryne Mejri, Inder Pal Singh, Polina Kuleshova, Marcella Astrid, Anis Kacem, Enjie Ghorbel, and Djamila Aouada. Laa-net: Localized artifact attention network for quality-agnostic and generalizable deepfake detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17395–17405, 2024. +[50] Yunsheng Ni, Depu Meng, Changqian Yu, Chengbin Quan, Dongchun Ren, and Youjian Zhao. Core: Consistent representation learning for face forgery detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12-21, 2022. +[51] Yuval Nirkin, Yosi Keller, and Tal Hassner. Fsgan: Subject agnostic face swapping and reenactment. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7184-7193, 2019. +[52] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics, pages 311-318, 2002. +[53] Gan Pei, Jiangning Zhang, Menghan Hu, Zhenyu Zhang, Chengjie Wang, Yunsheng Wu, Guangtao Zhai, Jian Yang, Chunhua Shen, and Dacheng Tao. Deepfake generation and detection: A benchmark and survey. arXiv preprint arXiv:2403.17881, 2024. +[54] Yuyang Qian, Guojun Yin, Lu Sheng, Zixuan Chen, and Jing Shao. Thinking in frequency: Face forgery detection by mining frequency-aware clues. In European conference on computer vision, pages 86-103. Springer, 2020. +[55] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. + +[56] Anirudh Sundara Rajan, Utkarsh Ojha, Jedidiah Schroesser, and Yong Jae Lee. Aligned datasets improve detection of latent diffusion-generated images. In The Thirteenth International Conference on Learning Representations, 2025. +[57] Felix Rosberg, Eren Erdal Aksoy, Fernando Alonso-Fernandez, and Cristofer Englund. Facedancer: Pose-and occlusion-aware high fidelity face swapping. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 3454-3463, 2023. +[58] Andreas Rossler, Davide Cozzolino, Luisa Verdoliva, Christian Riess, Justus Thies, and Matthias Nießner. Faceforensics++: Learning to detect manipulated facial images. In Proceedings of the IEEE/CVF Conference on International Conference on Computer Vision, 2019. +[59] Somdev Sangwan. Roop. https://github.com/s0md3v/roop, 2020. [GitHub repository]. +[60] Yichen Shi, Yuhao Gao, Yingxin Lai, Hongyang Wang, Jun Feng, Lei He, Jun Wan, Changsheng Chen, Zitong Yu, and Xiaochun Cao. Shield: An evaluation benchmark for face spoofing and forgery detection with multimodal large language models. arXiv preprint arXiv:2402.04178, 2024. +[61] Kaede Shiohara and Toshihiko Yamasaki. Detecting deepfakes with self-blended images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18720-18729, 2022. +[62] Kaede Shiohara, Xingchao Yang, and Takafumi Taketomi. Blendface: Re-designing identity encoders for face-swapping. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7634-7644, 2023. +[63] Ke Sun, Shen Chen, Taiping Yao, Ziyin Zhou, Jiayi Ji, Xiaoshuai Sun, Chia-Wen Lin, and Rongrong Ji. Towards general visual-linguistic face forgery detection (v2). arXiv preprint arXiv:2502.20698, 2025. +[64] Ke Sun, Taiping Yao, Shen Chen, Shouhong Ding, Jilin Li, and Rongrong Ji. Dual contrastive learning for general face forgery detection. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 2316-2324, 2022. +[65] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In Proceedings of the International Conference on Machine Learning, pages 6105-6114. PMLR, 2019. +[66] Ramakrishna Vedantam, C Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4566-4575, 2015. +[67] Chengrui Wang and Weihong Deng. Representative forgery mining for fake face detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14923-14932, 2021. +[68] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. +[69] Zhendong Wang, Jianmin Bao, Wengang Zhou, Weilun Wang, and Houqiang Li. Altfreezing for more general video face forgery detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4129-4138, 2023. +[70] Siwei Wen, Junyan Ye, Peilin Feng, Hengrui Kang, Zichen Wen, Yize Chen, Jiang Wu, Wenjun Wu, Conghui He, and Weijia Li. Spot the fake: Large multimodal model-based synthetic image detection with artifact explanation. arXiv preprint arXiv:2503.14905, 2025. +[71] Jiayang Wu, Wensheng Gan, Zefeng Chen, Shicheng Wan, and Philip S Yu. Multimodal large language models: A survey. In 2023 IEEE International Conference on Big Data (BigData), pages 2247-2256. IEEE, 2023. +[72] Yuting Xu, Jian Liang, Gengyun Jia, Ziming Yang, Yanhao Zhang, and Ran He. Tall: Thumbnail layout for deepfake video detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 22658-22668, 2023. +[73] Zhipei Xu, Xuanyu Zhang, Runyi Li, Zecheng Tang, Qing Huang, and Jian Zhang. Fakeshield: Explainable image forgery detection and localization via multi-modal large language models. arXiv preprint arXiv:2410.02761, 2024. +[74] Zhiyuan Yan, Yuhao Luo, Siwei Lyu, Qingshan Liu, and Baoyuan Wu. Transcending forgery specificity with latent space augmentation for generalizable deepfake detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8984-8994, 2024. + +[75] Zhiyuan Yan, Jiangming Wang, Zhendong Wang, Peng Jin, Ke-Yue Zhang, Shen Chen, Taiping Yao, Shouhong Ding, Baoyuan Wu, and Li Yuan. Effort: Efficient orthogonal modeling for generalizable ai-generated image detection. arXiv preprint arXiv:2411.15633, 2024. +[76] Zhiyuan Yan, Taiping Yao, Shen Chen, Yandan Zhao, Xinghe Fu, Junwei Zhu, Donghao Luo, Chengjie Wang, Shouhong Ding, Yunsheng Wu, et al. Df40: Toward next-generation deepfake detection. arXiv preprint arXiv:2406.13495, 2024. +[77] Zhiyuan Yan, Junyan Ye, Weijia Li, Zilong Huang, Shenghai Yuan, Xiangyang He, Kaiqing Lin, Jun He, Conghui He, and Li Yuan. Gpt-imgeval: A comprehensive benchmark for diagnosing gpt4o in image generation. arXiv preprint arXiv:2504.02782, 2025. +[78] Zhiyuan Yan, Yong Zhang, Yanbo Fan, and Baoyuan Wu. Ucf: Uncovering common features for generalizable deepfake detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22412-22423, 2023. +[79] Zhiyuan Yan, Yong Zhang, Xinhang Yuan, Siwei Lyu, and Baoyuan Wu. Deepfakebench: A comprehensive benchmark of deepfake detection. arXiv preprint arXiv:2307.01426, 2023. +[80] Zhiyuan Yan, Yandan Zhao, Shen Chen, Mingyi Guo, Xinghe Fu, Taiping Yao, Shouhong Ding, and Li Yuan. Generalizing deepfake video detection with plug-and-play: Video-level blending and spatiotemporal adapter tuning. arXiv preprint arXiv:2408.17065, 2024. +[81] Tianyun Yang, Juan Cao, Qiang Sheng, Lei Li, Jiaqi Ji, Xirong Li, and Sheng Tang. Learning to disentangle gan fingerprint for fake image attribution. arXiv preprint arXiv:2106.08749, 2021. +[82] Xin Yang, Yuezun Li, and Siwei Lyu. Exposing deep fakes using inconsistent head poses. In ICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 8261-8265. IEEE, 2019. +[83] Zheng Yang, Ruoxin Chen, Zhiyuan Yan, Ke-Yue Zhang, Xinghe Fu, Shuang Wu, Xiujun Shu, Taiping Yao, Shouhong Ding, and Xi Li. All patches matter, more patches better: Enhance ai-generated image detection via panoptic patch learning. arXiv preprint arXiv:2504.01396, 2025. +[84] Junyan Ye, Baichuan Zhou, Zilong Huang, Junan Zhang, Tianyi Bai, Hengrui Kang, Jun He, Honglin Lin, Zihao Wang, Tong Wu, et al. Loki: A comprehensive synthetic data detection benchmark using large multimodal models. arXiv preprint arXiv:2410.09732, 2024. +[85] Peipeng Yu, Jianwei Fei, Hui Gao, Xuan Feng, Zhihua Xia, and Chip Hong Chang. Unlocking the capabilities of vision-language models for generalizable and explainable deepfake detection. arXiv preprint arXiv:2503.14853, 2025. +[86] Daichi Zhang, Zihao Xiao, Shikun Li, Fanzhao Lin, Jianmin Li, and Shiming Ge. Learning natural consistency representation for face forgery video detection. In European Conference on Computer Vision, pages 407-424. Springer, 2024. +[87] Yue Zhang, Ben Colman, Xiao Guo, Ali Shahriyari, and Gaurav Bharaj. Common sense reasoning for deepfake detection. In European Conference on Computer Vision, pages 399-415. Springer, 2024. +[88] Hanqing Zhao, Wenbo Zhou, Dongdong Chen, Tianyi Wei, Weiming Zhang, and Nenghai Yu. Multi-attentional deepfake detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2185–2194, 2021. +[89] Tianchen Zhao, Xiang Xu, Mingze Xu, Hui Ding, Yuanjun Xiong, and Wei Xia. Learning self-consistency for deepfake detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 15023-15033, 2021. +[90] Yinglin Zheng, Jianmin Bao, Dong Chen, Ming Zeng, and Fang Wen. Exploring temporal coherence for more general video face forgery detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 15044-15054, 2021. +[91] Jiancan Zhou, Xi Jia, Qiufu Li, Linlin Shen, and Jinming Duan. Uniface: Unified cross-entropy loss for deep face recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20730–20739, 2023. +[92] Tianfei Zhou, Wenguan Wang, Zhiyuan Liang, and Jianbing Shen. Face forensics in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5778-5788, 2021. + +[93] Ziyin Zhou, Yunpeng Luo, Yuanchen Wu, Ke Sun, Jiayi Ji, Ke Yan, Shouhong Ding, Xiaoshuai Sun, Yunsheng Wu, and Rongrong Ji. Aigi-holmes: Towards explainable and generalizable ai-generated image detection via multimodal large language models. arXiv preprint arXiv:2507.02664, 2025. +[94] Wanyi Zhuang, Qi Chu, Zhentao Tan, Qiankun Liu, Haojie Yuan, Changtao Miao, Zixiang Luo, and Nenghai Yu. Uia-vit: Unsupervised inconsistency-aware method based on vision transformer for face forgery detection. In European conference on computer vision, pages 391-407. Springer, 2022. +[95] Bojia Zi, Minghao Chang, Jingjing Chen, Xingjun Ma, and Yu-Gang Jiang. Wilddeepfake: A challenging real-world dataset for deepfake detection. In Proceedings of the 28th ACM International Conference on Multimedia, pages 2382-2390, 2020. + +# A Overview + +Due to space constraints, we have included additional important content in the supplementary materials. Below is a brief outline of the supplementary content to facilitate readers easily locate the relevant sections: + +- Appendix B: Experiment Setting Details +- Appendix B.1: Details of Datasets +- Appendix B.2: Implementation Details + +- Appendix C: Additional Experimental Results + +- Appendix C.1: Results of Robustness Against Unseen Perturbations +- Appendix C.2: Results of Training and In-domain Testing on FF++ +- Appendix C.3: Results of Cross-manipulation Evaluation on DF40 +- Appendix C.4: Results of Experiments on Different LLMs/MLLMs +- Appendix C.5: Results of Experiments on GenAI Images + +- Appendix D: Human Study and GPT4 Evaluation + +- Appendix D.1: Human Study Details +- Appendix D.2: GPT-4 Evaluation Details + +- Appendix E: Additional Analysis of Model Feature Assessment + +- Appendix E.1: Evaluation Setup of MFA +- Appendix E.2: Evaluation of the Overall Detection Performance Details +- Appendix E.3: Deeply Investigation of Individual Feature's Discrimination + +- Appendix F: Additional Analysis of Strong Feature Strengthening +- Appendix G: Additional Analysis of Weak Feature Supplementing +- Appendix H: Additional Analysis of Ablation Study +- Appendix I: Sample Showing + +# B Experiment Setting Details + +# B.1 Details of Datasets + +VQA Datasets Construction Details In this part, we detail the data construction process for our deepfake detection framework, $\mathcal{X}^2$ -DFD, which integrates multiple stages: Multi-Feature Analysis (MFA), Strong Feature Selection (SFS), Weak Feature Selection (WFS), and Model Inference (MI). The process, illustrated in Figure 4, leverages multimodal large language models (MLLMs) and specific feature detectors (SFD) to construct a robust dataset for training and fine-tuning. + +The data construction pipeline begins with the MFA stage, where an MLLM generates yes-or-no questions (e.g., "Is the image blurry?" or "Are there blending artifacts?") to identify deepfake characteristics in a given dataset. These questions are evaluated by the MLLM, producing a feature score list that ranks features by their relevance to deepfake detection. In the SFS stage, the top-K features are selected to generate real and fake prompts, such as providing reasoning for why an image might be deepfake or real. These prompts are paired with the deepfake dataset to construct a fine-tuning dataset, where the MLLM answers questions like "Is this image real or fake?" to generate labeled data. The WFS stage further refines this dataset by incorporating an External Dedicated Detector (SFD), which assigns confidence scores (e.g., SFD score = 0.901 for fake samples) to filter out unreliable samples, ensuring the dataset's quality for fine-tuning. Finally, in the MI stage, the fine-tuned MLLM, now strengthened as $\mathcal{X}^2$ -DFD, performs inference on new samples, producing accurate deepfake detection results (e.g., "This image is fake" with detailed reasoning about blending artifacts and omitted features). This multi-stage pipeline ensures that the constructed dataset is both comprehensive and reliable, enabling the $\mathcal{X}^2$ -DFD model to achieve robust performance across various deepfake detection scenarios. + +Training and Testing Image Datasets We trained our model using the FF++ dataset [58]. For preprocessing and cropping, we adopted the methods from DeepfakeBench [79]. We utilized 8 frames per video for training and 32 frames per video for testing. + +![](images/a61ae174cedb2a0f494c264cff63f68f8ad9283b74b7127db90d32104ef88d2c.jpg) +Figure 4: Overview of the data construction pipeline for the $\mathcal{X}^2$ -DFD framework. The pipeline consists of four stages: (1) Multi-Feature Analysis (MFA) for generating yes-or-no questions to identify deepfake characteristics, resulting in a feature list; (2) Strong Feature Selection (SFS) for creating real and fake prompts based on top-K features, followed by dataset construction using MLLM-generated answers; (3) Weak Feature Selection (WFS) for fine-tuning the dataset with SFD scores, filtering samples based on reliability; and (4) Model Inference (MI) for final evaluation. + +# B.2 Implementation Details + +The data splits and proposing follow the DeepfakeBench [79], the with a learning rate of $2 \times 10^{-5}$ in two layer mlp projector and $2 \times 10^{-4}$ in others which is borrow from the official implementation of LLava [43], a rank of 16, and an alpha value set conventionally to twice the rank at 32. We training for three epochs, for each epoch on NVIDIA 4090 (Driver Version: 535.247.01; CUDA Version: 12.2), AMD 32-Corecost for 4 hours by training on FF++ [58], each video we take 8 frames for training. AUC is calculated by directly obtaining the token probabilities. Previous AUC calculations for large models mostly relied on averaging methods, such as in [28], but this approach is not very accurate because: (1) multiple samplings are needed to approximate the true probability distribution, and (2) large models inherently perform inference with a default temperature, which itself involves sampling over probabilities. Averaging over multiple samples effectively results in a second layer of sampling, making the evaluation less accurate. Therefore, in this paper, we calculate AUC by directly obtaining token probabilities. + +# C Additional Experimental Results + +# C.1 Results of Robustness Against Unseen Perturbations + +To evaluate the robustness of our model to random perturbations, we follow the methodology outlined in previous studies [24, 90], which examines four types of degradation: Gaussian blur, block-wise distortion, contrast changes, and JPEG compression. Each perturbation is applied at five different levels to assess the model's performance under varying degrees of distortion. + +To highlight the advantages of our approach over conventional detectors like FWA [34], SBI [61], and X-ray [32], we conducted multiple evaluations. As illustrated in Figure 5, which shows the video-level AUC results for these unseen perturbations using a model trained on $\mathrm{FF}++$ c23, our method consistently demonstrates superior robustness compared to other RGB-based methods. + +![](images/a2212a2b8eabbecea066371edcdd8ab81786f9dd04ba3a04e6220b0e11890693.jpg) +Figure 5: Robustness evaluation. We adopt four types of degradation for examining the robustness of our model: Gaussian blur, block-wise distortion, contrast changes, and JPEG compression. Our model shows superior robustness over other compared models. + +# C.2 Results of Training and In-domain Testing on $\mathbf{FF} + +$ + +In our manuscript, we mainly focus on the cross-domain evaluation to assess the generalization performance of different models. Here, we conduct the in-domain evaluation on the FF++ dataset and compare our approach with the other four SOTA methods: FWA, Face X-ray, SRM, and CDFA. Following DeepfakeBench [79], we train all models on FF++ (c23) and test them on FF++ (c23), FF++ (c40), FF-DF, FF-F2F, FF-FS, and FF-NT. As shown in Table 9, the in-domain results demonstrate that our framework achieves the best performance, outperforming all other methods. + +Table 9: In-domain results in the FF++ dataset (AUC) + +
DetectorFF++c23FF++c40FF-DFFF-F2FFF-FSFF-NTAVG
FWA [34]87.773.692.190.088.481.285.5
Face X-ray [32]95.979.397.998.798.792.993.9
SRM [48]95.881.197.397.097.493.093.6
CDFA [41]90.269.099.986.993.380.790.2
Ours96.682.699.997.298.191.094.2
+ +# C.3 Results of Cross-manipulation Evaluation on DF40 + +Evaluating our model's performance on cross-manipulation tasks helps assess whether it can handle previously unseen fake types. We use the recently released DF40 dataset [76] for evaluation. Our method generally outperforms other models on average, particularly the e4s, Inswap, and SimSwap methods (see Table 10). This shows that our method effectively learns more generalizable features for detection, even against the latest techniques. + +# C.4 Results of Experiments on Different LLMs/MLLMs + +We conducted experiments using various models, including GPT4o [1], Claude3.5-Sonnet [3], and LLaVa [44], to evaluate the adaptability and robustness of our framework. + +Table 10: Cross-manipulation evaluations within the FF++ domain (frame-level AUC only). We leverage the DF40 dataset [76] and select six representative face-swapping methods generated within the FF++ domain, keeping the data domain unchanged. The top two results are highlighted, with the best result in **bold** and the second-best **underlined**. + +
MethodVenuesunifacee4sfacedancerfsganinswapsimswapAvg.
RECCE [5]CVPR 202284.265.278.388.479.573.078.1
SBI [61]CVPR 202264.469.066.787.963.356.868.0
CORE [50]CVPRW 202281.763.471.791.179.469.376.1
IID [26]CVPR 202379.571.079.086.474.464.075.7
UCF [78]ICCV 202378.769.280.088.176.864.977.5
LSDA [74]CVPR 202485.468.475.983.281.072.777.8
CDFA [41]ECCV 202476.567.475.484.872.076.175.9
ProDet [10]NIPS 202484.571.073.686.578.877.878.7
Ours-85.590.382.691.181.285.185.9
+ +Different LLMs to Generate Questions in MFA. In the MFA stage, we employed different LLMs, such as GPT4o and Claude 3.5-Sonnet, to generate forgery-related questions and test the adaptability of our framework. The results, shown in $GPT4o + LLaVa - 7B$ and Claude 3.5-Sonnet + LLaVa-7B, demonstrate consistent performance regardless of the LLM used. Questions from Claude 3.5-Sonnet were also effective (see Table 21 and 22). + +Table 11: Experiments on different LLMs/MLLMs were conducted to evaluate their performance under various conditions. The evaluation metric used for these experiments is the Area Under the Curve (AUC) + +
VariantCDFDFDCPDFDCDFDUnifacee4sFacedancerFSGANInswapSimswapAvg
Human Writing + LLaVa-7B [43]89.189.783.592.583.386.182.589.178.787.086.2
GPT4o[1] + LLaVa-7B [43]90.389.783.592.585.291.283.889.978.584.987.0
GPT4o[1] + Qwen2VL-7B [68]89.890.382.993.384.991.281.690.179.984.086.8
Claude3.5-Sonnet[3] + LLaVa-7B [43]90.188.583.593.084.990.683.890.080.085.687.0
GPT4o[1] + LLaVa-13B [43]91.390.383.492.586.092.584.591.080.685.487.8
+ +Different MLLMs for Fine-tuning in SFS and WFS. In the SFS and WFS stages, we investigate the impact of using different sizes of MLLMs with the same architecture during fine-tuning. For instance, we compare $GPT4o + LLaVa - 7B$ and $GPT4o + LLaVa - 13B$ . The results indicate that as the size of the model increases, the performance of the framework in Table 11 improves proportionally, benefiting from the enhanced capabilities of the larger MLLMs. Additionally, we tested Qwen2VL-7B, which yielded performance comparable to $LLaVa - 7B$ . + +# Comparison of Human-generated and Model-generated questions. + +We also compared the effectiveness of human-written questions with those generated by advanced LLMs, such as GPT4o and Claude 3.5-Sonnet, in the MFA stage. The results on Table 11, exemplified by Human Writing + LLaVa-7B (Avg AUC 86.2) versus GPT4o + LLaVa-7B (Avg AUC 87.0) and Claude 3.5-Sonnet + LLaVa-7B (Avg AUC 87.0), reveal that human-generated questions perform comparably to model-generated ones. Notably, questions produced by LLMs often exhibit greater competitiveness, leveraging their ability to generate diverse and nuanced forgery-related prompts, which further enhances the framework's adaptability and performance. + +Summary of Findings. These experiments collectively highlight the robustness of our framework. It is not dependent on specific LLMs or MLLMs, making it adaptable to a wide range of models. Furthermore, as the performance and size of the underlying models improve, our framework effectively leverages these advancements to achieve enhanced results. + +# C.5 Results of Experiments on GenAI Images + +In addition to face swap forgery, we further explored the effectiveness of our framework on GenAI Image. Please refer to the supplementary material for more details and examples. + +# D Human Study and GPT4 Evaluation + +# D.1 Human Study Details + +The human study was designed to evaluate the explainability of our deepfake detection model by involving human evaluators in assessing three key dimensions: detection ability, reasonableness of explanations, and level of detail. Below are the key details of the experimental setup: + +Participant Recruitment. We recruited 15 well-educated participants and provided them with a detailed guideline to ensure a clear understanding of the experimental task. Participants were aged between 20 and 40. + +Task Description. Participants were presented with a set of 100 samples, each consisting of a deepfake image, the model's detection output ( $\mathcal{X}^2$ -DFD), and an associated explanation generated by a Multimodal Large Language Model (MLLM). They were tasked with scoring the model's performance across three dimensions on a scale from 0 to 5: + +- Detection Ability: How accurately does the model identify the deepfake? (0 = completely incorrect, 5 = perfectly accurate) +- Reasonableness of Explanations: How logical and understandable is the explanation? (0 = completely unreasonable, 5 = highly reasonable) +- Level of Detail: How detailed and specific is the explanation? (0 = no detail, 5 = very detailed) + +Study Procedure. The study was conducted using a custom-built GUI. Participants completed an initial 10-minute training session to familiarize themselves with the task and scoring criteria. Each participant evaluated random 50 samples, ensuring diverse coverage of the dataset. The study took approximately 20 minutes per participant, and participants volunteered their time without compensation. + +Dataset. All samples were sourced from the testing datasets of DD-VQA. Each sample included both the raw image and the model-generated explanation, with a focus on unannotated regions. + +Evaluation Metrics. Scores for each dimension were averaged across all participants. Additionally, participants provided a final overall score for each sample on a scale from 0 to 5, categorized as follows: 0 (very poor), 1 (poor), 2 (fair), 3 (good), 4 (very good), and 5 (excellent). + +GUI of Human Study. he human study was conducted using a custom-designed Graphical User Interface (GUI) to facilitate efficient and user-friendly evaluation. The GUI was developed using Python with the flask library and hosted on a server. Below are the details of the GUI design: The GUI consisted of three main panels (see Figures 6 to 9 for screenshots): + +- Image Panel: Displayed the deepfake image on the left side of the screen, with zoom and pan functionality for detailed inspection. +- Explanation Panel: Presented the model's detection output $(\mathcal{X}^2$ -DFD) and other tested MLLLM's explanations on the right side. +Scoring Panel: Provided sliders for each answer allowing participants to select a score from 0 to 5. + +User Interaction. Participants could navigate between samples using "Next" and "Previous" buttons. A progress bar indicated the number of samples completed. The GUI automatically saved scores to a database after each submission. + +# D.2 GPT-4 Evaluation Details + +In addition to the human study, we conducted evaluations using GPT-4o to comprehensively assess the unannotated explainability of our deepfake detection model. This subsection provides the + +# Answer Evaluation + +Background: Deepfakes are synthetic media, typically videos or images, created using advanced artificial intelligence techniques like deep learning to manipulate or generate realistic-looking content, often by superimposing one person's face onto another's, raising concerns about misinformation and privacy. + +![](images/a7f7e2192a861c93830c388133ea66cdfb5df8589edc4fd91ee265b9f9cddd3e.jpg) + +# Ground Truth: fake + +Description: The image is fake because it is a manipulated or digitally altered photo of a woman with a nose that is too large for her face. This distortion is evident in the facial features and the overall appearance of the image. The alteration creates an unnatural and artificial look, which is not representative of a real person. This manipulation is typically done to create a specific visual effect or to exaggerate certain features, but in this case, it leads to an unnatural and unconvincing representation of the woman's face. + +Please evaluate the model's response based on three key dimensions: + +1. Detection Accuracy: Did the model answer the same as the Ground truth? +2. Reasonableness of Explanations: How plausible and well-founded are the model's explanations? +3. Level of Detail: Does the model provide sufficient and appropriate detail in its analysis? + +Based on the above criteria, please provide your final assessment of the model's response. + +Final Score: + +![](images/dd208c954f220831cfc9e158025ddc3611d771f05a74426e13438d5bcc6f620e.jpg) +Figure 6: Human evaluation + +Pr + +Next + +# Answer Evaluation + +Background: Deepfakes are synthetic media, typically videos or images, created using advanced artificial intelligence techniques like deep learning to manipulate or generate realistic-looking content, often by superimposing one person's face onto another's, raising concerns about misinformation and privacy. + +![](images/eeede118fb2f805e88242470a436f70a433ee162738159a87fb6e6d495fcb3dc.jpg) + +# Ground Truth: real + +Description: The image is real because it shows a man with a beard and mustache, wearing a white shirt, and making a surprised or shocked expression. The man's facial features, such as his eyes, nose, and mouth, appear to be in their natural state, and there are no noticeable glitches or artifacts in the image. The man's expression and the overall composition of the image suggest that it is a genuine, real-life moment captured in the photo. + +Please evaluate the model's response based on three key dimensions: + +1. Detection Accuracy: Did the model answer the same as the Ground truth? +2. Reasonableness of Explanations: How plausible and well-founded are the model's explanations? +3. Level of Detail: Does the model provide sufficient and appropriate detail in its analysis? + +Based on the above criteria, please provide your final assessment of the model's response. + +Final Score: + +![](images/28a998e327c2e9eeebeb2bbdb29381a33ade48ea4cb2333655ace57ea1cdeb40.jpg) +Figure 7: Human evaluation + +Pr + +Next + +implementation details of the automated evaluation process using GPT-4o, including the code structure, prompt formatting, and result processing. + +The evaluation process was implemented in Python, leveraging the OpenAI API to interact with GPT-4o. The script performs the following key steps: + +Data Loading. The script loads a JSON file containing metadata for the evaluation samples, including image IDs, ground truth labels (e.g., "real" or "fake"), and model-generated descriptions. + +![](images/ea8a0c213b365d4b8c87942ff5d6d53de0d87e8997d20ee8cd506d80d5d4fa6b.jpg) +Ground Truth: fake + +Description: The image is fake because it features a woman with a distorted or altered nose, which is an unnatural facial feature in the context of the image. Additionally, the woman's eyes are described as having an orange tint, which is not a natural eye color. These factors suggest that the image has been digitally manipulated or altered, resulting in a fake or artificial appearance. + +Please evaluate the model's response based on three key dimensions: + +1. Detection Accuracy: Did the model answer the same as the Ground truth? +2. Reasonableness of Explanations: How plausible and well-founded are the model's explanations? +3. Level of Detail: Does the model provide sufficient and appropriate detail in its analysis? + +Based on the above criteria, please provide your final assessment of the model's response. + +Final Score: + +![](images/b597e8c04853ea4204fd177c6fe73364ad1822a97ab4ea179db39ca8804de060.jpg) +Figure 8: Human evaluation + +Previous + +Next + +Background: Deepfakes are synthetic media, typically videos or images, created using advanced artificial intelligence techniques like deep learning to manipulate or generate realistic-looking content, often by superimposing one person's face onto another's, raising concerns about misinformation and privacy. + +![](images/b4d4d26a88de1c29bccd52b1d920d145d61c0a7a20d00074189e0c28f70c69d1.jpg) +Ground Truth: real + +Description: The image is real, as it shows a woman with glasses on television. The woman is wearing glasses and appears to be talking or making a statement. The presence of the television in the background further supports the authenticity of the scene. + +Please evaluate the model's response based on three key dimensions: + +1. Detection Accuracy: Did the model answer the same as the Ground truth? +2. Reasonableness of Explanations: How plausible and well-founded are the model's explanations? +3. Level of Detail: Does the model provide sufficient and appropriate detail in its analysis? + +Based on the above criteria, please provide your final assessment of the model's response. + +Final Score: + +![](images/2b1fcf0c0089fe8807822d7937a2a9c985c729375be3f3e3c4a60c1878b0b6e0.jpg) +Figure 9: Human evaluation + +Previous + +Next + +Prompt Formatting. For each sample, the script formats an evaluation prompt that includes the ground truth label and the model's description (the blue guideline as shown in Figure 6). The prompt instructs GPT-4o to evaluate the model's response across three dimensions—detection accuracy, reasonableness of explanations, and level of detail—and to provide scores on a 0–5 scale ( $0 = \text{Very Poor}, 5 = \text{Excellent}$ ). Additionally, a final overall score is requested. + +Image Encoding. The script encodes each image into Base64 format to enable GPT-4o to process both the textual prompt and the visual content. Images are sourced from the directory + +Result Processing and Storage. The script processes all 100 samples, storing the results in a list. Each result entry includes the image ID, image path, ground truth, prompt, and GPT-4o's response. The final output JSON file provides a comprehensive record of GPT-4o's evaluations, which are later aggregated and compared with human scores. + +![](images/d5e15697f32473cc9cafcb8360850c2c005f186c5dd8a7141754d23ae975dc2e.jpg) +Ground Truth: real + +# Model Descriptions + +GPT40 + +Fake, [lips, unnatural smoothness; eyes, inconsistent shading]. + +Llama-3.2-11B-Vision + +This image is a still from a video, and it shows a blonde woman with a shocked expression. The image is of low quality and is distorted, which suggests that it was taken from a movie. The woman's expression is exaggerated, which is typical of a news anchor or actress. The background is blurry, which is also consistent with a video still. Overall, + +llava-v1.5-7b + +The image appears to be digitally altered or manipulated. The exaggerated features and the unnatural lighting suggest that this is not a real photograph but rather an edited image. + +Qwen2-VL-7B-Instruct + +The image appears to be digitally altered or manipulated. The exaggerated features and the unnatural lighting suggest that this is not a real photograph but rather an edited image. + +OURS + +The image is real because it shows a woman with blonde hair and a pink lipstick smile. The woman appears to be a news reporter, and her facial expression and appearance are consistent with a real person. There are no noticeable glitches, errors, or distortions in the image that would indicate it is not real. + +![](images/98c718fc470e038a5a430c9d8d698a17fe585aed3f8b5a680379c1fdb6b10cc4.jpg) +Figure 10: Answer from different model +Ground Truth: fake +Figure 11: Answer from different model + +# Model Descriptions + +GPT40 + +Real. + +Llama-3.2-11B-Vision + +This is a photo of a man, likely a news anchor or politician. The image is blurry, which suggests that it is a cropped image taken from a video, and there is a blue background behind him, which is common for news broadcasts. The image quality and context suggest that it is likely a screenshot from a video, rather than a photograph. The image appears to be still from a video, + +llava-v1.5-7b + +The image appears to be a screenshot from a television show or a video. The background and lighting suggest it is taken from a broadcast or recording. + +Qwen2-VL-7B-Instruct + +The image appears to be a screenshot from a television show or a video. The background and lighting suggest it is taken from a broadcast or recording. + +OURS + +The image is fake because it has been digitally manipulated to create an unnatural and artificial appearance. The man's face has been altered, with the skin appearing smooth and unnatural, and the facial features are out of proportion. The image also has a blurry background, which further emphasizes the artificial nature of the photo. These elements suggest that the image has been digitally manipulated to create a fake or altered appearance, rather than being a genuine photograph. + +![](images/1b7aeab5705672d5655f0c794edc32976741563095885af8c3781cbd9035ac8a.jpg) +Ground Truth: real + +# Model Descriptions + +GPT40 + +Fake [eyes, uneven blinking] + +Llama-3.2-11B-Vision + +The image is blurry and poorly lit, and the person in the image appears to be sleeping, which suggests that the image may have been taken in low light or with a flash, which can cause blurriness. Additionally, the image quality is not high, which could indicate that it was taken with a lower-quality camera or device. Overall, the image appears to be a candid shot of someone sleeping, + +llava-v1.5-7b + +The image appears to be real. There are no obvious signs of digital manipulation or distortion that would suggest it is fake. + +Qwen2-VL-7B-Instruct + +The image appears to be real. There are no obvious signs of digital manipulation or distortion that would suggest it is fake. + +OURS + +The image is real because it shows a man with a nose, mouth, and other facial features that are clearly visible and not distorted or altered. The man is wearing a jacket and has a close-up shot of his face, which indicates that the image is not a manipulated or digitally altered photo. The presence of a nose, mouth, and other facial features in the image suggests that it is a genuine representation of the man's appearance, rather than a manipulated or altered image. + +![](images/8d195e89fcb4f8591d028c604dc71816e31645f7a428809790ff81fe6f8fb4ba.jpg) +Figure 12: Answer from different model +Ground Truth: fake +Figure 13: Answer from different model + +# Model Descriptions + +GPT40 + +Fake, [face, unnatural texture and blending issues around the skin and eyes]. + +Llama-3,2-11B-Vision + +It's a bit blurry and low quality, but I can see that the woman's face is distorted and the eyes are not properly proportioned to the rest of her face. The image appears to be too far apart. The image appears to be too wide and the eyes appear to be too far apart. The image appears to be a fake image. It may have been created for a specific purpose. + +Ilya-y1.5-7b + +The image appears to be real. It shows a person with long hair, but the context and background are not clear enough to determine the authenticity of the image. + +Qwen2-VL-7B-Instruct + +The image appears to be real. It shows a person with long hair, but the context and background are not clear enough to determine the authenticity of the image. + +OURS + +The image is real because it shows a woman with a nose that is not perfectly aligned with the rest of her face. This imperfection is noticeable and adds to the authenticity of the Image. The woman's facial features are not perfectly symmetrical, and there are no signs of distortion or manipulation that would indicate this. + +# Project Summary: + +With the rapid development of Deepfake technology, the generation of fake images and videos has become increasingly realistic, posing significant threats to society and individual privacy. Despite the availability of various deep learning models for detecting Deepfake content, there are still notable shortcomings in terms of explainability and transparency. These differences directly impact human trust and understanding during the detection process. + +This project aims to evaluate the alignment between our model's + +explainability in Deepfake detection and human intuition through a human evaluation study. We will compare our model with current mainstream Deepfake detection models to examine the intuitiveness, accuracy, and user trust in the explanations provided. The study will involve participants evaluating the detection results and explanations from different models, assessing the effectiveness of each model in real-world applications. + +# Figure 14: Human study material part I + +# Project Significance: + +This research will reveal the strengths and weaknesses of different Deepfake detection models in terms of explainability, especially in alignment with human intuition. By comparing our model with other mainstream models, we hope to demonstrate that our model is not only competitive in detection accuracy but also achieves higher trust and satisfaction in terms of explainability. The project outcomes will offer new insights and directions for the future design and application of Deepfake detection models, promoting the development of more transparent and reliable detection technologies. + +Potential Side Effects, Hazards, and Emergency Plans: + +Side Effects: During the project, participants may experience confusion or reduced trust in real images due to exposure to a large amount of Deepfake content, leading to difficulty in distinguishing between real and fake images. + +Emergency Plan: We will inform participants that there are currently various effective Deepfake detection solutions, including the model we are developing, which can effectively counter Deepfake attacks to some extent. Through education and explanation, we will help participants understand the progress of the technology and the available protective measures, enhancing their trust in real images. + +# Figure 15: Human study material part2 + +Potential Ethical Issues and Countermeasures (including Informed Consent, Privacy Protection, Physical Harm, and Benefit Distribution): + +Informed Consent: Participants need to fully understand the research content, purpose, and potential impact. We will ensure that all participants voluntarily sign informed consent forms. Countermeasure: Provide detailed research explanations and Q&A sessions to ensure participants fully understand and agree to participate. + +Privacy Protection: All participant information and data involved in the research will be kept strictly confidential and will not be used for any purposes other than the research. Countermeasure: Implement data encryption and anonymization to ensure participant privacy is not + +compromised. + +Physical Harm: Although this study mainly involves psychological and cognitive assessments, we will minimize potential psychological impacts on participants and avoid any form of psychological stress or harm. + +Countermeasure: Conduct real-time monitoring during the study to ensure participant mental health, and provide participants with the right to withdraw from the study at any time. + +Benefit Distribution: Ensure that participants are not unfairly treated during the study and that the outcomes of the research do not disproportionately benefit or disadvantage any individual or group. + +Countermeasure: Fair and reasonable distribution of research outcomes, ensuring openness and transparency. The research results will primarily focus on academic and social contributions, not for personal or commercial gain. + +# Figure 16: Human study material part3 + +# E Additional Analysis of Model Feature Assessment + +# E.1 Evaluation Setup of MFA + +Model. We choose the mainstream MLLM, i.e., LLaVA [44] as the implementation instance of the pre-trained MLLMs. Additionally, we choose one classical detector, Xception [12], as a baseline model for comparison. + +Dataset. We evaluate the models on several widely-used deepfake datasets, including the Deepfake Detection Challenge (DFDC) [16], the preview version of DFDC (DFDCP) [15], DeepfakeDetection (DFD) [13], Celeb-DF-v2 (CDF-v2) [37], DF40 dataset [76], which incorporates state-of-the-art forgery techniques such as Facedancer [57], Fsgan [51], Inswap [59], e4s [33], Simswap [7], and Uniface [91]. providing a comprehensive foundation for evaluating overall detection performance. + +Evaluation Metrics. We use the Area Under the Curve (AUC) as the primary evaluation metric, enabling us to assess the model's ability to distinguish between real and fake images across the whole dataset. In this section, we use the frame-level AUC for evaluation. For individual feature discrimination, we focus on forgery-related features such as "Is the face layout unnatural?" with responses of either "yes" or "no." The proportions of "yes" and "no" answers for real and fake images are calculated as follows, with the ranking score $S^{(q)}$ defined based on the balanced accuracy of the responses: + +$$ +S ^ {(q)} = \frac {1}{2} \left(\frac {Y _ {\text {r e a l}} ^ {(q)}}{Y _ {\text {r e a l}} ^ {(q)} + N _ {\text {r e a l}} ^ {(q)}} + \frac {N _ {\text {f a k e}} ^ {(q)}}{Y _ {\text {f a k e}} ^ {(q)} + N _ {\text {f a k e}} ^ {(q)}}\right). \tag {3} +$$ + +Here, $Y_{\mathrm{real}}^{(q)}$ and $Y_{\mathrm{fake}}^{(q)}$ denote the number of "yes" answers, while $N_{\mathrm{real}}^{(q)}$ and $N_{\mathrm{fake}}^{(q)}$ represent the number of "no" answers for real and fake, respectively. This formulation ensures that both true positive and true negative rates are considered, providing a balanced measure of feature discrimination. + +# E.2 Evaluation of the Overall Detection Performance + +The comparison between LLaVA [44] and Xception [12] highlights a notable performance gap. Results in Figure 17 (left) indicate that the average AUC for LLaVA is $63.7\%$ , while Xception achieves $75.8\%$ , showing a notable gap of $12.1\%$ points. This suggests that, while the LLaVA has certain zero-shot capabilities in other tasks such as (general) image classification, it is still not as strong as the traditional detector in detecting deepfakes. + +However, LLaVA shows strong detection abilities in specific methods (eg., e4s), sometimes even surpassing Xception (see Figure 17 (left)). This motivates us to further investigate its intrinsic detection capabilities, and understand the model's "strengths and weaknesses" in deepfake detection. Below, we provide a detailed investigation of the discrimination of each forgery-related feature. + +![](images/556cbd479abc3d14237527917dc9f70e3fc67b87e11aa7c3e875128e92856098.jpg) +Figure 17: (Left) AUC comparison between (zero-shot) LLaVA (blue) and Xception (red) for deepfake detection across different datasets; (Right) Balance accuracy score for individual feature discrimination, with Strong features in the top-left corner and Weak features in the bottom-right corner based on discrimination scores. Full questions/features are provided in the Table. + +![](images/c522c36ae98d5e676a5a328b608d5e204eea1c792927bcdd94f83ed40ef5d14b.jpg) + +# E.3 Deeply Investigation of Individual Feature's Discrimination + +Step 1: Question Generation. For each candidate forgery-related feature, we formulate a corresponding interrogative statement. For instance, the feature "blurry" is transformed into the question "Is the image blurry?" Recognizing that the candidate features are not pre-specified by developers, we employ a Large Language Model (LLM), i.e., GPT-4o, to automatically generate a comprehensive list of $F_{i}$ questions. These questions target key forgery indicators, including but not limited to lighting anomalies, unnatural facial expressions, and mismatched textures, which are critical for identifying deepfakes. + +Step 2: Question Evaluation. Each generated question is paired with an image from the assessment dataset to form a prompt for constructing the fine-tuning dataset. The model responds with a binary output ("yes" or "no") based on its interpretation of the image in relation to the question. These responses are aggregated into a confusion matrix for each question, thereby quantifying the detection capability of the associated forgery-related features. Mathematically, for each question $F_{i}$ and image $x_{j}$ , the MLLM produces: + +$$ +R _ {i, j} = \mathcal {M} _ {\text {b a s e}} \left(F _ {i}, x _ {j}\right), \tag {4} +$$ + +where $R_{i,j} \in \{\text{yes, no}\}$ , representing the model's response for each image-question pair. + +Step 3: Question Ranking. According to the accuracy of all candidate questions, we obtain a descending ranking of questions, i.e., the ranking of forgery-related features. This ranking allows us to quantify how well each feature contributes to distinguishing between real and fake images. Specifically, the accuracy of each question is computed by evaluating the proportion of correct responses across the dataset. Specifically, for each question $Q_{i}^{k}$ , We calculate the true positive rate (TPR) and true negative rate (TNR), then take their average to obtain the Balanced Accuracy, as follows: + +$$ +\text {B a l a n c e d A c c u r a c y} _ {i} = \frac {1}{2} \left(\frac {\mathrm {T P} _ {i}}{\mathrm {T P} _ {i} + \mathrm {F N} _ {i}} + \frac {\mathrm {T N} _ {i}}{\mathrm {T N} _ {i} + \mathrm {F P} _ {i}}\right), \tag {5} +$$ + +where: $\mathrm{TP}_i$ denotes True Positives for question $F_i$ , $\mathrm{TN}_i$ the True Negatives for question $F_i$ , $\mathrm{FP}_i$ the False Positives for question $F_i$ , and $\mathrm{FN}_i$ the False Negatives for question $F_i$ . + +Subsequently, questions are ranked in descending order based on their balance accuracy scores, thereby prioritizing forgery features that effectively discriminate between real and fake images. + +Strong Features. Strong features typically involve semantic-level facial structural or appearance anomalies. As shown in the strong feature section of Fig. 17 (right), which primarily includes facial irregularities such as unusual facial layouts (eg., Rank 9, 11, 17) or distorted facial features (eg., Rank 3, 4, 14), eg., the nose, eyes, or mouth. Since the pre-trained MLLM is good at extracting and utilizing these features for detection, it can provide a more reliable and accurate explanation. + +Weak Features. Weak features typically involve fine-grained, low-level textures, such as blending anomalies. As shown in Fig. 17 (right), these weak features are primarily subtle details related to texture, reflection, shadow, and blending. Examples of texture issues include rough or overly smooth surfaces (eg., Rank 68, 77, 83). Furthermore, inconsistencies in lighting and shadows (eg., Rank 85, 86, 90, 96) and blending artifacts on the face (eg., Rank 54, 84, 88) are also prominent. Since these signal-level anomalies are challenging for pre-trained MLLMs to detect, the pre-trained MLLM is likely to struggle in reliably distinguishing between real and manipulated content when relying on these weak features for detection and explanation. + +Feature Score. The scores for different forgery-related features are presented, where Table 19 highlights the top 50 strong features, and Table 20 shows the 50 weakest features based on their scores. + +Does the model know these features are related to deepfake? + +We used a series of questions to query the model, applying simple prompt augmentation with the feature-related questions mentioned above. A "yes" indicates the model knows these features are related to deepfake detection, while a "no" indicates the model does not. Detailed results are shown in Table 17 and Table 18. + +# F Additional analysis of Strong Feature Strengthening + +Following the Model Feature Assessment (MFA) module, we observed significant performance improvements in forgery detection after applying the Strong Feature Strengthening (SFS) module. The generalization performance is improved by $20\%$ AUC (see Table 14) on average compared to the pretrain model. + +we conducted a detailed analysis by re-evaluating the model using the Model Feature Assessment (MFA) module to compare the discriminative capability of forgery-related features before and after + +applying the SFS module. As illustrated in Figure 18, over $60\%$ of the features exhibited enhanced discriminative power post-SFS, with particularly notable improvements in strong features. + +![](images/4872ff6a0358776f777c93d3dc3a8ea6b8c346f99318be9258fc10e512db723a.jpg) +Feature Capabilities Ablation +Figure 18: Comparison of feature capability before and after SFS. After adding the external detector to supplement the MLLM, the model's feature capabilities (almost all) can be further improved. + +# G Additional Analysis of Weak Feature Supplementing + +In addition to the used blending model [41], we also try other instances to implement the SFDs in the WFS module of our framework, each targeting specific types of artifacts, where SRI focusing on the generative artifacts by deep nets, F3-Net focusing on the frequency-level anomalies, and SBI and CFDA focusing on the blending boundaries. Based on these empirical attempts, We summarize the general criteria under which conditions the selected SFD instance can be used in our framework. Specifically, the integrated SFD instance should meet the following criteria: + +- Criteria-1: Each SFD instance should focus on only one type of feature that is positively correlated with fake; +- Criteria-2: The score given by the SFD instance can accurately reflect the characteristics of the corresponding feature; +- Criteria-3: The data distribution of this feature in the dataset is relatively uniform. + +Below, we show a detailed illustration of using other SFD instances for implementation one by one. + +AIGC Expert Integration. We first consider implementing an AIGC expert to learn the deep generative artifacts. For implementation, we introduce the SRI model, based on self-reconstruction images generated by Simswap [7] and train on the Xception model, designed to capture self-reconstruction generative features. However, from Figure 19, integrating this model into our framework results in only a minor performance improvement of $0.3\%$ . Further analysis reveals a negative correlation between the model's features and fake labels in the training set (do NOT meet the Criteria-1), indicating that these artifacts are poorly represented in the training data. Consequently, the model struggles to leverage the expert-provided features effectively, offering limited benefits over not using the expert model. + +![](images/480d0a6639ffa19e5f1e0bb722c2822513430a136590c1d0ffb01ff585617b93.jpg) +Figure 19: The probability distributions of different expert models on the FF++ training dataset. From left to right, the models are SRI, F3-Net, SBI, and CFDA, corresponding to experts in capturing self-reconstruction, frequency anomalies, and self-blending artifacts, respectively. The blending here directly uses the trained weights. + +![](images/4ce90c81889379cba2190c74dee374dce68baaff8713bcea769b55383244ffce.jpg) + +![](images/5ee8dd1b5e74f56d46adf5a0513df1e54f9e213aa7744867b5ed6101bbe20d1a.jpg) + +![](images/e2a9bf1e443579acacb8f1efae772f6524315f31a6444a20945e257632c6aa20.jpg) + +Frequency Expert Integration. We then integrate a frequency-based model $F3-Net$ and train it on the FF++ dataset [58] to capture frequency anomalies. However, from Figure 19, the overall model's performance is identical to that of the expert, with no improvement. Although the expert features are positively correlated with fake labels, the frequency-based scores are overfitted to the training set and do not accurately reflect the true feature quantity, with only near-1 (1 for fake) and near-0 (0 for real) predictions (Not satisfy the Criteria-2) This leads to a shortcut, where the model relies solely on the + +expert's output without learning from the feature information, thus limiting the extendability of the integrated model. + +Table 12: Comparison of methods across datasets with values rounded to two decimal places, where the evaluation metric is AUC, The "Diff" column shows the difference from the mllms average. + +
VariantCDFDFDCPDFDCDFDUnifacee4sFacedancerFSGANInswapSimswapAvgDiff
MLLM83.382.079.291.484.594.179.988.077.283.384.30.0
SRI42.949.352.950.997.365.771.380.180.599.969.1-15.2
SRI+MLLM83.282.577.688.885.695.881.988.577.984.684.6+0.3
F3Net77.077.272.882.387.571.675.489.283.977.279.4-4.9
F3Net+MLLM76.877.873.183.388.475.576.689.884.678.480.4-3.9
SBI82.182.370.585.583.476.868.583.277.487.779.7-4.6
SBI+MLLM88.685.575.690.888.793.777.688.281.691.186.2+1.9
CDFA87.986.683.590.976.567.475.484.872.076.180.1-4.2
CFDA+MLLM90.389.783.592.585.291.283.889.978.584.987.0+2.7
+ +SBI and CFDA Models Integration. We also integrate another blending-based expert model, SBI [61], which specializes in detecting blending artifacts. From Figure 19, we can see that trained using self-blending techniques on real images to prevent overfitting, the SBI model's expert features show a strong correlation with fake labels, and its scoring effectively quantifies the extent of blending artifacts. Similarly, the incorporation of the CFDA model [41], an enhanced version of the SBI model, results in an additional performance boost, indicating that as the expert model's ability to capture blending features improved, the overall model's generalization capability also increases. + +To explain criteria-3, we conducted additional experiments using non-uniform data distribution. Specifically, we created an extremely imbalanced dataset by removing a large portion of fake samples that do not contain the blending feature. As the imbalance increased, the model's performance degraded, and in extreme cases, it began to rely on shortcut solutions. In the remove 95 and remove 99.5 cases, we removed $95\%$ and $99.5\%$ of samples close to the real distribution, respectively, resulting in highly imbalanced datasets with mostly fake samples remaining. + +Table 13: Performance Comparison of Different Models on Various Datasets. The remove 95 and remove 99.5 scenarios represent extreme cases of data imbalance by removing $95\%$ and $99.5\%$ of the samples near the real distribution, respectively. + +
VariantCeleb-DF-v2DFDCPE4SFacedancerFsganInswapSimswapAverage
remove 99.575.679.063.667.280.263.065.470.6
remove 9579.381.468.969.782.165.770.373.9
CDFA+MLLM90.389.691.283.889.978.584.986.9
+ +# H Additional Analysis of Ablation Study + +Effects of Strong Feature Strengthening Module. We observed a substantial performance boost in the model after fine-tuning it with a dataset constructed using strong features, as evidenced by the leap from Variant-1 to Variant-5. Remarkably, this significant improvement occurred even without the aid of specific feature detectors (WFS), underscoring the potency of strong feature utilization. This finding prompted us to investigate the underlying causes. We reexamined the feature capabilities of the pre-trained model (Variant-1) and compared them to those of the model enhanced with strong features (Variant-5). As illustrated in Figure 18, the majority of feature capabilities exhibited marked enhancement following the application of strong feature strengthening. Intriguingly, even some initially weaker features demonstrated noticeable improvement post-enhancement. + +Effect of Model Feature Assessment Module. Without Model Feature Assessment (MFA), identifying the model's strong features becomes impossible, and relying on weaker features undermines both the reliability of the model's explanations and its overall performance. To explore this, we allowed the model to construct the dataset using the entire list of forgery-related features rather than prioritizing strong ones. In Variant-3, which adopts this approach without enforcing strong feature use, performance lags significantly behind Variant-5, where MFA pinpoints and leverages strong features for fine-tuning, as evidenced in Table 14. To further investigate the absence of + +Table 14: Ablation study regarding the effectiveness of each proposed module via cross-dataset evaluations. All models are trained on the FF++ c23 dataset and evaluated with metrics in the order of AUC || AP || EER (frame-level). The results show an incremental benefit in each module. We use $\checkmark$ to indicate the presence of a module and $X$ to indicate its absence. + +
OursCDFDFDDFDCSimswapUnifaceAvg.
#MFASFSWFSAUC | AP | EERAUC | AP | EERAUC | AP | EERAUC | AP | EERAUC | AP | EERAUC | AP | EER
1XXX52.1 | 68.2 | 48.769.8 | 95.2 | 36.457.8 | 59.9 | 44.664.0 | 64.1 | 40.465.5 | 65.6 | 39.061.8 | 70.6 | 41.8
2✓XX52.3 | 67.4 | 49.475.0 | 96.0 | 31.563.3 | 66.0 | 39.959.3 | 59.6 | 43.757.8 | 58.5 | 44.161.5 | 69.5 | 41.8
3X✓X79.0 | 88.3 | 28.988.9 | 98.7 | 18.077.8 | 81.9 | 28.982.0 | 84.0 | 25.982.3 | 84.8 | 25.282.0 | 87.3 | 25.6
4XX✓87.9 | 93.6 | 20.590.9 | 98.9 | 17.683.5 | 86.1 | 24.876.0 | 74.2 | 29.876.5 | 75.1 | 29.883.0 | 85.6 | 24.5
5✓✓X83.2 | 90.5 | 24.691.4 | 99.0 | 15.879.2 | 82.1 | 27.683.3 | 85.0 | 24.884.5 | 86.2 | 22.484.9 | 88.5 | 23.0
6X✓✓88.1 | 93.6 | 20.491.1 | 98.6 | 16.882.0 | 84.5 | 26.178.0 | 76.5 | 28.878.7 | 77.1 | 28.683.6 | 86.1 | 24.1
7✓X✓87.3 | 93.2 | 21.090.2 | 98.7 | 18.082.0 | 82.1 | 26.676.7 | 75.0 | 29.577.2 | 75.8 | 29.082.6 | 84.9 | 24.8
8✓✓✓90.4 | 94.9 | 17.792.3 | 99.1 | 15.583.7 | 86.1 | 24.885.1 | 85.7 | 23.285.5 | 86.4 | 22.487.4 | 90.4 | 20.7
+ +MFA, we introduced Variant-6, which integrates specific feature detectors (WFS) without MFA, and compared it to Variant-8, the full model with MFA, SFS, and WFS. The results clearly show Variant-6 underperforming Variant-8, highlighting a notable disadvantage when model assessment is omitted. These experiments Variant-3 versus Variant-5 and Variant-6 versus Variant-8 collectively demonstrate that assessing and utilizing strong features via MFA is critical for optimizing model effectiveness in deepfake detection. Strikingly, despite the pre-trained Variant-1's limited detection ability, Variant-2 reveals the model's feature capability by using a single MFA-identified strong feature (e.g., distortion—present for fake, absent for real) for direct judgment without training. This underscores the model's strong feature-identification capability and the potential of tapping these features, which combinations in Variants 5 and 8 further amplify. + +We evaluate the generalization performance of our model in cross-dataset evaluation scenarios through an ablation study involving several variants to systematically assess the contributions of different components. The variants include: Variant-1, a pre-trained MLLM LlaVa-1.5-7B[44] as baseline, without any feature strengthening or supplementation; Variant-2, which utilizes the single strongest feature identified by Model Feature Assessment (MFA) for deepfake detection; Variant-3, Assumes all features are strong and uses them to construct the dataset without prior evaluation or ranking, followed by fine-tuning the model; Variant-4, a simple test of the specific feature detectors ability, for this model can only output a single probability, as proposed in [41]; Variant-5, only use strong features to construct datasets after feature assessment, and use this dataset to fine-tune; Variant-6, Uses all strong features without feature assessment, integrating the specific feature detectors (SFD) to construct datasets, followed by fine-tuning the model.; and Variant-7, Constructs datasets solely using the specific feature detectors (SFD), masking visual information to prevent strong feature leakage, followed by fine-tuning the model. Variant-8, Full model, integrating MFA, SFS, and WFS. The results, presented in Table 14, demonstrate incremental improvements across various datasets. For instance, Variant-8 achieves the highest average AUC $(87.4\%)$ , AP $(90.3\%)$ , and lowest EER $(20.7\%)$ , highlighting the synergy of MFA, SFS, and WFS in optimizing deepfake detection performance. + +Feature Capability. We also conduct a comparative study of feature capabilities before and after feature strengthening. + +Effect of inconsistent use of supplementary features in training and inference. The model performs best when supplementary features are used consistently during both training and inference (average auc: 87.8), indicating that these features significantly enhance performance. When supplementary features are omitted entirely from both stages, the performance drops (average auc: 0.83.3), though it remains better than when features are used inconsistently. Specifically, when features are used during training but not inference, the performance suffers greatly (average auc: 76.6), suggesting the model relies on these features and struggles without them at inference time. On the other hand, when features are introduced at inference but not used during training, the model achieves slightly better results (average auc: 82.5), but it cannot fully leverage unseen features, showing the importance of using supplementary features consistently across both phases. + +Extension in new datasets. Our model, trained on a mixture of datasets including FF++, showed improved overall performance when we added a new dataset without blending artifacts to the training process. This demonstrates that incorporating diverse datasets with supplementary features, even + +Table 15: Impact of omitting supplementary features during Training and Adding Them During Inference, on Model Performance + +
VariantCeleb-DF-v2DFDDFDCDFDCPDFRWDFFFIWAvg.
no train + no infer83.291.479.282.093.776.876.683.3
train + no infer76.584.771.072.089.571.371.376.6
no train + infer81.790.679.181.392.674.777.382.5
train + infer90.492.383.787.397.681.481.787.8
+ +from different domains, as shown in Table 16 enhances the model's generalization and comprehensive performance. + +Table 16: Comparison of Model Performance When Trained on FF++ Alone vs. FF++ and SRI Across Different Datasets + +
VariantCDFDFDCPDFDCDFDUnifacee4sFacedancerFsganInswapSimswap
Train FF++90.487.383.792.385.591.283.891.181.285.1
Train FF++ and SRI91.589.383.992.787.493.084.689.981.086.1
+ +# I Sample Showing + +Here we present some failure cases of pre-trained MLLMs and then compare them with our results. See the supplementary material for more examples and details. + +![](images/bafb1223a7581fb11f4a9a2f2814a7adbd1290bd66fa61daa49eafc4549ef46a.jpg) +Label: real + +USER: I want you to work as an image forensic expert for Deepfake faces. Check if the image has the artifact attribute listed in the following list and ONLY return the attribute number in this image. The artifact list is [1-asymmetric eye iris; 2-irregular glasses shape or reflection; 3-irregular teeth shape or texture; 4-irregular ears or earrings; 5-strange hair texture; 6-inconsistent skin texture; 7-inconsistent lighting and shading; 8-strange background; 9-weird hands; 10-unnatural edges]. + +Detection + +GPT: Attribute number: Based on the provided list of artifacts, I reviewed the image and found that it displays the following attribute:5-strange hair texture. The texture of the hair appears somewhat artificial, a common characteristic of AI-generated images. + +USER: This image is real, tell me the reason. + +Explanation + +GPT: The image you provided appears to be a real person, possibly from a television interview or public appearance. The reason it seems real is due to the natural lighting, human facial expressions, and other small details such as skin texture and natural positioning, which are characteristics of an authentic photograph of a person in a live setting. However, based on this image alone, I cannot confirm its exact source or context. + +USER: Is this image is real or deepfake, tell me the reason. + +Detection& Explanation + +GPT: The image you provided appears to be a real person, possibly from a television interview or public appearance. The reason it seems real is due to the natural lighting, human facial expressions, and other small details such as skin texture and natural positioning, which are characteristics of an authentic photograph of a person in a live setting. However, based on this image alone, I cannot confirm its exact source or context. + +Figure 20: Diagram showing how the GPT4o generates results for detection, explanation, and detection with explanation. + +For LLaVa, we use the same prompts as GPT-4 to ensure fairness in the evaluation process. The robustness of Llava in these tasks is illustrated in Figure 21. + +Table 17: Relationship between various facial features and deepfake detection (Part 1) + +
No.QuestionPretrain
1Is the face color related to deepfake detection?No
2Are the eyes related to deepfake detection?No
3Are the facial features related to deepfake detection?No
4Is the nose contour related to deepfake detection?No
5Is face blurriness related to deepfake detection?No
6Is the skin tone related to deepfake detection?No
7Are the cheeks related to deepfake detection?No
8Is the skin tone pattern related to deepfake detection?No
9Is the placement of facial features related to deepfake detection?No
10Are the lips related to deepfake detection?Yes
11Is facial symmetry related to deepfake detection?No
12Is the lighting on the cheeks related to deepfake detection?Yes
13Is the facial lighting related to deepfake detection?No
14Are the shapes of facial features related to deepfake detection?No
15Is facial evenness related to deepfake detection?Yes
16Are the cheekbones related to deepfake detection?No
17Is the face layout related to deepfake detection?No
18Are the lip edges related to deepfake detection?No
19Is facial detail related to deepfake detection?Yes
20Is cheek smoothness related to deepfake detection?No
21Is the forehead shape related to deepfake detection?No
22Is face-background blending related to deepfake detection?Yes
23Is skin texture related to deepfake detection?No
24Are the eyelashes related to deepfake detection?No
25Are facial lines related to deepfake detection?No
26Is facial expression related to deepfake detection?No
27Is the nose shape related to deepfake detection?No
28Are color changes on the face related to deepfake detection?No
29Is the mouth shape related to deepfake detection?No
30Are the face edges related to deepfake detection?No
31Is facial rigidity related to deepfake detection?No
32Are sharp facial lines related to deepfake detection?No
33Is skin perfection related to deepfake detection?No
34Is forehead shininess related to deepfake detection?Yes
35Are sharp face edges related to deepfake detection?Yes
36Is skin smoothness related to deepfake detection?No
37Are eye details related to deepfake detection?No
38Are smooth facial lines related to deepfake detection?No
39Is lip texture related to deepfake detection?Yes
40Is forehead shine evenness related to deepfake detection?No
41Are the eyebrows related to deepfake detection?No
42Are unusual eye appearances related to deepfake detection?No
43Are facial transitions related to deepfake detection?Yes
44Is face color related to deepfake detection?No
45Is facial emotion exaggeration related to deepfake detection?No
46Is unusual face layout related to deepfake detection?No
47Are eye reflections related to deepfake detection?No
48Is skin texture roughness related to deepfake detection?No
49Is the jawline related to deepfake detection?No
50Is facial expression stiffness related to deepfake detection?Yes
+ +Table 18: Relationship between various facial features and deepfake detection (Part 2) + +
No.QuestionPretrain
51Is nose texture related to deepfake detection?No
52Is skin shininess under the nose related to deepfake detection?No
53Is uneven facial sharpness related to deepfake detection?Yes
54Is facial blending related to deepfake detection?No
55Is facial lighting evenness related to deepfake detection?No
56Is nose bridge smoothness related to deepfake detection?No
57Is the hairline related to deepfake detection?No
58Is skin texture evenness related to deepfake detection?No
59Is facial feature balance related to deepfake detection?No
60Is facial symmetry related to deepfake detection?No
61Is forced facial expression related to deepfake detection?No
62Are the nostrils related to deepfake detection?No
63Are unnatural lip appearances related to deepfake detection?No
64Is partial skin smoothness related to deepfake detection?No
65Is lip texture related to deepfake detection?No
66Is lighting around the nose related to deepfake detection?Yes
67Are facial feature proportions related to deepfake detection?Yes
68Is skin smoothness around the nose related to deepfake detection?No
69Are soft facial creases related to deepfake detection?No
70Are teeth appearances related to deepfake detection?No
71Is neck-face transition related to deepfake detection?No
72Is skin tone variation related to deepfake detection?No
73Is face edge sharpness related to deepfake detection?No
74Is chin outline visibility related to deepfake detection?No
75Is facial lighting evenness related to deepfake detection?Yes
76Are ear details related to deepfake detection?No
77Is chin smoothness related to deepfake detection?No
78Are bright facial areas related to deepfake detection?No
79Is skin brightness near the mouth related to deepfake detection?No
80Are nostril appearances related to deepfake detection?No
81Are dimples related to deepfake detection?Yes
82Is jawline prominence related to deepfake detection?No
83Is under-eye texture related to deepfake detection?No
84Is facial blending related to deepfake detection?Yes
85Is chin shadow related to deepfake detection?No
86Are forehead shadows related to deepfake detection?No
87Is nose light reflection related to deepfake detection?No
88Is face-background transition related to deepfake detection?No
89Is forehead light reflection related to deepfake detection?No
90Are nose shadows related to deepfake detection?No
91Is lighting around the mouth related to deepfake detection?No
92Is neck smoothness related to deepfake detection?No
93Are face outlines related to deepfake detection?No
94Are face edges related to deepfake detection?No
95Are skin details related to deepfake detection?No
96Are under-eye shadows related to deepfake detection?No
97Are cheek shadows related to deepfake detection?No
98Are cheekbone appearances related to deepfake detection?No
99Is facial lighting related to deepfake detection?No
100Are facial wrinkle details related to deepfake detection?No
+ +Table 19: Top Strong 50 Features + +
RankQuestionPretrainStrengthened
1Is the face color unusual?0.63400.7486
2Is there something wrong with the eyes?0.63090.6320
3Do the facial features look oddly shaped?0.62920.6636
4Is the contour of the nose incorrect?0.62780.5817
5Is part of the face blurry?0.62310.7643
6Does the skin tone make the face look fake?0.61650.6479
7Is there something wrong with the cheek?0.61440.8082
8Are there strange patterns in the skin tone?0.61440.8075
9Are the face parts out of place?0.61300.7919
10Do the lips seem out of place or strangely shaped?0.61270.7408
11Is one side of the face uneven with the other?0.61230.7622
12Are there strange lighting spots on the cheeks?0.61110.8029
13Does the lighting change strangely on the face?0.60920.8006
14Are the shapes of the eyes, nose, or mouth unnatural?0.60540.6714
15Does the face look uneven or off?0.60480.6732
16Does the cheekbone appear too flat?0.60140.7406
17Does the face layout look wrong?0.59860.5422
18Are the edges of the lips too smooth?0.59790.6264
19Is part of the face lacking detail?0.59420.6843
20Are the cheeks too smooth?0.59340.6728
21Does the forehead look odd in shape?0.59110.7493
22Does the face mix poorly with the background?0.59020.6382
23Is the skin texture uneven?0.58610.7158
24Are the eyelashes missing or blurred?0.58570.6733
25Are the face lines uneven or changing in different areas?0.58260.6546
26Does the face lack expression?0.58220.6679
27Does the nose shape look odd?0.58120.5306
28Are the color changes on the face and skin sudden?0.58070.6643
29Does the mouth appear too flat?0.57750.6542
30Are the edges of the face too sharp?0.57740.8188
31Does the face appear too rigid?0.57700.7446
32Are the face lines too sharp?0.57610.7724
33Does the skin look too perfect, like it was edited?0.57550.5749
34Is the forehead too shiny?0.57370.8168
35Are the face edges too sharp?0.57200.8162
36Does the face skin look too smooth?0.56400.5306
37Are the eyes blurry or lacking detail?0.56360.5362
38Are the face lines too smooth?0.55490.5927
39Are the lips too smooth or lacking texture?0.55370.5475
40Is the forehead's shine uneven?0.55150.7208
41Are the eyebrows too dark or too light?0.54540.5075
42Do the eyes look odd?0.54330.5389
43Are transitions on the face poorly blended?0.54100.5854
44Do the face colors look strange?0.53820.6163
45Does the face show emotions that seem exaggerated?0.53550.6337
46Does the face layout look unusual?0.53440.5377
47Do the eyes have unnatural reflections?0.53230.6417
48Does the face have rough or uneven skin texture?0.52920.7973
49Does the jawline appear too sharp or unclear?0.52920.5017
50Does the facial expression look stiff?0.52890.5346
+ +Table 20: Bottom 50 Weak Features + +
RankQuestionPretrainedStrengthened
51Does the nose lack texture?0.52310.5111
52Is the skin too shiny under the nose?0.52230.7458
53Is the sharpness of the face uneven in parts?0.52140.5701
54Does the blending on the face look unnatural or uneven?0.52120.5151
55Is the lighting on the face strange or uneven?0.52000.5968
56Does the nose bridge appear too smooth?0.51720.5395
57Does the hairline seem unnatural?0.51480.5495
58Does the face skin texture look uneven?0.51440.5489
59Do the face parts look out of balance?0.51370.5887
60Are the facial features too symmetrical?0.51300.7300
61Does the facial expression look forced?0.51160.5562
62Are the nostrils hard to see?0.51150.6535
63Do the lips look unnatural?0.51100.5555
64Does the face skin look too smooth in some areas?0.50890.5287
65Do the lips lack natural texture?0.50830.5855
66Is the lighting around the nose inconsistent?0.50800.7257
67Do the sizes of the eyes, nose, and mouth seem off?0.50380.5275
68Does the skin around the nose look unnaturally smooth?0.50300.5309
69Are the facial creases too soft?0.50280.7943
70Do the teeth appear blurry or unrealistic?0.50280.5210
71Is the transition between the neck and the face not smooth?0.50260.5330
72Is the skin tone different in parts of the face?0.50230.5749
73Does the face lack sharpness around the edges?0.50210.5311
74Is the chin outline hard to see?0.50210.6160
75Is the lighting uneven on the face?0.50120.6351
76Are the details around the ears unclear?0.50100.6751
77Is the chin too smooth compared to the rest of the face?0.50100.5664
78Do the bright areas on the face seem odd?0.50070.5196
79Is the skin near the mouth unnaturally bright?0.50070.5930
80Are the nostrils blurry or unclear?0.50070.5125
81Are the dimples missing or poorly defined?0.50050.5000
82Is the jawline too pronounced or too faint?0.50000.5003
83Is the area under the eyes missing natural texture?0.50000.5111
84Is there blending on the face that looks edited?0.50000.5014
85Does the shadow under the chin seem unnatural?0.50000.5090
86Is the forehead missing natural shadows?0.50000.5000
87Does the light reflection on the nose look strange?0.50000.5049
88Are the transitions between the face and the background poorly blended?0.50000.5447
89Does the light reflection on the forehead look artificial?0.50000.5007
90Are there missing shadows around the nose?0.50000.5217
91Does the lighting around the mouth look unusual?0.50000.5301
92Does the neck look unnaturally smooth compared to the face?0.50000.6259
93Do the face outlines look off?0.50000.5247
94Do the edges around the face look unnatural?0.50000.5299
95Are the fine details on the skin missing?0.50000.5165
96Are the shadows under the eyes missing?0.50000.5000
97Are the cheeks lacking shadows?0.50000.5000
98Do the cheekbones appear unnaturally smooth?0.50000.5709
99Does the face appear overly lit in certain areas?0.50000.6758
100Are the wrinkles on the face lacking detail?0.50000.5014
+ +Table 21: Questions list generated by Claude3.5-Sonnet (part1) + +
No.Question
1Are there noticeable inconsistencies in facial symmetry? Return me yes or no
2Does the skin texture appear artificially smooth or lacking natural details? Return me yes or no
3Are the eyes misaligned or disproportionate? Return me yes or no
4Is there unnatural blending between facial features and background? Return me yes or no
5Do shadows and lighting appear inconsistent across the face? Return me yes or no
6Are facial expressions unnatural or mechanically rigid? Return me yes or no
7Does the hairline show signs of artificial blending? Return me yes or no
8Are there visible artifacts or glitches in the image? Return me yes or no
9Do reflections in the eyes match the environment? Return me yes or no
10Is there proper alignment of facial features? Return me yes or no
11Does the skin show natural imperfections and pores? Return me yes or no
12Are teeth shapes and alignment realistic? Return me yes or no
13Is there consistent image quality across all facial areas? Return me yes or no
14Do facial proportions follow natural human anatomy? Return me yes or no
15Are shadows cast appropriately based on lighting? Return me yes or no
16Does facial hair follow natural growth patterns? Return me yes or no
17Is there proper depth and dimension to facial features? Return me yes or no
18Are color tones consistent throughout the face? Return me yes or no
19Do glasses and accessories appear properly attached? Return me yes or no
20Is there natural variation in skin texture? Return me yes or no
21Are facial contours anatomically correct? Return me yes or no
22Does the head size match body proportions? Return me yes or no
23Is there appropriate detail in fine features? Return me yes or no
24Are transitions between features naturally blended? Return me yes or no
25Do facial movements appear fluid and natural? Return me yes or no
26Are ear shapes and positions symmetrical? Return me yes or no
27Do eyebrows have natural hair patterns? Return me yes or no
28Is there consistent resolution between face and background? Return me yes or no
29Are nose contours anatomically accurate? Return me yes or no
30Does makeup application appear natural? Return me yes or no
31Are facial wrinkles and lines age-appropriate? Return me yes or no
32Do eyelashes appear realistic and properly attached? Return me yes or no
33Is there natural skin coloration variation? Return me yes or no
34Are facial highlights consistent with lighting? Return me yes or no
35Do lips have natural texture and color? Return me yes or no
36Is there proper depth in eye sockets? Return me yes or no
37Are facial moles and marks naturally placed? Return me yes or no
38Do teeth have individual characteristics? Return me yes or no
39Is there natural asymmetry in facial features? Return me yes or no
40Are skin pores visible where expected? Return me yes or no
41Do facial muscles move naturally? Return me yes or no
42Is there consistent focus across the image? Return me yes or no
43Are shadows under facial features natural? Return me yes or no
44Do earrings and jewelry sit naturally? Return me yes or no
45Is there proper skin subsurface scattering? Return me yes or no
46Are facial proportions consistent in different angles? Return me yes or no
47Do eye corners have natural creases? Return me yes or no
48Is there natural variation in lip texture? Return me yes or no
49Are facial hair shadows realistic? Return me yes or no
50Do glasses cast appropriate shadows? Return me yes or no
+ +Table 22: Questions list generated by Claude3.5-Sonnet (part2) + +
No.Question
51Is there natural skin translucency? Return me yes or no
52Are facial expressions emotionally consistent? Return me yes or no
53Do neck muscles align naturally? Return me yes or no
54Is there proper depth in smile lines? Return me yes or no
55Are eye reflections consistent with scene lighting? Return me yes or no
56Do facial features maintain proportion when moving? Return me yes or no
57Is there natural skin aging present? Return me yes or no
58Are hair strands individually visible? Return me yes or no
59Do facial veins appear natural where visible? Return me yes or no
60Is there consistent skin tone across transitions? Return me yes or no
61Are nostril shapes symmetrical? Return me yes or no
62Do ears have natural internal structure? Return me yes or no
63Is there proper depth in nasolabial folds? Return me yes or no
64Are eye bags and circles age-appropriate? Return me yes or no
65Do facial piercings sit naturally? Return me yes or no
66Is there natural variation in beard density? Return me yes or no
67Are lip lines naturally defined? Return me yes or no
68Do cheekbones have natural contours? Return me yes or no
69Is there proper temple definition? Return me yes or no
70Are eye whites naturally textured? Return me yes or no
71Do facial scars appear authentic? Return me yes or no
72Is there natural jaw definition? Return me yes or no
73Are facial dimples naturally placed? Return me yes or no
74Do eyebrow hairs have direction variation? Return me yes or no
75Is there proper chin definition? Return me yes or no
76Are facial freckles naturally distributed? Return me yes or no
77Do eyelids have natural creases? Return me yes or no
78Is there consistent skin shininess? Return me yes or no
79Are facial tattoos properly embedded? Return me yes or no
80Do smile lines appear natural? Return me yes or no
81Is there proper forehead texture? Return me yes or no
82Are eye corners naturally aged? Return me yes or no
83Do facial muscles show proper definition? Return me yes or no
84Is there natural lip symmetry? Return me yes or no
85Are ear lobes naturally shaped? Return me yes or no
86Do facial shadows have color variation? Return me yes or no
87Is there proper nose bridge definition? Return me yes or no
88Are facial pores consistently sized? Return me yes or no
89Do eyebrows have natural thickness variation? Return me yes or no
90Is there natural skin elasticity? Return me yes or no
91Are facial creases movement-appropriate? Return me yes or no
92Do teeth have natural translucency? Return me yes or no
93Is there proper cheek coloring? Return me yes or no
94Are eye bags naturally shadowed? Return me yes or no
95Do facial features maintain proper scale? Return me yes or no
96Is there natural skin undertone? Return me yes or no
97Are facial expressions muscle-consistent? Return me yes or no
98Do wrinkles have proper depth? Return me yes or no
99Is there natural facial bone structure? Return me yes or no
100Are skin textures consistently detailed? Return me yes or no
+ +![](images/5f123f529ff62ddcdbc5831ec209687692ef7f243cbe0120ead66ebf65eb50c7.jpg) +Figure 21: Diagram showing how the LLaVa generates results for detection, explanation, and detection with explanation. + +Label: real + +USER: I want you to work as an image forensic expert for Deepfake faces. Check if the image has the artifact attribute listed in the following list and ONLY return the attribute number in this image. The artifact list is [1-asymmetric eye iris; 2-irregular glasses shape or reflection; 3-irregular teeth shape or texture; 4-irregular ears or earrings; 5-strange hair texture; 6-inconsistent skin texture; 7-inconsistent lighting and shading; 8-strange background; 9-weird hands; 10-unnatural edges]. + +Detection + +LLaVa:1 + +Explanation + +USER: This image is real, tell me the reason. + +LLaVa: The reason for the image being real is related to the subject matter depicted in the image, which is a man with a beard and mustache. The man appears to be the famous actor, John Wayne, known for his distinctive facial appearance in movies. The image is a close-up, and John Wayne's facial features are clearly visible and recognizable. It's not a photo-shopped or digitally altered image. The context of the image is John Wayne talking to someone, which is also consistent with his well-known roles. + +Detection& Explanation + +USER: Is this image is real or deepfake, tell me the reason. + +LLaVa: The image is real, as it is an authentic photo of John Wayne, a famous actor known for his distinctive beard and mustache. The close-up view of John Wayne's face in the image allows for clear recognition of his facial features, making it a genuine representation of him. + +![](images/3de904cfe94667eac9ef05a1e29d1565b2fe2dddd340295346090f6b24ee50d3.jpg) +Figure 22: Diagram showing how the ours model generates results for detection, explanation, and detection with explanation. + +Label: real + +Detection + +USER:Is this image real or fake? + +OURS: This image is real. The image appears to be real, as the man's facial features, such as his nose, eyes, and mouth, are well-defined and not distorted. The skin tone appears natural, and there are no signs of blurriness or unusual lighting. The man's facial expressions and overall appearance seem genuine, making it likely that the image is a true representation of the man. And this image does not contain any blending artifacts. + +Explanation + +USER: This image is real, tell me the reason. + +OURS: The image of the man's face appears to be real, as it displays a well-defined facial structure with clear skin, prominent eyes, and a nose that is neither overly smooth nor has flat cheekbones. The man's facial features are well-defined, and there is no noticeable blurriness or distortion in the image. The overall appearance of the man's face suggests that it is a genuine representation of the individual. And this image does not contain any blending artifacts. + +Detection& Explanation + +USER: Is this image real or fake? + +Ours: This image is real. The image appears to be real, as the man's facial features, such as his nose, eyes, and mouth, are well-defined and not distorted. The skin tone appears natural, and there are no signs of blurriness or unusual lighting. The man's facial expressions and overall appearance seem genuine, making it likely that the image is a true representation of the man. And this image does not contain any blending artifacts. + +![](images/acd594d2b91d0779dcb56f92538c15beffdcc8acbf0d2cae2a0e82dcddf293b4.jpg) +Fake + +USER: Is this image real or fake? Tell the reason in detail? + +MLLM: The image is real, featuring a bald man wearing a blue shirt and a white collar. He is talking and has a mustache. The image is not photoshopped. The man appears to be a news reporter or a presenter, and his expression suggests that he is engaged in a discussion or interview. The image is a genuine representation of the man and his attire. + +![](images/21a8445f45940f6b38877747fd7cc98af6b9281fb458fcad1f7b9bfc105acb96.jpg) +Fake +Figure 23: Bad sample of Pre-trained model (part1) + +USER: This is fake image, tell the reason in detail. + +MLLM: The image is a fake or manipulated picture of a bald man with a goatee, likely created for entertainment or artistic purposes. The reason for this manipulation could be to create a humorous or unconventional visual effect, to make a statement, or to generate curiosity and discussion. The image may have been created using photoshop software, resulting in the unusual combination of a bald man with a goatee. This manipulation goes beyond the typical representation of a bald man, as it adds a distinctive element to the image that is not commonly seen in real life. + +![](images/d700527463d97623cfb3191f63ef15781e596d93a7bb952c103f7f92512f2bdb.jpg) + +USER: Is this image real or fake? Tell the reason in detail. + +MLLM: The image is real, as it shows a woman with blue eyes and red lips. The image also features a TV in the background, which is an indication that the setting is likely a TV studio or a live broadcast. The presence of the TV and the woman's makeup suggest that this is a real, live situation, rather than a fake or manipulated image. + +![](images/dd2bae790e5ff2c03f083a697f4c5499f8aad637a0ca85336d85edab34823f0a.jpg) +Fake +Figure 24: Bad sample of Pre-trained model (part2) + +USER: Why this image is fake? + +MLLM: The image is fake because the person's face appears to be photoshopped, This can be inferred from the fact that her eyes do not appear to be looking at the camera, and the lighting and background are not consistent. Additionally, the woman's face is distorted, and she has fake eyelashes and a blurry face. The face manipulation creates an unrealistic and artificial appearance, making it clear that the image is not a genuine photograph. + +![](images/ffcdece0a06507808607c5de915108802819c35b021a9839ec6cabeb315fc3eb.jpg) +Fake + +Q: What is the feature of fake in this image? + +A: In this image, the photoshop feature being used is the "face swap" technique. This technique involves replacing one person's face with another person's face, creating a blend of the two individuals. In this case, a man's face has been swapped with another man's face, resulting in a unique and unconventional appearance. And you can notice the blur in image. The image is a collage of two different people, one with a beard and the other without. creating a mix of their facial features. + +![](images/4779dd33d4a3f6d68fcb2b793f08d99833b2237a8f725bf22d7d54c63061d7e8.jpg) +Fake + +Q1: Is this image photoshopped? +Return yes or no. + +A1: Yes. + +Q4: Was man's face unconventional appearance? Return yes or no. + +A4: Yes. + +Q2: Is the "face swap" used in the image? Return yes or no. + +A2: Yes, face swap is used. + +Q5: Is there any blur in the image? +Return yes or no. + +A5: Yes, the image is blurry. + +Q3: Does the image show blend of two face? Return yes or no. + +A3: Yes, image shows a blend of two faces. + +Q6: Does the image mix facial features? Return yes or no. + +A6: Yes. + +![](images/f0bc887ad2197b8e4f4b1a796cc05f85a2acf94b68adc4bd46f8e4c535c00ef2.jpg) +Real +Figure 25: Bad sample of Pre-trained model (part3) + +Q1: Is this image photoshopped? +Return yes or no. + +A1: No. + +Q4: Was man's face unconventional appearance? Return yes or no. + +A4: No. + +Q2: Is the "face swap" technique used in this image? Return yes or no. + +A2: No, face swap is not used. + +Q5: Is there any blur in the image? Return yes or no. + +A5: Yes, the image is blurry. + +Q3: Does the image show blend of two face? Return yes or no. + +A3: No. + +Q6: Does the image mix facial features? Return yes or no. + +A6: No. + +![](images/78ed6d94cb9888257e8d04cac6add680f285f6890975b37c4ddbab4d17015c22.jpg) +Figure 26: Comparison between Pre-traind LLaVa and ours + +USER: Is this image real or fake? Tell the reason in detail? + +Pretrained MLLM: This image appears real. It shows a woman speaking into a camera with her mouth wide open, possibly in an interview setting. Deepfake images often involve more discrepancies in facial features, lighting, and overall content when compared to a regular image with a genuine, real-world scenario. The image seems to capture a natural moment, so it does not suggest any deepfake content as a possibility. + +USER: This image is fake, tell me the reasons. + +Pretrained MLLM: The image is fake because it is a manipulated photo of a woman with a different background. The woman's face is placed onto a different body, creating an unnatural and unrealistic appearance. This manipulation is commonly done for entertainment, artistic purposes, or to create a humorous effect. The image does not depict a real-life scenario, and the altered elements make it a fake representation of a person. + +USER: Is this image real or fake? + +$\chi^2$ -DFD : This image is fake. The image appears to be a deepfake as the woman's face appears to be distorted, with an unusual layout and unnatural skin tone. The facial features seem to be mismatched, and the overall appearance of the face is not consistent with a real person. And image show the obvious blending artifact. + +![](images/67d220b3a3945a9602293baaed0b74fcad310b17397e22db8925491d95db8201.jpg) +Fake + +![](images/48ec91d4a37dcd98cb0f9d4b1654ebc1d38b28a6cc59447531bcbba519c50a01.jpg) +Fake + +![](images/d2b5ece6d49513535fe571cc52b1f9229c06ad9b4ab81146dac5ad0448f8f1bc.jpg) +Real +Figure 27: Feature related questions + +Q: What is the feature of fake in this image? + +A: In this image, the photoshop feature being used is the "face swap" technique. This technique involves replacing one person's face with another person's face, creating a blend of the two individuals. In this case, a man's face has been swapped with another man's face, resulting in a unique and unconventional appearance. And you can notice the blur in image. The image is a collage of two different people, one with a beard and the other without, creating a mix of their facial features. + +Q1: Is this image photoshopped? +Return yes or no. + +A1: Yes. + +Q4: Was man's face unconventional appearance? Return yes or no. + +A4: Yes. + +Q1: Is this image photoshopped? +Return yes or no. + +A1: No. + +Q4: Was man's face unconventional appearance? Return yes or no. + +A4: No. + +Q2: Is the "face swap" used in the image? Return yes or no. + +A2: Yes, face swap is used. + +Q5: Is there any blur in the image? +Return yes or no. + +A5: Yes, the image is blurry. + +Q2: Is the "face swap" technique used in this image? Return yes or no. + +A2: No, face swap is not used. + +Q5: Is there any blur in the image? +Return yes or no. + +A5: Yes, the image is blurry. + +Q3: Does the image show blend of two face? Return yes or no. + +A3: Yes, image shows a blend of two faces. + +Q6: Does the image mix facial features? Return yes or no. + +A6: Yes. + +Q3: Does the image show blend of two face? Return yes or no. + +A3: No. + +Q6: Does the image mix facial features? Return yes or no. + +A6: No. \ No newline at end of file diff --git a/NeurIPS/2025/$_mathcal{X}^2$-DFD_ A framework for e$_mathcal{X}$plainable and e$_mathcal{X}$tendable Deepfake Detection/images.zip b/NeurIPS/2025/$_mathcal{X}^2$-DFD_ A framework for e$_mathcal{X}$plainable and e$_mathcal{X}$tendable Deepfake Detection/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f132a4b827fc9d91336e28d2607ea15f426c8653 --- /dev/null +++ b/NeurIPS/2025/$_mathcal{X}^2$-DFD_ A framework for e$_mathcal{X}$plainable and e$_mathcal{X}$tendable Deepfake Detection/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:049b07da2b98698e31381fba2a20e789343638d2204c26c7dcbd59605f7cc957 +size 3316867 diff --git a/NeurIPS/2025/$_mathcal{X}^2$-DFD_ A framework for e$_mathcal{X}$plainable and e$_mathcal{X}$tendable Deepfake Detection/layout.json b/NeurIPS/2025/$_mathcal{X}^2$-DFD_ A framework for e$_mathcal{X}$plainable and e$_mathcal{X}$tendable Deepfake Detection/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c058dd52d7525d9a6e871085b3ebf1e29b7cdefa --- /dev/null +++ b/NeurIPS/2025/$_mathcal{X}^2$-DFD_ A framework for e$_mathcal{X}$plainable and e$_mathcal{X}$tendable Deepfake Detection/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:856713a41d8ae34ba531bcc27a71af0842aeef77986fd54f07c2ab00abee9145 +size 1138415 diff --git a/NeurIPS/2025/$_mu$PC_ Scaling Predictive Coding to 100+ Layer Networks/fbf329ca-83e7-4d17-ac37-18f99b180880_content_list.json b/NeurIPS/2025/$_mu$PC_ Scaling Predictive Coding to 100+ Layer Networks/fbf329ca-83e7-4d17-ac37-18f99b180880_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d1f6b3c7611ed78f4869ef295600b6ba729ff91d --- /dev/null +++ b/NeurIPS/2025/$_mu$PC_ Scaling Predictive Coding to 100+ Layer Networks/fbf329ca-83e7-4d17-ac37-18f99b180880_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2970bb69efc1380dd85622d56e2dbf3697b4f93503b0f340161ed773c56fa31f +size 182910 diff --git a/NeurIPS/2025/$_mu$PC_ Scaling Predictive Coding to 100+ Layer Networks/fbf329ca-83e7-4d17-ac37-18f99b180880_model.json b/NeurIPS/2025/$_mu$PC_ Scaling Predictive Coding to 100+ Layer Networks/fbf329ca-83e7-4d17-ac37-18f99b180880_model.json new file mode 100644 index 0000000000000000000000000000000000000000..f7b7d1abdc4291e1c4770319c730d92e3e2bb02d --- /dev/null +++ b/NeurIPS/2025/$_mu$PC_ Scaling Predictive Coding to 100+ Layer Networks/fbf329ca-83e7-4d17-ac37-18f99b180880_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6f69d606a2599630425595ab6585e07b38387158f72363f39fb7d924a79f2fa9 +size 218599 diff --git a/NeurIPS/2025/$_mu$PC_ Scaling Predictive Coding to 100+ Layer Networks/fbf329ca-83e7-4d17-ac37-18f99b180880_origin.pdf b/NeurIPS/2025/$_mu$PC_ Scaling Predictive Coding to 100+ Layer Networks/fbf329ca-83e7-4d17-ac37-18f99b180880_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..b92656bfc9d144b03e2be5a10b2af95d2346097b --- /dev/null +++ b/NeurIPS/2025/$_mu$PC_ Scaling Predictive Coding to 100+ Layer Networks/fbf329ca-83e7-4d17-ac37-18f99b180880_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ccf0af576d32bf90246139bc3e094e8834373b60bd090d5dc2af0031b39c966f +size 19707802 diff --git a/NeurIPS/2025/$_mu$PC_ Scaling Predictive Coding to 100+ Layer Networks/full.md b/NeurIPS/2025/$_mu$PC_ Scaling Predictive Coding to 100+ Layer Networks/full.md new file mode 100644 index 0000000000000000000000000000000000000000..0c3c2ad4288cce7e761edc3a9e6ef53e573de9f1 --- /dev/null +++ b/NeurIPS/2025/$_mu$PC_ Scaling Predictive Coding to 100+ Layer Networks/full.md @@ -0,0 +1,795 @@ +# $\mu \mathbf{PC}$ : Scaling Predictive Coding to 100+ Layer Networks + +Francesco Innocenti + +School of Engineering and Informatics + +University of Sussex, UK + +F. Innocenti@sussex.ac.uk + +El Mehdi Achour + +UM6P College of Computing + +Rabat, Morocco + +elmehdi.achour@um6p.ma + +Christopher L. Buckley + +School of Engineering and Informatics + +University of Sussex, UK + +VERSES AI Research Lab + +Los Angeles, CA, USA + +c.l.buckley@sussex.ac.uk + +# Abstract + +The biological implausibility of backpropagation (BP) has motivated many alternative, brain-inspired algorithms that attempt to rely only on local information, such as predictive coding (PC) and equilibrium propagation. However, these algorithms have notoriously struggled to train very deep networks, preventing them from competing with BP in large-scale settings. Indeed, scaling PC networks (PCNs) has recently been posed as a challenge for the community [48]. Here, we show that $100+$ layer PCNs can be trained reliably using a Depth- $\mu$ P parameterisation [72, 3] which we call “ $\mu$ PC”. By analysing the scaling behaviour of PCNs, we reveal several pathologies that make standard PCNs difficult to train at large depths. We then show that, despite addressing only some of these instabilities, $\mu$ PC allows stable training of very deep (up to 128-layer) residual networks on simple classification tasks with competitive performance and little tuning compared to current benchmarks. Moreover, $\mu$ PC enables zero-shot transfer of both weight and activity learning rates across widths and depths. Our results serve as a first step towards scaling PC to more complex architectures and have implications for other local algorithms. Code for $\mu$ PC is made available as part of a JAX library for PCNs. $^{1}$ + +# 1 Introduction + +Backpropagation (BP) is arguably the core algorithm behind the success of modern AI and deep learning [52, 29]. Yet, it is widely believed that the brain cannot implement BP due to its non-local nature [34], in that the update of any weight requires knowledge of all the weights deeper or further downstream in the network. This fundamental biological implausibility of BP has motivated the study of many local algorithms, including predictive coding (PC) [37, 36, 54, 63], equilibrium propagation [59, 74], and forward learning [20], among others [33, 43, 8]. These algorithms offer the potential for more energy efficient AI and have been argued to outperform BP in more biologically relevant + +![](images/ce0cd5392ce0b9f9f8a7318a61c5b90cf9f27a5cf9cdd3aba29ea98600a56e6c.jpg) +Figure 1: $\mu$ PC enables stable training of $100+$ layer ResNets with zero-shot learning rate transfer. (Right) Test accuracy of ReLU ResNets with depths $H = \{8,16,32,64,128\}$ trained to classify MNIST for one epoch with standard PC, $\mu$ PC and BP with Depth- $\mu$ P (see §A.4 for details). Solid lines and shaded regions indicate the mean and $\pm 1$ standard deviation across 3 different random seeds. These results hold across other activation functions (see Fig. A.16). See also Figs. A.17-A.19 for asymptotic results with 128-layer ReLU networks trained for multiple epochs on MNIST, Fashion-MNIST and CIFAR10. (Left) Example of zero-shot transfer of the weight and activity learning rates from 16- to 128-layer Tanh networks. See Figs. 5 & A.31-A.32 for an explanation and the complete transfer results across widths as well as depths. + +settings such as online and continual learning [61]. However, local learning rules have notoriously struggled to train large and especially deep models on the scale of modern AI applications.2 + +For the first time, we show that very deep $(100+$ layer) networks can be trained reliably using a Depth- $\mu \mathrm{P}$ -inspired parameterisation [72, 3] of PC which we call “ $\mu \mathrm{PC}$ ” (Fig. 1). To our knowledge, no networks of such depth have been trained before with a local algorithm. Indeed, this has recently been posed as a challenge for the PC community [48]. We start by showing that the standard parameterisation of PC networks (PCNs) is inherently unscalable in that (i) the inference landscape becomes increasingly ill-conditioned with model size and training time, and (ii) the forward initialisation of the activities vanishes or explodes with the depth. We then show that, despite addressing only the second instability, $\mu \mathrm{PC}$ is capable of training up to 128-layer fully connected residual networks (ResNets) on standard classification tasks with competitive performance and little tuning compared to current benchmarks (Fig. 1). Moreover, $\mu \mathrm{PC}$ enables zero-shot transfer of both the weight and activity learning rates across widths and depths (Fig. 5). We make code for $\mu \mathrm{PC}$ available as part of a JAX library for PCNs at https://github.com/thebuckleylab/jpc [23]. + +The rest of the paper is structured as follows. Following a brief review of the maximal update parameterisation $(\mu \mathrm{P})$ and PCNs (§2), Section 3 exposes two distinct pathologies in standard PCNs which make training at large scale practically impossible. Motivated by these findings, we then suggest a minimal set of desiderata for a more scalable PCN parameterisation (§4). Section 5 presents experiments with $\mu \mathrm{PC}$ , and Section 6 studies a specific regime where $\mu \mathrm{PC}$ converges to BP. We conclude with the limitations of this work and promising directions for future research (§7). For space reasons, we include related work and additional experiments in Appendix A, along with derivations, experimental details and supplementary figures. + +# 1.1 Summary of contributions + +- We show that $\mu \mathrm{PC}$ , which reparameterises PCNs using Depth- $\mu \mathrm{P}$ [72, 3], allows stable training of very deep (100+ layer) ResNets on simple classification tasks with competitive performance and little tuning compared to current benchmarks [48] (Figs. 1 & A.17-A.18). +- $\mu$ PC also empirically enables zero-shot transfer of both the weight and activity learning rates across widths and depths (Figs. 5 & A.31-A.32). +- We achieve these results by a theoretical and empirical analysis of the scaling behaviour of the inference landscape and dynamics of PCNs (§3), revealing the following two pathologies: +- the inference landscape becomes increasingly ill-conditioned with model size (Fig. 2) and training time (Fig. 3) ( $\S 3.1$ ); and +- the forward pass of standard PCNs vanishes or explodes with the depth (§3.2). +- To address these instabilities, we propose a minimal set of desiderata that PCNs should aim to satisfy to be trainable at scale (§4), revealing an apparent trade-off between the conditioning of the inference landscape and the stability of the forward pass (Fig. 4). This analysis can be applied to other inference-based algorithms (§A.2.5). +- To better understand $\mu \mathrm{PC}$ , we study a theoretical regime where the $\mu \mathrm{PC}$ energy converges to the mean squared error (MSE) loss and so PC effectively implements BP (Theorem 1, Fig. 6). However, we find that $\mu \mathrm{PC}$ can successfully train deep networks far from this regime. + +# 2 Background + +# 2.1 The maximal update parameterisation $(\mu \mathbf{P})$ + +The maximal update parameterisation was first introduced by [70] to ensure that the order of the activation or feature updates at each layer remains stable with the width $N$ . This was motivated by the lack of feature learning in the neural tangent kernel or "lazy" regime [27], where the activations remain practically unchanged during training [6, 31]. More formally, $\mu \mathrm{P}$ can be derived from the following 3 desiderata [70]: (i) the layer preactivations are $\mathcal{O}_N(1)$ at initialisation, (ii) the network output is $\mathcal{O}_N(1)$ during training, and (iii) the layer features are also $\mathcal{O}_N(1)$ during training. + +Satisfying these desiderata boils down to solving a system of equations for a set of scalars (commonly referred to as "abcd") parameterising the layer transformation, the (Gaussian) initialisation variance, and the learning rate [71, 44]. Different optimisers and types of layer lead to different scalings. One version of $\mu \mathrm{P}$ (and the version we will be using here) initialises all the weights from a standard Gaussian and rescales each layer transformation by $1 / \sqrt{N_{\ell - 1}}$ , with the exception of the output which is scaled by $1 / N_{L - 1}$ . Remarkably, $\mu \mathrm{P}$ allows not only for more stable training dynamics but also for zero-shot hyperparameter transfer: tuning a small model parameterised with $\mu \mathrm{P}$ guarantees that optimal hyperparameters such as the learning rate will transfer to a wider model [69, 42]. + +More recently, $\mu \mathrm{P}$ has been extended to depth for ResNets ("Depth- $\mu \mathrm{P}$ " [72, 3], such that transfer is also conserved across depths $L$ . This is done by mainly introducing a $1 / \sqrt{L}$ scaling before each residual block. Extensions of standard $\mu \mathrm{P}$ for other algorithms have also been proposed [25, 26, 14, 9]. + +# 2.2 Predictive coding networks (PCNs) + +We consider the following general parameterisation of the energy function of $L$ -layered PCNs [5]: + +$$ +\mathcal {F} = \sum_ {\ell = 1} ^ {L} \frac {1}{2} \left\| \mathbf {z} _ {\ell} - a _ {\ell} \mathbf {W} _ {\ell} \phi_ {\ell} \left(\mathbf {z} _ {\ell - 1}\right) - \tau_ {\ell} \mathbf {z} _ {\ell - 1} \right\| ^ {2} \tag {1} +$$ + +with weights $\mathbf{W}_{\ell} \in \mathbb{R}^{N_{\ell} \times N_{\ell-1}}$ , activities $\mathbf{z}_{\ell} \in \mathbb{R}^{N_{\ell}}$ and activation function $\phi_{\ell}(\cdot)$ . Dense weight matrices could be replaced by convolutions, all assumed to be initialised i.i.d. from a Gaussian $(\mathbf{W}_{\ell})_{ij} \sim \mathcal{N}(0, b_{\ell})$ with variance scaled by $b_{\ell}$ . We omit multiple data samples to simplify the notation, and ignore biases since they do not affect the main analysis, as explained in §A.2.1. We also add scalings $a_{\ell} \in \mathbb{R}$ and optional skip or residual connections set by $\tau_{\ell} \in \{0, 1\}$ . + +![](images/7815c82475bddf592e5ae1518b71b519fc37f46799bb2cc3c06ccfa306ffd64a.jpg) +Figure 2: Wider and particulary deeper PCNs have a more ill-conditioned inference landscape. We plot the condition number of the activity Hessian $\kappa (\mathbf{H}_{\mathbf{z}})$ (lower is better) of randomly initialised fully connected networks as a function of the width $N$ and depth $H$ (see $\S A.4$ for details). Insets show 2D projections of the landscape of selected networks around the linear solution (Eq. 4) along the maximum and minimum eigenvectors of the Hessian $\mathcal{F}(\mathbf{z}^{*} + \alpha \hat{\mathbf{v}}_{\mathrm{min}} + \beta \hat{\mathbf{v}}_{\mathrm{max}})$ . Note that the ill-conditioning is much more extreme for ResNets (see Fig. A.22). Results were similar across different seeds. + +![](images/0645e3a05bbaff5bbe96998cb264a69720a83e31a6f44cd098065e64c0bec956.jpg) + +![](images/52b8d4edbe1adedb133f83aab7dd70421c39f327434c9cc3e38cb7fdc77c8431.jpg) + +The energy of the last layer is defined as $\mathcal{F}_L = \frac{1}{2} ||\mathbf{z}_L - a_L\mathbf{W}_L\phi_L(\mathbf{z}_{L - 1})||^2$ for some target $\mathbf{z}_L\coloneqq \mathbf{y}\in \mathbb{R}^{d_{\mathrm{out}}}$ , while the energy of the first layer is $\mathcal{F}_1 = \frac{1}{2} ||\mathbf{z}_1 - a_1\mathbf{W}_1\mathbf{z}_0||^2$ , with some optional input $\mathbf{z}_0\coloneqq \mathbf{x}\in \mathbb{R}^{d_{\mathrm{in}}}$ for supervised (vs unsupervised) training. We will refer to PC or SP as the "standard parameterisation" with unit premultipliers $a_{\ell} = 1$ for all $\ell$ and standard initialisations [30, 11, 18] such as $b_{\ell} = 1 / N_{\ell -1}$ , and to $\mu \mathrm{PC}$ as that which uses (some of) the scalings of Depth- $\mu \mathrm{P}$ ( $\S 2.1$ ). See Table 1 for a summary. + +We fix the width of all the hidden layers $N = N_{1} = \dots = N_{H}$ where $H = L - 1$ is the number of hidden layers. We use $\pmb{\theta} \coloneqq \{\mathrm{vec}(\mathbf{W}_{\ell})\}_{\ell=1}^{L} \in \mathbb{R}^{p}$ to represent all the weights with $p$ as the total number of parameters and $\mathbf{z} \coloneqq \{\mathbf{z}_{\ell}\}_{\ell=1}^{H} \in \mathbb{R}^{NH}$ to denote all the activities free to vary. Note that, depending on the context, we will use both $H$ and $L$ to refer to the network depth. + +PCNs are trained by minimising the energy (Eq. 1) in two separate phases: first with respect to the activities (inference) and then with respect to the weights (learning), + +Infer: + +minF + +(2) + +Learn: $\min_{\theta} \mathcal{F}$ . + +(3) + +Inference acts on a single data point and is generally performed by gradient descent (GD), $\mathbf{z}_{t + 1} = \mathbf{z}_t - \beta \nabla_{\mathbf{z}}\mathcal{F}$ with step size $\beta$ . The weights are often updated at numerical convergence of the inference dynamics, when $\nabla_{\mathbf{z}}\mathcal{F} \approx 0$ . Our theoretical results will mainly address the first optimisation problem (Eq. 2), namely the inference landscape and dynamics, but we discuss and numerically investigate the impact on the learning dynamics (Eq. 3) wherever relevant. + +# 3 Instability of the standard PCN parameterisation + +In this section, we reveal through both theory and experiment that the standard parameterisation (SP) of PCNs suffers from two instabilities that make training and convergence of the PC inference dynamics (Eq. 2) at large scale practically impossible. First, the inference landscape of standard PCNs becomes increasingly ill-conditioned with model size and training time (§3.1). Second, depending + +![](images/01465e4c303b5a394438d53b14afc7a7bdd580302fae1d1e0cb8b6a55f240b85.jpg) + +![](images/c68a7dcb824c82f3acc67ec451000d3ecb521184e381484ad8f76550595785ca.jpg) + +![](images/282d079701bcc3f116af80c6acb551cd06b62251c420029f03d5eca6ed18d585.jpg) + +![](images/ae1043e249c99870a0dac696d9bed41974b3a88e79f7cbf842643daf31ea5782.jpg) +Figure 3: The inference landscape of PCNs grows increasingly ill-conditioned with training. We plot the condition number of the activity Hessian (Eq. 5) (top) as well as test accuracies (bottom) for fully connected networks of depths $H \in \{8, 16, 32\}$ during one epoch of training. All networks had width $N = 128$ and were trained to classify MNIST (see §A.4 for more details). Similar results are observed for ResNets (Fig. A.9) and Fashion-MNIST (Fig. A.23). Solid lines and shaded regions indicate the mean and standard deviation over 3 random seeds. + +![](images/51a202075ae4bc708d983976fbb6ee953db20e25e06affb79dcc63efe3eb21a8.jpg) + +![](images/7e202f49cc69f81e666358a42ed118586caad9ba4d097b3d18cc112d2744c621.jpg) + +on the model, the feedforward pass either vanishes or explodes with the depth (§3.2). The second problem is shared with BP-trained networks, while the first instability is unique to PC and likely any other algorithm performing inference minimisation (§A.2.5). + +# 3.1 Ill-conditioning of the inference landscape + +Here we show that the inference landscape of standard PCNs becomes increasingly ill-conditioned with network width, depth and training time. As reviewed in §2.2, the inference phase of PC (Eq. 2) is commonly performed by GD. For a deep linear network (DLN, Eq. 1 with $\phi_{\ell} = \mathbf{I}$ for all $\ell$ ), one can solve for the activities in closed form as shown by [26], + +$$ +\nabla_ {\mathbf {z}} \mathcal {F} = \mathbf {H} _ {\mathbf {z}} \mathbf {z} - \mathbf {b} = 0 \quad \Longrightarrow \quad \mathbf {z} ^ {*} = \mathbf {H} _ {\mathbf {z}} ^ {- 1} \mathbf {b} \tag {4} +$$ + +where $(\mathbf{H}_{\mathbf{z}})_{\ell k} \coloneqq \partial^{2}\mathcal{F} / \partial \mathbf{z}_{\ell}\partial \mathbf{z}_{k} \in \mathbb{R}^{(NH)\times (NH)}$ is the Hessian of the energy with respect to the activities, and $\mathbf{b} \in \mathbb{R}^{NH}$ is a sparse vector depending only on the data and associated weights (see §A.2.1 for details). Eq. 4 shows that for a DLN, PC inference is a well-determined linear problem. $^6$ + +For arbitrary DLNs, one can also prove that the inference landscape is strictly convex as the Hessian is positive definite $^7$ , $\mathbf{H}_{\mathbf{z}} \succ 0$ (Theorem A.1; see §A.2.2 for proof). This makes intuitive sense since the energy (Eq. 1) is quadratic in $\mathbf{z}$ . The result is empirically verified for DLNs in Figs. A.5-A.7 and appears to generally hold for nonlinear networks (see Figs. A.7 & A.22). + +For such convex problems, the convergence rate of GD is known to be given by the condition number of the Hessian [4, 41], $\kappa (\mathbf{H}_{\mathbf{z}}) = |\lambda_{\mathrm{max}}| / |\lambda_{\mathrm{min}}|$ . Intuitively, the higher the condition number, the more elliptic the level sets of the energy $\mathcal{F}(\mathbf{z})$ become, and the more iterations GD will need to reach the solution (see Fig. A.21), with the step size bounded by the highest curvature direction $\beta < 2 / \lambda_{\mathrm{max}}$ (see Fig. A.10 for an example). For non-convex problems, it can still be useful to have a notion of local conditioning [e.g. 73]. + +What determines the condition number of $\mathbf{H}_{\mathbf{z}}$ ? Looking more closely at the structure of the Hessian + +$$ +\frac {\partial^ {2} \mathcal {F}}{\partial \mathbf {z} _ {\ell} \partial \mathbf {z} _ {k}} = \left\{ \begin{array}{l l} \mathbf {I} + a _ {\ell + 1} ^ {2} \mathbf {W} _ {\ell + 1} ^ {T} \mathbf {W} _ {\ell + 1}, & \ell = k \\ - a _ {k + 1} \mathbf {W} _ {k + 1}, & \ell - k = 1 \\ - a _ {\ell + 1} \mathbf {W} _ {\ell + 1} ^ {T}, & \ell - k = - 1 \\ \mathbf {0}, & \text {e l s e} \end{array} , \right. \tag {5} +$$ + +one realises that it depends on two main factors: (i) the network architecture, including the width $N$ , depth $L$ and connectivity; and (ii) the value of the weights at any time during training $\theta_{t}$ . We first find that the inference landscape of standard PCNs becomes increasingly ill-conditioned with the width and particularly depth (Fig. 2), and extremely so for ResNets (Fig. A.22). See also §A.2.3 for a random matrix theory analysis of the scaling behaviour of the initialised Hessian eigenspectrum with $N$ and $L$ . In addition, we observe that the ill-conditioning grows and spikes during training (Figs. 3, A.9, A.23 & A.25), and using an adaptive optimiser such as Adam [28] does not seem to help (Figs. A.8 & A.24). Together, these findings help to explain why the convergence of the GD inference dynamics (Eq. 2) can dramatically slow down on deeper models [23, 48], while also highlighting that small inference gradients—which are commonly used to determine convergence—do not necessarily imply closeness to a solution. + +# 3.2 Vanishing/exploding forward pass + +In the previous section (§3.1), we saw that the growing ill-conditioning of the inference landscape with the model size and training time is one likely reason for the challenging training of PCNs at large scale. Another reason—and as we will see the key reason—is that the forward initialisation of the activities can vanish or explode with the depth. This is a classic finding in the neural network literature that has been surprisingly ignored for PCNs. For fully connected networks with standard initialisations [30, 11, 18], the forward pass vanishes with the depth, leading to vanishing gradients. This issue can be addressed with residual connections [19] and various forms of activity normalisation [24, 1], both of which remain key components of the modern transformer block [64]. + +However, while there have been attempts to train ResNets with PC [48], they have been without activity normalisation. This is likely because any kind of normalisation of the activities seems at odds with convergence of the inference dynamics to a solution (Eq. 2). Without normalisation, however, the activations (and gradients) of vanilla ResNets explode with the depth (see Fig. A.30). A potential remedy would be to normalise only the forward pass, but here we will aim to take advantage of more principled approaches with stronger guarantees about the stability of the forward pass (\$4). + +# 4 Desiderata for stable PCN parameterisation + +In §3, we exposed two main pathologies in the scaling behaviour of standard PCNs: (i) the growing ill-conditioning of the inference landscape with model size and training time (§3.1), and (ii) the instability of the forward pass with depth (§3.2). These instabilities motivate us to specify a minimal set of desiderata that we would like a PCN to satisfy to be trainable at large scale.8 + +Desideratum 1. Stable forward pass at initialisation. At initialisation, all the layer preactivations are stable independent of the network width and depth, $||\mathbf{z}_{\ell}|| \sim \mathcal{O}_{N,H}(1)$ for all $\ell$ , where $\mathbf{z}_{\ell} = h_{\ell}(\ldots h_{1}(\mathbf{x}))$ with $h_\ell (\cdot)$ as the map relating one layer to the next. + +To our knowledge, there are two approaches that provide strong theoretical guarantees about this desideratum: (i) orthogonal weight initialisation for both fully connected [58, 46, 47, 68] and convolutional networks [68], ensuring that $\mathbf{W}_{\ell}^{T}\mathbf{W}_{\ell} = \mathbf{I}$ at every layer $\ell$ ; and (ii) the recent Depth- $\mu$ P parameterisation [72, 3] (see §2.1 for a review). For a replication of these results, see Fig. A.30. To apply Depth- $\mu$ P to PC, we simply reparameterise the PC energy for ResNets (Eq. 1 with $\tau_{\ell} = 1$ for $\ell = 2, \ldots, H$ and $\tau_{\ell} = 0$ otherwise) with the layer scalings of Depth- $\mu$ P (see Table 1). We call this reparameterisation $\mu$ PC. + +![](images/26a641d34078dc8c24313b4481f895fc1dbc43bc5539274d2e4f6ecc14629693.jpg) +Figure 4: Parameterisations with stable forward passes induce highly ill-conditioned inference landscapes with depth. We plot the conditioning of the activity Hessian of randomly initialised networks over width $N$ and depth $H$ for the $\mu \mathrm{PC}$ and orthogonal parameterisations. Networks with and without residual connections were used for these respective parameterisations. Note that ReLU networks with orthogonal initialisation cannot achieve stable forward passes (see Fig. A.30). Results were similar across different seeds. + +Table 1: Summary of parameterisations. Standard PC has unit layer premultipliers and weights initialised from a Gaussian with variance scaled by the input width at every layer $N_{\ell -1}$ . $\mu \mathrm{PC}$ uses a standard Gaussian initialisation and adds width- and depth-dependent scalings at every layer. + +
a1(input weights)aℓ(hidden weights)aL(output weights)bℓ(init. variance)
PC111N-1ℓ-1
μPCN-1/20(Nℓ-1L)-1/2N-1L-11
+ +We would like Desideratum 1 to hold throughout training as we state in the following desideratum. + +Desideratum 2. Stable forward pass during training. The forward pass is stable during training such that Desideratum 1 is true for all training steps $t = 1, \ldots, T$ . + +Depth- $\mu$ P ensures this desideratum for BP, but we do not know whether the same will apply to $\mu$ PC. We return to this point in §7. For the orthogonal parameterisation, the weights should remain orthogonal during training to satisfy Desideratum 2, which could be encouraged with some kind of regulariser. Next, we address the ill-conditioning of the inference landscape (§3.1), again first at initialisation. + +Desideratum 3. Stable conditioning of the inference landscape at initialisation. The condition number of the activity Hessian (Eq. 5) at initialisation stays constant with the network width and depth, $\kappa (\mathbf{H}_{\mathbf{z}})\sim \mathcal{O}_{N,H}(1)$ . + +Ideally, we would like the PC inference landscape to be perfectly conditioned, i.e. $\kappa (\mathbf{H}_{\mathbf{z}}) = 1$ . However, this cannot be achieved without zeroing out the weights, $\mathbf{H}_{\mathbf{z}}(\theta = \mathbf{0}) = \mathbf{I}$ , since the Hessian is symmetric and so it can only have all unit eigenvalues if it is the identity. Starting with small weights $(\mathbf{W}_{\ell})_{ij} \ll 1$ at the cost of slightly imperfect conditioning is not a solution, since the forward pass vanishes, thus violating Desideratum 1. See §A.3.3 for another intervention that appears to come at the expense of performance. + +What about the above parameterisations ensuring stable forward passes? Interestingly, both orthogonal initialisation and $\mu \mathrm{PC}$ induce highly ill-conditioned inference landscapes with the depth (Fig. 4), similar to standard PC ResNets (Fig. A.22). This highlights a potential trade-off between the stability of the forward pass (technically, the conditioning of the input-output Jacobian) and the conditioning of the activity Hessian. Because PCNs with ill-conditioned inference landscapes can still be trained (e.g. see Fig. 3), we will choose to satisfy Desideratum 1 at the expense of Desideratum 3, while seeking to prevent the condition number from exploding during training. + +Desideratum 4. Stable conditioning of the inference landscape during training. The condition number of the activity Hessian (Eq. 5) is stable throughout training such that $\kappa(\mathbf{H}_{\mathbf{z}}(t)) \approx \kappa(\mathbf{H}_{\mathbf{z}}(t - 1))$ for all training steps $t = 1, \ldots, T$ . + +# 5 Experiments + +We performed experiments with parameterisations ensuring stable forward passes at initialisation (Desideratum 1), namely $\mu \mathrm{PC}$ and orthogonal, despite their inability to solve the ill-conditioning of the inference landscape with depth (Desideratum 3; Fig. 4). Due to limited space, we report results only for $\mu \mathrm{PC}$ since orthogonal initialisation was not found to be as effective (see §A.3.4). We trained fully connected residual PCNs on standard image classification tasks (MNIST, Fashion-MNIST and CIFAR10). This simple setup was chosen because the main goal was to test whether $\mu \mathrm{PC}$ is capable of training deep PCNs—a task that has proved challenging with more complex datasets and architectures [48]. We note that all the networks used as many inference steps as hidden layers (see Figs. A.14 & A.27 for results with one step). + +First, we trained ResNets of varying depth (up to 128 layers) to classify MNIST for a single epoch. Remarkably, we find that $\mu \mathrm{PC}$ allows stable training of networks of all depths across different activation functions (Figs. 1 & A.16). These networks were tuned only for the weight and activity learning rates, with no other optimisation techniques such as momentum, weight decay, and nudging, as used in previous studies [48]. Competitive performance ( $\approx 98\%$ ) is achieved in 5 epochs (Fig. A.17), $5\times$ faster than the current benchmark [48]. Similar results are observed on Fashion-MNIST, where competitive accuracy ( $\approx 89\%$ ) is reached in fewer than 15 epochs (Fig. A.18). On CIFAR10, performance is far from SOTA because of the fully connected (as opposed to convolutional) architectures used, but $\mu \mathrm{PC}$ remains trainable at large depth (Fig. A.19). + +![](images/fe4dce0f2b16d9ed8ad0a54bdc1720599158db7e77094788cc7a2c568994b126.jpg) +Figure 5: $\mu$ PC enables zero-shot transfer of the weight and activity learning rates across widths $N$ and depths $H$ . Minimum training loss (log) achieved by ResNets of varying width and depth trained with $\mu$ PC on MNIST across different weight and activity learning rates. All networks had Tanh as nonlinearity (see Figs. A.31-A.32 for other activation functions); those with varying width (first row) had 8 hidden layers, and those with varying the depth (second row) had 512 hidden units (see §A.4 for details). Each contour was averaged over 3 random seeds. + +Strikingly, we also find that $\mu \mathrm{PC}$ enables zero-shot transfer of both the weight and activity learning rates across widths and depths (Figs. 5 & A.31-A.32), consistent with recent results with Depth- $\mu \mathrm{P}$ [72, 3]. This means that one can tune a small PCN and then transfer the optimal learning rates to wider and/or deeper PCNs—a process that is particularly costly for PC since it requires two separate learning rates. In fact, this is precisely how we obtained the Fashion-MNIST (Fig. A.18) and CIFAR10 (Fig. A.19) results: by performing transfer from 8- to 128-layer networks, avoiding the expensive tuning at large scale. + +# 6 Is $\mu$ PC BP? + +Why does $\mu \mathrm{PC}$ seem to work so well despite failing to solve the ill-conditioning of the inference landscape with depth (Fig. 4)? Depth- $\mu \mathrm{P}$ also satisfies other, BP-specific desiderata that PC might not require or benefit from. Here we show that while there is a practical regime where $\mu \mathrm{PC}$ approximates BP, it turns out to be brittle, and so BP cannot explain the success of $\mu \mathrm{PC}$ (at least on the tasks considered). In particular, it is possible to show that, when the width is much larger than the depth $N \gg L$ , at initialisation the $\mu \mathrm{PC}$ energy at the inference equilibrium converges to the MSE loss. In this regime, PC computes the same gradients as BP and all the Depth- $\mu \mathrm{P}$ theory applies. + +Theorem 1 (Limit Convergence of $\mu$ PC to BP.). Let $\mathcal{F}_{\mu PC}(\pmb{\theta},\mathbf{z})$ be the PC energy of a randomly initialised linear ResNet (Eq. 1 with $\tau_{\ell} = 1$ for $\ell = 2,\dots ,H$ and $\tau_{\ell} = 0$ otherwise) parameterised with Depth- $\mu P$ (Table 1) and $\mathcal{L}_{\mu P}(\pmb {\theta})$ its corresponding MSE loss. Then, as the aspect ratio of the network $r\coloneqq L / N$ vanishes, the equilibrated energy (Eq. 31) converges to the loss (see §A.2.6 for proof) + +$$ +r \rightarrow 0, \quad \mathcal {F} _ {\mu P C} (\boldsymbol {\theta}, \mathbf {z} ^ {*}) = \mathcal {L} _ {\mu P} (\boldsymbol {\theta}). \tag {6} +$$ + +The result relies on a recent derivation of the equilibrated energy as a rescaled MSE loss for DLNs [22]. We simply extend this to linear ResNets and show that the rescaling approaches the identity with $\mu \mathrm{PC}$ in the above limit. Fig. 6 shows that the result holds at initialisation $(t = 0)$ , with the equilibrated energy converging to the loss when the width is around $32\times$ the depth. (Note that the deepest networks $(H = 128, N = 512)$ we tested in the previous section had a much smaller aspect ratio, $r = 4$ .) Nevertheless, we observe that the equilibrated energy starts to diverge from the loss with training at large width and depth (Fig. 6). Note also that we do not know the inference solution for nonlinear networks. We therefore leave further theoretical study of $\mu \mathrm{PC}$ to future work. See also §A.1 for a discussion of how Theorem 1 relates to previous correspondences between PC and BP. + +![](images/6f40c31ef2a7bddef5dd9eb6faf725c6e066a3706ca18b39ec91423ab88b976d.jpg) +Figure 6: Convergence/Divergence of $\mu \mathbf{PC}$ to BP for linear ResNets. To verify Theorem 1 (Eq. 6), we plot the ratio between the MSE loss and the equilibrated $\mu \mathrm{PC}$ energy of linear ResNets (Eq. 31) at different training points $t$ as a function of the width $N$ and depth $H$ (see §A.4 for details). We observe that while at initialisation ( $t = 0$ ) the equilibrated energy converges to the loss as the the width grows relative to the depth (verifying Theorem 1), the correspondence breaks down with training at large depth and width. Results were similar across different runs. + +# 7 Discussion + +In summary, we showed that it is possible to reliably train very deep $(100+$ layer) networks with a local learning algorithm. We achieved this via a Depth- $\mu \mathrm{P}$ -like reparameterisation of PCNs which we labelled $\mu \mathrm{PC}$ . We found that $\mu \mathrm{PC}$ is capable of training very deep networks with little tuning and competitive performance on simple classification tasks (Fig. 1), while also enabling zero-shot transfer of weight and activity learning rates across widths and depths (Fig. 5). + +$\mu \mathrm{PC}$ and inference ill-conditioning. Despite its relative success, $\mu \mathrm{PC}$ failed to solve the growing ill-conditioning of the inference landscape with the network depth (Desideratum 3; Fig. 4). This can be explained by two additional findings. First, the forward pass of $\mu \mathrm{PC}$ seems to initialise the activities much closer to the analytical solution (Eq. 4) for DLNs than standard PC (Fig. A.35). Second, training $\mu \mathrm{PC}$ networks with a single inference step (as opposed to as many as hidden layers) led to performance degradation not only during training, but also with depth (Figs. A.14 & A.27). Together, these results suggest that a stable forward pass, as ensured by $\mu \mathrm{PC}$ , is critical not only for performance but also for dealing with landscape ill-conditioning, by initialising the activities closer to a solution such that only a few (empirically determined) inference steps are needed. This is also consistent with the finding that while inference convergence is necessary for successful training of the SP, it does not appear sufficient for good generalisation (see §A.3.6). It would be interesting to study $\mu \mathrm{PC}$ in more detail in linear networks given their analytical tractability. + +Another recent study investigated the problem of training deep PCNs [12], showing an exponential decay in the activity gradients over depth. This result can be seen as a consequence of the ill-conditioning of the inference landscape with depth (Fig. 2), since flat regions where the forward pass seems to initialise the activities (see §A.3.2) have small gradients, and depth drives ill-conditioning. [12] proposed a reparameterisation of PCNs leveraging BP for faster inference convergence on GPUs, and it could be interesting to combine this approach with $\mu$ PC, especially for generation tasks or more complex datasets where more inference steps might be necessary for good performance. + +$\mu \mathrm{PC}$ and the other Desiderata. Did $\mu \mathrm{PC}$ satisfy some other Desiderata (§4) besides the stability of the forward pass at initialisation (Desideratum 1)? When experimenting with $\mu \mathrm{PC}$ , we tried including the Depth- $\mu \mathrm{P}$ scalings only in the forward pass (i.e. removing them from the energy or even just the inference or weight gradients). However, this always led to non-trainable networks even at small depths, suggesting that the Depth- $\mu \mathrm{P}$ scalings are also beneficial for the PC inference and learning dynamics and that the resulting updates are likely to keep the forward pass stable during training (Desideratum 2). Deriving principled scalings specific to PC could help explain these findings or even lead to better scalings. Finally, $\mu \mathrm{PC}$ did not seem to prevent the ill-conditioning of the inference landscape from growing with training (see Figs. A.28 & A.29), thus violating Desideratum 4. + +Is $\mu \mathrm{PC}$ optimal? $\mu \mathrm{PC}$ unlikely to be the optimal parameterisation for PCNs. This is because we adapted, rather than derived, principled (Depth- $\mu \mathrm{P}$ ) scalings for BP, with only guarantees about the stability of the forward pass. Indeed, we did not rescale the learning rate of Adam (used in all our experiments) by $\sqrt{NL}$ as prescribed by Depth- $\mu \mathrm{P}$ [72], since this scaling always led to non-trainable networks. We note that depth transfer has also been achieved without this scaling [3, 42] and that the optimal depth scaling is still an active area of research [10]. It would also be useful to better understand the relationship between $\mu \mathrm{PC}$ and the (width-only) $\mu \mathrm{P}$ parameterisation for PC proposed by [26] (see §A.1 for a comparison). More generally, it would therefore be potentially impactful to derive principled scalings specific to PC. While an analysis far from inference equilibrium appears challenging, one could start with the order of the weight updates of the equilibrated energy of linear ResNets (Eq. 31). + +Other future directions. Given the recent successful application of Depth- $\mu$ P to convolutional networks and transformers [3, 42], it would be interesting to investigate whether these more complex architectures can be successfully trained on large-scale datasets with $\mu$ PC. Our analysis of the inference landscape can also be applied to any other algorithm performing some kind of inference minimisation (see §A.2.5 for a preliminary investigation of equilibrium propagation), and it could be interesting to see whether these algorithms could also benefit from $\mu$ P-like parameterisation. + +# Acknowledgements + +FI is funded by the Sussex Neuroscience 4-year PhD Programme. EMA acknowledges funding by UM6P and the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) - Project number 442047500 through the Collaborative Research Center "Sparsity and Singular Structures" (SFB 1481) as he started this project at RWTH Aachen University. CLB was partially supported by the European Innovation Council (EIC) Pathfinder Challenges, Project METATOOL with Grant Agreement (ID: 101070940). FI would like to thank Alexandru Meterez and Lorenzo Noci for their help in better understanding $\mu$ P, and Ivor Simpson for providing access to GPUs used to run some of the experiments. + +# References + +[1] J. L. Ba, J. R. Kiros, and G. E. Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016. +[2] N. P. Baskerville, J. P. Keating, F. Mezzadri, J. Najnudel, and D. Granziol. Universal characteristics of deep neural network loss surfaces from random matrix theory. Journal of Physics A: Mathematical and Theoretical, 55(49):494002, 2022. +[3] B. Bordelon, L. Noci, M. B. Li, B. Hanin, and C. Pehlevan. Depthwise hyperparameter transfer in residual networks: Dynamics and scaling limit. arXiv preprint arXiv:2309.16620, 2023. +[4] S. Boyd and L. Vandenberghe. Convex optimization. Cambridge university press, 2004. +[5] C. L. Buckley, C. S. Kim, S. McGregor, and A. K. Seth. The free energy principle for action and perception: A mathematical review. Journal of Mathematical Psychology, 81:55-79, 2017. +[6] L. Chizat, E. Oyallon, and F. Bach. On lazy training in differentiable programming. Advances in neural information processing systems, 32, 2019. +[7] A. Choromanska, M. Henaff, M. Mathieu, G. B. Arous, and Y. LeCun. The loss surfaces of multilayer networks. In Artificial intelligence and statistics, pages 192-204. PMLR, 2015. +[8] G. Dellaferrella and G. Kreiman. Error-driven input modulation: solving the credit assignment problem without a backward pass. In International Conference on Machine Learning, pages 4937-4955. PMLR, 2022. +[9] N. Dey, S. Bergsma, and J. Hestness. Sparse maximal update parameterization: A holistic approach to sparse training dynamics. arXiv preprint arXiv:2405.15743, 2024. +[10] N. Dey, B. C. Zhang, L. Noci, M. Li, B. Bordelon, S. Bergsma, C. Pehlevan, B. Hanin, and J. Hestness. Don't be lazy: Completestep enables compute-efficient deep transformers. arXiv preprint arXiv:2505.01618, 2025. +[11] X. Glorot and Y. Bengio. Understanding the difficulty of training deep feedforward neural networks. In Proceedings of the thirteenth international conference on artificial intelligence and statistics, pages 249-256. JMLR Workshop and Conference Proceedings, 2010. +[12] C. Goemaere, G. Oliviers, R. Bogacz, and T. Demeester. Error optimization: Overcoming exponential signal decay in deep predictive coding networks. arXiv preprint arXiv:2505.20137, 2025. +[13] D. Granziol. Beyond random matrix theory for deep networks. arXiv preprint arXiv:2006.07721, 2020. +[14] M. Haas, J. Xu, V. Cevher, and L. C. Vankadara. Effective sharpness aware minimization requires layerwise perturbation scaling. In High-dimensional Learning Dynamics 2024: The Emergence of Structure and Reasoning, 2024. +[15] M. Hardt and T. Ma. Identity matters in deep learning. arXiv preprint arXiv:1611.04231, 2016. + +[16] S. Hayou. Commutative scaling of width and depth in deep neural networks. Journal of Machine Learning Research, 25(299):1-41, 2024. +[17] S. Hayou and G. Yang. Width and depth limits commute in residual networks. In International Conference on Machine Learning, pages 12700-12723. PMLR, 2023. +[18] K. He, X. Zhang, S. Ren, and J. Sun. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In Proceedings of the IEEE international conference on computer vision, pages 1026-1034, 2015. +[19] K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. +[20] G. Hinton. The forward-forward algorithm: Some preliminary investigations. arXiv preprint arXiv:2212.13345, 2022. +[21] R. A. Horn and C. R. Johnson. Matrix analysis. Cambridge university press, 2012. +[22] F. Innocenti, E. M. Achour, R. Singh, and C. L. Buckley. Only strict saddles in the energy landscape of predictive coding networks? Advances in Neural Information Processing Systems, 37:53649-53683, 2025. +[23] F. Innocenti, P. Kinghorn, W. Yun-Farmbrough, M. D. L. Varona, R. Singh, and C. L. Buckley. Jpc: Flexible inference for predictive coding networks in jax. arXiv preprint arXiv:2412.03676, 2024. +[24] S. Ioffe. Batch normalization: Accelerating deep network training by reducing internal covariate shift. arXiv preprint arXiv:1502.03167, 2015. +[25] S. Ishikawa and R. Karakida. On the parameterization of second-order optimization effective towards the infinite width. arXiv preprint arXiv:2312.12226, 2023. +[26] S. Ishikawa, R. Yokota, and R. Karakida. Local loss optimization in the infinite width: Stable parameterization of predictive coding networks and target propagation. arXiv preprint arXiv:2411.02001, 2024. +[27] A. Jacot, F. Gabriel, and C. Hongler. Neural tangent kernel: Convergence and generalization in neural networks. Advances in neural information processing systems, 31, 2018. +[28] D. P. Kingma and J. Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. +[29] Y. LeCun, Y. Bengio, and G. Hinton. Deep learning. nature, 521(7553):436-444, 2015. +[30] Y. LeCun, L. Bottou, G. B. Orr, and K.-R. Müller. Efficient backprop. In Neural networks: Tricks of the trade, pages 9-50. Springer, 2002. +[31] J. Lee, L. Xiao, S. Schoenholz, Y. Bahri, R. Novak, J. Sohl-Dickstein, and J. Pennington. Wide neural networks of any depth evolve as linear models under gradient descent. Advances in neural information processing systems, 32, 2019. +[32] Z. Liao and M. W. Mahoney. Hessian eigenspectra of more realistic nonlinear models. Advances in Neural Information Processing Systems, 34:20104-20117, 2021. +[33] T. P. Lillicrap, D. Cownden, D. B. Tweed, and C. J. Akerman. Random synaptic feedback weights support error backpropagation for deep learning. Nature communications, 7(1):13276, 2016. +[34] T. P. Lillicrap, A. Santoro, L. Marris, C. J. Akerman, and G. Hinton. Backpropagation and the brain. Nature Reviews Neuroscience, 21(6):335-346, 2020. +[35] V. A. Marchenko and L. A. Pastur. Distribution of eigenvalues for some sets of random matrices. Matematicheskii Sbornik, 114(4):507-536, 1967. + +[36] B. Millidge, T. Salvatori, Y. Song, R. Bogacz, and T. Lukasiewicz. Predictive coding: towards a future of deep learning beyond backpropagation? arXiv preprint arXiv:2202.09467, 2022. +[37] B. Millidge, A. Seth, and C. L. Buckley. Predictive coding: a theoretical and experimental review. arXiv preprint arXiv:2107.12979, 2021. +[38] B. Millidge, Y. Song, T. Salvatori, T. Lukasiewicz, and R. Bogacz. Backpropagation at the infinitesimal inference limit of energy-based models: Unifying predictive coding, equilibrium propagation, and contrastive hebbian learning. arXiv preprint arXiv:2206.02629, 2022. +[39] B. Millidge, Y. Song, T. Salvatori, T. Lukasiewicz, and R. Bogacz. A theoretical framework for inference and learning in predictive coding networks. arXiv preprint arXiv:2207.12316, 2022. +[40] B. Millidge, A. Tschantz, and C. L. Buckley. Predictive coding approximates backprop along arbitrary computation graphs. Neural Computation, 34(6):1329-1368, 2022. +[41] Y. Nesterov. Introductory lectures on convex optimization: A basic course, volume 87. Springer Science & Business Media, 2013. +[42] L. Noci, A. Meterez, T. Hofmann, and A. Orvieto. Super consistency of neural network landscapes and learning rate transfer. Advances in Neural Information Processing Systems, 37:102696-102743, 2025. +[43] A. Payeur, J. Guerguiev, F. Zenke, B. A. Richards, and R. Naud. Burst-dependent synaptic plasticity can coordinate learning in hierarchical circuits. Nature neuroscience, 24(7):1010-1019, 2021. +[44] C. Pehlevan and B. Bordelon. Lecture notes on infinite-width limits of neural networks. 2023. +[45] J. Pennington and Y. Bahri. Geometry of neural network loss surfaces via random matrix theory. In International conference on machine learning, pages 2798-2806. PMLR, 2017. +[46] J. Pennington, S. Schoenholz, and S. Ganguli. Resurrecting the sigmoid in deep learning through dynamical isometry: theory and practice. Advances in neural information processing systems, 30, 2017. +[47] J. Pennington, S. Schoenholz, and S. Ganguli. The emergence of spectral universality in deep networks. In International Conference on Artificial Intelligence and Statistics, pages 1924-1932. PMLR, 2018. +[48] L. Pinchetti, C. Qi, O. Lokshyn, G. Olivers, C. Emde, M. Tang, A. M'Charrak, S. Frieder, B. Menzat, R. Bogacz, et al. Benchmarking predictive coding networks-made simple. arXiv preprint arXiv:2407.01163, 2024. +[49] C. Qi, T. Lukasiewicz, and T. Salvatori. Training deep predictive coding networks. In New Frontiers in Associative Memories, 2025. +[50] D. A. Roberts, S. Yaida, and B. Hanin. The principles of deep learning theory, volume 46. Cambridge University Press Cambridge, MA, USA, 2022. +[51] R. Rosenbaum. On the relationship between predictive coding and backpropagation. Plos one, 17(3):e0266102, 2022. +[52] D. E. Rumelhart, G. E. Hinton, and R. J. Williams. Learning representations by back-propagating errors. nature, 323(6088):533-536, 1986. +[53] D. K. Salkuyeh. Comments on "a note on a three-term recurrence for a tridiagonal matrix". Applied mathematics and computation, 176(2):442-444, 2006. +[54] T. Salvatori, A. Mali, C. L. Buckley, T. Lukasiewicz, R. P. Rao, K. Friston, and A. Ororbia. Brain-inspired computational intelligence via predictive coding. arXiv preprint arXiv:2308.07870, 2023. + +[55] T. Salvatori, L. Pinchetti, B. Millidge, Y. Song, T. Bao, R. Bogacz, and T. Lukasiewicz. Learning on arbitrary graph topologies via predictive coding. Advances in neural information processing systems, 35:38232-38244, 2022. +[56] T. Salvatori, Y. Song, T. Lukasiewicz, R. Bogacz, and Z. Xu. Predictive coding can do exact backpropagation on convolutional and recurrent neural networks. arXiv preprint arXiv:2103.03725, 2021. +[57] T. Salvatori, Y. Song, B. Millidge, Z. Xu, L. Sha, C. Emde, R. Bogacz, and T. Lukasiewicz. Incremental predictive coding: A parallel and fully automatic learning algorithm. arXiv preprint arXiv:2212.00720, 2022. +[58] A. M. Saxe, J. L. McClelland, and S. Ganguli. Exact solutions to the nonlinear dynamics of learning in deep linear neural networks. arXiv preprint arXiv:1312.6120, 2013. +[59] B. Scellier and Y. Bengio. Equilibrium propagation: Bridging the gap between energy-based models and backpropagation. Frontiers in computational neuroscience, 11:24, 2017. +[60] Y. Song, T. Lukasiewicz, Z. Xu, and R. Bogacz. Can the brain do backpropagation?—exact implementation of backpropagation in predictive coding networks. Advances in neural information processing systems, 33:22566–22579, 2020. +[61] Y. Song, B. Millidge, T. Salvatori, T. Lukasiewicz, Z. Xu, and R. Bogacz. Inferring neural activity before plasticity: A foundation for learning beyond backpropagation. bioRxiv, pages 2022-05, 2022. +[62] R. Van Handel. Structured random matrices. Convexity and concentration, pages 107-156, 2017. +[63] B. van Zwol, R. Jefferson, and E. L. Broek. Predictive coding networks and inference learning: Tutorial and survey. arXiv preprint arXiv:2407.04117, 2024. +[64] A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. +[65] H. Weyl. Das asymptotische verteilungsgesetz der eigenwerte linearer partieller differentialgleichungen (mit einer anwendung auf die theorie der hohlraumstrahlung). Mathematische Annalen, 71(4):441-479, 1912. +[66] J. C. Whittington and R. Bogacz. An approximation of the error backpropagation algorithm in a predictive coding network with local hebbian synaptic plasticity. Neural computation, 29(5):1229-1262, 2017. +[67] E. P. Wigner. Characteristic vectors of bordered matrices with infinite dimensions i. The Collected Works of Eugene Paul Wigner: Part A: The Scientific Papers, pages 524-540, 1993. +[68] L. Xiao, Y. Bahri, J. Sohl-Dickstein, S. Schoenholz, and J. Pennington. Dynamical isometry and a mean field theory of cnns: How to train 10,000-layer vanilla convolutional neural networks. In International Conference on Machine Learning, pages 5393-5402. PMLR, 2018. +[69] G. Yang, E. Hu, I. Babuschkin, S. Sidor, X. Liu, D. Farhi, N. Ryder, J. Pachocki, W. Chen, and J. Gao. Tuning large neural networks via zero-shot hyperparameter transfer. Advances in Neural Information Processing Systems, 34:17084-17097, 2021. +[70] G. Yang and E. J. Hu. Tensor programs iv: Feature learning in infinite-width neural networks. In International Conference on Machine Learning, pages 11727-11737. PMLR, 2021. +[71] G. Yang and E. Littwin. Tensor programs ivb: Adaptive optimization in the infinite-width limit. arXiv preprint arXiv:2308.01814, 2023. +[72] G. Yang, D. Yu, C. Zhu, and S. Hayou. Tensor programs vi: Feature learning in infinite-depth neural networks. arXiv preprint arXiv:2310.02244, 2023. + +[73] J. Zhao, S. P. Singh, and A. Lucchi. Theoretical characterisation of the gauss-newton conditioning in neural networks. arXiv preprint arXiv:2411.02139, 2024. +[74] N. Zucchet and J. Sacramento. Beyond backpropagation: bilevel optimization through implicit differentiation and equilibrium propagation. Neural Computation, 34(12):2309-2346, 2022. + +# A Appendix + +# Contents + +A.1 Related work 16 +A.2 Proofs and derivations 17 + +A.2.1 Activity gradient (Eq. 4) and Hessian (Eq. 5) of DLNs 17 +A.2.2 Positive definiteness of the activity Hessian 18 +A.2.3 Random matrix theory of the activity Hessian 19 +A.2.4 Activity Hessian of linear ResNets 21 +A.2.5 Extension to other energy-based algorithms 22 +A.2.6 Limit convergence of $\mu \mathrm{PC}$ to BP (Thm. 1) 23 + +A.3 Additional experiments 23 + +A.3.1 Ill-conditioning with training 23 +A.3.2 Activity initialisations 24 +A.3.3 Activity decay 25 +A.3.4 Orthogonal initialisation 25 +A.3.5 $\mu \mathrm{PC}$ with one inference step 26 +A.3.6 Is inference convergence sufficient for good generalisation? 26 + +A.4 Experimental details 27 +A.5 Compute resources 28 +A.6 Supplementary figures 29 + +# A.1 Related work + +$\mu \mathbf{P}$ for PC [26]. The study closest to our work is [26], who derived a $\mu \mathrm{P}$ parameterisation for PC (as well as target propagation), also showing hyperparameter transfer across widths. This work differs from ours in the following three important aspects: (i) it derives $\mu \mathrm{P}$ for PC only for the width, (ii) it focuses on regimes where PC approximates or is equivalent to other algorithms (including BP) so that all the $\mu \mathrm{P}$ theory can be applied, and (iii) it considers layer-wise scalar precisions $\gamma_{\ell}$ for each layer energy term, which are not standard in how PCNs are trained (but are nevertheless interesting to study). By contrast, we propose to apply Depth- $\mu \mathrm{P}$ to PC, showing transfer for depth as well as width (Figs. 5 & A.31-A.32). We also study a regime where this parameterisation reduces to BP (Fig. 6) while showing that successful training is still possible far from this regime (Fig. 1). + +Training deep PCNs [49, 48]. Our work is also related to [49], who following [48] showed that the PC energy (Eq. 1) is disproportionately concentrated at the output layer $\mathcal{F}_L$ (closest to the target) for deep PCNs. They conjecture that this is problematic for two reasons: first, it does not allow the model to use (i.e. update) all of its layers; and second, it makes the latents diverge from the forward pass, which they claim leads to suboptimal weight updates. The first point is consistent with our theory and experiments. In particular, because the activities of standard PCNs vanish or explode with the depth ( $\S 3.2$ ) and stay almost constant during inference due to the ill-conditioning of the landscape ( $\S 3.1$ ) (Figs. A.10-A.11 & A.36), the weight updates are likely to be imbalanced across layers. However, the ill-conditioning contradicts the second point, in that the activities barely move during inference and stay close to the forward pass (see $\S 3.2$ for relevant experiments). Moreover, divergence from the forward pass does not necessarily lead to suboptimal weight updates and worse performance. For standard PC, deep networks cannot achieve good performance regardless of whether one stays close to the forward pass (see $\S 3.6$ ). For $\mu \mathrm{PC}$ , on the other hand, as many steps as the number of hidden layers (e.g. Fig. 1) leads to depth-stable and much better accuracy than a single step (e.g. Fig. A.14). + +PC and BP. Our theoretical result about the convergence of $\mu \mathrm{PC}$ to BP (Theorem 1) relates to a relatively well-established series of correspondences between PC and BP [66, 40, 60, 51, 56, 38]. In brief, if one makes some rather biologically implausible assumptions (such as precisely timed inference updates), it can be shown that PC can approximate or even compute exactly the same gradients as BP. In stark contrast to these results and also the work of [26] (which requires arbitrarily specific precision values at different layers), Theorem 1 applies to standard PC, with arguably interpretable width- and depth-dependent scalings. $^{10}$ + +Theory of PC inference (Eq. 2) & learning (Eq. 3). Finally, our work can be seen as a companion paper to [22], who provided the first rigorous, explanatory and predictive theory of the learning landscape and dynamics of practical PCNs (Eq. 3). They first show that for DLNs the energy at the inference equilibrium is a rescaled MSE loss with a weight-dependent rescaling, a result that we build on here for Theorem 1. They then characterise the geometry of the equilibrated energy (the effective landscape on which PC learns), showing that many highly degenerate saddles of the loss including the origin become much easier to escape in the equilibrated energy. Here, by contrast, we focus on the geometry of the inference landscape and dynamics (Eq. 2). As an aside, we note that the origin saddle result of [22] probably breaks down for ResNets, where for the linear case it has been shown that the saddle is effectively shifted and the origin becomes locally convex [15]. We suspect that the results generalise, but it could still be interesting to extend the theory of [22] to ResNets, especially by also looking at the geometry of minima. + +$\mu \mathbf{P}$ . For a full treatment of $\mu \mathrm{P}$ and its extensions, we refer the reader to key works of the "Tensor Programs" series [70, 69, 71, 72]. $\mu \mathrm{P}$ effectively puts feature learning back into the infinite-width limit of neural networks, lacking from the neural tangent kernel (NKT) or "lazy" regime [27, 6, 31]. In particular, in the NTK the layer preactivations evolve in $\mathcal{O}(N^{-1/2})$ time. In $\mu \mathrm{P}$ , the features instead change in a "maximal" sense (hence "μ"), in that they vary as much as possible without diverging with the width, which occurs for the output predictions under SP [70]. More formally, $\mu \mathrm{P}$ can be derived from the 3 desiderata stated in §2.1. $\mu \mathrm{P}$ was extended to depth (Depth-μP) for ResNets by mainly introducing a $1/\sqrt{L}$ scaling before each residual block [72, 3]. This breakthrough was enabled by the commutativity of the infinite-width and infinite-depth limit of ResNets [17, 16]. Standard $\mu \mathrm{P}$ has also been extended to local algorithms including PC [26] (see $\mu \mathbf{P}$ for PC above), sparse networks [9], second-order methods [25], and sharpness-aware minimisation [14]. + +# A.2 Proofs and derivations + +All the theoretical results below are derived for linear networks of some form. + +# A.2.1 Activity gradient (Eq. 4) and Hessian (Eq. 5) of DLNs + +The gradient of the energy with respect to all the PC activities of a DLN (Eq. 4) can be derived by simple rearrangement of the partials with respect to each layer, which are given by + +$$ +\partial \mathcal {F} / \partial \mathbf {z} _ {1} = \mathbf {z} _ {1} - a _ {1} \mathbf {W} _ {1} \mathbf {x} - a _ {2} \mathbf {W} _ {2} ^ {T} \mathbf {z} _ {2} + a _ {2} ^ {2} \mathbf {W} _ {2} ^ {T} \mathbf {W} _ {2} \mathbf {z} _ {1} \tag {7} +$$ + +$$ +\partial \mathcal {F} / \partial \mathbf {z} _ {2} = \mathbf {z} _ {2} - a _ {2} \mathbf {W} _ {2} \mathbf {z} _ {1} - a _ {3} \mathbf {W} _ {3} ^ {T} \mathbf {z} _ {3} + a _ {3} ^ {2} \mathbf {W} _ {3} ^ {T} \mathbf {W} _ {3} \mathbf {z} _ {2} \tag {8} +$$ + +$$ +\vdots \tag {9} +$$ + +$$ +\partial \mathcal {F} / \partial \mathbf {z} _ {H} = \mathbf {z} _ {H} - a _ {L - 1} \mathbf {W} _ {L - 1} \mathbf {z} _ {H - 1} - a _ {L} \mathbf {W} _ {L} ^ {T} \mathbf {y} + a _ {L} ^ {2} \mathbf {W} _ {L} ^ {T} \mathbf {W} _ {L} \mathbf {z} _ {H}. \tag {10} +$$ + +Factoring out the activity of each layer + +$$ +\partial \mathcal {F} / \partial \mathbf {z} _ {1} = \mathbf {z} _ {1} \left(\mathbf {1} + a _ {2} ^ {2} \mathbf {W} _ {2} ^ {T} \mathbf {W} _ {2}\right) - a _ {1} \mathbf {W} _ {1} \mathbf {x} - a _ {2} \mathbf {W} _ {2} ^ {T} \mathbf {z} _ {2} \tag {11} +$$ + +$$ +\partial \mathcal {F} / \partial \mathbf {z} _ {2} = \mathbf {z} _ {2} \left(\mathbf {1} + a _ {3} ^ {2} \mathbf {W} _ {3} ^ {T} \mathbf {W} _ {3}\right) - a _ {2} \mathbf {W} _ {2} \mathbf {z} _ {1} - a _ {3} \mathbf {W} _ {3} ^ {T} \mathbf {z} _ {3} \tag {12} +$$ + +$$ +\vdots \tag {13} +$$ + +$$ +\partial \mathcal {F} / \partial \mathbf {z} _ {H} = \mathbf {z} _ {H} \left(\mathbf {1} + a _ {L} ^ {2} \mathbf {W} _ {L} ^ {T} \mathbf {W} _ {L}\right) - a _ {L - 1} \mathbf {W} _ {L - 1} \mathbf {z} _ {H - 1} - a _ {L} \mathbf {W} _ {L} ^ {T} \mathbf {y}, \tag {14} +$$ + +one realises that this can be rearranged in the form of a linear system + +$$ +\nabla_ {\mathbf {z}} \mathcal {F} = \underbrace {\left[ \begin{array}{c c c c c} \mathbf {I} + a _ {2} ^ {2} \mathbf {W} _ {2} ^ {T} \mathbf {W} _ {2} & - a _ {2} \mathbf {W} _ {2} ^ {T} & \mathbf {0} & \dots & \mathbf {0} \\ - a _ {2} \mathbf {W} _ {2} & \mathbf {I} + a _ {3} ^ {2} \mathbf {W} _ {3} ^ {T} \mathbf {W} _ {3} & - a _ {3} \mathbf {W} _ {3} ^ {T} & \dots & \mathbf {0} \\ \mathbf {0} & - a _ {3} \mathbf {W} _ {3} & \mathbf {I} + a _ {4} ^ {2} \mathbf {W} _ {4} ^ {T} \mathbf {W} _ {4} & \ddots & \mathbf {0} \\ \vdots & \vdots & \ddots & \ddots & - a _ {L - 1} \mathbf {W} _ {L - 1} ^ {T} \\ \mathbf {0} & \mathbf {0} & \mathbf {0} & - a _ {L - 1} \mathbf {W} _ {L - 1} & \mathbf {I} + a _ {L} ^ {2} \mathbf {W} _ {L} ^ {T} \mathbf {W} _ {L} \\ \hline & & & & \end{array} \right] \underbrace {\left[ \begin{array}{c} \mathbf {z} _ {1} \\ \mathbf {z} _ {2} \\ \vdots \\ \mathbf {z} _ {H - 1} \\ \mathbf {z} _ {H} \end{array} \right]} _ {\mathbf {z}} - \underbrace {\left[ \begin{array}{c} a _ {1} \mathbf {W} _ {1} \mathbf {x} \\ \mathbf {0} \\ \vdots \\ \mathbf {0} \\ a _ {L} \mathbf {W} _ {L} ^ {T} \mathbf {y} \end{array} \right]} _ {\mathbf {b}}} (1 5) +$$ + +where the matrix of coefficients corresponds to the Hessian of the energy with respect to the activities $(\mathbf{H}_{\mathbf{z}})_{\ell k} \coloneqq \partial^2 \mathcal{F} / \partial \mathbf{z}_\ell \partial \mathbf{z}_k$ . We make the following side remarks about how different training and architecture design choices impact the structure of the activity Hessian: + +- In the unsupervised case where $\mathbf{z}_0$ is left free to vary like any other hidden layer, the Hessian gets the additional terms $a_1^2\mathbf{W}_1^T\mathbf{W}_1$ as the first diagonal block, $-a_{1}\mathbf{W}_{1}$ as the superdiagonal block (and its transpose as the subdiagonal block), and $\mathbf{b}_1 = \mathbf{0}$ . This does not fundamentally change the structure of the Hessian; in fact, in the next section we show that convexity holds for both the unsupervised and supervised cases. +- Turning on biases at each layer such that $\mathcal{F}_{\ell} = \frac{1}{2}\|\mathbf{z}_{\ell} - a_{\ell}\mathbf{W}_{\ell}\mathbf{z}_{\ell - 1} - \mathbf{b}_{\ell}\|^2$ does not impact the Hessian and simply makes the constant vector of the linear system more dense: $\mathbf{b} = [a_1\mathbf{W}_1\mathbf{x} + \mathbf{b}_1 - a_2\mathbf{W}_2^T\mathbf{b}_2, \mathbf{b}_2 - a_3\mathbf{W}_3^T\mathbf{b}_3, \dots, a_L\mathbf{W}_L^T\mathbf{y} + \mathbf{b}_{L - 1} - a_L\mathbf{W}_L^T\mathbf{b}_L]^T$ . +- Adding an $\ell^2$ norm regulariser to the activities $\frac{1}{2} ||\mathbf{z}_{\ell}||^2$ scales the identity in each diagonal block by 2. This induces a unit shift in the Hessian eigenspectrum such that the minimum eigenvalue is lower bounded at one rather than zero (see §A.2.3), as shown in Fig. A.12. +- Adding "dummy" latents at either end of the network, such that $\mathcal{F}_0 = \frac{1}{2} ||\mathbf{x} - \mathbf{z}_0||^2$ or $\mathcal{F}_L = \frac{1}{2} ||\mathbf{y} - \mathbf{z}_L||^2$ , simply adds one layer to the Hessian with a block diagonal given by 2I. +- Compared to fully connected networks, the activity Hessian of convolutional networks is sparser in that (dense) weight matrices are replaced by (sparser) Toeplitz matrices. The activity Hessian of ResNets is derived and discussed in §A.2.4. + +We also note that Eq. 15 can be used to provide an alternative proof of the known convergence of PC inference to the feedforward pass [39] $\mathbf{z}^{*} = \mathbf{H}_{\mathbf{z}}^{-1}\mathbf{b} = f(\mathbf{x}) = a_{L}\mathbf{W}_{L}\dots a_{1}\mathbf{W}_{1}\mathbf{x}$ when the output layer is unclamped or free to vary with $\partial^2\mathcal{F} / \partial \mathbf{z}_L^2 = \mathbf{I}$ and $\mathbf{b}_H = \mathbf{0}$ . + +# A.2.2 Positive definiteness of the activity Hessian + +Here we prove that the Hessian of the energy with respect to the activities of arbitrary DLNs (Eq. 5) is positive definite (PD), $\mathbf{H}_{\mathbf{z}}\succ 0$ . The result is empirically verified for DLNs in §A.2.3 and also appears to generally hold for nonlinear networks, where we observe small negative Hessian eigenvalues only for very shallow Tanh networks with no skip connections (see Figs. A.7 & A.22). + +Theorem A.1 (Convexity of the PC inference landscape of DLNs.). For any DLN parameterised by $\pmb{\theta} \coloneqq (\mathbf{W}_1, \dots, \mathbf{W}_L)$ with input and output $(\mathbf{x}, \mathbf{y})$ , the activity Hessian of the PC energy (Eq. 1) is positive definite + +$$ +\mathbf {H} _ {\mathbf {z}} (\theta) \succ 0, \tag {16} +$$ + +showing that the inference or activity landscape $\mathcal{F}(\mathbf{z})$ is strictly convex. + +To prove this, we will show that the Hessian satisfies Sylvester's criterion, which states that a Hermitian matrix is PD if all of its leading principal minors (LPMs) are positive, i.e. if the determinant of all its square top-left submatrices is positive [21]. Recall that an $n \times n$ square matrix $\mathbf{A}$ has $n$ LPMs $\mathbf{A}_h$ of size $h \times h$ for $h = 1, \ldots, n$ . For a Hermitian matrix, showing that the determinant of all its LPMs is positive is a necessary and sufficient condition to determine whether the matrix is PD ( $\mathbf{A} \succ 0$ ), and this result can be generalised to block matrices. + +We now show that the activity Hessian of arbitrary DLNs (Eq. 5) satisfies Sylvester's criterion. We drop the Hessian subscript $\mathbf{H}$ for brevity of notation. The proof technique lies in a Laplace or cofactor expansion of the LPMs along the last row. This has an intuitive interpretation in that it starts by proving that the inference landscape of one-hidden-layer PCNs is (strictly) convex, and then proceeds by induction to show that adding layers does not change the result. + +The activity Hessian has $NH$ LPMs of size $N\ell \times N\ell$ for $\ell = 1,\dots ,H$ . Let $[\mathbf{H}]_{\ell}$ denote the $\ell$ th LPM of $\mathbf{H}$ , $\Delta_{\ell} = |[\mathbf{H}]_{\ell}|$ its determinant, and $\mathbf{D}_{\ell}$ and $\mathbf{O}_{\ell}$ the $\ell$ th diagonal and off-diagonal blocks of $\mathbf{H}$ , respectively. Now note that $\mathbf{H}$ is a block tridiagonal symmetric matrix, as can be clearly seen from Eq. 15. There is a known two-term recurrence relation that can be used to calculate the determinant of such matrices through their LPMs [53] + +$$ +\Delta_ {\ell} = | \mathbf {D} _ {\ell} | \Delta_ {\ell - 1} - | \mathbf {O} _ {\ell - 1} | ^ {2} \Delta_ {\ell - 2}, \quad \ell = 2, \dots , H \tag {17} +$$ + +with $\Delta_0 = 1$ and $\Delta_{1} = |\mathbf{D}_{1}|$ . The first LPM is clearly PD and so its determinant is positive, $\mathbf{D}_1 = \mathbf{I} + a_2^2\mathbf{W}_2^T\mathbf{W}_2\succ 0\Rightarrow \Delta_1 > 0$ , showing that the inference landscape of one-hidden-layer linear PCNs is strictly convex. For $\ell = 2$ , the first term of the recursion (Eq. 17) is positive, since $|\mathbf{D}_2| = |\mathbf{I} + a_3^2\mathbf{W}_3^T\mathbf{W}_3| > 0$ and, $\Delta_{1} > 0$ as we just saw. The second term is negative, but it is strictly less than the positive term, $|a_{2}\mathbf{W}_{2}|^{2} < |\mathbf{I} + a_{3}^{2}\mathbf{W}_{3}^{T}\mathbf{W}_{3}||\mathbf{I} + a_{2}^{2}\mathbf{W}_{2}^{T}\mathbf{W}_{2}|$ and so $\Delta_{2} > 0$ . Hence, the activity landscape of 2-hidden-layer linear PCNs remains convex. The same holds for three hidden layers where $|\mathbf{O}_2|\Delta_1 < |\mathbf{D}_3|\Delta_2\Rightarrow \Delta_3 > 0$ . + +We can keep iterating this argument, showing by induction that the inference landscape is (strictly) convex for arbitrary DLNs. More formally, the positive term of the recurrence relation is always strictly greater than the negative term, + +$$ +\left| \mathbf {D} _ {\ell} \right| \Delta_ {\ell - 1} > 0 \tag {18} +$$ + +$$ +\left| \mathbf {D} _ {\ell} \right| \Delta_ {\ell - 1} > \left| \mathbf {O} _ {\ell - 1} \right| ^ {2} \Delta_ {\ell - 2} \tag {19} +$$ + +and so $\Delta_{\ell} > 0$ and $\mathbf{H} \succ 0$ for all $\ell$ . Convexity holds for the unsupervised case, where the activity Hessian is now positive semidefinite since the term $a_1^2 \mathbf{W}_1^T \mathbf{W}_1$ is introduced (see §A.2.1). The result can also be extended to any other linear layer transformation $\mathbf{B}_{\ell}$ including ResNets where $\mathbf{B}_{\ell} = \mathbf{I} + \mathbf{W}_{\ell}$ . + +# A.2.3 Random matrix theory of the activity Hessian + +Here we analyse the Hessian of the energy with respect to the activities of DLNs (Eq. 5) using random matrix theory (RMT). This analysis follows a line of work using RMT to study the Hessian of neural networks, specifically the Hessian of the loss with respect to the parameters [7, 45, 13, 32, 2]. We note that the structure of the activity Hessian is much simpler than the weight or parameter Hessian, in that for linear networks the former is positive definite (Theorem A.1, §A.2.2), while for the latter this is only true for one hidden layer [22]. + +In what follows, we recall from §2.2 that the PC energy (Eq. 1) has layer-wise scalings $a_{\ell}$ for all $\ell$ , and the weights are assumed to be drawn from a zero-mean Gaussian $(\mathbf{W}_{\ell})_{ij} \sim \mathcal{N}(0, b_{\ell})$ with variance set by $b_{\ell}$ . + +Hessian decomposition. The activity Hessian (Eq. 5) is a challenging matrix to study theoretically as its entries are not i.i.d. even at initialization due to the off-diagonal couplings between layers. However, we can decompose the matrix into its diagonal and off-diagonal components: + +$$ +\mathbf {H} _ {\mathbf {z}} = \mathbf {D} + \mathbf {O} \tag {20} +$$ + +with $\mathbf{D} := \mathrm{diag}(\mathbf{I} + a_2^2\mathbf{W}_2^T\mathbf{W}_2, \ldots, \mathbf{I} + a_L^2\mathbf{W}_L^T\mathbf{W}_L)$ and $\mathbf{O} := \mathrm{offdiag}(-a_2\mathbf{W}_2, \ldots, -a_{L-1}\mathbf{W}_{L-1})$ , where the off-diagonal part can be seen as a perturbation. Since these matrices are on their own i.i.d. at initialisation, we can use standard RMT results to analyse their respective eigenvalue distributions in the regime of large width $N$ and depth $H$ we are interested in. We will then use these results to gain some qualitative insights into the overall spectrum of $\mathbf{H}_{\mathbf{z}}$ . + +Analysis of $\mathbf{D}$ . As a block diagonal matrix, the eigenvales of $\mathbf{D}$ are given by those of its blocks $\mathbf{D}_{\ell} = \mathbf{I} + a_{\ell +1}^{2}\mathbf{W}_{\ell +1}^{T}\mathbf{W}_{\ell +1} \in \mathbb{R}^{N\times N}$ for $\ell = 1,\dots ,H$ . Note that the size of each block depends only on the network width $N$ . It is easy to see that each block is a positively shifted Wishart matrix. As $N \to \infty$ , the eigenspectrum of such matrices converges to the well-known Marchenko-Pastur (MP) distribution [35] if properly normalised such that $a_{\ell +1}^{2}\mathbf{W}_{\ell +1}^{T}\mathbf{W}_{\ell +1} \sim \mathcal{O}(1 / N)$ . + +As shown in Figs. A.1-A.2, this normalisation can be achieved in two distinct but equivalent ways: (i) by initialising from a standard Gaussian with $b_{\ell} = 1$ and setting the layer scaling to $a_{\ell} = 1 / \sqrt{N}$ , or (ii) by setting $a_{\ell} = 1$ and $b_{\ell} = 1 / N$ as done by standard initialisations [30, 11, 18]. In either case, in the infinite-width limit the eigenvalues of each diagonal block will converge to a unit-shifted MP density with extremes + +$$ +\begin{array}{l} \lim _ {N \rightarrow \infty} \quad \lambda_ {\pm} (\mathbf {D} _ {\ell}) = 1 + (1 \pm \sqrt {N / N}) ^ {2} (21) \\ = \{1, 5 \}. (22) \\ \end{array} +$$ + +While the spectrum of $\mathbf{D}$ will be a combination of these independent MP densities, its extremes will be the same of $\mathbf{D}_{\ell}$ since all of the blocks are i.i.d. and grow at the same rate as $N\to \infty$ This is empirically verified in Figs.A.1-A.2, which also confirm that the spectrum of $\mathbf{D}$ is only affected by the width and not the depth. + +Analysis of O. The off-diagonal component of the Hessian $\mathbf{O}$ is a sparse Wigner matrix whose size depends on both the width and the depth and so the correct limit should take both $N, H \to \infty$ at some constant ratio. Note that the sparsity of $\mathbf{O}$ grows much faster with the depth. Because sparse Wigner matrices are poorly understood and still an active area of research [62], we make the simplifying assumption that $\mathbf{O}$ is dense. + +If properly normalised as above, we know that in the limit the eigenspectrum of dense Wigner matrices converges the classical Wigner semicircle distribution [67] with extremes + +$$ +\lim _ {H / N \rightarrow \infty} \lambda_ {\pm} (\mathbf {O}) = \pm 2. \tag {23} +$$ + +We find that the empirical eigenspectrum of $\mathbf{O}$ is slightly broader than the semicircle and, as expected, is affected by both the width and the depth (Figs. A.3-A.4). + +Analysis of $\mathbf{H}_{\mathbf{z}}$ Given the above asymptotic results on $\mathbf{D}$ and $\mathbf{O}$ , we can use Weyl's inequalities [65] to lower and upper bound the minimum and maximum eigenvalues (and so the condition number) of the overall Hessian at initialisation: $\lambda_{\mathrm{max}}(\mathbf{D} + \mathbf{O}) \leq \lambda_{\mathrm{max}}(\mathbf{D}) + \lambda_{\mathrm{max}}(\mathbf{O})$ and $\lambda_{\mathrm{min}}(\mathbf{D} + \mathbf{O}) \geq \lambda_{\mathrm{min}}(\mathbf{D}) + \lambda_{\mathrm{min}}(\mathbf{O})$ . The upper bound $(\tilde{\lambda}_{\mathrm{max}} = 7)$ appears tight, as shown in Figs. A.5-A.7. However, the lower bound predicts a negative minimum eigenvalue $(\tilde{\lambda}_{\mathrm{min}} = -1)$ , which is not possible since the Hessian is positive definite as we proved in §A.2.2. + +Nevertheless, we can still gain some insights + +into the interaction between $\bar{\mathbf{D}}$ and $\mathbf{O}$ by looking at the empirical eigenspectrum of $\mathbf{H}_{\mathbf{z}}$ . In particular, + +![](images/b84955bb34f79f9fe5016a4f1ae2ed970dfed37610a1d1473098758c17bdfd9e.jpg) +Figure A.1: Empirical eigenspectra of D at initialisation, holding the network width constant $(N = 128)$ and varying the depth $H$ . $a_{\ell}$ indicates the premultiplier at each network layer (Eq. 1), while $b_{\ell}$ is the variance of Gaussian initialisation, with $a_{\ell} = 1$ and $b_{\ell} = 1 / N$ corresponding to the "standard parameterisation" (SP). + +![](images/63ef2d91aedd57ae8f0726941dab51473c1f9a224705f1684d862063097d58fa.jpg) +Figure A.2: Empirical eigenspectra of D at initialisation, holding the network depth constant $(H = 128)$ and varying the width $N$ . + +![](images/e81f603583cb35eb56effc589dfcc7089a9ddc797a445cecb3942670f7c4f035.jpg) +Figure A.3: Empirical eigenspectra of O at initialisation, holding the network width constant $(N = 128)$ and varying the depth $H$ . + +![](images/2efe98abc162c56cbbbc759ebd3131faed923879e4fa6bc6659e8c78a1c24bbe.jpg) +Figure A.4: Empirical eigenspectra of O at initialisation, holding the network depth constant $(H = 128)$ and varying the width $N$ . + +we observe that the maximum and especially the minimum eigenvalue of the Hessian scale with the network depth (Figs. A.7 & A.22), thus driving the growth of the condition number. + +![](images/492b0e98e384f695814c7e02e331dd00a5c23011eccecc14da4867db9585de2d.jpg) +Figure A.5: Empirical eigenspectra of $\mathbf{H}$ at initialisation, holding the network width constant $(N = 128)$ and varying the depth $H$ . + +![](images/8041528900107ac17c8b62422212f964476c03375277fd5f6162b74e791d2896.jpg) +Figure A.6: Empirical eigenspectra of $\mathbf{H}$ at initialisation, holding the network depth constant $(H = 128)$ and varying the width $N$ . + +![](images/057d6d37fbc06103ce7b6a261b63a23c9d141341088d37eccfa90454d8ff6aca.jpg) + +![](images/7d307247f6ac9dae3e0207c67b8f22b096cdefa68223ec9be86aa9a1acce7865.jpg) + +![](images/b218944c07b88f21d5085fa504ddd7c8781a37311bddf29dc911624e3a076000.jpg) + +![](images/a9fac5ac8192d04420133a0cbf373c3c10d25afef64efdfa889e3f11c8ed5b69.jpg) +Figure A.7: Maximum and minimum eigenvalues of $\mathbf{H}_{\mathbf{z}}$ at initialisation as a function of network width $N$ and depth $L$ . + +![](images/06b0d1c6bcf511e70a8d609d628c6bb8a9edac5c6326a8f623adf931779f3b5b.jpg) + +![](images/0d95d15a93e599bf8a68663546a685d03d649a0a3d83375ed872eb18a3e83316.jpg) + +# A.2.4 Activity Hessian of linear ResNets + +Here we derive the activity Hessian for linear ResNets [19], extending the derivation in §A.2.1 for DLNs. Following the Depth- $\mu \mathrm{P}$ parameterisation [72, 3], we consider ResNets with identity skip connections at every layer except from the input and to the output. The PC energy for such ResNets is given by + +$$ +\mathcal {F} _ {1 - \operatorname {s k i p}} = \frac {1}{2} \| \boldsymbol {\epsilon} _ {L} \| ^ {2} + \frac {1}{2} \| \boldsymbol {\epsilon} _ {1} \| ^ {2} + \sum_ {\ell = 2} ^ {H} \frac {1}{2} \| \mathbf {z} _ {\ell} - a _ {\ell} \mathbf {W} _ {\ell} \mathbf {z} _ {\ell - 1} - \underbrace {\mathbf {z} _ {\ell - 1}} _ {1 - \operatorname {s k i p}} \| ^ {2}, \tag {24} +$$ + +![](images/771880ed79aa447fa05346eb4b2efd31981aff364b62e37c2a656dbafb60bd1a.jpg) + +where recall that $\epsilon_{\ell} = \mathbf{z}_{\ell} - a_{\ell}\mathbf{W}_{\ell}\mathbf{z}_{\ell -1}$ and $\mathbf{z}_0\coloneqq \mathbf{x}$ , $\mathbf{z}_L\coloneqq \mathbf{y}$ . We refer to this model as "1-skip" since the residual is added to every layer. Its activity Hessian is given by + +$$ +\mathbf {H} _ {\mathbf {z}} ^ {1 - \mathrm {s k i p}} := \frac {\partial^ {2} \mathcal {F} _ {1 - \mathrm {s k i p}}}{\partial \mathbf {z} _ {\ell} \partial \mathbf {z} _ {k}} = \left\{ \begin{array}{l l} 2 \mathbf {I} + a _ {\ell + 1} ^ {2} \mathbf {W} _ {\ell + 1} ^ {T} \mathbf {W} _ {\ell + 1} + a _ {\ell + 1} \left(\mathbf {W} _ {\ell + 1} ^ {T} + \mathbf {W} _ {\ell + 1}\right), & \ell = k \neq H \\ \mathbf {I} + a _ {\ell + 1} ^ {2} \mathbf {W} _ {\ell + 1} ^ {T} \mathbf {W} _ {\ell + 1}, & \ell = k = H \\ - a _ {k + 1} \mathbf {W} _ {k + 1} - \mathbf {I}, & \ell - k = 1 \\ - a _ {\ell + 1} \mathbf {W} _ {\ell + 1} ^ {T} - \mathbf {I}, & \ell - k = - 1 \\ \mathbf {0}, & \text {e l s e} \end{array} . \right. \tag {25} +$$ + +We find that this Hessian is much more ill-conditioned (Fig. A.22) than that of networks without skips (Fig. 2), across different parameterisations (Fig. 4). We note that one can extend these results to $n$ -skip linear ResNets with energy + +$$ +\mathcal {F} _ {n - \operatorname {s k i p}} = \frac {1}{2} \| \boldsymbol {\epsilon} _ {L} \| ^ {2} + \sum_ {\ell = 1} ^ {n} \frac {1}{2} \| \boldsymbol {\epsilon} _ {\ell} \| ^ {2} + \sum_ {\ell = n + 1} ^ {H} \frac {1}{2} \| \mathbf {z} _ {\ell} - a _ {\ell} \mathbf {W} _ {\ell} \mathbf {z} _ {\ell - 1} - \underbrace {\mathbf {z} _ {\ell - n}} _ {n - \operatorname {s k i p}} \| ^ {2} \tag {26} +$$ + +or indeed arbitrary computational graphs [55]. It could be interesting to investigate whether there exist architectures with better conditioning of the inference landscape that do not sacrifice the stability of the forward pass (see §4, Fig. 4). + +# A.2.5 Extension to other energy-based algorithms + +Here we include a preliminary investigation of the inference dynamics of other energy-based local learning algorithms. As an example, we consider equilibrium propagation (EP) [59], whose energy for a DLN is given by + +$$ +E = \frac {1}{2} \| \mathbf {z} _ {\ell} \| ^ {2} - \sum_ {\ell = 1} ^ {L} \mathbf {z} _ {\ell} ^ {T} \mathbf {W} _ {\ell} \mathbf {z} _ {\ell - 1} + \frac {\beta}{2} \| \mathbf {y} - \mathbf {z} _ {L} \| ^ {2}, \tag {27} +$$ + +where $\mathbf{z}_0\coloneqq \mathbf{x}$ for supervised learning (as for PC), and it is also standard to include an $\ell^2$ regulariser on the activities. Unlike PC, EP has two inference phases: a free phase where the output layer $\mathbf{z}_L$ is free to vary like any other hidden layer with $\beta = 0$ ; and a clamped or nudged phase where the output is fixed to some target $\mathbf{y}$ with $\beta >0$ . The activity gradient and Hessian of the EP energy (Eq. 27) are given by + +$$ +\frac {\partial E}{\partial \mathbf {z} _ {\ell}} = \left\{ \begin{array}{l l} \mathbf {z} _ {\ell} - \mathbf {W} _ {\ell} \mathbf {z} _ {\ell - 1} - \mathbf {z} _ {\ell + 1} ^ {T} \mathbf {W} _ {\ell + 1}, & \ell \neq L \\ \mathbf {z} _ {\ell} - \mathbf {W} _ {\ell} \mathbf {z} _ {\ell - 1} - \beta (\mathbf {y} - \mathbf {z} _ {\ell}), & \ell = L \end{array} \right. \tag {28} +$$ + +and + +$$ +\mathbf {H} _ {\mathbf {z}} := \frac {\partial^ {2} E}{\partial \mathbf {z} _ {\ell} \partial \mathbf {z} _ {k}} = \left\{ \begin{array}{l l} \mathbf {I}, & \ell = k \neq L \\ \mathbf {I} + \beta , & \ell = k = L \\ - \mathbf {W} _ {\ell + 1}, & \ell - k = 1 \\ - \mathbf {W} _ {k + 1} ^ {T}, & \ell - k = - 1 \\ \mathbf {0}, & \text {e l s e} \end{array} \right. \tag {29} +$$ + +where we abuse notation by denoting the Hessian in the same way as that of the PC energy. We observe that the off-diagonal blocks are equal to those of the PC activity Hessian (Eq. 5). Similar to PC, one can also rewrite the EP activity gradient (Eq. 28) as a linear system + +$$ +\nabla_ {\mathbf {z}} E = \underbrace {\left[ \begin{array}{c c c c c} \mathbf {I} & - \mathbf {W} _ {2} ^ {T} & \mathbf {0} & \dots & \mathbf {0} \\ - \mathbf {W} _ {2} & \mathbf {I} & - \mathbf {W} _ {3} ^ {T} & \dots & \mathbf {0} \\ \mathbf {0} & - \mathbf {W} _ {3} & \mathbf {I} & \ddots & \mathbf {0} \\ \vdots & \vdots & \ddots & \ddots & - \mathbf {W} _ {L} ^ {T} \\ \mathbf {0} & \mathbf {0} & \mathbf {0} & - \mathbf {W} _ {L} & \mathbf {I} + \beta \end{array} \right]} _ {\mathbf {H} _ {\mathbf {z}}} \underbrace {\left[ \begin{array}{c} \mathbf {z} _ {1} \\ \mathbf {z} _ {2} \\ \vdots \\ \mathbf {z} _ {L - 1} \\ \mathbf {z} _ {L} \end{array} \right]} _ {\mathbf {z}} - \underbrace {\left[ \begin{array}{c} \mathbf {W} _ {1} \mathbf {x} \\ \mathbf {0} \\ \vdots \\ \mathbf {0} \\ \beta \mathbf {y} \end{array} \right]} _ {\mathbf {b}} +$$ + +with solution $\mathbf{z}^{*} = \mathbf{H}_{\mathbf{z}}^{-1}\mathbf{b}$ . Interestingly, unlike for PC, the EP inference landscape is not necessarily convex, which can be easily seen for a shallow 2-layer scalar network where $\exists \lambda (\mathbf{H}_{\mathbf{z}}(w_2 > 1)) < 0$ . This is always true without the activity regulariser, in which case the identity in each diagonal block vanishes. + +# A.2.6 Limit convergence of $\mu \mathbf{PC}$ to BP (Thm. 1) + +Here we provide a simple proof of Theorem 1. Consider a slight generalisation to linear ResNets (Eq. 24) of the PC energy at the inference equilibrium derived by [22] for DLNs: + +$$ +\mathcal {F} \left(\mathbf {z} ^ {*}\right) = \frac {1}{2 B} \sum_ {i = 1} ^ {B} \mathbf {r} _ {i} ^ {T} \mathbf {S} ^ {- 1} \mathbf {r} _ {i}, \tag {31} +$$ + +$$ +\text {w h e r e} \quad \mathbf {S} = \mathbf {I} _ {d _ {y}} + a _ {L} ^ {2} \mathbf {W} _ {L} \mathbf {W} _ {L} ^ {T} + \sum_ {\ell = 2} ^ {H} \left(a _ {L} \mathbf {W} _ {L} \prod_ {\ell} ^ {H} \mathbf {I} + a _ {\ell} \mathbf {W} _ {\ell}\right) \left(a _ {L} \mathbf {W} _ {L} \prod_ {\ell} ^ {H} \mathbf {I} + a _ {\ell} \mathbf {W} _ {\ell}\right) ^ {T} \tag {32} +$$ + +and the residual error is $\mathbf{r}_i = \mathbf{y}_i - a_L\mathbf{W}_L\left(\prod_{\ell = 2}^H\mathbf{I} + a_\ell \mathbf{W}_\ell\right)a_1\mathbf{W}_1\mathbf{x}_i$ . $B$ can stand for the batch or dataset size. Note that Eq. 31 is an MSE loss with a weight-dependent rescaling (Eq. 32). Now, we know that, for Depth- $\mu$ P, the forward pass of this model has $\mathcal{O}_{N,H}(1)$ preactivations at initialisation and so the residual will also be of order 1. Note that, by contrast, for SP ( $a_\ell = 1$ for all $\ell$ and $b_\ell = 1 / N_{\ell-1}$ ) the preactivations explode with the depth (Fig. A.30). + +The key question, then, is what happens to the rescaling $\mathbf{S}$ in the limit of large depth and width. Recall that for $\mu \mathrm{PC}$ , $a_{L} = 1 / N$ and $a_{\ell} = 1 / \sqrt{NL}$ for $\ell = 2,\dots ,H$ (see Table 1). Because the output weights factor in every term of the rescaling $\mathbf{S}$ except for the identity, these terms will all vanish at a $1 / N$ rate as $N\to \infty$ , i.e. $\mathbf{W}_L\mathbf{W}_L^T /N^2\sim \mathcal{O}(1 / N)$ . The depth, on the other hand, scales the number of terms in $\mathbf{S}$ . Therefore, the width will have to grow with the depth at some constant ratio $L / N$ which can be thought of as the aspect ratio of the network [50]—to make the contribution of each term as small as possible. In the limit of this ratio $r\rightarrow 0$ , the energy rescaling (Eq. 32) approaches the identity $\mathbf{S} = \mathbf{I}$ , the equilibrated energy converges to the MSE $\mathcal{F}_{\mu \mathrm{PC}}(\mathbf{z}^{*},\pmb {\theta}) = \mathcal{L}_{\mu \mathrm{P}}(\pmb {\theta})$ and so PC computes the same gradients as BP. + +# A.3 Additional experiments + +# A.3.1 Ill-conditioning with training + +For the setting in Fig. 3, we also ran experiments with Adam as inference algorithm and ResNets with standard GD. All the results were tuned for the weight learning rate (see §A.4 for more details). We found that Adam led to more ill-conditioned inference landscapes associated with significantly lower and more unstable performance than GD (Figs. 3 & A.23). + +![](images/ea2b07a6c36707aff7f26abe28e0ea21cd536bcff223ee9745af96a90203bb89.jpg) +Figure A.8: Same results as Fig. 3 with Adam as inference algorithm (MNIST). + +![](images/2e9c684ecc39afc141c87c815e5d5c22ad1d9f25ecd09f9c986be8b9b51cda19.jpg) + +![](images/a96e55f3d6af6c093d9825199459eef077199dac4588d9819f5fa589c3af28f2.jpg) + +Interestingly, while skip connections induced much more extreme ill-conditioning (Fig. A.22), performance was equal to, and sometimes significantly better than, networks without skips (Figs. A.9 & A.25), suggesting a complex relationship between trainability and the geometry of the inference landscape which we return to in §A.3.6. + +![](images/346c7197fc21ef6435ed951fd1f7dd0362bbafbbedac40986eff415c0aa8a4b8.jpg) + +![](images/f3f1fb5bfc6acf517ac6aecb4ab11f81b54abfc1867fdfc99004774d10ce91c2.jpg) + +![](images/b389c0fe38fe1a507d604911f2668d9079ca1e2c80f5322e4ca5b04860c50abd.jpg) + +![](images/506d53cffe7af7292eb01d8503dce52679495b0dcd2dbd10304bd73ede51126c.jpg) +Figure A.9: Same results as Fig. 3 with skip connections (MNIST). + +![](images/c764e23d38f749c23b7c3548af45474b11f3429577a3ac4711119aaf9bd8fae6.jpg) + +![](images/02bc3687c850aadc93284e714eb315bc21f83acceb6f9141392da7c1de5421a7.jpg) + +# A.3.2 Activity initialisations + +Here we present some additional results on the initialisation of the activities of PCNs. All experiments used fully connected ResNets, GD as activity optimiser, and as many inference steps as the number of hidden layers. For intuition, we start with linear scalar PCNs or chains. First, we verify that the ill-conditioning of the inference landscape (§3.1) causes the activities to barely move during inference, and increasing the activity learning rate leads to divergence for both forward and random initialisation (Fig. A.10). Similar results are observed for $\mu$ PC (see Fig. A.35). + +![](images/322559a6c1a2a6e7a7010cca0a9809f6e09f4073732328b9d392a44aee9a2c20.jpg) +Figure A.10: Ill-conditioning of the inference landscape prevents convergence to the analytical solution regardless of initialisation. For different initialisations (forward and random) and activity learning rates $\beta$ , we plot the activities of a 64-layer scalar PCN over inference at the start of training. The theoretical activities were computed using Eq. 4. The task was a simple toy regression with $y = -x + \epsilon$ with $x \sim \mathcal{N}(1,1)$ and $\epsilon \sim \mathcal{N}(0,0.5)$ . A standard Gaussian was used for random initialisation, $z_{\ell} \sim \mathcal{N}(0,1)$ . Results were similar across different random seeds. + +For wide linear PCNs with forward initialisation, we find similar results except that $\mu \mathrm{PC}$ seems to initialise the activities close to the analytical solution (Fig. A.11). The same pattern of results is observed for nonlinear networks (Fig. A.36), although note that in this case we do have an analytical solution. These results might suggest that one does not need to perform many inference steps to achieve good performance with $\mu \mathrm{PC}$ . However, we found that one inference step led to worse performance (including as a function of depth) (Figs. A.14 & A.27) compared to as many steps as number of hidden layers (Figs. A.16 & A.18). + +![](images/c1bdc782654c49d28ae3d1f55154fe6b02d64a41ad2aff9277fc30dc89d71c3d.jpg) +Figure A.11: The forward pass of $\mu \mathbf{PC}$ seems to initialise the activities close to the analytical solution (Eq. 4). Similar to Fig. A.10, we plot the $\ell^2$ norm of the activities over inference of 16-layer linear PCNs ( $N = 128$ ) at the start of training (MNIST). Again, results were similar across different random initialisations. + +# A.3.3 Activity decay + +In §4, we discussed how it seems impossible to achieve good conditioning of the inference landscape without making the forward pass unstable (e.g. by zeroing out the weights). We identified one way of inducing relative well-conditionness at initialisation without affecting the forward pass, namely adding an $\ell^2$ norm regulariser on the activities $\frac{\alpha}{2}\sum_{\ell}^{H}||\mathbf{z}_{\ell}||^{2}$ with $\alpha = 1$ . This effectively induces a unit shift in the Hessian spectrum and bounds the minimum eigenvalue at one rather than zero (see §A.2.3). However, we find that PCNs with any degree of activity regularisation $\alpha$ are untrainable (Fig. A.12). + +![](images/c741a6b06d9330b13c12c0b3d36aa8c5fc444a5a3ff308b42bcc1e67e109dd0b.jpg) +Figure A.12: Activity decay induces well-conditioned inference at the cost of performance. Left: Same plot as Fig. 2 with an added activity regulariser $\frac{\alpha}{2} ||\mathbf{z}_{\ell}||^2$ with $\alpha = 1$ . Right: Maximum test accuracy on MNIST achieved by a linear PCN with $N = 128$ and $H = 8$ over activity regularisers of varying strength $\alpha$ . Solid lines and (barely visible) shaded regions indicate the mean and standard deviation across 3 random seeds, respectively. + +# A.3.4 Orthogonal initialisation + +As mentioned in §5, in addition to $\mu \mathrm{PC}$ we also tested PCNs with orthogonal initialisation as a parameterisation ensuring stable forward passes at initialisation for some activation functions (§4; Fig. A.30). We found that this initialisation was not as effective as $\mu \mathrm{PC}$ (Figs. A.13 & A.26), likely due to loss of orthogonality of the weights during training. Adding an orthogonal regulariser could help, but at the cost of an extra hyperparameter to tune. We also find that, except for linear networks, the ill-conditioning of the inference landscape still grows and spikes during training, similar to other parameterisations (e.g. Fig. 3). + +![](images/4376bde30bde3590b25c104cc9f12989e75f25fea1fc8fbd9f6db22f609aa2d1.jpg) +Figure A.13: Test accuracies in Fig. 1 for orthogonal initialisation. Note that performance is expected to drop for ReLU networks which cannot have stable forward passes with orthogonal weights (Fig. A.30). We also plot the condition number of the activity Hessian over training. + +![](images/74230a5aa1676a0452ae15a25f84185b8987a44edbe940c5a4b0256ce176bd51.jpg) + +![](images/3629700e1982b5ed07c9b046f338c898750c25be21dac85ad23dfc0b14f9081f.jpg) + +# A.3.5 $\mu$ PC with one inference step + +All the experiments with $\mu \mathrm{PC}$ (e.g. Fig. 1) used as many inference steps as hidden layers. Motivated by the results of §A.3.2 showing that the forward pass of $\mu \mathrm{PC}$ seems to initialise the activities close to the analytical solution for DLNs (Eq. 4), we also performed experiments with a single inference step. We found that this led to a degradation in performance not only at initialisation but also as a function of depth (Figs. A.14 & A.27), suggesting that some number of steps is still necessary despite $\mu \mathrm{PC}$ appearing to initialise the activities close to the inference solution (Fig. A.11). Similar to other parameterisations, we find that the ill-conditioning of the inference landscape grows and spikes during training. + +![](images/913fb06a1c4546836dd20350cc5e3f2d1f3f34e14e0f29d03702217ca27fdea1.jpg) +Figure A.14: $\mu$ PC test accuracies in Fig. 1 with one inference step. We also plot the condition number of the activity Hessian during training. + +![](images/d49e9300cc29146f8175a755333ee5cabfb5956684f8b926a34dc9ac70135d35.jpg) + +![](images/aa6f68a053cd324b115610e3482fe2e9ee2ffe214dbc6c0ba220e8d7b2f6e92a.jpg) + +# A.3.6 Is inference convergence sufficient for good generalisation? + +Our analysis of the conditioning of the inference landscape ( $\S 3.1$ ) could be argued to rely on the assumption that converging to a solution of the inference dynamics is beneficial for learning and ultimately performance. This question has yet to be resolved, with some works showing both theoretical and empirical benefits for learning close to the inference equilibrium [61, 22], while others argue to take only one step [57]. As discussed in $\S 7$ , our results suggest that convergence close to a solution is necessary for successful training (or monotonic decrease of the loss), which for brevity we will refer to as "trainability". In particular, $\mu \mathrm{PC}$ seems to the activities much closer to the analytical solution (Eq. 4) than the SP ( $\S \mathrm{A}.3.2$ ), and training $\mu \mathrm{PC}$ with one inference step leads to worse performance (e.g. Fig. A.14) than with as many as hidden layers (e.g. Fig. 1). + +Here we report another experiment that speaks to this question and in particular suggests that while inference convergence is necessary for trainability, it is insufficient for good generalisation, at least + +for standard PC. Training linear ResNets of varying depth on MNIST with "perfect inference" (using Eq. 4), we observe that even the deepest $(H = 32)$ networks now become trainable with standard PC in the sense that the training and test losses decrease monotonically (Fig. A.15). However, the starting point of the test losses substantially increases with the depth, and the test accuracies of the deepest networks remain at chance level. These results do not contradict our analysis but highlight the important distinction between trainability and generalisation. Our analysis addresses the former, while the latter is beyond the scope of this work. + +![](images/c417b1a7c3c5d55058658feee8ddcc380e4147c763e30d8ea61e93a56004eb65.jpg) +Figure A.15: Train and test metrics of standard PCNs of varying depth trained with analytical inference (Eq. 4). We plot the training loss, test loss and test accuracy of ResNets ( $N = 128$ ) trained with standard PC on MNIST by solving for inference analytically (using Eq. 4). All experiments used Adam as optimiser with learning rate $\eta = 1e^{-3}$ . Solid lines and shaded regions represent the mean and standard deviation across 3 random initialisations. + +![](images/8364307b06cb087ab903faa2c585480734b8ecbe584335a953c1bc07e2f701d3.jpg) + +![](images/0a02a3da5dd45bc5684c2c27eb125fa2628b7cd26c25c76031f0744dee4699ea.jpg) + +# A.4 Experimental details + +Code to reproduce all the experiments is available at https://github.com/thebuckleylab/jpc/experiments/mupc_paper. We always used no biases, batch size $B = 64$ , Adam as parameter optimiser, and GD as inference optimiser (with the exception of Figs. A.8 & A.24). For the SP, all networks used Kaiming Uniform $(\mathbf{W}_{\ell})_{ij} \sim \mathcal{U}(-1 / N_{\ell - 1}, 1 / N_{\ell})$ as the standard (PyTorch) initialisation used to train PCNs. + +$\mu \mathrm{PC}$ experiments (e.g. Fig. 1). For the test accuracies in Figs. 1 & A.16, we trained fully connected ResNets (Eq. 24) to classify MNIST with standard PC, $\mu \mathrm{PC}$ and BP with Depth- $\mu \mathrm{P}$ . To ensure fair comparison, BP with Depth- $\mu \mathrm{P}$ employed the same scalings as $\mu \mathrm{PC}$ . All networks had width $N = 512$ and always used as many GD inference iterations as the number of hidden layers $H \in \{2^i\}_{i=3}^7$ . To save compute, we trained only for one epoch and evaluated the test accuracy every 300 iterations. For $\mu \mathrm{PC}$ , we selected runs based on the best results from the depth transfer (see Hyperparameter transfer below). For standard PC, we conducted the same grid search over the weight and activity learning rates as used for $\mu \mathrm{PC}$ . For BP, we performed a sweep over learning rates $\eta \in \{1e^0, 5e^{-1}, 1e^{-1}, 5e^{-2}, 1e^{-2}, 5e^{-3}, 1e^{-3}, 5e^{-4}, 1e^{-4}\}$ at depth $H = 8$ , and transferred the optimal value to the deepest ( $H = 128$ ) networks presented. + +Fig. A.20 shows similar results for $\mu \mathrm{PC}$ based on the width transfer results. Fig. A.17 was obtained by extending the training of the 128 ReLU networks in Fig. 1 to 5 epochs. Figs. A.14 & A.27 were obtained with the same setup as Fig. 1 by running $\mu \mathrm{PC}$ for a single inference step. As noted in §5, the results on Fashion-MNIST (Fig. A.18) were obtained with depth transfer by tuning 8-layer networks and transferring the optimal learning rates to 128 layers. + +Hessian condition number at initialisation (e.g. Fig. 2). For different activation functions (Fig. 2), architectures (Fig. A.22) and parameterisations (Fig. 4), we computed the condition number of the activity Hessian (Eq. 5) at initialisation over widths and depths $N, H \in \{2^i\}_{i=1}^7$ . This was the maximum range we could achieve to compute the full Hessian matrix given our memory resources. No biases were used since these do not affect the Hessian as explained in §A.2.1. Results did not differ significantly across different seeds or input and output data dimensions, as predicted from the structure of the activity Hessian (Eq. 5). + +For the landscape insets of Fig. 2, the energy landscape was sampled around the linear solution of the activities (Eq. 4) along the maximum and minimum eigenvectors of the Hessian $\mathcal{F}(\mathbf{z}^{*} + \alpha \hat{\mathbf{v}}_{\mathrm{min}} + \beta \hat{\mathbf{v}}_{\mathrm{min}})$ , with domain $\alpha, \beta \in [-2,2]$ and $30 \times 30$ resolution. + +Hessian condition number over training (e.g. Fig. 3). For different activations (e.g. Fig. 3), architectures (e.g. Fig. A.9), algorithms (e.g. Fig. A.8) and parameterisations (e.g. Fig. A.13), we trained networks of width $N = 128$ and hidden layers $H \in \{8, 16, 32\}$ to perform classification on MNIST and Fashion-MNIST. This set of widths and depths was chosen to allow for tractable computation of the full activity Hessian (Eq. 5). Training was stopped after one epoch to illustrate the phenomenon of ill-conditioning. All experiments used weight learning rate $\eta = 1e^{-3}$ and performed a grid search over activity learning rates $\beta \in \{5e^{-1}, 1e^{-1}, 5e^{-2}\}$ . A maximum number of $T = 500$ steps was used, and inference was stopped when the norm of the activity gradients reached some tolerance. + +Hyperparameter transfer (e.g. Fig. 5). For the ResNets trained on MNIST with $\mu \mathrm{PC}$ (e.g. Fig. 1), we performed a 2D grid search over the following learning rates: $\eta \in \{5e^{-1}, 1e^{-1}, 5e^{-2}, 1e^{-2}\}$ for the weights, and $\beta \in \{1e^3, 5e^2, 1e^2, 5e^1, 1e^1, 5e^0, 1e^0, 5e^{-1}, 1e^{-1}, 5e^{-2}, 1e^{-2}\}$ for the activities. We trained only for one epoch, in part to save compute and in part based on the results of [3, Fig. B.3] showing that the optimal learning rate could be decided after just 3 epochs on CIFAR-10. The number of (GD) inference iterations was always the same as the number of hidden layers. For the width transfer results, we trained networks of 8 hidden layers and widths $N \in \{2^i\}_{i=6}^{10}$ , while for the depth transfer we fixed the width to $N = 512$ and varied the depth $H \in \{2^i\}_{i=3}^7$ . Note that this means that the plots with title $N = 512$ and $H = 8$ in Figs. 5 & A.31-A.32 are the same. The landscape contours were averaged over 3 different random seeds, and the training loss is plotted on a log scale to aid interpretation. + +Loss vs energy ratios (e.g. Fig. 6). We trained ResNets (Eq. 24) to classify MNIST for one epoch with widths and depths $N$ , $H \in \{2^i\}_{i=1}^6$ . To replicate the successful setup of Fig. 1, we used the same learning rate for the optimal linear networks trained on MNIST, $\eta = 1e^{-1}$ . To verify Theorem 1, at every training step we computed the ratio between the Depth-µP MSE loss $\mathcal{L}(\theta)$ and the equilibrated $\mu \mathrm{PC}$ energy $\mathcal{F}(\mathbf{z}^{*}, \boldsymbol{\theta})$ (Eq. 31), where $\mathbf{z}^{*}$ was computed using Eq. 4. All experiments used the weight learning rate $\eta = 1e^{-4}$ . Fig. A.33 shows the same results for the SP, which used a smaller learning rate $\eta = 1e^{-4}$ to avoid divergence at large depth. All the phase diagrams are plotted on a log scale for easier visualisation. Fig. A.34 shows an example of the ratio dynamics of $\mu \mathrm{PC}$ vs PC for a ResNet with 4 hidden layers and different widths. Results were similar across different random initialisations. + +# A.5 Compute resources + +The experiments involving $\mu \mathrm{PC}$ , hyperparameter transfer, and the monitoring of the condition number of the Hessian during training were all run on an NVIDIA RTX A6000. The runtime varied by experiment, with the 128-layer networks trained for multiple epochs (Figs. A.17-A.18) taking several days. All other experiments were run on a CPU and took between one hour and half a day, depending on the specific experiment. + +# A.6 Supplementary figures + +![](images/f122c17db9562cdd7709d552e670edab1eb8f7b79867c324013662352b6aa015.jpg) + +![](images/e1ceeff659c958aaf2f6ade5b41ad0695f1373786deae7dc7ba52fb6dc5f516e.jpg) + +![](images/7994df71fd92783a87b799162c5c795a13248b24a42708b73419ec3a2d85e8c3.jpg) + +![](images/575326242af2fba48b3715834d6a99b08ef9a7bafd30a93ead507a98c4afc07e.jpg) +Figure A.16: Test accuracies in Fig. 1 for different activation functions. Solid lines and shaded regions indicate the mean and standard deviation across 3 random seeds, respectively. BP represents BP with Depth- $\mu$ P. + +![](images/f1ef7ca7da60705c4fcf086dc23a57f881fc0f5ab22f9651b207c03194748827.jpg) +Figure A.17: 128-layer residual ReLU network trained with $\mu$ PC on MNIST for 5 epochs. Solid lines and (barely visible) shaded regions indicate the mean and standard deviation across 5 random seeds, respectively. BP represents BP with Depth- $\mu$ P. + +![](images/f05531b422b77909c75c30179e83d9a14bfd3ad414255ab80004e3cb561be9c1.jpg) +Figure A.18: 128-layer residual ReLU network trained with $\mu$ PC on Fashion-MNIST. Solid lines and (barely visible) shaded regions indicate the mean and standard deviation across 3 random seeds, respectively. BP represents BP with Depth- $\mu$ P. + +![](images/0a29a8f88f17f92dd35f9a388db04dc8ff9e12145ed296dc9aa93e19fe698b6c.jpg) +Figure A.19: 128-layer fully connected residual ReLU network trained with $\mu$ PC on CIFAR10. Solid lines and (barely visible) shaded regions indicate the mean and standard deviation across 3 random seeds, respectively. BP represents BP with Depth- $\mu$ P. As for other datasets, we see that $\mu$ PC remains capable of training such deep networks, although performance slightly lags behind BP. Note that accuracies for all algorithms are far from SOTA because of the fully connected (as opposed to convolutional) architecture used. + +![](images/b86a6e19a91c4a411c077797bf4f4357e7d1149cf0cfacf785c7bd8bdd86e17a.jpg) +Figure A.20: Same results as Fig. 1 varying the width $N$ and fixing the depth at $H = 8$ , showing that "wider is better" [69, 26]. + +![](images/80eea2f00fcd560e062d253515ee8ccb8fd93786108f73d764b3e1e7d5d6587f.jpg) + +![](images/0b718378994d645cdff0b1be611fed854ce1bf22327818c60ef82bf8a430376b.jpg) + +![](images/b8d2fcb633f548730705b94066dde0d767f062a360e2c59761b3157a03c851f0.jpg) +Figure A.21: Toy illustration of the ill-conditioning of the inference landscape. Plotted is the activity or inference landscape $\mathcal{F}(z_1, z_2)$ for a toy linear network with two hidden units $f(x) = w_3w_2w_1x$ , along with the GD dynamics. One weight was artificially set to a much higher value than the others to induce ill-conditioning. + +![](images/08862a7565c0bdac4f9b221d172a121c9a8e62357da4b9d4d562857b2be4328a.jpg) +Figure A.22: Same results as Fig. 2 for the activity Hessian of ResNets (Eq. 25). + +![](images/bd84396e68d9dd2df02dd89ddc359fd911917d57659e7ad02ee0d182e33d34c1.jpg) + +![](images/af03d02287089f420ed44c17550229ebd39c088bd6164d302ba4883cc111ed93.jpg) + +![](images/6fc709439131ed1cc59d8f6692897af30c29e36db61625020ea09fd2f12120d8.jpg) +Figure A.23: Same results as Fig. 3 for Fashion-MNIST. + +![](images/a10e6c9ec197843177671d5400b54a4bc07ba43976ef4c2cad633e4dfd798f33.jpg) + +![](images/0e6e14842badb8a51bdd4e3c46f7512e92493c47908522f15a0dd255b2a41e81.jpg) + +![](images/5d8220c99ae206cc03f4886dad85eb34474afc09cca9c1c12f35f3de096f2b74.jpg) +Figure A.24: Same results as Fig. A.8 for Fashion-MNIST. + +![](images/5adc53001160520aebfa4f3a8ab200a47060678e957934f1fd502dba21cae926.jpg) + +![](images/60e998501fdac60dcd4f820e9ef256007347282ff9832278d5ea5749719ecf0b.jpg) + +![](images/9afef336b03be2989fe63a56e12da8d2572b33e0fa0f52c831c109bf6e0c62b2.jpg) +Figure A.25: Same results as Fig. A.9 for Fashion-MNIST. + +![](images/0ff9ef19abae8eb59f4459cf59fc8f7d70ea91a030a7062f4819234fdfb5082d.jpg) + +![](images/d787c8526b79750d910d8417f0b3dcf20b21d462b00ead431f793419b4feb3d1.jpg) + +![](images/219d4adbe339e238c0c01dbf1db918dbd37655aa73fb6a6d59e994dcffc1368a.jpg) + +![](images/eabb81e11368100fa62e9dec86a783dc29fab87706aa71f469c8aeab40dc54f7.jpg) + +![](images/d2547fffbbbb4b2beaab57db9b935e2361064628d23795b7e4aebbc4532c1183.jpg) + +![](images/47928bdcfa7fa9e7789812fcc6f0767d2c1a0cf0d1ed0d1464fab7044c5908ba.jpg) +Figure A.26: Same results as Fig. A.13 for Fashion-MNIST. + +![](images/b69c86f7cc4f3453bbf75f9fba87c41cc18029a87dffae4d5d2c10cdd04757c1.jpg) + +![](images/1cf67dd210915d33eb0f990b7add324f32f13cd7e80b672008d683d89f4b466e.jpg) + +![](images/6739d85163ffdd2c367ba01d231c075b4fe0990154d931c9b6ed49cd8348d8a2.jpg) +Figure A.27: Same results as Fig. A.14 for Fashion-MNIST. + +![](images/9f907b41bf80a74f9aaa9707a1909874208dfebd0b67c221647224b34e5517c5.jpg) + +![](images/06b27bdd1159fb9d128ca4b33be6abddd5da00aa9b1aef49f53b6bc20b47149d.jpg) + +![](images/bfaff107dedd53fc5a412e3ed9264febc83734d1d67ec3a26feed85ae081e04b.jpg) +Figure A.28: Inference conditioning during training for some $\mu \mathrm{PC}$ networks in Fig. 1. + +![](images/fcca9c25060e4f3bb39d5eaf24236664ff5b6ad7b03c9334b78f359803f6726b.jpg) + +![](images/1a8ddcd3d8a0770de3a8f70a8cd00156098cb733f92a44a5290411d3a47a63e8.jpg) + +![](images/eb6311358200dae3db6c1e96dff45981256697603be284d5f5b87128847f2c31.jpg) +Figure A.29: Same results as Fig. A.28 for Fashion-MNIST. + +![](images/5ed94401b3ddc1b5f3a53c3005e34541303afbb200d74729fca05079e10a919a.jpg) + +![](images/96aa4747fe5bfc487f8f73b96b722c95a6cd99e064a29105f5b1d1a5949d28f6.jpg) + +![](images/35d57142a10720014488e3948cc2686c03557b7b05249ae4f32729ca0590e66c.jpg) +Figure A.30: Forward pass (in)stability with network depth for different parameterisations. For different activation functions and parameterisations, we plot the mean $\ell^1$ norm of the feedforward pass activities at initialisation as a function of the network depth $L$ . Networks ( $N = 1024$ ) had skip connections for the standard parameterisation (SP) and Depth- $\mu$ P but not orthogonal. Results were similar across different seeds. + +![](images/2b6f7c994558e53f5c9367a62f354ec4768afb5afd3d5d9b11c3fc37f794bd07.jpg) +Figure A.31: Same results as Fig. 5 for Linear. + +![](images/152113ef25ed4122ec70a5f61c26d29eb80c7a65feea62add098c280f45b3e22.jpg) +Figure A.32: Same results as Fig. 5 for ReLU. + +![](images/c8a6d043f2d0071072fc11804e693678dfe9827cb65b222f259664456c91f715.jpg) +Figure A.33: Same results as Fig. 6 for the standard parameterisation (SP). + +![](images/2c49970536335a468a723addd6124f77e34ba8d61e87cf5638af6184a0edbf8a.jpg) +Figure A.34: Example of the loss vs energy ratio dynamics of SP and $\mu \mathbf{PC}$ for $H = 4$ . + +![](images/f492a2519703f2bf98d4ee29877445c7acf585e35ad50960c2b7bea4016c6457.jpg) +Figure A.35: Same results as Fig. A.10 for $\mu \mathbf{PC}$ + +![](images/7650ca6667f2383e0434e93a4720be525f59699ca6e1b7e02af8e2a1bc506703.jpg) +Figure A.36: Same results as Fig. A.11 for a ReLU network. \ No newline at end of file diff --git a/NeurIPS/2025/$_mu$PC_ Scaling Predictive Coding to 100+ Layer Networks/images.zip b/NeurIPS/2025/$_mu$PC_ Scaling Predictive Coding to 100+ Layer Networks/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f37889d9a795614f351c46acf15faddf67c9f9ce --- /dev/null +++ b/NeurIPS/2025/$_mu$PC_ Scaling Predictive Coding to 100+ Layer Networks/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:90b100db9f59dd7ac433526912851f13379353b7fe7b5f116348d9bafcddb193 +size 2066241 diff --git a/NeurIPS/2025/$_mu$PC_ Scaling Predictive Coding to 100+ Layer Networks/layout.json b/NeurIPS/2025/$_mu$PC_ Scaling Predictive Coding to 100+ Layer Networks/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..7a421ae467313ef304d53faabeff4b64d09ebb2a --- /dev/null +++ b/NeurIPS/2025/$_mu$PC_ Scaling Predictive Coding to 100+ Layer Networks/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8ddd369fca73e6f17412a788829e741d851ca5a23e32396691cb5f587ef33da2 +size 1195800 diff --git a/NeurIPS/2025/$_textit{HiMaCon_}$ Discovering Hierarchical Manipulation Concepts from Unlabeled Multi-Modal Data/1c795ae1-a64d-41eb-8ed5-e0f5ebeac451_content_list.json b/NeurIPS/2025/$_textit{HiMaCon_}$ Discovering Hierarchical Manipulation Concepts from Unlabeled Multi-Modal Data/1c795ae1-a64d-41eb-8ed5-e0f5ebeac451_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..43d20230e8ee2e39823f8fb1196f859728bf255d --- /dev/null +++ b/NeurIPS/2025/$_textit{HiMaCon_}$ Discovering Hierarchical Manipulation Concepts from Unlabeled Multi-Modal Data/1c795ae1-a64d-41eb-8ed5-e0f5ebeac451_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3df2d622af1c74944e217ad8bed7ea49def29d66862428ef2521cf5514728363 +size 221988 diff --git a/NeurIPS/2025/$_textit{HiMaCon_}$ Discovering Hierarchical Manipulation Concepts from Unlabeled Multi-Modal Data/1c795ae1-a64d-41eb-8ed5-e0f5ebeac451_model.json b/NeurIPS/2025/$_textit{HiMaCon_}$ Discovering Hierarchical Manipulation Concepts from Unlabeled Multi-Modal Data/1c795ae1-a64d-41eb-8ed5-e0f5ebeac451_model.json new file mode 100644 index 0000000000000000000000000000000000000000..eb285ccfe5e749ca0d0f93962d88b7ad1b0dd5cb --- /dev/null +++ b/NeurIPS/2025/$_textit{HiMaCon_}$ Discovering Hierarchical Manipulation Concepts from Unlabeled Multi-Modal Data/1c795ae1-a64d-41eb-8ed5-e0f5ebeac451_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a31698edfce1beab2ab798d696ced4adc77418e881a10bd710612967ce559e54 +size 285363 diff --git a/NeurIPS/2025/$_textit{HiMaCon_}$ Discovering Hierarchical Manipulation Concepts from Unlabeled Multi-Modal Data/1c795ae1-a64d-41eb-8ed5-e0f5ebeac451_origin.pdf b/NeurIPS/2025/$_textit{HiMaCon_}$ Discovering Hierarchical Manipulation Concepts from Unlabeled Multi-Modal Data/1c795ae1-a64d-41eb-8ed5-e0f5ebeac451_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..fcbba4eb8016bcc9bfa6207f5fcb4a77046cda04 --- /dev/null +++ b/NeurIPS/2025/$_textit{HiMaCon_}$ Discovering Hierarchical Manipulation Concepts from Unlabeled Multi-Modal Data/1c795ae1-a64d-41eb-8ed5-e0f5ebeac451_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:95b45bf29e50bbc6a3d6804566afd1835f24c1702454aace51e97b621dd173b3 +size 17601020 diff --git a/NeurIPS/2025/$_textit{HiMaCon_}$ Discovering Hierarchical Manipulation Concepts from Unlabeled Multi-Modal Data/full.md b/NeurIPS/2025/$_textit{HiMaCon_}$ Discovering Hierarchical Manipulation Concepts from Unlabeled Multi-Modal Data/full.md new file mode 100644 index 0000000000000000000000000000000000000000..65c79b8cdb74217735e1e75a192563345aceb7d8 --- /dev/null +++ b/NeurIPS/2025/$_textit{HiMaCon_}$ Discovering Hierarchical Manipulation Concepts from Unlabeled Multi-Modal Data/full.md @@ -0,0 +1,973 @@ +# HiMaCon: Discovering Hierarchical Manipulation Concepts from Unlabeled Multi-Modal Data + +Ruizhe Liu1 Pei Zhou1 Qian Luo1,4 Li Sun1 Jun Cen3 Yibing Song3 Yanchao Yang1,2 + +$^{1}$ HKU Musketeers Foundation Institute of Data Science, The University of Hong Kong + $^{2}$ Department of Electrical and Electronic Engineering, The University of Hong Kong + $^{3}$ DAMO Academy, Alibaba Group $^{4}$ Transcengram {zrl1rz360,pezhou,qianluo,sunlids}@connect.hku.hk {cenjun.cen,songyibing.syb}@alibaba-inc.com, yanchaoy@hku.hk + +# Abstract + +Effective generalization in robotic manipulation requires representations that capture invariant patterns of interaction across environments and tasks. We present a self-supervised framework for learning hierarchical manipulation concepts that encode these invariant patterns through cross-modal sensory correlations and multilevel temporal abstractions without requiring human annotation. Our approach combines a cross-modal correlation network that identifies persistent patterns across sensory modalities with a multi-horizon predictor that organizes representations hierarchically across temporal scales. Manipulation concepts learned through this dual structure enable policies to focus on transferable relational patterns while maintaining awareness of both immediate actions and longer-term goals. Empirical evaluation across simulated benchmarks and real-world deployments demonstrates significant performance improvements with our concept-enhanced policies. Analysis reveals that the learned concepts resemble human-interpretable manipulation primitives despite receiving no semantic supervision. This work advances both the understanding of representation learning for manipulation and provides a practical approach to enhancing robotic performance in complex scenarios. Code is available at: https://github.com/zrllrz/HiMaCon. + +# 1 Introduction + +Robot manipulation in diverse, unstructured environments remains a fundamental challenge. Despite advances in policy learning and architectures [4, 10, 19, 24], current approaches often fail when encountering unexpected variations or novel scenarios. As illustrated in Fig. 1, a policy trained to place cups into containers may succeed in familiar settings but fail when encountering unexpected barriers—revealing a critical generalization gap limiting real-world deployment. + +We propose that addressing this challenge requires learning transferable manipulation concepts—hierarchical abstractions capturing fundamental manipulation patterns. These concepts connect low-level actions to high-level goals, enabling robust generalization. For example, the concept of "placing an object inside a container" encompasses invariant relational patterns that persist whether the container has barriers or not, allowing adaptation while maintaining core manipulation strategy. + +To acquire these manipulation concepts, we propose a self-supervised framework that learns hierarchical latent representations without requiring labor-intensive human annotations [13, 28, 40]. Our approach operates through two complementary mechanisms: 1) Cross-modal correlation learning captures invariant patterns across different sensory modalities (vision, proprioception), enabling generalization across visual variations while preserving functional relationships. When placing ob + +![](images/719c23536f8e1c5725c5819fe923c8b3f5b1f5796aaf5c22f6ef6d73f69ffb62.jpg) +Figure 1: Manipulation concepts enhance generalization. Top: Training data with cups and containers without barriers. Middle: Without manipulation concepts, policies fail when encountering barriers. Bottom: With our concept enhancement, policies adapt accordingly. + +jects in containers, these correlations encode the relationship between visual perception of container boundaries and proprioceptive feedback during placement, regardless of container appearance. 2) Multi-horizon sub-goal organization structures concepts hierarchically across temporal scales, from immediate actions (e.g., "align gripper with object") to extended sequences (e.g., "transport object to container"). This hierarchical representation enables policies to simultaneously reason about immediate actions and longer-term goals, maintaining task coherence even when specific execution paths require adaptation. + +Our experiments across both simulated benchmark tasks and real-world robot deployments demonstrate that policies enhanced with these manipulation concepts consistently outperform conventional approaches, particularly in challenging scenarios requiring adaptation to novel objects, unexpected obstacles, and environmental variations (Fig. 1). The learned concepts form interpretable clusters that resemble meaningful manipulation primitives, providing insights into how robots perceive and reason about manipulation tasks. + +In summary, our key contributions include: (1) a self-supervised framework that extracts structured hierarchical manipulation concepts from unlabeled multi-modal demonstrations, capturing both cross-modal correlations and multi-level temporal abstractions without human annotation; (2) an effective policy enhancement approach that integrates these concepts through joint prediction, maintaining compatibility with diverse policy architectures; and (3) comprehensive empirical evidence demonstrating significant performance improvements across diverse settings, with analyses revealing how learned concepts enable more robust generalization to novel environments. + +# 2 Related Work + +Representation Learning in Robotics Self-supervised representation learning has emerged as a powerful approach for extracting meaningful skills [29, 32, 48] from robotic data, avoiding the need for manual annotation in methods such as [13, 28, 37, 40]. Initial efforts explored single-modality approaches for vision-based [7, 11, 51, 65, 68] and proprioception-based [26, 39, 45, 52] representation learning. Recent work integrates multiple modalities, combining vision with language [22, 36, 42, 50, 64], proprioception with vision [47, 62, 66], even richer [6, 53, 69]. + +These approaches typically focus on cross-modal alignment but often overlook the structured temporal patterns inherent in manipulation tasks. Parallel developments in temporal representation learning have addressed this challenge through various approaches: time-contrastive learning [27, 34, 35, 42, 65], temporal masked auto-encoding [47], and explicit modeling of state transitions across different + +![](images/daf3f3599f5ad9ad1662fc78636bca535892fe1604b1f6a93359c15f87ee6f5f.jpg) +1. Manipulation Concept Discovery +Figure 2: The proposed self-supervised manipulation concept discovery and policy enhancement. Stage 1: The concept encoder $(\mathcal{E})$ processes multi-modal robot demonstrations to extract concept latents. These latents are refined through two objectives: (1) the Cross-Modal Correlation Network $(\mathcal{C})$ employs a mask-and-predict strategy to capture persistent patterns across sensing modalities (Sec. 3.2); (2) the Multi-Horizon Future Predictor $(\mathcal{F})$ enables concept latents to organize hierarchically into multi-horizon sub-goals based on coherence thresholds $(\epsilon)$ (Sec. 3.3). Stage 2: The learned concepts are integrated into policy learning through a backbone network $(\pi_h)$ with concept $(\pi_z)$ and action $(\pi_a)$ prediction heads, regularizing action generation with structured manipulation knowledge (Eq. 9). + +timescales [25, 44, 54, 55]. Our work advances the field by simultaneously addressing both multimodal integration and hierarchical temporal structures, creating representations that naturally align with manipulation sub-goals at varying time horizons while leveraging cross-modal correlational patterns that persist across different objects and contexts. + +Concept-Guided Robotic Policies Concept-guided approaches enhance robotic policy performance by leveraging intermediate representations that bridge perception and action. These methods generally fall into two categories. First, two-module frameworks [3, 8, 29, 62, 68, 73] employ a dual-model architecture where one component extracts high-level task concepts while another generates the corresponding actions. While effective, these approaches often require specialized architectural designs that limit their applicability across different policy classes. + +Second, in contrast, joint prediction approaches [18, 20, 56, 58, 67] integrate concept guidance by training policies to simultaneously predict both concepts and actions. This creates an implicit information pathway where concept understanding regularizes action generation. Our work adopts this more flexible approach, enabling seamless integration with diverse policy architectures while maintaining the interpretability benefits of explicit concept representations. + +# 3 Method + +We aim to encode robotic manipulation demonstrations into latent representations that capture task-induced patterns in multi-modal sensory-motor data. These representations should naturally cluster according to functional sub-goals, providing insights into manipulation objectives and enhancing policy learning. We term these clusters manipulation concepts—each representing action sequences targeting specific sub-goals—and call the learning process manipulation concept discovery. + +Our self-supervised approach works without explicit sub-goal annotations, addressing the challenge of capturing meaningful manipulation patterns without labels. We design objective functions enforcing latent representations that reflect both temporal structure and cross-modal correlations. Our approach ensures: (1) integration of modality-specific features while encoding cross-modal correlations that persist across objects and contexts; (2) hierarchical organization of sub-processes representing sub-goals across temporal horizons and enabling action prediction guided by immediate and long-term + +objectives. We validate these concepts through policy performance improvements and multiple analysis methods that demonstrate correspondence with meaningful manipulation primitives. + +# 3.1 Problem Setup & Manipulation Concept Encoder + +Given a dataset $D = \{\tau_i\}_{i=1}^N$ of $N$ manipulation trajectories, each $\tau_i = \{(\mathbf{o}_i^t, a_i^t)\}_{t=1}^{T_i}$ contains observations $\mathbf{o}_i^t$ and actions $a_i^t$ at time $t$ . For $M$ modalities, $\mathbf{o}_i^t = \{o_i^{1,t}, o_i^{2,t}, \ldots, o_i^{M,t}\}$ , where $o_i^{m,t}$ is the observation of the modality $m$ . We denote $\mathbf{o}_i^{S,t} = \{o_i^{m,t} \mid m \in S\}$ as observations of modalities in $S \subseteq [M] = \{1, 2, \dots, M\}$ . We treat observations of both the same sensory modes (e.g., multiple views) and different modes as distinct modalities, as they are functionally different modalities in terms of complementary information. + +The Manipulation Concept Discovery process assigns latent representation $z_{i}^{t} \in \mathbb{R}^{Z}$ to each timestep $t$ of the trajectory $\tau_{i}$ , where $z_{i}^{t}$ can be viewed as a noisy sampling of the underlying manipulation concept active at $t$ . Since these representations cluster based on sub-goals, we refer to $z_{i}^{t}$ as manipulation concept latents or simply manipulation concepts. We use continuous representations for differentiability and to avoid constraints of codebook-based discrete representations (e.g., finite capacity). To learn $z_{i}^{t}$ , we introduce a manipulation concept encoder $\mathcal{E}$ parameterized by $\Theta_{\mathcal{E}}$ , which maps observation sequence $\mathbf{o}_i = \{\mathbf{o}_i^t\}_{t=1}^{T_i}$ from trajectory $\tau_{i}$ to concept sequence $\mathbf{z}_i = \{z_i^t\}_{t=1}^{T_i}$ : + +$$ +\mathbf {z} _ {i} \leftarrow \mathcal {E} \left(\mathbf {o} _ {i}; \Theta_ {\varepsilon}\right) \tag {1} +$$ + +We implement $\mathcal{E}$ using a transformer to encode temporal dependencies (details in Sec. A.1). Next, we elaborate on the training strategies optimizing cross-modal and multi-horizon temporal correlation metrics (Sec. 3.2 and 3.3). + +# 3.2 Capturing Multi-Modal Correlations + +To enhance the utility of multi-modal information, we propose that manipulation concepts should capture cross-modal correlations rather than simply aggregating features from different modalities (e.g., concatenating multi-modal signals [9, 71]). Physiological evidence suggests that concept formation often occurs when correlations across sensory modalities are high [1, 15, 38, 57, 63]. These correlations remain consistent across scenarios involving the same concept, facilitating generalization. For instance, in container opening tasks, the correlated patterns between visual lid rotation, characteristic force feedback, and audio cues persist across different container types, enabling the transfer of the "opening" concept despite variations in object appearances. + +To learn manipulation concepts that capture cross-modal correlations, we propose maximizing mutual information—a metric capable of modeling diverse correlations—between observations from different modalities, conditioned on the associated manipulation concept. Specifically, we maximize the conditional mutual information over bipartitions of modality observations: + +$$ +\max _ {\mathbf {Z}} \sum_ {S \subsetneq [ M ], S \neq \emptyset} \mathbb {I} \left(\mathbf {O} _ {S}: \mathbf {O} _ {[ M ] \backslash S} \mid \mathbf {Z}\right), \tag {2} +$$ + +where $\mathbf{O}_S$ are observations from a subset of modalities, $\mathbf{O}_{[M]\backslash S}$ are observations from remaining modalities, and $\mathbf{Z}$ is the manipulation concept. We implement Eq. 2 using a computationally efficient self-supervised mask-and-predict approach that stochastically samples bipartitions during training. This ensures scalability despite exponentially increasing bipartition numbers while integrating cross-modal correlation learning with multi-modal information compression. + +Specifically, a Cross-Modal Correlation Network $\mathcal{C}$ (CMCN) with parameters $\Theta_{c}$ reconstructs full-modality observations from partial observations guided by manipulation concepts. During training, we mask observations from a random subset $S$ of modalities and reconstruct all observations $\mathbf{o}_i^t$ using the unmasked subset $\mathbf{o}_i^{[M]\backslash S,t}$ and concept $z_i^t$ : + +$$ +\mathcal {L} _ {\mathrm {m m}} (t, \tau_ {i}) = \mathbb {E} _ {S} \left\| \mathcal {C} \left(\mathbf {o} _ {i} ^ {[ M ] \backslash S, t}, z _ {i} ^ {t}; \Theta_ {c}\right) - \mathbf {o} _ {i} ^ {t} \right\|, \tag {3} +$$ + +where $S \sim \mathrm{U}\left(2^{[M]} \setminus \{\emptyset\}\right)$ is a uniformly sampled non-empty subset of modality indices. By predicting full observations from partial inputs, we maximize the conditional mutual information in Eq. 2, forcing manipulation concepts $z_{i}^{t}$ to capture cross-modal correlations. Additionally, when all modalities are masked, reconstruction solely from $z_{i}^{t}$ ensures these representations compress and preserve essential multi-modal information (please see Sec. A.1 for more details). + +# 3.3 Representing Multi-Horizon Sub-Goals + +To complete tasks with hierarchical structures, manipulation concepts must encode multi-horizon sub-goal information. Physiological evidence shows human actions are hierarchically organized [17, 41], with coarse-grained goals defining overall tasks and fine-grained goals informing immediate actions. These multi-horizon sub-goals link ultimate goals with low-level actions, enabling smooth transitions while enhancing robustness. + +We aim to make manipulation concepts organized to encode sub-processes across multiple temporal horizons without explicit annotations. Since concepts cluster by sub-goals, hierarchical sub-goals can emerge from these clusters at varying temporal scales. We propose that the temporal extent of a sub-process is determined by concept latent coherence within clusters, yielding a natural spectrum from short-horizon to long-horizon sub-goals. Specifically, given manipulation concept latents $\mathbf{z}_i = \{z_i^t\}_{t=1}^{T_i}$ from trajectory $\tau_i$ , we quantify their similarities using spherical distance: $\mathrm{dist}(z,u) = \frac{1}{\pi}\arccos \left\langle \frac{z}{\|z\|_2}, \frac{u}{\|u\|_2} \right\rangle$ . Concepts belong to the same sub-process if their distance falls below a coherence threshold $\epsilon \in [0,1]$ . More explicitly, sub-processes are derived as: + +$$ +\left. \mathrm {h} \left(\mathbf {z} _ {i}; \epsilon\right) = \left\{\left[ g _ {k}, g _ {k + 1}\right) \mid k = 1, 2, \dots , K \left(\mathbf {z} _ {i}; \epsilon\right) \right\}, \right. +$$ + +$$ +\text {w h e r e} g _ {1} = 1, \quad g _ {k + 1} = \max _ {g} \left\{g \mid g \in \left(g _ {k}, T _ {i} + 1 \right] \cap \mathbb {N} ^ {+} \wedge \forall t, t ^ {\prime} \in \left[ g _ {k}, g\right), \operatorname {d i s t} \left(z _ {i} ^ {t}, z _ {i} ^ {t ^ {\prime}}\right) < \epsilon \right\}, \tag {4} +$$ + +where $K(\mathbf{z}_i;\epsilon)$ is the number of clusters determined by $\epsilon$ , and increasing $\epsilon$ yields sub-processes spanning from short-horizon to long-horizon. Please see Alg. 1 for more details. + +Furthermore, we propose learning objectives to ensure multi-horizon sub-processes from Eq. 4 align with meaningful sub-goal completion processes. Specifically, the manipulation concept guiding each sub-process should be informative about the state achieved upon sub-task completion [5, 33, 72]. For all coherence thresholds $\epsilon$ , current observation $\mathbf{O}$ and its associated concept $\mathbf{Z}$ should be informative of the terminal observation $\mathbf{O}^{\mathrm{goal}(\epsilon)}$ , characterized by minimizing the following conditional entropy: + +$$ +\forall \epsilon , \min _ {\mathbf {Z}} \mathbb {H} \left(\mathbf {O} ^ {\text {g o a l} (\epsilon)} \mid \mathbf {O}, \mathbf {Z}\right), \tag {5} +$$ + +To implement Eq. 5, we train a Multi-Horizon Future Predictor $\mathcal{F}$ (MHFP) to hallucinate terminal observations of different sub-processes. For time step $t$ in trajectory $\tau_{i}$ , the terminal observation is determined by the ending time step of the interval containing $t$ : + +$$ +\mathrm {g} (t; \mathbf {z} _ {i}, \epsilon) = \min \left\{T _ {i}, g _ {k + 1} \right\}, \text {w h e r e} t \in \left[ g _ {k}, g _ {k + 1}\right) \in \mathrm {h} (\mathbf {z} _ {i}; \epsilon), \tag {6} +$$ + +During training, the network $\mathcal{F}$ , parameterized by $\Theta_f$ , predicts this terminal observation based on current observation $\mathbf{o}_i^t$ , manipulation concept $z_i^t$ , and coherence threshold $\epsilon$ : + +$$ +\mathcal {L} _ {\mathrm {m h}} (t, \tau_ {i}) = \mathbb {E} _ {\epsilon} \left\| \mathcal {F} \left(\mathbf {o} _ {i} ^ {t}, z _ {i} ^ {t}, \epsilon ; \Theta_ {f}\right) - \mathbf {o} _ {i} ^ {\mathrm {g} (t; \mathbf {z} _ {i}, \epsilon)} \right\|, \tag {7} +$$ + +where $\epsilon \sim \mathrm{U}([0,1])$ is sampled uniformly per iteration to improve efficiency by avoiding training over all $\epsilon$ values. This training process iteratively improves both latents and sub-process derivation: we compute manipulation concepts using the encoder (Eq. 1), determine sub-process boundaries, then update all networks, including $\mathcal{F}$ and the concept encoder. This improves future observation prediction and concept latents, which in turn refines sub-process derivation. By minimizing Eq. 7, $z_i^t$ is ensured to encode multi-horizon sub-goal information, indicating hierarchical transitions to terminal states under various $\epsilon$ while adjusting sub-processes by shaping concept latents for terminal state predictability. More details can be found in Sec. A.1. + +Final Objective for Manipulation Concept Discovery. We jointly optimize the multi-modal correlation objective (Eq. 3) and multi-horizon sub-goal prediction objective (Eq. 7) to ensure manipulation concepts generated by the encoder $\mathcal{E}$ (Eq. 1) satisfy both key properties: + +$$ +\mathcal {L} _ {\mathrm {z}} (t, \tau_ {i}) = \lambda_ {\mathrm {m m}} \mathcal {L} _ {\mathrm {m m}} (t, \tau_ {i}) + \lambda_ {\mathrm {m h}} \mathcal {L} _ {\mathrm {m h}} (t, \tau_ {i}), \tag {8} +$$ + +where $\lambda_{\mathrm{mm}}, \lambda_{\mathrm{mh}} > 0$ balance the two loss terms. + +# 3.4 Enhancing Imitation Learning with Manipulation Concepts + +After learning manipulation concepts through our self-supervised framework, we address how these concepts enhance policy learning. Unlike previous approaches that learn task-specific policies + +directly from demonstrations [12, 28], we propose to leverage the learned manipulation concepts as an informative representation that bridges low-level actions and high-level goals. + +Specifically, with manipulation concepts $\mathbf{z}_i$ generated by encoder $\mathcal{E}$ , we augment imitation learning by training policies to predict both ground-truth actions and corresponding concepts [21, 67, 70]. This approach uses concept prediction as a regularization that guides the policy to encode conceptual understanding alongside action planning: + +$$ +h _ {i} ^ {t} = \pi_ {h} \left(\mathbf {o} _ {i} ^ {t}, \ell_ {i}; \Theta_ {\pi} ^ {h}\right), \quad \hat {z} _ {i} ^ {t} = \pi_ {z} \left(h _ {i} ^ {t}; \Theta_ {\pi} ^ {z}\right), \quad \hat {a} _ {i} ^ {t} = \pi_ {a} \left(h _ {i} ^ {t}; \Theta_ {\pi} ^ {a}\right), \tag {9} +$$ + +$$ +\mathcal {L} _ {\pi} (t, \tau_ {i}, \ell_ {i}) = \| \hat {a} _ {i} ^ {t} - a _ {i} ^ {t} \| + \lambda_ {\mathrm {m c}} \| \hat {z} _ {i} ^ {t} - z _ {i} ^ {t} \|. +$$ + +The policy consists of: (1) A backbone $\pi_h$ processing task descriptions $\ell_i$ and observations $\mathbf{o}_i^t$ to produce a shared representation $h_i^t$ ; (2) A concept predictor $\pi_z$ mapping $h_i^t$ to predicted concepts $\hat{z}_i^t$ ; and (3) An action decoder $\pi_a$ mapping $h_i^t$ to predicted actions $\hat{a}_i^t$ . This joint objective enforces the policy to leverage concept information encoded within $h_i^t$ while predicting actions. Even though concepts are learned task-agnostically for generalization, the policy receives task descriptions in a multi-task setting, serving as a mechanism to learn the reuse of concepts. The learning objective balances action and concept prediction using $\lambda_{\mathrm{mc}} > 0$ . More details are provided in Sec. A.2. + +# 4 Experiments + +We evaluate our manipulation concept discovery approach through experiments addressing four key questions: (1) Do learned concepts enhance policy performance on tasks used for concept discovery, validating our strategies for encoding cross-modal correlations (Sec. 3.2) and multi-horizon sub-goals (Sec. 3.3)? (2) Can concepts learned from one task set transfer effectively to different tasks sharing underlying manipulation patterns? (3) Does our concept discovery mechanism generalize to novel tasks with decreased overlap in manipulation patterns? (4) What interpretable properties emerge in the learned concepts that explain their effectiveness for robotic manipulation? Through these investigations, we demonstrate both the immediate benefits of our approach for imitation learning and its broader applicability for transfer learning and generalization in manipulation tasks. + +# 4.1 Experimental Setup + +Dataset and Environment Sec. 4.2 and 4.3 conduct experiments using the LIBERO benchmark [30], a comprehensive platform for robotic learning built on Robosuite [75]. We utilize three distinct task sets: + +- LIBERO-90: A diverse collection of 90 manipulation tasks serving as our primary training domain for concept discovery and initial policy learning. +- LIBERO-LONG: 10 novel long-horizon tasks, each composed of two LIBERO-90 tasks in sequence, designed to evaluate transfer to more complex task structures. +- LIBERO-GOAL: 10 tasks in an entirely novel environment unseen during concept discovery, used to evaluate the generalization of learned concepts to unfamiliar contexts. + +Each task includes a natural language description and 50 expert demonstrations. For multi-modal observations, we use: Agentview vision: $128 \times 128$ RGB third-person camera capturing the entire environment; Eye-in-hand vision: $128 \times 128$ RGB gripper-mounted camera; Proprioceptive state: 9D vector encoding gripper position, rotation, and physical states. + +Manipulation Concept Discovery Methods We compare our approach with several state-of-the-art concept discovery baselines (implementation details in Sec. A.3): + +- InfoCon [31]: A VQ-VAE type of method for single-hierarchy concept discovery. +- XSkill [65]: Contrastive learning for manipulation skill extraction from demonstration videos. +- DecisionNCE [27]: Learns reward-relevant representations from demonstrations with language annotations, evaluated in two variants: using task instructions (DecisionNCE-task) and using elementary action labels (DecisionNCE-motion). +- RPT [47]: Temporally and modality-masked autoencoder for multi-modal sequence modeling. +- All: A simplified variant of our approach that predicts all modalities from concepts without modeling cross-modal correlations. + +Table 1: Evaluation of manipulation concept discovery methods across different task settings. Success rates $(\%)$ of ACT and Diffusion Policy (DP) models when enhanced with manipulation concepts from various discovery methods. All concept encoders were trained only on LIBERO-90, and evaluated on: original tasks (L90-90), novel long-horizon compositions (L90-L), and entirely new environments (L90-G). Values in parentheses show standard deviations across 4 seeds. Bold and underlined values indicate best and second-best results. + +
L90-90InfoConXSkillRPTAllNextCLIPDINOv2DecisionNCE taskmotionPlainOurs
ACT66.5 (0.8)73.4 (0.8)68.8 (0.8)64.1 (2.0)68.0 (0.4)63.8 (0.5)71.9 (0.3)69.0 (0.1)66.8 (0.8)46.6 (1.9)74.8 (0.8)
DP78.2 (0.6)87.7 (0.6)84.3 (0.1)81.5 (0.5)82.6 (0.1)80.7 (0.9)79.4 (0.1)75.7 (0.8)82.7 (0.6)75.1 (0.6)89.6 (0.6)
L90-LInfoConXSkillRPTAllNextCLIPDINOv2DecisionNCE taskmotionPlainOurs
ACT55.5 (0.9)55.0 (1.0)59.0 (1.0)55.5 (0.9)55.0 (1.0)51.0 (1.0)55.0 (1.0)53.0 (1.0)49.3 (0.9)54.0 (0.9)63.0 (1.0)
DP75.0 (1.0)73.0 (1.0)61.3 (0.9)79.3 (0.9)83.0 (1.0)67.0 (1.0)63.0 (1.0)58.7 (0.9)52.7 (0.9)34.1 (1.1)89.0 (1.0)
L90-GInfoConXSkillRPTAllNextCLIPDINOv2DecisionNCE taskmotionPlainOurs
ACT67.0 (1.0)77.0 (1.0)75.0 (1.0)69.0 (1.0)71.0 (1.0)77.0 (1.0)77.3 (0.9)70.0 (0.9)75.0 (0.5)57.0 (1.0)81.0 (1.0)
DP92.7 (0.9)93.0 (1.0)91.5 (0.9)91.0 (1.0)91.3 (0.9)92.0 (0.9)91.0 (0.7)92.0 (0.8)93.0 (1.0)90.7 (0.9)95.7 (0.7)
+ +- Next: Predicts adjacent time-step observations, a common approach adopted in [7, 68]. +- CLIP [46]: Language-aligned visual features from a pretrained foundation model. +- DINOv2 [43]: Self-supervised visual representations without temporal modeling. +- Plain: Standard imitation learning without manipulation concepts. + +Policies for Concept-Enhanced Imitation Learning To evaluate the effectiveness of our discovered manipulation concepts, we integrate them into two established imitation learning frameworks using the joint prediction approach described in Sec. 3.4: + +- ACT [71]: A transformer-based conditional variational autoencoder that predicts action chunks. +- Diffusion Policy (DP) [9]: A 1D convolutional UNet that generates actions through denoising. + +For both policy architectures, we add the concept prediction head $(\pi_z$ in Eq. 9) to predict manipulation concepts from the shared concept-aware representations. Implementation details appear in Sec. A.2. All experiments are reported with results aggregated across 4 random seeds. + +# 4.2 Evaluating Policy Performance with Learned Manipulation Concepts + +- Performance on Original Training Tasks We first evaluate our concept discovery method on the same tasks used for concept training. As shown in the L90-90 results (Tab. 1), our approach consistently outperforms all baselines with both policy architectures. The performance gap between our method and Next/InfoCon demonstrates the importance of multi-hierarchical sub-goal modeling, while improvements over All highlight the value of explicitly capturing cross-modal correlations. Our method also surpasses DecisionNCE variants despite not requiring language supervision, validating the effectiveness of our self-supervised objectives. +- Transfer to Long-Horizon Tasks To evaluate concept transferability to more complex compositions, we apply concept encoders trained on LIBERO-90 directly to LIBERO-LONG demonstrations featuring novel long-horizon tasks. The L90-L results show our method maintains its performance advantage in this challenging transfer setting. This demonstrates that our approach learns manipulation concepts that effectively decompose hierarchical tasks, enabling policies to better handle novel complex task compositions requiring sequential execution of multiple sub-goals. + +Table 2: Impact of modality combinations on concept discovery performance. Success rates $(\%)$ of ACT and DP policies using manipulation concepts discovered with different input modality combinations. All models were trained and evaluated on LIBERO-90, with specific modalities excluded (marked with “-”). A: agentview vision, H: eye-in-hand vision, P: proprioceptive state. + +
Ours-HPA-PAH---P-H-A--
ACT74.8±0.870.5±1.871.3±0.370.1±1.267.5±0.868.7±0.669.4±0.4
DP89.6±0.685.8±0.285.6±0.384.3±0.584.8±0.183.7±0.185.3±0.5
+ +- Generalization to Novel Environments We further test generalization by applying concept encoders trained on LIBERO-90 directly to LIBERO-GOAL demonstrations featuring unseen environments and tasks. The $L90-G$ results show our method continues to outperform all baselines in this challenging scenario. This indicates our approach discovers fundamental manipulation primitives that transfer effectively across environmental variations. +- Impact of Multi-Modal Observations Our ablation study (Tab. 2) shows that performance consistently improves as more modalities are incorporated. The most significant drops occur when removing proprioceptive information, highlighting its importance for grounding visual observations with physical interaction states and confirming the value of our cross-modal correlation approach. + +# 4.3 Analyzing Manipulation Concept Properties + +Enhanced Cross-Modal Correlation To verify our Cross-Modal Correlation Network's effectiveness (Sec. 3.2), we measure mutual information between modalities conditioned on concept latents (Sec. A.4). Tab. 3 shows that our approach achieves higher conditional mutual information than the All baseline. This confirms that our mask-and-predict strategy enables the concept encoder to capture persistent cross-modal patterns that generalize across different objects and contexts, providing a robust representational basis for policies. + +Alignment with Semantic Sub-Goals We evaluate whether our concepts align with human-understandable semantics by grouping latents from different demonstrations based on human-identified sub-goals and computing similarities between these groupings: + +Table 3: Conditional mutual information between modality pairs. Values conditioned on concept latents from our method versus the All baseline that does not model cross-modal correlations. A: agentview, H: eye-in-hand vision, P: proprioception. + +
OursAll
I(oH: oA | z)3.79992.0080
II(oP: oA | z)4.83193.1312
III(oP: oH | z)4.82553.1322
+ +$$ +\left\langle C _ {i}, C _ {j} \right\rangle = \frac {1}{| C _ {i} | | C _ {j} |} \sum_ {z _ {i} \in C _ {i}} \sum_ {z _ {j} \in C _ {j}} \left\langle \frac {z _ {i}}{\| z _ {i} \| _ {2}}, \frac {z _ {j}}{\| z _ {j} \| _ {2}} \right\rangle , \tag {10} +$$ + +where $C_i$ , $C_j$ represent human-identified sub-goal categories, and $z_i$ , $z_j$ are latents within each category (details in Sec. C.2). As shown in Fig. 4, similarity matrices consistently show the highest values along the diagonal, demonstrating that our approach discovers concepts that exhibit clustering patterns corresponding to meaningful manipulation primitives. + +Multi-Level Hierarchical Structure Varying the coherence threshold $\epsilon$ in Eq. 4 reveals the hierarchical organization of our learned concepts. Fig. 3 (and Sec. C.5) shows larger $\epsilon$ values identify coarse-grained phases, while smaller values capture fine-grained actions. This emergent hierarchy enables policies to simultaneously reason about immediate actions and longer-term goals without explicit hierarchical supervision, contributing to improved performance on complex sequential tasks that require coordinated execution across multiple temporal scales. + +# 4.4 Real-World Validation + +Real-World Generalization Study To study generalization capabilities, we deploy concept-enhanced policies on a Mobile ALOHA robot [16] in "cleaning cup" tasks (Fig. 5). Training data includes only simple container arrangements with consistent color pairings. + +![](images/9d9db51387c838315c132507a374f426961abecd98a271bab7b39b673f8fb7a9.jpg) +Task: "Open the top drawer of the cabinet and put the bowl in it" + +![](images/56b489dc2cb7430550127832cceec2db70d1743e6f24c84e445e051e5cfdb859.jpg) + +![](images/176798c3342ebbeac83415b10df9a6b915293e46e4377a6a9efe27a3f2e6dac3.jpg) + +![](images/be94ec36047eed4b62db418ceb08780a3456c8c6714a3ffa341a5e271df2f5b3.jpg) + +![](images/7e3eed407810ae3952e601e428b0374029fd68b9f7bae0f73891aa8c3d71977d.jpg) +Discovered Manipulation Concepts at Multiple Granularities + +![](images/80fce9c7364423b1d632469f75755920759b662bf6bdb18f4601d7cf464a0055.jpg) +Figure 3: Multi-granular task decomposition through concept latent clustering. Visualization of sub-processes derived by clustering manipulation concept latents at different coherence thresholds $(\epsilon)$ for the task "open the top drawer and put the bowl in it." Higher $\epsilon$ values (top rows) produce coarser decompositions, while lower values (bottom rows) yield finer-grained segmentation. The emergent sub-processes naturally align with semantic task components, for example, the third segment in row 2 corresponds to "put bowl in drawer," while the second segment in row 4 corresponds to "pull drawer open." This demonstrates our method's ability to discover hierarchical, human-interpretable task structures without explicit supervision. + +We test on six increasingly challenging variations: (1) Novel Placements: Cups and containers in unseen arrangements; (2) Color Composition: Altered cup-container color pairings; (3) Novel Objects: Entirely unseen containers, cups, and plates; (4) Obstacles: Objects between the robot and the cups obstructing vision; (5) Barriers: Internal dividers within containers impeding placement; (6) Grasping Together: Two adjacent cups requiring simultaneous grasp. + +As shown in Tab. 4, policies enhanced with our manipulation concepts consistently outperform baselines across all scenarios, with advantages in challenging conditions. We suggest that the following two mechanisms behind learned manipulation concepts improve generalization: + +![](images/8b46ea2edf681d50280d784d381013390f59a63f8ef205730f89c9bfd28ad607.jpg) +Figure 4: Semantic alignment of learned concepts. Cosine similarity between concept latents grouped by human-defined sub-goals. Diagonal patterns demonstrate that our approach discovers concepts that exhibit clustering patterns corresponding to meaningful manipulation primitives. + +# 1. Relational focus: Concept-enhanced policies + +prioritize transferable relational patterns (e.g., "object inside container") over surface features. Our cross-modal correlation learning (Sec. 3.2) enables this capability by identifying patterns that remain invariant across modalities. This relational emphasis explains the stronger performance on scenarios that alter visual appearance while preserving task structure. For instance, while Novel Placement tests spatial variation alone, the other protocols introduce substantial visual perturbations (different colors, objects, or occlusions) that shift the appearance distribution. The consistent performance gains across these visually diverse scenarios (Tab. 4) suggest that the learned concepts successfully capture the underlying relational invariant —placing cups into containers —rather than memorizing superficial visual patterns. + +2. Hierarchical awareness: Concept-enhanced policies exhibit more systematic failure recovery than baselines, suggesting better tracking of sub-goal completion. Baseline failures frequently exhibit + +![](images/867823d6736674af731b3dbc381d85de66c9f55e3eafaa2879df71efe0b0dd71.jpg) +Real Robot +Figure 5: Real-world generalization evaluation with Mobile ALOHA robot. Left: Mobile ALOHA robot setup for cup cleaning tasks. Center: Training conditions with simple, consistent cup-container color pairings. Right: Six test variations with increasing complexity: novel placements, altered color combinations, unfamiliar objects, external obstacles, internal barriers, and simultaneous grasping of multiple cups. These variations test the policy's ability to generalize beyond training conditions by systematically introducing new challenges. +Fig. 7 (Sec. C.7) shows predicted goal states when conditioned on the current observation, manipulation concept, and various coherence thresholds $(\epsilon)$ . + +![](images/87f48e8cccefd0e0ab9912706b97ec57fcb3a4d25a9cdbb56d33ccdb7038b375.jpg) +Training + +![](images/ec2549edff82749a85abdf8b225296c93935bb85d7c49293b8cbf37fb12e9b71.jpg) + +![](images/8b4a4b692f2957b192f45ca48df8f13d133f123cb09dce5d6eadac850ac0c12c.jpg) +Test on Unseen Setups + +![](images/2423a8c01d43f291ad342b734c138340a0c03ac3fc15763fb846f5225bb498b4.jpg) + +![](images/4d1dd136555de912c65a90321aa07db8a0ac42be66e1cfcf42bd48543bcdf7a0.jpg) + +![](images/e638f908683692de3e597a36a92dd6bab1f54fd7b17962e7758cc115cfadcda8.jpg) + +![](images/1113312c55390d82c9d8d15c3191eb7ca0a134dba2af578675b5d902d1b215fb.jpg) + +![](images/dca0926b81792dab7364c4cbb832cd24a13ae8baea3f98641d0c9384a5d45d1f.jpg) + +premature task abandonment: the robot moves toward containers without having grasped objects, or hovers near placement locations without executing placement. In contrast, when concept-enhanced policies fail initial grasp attempts, they consistently retry grasping (typically 2-3 attempts) before proceeding, demonstrating recognition of incomplete sub-goals. Although these recoveries ultimately fail due to time limits or object displacement, they reveal structured task progression rather than blind action execution. + +These mechanisms may enable manipulation concepts to promote policy generalization by encoding fundamental spatial and functional relationships that remain consistent across environmental variations. Details are provided in Sec. C.6. + +Multi-Horizon Goal Prediction Visualization To visualize the temporal information encoded in our manipulation concepts, we examine outputs from our Multi-Horizon Goal Predictor (MHGP, $\mathcal{F}$ in Eq. 7) using the BridgeDataV2 dataset [60]. + +Table 4: Real-world generalization success rates $(\%)$ for ACT policies with and without manipulation concepts (MC). Test conditions: Placements (novel layouts), Color (new pairings), Objects (unseen items), Obstacles (external barriers), Barrier (internal dividers), and Multigrasp (two cups simultaneously). + +
PlaceColorObj.Obsy.Barr.Multi
w/o MC53.346.740.020.00.00.0
w/ MC73.360.053.333.320.013.3
+ +The predictions capture essential task structures – such as anticipated arm trajectories and object interactions – rather than attempting pixel-perfect reconstructions. This abstraction of scene-specific details in favor of functional relationships is crucial for cross-environment generalization. Importantly, as $\epsilon$ increases, the predictions correspond to states progressively further into the future, with smaller values showing immediate next steps and larger values revealing final goal states. This demonstrates that our learned concepts encode meaningful temporal structures at multiple time horizons, enabling policies to simultaneously reason about immediate actions and longer-term objectives. Details are provided in Sec. C.7. + +# 5 Discussion + +We demonstrate that self-supervised discovery of hierarchical manipulation concepts significantly enhances robot policy performance across original tasks, novel compositions, and entirely new environments. Three key strengths emerge: (1) our representations naturally resemble semantically meaningful manipulation primitives without requiring explicit labels, as evidenced by diagonal clustering in similarity matrices; (2) the concepts bridge low-level actions and high-level goals through hierarchical organization, enabling reasoning at multiple temporal scales; and (3) concept-enhanced policies focus on transferable relational patterns rather than superficial features, explaining their robust generalization to scenarios with substantial distribution shifts. These findings highlight the potential of learning manipulation concepts from unlabeled multi-modal demonstrations for creating more adaptable and interpretable robotic systems. Limitations are discussed in Sec. D. + +# Acknowledgments and Disclosure of Funding + +This work is supported by the Early Career Scheme of the Research Grants Council (RGC) grant # 27207224, the HKU-100 Award, a donation from the Musketeers Foundation, in part by the JC STEM Lab of Robotics for Soft Materials funded by The Hong Kong Jockey Club Charities Trust, and DAMO Academy through the Alibaba Innovative Research Program. + +# References + +[1] Nikolai Axmacher, Florian Mormann, Guillen Fernandez, Christian E Elger, and Juergen Fell. Memory formation by neuronal synchronization. Brain research reviews, 52(1):170-182, 2006. +[2] Mohamed Ishmael Belghazi, Aristide Baratin, Sai Rajeshwar, Sherjil Ozair, Yoshua Bengio, Aaron Courville, and Devon Hjelm. Mutual information neural estimation. In International conference on machine learning, pages 531-540. PMLR, 2018. +[3] Suneel Belkhale, Tianli Ding, Ted Xiao, Pierre Sermanet, Quon Vuong, Jonathan Tompson, Yevgen Chebotar, Debidatta Dwibedi, and Dorsa Sadigh. Rt-h: Action hierarchies using language. arXiv preprint arXiv:2403.01823, 2024. +[4] Kevin Black, Noah Brown, Danny Driess, Adnan Esmail, Michael Equi, Chelsea Finn, Niccolo Fusai, Lachy Groom, Karol Hausman, Brian Ichter, et al. $\pi_0$ : A vision-language-action flow model for general robot control. arXiv preprint arXiv:2410.24164, 2024. +[5] Kevin Black, Mitsuhiro Nakamoto, Pranav Atreya, Homer Rich Walke, Chelsea Finn, Aviral Kumar, and Sergey Levine. Zero-shot robotic manipulation with pre-trained image-editing diffusion models. In The Twelfth International Conference on Learning Representations, 2024. +[6] Rogerio Bonatti, Sai Vemprala, Shuang Ma, Felipe Frujeri, Shuhang Chen, and Ashish Kapoor. Pact: Perception-action causal transformer for autoregressive robotics pre-training. In 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 3621-3627. IEEE, 2023. +[7] Jake Bruce, Michael D Dennis, Ashley Edwards, Jack Parker-Holder, Yuge Shi, Edward Hughes, Matthew Lai, Aditi Mavalankar, Richie Steigerwald, Chris Apps, et al. Genie: Generative interactive environments. In Forty-first International Conference on Machine Learning, 2024. +[8] Qingwen Bu, Hongyang Li, Li Chen, Jisong Cai, Jia Zeng, Heming Cui, Maoqing Yao, and Yu Qiao. Towards synergistic, generalized, and efficient dual-system for robotic manipulation. arXiv preprint arXiv:2410.08001, 2024. +[9] Cheng Chi, Zhenjia Xu, Siyuan Feng, Eric Cousineau, Yilun Du, Benjamin Burchfiel, Russ Tedrake, and Shuran Song. Diffusion policy: Visuomotor policy learning via action diffusion. arXiv preprint arXiv:2303.04137, 2023. +[10] Embodiment Collaboration, Abby O'Neill, Abdul Rehman, Abhinav Gupta, Abhiram Madukuri, Abhishek Gupta, Abhishek Padalkar, Abraham Lee, Acorn Pooley, Agrim Gupta, Ajay Mandlekar, Ajinkya Jain, Albert Tung, Alex Bewley, Alex Herzog, Alex Irpan, Alexander Khazatsky, Anant Rai, Anchit Gupta, Andrew Wang, Andrey Kolobov, Anikait Singh, Animesh Garg, Aniruddha Kembhavi, Annie Xie, Anthony Brohan, Antonin Raffin, Archit Sharma, Arefeh Yavary, Arhan Jain, Ashwin Balakrishna, Ayzaan Wahid, Ben Burgess-Limerick, Beomjoon Kim, Bernhard Schölkopf, Blake Wulfe, Brian Ichter, Cewu Lu, Charles Xu, Charlotte Le, Chelsea Finn, Chen Wang, Chenfeng Xu, Cheng Chi, Chenguang Huang, Christine Chan, Christopher Agia, Chuer Pan, Chuyuan Fu, Coline Devin, Danfei Xu, Daniel Morton, Danny Driess, Daphne Chen, Deepak Pathak, Dhruv Shah, Dieter Büchler, Dinesh Jayaraman, Dmitry Kalashnikov, Dorsa Sadigh, Edward Johns, Ethan Foster, Fangchen Liu, Federico Ceola, Fei Xia, Feiyu Zhao, Felipe Vieira Frujeri, Freek Stulp, Gaoyue Zhou, Gaurav S. Sukhatme, Gautam Salhotra, Ge Yan, Gilbert Feng, Giulio Schiavi, Glen Berseth, Gregory Kahn, Guangwen Yang, Guanzhi Wang, Hao Su, Hao-Shu Fang, Haochen Shi, Henghui Bao, Heni Ben Amor, Henrik I Christensen, Hiroki Furuta, Homanga Bharadhwaj, Homer Walke, Hongjie Fang, Huy Ha, Igor Mordatch, Iija Radosavovic, Isabel Leal, Jacky Liang, Jad Abou-Chakra, Jaehyung Kim + +Jaimyn Drake, Jan Peters, Jan Schneider, Jasmine Hsu, Jay Vakil, Jeannette Bohg, Jeffrey Bingham, Jeffrey Wu, Jensen Gao, Jiaheng Hu, Jiajun Wu, Jialin Wu, Jiankai Sun, Jianlan Luo, Jiayuan Gu, Jie Tan, Jihoon Oh, Jimmy Wu, Jingpei Lu, Jingyun Yang, Jitendra Malik, Joao Silverio, Joey Hejna, Jonathan Booher, Jonathan Tompson, Jonathan Yang, Jordi Salvador, Joseph J. Lim, Junhyek Han, Kaiyuan Wang, Kanishka Rao, Karl Pertsch, Karol Hausman, Keegan Go, Keerthana Gopalakrishnan, Ken Goldberg, Kendra Byrne, Kenneth Oslund, Kento Kawaharazuka, Kevin Black, Kevin Lin, Kevin Zhang, Kiana Ehsani, Kiran Lekkala, Kirsty Ellis, Krishan Rana, Krishnan Srinivasan, Kuan Fang, Kunal Pratap Singh, Kuo-Hao Zeng, Kyle Hatch, Kyle Hsu, Laurent Itti, Lawrence Yunliang Chen, Lerrel Pinto, Li Fei-Fei, Liam Tan, Linxi "Jim" Fan, Lionel Ott, Lisa Lee, Luca Weihs, Magnum Chen, Marion Lepert, Marius Memmel, Masayoshi Tomizuka, Masha Itkina, Mateo Guaman Castro, Max Spero, Maximilian Du, Michael Ahn, Michael C. Yip, Mingtong Zhang, Mingyu Ding, Minho Heo, Mohan Kumar Srirama, Mohit Sharma, Moo Jin Kim, Naoaki Kanazawa, Nicklas Hansen, Nicolas Heess, Nikhil J Joshi, Niko Suenderhauf, Ning Liu, Norman Di Palo, Nur Muhammad Mahi Shafiullah, Oier Mees, Oliver Kroemer, Osbert Bastani, Pannag R Sanketi, Patrick "Tree" Miller, Patrick Yin, Paul Wohlhart, Peng Xu, Peter David Fagan, Peter Mitrano, Pierre Sermanet, Pieter Abbeel, Priya Sundaresan, Qiuyu Chen, Quan Vuong, Rafael Rafailov, Ran Tian, Ria Doshi, Roberto Mart'in-Mart'in, Rohan Baijal Rosario Scalise Rose Hendrix Roy Lin Runjia Qian Ruohan Zhang Russell Mendonca Rutav Shah Ryan Hoque Ryan Julian Samuel Bustamante Sean Kirmani Sergey Levine Shan Lin Sherry Moore Shikhar Bahl Shivin Dass Shubham Sonawani, Shubham Tulsiani Shuran Song Sichun Xu Siddhant Haldar Siddharth Karamcheti Simeon Adebola Simon Guist Soroush Nasiriany Stefan Schaal Stefan Welker Stephen Tian Subramanian Ramamoorthy Sudeep Dasari Suneel Belkhale Sungjae Park Suraj Nair Suvir Mirchandani Takayuki Osa Tanmay Gupta Tatsuya Harada Tatsuya Matsushima Ted Xiao Thomas Kollar Tianhe Yu Tianli Ding,Todor DavchevTony Z.Zhao Travis ArmstrongTrevor Darrell Trinity Chung,Vidhi Jain,Vikash Kumar Vincent VanhouckeWei Zhan Wenxuan Zhou,Wolfram Burgard Xi Chen Xiangyu Chen Xiaolong Wang Xinghao Zhu Xinyang GengXiyuan Liu Xu Liangwei Xuanlin Li Yansong Pang Yao Lu Yecheng Jason Ma Yejin KimYevgen Chebotar Yifan ZhouYifeng ZhuYilin Wu Ying XuYixuan WangYonatan Bisk Yongqiang Dou Yoonyoung Cho Youngwoo Lee Yuchen Cui Yue Cao Yueh-Hua WuYujin Tang,Yuke ZhuYunchu ZhangYunfan Jiang Yunshuang Li Yunzhu Li Yusuke Iwasawa Yutaka Matsuo Zehan Ma,Zhuo XuZichen Jeff Cui Zichen ZhangZipeng Fu and Zipeng Lin.Open x-embodiment: Robotic learning datasets and rt-x models 2024. + +[11] Sudeep Dasari, Mohan Kumar Srirama, Unnat Jain, and Abhinav Gupta. An unbiased look at datasets for visuo-motor pre-training. In Conference on Robot Learning, pages 1183-1198. PMLR, 2023. +[12] Yan Ding, Xiaohan Zhang, Chris Paxton, and Shiqi Zhang. Task and motion planning with large language models for object rearrangement. In 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 2086-2092. IEEE, 2023. +[13] Ben Eisner, Harry Zhang, and David Held. Flowbot3d: Learning 3d articulation flow to manipulate articulated objects. arXiv preprint arXiv:2205.04382, 2022. +[14] Haoquan Fang, Markus Grotz, Wilbert Pumacay, Yi Ru Wang, Dieter Fox, Ranjay Krishna, and Jiafei Duan. Sam2act: Integrating visual foundation model with a memory architecture for robotic manipulation. In 7th Robot Learning Workshop: Towards Robots with Human-Level Abilities. +[15] Juergen Fell and Nikolai Axmacher. The role of phase synchronization in memory processes. Nature reviews neuroscience, 12(2):105-118, 2011. +[16] Zipeng Fu, Tony Z Zhao, and Chelsea Finn. Mobile aloha: Learning bimanual mobile manipulation with low-cost whole-body teleoperation. arXiv preprint arXiv:2401.02117, 2024. +[17] Scott T Grafton and Antonia F de C Hamilton. Evidence for a distributed hierarchy of action representation in the brain. Human movement science, 26(4):590-616, 2007. +[18] Yanjiang Guo, Yucheng Hu, Jianke Zhang, Yen-Jen Wang, Xiaoyu Chen, Chaochao Lu, and Jianyu Chen. Prediction with action: Visual policy learning via joint denoising process. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. + +[19] Physical Intelligence, Kevin Black, Noah Brown, James Darpinian, Karan Dhabalia, Danny Driess, Adnan Esmail, Michael Equi, Chelsea Finn, Niccolo Fusai, et al. $\pi_{0.5}$ : a vision-language-action model with open-world generalization. arXiv preprint arXiv:2504.16054, 2025. +[20] Zhiwei Jia, Vineet Thumuluri, Fangchen Liu, Linghao Chen, Zhiao Huang, and Hao Su. Chain-of-thought predictive control. In _Forty-first International Conference on Machine Learning_, 2024. +[21] Zhiwei Jia, Vineet Thumuluri, Fangchen Liu, Linghao Chen, Zhiao Huang, and Hao Su. Chain-of-thought predictive control. In International Conference on Machine Learning, pages 21768-21790. PMLR, 2024. +[22] Siddharth Karamcheti, Suraj Nair, Annie S Chen, Thomas Kollar, Chelsea Finn, Dorsa Sadigh, and Percy Liang. Language-driven representation learning for robotics. arXiv preprint arXiv:2302.12766, 2023. +[23] Moo Jin Kim, Chelsea Finn, and Percy Liang. Fine-tuning vision-language-action models: Optimizing speed and success. arXiv preprint arXiv:2502.19645, 2025. +[24] Moo Jin Kim, Karl Pertsch, Siddharth Karamcheti, Ted Xiao, Ashwin Balakrishna, Suraj Nair, Rafael Rafailov, Ethan Foster, Grace Lam, Pannag Sanketi, et al. Openvla: An open-source vision-language-action model. arXiv preprint arXiv:2406.09246, 2024. +[25] Thomas Kipf, Yujia Li, Hanjun Dai, Vinicius Zambaldi, Alvaro Sanchez-Gonzalez, Edward Grefenstette, Pushmeet Kohli, and Peter Battaglia. Compile: Compositional imitation learning and execution. In International Conference on Machine Learning, pages 3418-3428. PMLR, 2019. +[26] Seungjae Lee, Yibin Wang, Haritheja Etukuru, H Jin Kim, Nur Muhammad Mahi Shafiullah, and Lerrel Pinto. Behavior generation with latent actions. arXiv preprint arXiv:2403.03181, 2024. +[27] Jianxiong Li, Jinliang Zheng, Yinan Zheng, Liyuan Mao, Xiao Hu, Sijie Cheng, Haoyi Niu, Jihao Liu, Yu Liu, Jingjing Liu, et al. Decisionne: Embodied multimodal representations via implicit preference learning. In International Conference on Machine Learning, pages 29461-29488. PMLR, 2024. +[28] Xiaoqi Li, Mingxu Zhang, Yiran Geng, Haoran Geng, Yuxing Long, Yan Shen, Renrui Zhang, Jiaming Liu, and Hao Dong. Manipllm: Embodied multimodal large language model for object-centric robotic manipulation. arXiv preprint arXiv:2312.16217, 2023. +[29] Zhixuan Liang, Yao Mu, Hengbo Ma, Masayoshi Tomizuka, Mingyu Ding, and Ping Luo. Skilldiffuser: Interpretable hierarchical planning via skill abstractions in diffusion-based task execution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16467-16476, 2024. +[30] Bo Liu, Yifeng Zhu, Chongkai Gao, Yihao Feng, Qiang Liu, Yuke Zhu, and Peter Stone. Libero: Benchmarking knowledge transfer for lifelong robot learning. In A. Oh, T. Naumann, A. Globerson, K. Saenko, M. Hardt, and S. Levine, editors, Advances in Neural Information Processing Systems, volume 36, pages 44776-44791. Curran Associates, Inc., 2023. +[31] Ruizhe Liu, Qian Luo, and Yanchao Yang. Infocon: Concept discovery with generative and discriminative informativeness. In The Twelfth International Conference on Learning Representations, 2024. +[32] Yuyao Liu, Jiayuan Mao, Joshua B Tenenbaum, Tomás Lozano-Pérez, and Leslie Pack Kaelbling. One-shot manipulation strategy learning by making contact analogies. In 2025 IEEE International Conference on Robotics and Automation (ICRA), pages 15387-15393. IEEE, 2025. +[33] Corey Lynch, Mohi Khansari, Ted Xiao, Vikash Kumar, Jonathan Thompson, Sergey Levine, and Pierre Sermanet. Learning latent plans from play. In Conference on robot learning, pages 1113-1132. PMLR, 2020. + +[34] Yecheng Jason Ma, Vikash Kumar, Amy Zhang, Osbert Bastani, and Dinesh Jayaraman. Liv: Language-image representations and rewards for robotic control. In International Conference on Machine Learning, pages 23301-23320. PMLR, 2023. +[35] Yecheng Jason Ma, Shagun Sodhani, Dinesh Jayaraman, Osbert Bastani, Vikash Kumar, and Amy Zhang. Vip: Towards universal visual reward and representation via value-implicit pre-training. arXiv preprint arXiv:2210.00030, 2022. +[36] Arjun Majumdar, Karmesh Yadav, Sergio Arnaud, Jason Ma, Claire Chen, Sneha Silwal, Aryan Jain, Vincent-Pierre Berges, Tingfan Wu, Jay Vakil, et al. Where are we in the search for an artificial visual cortex for embodied intelligence? Advances in Neural Information Processing Systems, 36:655-677, 2023. +[37] Jiayuan Mao, Tomás Lozano-Pérez, Joshua B Tenenbaum, and Leslie Pack Kaelbling. Learning reusable manipulation strategies. In Conference on Robot Learning, pages 1467-1483. PMLR, 2023. +[38] Lucia Melloni, Carlos Molina, Marcela Pena, David Torres, Wolf Singer, and Eugenio Rodriguez. Synchronization of neural activity across cortical areas correlates with conscious perception. Journal of neuroscience, 27(11):2858-2865, 2007. +[39] Atharva Mete, Haotian Xue, Albert Wilcox, Yongxin Chen, and Animesh Garg. Quest: Self-supervised skill abstractions for learning continuous control. arXiv preprint arXiv:2407.15840, 2024. +[40] Kaichun Mo, Shilin Zhu, Angel X Chang, Li Yi, Subarna Tripathi, Leonidas J Guibas, and Hao Su. Partnet: A large-scale benchmark for fine-grained and hierarchical part-level 3d object understanding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 909–918, 2019. +[41] John D Murray, Alberto Bernacchia, David J Freedman, Ranulfo Romo, Jonathan D Wallis, Xinying Cai, Camillo Padoa-Schioppa, Tatiana Pasternak, Hyojung Seo, Daeyeol Lee, et al. A hierarchy of intrinsic timescales across primate cortex. Nature neuroscience, 17(12):1661-1663, 2014. +[42] Suraj Nair, Aravind Rajeswaran, Vikash Kumar, Chelsea Finn, and Abhinav Gupta. R3m: A universal visual representation for robot manipulation. arXiv preprint arXiv:2203.12601, 2022. +[43] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. +[44] Karl Pertsch, Oleh Rybkin, Jingyun Yang, Shenghao Zhou, Konstantinos Derpanis, Kostas Daniilidis, Joseph Lim, and Andrew Jaegle. Keyframing the future: Keyframe discovery for visual prediction and planning. In Learning for Dynamics and Control, pages 969-979. PMLR, 2020. +[45] Karl Pertsch, Kyle Stachowicz, Brian Ichter, Danny Driess, Suraj Nair, Quan Vuong, Oier Mees, Chelsea Finn, and Sergey Levine. Fast: Efficient action tokenization for vision-language-action models. arXiv preprint arXiv:2501.09747, 2025. +[46] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. +[47] Ilija Radosavovic, Baifeng Shi, Letian Fu, Ken Goldberg, Trevor Darrell, and Jitendra Malik. Robot learning with sensorimotor pre-training. In Conference on Robot Learning, pages 683-693. PMLR, 2023. +[48] Seungeun Rho, Laura Smith, Tianyu Li, Sergey Levine, Xue Bin Peng, and Sehoon Ha. Language guided skill discovery. In The Thirteenth International Conference on Learning Representations. + +[49] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2021. +[50] Dominik Schmidt and Minqi Jiang. Learning to act without actions. In The Twelfth International Conference on Learning Representations, 2024. +[51] Younggyo Seo, Danijar Hafner, Hao Liu, Fangchen Liu, Stephen James, Kimin Lee, and Pieter Abbeel. Masked world models for visual control. In Conference on Robot Learning, pages 1332-1344. PMLR, 2023. +[52] Nur Muhammad Shafiullah, Zichen Cui, Ariuntuya Arty Altanzaya, and Lerrel Pinto. Behavior transformers: Cloning $k$ modes with one stone. Advances in neural information processing systems, 35:22955-22968, 2022. +[53] Rutav Shah, Roberto Martín-Martín, and Yuke Zhu. Mutex: Learning unified policies from multimodal task specifications. arXiv preprint arXiv:2309.14320, 2023. +[54] Archit Sharma, Shixiang Gu, Sergey Levine, Vikash Kumar, and Karol Hausman. Dynamics-aware unsupervised discovery of skills. In International Conference on Learning Representations, 2020. +[55] Shashank Sharma, Vinay Namboodiri, and Janina Hoffmann. Multi-resolution skill discovery for hierarchical reinforcement learning. In NeurIPS 2023 Workshop on Goal-Conditioned Reinforcement Learning, 2023. +[56] Lucy Xiaoyang Shi, Archit Sharma, Tony Z Zhao, and Chelsea Finn. Waypoint-based imitation learning for robotic manipulation. In Conference on Robot Learning, pages 2195-2209. PMLR, 2023. +[57] Wolf Singer. Consciousness and neuronal synchronization. The neurology of consciousness, pages 43-52, 2011. +[58] Yang Tian, Sizhe Yang, Jia Zeng, Ping Wang, Dahua Lin, Hao Dong, and Jiangmiao Pang. Predictive inverse dynamics models are scalable learners for robotic manipulation. In The Thirteenth International Conference on Learning Representations, 2025. +[59] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. +[60] Homer Rich Walke, Kevin Black, Tony Z Zhao, Quan Vuong, Chongyi Zheng, Philippe Hansen-Estruch, Andre Wang He, Vivek Myers, Moo Jin Kim, Max Du, et al. Bridgedata v2: A dataset for robot learning at scale. In Conference on Robot Learning, pages 1723-1736. PMLR, 2023. +[61] Weikang Wan, Yifeng Zhu, Rutav Shah, and Yuke Zhu. Lotus: Continual imitation learning for robot manipulation through unsupervised skill discovery. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 537-544. IEEE, 2024. +[62] Lirui Wang, Xinlei Chen, Jialiang Zhao, and Kaiming He. Scaling proprioceptive-visual learning with heterogeneous pre-trained transformers. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. +[63] Thilo Womelsdorf and Pascal Fries. The role of neuronal synchronization in selective attention. Current opinion in neurobiology, 17(2):154-160, 2007. +[64] Hongtao Wu, Ya Jing, Chilam Cheang, Guangzeng Chen, Jiafeng Xu, Xinghang Li, Minghuan Liu, Hang Li, and Tao Kong. Unleashing large-scale video generative pre-training for visual robot manipulation. arXiv preprint arXiv:2312.13139, 2023. +[65] Mengda Xu, Zhenjia Xu, Cheng Chi, Manuela Veloso, and Shuran Song. Xskill: Cross embodiment skill discovery. In Conference on Robot Learning, pages 3536-3555. PMLR, 2023. +[66] Jonathan Yang, Dorsa Sadigh, and Chelsea Finn. Polybot: Training one policy across robots while embracing variability. arXiv preprint arXiv:2307.03719, 2023. + +[67] Mengjiao Sherry Yang, Dale Schuurmans, Pieter Abbeel, and Ofir Nachum. Chain of thought imitation with procedure cloning. Advances in Neural Information Processing Systems, 35:36366-36381, 2022. +[68] Seonghyeon Ye, Joel Jang, Byeongguk Jeon, Sejune Joo, Jianwei Yang, Baolin Peng, Ajay Mandlekar, Reuben Tan, Yu-Wei Chao, Bill Yuchen Lin, et al. Latent action pretraining from videos. arXiv preprint arXiv:2410.11758, 2024. +[69] Jia Zeng, Qingwen Bu, Bangjun Wang, Wenke Xia, Li Chen, Hao Dong, Haoming Song, Dong Wang, Di Hu, Ping Luo, et al. Learning manipulation by predicting interaction. arXiv preprint arXiv:2406.00439, 2024. +[70] Qingqing Zhao, Yao Lu, Moo Jin Kim, Zipeng Fu, Zhuoyang Zhang, Yecheng Wu, Zhaoshuo Li, Qianli Ma, Song Han, Chelsea Finn, et al. Cot-vla: Visual chain-of-thought reasoning for vision-language-action models. arXiv preprint arXiv:2503.22020, 2025. +[71] Tony Z. Zhao, Vikash Kumar, Sergey Levine, and Chelsea Finn. Learning fine-grained bimanual manipulation with low-cost hardware. In ICML Workshop on New Frontiers in Learning, Control, and Dynamical Systems, 2023. +[72] Haoyu Zhen, Xiaowen Qiu, Peihao Chen, Jincheng Yang, Xin Yan, Yilun Du, Yining Hong, and Chuang Gan. 3d-vla: A 3d vision-language-action generative world model. In _Forty-first International Conference on Machine Learning_, 2024. +[73] Pei Zhou, Ruizhe Liu, Qian Luo, Fan Wang, Yibing Song, and Yanchao Yang. Autocgp: Closed-loop concept-guided policies from unlabeled demonstrations. In The Thirteenth International Conference on Learning Representations. +[74] Yifeng Zhu, Peter Stone, and Yuke Zhu. Bottom-up skill discovery from unsegmented demonstrations for long-horizon robot manipulation. IEEE Robotics and Automation Letters, 7(2):4126-4133, 2022. +[75] Yuke Zhu, Josiah Wong, Ajay Mandlekar, Roberto Martín-Martín, Abhishek Joshi, Soroush Nasiriany, Yifeng Zhu, and Kevin Lin. robosuite: A modular simulation framework and benchmark for robot learning. In arXiv preprint arXiv:2009.12293, 2020. + +# A Implementation details + +# A.1 Manipulation concept discovery (Ours) + +This section details the neural network architectures and training procedures employed in our manipulation concepts discovery framework (Sec. 3) as implemented on the LIBERO benchmark. + +Manipulation Concept Encoder (Sec. 3.1) The manipulation concepts encoder $\mathcal{E}$ (Eq. 1) first encodes the multi-modal observations at each time step of the input demonstration into an encoded vector. It then utilizes a self-attention transformer to process the sequence of encoded vectors into a sequence of manipulation concepts. For the observation encoding process, our experiments on LIBERO incorporate two vision observations: agent-view vision and eye-in-hand vision. The original images are tensors of shape $128 \times 128 \times 3$ . To enhance processing efficiency, we preprocess the images for each time step using the VAE encoder from stable diffusion [49], compressing each image into a tensor of shape $16 \times 16 \times 4$ , which is then flattened into a 1024-dimensional vector. In addition to the two vision observations, we include a 9-dimensional robot state at each time step of each demonstration as the proprioceptive state observation. For these three observations at each time step, we employ three distinct 2-layer MLPs to process each observation into a feature vector of the hidden size (256) used by the subsequent transformer. The encoded features from these observations are then summed to form a 256-dimensional representation that encapsulates the sensing information from the three modalities. + +$$ +h _ {\mathrm {a v}} = \mathrm {M L P} _ {\mathrm {a v}} \left(I _ {\mathrm {a v - c o m p r e s s}}\right) \quad h _ {\mathrm {e v}} = \mathrm {M L P} _ {\mathrm {e v}} \left(I _ {\mathrm {e v - c o m p r e s s}}\right) \quad h _ {\text {p r o p}} = \mathrm {M L P} _ {\text {p r o p}} \left(s _ {\text {p r o p}}\right) \tag {11} +$$ + +$$ +h = h _ {\mathrm {a v}} + h _ {\mathrm {e v}} + h _ {\mathrm {p r o p}} +$$ + +Here, $I_{\mathrm{av - compress}}$ represents the 1024-dimensional compressed agent-view vision, and $I_{\mathrm{ev - compress}}$ represents the 1024-dimensional compressed eye-in-hand vision. $s_{\mathrm{prop}}$ denotes the 9-dimensional proprioceptive state observation. The output of the hidden layers from the three MLPs is 1024 dimensions. The $h$ in Eq. 11 represents the encoded observation feature at each time step of a given demonstration $\tau_{i}$ : $(h_{i}^{1}, h_{i}^{2}, \dots, h_{i}^{T_{i}})$ . The next module in $\mathcal{E}$ is a 12-layer self-attention (MHA in Eq. 12) transformer, enabling each time step to aggregate information from every other time step in the input sequence. In our implementation, we do not input the entire demonstration; instead, the transformer processes a fixed input sequence length of $T_{\mathrm{context}} = 60$ . A learnable temporal embedding, represented as a tensor of shape $60 \times 256$ , is added to the input sequence to enhance temporal representation. The hidden feature dimension at each time step is 256, and each self-attention layer consists of 8 heads. Moreover, since spherical distance is utilized in Sec. 3.3, the output manipulation concepts are normalized to have a unit length with respect to the 2-norm: + +$$ +\left. \left(z _ {i} ^ {t}, z _ {i} ^ {t + 1}, \dots , z _ {i} ^ {t + T _ {\text {c o n t e x t}} - 1}\right) \leftarrow \operatorname {N o r m} _ {2} \left(\left[ \mathrm {M H A} \right] _ {\times 1 2} \left(h _ {i} ^ {t}, h _ {i} ^ {t + 1}, \dots , h _ {i} ^ {t + T _ {\text {c o n t e x t}} - 1}\right)\right) \right. \tag {12} +$$ + +The output manipulation concept sequence in Eq. 1 represents the predicted manipulation concepts at time-steps $t, t + 1, \dots, (t + T_{\text{context}} - 1)$ of the demonstration $\tau_i$ . During training, demonstrations with lengths shorter than $T_{\text{context}}$ are padded to $T_{\text{context}}$ by repeating the observations from the last time-step at the end of each demonstration. During inference, when $\mathcal{E}$ is used to label the demonstrations in the original dataset, the manipulation concepts at each time step are designed to incorporate information from as many future time steps as possible. This approach aims to better capture motion pattern dynamics, aligning with prior works that generate the dynamics at the current time step based on information derived from the dynamics spanning the current to future time steps [68]. Specifically: + +- For each time-step $t \leq T_i - T_{\text{context}}$ , the corresponding manipulation concepts are derived when the input to Eq. 12 starts from this time-step and spans a length of $T_{\text{context}}$ : $\left(h_i^t, h_i^{t+1}, \dots, h_i^{t + T_{\text{context}} - 1}\right)$ . +- For each time step $t > T_{i} - T_{\text{context}}$ , the corresponding manipulation concepts are derived when the input to Eq. 12 begins at time step $h_{i}^{T_{i} - T_{\text{context}} + 1}$ and spans a length of $T_{\text{context}}$ , ensuring that the final time step corresponds to the end of the demonstration: $\left(h_{i}^{T_{i} - T_{\text{context}} + 1}, h_{i}^{T_{i} - T_{\text{context}} + 2}, \dots, h_{i}^{T_{i}}\right)$ . +- If the original demonstration length is smaller than $T_{\text{context}}$ , the manipulation concepts correspond to the input appended with repeated observations as described earlier. + +However, we do not firmly believe this is the optimal approach for labeling manipulation concepts. Further exploration of inference-time strategy design is left for future work, as it is not a core focus of the manipulation concept discovery methodology presented. + +Learning Multi-Modal Features and Correlations (Sec. 3.2) The Cross-Modal Correlation Network $\mathcal{C}$ (Eq. 3) shares a similar structure with $\mathcal{E}$ (Eq. 1). First, it includes four 2-layer MLP encoders, analogous to the three encoders in Eq. 11, with an additional encoder for processing the manipulation concepts. Each of these four MLPs outputs a hidden feature of dimension 1024, which is then reduced to a 256-dimensional encoded feature. These encoded features are summed to represent the combined information from the three observations and the manipulation concepts. Second, it incorporates a 4-layer self-attention transformer to process the sequence of features (with the same fixed length $T_{\mathrm{context}} = 60$ ) produced by the four MLPs. Following this, three 3-layer MLP decoders map the transformer's output to the reconstructed observations at each time step. Unlike in Eq. 12, the transformer's output does not require normalization. Each decoder MLP has hidden layers with a dimension of 1024. As described in Eq. 3, for the three observations—agent-view camera vision, eye-in-hand camera vision, and proprioceptive state observation—we randomly mask these modalities, ensuring that at least one modality is masked during each iteration. The $2^{3} - 1 = 7$ possible masking scenarios follow a uniform distribution, with each scenario appearing with a probability of $\frac{1}{7}$ . For the sampled masks, all observations of the corresponding masked modalities in the input sequence are replaced with zero tensors. The loss is applied separately to the reconstruction of the three different observations. Specifically, L2 loss is applied to the two vision observations, while L1 loss is applied to the proprioceptive state observations. The loss weight $\lambda_{\mathrm{mm}}$ in Eq. 8 is set to 1.0. + +Learning Multi-Hierarchical Sub-goals (Sec. 3.3) The Multi-Horizon Future Predictor $\mathcal{F}$ (Eq. 7) shares a similar structure with $\mathcal{C}$ (Eq. 3). The key differences are: + +- $\mathcal{F}$ does not require a masking strategy. +- The transformer in $\mathcal{F}$ is a 4-layer causal self-attention transformer. Causal attention is used because, in Eq. 7, the prediction is made from each current time step to certain future time steps. Therefore, for each time-step input in $\mathcal{F}$ , access to information from subsequent time steps is restricted. +- To incorporate the granularity parameter $\epsilon \in [0,1]$ , we discretize the continuous range into 1000 uniform bins $\{0.000, 0.001, \dots, 0.999\}$ and learn a corresponding VQ-VAE codebook [59] with 1000 entries, each represented as a 256-dimensional embedding vector. In each transformer block, the feed-forward layer receives the concatenation of the attention output and the embedding corresponding to the sampled $\epsilon$ value. +- The output predictions correspond to the observations at the time steps determined by the rules described in Sec. 3.3 (Eqs. 4 and 6). Still, the loss is applied separately to the reconstruction of the three types of observations. Specifically, L2 loss is used for the two vision observations, while L1 loss is applied to the proprioceptive state observations. The loss weight $\lambda_{\mathrm{mh}}$ in Eq. 8 is set to 1.0. + +Training Details We train the manipulation concept discovery process for 200,000 iterations with a batch size of 512. Each item in the batch is a segment of demonstration with a fixed length of $T_{\mathrm{context}} = 60$ . The training process uses the AdamW optimizer with a weight decay of 0.001 and momentum parameters $\beta_{1} = 0.9$ and $\beta_{2} = 0.95$ . The base learning rate is set to 0.001. Initially, the model is trained with a 100-iteration warmup phase, during which the learning rate increases linearly from 0.0001 to 0.001. After the warmup, the model is trained for the remaining iterations using a cosine decay schedule, gradually reducing the learning rate back to 0.0001. This training setup is compatible with GPUs such as the GeForce RTX 3090 or 4090. However, we leverage the A800 GPU for improved efficiency, completing the training process in 1.5 days. + +# A.2 Enhancing Imitation Learning + +![](images/3c9d6a70325153a69a22dc26c1c7903df39ad5a7ab716f4c800bfe03469534aa.jpg) + +![](images/4c4bc019561d625b041878a2b5da74130f00f91e342c395ea712592e0a705b5d.jpg) +Figure 6: Upper: Enhanced ACT (decoder part); lower: Enhanced Diffusion Policy + +This section introduces the neural network architectures and training details used in the Enhancing Imitation Learning process (Sec. 3.4). It focuses on modifying the original neural network policy to enable the prediction of manipulation concepts, thereby enhancing performance. Moreover, the implementation of the base policy follows [39]. + +ACT The pipeline (we focus on the CVAE decoder as it is the only modified component) is shown in the upper part of Fig. 6. Following [39], the transformer encoder in ACT's CVAE decoder is modified to incorporate task embeddings provided by CLIP. The transformer decoder in ACT's CVAE decoder is adapted to predict manipulation concepts. Specifically, the output of the $L$ -th layer in the transformer decoder is processed by an additional decoding head, which is nearly identical to the one used for outputting action chunks, with the only modification being the output dimension. This decoding head outputs manipulation concept chunks corresponding to the same time steps as the action chunks, with its parameters adjusted to match the dimensionality of the manipulation concepts at each time step (256). Other training and testing settings follow [39]. Moreover, the transformer decoder in ACT's CVAE decoder, as implemented by [39], consists of 7 layers. During our experiments, we tested various combinations of $L$ -th layers to determine the optimal layer for processing by the manipulation concept decoding head. Our results indicate that $L = 2$ provides slightly better performance than other configurations. We present the ablation study on $L$ and the weight $\lambda_{\mathrm{mc}}$ in Eq. 9 for ACT on LIBERO-90 tasks, as shown in Tab. 51. However, we believe this raises an interesting and challenging direction for future work: systematically investigating the rationale and insights behind the selection of $L$ , even beyond the context of our setting. + +Table 5: Ablation study on the intermediate layer outputs $(L)$ used as inputs to the manipulation concept decoder and the loss weight $\lambda_{\mathrm{mc}}$ in Eq. 9 for Enhancing Imitation Learning in ACT, evaluated on LIBERO-90 tasks. + +
ACTλmc=1.0λmc=0.1λmc=0.01λmc=0.001
L=274.8±0.870.6±0.869.0±0.168.7±0.5
L=370.0±0.469.9±0.268.8±1.068.7±0.6
L=472.6±0.569.9±0.369.6±0.267.3±0.5
+ +Diffusion Policy The pipeline is illustrated in the lower part of Fig. 6. Following [39], the convolution-based Diffusion Policy is modified to concatenate the noise level ( $k$ and corresponding $\epsilon_{k}$ ) embedding, observation, and task embedding as the conditional input to the diffusion model network, using the FiLM strategy. We further introduce an additional manipulation concept decoding up-sampling module, nearly identical to the one used for outputting action chunks, with the only modification being the output dimension, to decode intermediate outputs from the corresponding upsampling layer of the diffusion model. This decoding head can be configured to process intermediate outputs to predict manipulation concept chunks corresponding to the time steps of the predicted (noise of) action chunk outputs. The figure illustrates the cases for $L = 0$ and $L = 1$ . During our experiments, we tested various combinations of $L$ -th layers to identify the optimal layer for processing by the manipulation concept decoding head. Our results suggest that $L = 1$ achieves better performance than other configurations. We present the ablation study on $L$ and the weight $\lambda_{\mathrm{mc}}$ in Eq. 9 for Diffusion Policy on LIBERO-90 tasks, as shown in Tab. 6. Similar to ACT, we believe this topic needs further systematic study to uncover deeper insights. Other training and testing settings follow [39]. + +Table 6: Ablation study on the intermediate layer outputs $(L)$ used as inputs to the manipulation concept decoder and the loss weight $\lambda_{\mathrm{mc}}$ in Eq. 9 for Enhancing Imitation Learning in Diffusion Policy, evaluated on LIBERO-90 tasks. + +
DPλmc=1.0λmc=0.1λmc=0.01λmc=0.001
L=083.5±0.878.9±0.478.7±0.375.6±0.6
L=180.0±0.489.6±0.682.0±0.279.9±0.1
+ +Our future work includes a deeper study of modification strategies for various policies to adapt to the Enhancing Imitation Learning framework, following the methodology outlined in Sec. 3.4. + +# A.3 Manipulation concept discovery (Baselines) + +- InfoCon. Based on the design of InfoCon [31], All the size of hidden features output by transformers and concept features is 256. The state encoder (also process video clips consisting of concatenated, compressed vision observations and proprioceptive states, as outlined in Sec. A.1) uses a 12-layer transformer. The state reconstructor uses a 4-layer transformer. The goal-based policy uses a 4-layer transformer. The predictor for the generative goal uses a 4-layer transformer. For hyper-network used for discriminative goals, we use 2 hidden layers in the goal function. The number of concepts is fixed, the maximum number of 30 manipulation concepts for all the tasks. We employ the AdamW optimizer, coupled with a warm-up cosine annealing scheduler same as Sec. A.1. The weight decay is always $1.0 \times 10^{-3}$ . We use a batch size of 512 during training. We train our model for 200,000 iterations with a base learning rate of $1.0 \times 10^{-3}$ on a single A800 GPU within 1.5 days. +- XSkill. Following the design of XSkill [65], we implement its skill discovery framework on LIBERO-90, focusing exclusively on the "robot" embodiment and the Skill Discovery component from the XSkill pipeline. To ensure comparable model capacity and support multi-modality, our implementation employs a 12-layer Transformer as the temporal skill encoder. This encoder processes video clips consisting of concatenated, compressed vision observations and proprioceptive states, as outlined in Sec. A.1, along with a trainable token to predict skill representations, which are subsequently used for skill prototype prediction. To augment the concatenated video clips containing multi-modality information, Gaussian noise with $\sigma = 1.0 \times 10^{-3}$ is applied. This unified augmentation approach accommodates the nature of proprioceptive states, as standard image augmentation techniques are not directly suitable for robotic proprioception. The training process employs a batch size of 512 and a learning rate of $1.0 \times 10^{-3}$ for 200,000 iterations on a single A800 GPU within 1.5 days. +- DecisionNCE We fine-tune the DecisionNCE-T model (https://github.com/2toinf/DecisionNCE) on our dataset, as it outperforms DecisionNCE-P in our analysis of the experimental results in [27]. We use two types of language annotations: (1) the original task descriptions (Decision-task), and (2) detailed subtask labels derived by decomposing each task into meaningful subprocesses (Decision-motion). To construct the latter, we manually segment each demonstration based on changes in the robot's proprioceptive state (e.g., movement direction, gripper open/close status). Segments corresponding to the same task are then assigned unified subtask labels across demonstrations, with remaining inconsistencies resolved through manual adjustment. +- RPT. We modify the original RPT design [47] to adapt it for our task of discovering manipulation concept latents in the LIBERO-90 setting. We employ a 16-layer self-attention transformer to process inputs consisting of 60 consecutive, interleaved agent-view and eye-in-hand vision frames. Vision inputs are compressed using a stable diffusion VAE encoder, similar to the method in Sec. A.1. The total sequence length processed by the transformer is $60 \times 3 = 180$ . Each modality is mapped to a 256-dimensional embedding vector using an MLP, as defined in Eq. 11. The transformer's output is then decoded to reconstruct the original inputs using a 3-layer MLP with 1024-dimensional hidden layers. We follow the masking strategy outlined in [47] to perform temporal MAE training for the transformer. To label manipulation concept latents using the trained transformer, we extract the intermediate output of the 12th layer when the input consists of the full observation without masking. Notice that we select the output at the proprioceptive state input positions of the transformer to represent the manipulation concept latent at each time-step. The labeling process follows the procedure introduced in Sec. A.1. For training, we use a batch size of 512 and a learning rate of $1.0 \times 10^{-3}$ , running for 200,000 iterations on a single A800 GPU, which completes within 3 days. +- All. This is an ablation version of our manipulation concept discovery method, focusing on the design for capturing multi-modal correlations (Sec. 3.2). Specifically, this baseline replaces the loss in Eq. 3 with a loss that does not use partial masking but instead always masks all modalities: $\mathcal{L}_{\mathrm{all}}(t,\tau_i) = \left\| \mathcal{C}\left(\emptyset \big|z_i^t;\Theta_c\right) - \mathbf{o}_i^t\right\|$ . Based on our reasoning and the experiment results show in Tab. 3, we think this method may not be good at learning correlation between different modalities. Other settings follow Sec. A.1. + +- Next. This is an ablation version of our manipulation concept discovery method, focusing on the design for representing multi-horizon subgoals (Sec. 3.3). Specifically, this baseline replaces the loss in Eq. 7 with a loss that always predicts the next adjacent time-step observation: $\mathcal{L}_{\mathrm{next}}(t,\tau_i) = \left\| \mathcal{F}(\mathbf{o}_i^t,z_i^t,\Theta_f) - \mathbf{o}_i^{t + 1}\right\|$ . We observe that this setting is commonly used in recent works [7, 68], which learn representations based on adjacent time-step observations or observations separated by a fixed time horizon. We suggest that learning based on a fixed time horizon is conceptually similar to adjacent time-step settings, as the fixed time horizon can be interpreted as a unified time step. Our method differs by considering the temporal correlation across multiple variable horizons, which is also addressed by baseline methods like RPT. Other settings follow Sec. A.1. +- CLIP. To ensure compatibility with other baselines, which have an output dimension of 256, we select the ViT-B/32 CLIP model from the original source (https://github.com/openai/CLIP). This model outputs a 512-dimensional feature vector, the closest to 256 among the accessible CLIP models from this codebase when given an image. +- DINOv2. To match the output dimension of 256 used by other baselines, we select the dinov2-small DINOv2 model from the source at https://huggingface.co/facebook/dinov2-small1. This model produces a 384-dimensional feature vector when given an image. + +Note that DecisionNCE, CLIP, and DINOv2 baselines use only vision (and language) information for concept discovery. We preserve their original modality structure rather than adapting them to include proprioceptive states, as this would deviate from their pretraining foundations. + +# A.4 Mutual information estimation + +The estimation of mutual information is based on MINE [2], which uses batchwise samples drawn from a joint distribution and employs a neural network to estimate the mutual information. To extend this approach for estimating conditional mutual information (CMI), we reformulate CMI by decomposing it into mutual information terms, as shown below: + +$$ +\mathbb {I} (X: Y \mid Z) = \mathbb {I} (X: Y) + \mathbb {I} (X Y: Z) - \mathbb {I} (X: Z) - \mathbb {I} (Y: Z), \tag {13} +$$ + +where $XY$ denotes the random variable sampled from the joint distribution of $X$ and $Y$ and is represented as the concatenation of their encoded vectors. The neural network in MINE has two layers, with the hidden layer size set to 1.5 times of the dimensions of the two random variables. + +# B Pseudocode + +Here we provide pseudocode for (i) Deriving subprocess from manipulation concept latents (Alg. 1). (ii) Manipulation concept discovery training process of our method (Alg. 2). + +# C More Study on Learned Manipulation Concepts + +# C.1 Additional Experiments on Enhanced Imitation Learning + +Sampling Strategies In this part, we focus on methodology for deriving hierarchical structures from learned representations (Sec. 3.3). While we adopt a threshold-based hierarchy derivation method (Eq. 4) as a proof of concept, we acknowledge that alternative derivation methodologies warrant further investigation (see Sec. D). For the threshold-based approach, we employ uniform sampling of the threshold $\epsilon$ during training. This choice ensures full coverage of all possible hierarchical structures, as we do not know a priori which threshold values might be suboptimal. To validate this design choice, we conduct an ablation study comparing different sampling strategies for $\epsilon$ in Eq. 7: + +As shown in Tab. 7, uniform sampling currently achieves the best performance across both policy architectures. We hypothesize that while task-specific sampling strategies might excel on particular subsets, uniform sampling provides robust performance across diverse tasks due to its comprehensive coverage of the threshold space. Future work could explore adaptive sampling strategies tailored to specific task distributions. + +Table 7: Sampling Strategies Ablation. We compare different sampling strategies for $\epsilon$ in Eq. 7. Manipulation concepts are learned from LIBERO-90 and applied to policy learning on LIBERO-90. We report success rates (\%). + +
Sampling StrategyDescriptionACTDP
Uniform (Ours)ε ~ U(0,1)74.8±0.889.6±0.6
Sparseε ~ {0.1, 0.2, ···, 1.0}67.6±0.581.1±0.8
Biasedε ~ U(1/3, 2/3)65.6±0.778.7±0.4
+ +Learning Methodology Contribution We conduct an ablation study to isolate the contributions of our two core learning methodologies: Capturing Multi-Modal Correlations (Sec. 3.2) and Representing Multi-Horizon Sub-Goals (Sec. 3.3). Tab. 8 compares three configurations: (1) Cross-modal only: learning with only cross-modal alignment objectives in Eq. 3, (2) Multi-horizon only: learning with multi-horizon sub-goal prediction in Eq. 7 but without cross-modal alignment, and (3) Full method: combining both cross-modal alignment and multi-horizon prediction. + +Table 8: Methodology Contribution Ablation. We evaluate the contribution of each learning component by training manipulation concepts on LIBERO-90 and applying them to policy learning on LIBERO-90. We report success rates $(\%)$ . + +
MethodACTDP
Cross-modal only69.1±0.682.8±1.0
Multi-horizon only71.6±0.480.5±0.5
Ours (Full method)74.8±0.889.6±0.6
+ +The results in Tab. 8 reveal that both components make substantial and complementary contributions to performance. We attribute this synergy to the distinct roles of each component: cross-modal alignment grounds the understanding of correlations across different modalities, while multi-horizon prediction captures hierarchical temporal structure. Together, they enable the learning of manipulation concepts that are both correlationally coherent and temporally structured, leading to more robust policy learning. + +Data Constraint Experiments We evaluate whether manipulation concepts can help mitigate the challenges of imitation learning under limited data. Specifically, we vary the amount of data available for training both the manipulation concept encoder (Eq. 1) and the enhanced imitation learning framework (Sec. 3.4) to assess their impact on policy success rates. We conduct experiments on LIBERO-90 tasks using the diffusion policy. As shown in Tab. 9, incorporating manipulation concepts consistently improves policy performance compared to settings without them, even under restricted data conditions. This demonstrates that learning and leveraging manipulation concepts can make imitation learning more data-efficient and effective. + +Table 9: Performance under data constraints. Success rates of diffusion policies with and without manipulation concept enhancement, evaluated on LIBERO-90 (L90-90). In each setting, the number of demonstrations per task available for training both the manipulation concept encoder and the policy is limited as indicated. + +
50 demos/task25 demos/task10 demos/task
Ours89.6 ± 0.677.6 ± 0.561.2 ± 1.1
Plain75.1 ± 0.670.1 ± 0.359.1 ± 0.9
+ +Distance Metric We conduct an ablation study comparing spherical distance and cosine distance $\frac{1 - \cos(\cdot)}{2}$ for $\mathrm{dist}(\cdot, \cdot)$ in Eq. 4. Tab. 10 reports the performance when concepts are learned and applied + +to LIBERO-90 tasks. Further investigation into distance-threshold-based subprocess derivation methods represents a promising direction for future work. + +Table 10: Ablation study on distance metrics for concept learning on LIBERO-90. Spherical distance consistently outperforms cosine distance across both baseline methods. + +
Cosine DistanceSpherical Distance (Ours)
ACT67.8±0.574.8±0.8
DP82.0±0.489.6±0.6
+ +Sub-process Derivation We conduct an ablation study comparing two approaches for constraining manipulation concept latents within each sub-process in Eq. 4. Our proposed method enforces proximity among all concept latents throughout the sub-process ("Sequential Constraint"), while the baseline only constrains the distance between the initial and final concept latents ("Endpoint Constraint"). We evaluate both approaches on LIBERO-90, where concept discovery and policy enhancement are performed. Tab. 11 reports the task success rates when integrating the learned manipulation concepts with different policy architectures. + +Table 11: Ablation study on sub-process derivation constraints. We compare enforcing proximity among all manipulation concept latents within each sub-process (Sequential Constraint) versus constraining only the initial and final latents (Endpoint Constraint). Results show average success rates (\%) with standard errors across LIBERO-90 tasks. + +
Sequential ConstraintEndpoint Constraint
ACT74.8±0.868.4±0.8
DP89.6±0.679.8±0.5
+ +Future Prediction Strategy Apart from the different sub-goal determination strategies we compared (Next and InfoCon in Sec. 4.1), we evaluate two additional future prediction strategies. + +- Next-n. Unlike our sub-process derivation strategy (Eq. 4), this baseline encodes future observations at varying time horizons by randomly sampling a future timestep. Specifically: $\mathcal{L}_{\mathrm{next - n}}(t,\tau_i) = \mathbb{E}_{n\sim \mathrm{U}\{1,2,\dots ,T_i - t\}}\| \mathcal{F}(\mathbf{o}_i^t,z_i^t,n;\Theta_f) - \mathbf{o}_i^{t + n}\|$ . +- Next-random. This strategy builds upon Next-n but differs in how future targets are selected. We first randomly segment training demonstrations into sub-processes for concept discovery. Then, for a state at time-step $t$ , the prediction target is randomly selected from among the end-states of subsequent sub-processes. For example, if a demonstration is segmented into 5 sub-processes and time-step $t$ is in the 2nd sub-process, the model will randomly predict one of the end-states from the 2nd, 3rd, 4th, or 5th sub-processes during concept discovery learning. + +We evaluated diffusion policies enhanced by these strategies, with results presented in Tab. 12. The data demonstrates that our manipulation concepts yield better policy enhancement compared to the alternative strategies. This highlights the importance of carefully designing which future observations to predict and validates the effectiveness of our self-supervised sub-goal derivation and learning method. Specifically, the performance decrease observed with Next-n and Next-random, despite their consideration of multi-horizon futures, likely stems from the fact that not all future states effectively represent sub-goal completion. Intermediate movement states may be reached through multiple alternative trajectories that ultimately achieve the same sub-goal, thus providing limited information about the underlying task structure. + +Usage of Manipulation Concept Encoder We investigate two strategies for leveraging the manipulation concept encoder from Eq. 1 in downstream policy learning. The encoder serves as an intermediate module that extracts manipulation concept representations from demonstrations. We compare the following approaches: (1) Direct Conditioning: The trained encoder directly processes current observations to generate manipulation concepts, which are then concatenated with observations as additional input features to the policy network. (2) Joint Prediction (Ours): The policy + +Table 12: Comparison of Additional Future Prediction Strategies. Success rates of diffusion policies enhanced with manipulation concepts discovered using our method versus two alternative future prediction strategies on the LIBERO-90 benchmark. + +
L90-90OursNext-nNext-random
DP89.6 ± 0.683.0±0.382.8 ± 0.4
+ +network is trained to jointly predict both future actions and future manipulation concepts from current observations, as described in Sec. 3.4. Tab. 13 presents the comparative results across two policy architectures. + +Table 13: Comparison of Manipulation Concept Usage Strategies. + +
PolicyDirect ConditioningJoint Prediction (Ours)
ACT71.1±0.474.8±0.8
DP79.3±0.989.6±0.6
+ +The performance gap stems from a temporal alignment mismatch between concept representations and action predictions. In Direct Conditioning, the encoder extracts concepts from current or past observations, creating a temporal lag: the policy receives historical concept information when planning future actions. In contrast, Joint Prediction enforces temporal coherence by training the policy to predict future manipulation concepts alongside future actions, ensuring that the predicted concepts align temporally with the planned action sequence. + +This temporal alignment is critical in multi-phase manipulation tasks. For example, consider a pick-and-place scenario: immediately after grasping an object, the current observation encodes grasping-related dynamics. However, to execute the subsequent placement action, the policy requires placement-relevant information. Joint Prediction learns to anticipate these future task-phase concepts, providing the policy with forward-looking contextual information. Direct Conditioning, by contrast, conditions the policy on backward-looking grasping concepts that offer limited guidance for placement planning. + +While our results demonstrate the advantages of temporal alignment through joint prediction, we acknowledge that direct conditioning on historical concepts may benefit tasks requiring explicit long-horizon memory or reactive behaviors based on past states [14]. Future work will explore hybrid architectures that combine both strategies. + +# C.2 Alignment with Semantic Sub-Goals + +We evaluate whether the manipulation concept latents learned by our method resemble human-interpretable semantics. Specifically, we assess whether latents assigned to time steps of demonstrations (Sec. 3.1) exhibit higher pairwise similarity when those steps belong to sub-processes pursuing the same human-defined sub-goal. + +To analyze the learned representations, we first group manipulation concept latents according to human-annotated sub-goals. For instance, in the task "open the top drawer", latents from time steps where the robot reaches for the top drawer handle are categorized as "reach the top drawer". Latents from other demonstrations and tasks involving identical processes (reaching the top drawer) are placed in the same category. We then quantify the similarity between two categories by calculating the average cosine similarity between their respective latents, as defined in Eq. 10. + +Fig. 9 shows results from analyzing demonstrations from three tasks: + +- Task #1: Open the top drawer of the cabinet and put the bowl in it; +- Task #2: Close the bottom drawer of the cabinet and open the top drawer; +- Task #3: Close the top drawer of the cabinet and put the black bowl on top. + +We selected these tasks because they clearly demonstrate overlapping subgoals across different tasks (e.g., Task #1 and Task #2 both include "opening the top drawer"). This enables testing whether the + +latents capture similar subgoal semantics across different tasks—an essential capability for cross-task learning efficiency (Sec. 1). Manipulation concept latents are grouped based on human-defined sub-goals, with similarities between category pairs visualized as heatmaps. Three heatmaps are presented, each using a different granularity of sub-goal annotation: + +1. Top-1st heatmap: Omits task-specific distinctions, merging similar manipulation processes across tasks into the same category +2. Top-2nd heatmap: Further merges similar manipulation processes, disregarding distinctions like "top drawer" versus "bottom drawer" +3. Top-3rd heatmap: Consolidates manipulation processes further, treating actions like bowl transitions as the same concept regardless of context + +In each heatmap, the entry at position $(i,j)$ represents the average similarity $(\times 10.0)$ between categories $i$ and $j$ . For readability, only the top three similarity values in each row are displayed. + +We emphasize that testing semantic capture at different "description granularity levels" is important because semantics naturally exist at multiple levels of abstraction, from highly specific details to broadly generalizable patterns. Finer-grained descriptions provide more precise details but limited generalization, while coarser-grained descriptions capture more general features applicable across diverse scenarios. For example, the general instruction "close the drawer" applies broadly to subprocesses in both Task #2 and Task #3, whereas the more specific "close the top drawer" incorporates spatial features that make it applicable in Task #3 but not in Task #2. Through this multi-granularity analysis, we evaluate whether our manipulation concept latents successfully capture both fine-grained semantics needed for specific scenarios and coarse-grained semantics that enable transfer across more scenarios. + +What we observe is that the highest similarity values consistently appear along the diagonal in each heatmap in Fig. 9, so concept latents from the same category show higher similarity compared with different categories. This indicates that the learned latent clusters resemble clusters derived from human-interpretable sub-goal classifications, suggesting that our model captures meaningful semantic structure in the manipulation processes. Moreover, the patterns observed across the three heatmaps with different description granularities reveal that the latents encode semantics at multiple levels of abstraction. They capture both generalizable semantics applicable across tasks and scenes, while simultaneously preserving fine-grained scene-specific details. + +Furthermore, Fig. 10(b) provides a t-SNE visualization of manipulation concept latents from all 90 tasks in LIBERO-90. For each task, latents $(z_{i}^{t})$ were extracted at every time step of demonstrations. In the plot, latents are color-coded by their originating tasks. We observe that clusters often contain latents from diverse tasks, as indicated by the mixed colors in each cluster. This further supports our finding that the learned latents generalize across tasks and capture shared semantic structures.2 + +# C.3 Motion Study + +We evaluate whether the learned manipulation concept latents capture the robot's motion. Using Eq. 10, we calculate the average similarity $(\times 100.0)$ between movements based on manipulation concept latents corresponding to specific gripper actions. Specifically, we collect latents for the following movements from task demonstrations in LIBERO-90: + +1. Forward-backward motion: Latents for time-steps where the robot moves forward, backward, or remains still along the forward-backward axis. +2. Left-right motion: Latents for time-steps where the robot moves left, right, or remains still along the left-right axis. +3. Up-down motion: Latents for time-steps where the robot moves up, down, or remains still along the up-down axis. +4. Gripper state: Latents for time-steps where the gripper opens or closes. + +Movements with velocities below $20\%$ of the maximum observed velocity are classified as "still". Using these collected latents, we generate heatmaps (similar to Fig. 9) to visualize the average cosine similarity across different movement directions and gripper states (Fig. 10(a)). + +The heatmaps reveal that the highest cosine similarity values often appear along the diagonal. This demonstrates that latents corresponding to the same motion patterns exhibit greater similarity to each other than to those from different motion patterns, indicating that the latents effectively capture different movement directions and gripper states. However, we observe that forward-backward motion is captured with lower accuracy compared to other dimensions. We hypothesize that incorporating additional 3D-informative modalities, such as depth maps, beyond the current proprioceptive states could improve the representation of motion along the forward-backward axis. We leave the exploration of such modality incorporation to future work. + +# C.4 Diversity & Discrimination Study + +We analyze the diversity and discriminability of learned manipulation concepts by comparing concept latents from our method (Sec. 3) and the baselines in Manipulation Concept Discovery Baselines (Sec. 4.1). Specifically, we cluster latents from these methods and examine the number of clusters under varying granularities. The number of clusters reflects concept diversity: more clusters indicate a wider variety of concepts. Clustering granularity determines whether clusters are fine-grained (fine granularity) or general (coarse granularity). Additionally, small granularity perturbations test discriminability, as less discriminative latents lead to significant clustering changes under small granularity variations. For each method, We collect manipulation concept latents from 90 LIBERO-90 tasks (one demonstration per task) and use DBSCAN to cluster them while varying the density parameter Eps, which controls clustering granularity. Fig. 11 shows the cluster counts across different Eps values. From Fig. 11, our manipulation concept discovery method (Ours) demonstrates two key advantages: 1) At higher granularities $(\mathrm{Eps} > 0.2)$ , Ours maintains a higher number of clusters. 2) The decline in cluster count is relatively smooth and gradual, showing stability under small Eps changes. These results highlight the superior diversity and discriminability of our manipulation concept discovery method. + +# C.5 Multi-Level Hierarchical Structure + +In Fig. 3, we present a visualization example of the Multi-Level Hierarchical Structure described in Sec. 4.3. Additional visualization results are available in the supplementary materials under the directory supplementary/vis-multi_h. + +# C.6 Real Robot Experiments Details + +Training Data. As shown in Fig. 5, the training data for the "cleaning cup" task consists of demonstrations using mobile ALOHA [16] to place the cup from the table into the container. Each demonstration features a scene containing exactly one cup and one container. There are two pairings of color combinations: blue cups with green containers and yellow cups with pink containers. For each pairing, we collect 27 demonstrations with varied spatial arrangements. + +Evaluation Setting. For evaluation, we test our model on six scenarios that introduce variations absent from the training data: + +- Novel Placements. Objects maintain the same color pairings as in training but appear in previously unseen spatial arrangements. +- Color Composition. We rearrange color pairings (blue cups with pink containers and yellow cups with green containers) to test generalization across color combinations. +- Novel Objects. We introduce unseen objects, such as bamboo-woven containers, pink cups not present in training, or cups initially placed on plates rather than directly on the table. +- Obstacles. We position obstacles in front of cups to challenge visual perception. +- Barriers. We place a plate inside the container, requiring the robot to lift the cup high enough to clear this barrier when depositing it. +- Grasping Together. We position two cups adjacent to one another, requiring the robot to grasp both simultaneously at their contact point and deposit them together in the container. + +![](images/40650da8d648beb10efb28dfcf43c748fd75b12d7bf7a08f77c16e6eb74fa249.jpg) +Figure 7: Multi-horizon goal prediction with learned manipulation concepts. Visualization of future states predicted by our Multi-Horizon Goal Predictor (MHGP, Eq. 7) when conditioned on the current observation, a manipulation concept latent $(z)$ , and varying coherence thresholds $(\epsilon)$ . From left to right, as $\epsilon$ increases from 0 to 1, predictions extend progressively further into the future, demonstrating how our manipulation concepts encode temporal abstraction at multiple horizons. Note that predictions capture essential functional relationships (robot-object interactions) rather than pixel-perfect reconstructions, facilitating generalization across environments. + +Manipulation Concept Discovery. The model architecture and hyperparameter configuration for manipulation concept discovery follow the methodology described in Sec. A.1. Since the dataset is relatively small, we adapt smaller transformers: a 4-layer concept encoder $(\mathcal{E},$ Eq. 1), a 4-layer Cross-Modal Correlation Network (C, Eq. 3), and a 4-layer Multi-Horizon Future Predictor $(\mathcal{F},$ Eq. 7). For data collected using mobile ALOHA [16], we incorporate the following modalities: three $640\times 480$ resolution cameras (left-gripper, right-gripper, and upper-gripper) and 42-dimensional proprioception states (comprising 14-dimensional joint torque, position, and velocity measurements). All image data undergoes preprocessing as detailed in Sec. A.1. + +Enhancing Imitation Learning. Please refer to ACT section in Sec. A.2. + +# C.7 Multi-Horizon Goal Prediction Visualization + +We provide visualization results of the Multi-Horizon Goal Prediction Visualization (Sec. 4.4) in Fig. 7 and supplementary materials under the directory supplementary/prediction. Below are the details of the experiments: + +Dataset. For our experiments, we utilized the BridgeDataV2 dataset [60]. Since multi-view data is not universally available across all demonstrations, we selected two specific modalities: the robot's proprioceptive states (7DoF) and the third-person camera view. The camera images were preprocessed to $128 \times 128$ resolution following the procedure outlined in Sec. A.2. + +Manipulation Concept Discovery. We implemented the model architecture and hyperparameter configuration as detailed in Sec. A.1, adapting it specifically to operate with the two modalities described in the Dataset section above. + +# C.8 Preliminary VLA Integration + +We present a preliminary exploration of integrating manipulation concepts with vision-language-action models (VLAs). We build upon OpenVLA-OFT [23], which fine-tunes OpenVLA using pretrained parameters and a novel action adapter for downstream tasks. The action adapter processes hidden layer features from the original pretrained VLA model. Following this architecture, we introduce an additional "concept adapter" that implements the method described in Sec. 3.4, enabling the integration of manipulation concepts into the VLA. + +To evaluate the data efficiency gains from manipulation concepts, we fine-tune the enhanced VLA on $50\%$ of the training data used for LIBERO-10 tasks in the original OpenVLA-OFT study [23]. We compare fine-tuning performance with and without manipulation concept integration. Fig. 8 presents the results, where the x-axis indicates training epochs and the y-axis shows success rates for checkpoints at each epoch. The solid lines labeled "best" represent the highest success rate achieved up to that epoch. + +The results demonstrate that manipulation concepts improve data utilization. With only half the training data, the concept-enhanced approach consistently achieves higher success rates throughout training. Notably, the original OpenVLA-OFT achieved $94.5\%$ success with the full dataset [23], while our concept-enhanced model with half the data reaches comparable performance levels, indicating substantially improved data efficiency. + +![](images/ac585525e463ec12ccc3801e795c7339d63dbcc9fab45d204d594b4cc9b2676b.jpg) +Figure 8: Data efficiency comparison on LIBERO-10 tasks with $50\%$ training data. Solid lines show best performance up to each epoch for models with and without manipulation concepts. + +We hypothesize that this improvement stems from HiMaCon's ability to capture manipulation dynamics at multiple abstraction levels. The learned concepts provide explicit intermediate representations that bridge high-level task instructions and low-level control actions, thereby reducing the learning burden on VLAs by supplying structured manipulation knowledge rather than requiring learning of complex sensorimotor patterns from scratch. Further investigation of this integration will be pursued in future work. + +# D Limitations & Future works + +Further Exploration of multi-modality. We propose enhancing robotic data collection with richer modalities and studying how these modalities can derive more effective manipulation concepts. While current robotics research primarily focuses on visual information, human manipulation relies on multiple sensory inputs, particularly tactile feedback to complement vision. This is especially crucial for robotic systems with limited tactile capabilities. Future work should investigate which modalities contribute most significantly to performance improvements and how to fully leverage their potential. + +Further Exploration of multi-horizon sub-goal. Our work proposes methods to derive sub-processes for achieving sub-goals across multiple horizons, though several improvements remain possible. Current methods inadequately capture relationships between different values of $\epsilon$ in Eq. 4, failing to reflect the natural tree structure of hierarchical sub-goals. Future research could explicitly derive tree structures [61, 74] where long-horizon sub-goals serve as parent nodes to short-horizon child nodes. Additionally, our cosine similarity approach for determining sub-goal correspondence could be refined with more sophisticated metrics. + +Scaling up. Computational constraints have limited our exploration of how our method scales with larger datasets. We plan to leverage pretrained multi-modal foundation models, adopting structures inspired by [7] and extending pretraining beyond robotics data as in [68]. We also aim to further investigate whether our manipulation concepts can enhance advanced policies like Vision-Language-Action models [4, 19, 23, 24]. + +Algorithm 1 Derive Subprocess $\mathrm{h}(\mathbf{z}_i;\epsilon)$ +Input: manipulation concept vectors $\mathbf{z}_i = \{z_i^t\}_{t=1}^{T_i}$ , coherence parameter $\epsilon \in [0,1]$ . +Initialize: $End = [], g_b = 1$ +while $g_b \leq T_i$ do + $g_e = g_b + 1$ +while true do +if $\exists u \in [g_b, g_e)$ , s.t. $\mathrm{dist}(z_i^u, z_i^{g_e}) \geq \epsilon$ or $g_e > T_i$ then +break +end if + $g_e = g_e + 1$ +end while +End.append([g_b, g_e]) + $g_b = g_e$ +end while +Return End + +Algorithm 2 Manipulation Concept Discovery Training (one demonstration per batch) +Input: demonstrations $\tau_{i}\in D$ , where $\tau_{i} = \{(o_{i}^{1,t},o_{i}^{2,t},\dots,o_{i}^{M,t},a_{i}^{t})\}_{t = 1}^{T_{i}}$ +Initialize: Manipulation concept assignment encoder $\mathcal{E}(\cdot ;\Theta_{\mathcal{E}})$ +Initialize: Modality Correlation Learner $\mathcal{C}(\cdot ;\Theta_c)$ , Subgoal Learner $\mathcal{F}(\cdot ;\Theta_f)$ +while true do +for $\tau_{i}$ in $D$ do + $(z_i^1,\dots ,z_i^{T_i})\gets \mathcal{E}\left((o_i^{1,1},\dots,o_i^{M,1}),(o_i^{1,2},\dots,o_i^{M,2}),\dots ,(o_i^{1,T_i},\dots,o_i^{M,T_i});\Theta_\mathcal{E}\right)$ +while True do +Randomly generate a tuple $(m_1,m_2,\ldots ,m_M)$ , where $m_i\in \{0,1\}$ +if $\sum_{i = 1}^{M}m_{i} < M$ then break +end if +end while + $(\hat{o}_i^{1,t},\dots ,\hat{o}_i^{M,t})_{t = 1}^{T_i}\gets \mathcal{C}\left((o_i^{1,t}\cdot m_1,o_i^{2,t}\cdot m_2,\dots,o_i^{M,t}\cdot m_M,z_i^t)_{t = 1}^{T_i};\Theta_c\right)$ $\mathcal{L}_{\mathrm{mm}} = \sum_{t = 1}^{T_i}\sum_{m = 1}^{M}\| \hat{o}_i^{m,t} - o_i^{m,t}\|$ $\epsilon \sim \mathrm{U}([0,1])$ +End $= \mathrm{h}(z_i^1,\dots ,z_i^{T_i};\epsilon)$ {Alg. 1} +for $t = 1$ to $T_{i}$ do + ${\bf g}_t = \min \left(\left\{g_e\mid [g_b,g_e)\in End,g_e > t\right\} \cup \{T_i\}\right)$ +end for + $(\overline{o}_i^{1,t},\dots ,\overline{o}_i^{M,t})_{t = 1}^{T_i}\gets \mathcal{F}\left((o_i^{1,t},o_i^{2,t},\dots,o_i^{M,t},z_i^t,\epsilon)_{t = 1}^{T_i};\Theta_f\right)$ $\mathcal{L}_{\mathrm{mh}} = \sum_{t = 1}^{T_i}\sum_{m = 1}^{M}\left\| \overline{o}_i^{m,t} - o_i^{m,{\bf g}_t}\right\|$ +end for +end while + +![](images/64aa819b166a30c81725107f28d6bf8e5dd436686da32829fcd9c9ffaf4f3bc9.jpg) + +![](images/f2bfa1d2a3afaa73ef7e01709a4f99095c80173d01372f573d80afcb87ac9deb.jpg) + +![](images/0a227f5b1ece588af90d14e16c8bac68b5b0aee6975dc46a2eaeaa2d5539e3c9.jpg) +Figure 9: Average cosine similarity between pairs of sub-goal categories (defined by human semantics) computed using manipulation concept latents learned by our method (Sec.3). In each heatmap, the value at the $i$ -th row and $j$ -th column represents the average cosine similarity between latent vectors from the $i$ -th and $j$ -th categories. Three levels of labeling are provided across the heatmaps; please refer to Sec. C.2 for details. + +![](images/e752e259abaeadc92314dafffafaf2cf7de5dfaa45ce61e06ed548b4e78f5ea0.jpg) + +![](images/94ce6a57c7eebc55b37be29b7c7969125032bc859ea115579e9d4efeac82a58c.jpg) + +Figure 10 +![](images/7d3d8f5d565385235989fac8566d1aa387351d6556b7f9eae4a3458fd7d492f9.jpg) +(a) Average cosine similarity between pairs of movement categories (defined by human semantics) computed using manipulation concept latents learned by our method (Sec.3). +(b) t-SNE Clustering of Manipulation Concept Latents corresponding to tasks. We perform t-SNE clustering on the manipulation concepts at each time step. These concepts are generated by our method (Sec. 3). Each sample is colored according to its task, representing one of 90 possible tasks as indicated by the colorbar. + +![](images/50191299bab3ca54963ee3f47f031cded4cdcbb14f7a69bf6333bcb62a9924af.jpg) +Figure 11: DBSCAN Clustering Analysis of Manipulation Concept Latents' Diversity and Discrimination. Clustering is performed on manipulation concept latents generated by our method and the baseline methods described in Manipulation Concept Discovery Baselines (Sec. 4.1), across 90 tasks from the LIBERO-90 dataset. The figure shows the $(\log)$ number of clusters obtained using DBSCAN for clustering density $\epsilon \in [0,1]$ , with no points classified as noise. + +# NeurIPS Paper Checklist + +The checklist is designed to encourage best practices for responsible machine learning research, addressing issues of reproducibility, transparency, research ethics, and societal impact. Do not remove the checklist: The papers not including the checklist will be desk rejected. The checklist should follow the references and follow the (optional) supplemental material. The checklist does NOT count towards the page limit. + +Please read the checklist guidelines carefully for information on how to answer these questions. For each question in the checklist: + +- You should answer [Yes], [No], or [NA]. +- [NA] means either that the question is Not Applicable for that particular paper or the relevant information is Not Available. +- Please provide a short (1–2 sentence) justification right after your answer (even for NA). + +The checklist answers are an integral part of your paper submission. They are visible to the reviewers, area chairs, senior area chairs, and ethics reviewers. You will be asked to also include it (after eventual revisions) with the final version of your paper, and its final version will be published with the paper. + +The reviewers of your paper will be asked to use the checklist as one of the factors in their evaluation. While "[Yes]" is generally preferable to "[No]", it is perfectly acceptable to answer "[No]" provided a proper justification is given (e.g., "error bars are not reported because it would be too computationally expensive" or "we were unable to find the license for the dataset we used"). In general, answering "[No]" or "[NA]" is not grounds for rejection. While the questions are phrased in a binary way, we acknowledge that the true answer is often more nuanced, so please just use your best judgment and write a justification to elaborate. All supporting evidence can appear either in the main paper or the supplemental material, provided in appendix. If you answer [Yes] to a question, in the justification please point to the section(s) where related material for the question can be found. + +IMPORTANT, please: + +- Delete this instruction block, but keep the section heading "NeurIPS Paper Checklist", +- Keep the checklist subsection headings, questions/answers and guidelines below. +- Do not modify the questions and only use the provided macros for your answers. + +# 1. Claims + +Question: Do the main claims made in the abstract and introduction accurately reflect the paper's contributions and scope? + +Answer: [Yes] + +Justification: The abstract gives a summary of our contribution on self-supervised learning of manipulation concepts. + +Guidelines: + +- The answer NA means that the abstract and introduction do not include the claims made in the paper. +- The abstract and/or introduction should clearly state the claims made, including the contributions made in the paper and important assumptions and limitations. A No or NA answer to this question will not be perceived well by the reviewers. +- The claims made should match theoretical and experimental results, and reflect how much the results can be expected to generalize to other settings. +- It is fine to include aspirational goals as motivation as long as it is clear that these goals are not attained by the paper. + +# 2. Limitations + +Question: Does the paper discuss the limitations of the work performed by the authors? + +Answer: [Yes] + +Justification: We discuss limitations including improvements to hierarchy derivation, further work on scaling up, and modality balance. + +# Guidelines: + +- The answer NA means that the paper has no limitation while the answer No means that the paper has limitations, but those are not discussed in the paper. +- The authors are encouraged to create a separate "Limitations" section in their paper. +- The paper should point out any strong assumptions and how robust the results are to violations of these assumptions (e.g., independence assumptions, noiseless settings, model well-specification, asymptotic approximations only holding locally). The authors should reflect on how these assumptions might be violated in practice and what the implications would be. +- The authors should reflect on the scope of the claims made, e.g., if the approach was only tested on a few datasets or with a few runs. In general, empirical results often depend on implicit assumptions, which should be articulated. +- The authors should reflect on the factors that influence the performance of the approach. For example, a facial recognition algorithm may perform poorly when image resolution is low or images are taken in low lighting. Or a speech-to-text system might not be used reliably to provide closed captions for online lectures because it fails to handle technical jargon. +- The authors should discuss the computational efficiency of the proposed algorithms and how they scale with dataset size. +- If applicable, the authors should discuss possible limitations of their approach to address problems of privacy and fairness. +- While the authors might fear that complete honesty about limitations might be used by reviewers as grounds for rejection, a worse outcome might be that reviewers discover limitations that aren't acknowledged in the paper. The authors should use their best judgment and recognize that individual actions in favor of transparency play an important role in developing norms that preserve the integrity of the community. Reviewers will be specifically instructed to not penalize honesty concerning limitations. + +# 3. Theory assumptions and proofs + +Question: For each theoretical result, does the paper provide the full set of assumptions and a complete (and correct) proof? + +Answer: [NA] + +Justification: We mainly make use of established theoretical frameworks (such as mutual information) for clarification and modeling of our method. + +# Guidelines: + +- The answer NA means that the paper does not include theoretical results. +- All the theorems, formulas, and proofs in the paper should be numbered and cross-referenced. +- All assumptions should be clearly stated or referenced in the statement of any theorems. +- The proofs can either appear in the main paper or the supplemental material, but if they appear in the supplemental material, the authors are encouraged to provide a short proof sketch to provide intuition. +- Inversely, any informal proof provided in the core of the paper should be complemented by formal proofs provided in appendix or supplemental material. +- Theorems and Lemmas that the proof relies upon should be properly referenced. + +# 4. Experimental result reproducibility + +Question: Does the paper fully disclose all the information needed to reproduce the main experimental results of the paper to the extent that it affects the main claims and/or conclusions of the paper (regardless of whether the code and data are provided or not)? + +Answer: [Yes] + +Justification: We provide details in the appendix and supplementary materials. + +Guidelines: + +- The answer NA means that the paper does not include experiments. + +- If the paper includes experiments, a No answer to this question will not be perceived well by the reviewers: Making the paper reproducible is important, regardless of whether the code and data are provided or not. + +- If the contribution is a dataset and/or model, the authors should describe the steps taken to make their results reproducible or verifiable. + +- Depending on the contribution, reproducibility can be accomplished in various ways. For example, if the contribution is a novel architecture, describing the architecture fully might suffice, or if the contribution is a specific model and empirical evaluation, it may be necessary to either make it possible for others to replicate the model with the same dataset, or provide access to the model. In general, releasing code and data is often one good way to accomplish this, but reproducibility can also be provided via detailed instructions for how to replicate the results, access to a hosted model (e.g., in the case of a large language model), releasing of a model checkpoint, or other means that are appropriate to the research performed. + +- While NeurIPS does not require releasing code, the conference does require all submissions to provide some reasonable avenue for reproducibility, which may depend on the nature of the contribution. For example + +(a) If the contribution is primarily a new algorithm, the paper should make it clear how to reproduce that algorithm. +(b) If the contribution is primarily a new model architecture, the paper should describe the architecture clearly and fully. +(c) If the contribution is a new model (e.g., a large language model), then there should either be a way to access this model for reproducing the results or a way to reproduce the model (e.g., with an open-source dataset or instructions for how to construct the dataset). +(d) We recognize that reproducibility may be tricky in some cases, in which case authors are welcome to describe the particular way they provide for reproducibility. In the case of closed-source models, it may be that access to the model is limited in some way (e.g., to registered users), but it should be possible for other researchers to have some path to reproducing or verifying the results. + +# 5. Open access to data and code + +Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material? + +Answer: [No] + +Justification: We will open source all code and newly-created datasets upon acceptance. + +Guidelines: + +- The answer NA means that paper does not include experiments requiring code. +- Please see the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details. +- While we encourage the release of code and data, we understand that this might not be possible, so "No" is an acceptable answer. Papers cannot be rejected simply for not including code, unless this is central to the contribution (e.g., for a new open-source benchmark). +- The instructions should contain the exact command and environment needed to run to reproduce the results. See the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details. +- The authors should provide instructions on data access and preparation, including how to access the raw data, preprocessed data, intermediate data, and generated data, etc. +- The authors should provide scripts to reproduce all experimental results for the new proposed method and baselines. If only a subset of experiments are reproducible, they should state which ones are omitted from the script and why. +- At submission time, to preserve anonymity, the authors should release anonymized versions (if applicable). + +- Providing as much information as possible in supplemental material (appended to the paper) is recommended, but including URLs to data and code is permitted. + +# 6. Experimental setting/details + +Question: Does the paper specify all the training and test details (e.g., data splits, hyperparameters, how they were chosen, type of optimizer, etc.) necessary to understand the results? + +Answer: [Yes] + +Justification: Details are specified in the Experiments section and in the appendix and supplementary materials. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The experimental setting should be presented in the core of the paper to a level of detail that is necessary to appreciate the results and make sense of them. +- The full details can be provided either with the code, in appendix, or as supplemental material. + +# 7. Experiment statistical significance + +Question: Does the paper report error bars suitably and correctly defined or other appropriate information about the statistical significance of the experiments? + +Answer: [No] + +Justification: For the policy success rates we currently include standard deviation. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The authors should answer "Yes" if the results are accompanied by error bars, confidence intervals, or statistical significance tests, at least for the experiments that support the main claims of the paper. +- The factors of variability that the error bars are capturing should be clearly stated (for example, train/test split, initialization, random drawing of some parameter, or overall run with given experimental conditions). +- The method for calculating the error bars should be explained (closed form formula, call to a library function, bootstrap, etc.) +- The assumptions made should be given (e.g., Normally distributed errors). +- It should be clear whether the error bar is the standard deviation or the standard error of the mean. +- It is OK to report 1-sigma error bars, but one should state it. The authors should preferably report a 2-sigma error bar than state that they have a $96\%$ CI, if the hypothesis of Normality of errors is not verified. +- For asymmetric distributions, the authors should be careful not to show in tables or figures symmetric error bars that would yield results that are out of range (e.g. negative error rates). +- If error bars are reported in tables or plots, The authors should explain in the text how they were calculated and reference the corresponding figures or tables in the text. + +# 8. Experiments compute resources + +Question: For each experiment, does the paper provide sufficient information on the computer resources (type of compute workers, memory, time of execution) needed to reproduce the experiments? + +Answer: [Yes] + +Justification: Please refer to the details provided in the appendix. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The paper should indicate the type of compute workers CPU or GPU, internal cluster, or cloud provider, including relevant memory and storage. + +- The paper should provide the amount of compute required for each of the individual experimental runs as well as estimate the total compute. +- The paper should disclose whether the full research project required more computer than the experiments reported in the paper (e.g., preliminary or failed experiments that didn't make it into the paper). + +# 9. Code of ethics + +Question: Does the research conducted in the paper conform, in every respect, with the NeurIPS Code of Ethics https://neurips.cc/public/EthicsGuidelines? + +Answer: [Yes] + +Justification: Current experiments and topics do not conflict with the Code of Ethics. + +Guidelines: + +- The answer NA means that the authors have not reviewed the NeurIPS Code of Ethics. +- If the authors answer No, they should explain the special circumstances that require a deviation from the Code of Ethics. +- The authors should make sure to preserve anonymity (e.g., if there is a special consideration due to laws or regulations in their jurisdiction). + +# 10. Broader impacts + +Question: Does the paper discuss both potential positive societal impacts and negative societal impacts of the work performed? + +Answer: [NA] + +Justification: Currently, the experiments are carried out in simulations and on robots in laboratories. + +Guidelines: + +- The answer NA means that there is no societal impact of the work performed. +- If the authors answer NA or No, they should explain why their work has no societal impact or why the paper does not address societal impact. +- Examples of negative societal impacts include potential malicious or unintended uses (e.g., disinformation, generating fake profiles, surveillance), fairness considerations (e.g., deployment of technologies that could make decisions that unfairly impact specific groups), privacy considerations, and security considerations. +- The conference expects that many papers will be foundational research and not tied to particular applications, let alone deployments. However, if there is a direct path to any negative applications, the authors should point it out. For example, it is legitimate to point out that an improvement in the quality of generative models could be used to generate deepfakes for disinformation. On the other hand, it is not needed to point out that a generic algorithm for optimizing neural networks could enable people to train models that generate Deepfakes faster. +- The authors should consider possible harms that could arise when the technology is being used as intended and functioning correctly, harms that could arise when the technology is being used as intended but gives incorrect results, and harms following from (intentional or unintentional) misuse of the technology. +- If there are negative societal impacts, the authors could also discuss possible mitigation strategies (e.g., gated release of models, providing defenses in addition to attacks, mechanisms for monitoring misuse, mechanisms to monitor how a system learns from feedback over time, improving the efficiency and accessibility of ML). + +# 11. Safeguards + +Question: Does the paper describe safeguards that have been put in place for responsible release of data or models that have a high risk for misuse (e.g., pretrained language models, image generators, or scraped datasets)? + +Answer: [NA] + +Justification: Currently, we have not encountered any safeguard issues. + +Guidelines: + +- The answer NA means that the paper poses no such risks. +- Released models that have a high risk for misuse or dual-use should be released with necessary safeguards to allow for controlled use of the model, for example by requiring that users adhere to usage guidelines or restrictions to access the model or implementing safety filters. +- Datasets that have been scraped from the Internet could pose safety risks. The authors should describe how they avoided releasing unsafe images. +- We recognize that providing effective safeguards is challenging, and many papers do not require this, but we encourage authors to take this into account and make a best faith effort. + +# 12. Licenses for existing assets + +Question: Are the creators or original owners of assets (e.g., code, data, models), used in the paper, properly credited and are the license and terms of use explicitly mentioned and properly respected? + +Answer: [Yes] + +Justification: We have checked the sources we used. + +# Guidelines: + +- The answer NA means that the paper does not use existing assets. +- The authors should cite the original paper that produced the code package or dataset. +- The authors should state which version of the asset is used and, if possible, include a URL. +- The name of the license (e.g., CC-BY 4.0) should be included for each asset. +- For scraped data from a particular source (e.g., website), the copyright and terms of service of that source should be provided. +- If assets are released, the license, copyright information, and terms of use in the package should be provided. For popular datasets, paperswithcode.com/datasets has curated licenses for some datasets. Their licensing guide can help determine the license of a dataset. +- For existing datasets that are re-packaged, both the original license and the license of the derived asset (if it has changed) should be provided. +- If this information is not available online, the authors are encouraged to reach out to the asset's creators. + +# 13. New assets + +Question: Are new assets introduced in the paper well documented and is the documentation provided alongside the assets? + +Answer: [NA] + +Justification: We will release all new assets we created (code/models/datasets) upon acceptance. + +# Guidelines: + +- The answer NA means that the paper does not release new assets. +- Researchers should communicate the details of the dataset/code/model as part of their submissions via structured templates. This includes details about training, license, limitations, etc. +- The paper should discuss whether and how consent was obtained from people whose asset is used. +- At submission time, remember to anonymize your assets (if applicable). You can either create an anonymized URL or include an anonymized zip file. + +# 14. Crowdsourcing and research with human subjects + +Question: For crowdsourcing experiments and research with human subjects, does the paper include the full text of instructions given to participants and screenshots, if applicable, as well as details about compensation (if any)? + +Answer: [NA] + +Justification: We currently do not have crowdsourcing experiments. + +# Guidelines: + +- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. +- Including this information in the supplemental material is fine, but if the main contribution of the paper involves human subjects, then as much detail as possible should be included in the main paper. +- According to the NeurIPS Code of Ethics, workers involved in data collection, curation, or other labor should be paid at least the minimum wage in the country of the data collector. + +# 15. Institutional review board (IRB) approvals or equivalent for research with human subjects + +Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals (or an equivalent approval/review based on the requirements of your country or institution) were obtained? + +Answer: [NA] + +Justification: We currently do not involve crowdsourcing nor research with human subjects. + +# Guidelines: + +- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. +- Depending on the country in which research is conducted, IRB approval (or equivalent) may be required for any human subjects research. If you obtained IRB approval, you should clearly state this in the paper. +- We recognize that the procedures for this may vary significantly between institutions and locations, and we expect authors to adhere to the NeurIPS Code of Ethics and the guidelines for their institution. +- For initial submissions, do not include any information that would break anonymity (if applicable), such as the institution conducting the review. + +# 16. Declaration of LLM usage + +Question: Does the paper describe the usage of LLMs if it is an important, original, or non-standard component of the core methods in this research? Note that if the LLM is used only for writing, editing, or formatting purposes and does not impact the core methodology, scientific rigorousness, or originality of the research, declaration is not required. + +Answer: [NA]. + +Justification: The core method development in this research does not involve LLMs. + +# Guidelines: + +- The answer NA means that the core method development in this research does not involve LLMs as any important, original, or non-standard components. +- Please refer to our LLM policy (https://neurips.cc/Conferences/2025/LLM) for what should or should not be described. \ No newline at end of file diff --git a/NeurIPS/2025/$_textit{HiMaCon_}$ Discovering Hierarchical Manipulation Concepts from Unlabeled Multi-Modal Data/images.zip b/NeurIPS/2025/$_textit{HiMaCon_}$ Discovering Hierarchical Manipulation Concepts from Unlabeled Multi-Modal Data/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..e4ac943d93f52351409cb547d858c3f8e5ce3faf --- /dev/null +++ b/NeurIPS/2025/$_textit{HiMaCon_}$ Discovering Hierarchical Manipulation Concepts from Unlabeled Multi-Modal Data/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7c7b575e990e300d3c68ef5c38193a3d63d0710e072c0ad9caca4b0258174917 +size 1159697 diff --git a/NeurIPS/2025/$_textit{HiMaCon_}$ Discovering Hierarchical Manipulation Concepts from Unlabeled Multi-Modal Data/layout.json b/NeurIPS/2025/$_textit{HiMaCon_}$ Discovering Hierarchical Manipulation Concepts from Unlabeled Multi-Modal Data/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..d40217b114dbf17097710b44f63699d48491e549 --- /dev/null +++ b/NeurIPS/2025/$_textit{HiMaCon_}$ Discovering Hierarchical Manipulation Concepts from Unlabeled Multi-Modal Data/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8c9ec43f0aafbb2f6730cf0fce645478d9a318829faa58c4cad1883c69624990 +size 1159180 diff --git a/NeurIPS/2025/$_textit{Hyper-GoalNet}$_ Goal-Conditioned Manipulation Policy Learning with HyperNetworks/c7074468-e0ed-4a67-900c-b1a066f37372_content_list.json b/NeurIPS/2025/$_textit{Hyper-GoalNet}$_ Goal-Conditioned Manipulation Policy Learning with HyperNetworks/c7074468-e0ed-4a67-900c-b1a066f37372_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..ce5fd18d463ad74b1033bdf90487033a08d86c7a --- /dev/null +++ b/NeurIPS/2025/$_textit{Hyper-GoalNet}$_ Goal-Conditioned Manipulation Policy Learning with HyperNetworks/c7074468-e0ed-4a67-900c-b1a066f37372_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2ada6de2b991142dd37a96202487cfc15cd82e1226b7ef6d5f377dec3c5fb4ba +size 200658 diff --git a/NeurIPS/2025/$_textit{Hyper-GoalNet}$_ Goal-Conditioned Manipulation Policy Learning with HyperNetworks/c7074468-e0ed-4a67-900c-b1a066f37372_model.json b/NeurIPS/2025/$_textit{Hyper-GoalNet}$_ Goal-Conditioned Manipulation Policy Learning with HyperNetworks/c7074468-e0ed-4a67-900c-b1a066f37372_model.json new file mode 100644 index 0000000000000000000000000000000000000000..3c572fcca33ed8fd799191e9281992f0286f86cb --- /dev/null +++ b/NeurIPS/2025/$_textit{Hyper-GoalNet}$_ Goal-Conditioned Manipulation Policy Learning with HyperNetworks/c7074468-e0ed-4a67-900c-b1a066f37372_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a6a2be193f4619b65bfc4ab843a0e78be3733c71917da3728ca8d323e3ac6935 +size 262910 diff --git a/NeurIPS/2025/$_textit{Hyper-GoalNet}$_ Goal-Conditioned Manipulation Policy Learning with HyperNetworks/c7074468-e0ed-4a67-900c-b1a066f37372_origin.pdf b/NeurIPS/2025/$_textit{Hyper-GoalNet}$_ Goal-Conditioned Manipulation Policy Learning with HyperNetworks/c7074468-e0ed-4a67-900c-b1a066f37372_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..2c131382189bc6cd5a7b3b238517d1acf3889c51 --- /dev/null +++ b/NeurIPS/2025/$_textit{Hyper-GoalNet}$_ Goal-Conditioned Manipulation Policy Learning with HyperNetworks/c7074468-e0ed-4a67-900c-b1a066f37372_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ae99d3a048178225550f350fbabcf7646d0ae1ec873964c579dcc6a97bfd921c +size 7716312 diff --git a/NeurIPS/2025/$_textit{Hyper-GoalNet}$_ Goal-Conditioned Manipulation Policy Learning with HyperNetworks/full.md b/NeurIPS/2025/$_textit{Hyper-GoalNet}$_ Goal-Conditioned Manipulation Policy Learning with HyperNetworks/full.md new file mode 100644 index 0000000000000000000000000000000000000000..a540f7904411f2a2f0afb4a7ab6f856e40de2a45 --- /dev/null +++ b/NeurIPS/2025/$_textit{Hyper-GoalNet}$_ Goal-Conditioned Manipulation Policy Learning with HyperNetworks/full.md @@ -0,0 +1,975 @@ +# Hyper-GoalNet: Goal-Conditioned Manipulation Policy Learning with HyperNetworks + +Pei Zhou1 Wanting Yao1,2* Qian Luo1 Xunzhe Zhou1 Yanchao Yang1 + +1InfoBodied AI Lab, The University of Hong Kong 2University of Pennsylvania {pezhou, qianluo1, xunzhe_zhou}@connect.hku.hk wtyao@seas.upenn.edu, yanchaoy@hku.hk + +# Abstract + +Goal-conditioned policy learning for robotic manipulation presents significant challenges in maintaining performance across diverse objectives and environments. We introduce Hyper-GoalNet, a framework that generates task-specific policy network parameters from goal specifications using hypernetworks. Unlike conventional methods that simply condition fixed networks on goal-state pairs, our approach separates goal interpretation from state processing – the former determines network parameters while the latter applies these parameters to current observations. To enhance representation quality for effective policy generation, we implement two complementary constraints on the latent space: (1) a forward dynamics model that promotes state transition predictability, and (2) a distance-based constraint ensuring monotonic progression toward goal states. We evaluate our method on a comprehensive suite of manipulation tasks with varying environmental randomization. Results demonstrate significant performance improvements over state-of-the-art methods, particularly in high-variability conditions. Real-world robotic experiments further validate our method's robustness to sensor noise and physical uncertainties. Code is available at: https://github.com/wantingyao/hyper-goalnet. + +# 1 Introduction + +Goal-conditioned policy learning enables embodied agents to adjust their actions based on current state observations and specified goals [8, 27, 46]. By integrating goal information into decision making, agents leverage knowledge across various tasks, enhancing adaptability [7, 12, 35] in hierarchical reinforcement learning and complex imitation learning [4, 17, 49]. + +Conventional approaches typically concatenate goal observations with current states as input to a fixed-parameter network [9, 56, 53]. This design creates a fundamental limitation: the network must process all possible goal-current state combinations using the same fixed weights, conflating "what" to process (current state) with "how" to process it (goal-dependent strategy). Consequently, these architectures often struggle with generalization to novel goals and complex manipulation tasks that require different processing strategies depending on the goal specification. + +We aim to rethink this relationship by treating goals not as additional input features but as specifications that determine how current observations should be processed. Hypernetworks – neural networks that generate weights for another network – offer a natural implementation of this perspective. By explicitly modeling goals as determinants of policy parameters rather than as inputs, hypernetworks effectively disentangle task-dependent processing (defined by goals) from state-dependent processing (applied to current observations) [19, 45]. This approach better aligns with biological goal-directed + +![](images/ea3d462609b1570b11f41e26cfa96a4ebc9f99f5bcb8bcd158b49e2fbdaf9bf0.jpg) +Figure 1: The proposed Goal-Conditioned Policy Generation framework (Hyper-GoalNet) and conventional goal-conditioned policies. Existing methods typically employ a fixed-parameter policy network that processes concatenated current observations and goal states, treating goals mostly as additional inputs. In contrast, our approach formulates policy learning as an adaptive generation task, where the goal image determines the parameters of the policy network itself – transforming goals from inputs into specifications that define how current observations should be processed. This allows for more effective handling of diverse goals and complex manipulation tasks. + +![](images/e9508fd999f648b982f1836b730bc482ceab306654f151609dad341c943487f1.jpg) + +behavior, where prefrontal regions interpret task goals and dynamically modulate processing in sensorimotor circuits accordingly [32, 51]. + +To address these challenges, we present Hyper-GoalNet, a hypernetwork-based framework for robotic manipulation illustrated in Fig. 1. Our approach employs a hypernetwork to dynamically generate target policy parameters conditioned on specified goals. When loaded into the policy network, these parameters enable the processing of current observations without requiring further access to goal information. This architecture creates a clear separation of concerns: the hypernetwork interprets what the goal means for processing strategy, while the generated policy focuses exclusively on transforming current observations into appropriate actions. A key advantage of this goal-aware design is that it enables the system to autonomously detect task completion during execution. By training the hypernetwork to model the conditional distribution of effective policy weights given goal specifications, we obtain a system that can adapt its processing strategy to diverse manipulation tasks. + +Our technical contributions center on effectively applying hypernetworks for parameter-adaptive goal-conditioned policies learning. First, we adapt optimization-inspired hypernetwork architectures for generating policy parameters conditioned on goal specifications, creating a framework that dynamically determines how current observations should be processed. Second, we introduce an effective latent space shaping technique that imposes two critical properties: (1) predictability of future states through a learned dynamics model, and (2) preservation of physical relationships through distance constraints that ensure monotonic progression toward goals. These properties create an ideal representation space for our hypernetwork, providing clear signals about how policy parameters should change as states approach goals. + +Our extensive experiments across multiple manipulation tasks show that Hyper-GoalNet significantly outperforms state-of-the-art methods, achieving higher success rates on complex contact-rich manipulations. Notably, while conventional approaches fail almost completely in high-variability environments, our method maintains robust performance. Ablation studies confirm the critical importance of our proposed components in the parameter-adaptive goal-conditioned policy leaning framework. Finally, real-robot experiments demonstrate that our parameter-adaptive approach succeeds in physical environments where conventional methods struggle with sensor noise and environmental variations. These results confirm that explicitly modeling goals as determinants of processing strategy rather than as additional inputs creates a more effective and robust framework for goal-conditioned manipulation. + +# 2 Related Work + +Goal-conditioned policy. Goal-conditioned policy learning has attracted significant attention for developing versatile and generalizable agents [24, 52, 50, 57, 25]. Traditional methods augment + +state spaces with goal information and train policies that condition on these augmented states [13, 44, 30, 54]. Hindsight Experience Replay (HER) [1] exemplifies this approach by allowing agents to learn from failures by reinterpreting unsuccessful outcomes as alternative goals. However, these methods typically suffer from increased complexity and require extensive tuning to manage the parameters associated with goal conditioning [39, 33, 48]. Our approach distinguishes itself by utilizing hypernetworks to dynamically generate policy weights, thereby reducing parameters in the policy network while enhancing scalability without extensive retraining [20, 43]. + +In the realm of imitation learning for goal-conditioned policies, several frameworks have demonstrated promising results by learning from pre-collected demonstrations [29, 9, 53]. Recent goal-conditioned behavior cloning approaches such as C-BeT [9] and MimicPlay [53] have advanced long-horizon manipulation tasks. However, these methods typically require sequences of achievable goal images, which are challenging to obtain in practice. Moreover, while performing well in basic pick-and-place scenarios, they often struggle with contact-rich tasks that demand precise environmental awareness [31]. Our method overcomes these limitations through effective latent space shaping, requiring only a single goal image while maintaining robustness across diverse manipulation scenarios. + +Hypernetworks and Cognitive science insights for goal-directed behavior. Our work draws inspiration from cognitive science research on human goal-directed behavior, where meta-cognitive strategies and higher-level planning mechanisms enable adaptive actions [6, 36, 37]. Studies show that humans efficiently manage cognitive resources and flexibly adapt to different goals through higher-level representations [18, 11, 10]. Current policy learning methods incorporating cognitive principles often focus on imitation learning to mimic human strategies [2, 16, 14], but can be limited by demonstration quality and diversity [38, 42, 22]. Hypernetworks have been explored in robotic control [21, 55, 3], though primarily within reward-driven reinforcement learning (RL) settings [23, 5]. This fundamental difference in training paradigms, RL versus our reward-free behavior cloning (BC), means their end-to-end algorithms are not directly adaptable. We clarify, however, that their core hypernetwork architectures can be decoupled from the RL framework. By embedding cognitive insights into our hypernetwork architecture, we emulate human-like flexible adaptation [6, 36, 37]. Hyper-GoalNet's capacity to generate goal-specific policy parameters without extensive retraining addresses practical challenges of goal-conditioned learning [20] while mirroring key cognitive mechanisms, offering a biologically plausible framework for adaptable policy generation. + +# 3 Method + +Let $\mathcal{D} = \{\tau_i\}_{i=1}^M$ be a dataset consisting of $M$ robotic manipulation demonstrations, where each trajectory comprises a sequence of observation-action pairs, i.e., $\tau_i = \{(o_j^i, a_j^i)\}_{j=1}^{N_i}$ , with $a_j$ denoting a continuous-valued action and $o_j$ representing a tuple containing high-dimensional state observations. Specifically, $o_j$ includes an RGB image $I_j$ captured by a single front-view camera, as well as the proprioceptive information $s_j$ of the embodied agent. Given this formulation, our objective is to develop a generalizable goal-conditioned policy learning framework that enables efficient adaptation to diverse manipulation tasks. + +We propose a shift from conventional goal-conditioned policies, which typically use fixed parameters while processing both current and goal images. Our key insight is that the goal image inherently specifies how the current image should be processed to generate appropriate actions. Therefore, we argue that the policy parameters themselves – which determine the processing mechanism – should adapt based on different goal specifications. To realize this insight, we leverage hypernetworks to dynamically generate task-specific policy parameters conditioned on goal images, rather than directly conditioning a fixed policy network on both current and goal observations. This approach creates a more flexible and efficient framework where the processing of current states is explicitly tailored to the specified goals. + +The full pipeline is illustrated in Fig. 2. Next, we elaborate on the key components of our approach. In Sec. 3.1, we describe how we adapt hypernetwork architectures to effectively generate varying goal-reaching policies. In Sec. 3.2, we introduce an effective latent space shaping techniques that significantly enhance the performance of goal-conditioned policy generation by enforcing meaningful geometric structure in the representation space. Finally, in Sec. 3.3, we present the test-time inference pipeline that integrates our trained model (Hyper-GoalNet) to accomplish diverse manipulation tasks, highlighting the practical advantages of our parameter-adaptive approach. + +![](images/7a6af8488eb2620867c28e0b9cec8f9340510209666019ce4c67e000c5b9a89a.jpg) +Figure 2: An overview of the proposed Hyper-GoalNet framework. (a) Adaptive Policy Generation: Unlike conventional approaches with fixed parameters, our hypernetwork dynamically generates task-specific policy parameters conditioned on goal images. This creates a parameter-adaptive target policy that processes current observations (RGB images and proprioception) through a multimodal encoder to produce actions tailored to specific goals. (b) Latent Shaping: Our approach enhances performance by explicitly structuring the latent space in two ways: a predictive network models state transitions to improve temporal dynamics, while geometric constraints ensure distances to goal states monotonically decrease during successful trajectories (detailed in Sec. 3.2). + +![](images/7ff97a2ad594b63e830eb36144d3ff4f6987c1054472e4e401444038a9c2d63e.jpg) + +# 3.1 Goal-Conditioned Hypernetworks + +We formulate goal-conditioned policy learning as a parameter generation task rather than a direct conditioning problem. This formulation reframes the challenge from "what action to take given current and goal observations" to "what processing parameters to use given the goal." + +More formally, given a current observation $o_c$ and a desired goal observation $o_g$ , we model the conditional distribution over target policy weights that will transform the scene into the goal configuration. With the set of robotic manipulation demonstrations $\mathcal{D}$ , we learn the distribution $\mathcal{H}(\theta \mid o_c, o_g)$ : + +$$ +\mathcal {H} \left(\theta \mid o _ {c}, o _ {g}\right) := \mathbb {P} _ {\mathcal {D}} \left(\theta \mid o _ {c} = I _ {t}, o _ {g} = I _ {t ^ {\prime}}\right), \text {w h e r e} I _ {t}, I _ {t ^ {\prime}} \in \tau_ {i} \in \mathcal {D} \text {a n d} t ^ {\prime} > t. \tag {1} +$$ + +For practical implementation, we focus on a single goal state rather than a sequence of goals, and use RGB image observations to condition the hypernetwork. Since our primary objective is to investigate the efficacy of goal-conditioned policy generation for manipulation tasks rather than developing a full probabilistic model, we approximate this as a deterministic mapping: + +$$ +\mathcal {H}: \mathcal {O} \times \mathcal {O} \rightarrow \Theta , \tag {2} +$$ + +where $\mathcal{O}$ denotes the observation space and $\Theta$ represents the space of target policy parameters. With a current-goal observation pair $(o_c, o_g)$ , the hypernetwork $\mathcal{H}$ produces a policy that guides the transition from current state $o_c$ to goal state $o_g$ through action execution. + +Hypernetwork architecture. To implement our parameter-adaptive approach, we adopt a hypernetwork architecture that efficiently generates policy parameters for achieving specified goals. The architecture must be capable of capturing the complex relationships between current states, goal states, and the required actions to bridge them. + +We leverage an optimization-inspired architecture following [40], which provides beneficial inductive bias for our parameter generation task. This approach mimics iterative optimization by refining policy parameters through multiple feed-forward steps: + +$$ +\theta^ {K} = \mathcal {H} \left(o _ {c}, o _ {g}\right), \tag {3} +$$ + +where $\theta^K$ represents the final policy parameters after $K$ refinement iterations, with each update computed as: + +$$ +\theta^ {k} = \theta^ {k - 1} + \lambda^ {k} \left(\theta^ {k - 1}, \alpha\right) \psi^ {k} \left(\theta^ {k - 1}, \alpha\right), \alpha = \phi \left(o _ {c}, o _ {g}\right). \tag {4} +$$ + +Here, neural modules $\lambda^k$ and $\psi^k$ serve as learned analogs to step sizes and gradients in optimization, operating on embeddings $\phi(o_c, o_g)$ of the current and goal observations. This mechanism enhances the hypernetwork's ability to generate effective task-specific policy parameters and improves generalization to new goal specifications. Once we obtain the goal-conditioned policy weights $\theta$ , we can process the current observation through the generated policy to predict appropriate actions. + +Hypernetwork Training. We train our hypernetwork $\mathcal{H}:\mathcal{O}\times \mathcal{O}\to \Theta$ to generate parameters for the target visuomotor policy $\pi (\cdot ;\theta)$ using behavior cloning (BC) on demonstration data. The generated policy takes the current observation $o_t$ (comprising image $I_{t}$ and proprioception $s_t$ ) and outputs actions for execution. + +To enhance robustness, we utilize a sequence of $L$ consecutive observations as input to the policy, capturing temporal dependencies under a non-Markovian assumption. The training objective minimizes the BC loss between demonstrated actions $a_{t}^{i}$ and predicted actions $\hat{a}_{t}^{i}$ : + +$$ +\mathcal {L} _ {\text {p o l i c y}} = \sum_ {i = 1} ^ {M} \sum_ {1 \leq t < t ^ {\prime} \leq N _ {i}} \ell \left(a _ {t} ^ {i}, \hat {a} _ {t} ^ {i}\right), \quad \hat {a} _ {t} ^ {i} = \pi \left(o _ {t - L: t} ^ {i}; \mathcal {H} \left(o _ {t} ^ {i}, o _ {t ^ {\prime}} ^ {i}\right)\right). \tag {5} +$$ + +Here, $\ell$ represents the Mean Squared Error (MSE) loss. The end-to-end training procedure works as follows: for a given current observation $o_{t}^{i}$ and a goal image $o_{t'}^{i}$ , the hypernetwork $\mathcal{H}$ generates the weights for the policy network $\pi$ . This goal-conditioned policy then processes the observation sequence $o_{t-L:t}^{i}$ to predict the action $\hat{a}_{t}^{i}$ . The resulting loss is backpropagated through both the policy network and the hypernetwork. We restrict goal representations to image inputs since proprioceptive goal states may not always be available in practical applications. This formulation allows the hypernetwork to learn how to generate goal-specific processing mechanisms (target policy parameters) from visual goals, embodying our key insight that goal images determine how current observations should be processed. + +# 3.2 Latent Space Shaping + +A critical insight in our approach is that the effectiveness of parameter-adaptive policies depends significantly on the quality of the representation space in which observations are embedded. While our hypernetwork can directly generate policy parameters from raw observations, we find that explicitly shaping the latent representation space substantially enhances performance. Given the high dimensionality and redundant information in RGB images, we employ an image encoder $\mathcal{E}$ to extract task-relevant features and compress them into low-dimensional latents $z = \mathcal{E}(I)$ . + +We identify two fundamental properties that, when enforced in the latent space, particularly benefit our parameter-adaptive approach: predictability and physical structure preservation. The first property ensures the latent space facilitates modeling of state transitions, making the hypernetwork's task of generating appropriate policy parameters more tractable. The second property ensures that the geometric relationships in the latent space meaningfully reflect physical relationships between states, enabling the generated policies to exploit these structured representations. + +Enhancing predictability through dynamic modeling. To improve the predictability of latent representations, we introduce a dynamics model that forecasts future states in the latent space. By training this model to predict state transitions while simultaneously shaping the encoder $\mathcal{E}$ through backpropagation, we create a latent space where sequential relationships are explicitly captured. This significantly benefits our hypernetwork, as it needs to generate policy parameters that leverage these sequential relationships to guide transitions from current to goal states. + +Formally, consider a discrete-time dynamical system with state representation $z_{t} \in \mathcal{Z}$ and control input $a_{t} \in \mathcal{A}$ . The forward dynamic model is: + +$$ +\hat {z} _ {t + 1} \sim p _ {\Phi} \left(z _ {t + 1} \mid z _ {t}, a _ {t}\right), \tag {6} +$$ + +where $z_{t} = \mathcal{E}(I_{t})$ represents the latent encoding at time $t$ , $a_{t}$ is the executed action, and $p_{\Phi}$ is the transition dynamics parameterized by $\Phi$ . For practical implementation, we approximate this with a deterministic model $\Phi : \mathcal{Z} \times \mathcal{A} \rightarrow \mathcal{Z}$ . The corresponding learning objective minimizes the prediction loss: + +$$ +\mathcal {L} _ {\text {p r e d}} = \mathbb {E} _ {\tau \sim \mathcal {D}} \left[ \ell \left(\Phi \left(z _ {t}, a _ {t}\right), z _ {t + 1}\right) \right], \tag {7} +$$ + +where $\ell$ is a distance metric in the latent space. When $\Phi$ is well-trained, we further finetune $\mathcal{E}$ through $\Phi$ to shape representations that capture both current state and potential transition information. + +Preserving physical structure through distance constraints. For our parameter-adaptive approach to be effective, the latent space must preserve the physical structure of the task, particularly the progression towards goals. We formalize this as a requirement that the distance between the current state and goal state should monotonically decrease along goal-reaching trajectories. This property is + +especially valuable for our hypernetwork, as it provides a clear signal about how policy parameters should change as states approach goals. + +Specifically, for any goal-reaching trajectory $\tau_{i} \in \mathcal{D}$ , where $\tau_{i} = \{(o_{j}^{i}, a_{j}^{i})\}_{j=1}^{N_{i}}$ , we enforce: + +$$ +d _ {\mathcal {E}} \left(o _ {j} ^ {i}, o _ {j ^ {\prime}} ^ {i}\right) \geq d _ {\mathcal {E}} \left(o _ {j + 1} ^ {i}, o _ {j ^ {\prime}} ^ {i}\right), \quad \forall j < j ^ {\prime}, \tag {8} +$$ + +where $d_{\mathcal{E}}$ denotes a distance metric in the latent space. To explicitly model this behavior, we propose the following loss function: + +$$ +\mathcal {L} _ {\text {d i s t}} = \mathbb {E} _ {\tau \sim \mathcal {D}} \sum_ {j} \max \left(0, \beta + d \left(z _ {j + 1}, z _ {g}\right) - d \left(z _ {j}, z _ {g}\right)\right), \tag {9} +$$ + +where $d(z_{1},z_{2}) = \| z_{1} - z_{2}\|_{2}$ represents the Euclidean distance between latent features, $z_{j} = \mathcal{E}(I_{j})$ and $z_{g} = \mathcal{E}(I_{g})$ denote the image and goal image latents, respectively. The margin parameter $\beta \geq 0$ enforces a minimum decrease in distance between consecutive states and the goal. Empirically, setting $\beta = 0$ suffices to induce the desired monotonic progression. + +This shaped latent space creates an ideal foundation for our parameter-adaptive approach, as it encodes both the predictive dynamics and geometric structure needed for the hypernetwork to effectively generate goal-tailored policy parameters. Figure 3 illustrates how our latent space shaping approach compares to alternative methods, showing the enhanced structure that benefits our goal-conditioned policy generation. + +# 3.3 Hyper-GoalNet for Manipulation + +Having established our parameter-adaptive architecture and latent space shaping techniques, we now present our complete framework, Hyper-GoalNet, which synthesizes these components for effective goal-conditioned manipulation. The overall training objective combines our policy generation loss with the latent space shaping terms: + +$$ +\mathcal {L} _ {\text {H y p e r - G o a l N e t}} = \mathcal {L} _ {\text {p o l i c y}} + \lambda_ {\text {p r e d}} \mathcal {L} _ {\text {p r e d}} + \lambda_ {\text {d i s t}} \mathcal {L} _ {\text {d i s t}}, \tag {10} +$$ + +where $\lambda_{\mathrm{pred}}$ and $\lambda_{\mathrm{dist}}$ are weight coefficients balancing the contributions of predictability and structural constraints. The framework is trained end-to-end using gradient descent, allowing all components to co-adapt for optimal performance. + +During inference for task completion, Hyper-GoalNet generates goal-specific policy parameters by feeding the concatenated latent features $[z_g,z_t]$ into the hypernetwork $\mathcal{H}$ , where $z_{g} = \mathcal{E}(I_{g})$ and $z_{t} = \mathcal{E}(I_{t})$ are the latent representations of the goal and current observations. The generated goal-specific parameters $\theta = \mathcal{H}(z_t,z_g)$ are then loaded into the target policy $\pi (\cdot ;\theta)$ , which processes the current observation sequence to produce actions that guide the agent toward the goal. + +Hyper-GoalNet offers two principal advantages over conventional goal-conditioned policies: + +1) Parameter-adaptive policy generation: By dynamically synthesizing policy parameters based on goal specifications, our approach effectively changes how goal-conditioned policies operate. Rather than relying on a fixed network with static parameters to handle all possible goals, Hyper-GoalNet generates compact and efficient processing pathways that are tailored to specific goals. + +2) Natural goal completion detection: Our shaped latent space provides an elegant solution to the challenging problem of goal completion detection. The distance metric $d_{\mathcal{E}}(o_t, o_g)$ in the latent space serves as a natural criterion for determining when a goal has been achieved, enabling autonomous goal transitions without external supervision. + +The test-time task evaluation procedure for Hyper-GoalNet is formalized in Algorithm 1. Given an initial observation $I_0$ and goal observation $I_g$ , the algorithm iteratively generates policy parameters, samples actions, and applies them until either the goal is reached (as determined by the latent distance falling below a threshold $\epsilon$ ) or a maximum number of steps $T$ is achieved. This simple yet effective procedure demonstrates how our parameter-adaptive approach seamlessly integrates into practical robotic control scenarios. + +# 4 Experiments + +In this section, we present a comprehensive experimental evaluation across a suite of simulated and real-robot manipulation tasks designed to address the following questions: 1) How effectively does + +Algorithm 1 Hyper-GoalNet: Test-Time Task Evaluation +Input: Initial observation $I_0$ , Goal observation $I_g$ +Modules: Encoder $\mathcal{E}$ , Policy generation hypernetwork $\mathcal{H}$ +Parameters: Max steps $T$ , Goal completion threshold $\epsilon$ +1: $I_t \gets I_0$ , $t \gets 0$ , done $\gets$ false +2: while $t < T$ and not done do +3: $\theta \gets \mathcal{H}(\mathcal{E}(I_t), \mathcal{E}(I_g))$ +4: $\hat{a}_t \gets \pi(o_{t-L:t}; \theta)$ +5: $I_{t+1}$ , done $\gets \mathrm{Env}(\hat{a}_t)$ +6: if $d(\mathcal{E}(I_{t+1}), \mathcal{E}(I_g)) < \epsilon$ or done then +7: return SUCCESS +8: end if +9: $I_t \gets I_{t+1}$ , $t \gets t+1$ +10: end while +11: return TIMEOUT + +Hyper-GoalNet's parameter-adaptive approach generate successful policies for diverse manipulation tasks? 2) To what extent does our latent space shaping enhance the performance of the hypernetwork for goal-conditioned policy generation? 3) How does Hyper-GoalNet compare with conventional goal-conditioned methods and alternative representation learning approaches? Through extensive empirical analysis, we validate the effectiveness of our parameter-adaptive framework. + +# 4.1 Experiment Setup + +Simulation Environment. We evaluate our approach using Robosuite, a comprehensive robotics benchmark designed for both short and long-horizon manipulation tasks [31, 58]. This framework provides a standardized suite of environments, from which we select multiple contact-rich tabletop manipulation tasks: coffee manipulation, threading, mug cleanup, nut assembly, three-piece assembly, and several long-horizon tasks including coffee preparation and kitchen manipulation. To assess robustness across varying initial conditions, we use three difficulty levels $(\mathrm{d}_0, \mathrm{d}_1, \mathrm{d}_2)$ , where higher indices correspond to increased environmental variability, particularly in object pose initialization (position and orientation). Each experimental environment features a robotic manipulator positioned adjacent to a workspace containing task-specific manipulable objects. + +Training Protocol. Our approach follows the behavior cloning paradigm, utilizing a dataset based on MimicGen [31]. For each task, we employ a training dataset of 950 demonstrations, where each timestep comprises front-view RGB images ( $128 \times 128$ resolution), robot proprioceptive states $s_t \in S$ , and corresponding ground-truth actions $a_t \in \mathcal{A}$ . The training procedure employs the Adam optimizer [26] with a cosine learning rate schedule [28]. We initialize the learning rate at $5 \times 10^{-4}$ and maintain uniform loss balancing coefficients ( $\lambda_i = 1$ for all components). Our model is trained for 500 epochs with a batch size of 256, by default. Detailed implementation and training protocols are provided in the Sec. D. + +# 4.2 Main Results + +Baselines. We compare our parameter-adaptive approach against state-of-the-art goal-conditioned methods that use fixed network parameters. All methods are trained on pre-collected demonstrations from MimicGen [31] and modified to operate with a single future image as the goal specification for fair comparison: + +- GCBC [29, 15]: Goal-Conditioned Behavioral Cloning concatenates current and goal observations as input to a fixed policy network, learning a direct mapping from this concatenated representation to actions through supervised learning on demonstration data. +- Play-LMP [29]: Play-supervised Latent Motor Plans learns a latent plan space from demonstration data, then trains a fixed-parameter policy conditioned on both the current state and the inferred latent plan for the specified goal. + +Table 1: Comparison with state-of-the-art goal-conditioned methods. Success rates (higher is better) are computed over 50 rollouts across various manipulation tasks with increasing difficulty levels (d0-d2). The experimental setup uses only two historical observations and a single goal image, representing a practical deployment scenario. Our method consistently outperforms conventional fixed-parameter approaches, demonstrating the effectiveness of dynamically generating policy parameters based on goal specifications. + +
MethodCoffee ↑Mug-cleanup ↑Three piece Assemb. ↑Threading ↑Nut Assemb. ↑
d0d1d2Avg.d0d1Avg.d0d1d2Avg.d0d1d2Avg.d0Avg.
GCBC0.000.000.000.000.000.000.000.000.000.000.000.000.000.000.000.000.00
Play-LMP0.000.000.000.000.000.000.000.000.000.000.000.000.000.000.000.000.00
MimicPlay0.280.280.160.240.260.060.160.060.060.000.040.180.020.000.070.030.12
C-BeT0.920.000.740.550.300.500.400.000.020.000.010.620.220.120.320.340.32
Ours0.940.760.620.770.780.460.620.520.200.040.250.820.320.240.460.550.52
+ +Table 2: Long-horizon task performance. Our parameter-adaptive approach excels in complex sequential tasks, outperforming fixed-parameter methods across difficulty levels. + +
MethodCoffeeKitchenAvg.
d0d1Avg.d0d1Avg.
GCBC0.000.000.000.000.000.000.00
Play-LMP0.000.000.000.000.000.000.00
MimicPlay0.340.000.170.860.180.520.35
C-BeT0.820.040.430.780.700.740.59
Ours0.800.500.651.000.800.900.78
+ +Table 3: Component ablation analysis. Latent space shaping significantly enhances performance, while proper distance metrics and training are crucial for robustness. + +
MethodCoffee ↑
d0d1d2Avg.
Ours(uf. at epoch 0)0.920.000.620.51
Ours(w/o shaping)0.920.000.000.31
Ours(dist ↔ start img.)0.500.520.320.45
Ours(cos dist)0.940.360.480.59
C-BeT(w/ shaping)0.800.640.640.69
Ours0.940.760.620.77
+ +- C-BeT [9]: Conditional Behavior Transformer uses self-attention to compress observation history into a latent representation, which is combined with the goal state to condition a fixed-parameter transformer that predicts actions. +- MimicPlay [53]: MimicPlay is a self-supervised approach that learns general robotic skills from unstructured teleoperation data, which consists of continuous sequences of observations and actions from a human video. For our experiments, this method is adapted into a goal-image-conditioned policy, with implementation details provided in the Appendix. + +Evaluation Metrics. We evaluate performance using task completion success rates over 50 independent rollouts with randomly initialized, previously unseen environmental configurations. We impose maximum trajectory lengths of $T = 600$ or 800 steps for contact-rich tasks and $T = 1600$ for long-horizon tasks. While our parameter-adaptive approach enables autonomous task completion detection through latent space metrics (Algorithm 1), we use environment-provided terminal signals for standardized evaluation across all methods. Success is indicated as $S_{i} = 1$ if rollout $i$ completes within $T$ steps, and $S_{i} = 0$ otherwise. + +Quantitative Results. Tables 1 and 2 present success rates across multi-step and long-horizon tasks, respectively. For each task, success is determined by task-specific criteria provided by the environment – such as correct object placement, successful insertion, proper assembly configuration, or completion of a sequence of subtasks for long-horizon scenarios. Our parameter-adaptive approach outperforms fixed-parameter methods across these diverse evaluation criteria and difficulty levels. This superior performance stems from our hypernetwork's ability to dynamically generate task-specific policy parameters tailored to each goal, resulting in more effective goal-directed behavior. Particularly in high-variability environments (difficulty levels d1-d2), our method demonstrates greater robustness and adaptability compared to conventional approaches – highlighting the advantage of having policy parameters explicitly conditioned on goals rather than using fixed parameters for all scenarios. + +Analysis of Likelihood-Based Baselines. We diagnose the poor performance of GCBC and PlayLMP as a fundamental issue of their learning objective, not implementation. These methods, adapted from reputable third-party code, aim to maximize the log-likelihood of expert actions. We found this leads to severe overfitting on the training data, evidenced by a large gap between low training loss and high validation loss. Such memorization-based learning fails to generalize to the subtle variations + +present in our high-precision test scenarios. This limitation of likelihood-based models in complex settings is corroborated by prior work [9]. In contrast, our hypernetwork's design imposes a beneficial structural bias: by emulating an optimization process to generate parameters, it is incentivized to learn a functional mapping from goal to policy, ensuring better generalization and avoiding the overfitting issues that plague the baselines. + +# 4.3 Ablation Study + +Hypernetwork Architecture Analysis. We evaluate our optimization-inspired hypernetwork design against HyperZero [41], a prominent alternative architecture. Architecturally, HyperZero encodes conditioning information into a meta-embedding that is then transformed to produce the parameters for the target policy network. Because this direct mapping can produce parameters with a numerical range misaligned with that of an optimally trained network, it often requires special initialization to stabilize training. To ensure a stable and robust + +Table 4: Ablating hypernetwork architectures. Our method with standard initialization is compared against HyperZero variants stabilized with enhanced initializations. + +
MethodCoffee Task (%) ↑
d0d1d2Avg.
HyperZero + ScalarInit16181416
HyperZero + Bias-Init3018016
Ours (Standard Init.)94766277
+ +comparison, we therefore implemented two enhanced initialization schemes for HyperZero. The first, ScalarInit, introduces a learnable scalar to control the initial scale of the hypernetwork's output. The second, Bias-Init [5], is designed for high-dimensional conditioning inputs and constrains the parameter range by incorporating learnable biases alongside weights that are initialized to zero. As shown in Table 4, even with these stabilization techniques, our architecture, which requires no special initialization, vastly outperforms HyperZero. This demonstrates that the iterative refinement mechanism in our design is inherently more effective at capturing the relationship between goals and appropriate policy parameters. These results highlight the critical role of architectural choice in developing robust parameter-adaptive policies. More details can be found in Sec. C. + +Latent Space Shaping Analysis. Table 3 shows that proper latent space shaping is critical for our parameter-adaptive framework. Removing shaping ("Ours w/o shaping") severely degrades performance on harder difficulty levels, while our specific choices of using goal-relative distances and Euclidean metrics prove superior to alternatives. Figure 3 visually confirms how our approach creates more consistent monotonic progression toward goals compared to unshaped representations. Notably, applying our shaping techniques to the baseline C-BeT improves its performance, yet its performance lags compared to ours, which in turn signifies the importance of the parameter-adaptive framework as well as the synergy between policy generation and latent shaping. Our training strategy also matters – unfreezing the R3M visual encoder [34] only after 20 epochs ensures stable parameter generation. These results validate our insight that latent spaces should reflect physical progression toward goals to effectively support parameter-adaptive policy learning. + +Visualization. Figures 3 and 6 illustrate how our latent space shaping creates representations that directly benefit our parameter-adaptive approach. The plot shows the latent distance to the goal (y-axis) over the execution timesteps of a robotic task rollout (x-axis), where later steps are progressively closer to the goal. Unlike R3M embeddings which show significant fluctuations, our method produces consistently monotonic distance reductions toward goal states. This structured latent space offers two key advantages for our hypernetwork: (1) it provides clearer signals for generating appropriate policy parameters as the agent progresses toward goals, and (2) it enables reliable autonomous detection of task completion based on latent distances. The + +![](images/0584af9c17fb6735ae95456daee1e8d96322a93f590fb28edfb86188866ff78b.jpg) +Figure 3: Comparison of (a) unshaped R3M embeddings versus (b) our shaped latent space, showing $L_{2}$ distances to goal states along multiple trajectories. Our shaping creates consistent monotonic decreases in distance-to-goal, facilitating more effective parameter generation. + +visualization confirms that our combined predictive modeling and distance constraints create opti + +mal representations for goal-conditioned parameter generation, enhancing both performance and interpretability. + +Goal Completion Detection. To substantiate our claim that the shaped latent space enables autonomous goal completion detection, we supplement the qualitative evidence from Figures 3, 9, and 10. We evaluate this capability by comparing our autonomous detection, where success is determined by a latent distance threshold (Auto SR), against the environment's ground-truth signal (Env SR). Table 5 presents the results using two key metrics: Accuracy, which measures the agreement between the two signals, and Recall, which measures our + +Table 5: Quantitative validation of autonomous goal completion detection. Our method's latent distance-based success rate (Auto SR) is compared against the environment's ground truth (Env SR) on the Coffee tasks. High Accuracy and Recall validate its reliability. + +
TaskAuto SREnv SRAccuracyRecall
D00.960.9494%98%
D10.780.7690%95%
D20.740.6276%90%
Mean--86.6%94.3%
+ +method's ability to identify true successes reported by the environment. The strong alignment, evidenced by an average accuracy of $86.6\%$ and recall of $94.3\%$ , provides strong empirical evidence that our latent-distance-based approach is a reliable autonomous completion detector. + +# 4.4 Real Robot Experiments + +We validate our parameter-adaptive approach on physical hardware using the Realman Robotics Platform, featuring a 7-DoF manipulator with a 1-DoF parallel gripper (Figure 7). We evaluate four diverse manipulation tasks - sweep, pick&place, pull, and stack - with 15 trials per task. Due to hardware constraints limiting control to joint angles without end-effector pose information, we exclude MimicPlay, which requires precise 3D end-effector trajectories. + +Table 6: Real-robot experiment results. (successes/total trials). + +
MethodPickplacePullStackSweep
GCBC0/150/150/152/15
Play-LMP0/150/150/155/15
C-BeT2/156/155/158/15
Ours14/1515/1514/1515/15
+ +As shown in Table 6, conventional fixed-parameter approaches struggle significantly in real-world conditions where environmental noise, perception uncertainties, and imperfect demonstrations create substantial challenges. In contrast, our method maintains high success rates across all tasks, including those requiring precise contact-rich interactions. This real-world performance gap highlights a key advantage of our parameter-adaptive approach: by dynamically generating task-specific policy parameters based on goal images, our method better adapts to real-world variations and demonstration imperfections that weren't encountered during training. Detailed experimental protocols are provided in the Sec. F. + +# 5 Discussion + +Conclusion. Our parameter-adaptive approach represents an effective move in goal-conditioned policy learning by dynamically generating policy parameters based on goal information rather than using fixed parameters with conditioning. The consistent performance improvements across tasks demonstrate that "how" observations should be processed is inherently dependent on the goal specification. Our latent space shaping techniques prove critical for this architecture - imposing physical structure and predictive capacity provides clearer signals for the hypernetwork to generate effective policy parameters. Overall, our results suggest that explicitly modeling the relationship between goals and processing mechanisms offers a promising direction for more flexible and robust robotic control. + +Limitations. Our method's primary limitation is its reliance on a well-structured latent space, which is challenging to form for highly complex tasks and requires demonstration data with clear goal progression. This data dependency creates a key failure mode: out-of-distribution goals can cause the hypernetwork to generate erratic policies. Furthermore, as an offline-trained method, its dynamic generation of parameters for novel goals lacks explicit safety guarantees against unforeseen states, making the integration of robust safety constraints a critical direction for future research. + +# Acknowledgments and Disclosure of Funding + +This work is supported by the Early Career Scheme of the Research Grants Council (RGC) grant # 27207224, the HKU-100 Award, a donation from the Musketeers Foundation, and in part by the JC STEM Lab of Autonomous Intelligent Systems funded by The Hong Kong Jockey Club Charities Trust. + +# References + +[1] Marcin Andrychowicz, Filip Wolski, Alex Ray, Jonas Schneider, Rachel Fong, Peter Welinder, Bob McGrew, Josh Tobin, OpenAI Pieter Abbeel, and Wojciech Zaremba. Hindsight experience replay. Advances in neural information processing systems, 30, 2017. +[2] Brenna D Argall, Sonia Chernova, Manuela Veloso, and Brett Browning. A survey of robot learning from demonstration. Robotics and autonomous systems, 57(5):469-483, 2009. +[3] Sayantan Auddy, Jakob Hollenstein, Matteo Saveriano, Antonio Rodríguez-Sánchez, and Justus Piater. Scalable and efficient continual learning from demonstration via a hypernetwork-generated stable dynamics model. arXiv preprint arXiv:2311.03600, 2023. +[4] Pierre-Luc Bacon, Jean Harb, and Doina Precup. The option-critic architecture. In Proceedings of the AAAI conference on artificial intelligence, volume 31, 2017. +[5] Jacob Beck, Matthew Thomas Jackson, Risto Vuorio, and Shimon Whiteson. Hypernetworks in meta-reinforcement learning. In Conference on Robot Learning, pages 1478-1487. PMLR, 2023. +[6] Matthew M Botvinick, Yael Niv, and Andrew G Barto. Hierarchically organized behavior and its neural foundations: A reinforcement learning perspective. cognition, 113(3):262-280, 2009. +[7] Qingwen Bu, Jia Zeng, Li Chen, Yanchao Yang, Guyue Zhou, Junchi Yan, Ping Luo, Heming Cui, Yi Ma, and Hongyang Li. Closed-loop visuomotor control with generative expectation for robotic manipulation. In The Thirty-eighth Annual Conference on Neural Information Processing Systems. +[8] Xiaoyu Chen, Junliang Guo, Tianyu He, Chuheng Zhang, Pushi Zhang, Derek Cathera Yang, Li Zhao, and Jiang Bian. Igor: Image-goal representations are the atomic control units for foundation models in embodied ai. arXiv preprint arXiv:2411.00785, 2024. +[9] Zichen Jeff Cui, Yibin Wang, Nur Muhammad Mahi Shafiullah, and Lerrel Pinto. From play to policy: Conditional behavior generation from uncurated robot data. arXiv preprint arXiv:2210.10047, 2022. +[10] Nathaniel D Daw, Yael Niv, and Peter Dayan. Uncertainty-based competition between prefrontal and dorsolateral striatal systems for behavioral control. Nature neuroscience, 8(12):1704-1711, 2005. +[11] Peter Dayan and Nathaniel D Daw. Decision theory, reinforcement learning, and the brain. Cognitive, Affective, & Behavioral Neuroscience, 8(4):429-453, 2008. +[12] Yiming Ding, Carlos Florensa, Pieter Abbeel, and Mariano Pielipp. Goal-conditioned imitation learning. Advances in neural information processing systems, 32, 2019. +[13] Yilun Du, Sherry Yang, Bo Dai, Hanjun Dai, Ofir Nachum, Josh Tenenbaum, Dale Schuurmans, and Pieter Abbeel. Learning universal policies via text-guided video generation. Advances in neural information processing systems, 36:9156-9172, 2023. +[14] Yan Duan, Xi Chen, Rein Houthooft, John Schulman, and Pieter Abbeel. Benchmarking deep reinforcement learning for continuous control. In International conference on machine learning, pages 1329-1338. PMLR, 2016. +[15] Scott Emmons, Benjamin Eysenbach, Ilya Kostrikov, and Sergey Levine. Rvs: What is essential for offline rl via supervised learning? arXiv preprint arXiv:2112.10751, 2021. + +[16] Chelsea Finn, Ian Goodfellow, and Sergey Levine. Unsupervised learning for physical interaction through video prediction. Advances in neural information processing systems, 29, 2016. +[17] Carlos Florensa, Yan Duan, and Pieter Abbeel. Stochastic neural networks for hierarchical reinforcement learning. arXiv preprint arXiv:1704.03012, 2017. +[18] Michael J Frank, Lauren C Seeberger, and Randall C O'reilly. By carrot or by stick: cognitive reinforcement learning in parkinsonism. Science, 306(5703):1940-1943, 2004. +[19] Tomer Galanti and Lior Wolf. On the modularity of hypernetworks. Advances in Neural Information Processing Systems, 33:10409-10419, 2020. +[20] David Ha and Jürgen Schmidhuber. World models. arXiv preprint arXiv:1803.10122, 2018. +[21] Shashank Hegde, Zhehui Huang, and Gaurav S Sukhatme. Hyperppo: A scalable method for finding small policies for robotic control. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 10821-10828. IEEE, 2024. +[22] Jonathan Ho and Stefano Ermon. Generative adversarial imitation learning. Advances in neural information processing systems, 29, 2016. +[23] Yizhou Huang, Kevin Xie, Homanga Bharadhwaj, and Florian Shkurti. Continual model-based reinforcement learning with hypernetworks. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 799-805. IEEE, 2021. +[24] Leslie Pack Kaelbling, Michael L Littman, and Andrew W Moore. Reinforcement learning: A survey. Journal of artificial intelligence research, 4:237-285, 1996. +[25] Moo Jin Kim, Karl Pertsch, Siddharth Karamcheti, Ted Xiao, Ashwin Balakrishna, Suraj Nair, Rafael Rafailov, Ethan Foster, Grace Lam, Pannag Sanketi, et al. Openvla: An open-source vision-language-action model. arXiv preprint arXiv:2406.09246, 2024. +[26] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. +[27] Minghuan Liu, Menghui Zhu, and Weinan Zhang. Goal-conditioned reinforcement learning: Problems and solutions. arXiv preprint arXiv:2201.08299, 2022. +[28] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016. +[29] Corey Lynch, Mohi Khansari, Ted Xiao, Vikash Kumar, Jonathan Tompson, Sergey Levine, and Pierre Sermanet. Learning latent plans from play. In Conference on robot learning, pages 1113-1132. PMLR, 2020. +[30] Jason Yecheng Ma, Jason Yan, Dinesh Jayaraman, and Osbert Bastani. Offline goal-conditioned reinforcement learning via $f$ -advantage regression. Advances in Neural Information Processing Systems, 35:310-323, 2022. +[31] Ajay Mandlekar, Soroush Nasiriany, Bowen Wen, Iretiayo Akinola, Yashraj Narang, Linxi Fan, Yuke Zhu, and Dieter Fox. Mimicgen: A data generation system for scalable robot learning using human demonstrations. arXiv preprint arXiv:2310.17596, 2023. +[32] Earl K Miller and Jonathan D Cohen. An integrative theory of prefrontal cortex function. Annual review of neuroscience, 24(1):167-202, 2001. +[33] Ashvin V Nair, Vitchyr Pong, Murtaza Dalal, Shikhar Bahl, Steven Lin, and Sergey Levine. Visual reinforcement learning with imagined goals. Advances in neural information processing systems, 31, 2018. +[34] Suraj Nair, Aravind Rajeswaran, Vikash Kumar, Chelsea Finn, and Abhinav Gupta. R3m: A universal visual representation for robot manipulation. arXiv preprint arXiv:2203.12601, 2022. + +[35] Soroush Nasiriany, Vitchyr Pong, Steven Lin, and Sergey Levine. Planning with goal-conditioned policies. Advances in neural information processing systems, 32, 2019. +[36] Randall C O'Reilly and Michael J Frank. Making working memory work: a computational model of learning in the prefrontal cortex and basal ganglia. *Neural computation*, 18(2):283-328, 2006. +[37] Giovanni Pezzulo and Paul Cisek. Navigating the affordance landscape: feedback control as a process model of behavior and cognition. Trends in cognitive sciences, 20(6):414-424, 2016. +[38] Dean A Pomerleau. Alvinn: An autonomous land vehicle in a neural network. Advances in neural information processing systems, 1, 1988. +[39] Vitchyr Pong, Shixiang Gu, Murtaza Dalal, and Sergey Levine. Temporal difference models: Model-free deep rl for model-based control. arXiv preprint arXiv:1802.09081, 2018. +[40] Hanxiang Ren, Li Sun, Xulong Wang, Pei Zhou, Zewen Wu, Siyan Dong, Difan Zou, Youyi Zheng, and Yanchao Yang. Hypogen: Optimization-biased hypernetworks for generalizable policy generation. In The Thirteenth International Conference on Learning Representations. +[41] Sahand Rezaei-Shoshtari, Charlotte Morissette, Francois R Hogan, Gregory Dudek, and David Meger. Hypernetworks for zero-shot transfer in reinforcement learning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 37, pages 9579-9587, 2023. +[42] Stéphane Ross, Geoffrey Gordon, and Drew Bagnell. A reduction of imitation learning and structured prediction to no-regret online learning. In Proceedings of the fourteenth international conference on artificial intelligence and statistics, pages 627-635. JMLR Workshop and Conference Proceedings, 2011. +[43] Cicero Nogueira dos Santos, Youssef Mroueh, Inkit Padhi, and Pierre Dognin. Learning implicit generative models by matching perceptual features. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4461-4470, 2019. +[44] Tom Schaul, Daniel Horgan, Karol Gregor, and David Silver. Universal value function approximators. In International conference on machine learning, pages 1312-1320. PMLR, 2015. +[45] Simon Schug, Seijin Kobayashi, Yassir Akram, João Sacramento, and Razvan Pascanu. Attention as a hypernetwork. arXiv preprint arXiv:2406.05816, 2024. +[46] Daniel Seita, Pete Florence, Jonathan Tompson, Erwin Coumans, Vikas Sindhwani, Ken Goldberg, and Andy Zeng. Learning to rearrange deformable cables, fabrics, and bags with goal-conditioned transporter networks. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 4568-4575. IEEE, 2021. +[47] Nur Muhammad Shafiullah, Zichen Cui, Ariuntuya Arty Altanzaya, and Lerrel Pinto. Behavior transformers: Cloning $k$ modes with one stone. Advances in neural information processing systems, 35:22955-22968, 2022. +[48] Ajay Sridhar, Dhruv Shah, Catherine Glossop, and Sergey Levine. Nomad: Goal masked diffusion policies for navigation and exploration. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 63-70. IEEE, 2024. +[49] Richard S Sutton. Reinforcement learning: An introduction. A Bradford Book, 2018. +[50] Octo Model Team, Dibya Ghosh, Homer Walke, Karl Pertsch, Kevin Black, Oier Mees, Sudeep Dasari, Joel Hejna, Tobias Kreiman, Charles Xu, et al. Octo: An open-source generalist robot policy. arXiv preprint arXiv:2405.12213, 2024. +[51] I Toni. The cerebellum and parietal cortex play a specific role in coordination: a pet study. Neuroimage, 14(4):899-911, 2001. + +[52] Quan Vuong, Sergey Levine, Homer Rich Walke, Karl Pertsch, Anikait Singh, Ria Doshi, Charles Xu, Jianlan Luo, Liam Tan, Dhruv Shah, et al. Open x-embodiment: Robotic learning datasets and rt-x models. In Towards Generalist Robots: Learning Paradigms for Scalable Skill Acquisition@ CoRL2023, 2023. +[53] Chen Wang, Linxi Fan, Jiankai Sun, Ruohan Zhang, Li Fei-Fei, Danfei Xu, Yuke Zhu, and Anima Anandkumar. Mimicplay: Long-horizon imitation learning by watching human play. arXiv preprint arXiv:2302.12422, 2023. +[54] Huilin Xu, Jian Ding, Jiakun Xu, Ruixiang Wang, Jun Chen, Jinjie Mai, Yanwei Fu, Bernard Ghanem, Feng Xu, and Mohamed Elhoseiny. Diffusion-based imaginative coordination for bimanual manipulation. arXiv preprint arXiv:2507.11296, 2025. +[55] Hongxiang Yu, Anzhe Chen, Kechun Xu, Zhongxiang Zhou, Wei Jing, Yue Wang, and Rong Xiong. A hyper-network based end-to-end visual servoing with arbitrary desired poses. IEEE Robotics and Automation Letters, 8(8):4769-4776, 2023. +[56] Tianhe Yu, Deirdre Quillen, Zhanpeng He, Ryan Julian, Karol Hausman, Chelsea Finn, and Sergey Levine. Meta-world: A benchmark and evaluation for multi-task and meta reinforcement learning. In Conference on robot learning, pages 1094-1100. PMLR, 2020. +[57] Pei Zhou, Ruizhe Liu, Qian Luo, Fan Wang, Yibing Song, and Yanchao Yang. Autocgp: Closed-loop concept-guided policies from unlabeled demonstrations. In The Thirteenth International Conference on Learning Representations. +[58] Yuke Zhu, Josiah Wong, Ajay Mandlekar, Roberto Martín-Martín, Abhishek Joshi, Soroush Nasiriany, and Yifeng Zhu. robosuite: A modular simulation framework and benchmark for robot learning. arXiv preprint arXiv:2009.12293, 2020. + +# Appendix + +In this section, we present supplementary materials detailing the methodological framework and experimental procedures used in this study. + +# A Manipulation Task Details + +Task Details. Our experimental evaluation was conducted within the Robosuite [58] simulation environment, utilizing the benchmark dataset from Mimicgen [31]. We primarily investigated complex tabletop manipulation tasks that encompass diverse robotic skills. The selected tasks are characterized as follows: + +- Coffee: A multi-step manipulation task requiring precise object handling, where the robot must grasp a coffee capsule, insert it into the designated slot of the coffee machine, and securely close the machine's lid. +- Mug cleanup: A sequential task involving both articulated object interaction and object placement. The robot must coordinate drawer manipulation and object transportation, culminating in storage of a mug. +- Three piece assembly: A multi-step assembly task demanding spatial reasoning and precise manipulation. The robot must stack three components in a specific sequence to achieve compact assembly configuration. +- Threading: A high-precision manipulation task requiring fine motor control. The robot must accurately orient and manipulate a needle for successful insertion through a minimal aperture. +- Nut Assembly: A manipulation task involving precise grip control and spatial alignment for successful mechanical assembly. In this task, the success rates for the two nuts are measured separately, and the overall success rate is then calculated. +- Coffee Preparation: An extended sequential task combining multiple sub-goals, including cup positioning, drawer manipulation, capsule retrieval, and coffee maker operation, culminating in a fully prepared coffee setup. +- Kitchen: A complex sequence involving appliance interaction, object manipulation, and spatial reasoning. The task includes stove operation, cookware handling, and precise object placement. + +These tasks are specifically selected for their comprehensive representation of challenging robotic manipulation scenarios, featuring contact-rich interactions, precise object manipulation, and complex multi-step sequences. Each task requires a combination of skills including spatial reasoning, and sequential decision-making. The visual representation of these manipulation tasks and their key phases are illustrated in Figure 4. + +Data Processing and Observation Space. The demonstration data from Mimicgen is initially preprocessed by segmenting complete demonstrations into trajectory subsequences to facilitate learning. Our framework implements a constrained observation context with a length of 2, including front-view RGB images and the agent's proprioceptive state information. This design choice means that the agent's decision-making process is based solely on the current frame and one historical frame, deliberately limiting the temporal horizon to enhance real-world applicability. Given our focus on goal-conditioned policy learning, the observation space of the hypernetwork is augmented with a single RGB goal image representing a feasible target state. The input modalities are structured as follows: + +- Visual observations: RGB images with dimensions $128 \times 128$ pixels for both contextual and goal representations. +- Proprioceptive state: A compact 9-dimensional vector encoding essential agent state information. + +This deliberately constrained observation space creates a partially observable environment that closely aligns with real-world robotics scenarios, where complete state information is rarely available. While + +![](images/e3a25c36511dcf5739784efef172de4e5d31044a579e52aa0bdd3ef8eba2a34a.jpg) +(a) Coffee d0 + +![](images/a4a422b1bda32b1879470f198b9b2fcead7a735726b6794ba8aefa443fd912b4.jpg) +(b) Coffee d1 + +![](images/8a1ba3b3cfa9933175fa2872be617a119a2b4ee55785fc915d171920b5a49378.jpg) +(c) Coffee d2 + +![](images/3704690c4543f86fc76e4ff2a5f73296ddcafa48eec1b4794fd97d84c7f28fdc.jpg) +(d) Mug cleanup d0 + +![](images/63e063709c51e2f3a93f4eb9bcf743b89e172dff23193060fc18b3cd3e3453ab.jpg) +(e) Mug cleanup d1 + +![](images/b1da380a04ee92cd26fd0328b9704be79c44a4f2d23cfacdc63ef2c360600fbd.jpg) +(f) Three piece assembly d0 + +![](images/b1f23efcbf9ddd47703e020d67aeefad61e1673118f78996470e50a9aadb6270.jpg) +(g) Three piece assembly d1 + +![](images/ad5881189b416147b35501ae7ef1731a057fcd4aba259c782f9f9474be978e83.jpg) +(h) Three piece assembly d2 + +![](images/9aa489870c71d81e8e41b0c0824db68ccd4c858b75660cfd67ef886eb3d820b0.jpg) +(i) Threading d0 + +![](images/039da30fcdbf13b0d6f5e11ddeb9a200fda6568433d33122d33e2315d3d9e5c4.jpg) +(j) Threading d1 + +![](images/b63b5ba49906d9b1fe96caf22681bb5107a99c8b8c8116d68d4d10e9d28e0839.jpg) +(k) Threading d2 + +![](images/77cf41b592e01c7b83de844d670164cae3da423413ae03a99dbab9714e6c704b.jpg) +(1) Nut assembly d0 + +![](images/688db14d28039c16a50c9d6a34397c66e5f97c77a7c9536877f5920815596228.jpg) +(m) Coffee preparation d0 +Figure 4: Key phases of diverse manipulation tasks in experimental evaluation. + +![](images/6427e29a199ea401aaa82c37d69754b58a3d03241c9dbad759fa7ea211112889.jpg) +(n) Coffee preparation d1 + +![](images/8bddef46f43a727f782c02f873aabcdde89b4fff195496349cddafccae1eb1a9.jpg) +(o) Kitchen d0 + +![](images/0c5be42897656f7dacb46da3a705d020836c6de1d5c0ecf0769ad5a9cb353a54.jpg) +(p) Kitchen d1 + +this design choice enhances the practical applicability of our approach, it also introduces significant challenges: + +- Single-goal scenarios necessitate hypernetwork architectures to efficiently handle the demanding requirements of goal-conditioned policy generation. +- Limited temporal context requiring efficient use of historical information. +- Partial observability demanding robust state estimation and feature extraction. +- Complex vision-based reasoning with constrained visual information. + +Such challenging conditions serve to validate our method's effectiveness under realistic constraints, demonstrating its potential for real-world deployment. + +Goal Specification Strategy. Our approach implements a systematic strategy for goal specification across both training and evaluation phases. During training, we employ a dynamic goal sampling mechanism where the goal image is stochastically selected from future frames within the same demonstration, subsequent to the current timestep. This design offers two key advantages: + +- Goal Feasibility: By sampling from actual demonstration frames, we inherently guarantee the physical feasibility and reachability of the specified goals. +- Goal Diversity: The random sampling mechanism ensures sufficient variation in goal states, promoting the learning of a robust and generalizable policy. + +During the evaluation phase, the goal specification mechanism leverages the Mimicgen framework to generate feasible goal states, facilitating potential transfer from simulation to physical systems. + +# B Comparison with Baselines + +Given the challenging nature of our experimental setting as shown in Table 1, existing baseline methods demonstrate limited success in task completion. To ensure comprehensive evaluation, we introduce modified versions of existing approaches and additional baseline methods adapted for our scenario. The following section details these enhanced baseline implementations and their comparative performance. + +Table 7: Additional quantitative experiments with more baseline methods across different tasks. $\dagger$ indicates methods using a sequence of visual frames as goals rather than a single goal image, with an extended observation length of 10 frames versus the standard 2. $\ddagger$ indicates methods with access to extra wrist-mounted camera images. The wrist views, extended observation sequence, and extended goal sequences provide richer observational and task-guidance information for augmented baselines. Hyper-GoalNet leverages only one goal image and two front-view observations, indicating a more easy-to-be-applied setup while being more effective. + +
MethodCoffee ↑Mug-cleanup ↑Three piece Assemb. ↑Threading ↑Nut Assemb. ↑
d0d1d2Avg.d0d1Avg.d0d1d2Avg.d0d1d2Avg.
MimicPlay-O†‡0.800.840.880.840.680.580.630.500.380.020.300.320.060.060.15
MimicPlay-M†0.280.280.160.240.260.060.160.060.060.000.040.180.020.000.07
C-BeT0.920.000.740.550.300.500.400.000.020.000.010.620.220.120.32
Hyper-GoalNet(G)0.940.600.620.720.720.540.630.460.220.020.230.780.200.180.39
Hyper-GoalNet0.940.760.620.770.780.460.620.520.200.040.250.820.320.240.46
+ +We also conducted additional comparative experiments with BeT [47], C-BeT [9], and MimicPlay [53] (implemented in both its original setting and a modified setting). And a variant of our method, HyperGoalNet(G) is also introduced. Experiments were performed across all 16 MimicGen tasks, with results for contact-rich tasks and long-horizon tasks presented in Tables 7 and 8, respectively. Please note that the success rates reported in Table 8 reflect a modified evaluation criterion in the simulation environment, resulting in slight variations from the kitchen task results presented in Table 1. The implementation details are explained below. + +Details about the Baselines. We selected four task-specific baselines and reimplemented them under settings generally consistent with Hyper-GoalNet. The reimplementation details are as follows. + +- GCBC [29, 15]: Goal-Conditioned Behavioral Cloning (GCBC) is the most general framework for learning goal-conditioned policies. It consists of a perception module, a visual encoder, and a RNN-based goal-conditioned policy module. GCBC takes in a 9-dimensional proprioceptive state, a current front-view RGB image, and a goal RGB image as input and predicts the action distribution to transfer the current state to the goal state. The model is trained end-to-end with the objective of maximizing the log-likelihood of the ground-truth action in the predicted distribution. The observation sequence length and the predicted action sequence length are both restricted to 5 steps. +- Play-LMP [29]: Play-Supervised Latent Motor Plans (Play-LMP) builds upon the foundation of GCBC, aiming to learn reusable plan representations and task-agnostic control from play data. Play-LMP consists of three main components: 1) Plan recognition module: maps the input sequence to a distribution in the latent plan space. 2) Plan proposal module: generates multiple conditional prior solutions based on the current and goal states. 3) Plan and goal-conditioned policy: predicts actions conditioned on the current state, goal state, and a latent plan sampled from the plan proposals. Similar to GCBC, both the observation sequence length and the predicted action sequence length are restricted to 5 steps. The model is trained end-to-end. +- MimicPlay [53]: MimicPlay employs a hierarchical learning framework consisting of two training stages. In the high-level training stage, the model takes the robot's end-effector pose, along with the current visual observation and goal observation, as input to predict the future pose trajectory of the robot's end-effector. This component is referred to as the high-level planner. In the low-level training stage, the high-level planner with the best validation performance from the previous stage is loaded and its parameters are frozen. The model then continues training using a 9-dimensional robot proprioceptive state and visual observations (both current and goal) as input to predict the robot's actions. + +Please note that in the original MimicPlay low-level training setup, in addition to the current front-view RGB image, a wrist-mounted RGB image is also used as input, which may contribute to its higher success rate. To ensure a fair comparison with our method, we modify this setup by replacing the wrist-mounted image with a duplicate front-view image + +Table 8: Additional evaluation on long-horizon tasks. $\dagger$ indicates methods using a sequence of visual frames as goals rather than a single goal image, with an extended observation length of 10 frames versus the standard 2. $\ddagger$ indicates methods with access to extra wrist-mounted camera images. The wrist views, extended observation sequence, and extended goal sequences provide richer observational and task-guidance information for augmented baselines. Ours achieves similar performance compared to the baseline with access to wrist view images, extended observation sequences and a sequence of goal images, which demonstrates the effectiveness of our method in achieving long-horizon tasks with much less guidance information (effort). + +
MethodCoffee Prep.↑Kitchen↑Avg.
d0d1Avg.d0d1Avg.
MimicPlay-O†‡0.860.680.771.000.700.850.81
MimicPlay-M†0.340.000.170.860.180.520.35
C-BeT0.820.040.430.780.700.740.59
Hyper-GoalNet(G)0.800.500.651.000.880.940.80
Hyper-GoalNet0.800.500.651.000.800.900.78
+ +Table 9: Comparison with augmented BeT. † represents extended context length and additional wrist-view images. Although BeT is not originally designed as a goal-conditioned policy method, its augmented version serves as a strong baseline. Our method still outperforms the augmented BeT. + +
MethodCof. d0 ↑Cof. d2 ↑Mug. d1 ↑Avg.
BeT†0.660.420.260.45
Ours0.940.620.460.67
+ +during the low-level training process. Additionally, we adopt the same goal-specified strategy as described above, rather than providing the entire prompt video as used in MimicPlay's original test-time evaluation setting. + +Hyper-GoalNet(G) We introduce another variant of our method, Hyper-GoalNet(G), where the hypernetwork backbone takes only a single goal image as input, without requiring the current image during both training and testing phases. All other settings remain unchanged. During inference, Hyper-GoalNet(G) generates the weights for the lightweight target policy only once and maintains them fixed during rollouts, resulting in improved computational efficiency. As shown in Table 7 and Table 8, Hyper-GoalNet(G) still outperforms the baselines while utilizing a much smaller lightweight policy network, demonstrating both the effectiveness and efficiency of our approach. + +Comparison with BeT [47]. We conducted comparative experiments with Behavior Transformer (BeT), a state-of-the-art approach for multi-modal behavioral learning. BeT employs k-means clustering to discretize continuous actions and utilizes transformers to model categorical distributions across action bins, incorporating an action correction head to refine discretized actions into continuous ones. Despite not being explicitly designed as a goal-conditioned policy, BeT has emerged as a robust baseline in current robot learning literature. Given that the original BeT implementation for the Franka Kitchen task was limited to state space observations and lacked compatibility with Robosuite tasks, we enhanced its architecture by incorporating a pretrained image encoder [34]. To strengthen the baseline comparison, we augmented BeT with additional wrist camera observations—a feature absent in our method—and extended the context length from 2 (used in our approach) to 4, which typically facilitates more effective policy learning for BeT. Consequently, BeT acquires more image observations per timestep than our method, while maintaining the same image resolution of $128 \times 128$ pixels. + +Table 9 presents experimental results for some randomly selected tasks, where Cof. d0 and Cof. d2 represent Coffee d0 and Coffee d2 tasks, respectively, and Mug. d1 denotes the Mug cleanup d1 task. Notably, despite utilizing reduced contextual information, our method demonstrates superior performance across all tasks in terms of success rates. These results substantiate our method's robust capability in behavior learning, even under more constrained observational conditions. + +Comparison with C-BeT [9]. Conditional Behavior Transformer (C-BeT) is a goal-conditioned version of BeT, a behavior prediction model that compresses an agent's observation history and the goal state into a compact latent representation using self-attention, which is then transformed along with discrete action representations to efficiently predict the agent's future behaviors. To ensure fair comparison, we configured C-BeT with identical observation settings to our method, maintaining a context length of 2 and utilizing a single goal image. The comparative results are presented in Table 7 and Table 8. + +Comparison with Original MimicPlay (MimicPlay-O). In the original MimicPlay experiment setting, we retained the wrist image as an input during the low-level training stage. For the test-time evaluation process, we provided the pretrained model with prompt videos in HDF5 format generated by MimicGen. In contrast to our approach, MimicPlay-O has access to wrist-mounted camera images, uses an extended observation length of 10 frames instead of 2, and utilizes a sequence of visual frames as goals, rather than a single goal image configuration. + +Comparison with another Modified MimicPlay (MimicPlay-M). In this experimental setting, we removed the wrist image as an input during the low-level training stage, since the wrist image is not easy to obtain for goal specification. For the test-time evaluation process, we provided the pretrained model with prompt videos in HDF5 format generated by MimicGen. Please note that this experimental setup slightly differs from the one in our baseline setting. In contrast to our approach, MimicPlay-M uses an extended observation length of 10 frames instead of 2, and utilizes a sequence of visual frames as goals, rather than a single goal image configuration. + +Efficiency Analysis. We evaluate the computational efficiency by measuring the average inference time per action step during deployment. Table 10 presents the average inference latency per step across different methods, measured over 40,000 steps on a single NVIDIA RTX 3090 GPU. Our proposed Hyper-GoalNet(G) demonstrates superior computational efficiency while maintaining state-of-the-art performance. This efficiency stems from our novel approach of dynamically generating weights for a lightweight target policy. Specifically, Hyper-GoalNet(G) generates a suitable set of policy weights at the beginning of each rollout based on the goal image. These weights remain fixed throughout the execution, eliminating the need for repeated weight generation and thus significantly reducing the computational overhead during deployment. + +Table 10: Average Inference Time Per Step + +
MethodGCBCPlay-LMPC-BeTHyper-GoalNetHyper-GoalNet(G)
Time (ms)15.4722.7813.616.331.46
+ +# C Comparison with Other Hypernetworks + +This section provides additional details on the comparison between our proposed hypernetwork architecture and HyperZero, particularly regarding computational requirements. + +# C.1 Training Time & Memory Requirements + +To ensure a fair and direct comparison, all experiments were conducted on a single NVIDIA RTX 4090 GPU. The training hyperparameters, including batch size, learning rate, and optimizer settings, were kept identical for both our method and the HyperZero baseline. The only modification was the hypernetwork architecture itself. This controlled setup ensures that any observed differences in resource consumption are directly attributable to the design of the hypernetwork module. + +As detailed in Table 11, our method requires slightly more resources than HyperZero in terms of per-epoch training time and memory usage. However, this modest computational overhead is coupled with the substantial performance gains documented in the main paper, highlighting the efficiency and effectiveness of our architectural design. + +Table 11: Computational resource comparison. The table shows per-epoch training time and memory footprint for our method versus HyperZero under identical hyperparameter settings. "Frozen" and "Unfrozen" refer to the state of the visual encoder. + +
MetricHyperZeroOurs
Training Time / Epoch~90s~104s
Memory (Frozen Encoder)3,038 MB4,916 MB
Memory (Unfrozen Encoder)13,452 MB14,844 MB
+ +# D Policy Learning Details + +Overview. A conventional sequential decision-making problem can be formalized as a discrete-time finite Markov decision process (MDP) defined by a 7-tuple $M = (\mathcal{O}, \mathcal{A}, \mathcal{P}, r, \rho_0, \gamma, H)$ , where: + +- $\mathcal{O}$ denotes the observation space, +- $\mathcal{A}$ represents the action space, +- $\mathcal{P}:\mathcal{O}\times \mathcal{A}\times \mathcal{O}\to \mathbb{R}_{+}$ defines the transition probability distribution, +- $\gamma \in [0,1]$ is the discount factor, +- $H$ specifies the temporal horizon of the process. + +In the context of imitation learning, we define a complete state-action trajectory as $\tau = (o_0, a_0, \dots, o_t, a_t)$ , where the initial state is sampled as $o_0 \sim \rho_0(o_0)$ , actions are generated by the policy $a_t \sim \pi_\theta(\cdot | o_t)$ , and state transitions follow $o_{t+1} \sim \mathcal{P}(\cdot | o_t, a_t)$ . + +Traditionally, the objective in goal-conditioned decision-making problems is to identify an optimal goal-conditioned policy $\pi_{\theta}$ that maximizes the expected discounted reward: + +$$ +\eta \left(\pi_ {\theta}\right) = \mathbb {E} _ {\tau} \left[ \sum_ {t = 0} ^ {T} \gamma^ {t} r \left(o _ {t}, a _ {t}, o _ {t + 1} \mid o _ {g}\right) \right] \tag {11} +$$ + +However, our approach diverges from this conventional framework in several key aspects: Reward-Free Learning: Operating within a behavior cloning paradigm, we lack access to explicit reward signals. Instead, we aim to learn a goal-specific policy $\pi_{\theta}$ that maps states to optimal actions purely from demonstrations. Goal-Specific Policy Generation: Rather than learning a universal goal-conditioned policy, our hypernetwork architecture generates specialized policies for specific goals, conditioned on the current RGB observation and a target goal image. Non-Markovian Extension: We relax the Markovian assumption to incorporate temporal dependencies. The resulting policy formulation becomes: + +$$ +\pi_ {\theta} \left(a _ {t} \mid o _ {t - 1}, o _ {t}\right) \tag {12} +$$ + +This extended formulation enables the policy to leverage information from a context window of length 2, enhancing its capacity to handle complex, temporally-dependent manipulation sequences. + +Model Architecture. Our architectural design addresses the challenges of visuomotor manipulation tasks, which require processing of high-dimensional visual inputs rather than simple state-based representations. The architecture comprises several key components integrated to handle visual and proprioceptive information effectively. Visual Processing Pipeline: To bridge the gap between high-dimensional visual inputs and hypernetwork processing capabilities, we employ a pre-trained visual encoder [34] to compress RGB images into compact latent representations. This encoder's training follows a two-phase strategy: Initial phase (first 20 epochs): Parameters remain frozen to establish stable feature representations; Fine-tuning phase: Parameters become trainable to optimize task-specific visual features. + +Hypernetwork Configuration. Our hypernetwork uses the HyPoGen architecture [40] with 8 optimization blocks. It processes encoded current and goal images to generate the parameters for a 3-layer MLP target policy, an architecture that can be flexibly extended in depth and width. This + +![](images/74355f669f426b299a4084821831816580a444b21465c62baf775509e710bddc.jpg) +Figure 5: Normalized distance between current and goal states computed with different latent spaces along policy rollouts. + +![](images/c4727fbd7b96c3c26ac2b9df717e4c811ceea46cf2c5b3b600be385f119ea1c0.jpg) +(a)coffeed2 + +![](images/d61ec01985eb0f3b65ac39503eb0b20bb99e603619eb4569fb901775ba7cd0ad.jpg) +(b) coffee preparation d1 +Figure 6: Distance to the goal in the latent space along policy rollouts across different manipulation tasks. The consistent decreasing trend across diverse tasks demonstrates that our learned latent representations effectively capture the physical progress toward task goals, establishing meaningful correspondence between latent-space distances and real-world task completion. + +![](images/2c53988cc4e567e6e0d37ab77b4cac762d7a324a1f39b40571114b5f76ae8f43.jpg) +(c) threading d1 + +meta-learning approach enables dynamic policy adaptation based on specified goals while maintaining computational efficiency. Beside the image encoder mentioned above, the action generation incorporates several specialized components: Predictive Model: Implemented as an MLP operating in the compressed latent space, leveraging the reduced dimensionality for efficient dynamics modeling, Proprioceptive Encoder: A compact MLP processes low-dimensional proprioceptive states, providing essential agent state information, Feature Integration: Temporal image features are concatenated with proprioceptive information at each timestep, Target Policy: A lightweight MLP processes the integrated features to generate appropriate control actions. This architecture efficiently handles the complexity of visuomotor tasks while maintaining computational tractability through dimensionality reduction and feature integration. + +**Training Details.** Our framework implements an end-to-end training paradigm with controlled experimental conditions for reproducibility. Using a fixed random seed, we partition the dataset into 950 training and 50 validation demonstrations across all tasks. Training employs a batch size of 256 and the Adam optimizer with an initial learning rate of $5 \times 10^{-4}$ , coupled with cosine annealing for learning rate decay. The model trains for 500 epochs without weight decay or dropout regularization in the hypernetwork component. The training and evaluation procedures were performed on a single NVIDIA GeForce RTX 3090 or RTX 4090 GPU. To ensure a fair comparison, all methods evaluated in our experiments were trained using this identical configuration. + +![](images/413b839c82d645bedfc544962cc3f4f3dcf78b91daf44ccdeeb1ead10d3360da.jpg) +Figure 7: The real robot workspace. + +# E Ablation Studies + +We conducted ablation experiments to evaluate the impact of various methodological choices, with quantitative results presented in Table 3. Figure 6 and Figure 5 illustrates the normalized distance between current and goal states under different experimental configurations. Our analysis includes several key variations: + +- uf. at epoch 0. Full parameter unfreezing from epoch 0, where all model components are trainable from initialization. +- w/o shapping. Removal of latent shaping technique detailed in Section 3.2. +- dist $\leftrightarrow$ start img. Alternative distance computation between current and start images, rather than current and goal images. +- cos. dist. Implementation of cosine distance metric in place of Euclidean distance. +- C-Bet(w/ shaping). We implement C-Bet and incorporate our proposed additional shaping method while maintaining the same experimental setup. + +These systematic variations enable us to quantify the contribution of each design choice to the overall system performance. + +# F Real Robot Experiment Setting + +Real Robot Platform. All real-world experiments were conducted on a RealMan RMC-DA dual-arm manipulator, which comprises two 7-degree-of-freedom robotic arms, each rated for a $5\mathrm{kg}$ payload and fitted with a parallel-jaw gripper. While the platform supports bimanual manipulation, this work focuses specifically on evaluating Hyper-GoalNet's capabilities in single-arm tabletop manipulation tasks. The extension to bimanual manipulation remains as future work. The robot is mounted in front of a $1.25\mathrm{m} \times 0.75\mathrm{m}$ tabletop, which serves as the exclusive workspace for all manipulation tasks. A standardized set of test objects—ranging from simple geometric primitives (e.g., cubes) to more complex shapes—is placed on the table according to predefined configurations. And on the tabletop there might be some distractor object. To support perception, we employ an overhead RGB-D sensor (Intel RealSense D435i). The entire workspace is shown in Figure 7 + +Task Details. To validate the versatility and robustness of our method across a broad spectrum of manipulation skills, we selected four representative tabletop tasks. Each task emphasizes a different + +![](images/cce29fc0ee88bc923c4a4e2af2df4fd0afbd127f8c7698c79b4fefff9f159544.jpg) +(a) Stacking + +![](images/841e23ad5102c678357cac1fc1e141f48a989de5cd34ff75c253852999f2d12d.jpg) + +![](images/2a83e76bf8961b55f7c97eb609f691bf74fe3cdf9def784a07d68ce2295505ab.jpg) +(b) Pick-and-Place + +![](images/b04c88a5a9ef17794b048ba1a177416823d632e0bfa497693b49ac0d632de464.jpg) + +![](images/27a690877e2970d1d8ff782c8a4052ad4b624940fcea4af92d39f76b55da183d.jpg) +(c)Drawer Pulling + +![](images/2ff01acf7ce678e7f4a24c5fedaeed347dee48a455897d0c7c16cf4fc574b8cc.jpg) +Figure 8: Key phases of diverse manipulation tasks in real robot experimental evaluation. + +![](images/161143fecc8bfc530fa3b3974a96afd7e1b89498b764ab10dd7ab51513a3b14e.jpg) +(d) Sweeping + +![](images/eb1c2cc15b863d2b41c2e8abff15b815f4fc48ada7d09d9613bf360fd192df8d.jpg) + +core competency—object localization, precision grasping, and surface contact manipulation—and is defined as follows: + +- Pick-and-Place. The robot must perceive and localize a specified cubic object within the workspace, plan a collision-free trajectory, execute a stable grasp, and transport the object to a predefined target location (e.g., a plate). Success is measured by the accuracy of the final placement and the repeatability across trials. +- Stacking. Extending the pick-and-place paradigm, this task requires the robot to grasp a source cube, position it directly above a target cube resting on the tabletop, lower it until gentle contact is detected via proximity or vision-based cues, and then release to complete the stack. Success is defined by the source cube being neatly aligned atop the target cube. +- **Drawer Pulling.** The robot must detect drawer handle, plan an approach to engage the handle with its gripper, and execute a controlled pulling maneuver to extend the drawer along its linear guide. Performance is evaluated by the final extension distance achieved without stalling. +- Sweeping. The end-effector is equipped with a broom attachment. The robot must locate and gather a target object scattered on the tabletop, then sweep it into a designated collection zone (e.g., a dustpan). Success is defined by the target object being fully contained within the collection zone at the end of each trial. + +Data Processing and Observation Space. Our real-robot evaluation follows the same protocol as in simulation: at each timestep, the policy receives the two most recent observations and a single goal image. We collect approximately 70-100 human teleoperation trajectories per task for training. Vision is acquired with an Intel RealSense D435i depth camera; we concatenate its depth channel with the RGB channels to form 4-channel images of resolution $128 \times 128$ , which are normalized to $[0,1]$ before input to the encoder. Since explicit end-effector poses are unavailable, we represent the current proprioceptive state by the previous action—comprising seven joint angles and one gripper command—resulting in an 8-dimensional vector. All modalities are synchronized identically to the simulation setting, ensuring a seamless transfer between simulated and real-world experiments. + +# G Additional Visualization + +We provide additional visualization of the shaping. Figure 9 and Figure 10 demonstrate the approximate monotonic trend of the distance between the current state and the goal state is consistent across different task scenarios. This consistent pattern substantiates the robustness of our shaping mechanism and validates its task-agnostic applicability. + +![](images/5939fe90b471e0dcbdd24d756426ba53a4ba719fbbc480d2c197a9f475e03878.jpg) +(a) Coffee dO + +![](images/29643c026491b791d409c41fbe546850c262855253e9a461dc0567a89e2f86c9.jpg) +(b) Coffee d1 + +![](images/37bd445ef8cdcbe6c6af78db8f666877ec5b1ee96d6a9da2df2e7893d3e7996a.jpg) +(c) Coffee d2 + +![](images/8c22d2ad70aeea7cb88ca0dda70d6531f7bce4c8c2a0b4f01c83c7acbb282cd3.jpg) +(d) Mug cleanup d0 + +![](images/dee66f9885b1ae1566ef339ab0bbb4a39e6c37cedb7056a4443a5b0cf55a14e2.jpg) +(e) Mug cleanup d1 + +![](images/698acfe1d47e3b58eae09b7b609e9691baa5eb68dcacf16ffca4b6984f390d13.jpg) +(f) Three piece assembly d0 + +![](images/3e58ac696575fdfdd58ac2bb6b64f8b409c8e63f1e9c0466d9eb13481e9675cb.jpg) +(g) Three piece assembly d1 + +![](images/2142eef57dc58b04d6efa9863b4825c629dc846f98065751adf1c793868aa9b2.jpg) +(h) Three piece assembly d2 +Figure 9: Additional visualization: Distance to the goal in the latent space along policy rollouts across different manipulation tasks. The consistent decreasing trend across diverse tasks demonstrates that our learned latent representations effectively capture the physical progress toward task goals, establishing meaningful correspondence between latent-space distances and real-world task completion. + +# H Future Work + +Building upon our current findings, we identify three primary directions for future research: First, we aim to incorporate multi-goal reasoning to handle complex sequential tasks. Second, we plan to extend our framework by integrating foundation models to develop a more generalized policy generation framework, potentially enabling broader task generalization and enhanced adaptability + +![](images/1966a27de7835f9e8f384b27021c58caa225b8bdb6184d5b0c91cfdc0a8bad97.jpg) +Initial State + +![](images/5f9e8a1f5c817215a96453536bbae3afd3163b15cfaf8921ad8e99c0001efd54.jpg) +Goal State + +![](images/31f0aeef56decccb3492ab9f6071b0c252848a7a3d9edb576c96c5017b40c836.jpg) +Rollout by Generated Policy +(i) Threading d0 + +![](images/494e3d91a9034ac6c33694b8cd812ea81ee98300dc6a1068848305c1d1d151f3.jpg) +Initial State + +![](images/29136d4d3e428bb8ad5b57053e7a7a60f8101e292e823a919044217afb4f9397.jpg) +Goal State + +![](images/d0b4a00f7b09c5a892144bab1f5bea3773a9e9e3273fd156ba60b49c7efc896d.jpg) +Rollout by Generated Policy +(j) Threading d1 + +![](images/9ef43eb236fca7411324c74d1e8f5dd886ef47acfa46565b8b1e1725552c8cdd.jpg) +Initial State + +![](images/1ddfa90b4025ab30b7f7065011da69d41a9622c52935cf919efb1f011de0711c.jpg) +Goal State + +![](images/4a579fc0dd03871e9ea15b8579c1ef24cdf546e0d56120a0069f0c8d2ff55d4d.jpg) +Rollout by Generated Policy +(k) Threading d2 +Rollout by Generated Policy + +![](images/d59c14eb81562c0fe70af52754834ff7db88ba2c359a2349801d4898de8a7e15.jpg) + +![](images/e4bc954528dc9f36386f43b529f47e22e346acd207391ed7ea331f7a5ead5d32.jpg) +Goal State + +![](images/8bdfa3c592c4418a118e449ba22f9d35fd58831786b77ebc96d5b80fffe79319.jpg) +Rollout by Generated Policy +(1) Nut assembly d0 + +![](images/74dbaabfdd4102e3b177c202c6ac037507d0e54add9ab24ccb81fef490353fb4.jpg) +Initial State + +![](images/aa4b7bbc0120da5823d5dca5079b560c26ef3ad1afc2ac144206f366e83db3a3.jpg) +Goal State + +![](images/131844e8a5bb007d61e37c42339512014f95f02638cd3a492a3c36b986f8adf1.jpg) +(m) Coffee preparation d0 + +![](images/c1dfb3b65e74c3454270e47528d7c76edbc84d0199980d31e18485cb77ead751.jpg) + +![](images/3850e5b6a89c44605dc9a4db2a8d3e0dcb7ba35bfed68a5bddcc326aa1c8676c.jpg) +Goal State + +![](images/23cc22255b1fa2bf01c26b1320455edb1d9cad198af8077e8a725b006e3d60ba.jpg) +Rollout by Generated Policy +(n) Coffee preparation d1 + +![](images/c78ba06f880dfaed2ee771648ed22cbc5e209a1dd7425338a04f7b6c2a0f41b4.jpg) +Initial State + +![](images/f6f8610bdedfb6c06887e557c37be049fc8366a5cc07a8660e42d9dd9bf0681b.jpg) +Goal State + +![](images/bb32a9c45127dfc8f683d1d5fc079df9c80059cbd1e897af5819bf43c1c4fabb.jpg) +Rollout by Generated Policy +(o) Kitchen d0 +Figure 10: Additional visualization: Distance to the goal in the latent space along policy rollouts across different manipulation tasks. The consistent decreasing trend across diverse tasks demonstrates that our learned latent representations effectively capture the physical progress toward task goals, establishing meaningful correspondence between latent-space distances and real-world task completion. + +![](images/7e55cb517f586bce68e9a1ef0b65093ed4738c456d40ad7ef3ece8655d87845a.jpg) + +![](images/30a47f1e76b7166b4e637e04cc235e1611be7982e4ed28a0c36868f5e15374ac.jpg) +Goal State + +![](images/2a95428fd1c435b3eb520ce8546b9775dd07f459ac2eac79450be62b68607c9a.jpg) +Rollout by Generated Policy +(p) Kitchen d1 + +across diverse manipulation scenarios. Third, we plan to combine our parameter-adaptive approach with reinforcement learning to reduce dependence on demonstration data. + +# NeurIPS Paper Checklist + +The checklist is designed to encourage best practices for responsible machine learning research, addressing issues of reproducibility, transparency, research ethics, and societal impact. Do not remove the checklist: The papers not including the checklist will be desk rejected. The checklist should follow the references and follow the (optional) supplemental material. The checklist does NOT count towards the page limit. + +Please read the checklist guidelines carefully for information on how to answer these questions. For each question in the checklist: + +- You should answer [Yes], [No], or [NA]. +- [NA] means either that the question is Not Applicable for that particular paper or the relevant information is Not Available. +- Please provide a short (1–2 sentence) justification right after your answer (even for NA). + +The checklist answers are an integral part of your paper submission. They are visible to the reviewers, area chairs, senior area chairs, and ethics reviewers. You will be asked to also include it (after eventual revisions) with the final version of your paper, and its final version will be published with the paper. + +The reviewers of your paper will be asked to use the checklist as one of the factors in their evaluation. While "[Yes]" is generally preferable to "[No]", it is perfectly acceptable to answer "[No]" provided a proper justification is given (e.g., "error bars are not reported because it would be too computationally expensive" or "we were unable to find the license for the dataset we used"). In general, answering "[No]" or "[NA]" is not grounds for rejection. While the questions are phrased in a binary way, we acknowledge that the true answer is often more nuanced, so please just use your best judgment and write a justification to elaborate. All supporting evidence can appear either in the main paper or the supplemental material, provided in appendix. If you answer [Yes] to a question, in the justification please point to the section(s) where related material for the question can be found. + +IMPORTANT, please: + +- Delete this instruction block, but keep the section heading "NeurIPS Paper Checklist", +- Keep the checklist subsection headings, questions/answers and guidelines below. +- Do not modify the questions and only use the provided macros for your answers. + +# 1. Claims + +Question: Do the main claims made in the abstract and introduction accurately reflect the paper's contributions and scope? + +Answer: [Yes] + +Justification: We claim our contribution in the introduction and method section. + +Guidelines: + +- The answer NA means that the abstract and introduction do not include the claims made in the paper. +- The abstract and/or introduction should clearly state the claims made, including the contributions made in the paper and important assumptions and limitations. A No or NA answer to this question will not be perceived well by the reviewers. +- The claims made should match theoretical and experimental results, and reflect how much the results can be expected to generalize to other settings. +- It is fine to include aspirational goals as motivation as long as it is clear that these goals are not attained by the paper. + +# 2. Limitations + +Question: Does the paper discuss the limitations of the work performed by the authors? + +Answer: [Yes] + +Justification: We discuss limitations including challenges in long-term planning, further work on scaling up, and multi-goal reasoning design. + +# Guidelines: + +- The answer NA means that the paper has no limitation while the answer No means that the paper has limitations, but those are not discussed in the paper. +- The authors are encouraged to create a separate "Limitations" section in their paper. +- The paper should point out any strong assumptions and how robust the results are to violations of these assumptions (e.g., independence assumptions, noiseless settings, model well-specification, asymptotic approximations only holding locally). The authors should reflect on how these assumptions might be violated in practice and what the implications would be. +- The authors should reflect on the scope of the claims made, e.g., if the approach was only tested on a few datasets or with a few runs. In general, empirical results often depend on implicit assumptions, which should be articulated. +- The authors should reflect on the factors that influence the performance of the approach. For example, a facial recognition algorithm may perform poorly when image resolution is low or images are taken in low lighting. Or a speech-to-text system might not be used reliably to provide closed captions for online lectures because it fails to handle technical jargon. +- The authors should discuss the computational efficiency of the proposed algorithms and how they scale with dataset size. +- If applicable, the authors should discuss possible limitations of their approach to address problems of privacy and fairness. +- While the authors might fear that complete honesty about limitations might be used by reviewers as grounds for rejection, a worse outcome might be that reviewers discover limitations that aren't acknowledged in the paper. The authors should use their best judgment and recognize that individual actions in favor of transparency play an important role in developing norms that preserve the integrity of the community. Reviewers will be specifically instructed to not penalize honesty concerning limitations. + +# 3. Theory assumptions and proofs + +Question: For each theoretical result, does the paper provide the full set of assumptions and a complete (and correct) proof? + +Answer: [NA] + +Justification: This paper does not do not involve the theoretical part and the proofs. + +Guidelines: + +- The answer NA means that the paper does not include theoretical results. +- All the theorems, formulas, and proofs in the paper should be numbered and cross-referenced. +- All assumptions should be clearly stated or referenced in the statement of any theorems. +- The proofs can either appear in the main paper or the supplemental material, but if they appear in the supplemental material, the authors are encouraged to provide a short proof sketch to provide intuition. +- Inversely, any informal proof provided in the core of the paper should be complemented by formal proofs provided in appendix or supplemental material. +- Theorems and Lemmas that the proof relies upon should be properly referenced. + +# 4. Experimental result reproducibility + +Question: Does the paper fully disclose all the information needed to reproduce the main experimental results of the paper to the extent that it affects the main claims and/or conclusions of the paper (regardless of whether the code and data are provided or not)? + +Answer: [Yes] + +Justification: We include implementation details in the experiment section and training details in the appendix. + +Guidelines: + +- The answer NA means that the paper does not include experiments. + +- If the paper includes experiments, a No answer to this question will not be perceived well by the reviewers: Making the paper reproducible is important, regardless of whether the code and data are provided or not. +- If the contribution is a dataset and/or model, the authors should describe the steps taken to make their results reproducible or verifiable. +- Depending on the contribution, reproducibility can be accomplished in various ways. For example, if the contribution is a novel architecture, describing the architecture fully might suffice, or if the contribution is a specific model and empirical evaluation, it may be necessary to either make it possible for others to replicate the model with the same dataset, or provide access to the model. In general, releasing code and data is often one good way to accomplish this, but reproducibility can also be provided via detailed instructions for how to replicate the results, access to a hosted model (e.g., in the case of a large language model), releasing of a model checkpoint, or other means that are appropriate to the research performed. +- While NeurIPS does not require releasing code, the conference does require all submissions to provide some reasonable avenue for reproducibility, which may depend on the nature of the contribution. For example +(a) If the contribution is primarily a new algorithm, the paper should make it clear how to reproduce that algorithm. +(b) If the contribution is primarily a new model architecture, the paper should describe the architecture clearly and fully. +(c) If the contribution is a new model (e.g., a large language model), then there should either be a way to access this model for reproducing the results or a way to reproduce the model (e.g., with an open-source dataset or instructions for how to construct the dataset). +(d) We recognize that reproducibility may be tricky in some cases, in which case authors are welcome to describe the particular way they provide for reproducibility. In the case of closed-source models, it may be that access to the model is limited in some way (e.g., to registered users), but it should be possible for other researchers to have some path to reproducing or verifying the results. + +# 5. Open access to data and code + +Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material? + +Answer: [No] + +Justification: We don't provide open access to the code now, but the data is publicly available. We will open the source code upon acceptance. + +Guidelines: + +- The answer NA means that paper does not include experiments requiring code. +- Please see the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details. +- While we encourage the release of code and data, we understand that this might not be possible, so "No" is an acceptable answer. Papers cannot be rejected simply for not including code, unless this is central to the contribution (e.g., for a new open-source benchmark). +- The instructions should contain the exact command and environment needed to run to reproduce the results. See the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details. +- The authors should provide instructions on data access and preparation, including how to access the raw data, preprocessed data, intermediate data, and generated data, etc. +- The authors should provide scripts to reproduce all experimental results for the new proposed method and baselines. If only a subset of experiments are reproducible, they should state which ones are omitted from the script and why. +- At submission time, to preserve anonymity, the authors should release anonymized versions (if applicable). + +- Providing as much information as possible in supplemental material (appended to the paper) is recommended, but including URLs to data and code is permitted. + +# 6. Experimental setting/details + +Question: Does the paper specify all the training and test details (e.g., data splits, hyperparameters, how they were chosen, type of optimizer, etc.) necessary to understand the results? + +Answer: [Yes] + +Justification: We show the training and testing details, as well as model architecture in the appendix. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The experimental setting should be presented in the core of the paper to a level of detail that is necessary to appreciate the results and make sense of them. +- The full details can be provided either with the code, in appendix, or as supplemental material. + +# 7. Experiment statistical significance + +Question: Does the paper report error bars suitably and correctly defined or other appropriate information about the statistical significance of the experiments? + +Answer: [No] + +Justification: We do not show error bars in paper. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The authors should answer "Yes" if the results are accompanied by error bars, confidence intervals, or statistical significance tests, at least for the experiments that support the main claims of the paper. +- The factors of variability that the error bars are capturing should be clearly stated (for example, train/test split, initialization, random drawing of some parameter, or overall run with given experimental conditions). +- The method for calculating the error bars should be explained (closed form formula, call to a library function, bootstrap, etc.) +- The assumptions made should be given (e.g., Normally distributed errors). +- It should be clear whether the error bar is the standard deviation or the standard error of the mean. +- It is OK to report 1-sigma error bars, but one should state it. The authors should preferably report a 2-sigma error bar than state that they have a $96\%$ CI, if the hypothesis of Normality of errors is not verified. +- For asymmetric distributions, the authors should be careful not to show in tables or figures symmetric error bars that would yield results that are out of range (e.g. negative error rates). +- If error bars are reported in tables or plots, The authors should explain in the text how they were calculated and reference the corresponding figures or tables in the text. + +# 8. Experiments compute resources + +Question: For each experiment, does the paper provide sufficient information on the computer resources (type of compute workers, memory, time of execution) needed to reproduce the experiments? + +Answer: [Yes] + +Justification: We mention the compute resources in the appendix. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The paper should indicate the type of compute workers CPU or GPU, internal cluster, or cloud provider, including relevant memory and storage. + +- The paper should provide the amount of compute required for each of the individual experimental runs as well as estimate the total compute. +- The paper should disclose whether the full research project required more computer than the experiments reported in the paper (e.g., preliminary or failed experiments that didn't make it into the paper). + +# 9. Code of ethics + +Question: Does the research conducted in the paper conform, in every respect, with the NeurIPS Code of Ethics https://neurips.cc/public/EthicsGuidelines? + +Answer: [Yes] + +Justification: The research conducted in the paper conform, in every respect, with the NeurIPS Code of Ethics. + +Guidelines: + +- The answer NA means that the authors have not reviewed the NeurIPS Code of Ethics. +- If the authors answer No, they should explain the special circumstances that require a deviation from the Code of Ethics. +- The authors should make sure to preserve anonymity (e.g., if there is a special consideration due to laws or regulations in their jurisdiction). + +# 10. Broader impacts + +Question: Does the paper discuss both potential positive societal impacts and negative societal impacts of the work performed? + +Answer: [NA] + +Justification: There is no societal impact of the work performed. + +Guidelines: + +- The answer NA means that there is no societal impact of the work performed. +- If the authors answer NA or No, they should explain why their work has no societal impact or why the paper does not address societal impact. +- Examples of negative societal impacts include potential malicious or unintended uses (e.g., disinformation, generating fake profiles, surveillance), fairness considerations (e.g., deployment of technologies that could make decisions that unfairly impact specific groups), privacy considerations, and security considerations. +- The conference expects that many papers will be foundational research and not tied to particular applications, let alone deployments. However, if there is a direct path to any negative applications, the authors should point it out. For example, it is legitimate to point out that an improvement in the quality of generative models could be used to generate deepfakes for disinformation. On the other hand, it is not needed to point out that a generic algorithm for optimizing neural networks could enable people to train models that generate Deepfakes faster. +- The authors should consider possible harms that could arise when the technology is being used as intended and functioning correctly, harms that could arise when the technology is being used as intended but gives incorrect results, and harms following from (intentional or unintentional) misuse of the technology. +- If there are negative societal impacts, the authors could also discuss possible mitigation strategies (e.g., gated release of models, providing defenses in addition to attacks, mechanisms for monitoring misuse, mechanisms to monitor how a system learns from feedback over time, improving the efficiency and accessibility of ML). + +# 11. Safeguards + +Question: Does the paper describe safeguards that have been put in place for responsible release of data or models that have a high risk for misuse (e.g., pretrained language models, image generators, or scraped datasets)? + +Answer: [NA] + +Justification: In our manipulation tasks, we have not encountered any safeguard issues. + +Guidelines: + +- The answer NA means that the paper poses no such risks. +- Released models that have a high risk for misuse or dual-use should be released with necessary safeguards to allow for controlled use of the model, for example by requiring that users adhere to usage guidelines or restrictions to access the model or implementing safety filters. +- Datasets that have been scraped from the Internet could pose safety risks. The authors should describe how they avoided releasing unsafe images. +- We recognize that providing effective safeguards is challenging, and many papers do not require this, but we encourage authors to take this into account and make a best faith effort. + +# 12. Licenses for existing assets + +Question: Are the creators or original owners of assets (e.g., code, data, models), used in the paper, properly credited and are the license and terms of use explicitly mentioned and properly respected? + +Answer: [Yes] + +Justification: All the methods, datasets and models originated from other works are cited in reference. + +Guidelines: + +- The answer NA means that the paper does not use existing assets. +- The authors should cite the original paper that produced the code package or dataset. +- The authors should state which version of the asset is used and, if possible, include a URL. +- The name of the license (e.g., CC-BY 4.0) should be included for each asset. +- For scraped data from a particular source (e.g., website), the copyright and terms of service of that source should be provided. +- If assets are released, the license, copyright information, and terms of use in the package should be provided. For popular datasets, paperswithcode.com/datasets has curated licenses for some datasets. Their licensing guide can help determine the license of a dataset. +- For existing datasets that are re-packaged, both the original license and the license of the derived asset (if it has changed) should be provided. +- If this information is not available online, the authors are encouraged to reach out to the asset's creators. + +# 13. New assets + +Question: Are new assets introduced in the paper well documented and is the documentation provided alongside the assets? + +Answer: [NA] + +Justification: We will release relative assets upon acceptance. We welcome new creation based on our system. + +Guidelines: + +- The answer NA means that the paper does not release new assets. +- Researchers should communicate the details of the dataset/code/model as part of their submissions via structured templates. This includes details about training, license, limitations, etc. +- The paper should discuss whether and how consent was obtained from people whose asset is used. +- At submission time, remember to anonymize your assets (if applicable). You can either create an anonymized URL or include an anonymized zip file. + +# 14. Crowdsourcing and research with human subjects + +Question: For crowdsourcing experiments and research with human subjects, does the paper include the full text of instructions given to participants and screenshots, if applicable, as well as details about compensation (if any)? + +# Answer: [NA] + +Justification: Our work does not involve crowdsourcing nor research with human subjects. + +# Guidelines: + +- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. +- Including this information in the supplemental material is fine, but if the main contribution of the paper involves human subjects, then as much detail as possible should be included in the main paper. +- According to the NeurIPS Code of Ethics, workers involved in data collection, curation, or other labor should be paid at least the minimum wage in the country of the data collector. + +# 15. Institutional review board (IRB) approvals or equivalent for research with human subjects + +Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals (or an equivalent approval/review based on the requirements of your country or institution) were obtained? + +# Answer: [NA] + +Justification: Our work does not involve crowdsourcing nor research with human subjects. + +# Guidelines: + +- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. +- Depending on the country in which research is conducted, IRB approval (or equivalent) may be required for any human subjects research. If you obtained IRB approval, you should clearly state this in the paper. +- We recognize that the procedures for this may vary significantly between institutions and locations, and we expect authors to adhere to the NeurIPS Code of Ethics and the guidelines for their institution. +- For initial submissions, do not include any information that would break anonymity (if applicable), such as the institution conducting the review. + +# 16. Declaration of LLM usage + +Question: Does the paper describe the usage of LLMs if it is an important, original, or non-standard component of the core methods in this research? Note that if the LLM is used only for writing, editing, or formatting purposes and does not impact the core methodology, scientific rigorousness, or originality of the research, declaration is not required. + +# Answer: [NA] + +Justification: The development of our core methods does not incorporate LLMs. + +# Guidelines: + +- The answer NA means that the core method development in this research does not involve LLMs as any important, original, or non-standard components. +- Please refer to our LLM policy (https://neurips.cc/Conferences/2025/LLM) for what should or should not be described. \ No newline at end of file diff --git a/NeurIPS/2025/$_textit{Hyper-GoalNet}$_ Goal-Conditioned Manipulation Policy Learning with HyperNetworks/images.zip b/NeurIPS/2025/$_textit{Hyper-GoalNet}$_ Goal-Conditioned Manipulation Policy Learning with HyperNetworks/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..c9f9b55fef07c6c784179627974192a859a6c9b8 --- /dev/null +++ b/NeurIPS/2025/$_textit{Hyper-GoalNet}$_ Goal-Conditioned Manipulation Policy Learning with HyperNetworks/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c2c11e5e81109bc1ea052337cd8f44fa06c43046ea24a20cd467eedcc6b00904 +size 1028638 diff --git a/NeurIPS/2025/$_textit{Hyper-GoalNet}$_ Goal-Conditioned Manipulation Policy Learning with HyperNetworks/layout.json b/NeurIPS/2025/$_textit{Hyper-GoalNet}$_ Goal-Conditioned Manipulation Policy Learning with HyperNetworks/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..d7b2e3dae71a0d49d7927587aaafbf7ace3bc9e4 --- /dev/null +++ b/NeurIPS/2025/$_textit{Hyper-GoalNet}$_ Goal-Conditioned Manipulation Policy Learning with HyperNetworks/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:394b4ec8282eac4f4f49eb90e46cc984f3b7647601fca1790f2ef4dc88dd545c +size 1074152 diff --git a/NeurIPS/2025/$_texttt{BetaConform}$_ Efficient MAP Estimation of LLM Ensemble Judgment Performance with Prior Transfer/ec15ba4b-643b-443b-a5cc-470c3ec49ae5_content_list.json b/NeurIPS/2025/$_texttt{BetaConform}$_ Efficient MAP Estimation of LLM Ensemble Judgment Performance with Prior Transfer/ec15ba4b-643b-443b-a5cc-470c3ec49ae5_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..6f5062f2883c0689c5bace45d6189fec2ae90500 --- /dev/null +++ b/NeurIPS/2025/$_texttt{BetaConform}$_ Efficient MAP Estimation of LLM Ensemble Judgment Performance with Prior Transfer/ec15ba4b-643b-443b-a5cc-470c3ec49ae5_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:19e5cf670ebb71ed7e2045f7603a4afd55ef410ec166cc2b0b728922134302d7 +size 194273 diff --git a/NeurIPS/2025/$_texttt{BetaConform}$_ Efficient MAP Estimation of LLM Ensemble Judgment Performance with Prior Transfer/ec15ba4b-643b-443b-a5cc-470c3ec49ae5_model.json b/NeurIPS/2025/$_texttt{BetaConform}$_ Efficient MAP Estimation of LLM Ensemble Judgment Performance with Prior Transfer/ec15ba4b-643b-443b-a5cc-470c3ec49ae5_model.json new file mode 100644 index 0000000000000000000000000000000000000000..0b9379e042769b587dc2d6895f729c1b2dd54660 --- /dev/null +++ b/NeurIPS/2025/$_texttt{BetaConform}$_ Efficient MAP Estimation of LLM Ensemble Judgment Performance with Prior Transfer/ec15ba4b-643b-443b-a5cc-470c3ec49ae5_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f1aaa4221138b9cf6cbb7c19765aefc0d5bb84cf3f783aa2fc40399c54b53578 +size 247685 diff --git a/NeurIPS/2025/$_texttt{BetaConform}$_ Efficient MAP Estimation of LLM Ensemble Judgment Performance with Prior Transfer/ec15ba4b-643b-443b-a5cc-470c3ec49ae5_origin.pdf b/NeurIPS/2025/$_texttt{BetaConform}$_ Efficient MAP Estimation of LLM Ensemble Judgment Performance with Prior Transfer/ec15ba4b-643b-443b-a5cc-470c3ec49ae5_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..482457915b1a67e16d6d3df6acb7de92c39599a2 --- /dev/null +++ b/NeurIPS/2025/$_texttt{BetaConform}$_ Efficient MAP Estimation of LLM Ensemble Judgment Performance with Prior Transfer/ec15ba4b-643b-443b-a5cc-470c3ec49ae5_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ddd05c9557841010677d62a6548693fd576c4b69414745276879f8fedfc977e8 +size 1111972 diff --git a/NeurIPS/2025/$_texttt{BetaConform}$_ Efficient MAP Estimation of LLM Ensemble Judgment Performance with Prior Transfer/full.md b/NeurIPS/2025/$_texttt{BetaConform}$_ Efficient MAP Estimation of LLM Ensemble Judgment Performance with Prior Transfer/full.md new file mode 100644 index 0000000000000000000000000000000000000000..cc20cf460a0122e66f7c33634e3fa76d151524f6 --- /dev/null +++ b/NeurIPS/2025/$_texttt{BetaConform}$_ Efficient MAP Estimation of LLM Ensemble Judgment Performance with Prior Transfer/full.md @@ -0,0 +1,1128 @@ +# BetaConform: Efficient MAP Estimation of LLM Ensemble Judgment Performance with Prior Transfer + +Huaizhi Qu1, Inyoung Choi2, Zhen Tan3, Song Wang4, Sukwon Yun1, Qi Long2, Faizan Siddiqui5, Kwonjoon Lee5, Tianlong Chen1 + +1University of North Carolina at Chapel Hill, 2University of Pennsylvania +3Arizona State University 4University of Virginia 5Honda Research Institute USA {huaizhiq, tianlong}@cs.unc.edu + +# Abstract + +LLM ensembles are widely used for LLM judges. However, how to estimate their accuracy, especially in an efficient way, is unknown. In this paper, we present a principled maximum a posteriori (MAP) framework for an economical and precise estimation of the performance of LLM ensemble judgment. We first propose a mixture of Beta-Binomial distributions to model the judgment distribution, revising from the vanilla Binomial distribution. Next, we introduce a conformal prediction-driven approach that enables adaptive stopping during iterative sampling to balance accuracy with efficiency. Furthermore, we design a prior transfer mechanism that utilizes learned distributions on open-source datasets to improve estimation on a target dataset when only scarce annotations are available. Finally, we present BetaConform, a framework that integrates our distribution assumption, adaptive stopping, and the prior transfer mechanism to deliver a theoretically guaranteed distribution estimation of LLM ensemble judgment with minimum labeled samples. BetaConform is also validated empirically. For instance, with only 10 samples from the TruthfulQA dataset, for a Llama ensembled judge, BetaConform gauges its performance with an error margin as small as $3.37\%$ . + +# 1 Introduction + +With the improving performance of large language models (LLMs), there is a proliferation of adopting LLMs as judges for various tasks [Liang et al., 2023, Yuan et al., 2024b, Zhang et al., 2025]. In applications of LLM judge ensembles, the judgment distribution is critical to the service quality [Chen et al., 2024, Schoenegger et al., 2024, Qiu et al., 2025]. Many datasets [Zheng et al., 2023, Zeng et al., 2023, Yuan et al., 2024a] have been employed to evaluate the performance of LLM judges. However, these datasets rely on human annotations, which are impractical at a large scale due to the substantial time and financial costs of annotating. This + +![](images/60fda3fc56e03a6b14cf1b2c076ef482871ff37493a3fdb6bd2752238a6d7310.jpg) +Figure 1: In this paper, we aim to answer (1) how to estimate the judgment distribution of LLM ensemble on a dataset, and (2) how to achieve efficient estimation to reduce annotation effort. + +challenge highlights the need of how to estimate the LLM ensemble judging performance efficiently. + +In this work, we consider the following judgment distribution estimation problem: + +$$ +\mathbb {P} (\# \text {c o r r e c t j u d g m e n t s} = n \mid k \text {L L M s j u d g e s a m p l e} x). +$$ + +We propose an efficient method for MAP estimation of the distribution of LLM ensemble judgment to answer two research questions shown in Figure 1. + +- RQ1: How to efficiently and accurately estimate the judgment distribution? +- RQ2: How many samples are needed for estimation under given error margin threshold? + +Given a small number of samples, one intuitive estimation is to directly adopt the distribution of the samples as the judgment distribution on the entire dataset. However, this is susceptible to the sampling bias. To avoid this, one common practice is to first calculate the single LLM accuracy on the samples and then model the distribution on the full dataset as Binomial. We first posit that the judgment distribution is not Binomial. Theoretically, a Binomial distribution implies increasing accuracy in majority voting as the ensemble size grows [De Condorcet et al., 2014, Austen-Smith and Banks, 1996]. However, this is unrealistic since the accuracy of LLM ensembles remains bounded even with a large number of judges. To testify to this, we start by observing the distribution of LLM ensemble judges on various benchmarks. We find marked deviations from the Binomial distribution and show a stratification between questions that can be classified as "easy" and "hard". To this end, we propose to model the judgment distribution with a mixture of Beta-Binomial distributions to reflect the stratification. We show that under this assumption, by utilizing an expectation maximization (EM) estimation method, it can achieve accurate judgment distribution estimation with high data efficiency. + +To rigorously guide the sampling process and determine how many samples to use for the estimation, we draw inspiration from the conformal prediction (CP) [Shafer and Vovk, 2008, Fontana et al., 2023] that can efficiently estimate the sampling deviation. Based on this, we propose a novel adaptive stopping strategy for iterative sampling, designed to meet a pre-defined deviation threshold. Our experiments demonstrate the effectiveness of this method for limiting the sample amount while maintaining high estimation precision. + +Moreover, we hypothesize that the prior knowledge of judgment distribution on open-source datasets can benefit the estimation of a new dataset when only a few samples are available. To achieve this, we propose a text similarity-based distribution prior transfer mechanism. This method embeds text inputs from both source and target datasets and calculates embedding similarities to determine the transfer weight. Our design greatly improves the estimation accuracy when transferring from similar datasets and avoids performance degradation when the datasets are distinct. Notably, this method relies + +solely on the text inputs, making it practical for application to vast amounts of unlabeled data. + +![](images/815d52b097def833e5fe690ec479c2822bfc3546b56fc5b7beb413353804bd04.jpg) +Figure 2: Overview of BetaConform. Given a target dataset, adaptive stopping is adopted to determine the sample amount (b, Section 5). During iterative sampling, the sampling deviation is monitored by using conformal prediction. The sampling process stops when the deviation is sufficiently low. Next, the estimation of the small number of samples from the previous step is further enhanced by transferring distribution priors from source datasets (c, Section 6). The transfer will assign a larger weight to the dataset that is textually closer to the target dataset. + +Our contribution can be summarized as follows: + +- We present pioneering work in judgment distribution estimation. We point out that the Binomial assumption of judgment distribution is inaccurate. By replacing it with a mixture of Beta-Binomial distributions, we could achieve efficient and accurate estimation. +- We design a rigorous conformal prediction-based adaptive stopping strategy during iterative sampling when the sampling deviation is sufficiently low. +- We introduce a distribution prior transfer mechanism that leverages judgment distributions on open-source datasets to improve few-sample estimations. +- Extensive experiments show BetaConform's high estimation efficiency. For example, using only 10 samples could result in an average of $10.84\%$ error margin. + +# 2 Related Works + +LLMs for Judgment. Reliable model evaluation is a critical problem. Traditional human evaluations remain the gold standard, but their scalability is a significant bottleneck in large-scale + +applications. Thus, recent works have proposed leveraging LLMs to evaluate the text quality, ranking outputs, and ensuring alignment with human preferences [Zheng et al., 2023, Liu et al., 2023, Dubois et al., 2024]. While initially focused on text generation evaluation, the use of LLMs as judges has expanded to diverse applications including model alignment and safety assessment [Lee et al., 2024], code quality evaluation [Zhao et al., 2024b], and knowledge verification [Min et al., 2023], etc. + +Challenges and Limitations. The reliability of such frameworks is not without concerns. Studies have found that even advanced models like GPT-4 often exhibit systematic biases such as position bias and egocentric bias [Zeng et al., 2023, Wang et al., 2023], overconfidence in their judgments [Koo et al., 2024], and self-preference effects [Panickssery et al., 2024]. Moreover, many studies employing LLM annotations do not explicitly measure the alignment between LLMs and humans, thus further raising questions about their dependability [Calderon et al., 2025]. While researchers have proposed various solutions, including dynamic evaluation pipelines [Yu et al., 2024, Zhao et al., 2024a, Moniri et al., 2024], self-reflection mechanisms [Wu et al., 2024, Li et al., 2023b, Wang et al., 2024], and specialized benchmarks for assessing judge performance [Zheng et al., 2023, Tan et al., 2024, Park et al., 2024, Li et al., 2024, Zhao et al., 2024b], these methods often fall short in offering rigorous guarantees of their outcomes. A related line of research is Item Response Theory (IRT) [Cai et al., 2016, Baker, 2001, Harvey and Hammer, 1999], which assesses respondents' latent abilities using responses to calibrated questions. However, the requirement for calibrated questions limits the direct applicability of IRT in the context of judgment distribution estimation, as datasets in this domain are frequently unlabeled. + +Statistical Approaches. Another direction of research focuses on providing statistical guarantees for LLM performance. Researchers have explored conformal methods [Angelopoulos et al., 2023] to ensure correctness and factuality [Mohri and Hashimoto, 2024] and to determine when LLMs should abstain from responding [Yadkori et al., 2024]. While these methods provide some statistical rigor, there is still a need for a unified framework that establishes reliable, theoretically grounded approaches for assessing LLM performance across diverse applications. + +# 3 Problem Setup + +We consider the task of using an LLM ensemble to evaluate and judge samples by discerning, choosing, or scoring. Let: + +- $n$ : Total number of samples in the dataset to be judged. +- $k$ : Number of LLMs in an ensemble. +- $S$ : The random variable of correct judgments. +- $r$ : Number of samples to estimate $S$ . +- $D$ : A dataset to estimate the judgment distribution. + +Definition 1 (LLM Ensemble Judgment). Let $\mathcal{J} = \{J_1, J_2, \dots, J_k\}$ be an ensemble of $k$ LLM judges. For a given input $x$ , each LLM $J_i$ generates an output $o_i = J_i(x)$ , yielding the set of all judgments $\mathcal{O} = \{o_1, o_2, \dots, o_k\}$ . In this paper, we focus on binary and scoring judgments. We consider the LLM ensemble to be composed of multiple instances of the same underlying model (e.g., $k = 11$ Llama models). Variations in their judgments for a given input are due to Top-P token sampling [Zhou et al., 2024] and the difference in random seeds. + +Definition 2 (LLM Ensemble Correct Judgment). For an ensemble of $k$ LLMs, the random variable $S = \sum_{i=1}^{k} \text{Match}(o_i, y)$ represents the number of correct judgments. $y$ denotes the ground truth, and $\text{Match}(\cdot)$ is the criterion for a correct judgment. For instance, for binary classification judgments, $\text{Match}(\cdot)$ could be an exact match; for scoring judgments, it could be whether the score falls within a predefined range of the human average score. The ensemble's decision is deemed correct if $S \geq \lceil k/2 \rceil$ . To prevent ties, which can occur if $k$ is an even integer and $S = \lceil k/2 \rceil$ , we stipulate that $k$ must be an odd integer. + +# 4 Mixture of Beta-Binomial Distribution + +# 4.1 Examination of Binomial Distribution + +We start by examining the common assumption of $S$ follows a Binomial distribution, i.e. the probability of having $s$ correct judgments when a single judge accuracy $\hat{p}$ is, + +![](images/44b31a438ade4af1c2c358eacbd55e5f384f23f1d4bb9b33e309dde685672150.jpg) +Figure 3: Comparison of judgment distributions among actual, Binomial, and ours. Llama-3.3-70B and GPT-4 ensembles of 11 models are tested on HaluEval and JudgeBench, respectively. The Binomial distribution is estimated by using single judge accuracy $p$ . Our mixture distribution is estimated with 100 samples and scaled to the full dataset. Our distribution is consistently better. + +![](images/0100673f73d1079195fa29a5352e45f7ea25fec4498b4d055a0fc72b2dc1d4fe.jpg) +Figure 4: Majority voting error rate of actual, Binomial, and our mixture distribution. Binomial uses single judge accuracy $p$ . Our distribution is estimated with 100 random samples and tested for 3 times. The line denotes the average error rate and the shadow represents the standard variance. Binomial shows decreasing error rate, while our distribution captures the actual trend. + +$$ +\mathbb {P} _ {\operatorname {B i n}} (S = s) = \operatorname {B i n} (s \mid k, \hat {p}) = \binom {k} {s} \hat {p} ^ {s} (1 - \hat {p}) ^ {k - s}. \tag {1} +$$ + +The error rate $\tilde{P}_{Bin}$ of ensemble judgment is: + +$$ +\tilde {P} _ {\mathrm {B i n}} = \mathbb {P} _ {\mathrm {B i n}} (S < \lceil k / 2 \rceil) = \sum_ {s = 0} ^ {\lceil k / 2 \rceil - 1} \binom {k} {s} \hat {p} ^ {s} (1 - \hat {p}) ^ {k - s}. \tag {2} +$$ + +We first examine the common assumption that $S$ follows a Binomial distribution in Equation (1). Specifically, we ① evaluate individual LLMs on datasets across domains and ② use the single LLM accuracy $p$ in Equation (1) and (2) to estimate both the distribution of LLM ensembles on these datasets and the majority voting error rate for different numbers $k$ of LLMs. Specifically, we evaluate GPT-4 [OpenAI et al., 2024] and Llama-3.3-70B [Dubey et al., 2024] on hallucination detection (HaluEval, Li et al., 2023a) and Human alignment (JudgeBench, Tan et al., 2024) datasets. Results are shown in Figure 3 and Figure 4. + +The results in Figure 3 and Figure 4 demonstrate the large deviation of Binomial distribution to the real distribution. On both datasets, the real distributions of LLM ensemble judgments consistently show two peaks centering at the two ends, while Binomial distribution results in a single peak with a large shift to either of the two peaks. Notably, in Figure 4, the assumption of a Binomial distribution leads to an always decreasing majority voting error rate, which is in sharp contrast with the actual error rate that remains at the same level when the ensemble becomes larger. + +# 4.2 Mixture of Beta-Binomial Distributions + +Assumption 1 (Mixture of Beta-Binomial Distributons). + +$$ +S \sim w \mathrm {B B} (k, \alpha_ {1}, \beta_ {1}) + (1 - w) \mathrm {B B} (k, \alpha_ {2}, \beta_ {2}), \tag {3} +$$ + +where $\mathrm{BB}(\cdot ,\cdot ,\cdot)$ is the Beta-Binomial distribution, $k$ is the number of judges in the ensemble, $\alpha_{1},\beta_{1},\alpha_{2},\beta_{2}$ are parameters of the two distributions, and $w$ is the mixture weight. + +Corollary 1 (Mixture Distribution Error Rate). The error rate of the mixture of Beta-Binomial distributions is + +$$ +\tilde {P} _ {\mathrm {B B}} = w \sum_ {s = 0} ^ {\lceil k / 2 \rceil - 1} \binom {k} {s} \frac {\mathrm {B} (s + \alpha_ {1} , k - s + \beta_ {1})}{\mathrm {B} (\alpha_ {1} , \beta_ {1})} + (1 - w) \sum_ {s = 0} ^ {\lceil k / 2 \rceil - 1} \binom {k} {s} \frac {\mathrm {B} (s + \alpha_ {2} , k - s + \beta_ {2})}{\mathrm {B} (\alpha_ {2} , \beta_ {2})}, \tag {4} +$$ + +where $\mathrm{B}(\cdot ,\cdot)$ is the Beta function. + +After examining the common Binomial distribution assumption in Figure 3 and Figure 4, we notice that the real distribution keeps showing two peaks centering near all wrong and all correct. Motivated by this observation, in Assumption 1 we model the distribution as a mixture of two Beta-Binomial distributions, where one distribution models the LLM ensemble judgments on simple questions and the other one for hard problems. To derive all the parameters, we utilize labeled samples from the dataset and design a distribution-tailored expectation maximization (EM) algorithm. + +# 4.3 Expectation Maximization + +Samples as Distribution Evidence. Given $r$ samples, each containing judgments from $k$ LLMs, $S_{i}$ is the number of correct judgments in the $i$ -th sample and $p_{i} = S_{i} / k$ as the estimated probability of success for the $i$ -th sample. + +For the $i$ -th sample, considering the first Beta-Binomial distribution, a responsibility $\gamma_1^i$ is assigned as + +$$ +\gamma_ {1} ^ {i} = \frac {w \operatorname {B e t a} \left(p _ {i} \mid \alpha_ {1} , \beta_ {1}\right)}{w \operatorname {B e t a} \left(p _ {i} \mid \alpha_ {1} , \beta_ {1}\right) + (1 - w) \operatorname {B e t a} \left(p _ {i} \mid \alpha_ {2} , \beta_ {2}\right)}, \tag {5} +$$ + +where $\mathrm{Beta}(p_i\mid \alpha ,\beta)$ is the probability density of beta distribution at $p_i$ for the $i$ -th sample under the corresponding Beta component. $\gamma_1^i$ represents the probability that the $i$ -th sample belongs to the first Beta component, and $\gamma_2^i = 1 - \gamma_1^i$ is the probability for the second component. + +Parameters Update. The parameters are updated based on the weighted contributions of samples. The parameters of two distributions $j = \{1,2\}$ are updated as + +$$ +\alpha_ {j} ^ {\prime} = \sum_ {i = 1} ^ {r} \gamma_ {1} ^ {i} \cdot S _ {i}, \beta_ {j} ^ {\prime} = \sum_ {i = 1} ^ {r} \gamma_ {1} ^ {i} \cdot (k - S _ {i}), w ^ {\prime} = \frac {1}{r} \sum_ {i = 1} ^ {r} \gamma_ {1} ^ {i} \tag {6} +$$ + +We verify our distribution assumption by first sampling $r = 100$ judgments made by two models on two datasets and apply our distribution-tailored EM algorithm to estimate the parameters. Our method is evaluated in two scenarios: ① In Figure 3, we fix the ensemble size $k = 11$ and compare the estimated distribution against the real distribution and Binomial distribution, and ② in Figure 4 we estimate the error rate of majority voting with different ensemble sizes. + +In Figure 3, the mixture of Beta-Binomial distributions is significantly closer to the real distribution compared to the Binomial, with clear two-peak patterns that are analogous to the observation. In Figure 4 it shows that our distribution is consistently close to the real majority voting error rate across all ensemble sizes. Contrary to the Binomial distribution that produced a decreasing error rate, our distribution successfully modeled the stable error rate when the ensemble becomes larger. Additionally, the narrow confidence interval demonstrates the high stability of our method. + +# 5 Guide Sampling via Conformal Prediction + +In the experiments above, we used a fixed number of samples. However, in practical settings where datasets are unannotated and being labeled, it is essential to determine when the number of annotated samples is sufficient for accurate estimation. Inspired by conformal prediction (CP), which does not rely on prior knowledge of the dataset distribution and can rigorously estimate the sampling deviation, we propose leveraging its principles to address this challenge. + +# 5.1 Conformal Prediction for Adaptive Stopping + +CP provides a principled approach to dynamically evaluate the sampling deviation in the distribution of the number of correct judgments $S$ , which can be used as guidance. + +Nonconformity Scores. A major part of CP is the nonconformity score, which measures how a test sample differs from the rest of the data. In our implementation, we set the nonconformity score as + +$$ +\operatorname {s c o r e} \left(S _ {i}\right) = \left| S _ {i} - \mathbb {E} [ S ] \right|, \tag {7} +$$ + +![](images/899ba01ab10c5d593300edd4b75a2475db66fbb9e01bebe10f5a09ec200540f3.jpg) +HaluEval Embedding + +![](images/c9423851dc3f2596b1987f08b6a9369a8759f04a3c8d74e678bfdf11e46ae3e3.jpg) +HaluEval Estimation Margin + +![](images/a1f783b825a85d00cd56bb52a07aa70e6e41f08d9766cedf87ed90e9fe93c182.jpg) +TruthfulQA Embedding +Figure 5: Examples of distribution prior transfer. Splits from HaluEval form distinct clusters in the embedding space, and transfer does not degrade performance compared to only using target dataset samples. In contrast, topics in TruthfulQA exhibit closer proximity, where transfer leads to significant performance improvements compared to solely using the limited samples of the target dataset. + +![](images/874bac15cca8d6500b047ddcdf371e5389bb86d4a9f275786f185701d7f7e505.jpg) +TruthfulQA Estimation Margin + +which quantifies the deviation of each observed value of $S$ from the expected value. + +Calibration Data and Quantile Computation. Suppose $r$ samples have been used to test the LLM ensemble with $S_{1}, S_{2}, \ldots, S_{r}$ correct judgments, the CP sampling computes the nonconformity scores for all calibration data as $s_{i} = \mathrm{score}(S_{i})$ and these scores are sorted in ascending order as $s_{1} < \ldots < s_{r}$ . For a desired estimation confidence $1 - \epsilon$ , the $(1 - \epsilon)$ -quantile with $r$ samples $q_{1 - \epsilon}^{r}$ is + +$$ +q _ {1 - \epsilon} ^ {r} = s _ {\lceil (1 - \epsilon) \cdot (r + 1) \rceil}. \tag {8} +$$ + +Adaptive Stopping Criteria. Adaptive stopping is achieved by monitoring the variation of the conformal prediction quantile. After $r$ samples, the $(1 - \epsilon)$ -quantile is recomputed and compared with the one from $r - 1$ samples. The sampling process stops when the quantile satisfies + +$$ +\left| q _ {1 - \epsilon} ^ {r} - q _ {1 - \epsilon} ^ {r - 1} \right| \leq \xi \tag {9} +$$ + +where $\xi$ is a predefined threshold. + +Proposition 1 (Sample Amount with Adaptive Stopping). For a given sampling deviation threshold $\xi$ and a scale $\tau$ , the sample amount $r$ should satisfy + +$$ +\tau \left(\frac {1}{\sqrt {r - 1}} - \frac {1}{\sqrt {r}}\right) \leq \xi , \tag {10} +$$ + +This proposition offers an estimation of the sample amount under the threshold $\xi$ . + +Proposition 2 (Error Rate with Adaptive Stopping). Under the sampling threshold $\xi$ , the majority voting error rate of the mixture distribution becomes + +$$ +\left(1 - \min \left(\xi , \frac {\tau}{\sqrt {r}}\right)\right) \tilde {P} _ {\mathrm {B B}} < \tilde {P} _ {\text {a d a p t}} < \left(1 + \min \left(\xi , \frac {\tau}{\sqrt {r}}\right)\right) \tilde {P} _ {\mathrm {B B}} \tag {11} +$$ + +This proposition provides a theoretical error bound for estimation under adaptive stopping, suggesting the mild degradation of estimation performance. + +We leave the proofs of Proposition 1 and 2 in Appendix B.1 and B.2, respectively. In our experiments, we set $\xi = 0.03$ , and $\tau = 25$ , which leads to $r \geq 56$ . + +# 6 Text Similarity for Distribution Prior Transfer + +To further improve the data efficiency when only a few samples are available and enhance estimation accuracy, we propose to incorporate prior knowledge about the LLM ensemble on other open-source datasets and transfer the estimated judgment distributions to the target dataset. However, one challenge is that the prior transfer could bring performance degradation if the distributions of the source datasets and the target dataset are very different. To resolve this challenge, we design text similarity-based distribution prior transfer, which leverages the strong text embedding capability of the recent models to understand and measure the textual difference among datasets. + +Text Embedding. To embed the text inputs of the LLM ensemble, we use NV-Embed-V2 [Lee et al., 2025]. Given sets of samples $\{D_1, D_2, \ldots, D_m\}$ from $m$ source datasets, the embedding model $\mathcal{E}(\cdot)$ is utilized to transform the sets of samples to sets of embeddings for the source datasets + +$$ +\left\{E _ {1}, E _ {2}, \dots , E _ {m} \right\} = \left\{\mathcal {E} \left(D _ {1}\right), \mathcal {E} \left(D _ {2}\right), \dots , \mathcal {E} \left(D _ {m}\right) \right\}. \tag {12} +$$ + +The average embedding $\bar{E}_i = \frac{1}{r_i}\sum_{j=1}^{r_i} E_i^j$ of the $i$ -th dataset is used to represent it. + +Distribution Prior Transfer. To transfer the distribution from source datasets to the target dataset $D_0$ , the process starts by embedding the target dataset $E_0 = \mathcal{E}(D_0)$ and acquiring its average embedding $\bar{E}_0$ . For the dataset $D_i$ , its transfer weight is + +$$ +\lambda_ {i} = \log (r _ {i}) \cdot \sigma \left(\rho_ {1} \cdot \left(\operatorname {C o s S i m} (\bar {E} _ {0}, \bar {E} _ {i}) - \rho_ {2}\right)\right), \tag {13} +$$ + +where $\sigma (\cdot)$ is the sigmoid function, $r_i$ is the number of samples and $\rho_{1}$ and $\rho_{2}$ are hyperparameters. We adopt this design to avoid the degradation of estimation caused by transferring datasets with dissimilar text inputs. This is achieved by setting a threshold and applying the sigmoid function to suppress the weight when the similarity is low. $\log (r_i)$ is included as datasets with more samples could produce a more accurate estimation and thus should have a higher impact on the transfer. The transfer from the source datasets to the target dataset is performed as + +$$ +w _ {0} ^ {t r} = \frac {\sum_ {i = 0} ^ {m} \lambda_ {i} \cdot w _ {i}}{\sum_ {i = 0} ^ {m} \lambda_ {i}}, \alpha_ {0, j} ^ {t r} = \frac {\sum_ {i = 0} ^ {m} \lambda_ {i} \cdot \alpha_ {i , j}}{\sum_ {i = 0} ^ {m} \lambda_ {i}}, \beta_ {0, j} ^ {t r} = \frac {\sum_ {i = 0} ^ {m} \lambda_ {i} \cdot \beta_ {i , j}}{\sum_ {i = 0} ^ {m} \lambda_ {i}}, \quad j \in \{1, 2 \}. \tag {14} +$$ + +In Equation (14), $\alpha_{i,j}$ and $\beta_{i,j}$ are the $j$ -th parameter in the mixture distribution of $i$ -th dataset. The parameters in the weighted sum with index 0 denote direct estimation on the target dataset. + +Examples. To verify our distribution design, we evaluate the distribution within splits of HaluEval [Li et al., 2023a] and TruthfulQA [Lin et al., 2021] datasets. For HaluEval, we use Dialogue and Summarization splits as source datasets and transfer to QA split; for TruthfulQA, we transfer from topics of Health and Law to Misconceptions. As shown in Figure 5, the embeddings form distant clusters in HaluEval, as the text inputs of the three splits have different hallucination detection requirements, and embeddings from TruthfulQA overlap due to the similarity of judgment format. When clusters are separated, our method will not bring performance degradation compared to solely using samples from the target dataset, while when clusters are overlapping, our method brings a significantly lower estimation error rate margin compared to only using target dataset samples. This supports the effectiveness of our distribution transfer design. + +We present the algorithm and Python implementation of BetaConform in Section A and Section E. + +# 7 Experiments + +# 7.1 Estimation Accuracy + +We begin by evaluating BetaConform with adaptive stopping on datasets to verify its accuracy. We choose Binomial distribution and a single Beta-Binomial distribution as baselines and compare the error margin, which is the absolute difference between the estimation error rate and the actual value. The results are reported in Table 1. Please see Section E for implementation details. + +From the results, the following observations can be drawn: ① Compared to the Binomial distribution, BetaConform achieves consistently lower error margin, with $32.4\% \sim 54.1\%$ improvements of average error margin of all models. This demonstrates an effective answer to RQ1 by modeling judgment distribution as a mixture of Beta-Binomial distributions. ② The number of samples is close to the theoretical estimation. The average sample amount of models on all datasets exhibit a slight deviation of the estimated value 56 by $3.14 \sim 12.86$ samples. This validates our design of using the distribution-free CP for adaptive stopping, which effectively solved RQ2. + +# 7.2 Distribution Prior Transfer + +We then verify our text similarity-based distribution prior transfer when only limited samples are available. We constrain to 10 samples from the target dataset and assume the full source datasets are accessible. Transfer is compared with estimating only on the target dataset samples (w/o Transfer). Error margins are shown in Table 2. We also conduct ablation studies of the transfer design in Table 4 + +Table 1: The comparison of error margins between our mixture of Beta-Binomial distributions and Binomial distribution. The Err. Margin and # Samples answer RQ1 and RQ2, respectively. The error margin is calculated as the absolute difference between the actual error rate and the estimation. Estimations using both distributions are done on samples obtained through iterative sampling with adaptive stopping. For each run, the error margin is computed from $k = 1$ to 11, and the average margin of ensemble sizes is used as the result for that run. We conduct 30 runs and report the average and standard deviation. The average number of samples across runs is also reported. + +
DatasetMethodLlama-3.3-70BQwen-2.5-72BInternLM-20BGPT-3.5GPT-4
Error Margin (↓)# Samples (↓)Error Margin (↓)# Samples (↓)Error Margin (↓)# Samples (↓)Error Margin (↓)# Samples (↓)Error Margin (↓)# Samples (↓)
Hallucination Detection
HaluEvalBinomial17.62 ± 0.7312.45 ± 1.0416.67 ± 0.385.78 ± 0.089.16 ± 0.18
Single BB14.46 ± 0.165.14 ± 0.2115.92 ± 0.115.27 ± 0.099.77 ± 0.84
Ours6.68 ± 0.534.72 ± 0.385.48 ± 0.415.10 ± 0.246.28 ± 0.39
TruthfulQABinomial14.00 ± 0.6519.86 ± 0.4019.55 ± 0.6514.44 ± 0.4015.20 ± 0.55
Single BB8.83 ± 1.027.84 ± 0.266.79 ± 0.2512.17 ± 0.9911.31 ± 0.52
Ours7.53 ± 0.557.18 ± 0.446.24 ± 0.596.75 ± 0.586.73 ± 0.38
HalluDialBinomial13.10 ± 0.3713.42 ± 0.5414.84 ± 0.428.79 ± 0.219.25 ± 0.27
Single BB11.33 ± 0.6416.75 ± 0.907.95 ± 0.349.24 ± 0.598.43 ± 0.45
Ours7.94 ± 0.686.96 ± 0.476.43 ± 0.506.27 ± 0.365.22 ± 0.59
Reasoning
PRM800KBinomial10.11 ± 0.299.14 ± 0.179.12 ± 0.208.83 ± 0.2514.52 ± 0.73
Single BB16.45 ± 1.3510.30 ± 0.609.81 ± 0.619.45 ± 0.7212.31 ± 0.31
Ours9.37 ± 0.647.82 ± 0.694.52 ± 0.508.46 ± 0.516.17 ± 0.48
BIG-benchBinomial13.29 ± 0.7814.17 ± 0.4014.68 ± 0.2414.83 ± 0.5312.15 ± 0.74
Single BB13.15 ± 0.6812.32 ± 0.609.51 ± 0.5617.93 ± 0.8911.50 ± 0.91
Ours11.15 ± 0.606.97 ± 0.585.54 ± 0.5112.59 ± 0.488.02 ± 0.59
TRAMBinomial14.79 ± 0.8213.13 ± 0.6413.06 ± 0.774.99 ± 0.135.14 ± 0.11
Single BB11.75 ± 0.745.72 ± 0.396.01 ± 0.447.42 ± 0.144.01 ± 0.30
Ours8.39 ± 0.636.20 ± 0.346.10 ± 0.583.94 ± 0.174.81 ± 0.23
Alignment
JudgeBenchBinomial12.06 ± 0.7813.45 ± 0.5410.31 ± 1.038.85 ± 0.3310.98 ± 0.32
Single BB7.60 ± 0.377.64 ± 0.545.11 ± 0.2411.85 ± 0.787.62 ± 0.25
Ours6.98 ± 0.565.39 ± 0.395.26 ± 0.397.03 ± 0.616.45 ± 0.53
RewardBenchBinomial8.40 ± 0.198.93 ± 0.2217.36 ± 1.4111.42 ± 0.3313.98 ± 0.29
Single BB16.29 ± 1.3911.40 ± 1.206.15 ± 0.278.79 ± 0.218.80 ± 0.40
Ours11.30 ± 0.624.68 ± 0.566.58 ± 0.406.90 ± 0.457.65 ± 0.51
LLMBarBinomial13.61 ± 0.5814.63 ± 0.5113.66 ± 1.1413.19 ± 0.5510.36 ± 0.33
Single BB14.21 ± 0.677.97 ± 0.585.46 ± 0.3013.46 ± 0.8311.72 ± 0.48
Ours10.18 ± 0.717.52 ± 0.636.38 ± 0.5313.71 ± 0.548.16 ± 0.50
Scoring
ICE-ScoreBinomial8.91 ± 0.259.27 ± 0.2322.24 ± 1.023.61 ± 0.063.66 ± 0.07
Single BB16.71 ± 1.119.24 ± 0.5910.97 ± 0.273.54 ± 0.224.69 ± 0.10
Ours8.97 ± 0.456.91 ± 0.5918.19 ± 0.373.39 ± 0.325.78 ± 0.08
COMP-AnalysisBinomial14.45 ± 0.7115.88 ± 0.7213.28 ± 0.7312.87 ± 0.3215.64 ± 0.68
Single BB8.56 ± 0.666.93 ± 0.344.61 ± 0.277.82 ± 0.2911.32 ± 0.43
Ours6.50 ± 0.636.95 ± 0.504.86 ± 0.486.66 ± 0.387.07 ± 0.48
Average
AverageBinomial12.76 ± 0.5613.12 ± 0.4914.98 ± 0.739.78 ± 0.2910.91 ± 0.39
Single BB12.67 ± 0.809.20 ± 0.568.03 ± 0.339.72 ± 0.5211.11 ± 0.45
Ours8.63 ± 0.606.48 ± 0.516.87 ± 0.487.35 ± 0.426.38 ± 0.44
+ +![](images/92395dae91ceadb1f476e9c48e07a295fd75fcb682315b51cf95075dd98a6448.jpg) +Figure 6: The actual number of samples under various thresholds $\xi$ versus the theoretical value from Equation (10). The actual sample numbers match with the theoretical bound. + +![](images/c6357bb5b1061257e7f7abce5a171129dd1de3f5955a8f7e2a1e497d0f0a090f.jpg) +Figure 7: The actual number of samples from different datasets under three $\xi$ values. Our sampling with adaptive stopping shows consistent results on all the datasets. + +From the results, we observe that by transferring from other datasets in the same category (e.g., from TruthfulQA and HalluDial to HaluEval), the average error margin across all datasets is reduced by $5.0\% \sim 25.0\%$ and is consistently lower compared to no transfer, suggesting the effectiveness of using prior knowledge of the judgment distributions on open-source datasets can benefit estimation. + +# 7.3 More Research Questions + +RQ3: Is sampling with adaptive stopping consistent to the theory? We examine our adaptive stopping to see if Equation (10) matches the real sampling amount. We set a series of $\xi$ while keeping $\tau = 25$ and sample with adaptive stopping from judgment samples produced by Llama, Qwen, and GPT-4, and compare with the theoretical value of Equation (10). The actual sample amounts under different thresholds in Figure 6 match closely with the theoretical estimation, which proves the effectiveness of quantifying sampling deviation through CP and the Proposition 1. + +RQ4: Is adaptive stopping really distribution-free? One benefit of adopting CP to quantify sampling deviation is distribution irrelevance. To testify to this, we consider sampling with various thresholds on all datasets to see if the sample amount remains consistent. The results in Figure 7 + +Table 2: The comparison of error margins with and without distribution prior transfer. Estimations are performed using the mixture of Beta-Binomial distributions, with 10 samples randomly drawn for estimation. In experiments, each dataset is chosen as the target dataset, and the left datasets in the same domain are used as source datasets. Bold denotes lower margin. Scores are in percent (\%). + +
DatasetMethodLlama-3.3-70BQwen-2.5-72BInternLM-2.5-20BGPT-3.5GPT-4
Hallucination Detection Datasets
HaluEvalw/o Transfer12.43 ± 0.8712.50 ± 0.9210.09 ± 0.6414.07 ± 0.7512.85 ± 0.83
w/ Transfer8.82 ± 0.429.19 ± 0.758.60 ± 0.648.88 ± 0.718.88 ± 0.86
TruthfulQAw/o Transfer15.30 ± 0.8113.88 ± 0.8513.17 ± 1.1112.54 ± 0.7013.21 ± 1.03
w/ Transfer3.37 ± 0.108.55 ± 0.0710.18 ± 0.1010.18 ± 0.829.66 ± 0.70
HalluDialw/o Transfer17.53 ± 0.8116.15 ± 0.6011.35 ± 0.8316.62 ± 0.7014.64 ± 0.85
w/ Transfer12.89 ± 0.7713.42 ± 0.538.72 ± 0.5423.79 ± 0.8418.77 ± 0.92
Reasoning Datasets
PRM800Kw/o Transfer15.02 ± 0.7812.85 ± 0.888.22 ± 0.589.27 ± 0.849.97 ± 0.53
w/ Transfer15.11 ± 0.6210.96 ± 0.998.46 ± 0.6010.55 ± 0.849.71 ± 1.00
BIG-benchw/o Transfer15.22 ± 0.7413.81 ± 0.829.44 ± 0.5314.39 ± 0.7413.31 ± 1.15
w/ Transfer12.69 ± 0.7414.28 ± 0.7910.00 ± 0.629.98 ± 0.6713.22 ± 0.69
TRAMw/o Transfer14.77 ± 0.8412.27 ± 0.6911.67 ± 0.7613.52 ± 0.8112.69 ± 1.26
w/ Transfer12.52 ± 0.9211.03 ± 1.0410.85 ± 0.9711.81 ± 1.0011.25 ± 0.57
Alignment Datasets
JudgeBenchw/o Transfer14.05 ± 0.8812.41 ± 0.6611.37 ± 0.798.23 ± 0.7512.32 ± 0.69
w/ Transfer9.45 ± 0.598.19 ± 0.668.03 ± 0.5414.36 ± 0.6815.30 ± 1.19
RewardBenchw/o Transfer12.73 ± 0.689.47 ± 1.0710.34 ± 0.6715.17 ± 0.9213.30 ± 0.77
w/ Transfer12.72 ± 0.3012.84 ± 0.4816.35 ± 0.3618.12 ± 0.3412.57 ± 0.38
LLMBarw/o Transfer16.97 ± 1.1015.91 ± 0.7010.03 ± 0.8817.00 ± 0.6412.90 ± 0.97
w/ Transfer8.03 ± 0.399.95 ± 0.308.61 ± 0.4121.94 ± 0.4217.70 ± 0.40
Scoring Datasets
ICE-Scorew/o Transfer14.08 ± 0.5311.90 ± 1.0519.59 ± 0.7812.11 ± 0.8213.98 ± 0.88
w/ Transfer11.32 ± 0.6611.99 ± 0.7619.25 ± 1.0510.63 ± 0.6612.30 ± 0.67
COMP-Analysisw/o Transfer14.85 ± 1.4510.83 ± 0.6010.29 ± 0.6010.22 ± 0.5316.18 ± 1.00
w/ Transfer15.29 ± 0.9112.28 ± 1.3810.23 ± 0.729.62 ± 0.5314.97 ± 0.82
Average
Averagew/o Transfer14.81 ± 0.8612.91 ± 0.8011.41 ± 0.7413.01 ± 0.7513.21 ± 0.91
w/ Transfer11.11 ± 0.5811.15 ± 0.7010.84 ± 0.6013.62 ± 0.6813.12 ± 0.74
+ +show only a slight variance of sampling amounts across datasets, demonstrating superior stability. This verifies that our adaptive stopping is truly distribution-free, and stable on diverse datasets. + +RQ5: Is CP-based Adaptive Stopping efficient? To validate the effectiveness of our CP-based adaptive stopping, we compare it against variance-based stopping. Specifically, we calculate the variance of sampling as + +$$ +\operatorname {V a r} (s a m p l i n g) = \frac {\alpha_ {r} \beta_ {r}}{(\alpha_ {r} + \beta_ {r}) ^ {2} (\alpha_ {r} + \beta_ {r} + 1)}, \tag {15} +$$ + +where $\alpha_{r}$ and $\beta_{r} = r - \alpha_{r}$ are the number of correct and wrong judgments in $r$ samples, respectively. + +As shown in Table 3, is consistently more effective for adaptive stopping under the same deviation threshold $\xi$ , which results in a reduced number of samples and achieves a reduction of up to $46.3\%$ . + +# 8 Conclusion + +We present BetaConform, a framework for efficient estimation of LLM ensemble judge distribution. As part of our framework, we propose a mixture of Beta-Binomial distributions to model the judgment distribution after examining the inaccuracy of the Binomial assumption. We design conformal prediction-based adaptive stopping for sampling, which monitors the sampling deviation and effectively determines the sample amount for estimation. When only limited sam + +Table 3: Comparison of variance-based adaptive stopping and ours. We compare the sample amount of both methods under the same threshold. Bold denotes less samples. + +
ThresholdξMethodsHaluEvalJudgeBenchPRM800KICE-Score
# Samples (↓)# Samples (↓)# Samples (↓)# Samples (↓)# Samples (↓)
ξ=0.06Variance36.8736.8726.0024.77
Ours35.3736.3730.4731.53
ξ=0.03Variance82.0974.4379.7681.47
Ours54.7253.9043.3245.27
ξ=0.01Variance194.72198.56147.22151.44
Ours109.06106.56101.2896.50
+ +plies are available, we incorporate a text similarity-based distribution prior transfer mechanism to improve the estimation accuracy. As shown by experiments, the conformal prediction-based adaptive stopping effectively guided the sampling. Our mixture of Beta-Binomial distributions significantly outperforms the common Binomial assumption. With the transfer mechanism, BetaConform can achieve high estimation precision with as few as 10 samples from the target dataset. + +# References + +Anastasios N. Angelopoulos, Stephen Bates, Adam Fisch, Lihua Lei, and Tal Schuster. Conformal risk control, 2023. URL https://arxiv.org/abs/2208.02814. +David Austen-Smith and Jeffrey S Banks. Information aggregation, rationality, and the condorcet jury theorem. American political science review, 90(1):34-45, 1996. +Frank B Baker. The basics of item response theory. ERIC, 2001. +Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. +Li Cai, Kilchan Choi, Mark Hansen, and Lauren Harrell. Item response theory. Annual Review of Statistics and Its Application, 3(1):297-321, 2016. +Zheng Cai, Maosong Cao, Haojiong Chen, Kai Chen, Keyu Chen, Xin Chen, Xun Chen, Zehui Chen, Zhi Chen, Pei Chu, et al. Internl m2 technical report. arXiv preprint arXiv:2403.17297, 2024. +Nitay Calderon, Roi Reichart, and Rotem Dror. The alternative annotator test for llm-as-a-judge: How to statistically justify replacing human annotators with llms, 2025. URL https://arxiv.org/abs/2501.10970. +Beiduo Chen, Xinpeng Wang, Siyao Peng, Robert Litschko, Anna Korhonen, and Barbara Plank. "seeing the big through the small": Can llms approximate human judgment distributions on nli from a few explanations? arXiv preprint arXiv:2406.17600, 2024. +Nicolas De Condorcet et al. Essai sur l'application de l'analyse à la probabilité des décisions rendues à la pluralité des voix. Cambridge University Press, 2014. +Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. +Yann Dubois, Balázs Galambosi, Percy Liang, and Tatsunori B. Hashimoto. Length-controlled alpacaeval: A simple way to debias automatic evaluators, 2024. URL https://arxiv.org/abs/2404.04475. +Matteo Fontana, Gianluca Zeni, and Simone Vantini. Conformal prediction: a unified review of theory and new challenges. Bernoulli, 29(1):1-23, 2023. +Robert J Harvey and Allen L Hammer. Item response theory. The Counseling Psychologist, 27(3): 353-383, 1999. +Ryan Koo, Minhwa Lee, Vipul Raheja, Jong Inn Park, Zae Myung Kim, and Dongyeop Kang. Benchmarking cognitive biases in large language models as evaluators, 2024. URL https://arxiv.org/abs/2309.17012. +Nathan Lambert, Valentina Pyatkin, Jacob Morrison, LJ Miranda, Bill Yuchen Lin, Khyathi Chandu, Nouha Dziri, Sachin Kumar, Tom Zick, Yejin Choi, et al. Rewardbench: Evaluating reward models for language modeling. arXiv preprint arXiv:2403.13787, 2024. +Chankyu Lee, Rajarshi Roy, Mengyao Xu, Jonathan Raiman, Mohammad Shoeybi, Bryan Catanzaro, and Wei Ping. Nv-embed: Improved techniques for training llms as generalist embedding models, 2025. URL https://arxiv.org/abs/2405.17428. +Harrison Lee, Samrat Phatale, Hassan Mansoor, Thomas Mesnard, Johan Ferret, Kellie Lu, Colton Bishop, Ethan Hall, Victor Carbune, Abhinav Rastogi, and Sushant Prakash. Rlaif vs. rlhf: Scaling reinforcement learning from human feedback with ai feedback, 2024. URL https://arxiv.org/abs/2309.00267. +Junyi Li, Xiaoxue Cheng, Wayne Xin Zhao, Jian-Yun Nie, and Ji-Rong Wen. Halueval: A large-scale hallucination evaluation benchmark for large language models. arXiv preprint arXiv:2305.11747, 2023a. + +Lei Li, Yuancheng Wei, Zhihui Xie, Xuqing Yang, Yifan Song, Peiyi Wang, Chenxin An, Tianyu Liu, Sujian Li, Bill Yuchen Lin, Lingpeng Kong, and Qi Liu. Vlrewardbench: A challenging benchmark for vision-language generative reward models, 2024. URL https://arxiv.org/abs/2411.17451. +Yuhui Li, Fangyun Wei, Jinjing Zhao, Chao Zhang, and Hongyang Zhang. Rain: Your language models can align themselves without finetuning, 2023b. URL https://arxiv.org/abs/2309.07124. +Tian Liang, Zhiwei He, Wenxiang Jiao, Xing Wang, Yan Wang, Rui Wang, Yujiu Yang, Shuming Shi, and Zhaopeng Tu. Encouraging divergent thinking in large language models through multi-agent debate. arXiv preprint arXiv:2305.19118, 2023. +Hunter Lightman, Vineet Kosaraju, Yura Burda, Harri Edwards, Bowen Baker, Teddy Lee, Jan Leike, John Schulman, Ilya Sutskever, and Karl Cobbe. Let's verify step by step. arXiv preprint arXiv:2305.20050, 2023. +Stephanie Lin, Jacob Hilton, and Owain Evans. Truthfulqa: Measuring how models mimic human falsehoods. arXiv preprint arXiv:2109.07958, 2021. +Yang Liu, Dan Iter, Yichong Xu, Shuohang Wang, Ruochen Xu, and Chenguang Zhu. G-eval: NLG evaluation using gpt-4 with better human alignment. In Houda Bouamor, Juan Pino, and Kalika Bali, editors, Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 2511-2522, Singapore, December 2023. Association for Computational Linguistics. doi: 10.18653/v1/2023.emnlp-main.153. URL https://aclanthology.org/2023.emnlp-main.153/. +Wen Luo, Tianshu Shen, Wei Li, Guangyue Peng, Richeng Xuan, Houfeng Wang, and Xi Yang. Halludial: A large-scale benchmark for automatic dialogue-level hallucination evaluation. arXiv preprint arXiv:2406.07070, 2024. +Sewon Min, Kalpesh Krishna, Xinxi Lyu, Mike Lewis, Wen-tau Yih, Pang Koh, Mohit Iyyer, Luke Zettlemoyer, and Hannaneh Hajishirzi. FActScore: Fine-grained atomic evaluation of factual precision in long form text generation. In Houda Bouamor, Juan Pino, and Kalika Bali, editors, Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 12076-12100, Singapore, December 2023. Association for Computational Linguistics. doi: 10.18653/v1/2023.emnlp-main.741. URL https://aclanthology.org/2023.emnlp-main.741/. +Christopher Mohri and Tatsunori Hashimoto. Language models with conformal factuality guarantees, 2024. URL https://arxiv.org/abs/2402.10978. +Behrad Moniri, Hamed Hassani, and Edgar Dobriban. Evaluating the performance of large language models via debates, 2024. URL https://arxiv.org/abs/2406.11044. +OpenAI, Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, Red Avila, Igor Babuschkin, Suchir Balaji, Valerie Balcom, Paul Baltescu, Haiming Bao, Mohammad Bavarian, Jeff Belgium, Irwan Bello, Jake Berdine, Gabriel Bernadett-Shapiro, Christopher Berner, Lenny Bogdonoff, Oleg Boiko, Madelaine Boyd, Anna-Luisa Brakman, Greg Brockman, Tim Brooks, Miles Brundage, Kevin Button, Trevor Cai, Rosie Campbell, Andrew Cann, Brittany Carey, Chelsea Carlson, Rory Carmichael, Brooke Chan, Che Chang, Fotis Chantzis, Derek Chen, Sully Chen, Ruby Chen, Jason Chen, Mark Chen, Ben Chess, Chester Cho, Casey Chu, Hyung Won Chung, Dave Cummings, Jeremiah Currier, Yunxing Dai, Cory Decareaux, Thomas Degry, Noah Deutsch, Damien Deville, Arka Dhar, David Dohan, Steve Dowling, Sheila Dunning, Adrien Ecoffet, Atty Eleti, Tyna Eloundou, David Farhi, Liam Fedus, Niko Felix, Simón Posada Fishman, Juston Forte, Isabella Fulford, Leo Gao, Elie Georges, Christian Gibson, Vik Goel, Tarun Gogineni, Gabriel Goh, Rapha Gontijo-Lopes, Jonathan Gordon, Morgan Grafstein, Scott Gray, Ryan Greene, Joshua Gross, Shixiang Shane Gu, Yufei Guo, Chris Hallacy, Jesse Han, Jeff Harris, Yuchen He, Mike Heaton, Johannes Heidecke, Chris Hesse, Alan Hickey, Peter Hoeschele, Brandon Houghton, Kenny Hsu, Shengli Hu, Xin Hu, Joost Huizinga, Shantanu Jain, Shawn Jain, Joanne Jang, Angela Jiang, Roger Jiang, Haozhun Jin, Denny Jin, Shino Jomoto, Billie Jonn, Heewoo + +Jun, Tomer Kaftan, Łukasz Kaiser, Ali Kamali, Ingmar Kanitscheider, Nitish Shirish Keskar, Tabarak Khan, Logan Kilpatrick, Jong Wook Kim, Christina Kim, Yongjik Kim, Jan Hendrik Kirchner, Jamie Kiros, Matt Knight, Daniel Kokotajlo, Łukasz Kondraciuk, Andrew Kondrich, Aris Konstantinidis, Kyle Kosic, Gretchen Krueger, Vishal Kuo, Michael Lampe, Ikai Lan, Teddy Lee, Jan Leike, Jade Leung, Daniel Levy, Chak Ming Li, Rachel Lim, Molly Lin, Stephanie Lin, Mateusz Litwin, Theresa Lopez, Ryan Lowe, Patricia Lue, Anna Makanju, Kim Malfacini, Sam Manning, Todor Markov, Yaniv Markovski, Bianca Martin, Katie Mayer, Andrew Mayne, Bob McGrew, Scott Mayer McKinney, Christine McLeavey, Paul McMillan, Jake McNeil, David Medina, Aalok Mehta, Jacob Menick, Luke Metz, Andrey Mishchenko, Pamela Mishkin, Vinnie Monaco, Evan Morikawa, Daniel Mossing, Tong Mu, Mira Murati, Oleg Murk, David Mely, Ashvin Nair, Reiichiro Nakano, Rajeev Nayak, Arvind Neelakantan, Richard Ngo, Hyeonwoo Noh, Long Ouyang, Cullen O'Keefe, Jakub Pachocki, Alex Paino, Joe Palermo, Ashley Pantuliano, Giambattista Parascandolo, Joel Parish, Emy Parparita, Alex Passos, Mikhail Pavlov, Andrew Peng, Adam Perelman, Filipe de Avila Belbute Peres, Michael Petrov, Henrique Ponde de Oliveira Pinto, Michael, Pokorny, Michelle Pokrass, Vitchyr H. Pong, Tolly Powell, Alethea Power, Boris Power, Elizabeth Proehl, Raul Puri, Alec Radford, Jack Rae, Aditya Ramesh, Cameron Raymond, Francis Real, Kendra Rimbach, Carl Ross, Bob Rotsted, Henri Roussez, Nick Ryder, Mario Saltarelli, Ted Sanders, Shibani Santurkar, Girish Sastry, Heather Schmidt, David Schnurr, John Schulman, Daniel Selsam, Kyla Sheppard, Toki Sherbakov, Jessica Shieh, Sarah Shoker, Pranav Shyam, Szymon Sidor, Eric Sigler, Maddie Simens, Jordan Sitkin, Katarina Slama, Ian Sohl, Benjamin Sokolowsky, Yang Song, Natalie Staudacher, Felipe Petroski Such, Natalie Summers, Ilya Sutskever, Jie Tang, Nikolas Tezak, Madeleine B. Thompson, Phil Tillet, Amin Tootoonchian, Elizabeth Tseng Preston Tuggle Nick Turley Jerry Tworek Juan Felipe Ceron Uribe Andrea Vallone Arun Vijayvergiya Chelsea Voss Carroll Wainwright Justin Jay Wang Alvin Wang Ben Wang Jonathan Ward Jason Wei CJ Weinmann Akila Welihinda Peter Welinder Jiayi Weng Lilian Weng Matt Wiethoff Dave Willner Clemens Winter Samuel Wolrich Hannah Wong Lauren Workman Sherwin Wu Jeff Wu Michael Wu Kai Xiao Tao Xu Sarah Yoo Kevin Yu Qiming Yuan Wojciech Zaremba Rowan Zellers Chong Zhang Marvin Zhang Shengjia Zhao Tianhao Zheng Juntang Zhuang William Zhuk and Barret Zoph. Gpt-4 technical report 2024. URL https://arxiv.org/abs/2303.08774. + +Arjun Panickssery, Samuel R. Bowman, and Shi Feng. Llm evaluators recognize and favor their own generations, 2024. URL https://arxiv.org/abs/2404.13076. +Junsoo Park, Seungyeon Jwa, Meiying Ren, Daeyoung Kim, and Sanghyuk Choi. Offsetbias: Leveraging debiased data for tuning evaluators, 2024. URL https://arxiv.org/abs/2407.06551. +Jiaxing Qiu, Dongliang Guo, Papini Natalie, Peace Noelle, Levinson Cheri, and Teague R. Henry. Ensemble of large language models for curated labeling and rating of free-text data, 2025. URL https://arxiv.org/abs/2501.08413. +Philipp Schoenegger, Indre Tuminauskaite, Peter S. Park, and Philip E. Tetlock. Wisdom of the silicon crowd: LIm ensemble prediction capabilities rival human crowd accuracy, 2024. URL https://arxiv.org/abs/2402.19379. +Glenn Shafer and Vladimir Vovk. A tutorial on conformal prediction. Journal of Machine Learning Research, 9(3), 2008. +Aarohi Srivastava, Abhinav Rastogi, Abhishek Rao, Abu Awal Md Shoeb, Abubakar Abid, Adam Fisch, Adam R Brown, Adam Santoro, Aditya Gupta, Adrià Garriga-Alonso, et al. Beyond the imitation game: Quantifying and extrapolating the capabilities of language models. arXiv preprint arXiv:2206.04615, 2022. +Sijun Tan, Siyuan Zhuang, Kyle Montgomery, William Y Tang, Alejandro Cuadron, Chenguang Wang, Raluca Ada Popa, and Ion Stoica. Judgebench: A benchmark for evaluating llm-based judges. arXiv preprint arXiv:2410.12784, 2024. +Peiyi Wang, Lei Li, Liang Chen, Zefan Cai, Dawei Zhu, Binghuai Lin, Yunbo Cao, Qi Liu, Tianyu Liu, and Zhifang Sui. Large language models are not fair evaluators, 2023. URL https://arxiv.org/abs/2305.17926. + +Tianlu Wang, Ilia Kulikov, Olga Golovneva, Ping Yu, Weizhe Yuan, Jane Dwivedi-Yu, Richard Yuanzhe Pang, Maryam Fazel-Zarandi, Jason Weston, and Xian Li. Self-taught evaluators, 2024. URL https://arxiv.org/abs/2408.02666. +Yuqing Wang and Yun Zhao. Tram: Benchmarking temporal reasoning for large language models. arXiv preprint arXiv:2310.00835, 2023. +Tianhao Wu, Weizhe Yuan, Olga Golovneva, Jing Xu, Yuandong Tian, Jiantao Jiao, Jason Weston, and Sainbayar Sukhbaatar. Meta-rewarding language models: Self-improving alignment with llm-as-a-meta-judge, 2024. URL https://arxiv.org/abs/2407.19594. +Yasin Abbasi Yadmori, Ilja Kuzborskij, David Stutz, András György, Adam Fisch, Arnaud Doucet, Iuliya Beloshapka, Wei-Hung Weng, Yao-Yuan Yang, Csaba Szepesvári, Ali Taylan Cemgil, and Jenad Tomasev. Mitigating llm hallucinations via conformal abstention, 2024. URL https://arxiv.org/abs/2405.01563. +An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, et al. Qwen2. 5 technical report. arXiv preprint arXiv:2412.15115, 2024. +Zhuohao Yu, Chang Gao, Wenjin Yao, Yidong Wang, Wei Ye, Jindong Wang, Xing Xie, Yue Zhang, and Shikun Zhang. Kieval: A knowledge-grounded interactive evaluation framework for large language models, 2024. URL https://arxiv.org/abs/2402.15043. +Tongxin Yuan, Zhiwei He, Lingzhong Dong, Yiming Wang, Ruijie Zhao, Tian Xia, Lizhen Xu, Binglin Zhou, Fangqi Li, Zhuosheng Zhang, et al. R-judge: Benchmarking safety risk awareness for llm agents. arXiv preprint arXiv:2401.10019, 2024a. +Weizhe Yuan, Richard Yuanzhe Pang, Kyunghyun Cho, Sainbayar Sukhbaatar, Jing Xu, and Jason Weston. Self-rewarding language models. arXiv preprint arXiv:2401.10020, 2024b. +Zhiyuan Zeng, Jiatong Yu, Tianyu Gao, Yu Meng, Tanya Goyal, and Danqi Chen. Evaluating large language models at evaluating instruction following. arXiv preprint arXiv:2310.07641, 2023. +Chen Zhang, Luis Fernando D'Haro, Yiming Chen, Malu Zhang, and Haizhou Li. A comprehensive analysis of the effectiveness of large language models as automatic dialogue evaluators. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 19515-19524, 2024. +Zhenru Zhang, Chujie Zheng, Yangzhen Wu, Beichen Zhang, Runji Lin, Bowen Yu, Dayiheng Liu, Jingren Zhou, and Junyang Lin. The lessons of developing process reward models in mathematical reasoning, 2025. URL https://arxiv.org/abs/2501.07301. +Ruochen Zhao, Wenxuan Zhang, Yew Ken Chia, Weiwen Xu, Deli Zhao, and Lidong Bing. Autoarena: Automating llm evaluations with agent peer battles and committee discussions, 2024a. URL https://arxiv.org/abs/2405.20267. +Yuwei Zhao, Ziyang Luo, Yuchen Tian, Hongzhan Lin, Weixiang Yan, Annan Li, and Jing Ma. Codejudge-eval: Can large language models be good judges in code understanding?, 2024b. URL https://arxiv.org/abs/2408.10718. +Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, et al. Judging llm-as-a-judge with mt-bench and chatbot arena. Advances in Neural Information Processing Systems, 36:46595-46623, 2023. +Yuxuan Zhou, Margret Keuper, and Mario Fritz. Balancing diversity and risk in llm sampling: How to select your method and parameter for open-ended text generation. arXiv preprint arXiv:2408.13586, 2024. +Terry Yue Zhuo. Ice-score: Instructing large language models to evaluate code. arXiv preprint arXiv:2304.14317, 2023. + +# A BetaConform + +In this section, we introduce BetaConform, a framework designed for the efficient estimation of judgment distributions, as illustrated in Figure 2 and Algorithm 1. The framework operates in two scenarios: when only limited samples are available on the target dataset, and when a larger number of samples can be collected. In the former case, BetaConform leverages prior distributions from source datasets to enhance estimation. In the latter, it employs adaptive stopping during iterative sampling to balance sample efficiency and estimation accuracy. + +(1) When only a small number of samples are available from the target dataset, BetaConform follows + +# Algorithm 1 BetaConform + +1: Input: target dataset $D_0$ , source datasets $D_1, \ldots, D_m$ , judges $\mathcal{J} = \{J_1, \ldots, J_k\}$ , EM algorithm $\mathrm{EM}(\cdot)$ +2: Output: distribution parameters $\Omega$ on the target dataset +3: if limited samples in $D_0$ then +4: Compute distribution parameters on $D_0$ +5: Compute parameters of distributions on $D_{1},\ldots ,D_{m}$ +6: Compute transfer weights by Equation (13) +7: $\Omega \gets$ Compute transferred parameters by Eq. (14) +8: else +9: Initial $D \gets \{\}$ , $q_{1 - \epsilon}^{0} \gets -\infty$ +10: while Equation (9) is not satisfied do +11: Add a sample from $D_0$ to $D$ and update $q_{1 - \epsilon}^{|D|}$ +12: end while +13: $\Omega \gets$ Compute distribution parameters on samples $D$ +14: end if +15: return $\Omega$ + +these steps: 1 First, it estimates the mixture of Beta-Binomial distributions using the available samples. 2 Next, it incorporates prior knowledge by transferring distributions from source datasets. Specifically, it estimates the distributions on the source datasets using all available samples and calculates transfer weights based on Equation 13. 3 Finally, the distributions from the source datasets are aggregated using Equation 14 to produce an enhanced estimation for the target dataset. + +(2) When the target dataset contains a large number of unlabeled samples, BetaConform employs the following process: ① It uses a conformal prediction (CP)-based adaptive stopping strategy to guide the labeling process. ② During iterative sampling, batches of samples are drawn and labeled, while the variation in the nonconformity score is monitored. The sampling process stops when the variation falls below a predefined threshold. ③ Once sufficient labeled samples are collected, the mixture of Beta-Binomial distributions is directly estimated using these samples. + +# B Proofs + +# B.1 Determination of Sample Amount. + +To derive a theoretical estimation of the sample amount for the adaptive stopping criteria above, we utilize the fundamental statistical properties of variance reduction with increasing sample size. Specifically, for i.i.d samples, the variance of the quantile decreases as: + +$$ +\operatorname {V a r} \left(q _ {1 - \epsilon} ^ {r}\right) \propto \frac {1}{r \cdot f \left(q _ {1 - \epsilon}\right) ^ {2}}, \tag {16} +$$ + +where $f(q_{1 - \epsilon})$ is the density function at the quantile. The standard deviation of the estimator, which determines the variability of the quantile estimate, thus decays as: + +$$ +\operatorname {S t d D e v} \left(q _ {1 - \epsilon} ^ {r}\right) \propto \frac {1}{\sqrt {r}}. \tag {17} +$$ + +By the asymptotic theory of quantile estimation, for a large enough number of samples $r$ , the empirical quantile $q_{1 - \epsilon}^{r}$ converges to the quantile on the whole dataset $q_{1 - \epsilon}$ with a known distribution based on Bahadur's representation: + +$$ +\sqrt {r} \left(q _ {1 - \epsilon} ^ {r} - q _ {1 - \epsilon}\right) \sim \mathcal {N} \left(0, \frac {\epsilon (1 - \epsilon)}{f \left(q _ {1 - \epsilon}\right) ^ {2}}\right), \tag {18} +$$ + +This implies: + +$$ +q _ {1 - \epsilon} ^ {r} = q _ {1 - \epsilon} + O _ {p} \left(\frac {1}{\sqrt {r}}\right), \tag {19} +$$ + +where $O_{p}(\cdot)$ denotes the order in probability. Thus, we can determine that the quantile itself decays as: + +$$ +q _ {1 - \epsilon} ^ {r} - q _ {1 - \epsilon} = O _ {p} \left(\frac {1}{\sqrt {r}}\right). \tag {20} +$$ + +This decay behavior shows that as $r$ increases, the estimated quantile approaches the theoretical quantile $q_{1 - \epsilon}$ , reflecting decreasing sampling deviation by using more samples. We will use this property to derive the relationship between the stopping criteria and the sample size $r$ . From the stopping criteria in Equation (9), + +$$ +\left| q _ {1 - \epsilon} ^ {r} - q _ {1 - \epsilon} ^ {r - 1} \right| \leq \xi . \tag {21} +$$ + +According to the calculations in Equation (20), we can rewrite the bound for $q_{1 - \epsilon}^{r - 1}$ as + +$$ +q _ {1 - \epsilon} ^ {r - 1} - q _ {1 - \epsilon} = O _ {p} \left(\frac {1}{\sqrt {r - 1}}\right). \tag {22} +$$ + +Thus we have + +$$ +\left| q _ {1 - \epsilon} ^ {r} - q _ {1 - \epsilon} ^ {r - 1} \right| = O _ {p} \left(\frac {1}{\sqrt {r}} - \frac {1}{\sqrt {r - 1}}\right). \tag {23} +$$ + +This suggests to meet Equation (9), it requires + +$$ +\tau \left(\frac {1}{\sqrt {r - 1}} - \frac {1}{\sqrt {r}}\right) < \xi , \tag {24} +$$ + +which proves Equation (10). + +# B.2 Error Rate with Adaptive Sampling + +In this section we develop a theoretical estimation of the error bound for adaptive sampling. We first consider the base case and as shown in Equation (4), we know that the mixture distribution error rate is: + +$$ +\tilde {P} _ {\mathrm {B B}} = w \sum_ {s = 0} ^ {\lceil k / 2 \rceil - 1} \binom {k} {s} \frac {\mathrm {B} (s + \alpha_ {1} , k - s + \beta_ {1})}{\mathrm {B} (\alpha_ {1} , \beta_ {1})} + (1 - w) \sum_ {s = 0} ^ {\lceil k / 2 \rceil - 1} \binom {k} {s} \frac {\mathrm {B} (s + \alpha_ {2} , k - s + \beta_ {2})}{\mathrm {B} (\alpha_ {2} , \beta_ {2})} \tag {25} +$$ + +The adaptive stopping criterion is given by Equation (9): + +$$ +\left| q _ {1 - \epsilon} ^ {r} - q _ {1 - \epsilon} ^ {r - 1} \right| \leq \xi . \tag {26} +$$ + +The sample size requirement is given by Equation (10): + +$$ +\tau \left(\frac {1}{\sqrt {r - 1}} - \frac {1}{\sqrt {r}}\right) \leq \xi . \tag {27} +$$ + +Based on the two equations and large number theory, we know that the difference between the quantile on samples $q_{1 - \epsilon}^{r}$ and the quantile on the whole dataset $q_{1 - \epsilon}$ decays proportionally to $\frac{\tau}{\sqrt{r}}$ . In addition, the non-conformity score $s_i$ is defined in Equation (7): + +$$ +s _ {i} = \operatorname {s c o r e} \left(S _ {i}\right) = \left| S _ {i} - \mathbb {E} [ S ] \right|, \tag {28} +$$ + +where $S_{i}$ is the number of correct judgments in the $i$ -th sample. As the $(1 - \epsilon)$ -quantile of the sorted scores $s_1 < \ldots < s_r$ at stopping time with $r$ samples is: + +$$ +q _ {1 - \epsilon} ^ {r} = s _ {\lceil (1 - \epsilon) \cdot (r + 1) \rceil}. \tag {29} +$$ + +When the stopping criterion is met, this implies the confidence region for $\mathbb{E}[S]$ has stabilized and the following holds: + +$$ +\mathbb {P} \left(\left| S _ {i} - \mathbb {E} [ S ] \right| \leq q _ {1 - \epsilon} ^ {r}\right) = 1 - \epsilon . \tag {30} +$$ + +For the Beta-Binomial mixture model, $\mathbb{E}[S]$ relates to the error rate via: + +$$ +\bar {P} _ {\mathrm {B B}} = \mathbb {P} (S < \lceil k / 2 \rceil). \tag {31} +$$ + +We will use the quantile stability argument as follows. For a sequence of independent samples $\{S_1,\dots,S_r\}$ , let $s_i$ be the non-conformity score defined as: + +$$ +s _ {i} = \operatorname {s c o r e} \left(S _ {i}\right) = \left| S _ {i} - \mathbb {E} [ S ] \right|, \tag {32} +$$ + +where $S_{i}$ is the number of correct judgments in the $i$ -th sample. By the theory of quantile estimation, for a large enough number of samples $r$ , the empirical quantile $q_{1 - \epsilon}^{r}$ converges to the population quantile $q_{1 - \epsilon}$ with a known distribution: + +$$ +\sqrt {r} \left(q _ {1 - \epsilon} ^ {r} - q _ {1 - \epsilon}\right) \sim \mathcal {N} \left(0, \frac {\epsilon (1 - \epsilon)}{f \left(q _ {1 - \epsilon}\right) ^ {2}}\right), \tag {33} +$$ + +where $f(\cdot)$ is the density function. This implies: + +$$ +q _ {1 - \epsilon} ^ {r} = q _ {1 - \epsilon} + O _ {p} \left(\frac {1}{\sqrt {r}}\right), \tag {34} +$$ + +where $O_{p}(\cdot)$ denotes the order in probability. As the $(1 - \epsilon)$ -quantile of the sorted scores $s_1 < \ldots < s_r$ at stopping time with $r$ samples is: + +$$ +q _ {1 - \epsilon} ^ {r} = s _ {\lceil (1 - \epsilon) \cdot (r + 1) \rceil}. \tag {35} +$$ + +When the stopping criterion is met, this implies the confidence region for $\mathbb{E}[S]$ has stabilized and the following holds: + +$$ +\mathbb {P} \left(\left| S _ {i} - \mathbb {E} [ S ] \right| \leq q _ {1 - \epsilon} ^ {r}\right) = 1 - \epsilon . \tag {36} +$$ + +For the Beta-Binomial mixture model, $\mathbb{E}[S]$ relates to the error rate via: + +$$ +\tilde {P} _ {\mathrm {B B}} = \mathbb {P} (S < \lceil k / 2 \rceil). \tag {37} +$$ + +By the quantile stability argument above, we have the bound: + +$$ +(1 - \min (\xi , \frac {\tau}{\sqrt {r}})) \mathbb {E} [ S ] _ {\mathrm {B B}} < \mathbb {E} [ S ] _ {\text {a d a p t}} < (1 + \min (\xi , \frac {\tau}{\sqrt {r}})) \mathbb {E} [ S ] _ {\mathrm {B B}} \tag {38} +$$ + +The error probability of $\tilde{P}_{\mathrm{BB}}$ is defined using the Beta-Binomial cumulative distribution function: + +$$ +\tilde {P} _ {\mathrm {B B}} = \mathbb {P} (S < \lceil k / 2 \rceil) = F _ {\mathrm {B B}} (\lceil k / 2 \rceil - 1), \tag {39} +$$ + +where $F_{\mathrm{BB}}$ is the Beta-Binomial cumulative distribution function. Since $F_{\mathrm{BB}}$ is monotonically increasing, the error probability $\tilde{P}_{\mathrm{adapt}}$ follows the same proportional bound. + +$$ +\left(1 - \min \left(\xi , \frac {\tau}{\sqrt {r}}\right)\right) \tilde {P} _ {\mathrm {B B}} < \tilde {P} _ {\text {a d a p t}} < \left(1 + \min \left(\xi , \frac {\tau}{\sqrt {r}}\right)\right) \tilde {P} _ {\mathrm {B B}}. \tag {40} +$$ + +Therefore, we have: + +$$ +\tilde {P} _ {\text {a d a p t}} = \left(1 \pm \min \left(\xi , \frac {\tau}{\sqrt {r}}\right)\right) \tilde {P} _ {\mathrm {B B}}. \tag {41} +$$ + +# C Implementation Details + +In this section, we elaborate on the implementation details of BetaConform. + +We evaluate LLM ensembles with $k \in 1,3,5,7,9,11$ models, including GPT-3.5 [Brown et al., 2020], GPT-4 [OpenAI et al., 2024], Llama-3.3-70B [Dubey et al., 2024], Qwen-2.5-72B [Yang et al., 2024], and InternLM-2.5-20B [Cai et al., 2024]. The experiments cover four domains: hallucination detection (HaluEval Li et al., 2023a, TruthfulQA Lin et al., 2021, HalluDial Luo et al., 2024), reasoning (PRM800K Lightman et al., 2023, BIG-bench Srivastava et al., 2022, TRAM Wang and Zhao, 2023), scoring (ICE-Score Zhuo, 2023, Comp-Analysis Zhang et al., 2024), and alignment (JudgeBench Tan et al., 2024, RewardBench Lambert et al., 2024, LLMBar Zeng et al., 2023). + +For all experiments, the sampling temperature of LLMs is set to 1, and the random seeds are not fixed. The randomness comes from the Top - P sampling of token generation. Each experiment is repeated 30 times to compute the mean and standard deviation of the error margin. The adaptive stopping threshold is set to $\xi = 0.03$ and $\tau = 25$ , requiring at least $r \geq 56$ samples to meet the stopping criteria. + +# D Additional Experiments + +In Table 4, we conduct ablation studies on our distribution transfer. Compared to ablated variants, our full design achieves the smallest error margin, indicating the effectiveness of our transfer design. + +Table 4: The ablation study of BetaConform distribution prior transfer. $\mathbf{1}\log (r_i)\rightarrow r_i$ means the first term $\log (r_i)$ in Eq. 14 is replaced with $r_i$ to still assign a larger dataset higher weight while not considering source datasets could be magnitudes larger. $\bullet \mathrm{CosSim}(\bar{E}_0,\bar{E}_i)\to \frac{1}{|\bar{E}_0 - \bar{E}_i|_2}$ refers to replacing the cosine similarity to measure the source datasets and the target dataset with the reciprocal of the Euclidean distance between the embeddings of the two datasets. This still assigns more similar datasets higher weights. $\bullet \mathrm{No}\sigma (\cdot)$ means the transfer weight is computed as $\lambda_{i} = \log (r_{i})\cdot \mathrm{CosSim}(\bar{E}_{0},\bar{E}_{i})$ , without using the sigmoid function $\sigma (\cdot)$ to reduce the weight of low similarity datasets + +
DatasetAblationLlama-3.3-70BQwen-2.5-72BInternLM-20B
Error MarginError MarginError Margin
HaluEvallog(r_i) → ri10.94 ± 0.579.53 ± 0.7011.16 ± 0.75
CosSim(¯E_0,¯E_i) → 1/¯E_0–¯E_i|211.90 ± 0.8513.17 ± 0.6810.29 ± 0.80
No σ(·)10.04 ± 0.2323.03 ± 0.128.45 ± 0.10
Ours8.82 ± 0.429.19 ± 0.758.60 ± 0.64
TruthfulQAlog(r_i) → ri13.47 ± 0.6611.17 ± 1.1510.65 ± 0.89
CosSim(¯E_0,¯E_i) → 1/¯E_0–¯E_i|215.13 ± 0.7113.14 ± 0.9611.03 ± 0.80
No σ(·)6.87 ± 0.0116.52 ± 0.0312.47 ± 0.06
Ours3.37 ± 0.108.55 ± 0.0710.18 ± 0.10
HalluDiallog(r_i) → ri13.55 ± 0.5815.43 ± 0.8610.42 ± 1.00
CosSim(¯E_0,¯E_i) → 1/¯E_0–¯E_i|215.54 ± 0.5915.89 ± 0.6510.47 ± 0.67j
No σ(·)12.39 ± 0.0016.61 ± 0.0913.00 ± 0.07
Ours12.89 ± 0.7713.42 ± 0.538.72 ± 0.54
JudgeBenchlog(r_i) → ri25.97 ± 0.0321.23 ± 0.0415.46 ± 0.06
CosSim(¯E_0,¯E_i) → 1/¯E_0–¯E_i|214.43 ± 1.1211.26 ± 0.9911.47 ± 0.74
No σ(·)24.57 ± 0.4419.26 ± 0.1310.42 ± 0.08
Ours9.45 ± 0.598.19 ± 0.668.03 ± 0.54
RewardBenchlog(r_i) → ri15.00 ± 0.0117.33 ± 0.0220.32 ± 0.01
CosSim(¯E_0,¯E_i) → 1/¯E_0–¯E_i|213.29 ± 0.8714.48 ± 0.4516.75 ± 0.34
No σ(·)12.88 ± 0.5913.74 ± 0.4816.45 ± 0.26
Ours12.72 ± 0.3012.84 ± 0.4816.35 ± 0.36
LLMBarlog(r_i) → ri13.88 ± 0.0115.88 ± 0.0115.45 ± 0.01
CosSim(¯E_0,¯E_i) → 1/¯E_0–¯E_i|216.27 ± 0.8115.55 ± 0.8311.90 ± 1.07
No σ(·)9.53 ± 0.1113.65 ± 0.0112.58 ± 0.01
Ours8.03 ± 0.399.95 ± 0.308.61 ± 0.41
+ +# E Python Implementation + +Below we provide the Python-style code for the implementation of our methods + +Listing 1: Adaptive Conformal Sampling +```python +import math +import random +import numpy as np +#---Helper Function for Conformal Sampling --- +def _nonconformity_score_abs_diff_mean(value，mean_value)： ""Calculates L1 distance between a value and the mean as a nonconformity measure. return abs(value - mean_value) +#---Core Function 2:Adaptive Conformal Sampling--- +def run_adaptive_conformal_sampling_for_k_value( full_dataset_items, k_value_num_models, num_samples_per_batch, +``` + +```txt +max_batches, epsilon_conformal $= 0.05$ convergence_threshold_q_diff $= 0.01$ min_batches_before_stopping_check $= 5$ +} : +" Performas adaptive sampling for a fixed k-value (number of models) using conformal prediction. Samples are drawn in batches until the width of the conformal interval (related to q-value) stabilizes. +Args: full_dataset_items (list of lists): Each inner list contains binary outcomes for a data point across all available models. k_value_num_models (int): Number of models/outcomes to consider from the start of each item. num_samples_per_batch (int): Number of items to sample per batch. max_batches (int): Maximum number of batches to draw. epsilon_conformal(float): Significance level for conformal prediction (e.g., 0.05 for $95\%$ interval). convergence_threshold_q_diff(float): Threshold for q-value change to determine stopping. min_batches_before_stopping_check (int): Minimum batches before checking q-value convergence. +Returns: tuple: (collected_success_counts_S, final_q_value, num_batchesProcessed,sampledIndices_overall) -collected_success_counts_S:List of success counts for all sampled items. final_q_value: q-value from conformal prediction at stopping or max batches. -num_batches_processed: Actual number of batches processed. -sampledIndices_overall: List of original indices of the sampled items. +if not full_dataset_items: return [], None, 0, [] if not $(0 < \mathrm{k\_value\_num\_models}\leq = \mathrm{len}(f u l l_{-}$ dataset_items[0]): raise ValueError(f"Invalid k_value_num_models:{ k_value_num_models}]") +all_collected_S_values $= [ ]$ #Stores S_i $=$ sum(item[: k_value_num_models]) for calibration set +qprevious $=$ None +final_q_value $=$ None +indexed_full_dataset $=$ list(indexerate(full_dataset_items)) available Indices_for_sampling $=$ list(range(len( indexed_full_dataset))) +sampledIndices_overall $=$ [] +for batchidx in range(max_batches): if lenavailable Indices_for_sampling)< num_samples_per_batch : if not available Indices_for_sampling: break # No more samples available # If remaining samples are less than a batch, sample all remaining actual_samples.this_batch $=$ len( availableIndices_for_sampling) +``` + +Listing 2: Mixture of Beta Distributions Fitting via EM +```python +else: actual_samples.this_batch = num_samples_per_batch +# Sample indices for the current batch without replacement from available indices +chosen_poolIndices = random.sample(available Indices_for_sampling, actual_samples.this_batch) +current_batch_items = [] +current_batch ORIGINAL Indices = [] +temp-availableIndices = [] # To update available indices for the next round +# Build a set for quick removal of chosen indices +chosen_poolIndices_set = set(chosen_poolIndices) +for poolidx in availableIndices_for_sampling: + if poolidx in chosen_poolIndices_set: + original_dataidx, item = indexed_full_dataset[poolidx] + current_batch_items.append(item) + current_batch ORIGINAL Indices.append( + original_dataidx) + else: + temp-availableIndices.append(poolidx) + availableIndices_for_sampling = temp-availableIndices + sampledIndices_overall.append(current_batch ORIGINAL Indices) +for item in current_batch_items: + s_value = sum(item[:k_value_num_models]) + all_collected_S_values.append(s_value) +if not all_collected_S_values: continue +s_mean = np.mean(all_collected_S_values) +nonconformity Scores = [_nonconformity_score.abs_diff_mean(s,s_mean) for s in all_collected_S_values] +nonconformity Scores Sorted = sorted(nonconformity Scores) +r_calib_size = len(nonconformity Scores Sorted) +quantileidx = int(math.ceil((r_calib_size + 1) * (1 - epsilon_conformal))) - 1 +quantileidx = min(max(quantileidx, 0), r_calib_size - 1) # Ensure index is valid +current_q_value = nonconformity Scores Sorted[quantileidx] +final_q_value = current_q_value +if batchidx >= min_batches_before_stopping_check -1 : # batchidx is 0-indexed +if qprevious is not None: + if abs(current_q_value - qprevious) < convergence_threshold_q_diff: + return all_collected_S_values, final_q_value, batchidx + 1, sampled Indices_overall +qprevious = current_q_value +elif batchidx == 0: # Set qprevious for the first iteration +qprevious = current_q_value +return all_collected_S_values, final_q_value, max_batches, +sampled Indices_overall +``` + +```python +import numpy as np +#---Helper Function for Distribution Transfer --- +def normalize_vector(v): ""L2 normalizes a vector."" norm $=$ np.linalg.norm(v) return v / norm if norm>0 else v +#---Core Function 3:Distribution Transfer for Beta Mixture Parameters--- +def transfer_beta_mixture_parameters( target_direct_parameters, source_parameters_list, target_mean_embedding, source_mean_embedding_list, target_data_size, source_data_sizes_list, embedding_sincerity_threshold=0.9, similarity_scaling_factor=10.0, min_source_weight_factor=0.0 +) : "" Transfers/adjusts Beta mixture parameters from source domains to a target domain based on embedding similarity and data size. Args: target_direct_parameters (tuple): (a1_t,b1_t,a2_t,b2_t,w1_t)- Directly estimated parameters for the target domain. source_parameters_list(list of tuples):Each tuple contains parameters for a source domain. target_mean_embedding(np.array):Mean embedding vector for the target domain. source_mean_embedding_list(list of np.array):List of mean embedding vectors for source domains. target_data_size(int):Number of samples in the target domain . source_data Sizes list(list of int):List of data sizes for source domains. embeddingSimilarity_threshold(float):Threshold for cosine similarity. similarity_scaling_factor(float):Scaling factor for the similarity score. min_source_weight_factor(float): Minimum source weight factor ,ensuring non-negativity. +Returns: tuple:Transferred parameters(a1_f,b1_f,a2_f,b2_f,w1_f). "" if not source_parameters_list:# No source,return target's own parameters return target_direct_parameters if not (len.source_parameters_list) $= =$ len.source_mean_embedding_list $) = =$ len.source_data_sizes_list)): raise ValueError("Lengths of source parameters,embeddings, and size lists must match.") norm_target_emb $=$ _normalize_vector(np.asarray( target_mean_embedding,dtype $\equiv$ float)) +weight_target $=$ float(target_data_size) source_finalweights $= [ ]$ +for i in range(len.source_parameters_list): +``` + +Listing 3: Distribution Transfer for Beta Mixture Parameters +```python +norm_source_emb_i = _normalize_vector(np.asarray( source_mean_embeddingings_list[i], dtype=float)) similarity $=$ np.dot(norm_target_emb, norm_source_emb_i) # Calculate similarity-based weight factor, ensuring non- negativity similarity_based_factor $=$ similarity_scaling_factor \* ( similarity - embedding_similarity_threshold) similarity_based_factor $=$ max(min_source_weight_factor, similarity_based_factor) current_source_weight $=$ source_data_sizes_list[i] \* similarity_based_factor source_finalweights.append(current_source_weight) total Combined weight $=$ weight_target $^+$ sum.source_finalweights) if total Combined weight $< =$ 1e-9: # If total weight is too small, return target's own parameters return target_direct_parameters num.params_to_transfer $=$ len(target_direct_parameters) final_transferred.params_list $= [0.0]$ \* num.params_to_transfer # Contribution from target parameters for i in range(num.params_to_transfer): final_transferred.params_list[i] $+ =$ weight_target \* target_direct_parameters[i] # Contribution from source parameters for i, src.params_tuple in enumeratesource.params_list): if len(src.params_tuple)! $=$ num.params_to_transfer: raise ValueError(f"Source parameter tuple{i} length mismatch with target parameters.") for j in range(num.params_to_transfer): final_transferred.params_list[j] $+ =$ source_finalweights[i ] \* src.params_tuple[j] final.params_values $= [\mathfrak{p}$ /total Combined_weight for p in final_transferred.params_list] # Post-process parameters: ensure alpha, beta are positive, and w1 is in [0,1] # Assuming the order is (a1,b1,a2,b2,w1) a1_f,b1_f,a2_f,b2_f,w1_f $=$ final.params_values a1_f $=$ max(a1_f,1e-6) b1_f $=$ max(b1_f,1e-6) a2_f $=$ max(a2_f,1e-6) b2_f $=$ max(b2_f,1e-6) w1_f $=$ npclip(w1_f,1e-6,1.0 - 1e-6) return (a1_f,b1_f,a2_f,b2_f,w1_f) +``` + +```python +import math +import random +import numpy as np +from scipy.stats import beta +from scipy especial import betaln, gamm as lgamma # gamm is scipy 's log gamma +from math import comb # math.comb for combinations +#---Helper Functions for Beta Mixture and Beta-Binomial --- +``` + +```python +def _replace_elements_for_beta(pdf(probabilities): + Replaces 0s and 1s in a list of probabilities with close values to avoid issues with beta.pdf calculations. + return [0.999999 if x >= 1.0 else 0.000001 if x <= 0.0 else x for x in probabilities] +def _beta_binomial_pmf_log(k_trials, num_successes, alpha, beta_param): + Calculates the log of the Beta-Binomial PMF: log(P(X=num_successes)) + where X ~ BB(k_trials, alpha, beta-param). + P(X=x) = C(k,x) * Beta(alpha+x, beta+k-x) / Beta(alpha,beta) + if not (0 <= num_successes <= k_trials): + return -np.inf # Log probability of zero + # Ensure alpha and beta-param are positive + alpha_stable = max(alpha, 1e-9) + beta_stable = max(betaparam, 1e-9) + log_C_k_x = lgamma(k_trials + 1) - (lgamma(num_successes + 1) + lgamma(k_trials - num_successes + 1)) + log_beta_num = betaln(alpha_stable + num_successes, beta_stable + k_trials - num_successes) + log_beta_den = betaln(alpha_stable, beta_stable) + return log_C_k_x + log_beta_num - log_beta_den +def _mixture_beta_binomial_pmf(num_successes, alpha1, beta1, alpha2, beta2, w1, k_trials): + PMF of the mixture Beta-Binomial model: + P_mix(x=x) = w1 * BB(k, alpha1, beta1) + (1-w1) * BB(k, alpha2, beta2) + log_p1 = _beta_binomial_pmf_log(k_trials, num_successes, alpha1, beta1) + log_p2 = _beta_binomial_pmf_log(k_trials, num_successes, alpha2, beta2) + p1 = np.exp(log_p1) + p2 = np.exp(log_p2) + return w1 * p1 + (1 - w1) * p2 +# --- Core Function 1: Mixture of Beta Distributions Fitting via EM +def fit_mixture_of_betas_em( raw_samples_outcomes, + num,trials_per_sample, + max_iters=100, + tol=1e-6, + alpha1_init=None, beta1_init=None, + alpha2_init=None, beta2_init=None, + w1_init=None +): " " +Fits a mixture of two Beta distributions using the EM algorithm. This model is used for modeling observed success rates p_i = ( + successes for sample i) / num_trials_per_sample. +``` + +Args: + +raw_samples_outcomes (list of lists): Each inner list contains binary outcomes (0 or 1) for a data point. + +num_trials_per_sample (int): Number of trials/outcomes to consider from the start of each inner list (K or m). + +max_iters (int): Maximum number of iterations for the EM algorithm. + +tol (float): Tolerance for convergence. + +alpha1_init, beta1_init, alpha2_init, beta2_init, w1_init: +Optional initial parameters. + +Returns: + +tuple: (alpha1, beta1, alpha2, beta2, w1) - The estimated parameters. + +111 + +num_data_points = len(raw_samples_outcomes) + +if num_data_points == 0: + +raise ValueError("Input raw_samples_outcomes cannot be empty.") +) + +if num_trials_per_sample $<= 0$ + +raise ValueError("num_trials_per_sample must be positive.") + +Initialize parameters (heuristic based on original code) + +alpha1 = alpha1_init if alpha1_init is not None else 10 * num_trials_per_sample + +beta1 = beta1_init if beta1_init is not None else 1 * num_trials_per_sample + +alpha2 = alpha2_init if alpha2_init is not None else 1 * num_trials_per_sample + +beta2 = beta2_init if beta2_init is not None else 10 * num_trials_per_sample + +w1 = w1_init if w1_init is not None else 0.5 + +alpha1, beta1 = max(alpha1, 1e-6), max(beta1, 1e-6) + +alpha2, beta2 = max(alpha2, 1e-6), max(beta2, 1e-6) + +w1 = np.clip(w1, 1e-6, 1.0 - 1e-6) + +observed_successes = np.array([sum(sample[:num_trials_per_sample]) + +for sample in raw_samples_outcomes]) + +proportions $=$ observed_successes / num_trials_per_sample + +proportions_for(pdf $=$ np.array(_replace_elements_for_beta(pdf(proportions.tolist())))) + +for iteration in range(maxitaire): + +E-Step: Calculate responsibilities + +pdf_vals1 = beta.pdf(proportions_for(pdf, alpha1 + 1e-9, beta1 + 1e-9) # Add small epsilon for stability + +pdf_vals2 = beta.pdf(proportions_for(pdf, alpha2 + 1e-9, beta2 + 1e-9) + +numerator1 = w1 * pdf_vals1 + +numerator2 = (1 - w1) * pdf_vals2 + +denominator = numerator1 + numerator2 + +denominator [denominator < 1e-9] = 1e-9 # Avoid division by zero + +resp1 = numerator1 / denominator +resp2 = numerator2 / denominator + +M-Step: Update parameters (using weighted method of moments for Beta parameters) + +w1_new = np.mean(resp1) +w1_new = np.clip(w1_new, 1e-6, 1.0 - 1e-6) + +Update alpha, beta for component 1 + +```python +sum_resp1 = np.sum(resp1) +if sum_resp1 < 1e-6: + alpha1_new, beta1_new = alpha1, beta1 # Keep old if weight is too small +else: + mean_p1_w = np.sum(resp1 * proportions) / sum_resp1 + var_p1_w = np.sum(resp1 * ((proportions - mean_p1_w)**2)) / sum_resp1 + mean_p1_w = np.clip(mean_p1_w, 1e-6, 1.0 - 1e-6) + if var_p1_w <= 1e-9 or var_p1_w >= mean_p1_w * (1.0 - mean_p1_w) * (1-1e-6): # Check if variance is valid + # Invalid or too small variance, use heuristic (e.g., high confidence) + alpha1_new = mean_p1_w * (num_trials_per_sample * 10) + # Larger concentration + beta1_new = (1.0 - mean_p1_w) * (num_trials_per_sample * 10) + else: + common_factor = (mean_p1_w * (1.0 - mean_p1_w) / var_p1_w) - 1.0 + alpha1_new = mean_p1_w * common_factor + beta1_new = (1.0 - mean_p1_w) * common_factor +# Update alpha, beta for component 2 +sum_resp2 = np.sum(resp2) +if sum_resp2 < 1e-6: + alpha2_new, beta2_new = alpha2, beta2 +else: + mean_p2_w = np,sum(resp2 * proportions) / sum_resp2 + var_p2_w = np,sum(resp2 * ((proportions - mean_p2_w)**2)) / sum_resp2 + mean_p2_w = np.clip(mean_p2_w, 1e-6, 1.0 - 1e-6) + if var_p2_w <= 1e-9 or var_p2_w >= mean_p2_w * (1.0 - mean_p2_w) * (1-1e-6): + alpha2_new = mean_p2_w * (num_trials_per_sample * 10) + beta2_new = (1.0 - mean_p2_w) * (num_trials_per_sample * 10) + else: + common_factor2 = (mean_p2_w * (1.0 - mean_p2_w) / var_p2_w) - 1.0 + alpha2_new = mean_p2_w * common_factor2 + beta2_new = (1.0 - mean_p2_w) * common_factor2 +alpha1_new, beta1_new = max(alpha1_new, 1e-6), max(beta1_new, 1e-6) +alpha2_new, beta2_new = max(alpha2_new, 1e-6), max(beta2_new, 1e-6) +# Check for convergence +param_diff = (abs(alpha1 - alpha1_new) + abs(beta1 - beta1_new) + abs(alpha2 - alpha2_new) + abs(beta2 - beta2_new) + abs(w1 - w1_new)) +if param_diff < tol: + alpha1, beta1, alpha2, beta2, w1 = alpha1_new, beta1_new, alpha2_new, beta2_new, w1_new +break +alpha1, beta1, alpha2, beta2, w1 = alpha1_new, beta1_new, alpha2_new, beta2_new, w1_new +return alpha1, beta1, alpha2, beta2, w1 +``` + +```python +def calculate Majority Vote success prob_from_mixture( k_trials_for_vote, alpha1, beta1, alpha2, beta2, w1_mixture_weight +): "Calculates the probability of achieving majority success given Beta-Binomial mixture parameters. +Majority success is defined as number of successes >= ceil( k_trials_for_vote / 2). +Args: + k_trials_for_vote (int): Total number of trials (e.g., number of LLMs). + alpha1, beta1: Parameters for the first Beta-Binomial component. + alpha2, beta2: Parameters for the second Beta-Binomial component. + w1_mixture_weight (float): Mixture weight for the first component. +Returns: + float: Probability of majority vote success. +"if k_trials_for_vote <= 0: return 0.0 +majority_threshold = math.ceil(k_trials_for_vote / 2.0) +prob_sum_for Majority = 0.0 +for num_successes in range(int(majority_threshold), k_trials_for_vote + 1): + prob_sum_for Majority += _mixture_beta_binomial_pmf( num_successes, alpha1, beta1, alpha2, beta2, w1_mixture_weight, k_trials_for_vote ) +return prob_sum_for Majority +``` + +# F Limitations and Future Work + +The two-component Beta-Binomial mixture improves over simpler models but may still underfit complex judgment distributions. Prior transfer depends on text embedding quality and assumes textual similarity implies similar judgments—an assumption that may not always hold. The current design also focuses on binary/scoring tasks and requires an odd number of annotators. + +Future work could explore more flexible mixture models, robust prior transfer methods beyond textual similarity, task-specific features, and extensions to diverse judgment formats and ensemble sizes. + +# G Broader Impacts + +BetaConform can reduce the cost of LLM ensemble evaluations, supporting broader use in QA, benchmarking, annotation, and MLOps. It enables scalable, reliable assessment but requires careful attention to estimation error and modeling assumptions, especially in high-stakes applications. + +# NeurIPS Paper Checklist + +The checklist is designed to encourage best practices for responsible machine learning research, addressing issues of reproducibility, transparency, research ethics, and societal impact. Do not remove the checklist: The papers not including the checklist will be desk rejected. The checklist should follow the references and follow the (optional) supplemental material. The checklist does NOT count towards the page limit. + +Please read the checklist guidelines carefully for information on how to answer these questions. For each question in the checklist: + +- You should answer [Yes], [No], or [NA]. +- [NA] means either that the question is Not Applicable for that particular paper or the relevant information is Not Available. +- Please provide a short (1–2 sentence) justification right after your answer (even for NA). + +The checklist answers are an integral part of your paper submission. They are visible to the reviewers, area chairs, senior area chairs, and ethics reviewers. You will be asked to also include it (after eventual revisions) with the final version of your paper, and its final version will be published with the paper. + +The reviewers of your paper will be asked to use the checklist as one of the factors in their evaluation. While "[Yes]" is generally preferable to "[No]", it is perfectly acceptable to answer "[No]" provided a proper justification is given (e.g., "error bars are not reported because it would be too computationally expensive" or "we were unable to find the license for the dataset we used"). In general, answering "[No]" or "[NA]" is not grounds for rejection. While the questions are phrased in a binary way, we acknowledge that the true answer is often more nuanced, so please just use your best judgment and write a justification to elaborate. All supporting evidence can appear either in the main paper or the supplemental material, provided in appendix. If you answer [Yes] to a question, in the justification please point to the section(s) where related material for the question can be found. + +IMPORTANT, please: + +- Delete this instruction block, but keep the section heading "NeurIPS Paper Checklist", +- Keep the checklist subsection headings, questions/answers and guidelines below. +- Do not modify the questions and only use the provided macros for your answers. + +# 1. Claims + +Question: Do the main claims made in the abstract and introduction accurately reflect the paper's contributions and scope? + +Answer: [Yes] + +Justification: Our method sections and the experiment section matches the description of our method in the abstract and introduction. + +Guidelines: + +- The answer NA means that the abstract and introduction do not include the claims made in the paper. +- The abstract and/or introduction should clearly state the claims made, including the contributions made in the paper and important assumptions and limitations. A No or NA answer to this question will not be perceived well by the reviewers. +- The claims made should match theoretical and experimental results, and reflect how much the results can be expected to generalize to other settings. +- It is fine to include aspirational goals as motivation as long as it is clear that these goals are not attained by the paper. + +# 2. Limitations + +Question: Does the paper discuss the limitations of the work performed by the authors? + +Answer: [Yes] + +Justification: In the Section F, we discuss about the limitations and future work of BetaConform. + +# Guidelines: + +- The answer NA means that the paper has no limitation while the answer No means that the paper has limitations, but those are not discussed in the paper. +- The authors are encouraged to create a separate "Limitations" section in their paper. +- The paper should point out any strong assumptions and how robust the results are to violations of these assumptions (e.g., independence assumptions, noiseless settings, model well-specification, asymptotic approximations only holding locally). The authors should reflect on how these assumptions might be violated in practice and what the implications would be. +- The authors should reflect on the scope of the claims made, e.g., if the approach was only tested on a few datasets or with a few runs. In general, empirical results often depend on implicit assumptions, which should be articulated. +- The authors should reflect on the factors that influence the performance of the approach. For example, a facial recognition algorithm may perform poorly when image resolution is low or images are taken in low lighting. Or a speech-to-text system might not be used reliably to provide closed captions for online lectures because it fails to handle technical jargon. +- The authors should discuss the computational efficiency of the proposed algorithms and how they scale with dataset size. +- If applicable, the authors should discuss possible limitations of their approach to address problems of privacy and fairness. +- While the authors might fear that complete honesty about limitations might be used by reviewers as grounds for rejection, a worse outcome might be that reviewers discover limitations that aren't acknowledged in the paper. The authors should use their best judgment and recognize that individual actions in favor of transparency play an important role in developing norms that preserve the integrity of the community. Reviewers will be specifically instructed to not penalize honesty concerning limitations. + +# 3. Theory assumptions and proofs + +Question: For each theoretical result, does the paper provide the full set of assumptions and a complete (and correct) proof? + +Answer: [Yes] + +Justification: For Proposition 1 and Proposition 2, we provide the assumption and proof in Section B.1 and Section B.2. + +Guidelines: + +- The answer NA means that the paper does not include theoretical results. +- All the theorems, formulas, and proofs in the paper should be numbered and cross-referenced. +- All assumptions should be clearly stated or referenced in the statement of any theorems. +- The proofs can either appear in the main paper or the supplemental material, but if they appear in the supplemental material, the authors are encouraged to provide a short proof sketch to provide intuition. +- Inversely, any informal proof provided in the core of the paper should be complemented by formal proofs provided in appendix or supplemental material. +- Theorems and Lemmas that the proof relies upon should be properly referenced. + +# 4. Experimental result reproducibility + +Question: Does the paper fully disclose all the information needed to reproduce the main experimental results of the paper to the extent that it affects the main claims and/or conclusions of the paper (regardless of whether the code and data are provided or not)? + +Answer: [Yes] + +Justification: In Section E, we provide the implementation detail of our method and experiments. In Section A, we provide the detailed description of our method. + +# Guidelines: + +- The answer NA means that the paper does not include experiments. +- If the paper includes experiments, a No answer to this question will not be perceived well by the reviewers: Making the paper reproducible is important, regardless of whether the code and data are provided or not. +- If the contribution is a dataset and/or model, the authors should describe the steps taken to make their results reproducible or verifiable. +- Depending on the contribution, reproducibility can be accomplished in various ways. For example, if the contribution is a novel architecture, describing the architecture fully might suffice, or if the contribution is a specific model and empirical evaluation, it may be necessary to either make it possible for others to replicate the model with the same dataset, or provide access to the model. In general, releasing code and data is often one good way to accomplish this, but reproducibility can also be provided via detailed instructions for how to replicate the results, access to a hosted model (e.g., in the case of a large language model), releasing of a model checkpoint, or other means that are appropriate to the research performed. +- While NeurIPS does not require releasing code, the conference does require all submissions to provide some reasonable avenue for reproducibility, which may depend on the nature of the contribution. For example +(a) If the contribution is primarily a new algorithm, the paper should make it clear how to reproduce that algorithm. +(b) If the contribution is primarily a new model architecture, the paper should describe the architecture clearly and fully. +(c) If the contribution is a new model (e.g., a large language model), then there should either be a way to access this model for reproducing the results or a way to reproduce the model (e.g., with an open-source dataset or instructions for how to construct the dataset). +(d) We recognize that reproducibility may be tricky in some cases, in which case authors are welcome to describe the particular way they provide for reproducibility. In the case of closed-source models, it may be that access to the model is limited in some way (e.g., to registered users), but it should be possible for other researchers to have some path to reproducing or verifying the results. + +# 5. Open access to data and code + +Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material? + +Answer: [No] + +Justification: We do not release the code. All the datasets used in this paper are open-source datasets and can be found online with their names. + +# Guidelines: + +- The answer NA means that paper does not include experiments requiring code. +- Please see the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details. +- While we encourage the release of code and data, we understand that this might not be possible, so "No" is an acceptable answer. Papers cannot be rejected simply for not including code, unless this is central to the contribution (e.g., for a new open-source benchmark). +- The instructions should contain the exact command and environment needed to run to reproduce the results. See the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details. +- The authors should provide instructions on data access and preparation, including how to access the raw data, preprocessed data, intermediate data, and generated data, etc. +- The authors should provide scripts to reproduce all experimental results for the new proposed method and baselines. If only a subset of experiments are reproducible, they should state which ones are omitted from the script and why. + +- At submission time, to preserve anonymity, the authors should release anonymized versions (if applicable). +- Providing as much information as possible in supplemental material (appended to the paper) is recommended, but including URLs to data and code is permitted. + +# 6. Experimental setting/details + +Question: Does the paper specify all the training and test details (e.g., data splits, hyperparameters, how they were chosen, type of optimizer, etc.) necessary to understand the results? + +Answer: [Yes] + +Justification: Our work do not involve training, and we directly use the designed validation/test splits of each dataset. The hyperparameters of experiments are described in Section E. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The experimental setting should be presented in the core of the paper to a level of detail that is necessary to appreciate the results and make sense of them. +- The full details can be provided either with the code, in appendix, or as supplemental material. + +# 7. Experiment statistical significance + +Question: Does the paper report error bars suitably and correctly defined or other appropriate information about the statistical significance of the experiments? + +Answer: [Yes] + +Justification: We provide the mean and standard deviation of our experimental results. The setting is described in Section E. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The authors should answer "Yes" if the results are accompanied by error bars, confidence intervals, or statistical significance tests, at least for the experiments that support the main claims of the paper. +- The factors of variability that the error bars are capturing should be clearly stated (for example, train/test split, initialization, random drawing of some parameter, or overall run with given experimental conditions). +- The method for calculating the error bars should be explained (closed form formula, call to a library function, bootstrap, etc.) +- The assumptions made should be given (e.g., Normally distributed errors). +- It should be clear whether the error bar is the standard deviation or the standard error of the mean. +- It is OK to report 1-sigma error bars, but one should state it. The authors should preferably report a 2-sigma error bar than state that they have a $96\%$ CI, if the hypothesis of Normality of errors is not verified. +- For asymmetric distributions, the authors should be careful not to show in tables or figures symmetric error bars that would yield results that are out of range (e.g. negative error rates). +- If error bars are reported in tables or plots, The authors should explain in the text how they were calculated and reference the corresponding figures or tables in the text. + +# 8. Experiments compute resources + +Question: For each experiment, does the paper provide sufficient information on the computer resources (type of compute workers, memory, time of execution) needed to reproduce the experiments? + +Answer: [NA] + +Justification: The judging process only require inference of LLMs. The distribution estimation solely uses CPU. + +# Guidelines: + +- The answer NA means that the paper does not include experiments. +- The paper should indicate the type of compute workers CPU or GPU, internal cluster, or cloud provider, including relevant memory and storage. +- The paper should provide the amount of compute required for each of the individual experimental runs as well as estimate the total compute. +- The paper should disclose whether the full research project required more compute than the experiments reported in the paper (e.g., preliminary or failed experiments that didn't make it into the paper). + +# 9. Code of ethics + +Question: Does the research conducted in the paper conform, in every respect, with the NeurIPS Code of Ethics https://neurips.cc/public/EthicsGuidelines? + +Answer: [Yes] + +Justification: The anonymity is preserved and we follow the NeurIPS Code of Ethics. + +# Guidelines: + +- The answer NA means that the authors have not reviewed the NeurIPS Code of Ethics. +- If the authors answer No, they should explain the special circumstances that require a deviation from the Code of Ethics. +- The authors should make sure to preserve anonymity (e.g., if there is a special consideration due to laws or regulations in their jurisdiction). + +# 10. Broader impacts + +Question: Does the paper discuss both potential positive societal impacts and negative societal impacts of the work performed? + +Answer: [Yes] + +Justification: In Section G we discuss the broader impact of our method. + +# Guidelines: + +- The answer NA means that there is no societal impact of the work performed. +- If the authors answer NA or No, they should explain why their work has no societal impact or why the paper does not address societal impact. +- Examples of negative societal impacts include potential malicious or unintended uses (e.g., disinformation, generating fake profiles, surveillance), fairness considerations (e.g., deployment of technologies that could make decisions that unfairly impact specific groups), privacy considerations, and security considerations. +- The conference expects that many papers will be foundational research and not tied to particular applications, let alone deployments. However, if there is a direct path to any negative applications, the authors should point it out. For example, it is legitimate to point out that an improvement in the quality of generative models could be used to generate deepfakes for disinformation. On the other hand, it is not needed to point out that a generic algorithm for optimizing neural networks could enable people to train models that generate Deepfakes faster. +- The authors should consider possible harms that could arise when the technology is being used as intended and functioning correctly, harms that could arise when the technology is being used as intended but gives incorrect results, and harms following from (intentional or unintentional) misuse of the technology. +- If there are negative societal impacts, the authors could also discuss possible mitigation strategies (e.g., gated release of models, providing defenses in addition to attacks, mechanisms for monitoring misuse, mechanisms to monitor how a system learns from feedback over time, improving the efficiency and accessibility of ML). + +# 11. Safeguards + +Question: Does the paper describe safeguards that have been put in place for responsible release of data or models that have a high risk for misuse (e.g., pretrained language models, image generators, or scraped datasets)? + +# Answer: [NA] + +Justification: Our paper does not pose such risks. + +# Guidelines: + +- The answer NA means that the paper poses no such risks. +- Released models that have a high risk for misuse or dual-use should be released with necessary safeguards to allow for controlled use of the model, for example by requiring that users adhere to usage guidelines or restrictions to access the model or implementing safety filters. +- Datasets that have been scraped from the Internet could pose safety risks. The authors should describe how they avoided releasing unsafe images. +- We recognize that providing effective safeguards is challenging, and many papers do not require this, but we encourage authors to take this into account and make a best faith effort. + +# 12. Licenses for existing assets + +Question: Are the creators or original owners of assets (e.g., code, data, models), used in the paper, properly credited and are the license and terms of use explicitly mentioned and properly respected? + +# Answer: [Yes] + +Justification: We make proper citation of each dataset used in our paper, and we follow the license of each dataset. + +# Guidelines: + +- The answer NA means that the paper does not use existing assets. +- The authors should cite the original paper that produced the code package or dataset. +- The authors should state which version of the asset is used and, if possible, include a URL. +- The name of the license (e.g., CC-BY 4.0) should be included for each asset. +- For scraped data from a particular source (e.g., website), the copyright and terms of service of that source should be provided. +- If assets are released, the license, copyright information, and terms of use in the package should be provided. For popular datasets, paperswithcode.com/datasets has curated licenses for some datasets. Their licensing guide can help determine the license of a dataset. +- For existing datasets that are re-packaged, both the original license and the license of the derived asset (if it has changed) should be provided. +- If this information is not available online, the authors are encouraged to reach out to the asset's creators. + +# 13. New assets + +Question: Are new assets introduced in the paper well documented and is the documentation provided alongside the assets? + +# Answer: [NA] + +Justification: The paper does not release new assets. + +# Guidelines: + +- The answer NA means that the paper does not release new assets. +- Researchers should communicate the details of the dataset/code/model as part of their submissions via structured templates. This includes details about training, license, limitations, etc. +- The paper should discuss whether and how consent was obtained from people whose asset is used. +- At submission time, remember to anonymize your assets (if applicable). You can either create an anonymized URL or include an anonymized zip file. + +# 14. Crowdsourcing and research with human subjects + +Question: For crowdsourcing experiments and research with human subjects, does the paper include the full text of instructions given to participants and screenshots, if applicable, as well as details about compensation (if any)? + +Answer: [NA] + +Justification: The paper does not involve crowdsourcing nor research with human subjects. + +Guidelines: + +- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. +- Including this information in the supplemental material is fine, but if the main contribution of the paper involves human subjects, then as much detail as possible should be included in the main paper. +- According to the NeurIPS Code of Ethics, workers involved in data collection, curation, or other labor should be paid at least the minimum wage in the country of the data collector. + +# 15. Institutional review board (IRB) approvals or equivalent for research with human subjects + +Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals (or an equivalent approval/review based on the requirements of your country or institution) were obtained? + +Answer: [NA] + +Justification: The paper does not involve crowdsourcing nor research with human subjects. + +Guidelines: + +- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. +- Depending on the country in which research is conducted, IRB approval (or equivalent) may be required for any human subjects research. If you obtained IRB approval, you should clearly state this in the paper. +- We recognize that the procedures for this may vary significantly between institutions and locations, and we expect authors to adhere to the NeurIPS Code of Ethics and the guidelines for their institution. +- For initial submissions, do not include any information that would break anonymity (if applicable), such as the institution conducting the review. + +# 16. Declaration of LLM usage + +Question: Does the paper describe the usage of LLMs if it is an important, original, or non-standard component of the core methods in this research? Note that if the LLM is used only for writing, editing, or formatting purposes and does not impact the core methodology, scientific rigorousness, or originality of the research, declaration is not required. + +Answer: [NA] + +Justification: The core method development in this research does not involve LLMs as any important, original, or non-standard components. + +- The answer NA means that the core method development in this research does not involve LLMs as any important, original, or non-standard components. +- Please refer to our LLM policy (https://neurips.cc/Conferences/2025/LLM) for what should or should not be described. \ No newline at end of file diff --git a/NeurIPS/2025/$_texttt{BetaConform}$_ Efficient MAP Estimation of LLM Ensemble Judgment Performance with Prior Transfer/images.zip b/NeurIPS/2025/$_texttt{BetaConform}$_ Efficient MAP Estimation of LLM Ensemble Judgment Performance with Prior Transfer/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..70fd741662bdfee62b037847bcee0e93098dda78 --- /dev/null +++ b/NeurIPS/2025/$_texttt{BetaConform}$_ Efficient MAP Estimation of LLM Ensemble Judgment Performance with Prior Transfer/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:64eb8926da976a2613cb630fec95006c1a648be662dac7547f1df6ea3eca4101 +size 900998 diff --git a/NeurIPS/2025/$_texttt{BetaConform}$_ Efficient MAP Estimation of LLM Ensemble Judgment Performance with Prior Transfer/layout.json b/NeurIPS/2025/$_texttt{BetaConform}$_ Efficient MAP Estimation of LLM Ensemble Judgment Performance with Prior Transfer/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..dc6aece8ad51294dfa8a0d1f7479063d2164f5f4 --- /dev/null +++ b/NeurIPS/2025/$_texttt{BetaConform}$_ Efficient MAP Estimation of LLM Ensemble Judgment Performance with Prior Transfer/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a63304d27c9d488922414957997c9c64d4bbf92983b4d76c412a6d0851bb3272 +size 1016328 diff --git a/NeurIPS/2025/$_texttt{G1}$_ Teaching LLMs to Reason on Graphs with Reinforcement Learning/34acc6fe-738e-4fc8-bf3b-dc455aa6102a_content_list.json b/NeurIPS/2025/$_texttt{G1}$_ Teaching LLMs to Reason on Graphs with Reinforcement Learning/34acc6fe-738e-4fc8-bf3b-dc455aa6102a_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..bf8f13266ad41def6af84f3822dbde0d94aeccf1 --- /dev/null +++ b/NeurIPS/2025/$_texttt{G1}$_ Teaching LLMs to Reason on Graphs with Reinforcement Learning/34acc6fe-738e-4fc8-bf3b-dc455aa6102a_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:71b5520c02ecbba94d1c847201132c08c227a3c2ec1bd4a000fb6361e916ad30 +size 211922 diff --git a/NeurIPS/2025/$_texttt{G1}$_ Teaching LLMs to Reason on Graphs with Reinforcement Learning/34acc6fe-738e-4fc8-bf3b-dc455aa6102a_model.json b/NeurIPS/2025/$_texttt{G1}$_ Teaching LLMs to Reason on Graphs with Reinforcement Learning/34acc6fe-738e-4fc8-bf3b-dc455aa6102a_model.json new file mode 100644 index 0000000000000000000000000000000000000000..33470861de6086400ac5a5d61f504cab59c12100 --- /dev/null +++ b/NeurIPS/2025/$_texttt{G1}$_ Teaching LLMs to Reason on Graphs with Reinforcement Learning/34acc6fe-738e-4fc8-bf3b-dc455aa6102a_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dd937a4a27775c112b6155a41b61cadfb3f48e67c413f8fbf3ca30e3d17b040c +size 246110 diff --git a/NeurIPS/2025/$_texttt{G1}$_ Teaching LLMs to Reason on Graphs with Reinforcement Learning/34acc6fe-738e-4fc8-bf3b-dc455aa6102a_origin.pdf b/NeurIPS/2025/$_texttt{G1}$_ Teaching LLMs to Reason on Graphs with Reinforcement Learning/34acc6fe-738e-4fc8-bf3b-dc455aa6102a_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..9bf19e02b563a70f9ac21f5884e8c7b867a9cc13 --- /dev/null +++ b/NeurIPS/2025/$_texttt{G1}$_ Teaching LLMs to Reason on Graphs with Reinforcement Learning/34acc6fe-738e-4fc8-bf3b-dc455aa6102a_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d2ae310d38eb8ec698715faa7ab592ca8d629fd164ec85072776942e3410e6b7 +size 574707 diff --git a/NeurIPS/2025/$_texttt{G1}$_ Teaching LLMs to Reason on Graphs with Reinforcement Learning/full.md b/NeurIPS/2025/$_texttt{G1}$_ Teaching LLMs to Reason on Graphs with Reinforcement Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..2f6a20dc48f887a9f6be56afbdfde4b73c900975 --- /dev/null +++ b/NeurIPS/2025/$_texttt{G1}$_ Teaching LLMs to Reason on Graphs with Reinforcement Learning/full.md @@ -0,0 +1,754 @@ +# G1: Teaching LLMs to Reason on Graphs with Reinforcement Learning + +Xiaojun Guo $^{1*}$ Ang Li $^{1*}$ Yifei Wang $^{3*}$ + +Stefanie Jegelka4,3 Yisen Wang1,2 + +$^{1}$ State Key Lab of General Artificial Intelligence, + +School of Intelligence Science and Technology, Peking University + +$^{2}$ Institute for Artificial Intelligence, Peking University + +$^{3}$ MIT + +4TUM + +# Abstract + +Although Large Language Models (LLMs) have demonstrated remarkable progress, their proficiency in graph-related tasks remains notably limited, hindering the development of truly general-purpose models. Previous attempts, including pretraining graph foundation models or employing supervised fine-tuning, often face challenges such as the scarcity of large-scale, universally represented graph data. We introduce G1, a simple yet effective approach demonstrating that Reinforcement Learning (RL) on synthetic graph-theoretic tasks can significantly scale LLMs' graph reasoning abilities. To enable RL training, we curate Erdős, the largest graph reasoning dataset to date comprising 50 diverse graph-theoretic tasks of varying difficulty levels, 100k training data and 5k test data, all driven from real-world graphs. With RL on Erdős, G1 obtains substantial improvements in graph reasoning, where our finetuned 3B model even outperforms Qwen2.5-72B-Instruct (24x size). RL-trained models also show strong zero-shot generalization to unseen tasks, domains, and graph encoding schemes, including other graph-theoretic benchmarks as well as real-world node classification and link prediction tasks, without compromising general reasoning abilities. Our findings offer an efficient, scalable path for building strong graph reasoners by finetuning LLMs with RL on graph-theoretic tasks, which combines the strengths of pretrained LLM capabilities with abundant, automatically generated synthetic data, suggesting that LLMs possess graph understanding abilities that RL can elicit successfully. Our implementation is open-sourced at https://github.com/PKU-ML/G1, with models and datasets hosted on Hugging Face collections PKU-ML/G1 for broader accessibility. + +# 1 Introduction + +Large Language Models (LLMs) have achieved widespread success [2, 13] but exhibit notable limitations in reasoning about graph-structured data, a critical capability for achieving general-purpose intelligence. Proficient graph reasoning is essential for numerous applications, yet even state-of-the-art LLMs like OpenAI's o1 [30] demonstrate significant deficiencies, with reported accuracies as low as $58.49\%$ on graph connectivity tests [55]. + +Initial efforts to enhance LLMs' graph understanding explored various natural language encoding schemes [11, 5, 9], but these yielded only modest improvements. Alternative strategies have involved + +Table 1: An overview of 50 graph-theoretic tasks in our dataset Erdős (100k train, 5k test), alongside with the difficulty distribution, and the accuracy of the base model Qwen2.5-7B-Instruct and our RL-trained G1-7B model. A complete description of tasks are in Appendix K.2. + +
DifficultyTasksRatioBase Model AccG1 Acc
EasyNode Number, Dominating Set, Common Neighbor, Edge Num- +ber, Neighbor, BFS, Has Cycle, DFS, Minimum Spanning Tree, +Edge Existence, Is Regular, Degree, Is Tournament, Density29.16%57.16%95.07%
MediumAdamic Adar Index, Clustering Coefficient, Connected Compo- +nent Number, Bipartite Maximum Matching, Local Connectivity, +Jaccard Coefficient, Min Edge Covering, Is Eularian, Degree +Centrality, Is Bipartite, Resource Allocation Index22.91%42.55%88.91%
HardMax Weight Matching, Closeness Centrality, Traveling Sales- +man Problem, Strongly Connected Number, Shortest Path, Cen- +ter, Diameter, Barycenter, Radius, Topological Sort, Periphery, +Betweenness Centrality, Triangles, Average Neighbor Degree, +Harmonic Centrality, Bridges33.33%18.87%50.44%
ChallengingIsomorphic Mapping, Global Efficiency, Maximal Independent +Set, Maximum Flow, Wiener Index, Hamiltonian Path, Min Ver- +tex Cover14.58%3.29%23.57%
+ +instruction tuning [27, 53] or preference tuning [3, 42] on curated graph datasets. Others attempted to build specialized graph foundation models through pretraining [28, 21, 26]; however, these are often limited by the lack of large-scale, universal graph representations suitable for diverse graph types. Different from prior work, we believe LLMs pretrained on Internet-scale data already possess graph reasoning ability, and we can elicit it through their own trial and error without human data. + +In this work, we are the first to explore the use of Reinforcement Learning (RL) to solve graph reasoning tasks. We chose graph-theoretic problems as a testbed as they allow direct verification of generated answers to produce rule-based rewards for RL training, which is shown to be key for the success of DeepSeek R1 in math and coding problems [13]. We collect the largest-to-date graph-theoretic problem set, Erdős, with either groundtruth answers or automatic verification programs. As illustrated in Table 1, these tasks span a wide spectrum of difficulty levels, from basic graph properties like node counting to NP-hard problems such as finding the maximal independent set. Another advantage of adopting graph-theoretic tasks is its circumvention of scarce human-annotated data; the model learns through exploration and reinforcement on synthetic tasks where ground-truth outcomes provide direct reward signals, similar to the AlphaGo-Zero paradigm [37]. Besides data construction, we also study various aspects of the training process, such as the influence of data mixture, supervised initialization, and the use of chain-of-thought [45]. Our results confirm that RL with synthetic graph-theoretic task is a powerful and scalable approach to improving graph reasoning abilities of LLMs. + +Our work makes the following key contributions: + +- We are the first to apply reinforcement learning (RL) framework to improving LLMs on graph reasoning tasks. The resulting model G1 significantly enhances the graph reasoning abilities of LLMs across a diverse set of synthetic tasks, demonstrating that appropriately finetuned LLMs can become stronger graph reasoners. +- We introduce Erdős, the largest-scale and most comprehensive graph-theoretic dataset that comprises 50 distinct tasks of varying complexities, uniquely constructed from diverse real-world graphs, providing a reliable platform for training and evaluating graph reasoning. +- We empirically demonstrate that G1 achieves substantial performance improvements on our Erdős benchmark, with gains of up to $46\%$ over baseline models. Notably, our fintuned G1-7B model attains competitive performance with state-of-the-art reasoning models like OpenAI's o3-mini and G1-3B easily rivals Qwen2.5-72B-Instruct by noticeable margins. +- G1 models exhibit strong zero-shot generalization on unseen graph tasks and domains, improving base models' performance on other graph-theoretic benchmarks (GraphWiz and GraphArena) and real-world graphs (Cora and PubMed) without deteriorating general reasoning ability (GSM8K, MATH, and MMLU-pro), indicating a synergetic improvement of LLMs' graph reasoning abilities through RL. + +G1 charts a data-efficient and scalable course for developing LLMs with strong graph reasoning. By demonstrating that RL can unlock latent graph understanding within general-purpose LLMs using synthetic data, our work suggests a possible paradigm shift away from reliance on heterogeneous real-world graphs to build graph foundation models. This paves the way for more versatile AI systems capable of sophisticated reasoning across diverse data modalities. + +# 2 Related Work + +Graph Reasoning. Graph reasoning problems fall into two categories: domain-specific, which require understanding both graph structures and node/link attributes, e.g., node classification, link prediction, and knowledge-based QA [15, 56, 18]; and domain-agnostic, also called graph-theoretic problems, which focus solely on structural reasoning but find a lot of practical uses in various domains, e.g., shortest paths, Hamiltonian paths, graph isomorphism [50, 35]. For the latter problems that we study in this paper, people have studied the use of RL [29, 40] or unsupervised learning [19], often in conjunction with Graph Neural Networks (GNNs) [20, 49] that align with the solution structure [51]. Yet these models are often built to solve each problem alone. Recently, Sanford et al. [34] prove and validate the priority of the transformer models compared to GNNs on complex graph reasoning tasks requiring long-range dependencies. In this work, we focus on building general-purpose graph reasoners that could solve a range of graph-theoretic problems by exploiting the strength of LLM pretraining, and find that the ability also generalizes to the former domain-specific graph tasks. + +Benchmarking LLMs on Graph Reasoning. There is a growing interest in evaluating LLMs' graph reasoning abilities. NLGraph [41] evaluate LLMs on graph-theoretic tasks and discover preliminary yet brittle reasoning abilities in the face of spurious correlations and large graphs. Later, GraphArena [38] and GraCoRe [55] include a broader task coverage and recently released LLMs, finding that even OpenAI o1-mini struggles a lot with complex tasks. Moreover, GraphEval2000 [46] and ProGraph [23] emphasize code-oriented problem solving using library-based prompts, and GraphOmni [48] unify varying graph types, encodings, and prompt styles for a comprehensive evaluation. Overall, these benchmarks suggest that LLMs overall demonstrate moderate success on simple tasks but struggle with abstraction, generalization, and larger or more complex graph instances. Nevertheless, these datasets are either too small (e.g., thousands of examples) or not diverse enough (e.g., 8 tasks in NLGraph) for training general-purpose graph reasoners, which motivates the design of Erdős. + +Improving LLMs on Graph Reasoning. A major concern when using LLMs for graph tasks is the mismatch of data structure: LLMs take text sequences as input, while graphs have no natural order. Fatemi et al. [11] analyzed different graph encoding schemes for LLMs, such as adjacency lists and real-name networks, revealing that no single strategy proved universally optimal across all tasks and models. Subsequent explorations with different linearization orders [5], graph embeddings [31], or input modalities [9] have generally resulted in only modest improvements. Another thread of research proposes post-training LLMs using instruction tuning [27, 53] or preference tuning [3, 42, 39] on curated datasets of graph problems. However, the creation of diverse, high-quality instruction datasets at scale is challenging and expensive and requires extra supervision. Furthermore, models trained via distillation may only learn to memorize patterns and overfit to graph tasks [4]; in Section 5.2, we show that previous instruction-tuned models exhibit dramatic failures when generalizing to other data formats and reasoning tasks, while our RL training yields consistently better performance. + +Reinforcement Learning for LLMs Reasoning. Recent advances have demonstrated that LLMs can attain strong reasoning abilities in math and coding domains through RL, with representative work like OpenAI o1 [30] and DeepSeek R1 [13]. However, as discussed above, even o1 struggles a lot with graph reasoning tasks [55] and it is thus yet unclear whether RL can reliably and scalably improve LLMs' graph reasoning abilities. Our findings on G1 first confirm the effectiveness of RL on graph reasoning as well and suggest that applying RL to diverse graph-theoretic tasks with verifiable rewards is a scalable path for eliciting generalizable graph reasoning abilities of LLMs. + +# 3 Erdős: A Comprehensive Collection of Graph-theoretic Reasoning Tasks on Real-world Graphs + +To facilitate rule-based Reinforcement Learning of LLMs (aka. Reinforcement Learning from Verifiable Rewards (RLVR)) on graphs, we construct a diverse, large-scale collection of graph-theoretic reasoning tasks. We name it Erdős to remember Paul Erdős, a seminal figure with diverse contribu + +tions to graph theory. Compared to real-world graph tasks, these graph-theoretic tasks allow clear rule-based determination of rewards for the answers sampled from LLMs. We categorize these tasks into Easy, Medium, Hard, and Challenging, based on their inherent problem complexity as well as current LLMs' ability to solve them (see a full list in Table 1). For the training split, there are a total of 100,000 question-answer pairs, evenly distributed across tasks with 2,000 examples each. We also reserve 5,000 test pairs with different questions for evaluation. We include a detailed comparison of Erdős with other graph reasoning benchmarks in Appendix K.1. Erdős can serve as a dataset for training LLMs as well as a benchmark for evaluating LLMs on graph-theoretic tasks. We will release all task prompts, problems, chain-of-thought exemplars, and solution verification programs for public use. Below is a more detailed description of the data collection process. + +Graph-theoretic Tasks. We curate 50 graph-theoretic reasoning tasks available on NetworkX [14], one of the most widely used library for graph processing, and construct, as we know, the most comprehensive collection so far. In the difficulty level, the tasks vary from easy determination of graph attributes like node number counting, to well-known NP-hard problems like the traveling salesman problem. This collection includes both tasks for general graphs and tasks specific to directed graphs or weighted graphs, and covers a wide range of answer types including boolean, integer, float, node list, edge list, and node mapping. + +Answer Generation. To generate the golden answer for each problem, we utilize the default solvers of NetworkX to automatically solve the problem. If there are multiple solutions to each question, we use NetworkX-based programs to verify the correctness of each generated solution. The procedure ensures rigorous rewarding attribution, avoiding both costly human labeling and potential bias and hacking brought by LLM judges. + +Graph Sources. Most previous graph-theoretic datasets or benchmarks [41, 27, 3] consider random graphs, following Erdős-Rényi model [10] or Barabási-Albert model [1]. However, these random graph models are often far from graphs encountered in real-world practice. To mitigate this gap, we utilize the real-world graphs from the Network Repository [33], the largest network repository with thousands of donations in $30+$ domains. As these graphs can be very large and infeasible for LLMs, we downsample the graphs by random walk with a restart strategy, generating subgraphs with sizes from 5 to 35 nodes, following common settings in previous work [41, 55, 38]. + +Language Encoding. There are multiple ways to translate the graph structure into languages that LLMs can understand. Previous works explore serialized formats such as adjacency matrix, edge list, or graph embeddings [11, 8, 53], but fail to find a consistently good method. Here, we choose to describe the graph structure in a unified edge list format, e.g., (1, 2), (2, 3), ... In later experiments of Section 5.2, we show that our model trained on a single graph description method can even positively transfer to other formats. + +# 4 Training LLMs to Reason on Graphs + +In this section, we introduce the training pipeline that we explored for training G1. We design proper rule-based rewards for different graph tasks, while intentionally keeping the RL algorithm general and consistent with previous work. Similar to DeepSeek R1 [13], the training of G1 is very simple: it consists of a Reinforcement Learning phase for rewarding correct rollouts with the GRPO algorithm [36], and an optional SFT phase for warming up the model in the beginning (without which we call G1-Zero). We find that the SFT phase is generally beneficial for learning more challenging tasks, whose initial accuracy with the base model is close to zero. + +# 4.1 Reinforcement Learning of LLMs on Graphs + +Rule-based Rewards on Graphs. We design the following rule-based outcome reward model (ORM) for our training on graph-theoretic tasks, with a combination of value match, set matching, and algorithmic verification for different problems: + +- Strict value matching. For tasks that have a unique ground truth value, e.g., node counting, the policy receives a reward of $+1$ only when the generated answer is identical to the ground truth in terms of numerical value, e.g., 0.5 and $1/2$ , otherwise it receives a reward of 0. +- Jaccard Index for set matching. For problems whose answer is not a single value $\hat{s}$ but an unordered set, e.g., common neighbors of two nodes, the reward is defined as the Jaccard + +Index between the generated set $\hat{s}$ and the ground truth $s$ , i.e., $|s \cap \hat{s}| / |s \cup \hat{s}|$ . In this way, the model can receive intermediate rewards for imperfect solutions. + +- Algorithmic verification. Lastly, for problems that have multiple correct solutions (e.g., shortest paths) and it is not feasible to enumerate all of them, we implement algorithmic verifiers to check correctness of the proposed solutions. For instance, we determine the validity of a Hamiltonian path proposed by the policy by checking whether all the edges in the path exist and each node is visited exactly once. + +RL Algorithm. Following common practice [13], we use the Group Relative Policy Optimization (GRPO) [36] algorithm for RL training. Specifically, for each question $q \sim P(Q)$ drawn from the training set, GRPO first samples a set of responses $\{o_i\}_{i=1}^G$ from the policy model. The responses receive rewards $\{r_i\}_{i=1}^G$ , which enables calculating the group relative advantages $\{A_i\}_{i=1}^G$ : + +$$ +A _ {i} = \frac {r _ {i} - \operatorname {m e a n} \left(\left\{r _ {1} , r _ {2} , \cdots , r _ {G} \right\}\right)}{\operatorname {s t d} \left(\left\{r _ {1} , r _ {2} , \cdots , r _ {G} \right\}\right)}. \tag {1} +$$ + +Next, the policy model $\pi_{\theta}$ is updated by maximizing the following objective: + +$$ +\mathcal {J} _ {\mathrm {G R P O}} (\theta) = +$$ + +$$ +\mathbb {E} _ {q, \left\{o _ {i} \right\} _ {i = 1} ^ {G}} \frac {1}{G} \sum_ {i = 1} ^ {G} \left(\min \left(\frac {\pi_ {\theta} \left(o _ {i} \mid q\right)}{\pi_ {\theta_ {\text {o l d}}} \left(o _ {i} \mid q\right)} A _ {i}, \operatorname {c l i p} \left(\frac {\pi_ {\theta} \left(o _ {i} \mid q\right)}{\pi_ {\theta_ {\text {o l d}}} \left(o _ {i} \mid q\right)}, 1 - \epsilon , 1 + \epsilon\right) A _ {i}\right) - \beta \mathbb {D} _ {K L} \left(\pi_ {\theta} | | \pi_ {\text {r e f}}\right)\right), \tag {2} +$$ + +where the expectation is taken over $q \sim P(Q)$ and $\{o_i\}_{i=1}^G \sim \pi_{\theta_{old}}(O|q)$ . The KL divergence to the reference policy $\pi_{\mathrm{ref}}$ (base model) prevents large deviation from the pretrained model and circumvents severe overfitting. Besides, $\epsilon$ controls the clipping range of the probability ratios. + +# 4.2 Optional Warm-up with Supervised Fine-tuning + +During RL training, we have noticed that for some challenging tasks like isomorphic mapping (see Table 1), the initial accuracy of the base model is often so low that we frequently end up with only incorrect rollouts, producing no useful signal for RL training. This issue can be mitigated by using a stronger base model with higher initial performance; for example, R1 uses DeepSeek V3 (671B parameters) as its base model, although this inevitably increases compute cost. We find that introducing a short warm-up phase with supervised fine-tuning, aimed at teaching the model basic reasoning skills before the RL phase, effectively improves overall learning efficiency. Specifically, in this paper we consider two types of supervised fine-tuning. + +Direct-SFT. The first is direct supervised fine-tuning on question-answer pairs $(q, a)$ , where $q$ is the textual description of the problem and $a$ is the final answer without any intermediate reasoning steps. As discussed above, for graph-theoretic tasks, these question-answer pairs can often be synthesized by programming. However, this approach does not include the reasoning steps leading to the answers, meaning we cannot use it to explicitly teach the model reasoning processes. + +CoT-SFT. Secondly, we can collect reasoning trajectories via sampling $(q, c, a)$ triplets from another model [54], where $c$ represents the Chain-of-Thought (CoT) reasoning steps in natural language that lead to the final answer $a$ , and use them to fine-tune the base model. Specifically, we instruct a base model to generate potential solutions for each question $q$ , and only keep the correct responses that pass verification. This process is also called Rejection Sampling Fine-tuning (RFT) [54]. In practice, we use Qwen2.5-32B-Instruct [32], a more capable model for generating candidate solutions more reliably, ending up with around 4,500 training examples for the SFT phase. + +# 5 Experiments + +# 5.1 Benchmarking G1 on Graph-theoretic Reasoning Tasks + +Setup. As shown in Table 2, in the interest of academic compute budgets, we focus on comparing relatively small models. We include strong proprietary models (of unknown sizes) like GPT-4o-mini (non-reasoning) and OpenAI o3-mini (state-of-the-art reasoning), open-source instruction models like Qwen2.5-Instruct series (3B, 7B, 72B) [32], Qwen2.5-Math-Instruct [52], LLaMA-3 series (3B, 8B, + +Table 2: Test accuracy (%) comparison of different LLMs of varying sizes on our Erdős benchmark tasks. In all experiments we use Qwen2.5-Instruct models as our base model (marked below). We report the average accuracy across all tasks in the Average column, and full results for each task are provided in Appendix I.4. + +
ModelEasyMediumHardChallengingAverage
Proprietary (Unknown Parameters)
GPT-4o-mini76.2072.0728.813.3447.60
OpenAI o3-mini (w/ tool use)74.8383.4959.2843.2264.90
3B Parameters
Llama-3.2-3B-Instruct36.5021.456.811.1417.32
Qwen2.5-3B-Instruct (base model)45.7130.189.441.2922.72
Direct-SFT-3B (Ours)74.4375.2743.6914.4353.78
CoT-SFT-3B (Ours)65.5767.6429.444.5743.56
G1-3B (Ours)94.8684.6441.257.5759.76 (+37.04)
7B Parameters
Llama-3.1-8B-Instruct49.2130.4513.691.4325.10
Qwen2.5-7B-Instruct (base model)57.3642.5518.873.2932.06
Qwen2.5-Math-7B-Instruct52.7939.6414.822.4628.94
DeepSeek-R1-Distill-Qwen-7B71.7973.7339.1216.5751.64
GraphWiz-7B-RFT14.5713.731.380.477.70
GraphWiz-7B-DPO20.3619.091.440.7810.59
Direct-SFT-7B (Ours)73.5775.9139.1210.7151.76
CoT-SFT-7B (Ours)72.5775.7338.5011.0051.34
G1-7B (Ours)95.0788.9150.4423.5766.16 (+34.10)
70B Parameters
Llama-3.1-70B-Instruct68.0755.4531.874.4442.28
Qwen2.5-72B-Instruct71.7167.8133.378.2247.16
+ +70B) [12], and a strong baseline DeepSeek-R1-Distill-Qwen-7B [13] that is distilled from DeepSeek R1 with 671B parameters. Additionally, for reference, we incorporate previous training strategies for graph reasoning tasks such as GraphWiz-RFT and GraphWiz-DPO [3]. We finetune our model from Qwen2.5-Instruct models (3B and 7B) for 300 steps with batch size 512 on a cluster of $8 \times \mathrm{A}800$ GPUs, using our dataset Erdős. More experimental details can be found in Appendix A. + +Performance. As shown in Table 2, our proposed model G1-7B consistently outperforms most proprietary, open-source, and graph training counterparts by significant margins across all difficulty levels. With a notable average accuracy of $66.16\%$ , G1-7B outperforms GPT-4o-mini (47.60%) by $18.56\%$ , reaching competitive performance to a cutting-edge reasoning model like o3-mini (64.90%) that underwent much heavier training. Notably, our small variant G1-3B, delivers a strong average performance of $59.76\%$ , surpassing open-source models including Qwen2.5-72B-Instruct (47.16%) and Llama-3.1-70B-Instruct (42.28%) with $20 \times$ parameters. We also evaluate the GPT-4o model on a randomly sampled subset of Erdo's due to the cost budget. As shown in Appendix Table 14, G1-7B surpasses GPT-4o across all difficulty levels, exceeding it by over $10\%$ on average (65.29% vs. $55.13\%$ ), further validating the strong graph reasoning capabilities of the G1 models. + +Remark on SFT baselines. Interestingly, Direct-SFT emerges as a surprisingly strong baseline in Table 2. The 3B and 7B versions of Direct-SFT both outperform larger open-source models with $53.78\%$ and $51.76\%$ accuracy, suggesting that LLMs can discover some effective patterns by directly fitting targets. However, we also observe that with Direct-SFT, the 7B model yields no extra gain over the 3B model, while CoT-SFT and G1 (initialized with CoT-SFT) performance scales with larger models. This indicates that even though the CoT-SFT performance may appear low compared to Direct-SFT (possibly because of limited data size with about 100 examples per task), CoT-SFT could have better scaling and generalization properties. + +Robustness Analysis. To rigorously evaluate robustness, we conducted 32 repeated runs with different random seeds. The results in Appendix Table 17 demonstrate consistently small standard deviations ( $<1\%$ across all models and difficulty levels), confirming the stability of our method against potential randomness in LLM outputs. For prompt robustness, we rigorously test prompt sensitivity by having GPT-4o generate three semantically equivalent prompt variants. Appendix Table 18 shows minimal performance variance ( $<1.5\%$ standard deviation) across all models and difficulty levels, confirming our benchmark's stability to phrasing changes. + +Scaling G1 to 32B. To demonstrate the scalability of our approach, we extended the training methodology to develop G1-Zero-32B from Qwen2.5-32B-Instruct. As shown in Table 3, G1-Zero-32B achieves a $27.96\%$ improvement in average accuracy (from $47.10\%$ to $75.06\%$ ), with particularly notable gains in harder categories: $+31.87\%$ on Hard problems and $+26.43\%$ on Challenging problems. Furthermore, Appendix Table 16 demonstrates that G1-Zero-32B not only preserves but slightly enhances mathematical performance on standard benchmarks, which shows modest improvements on both GSM8K $(+0.08\%)$ and MATH $(+4.00\%)$ . + +Table 3: Test accuracy (%) of G1-Zero-32B and Qwen2.5-32B-Instruct on Erdős. + +
EasyMediumHardChallengingAverage
Qwen2.5-32B-Instruct70.5768.7333.389.0047.10
G1-Zero-32B97.7993.0065.2535.4375.06
+ +Scaling G1 to larger graphs. To verify the transferability of G1 to larger graphs, we construct a new test set of 5,000 graphs with 36-100 nodes, with other settings kept the same, which ensures there is no overlap between training and test data. Table 4 shows that both G1-3B and G1-7B achieve strong zero-shot generalization to these larger graphs without additional training, significantly outperforming the baselines across difficulty levels. These results demonstrate our method's effective scalability beyond the original training distribution. For larger graphs with thousands of nodes, G1 is bottle-necked by the context window limit of underlying LLMs, detailed in Appendix D. + +Table 4: Zero-shot generalization (accuracy in percentage) of G1 to larger graphs with 36-100 nodes. + +
EasyMediumHardChallengingAverage
Qwen2.5-3B-Instruct27.9828.535.260.2916.74
G1-3B79.3965.6618.463.7444.29
Qwen2.5-7B-Instruct37.8641.569.171.1723.94
G1-7B76.6570.6723.165.2246.46
+ +# 5.2 Transferability of G1 to Unseen Tasks and Domains + +In this section, we evaluate zero-shot generalization of G1 to unseen domains, tasks, and data formats. Detailed benchmark description and complete evaluation setups are provided in Appendix B. + +# 5.2.1 G1's Transferability to Other Graph Reasoning Benchmarks + +Setup. We consider two additional graph reasoning benchmarks, GraphWiz [3] and GraphArena [38], which bring three major shifts that challenge our model: 1) different distributions of the underlying graphs 2) tasks unseen during training 3) unfamiliar graph encoding formats, e.g., the GraphArena benchmark represents nodes with human names instead of integers. + +Results. The performance across models is reported in Table 5 and Table 6. On the GraphWiz benchmark, G1-7B achieves the highest overall accuracy (57.11%) among all models, outperforming DeepSeek-R1-Distill-Qwen-7B (51.86%) and even models specifically trained on GraphWiz data such as GraphWiz-7B-RFT (49.61%). The smaller variant G1-3B also achieves comparable performance with DeepSeek-R1-Distill-Qwen-7B. Similar results can be found on the GraphArena benchmark (Table 6) with a different graph encoding scheme. These results demonstrate that G1 has strong zero-shot generalization ability to unseen graph encoding methods, graph distributions, and graph tasks. Full results for GraphWiz and GraphArena are shown in Appendix I.1 and Appendix I.3. + +# 5.2.2 G1 on Real-world, Non-graph-theoretic Graph-reasoning Tasks + +Baseline. For real-world graph tasks, we consider two standard problems: node classification and link prediction. We adopt the benchmarks introduced by Wang et al. [44], which are constructed by subsampling from the widely used Cora and PubMed citation graphs. Each instance includes a description of the target node (or node pair) containing the paper ID and title, along with the textual + +Table 5: Test accuracy $(\%)$ by computational complexity on the GraphWiz benchmark. + +
ModelLinearPolyNP-CompleteAvg.
Llama-3.2-3B-Instruct29.803.002.5019.80
Qwen2.5-3B-Instruct (base)40.259.5869.1236.44
G1-3B58.0626.7569.1250.08
Llama-3.1-8B-Instruct54.005.6732.1233.03
DeepSeek-R1-Distill-Qwen-7B57.6931.4270.8851.86
GraphWiz-7B-RFT67.5629.8343.3849.61
GraphWiz-7B-DPO63.8836.2539.5049.25
Qwen2.5-7B-Instruct (base)49.0617.9276.1244.69
G1-7B68.0032.2572.6257.11
+ +Table 6: Test accuracy (%) by computational complexity on the GraphArena benchmark. + +
ModelPoly-TimeNP-CompleteAvg.
EasyHardEasyHard
Llama-3.2-3B-Instruct22.256.758.000.668.40
Qwen2.5-3B-Instruct (base)31.5014.5017.331.5014.85
G1-3B57.5026.7524.661.8324.80
Llama-3.1-8B-Instruct47.0021.2522.002.1620.90
DeepSeek-R1-Distill-Qwen-7B66.022.7534.831.5028.65
GraphWiz-7B-RFT2.250.750.830.000.85
GraphWiz-7B-DPO0.251.000.660.160.49
Qwen2.5-7B-Instruct (base)62.0035.7528.832.1628.84
G1-7B77.5044.2547.338.5041.10
+ +Table 7: Test accuracy $(\%)$ on Node Classification and Link Prediction benchmarks. + +
ModelNodeLinkAvg.
CoraPubMedCoraPubMed
Llama-3.2-3B-Instruct68.7775.2060.4057.6064.79
Qwen2.5-3B-Instruct (base)70.8375.0862.1558.3865.66
CoT-SFT-3B75.9781.4775.7071.5275.12
G1-3B77.2583.8878.9769.7575.16
Llama-3.1-8B-Instruct70.9075.0050.6046.1059.53
DeepSeek-R1-Distill-Qwen-7B76.5081.2568.0378.7278.80
Qwen2.5-7B-Instruct (base)79.3085.3588.2288.6785.50
CoT-SFT-7B73.2083.2564.7068.1273.17
G1-7B79.2086.2087.9891.8887.29
+ +Table 8: Test accuracy $(\%)$ on reasoning benchmarks beyond graph-related tasks. + +
ModelGSM8KMATHMMLU-pro
Llama-3.2-3B-Instruct71.0342.4013.50
Qwen2.5-3B-Instruct (base)81.9562.2038.53
CoT-SFT-3B75.3656.0034.85
G1-3B79.3061.8037.11
Llama-3.1-8B-Instruct74.4544.8032.02
DeepSeek-R1-Distill-Qwen-7B86.0387.2037.21
Qwen2.5-7B-Instruct (base)86.2769.8045.75
CoT-SFT-7B83.8565.8044.79
G1-7B87.4971.8048.56
+ +and structural information of neighboring nodes. These benchmarks emphasize the model's ability to leverage both local graph structure and textual attributes for effective prediction. + +Results. As shown in Table 7, our model G1 significantly outperforms both open-source and distilled baselines across tasks and model sizes. In the 3B model category, G1-3B surpasses the base model (Qwen2.5-3B-Instruct) by a large margin—especially in link prediction on Cora (+16.82%) and node classification on PubMed (+8.8%). In the 7B model category, G1-7B achieves the highest average score of 87.29%, ranking first on PubMed dataset in both node classification and link prediction tasks. Overall, G1 consistently demonstrates strong generalization across real-world graph tasks where graph-text reasoning is required. + +# 5.2.3 G1's Reasoning Ability beyond Graphs + +Setup. We next extend our investigations of G1's abilities beyond graph-based tasks. We consider two mathematics benchmarks, GSM8K [7] and MATH [16]. Additionally, we include MMLU-Pro [43], which is a massive multi-task benchmark covering disciplines such as chemistry, economics, and computer science. We believe the three benchmarks collectively provide a comprehensive assessment of G1's reasoning capabilities. + +Results. In table 8, we first notice that the CoT-SFT training on graph reasoning trajectories leads to a non-negligible degradation in general abilities, which could be attributed to the fact that SFT memorizes pattern instead of incentivizing truly generalizable skills [4]. Remarkably, the subsequent reinforcement learning stage—despite being trained exclusively on graph tasks—restores the reasoning abilities of both the 3B and the 7B model. G1-7B even surpasses the performance of the initial Qwen-7B checkpoint in all of the three benchmarks (87.49% v.s. 86.27% for GSM8K, 72.8% v.s. 69.8% for MATH, and 48.56% v.s. 45.75% for MMLU-pro). Interestingly, G1-7B also outperforms Qwen-7B-Instruct on several non-STEM tasks like Economy (68.76 v.s. 46.87), which are intuitively less related to graph reasoning (see Appendix I.2 for full MMLU-Pro results). We further provide a detailed analysis on the transferability of G1 to mathematics tasks in Appendix H, showing G1 mainly improves the numerical calculation and utilization of known information. + +# 5.3 Training Analysis + +In this section, we further analyze the influence of two training factors on G1's reasoning performance. + +Data Mixture. In Table 2, we observe that although G1-3B achieves strong overall performance, it is outperformed by Direct-SFT-3B on the Hard and Challenging subsets. We hypothesize that this gap arises from imbalanced reward signals across different difficulty levels during RL training. Since correct rollouts are much easier to obtain on simpler tasks, the policy tends to allocate more of its constrained probability ratios as well as KL budget to optimize for Easy and Medium tasks, thereby maximizing the overall rewarad. To test this hypothesis, we introduce G1-Hard-3B, which is trained exclusively on Hard and Challenging tasks during RL. As shown in Table 9, this model achieves the highest accuracy on Hard (48.50%) and Challenging (17.43%) tasks, surpassing both G1 and Direct-SFT. These results support our claim, suggesting that the suboptimal performance of G1-3B on challenging tasks is a natural consequence of the uniformly weighted reward function, rather than a shortcoming of G1 training pipeline. Notably, despite being trained only on hard tasks, G1-Hard-3B also generalizes to Easy and Medium tasks (69.36% and 70.64%), far exceeding the baseline Qwen2.5-3B-Instruct. This indicates that learning to solve difficult tasks confers transferable reasoning skills that benefit performance on simpler problems. To better balance the optimization process across difficulty levels, we further explore reward-weighting strategies in Appendix J. + +Table 9: Test accuracy (\%) on our benchmark. $\star$ denotes the tasks are excluded in model training. G1-Hard-3B is only RL-trained on Hard and Challenging tasks. + +
CategoryModelEasyMediumHardChallengingAverage
Base ModelQwen2.5-3B-Instruct45.7130.189.441.2922.72
OursDirect-SFT-3B74.4375.2743.6914.4353.78
G1-3B94.8684.6441.257.5759.76
G1-Hard-3B69.36*70.64*48.5017.4353.30
+ +SFT Warmup. We study the role of SFT as a cold-start mechanism for RL, evaluating its impact on both performance and response behavior. To isolate the effect of SFT, we compare two variants: G1-Zero-3B that is directly trained from the base model Qwen2.5-3B-Instruct with RL, and G1-3B that initializes RL from the CoT-SFT checkpoint. As shown in Figure 1, training RL directly from the base model achieves surprisingly strong performance, aligning with recent findings in Deepseek-R1-Zero [13]. Meanwhile, initializing RL with CoT-SFT provides clear and consistent improvements across all difficulty levels, with an average accuracy of $59.8\%$ compared to $50.1\%$ of G1-Zero-3B. Besides, we notice that relative improve + +![](images/bb5f4f0201dedd21b4c74ca15e0f5590eb3f3925b0b422f37d946b02e40d5046.jpg) +Figure 1: Test accuracy comparison of G1-3B and G1-Zero-3B on our benchmark. We include results for -7B in Appendix C. + +ments become larger as the difficulty increases. In addition to performance gains, we also observe that models initialized by CoT-SFT present more precise reasoning patterns, illustrated by the case study in the following section. + +# 5.4 Understanding the Benefits of RL Training for Graph Reasoning + +To understand how RL training helps graph reasoning, we take shortest path (a Hard task) as a case study. Specifically, we study the behaviors of three models: Qwen2.5-3B-Instruct (base), G1-Zero-3B (RL only), and G1-3B (SFT & RL). + +We identify three primary approaches adopted by the models to solve the problem: 1) Breadth-First Search (BFS), 2) Dijkstra's algorithm, and 3) Intuitive deductions. Figure 3a shows the distribution of these approaches alongside their corresponding accuracies for Qwen2.5-3B-Instruct. On unweighted graphs, BFS is the most efficient method and yields the highest performance. In contrast, Dijkstra's algorithm is best suited for weighted graphs, where it correctly accounts for edge costs. However, its reliance on a min-priority queue and a distance list introduces computational complexity, which appears to challenge Qwen2.5-3B-Instruct and results in its lowest observed accuracy. For example, as shown in Figure 2 (left), the model falsely states that node 4 has no edges (node 4 is connected + +Query: The task is to determine the shortest path between two nodes. Here is an undirected graph containing nodes from 1 to 7. The edges are: (1, 3), (1, 6), (1, 5), (3, 2), (3, 7), (5, 2), (7, 4). Question: What is the shortest path between node 6 and node 4? + +Ground Truth: [6, 1, 3, 7, 4] + +Question + +![](images/fc25d6d82d9d0966b23ac9e299435ed2526ed664dc1ea786a0e89219dd00c8b7.jpg) + +We can use Dijkstra's algorithm to find the shortest path. Let's start by initializing the distances from node 4 to all other nodes and setting the distance to the starting node (node 4) to 0. The distance to all other nodes is initially set to infinity. + +1. Initialize distances: -Distance to 4:0 .... Distance to $7:\infty$ + +2. Select the node with the smallest distance (node 4) and update the distances to its neighbors: + +8. Select the next smallest distance node (node 7), update the distances to its neighbors: .... + +Since we have reached node 4 and the distance to it is 5, we can now trace back the path: + +- From node 7 to node 4: distance is 2 (edge (7, 4)) .... + +The shortest path from node 6 to node 4 is: [6, 5, 2, 3, 7, 4] + +Therefore, the final answer is: \\(boxed{[6, 5, 2, 3, 7, 4]}. + +To determine the shortest path between node 6 and node 4, we will analyze the given graph step by step. +First, let's list the nodes and their connections: - Node 1 is connected to nodes 3, 6, and 5. +We need to find the path from node 6 to node 4. Let's start from node 6 and see which nodes we can reach next. +1. Node 6 is connected to node 1. +2. Node 1 is connected to nodes 3, 6, and 5. +3. Since we are already at node 6, we can try to reach the node 10 from node 6. We can do so by using node 4 so let's see if we can get to node 4 from there. +Now, let's look at the connections from node 3: +Now, let's look at the connections from node 7: +1. Node 7 is connected to nodes 3 and 4. +2. Since we are already at node 7, we can reach node 4 directly from here. +So, the path from node 6 to node 4 is: [6 to 1 to 3 to 7 to 4] +Therefore, the final answer is:boxed{[6, 1, 3, 7, 4]} + +To find the shortest path between node 6 and node 4 in the given undirected graph, we can use a graph traversal method such as Breadth-First Search (BFS). +Let's represent the graph as an adjacency list for easier visualization: - Node 1: connected to 3, 6, 5 +Starting from node 6, we will perform a BFS to find the shortest path to node 4. +1. Initialize the queue with the starting node 6. +2. Keep track of visited nodes to avoid revisiting. +3. Keep track of the node by backtracking from the end to the start node. +Let's execute the BFS: +- Start at node 6. +- Explore neighbors of 6: the only neighbor is 1 +- From node 1, explore its neighbors: 3, 5, 6 (but 6 is already visited)...... +- From node 7, explore its neighbors: 4 (this is the target node). +Thus, the shortest path found is: $6 \rightarrow 1 \rightarrow 3 \rightarrow 7 \rightarrow 4$ . +Therefore, the final answer is: $boxed{[6, 1, 3, 7, 4]}$ + +Qwen2.5-3B-Instruct + +G1-Zero-3B + +G1-3B + +to node 7) while updating the distance list. Interestingly, intuitive approaches—where the model attempts to visually estimate or heuristically trace paths—can also produce correct answers by a noticeable accuracy, particularly on small graphs. + +We proceed by observing that RL training significantly reshapes the models' graph reasoning strategies: RL-trained models largely abandon Dijkstra and prefer a combination of BFS and intuitive search. As shown in Figure 3b and Figure 2 (middle), G1-Zero-3B navigates the graph in a manner akin to human heuristics—sequentially checking neighbors and adjusting paths dynamically. G1-3B primarily adopts a neat BFS-style algorithm as in Figure 3b and Figure 2 (right), executing it with high precision, occasionally resorting to intuitive strategies for simple graphs. To conclude, + +our case study highlights how RL training enhances graph reasoning by guiding LLMs toward more model-aware strategies that are adaptive to their inherent capabilities [47]. + +Figure 3: Reasoning patterns for the shortest path task. +![](images/cf735ba70444b59c60be47c705e4ae366eeec137be316c37bc0cffe48bbd4a99.jpg) +(a) The accuracy of different graph reasoning patterns for shortest path on Qwen2.5-3B-Instruct. + +Figure 2: An intuitive illustration of the differences in solution strategies employed by Qwen2.5-3B-Instruct, G1-Zero-3B, and G1-3B for a shortest path problem. +![](images/2f9b38c732ee2af90dde852c0db5574fbe86ff727ee195c8942a915c0a04e992.jpg) +(b) Frequency of different graph reasoning patterns for Qwen2.5-3B-Instruct, G1-Zero-3B and G1-3B. + +# 6 Discussion + +In this paper, we explored the use of Reinforcement Learning to improve LLMs' reasoning abilities on graph reasoning and demonstrate significant improvements across a spectrum of tasks with various difficulty levels, showing that graph reasoning of LLMs can be elicited via RL training (even with only 300 steps). We also comprehensively evaluate the transferability of RL-trained models to unseen graph reasoning tasks, real-world graph tasks, and general reasoning tasks, observing strong zero-shot generalization. These results support our hypothesis that training LLMs on diverse synthetic graph-theoretic tasks via RL offers a scalable, generalizable path toward robust graph reasoning. As a first step, this approach may guide the development of efficient, general-purpose graph reasoners. + +In future work, we aim to explore dynamic difficulty scheduling during RL training to address the sample inefficiency issue. On a broader scale, we also plan to extend our approach to the following scenarios: (1) Handling larger graphs with thousands of nodes, aligning with the long-context reasoning challenges. (2) Incorporating visual inputs (e.g., images depicting graphs) to enhance real-world applicability. (3) Adapting G1 to more practical domains such as logistics, knowledge graph reasoning, and tabular problem-solving, where structured reasoning is critical. + +# 7 Limitations + +While G1 demonstrates significant improvements in graph reasoning through RL training, it inherits GRPO's sample inefficiency and requires extensive rollouts for challenging tasks (e.g., NP-hard problems), which might be solved with dynamic difficulty scheduling during training. Although G1 demonstrates strong generalization to real-world graph tasks, e.g., node classification and link prediction, its generalization to highly domain-specific applications (e.g., molecular property prediction) and other structured data (e.g., tabular, time-series) remains untested. + +# Social Impacts + +G1's graph reasoning capabilities offer practical benefits for education (e.g., adaptive graph theory tutoring) and scientific research. While the technology could amplify existing biases in training data, these risks can be mitigated through careful dataset design and human oversight. When developed responsibly, G1 has the potential to support human experts in solving complex graph problems. + +# Acknowledgement + +Yisen Wang was supported by Beijing Natural Science Foundation (L257007), Beijing Major Science and Technology Project under Contract no. Z251100008425006, National Natural Science Foundation of China (92370129, 62376010), Beijing Nova Program (20230484344, 20240484642), and State Key Laboratory of General Artificial Intelligence. Yifei Wang and Stefanie Jegelka were supported in part by the NSF AI Institute TILOS (NSF CCF-2112665), and an Alexander von Humboldt Professorship. + +# References + +[1] Barabási, A.-L. and Albert, R. Emergence of scaling in random networks. Science, 286(5439): 509-512, 1999. +[2] Brown, T. B., Mann, B., Ryder, N., Subbiah, M., Kaplan, J., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., Agarwal, S., Herbert-Voss, A., Krueger, G., Henighan, T., Child, R., Ramesh, A., Ziegler, D. M., Wu, J., Winter, C., Hesse, C., Chen, M., Sigler, E., Litwin, M., Gray, S., Chess, B., Clark, J., Berner, C., McCandlish, S., Radford, A., Sutskever, I., and Amodei, D. Language models are few-shot learners. In NeurIPS, 2020. +[3] Chen, N., Li, Y., Tang, J., and Li, J. Graphwiz: An instruction-following language model for graph computational problems. In SIGKDD, 2024. +[4] Chu, T., Zhai, Y., Yang, J., Tong, S., Xie, S., Schuurmans, D., Le, Q. V., Levine, S., and Ma, Y. Sft memorizes, rl generalizes: A comparative study of foundation model post-training. arXiv preprint arXiv:2501.17161, 2025. +[5] Chu, X., Xue, H., Tan, Z., Wang, B., Mo, T., and Li, W. Graphsos: Graph sampling and order selection to help lms understand graphs better. arXiv preprint arXiv:2501.14427, 2025. +[6] Cobbe, K., Kosaraju, V., Bavarian, M., Chen, M., Jun, H., Kaiser, L., Plappert, M., Tworek, J., Hilton, J., Nakano, R., Hesse, C., and Schulman, J. Training verifiers to solve math word problems. arXiv preprint arXiv:2110.14168, 2021. +[7] Cobbe, K., Kosaraju, V., Bavarian, M., Chen, M., Jun, H., Kaiser, L., Plappert, M., Tworek, J., Hilton, J., Nakano, R., et al. Training verifiers to solve math word problems. arXiv preprint arXiv:2110.14168, 2021. +[8] Dai, X., Qu, H., Shen, Y., Zhang, B., Wen, Q., Fan, W., Li, D., Tang, J., and Shan, C. How do large language models understand graph patterns? a benchmark for graph pattern comprehension. In ICLR, 2025. +[9] Das, D., Gupta, I., Srivastava, J., and Kang, D. Which modality should i use-text, motif, or image?: Understanding graphs with large language models. In NAACL, 2024. +[10] Erdős, P. Erdős-rényi model. Publ. Math. Debrecen, pp. 290-297, 1959. + +[11] Fatemi, B., Halcrow, J., and Perozzi, B. Talk like a graph: Encoding graphs for large language models. arXiv preprint arXiv:2310.04560, 2023. +[12] Grattafori, A., Dubey, A., Jauhri, A., Pandey, A., Kadian, A., Al-Dahle, A., Letman, A., Mathur, A., Schelten, A., Vaughan, A., Yang, A., Fan, A., Goyal, A., Hartshorn, A., Yang, A., Mitra, A., Sravankumar, A., Korenev, A., Hinsvark, A., Rao, A., Zhang, A., Rodriguez, A., Gregerson, A., Spataru, A., Roziere, B., Biron, B., Tang, B., Chern, B., Caucheteux, C., Nayak, C., Bi, C., Marra, C., McConnell, C., Keller, C., Touret, C., Wu, C., Wong, C., Ferrer, C. C., Nikolaidis, C., Allonsius, D., Song, D., Pintz, D., Livshits, D., Wyatt, D., Esiobu, D., Choudhary, D., Mahajan, D., Garcia-Olano, D., Perino, D., Hupkes, D., Lakomkin, E., AlBadawy, E., Lobanova, E., Dinan, E., Smith, E. M., Radenovic, F., Guzmán, F., Zhang, F., Synnaeve, G., Lee, G., Anderson, G. L., Thattai, G., Nail, G., Mialon, G., Pang, G., Cucurell, G., Nguyen, H., Korevaar, H., Xu, H., Touvron, H., Zarov, I., Ibarra, I. A., Kloumann, I., Misra, I., Evtimov, I., Zhang, J., Copet, J., Lee, J., Geffert, J., Vranes, J., Park, J., Mahadeokar, J., Shah, J., van der Linde, J., Billock, J., Hong, J., Lee, J., Fu, J., Chi, J., Huang, J., Liu, J., Wang, J., Yu, J., Bitton, J., Spisak, J., Park, J., Rocca, J., Johnstun, J., Saxe, J., Jia, J. Alwala, K. V. Prasad K. Upasani K. Plawiak K. Li K. Heafield K. Stone K. El-Arini K. Iyer K. Malik K. Chiu K. Bhalla K. Lakhotia K. Rantala-Yeary L. van der Maaten L. Chen L. Tan L. Jenkins L. Martin L. Madaan L. Malo L. Blecher L. Landzaat L. de Oliveira L. Muzzi M. Pasupuleti M. Singh M. Paluri M. Kardas M. Tsimpoukelli M. Oldham M. Rita M. Pavlova M. Kambadur M. Lewis M. Si M. Singh M. K. Hassan M. Goyal N. Torabi N. Bashlykov N. Bogoychev N. Chatterji N. Zhang N.Duchenne O.Celebi O.Arlassy P.Zhang P.Li P.Vasic P Weng P.Bhargava P.Dubal P.Krishnan P.Koura P.S.Xu P.He Q.Dong Q.Srinivasan R.Ganapathy R.Calderer R.Cabral R.S.Stojnic R.Raileanu R.Maheswari R.Girdhar R.Patel R.Sauvestre R.Polidoro R.Sumbaly R.Taylor R.Silva R.Hou R.Wang R Hosseini S.Chennabasappa S.Singh S.Bell S.Kim S.S.Edunov S.Nie S.Narang S Raparthy S.Shen S.Wan S.Bhosale S.Zhang S.Vandenhende S.BatraS.Whittman S.Sootla S.Collot S.Gururangan S.Borodinsky S.Herman T.Fowler T.Sheasha T Georgiou T.Scialom T.Speckbacher T.Mihaylov T.Xiao T.Karn U.Goswami V Gupta V.Ramanathan V.Kerkez V.Gonguet V.DoV.Vogeti V.Albiero V.Petrovic V.Chu W.XiongW.FuW.Meers W.Martinet X.WangX.WangX.Tan X.E.Xia X.Xie X.Jia X.Wang X.Goldschlag Y.Gaur Y.Babaei Y.Wen YSongY.Zhang Y.Li Y.Mao Y.Coudert Z.D.Yan Z.Chen Z.Papakipos Z.SinghA.Srivastava A.Jain A.Kelsey A.Shajnfeld A.Gangidi A.Victoria A.Goldstand A.Menon A. Sharma A.Boesenberg A.Baevski A.Feinstein A.Kallet A.Sangani A.Teo A.Yunus A.Lupu A.Alvarado A.Caples A.GuA.HoA.Poulton A.RyanA.Ramchandani A.DongA.FrancoA.GoyalA.SarafA.ChowdhuryA.GabrielA.BharambeA.Eisenman A.Yazdan A.James B.Maurer B.Leonhardi B.HuangB.Loyd B.Paola B.D.Paranjape B.Liu B.Wu B.Ni B.Hancock B.Wasti B.Spence B.Stojkovic B.Gamido B.Montalvo B.Parker C.Burton C.Mejia C.Liu C.WangC.KimC Zhou C.Hu C.Cai C.Tindal C.Feichtenhofer C.Gao C.Civin D.Beaty D.Kreymer D.Li D.Adkins D.Xu D.Testuggine DDavid D.Parikh D.Liskovich D.FossD.WangD.LeD.HollandD.DowlingE.JamilE.Montgomery E.Presani E.Hahn E.Wood E.LeE.-T.Brinkman E.Arcaute E.Dunbar E.Smothers E.Sun F.Kreuk F.Tian F.Kokkinos F.Ozgenel F.Caggioni F.Kanayet F.Seide F.Florez G.M.Schwarz G.Badeer G.Swee G.Halpern G.Herman G.Sizov G.Guangyi Zhang Lakshminarayanan G.Inan H.Shojanazeri H.Zou H.WangH.Zha H.Habeeb H. Rudolph H.Suk H.Aspegren H.Goldman H.Zhan H.Damlaj I.Molybog I.Tufanov I.Leontiadis I.Veliche I-E.Gat I.Weissman J.Geboski J.Kohl J.LamJ.AsherJ. GayaJ-B.MarcusJ.TangJChanJ.ZhenJ.ReizensteinJ.TeboulJ.ZhongJ.Jin J.YangJ.CummingsJ.CarvilleJ.ShepardJ.McPhieJ.TorresJ.GinsburgJ.Wang J.WuK.U.K.HaxenaK.Khandelwal K.ZandK.Matosich K.Veeraraghavan K Michelena K.Li K.Jagadeesh K.HuangK.ChawlaK.HuangK.ChenL.GargL.AL SilvaL.BellL.ZhangL.GuoL.YuL.Moshkovich L.Wehrstedt L.Khabsa M.Avalani M.Bhatt M.Kankus M.Hasson M.Lennie M.Reso M.Groshev M.Naumov M.Lathi M.Keneally M.Liu M.Seltzer M.L.LalkoM.Restrepo M.Patel M.VyatskovM.SamvelyanM.ClarkM.MaceyM.WangM.Hermoso M.J.Metanat M.Rastegari M.Bansal M.Santhanam N.Parks N.White N.Bawa N.Singhal N.Egebo N.Usunier N.Mehta N.Laptev N.P.DongN.ChengN.Chernoguz O.Hart O.Salpekar O.Kalinli O.Kent P.Parekh P.SaabP.Balaji P.Rittner P.Bontrager P.Roux P.Dollar P.Zvyagina P.Ratanchandani P.Yuvraj P.LiangQ.AlaoR.Rodriguez R + +Ayub, R., Murthy, R., Nayani, R., Mitra, R., Parthasarathy, R., Li, R., Hogan, R., Battey, R., Wang, R., Howes, R., Rinott, R., Mehta, S., Siby, S., Bondu, S. J., Datta, S., Chugh, S., Hunt, S., Dhillon, S., Sidorov, S., Pan, S., Mahajan, S., Verma, S., Yamamoto, S., Ramaswamy, S., Lindsay, S., Lindsay, S., Feng, S., Lin, S., Zha, S. C., Patil, S., Shankar, S., Zhang, S., Zhang, S., Wang, S., Agarwal, S., Sajuyigbe, S., Chintala, S., Max, S., Chen, S., Kehoe, S., Satterfield, S., Govindaprasad, S., Gupta, S., Deng, S., Cho, S., Virk, S., Subramanian, S., Choudhury, S., Goldman, S., Remez, T., Glaser, T., Best, T., Koehler, T., Robinson, T., Li, T., Zhang, T., Matthews, T., Chou, T., Shaked, T., Vontimitta, V., Ajayi, V., Montanez, V., Mohan, V., Kumar, V. S., Mangla, V., Ionescu, V., Poenaru, V., Mihalescu, V. T., Ivanov, V., Li, W., Wang, W., Jiang, W., Bouaziz, W., Constable, W., Tang, X., Wu, X., Wang, X., Wu, X., Gao, X., Kleinman, Y., Chen, Y., Hu, Y., Jia, Y., Qi, Y., Li, Y., Zhang, Y., Zhang, Y., Adi, Y., Nam, Y., Yu, Wang, Zhao, Y., Hao, Y., Qian, Y., Li, Y., He, Y., Rait, Z., DeVito, Z., Rosnbrick, Z., Wen, Z., Yang, Z., Zhao, Z., and Ma, Z. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. +[13] Guo, D., Yang, D., Zhang, H., Song, J., Zhang, R., Xu, R., Zhu, Q., Ma, S., Wang, P., Bi, X., et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948, 2025. +[14] Hagberg, A. A., Schult, D. A., and Swart, P. J. Exploring network structure, dynamics, and function using networkx. In SciPy, 2008. +[15] Hamilton, W., Ying, Z., and Leskovec, J. Inductive representation learning on large graphs. In NeurIPS, 2017. +[16] Hendrycks, D., Burns, C., Kadavath, S., Arora, A., Basart, S., Tang, E., Song, D., and Steinhardt, J. Measuring mathematical problem solving with the math dataset. In NeurIPS, 2021. +[17] Hsieh, C.-P., Sun, S., Kriman, S., Acharya, S., Rekesh, D., Jia, F., Zhang, Y., and Ginsburg, B. Ruler: What's the real context size of your long-context language models? arXiv preprint arXiv:2404.06654, 2024. +[18] Huang, X., Zhang, J., Li, D., and Li, P. Knowledge graph embedding based question answering. In WSDM, 2019. +[19] Karalias, N. and Loukas, A. Erdos goes neural: an unsupervised learning framework for combinatorial optimization on graphs. In NeurIPS, 2020. +[20] Kipf, T. N. and Welling, M. Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907, 2016. +[21] Kong, L., Feng, J., Liu, H., Huang, C., Huang, J., Chen, Y., and Zhang, M. Gofa: A generative one-for-all model for joint graph language modeling. In ICLR, 2025. +[22] Kwon, W., Li, Z., Zhuang, S., Sheng, Y., Zheng, L., Yu, C. H., Gonzalez, J. E., Zhang, H., and Stoica, I. Efficient memory management for large language model serving with pagedattention. In SIGOPS, 2023. +[23] Li, X., Chen, W., Chu, Q., Li, H., Sun, Z., Li, R., Qian, C., Wei, Y., Shi, C., Liu, Z., et al. Can large language models analyze graphs like professionals? a benchmark, datasets and models. In NeurIPS, 2024. +[24] Li, Y., Pan, Z., Lin, H., Sun, M., He, C., and Wu, L. Can one domain help others? a data-centric study on multi-domain reasoning via reinforcement learning. arXiv preprint arXiv:2507.17512, 2025. +[25] Lightman, H., Kosaraju, V., Burda, Y., Edwards, H., Baker, B., Lee, T., Leike, J., Schulman, J., Sutskever, I., and Cobbe, K. Let's verify step by step. arXiv preprint arXiv:2305.20050, 2023. +[26] Liu, H., Feng, J., Kong, L., Liang, N., Tao, D., Chen, Y., and Zhang, M. One for all: Towards training one graph model for all classification tasks. In ICLR, 2024. +[27] Luo, Z., Song, X., Huang, H., Lian, J., Zhang, C., Jiang, J., and Xie, X. Graphinstruct: Empowering large language models with graph understanding and reasoning capability. arXiv preprint arXiv:2403.04483, 2024. + +[28] Mao, H., Chen, Z., Tang, W., Zhao, J., Ma, Y., Zhao, T., Shah, N., Galkin, M., and Tang, J. Position: Graph foundation models are already here. In ICML, 2024. +[29] Mirhoseini, A., Goldie, A., Yazgan, M., Jiang, J. W., Songhori, E., Wang, S., Lee, Y.-J., Johnson, E., Pathak, O., Nova, A., et al. A graph placement methodology for fast chip design. Nature, 594(7862):207-212, 2021. +[30] OpenAI, J. Jaech, A., Kalai, A., Lerer, A., Richardson, A., El-Kishky, A., Low, A., Helyar, A., Madry, A., Beutel, A., Carney, A., Iftimie, A., Karpenko, A., Passos, A. T., Neitz, A., Prokofiev, A., Wei, A., Tam, A., Bennett, A., Kumar, A., Saraiva, A., Vallone, A., Duberstein, A., Kondrich, A., Mishchenko, A., Applebaum, A., Jiang, A., Nair, A., Zoph, B., Ghorbani, B., Rossen, B., Sokolowsky, B., Barak, B., McGrew, B., Minaiev, B., Hao, B., Baker, B., Houghton, B., McKinzie, B., Eastman, B., Lugaresi, C., Bassin, C., Hudson, C., Li, C. M., de Bourcy, C., Voss, C., Shen, C., Zhang, C., Koch, C., Orsinger, C., Hesse, C., Fischer, C., Chan, C., Roberts, D., Kappler, D., Levy, D., Selsam, D., Dohan, D., Farhi, D., Mely, D., Robinson, D., Tsipras, D., Li, D., Oprica, D., Freeman, E., Zhang, E., Wong, E., Proehl, E., Cheung, E., Mitchell, E., Wallace, E., Ritter, E., Mays, E., Wang, F., Such, F. P., Raso, F., Leoni, F., Tsimpourlas, F., Song, F., von Lohmann, F., Sulit, F., Salmon, G., Parascandolo, G., Chabot, G., Zhao, G., Brockman, G., Leclerc, G., Salman, H., Bao, H., Sheng, H., Andrin, H., Bagherinezhad, H., Ren, H., Lightman, H., Chung, H. W., Kivlichan, I., O'Connell, I., Osband, I., Gilaberte, I. C., Akkaya, I., Kostrikov, I., Sutskever, I., Kofman, I., Pachocki, J., Lennon, J., Wei, J., Harb, J., Twore, J., Feng, J., Yu, J., Weng, J., Tang, J., Yu, J., Candela, J. Q., Palermo, J., Parish, J., Heidecke, J., Hallman, J., Rizzoj, J., Gordon, J., Uesato, J., Ward, J. Huizinga, J. Wang. J. Chen K. Xiao K. Singhal K. Nguyen K.Cobbe K. Shi K.Wood K.RimbachK. Gu-Lemberg K.Liu K.Lu K.Stone K.Yu K.Ahmad L.Yang L.Liu L.Maksin L. HoL.FedusL.WengL.LiL.McallumL.HeldL.KuhnL.KondraciuK.L Kaiser L. Metz L. Boyd M.Trebacz M.Joglekar M.Chen M.Tintor M.Meyer M.Jones M. Kaufer M.Schwarzer M.Shah M.Yatbaz M.Guan M.Y.Xu M.Yan M.Glaese M. Chen M. Lampe M. Malek M. WangM.Fradin M.McClayM.PavlovM.Wang M.WangM.Murati M.Bavarian M.Rohaninejad M.McAleese N.Chowdhury N. Chowdhury N.Ryder N.Tezak N.Brown N.Nachum O.Boiko O.Murk O.Watkins O.Chao P.Ashbourne P.Izmailov P.Zhokhov P.Dias R.Arora R.Lin R.Lopes R.G Gaon R.Miyara R.Leike R.Hwang R.Garg R.Brown R.James R.Shu R.Cheu R. Greene R.Jain S.Altman S.Toizer S.Toyer S.Miserendino S.Agarwal S.Hernandez S.Baker S.McKinneyS.Yan S.Zhao S.Hu S.Santurkar S.Chaudhuri S.R.Zhang S.FuS.PapayS.LinS.BalajiS.Sanjeev S.SidorS.BrodaT.ClarkA.WangT. Gordon T.Sanders T.Patwardhan T.Sottiaux T.Degrty T.Dimson T.Zheng T.Garipov T.Stasi T.Bansal T.Creech T.Peterson T.Eloundou T.QiV.KosarajuV.MonacoV. PongV.FomenkoV.ZhengW.ZhouW.McCabeW.ZarembaW.DuboisY.LuY. Chen YChaY.BaiY.HeY.ZhangY.WangY.ShaoZ.and LiZ.Openai o1 system card.arXiv preprint arXiv:2412.167202024. +[31] Perozzi, B., Fatemi, B., Zelle, D., Tsitsulin, A., Kazemi, M., Al-Rfou, R., and Halcrow, J. Let your graph do the talking: Encoding structured data for llms. arXiv preprint arXiv:2402.05862, 2024. +[32] Qwen, :, Yang, A., Yang, B., Zhang, B., Hui, B., Zheng, B., Yu, B., Li, C., Liu, D., Huang, F., Wei, H., Lin, H., Yang, J., Tu, J., Zhang, J., Yang, J., Yang, J., Zhou, J., Lin, J., Dang, K., Lu, K., Bao, K., Yang, K., Yu, L., Li, M., Xue, M., Zhang, P., Zhu, Q., Men, R., Lin, R., Li, T., Tang, T., Xia, T., Ren, X., Ren, X., Fan, Y., Su, Y., Zhang, Y., Wan, Y., Liu, Y., Cui, Z., Zhang, Z., and Qiu, Z. Qwen2.5 technical report. arXiv preprint arXiv:2412.15115, 2025. +[33] Rossi, R. A. and Ahmed, N. K. The network data repository with interactive graph analytics and visualization. In AAAI, 2015. +[34] Sanford, C., Fatemi, B., Hall, E., Tsitsulin, A., Kazemi, M., Halcrow, J., Perozzi, B., and Mirrokni, V. Understanding transformer reasoning capabilities via graph algorithms. In NeurIPS, 2024. +[35] Sato, R., Yamada, M., and Kashima, H. Approximation ratios of graph neural networks for combinatorial problems. In NeurIPS, 2019. + +[36] Shao, Z., Wang, P., Zhu, Q., Xu, R., Song, J., Bi, X., Zhang, H., Zhang, M., Li, Y., Wu, Y., et al. Deepseekmath: Pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300, 2024. +[37] Silver, D., Schrittwieser, J., Simonyan, K., Antonoglou, I., Huang, A., Guez, A., Hubert, T., Baker, L., Lai, M., Bolton, A., Chen, Y., Lillicrap, T., Hui, F., Sifre, L., van den Driessche, G., Graepel, T., and Hassabis, D. Mastering the game of go without human knowledge. Nature, 550(7676):354-359, 2017. +[38] Tang, J., Zhang, Q., Li, Y., and Li, J. Grapharena: Benchmarking large language models on graph computational problems. In ICLR, 2025. +[39] Velicković, P., Ying, R., Padovano, M., Hadsell, R., and Blundell, C. Neural execution of graph algorithms. In ICLR, 2020. +[40] Wang, H., Wang, K., Yang, J., Shen, L., Sun, N., Lee, H.-S., and Han, S. Gcn-rl circuit designer: Transferable transistor sizing with graph neural networks and reinforcement learning. In DAC, 2020. +[41] Wang, H., Feng, S., He, T., Tan, Z., Han, X., and Tsvetkov, Y. Can language models solve graph problems in natural language? In NeurIPS, 2023. +[42] Wang, J., Wu, J., Hou, Y., Liu, Y., Gao, M., and McAuley, J. Instructgraph: Boosting large language models via graph-centric instruction tuning and preference alignment. In ACL, 2024. +[43] Wang, Y., Ma, X., Zhang, G., Ni, Y., Chandra, A., Guo, S., Ren, W., Arulraj, A., He, X., Jiang, Z., Li, T., Ku, M., Wang, K., Zhuang, A., Fan, R., Yue, X., and Chen, W. Mmlu-pro: A more robust and challenging multi-task language understanding benchmark. arXiv preprint arXiv:2406.01574, 2024. +[44] Wang, Y., Dai, X., Fan, W., and Ma, Y. Exploring graph tasks with pure llms: A comprehensive benchmark and investigation. arXiv preprint arXiv:2502.18771, 2025. +[45] Wei, J., Wang, X., Schuurmans, D., Bosma, M., Xia, F., Chi, E., Le, Q. V., Zhou, D., et al. Chain-of-thought prompting elicits reasoning in large language models. In NeurIPS, 2022. +[46] Wu, Q., Chen, Z., Corcoran, W., Sra, M., and Singh, A. K. Grapheval2000: Benchmarking and improving large language models on graph datasets. arXiv preprint arXiv:2406.16176, 2024. +[47] Wu, Y., Wang, Y., Du, T., Jegelka, S., and Wang, Y. When more is less: Understanding chain-of-thought length in llms. arXiv preprint arXiv:2502.07266, 2025. +[48] Xu, H., Jian, X., Zhao, X., Pang, W., Zhang, C., Wang, S., Zhang, Q., Monteiro, J., Sun, Q., and Yu, T. Graphomni: A comprehensive and extendable benchmark framework for large language models on graph-theoretic tasks. arXiv preprint arXiv:2504.12764, 2025. +[49] Xu, K., Hu, W., Leskovec, J., and Jegelka, S. How powerful are graph neural networks? arXiv preprint arXiv:1810.00826, 2018. +[50] Xu, K., Hu, W., Leskovec, J., and Jegelka, S. How powerful are graph neural networks? In ICLR, 2019. +[51] Xu, K., Li, J., Zhang, M., Du, S. S., Kawarabayashi, K.-i., and Jegelka, S. What can neural networks reason about? arXiv preprint arXiv:1905.13211, 2019. +[52] Yang, A., Zhang, B., Hui, B., Gao, B., Yu, B., Li, C., Liu, D., Tu, J., Zhou, J., Lin, J., Lu, K., Xue, M., Lin, R., Liu, T., Ren, X., and Zhang, Z. Qwen2.5-math technical report: Toward mathematical expert model via self-improvement. arXiv preprint arXiv:2409.12122, 2024. +[53] Ye, R., Zhang, C., Wang, R., Xu, S., and Zhang, Y. Language is all a graph needs. In ECAL, 2024. +[54] Yuan, Z., Yuan, H., Li, C., Dong, G., Lu, K., Tan, C., Zhou, C., and Zhou, J. Scaling relationship on learning mathematical reasoning with large language models. arXiv preprint arXiv:2308.01825, 2023. + +[55] Yuan, Z., Liu, M., Wang, H., and Qin, B. Gracore: Benchmarking graph comprehension and complex reasoning in large language models. In COLING, 2025. +[56] Zhang, M. and Chen, Y. Link prediction based on graph neural networks. In NeurIPS, 2018. + +# A Training Details + +# A.1 Rejection Sampling + +We randomly extract a subset with 100 examples per task from the training dataset, and use Qwen2.5-32B-Instruct to sample on the subset for $k = 8$ times with a temperature of 1.0. We filter the responses by keeping the reasoning steps that lead to the right answer. If the task is difficult and the filtered responses are insufficient, we resample the subset with a different random seed and repeat the process above. In the end, we obtain around 4,500 training examples ( $\sim 90$ per task) for the SFT phase. + +# A.2 Supervised Fine-tuning + +The detailed training configurations of Naive SFT and RFT are presented in Table 10. + +Table 10: Training configurations of Naive-SFT and RFT. In this table, batch size is abbreviated to BSZ, Max-Length refers to the maximum response length during training and Data Num. reports the number of training examples. + +
SettingLRWeight DecayBSZMax-LengthData Num.Epoch
Naive-SFT1e-5 w/ 1% warm-up1e-26451298.7k1
RFT1e-5 w/ 1% warm-up1e-26430724.4k2
+ +# A.3 Reinforcement Learning + +Configurations for training and evaluation. Our experiments primarily adopt Qwen-2.5-3B/7B-Instruct [32] for their moderate sizes and strong reasoning performance. For GRPO training, we set $\epsilon$ to be 0.02, $\beta$ to be 0.001, group size $G$ to be 5, and context length to be 4096 unless otherwise specified. We additionally incorporate an entropy loss of weight 0.001 to encourage the policy to explore. Lastly, we train the models on 8xA800 GPUs with batch size of 512. During evaluation, we use the vLLM [22] engine for efficient inference. For DeepSeek-R1-Distill-Qwen-7B, we set the maximum token generation length to 4096 tokens except for DeepSeek-R1-Distill-Qwen-7B, which is extended to 30768 for its prolonged thinking process. Sampling is configured with a temperature of 0.6, top-p of 0.95, and top-k of 30. + +The detailed RL training configurations are presented in Table 11. + +Table 11: Training configurations for Naive-SFT and RFT. For abbreviation, we refer the coefficient for entropy loss as Ent. in this table. We report (batch size)/(number of gradient accumulation steps) in the BSZ column, and the temperature for on-policy sampling as $T$ . + +
ModelLRε|G|βγTEnt.BSZMax-LengthData Num.Steps
RL-3B1e-60.251e-31.01.01e-3512/4409698.7k300
SFT-RL-3B1e-60.251e-31.01.01e-3512/4409698.7k300
SFT-RL-Hard-3B1e-60.2165e-41.01.05e-4512/8819249.3k150
SFT-RL-7B1e-60.251e-31.01.01e-3512/8409698.7k300
+ +# B Evaluation Details + +# B.1 Benchmark Introduction + +GraphWiz [3]. GraphWiz employs the Erdős-Rényi (ER) model to generate random graphs and describe graphs in the edge-list formation like $(u, v)$ . The tasks include four linear complexity tasks, Connectivity, Cycle Detection, Bipartite Graph Checking, and Topological Sort; three polynomial complexity tasks, Shortest Path, Maximum Triangle Sum, and Maximum Flow; and two NP-Complete tasks: Hamilton Path and Subgraph Matching. A prompt example is shown in the following: + +# Maximum Triangle Sum Example in GraphWiz + +Find the maximum sum of the weights of three interconnected nodes. In an undirected graph, $[i, k]$ means that node $i$ has the weight $k$ . $(i, j)$ means that node $i$ and node $j$ are connected with an undirected edge. Given a graph, you need to output the maximum sum of the weights of three interconnected nodes. Q: The nodes are numbered from 0 to 4, weights of nodes are: $[0, 8]$ , $[1, 5]$ , $[2, 3]$ , $[3, 6]$ , $[4, 3]$ , and the edges are: $(0, 4)$ , $(0, 3)$ , $(0, 1)$ , $(1, 3)$ , $(1, 2)$ , $(3, 4)$ . What is the maximum sum of the weights of three nodes? + +Node Classification and Link Prediction [44]. We adopt the benchmarks introduced by Wang et al. [44], which are constructed by subsampling from the widely used Cora and PubMed citation graphs. Each instance includes a description of the target node (or node pair) containing the paper ID and title, along with the textual and structural information of neighboring nodes. For node classification, we consider two cases that the description includes the attributes of the target node and those of its 2-hop neighbors, with or without labels. For link prediction, we consider two cases where target nodes are described using their own node attributes along with those of their 2-hop neighbors (excluding the other targeting node), with or without titles. For each task, we randomly sample 2,000 examples per case from the benchmark and report the average performance. A representative example for node classification is shown below: + +# Node Classification Example + +You are a good graph reasoner. Give you a graph language that describes a graph structure and node information from pubmed dataset. You need to understand the graph and the task definition and answer the question. + +## Target node: Paper id: 10695 Title: Haplotype structures and large-scale association testing of the 5' AMP-activated protein kinase genes PRKAA2, PRKAB1, and PRKAB2 [corrected] with type 2 diabetes. + +Known neighbor papers at hop 1 (partial, may be incomplete): + +Paper id: 1155 Title: Computational disease gene identification: a concert of methods prioritizes type 2 diabetes and obesity candidate genes. Label: Type 2 diabetes + +Known neighbor papers at hop 2 (partial, may be incomplete): + +Paper id: 9816 Title: Mitochondrial dysfunction and type 2 diabetes. Label: Type 2 diabetes +Paper id: 1683 Title: A genome-wide search for type II diabetes susceptibility genes in Chinese Hans. Label: Type 2 diabetes + +Paper id: 9916 Title: Genomewide search for type 2 diabetes-susceptibility genes in French whites: evidence for a novel susceptibility locus for early-onset diabetes on chromosome 3q27-qter and independent replication of a type 2-diabetes locus on chromosome 1q21-q24. Paper id: 3793 Title: Association of amino acid variants in the activating transcription factor 6 gene (ATF6) on 1q21-q23 with type 2 diabetes in Pima Indians. Label: Type 2 diabetes Paper id: 4788 Title: Altered glycolytic and oxidative capacities of skeletal muscle contribute to insulin resistance in NIDDM. Label: Type 2 diabetes + +Please predict the most likely type of the Target node. Your answer should be chosen from: Type 1 diabetes Type 2 diabetes Experimentally induced diabetes + +GraphArena [38]. GraphArena samples subgraphs from real-world graphs, including knowledge graphs, social networks, and molecular structures. The tasks include four polynomial-time tasks, Common Neighbor, Shortest Distance, Connected Component, Graph Diameter, and six NP-complete tasks, Maximum Clique Problem (MCP), Maximum Independent Set (MIS), Minimum Vertex Cover (MVC), Maximum Common Subgraph (MCS), Graph Edit Distance (GED), and Traveling Salesman Problem (TSP). Each problem is contextualized within the real-world setting of the graph with an example presented as below: + +# Connected Component Example in GraphArena + +You are required to identify all connected components in the given social network and output one representative node from each component. Within a connected component, any node can be reached from any other node through the edges in the graph. Different connected components are isolated from each other. + +**Problem to Solve** + +- Names in the network: Veronica Garcia, Katherine Brennan, Angel Chavez, Steven Martin, Brett Johnson, Megan Banks, Julia Dominguez, Rachel Mitchell - Friendship connections: Veronica Garcia to Brett Johnson, Veronica Garcia to Megan Banks, Katherine Brennan to Brett Johnson, Katherine Brennan to Megan Banks, Angel Chavez to Megan Banks, Angel Chavez to Rachel Mitchell, Steven Martin to Megan Banks, Brett Johnson to Megan Banks, Megan Banks to Julia Dominguez, Megan Banks to Rachel Mitchell. + +Identify all connected components in this network. Note that for each connected component, you should only output one of its nodes. Present your answer in the following format: [UserID, UserB, UserC, UserD, ...] + +GSM8K [6]. GSM8K is a dataset of 8.5K high quality linguistically diverse grade school math word problems created by human problem writers. We report the accuracies on the 1K test problems and the dataset is downloaded via https://huggingface.co/datasets/openai/gsm8k. + +# Example in GSM8K + +Natalia sold clips to 48 of her friends in April, and then she sold half as many clips in May. How many clips did Natalia sell altogether in April and May? + +MATH500. The dataset contains a subset of 500 problems from the MATH benchmark that OpenAI created in their Let's Verify Step by Step paper [25]. We download the dataset via https://huggingface.co/datasets/HuggingFaceH4/MATH-500. + +# Example in MATH500 + +```txt +Let $z = 2 + \sqrt{2} - (3 + 3\sqrt{2})i$ and let $c = 2 - 3i$ . Let $w$ be the result when $z$ is rotated around $c$ by $\frac{\pi}{4}$ counter-clockwise. +[asy] +unitsize(0.6 cm); +pair C, W, Z; + $Z = (2 + \mathrm{sqrt}(2), -3 - 3^{*}\mathrm{sqrt}(2))$ ; +C = (2,-3); +W = rotate(45,C)*(Z); +draw(Z-C-W); +dot("c", C, N); +dot("w", W, SE); +dot("z", Z, S); +label(" $\frac{\pi}{4}$ ", C + (0.6,-1)); +[/asy] +Find $w$ . +``` + +MMLU-Pro. MMLU-Pro is enhanced version of the Massive Multitask Language Understanding benchmark. It covers a wide range of disciplines, including Math, Law, Engineering, Health, Phycology, etc. We download the dataset via https://huggingface.co/datasets/TIGER-Lab/ MMLU-Pro/viewer/default/test?q=Health&row=5903. + +# Health Example in MMLU-pro + +Question: Food supplements, including trace minerals and vitamins are frequently advertised with promising health benefits. Which of the following substance could be consumed in excess, i.e. well above the recommended daily requirement? + +Options: [ "Vitamin C", "Vitamin D", "Zinc", "Vitamin A"] + +# B.2 Inference Configuration + +For inference, we adopt the vLLM framework [22]. We set the temperature to be 0.06 and the context window to be 4096 for our evaluations unless otherwise specified. + +# B.3 Prompt and Answer Extraction + +To facilitate answer extraction, we adopt the prompt shown in B.3 to guide the models to reason step by step and place their answers within \boxed{}\}. We extract the last \boxed{} shown in the model responses and do necessary format normalizations to retrieve the answer, which includes operations like converting LaTeX-style fraction numbers to float numbers. + +# Problem Instructions + +# {Question Description} + +Approach the problem methodically. Ensure all conclusions are based on precise calculations and logical deductions. Feel free to explore various solution methods and cross-check results for consistency. Maintain dynamic thinking and always verify each step of your reasoning. + +Present the final answer in \boxed{} format, like this: \boxed{\text{ANSWER}}\$, where ANSWER is the final result or expression. + +Think carefully and break down the problem step by step. + +# C Comparison between G1-Zero-7B and G1-7B + +In Section 5.3, we study the role of SFT as a cold-start mechanism for RL by comparing two variants: G1-Zero-3B that is directly trained from the base model Qwen2.5-3B-Instruct with RL, and G1-3B that initializes RL from the CoT-SFT checkpoint. We observe that G1-Zero-3B already achieves surprisingly strong performance, while G1-3B presents clear and consistent improvements across all difficulty levels. Here, we provide additional results for comparing G1-Zero-7B and G1-7B. As shown in Figure 4, for Easy and Medium tasks, the benefit brought by CoT-SFT initialization is marginal, with G1-Zero-7B (96.9%) even surpassing G1-7B (95.1%) on Easy tasks. However, + +![](images/946c8021f91c7ae32199ef184edd341016f6dd131f22a444efbee2fb3d8d2d16.jpg) +Figure 4: Test accuracy comparison of G1-7B and G1-Zero-7B on our benchmark. + +on Hard and Challenging tasks, CoT-SFT as a preliminary step has definite benefits by improving G1-Zero-7B from $13.7\%$ to $23.6\%$ on Challenging tasks. This observation agrees with the case in -3B. Moreover, the average gap between G1-Zero-7B and G1-7B is less than -3B case, indicating G1-7B can possibly be further improved with CoT-SFT generated by a stronger teacher model rather than Qwen2.5-32B-Instruct. We leave this exploration for further work. + +# D Transferability of G1 to Larger Graphs + +To verify the transferability of G1 to larger graphs, we construct a new test set of 5,000 graphs with 36-100 nodes, with other settings kept the same, which ensures there is no overlap between training and test data. Table 12 shows that both G1-3B and G1-7B achieve strong zero-shot generalization to these larger graphs without additional training, significantly outperforming the baselines across + +difficulty levels. These results demonstrate our method's effective scalability beyond the original training distribution. + +Currently, we limit our analysis to smaller graphs because of the context window limit of underlying LLMs (e.g., Qwen2.5-7B-Instruct). The token count scales quadratically with the number of nodes. As shown in Table 13, a 200-node graph often exceeds $32\mathrm{k}$ tokens, surpassing the maximum effective context window of many open-source LLMs [17]. Long-context understanding is actively studied in the LLM literature, and there are some cutting-edge proprietary variants (e.g., OpenAI's GPT-4.1) supporting inputs of over 1M tokens (though graphs of 2000 nodes need 4M). Due to computational constraints (demanding a huge GPU memory), it is hard for us to evaluate on a very long context. We believe our approach can be scaled to larger graphs with the rapid progress of long-context studies. + +Table 12: Zero-shot generalization (accuracy in percentage) of G1 to larger graphs with 36-100 nodes. + +
EasyMediumHardChallengingAverage
Qwen2.5-3B-Instruct27.9828.535.260.2916.74
G1-3B79.3965.6618.463.7444.29
Qwen2.5-7B-Instruct37.8641.569.171.1723.94
G1-7B76.6570.6723.165.2246.46
+ +Table 13: Token numbers of graph with different node sizes. We generate random graphs by Erdős–Rényi model with an edge probability of 0.2. For each node number, we generate 10 graphs and report the mean of the token numbers. For tokenization, we utilize the tokenizer of Qwen2.5-Instruct. + +
Node Number305010020050010002000
Token Number641.41941.57842.8~35k~230k~1M~4M
+ +# E GPT-4o Results on Erdős + +Due to the cost budget, we randomly sample 20 examples per task from Erdős's evaluation set to construct a subset of 1,000 samples. As shown in Table 14, G1-3B performs comparably to GPT-4o on average (57.37% vs. 55.13%), while G1-7B surpasses GPT-4o across all difficulty levels, exceeding it by over 10% on average. This comparison further validates the strong graph reasoning capabilities of the G1 models. + +Table 14: Test accuracy%) of GPT-4o and G1 variants on a subset of Erdős. + +
EasyMediumHardChallengingAverage
GPT-4o-2024-11-2082.5081.8244.0612.1455.13
G1-3B96.4385.4541.885.7157.37
G1-7B96.4388.6452.5023.5765.29
+ +# F Results of G1-Zero-32B + +To demonstrate the scalability of our approach, we extended our training methodology to develop G1-Zero-32B from Qwen2.5-32B-Instruct. The results showcase substantial improvements across all difficulty levels while maintaining computational efficiency. As shown in Table 15, G1-Zero-32B demonstrates remarkable improvements on the Erdős benchmark across all difficulty categories. The model achieves a $27.96\%$ improvement in average accuracy (from $47.10\%$ to $75.06\%$ ), with particularly notable gains in harder categories: +31.87 points on Hard problems and +26.43 points on Challenging problems. These results indicate that our method scales effectively to larger models while maintaining consistent performance improvements across varying problem complexities. + +Furthermore, Table 16 demonstrates that G1-Zero-32B not only preserves but slightly enhances mathematical performance on standard benchmarks. The model shows modest improvements on both + +GSM8K (+0.08 points) and MATH (+4.00 points), confirming that our reasoning-focused training does not compromise existing mathematical capabilities and may even provide synergistic benefits. The training process is completed in 40 hours on $32 \times \mathrm{A}100$ GPUs, demonstrating the practical feasibility of scaling our approach to larger model architectures without prohibitive computational costs. + +Table 15: Test accuracy (%) of G1-Zero-32B and Qwen2.5-32B-Instruct on Erdős. + +
EasyMediumHardChallengingAverage
Qwen2.5-32B-Instruct70.5768.7333.389.0047.10
G1-Zero-32B97.7993.0065.2535.4375.06
+ +Table 16: Test accuracy (%) of G1-Zero-32B and Qwen2.5-32B-Instruct on math tasks. + +
GSM8KMATH
Qwen2.5-32B-Instruct90.6776.80
G1-Zero-32B90.7580.80
+ +# G Experiments for Robustness Verification + +# G.1 Multi-runs Robustness + +To rigorously evaluate robustness, we conducted 32 repeated runs with different random seeds. The results in Table 17 demonstrate consistently small standard deviations ( $<1\%$ across all models and difficulty levels), confirming the stability of our method against potential randomness in LLM outputs. + +Table 17: Test accuracy $(\%)$ for 32 runs with different random seeds. + +
EasyMediumHardChallenging
Qwen2.5-3B-Instruct45.65 ± 0.5130.88 ± 0.3610.36 ± 0.211.54 ± 0.29
G1-3B94.96 ± 0.3283.22 ± 0.2441.48 ± 0.407.96 ± 0.64
Qwen2.5-7B-Instruct57.50 ± 0.1344.92 ± 0.0319.90 ± 0.313.45 ± 0.22
G1-7B95.66 ± 0.1288.89 ± 0.1650.76 ± 0.5324.46 ± 0.84
+ +# G.2 Prompt Robustness + +For prompt robustness, we rigorously test prompt sensitivity by having GPT-4o generate three semantically equivalent prompt variants. Table 18 shows minimal performance variance ( $<1.5\%$ standard deviation) across all models and difficulty levels, confirming our benchmark's stability to phrasing changes. + +# H Analysis of G1's Transferbility to Mathematics Tasks + +We compare the generation results of G1-7B and Qwen2.5-7B-Instruct on the math benchmarks case by case. As a summary, G1-7B improves over Qwen2.5-7B-Instruct on more accurate numerical calculation (Instance 1) and more comprehensive utilization of the given information (Instance 2). It's reasonable since the training of G1 includes complex numerical calculations (e.g., shortest path) and careful processing of questions (e.g., traversing the whole graph descriptions to find key edges). + +In fact, the transferability between different reasoning domains has also been observed in other works and is actively under research. For example, Chu et al. [5] demonstrates that reinforcement learning exhibits strong generalization even under distribution shifts, as it encourages models to learn robust, task-agnostic reasoning strategies rather than overfitting to surface-level patterns. Additionally, RL training has been shown to foster mutual reinforcement across seemingly distinct tasks, e.g., optimizing on puzzles can help reasoning on math [24]. Similarly, due to the similarity between math and graph problems, they would also require similar reasoning skills (such as guessing, backtracking, + +Table 18: Test accuracy $(\%)$ on different prompts, mean and standard deviation reported. + +
EasyMediumHardChallenging
Qwen2.5-3B-Instruct44.26 ± 0.6330.52 ± 1.0810.71 ± 0.411.24 ± 0.07
G1-3B94.86 ± 0.6983.03 ± 0.4241.08 ± 1.359.67 ± 1.05
Qwen2.5-7B-Instruct56.48 ± 1.1756.48 ± 1.1718.81 ± 0.793.10 ± 0.18
G1-7B95.48 ± 0.1888.39 ± 0.6751.27 ± 0.1325.14 ± 1.04
+ +and verification). Therefore, solving graph problems could have positive transferability to math domains. + +# Instance 1 (Calculation Correctness) + +Question: Toula went to the bakery and bought various types of pastries. She bought 3 dozen donuts, which cost 68 dollars per dozen, 2 dozen mini cupcakes, which cost 80 dollars per dozen, and 6 dozen mini cheesecakes for 55 dollars per dozen. How much was the total cost? + +# Ground Truth: 694 + +Answer of Qwen2.5-7B-Instruct: - Calculate the cost of the donuts: Toula bought 3 dozen donuts. The cost per dozen donuts is 68 dollars. Therefore, the total cost for the donuts is: 3 dozen $\times$ 68 dollars/dozen $= 180$ dollars. + +- Calculate the total cost: Add the costs of the donuts, mini cupcakes, and mini cheesecakes: 180 dollars + 160 dollars + 330 dollars = 670 dollars. + +Answer of G1-7B: - Cost of Donuts: Toula bought 3 dozen donuts. The cost per dozen donuts is 68 dollars. - Therefore, the total cost for the donuts is: 3 dozen $\times$ 68 per dozen $= 3 \times 68 = 204$ . + +- Now, we add up the costs of all the pastries to find the total cost: $204 (donuts) + $160 (mini cupcakes) + $330 (mini cheesecakes) = $694. + +# Instance 2 (Utilization of Known Conditions) + +Question: On Mondays, Wednesdays, and Fridays, college student Kimo has three 1-hour classes each day. On Tuesdays and Thursdays, he has two 2-hour classes each day. In one semester, there are 16 weeks of school. In one semester, how many hours does Kimo spend attending classes? + +# Ground Truth: 272 + +Answer of Qwen2.5-7B-Instruct: - Calculate the number of hours Kimo spends in classes each week: On Mondays, Wednesdays, and Fridays, Kimo has 3 classes, each lasting 1 hour. 3 classes $\times$ 1 hour/class $= 3$ hours. On Tuesdays and Thursdays, Kimo has 2 classes, each lasting 2 hours. 2 classes $\times$ 2 hours/class $= 4$ hours. Therefore, the total number of hours Kimo spends in classes each week is: 3 hours + 4 hours = 7 hours. - Calculate the number of hours Kimo spends in classes in one semester: There are 16 weeks in the semester. The total number of hours Kimo spends in classes in one semester is: 7 hours/week $\times$ 16 weeks = 112 hours. + +Answer of G1-7B: - Step 1: Calculate Weekly Class Hours. - Mondays, Wednesdays, and Fridays: Each day, Kimo has 3 classes, each lasting 1 hour. Therefore, for each of these days, the total class time is: 3 classes/day $\times$ 1 hour/class = 3 hours/day. Since there are 3 days (Monday, Wednesday, Friday): 3 hours/day $\times$ 3 days = 9 hours. - Tuesdays and Thursdays: Each day, Kimo has 2 classes, each lasting 2 hours. Therefore, for each of these days, the total class time is: 2 classes/day $\times$ 2 hours/class = 4 hours/day. Since there are 2 days (Tuesday, Thursday) 4 hours/day $\times$ 2 days = 8 hours. - Total Weekly Class Hours: Adding the hours from all the days: 9 hours + 8 hours = 17 hours/week. - Step 2: Calculate Total Class Hours in One Semester. There are 16 weeks in the semester. Therefore, the total class hours in one semester is: 17 hours/week $\times$ 16 weeks = 272 hours. + +# I Detailed Experimental Results + +# I.1 Detailed Results for GraphWiz + +We present the test accuracy for each task in the GraphWiz benchmark in Table 19. G1-7B achieves the highest overall accuracy (57.11%) among all models and reaches the top in 5/7 tasks. It outperforms DeepSeek-R1-Distill-Qwen-7B (51.86%) and even models specifically trained on GraphWiz data such as GraphWiz-7B-RFT (49.61%). Moreover, the smaller variant G1-3B ranks first on all tasks among models of similar parameters, surpassing the base model (Qwen2.5-3B-Instruct) by 13.64% on average and achieves comparable performance with DeepSeek-R1-Distill-Qwen-7B. The results in the GraphWiz benchmark verify the strong zero-shot generalization ability of our G1 models. + +Table 19: Test accuracy $(\%)$ on the GraphWiz benchmark. + +
ModelCycleconnecthippatictopologyshortestfringeflowhamiltonsubgraphAve.
Llama-3.2-3B-Instruct32.0053.7525.757.502.753.752.5038.2512.0019.80
Qwen2.5-3B-Instruct (base)58.0060.5038.504.005.7515.507.5075.0063.2536.44
G1-3B (Ours)91.0064.0064.2513.0014.0023.2543.0096.0042.2550.08
GraphWiz-RFT-7B88.0090.2572.2519.7528.0036.7524.752.5084.2549.61
GraphWiz-DPO-7B86.5082.2571.7515.0026.7537.0045.000.0079.0049.25
Llama-3.1-8B-Instruct64.7581.0058.7511.503.504.259.2519.2545.0033.03
DeepSeek-R1-Distill-Qwen-7B87.0090.0042.7511.0018.2536.0040.0084.7557.0051.86
Qwen2.5-7B-Instruct (base)79.0072.2540.754.2513.5028.7511.5091.2561.0044.69
G1-7B (Ours)92.0080.0075.7524.2521.0029.5046.2595.2550.0057.11
+ +# I.2 Detailed Results for MMLU-Pro + +We present the detailed results for our evaluations on MMLU-Pro in Table 20. We first notice that although G1 models share close accuracies with their base model on average, they excel at notably different disciplines: G1-3B does the best in Physics (56.18%) while G1-7B is good at CS (53.32%). Interestingly, RL training on graph problems in some cases improves G1over Qwen on non-STEM subjects such as Health (53.0% v.s. 37.65%) for 3B models and Business (62.76% v.s. 53.91%) for 7B models. + +Table 20: Test accuracy (%) on the MMLU-Pro benchmark. + +
ModelPhysicsChem.Econ.OtherMathPhil.HistoryBus.Psycho.LawEngl.HealthCSBio.Ave.
Llama-3.2-3B-Instruct7.1814.7915.9113.396.5013.6918.5411.2823.9115.409.8914.0313.259.7113.51
Qwen2.5-3B-Instruct (base)38.4931.1846.2137.3458.9231.0631.2345.2546.2418.0719.4037.6541.2254.2538.54
CoT-SFT-3B35.7013.9932.2538.7253.2934.4125.6530.0418.1642.7128.0839.2236.3446.0334.23
G1-3B (Ours)56.1842.4616.2643.7337.7844.5536.1031.8041.4620.9534.4253.0028.8630.1837.12
Llama-3.1-8B-Instruct28.7917.1333.9634.0332.2841.8324.9118.8043.8946.4535.2836.1031.7528.2632.02
DeepSeek-R1-Distill-Qwen-7B39.7511.7219.2049.8140.8019.9523.3525.6547.3930.3072.7636.8434.5949.5137.21
Qwen2.5-7B-Instruct (base)44.1748.5346.8755.8965.8021.4454.5353.9127.0449.5042.6453.6633.2735.9645.75
CoT-SFT-7B44.3655.5144.6129.8251.0864.8445.9741.4546.4237.0133.8745.6121.4452.0144.54
G1-7B (Ours)46.4351.1968.7640.9447.7053.9032.4062.7625.6149.8851.5051.7153.3236.0748.56
+ +# I.3 Detailed Results for GraphArena + +We report the detailed results for evaluations on the easy/hard problems from GraphArena in Table 21 and Table 22 respectively. We observe that G1 models perform equally or better compared to the other models on all tasks but Distance, in which G1 performs slightly worse than the Qwen models. + +# I.4 Detailed Results for Erdős + +In Table 23, we show the performance for each task in Erdősfor our models and baselines in detail. + +Table 21: Test accuracy (%) on the easy problems from the GraphArena benchmark. + +
ModelConnected DiameterDistanceNeighborGEDTSPMCPMCSMSMVC
Llama-3.2-3B-Instruct8.0016.0015.0050.009.002.0015.0010.007.005.00
Qwen2.5-3B-Instruct (base)20.0011.0047.0048.0037.0017.003.0041.004.002.00
G1-3B (Ours)52.0042.0047.0089.0030.0017.0027.0020.0032.0022.00
LLaMA2-7B-RFT0.007.001.001.004.000.000.001.000.000.00
LLaMA2-7B-DPO0.001.000.000.003.000.000.001.000.000.00
Llama-3.1-8B-Instruct33.0029.0045.0081.0024.0014.0032.0018.0024.0020.00
DeepSeek-R1-Distill-Qwen-7B77.0041.0064.0082.0022.0030.0044.0040.0056.0017.00
Qwen2.5-7B-Instruct (Ours)79.0015.0070.0084.0022.0022.0039.0041.0028.0021.00
G1-7B (Ours)86.0063.0062.0099.0030.0038.0052.0051.0050.0063.00
+ +Table 22: Test accuracy (%) on the hard problems from the GraphArena benchmark. + +
ModelConnected DiameterDistance NeighborGED TSP MCP MCS MJS MYC
Llama-3.2-3B-Instruct0.001.007.0019.003.000.000.000.000.001.00
Qwen2.5-3B-Instruct (base)4.004.0028.0022.007.000.001.000.000.001.00
G1-3B (Ours)19.0012.0025.0051.003.000.000.000.001.007.00
LLaMA2-7B-RFT0.002.000.001.000.000.000.000.000.000.00
LLaMA2-7B-DPO0.003.000.001.001.000.000.000.000.000.00
Llama-3.1-8B-Instruct8.004.0019.0054.003.001.002.000.000.007.00
DeepSeek-R1-Distill-Qwen-7B18.004.0033.0036.001.000.003.000.001.004.00
Qwen2.5-7B-Instruct (base)27.004.0044.0068.002.000.005.000.001.005.00
G1-7B (Ours)31.0027.0035.0084.003.000.003.000.006.0039.00
+ +Table 23: Accuracy comparison of models across tasks + +
TaskGPT-4oo3-miniLlama-3BQwen-3BDSFT-3BCSFT-3BG1-3BLlama-8BQwen-7BMath-7BR1-7BGWiz-RGWiz-DDSFT-7BCSFT-7BG1-7BLlama-70BQwen-72B
node_number100.00100.0083.0094.00100.0097.00100.0099.0099.0094.00100.000.000.00100.00100.00100.00100.00100.00
dominating_set29.4164.7157.0023.0072.0031.0099.0037.0027.0021.0027.0034.0028.0074.0068.0099.0024.0044.00
common_neighbor73.6873.6823.0044.0071.0071.0091.0056.0052.0048.0079.000.000.0076.0080.0093.0091.5289.99
edge_number72.2277.789.0031.0031.0059.0096.0016.0058.0038.0074.000.000.0039.0034.0097.0072.0066.00
neighbor84.2163.1626.0036.0087.0082.0091.0042.0065.0064.0094.004.002.0089.0089.0093.0093.0598.53
bfs52.1717.390.003.0052.0030.0095.005.0012.009.0012.000.000.0043.0044.0098.0025.0053.00
has_cycle80.00100.0051.0051.0098.0063.0089.0046.0055.0054.0083.0065.0083.0095.0093.0098.0064.0054.00
dfs73.3333.330.009.0061.0043.00100.0012.0027.0010.0023.000.000.0052.0050.0099.0029.0049.00
minimumspanning_tree38.1042.865.008.0062.0017.0081.0017.0015.0014.0046.000.000.0065.0065.0066.0028.0039.00
edgeExistence100.00100.0060.0080.00100.0097.00100.0073.0096.0082.00100.0052.0056.0098.0097.00100.0098.00100.00
is_normal100.0095.0088.0095.0098.0098.00100.0092.0096.0099.0098.0027.0058.0099.0099.00100.0099.00100.00
degree95.45100.0026.0058.0094.0093.0095.0072.0077.0079.0096.000.000.0088.0082.0099.0094.0082.00
is_tournament100.0088.8947.0075.0099.0099.0099.0080.0086.0087.00100.0022.0058.00100.00100.0099.0099.0094.00
density68.1890.9136.0033.0017.0038.0092.0042.0038.0040.0081.000.000.0012.0015.0097.0051.0051.00
adamic_adar_index92.3188.461.006.0074.0089.0094.0012.0039.0022.0082.003.001.0076.0075.0098.0052.0064.00
clustering_coefficient72.2294.4413.0031.0071.0056.0082.0025.0044.0036.0065.006.0010.0067.0066.0088.0049.0069.00
connected_component_number60.8782.619.0027.0085.0063.0079.0034.0035.0030.0079.000.000.0080.0081.0092.0064.0066.00
bipartite_maximum_MATCHing40.7448.153.0019.0053.0047.0082.0013.0012.003.0042.000.000.0076.0073.0087.0029.0037.00
local_connectivity96.15100.0057.0062.0093.0086.0090.0053.0074.0079.0082.0053.0066.0097.0098.0096.0077.0069.00
jaccard_coefficient100.00100.0023.0048.0081.0084.0095.0044.0077.0070.0095.003.005.0078.0076.00100.0087.0093.00
min_edgecovering10.5331.581.002.0023.0017.0051.000.001.001.0037.000.000.0018.0017.0050.0016.0015.00
is_ularian86.3695.4578.0081.0095.0089.0098.0082.0081.0089.0092.0033.0059.0093.0093.0097.0090.0080.00
degree_centrality71.4385.710.007.0081.0079.0089.004.008.0023.0087.000.000.0080.0084.0097.0049.0048.00
is_bipartite68.0092.0049.0039.0092.0055.0079.0053.0052.0043.0076.0051.0067.0093.0090.0080.0062.0067.00
resource.setLocation_index94.12100.002.0010.0080.0079.0092.0015.0045.0040.0086.002.002.0077.0080.0092.0036.0078.00
max_weight_MATCHing11.1127.782.003.0025.0022.0024.007.0012.002.0040.000.000.0025.0025.0043.0024.0026.00
closeness_centrality0.0031.580.001.008.006.009.004.003.003.0014.000.000.004.005.0011.0013.0011.00
traveling_salesmanProblem36.8489.478.0024.0029.0040.0043.0017.0041.0041.0062.003.001.0025.0020.0051.0047.0043.00
stronglyConnected_number13.3373.334.005.0063.0024.0058.003.0011.007.0035.000.000.0055.0056.0059.009.0010.00
shortest_path69.2338.4611.0019.0074.0051.0062.0031.0035.0011.0062.003.000.0077.0078.0070.0062.0060.00
center19.0566.674.008.0025.0013.0025.006.008.009.0026.000.000.0024.0025.0035.0029.7241.48
diameter17.6594.1212.008.0055.0031.0046.0014.0031.0027.0039.003.004.0041.0039.0049.005.000.00
barycenter7.6969.239.0015.0056.0026.0039.0020.0022.0011.1129.001.011.0149.0050.0047.0053.7147.61
radius68.7587.5012.0023.0066.0047.0056.0026.0034.0035.0052.001.002.0063.0058.0068.005.001.00
topological_sort60.0048.0010.0014.0076.0038.0067.0025.0025.0021.0064.005.0074.0071.0078.0073.0074.00
periphery29.4158.821.003.0033.0016.0022.001.0011.006.0025.000.000.0027.0029.0031.0050.0647.78
betweenness_centrality18.1850.004.004.0038.0030.0039.0024.001.005.006.001.002.0038.0037.0039.007.004.00
triangles35.2958.8213.004.0054.0048.0067.0012.0030.0021.0054.000.000.0042.0040.0079.0043.0055.00
avg_neighbour-degree66.6761.1116.0017.0036.0055.0068.0026.0036.0030.0029.0058.003.0061.0031.0036.0082.0064.00
harmonic_centrality7.6984.622.003.0017.0015.0019.003.005.008.0037.001.002.009.007.0030.007.0022.00
bridges0.009.091.000.0044.409.0016.001.003.003.001.005.000.0042.0042.0023.0028.5729.92
isomophic Mapping0.004.000.000.0010.001.001.001.001.001.001.001.1111.1111.1111.1112.1112.111.11
global_efficiency4.7671.430.001.0010.003.001.001.001.1111.1111.1111.1111.1111.1111.1112.1112.111.11
maximal_independent_set5.5655.562.001.0026.002.0013.002.002.1131.1131.1111.1111.1119.2224.2279.227.113.11
maximum_flow0.0080.955.002.008.0010.007.003.686.681.1212.123.355.499.727.924.124.12
wiener_index0.0073.680.001.0014.006.688.884.484.4822.2222.226.686.6831.1136.2282.227.117.11
harmonic_path0.004.760.001.113.353.352.68
minVertexCover13.0460.871.003.3522.25
+ +# J Discussion on Reward Weighting + +In Section 5.3, we analyze the factor of data mixture by introducing a model G1-Hard-3B trained exclusively on Hard and Challenging tasks. We observe that G1-Hard-3B effectively improves performance on hard tasks, while on easier tasks still lags behind G1-3B (Table 24). + +In this section, we further explore a soft data mixture strategy that scales the reward for each task according to its difficulty. In detail, we fix the scaling factor $s$ as 0.2, 0.4, 0.6, and 0.8 for Easy, Medium, Hard and Challenging tasks, respectively, and name the resulting model as G1-Soft-3B. As shown in Table 24, G1-Soft-3B achieves a balance between G1-3B and G1-Hard-3B. On easy tasks, G1-Soft-3B largely surpasses G1-Hard-3B and is on par with G1-3B which applies uniform scaling across all tasks. For hard tasks, G1-Soft-3B outperforms G1-3B (e.g., 11.71% v.s 7.57% for Challenging tasks), but there is still a gap to G1-Hard-3B. The results show the soft scaling method take effects, but the RL optimization remains dominated by easy tasks. This suggests that further reducing the reward scaling factor for easy tasks or a dynamic weighting strategy could be beneficial—a direction we leave for future work. + +Table 24: Test accuracy (%) on our benchmark. $\star$ denotes the tasks are excluded in model training. G1-Hard-3B is only RL-trained on Hard and Challenging tasks. G1-Soft-3B is trained on all tasks but with different reward scaling factors based on the task difficulty. + +
CategoryModelEasyMediumHardChallengingAverage
Base ModelQwen2.5-3B-Instruct45.7130.189.441.2922.72
OursDirect-SFT-3B74.4375.2743.6914.4353.78
G1-3B94.8684.6441.257.5759.76
G1-Hard-3B69.36*70.64*48.5017.4353.30
G1-Soft-3B96.0783.5540.8811.7160.38
+ +# K Detailed Description of Erdős + +# K.1 Comparing Erdős with Other Graph Reasoning Benchmarks for LLMs + +There is a growing interest in evaluating LLMs' graph reasoning abilities. NLGraph [41] evaluate LLMs on graph-theoretic tasks and discover preliminary yet brittle reasoning abilities in the face of spurious correlations and large graphs. Later, GraphArena [38] and GraCoRe [55] include a broader task coverage and recently released LLMs, finding that even OpenAI o1-mini struggles a lot with complex tasks. Moreover, GraphEval2000 [46] and ProGraph [23] emphasize code-oriented problem solving using library-based prompts, and GraphOmni [48] unify varying graph types, encodings, and prompt styles for a comprehensive evaluation. Overall, these benchmarks suggest that LLMs overall demonstrate moderate success on simple tasks but struggle with abstraction, generalization, and larger or more complex graph instances. Nevertheless, these datasets are either too small (e.g., thousands of examples) or not diverse enough (e.g., 8 tasks in NLGraph) for training general-purpose graph reasoners, which motivates the design of Erdős. We show the detailed comparison of existing graph reasoning benchmarks for LLM with our Erdős sin Table 25. + +Table 25: Comparison of existing graph-theoretic reasoning benchmarks for LLM with our Erdős. + +
Benchmark#Tasks# Q-A SamplesGraph TypesNode Size
NLGraph [41]85,902Synthetic5 to 35
GraphWiz [3]93,600Synthetic2 to 100
GraphArena [38]1010,000Real-world4 to 50
GraCoRe [55]195,140Synthetic & Real-world8 to 30
GraphOmni [48]6241,726Synthetic5 to 30
Erdős(ours)50100,000Real-world5 to 35
+ +# K.2 Full list of tasks in Erdős + +Table 26: Benchmark exemples + +
TaskPromptAnswer
adamic adar indexThe task is to determine the Adamic-Adar index of two nodes in a graph. +The Adamic-Adar index is the sum of the inverse logarithm of the degrees of the common neighbors of the two nodes. +The input graph is guaranteed to be undirected. +Here is an undirected graph containing nodes from 1 to 9. The edges are: (1, 5), (1, 4), (1, 8), (1, 2), (1, 3), (1, 7), (5, 2), (5, 3), (5, 4), (5, 9), (5, 6), (4, 8), (4, 9), (4, 7), (8, 2), (8, 3), (8, 6), (8, 7), (8, 9), (2, 3), (2, 7), (2, 6), (3, 9), (3, 7), (7, 6), (7, 9). +Question: What is the Adamic-Adar index between node 4 and node 6? +You need to format your answer as a float number.1.5859
avg neighbor degreeThe task is to determine the average degree of the neighbors of a node in the graph. +Here is an undirected graph containing nodes from 1 to 8. The edges are: (1, 7), (1, 8), (1, 4), (7, 8), (8, 5), (2, 3), (2, 6), (3, 5). +Question: What is the average neighbor degree of node 2 in the graph? +You need to format your answer as a float number.1.5
barycenterThe task is to determine the barycenter of a graph. +The barycenter of a graph is also called the median. It includes the node that minimizes the sum of shortest path lengths to all other nodes. +The input graph is guaranteed to be connected. +Here is an undirected graph containing nodes from 1 to 7. The edges are: (1, 2), (1, 6), (1, 5), (1, 7), (1, 4), (2, 6), (2, 5), (2, 7), (2, 4), (6, 4), (6, 5), (6, 7), (7, 3), (7, 4). +Question: What is the barycenter of the graph? +You need to format your answer as a list of nodes in ascending order, e.g., [node-1, node-2, ..., node-n].[1, 2, 6, 7]
betweenness centralityThe task is to determine the betweenness centrality of a node in the graph. +Betweenness centrality of a node *u* is the sum of the fraction of all-pairs shortest paths that pass through *u*. +Here is an undirected graph containing nodes from 1 to 9. The edges are: (1, 6), (1, 4), (1, 8), (1, 9), (6, 2), (6, 7), (4, 7), (4, 5), (8, 3), (8, 5), (8, 7), (9, 3), (9, 5), (2, 7). +Question: What is the betweenness centrality of node 5 in the graph? +You need to format your answer as a float number.0.0679
bfsThe task is to determine the breadth-first search (BFS) traversal order given a starting node. +Stop when the BFS cannot be continued. +Here is an undirected graph containing nodes from 1 to 7. The edges are: (1, 2), (1, 5), (2, 3), (2, 4), (5, 3), (5, 4), (3, 4), (4, 7), (7, 6). +Question: What is the breadth-first search (BFS) traversal order for the starting node 1? +You need to format your answer as a list of edges, e.g., [(u1, v1), (u2, v2), ..., (un, vn)].[(1, 2), (1, 5), (2, 3), (2, 4), (4, 7), (7, 6)]
+ +Continuing table 26 + +
TaskPromptAnswer
bipartite maximum matchingThe task is to determine the maximal matching in a bipartite graph.The input graph is guaranteed to be a bipartite graph.Here is an undirected graph containing nodes from 1 to 4. The edges are: (1, 3), (1, 4), (2, 3), (2, 4).Question: What is the bipartite maximal matching of the bipartite graph?You need to format your answer as a list of edges in ascending dictionary order, e.g., [(u1, v1), (u2, v2), ..., (un, vn)].[(1, 3), (2, 4)]
bridgesThe task is to find all bridges of a graph.A bridge is an edge in a graph whose removal increases the number of connected components.The input graph is guaranteed to be undirected.Here is an undirected graph containing nodes from 1 to 5. The edges are: (1, 2), (1, 3), (1, 4), (2, 3), (2, 4), (2, 5), (3, 4), (3, 5).Question: What are the bridges of the graph?You need to format your answer as a list of edges in ascending dictionary order, e.g., [(u1, v1), (u2, v2), ..., (un, vn)].[]
centerThe task is to determine the center of a graph.The center of a graph includes the node that minimizes the maximum distance to any other nodes in the graph.The input graph is guaranteed to be connected.Here is an undirected graph containing nodes from 1 to 6. The edges are: (1, 5), (5, 2), (2, 6), (6, 4), (3, 4).Question: What is the center of the graph?You need to format your answer as a list of nodes in ascending order, e.g., [node-1, node-2, ..., node-n].[2, 6]
closeness centralityThe task is to determine the closeness centrality of a node in the graph.For a node *u*, closeness centrality is the reciprocal of the average shortest path distance to *u* over all *n-1* reachable nodes. For directed graphs, it computes the incoming distance to *u*.Here is an undirected graph containing nodes from 1 to 8. The edges are: (1, 3), (3, 6), (2, 8), (2, 6), (8, 6), (8, 7), (4, 7), (7, 5).Question: What is the closeness centrality of node 2 in the graph?You need to format your answer as a float number.0.4667
clustering coefficientThe task is to compute the clustering coefficient for a given node.For unweighted graphs, the clustering of a node is the fraction of possible triangles through that node that exist.Here is an undirected graph containing nodes from 1 to 7. The edges are: (1, 4), (1, 5), (1, 3), (4, 2), (4, 3), (4, 5), (4, 6), (4, 7), (5, 2), (5, 3), (5, 6), (5, 7), (2, 6), (2, 7), (6, 7).Question: What is the clustering coefficient of node 6?You need to format your answer as a float number.1.0
common neighborThe task is to determine common neighbors between two nodes in the graph.The input graph is guaranteed to be undirected.Here is an undirected graph containing nodes from 1 to 7. The edges are: (1, 7), (1, 6), (1, 4), (1, 5), (7, 2), (7, 3), (6, 2), (4, 3), (5, 3).Question: What are the common neighbors between node 2 and node 3?You need to format your answer as a list of nodes in ascending order, e.g., [node-1, node-2, ..., node-n].[7]
+ +Continuing table 26 + +
TaskPromptAnswer
connected component numberThe task is to determine the number of connected components in an undirected graph. +A connected component is a subgraph where any two nodes are connected to each other by paths. +Here is an undirected graph containing nodes from 1 to 10. The edges are: (1, 4), (1, 7), (1, 5), (1, 9), (1, 10), (1, 6), (1, 2), (4, 2), (4, 3), (4, 8), (4, 5), (4, 9), (4, 10), (7, 2), (7, 3), (7, 5), (7, 6), (7, 8), (7, 9), (5, 2), (5, 3), (5, 8), (5, 9), (5, 10), (9, 2), (9, 3), (9, 6), (9, 8), (9, 10), (10, 2), (10, 3), (10, 8), (6, 2), (6, 3), (6, 8), (2, 8), (2, 3). +Question: How many connected components are there in the graph? +Your answer should be an integer.1
degreeThe task is to determine the degree of a node in the graph. +For the undirected graph, you should count the edge between two nodes only once. +Here is an undirected graph containing nodes from 1 to 6. The edges are: (1, 6), (6, 5), (2, 3), (2, 4), (3, 5). +Question: What is the degree of node 6 in the graph? +Your answer should be an integer.2
degree centralityThe task is to determine the degree centrality of a node in the graph. +Degree centrality for a node is the fraction of nodes it is connected to. +Here is an undirected graph containing nodes from 1 to 7. The edges are: (1, 2), (1, 4), (1, 5), (2, 3), (2, 4), (2, 5), (2, 6), (4, 3), (4, 5), (4, 7), (5, 3). +Question: What is the degree centrality of node 3 in the graph? +You need to format your answer as a float number.0.5
densityThe task is to determine the density of the graph. +Density is defined as the ratio of the number of edges in the graph to the number of possible edges. +Here is an undirected graph containing nodes from 1 to 5. The edges are: (1, 2), (1, 3), (2, 3), (2, 4), (2, 5), (3, 4), (4, 5). +Question: What is the density of the graph? +You need to format your answer as a float number.0.7
dfsThe task is to determine the depth-first search (DFS) traversal order given a starting node. +Stop when the DFS cannot be continued. +Here is an undirected graph containing nodes from 1 to 9. The edges are: (1, 2), (1, 3), (1, 6), (3, 9), (4, 8), (4, 5), (8, 7). +Question: What is the depth-first search (DFS) traversal order for the starting node 1? +You need to format your answer as a list of edges, e.g., [(u1, v1), (u2, v2), ..., (un, vn)].[(1, 2), (1, 3), (3, 9), (1, 6)]
diameterThe task is to determine the diameter of a graph. +The diameter of a graph is the longest shortest path between any two nodes in the graph. +The input graph is guaranteed to be connected. +Here is an undirected graph containing nodes from 1 to 7. The edges are: (1, 5), (1, 7), (1, 4), (5, 6), (2, 6), (2, 3). +Question: What is the diameter of the graph? +You need to format your answer as a float number.5
+ +Continuing table 26 + +
TaskPromptAnswer
dominating setThe task is to determine the dominating set of a graph. +A dominating set is a subset of nodes such that every node in the graph is either in the set or adjacent to a node in the set. +For directed graphs, any node not in the dominating set must be a successor of a node within the set. +Here is an undirected graph containing nodes from 1 to 7. The edges are: (1, 2), (1, 5), (1, 6), (1, 7), (2, 3), (2, 4), (5, 6), (7, 3), (7, 4). +Question: What is the dominating set of the graph? +You need to format your answer as a list of nodes in ascending order, e.g., [node-1, node-2, ..., node-n].[1, 3, 4]
edge existenceThe task is to determine if there is an edge connecting two nodes. +For an undirected graph, determine if there is an edge between nodes *u* and *v*. For a directed graph, determine if there is an edge from *u* to *v*. +Here is an undirected graph containing nodes from 1 to 8. The edges are: (1, 2), (1, 6), (3, 8), (3, 4), (8, 4), (8, 5), (8, 7), (4, 7), (4, 5), (7, 5). +Question: Is there an edge between node 5 and node 3? +Your answer should be Yes or No.No
edge numberThe task is to determine the number of edges in the graph. +For the undirected graph, you should count the edge between two nodes only once. +Here is an undirected graph containing nodes from 1 to 10. The edges are: (1, 10), (1, 8), (10, 7), (8, 6), (2, 5), (2, 4), (2, 6), (5, 4), (5, 9), (4, 3), (4, 9), (3, 7). +Question: How many edges are there in the graph? +Your answer should be an integer.12
global efficiencyThe task is to determine the global efficiency of a graph. +Global efficiency is the average efficiency of all pairs of nodes. +The efficiency of a pair of nodes is the multiplicative inverse of the shortest path distance between the nodes. +The input graph is guaranteed to be undirected. +Here is an undirected graph containing nodes from 1 to 7. The edges are: (1, 5), (1, 4), (5, 2), (2, 7), (7, 3), (3, 6). +Question: What is the global efficiency of the graph? +You need to format your answer as a float number.0.5310
hamiltonian pathThe task is to return a Hamiltonian path in a directed graph. +A Hamiltonian path is a path in a directed graph that visits each vertex exactly once. +The input graph is guaranteed to be directed and tournable. +Here is a directed graph containing nodes from 1 to 8. The edges are: (2, 1), (2, 4), (2, 5), (2, 6), (2, 7), (1, 3), (1, 4), (1, 7), (3, 2), (3, 7), (3, 8), (4, 3), (4, 5), (4, 7), (5, 1), (5, 3), (5, 8), (6, 1), (6, 3), (6, 4), (6, 5), (7, 5), (7, 6), (8, 1), (8, 2), (8, 4), (8, 6), (8, 7). +Question: Return a Hamiltonian path in the graph. +You need to format your answer as a list of nodes, e.g., [node-1, node-2, ..., node-n].[2, 1, 4, 5, 3, 8, 7, 6]
+ +Continuing table 26 + +
TaskPromptAnswer
harmonic centralityThe task is to determine the harmonic centrality of a node in the graph.Harmonic centrality of a node *u* is the sum of the reciprocal of the shortest path distances from all other nodes to u.Here is a directed graph containing nodes from 1 to 8. The edges are: (6, 2), (6, 1), (6, 4), (6, 5), (6, 3), (7, 8).Question: What is the harmonic centrality of node 3 in the graph?You need to format your answer as a float number.1.0
has cycleThe task is to determine if the graph has a cycle.Here is an undirected graph containing nodes from 1 to 9. The edges are: (1, 2), (1, 4), (1, 5), (2, 4), (2, 5), (4, 9), (5, 3), (3, 6), (3, 8), (6, 8), (9, 7).Question: Does the graph have a cycle?Your answer should be Yes or No.Yes
is bipartiteThe task is to determine if the graph is bipartite.A bipartite graph is a graph whose nodes can be divided into two disjoint sets such that no two graph vertices within the same set are adjacent.Here is an undirected graph containing nodes from 1 to 6. The edges are: (1, 4), (4, 3), (2, 5), (2, 3), (5, 6), (3, 6).Question: Is the graph bipartite?Your answer should be Yes or No.Yes
is eularianThe task is to determine if the graph is Eulerian.An Eulerian graph is a graph that contains an Eulerian circuit, which is a cycle that visits every edge exactly once.Here is an undirected graph containing nodes from 1 to 6. The edges are: (1, 5), (1, 3), (1, 2), (1, 4), (5, 2), (3, 2), (3, 4), (3, 6), (2, 4), (4, 6).Question: Is the graph Eulerian?Your answer should be Yes or No.Yes
is regularThe task is to determine if the graph is regular.A regular graph is a graph where every node has the same degree.Here is an undirected graph containing nodes from 1 to 10. The edges are: (1, 5), (1, 7), (1, 10), (5, 2), (5, 10), (7, 8), (7, 10), (3, 9), (3, 8), (3, 4), (9, 4), (4, 6).Question: Is the graph regular?Your answer should be Yes or No.No
is tournamentThe task is to determine if the graph is a tournament.A tournament is a directed graph where every pair of nodes is connected by a single directed edge.The input graph is guaranteed to be directed.Here is a directed graph containing nodes from 1 to 10. The edges are: (1, 2), (2, 1), (2, 4), (4, 2), (4, 3), (3, 1), (5, 2), (5, 4), (6, 2), (6, 5), (7, 8), (8, 6), (9, 7), (10, 7).Question: Is the graph a tournament?Your answer should be Yes or No.No
+ +Continuing table 26 + +
TaskPromptAnswer
isomorphic mappingGiven a pair of isomorphic graphs, determine the node correspondence between the two graphs.The first graph is: G describes an undirected graph among 0, 1, 2, 3, 4, 5, and 6. In this graph: Node 0 is connected to nodes 6, 3, 4. Node 1 is connected to nodes 4, 5, 6. Node 2 is connected to nodes 3, 4. Node 3 is connected to nodes 0, 2, 5. Node 4 is connected to nodes 0, 1, 2. Node 5 is connected to nodes 1, 3. Node 6 is connected to nodes 0, 1.The second graph is: G describes an undirected graph among 102, 106, 105, 101, 103, 100, and 104. In this graph: Node 100 is connected to nodes 106, 101. Node 101 is connected to nodes 102, 105, 100. Node 102 is connected to nodes 104, 101, 103. Node 103 is connected to nodes 102, 106, 105. Node 104 is connected to nodes 102, 106. Node 105 is connected to nodes 101, 103. Node 106 is connected to nodes 103, 100, 104.Provide a node matching dictionary such as {Graph1 #Node1: Graph2 #Node1, Graph1 #Node2: Graph2 #Node2,...}{0: 102, 3: 101, 2: 105, 4: 103, 1: 106, 5: 100, 6: 104}
jaccard coefficientThe task is to determine the Jaccard coefficient of two nodes in a graph.The Jaccard coefficient is the size of the intersection divided by the size of the union of the neighbors of the two nodes.The input graph is guaranteed to be undirected.Here is an undirected graph containing nodes from 1 to 5. The edges are: (1, 2), (1, 3), (2, 5), (2, 3), (3, 5), (5, 4).Question: What is the Jaccard coefficient between node 2 and node 4?You need to format your answer as a float number.0.3333
local connectivityThe task is to determine the local connectivity of two nodes in the graph.Local connectivity is whether there exists at least one path between the two nodes.Here is a directed graph containing nodes from 1 to 7. The edges are: (1, 7), (7, 6), (3, 1), (4, 3), (5, 4), (6, 2).Question: What is the local connectivity between node 7 and node 4 in the graph?Your answer should be Yes or No.No
max weight matchingThe task is to determine the maximum weight matching of a graph.A matching is a set of edges without common vertices. A maximal matching cannot add more edges and still be a matching.The weight of a matching is the sum of the weights of its edges.If not specified, all edges have equal edge weights.The input graph is guaranteed to be undirected.Here is an undirected graph containing nodes from 1 to 7. The edges are: (1, 7), (7, 5), (2, 4), (2, 5), (4, 3), (3, 6).Question: What is the maximum weight matching of the graph?You need to format your answer as a list of edges in ascending dictionary order, e.g., [(u1, v1), (u2, v2), ..., (un, vn)].[(2, 4), (5, 7), (6, 3)]
+ +Continuing table 26 + +
TaskPromptAnswer
maximal independent setThe task is to determine the maximal independent set guaranteed to contain a given node in the graph. An independent set is a set of nodes such that the subgraph induced by these nodes contains no edges. A maximal independent set is an independent set such that it is not possible to add a new node and still get an independent set. The input graph is guaranteed to be undirected. Here is an undirected graph containing nodes from 1 to 6. The edges are: (1, 2), (1, 6), (1, 3), (2, 3), (2, 4), (2, 5), (3, 5), (4, 5). Question: What is the maximal independent set that includes node 4 of the graph? You need to format your answer as a list of nodes in ascending order, e.g., [node-1, node-2, ..., node-n].[3, 4, 6]
maximum flowThe task is to determine the value of the maximum flow for the given source node and sink node. The maximum flow is the greatest amount of flow that can be sent from the source to the sink without violating capacity constraints. Here is a directed graph containing nodes from 1 to 5. The edges are: (2, 5, 8), (3, 1, 9), (3, 5, 3), (4, 2, 4). (u, v, w) denotes the edge from node *u* to node *v* has a capacity of *w*. Question: What is the value of the maximum flow from node 3 to node 2? You need to format your answer as a float number.0.0
min edge coveringThe task is to determine the minimum edge covering of a graph. An edge cover is a set of edges such that every vertex in the graph is incident to at least one edge in the set. The minimum edge cover is the edge cover with the smallest number of edges. The input graph is guaranteed to be undirected. Here is an undirected graph containing nodes from 1 to 9. The edges are: (1, 2), (1, 3), (1, 4), (2, 3), (2, 4), (2, 5), (3, 4), (3, 6), (3, 7), (3, 8), (3, 5), (4, 7), (4, 8), (5, 6), (5, 7), (6, 7), (6, 9), (7, 9). Question: What is the minimum edge covering of the graph? You need to format your answer as a list of edges in ascending dictionary order, e.g., [(u1, v1), (u2, v2), ..., (un, vn)].[(2, 1), (5, 2), (7, 4), (8, 3), (9, 6)]
min vertex coverThe task is to determine the minimum vertex cover of a graph. A vertex cover is a set of nodes such that every edge in the graph is incident to at least one node in the set. Here is an undirected graph containing nodes from 1 to 5. The edges are: (1, 2), (2, 3), (3, 5), (5, 4). Question: What is the minimum vertex cover of the graph? You need to format your answer as a list of nodes in ascending order, e.g., [node-1, node-2, ..., node-n].[2, 5]
minimum spanning treeThe task is to determine the minimum spanning tree of a graph. A minimum spanning tree is a subset of the edges that connects all vertices in the graph with the minimum possible total edge weight. If not specified, all edges have equal edge weights. The input graph is guaranteed to be undirected and connected. Here is an undirected graph containing nodes from 1 to 9. The edges are: (1, 2), (1, 8), (1, 5), (1, 6), (1, 4), (1, 7), (1, 9), (2, 5), (2, 6), (2, 4), (2, 7), (2, 3), (8, 3), (8, 4), (8, 6), (8, 7), (5, 3), (5, 4), (5, 6), (5, 7), (5, 9), (6, 3), (6, 4), (6, 7), (6, 9), (4, 3), (4, 7), (4, 9), (7, 9), (9, 3). Question: What is the minimum spanning tree of the graph? You need to format your answer as a list of edges in ascending dictionary order, e.g., [(u1, v1), (u2, v2), ..., (un, vn)].[(1, 2), (1, 4), (1, 5), (1, 6), (1, 7), (1, 8), (1, 9), (2, 3)]
+ +Continuing table 26 + +
TaskPromptAnswer
neighborThe task is to determine the neighbors of a node in the graph. For directed graph, you should return the successors of the node. Here is an undirected graph containing nodes from 1 to 10. The edges are: (1, 3), (1, 9), (1, 6), (1, 7), (3, 2), (3, 8), (3, 9), (6, 7), (2, 10), (10, 8), (4, 5). Question: What are the neighbors of node 2 in the graph? You need to format your answer as a list of nodes in ascending order, e.g., [node-1, node-2, ..., node-n].[3, 10]
node numberThe task is to determine the number of nodes in the graph. Here is an undirected graph containing nodes from 1 to 10. The edges are: (1, 10), (1, 3), (10, 6), (10, 8), (3, 7), (3, 4), (2, 7), (2, 5), (2, 9), (5, 9), (5, 8), (9, 4), (8, 6). Question: How many nodes are there in the graph? Your answer should be an integer.10
peripheryThe task is to determine the periphery of a graph. The periphery of a graph is the set of nodes with the maximum eccentricity. The eccentricity of a node is the maximum distance from this node to all other nodes in the graph. The input graph is guaranteed to be connected. Here is an undirected graph containing nodes from 1 to 6. The edges are: (1, 3), (3, 2), (3, 4), (3, 5), (3, 6). Question: What is the periphery of the graph? You need to format your answer as a list of nodes in ascending order, e.g., [node-1, node-2, ..., node-n].[1, 2, 4, 5, 6]
radiusThe task is to determine the radius of a graph. The radius of a graph is the minimum eccentricity of any node in the graph. The eccentricity of a node is the maximum distance from this node to all other nodes in the graph. The input graph is guaranteed to be connected. Here is an undirected graph containing nodes from 1 to 5. The edges are: (1, 2), (2, 3), (3, 4), (3, 5), (4, 5). Question: What is the radius of the graph? You need to format your answer as a float number.2
resource allocation indexThe task is to determine the resource allocation index of two nodes in a graph. The resource allocation index of two nodes is the sum of the inverse of the degrees of the common neighbors of the two nodes. The input graph is guaranteed to be undirected. Here is an undirected graph containing nodes from 1 to 5. The edges are: (1, 2), (1, 3), (2, 3), (3, 4), (3, 5), (4, 5). Question: What is the resource allocation index between node 1 and node 4? You need to format your answer as a float number.0.25
shortest pathThe task is to determine the shortest path between two nodes. The input nodes are guaranteed to be connected. Here is an undirected graph containing nodes from 1 to 6. The edges are: (1, 2), (1, 3), (2, 4), (2, 3), (2, 5), (3, 4), (3, 5), (4, 6). Question: What is the shortest path between node 1 and node 6? You need to format your answer as a list of nodes, e.g., [node-1, node-2, ..., node-n].[1, 2, 4, 6]
+ +Continuing table 26 + +
TaskPromptAnswer
strongly connected numberThe task is to determine the number of strongly connected components in a directed graph.A strongly connected component is a maximal subgraph where every node is reachable from every other node.Here is a directed graph containing nodes from 1 to 6. The edges are: (2,5), (5,1), (3,4), (6,2).Question: How many strongly connected components are there in the graph?Your answer should be an integer.6
topological sortThe task is to determine the topological sort of a directed acyclic graph (DAG).Here is a directed graph containing nodes from 1 to 6. The edges are: (1,6), (1,5), (1,4), (1,3), (1,2).Question: What is the topological sort of the directed acyclic graph (DAG)?You need to format your answer as a list of nodes, e.g., [node-1, node-2,..., node-n].[1,6,5,4,3,2]
traveling salesman problemThe task is to determine the minimal cost of the Traveling Salesman Problem (TSP).The Traveling Salesman Problem asks for the shortest possible route that visits each vertex exactly once and returns to the starting vertex.The input graph is guaranteed to be a complete graph.Here is an undirected graph containing nodes from 1 to 8. The edges are: (1,2,9), (1,3,3), (1,4,6), (1,5,8), (1,6,7), (1,7,4), (1,8,9), (2,3,10), (2,4,11), (2,5,5), (2,6,11), (2,7,1), (2,8,9), (3,4,11), (3,5,1), (3,6,9), (3,7,2), (3,8,9), (4,5,8), (4,6,3), (4,7,4), (4,8,8), (5,6,3), (5,7,3), (5,8,10), (6,7,8), (6,8,1), (7,8,10). (u ,v, w) denotes the edge from node *u* to node *v*.Question: What is the minimal cost of the Traveling Salesman Problem on the graph?You need to format your answer as a float number.27.0
trianglesThe task is to find the number of triangles that include a specific node as one vertex.A triangle is a set of three nodes that are all connected to each other.The input graph is guaranteed to be undirected.Here is an undirected graph containing nodes from 1 to 8. The edges are: (1,2), (1,3), (1,4), (1,5), (1,6), (1,7), (1,8), (2,3), (2,4), (2,5), (2,6), (2,7), (2,8), (3,4), (3,5), (3,6), (3,7), (3,8), (4,5), (4,6), (4,7), (4,8), (5,6), (5,7), (5,8), (6,7), (6,8), (7,8).Question: How many triangles include node 1 in the graph?Your answer should be an integer.21
+ +Continuing table 26 + +
TaskPromptAnswer
weighted minimum spanning treeThe task is to determine the minimum spanning tree of a weighted graph.A minimum spanning tree is a subset of the edges that connects all vertices in the graph with the minimum possible total edge weights. If not specified, all edges have equal edge weights.The input graph is guaranteed to be undirected and connected.Here is an undirected graph containing nodes from 1 to 5. The edges are: (1, 4, 5), (2, 4, 11), (2, 3, 10), (3, 4, 2), (3, 5, 2). (u ,v, w) denotes the edge from node *u* to node *v* has a weight of *w*.Question: What is the minimum spanning tree of the weighted graph?You need to format your answer as a list of edges in ascending dictionary order, e.g., [(u1, v1), (u2, v2), ..., (un, vn)].[(1, 4), (2, 3), (3, 4), (3, 5)]
weighted shortest pathThe task is to determine the shortest path between two nodes of a weighted graph.The input nodes are guaranteed to be connected.Here is a directed graph containing nodes from 1 to 8. The edges are: (1, 2, 5), (1, 4, 3), (1, 7, 9), (2, 3, 10), (2, 4, 10), (3, 1, 11), (3, 4, 2), (3, 5, 6), (4, 1, 1), (4, 2, 4), (4, 6, 8), (4, 8, 2), (5, 1, 7), (5, 2, 11), (5, 6, 2), (5, 7, 5), (5, 8, 11), (6, 1, 7), (6, 2, 11), (6, 3, 4), (6, 5, 1), (6, 8, 11), (7, 1, 3), (7, 2, 8), (7, 4, 7), (7, 6, 6), (7, 8, 3), (8, 1, 11), (8, 2, 7), (8, 4, 5), (8, 7, 5). (u ,v, w) denotes the edge from node *u* to node *v* has a weight of *w*.Question: What is the shortest path between node 1 and node 5?You need to format your answer as a list of nodes, e.g., [node-1, node-2, ..., node-n].[1, 4, 6, 5]
Wiener indexThe task is to determine the Wiener index of a connected graph.The Wiener index of a graph is the sum of the shortest-path distances between each pair of reachable nodes. For pairs of nodes in undirected graphs, only one orientation of the pair is counted.In the input graph, all node pairs are guaranteed to be reachable.Here is an undirected graph containing nodes from 1 to 5. The edges are: (1, 2), (1, 4), (2, 3), (4, 5), (3, 5).Question: What is the Wiener index of the graph?You need to format your answer as a float number.15.0
\ No newline at end of file diff --git a/NeurIPS/2025/$_texttt{G1}$_ Teaching LLMs to Reason on Graphs with Reinforcement Learning/images.zip b/NeurIPS/2025/$_texttt{G1}$_ Teaching LLMs to Reason on Graphs with Reinforcement Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..1968fb55a90f84725a4396b1365d6c3a8968aa82 --- /dev/null +++ b/NeurIPS/2025/$_texttt{G1}$_ Teaching LLMs to Reason on Graphs with Reinforcement Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5ad5b575c47112d4a0f6deb55c39d17430409fb9bb233609ed75679423086701 +size 4322174 diff --git a/NeurIPS/2025/$_texttt{G1}$_ Teaching LLMs to Reason on Graphs with Reinforcement Learning/layout.json b/NeurIPS/2025/$_texttt{G1}$_ Teaching LLMs to Reason on Graphs with Reinforcement Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..6d07cd22e5dd36c7aa7f232d258f6415136eaebe --- /dev/null +++ b/NeurIPS/2025/$_texttt{G1}$_ Teaching LLMs to Reason on Graphs with Reinforcement Learning/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:33a9a5446c17d0441142e4210f71362e6753e75621835fe1538afcf82dc69358 +size 790146 diff --git a/NeurIPS/2025/$_texttt{STRCMP}$_ Integrating Graph Structural Priors with Language Models for Combinatorial Optimization/3e772e3d-1de4-44c2-a5f8-ed39ed045157_content_list.json b/NeurIPS/2025/$_texttt{STRCMP}$_ Integrating Graph Structural Priors with Language Models for Combinatorial Optimization/3e772e3d-1de4-44c2-a5f8-ed39ed045157_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..e39d1404128f44e40446d9705f96730e1392c01f --- /dev/null +++ b/NeurIPS/2025/$_texttt{STRCMP}$_ Integrating Graph Structural Priors with Language Models for Combinatorial Optimization/3e772e3d-1de4-44c2-a5f8-ed39ed045157_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3643772ca371700daf8bec8eb0326530cac57a5bc8e2c36280ef5cae69c62f71 +size 242986 diff --git a/NeurIPS/2025/$_texttt{STRCMP}$_ Integrating Graph Structural Priors with Language Models for Combinatorial Optimization/3e772e3d-1de4-44c2-a5f8-ed39ed045157_model.json b/NeurIPS/2025/$_texttt{STRCMP}$_ Integrating Graph Structural Priors with Language Models for Combinatorial Optimization/3e772e3d-1de4-44c2-a5f8-ed39ed045157_model.json new file mode 100644 index 0000000000000000000000000000000000000000..202b88ab7d78b092c4d9f7c1ce78f7dcf7c3a9bd --- /dev/null +++ b/NeurIPS/2025/$_texttt{STRCMP}$_ Integrating Graph Structural Priors with Language Models for Combinatorial Optimization/3e772e3d-1de4-44c2-a5f8-ed39ed045157_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4fc58f36766b3878174be074d1488988c8dfd20aa04b4a8f60cd842262c784d1 +size 312458 diff --git a/NeurIPS/2025/$_texttt{STRCMP}$_ Integrating Graph Structural Priors with Language Models for Combinatorial Optimization/3e772e3d-1de4-44c2-a5f8-ed39ed045157_origin.pdf b/NeurIPS/2025/$_texttt{STRCMP}$_ Integrating Graph Structural Priors with Language Models for Combinatorial Optimization/3e772e3d-1de4-44c2-a5f8-ed39ed045157_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..d3971d519bbee89ab26be3cf459aede24fc3f561 --- /dev/null +++ b/NeurIPS/2025/$_texttt{STRCMP}$_ Integrating Graph Structural Priors with Language Models for Combinatorial Optimization/3e772e3d-1de4-44c2-a5f8-ed39ed045157_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:051f1f2600b6b9c714ab7ea65def3afe6c7871a105b994439d7af4eb61c0c84d +size 1714866 diff --git a/NeurIPS/2025/$_texttt{STRCMP}$_ Integrating Graph Structural Priors with Language Models for Combinatorial Optimization/full.md b/NeurIPS/2025/$_texttt{STRCMP}$_ Integrating Graph Structural Priors with Language Models for Combinatorial Optimization/full.md new file mode 100644 index 0000000000000000000000000000000000000000..6a6a1da69780fdb9c9e5cfd74fe4dc014c16a34b --- /dev/null +++ b/NeurIPS/2025/$_texttt{STRCMP}$_ Integrating Graph Structural Priors with Language Models for Combinatorial Optimization/full.md @@ -0,0 +1,1223 @@ +# STRCMP: Integrating Graph Structural Priors with Language Models for Combinatorial Optimization + +Xijun Li $^{1,2}$ , Jiexiang Yang $^{2}$ , Jinghao Wang $^{2}$ , Bo Peng $^{1,2}$ , Jianguo Yao $^{1,2}$ , Haibing Guan $^{1,2}$ + +$^{1}$ Shanghai Key Laboratory of Scalable Computing and Systems + +$^{2}$ School of Computer Science, Shanghai Jiao Tong University + +{lixijun,jianguo.yao}@sjtu.edu.cn + +# Abstract + +Combinatorial optimization (CO) problems, central to operation research and theoretical computer science, present significant computational challenges due to their $\mathcal{NP}$ -hard nature. While large language models (LLMs) have emerged as promising tools for CO—either by directly generating solutions or synthesizing solver-specific codes—existing approaches often neglect critical structural priors inherent to CO problems, leading to suboptimality and iterative inefficiency. Inspired by human experts' success in leveraging CO structures for algorithm design, we propose STRCMP, a novel structure-aware LLM-based algorithm discovery framework that systematically integrates structure priors to enhance solution quality and solving efficiency. Our framework combines a graph neural network (GNN) for extracting structural embeddings from CO instances with an LLM conditioned on these embeddings to identify high-performing algorithms in the form of solver-specific codes. This composite architecture ensures syntactic correctness, preserves problem topology, and aligns with natural language objectives, while an evolutionary refinement process iteratively optimizes generated algorithm. Extensive evaluations across Mixed Integer Linear Programming and Boolean Satisfiability problems, using nine benchmark datasets, demonstrate that our proposed STRCMP outperforms five strong neural and LLM-based methods by a large margin, in terms of both solution optimality and computational efficiency. The code is publicly available in the repository: https://github.com/Y-Palver/L2O-STRCMP. + +# 1 Introduction + +Making complex plans subject to multiple constraints is a time- and labor-intensive process, but is critical in many aspects of our lives such as scheduling [1], logistics [2], and robotics [3]. These problems are frequently modeled as combinatorial optimization (CO) problem, a cornerstone of operation research and theoretical computer science, where the objective is to identify optimal solutions within discrete, highly constrained search spaces. However, the inherent $\mathcal{NP}$ -hardness of many CO problems renders obtaining exact solutions computationally intractable, posing significant challenges in solving them efficiently and accurately. Consequently, substantial research efforts have endeavored to developing algorithms + +that balance solution quality with computational cost, typically requiring specialized domain expertise and rigorous analytical modeling of problem structure [4-6]. + +![](images/82815daed70a8694e31ca9188dcc91fc3e6b5443bdf0dfa1af6f65da24aec385.jpg) +Prompt +LLM +Code +Solver + +![](images/e4618fc5a5c6ae55d5da3f5860f7a20637cd0c4058bee4187d15f49ac8fc9162.jpg) +Previous work +Prompt +The proposed method +Figure 1: Intuitive comparison between prior work and the proposed framework. +ally requiring specialized domain expertise + +Machine learning (ML) techniques, particularly deep learning and reinforcement learning, have demonstrated significant potential in addressing CO problems [7]. As many CO problems arising from similar application domains exhibit inherent structural patterns, ML methods can leverage these patterns to reduce computational complexity in traditional CO approaches through data-driven strategies. The integration of ML and CO frameworks can be broadly categorized into three paradigms[7]: (1) End-to-end neural solvers [8-10], which train ML models to directly output CO solutions; (2) ML-augmented algorithm configuration [11-13], leveraging ML models to predict optimal algorithmic configurations for specific CO algorithm to solve associated problems; and (3) Hybrid methods [14- 16], embedding ML models as critical decision modules within traditional CO solvers. However, real-world adoption of above approaches remains limited due to reliance on training distributional alignment for generalization and inherently insufficient interpretability [17]. + +Building on the explosive advancements in large language models (LLMs) over recent years, combinatorial optimization tasks have shown potential for delegation to LLMs, particularly owing to their improved accountability and interpretability compared to conventional ML approaches. Capitalizing on their extensive world-knowledge priors and advanced code generation capabilities, numerous LLM-based approaches for CO problems have emerged. These methods bifurcate into two classes: (1) direct utilization of LLMs to solve CO problems [18, 19], and (2) employing LLMs to discover high-quality algorithm (in the form of solver-specific codes) for traditional solvers [20-22]. For example, Yang et al. [18] present Optimization by PROMpting (OPRO), which harnesses LLMs as black-box optimizers through iterative solution refinement via natural-language problem specifications and optimization trajectories. In contrast, Romera-Paredes et al. [20] proposes FunSearch that integrates a pre-trained LLM into an evolutionary algorithm to incrementally generate solver code, achieving state-of-the-art results on the cap set problem and discovering novel heuristics for online bin packing. Additionally, Huang et al. [19] demonstrates that multi-modal fusion of textual and visual prompts enhances LLM-driven optimization performance, as evidenced in capacitated vehicle routing problems. + +Despite these advances, current LLM-based approaches for solving CO problems remain in their infancy. As illustrated in Figure 1, they primarily rely on textual (or occasionally visual) prompting mechanisms to interface with LLMs, failing to effectively exploit the inherent topological structures of CO problems. Historically, human experts have successfully leveraged structural priors of CO problems to design sophisticated and efficient algorithms [4-6]. Furthermore, existing LLM-based methods [20, 21, 18] often require multiple iterations to produce high-quality solutions or algorithm implementations, primarily due to their lack of integration with CO-specific structural priors. As highlighted in Yao's insightful analysis [23], pre-training language model establishes strong priors for conversational tasks but weaker priors for structured domains like computer controls or video games — challenges exacerbated in solving CO problems, as these domains differ significantly from Internet text distributions. Given these challenges, developing a generative model that effectively integrates structural priors for CO problem solving becomes particularly compelling. As shown in Figure 1, the distinct insight lies in constructing a generative model that generates solver-specific code while respecting the intrinsic topological structure of CO problems. Theoretically, this approach aligns with the rising research directions of multi-modal generative model. Practically, such integration offers substantial benefits, dramatically reducing inference iterations while enhancing solution quality in solving CO problem. + +To tackle this challenge, we present STRCMP1, a novel structure-prior-aware LLM-based algorithm discovery framework for solving CO problems. To the best of our knowledge, STRCMP is the first framework to explicitly integrate structural priors of CO problems into LLM-driven algorithm discovery, jointly improving solution quality and computational efficiency. Our framework first constructs a composite architecture combining a graph neural network (GNN) and an LLM. The GNN extracts structural embeddings from input CO instances, which serves as inductive biases for subsequent algorithm discovery via code generation. The LLM then produces solver-specific code conditioned on these structural priors, ensuring adherence to solver syntax, preservation of CO problems' topological characteristics, and alignment with natural language optimization objectives. This composite model is further embedded within an evolutionary algorithm-based refinement process to iteratively enhance solution quality. Besides, we provide the theoretical analysis why fusing the structure prior into LLM can benefit solving CO problem from the perspective of information theory. + +Extensive experiments across two representative CO domains (Mixed Integer Linear Programming and Boolean Satisfiability), spanning nine benchmark datasets, demonstrate that STRCMP outperforms five well-recognized baselines by a large margin—including neural combinatorial optimization methods and LLM-based approaches—in both solution optimality and computational efficiency. + +# 2 Related Work + +Neural Combinatorial Optimization. Extensive research [14, 24, 25, 15, 16] has explored machine learning integration with combinatorial optimization for improved computational approaches to $\mathcal{NP}$ -hard problems, forming Neural Combinatorial Optimization (NCO) methods that either approximate expert heuristics or learn policies via reinforcement learning. Following Bengio et al. [7]'s taxonomy, NCO encompasses: (1) End-to-end neural solvers [8-10], (2) ML-augmented algorithm configuration [11-13], and (3) Hybrid methods integrating learned components into classical frameworks [14-16]. Pointer Networks [25] represent seminal work, employing attention mechanisms [8] as dynamic pointers for CO problems, eliminating dependence on fixed output dictionaries in conventional seq2seq models [26]. The framework achieves competitive performance on convex hulls, Delaunay triangulation, and small TSP instances while generalizing to unseen sizes. In hybrid approaches, Gasse et al. [14] introduced GCNs trained via imitation learning to replace strong branching heuristics [27] in branch-and-bound algorithms for MIP solvers. While avoiding manual feature engineering through graph representations, such approaches face scalability challenges under extreme problem sizes, a key limitation of neural solver architectures. Moreover, NCO's reliance on opaque machine learning models yields policies with insufficient interpretability—problematic for high-stakes decision-making systems demanding traceable logic [17]. Our approach addresses these gaps by furnishing optimality certificates or bounded sub-optimality guarantees via CO solver integration. By procedurally generating solver-executable code through iterative refinement, our framework ensures superior interpretability relative to black-box NCO policy learning. + +Language Models for Combinatorial Optimization Problem. Recent advances in large language models (LLMs) have spurred explorations of their applications to combinatorial optimization [20, 21, 28-30]. Current approaches fall into two categories: (1) direct solution generation via LLM inference [18, 19], and (2) automated generation of solver-compatible formalizations [20-22, 30]. In the first paradigm, Yang et al. [18] propose Optimization by PROempting (OPRO), leveraging LLMs as black-box optimizers through iterative refinement of natural-language problem descriptions. While achieving comparable performance to heuristics on small TSP instances, OPRO scales poorly beyond 50 nodes due to context limitations and solution-space complexity. For the second paradigm, Romero-Paredes et al. [20] introduce FunSearch, combining evolutionary methods with LLM-guided program synthesis to obtain breakthroughs on the cap set problem [31] and generate novel bin packing heuristics [32]. Similarly, [21] develop Evolution of Heuristics (EoH), integrating evolutionary algorithms with LLMs to co-optimize natural language "thoughts" and code implementations through iterative prompting, outperforming existing methods like FunSearch in query efficiency [20]. Existing approaches predominantly employ natural language (or occasionally visual [19]) prompts combined with evolutionary frameworks for iterative code generation to solve CO problems. However, these methods universally neglect the topological structural characteristics intrinsic to combinatorial optimization problems, which are systematically exploited by human experts during algorithm design [4-6]. Furthermore, evolutionary framework-based code generation with LLMs often necessitates multiple iterations for convergence while incurring significant computational overhead from repeated solver invocations during evaluation. To overcome these limitations, we propose a composite model that effectively incorporates structural priors of CO problems during algorithmic discovery. By developing composite architectures tailored to CO problem structures, our method substantially improves both solution quality and computational efficiency compared to existing LLM-based solving approaches. + +# 3 Problem Statement + +Combinatorial Optimization. Following [33], we formulate a general combinatorial optimization problem $Q$ (e.g., Boolean Satisfiability (SAT) and Mixed Integer Linear Programming (MILP)) as a constrained discrete optimization problem. For an instance $q$ of $Q$ , the formulation becomes: + +$$ +\min _ {x \in S (q)} c (x; q) + f (x; q), \tag {1} +$$ + +![](images/04b1469b41a9278736d219b0c93a927ac822aa467f65a65d2b3a922ac09b5e3d.jpg) +Figure 2: Overview of the proposed STRCMP. ① Combinatorial Structure Extraction. We utilize a graph neural network (GNN) to encode the topological structure of combinatorial optimization problems into latent embeddings, capturing problem-specific structural invariants. ② Structure-Aware Code Generation: (a) Data Curation: For a given CO problem's mathematical model, target solver and the specific prompt, an LLM generates candidate algorithm in the form of code snippets, which are automatically validated via the solver execution to generate performance metrics. This yields a curated dataset comprising (mathematical model, code snippet, metric) triplets. (b) Post Training: A composite architecture integrating a frozen pretrained GNN with an LLM is fine-tuned on this dataset, ensuring code generation respects both the CO's intrinsic topology and solver-specific syntax. ③ Evolutionary Code Refinement. Drawing on [20], the composite model is embedded within an evolutionary framework to evolve the algorithm discovery process through performance-driven iterations. + +where $x$ denotes decision variables subject to solution space $S(q); c(x;q)$ represents the objective function to minimize, and $f(x;q)$ corresponds to constraint penalties (zero when all constraints are satisfied). In SAT problems, the objective function $c(x;q)$ is absent since only constraint satisfaction is required. Conversely, MILP problems feature both objective and constraint functions. + +Generation Task. Given a combinatorial optimization problem $Q$ , our goal is to find an algorithm $A$ that can consistently solve $Q$ well: + +$$ +\min _ {A \sim A} \mathbb {E} _ {q \sim Q, x \sim A (q)} [ c (x; q) + f (x; q) ] \tag {2} +$$ + +where $\mathcal{A}$ denotes the algorithm search space. This work focuses on symbolic search spaces where $\mathcal{A}$ comprises algorithms representable as code snippets executable within a combinatorial optimization solver. We resort to a generative model to synthesize the target algorithm through code generation. + +# 4 Proposed Solution + +The proposed framework is illustrated in Figure 2. We first describe the technical interplay among components within the proposed framework, highlighting design motivations and synergistic interactions. Subsequently, we provide a theoretical analysis demonstrating that integrating structural priors into LLMs enhances performance in solving combinatorial optimization problems. + +# 4.1 Methodology + +1 Combinatorial Structure Extraction. Combinatorial optimization problem is equivalently converted to a bipartite graph. Thus, the intrinsic structure of the combinatorial optimization problem can be easily captured by a graph neural network, as has been done in many previous work [16, 24, 14]. + +As shown in the leftmost part of Figure 2, we continue to use this technique to extract structure embedding of given CO problem, benefiting the following algorithm discovery process. + +Specifically, a bipartite graph $G = (C, E, V)$ is constructed for the CO problem instance $q$ , where $C$ corresponds to the constraints of $q$ ; $V$ denotes the variables of $q$ ; and an edge $e_{ij} \in E$ between a constraint node $i$ and a variable node $j$ if they have a connection. A graph neural network $\theta_G$ takes as input the bipartite graph to generate the structural embedding $h_q \in \mathbb{R}^d$ of the problem instance $q$ . We simply train $\theta_G$ via a classification task, where a group of CO problems and their class label are given. The problems falling into the same class means that they originate from the same scenario/domain, such as traveling salesman problems [34], production planning [35], vehicle routing problem [36], etc. The data representation, training, implementation details are left to Appendix B.1. In this way, GNNs can capture structure priors of CO problem (such as symmetry, sparsity, and degeneracy), which are critical for solver decisions. The extracted embeddings transfers structural insights to downstream code generation task, facilitating generalization across problem classes. + +$\Theta$ Structure-Aware Code Generation. In this step, we construct a composite generative model to achieve structure-aware algorithm discovery for CO problem. The composite model is essentially a concatenation of a GNN and an LLM. GNN is obtained via training method mentioned in Step 1 and frozen in the composite model, which provides the structure embedding of given CO problem to LLM. The LLM takes as input the natural language description of the problem and of the code generation task. Then the LLM generates candidate algorithm in the form of code snippet conditioned on the natural language description and structure embedding for a specific solver, as shown in the middle part of Figure 2. The code generation procedure can be formulated below: + +$$ +P \left(w _ {1}, w _ {2}, \dots , w _ {T}\right) = \prod_ {t = 1} ^ {T} P _ {\theta_ {L}} \left(w _ {t} \mid w _ {< t}; \boldsymbol {h} _ {q}, N L\right), \tag {3} +$$ + +where $\{w_{i}|i = 1,\dots,T\}$ is the code snippet predicted by the composite model; $h_q$ is the structure embedding of given CO problem instance $q$ ; $NL$ is the natural language description of $q$ and of code generation task; $\theta_L$ is the parameter of LLM. + +As anticipated, the composite model exhibits suboptimal performance on code generation tasks for CO problems due to architectural incompatibilities that disrupt the native token prediction mechanism of the underlying LLM. To address this architectural mismatch, we implement a two-phase training protocol consisting of: (a) data curation through systematic problem sampling, and (b) parameter optimization via post-training. This phased approach enables effective adaptation of the composite architecture while preserving the structural integrity of the original CO problem representations. + +(a) Data Curation. To facilitate the post-training phase, we aim to collect four key categories of data: 1) mathematical formulations of CO problems; 2) natural language specifications detailing problem requirements alongside corresponding code generation objectives for targeted CO solvers; 3) executable code implementations derived from these specifications; and 4) quantitative performance metrics evaluating solution quality and computational efficiency of the implemented code. + +(b) Post-Training. In this phase, we only train the parameter $\theta_{L}$ of the composite model, which means that structural embeddings remain consistent during post-training. We conduct Supervised Fine-Tuning (SFT) on $\theta_{L}$ using the curated dataset, followed by Direct Preference Optimization (DPO) [37] initialized from the SFT checkpoint to derive the final composite model. All details, including prompt template, data curation and post-training, are presented in Appendix B.2. + +Evolutionary Code Refinement for Combinatorial Optimization Problem. Prior work [30, 20, 29] leverages LLMs' code generation and algorithm design capabilities within Evolutionary Algorithms (EAs) to address combinatorial optimization problems through iterative feedback frameworks. The iterative nature of evolutionary algorithms introduces significant computational overhead as the primary limitation of prior approaches [30, 20, 29] for identifying optimal code snippets or algorithms to solve combinatorial optimization problems. Convergence typically requires numerous iterations, exacerbated by the prolonged execution times required for solver-based evaluation of generated code or algorithm candidates. + +Our evolutionary code optimization framework for combinatorial optimization adopts the core principles of prior work[30, 20, 29], with the significant distinction lying in the composite model learned in Step ② (see Figure 2 right panel). Our framework jointly processes both textual problem descriptions with algorithm discovery objectives and formal mathematical model of CO problems. + +The composite model generates solver-specific, structure-aware code through standard EA operators (selection, crossover, mutation), producing higher-quality algorithm implementation within fewer iterations – a capability empirically validated in our experiments. By integrating this model into EA-based algorithm discovery frameworks, we achieve superior optimization performance while maintaining compatibility with existing evolutionary algorithm design paradigms, suggesting broader applicability across LLM-driven optimization methodologies. + +# 4.2 Theoretical Analysis + +In the following, we give the theoretical analysis why fusing the structure prior into the generative model helps algorithm discovery for solving CO problem on the basis of information theory [38] and multi-modal co-learning [39]. Specifically, we prove that a generative model with an additional prior will not lower the upper bound of its performance on the downstream task. Note that we leave all proofs into the Appendix A due to the space limitation. + +Definition 1 (Upper Bound of Model Performance). The performance upper bound of generative model, denoted by $\sup (\mathcal{P})$ , is utilized to measure the maximum expected performance of a generative model for a downstream task $p(\mathbf{w}|\mathbf{c})$ , where $\mathbf{w}$ is the generated content conditional on the different kinds of prior $\mathbf{c}$ . Formally, $\sup (\mathcal{P})$ associated with kinds of prior $\mathcal{C}$ can be expressed as + +$$ +\sup \left(\mathcal {P} _ {\mathcal {C}}\right) = \sum_ {\mathbf {C} \in \mathcal {C}} \sum_ {c \in \mathbf {C}} p (\mathbf {c}) \max _ {\mathbf {w}} p (\mathbf {w} | \mathbf {c}) \Phi (\mathbf {w}), \tag {4} +$$ + +where $\Phi$ is the performance evaluator for given generated content $\mathbf{w}$ ; $\mathcal{C}$ is the complete set of different priors and $\mathbf{C}$ is a type of prior belonging to the prior set $\mathcal{C}$ . + +Theorem 1. Given a prior dataset $\mathcal{D}$ whose data samples comprises $M$ types of prior $\mathcal{C} = \{\mathbf{C}_1,\dots,\mathbf{C}_M\}$ , the distinct priors follow respective true distribution $p(\mathbf{C}_i|\mathbf{w})$ . Let $\mathbf{c}_i$ be a sample drawn from the distribution, i.e., $\mathbf{c}_i\sim p(\mathbf{C}_i|\mathbf{w})$ . A generative model with an additional type of prior will not lower the upper bound of model performance $\sup (\mathcal{P})$ . + +Remark 1. Theorem 1 establishes that introducing additional priors cannot reduce a generative model's performance upper bound, while decreasing its entropy. Consequently, integrating combinatorial optimization structural priors into LLMs does not degrade their performance in generating code for solving CO problems. + +Definition 2 (Performance-Enhancing Prior). Assume a type of prior $\mathbf{C}$ can boost the performance of a generative model compared to the one without the prior, then $\mathbf{C}$ is the performance-enhancing prior to the generative model. It can be formally expressed as + +$$ +\exists \mathbf {w} ^ {\prime} \neq \mathbf {w} ^ {*}, p \left(\mathbf {w} ^ {\prime} | \mathbf {c}\right) \Phi \left(\mathbf {w} ^ {\prime}\right) _ {\mathbf {c} \in \mathcal {C}} > p \left(\mathbf {w} ^ {*} | \mathbf {c}\right) \Phi \left(\mathbf {w} ^ {*}\right) _ {\mathbf {c} \in \mathcal {C} \backslash \{\mathbf {C} \}}, \tag {5} +$$ + +where $\mathbf{w}^{*} = \arg \max_{\mathbf{w}}p(\mathbf{w}|\mathbf{c})\Phi (\mathbf{w})_{\mathbf{c}\in \mathcal{C}\backslash \{\mathbf{C}\}}$ + +Theorem 2. If prior $\mathbf{C}_{pe}$ is a performance-enhancing prior, a generative model neglecting prior $\mathbf{C}_{pe}$ will decrease the upper bound of model performance. It can be expressed as + +$$ +\sup \left(\mathcal {P} _ {\mathcal {C} \backslash \{\mathbf {C} _ {p e} \}}\right) < \sup \left(\mathcal {P} _ {\mathcal {C}}\right) \tag {6} +$$ + +Remark 2. Theorem 2 demonstrates that generative models equipped with performance-enhancing priors achieve superior performance relative to their prior-free counterparts. Specifically, the structural prior serves as such performance-enhancing prior for LLMs in code generation tasks targeting CO problems, which is empirically validated through our comprehensive experiments. + +# 5 Experimental Evaluation + +To empirically evaluate the efficacy of our proposed STRCMP in solving combinatorial optimization problems, we conduct extensive experiments across two fundamental CO problem classes: mixed-integer linear programming and Boolean satisfiability, evaluated on over nine benchmark datasets. We benchmark our approach against five well-established baselines encompassing both neural optimization methods and contemporary LLM-based approaches. We aim to answer the following research questions: + +# Research Questions + +RQ1: Does the proposed STRCMP identify superior algorithmic implementations compared to existing algorithm discovery approaches? + +RQ2: Does the proposed composite model effectively reduce computational overhead in existing algorithm discovery frameworks? + +RQ3: Does the structural prior benefit the generative model in solving combinatorial optimization problems? + +# 5.1 Settings + +Baselines. We evaluate our framework against two categories of baselines: neural combinatorial optimization methods and LLM-based evolutionary code optimization frameworks, covering mixed-integer linear programming (MILP) and Boolean satisfiability (SAT). More details of baselines and used backend solvers can be found in Appendix C. + +- Neural Combinatorial Optimization: For MILP, we compare with the seminal work L2B [14], which employs graph convolutional networks for variable selection to replace the strong branching policy, and HEM [15, 16], a hierarchical sequence model for cut selection in branch-and-bound solvers. For SAT, we include NeuroSAT [40], a message-passing neural network trained with single-bit supervision for SAT solving. +- Evolutionary Code Optimization: While numerous LLM-based evolutionary code optimization approaches exist for combinatorial optimization, we specifically compare with two methods specialized for SAT and MILP: AutoSAT [29] and LLM4Solver [30]. AutoSAT leverages LLMs to automate heuristic optimization in SAT solvers, minimizing manual intervention. LLM4Solver integrates LLMs with multi-objective evolutionary algorithms to automatically design effective diving heuristics for MILP solvers $^{2}$ . + +Dataset. We perform the empirical evaluation over ten widely-used benchmark dataset for SAT and MILP respectively. More details and statistics of the used datasets can be found in Appendix D. + +- Mixed-integer linear programming (MILP): The datasets for MILP solver evaluation contain three difficulty tiers [15, 16, 30]: (1) Easy features synthetic benchmarks (Set Covering [41], Maximum Independent Set [42], Multiple Knapsack [43]) generated using protocols from [44, 45]; (2) Medium includes MIK [46] and CORLAT [47]; (3) Hard contains the Google-inspired Load Balancing problem and the industrial-scale Anonymous problem [48]. +- Boolean satisfiability (SAT): The dataset comprises two sources: SAT Competition problems [49] and automatically generated instances via Picat [29]. SAT Competition data includes Profitable-Robust-Product (PRP) and Chromatic-Number-of-the-Plane (CNP) problems, while Picat-generated data contains CoinsGrid and Zamkeller instances. We adhere to the generation protocol established in [29]. + +Metrics. To evaluate the effectiveness of the proposed framework, we analyze solving time to optimality or solution quality attainable within a fixed time budget. For solving efficiency evaluation, we measure the number of iterations or training steps required for convergence. Specifically, for MILP domain, critical metrics include solving time and primal-dual (PD) integral which are widely used in benchmarking the MILP solvers [15, 16, 14]. For SAT domain, key metrics encompass solving time, PAR-2, and number of timeout frequently measured in evaluating SAT solvers [29, 40]. Metrics such as solving time, PD integral, number of timeout, and PAR-2 are minimized through optimization. More details of the used metrics are presented in Appendix E. + +# 5.2 Implementation + +Model Architecture. The proposed composite model comprises two components: a GNN and a structure-prior-aware LLM. We implement the GNN using graph convolution operators from torchgéometric, structured with three sequential convolutional layers terminated by global mean + +![](images/a4065eab2e905268d1c62c4b788819b8fef1a43296dc08970215f5075696eda7.jpg) +(a) PAR-2 in SAT domain + +![](images/e62608436d123455dc024b1a77781a7c0b593526f08c33c76693ce1d0806531c.jpg) +(b) PD integral in MILP domain +Figure 3: Optimization performance over different CO domains. Aligned with [29], we normalize each benchmark's indicator values using $1 - \frac{\text{value} - \min}{2 \times (\max - \min)}$ , where value represents the method's measured indicator, with min and max denoting respectively the minimum and maximum values observed during evaluation. A larger shaded area corresponds to superior performance. + +pooling. We adapt the LLM component based on Qwen2.5-Coder-7B-Instructor model through modifying its architecture, forward propagation dynamics and inference paradigm. Comprehensive implementation details including hyperparameter configurations, adaptations, and software dependencies are provided in Appendix B.1 and B.2. + +Training, Inference and Used Hardware. We first train domain-specific GNNs for SAT and MILP problems using aforementioned benchmark dataset. The GNNs are trained for three epochs on SAT instances and four epochs on MILP instances, followed by conducting post-training of the adapted LLM with a specialized corpus. All post-training data is collected via queries to Qwen2.5-Coder-7B-Instructor, aligned with the model used in the post training. The adapted LLM undergoes three epochs of post training per domain. Both GNN and LLM selecting optimal checkpoints based on validation prediction loss. We then integrate the composite model (i.e. the trained GNN and adapted LLM) into an EA-based framework to discover solver-specific algorithm configurations optimized for training instances. Finally, we evaluate the performance of identified algorithms on held-out test sets. All experiments are conducted on hardware platform with dual AMD EPYC 9534 64-core processors @ 2.45GHz and two NVIDIA H800 80GB GPUs connected via PCIe. More training specifics and data curation are detailed in Appendix B.1 and B.2 respectively. + +# 5.3 Results + +Optimization Performance (Answer to RQ1). The comparative optimization results are presented in Figure 3 and Table 1 ("CGD" and "ZAM" denotes CoinsGrid and Zamkeller dataset respectively) with respect to primary objective metrics (i.e. PAR-2 and Number of Timeout for SAT and PD integral for MILP domains). Comprehensive results for all evaluation metrics are provided in Appendix G.1). As evidenced in Figure 3 and Table 1, our STRCMP with various post-training variants consistently matches or exceeds all baseline performance across both SAT and MILP domains. Specifically for SAT, STRCMP demonstrates universal superiority over its closest counterpart AutoSAT, particularly achieving significant reductions in terms of timeout $(77.8\%)$ on Zamkeller: $18\rightarrow 4$ . $66.7\%$ on PRP: $9\rightarrow 3$ ) and solving time (on PRP: 22967 seconds $\rightarrow$ 21146 seconds; on + +Zamkeller: 20772 seconds $\rightarrow$ 6929 seconds). On MILP benchmarks, STRCMP maintains strong performance parity with NCO methods L2B/HEM and its direct competitor LLM4Solver. These + +Table 1: Optimization performance result w.r.t. Number of Timeout between different methods over SAT domain. + +
Compared MethodsNumber of Timeout (↓)
CNPCGDPRPZAM
AutoSAT [29]3216918
EasySAT32255018
NeuroSAT [40]44184629
STRCMP3117445
STRCMP (DPO Only)321635
STRCMP (SFT Only)3316454
STRCMP w/o GNN32164414
+ +empirical findings conclusively address RQ1: STRCMP successfully identifies superior algorithmic configurations compared to existing strong baselines. + +Efficiency Improvement (Answer to RQ2). Next, we assess the efficiency of our proposed STRCMP framework in discovering high-performance algorithms via iterative search for combinatorial optimization problems. The representative convergence comparison for SAT domain are presented in Figures 4. More evaluations over other SAT datasets and MILP domain across additional metrics are provided in Appendix G.2. Our experiments reveal that STRCMP achieves convergence significantly faster than existing evolutionary-based algorithm discovery frameworks. Furthermore, the framework + +![](images/9439b95f949f6a779f08ac325726fe6f1226b9f717ca949614216084b6be35ea.jpg) +Figure 4: Convergence comparison (w.r.t. PAR-2) between evolutionary-based algorithm discovery frameworks on Zamkeller dataset of SAT domain. + +attains markedly higher-quality convergence points compared to baseline methods AutoSAT and LLM4Solver. Notably, STRCMP achieves stable convergence while AutoSAT exhibits persistent oscillations even after convergence. These findings validate that our proposed STRCMP substantially reduces computational overhead in evolutionary-based algorithm discovery frameworks. Ablation Studies (Answer to RQ3). To address RQ3, we conduct comprehensive ablation studies on our composite model by systematically deactivating individual components and evaluating their impact on optimization performance and computational efficiency. Representative results are shown in Figure 5, with full ablation studies detailed in Appendix G.3. Analysis of Figure 5 reveals that the structural prior provides measurable benefits for code generation in combinatorial optimization tasks: the STRCMP w/o GNN variant (lacking structural guidance) exhibits consistently inferior optimization performance compared to counterparts incorporating the prior. Furthermore, this variant demonstrates increased solution variability during algorithmic search iterations, potentially resulting in higher computational costs. Counterintuitively, the full STRCMP model (with complete post-training) does not uniformly outperform its ablations STRCMP (SFT Only) and STRCMP (DPO Only) across all benchmarks, suggesting underlying conflicts within the post-training data distribution. We plan to investigate this phenomenon in subsequent research. + +# 5.4 Further Analysis + +Two-phase Training v.s. Joint Training? Besides the implementation of two-phase training strategy, we also design and implement an end-to-end joint training method for our proposed model. This new method directly aligns the learning objectives of the graph and text modalities with the same optimization goal: improving the performance of the generated code. Specifically, we do not discriminate between the weights of the GNN and the LLM, meaning they share the same loss function, and the gradient is backpropagated from the parameters of the LLM to those of the GNN. We implemented the end-to-end joint training method described above and tested the resulting models on two SAT datasets (i.e., Zamkeller and PRP). All the experimental settings keep the same except + +![](images/637867bed0f97ca7d42e2993a4d997ce30d52508fa63b94845a847a49a4524c5.jpg) +(a) PAR-2 in SAT domain + +![](images/4f5f32929702a5a08ac1e4f714d87a96b2e35d9e09e24b153ba7d1c43a507d1c.jpg) +(b) Convergence performance on Zamkeller dataset +Figure 5: Ablation studies in terms of optimization performance and convergence rate. + +for the training method. The detailed results of this new training and testing are presented show in Table 7 and Table 8 in Appendix I. We can conclude the followings from these results: 1) The end-to-end joint training is extremely unstable compared to our proposed two-stage training method. Compared with the joint training, the loss curve of our two-stage training method is stable and close to zero at convergence. The GNN and LLM are trained together, which makes it difficult for the models to learn optimal parameters simultaneously; 2) The model exhibits poor performance when tested on solving SAT instances. Unsurprisingly, due to its very unstable training and relatively high loss at convergence, the performance of the jointly trained model is much lower than our two-stage model on the two SAT datasets. + +Insight of Training Strategy. To provide a clearer and more comprehensive answer, we offer a detailed analysis focusing on dataset complexity and the iterative optimization behavior of our proposed method, STRCMP, and its variants, shown in Appendix J. Through the above thorough analysis and additional experiments, we can summarize our insights on selecting a training strategy: 1) For "easy" datasets: The SFT-only approach is often sufficient and efficient. On these problems, high-quality training data (i.e., optimal or near-optimal solutions) can be generated at a low cost, allowing SFT to quickly learn an effective policy. 2) For "hard" datasets: We strongly recommend a DPO-based approach (DPO-only or SFT+DPO). For these problems, obtaining optimal solutions for SFT is prohibitively expensive or impossible. However, generating preference pairs by comparing the performance of different candidate solutions is still feasible and relatively cheap. The superior generalization of DPO-trained models makes them better suited to navigating the vast and complex search spaces of these challenging instances. 3) For medium complexity datasets, the choice is less definitive. As the results show, the interplay between SFT and DPO is intricate. Our full STRCMP model (SFT+DPO) often acts as a robust default, leveraging SFT for a strong initialization and DPO for refinement and generalization. + +Dependence on the Underlying LLM. To investigate how much does the performance of STRCMP depend on the underlying LLM, we focused our evaluation on the Llama2 and Qwen2 families of models, selecting representatives of varying sizes. We performed training and testing on two datasets from the MILP domain and two from the SAT domain. The full result and analysis are presented in Appendix K. It is important to note that the entire Llama2 family of models failed on our code generation task, either by exceeding the context limitation of 4096 tokens or by being unable to generate syntactically correct, executable code. Similarly, the smaller Qwen2 models (0.5B and 1.5B) were also unable to produce viable code for the solvers. This underscores the complexity of the task, which requires a highly capable code-generating LLM. From the results in Table 18 and Table 19, we can draw the following conclusions: 1) The structural embedding from the GNN consistently and significantly contributes to the final performance; 2) STRCMP's performance is robust, provided a sufficiently capable LLM backbone is used, and 3) STRCMP's performance is sensitive to the size of the LLM backbone, but this effect is problem-dependent. + +# 6 Conclusion + +We propose STRCMP, a novel structure-aware LLM-based algorithm discovery framework for combinatorial optimization problems. To our knowledge, this represents the first methodology that explicitly incorporates structural priors of combinatorial optimization problems into LLM-driven algorithm discovery, simultaneously enhancing solution quality and computational efficiency. We validate the effectiveness of STRCMP through theoretical analysis and extensive empirical evaluations across SAT and MILP problems, demonstrating consistent improvements in both solving efficiency and solution optimality. Our framework offers a principled foundation for future research integrating additional modal priors into LLMs for solving combinatorial optimization problems. + +# Acknowledgements + +We thank all the reviewers and chairs for their reviews and valuable feedback. This work was funded by the National Key Research & Development Program of China (No. 2022YFB4500103), NSFC (No. 62506227, 62032008), STCSM (No. 25ZR1402224, 23511100100), Startup Fund for Young Faculty at SJTU (SFYF at SJTU, No. 25X010502616) and Aviation Key Laboratory of Science and Technology on Aerospace Vehicle. The corresponding authors are Bo Peng and Jianguo Yao. + +# References + +[1] David M Ryan and Brian A Foster. An integer programming approach to scheduling. Computer scheduling of public transport urban passenger vehicle and crew scheduling, pages 269-280, 1981. +[2] Xijun Li, Mingxuan Yuan, Di Chen, Jianguo Yao, and Jia Zeng. A data-driven three-layer algorithm for split delivery vehicle routing problem with 3d container loading constraint. In Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, pages 528-536, 2018. +[3] Bo Liu, Yuqian Jiang, Xiaohan Zhang, Qiang Liu, Shiqi Zhang, Joydeep Biswas, and Peter Stone. Llm+ p: Empowering large language models with optimal planning proficiency. arXiv preprint arXiv:2304.11477, 2023. +[4] Etienne De Klerk. Exploiting special structure in semidefinite programming: A survey of theory and applications. European Journal of Operational Research, 201(1):1-10, 2010. +[5] Maria Kandyba-Chimani. Exact algorithms for network design problems using graph orientations. PhD thesis, Citeseer, 2011. +[6] Bruce Hendrickson. The molecule problem: Exploiting structure in global optimization. SIAM Journal on Optimization, 5(4):835-857, 1995. +[7] Yoshua Bengio, Andrea Lodi, and Antoine Prouvost. Machine learning for combinatorial optimization: a methodological tour d'horizon. European Journal of Operational Research, 290(2):405-421, 2021. +[8] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. +[9] Irwan Bello, Hieu Pham, Quoc V Le, Mohammad Norouzi, and Samy Bengio. Neural combinatorial optimization with reinforcement learning. arXiv preprint arXiv:1611.09940, 2016. +[10] Elias B Khalil, Christopher Morris, and Andrea Lodi. Mip-gnn: A data-driven framework for guiding combinatorial solvers. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 10219-10227, 2022. +[11] Carlos Ansótegui, Britta Heymann, Josep Pon, Meinolf Sellmann, and Kevin Tierney. Hyperreactive tabu search for maxsat. In Learning and Intelligent Optimization: 12th International Conference, LION 12, Kalamata, Greece, June 10–15, 2018, Revised Selected Papers 12, pages 309–325. Springer, 2019. +[12] Markus Kruber, Marco E Lübbecke, and Axel Parmentier. Learning when to use a decomposition. In International conference on AI and OR techniques in constraint programming for combinatorial optimization problems, pages 202-210. Springer, 2017. +[13] Pierre Bonami, Andrea Lodi, and Giulia Zarpellon. Learning a classification of mixed-integer quadratic programming problems. In International conference on the integration of constraint programming, artificial intelligence, and operations research, pages 595–604. Springer, 2018. +[14] Maxime Gasse, Didier Chételat, Nicola Ferroni, Laurent Charlin, and Andrea Lodi. Exact combinatorial optimization with graph convolutional neural networks. arXiv preprint arXiv:1906.01629, 2019. +[15] Zhihai Wang, Xijun Li, Jie Wang, Yufei Kuang, Mingxuan Yuan, Jia Zeng, Yongdong Zhang, and Feng Wu. Learning cut selection for mixed-integer linear programming via hierarchical sequence model. arXiv preprint arXiv:2302.00244, 2023. +[16] Jie Wang, Zhihai Wang, Xijun Li, Yufei Kuang, Zhihao Shi, Fangzhou Zhu, Mingxuan Yuan, Jia Zeng, Yongdong Zhang, and Feng Wu. Learning to cut via hierarchical sequence/set model for efficient mixed-integer programming. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. + +[17] Weiwei Sun, Shengyu Feng, Shanda Li, and Yiming Yang. Co-bench: Benchmarking language model agents in algorithm search for combinatorial optimization. arXiv preprint arXiv:2504.04310, 2025. +[18] Chengrun Yang, Xuezhi Wang, Yifeng Lu, Hanxiao Liu, Quoc V Le, Denny Zhou, and Xinyun Chen. Large language models as optimizers. arXiv preprint arXiv:2309.03409, 2023. +[19] Yuxiao Huang, Wenjie Zhang, Liang Feng, Xingyu Wu, and Kay Chen Tan. How multimodal integration boost the performance of llm for optimization: Case study on capacitated vehicle routing problems. arXiv preprint arXiv:2403.01757, 2024. +[20] Bernardino Romera-Paredes, Mohammadamin Barekatain, Alexander Novikov, Matej Balog, M Pawan Kumar, Emilien Dupont, Francisco JR Ruiz, Jordan S Ellenberg, Pengming Wang, Omar Fawzi, et al. Mathematical discoveries from program search with large language models. Nature, 625(7995):468-475, 2024. +[21] Fei Liu, Xialiang Tong, Mingxuan Yuan, Xi Lin, Fu Luo, Zhenkun Wang, Zhichao Lu, and Qingfu Zhang. Evolution of heuristics: Towards efficient automatic algorithm design using large language model. arXiv preprint arXiv:2401.02051, 2024. +[22] Haoran Ye, Jiarui Wang, Zhiguang Cao, Federico Berto, Chuanbo Hua, Haeyeon Kim, Jinkyoo Park, and Guojie Song. Reevo: Large language models as hyper-heuristics with reflective evolution. arXiv preprint arXiv:2402.01145, 2024. +[23] Shunyu Yao. The Second Half. https://ysymyth.github.io/The-Second-Half/, 2025. +[24] Xijun Li, Qingyu Qu, Fangzhou Zhu, Mingxuan Yuan, Jia Zeng, and Jie Wang. Accelerating linear programming solving by exploiting the performance variability via reinforcement learning. 2023. +[25] Oriol Vinyals, Meire Fortunato, and Navdeep Jaitly. Pointer networks. Advances in neural information processing systems, 28, 2015. +[26] Ilya Sutskever, Oriol Vinyals, and Quoc V Le. Sequence to sequence learning with neural networks. Advances in neural information processing systems, 27, 2014. +[27] Santanu S Dey, Yatharth Dubey, Marco Molinaro, and Prachi Shah. A theoretical and computational analysis of full strong-branching. Mathematical Programming, 205(1):303-336, 2024. +[28] Fei Liu, Yiming Yao, Ping Guo, Zhiyuan Yang, Zhe Zhao, Xi Lin, Xialiang Tong, Mingxuan Yuan, Zhichao Lu, Zhenkun Wang, et al. A systematic survey on large language models for algorithm design. arXiv preprint arXiv:2410.14716, 2024. +[29] Yiwen Sun, Furong Ye, Xianyin Zhang, Shiyu Huang, Bingzhen Zhang, Ke Wei, and Shaowei Cai. Autosat: Automatically optimize sat solvers via large language models. arXiv preprint arXiv:2402.10705, 2024. +[30] Yuyan Zhou, Jie Wang, Yufei Kuang, Xijun Li, Weilin Luo, Jianye HAO, and Feng Wu. Llm4 solver: Large language model for efficient algorithm design of combinatorial optimization solver. +[31] Ernie Croot, Vsevolod F Lev, and Péter Pál Pach. Past and future of the cap set problem. arXiv preprint arXiv:2408.02328, 2024. +[32] Baoying Wang, Zhaohui Lin, Weijie Kong, and Huixu Dong. Bin packing optimization via deep reinforcement learning. IEEE Robotics and Automation Letters, 2025. +[33] Christos H Papadimitriou and Kenneth Steiglitz. Combinatorial optimization: algorithms and complexity. Courier Corporation, 1998. +[34] Karla L Hoffman, Manfred Padberg, Giovanni Rinaldi, et al. Traveling salesman problem. Encyclopedia of operations research and management science, 1:1573-1578, 2013. + +[35] Ludo F Gelders and Luk N Van Wassenhove. Production planning: a review. European Journal of Operational Research, 7(2):101-110, 1981. +[36] Paolo Toth and Daniele Vigo. The vehicle routing problem. SIAM, 2002. +[37] Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. Advances in Neural Information Processing Systems, 36:53728-53741, 2023. +[38] Robert B Ash. Information theory. Courier Corporation, 2012. +[39] Amir Zadeh, Paul Pu Liang, and Louis-Philippe Morency. Foundations of multimodal co-learning. Information Fusion, 64:188-193, 2020. +[40] Daniel Selsam, Matthew Lamm, Benedikt Bunz, Percy Liang, Leonardo de Moura, and David L Dill. Learning a sat solver from single-bit supervision. arXiv preprint arXiv:1802.03685, 2018. +[41] Egon Balas and Andrew Ho. Set covering algorithms using cutting planes, heuristics, and subgradient optimization: a computational study. In Combinatorial Optimization, pages 37-60. Springer, 1980. +[42] David Bergman, Andre A Cire, Willem-Jan Van Hoeve, and John Hooker. Decision diagrams for optimization, volume 1. Springer, 2016. +[43] Lara Scavuzzo, Feng Yang Chen, Didier Chételat, Maxime Gasse, Andrea Lodi, Neil Yorke-Smith, and Karen Aardal. Learning to branch with tree mdps. arXiv preprint arXiv:2205.11107, 2022. +[44] Maxime Gasse, Didier Chetelat, Nicola Ferroni, Laurent Charlin, and Andrea Lodi. Exact combinatorial optimization with graph convolutional neural networks. In H. Wallach, H. Larochelle, A. Beygelzimer, F. d'Alché-Buc, E. Fox, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 32. Curran Associates, Inc., 2019. +[45] Haoran Sun, Wenbo Chen, Hui Li, and Le Song. Improving learning to branch via reinforcement learning. In Learning Meets Combinatorial Algorithms at NeurIPS2020, 2020. +[46] Alper Atamtürk. On the facets of the mixed-integer knapsack polyhedron. Mathematical Programming, 98(1):145-175, 2003. +[47] Carla P Gomes, Willem-Jan van Hoeve, and Ashish Sabharwal. Connections in networks: A hybrid approach. In International Conference on Integration of Artificial Intelligence (AI) and Operations Research (OR) Techniques in Constraint Programming, pages 303-307. Springer, 2008. +[48] Simon Bowly, Quentin Cappart, Jonas Charfreitag, Laurent Charlin, Didier Chételat, Antonia Chmiela, Justin Dumouchelle, Maxime Gasse, Ambros Gleixner, Aleksandr M. Kazachkov, Elias B, Khalil, Pawel Lichocki, Andrea Lodi, Miles Lubin, Chris J, Maddison, Christopher Morris, Dimitri J, Papageorgiou, Augustin Parjadis, Sebastian Pokutta, Antoine Prouvost, Lara Scavuzzo, and Giulia Zarpellon. Machine learning for combinatorial optimization, 2021. URL https://www.ecole.ai/2021/ml4co-competition/. +[49] Tomas Balyo, Marijn Heule, Markus Iser, Matti Jarvisalo, and Martin Suda. Proceedings of sat competition 2023: Solver, benchmark and proof checker descriptions. 2023. +[50] Antoine Prouvost, Justin Dumouchelle, Lara Scavuzzo, Maxime Gasse, Didier Chételat, and Andrea Lodi. Ecole: A gym-like library for machine learning in combinatorial optimization solvers. arXiv preprint arXiv:2011.06069, 2020. +[51] Yang Li, Xinyan Chen, Wenxuan Guo, Xijun Li, Wanqian Luo, Junhua Huang, Hui-Ling Zhen, Mingxuan Yuan, and Junchi Yan. Hardsatgen: Understanding the difficulty of hard sat formula generation and a strong structure-hardness-aware baseline. In Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 4414-4425, 2023. + +[52] Ksenia Bestuzheva, Mathieu Besançon, Wei-Kun Chen, Antonia Chmiela, Tim Donkiewicz, Jasper van Doornmalen, Leon Eifler, Oliver Gaul, Gerald Gamrath, Ambros Gleixner, et al. The scip optimization suite 8.0. arXiv preprint arXiv:2112.08872, 2021. +[53] Vinod Nair, Sergey Bartunov, Felix Gimeno, Ingrid von Glehn, Pawel Lichocki, Ivan Lobov, Brendan O'Donoghue, Nicolas Sonnerat, Christian Tjandraatmadja, Pengming Wang, et al. Solving mixed integer programs using neural networks. arXiv preprint arXiv:2012.13349, 2020. +[54] He He, Hal Daume III, and Jason M Eisner. Learning to search in branch and bound algorithms. In Z. Ghahramani, M. Welling, C. Cortes, N. Lawrence, and K.Q. Weinberger, editors, Advances in Neural Information Processing Systems, volume 27. Curran Associates, Inc., 2014. +[55] Frank Hutter, Holger H Hoos, and Kevin Leyton-Brown. Automated configuration of mixed integer programming solvers. In International Conference on Integration of Artificial Intelligence (AI) and Operations Research (OR) Techniques in Constraint Programming, pages 186-202. Springer, 2010. + +# NeurIPS Paper Checklist + +# 1. Claims + +Question: Do the main claims made in the abstract and introduction accurately reflect the paper's contributions and scope? + +Answer: [Yes] + +Justification: We validated the effectiveness of the proposed method over 10 benchmarks compared with five baselines (please see Section 5.3 and Appendix G. Meanwhile, we provide the theoretical analysis to support our claim (please see Section 4.2 and corresponding proofs in Appendix A). + +Guidelines: + +- The answer NA means that the abstract and introduction do not include the claims made in the paper. +- The abstract and/or introduction should clearly state the claims made, including the contributions made in the paper and important assumptions and limitations. A No or NA answer to this question will not be perceived well by the reviewers. +- The claims made should match theoretical and experimental results, and reflect how much the results can be expected to generalize to other settings. +- It is fine to include aspirational goals as motivation as long as it is clear that these goals are not attained by the paper. + +# 2. Limitations + +Question: Does the paper discuss the limitations of the work performed by the authors? + +Answer: [Yes] + +Justification: we still discuss its limitation and future improvement room (see Appendix F). + +Guidelines: + +- The answer NA means that the paper has no limitation while the answer No means that the paper has limitations, but those are not discussed in the paper. +- The authors are encouraged to create a separate "Limitations" section in their paper. +- The paper should point out any strong assumptions and how robust the results are to violations of these assumptions (e.g., independence assumptions, noiseless settings, model well-specification, asymptotic approximations only holding locally). The authors should reflect on how these assumptions might be violated in practice and what the implications would be. +- The authors should reflect on the scope of the claims made, e.g., if the approach was only tested on a few datasets or with a few runs. In general, empirical results often depend on implicit assumptions, which should be articulated. +- The authors should reflect on the factors that influence the performance of the approach. For example, a facial recognition algorithm may perform poorly when image resolution is low or images are taken in low lighting. Or a speech-to-text system might not be used reliably to provide closed captions for online lectures because it fails to handle technical jargon. +- The authors should discuss the computational efficiency of the proposed algorithms and how they scale with dataset size. +- If applicable, the authors should discuss possible limitations of their approach to address problems of privacy and fairness. +- While the authors might fear that complete honesty about limitations might be used by reviewers as grounds for rejection, a worse outcome might be that reviewers discover limitations that aren't acknowledged in the paper. The authors should use their best judgment and recognize that individual actions in favor of transparency play an important role in developing norms that preserve the integrity of the community. Reviewers will be specifically instructed to not penalize honesty concerning limitations. + +# 3. Theory assumptions and proofs + +Question: For each theoretical result, does the paper provide the full set of assumptions and a complete (and correct) proof? + +# Answer: [Yes] + +Justification: In Section 4.2, we state all definitions, theorem and remarks. Besides, we provide the formal proofs to above theorems in Appendix A. + +# Guidelines: + +- The answer NA means that the paper does not include theoretical results. +- All the theorems, formulas, and proofs in the paper should be numbered and cross-referenced. +- All assumptions should be clearly stated or referenced in the statement of any theorems. +- The proofs can either appear in the main paper or the supplemental material, but if they appear in the supplemental material, the authors are encouraged to provide a short proof sketch to provide intuition. +- Inversely, any informal proof provided in the core of the paper should be complemented by formal proofs provided in appendix or supplemental material. +- Theorems and Lemmas that the proof relies upon should be properly referenced. + +# 4. Experimental result reproducibility + +Question: Does the paper fully disclose all the information needed to reproduce the main experimental results of the paper to the extent that it affects the main claims and/or conclusions of the paper (regardless of whether the code and data are provided or not)? + +# Answer: [Yes] + +Justification: We present the methodology of our proposed framework in Section 4.1 and then disclose all information related to model implementation, used dataset (all used dataset are publicly available and can be found in referred papers), metrics of interest and compared baselines in Section C.2. Besides, we provide all the details of above experimental elements. Please see Appendix B for model implementation details, Appendix D for description of used dataset, Appendix E for metrics of interest, Appendix C.1 for solver adoption, Appendix C for details of compared baselines and Appendix L for example prompts used in related experiments. + +# Guidelines: + +- The answer NA means that the paper does not include experiments. +- If the paper includes experiments, a No answer to this question will not be perceived well by the reviewers: Making the paper reproducible is important, regardless of whether the code and data are provided or not. +- If the contribution is a dataset and/or model, the authors should describe the steps taken to make their results reproducible or verifiable. +- Depending on the contribution, reproducibility can be accomplished in various ways. For example, if the contribution is a novel architecture, describing the architecture fully might suffice, or if the contribution is a specific model and empirical evaluation, it may be necessary to either make it possible for others to replicate the model with the same dataset, or provide access to the model. In general, releasing code and data is often one good way to accomplish this, but reproducibility can also be provided via detailed instructions for how to replicate the results, access to a hosted model (e.g., in the case of a large language model), releasing of a model checkpoint, or other means that are appropriate to the research performed. +- While NeurIPS does not require releasing code, the conference does require all submissions to provide some reasonable avenue for reproducibility, which may depend on the nature of the contribution. For example +(a) If the contribution is primarily a new algorithm, the paper should make it clear how to reproduce that algorithm. +(b) If the contribution is primarily a new model architecture, the paper should describe the architecture clearly and fully. +(c) If the contribution is a new model (e.g., a large language model), then there should either be a way to access this model for reproducing the results or a way to reproduce the model (e.g., with an open-source dataset or instructions for how to construct the dataset). + +(d) We recognize that reproducibility may be tricky in some cases, in which case authors are welcome to describe the particular way they provide for reproducibility. In the case of closed-source models, it may be that access to the model is limited in some way (e.g., to registered users), but it should be possible for other researchers to have some path to reproducing or verifying the results. + +# 5. Open access to data and code + +Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material? + +Answer: [Yes] + +Justification: The code, data and learned model will be publicly available upon the acceptance of the paper. + +Guidelines: + +- The answer NA means that paper does not include experiments requiring code. +- Please see the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details. +- While we encourage the release of code and data, we understand that this might not be possible, so "No" is an acceptable answer. Papers cannot be rejected simply for not including code, unless this is central to the contribution (e.g., for a new open-source benchmark). +- The instructions should contain the exact command and environment needed to run to reproduce the results. See the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details. +- The authors should provide instructions on data access and preparation, including how to access the raw data, preprocessed data, intermediate data, and generated data, etc. +- The authors should provide scripts to reproduce all experimental results for the new proposed method and baselines. If only a subset of experiments are reproducible, they should state which ones are omitted from the script and why. +- At submission time, to preserve anonymity, the authors should release anonymized versions (if applicable). +- Providing as much information as possible in supplemental material (appended to the paper) is recommended, but including URLs to data and code is permitted. + +# 6. Experimental setting/details + +Question: Does the paper specify all the training and test details (e.g., data splits, hyperparameters, how they were chosen, type of optimizer, etc.) necessary to understand the results? + +Answer: [Yes] + +Justification: We disclose all the details, including statistics of dataset, implementation of the proposed composite model, hyperparameters, etc. in Appendix C.2 and Appendix B. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The experimental setting should be presented in the core of the paper to a level of detail that is necessary to appreciate the results and make sense of them. +- The full details can be provided either with the code, in appendix, or as supplemental material. + +# 7. Experiment statistical significance + +Question: Does the paper report error bars suitably and correctly defined or other appropriate information about the statistical significance of the experiments? + +Answer: [Yes] + +Justification: Please see the result in Appendix G. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The authors should answer "Yes" if the results are accompanied by error bars, confidence intervals, or statistical significance tests, at least for the experiments that support the main claims of the paper. +- The factors of variability that the error bars are capturing should be clearly stated (for example, train/test split, initialization, random drawing of some parameter, or overall run with given experimental conditions). +- The method for calculating the error bars should be explained (closed form formula, call to a library function, bootstrap, etc.) +- The assumptions made should be given (e.g., Normally distributed errors). +- It should be clear whether the error bar is the standard deviation or the standard error of the mean. +- It is OK to report 1-sigma error bars, but one should state it. The authors should preferably report a 2-sigma error bar than state that they have a $96\%$ CI, if the hypothesis of Normality of errors is not verified. +- For asymmetric distributions, the authors should be careful not to show in tables or figures symmetric error bars that would yield results that are out of range (e.g. negative error rates). +- If error bars are reported in tables or plots, The authors should explain in the text how they were calculated and reference the corresponding figures or tables in the text. + +# 8. Experiments compute resources + +Question: For each experiment, does the paper provide sufficient information on the computer resources (type of compute workers, memory, time of execution) needed to reproduce the experiments? + +Answer: [Yes] + +Justification: All experiments are conducted on hardware platform with dual AMD EPYC 9534 64-core processors @ 2.45GHz and two NVIDIA H800 80GB GPUs connected via PCIe. More training specifics and data curation are detailed in Appendix B.1 and B.2 respectively. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The paper should indicate the type of compute workers CPU or GPU, internal cluster, or cloud provider, including relevant memory and storage. +- The paper should provide the amount of compute required for each of the individual experimental runs as well as estimate the total compute. +- The paper should disclose whether the full research project required more compute than the experiments reported in the paper (e.g., preliminary or failed experiments that didn't make it into the paper). + +# 9. Code of ethics + +Question: Does the research conducted in the paper conform, in every respect, with the NeurIPS Code of Ethics https://neurips.cc/public/EthicsGuidelines? + +Answer: [Yes] + +Justification: We confirm that the research conducted in the paper conform, in every respect, with the NeurIPS Code of Ethics. + +Guidelines: + +- The answer NA means that the authors have not reviewed the NeurIPS Code of Ethics. +- If the authors answer No, they should explain the special circumstances that require a deviation from the Code of Ethics. +- The authors should make sure to preserve anonymity (e.g., if there is a special consideration due to laws or regulations in their jurisdiction). + +# 10. Broader impacts + +Question: Does the paper discuss both potential positive societal impacts and negative societal impacts of the work performed? + +Answer: [NA] + +Justification: This paper does not explicitly show the positive/negative societal impacts. But we still discuss its limitation and future improvement room (see Appendix F). + +Guidelines: + +- The answer NA means that there is no societal impact of the work performed. +- If the authors answer NA or No, they should explain why their work has no societal impact or why the paper does not address societal impact. +- Examples of negative societal impacts include potential malicious or unintended uses (e.g., disinformation, generating fake profiles, surveillance), fairness considerations (e.g., deployment of technologies that could make decisions that unfairly impact specific groups), privacy considerations, and security considerations. +- The conference expects that many papers will be foundational research and not tied to particular applications, let alone deployments. However, if there is a direct path to any negative applications, the authors should point it out. For example, it is legitimate to point out that an improvement in the quality of generative models could be used to generate deepfakes for disinformation. On the other hand, it is not needed to point out that a generic algorithm for optimizing neural networks could enable people to train models that generate Deepfakes faster. +- The authors should consider possible harms that could arise when the technology is being used as intended and functioning correctly, harms that could arise when the technology is being used as intended but gives incorrect results, and harms following from (intentional or unintentional) misuse of the technology. +- If there are negative societal impacts, the authors could also discuss possible mitigation strategies (e.g., gated release of models, providing defenses in addition to attacks, mechanisms for monitoring misuse, mechanisms to monitor how a system learns from feedback over time, improving the efficiency and accessibility of ML). + +# 11. Safeguards + +Question: Does the paper describe safeguards that have been put in place for responsible release of data or models that have a high risk for misuse (e.g., pretrained language models, image generators, or scraped datasets)? + +Answer: [NA] + +Justification: The work poses no such risks. + +Guidelines: + +- The answer NA means that the paper poses no such risks. +- Released models that have a high risk for misuse or dual-use should be released with necessary safeguards to allow for controlled use of the model, for example by requiring that users adhere to usage guidelines or restrictions to access the model or implementing safety filters. +- Datasets that have been scraped from the Internet could pose safety risks. The authors should describe how they avoided releasing unsafe images. +- We recognize that providing effective safeguards is challenging, and many papers do not require this, but we encourage authors to take this into account and make a best faith effort. + +# 12. Licenses for existing assets + +Question: Are the creators or original owners of assets (e.g., code, data, models), used in the paper, properly credited and are the license and terms of use explicitly mentioned and properly respected? + +Answer: [Yes] + +Justification: We confirm that the baselines and datasets used in the paper all are publicly available. + +Guidelines: + +- The answer NA means that the paper does not use existing assets. +- The authors should cite the original paper that produced the code package or dataset. +- The authors should state which version of the asset is used and, if possible, include a URL. +- The name of the license (e.g., CC-BY 4.0) should be included for each asset. +- For scraped data from a particular source (e.g., website), the copyright and terms of service of that source should be provided. +- If assets are released, the license, copyright information, and terms of use in the package should be provided. For popular datasets, paperswithcode.com/datasets has curated licenses for some datasets. Their licensing guide can help determine the license of a dataset. +- For existing datasets that are re-packaged, both the original license and the license of the derived asset (if it has changed) should be provided. +- If this information is not available online, the authors are encouraged to reach out to the asset's creators. + +# 13. New assets + +Question: Are new assets introduced in the paper well documented and is the documentation provided alongside the assets? + +Answer: [NA] + +Justification: The paper does not release new assets. + +Guidelines: + +- The answer NA means that the paper does not release new assets. +- Researchers should communicate the details of the dataset/code/model as part of their submissions via structured templates. This includes details about training, license, limitations, etc. +- The paper should discuss whether and how consent was obtained from people whose asset is used. +- At submission time, remember to anonymize your assets (if applicable). You can either create an anonymized URL or include an anonymized zip file. + +# 14. Crowdsourcing and research with human subjects + +Question: For crowdsourcing experiments and research with human subjects, does the paper include the full text of instructions given to participants and screenshots, if applicable, as well as details about compensation (if any)? + +Answer: [NA] + +Justification: The paper does not involve crowdsourcing nor research with human subjects. + +Guidelines: + +- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. +- Including this information in the supplemental material is fine, but if the main contribution of the paper involves human subjects, then as much detail as possible should be included in the main paper. +- According to the NeurIPS Code of Ethics, workers involved in data collection, curation, or other labor should be paid at least the minimum wage in the country of the data collector. + +# 15. Institutional review board (IRB) approvals or equivalent for research with human subjects + +Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals (or an equivalent approval/review based on the requirements of your country or institution) were obtained? + +Answer: [NA] + +Justification: The paper does not involve crowdsourcing nor research with human subjects. + +# Guidelines: + +- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. +- Depending on the country in which research is conducted, IRB approval (or equivalent) may be required for any human subjects research. If you obtained IRB approval, you should clearly state this in the paper. +- We recognize that the procedures for this may vary significantly between institutions and locations, and we expect authors to adhere to the NeurIPS Code of Ethics and the guidelines for their institution. +- For initial submissions, do not include any information that would break anonymity (if applicable), such as the institution conducting the review. + +# 16. Declaration of LLM usage + +Question: Does the paper describe the usage of LLMs if it is an important, original, or non-standard component of the core methods in this research? Note that if the LLM is used only for writing, editing, or formatting purposes and does not impact the core methodology, scientific rigorousness, or originality of the research, declaration is not required. + +Answer: [NA] + +Justification: The core method development in this research does not involve LLMs as any important, original, or non-standard components. + +# Guidelines: + +- The answer NA means that the core method development in this research does not involve LLMs as any important, original, or non-standard components. +- Please refer to our LLM policy (https://neurips.cc/Conferences/2025/LLM) for what should or should not be described. + +# A Proofs + +Theorem 1. Given a prior dataset $\mathcal{D}$ whose data samples comprises $M$ types of prior $\mathcal{C} = \{\mathbf{C}_1, \dots, \mathbf{C}_M\}$ , the distinct priors follow respective true distribution $p(\mathbf{C}_i | \mathbf{w})$ . Let $\mathbf{c}_i$ be a sample drawn from the distribution, i.e., $\mathbf{c}_i \sim p(\mathbf{C}_i | \mathbf{w})$ . A generative model with an additional type of prior will not lower the upper bound of model performance $\sup(\mathcal{P})$ . + +Proof. Assume that there is a generative model, which is associated with a set of prior $\mathcal{C}_K = \{\mathbf{C}_1,\dots,\mathbf{C}_K\}$ . Its upper bound of model performance is denoted by $\sup (\mathcal{P}_{\mathcal{C}_K})$ . According to Definition 1, $\sup (\mathcal{P}_{\mathcal{C}_K})$ can be expanded as + +$$ +\begin{array}{l} \sup (\mathcal {P} _ {\mathcal {C} _ {K}}) = \sum_ {\mathbf {C} \in \mathcal {C} _ {K}} \sum_ {\mathbf {c} \in \mathbf {C}} p (\mathbf {c}) \max _ {\mathbf {w}} p (\mathbf {w} | \mathbf {c}) \Phi (\mathbf {w}) \\ = \sum_ {\mathbf {C} \in \mathcal {C} _ {K}} \sum_ {\mathbf {c} \in \mathbf {C}} p (\mathbf {c}) \max _ {\mathbf {w}} \sum_ {\mathbf {c} _ {i} \in \mathbf {C} _ {i}} p (\mathbf {c} _ {i} | \mathbf {c}) p (\mathbf {w} | \mathbf {c}, \mathbf {c} _ {i}) \Phi (\mathbf {w}) \\ \leq \sum_ {\mathbf {C} \in \mathcal {C} _ {K}} \sum_ {\mathbf {c} \in \mathbf {C}} \sum_ {\mathbf {c} _ {i} \in \mathbf {C} _ {i}} p (\mathbf {c}) p \left(\mathbf {c} _ {i} \mid \mathbf {c}\right) \max _ {\mathbf {w}} p \left(\mathbf {w} \mid \mathbf {c}, \mathbf {c} _ {i}\right) \Phi (\mathbf {w}) \tag {7} \\ = \sum_ {\mathbf {C} \in \mathcal {C} _ {K}} \sum_ {\mathbf {c} \in \mathbf {C}} \sum_ {\mathbf {c} _ {i} \in \mathbf {C} _ {i}} p (\mathbf {c} _ {i}, \mathbf {c}) \max _ {\mathbf {w}} p (\mathbf {w} | \mathbf {c}, \mathbf {c} _ {i}) \Phi (\mathbf {w}) \\ = \sum_ {\mathbf {C} \in \mathcal {C} _ {K} \bigcup \{\mathbf {C} _ {i} \}} \sum_ {\mathbf {c} \in \mathbf {C}} p (\mathbf {c} _ {i}, \mathbf {c}) \max _ {\mathbf {w}} p (\mathbf {w} | \mathbf {c}, \mathbf {c} _ {i}) \Phi (\mathbf {w}) \\ = \sup \left(\mathcal {P} _ {\mathcal {C} _ {K} \cup \{\mathbf {C} _ {i} \}}\right) \\ \end{array} +$$ + +Above proof indicates that for a given generative model with any additional modal prior $\mathbf{C}_i$ , its performance upper bound $\sup (\mathcal{P}_{\mathcal{C}_K} \cup \{\mathbf{C}_i\})$ is greater or equal to its original performance upper bound $\sup (\mathcal{P}_{\mathcal{C}_K})$ . + +Theorem 2. If prior $\mathbf{C}_{pe}$ is a performance-enhancing prior, a generative model neglecting prior $\mathbf{C}_{pe}$ will decrease the upper bound of model performance. It can be expressed as + +$$ +\sup \left(\mathcal {P} _ {\mathcal {C} \backslash \left\{\mathbf {C} _ {p e} \right\}}\right) < \sup \left(\mathcal {P} _ {\mathcal {C}}\right) \tag {6} +$$ + +Proof. If prior $\mathbf{C}_{pe}$ is a performance-enhancing prior, according to Definition 2, we have + +$$ +\begin{array}{l} \max _ {\mathbf {w}} p (\mathbf {w} | \mathbf {c}) \Phi (\mathbf {w}) _ {\mathbf {c} \in \mathcal {C} \backslash \{\mathbf {C} _ {p e} \}} < \max _ {w} p (\mathbf {w} | \mathbf {c}) \Phi (\mathbf {w}) _ {\mathbf {c} \in \mathcal {C}} \\ = \sum_ {\mathbf {c} _ {p e} \in \mathbf {C} _ {p e}} p (\mathbf {c} _ {p e} | \mathbf {c}) \max _ {\mathbf {w}} p (\mathbf {w} | \mathbf {c}) \Phi (\mathbf {w}) _ {\mathbf {c} \in \mathcal {C}} \tag {8} \\ \end{array} +$$ + +Then, according to Definition 1, $\sup (\mathcal{P}_{\mathcal{C}\setminus \{\mathbf{C}_{pe}\}})$ can be expanded as + +$$ +\sup \left(\mathcal {P} _ {\mathcal {C} \backslash \{\mathbf {C} _ {p e} \}}\right) = \sum_ {\mathbf {C} \in \mathcal {C} \backslash \{\mathbf {C} _ {p e} \}} \sum_ {\mathbf {c} \in \mathbf {C}} p (\mathbf {c}) \max _ {\mathbf {w}} p (\mathbf {w} | \mathbf {c}) \Phi (\mathbf {w}) \tag {9} +$$ + +Substitute Eq.(8) into Eq.(9) to derive + +$$ +\begin{array}{l} \sup (\mathcal {P} _ {\mathcal {C} \backslash \{\mathbf {C} _ {p e} \}}) = \sum_ {\mathbf {C} \in \mathcal {C} \backslash \{\mathbf {C} _ {p e} \}} \sum_ {\mathbf {c} \in \mathbf {C}} p (\mathbf {c}) \max _ {\mathbf {w}} p (\mathbf {w} | \mathbf {c}) \Phi (\mathbf {w}) \\ < \sum_ {\mathbf {C} \in \mathcal {C} \backslash \{\mathbf {C} _ {p e} \}} \sum_ {\mathbf {c} \in \mathbf {C}} p (\mathbf {c}) \sum_ {\mathbf {c} _ {p e} \in \mathbf {C} _ {p e}} p (\mathbf {c} _ {p e} | \mathbf {c}) \max _ {\mathbf {w}} p (\mathbf {w} | \mathbf {c}) \Phi (\mathbf {w}) \\ = \sum_ {\mathbf {C} \in \mathcal {C} \backslash \left\{\mathbf {C} _ {p e} \right\}} \sum_ {\mathbf {c} \in \mathbf {C}} \sum_ {\mathbf {c} _ {p e} \in \mathbf {C} _ {p e}} p (\mathbf {c}) p (\mathbf {c} _ {p e} | \mathbf {c}) \max _ {\mathbf {w}} p (\mathbf {w} | \mathbf {c}) \Phi (\mathbf {w}) \tag {10} \\ = \sum_ {\mathbf {C} \in \mathcal {C} \backslash \{\mathbf {C} _ {p e} \}} \sum_ {\mathbf {c} \in \mathbf {C}} \sum_ {\mathbf {c} _ {p e} \in \mathbf {C} _ {p e}} p (\mathbf {c} _ {p e}, \mathbf {c}) \max _ {\mathbf {w}} p (\mathbf {w} | \mathbf {c}) \Phi (\mathbf {w}) \\ = \sum_ {\mathbf {C} \in \mathcal {C}} \sum_ {\mathbf {c} \in \mathbf {C}} p (\mathbf {c}) \max _ {\mathbf {w}} p (\mathbf {w} | \mathbf {c}) \Phi (\mathbf {w}) \\ = \sup (\mathcal {P} _ {\mathcal {C}}) \\ \end{array} +$$ + +# B Implementation Details + +# B.1 Combinatorial Structure Extraction + +Data Representation of MILP problem. A mixed-integer linear programming (MILP) problem is formally defined as: + +$$ +\min _ {\boldsymbol {x} \in \mathbb {R} ^ {n}} \boldsymbol {w} ^ {\top} \boldsymbol {x}, \quad \text {s . t .} \boldsymbol {A} \boldsymbol {x} \leq \boldsymbol {b}, \boldsymbol {l} \leq \boldsymbol {x} \leq \boldsymbol {u}, x _ {j} \in \mathbb {Z}, \forall j \in \mathbb {I}, \tag {11} +$$ + +where $\pmb{w} \in \mathbb{R}^n$ , $\pmb{A} \in \mathbb{R}^{m \times n}$ , $\pmb{b} \in \mathbb{R}^m$ , $\pmb{l} \in (\mathbb{R} \cup \{-\infty\})^n$ , $\pmb{u} \in (\mathbb{R} \cup \{+\infty\})^n$ , and the index set $\mathbb{I} \subset \{1, 2, \dots, n\}$ specifies integer-constrained variables. + +To encode MILP instances, we design a bipartite graph $\mathcal{G} = (\mathcal{C} \cup \mathcal{V}, \mathcal{E})$ with constraint-variable interactions. The constraint node set $\mathcal{C} = \{c_1, \dots, c_m\}$ corresponds to rows of $A\mathbf{x} \leq \mathbf{b}$ , where each node $c_i$ is associated with a 1D feature vector $c_i = (b_i)$ representing its right-hand side value. The variable node set $\mathcal{V} = \{v_1, \dots, v_n\}$ represents decision variables, each equipped with a 9D feature vector $v_j$ containing the objective coefficient $w_j$ , variable type indicator (integer/continuous), and bound parameters $l_j, u_j$ . Edges $\mathcal{E} = \{e_{ij}\}$ connect constraint $c_i$ to variable $v_j$ iff $a_{ij} \neq 0$ , with edge features $e_{ij} = (a_{ij})$ encoding the constraint coefficients. + +Furthermore, a mixed-integer linear programming (MILP) instance is encoded as a weighted bipartite graph with feature matrices $\mathbf{G} = (\mathbf{C},\mathbf{V},\mathbf{E})$ , where $\mathbf{C},\mathbf{V}$ , and $\mathbf{E}$ aggregate constraint node features $\mathbf{c}_i$ , variable node features $\mathbf{v}_j$ , and edge features $\mathbf{e}_{ij}$ , respectively. The full specification of these features is summarized in Table 2, which preserves all structural and numerical information of the original MILP problem. Following standard practice, we utilize the observation function from Ecole [50] to generate these bipartite graph representations from MILP instances. + +Table 2: Description of the constraint, variable, and edge features in our bipartite graph representation for MILP instance. + +
TensorFeatureDescription
CConstraint coefficientAverage of all coefficients in the constraint.
Constraint degreeDegree of constraint nodes.
BiasNormalized right-hand-side of the constraint.
VObjectiveNormalized objective coefficient.
Variable coefficientAverage variable coefficient in all constraints.
Variable degreeDegree of the variable node in the bipartite graph representation.
Maximum variable coefficientMaximum variable coefficient in all constraints.
Minimum variable coefficientMinimum variable coefficient in all constraints.
ECoefficientConstraint coefficient.
+ +Data Representation of SAT problem. A Boolean satisfiability (SAT) problem consists of variables $x_{i}$ and logical operators $\land, \lor$ , and $\neg$ . A formula is satisfiable if there exists a variable assignment that makes all clauses evaluate to true. Following [51], we focus on formulas in conjunctive normal form (CNF) – conjunctions $(\land)$ of clauses, where each clause is a disjunction $(\lor)$ of literals (variables $x_{i}$ or their negations $\neg x_{i}$ ). Any SAT formula can be converted to an equisatisfiable CNF formula in linear time. For example, $(x_{1} \lor \neg x_{2}) \land (x_{2} \lor \neg x_{3})$ represents a CNF formula with two clauses. + +We utilize the Variable-Clause Graph (VCG) to represent SAT formulas. For a SAT formula, the VCG is constructed with nodes representing literals and clauses, and edges indicating the inclusion of a literal in a clause. The bipartite structure of VCGs ensures a one-to-one correspondence between CNF formulas and their graph representations. Formally, a bipartite graph $\mathcal{G} = (\mathcal{V}_1 \cup \mathcal{V}_2, \mathcal{E})$ is defined by its vertex set $\mathcal{V}_1 \cup \mathcal{V}_2 = \{v_1, \dots, v_n\}$ and edge set $\mathcal{E} \subseteq \{(v_i, v_j) | v_i, v_j \in \mathcal{V}\}$ . The vertex set is partitioned into two disjoint subsets $\mathcal{V}_1$ and $\mathcal{V}_2$ , with edges restricted to connections between nodes in distinct partitions: $\mathcal{E} \subseteq \{(v_i, v_j) | v_i \in \mathcal{V}_1, v_j \in \mathcal{V}_2\}$ . In the context of VCGs, a CNF formula with $n$ + +![](images/054ddff1ac6812756f21156565bc4c1d31e65a66a42d85f5a0f9eb2edbc5063f.jpg) +(a) Training Loss + +![](images/02b1a42c9695a95b6ef7374ffb62c38fe2b500996d6c58a41a1c7d09bc782867.jpg) +(b) Prediction Accuracy +Figure 6: The convergence curve of training GNN for SAT domain. + +literals and $m$ clauses induces a bipartitioned graph where $\mathcal{V}_1 = \{l_1, \dots, l_n\}$ denotes the literal nodes and $\mathcal{V}_2 = \{c_1, \dots, c_m\}$ represents the clause nodes. + +GNN Structure. The process of using GNN to extract combinatorial optimization problem can be formulated as: + +$$ +\mathbf {v} _ {i} ^ {(k + 1)} \leftarrow \mathbf {f} _ {\mathbf {V}} \left(\mathbf {v} _ {i} ^ {(k)}, \sum_ {j, i \neq j} ^ {(i, j) \in \mathcal {E}} \mathbf {g} _ {\mathbf {V}} \left(\mathbf {v} _ {i} ^ {(k)}, \mathbf {v} _ {j} ^ {(k)}, \mathbf {e} _ {i j}\right)\right), \quad (k = 0, 1, \dots , K - 1) \tag {12} +$$ + +$$ +\boldsymbol {h} _ {q} = \operatorname {P o o l} \left(\left\{\mathbf {v} _ {i} ^ {(K)} \right\}\right), \tag {13} +$$ + +where $\mathbf{f}_{\mathbf{V}}$ and $\mathbf{g}_{\mathbf{V}}$ are perceptrons for node representation; $K$ represents the total number of times that we perform the convolution; $Pool$ denotes pooling function that aggregates the embedding of each node in the graph, obtaining the structure embedding $h_q \in \mathbb{R}^d$ of the instance $q$ . We denote the parameters of GNN as $\theta_G$ . + +GNN Training. The loss function w.r.t. $\theta_{G}$ is given below: + +$$ +\mathcal {L} \left(\theta_ {G}\right) = - \frac {1}{N} \sum_ {i = 1} ^ {N} \sum_ {c = 1} ^ {C} y _ {i, c} \log p _ {\theta_ {G}} (c \mid q _ {i}), \tag {14} +$$ + +$$ +p _ {\theta_ {G}} (c \mid q _ {i}) = \frac {\exp \left(\boldsymbol {W} _ {c} ^ {T} \boldsymbol {h} _ {q _ {i}} + \boldsymbol {b} _ {c}\right)}{\sum_ {j = 1} ^ {C} \exp \left(\boldsymbol {W} _ {j} ^ {T} \boldsymbol {h} _ {q _ {i}} + \boldsymbol {b} _ {j}\right)}, \tag {15} +$$ + +where $N$ is the total number of CO problems in the training procedure; $C$ denotes the number of classes; $y_{i,c} \in \{0,1\}$ is the ground-truth label of problem instance $q_i$ ; and $\mathbf{W}_j \in \mathbb{R}^d$ , $\mathbf{b}_j \in \mathbb{R}$ , $(j = 1,\dots,C)$ is the parameters of final classifying layer, which are also part of $\theta_G$ . + +Implementation & Training Details. Following the training protocol outlined above, we implement separate GNN models for the SAT and MILP domains using the dataset described in Appendix D. Our GNN architecture consists of three convolutional layers ( $K = 3$ ) followed by a global mean pooling operation. We leverage the graph convolution operator from the torch_geometric library, with node embedding dimensions of 16, 32, and 64 for successive layers3. To capture global graph structure, we apply mean pooling to the final convolutional layer's node embeddings. For classification, a softmax layer processes the pooled representation to produce final predictions. The models are trained using the loss function defined in Eq.(14) with AdamW optimizer and cosine decay learning rate. Figure 6 illustrates the training convergence for SAT instance classification (5-way classification task). Upon model convergence, we compute structural prior representations for combinatorial optimization problem instances by processing their bipartite graph representations through the above GNN. The final convolutional layer's output embeddings are extracted as the structural prior for each instance in the target domain. + +# B.2 Structure-Aware Code Generation + +Data Curation. We begin by assembling a curated collection of mathematical models for CO problems, which are directly compatible with corresponding CO solvers, alongside their natural language descriptions. Subsequently, for each CO problem, we compile a prompt that integrates the natural language description with specific code generation requirements. For instance, the code generation requirements may include details such as the function name, input/output parameters, expected function behavior, relevant background knowledge, etc. Illustrative examples for SAT and MILP domain are provided in Appendix L. This prompt is then fed into an LLM to generate a code snippet. Note that to maximize the diversity of collected code snippets, we employ multiple LLM queries per prompt under high temperature settings to sample distinct candidate implementations. Each generated code snippet is evaluated by embedding it into the target solver and solving the corresponding CO problem, thereby obtaining performance metrics for the code snippet. Note that based on the principles of data curation, we collect the needed data via querying Qwen2.5-Coder-7B-Instructor same as the model adopted in the following post training procedure. + +Post Training. Specifically, we first curate the collected data by processing each CO problem $Q_{i}$ with its associated prompt $x_{i}$ and corresponding multiple generated code snippets $y_{i,j}(j = 1,\dots,M)$ . These code snippets are systematically ranked based on previously obtained performance metrics to establish quality ordering. The highest-performing code-prompt pairs are selected to form the SFT dataset $\mathcal{D}_{SFT} = \{(x_i,y_i^*)\}_{i = 1}^N$ . Additionally, by leveraging pairwise comparisons extracted from the established ranking hierarchy, we derive the preference dataset $\mathcal{D}_{DPO} = \{(x_i,y_w,y_l)\}_{i = 1}^N$ where $y_{w},y_{l}$ are preferred/dispreferred code snippets. The training process can be formulated as follows: + +$$ +\mathcal {L} _ {\mathrm {D P O}} = - \mathbb {E} _ {(x, y _ {w}, y _ {l}) \sim \mathcal {D} _ {D P O}} \left[ \log \sigma \left(\beta \ln \frac {\pi_ {\theta_ {L}} (y _ {w} | x)}{\pi_ {\operatorname {r e f}} (y _ {w} | x)} - \beta \ln \frac {\pi_ {\theta_ {L}} (y _ {l} | x)}{\pi_ {\operatorname {r e f}} (y _ {l} | x)}\right) \right], \tag {16} +$$ + +where $\sigma$ is the Sigmoid function; $\beta$ is the hyperparameter that governs the trade-off between reward maximization and KL divergence minimization; $\pi_{\mathrm{ref}}$ is the reference model trained on $D_{SFT}$ . Note that both $\pi_{\mathrm{ref}}$ and $\pi_{\theta_L}$ updates only the parameter $\theta_L$ of the composite model. Through above posttraining, we obtain a generative model capable of structure-aware code generation that simultaneously respects combinatorial optimization problems' inherent topological constraints and solver-specific syntactic requirements. Note that based on above principle of data curation and post-training, we collect 8k and 4k post-training instances for MILP and SAT instances respectively. + +Implementation & Training Details. To implement the proposed composite model, we first design a structure-prior-aware forward propagation mechanism and corresponding adapted inference framework based on the Qwen2.5-Coder-7B-Instructor $^4$ model via the Transformers library. + +Structure-Prior-Aware Forward Propagation Mechanism. Specifically, we process the input prompt through the tokenizer and the model's embedding layer to obtain input_embedding, which are then merged with structural feature vectors of combinatorial optimization problems extracted by the previous graph neural network. + +First, we align the dimensionality of the combinatorial optimization problem feature vector $\pmb{h}_q \in \mathbb{R}^d$ with the hidden layer dimensions of the large language model via zero-padding. Given the text embedding shape $\text{embeds} \in \mathbb{R}^{B \times S \times H}$ , where $B$ is the batch size, $S$ is the sequence length (number of tokens), and $H$ is the hidden dimension, the dimension-adapted feature vector is obtained as: + +$$ +\mathbf {H} _ {q} = \operatorname {Z e r o P a d d i n g} \left(\boldsymbol {h} _ {q} \oplus \boldsymbol {0} ^ {(H - d)}\right) \in \mathbb {R} ^ {B \times H} \tag {17} +$$ + +where ZeroPadding denotes zero-padding, and $\mathbf{H}_q$ represents the padded graph structural feature vector. Subsequently, we fuse text and structural features between the embedding layer and decoder layer by prepending the graph feature vector to the text embedding sequence, forming a hybrid input: + +$$ +\mathcal {E} \left(\text {e m b e d s}, \mathbf {H} _ {q}\right) = \left[ \mathrm {C L S} \right] \oplus \mathbf {H} _ {q} \oplus \text {e m b e d s} [ 1: ] \in \mathbb {R} ^ {B \times (1 + S) \times H} \tag {18} +$$ + +Here, $\text{embeds}$ is the input_embedding obtained from processing the textual prompt via the tokenizer and embedding layer, enabling the self-attention mechanism to jointly model textual semantics and graph structural features. A mask of all True values is constructed and merged with the attention_mask to match the shape of the combined input_embedding, ensuring $\mathbf{H}_q$ participates in attention computation. Let the original attention_mask shape be $M = \{0,1\}^{B\times S}$ ; the merged attention_mask becomes: + +$$ +M _ {i n} = \left[ M _ {0: 1}; \mathbf {1} ^ {B}; M _ {1: S} \right] \in \{0, 1 \} ^ {B \times (1 + S)} \tag {19} +$$ + +![](images/c1ea2823d4d000f0eafb754155daaecf9a521dc5cba92996403af4b9d09d21b4.jpg) +(a) SFT Loss + +![](images/37d95dd71577994399c49c6c4b7c99f8e1d35670b2231f2a496a661437a9c26d.jpg) +(b) DPO Loss + +![](images/6903c50657aacc2eda6de490e92b9a4b69a11f7a8a388c6f941844e07232cd51.jpg) +(c) DPO Reward Accuracies +Figure 7: The convergence curve of post-training composite model for SAT domain. + +The resulting input_embeds and attention_mask are then passed to the decoder layers and subsequent model structures for standard forward propagation. + +$\Theta$ Parameter-Efficient Finetuning. We adopt LoRA for parameter-efficient fine-tuning on an autoregressive language model task. Key hyperparameters are configured as: rank dimension $r = 16$ controlling the latent dimension of low-rank adaptation matrices; scaling factor lora_alpha $= 32$ for output normalization; dropout probability 0.05 for regularization; trainable low-rank adaptation layers injected exclusively into query $(q\_proj)$ and value $(v\_proj)$ projection submodules of the Transformer architecture while keeping other parameters frozen; bias parameters set to "none" to preserve original model biases. The adapted LLM is trained with AdamW optimizer and cosine decay learning rate. An illustrative training curve of the model for SAT domain is given in Figure 7. + +Adapted Inference Framework. Based on this modified architecture, we further adapt the LLM inference framework with structure-prior feature vector integration. During each autoregressive forward pass, logits are obtained via the feature-enhanced forward propagation, ensuring persistent influence of the structure-prior feature vector throughout the generation process rather than only affecting the initial output. We employ a hybrid sampling strategy with Top-k set to 20, Top-p to 0.8, and repetition penalty to 1.0. + +# C Baseline Details + +# C.1 Solver Adoption + +MILP Solver. In all experiments related to MILP domain, we employ SCIP 8.0.0 [52] as the MILP solver backend—a leading open-source solver widely adopted in machine learning research for combinatorial optimization [44, 53]. To ensure reproducibility and fair comparisons, all SCIP parameters remain at default values. We retain the solver's advanced features (e.g., presolve, heuristics), aligning our experimental setup with real-world applications. + +SAT Solver. In all experiments related to SAT domain, we use EasySAT5 to ensure direct comparability with AutoSAT[29]. The solver incorporates modern Conflict-Driven Clause Learning (CDCL) techniques including Literal Block Distance (LBD) heuristics, and VSIDS variable selection, and conflict-driven clause learning. All configurations (compiler versions, interfaces, time budgets) maintain parity with [29] for reproducibility. + +# C.2 Baselines Setting + +# C.2.1 Neural Combinatorial Optimization + +L2B [14]. The paper addresses the challenge of variable selection in the branch-and-bound (B&B) algorithm for solving mixed-integer linear programs (MILPs), where traditional expert-designed heuristics like strong branching incur prohibitive computational costs. To overcome this, the authors propose a graph convolutional neural network (GCNN) framework that leverages the natural bipartite + +graph representation of MILPs – with variable and constraint nodes connected via edges representing their coefficients – to learn branching policies through imitation learning. Key innovations include encoding MILP states as bipartite graphs with constraint/variable features, designing permutation-invariant sum-based graph convolutions with prenormalization layers to handle variable-sized inputs, and training via behavioral cloning of strong branching decisions using cross-entropy loss. All configurations take the default value used in [14] for fair comparison. + +HEM [15, 16]. The paper addresses the challenge of improving cut selection in mixed-integer linear programming (MILP) solvers by simultaneously optimizing three critical aspects: which cuts to select (P1), how many to choose (P2), and their ordering (P3). Existing learning-based methods focus primarily on scoring individual cuts while neglecting dynamic cut count determination and sequence dependencies. To overcome these limitations, the authors propose a novel hierarchical sequence model (HEM) that leverages a two-level architecture: a higher-level policy predicts the ratio of cuts to select using a tanh-Gaussian distribution, while a lower-level pointer network formulates cut selection as a sequence-to-sequence learning problem to output ordered subsets. The model is trained via hierarchical policy gradient optimization within a reinforcement learning framework, where states encode MILP relaxation features and candidate cut characteristics, actions represent ordered cut subsets, and rewards correspond to solver performance metrics like primal-dual gap integral. All hyperparameters take the default value used in [15, 16] for fair comparison. + +NeuroSAT [40]. The paper proposes a message-passing neural network (MPNN) to solve the Boolean satisfiability problem by training solely on binary labels indicating satisfiability. The key innovation lies in representing SAT instances as bipartite graphs where literals and clauses are nodes connected via edges, and leveraging permutation-invariant neural architectures to process these graphs. NeuroSAT iteratively refines node embeddings through bidirectional message passing: clauses aggregate information from their constituent literals, while literals update their states based on connected clauses and complementary literals. Trained on a distribution of random SAT problems generated by incrementally adding clauses until unsat, the model learns to predict satisfiability via a cross-entropy loss on the final aggregated literal embeddings. Crucially, the architecture enforces structural symmetries (variable/clause permutation invariance, negation equivalence) and generalizes to larger instances and entirely unseen domains (e.g., graph coloring, vertex cover) at test time by simply extending the message-passing iterations, despite being trained only on small random $n \leq 40$ problems. All configurations of this algorithm take the default used in [40] for fair comparison. Note that since NeuroSAT is a prediction framework for Boolean satisfiability problem, it is meaningless to measure the solving time for this framework over the SAT instances. Besides, the ground-truth dataset used to train the NeuroSAT model is those SAT instances can be proved/solved within timelimit $\tau$ . Thus, we only count the number predicted to be correct of NeuroSAT (this number must be less than the number of ground-truth labels). Then, the number of timeout for NeuroSAT is equal to the total number of test instances minus the number predicted to be correct. + +# C.2.2 Evolutionary Code Optimization + +AutoSAT [29]. The paper introduces a framework that leverages Large Language Models (LLMs) to automatically optimize heuristics in Conflict-Driven Clause Learning (CDCL) SAT solvers. The authors address the challenge of manually designing and tuning heuristic functions in modern CDCL solvers, which is time-consuming and expert-dependent. Instead of generating solvers from scratch, AutoSAT operates within a modular search space comprising nine key heuristic functions (e.g., branching, restart, and clause management) derived from an existing CDCL solver. The framework employs LLMs to iteratively refine these heuristics through two search strategies: a greedy hill climber (GHC) and a $(1 + 1)$ Evolutionary Algorithm (EA), where LLMs generate candidate code modifications guided by performance feedback. All configurations of this algorithm take the default used in [29] for fair comparison. + +LLM4Solver [30]. The paper proposes a novel framework that integrates large language models (LLMs) with evolutionary search to automate the design of high-performance diving heuristics for exact combinatorial optimization (CO) solvers. The key challenge addressed is the inefficiency of traditional manual and learning-based approaches in navigating the vast, discrete algorithm space for CO solver components like diving heuristics, which require domain expertise and suffer from poor generalization. The core methodology leverages LLMs as prior-knowledge-guided generators to produce candidate algorithms encoded as interpretable score functions, while a derivative-free evolutionary framework (single- or multi-objective) optimizes these candidates through iterative + +population-based search. The algorithm space is defined via 13 interpretable variable features, with LLMs enabling code-level manipulations during initialization, crossover, and mutation. We slightly adjusted LLM4Solver's optimization focus from diving heuristics to cut selection for MILP solver for fair comparison with previous work [15, 16]. All configurations of this algorithm take the default used in [15, 16] for fair comparison. + +# D Dataset Details + +# D.1 SAT Domain + +Chromatic-Number-of-the-Plane (CNP). This problem requires coloring graph vertices such that adjacent nodes have distinct colors. + +Profitable-Robust-Production (PRP). This problem seeks robust production plans maintaining profitability under uncertain/fluctuating conditions. The generation setting of PRP adhere to [29]. + +CoinsGrid. This problem, based on Tony Hurlimann's coin puzzle, involves arranging coins on a grid with row-wise, column-wise, and distance-based constraints. The generation setting of CoinsGrid adhere to [29]. + +Zamkeller. This problem requires finding a permutation of integers from 1 to $n$ that maximizes differential alternations in subsequences divisible by integers 1 to $k$ ( $1 < k < n$ ). Instance generation follows [29]. + +Table 3: The statistical description of used SAT datasets. Const. and Var. stand for constraints and variables respectively. + +
DatasetSize# Const. (mean)# Const. (std)# Var. (mean)# Var. (std)
CNP150261,260352,3798,7988,893
CoinsGrid781,116,9721,068,009154,828148,571
PRP802,120,9831,346,344317,635201,185
Zamkeller80310,804335,10224,59222,190
+ +# D.2 MILP Domain + +Easy. The SCIP 8.0.0 solver solves MILP instances in the Easy dataset within one minute. This dataset comprises three synthetic problems: Set Covering [41], Maximum Independent Set [42], and Multiple Knapsack [43]. We select these classes because: (1) they serve as standard benchmarks for MILP solver evaluation [44], and (2) they encompass diverse MILP problem structures encountered in practice. Following [44, 43], we generate set covering instances (500 rows $\times$ 1000 columns), Maximum Independent Set instances (500-node graphs with affinity 4), and multiple knapsack instances (60 items, 12 knapsacks). For each dataset within this category, we generate 1,000 training instances and hold 100 instances as test set. Besides, we set the timelimit as 60 seconds for solving problem instances within this dataset. + +Medium. The SCIP 8.0.0 solver requires at least five minutes to solve this set of instances optimally. Following [54, 55, 53], this dataset combines $\mathrm{MIK^6}$ [46] (MILPs with knapsack constraints) and CORLAT [47] (real-world grizzly bear corridor planning in Northern Rockies). Each problem set is partitioned into $80\%$ training and $20\%$ test instances. Furthermore, we set the timelimit as 120 seconds for solving problem instances within this dataset. + +Hard. The SCIP 8.0.0 solver requires more than one hour to solve Hard dataset instances to optimality. These datasets8 include Load Balancing from the NeurIPS 2021 ML4CO competition [48]. Load Balancing instances model server task allocation under resource constraints. Same as Medium dataset, each problem set in Hard dataset is also partitioned into $80\%$ training and $20\%$ test instances. Additionally, we set the timelimit as 360 seconds for solving problem instances within this dataset. + +$^{6}$ MIK data can be found at https://atamturk.ieor.berkeley.edu/data/mixed(integer.knapsack/. +$^{7}$ CORLAT data is available at https://bitbucket.org/mlindauer/aclib2/src/master/. +8All these data can be found at https://www.ecole.ai/2021/ml4co-competition/. + +Table 4: The statistical description of used MILP datasets. + +
Dataset# Constraints (mean)# Variables (mean)
Set Covering (SC)5001,000
Maximum Independent Set (MIS)1,953500
Knapsack72720
MIK346413
CORLAT486466
Load Balancing64,30461,000
+ +# E Metric Details + +# E.1 Metrics related to MILP + +For fair comparison, our evaluation employs metrics aligned with HEM [15, 16, 14], which are well-established in the MILP community. + +Solving Time. For MILP instances solved to optimality within time limit $T$ , we measure the solver runtime $t$ required to obtain certifiably optimal solutions. For those MILP instances that cannot be solved to optimality within time limit $T$ , we utilize the following metrics. + +Primal-Dual Gap. During the execution of MILP solvers, two critical bounds are maintained: the global primal bound and the global dual bound. The global primal bound represents the objective value of the incumbent feasible solution, serving as the tightest upper bound for the MILP. The global dual bound corresponds to the minimal lower bound among all active nodes in the search tree, constituting the strongest lower bound for the problem. The primal-dual gap is defined as the absolute difference between these two bounds. In SCIP 8.0.0 [52], the initial primal-dual gap is initialized to a constant value of 100. + +Primal-Dual (PD) Integral. The primal-dual gap integral is defined as the area enclosed between the primal and dual bound trajectories during solver execution. Formally, given a time limit $T$ , this integral is expressed as: + +$$ +\int_ {t = 0} ^ {T} \left(\boldsymbol {w} ^ {T} \boldsymbol {x} _ {t} ^ {*} - \boldsymbol {z} _ {t} ^ {*}\right) \mathrm {d} t, \tag {20} +$$ + +where $\mathbf{w}$ denotes the objective coefficient vector of the MILP instance; $\boldsymbol{x}_t^*$ represents the incumbent feasible solution at time $t$ , $\mathbf{z}_t^*$ corresponds to the tightest dual bound at time $t$ . This metric quantifies the cumulative optimality gap over the solving horizon and serves as an established performance benchmark for MILP solvers. Notably, the primal-dual gap integral was adopted as the primary evaluation criterion in the NeurIPS 2021 ML4CO competition [48]. + +# E.2 Metrics related to SAT + +To ensure fair comparison, our evaluation adopts metrics consistent with AutoSAT [29], which are widely recognized in the SAT community. + +Solving Time. For SAT instances that can be solved/proved within a timeout limitation $\tau$ , we record the running time $t$ of a SAT solver solving/proving these instances. For those SAT instances that cannot be solved/proved within timeout limitation $\tau$ , we measure the two following metrics. + +Number of Timeout. Given a timeout limitation $\tau$ and a set of SAT instances, we record the number of SAT instances that cannot be solved/proved within $\tau$ for different compared methods. The $\tau$ is set as 5000 seconds, same as [29]. + +Penalized Average Runtime with a factor of 2 score (PAR-2). PAR-2 is a standard evaluation metric in SAT solver competitions and benchmarks, employed to assess and compare solver performance under incomplete executions or timeout scenarios. Specifically, the PAR-2 score aggregates runtime across test instances as follows: 1) For solved instances, record the solver's runtime; 2) For instances where the solver exceeds the timeout limitation $\tau$ , assign twice the timeout threshold $\tau$ as the penalty—this is why we call this metric as PAR-2; 3) Compute the average across all instances by summing these values and dividing by the instance count. Lower PAR-2 scores indicate better overall solver efficiency. + +Intuitively, consider a benchmark set comprising $n$ SAT instances. Let $t_i$ denote the solver's runtime on the i-th instance. The PAR-2 score is formally defined as + +$$ +\operatorname {P A R} - 2 = \frac {1}{n} \sum_ {i = 1} ^ {n} t _ {i}, \tag {21} +$$ + +where + +$$ +t _ {i} = \left\{ \begin{array}{l l} t _ {i}, & \text {i f} t _ {i} \leq \tau ; \\ 2 \tau , & \text {i f} t _ {i} > \tau \text {o r t h e s o l v e r f a i l s t o r e t u r n a s o l u t i o n .} \end{array} \right. \tag {22} +$$ + +# F Discussion + +End-to-End Training. In this work, we adopt a two-stage training procedure for our composite model: first training a graph neural network (GNN), then training a large language model (LLM) conditioned on the frozen GNN embeddings. This approach stems from our empirical observation that end-to-end joint training of the composite model presents significant optimization challenges. We hypothesize that this limitation could be mitigated by exploring more sophisticated GNN architectures or developing enhanced alignment strategies between the GNN's structural representations and the LLM's semantic space. + +Explore Domain-specific Distributional Discrepancies. In our ablation studies, empirical results reveal that the full STRCMP model (with complete post-training) underperforms relative to its variants STRCMP (SFT Only) and STRCMP (DPO Only) across benchmark datasets, suggesting potential conflicts between optimization objectives during multi-stage post-training. This observation motivates future investigation into domain-specific distributional discrepancies that may arise from the heterogeneous nature of combinatorial optimization problems, which could inform improved alignment strategies for cross-domain post-training protocols. + +Interpretability of the generated code functions scale with the increasing complexity of the CO problem. We fully acknowledge that while code generation enhances transparency relative to neural black-box methods, interpretability does not scale linearly with problem complexity. For highly intricate CO problems or novel heuristics, generated code may involve nested logic, context-dependent optimizations, or non-intuitive transformations. This necessitates non-trivial effort for human experts to parse, reducing immediate interpretability. Nevertheless, STRCMP retains fundamental advantages over neural NCO methods: 1) Inspectability: solver logic remains open to direct inspection in its artifactual form; 2) Modifiability: code can be edited without model retraining; and 3) Debuggability: runtime validation and issue tracing via standard tools remain possible. We view the interpretability of complex generated code as an ongoing research challenge and are committed to addressing it in future work. Potential directions include simplifying the generated algorithms through post-processing or enhancing them with annotations to aid human understanding. For now, we believe STRCMP strikes a valuable balance, advancing interpretability over NCO methods while laying the groundwork for further improvements. + +Incorporating Additional Modal Priors. In this study, we focus specifically on integrating graph structural priors into LLM-based algorithm discovery for combinatorial optimization problems. While current methodologies rely on human expertise to predefine target components, we posit that LLMs' context-aware capabilities could assimilate additional modal priors (e.g., dynamic constraint patterns or solution quality metrics) to enhance combinatorial optimization problem-solving performance. Future extensions may enable LLMs to autonomously identify computational bottlenecks through techniques such as runtime complexity profiling or constraint violation pattern analysis, thereby prioritizing component optimization. + +# G Additional Experimental Results + +# G.1 Optimization Performance Result + +Table 5: The optimization performance result w.r.t. PAR-2 score between different methods over SAT domain. + +
Compared MethodsPAR-2 Score (↓)
CNPCoinsGridPRPZamkeller
AutoSAT644.721098.69639.34807.75
EasySAT649.421595.752000829.38
STRCMP624.151124.091804.46270.62
STRCMP (DPO Only)643.451098.69482.92265.94
STRCMP (SFT Only)643.231098.471837.86227.69
STRCMP w/o GNN645.671098.341820.38646.12
+ +Table 6: The optimization performance result w.r.t. Solving Time between different methods over SAT domain. + +
Compared MethodsSolving Time (↓)
CNPCoinsGridPRPZamkeller
AutoSAT32472191582296720772
EasySAT32942260645000021810
STRCMP3141518971462237990
STRCMP (DPO Only)3234519158211467765
STRCMP (SFT Only)3232319151468936929
STRCMP w/o GNN32567191474701917014
+ +![](images/ff7ca5360a344af8b80af912df27007c5cf455ba216ff28c9ae75fa0491bbc7e.jpg) +Figure 8: Convergence comparison (w.r.t. PAR-2) between evolutionary-based algorithm discovery frameworks on PRP dataset of SAT domain. + +# G.2 Convergence Comparison Result + +![](images/fd8491755cf7a95aac54843889ed83efd6aa01d89c4cbe97003fe2fc0a005986.jpg) +(a) CNP + +![](images/2d45d7cabecc9b5c3356a293c76fef2f1108bef4502d32bcb5aa77c3cb082569.jpg) +(b) CoinsGrid + +![](images/8f8a4ccfc0f7a43ad4c32b7ad64aa1572fe1e09d6701a32287104245485acbf0.jpg) +(c) CNP + +![](images/9128094e3947e449270c21798d14f472e32468d94f1edf329ca96c4cf4597c40.jpg) +(d) CoinsGrid + +![](images/4638c1cc056673ae2fc4b0ff4e64f7bddf81cae8f7a36061a2b8ef68c7fa5e48.jpg) +(e) PRP + +![](images/dae6058a27d12f43fa08ac74a9dddc48a8a637436aa950b0855977442b44492d.jpg) +(f) Zamkeller + +![](images/8172a0e4f8173e1af2a3225505a025a9025b5bb44ef41226110cadc394a04041.jpg) +(g) CNP + +![](images/4ada3211d196f8e8f0ec68337c0da7010c11ffdb26d18a7b4b836cefa14afb94.jpg) +(h) CoinsGrid + +![](images/0b1f66cdbda72bb703db88d25ad358ffab96d5e12cede24b2c2afe6a5dd324fc.jpg) +(i) PRP + +![](images/fddb2fcdb3418e1968befcfb05f4a5b731717b0527119b5f5105fd0e21ea68d8.jpg) +(j) Zamkeller +Figure 9: Convergence comparison (w.r.t. PAR-2, Solving Time, and Number of Timeout) between evolutionary-based algorithm discovery frameworks on SAT domain. + +![](images/5def7a858d4dd79e058be1ed4a3ac807de931e8a0d197b496579d0629e4a8d03.jpg) +(a) MIK w.r.t. PD Integral + +![](images/3b3f25d073b941d4e69cd73789b1b0a619cdcf93bac0a08490bc78b701f7ad64.jpg) +(b) MIK w.r.t. Solving Time + +![](images/06e3ca6f689e1ae4e920edda9741add759a51ef0bdc964f614a2cdcfd3b4c0fe.jpg) +(c) MIS w.r.t. PD Integral + +![](images/61a18659ff1fba0c99e4ba3853937847b029b3d2009c74a7507cc6e3efbcbdf2.jpg) +(d) MIS w.r.t. Solving Time + +![](images/05673ab77aa584dafb99dd4be3478c446d9b9c6b3f7bb920267bbc3bd02e6f48.jpg) +(e) CORLAT w.r.t. PD Integral + +![](images/eed3bab589ff4bf040fda3dab78b2de1dcf556eeed23e25bb910984bfd910f8e.jpg) +(f) CORLAT + +![](images/1791287a5283a1919ba13d122fd87c2550693cc55c13b136afd8f8fe10c46690.jpg) +(g) Knapsack w.r.t. PD Integral + +![](images/5dee61109fe42cca80eacd050a0a046353394e75ddeeeac2e96e27a03bad40d1.jpg) +(h) Knapsack w.r.t. Solving Time + +![](images/f9937aee67bb297335d455ef990c9dafd87e29792d810e5eef890c63d87667ea.jpg) +(i) Set Cover w.r.t. PD Integral + +![](images/933b7cbef919fa26d7a60abf640f101f9a76a6778c20ec4fe27cc1e7a0b53fa2.jpg) +(j) Set Cover w.r.t. Solving Time +Figure 10: Convergence comparison (w.r.t. PD Integral and Solving Time) between evolutionary-based algorithm discovery frameworks on MILP domain. + +# G.3 Ablation Studies Result + +![](images/ea5aed8027f058b5d367b2966142088916eee9b103092e469153fec704c918f0.jpg) +(a) CNP + +![](images/b4426ece226c7140537444f85a9ffc623888dac54737ae577036782ec67f3fc2.jpg) +(b) CoinsGrid + +![](images/c1c65f30db71398cc5faf0afbdee2dded0623d49b7b8ed6bb42f31ba867cd6f3.jpg) +(c) CNP + +![](images/56e20df8e8798debf3c5b22e9bcacf5515e0e7ded850206780d81343a2fbf26a.jpg) +(d) CoinsGrid + +![](images/6cae28e8bbd43301ba4f1e088525a45a0193018ce1e295448a42e12233947caa.jpg) +(e) PRP + +![](images/4d7c9ddcf742310b68a6a58a8a8c45aeaf456590d34914e833e6e13f86f645fd.jpg) +(f) Zamkeller + +![](images/a6b40c1f4f3c786788ac448c0d6b3b7029b7ffe982e407f04436706d512ea70c.jpg) +(g) CNP + +![](images/9a1ca8b531b3d6ec87f2e8ec58a40fe8041ae8350e8c0c4023defb23ab51d671.jpg) +(h) CoinsGrid + +![](images/28d875f2f2a6f06ed430b97e0ac26500659e19ca4d31c0edb638b8b33660d8ed.jpg) +(i) PRP + +![](images/26f34ca148a881c1961dbb4c5090378e0183f3d6ebead99d666f6c908b38732f.jpg) +(j) Zamkeller +Figure 11: Ablation studies (w.r.t. PAR-2, Solving Time, and Number of Timeout) during algorithm search on SAT domain. + +# H Experiment Statistical Significance + +![](images/1fe464c23eee65159128315ac5d8b79508793a16d0687252255c47049c1ecd0e.jpg) +(a) CNP + +![](images/2f3b7f54ac752ae1e77ae734dcb22bd68c3b1bc990ac5f52d8c198bf300c64ec.jpg) +(b) CoinsGrid + +![](images/1a0501d48cae0c672448172667f64ff1f1910de77a5fb419d0560427739a3d49.jpg) +(c) PRP + +![](images/2d0faa2e795e27bb702107d7e58283e524770777c1dc03ae351b8b1d19d52c43.jpg) +(d) Zamkeller + +![](images/64e2aaad1524cd576f34911b2e4b560b761474e8fc74c91b44dfe3106836b046.jpg) +(e) CNP + +![](images/a2af31d39149e5d41b2c3c79c35e02b21ad2e210a7c39c74bbdf419d78cd1b75.jpg) +(f) CoinsGrid + +![](images/2edc11d855e794a3b9efd90d943890d609a916d1181ea58b481be134e0a6f850.jpg) +(g) PRP + +![](images/5ebbcbcd7a26921b3570a423b8a7d8bda06ce1946ed2cd625ff24a5aff3ca613.jpg) +(h) Zamkeller + +![](images/2d720684158f87500b2a8570d2ab12ebdfe30d05631a9340a04b86a6e869dcf2.jpg) +(i) CNP + +![](images/38b57e0541af7a5a8aead61a0de6170ac529defb56cef32a1ef18ad72a91cbb3.jpg) +(j) CoinsGrid + +![](images/189782b3a6dd24234004a324b9de1b91d8a95c513276a49888733574d752dbf7.jpg) +(k) PRP + +![](images/959d7bb705590c0419ca16cc219496682a4db9973adbd96553ccdab7d9095a5e.jpg) +(1) Zamkeller +Figure 12: Convergence comparison with variance statistic (w.r.t. PAR-2, Solving Time, Number of Timeout) between evolutionary-based algorithm discovery frameworks on SAT domain. + +# I Comparison between Two-stage training and Joint End-to-End Training + +The detailed results of this new training and testing are presented show in Table 7 and Table 8. We can conclude the followings from these results: + +1 The end-to-end joint training is extremely unstable compared to our proposed two-stage training method. We infer that the training instability stems from the architectural and optimization differences between the GNN and the LLM. The gradients, originating from a sequence-level generation loss in the deep LLM, become noisy and attenuated when backpropagated through the entire LLM and into the much shallower GNN. This creates a challenging optimization landscape where a single learning rate is ineffective for both components, leading to unstable convergence (even with many more training steps) as shown in Table 7. Compared with the joint training, the loss curve of our two-stage training method is stable and close to zero at convergence (Please see the loss curve of our two-stage training method in Figure 7 in Appendix). The GNN and LLM are trained together, which makes it difficult for the models to learn optimal parameters simultaneously. + +The model exhibits poor performance when tested on solving SAT instances. Unsurprisingly, due to its very unstable training and relatively high loss at convergence, the performance of the jointly trained model is lower than our two-stage model on the two SAT datasets, as shown in Table 8. + +Table 7: The Training Loss of End-to-End Joint Training + +
Training StepEnd-to-End DPO LossEnd-to-End SFT Loss
1000.541.64
2000.601.53
3000.261.44
4000.431.28
5000.561.14
6000.131.04
7000.420.98
8000.290.88
9000.300.86
+ +Table 8: The Jointly Trained Model's Performance over Two SAT datasets w.r.t. PAR-2 Score + +
DatasetTwo-Stage JointEnd-to-End
Zamkeller270.62751.34
PRP1804.462092.71
+ +# J Experiment on the Training Strategy + +To provide a clearer and more comprehensive answer, we offer a detailed analysis focusing on dataset complexity and the iterative optimization behavior of our proposed method, strcmp, and its variants. + +Dataset Complexity. The complexity of combinatorial optimization problems can be characterized by various metrics. Drawing from established literature, we focus on several key graph-based metrics—Modularity, Density, and Clustering—to provide a quantitative perspective on the structural complexity of the datasets used in our experiments. While not absolute, higher values in these metrics generally suggest a more intricate problem structure and complexity. The problem sizes are detailed in Appendix D of our manuscript. The statistics for the MILP and SAT datasets are presented below. + +Table 9: Data Complexity of MILP Dataset + +
DatasetModularityDensityClustering
setcover0.180.070.34
knapsack0.070.040.36
mis0.480.010.02
mik0.110.230.30
loadbalancing0.490.010.01
+ +Table 10: Data Complexity of SAT Dataset + +
DatasetModularityDensityClustering
CNP0.81810.00160.2012
CoinsGrid0.93060.00050.5377
Zamkeller0.73830.00040.5659
PRP0.88650.00020.5429
+ +Iterative Behavior of STRCMP and its Variants. To provide a clearer picture, we have compiled tables showing the non-smoothed, iteration-by-iteration performance on both MILP and SAT datasets. We employ an early stopping strategy where the search terminates if no performance improvement is observed within three consecutive iterations. Hence the total iteration number of different method will be varied. These results illustrate the performance trajectory (Primal-Dual Integral and PAR-2 score) of STRCMP against its SFT-only and DPO-only ablations, as shown in the following tables. + +Table 11: The Performance v.s. Iteration Result w.r.t. Primal-Dual Integral between Different Methods over corlat Dataset + +
IterationsLLM4SolverSTRCMPSTRCMP (SFT Only)STRCMP (DPO Only)STRCMP w/o GNN
39939.488750.846880.257159.2410325.77
69860.208889.197235.316870.3110435.85
99673.478836.857364.476881.4810404.27
129900.066932.57
+ +Table 12: The Performance v.s. Iteration Result w.r.t. Primal-Dual Integral between Different Methods over mik Dataset + +
IterationsLLM4SolverSTRCMPSTRCMP (SFT Only)STRCMP (DPO Only)STRCMP w/o GNN
3535.04478.08394.25451.48492.96
9499.60455.18421.44434.09494.24
15470.48450.72408.55431.42481.97
21466.45438.51372.12
27471.22448.31380.74
+ +Table 13: The Performance v.s. Iteration Result w.r.t. Primal-Dual Integral between Different Methods over loadbalancing Dataset + +
IterationsLLM4SolverSTRCMPSTRCMP (SFT Only)STRCMP (DPO Only)STRCMP w/o GNN
31942.271881.742125.222193.621910.53
91885.311878.522065.612245.971905.49
151867.711859.142051.36
+ +Table 14: The Performance v.s. Iteration Result w.r.t. PAR-2 Score between Different Methods over CNP Dataset + +
IterationSTRCMPSTRCMP (SFT Only)STRCMP (DPO Only)
34993149395.166748963.5
1560433.333359101.333357184
............
5157108.166751935.666757688.67
5755257.833352890.166757352
............
7548779.5
7847443
+ +Analysis and Insights. From the iterative performance results, we draw two primary observations: + +SFT-only models often converge faster in terms of iteration count but may stabilize at a suboptimal performance level. This suggests that while SFT is effective at rapidly imitating strong solutions, it may overfit to the strategies present in the initial dataset, leading to premature convergence in a local optimum. +DPO-inclusive models (DPO-only and SFT+DPO) demonstrate a capacity for continued improvement over a longer training horizon, often surpassing the SFT-only model's peak performance. We posit that DPO enhances the model's generalization capabilities. By learning from preference pairs rather than absolute solutions, the model develops a more nuanced understanding of what constitutes a better search trajectory, allowing it to escape local optima and discover superior solutions. + +Based on these observations and the complexity analysis, we can summarize our insights on selecting a training strategy: + +1 For "easy" datasets (e.g., small-scale, sparse, solvable to optimality quickly): The SFT-only approach is often sufficient and efficient. On these problems, high-quality training data (i.e., optimal or near-optimal solutions) can be generated at a low cost, allowing SFT to quickly learn an effective policy. +2 For "hard" datasets (e.g., large-scale, dense, complex structure, intractable): We strongly recommend a DPO-based approach (DPO-only or SFT+DPO). For these problems, obtaining optimal solutions for SFT is prohibitively expensive or impossible. However, generating preference pairs by comparing the performance of different candidate solutions is still feasible and relatively cheap. The superior generalization of DPO-trained models makes them better suited to navigating the vast and complex search spaces of these challenging instances. +For medium complexity datasets, the choice is less definitive. As the results show, the interplay between SFT and DPO is intricate. Our full STRCMP model (SFT+DPO) often acts as a robust default, leveraging SFT for a strong initialization and DPO for refinement and generalization. However, we acknowledge that determining the optimal blend is a significant challenge. The task-specific training of large models has not yet fully clarified the precise relationship between SFT and DPO—when one is definitively superior, or how their data should be optimally combined. We consider this an important open problem for the field. + +Table 15: The Performance v.s. Iteration Result w.r.t. PAR-2 Score between Different Methods over CoinsGrid Dataset + +
IterationSTRCMPSTRCMP (SFT Only)STRCMP (DPO Only)
3603765944459735.67
965177.676530570214.83
............
3058516.8362026.1761296.83
3358091.6762460.8361622.17
............
5756775.1753737.67
6058212.83
+ +Table 16: The Performance v.s. Iteration Result w.r.t. PAR-2 Score between Different Methods over PRP Dataset + +
IterationSTRCMPSTRCMP (SFT Only)STRCMP (DPO Only)
353808.1753696.3354977.5
95452356456.6754462.5
............
3953015.8355141.6752635
4250812.55345952731.67
............
6652701.55362751030
6950352.33
............
8749281
+ +Table 17: The Performance v.s. Iteration Result w.r.t. PAR-2 Score between Different Methods over Zamkeller Dataset + +
IterationSTRCMPSTRCMP (SFT Only)STRCMP (DPO Only)
345613.54498845427
656487.1747457.551138
............
2756511.549499.1744826.5
3054933.675091445185.67
............
5151801.3353002.8342444.17
5451900.6749976.538346.17
............
7241710
7536365
+ +# K Dependence on the Underlying LLM + +Given the time and resource constraints, we focused our evaluation on the Llama2 and Qwen2 families of models, selecting representatives of varying sizes. We performed training and testing on two datasets from the MILP domain and two from the SAT domain. The results are presented below. + +It is important to note that the entire Llama2 family of models failed on our code generation task, either by exceeding the context limitation of 4096 tokens or by being unable to generate syntactically correct, executable code. Similarly, the smaller Qwen2 models (0.5B and 1.5B) were also unable to produce viable code for the solvers. This underscores the complexity of the task, which requires a highly capable code-generating LLM. From the results in the tables, we can draw the following conclusions: + +Table 18: The Performance of STRCMP w.r.t. Primal-Dual Integral with Varied-Size Backbone LLM over MILP Domain + +
STRCMPSTRCMP w/o GNN
DatasetQwen2(0.5B)Qwen2(1.5B)Qwen2(7B)Qwen2(14B)Qwen2(7B)Qwen2(14B)
mik--269.68233.56292.41262.74
setcover--34.0234.2934.0634.24
+ +Table 19: The Performance of STRCMP w.r.t. PAR-2 Score with Varied-Size Backbone LLM over SAT Domain + +
STRCMPSTRCMP w/o GNN
DatasetQwen2(0.5B)Qwen2(1.5B)Qwen2(7B)Qwen2(14B)Qwen2(7B)
PRP--470.63414.12500.00
Zamkeller--355.3127.7635.75
308.30
+ +1 The structural embedding from the GNN consistently and significantly contributes to the final performance. Across both MILP and SAT domains, and for both the 7B and 14B model sizes, STRCMP consistently outperforms its STRCMP w/o GNN variant. This confirms that the GNN provides a crucial inductive bias that guides the LLM toward better solutions, regardless of the LLM's scale. +STRCMP's performance is robust, provided a sufficiently capable LLM backbone is used. While our framework's effectiveness is contingent on the LLM's fundamental ability to comprehend the task and generate valid code, the performance is stable across capable models of different sizes. +3 STRCMP's performance is sensitive to the size of the LLM backbone, but this effect is problem-dependent. The results reveal a nuanced relationship between model scale and performance. For the more complex mik instances, scaling the LLM from 7B to 14B parameters yields a notable performance gain. However, for the setcover instances, the 7B and 14B models perform almost identically. This suggests that while larger models can unlock better performance on certain complex problem structures, a more moderately-sized model may be sufficient for others. + +# L Prompt and Corresponding Responses + +Prompt of data curation for SAT domain (adapted from AutoSAT [29]) +```txt +# Description of the role duty +You are a SAT solver researcher trying to write the bump_var to help $\rightarrow$ SAT solver escape from local optimum. +# Description of the task goal +Your goal is to improve bump_var for the SAT solver, after reading $\rightarrow$ and understanding the of SAT solver below. +# Description of the example codes +As an example, the format of bump_var can be: +// start +void Solver::bump_var(int var, double coeff) { ... } // end +To replace the original code: +// start +void Solver::bump_var(int var, double coeff) { if ((activity[var] += var_inc * coeff) > 1e100) { // Update score and prevent float overflow for (int i = 1; i <= vars; i++) activity[i] *= 1e-100; var_inc *= 1e-100;} if (vsids.inHeap(var)) vsids.update(var); // update heap } // end +# Description of the tips to the task +Tips: +1) Your bump_var code must start with '''// start''' and end with $\rightarrow$ '''// end'' +2) Feel free to unleash your creativity and provide your improved and $\rightarrow$ correct code. (Pay attention to correct namespace.) +3) Your bump_var code must be different from original codes. +4) You can only modify the current bump_var and use the existing $\rightarrow$ external variables (e.g. variables that defined in Solver). +5) You cannot add new external variables outside the bump_var. +6) The SAT problems that the SAT solver needs to solve have the $\rightarrow$ structural characteristics. +# The whole code of target solver of SAT solver is: +// Omit the key code of SAT solver here for space limitations. +Take a deep breath and think it carefully. Then respond in correct $\rightarrow$ format! +``` + +Collected Response to above Prompt +```cpp +// start +void Solver::bump_var(int var, double coeff) { + // Update the activity score with a dynamic coefficient based on LBD + double dynamic_coeff = coeff * (1.0 + (clause_DB.back().lbd / $\rightarrow$ 10.0)); + if ((activity[var] += var_inc * dynamic_coeff) > 1e100) { + // Update score and prevent float overflow for (int i = 1; i <= vars; i++) activity[i] *= 1e-100; + var_inc *= 1e-100; + } + if (vsids.inHeap(var)) vsids.update(var); + // update heap + // Implement a decay factor for activity scores based on the number of conflicts + static int conflict_count = 0; + conflict_count++; + if (conflict_count % 1000 == 0) { + for (int i = 1; i <= vars; i++) activity[i] *= 0.95; + // Decay activity scores + } +} +// end +``` + +# Prompt of data curation for MILP domain (adapted from previous work [15, 30]) + +# Description of the task background + +1. Introduction to MILP: Mixed Integer Linear Programming (MILP) is a + +$\hookrightarrow$ mathematical optimization or decision making methodology for +$\hookrightarrow$ mathematical planning problems in which the objective function +and constraints are linear and some of the decision variables are +$\hookrightarrow$ restricted to integers. + +2. MILP Definition: A mixed integer linear programming (MILP) problem + +has the following elements: + +2.1) A linear objective function fx, where f is a column vector +of constants and $\mathbf{x}$ is a column vector of unknowns. +2.2) Boundary and linear constraints, but no nonlinear +$\hookrightarrow$ constraints. +2.3) Restrictions on certain components of $\mathbf{x}$ so that they must +have integer values. + +3. Cutting Planes Definition: cutting planes are additional linear + +$\hookrightarrow$ inequality constraints added to the MILP problem. These +$\hookrightarrow$ inequalities attempt to restrict the feasible domain of the LP +$\hookrightarrow$ relaxation so that its solution is closer to an integer. + +4. Cutting Plane Algorithm: The Cutting Plane method solves the MILP +$\hookrightarrow$ problem by linearly relaxing the integer problem into a +$\hookrightarrow$ non-integer linear problem and solving it. The theory of linear +$\hookrightarrow$ programming states that under mild assumptions (if there exists +$\hookrightarrow$ an optimal solution to the linear programming and the feasible +$\hookrightarrow$ domain does not contain a line), there always exists an extremum +or vertex that is optimal. Test whether the optimal solution +$\hookrightarrow$ obtained is an integer solution. If not, then there must exist a +$\hookrightarrow$ linear inequality separating the optimal point from the convex +$\hookrightarrow$ envelope of the true feasible set. Finding such an inequality is +$\hookrightarrow$ the separation problem, and such an inequality is the cutting +$\hookrightarrow$ plane. The cutting plane can be added to the linear program being +relaxed such that the current abelian solution is no longer +$\hookrightarrow$ feasible for the relaxation. The process is repeated until an +$\hookrightarrow$ optimal integer solution is found. + +5. Instructions: You need to understand the above, especially about + +$\hookrightarrow$ the cut-plane algorithm, the next tasks will be closely related +to the cut-plane algorithm. + +# Description of task details + +1. Sample code for a cut plane selection function: + +# + +[ start ] + +import numpy as np + +import ... as ... + +def fun_name(cuts_features, sel_cuts_num): + +# + +return idxes + +[end] + +# + +The parameters of the function are the characteristics of all cuts + +and the number of cuts to be selected. After completing the +$\hookrightarrow$ selection, it returns an idx list of all selected cutting planes. +$\hookrightarrow$ In addition, each function is independent, each function needs to +$\hookrightarrow$ import the corresponding package such as numpy... and each +$\hookrightarrow$ function uses [start] and [end] as the start and end identifiers. + +# (Continued) Prompt of data curation for MILP solver + +```txt +2. Characterization: Pay attention to the cutting plane features in $\leftrightarrow$ the function input.The following are the specific contents $\rightarrow$ contained in these features: +cut_feature $= [$ +obj_parall, #Target Parallelism eff,# Cut-plane validity support, +integral_support, +normalizedViolation, +mean_coeff-cut, +max_coeffCut, +min_coeffCut, +std_coeffCut, +mean_coeff_obj, +max_coeff_obj, +min_coeff_obj, +std_coeff_obj] +``` + +```txt +3. No comments or annotations are required, only the code needs to be $\leftrightarrow$ provided. +``` + +# Collected Response to above Prompt + +```python +[begin] +import numpy as np +def hybrid_diversity_selector(cuts_features, sel_cuts_num): + features = np.array(cuts_features) + if len(features) <= sel_cuts_num: + return list(range(len(features))) + eff = features[:, 1] + viol = features[:, 4] + supp = features[:, 2] + features[:, 3] + coeff_std = features[:, 8] / (features[:, 5] + 1e-10) + quality = np.sqrt(eff * viol) * np.log1p(supp) + q_thresh = np-percentile(quality, 70) + var_thresh = np-percentile(coeff_std, 30) + mask = (quality >= q_thresh) & (coeff_std <= var_thresh) + candidates = np.where_mask)[0] + if len(candidates) <= sel_cuts_num: + return candidates.tolist() + active_features = features[candidates] + diversity = np.zeros(len(candidates)) + for i in range(len(candidates)): + coeff_diff = np.sum(np.abs(active_features[:, 5:9]) + → active_features[i, 5:9]), axis=1) + obj_diff = np.abs(active_features[:, 0] + → active_features[i, 0]) + diversity[i] = np.mean(coeff_diff * obj_diff) + scores = quality[candidates] * diversity + return candidates[np.argsort(-scores)[-sel_cuts_num]].tolist() +end +``` \ No newline at end of file diff --git a/NeurIPS/2025/$_texttt{STRCMP}$_ Integrating Graph Structural Priors with Language Models for Combinatorial Optimization/images.zip b/NeurIPS/2025/$_texttt{STRCMP}$_ Integrating Graph Structural Priors with Language Models for Combinatorial Optimization/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f597e551368d5187a91d6ed1785b0c9f8e0aca1c --- /dev/null +++ b/NeurIPS/2025/$_texttt{STRCMP}$_ Integrating Graph Structural Priors with Language Models for Combinatorial Optimization/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a2089f5e25fb5543c9899f056f9704e8690f941ffe1f0421518f85556eb94f69 +size 1859631 diff --git a/NeurIPS/2025/$_texttt{STRCMP}$_ Integrating Graph Structural Priors with Language Models for Combinatorial Optimization/layout.json b/NeurIPS/2025/$_texttt{STRCMP}$_ Integrating Graph Structural Priors with Language Models for Combinatorial Optimization/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b53b8e4ad7a84abfde82f13cc66988ed59bf296d --- /dev/null +++ b/NeurIPS/2025/$_texttt{STRCMP}$_ Integrating Graph Structural Priors with Language Models for Combinatorial Optimization/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:757263c8a6ba5c54b060f03276c22b0978b9e9a3f172bb073f07e258cecbfac1 +size 1312016 diff --git a/NeurIPS/2025/$_text{G}^2_text{M}$_ A Generalized Gaussian Mirror Method to Boost Feature Selection Power/34478a6b-d6ea-47d7-8e26-ceafa25687c9_content_list.json b/NeurIPS/2025/$_text{G}^2_text{M}$_ A Generalized Gaussian Mirror Method to Boost Feature Selection Power/34478a6b-d6ea-47d7-8e26-ceafa25687c9_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..533a24621e87662c2eacdfc0d2ffb76b36d36567 --- /dev/null +++ b/NeurIPS/2025/$_text{G}^2_text{M}$_ A Generalized Gaussian Mirror Method to Boost Feature Selection Power/34478a6b-d6ea-47d7-8e26-ceafa25687c9_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c1aebe67addb767c4bb7b011a9a70ca1f23f787bd6dc8dee0f7d81faa1cb5504 +size 240401 diff --git a/NeurIPS/2025/$_text{G}^2_text{M}$_ A Generalized Gaussian Mirror Method to Boost Feature Selection Power/34478a6b-d6ea-47d7-8e26-ceafa25687c9_model.json b/NeurIPS/2025/$_text{G}^2_text{M}$_ A Generalized Gaussian Mirror Method to Boost Feature Selection Power/34478a6b-d6ea-47d7-8e26-ceafa25687c9_model.json new file mode 100644 index 0000000000000000000000000000000000000000..279a5afb20d53c1522509a9c01b1f45e03790c7a --- /dev/null +++ b/NeurIPS/2025/$_text{G}^2_text{M}$_ A Generalized Gaussian Mirror Method to Boost Feature Selection Power/34478a6b-d6ea-47d7-8e26-ceafa25687c9_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f02a00ac366487a117f62fea9578de95a24e471779ddb8fa8d625aa4500199c +size 294653 diff --git a/NeurIPS/2025/$_text{G}^2_text{M}$_ A Generalized Gaussian Mirror Method to Boost Feature Selection Power/34478a6b-d6ea-47d7-8e26-ceafa25687c9_origin.pdf b/NeurIPS/2025/$_text{G}^2_text{M}$_ A Generalized Gaussian Mirror Method to Boost Feature Selection Power/34478a6b-d6ea-47d7-8e26-ceafa25687c9_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..b7d40b77772b5193804996af6ccc4d274c3e98c3 --- /dev/null +++ b/NeurIPS/2025/$_text{G}^2_text{M}$_ A Generalized Gaussian Mirror Method to Boost Feature Selection Power/34478a6b-d6ea-47d7-8e26-ceafa25687c9_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:231607cab9025cb87caa71039697373bab016dbbbcddb3fae2ab1c2a144983c4 +size 706355 diff --git a/NeurIPS/2025/$_text{G}^2_text{M}$_ A Generalized Gaussian Mirror Method to Boost Feature Selection Power/full.md b/NeurIPS/2025/$_text{G}^2_text{M}$_ A Generalized Gaussian Mirror Method to Boost Feature Selection Power/full.md new file mode 100644 index 0000000000000000000000000000000000000000..8b941a9a5f74b3ffd64861903e631ef80b56c7f3 --- /dev/null +++ b/NeurIPS/2025/$_text{G}^2_text{M}$_ A Generalized Gaussian Mirror Method to Boost Feature Selection Power/full.md @@ -0,0 +1,1084 @@ +# $\mathbf{G}^2 \mathbf{M}$ : A Generalized Gaussian Mirror Method to boost feature selection power + +Hongyu Shen1, Zhizhen Zhao1 + +$^{1}$ Department of Electrical and Computer Engineering + +University of Illinois Urbana-Champaign + +{hongyu2, zhizhenz}@illinois.edu + +# Abstract + +Recent advances in false discovery rate (FDR)-controlled feature selection methods have improved reliability by effectively limiting false positives, making them well-suited for complex applications. A popular FDR-controlled framework called data splitting uses the "mirror statistics" to select features. However, we find that the unit variance assumption on mirror statistics could potentially limit the feature selection power. To address this, we generalize the mirror statistics in the Gaussian mirror framework and introduce a new approach called "generalized Gaussian mirror" $(\mathrm{G}^2\mathrm{M})$ , which adaptively learns the variance and forms new test statistics. We demonstrate both theoretically and empirically that the proposed test statistics achieve higher power than those of Gaussian mirror and data splitting. Comparisons with other FDR-controlled frameworks on synthetic, semi-synthetic, and real datasets highlight the superior performance of the $\mathrm{G}^2\mathrm{M}$ method in achieving higher power while maintaining FDR control. These findings suggest the potential for the $\mathrm{G}^2\mathrm{M}$ method for practical applications in real-world problems. Code is available at: https://github.com/skyve2012/G2M. + +# 1 Introduction + +The complexity and high dimensionality of data in the era of big data make it challenging to develop a universal feature selection algorithm that performs well in diverse datasets [18]. An algorithm that works effectively for one type of data may fail for another. This variability complicates the verification of feature selection results, thereby hindering the practical use of selected features. Such challenges are especially prevalent in fields like healthcare, where there is no guarantee that identified genes or metabolites [8] correlate well with response variables (e.g., diseases), making downstream applications such as drug discovery even more difficult. + +To address these challenges, recent research has focused on false discovery rate (FDR)-controlled feature selection methods [4, 8]. These methods limit the number of false positives during feature selection, providing a guarantee of the maximum number of incorrect selections. This property is particularly advantageous in ultra-high-dimensional settings, such as genetic or RNA-sequencing data. + +Over the past decade, various methods have been proposed. Model-X knockoffs [4, 8] is a novel framework designed to select relevant features while controlling the FDR. It generates "knockoff" variables as references to the original design matrix and forms knockoff statistics. The symmetry about-zero property of these statistics for null features enables FDR control under specified nominal levels. Although theoretical guarantees exist for Gaussian design matrices [8, 37], challenges remain for non-Gaussian settings due to the lack of theoretical assurances. Methods like Deep Knockoff [30], KnockoffGAN [14], and DeepDRK [34] have been proposed to enhance selection power while + +controlling FDR in non-Gaussian data. However, as highlighted in Shen et al. [34], balancing reconstructability (which enhances power) and the swap property (which ensures FDR control) remains empirically difficult, and theoretical guarantees are still lacking. + +Another line of work is the conditional randomization test (CRT), first introduced in Candès et al. [8]. CRT relies on conditional sampling for each feature given the others to compute $p$ -value statistics, which are then used with the Benjamini-Hochberg (BH) procedure to control FDR. Unlike knockoff statistics, CRT avoids the reconstructability issue by sequentially considering individual features. However, its computational bottleneck arises from the need to repeatedly generate conditional samples, especially for non-Gaussian data where Markov Chain Monte Carlo (MCMC) techniques are required. This substantially increases the computational cost. Recent methods such as HRT [40], and Distillation-CRT [19] have sought to improve sampling accuracy and efficiency. Despite these advances, the inherent challenges of non-Gaussian data remain unresolved. + +Distinct from these approaches, data splitting [9] and Gaussian mirror [45] introduce a "mirror statistics" with properties similar to knockoff statistics but without relying on the swap property. This avoids the trade-off identified in Shen et al. [34]. The key difference between the two methods lies in how the mirror statistics are constructed. Data splitting divides the design matrix and response into two parts, generating paired estimates for each feature, which are then used to compute the mirror statistics. In contrast, the Gaussian mirror introduces two Gaussian noise perturbations to create additional columns in the design matrix, replacing the original feature column1. These perturbed columns are used to compute the mirror statistics. However, on the one hand, the data-splitting approach is constrained by its unit-variance assumption for the mirror statistics. On the other hand, the Gaussian mirror paper does not provide a discussion on the statistical power of the proposed method. + +To address these limitations, we generalize the variance assumptions in the mirror statistics introduced in the data-splitting framework and propose a generalized Gaussian mirror test statistics that considers the variance information of the fitting coefficients—an intermediate component that forms the mirror statistics. We demonstrate that this statistics is the entry-wise uniformly most powerful test statistics, leading to improved feature selection power compared to both the Gaussian mirror and data-splitting test statistics. We call the proposed method a "generalized Gaussian mirror" $(\mathrm{G}^2\mathrm{M})^2$ . With experiments considering synthetic, semi-synthetic, and real datasets, we demonstrate the practical value of the proposed approaches over a set of benchmarking methods. + +# 2 Related Work + +# 2.1 Data Splitting + +Data splitting is a feature selection method designed to control the FDR [9]. This method partitions the design matrix $X$ and the response vector $y$ into two disjoint datasets, denoted as $(X^{+},y^{+})$ and $(X^{-},y^{-})$ . Separate models are then fitted to each dataset using the same approach (e.g., both use ordinary least squares), yielding the coefficients $\beta_{j}^{+}$ and $\beta_{j}^{-}$ from two respective models for each feature $x_{j}$ . For each pair $(\beta_{j}^{+},\beta_{j}^{-})$ , the method constructs a test statistics $w_{j}$ , which is a function of these coefficients and reflects the importance of the corresponding feature $x_{j}$ . + +Notably, the statistics $w_{j}$ is designed as a "mirror statistics," ensuring symmetry about 0 for null features $(j \in S_0)$ while maintaining positive values for non-null features $(j \in S_1)$ . After calculating $w_{j}$ for all features, the selection procedure identifies the set of significant features $\hat{S}_1 = \{j : w_j \geq \tau_q\}$ , where the threshold $\tau_{q}$ is determined based on the desired FDR control. + +$$ +\tau_ {q} = \min _ {t > 0} \left\{t: \frac {1 + | \{j : w _ {j} \leq - t \} |}{\max (1 , | \{j : w _ {j} \geq t \} |)} \leq q \right\}. \tag {1} +$$ + +This setting controls feature selection FDR level at $q$ according to Theorem 2.1. + +Theorem 2.1. For a given set of mirror statistics $\{w_{j}\}_{j = 1}^{p}$ , the following properties hold: + +1. If $j$ is a null feature index, then for any real number $t$ , $\mathbb{P}(w_j \geq t) = \mathbb{P}(w_j \leq -t)$ . +2. If $j$ is a non-null feature index, then $\mathbb{P}(w_j \geq 0) > \mathbb{P}(w_j \leq 0)$ . + +Then, for a given nominal FDR level $q$ , the feature selection set + +$$ +\widehat {S} = \{j \in \{1, \dots , p \}: w _ {j} \geq \tau_ {q} \} +$$ + +controls the FDR, where $\tau_{q}$ is chosen via Eq. (1). + +Proof. Given the above two properties for the mirror statistics, we have $\mathbb{P}(w_j\geq t) = \mathbb{P}(w_j\leq -t)$ due to the symmetry-about-zero property of the mirror statistics $w_{j}$ under the null distribution. As a result, the selection rule $\widehat{S} = \{j:w_j\geq \tau_q\}$ (equivalently, the denominator of Eq. (1)) selects features based on the level- $q$ quantile on the right side of the null distribution, ensuring the FDR level is controlled at most $q$ . + +# 2.2 Gaussian Mirror + +The Gaussian mirror [45] is a different approach from the data splitting method given how $\beta_{j}^{+}$ and $\beta_{j}^{-}$ are generated. Specifically, given an index $j$ , the Gaussian mirror proposes two perturbed variables in place of the original $x_{j}$ , resulting in a new design matrix $X^{j} = (x_{j}^{+}, x_{j}^{-}, X_{-j})$ , where $x_{j}^{+} = x_{j} + c_{j}z_{j}$ , $x_{j}^{-} = x_{j} - c_{j}z_{j}$ , $z_{j} \sim \mathcal{N}(\mathbf{0}, \mathbf{I}_{n})$ , and $c_{j}$ is a scalar in $\mathbb{R}$ . $\beta_{j}^{+}$ and $\beta_{j}^{-}$ are the coefficient estimations for $x_{j}^{+}$ and $x_{j}^{-}$ , respectively. + +The Gaussian mirror adopts a similar notion of mirror statistics as the data splitting method to select features with controlled FDR. With $c_{j} = \frac{\|P_{\perp - j}x_{j}\|}{\|P_{\perp - j}z_{j}\|}$ , Xing et al. [45] shows that $\beta_j^+$ and $\beta_j^-$ are independent, allowing the choice of a specific form of mirror statistics $w_{j} = |\beta_{j}^{+} + \beta_{j}^{-}| - |\beta_{j}^{+} - \beta_{j}^{-}|$ to perform FDR-controlled feature selection at level $q$ , according to Eq. (1). Here, $P_{\perp -j}$ denotes the projection in the null space of $X_{-j}$ . + +# 2.3 Connection between Data Splitting and Gaussian Mirror in the $\mathbf{G}^2\mathbf{M}$ setting + +In this paper, the proposed $\mathrm{G}^2\mathrm{M}$ test statistics is based on a variance-generalized uniformly most powerful (UMP) statistics under the mirror statistics setting in the data splitting paper [9]. Specifically, it adopts the $\beta_{j}^{+}$ and $\beta_{j}^{-}$ from the Gaussian mirror paper [45] to form a generalized mirror statistics (see Lemma 3.5) over the one from the data splitting paper. The proposed statistics considers the variance information of $\beta_{j}^{+}$ and $\beta_{j}^{-}$ , resulting in higher feature selection power compared to that of the data splitting and Gaussian mirror. More importantly, it is theoretically verifiable. In addition, We provide evidence on the no-constant variance of $\beta_{j}^{+}$ and $\beta_{j}^{-}$ in Appendix A—the rationale on why such variance generalization is necessary. + +# 3 Method + +This section details the proposed approach in this paper. The arrangement is as follows: + +- Section 3.2 considers a more general setup of Gaussian mirror variable $x_{j}^{+}$ and $x_{j}^{-}$ . And provide the related form of $\beta_{j}^{+}$ and $\beta_{j}^{-}$ as a weighted sum of the true $\beta_{j}$ coefficient and the noise variables. We also provide two useful properties about $\beta_{j}^{+}$ and $\beta_{j}^{-}$ whose proofs are omitted in the original Gaussian mirror paper [45]. +- Section 3.3 presents the first major result (i.e., Lemma 3.5) on the UMP test statistics given the feature index $j$ . The setup we consider is more realistic compared to the one considered in Dai et al. [9], resulting in higher test power. +- Section 3.4 introduces the second major result that proves the proposed $\mathrm{G}^2\mathrm{M}$ test statistics (see Proposition 3.1) is a more powerful test statistics compared to the mirror statistics in Xing et al. [45], Dai et al. [9]. + +- In Section 3.5, we combine the results discussed in the previous sections and propose two algorithms—one exact algorithm that characterizes an ideal scenario, and one estimation algorithm that can be used in practice. Experiments in Section 4 are conducted w.r.t. the estimation version. + +# 3.1 Notation + +We first introduce the notation for the paper. Let $X \in \mathbb{R}^{n \times p}$ be the design matrix, $x_{j} = (x_{1j}, \ldots, x_{nj})^{\top}$ be the $j$ -th column of $X$ , and $X_{-j}$ be the submatrix of $X$ with the $j$ -th column removed. $n$ is the sample size and $p$ is the feature dimension. Following the setting in Xing et al. [45], we assume that $X$ is normalized such that $\sum_{i=1}^{n} x_{ij} = 0$ and $\| x_{j} \|_2 = n$ , $j \in [1, \ldots, p]$ . Let $y = (y_{1}, \ldots, y_{n})^{\top}$ be the vector of $n$ independent responses. We assume that the response variable $y$ only depends on a subset of features $X_{S_1} = \{X_j : j \in S_1\}$ , and the task of feature selection is to identify the set $S_1$ . And the response $y$ follows a linear model: $y = X\beta + \epsilon$ , where $\beta = (\beta_1, \ldots, \beta_p)$ indicates the coefficients for the features $x_{j}$ 's. Let $S_0 = \{1, \ldots, p\} \setminus S_1$ be the index set of the null features. Let $p_0 = |S_0|$ and $p_1 = |S_1|$ be the number of the null and the relevant features, respectively. + +# 3.2 A General form of Gaussian Mirror Coefficient Estimation + +We propose to represent $\beta_{j}^{+}$ and $\beta_{j}^{-}$ as a function of the true $\beta_{j}$ and the noise vector $\epsilon$ , given a generalized Gaussian mirror setup that considers additional variables (i.e., $d_{j}$ and $q_{j}$ ). This improves flexibility on the representation of $\beta_{j}^{+}$ and $\beta_{j}^{-}$ . Details are included in Proposition 3.1, which helps us to understand the behavior of $\beta_{j}^{+}$ and $\beta_{j}^{-}$ in terms of $q_{j}, z_{j}, c_{j}, d_{j}, \beta_{j}$ and $\epsilon$ . + +Proposition 3.1. Consider a generalized Gaussian mirror setup where $x_{j}^{+} = x_{j} + c_{j} \cdot z_{j}$ and $x_{j}^{-} = x_{j} - d_{j} \cdot q_{j}$ , and $z_{j}$ and $q_{j}$ are $n$ dimensional i.i.d. standard Gaussian. $c_{j}$ and $d_{j}$ are in $\mathbb{R}$ . Then the least square regression coefficients for $x_{j}^{+}$ and $x_{j}^{-}$ have the following forms: + +$$ +\beta_ {j} ^ {+} = \alpha \cdot \beta_ {j} + \gamma^ {\top} \cdot \epsilon , \quad \beta_ {j} ^ {-} = \zeta \cdot \beta_ {j} + \eta^ {\top} \cdot \epsilon , \tag {2} +$$ + +where $\epsilon$ stands for $n$ -dimensional i.i.d. standard Gaussian vector. And $\alpha, \gamma, \zeta$ , and $\eta$ are functions of $c_{j}$ , $d_{j}$ , $z_{j}$ , and $q_{j}$ (see Appendix B.1 for full forms). + +In addition to Proposition 3.1, we introduce three corollaries that reveal certain properties about $\beta_j^+$ and $\beta_j^-$ and terms associated with them (see Eq. (2)). These properties are crucial in developing the estimation algorithm in Section 3.5. + +Corollary 3.2. Given $z_{j} = q_{j}$ , $\alpha +\zeta = 1$ , where $\alpha$ and $\zeta$ are defined in Proposition 3.1. + +Corollary 3.3. Given $c_{j} = d_{j}$ , and $z_{j} = q_{j}$ , $\alpha = \zeta = 0.5$ , where $\alpha$ and $\zeta$ are defined in Proposition 3.1. + +Specifically, Corollary 3.2 describes a property on the behavior of $\alpha$ and $\eta$ that sum to one as long as the Gaussian mirror perturbation vectors $z_{j} = q_{j}$ . Notice that this does not require $c_{j} = d_{j}$ , indicating a more generalized setup beyond the Gaussian mirror. On the other hand Corollary 3.3 indicates that under the vanilla Gaussian mirror setup (see Section 2.2 and Xing et al. [45]), the mean of $\beta_{j}^{+}$ and $\beta_{j}^{-}$ is $0.5\beta$ . Note that the result is only mentioned in Xing et al. [45] without any proof. Here we provide formal proof of this fact, complementing the original Gaussian mirror work. The proof of Corollary 3.2 and 3.3 can be found in Appendix B. + +Corollary 3.4. $Var(\beta_j^+) = \|\gamma\|_2^2$ and $Var(\beta_j^-) = \|\eta\|_2^2$ . + +The proof is straightforward by realizing that $\epsilon$ is the only term that introduces randomness to $\beta_j^+$ and $\beta_j^-$ . Hence $\mathrm{Var}(\beta_j^+) = \mathrm{Var}(\gamma^\top \cdot \epsilon) = \| \gamma \|_2^2$ . Similar reasoning can be applied to $\mathrm{Var}(\beta_j^-)$ . + +# 3.3 Entry-wise Uniformly Most Powerful Mirror Statistics + +According to Section 2.3 and Appendix A, it is clear that the variances of $\beta_j^+$ and $\beta_j^-$ need not be 1. This violates the condition for the UMP mirror test statistics proposed in Dai et al. [9]. To consider a + +more realistic setup, we present the first result (i.e., Lemma 3.5) that reveals a mirror test statistics that is a function of the variance of $\beta_j^+$ and $\beta_j^-$ . We show that the proposed test statistics is UMP for every pair of $\beta_j^+$ and $\beta_j^-$ . Note that we do not assume a universal distribution for $\beta_j^+$ and $\beta_j^-$ across all $j$ 's. This flexibility directly leads to Theorem 3.6, which proves that the test statistics in Lemma 3.5 achieves higher power during feature selection compared to the existing mirror statistics [45, 9]. + +Lemma 3.5. Suppose that the set of coefficients $(\beta_j^+, \beta_j^-)$ is independent for all $j \in \{1, \ldots, p\}$ . Furthermore, suppose: + +(a) For $j \in S_1$ , the two coefficients $\beta_j^+$ and $\beta_j^-$ follow $N(\omega, \sigma_a^2)$ , $N(\omega, \sigma_b^2)$ independently. And $\omega \sim \delta \cdot \text{Rademacher}(0.5) (\delta > 0)$ , where $\delta \in \mathbb{R}$ is the absolute magnitude of $\omega^3$ ; +(b) For $j \in S_0$ , the two coefficients $\beta_j^+$ and $\beta_j^-$ follow $N(0, \sigma_a^2)$ and $N(0, \sigma_b^2)$ independently; +(c) $\frac{p_1}{p_0} \to r$ as $p \to \infty$ . + +Then the optimal choice of $f$ in the mirror statistics (i.e., $\mathrm{sign}(\beta_j^+ \beta_j^-) f(|\beta_j^+|, |\beta_j^-|)$ ) that yields the highest power follows the form: + +$$ +f (a, b) = U \left[ P _ {+} \exp (- S _ {-}) + P _ {-} \exp (- S _ {+}) \right], \tag {3} +$$ + +where $S_{-} = \frac{\delta(\delta - 2a)}{2\sigma_{a}^{2}} +\frac{\delta(\delta - 2b)}{2\sigma_{b}^{2}}$ $S_{+} = \frac{\delta(\delta + 2a)}{2\sigma_{a}^{2}} +\frac{\delta(\delta + 2b)}{2\sigma_{b}^{2}}$ and + +$$ +P _ {+} = \Phi \left(\frac {\delta}{\sigma_ {a}}\right) \Phi \left(\frac {\delta}{\sigma_ {b}}\right), P _ {-} = \left[ 1 - \Phi \left(\frac {\delta}{\sigma_ {a}}\right) \right] \left[ 1 - \Phi \left(\frac {\delta}{\sigma_ {b}}\right) \right], U = \frac {1}{P _ {E | H _ {1}}} = \frac {1}{P _ {-} + P _ {+}}, +$$ + +where $\Phi$ is the cumulative density function for the standard normal distribution. + +Lemma 3.5 reveals a closed-form test statistics that is UMP. This is used in developing the proposed feature selection algorithms in Section 3.5. + +# 3.4 More Powerful Test Setting + +Combining the results from Proposition 3.1 and Lemma 3.5, we present the following Theorem 3.6 that indicates a more powerful test statistics of the proposed over the existing ones in Xing et al. [45], Dai et al. [9]. + +Theorem 3.6. The proposed test statistics $f(a, b)$ in Eq. (3) achieves the highest power compared to the Gaussian mirror test statistics: $|a + b| - |a - b|$ [45], and the data splitting test statistics: $\mathrm{sign}(ab)(|a| + |b|)$ [9], under the same nominal FDR level $q$ . + +In Section 4, we will empirically observe evidence that supports Theorem 3.6. And the proof of Theorem 3.6 can be found in Appendix B. + +# 3.5 Algorithm + +With all the ingredients introduced in the previous sections, we present two algorithms: 1. the exact algorithm (Algorithm 1) when one has access to the scale of the true coefficients for the nonnull $\beta_{j}$ 's: $\delta$ or $\delta_{j}$ if it is index-dependent; 2. the estimation algorithm (Algorithm 2) that essentially proposes an estimation of $\delta$ or $\delta_{j}$ , resulting a generalized likelihood ratio test statistics given the estimation. Note that the proposed algorithm considers the vanilla Gaussian mirror setup to take advantage of the nice properties indicated in Corollary 3.2 and 3.3. This means $c_{j} = d_{j}$ and $z_{j} = q_{j}$ for all $j$ 's. And $c_{j} = \frac{\|P_{\perp - j}x_{j}\|}{\|P_{\perp - j}z_{j}\|}$ following the setup in Xing et al. [45]. + +Algorithm 1 introduces the exact algorithm when true $\delta$ presents. This is rarely the case as often we do not have prior knowledge about true $\beta_{j}$ 's. To overcome this issue, we propose an algorithm to estimate $\delta$ , which is Algorithm 2. + +Essentially, the estimation algorithm employs a $k$ -means-based algorithm to find $k$ potential modes for $\delta_j$ —a quantity that is assumed to be given in Algorithm 1, yet unavailable in practice. We + +# Algorithm 1 Exact algorithm + +1: Input: Design matrix $X$ , response $y$ , nominal FDR level $q$ and true scale $\delta$ or $\delta_j$ for the $\beta_j$ , $j \in S_1$ . +2: Output: $\tilde{S}_1 = \{j\mid w_j\geq \tau_q\}$ +3: for $j = 1$ to $p$ do +4: Sample $z_{j}\sim \mathcal{N}(\mathbf{0},\mathbf{I}_{n})$ +5: Calculate $c_{j} = \frac{\|P_{\perp - j}x_{j}\|}{\|P_{\perp - j}z_{j}\|}$ and form $x_{j}^{+} = x_{j} + c_{j}z_{j}, x_{j}^{-} = x_{j} - c_{j}z_{j}$ . +6: Obtain the ordinary least squares estimator of $\hat{\beta}_j^+$ and $\hat{\beta}_j^-$ : + +$$ +(\hat {\beta} _ {j} ^ {+}, \hat {\beta} _ {j} ^ {-}, \hat {\beta} _ {- j}) = \arg \min _ {\beta_ {j} ^ {+}, \beta_ {j} ^ {-}, \beta_ {- j}} \| y - X _ {- j} \beta_ {- j} - x _ {j} ^ {+} \beta_ {j} ^ {+} - x _ {j} ^ {-} \beta_ {j} ^ {-} \| _ {2} ^ {2}. +$$ + +7: Calculate $\operatorname{Var}(\beta_j^+) = \|\gamma\|_2^2$ and $\operatorname{Var}(\beta_j^-) = \|\eta\|_2^2$ . +8: Compute $w_{j} = \mathrm{sign}(\beta_{j}^{+}\beta_{j}^{-})f(\beta_{j}^{+},\beta_{j}^{-})$ , where $f(a,b)$ is specified in Eq. (3). +9: end for + +10: Calculate the threshold given the FDR level $q$ : + +$$ +\tau_ {q} = \min _ {t > 0} \left\{t: \frac {1 + | \{j : w _ {j} \leq - t \} |}{\max (1 , | \{j : w _ {j} \geq t \} |)} \leq q \right\}. +$$ + +choose $k$ -means given it has been widely considered as an efficient unsupervised algorithm due to no accessible to true values of $\delta_j$ . First, combining the results of Corollary 3.2 and 3.3, the algorithm obtains an unbiased estimate of the absolute value of $\beta_j$ ( $j \in (1, \dots, p)$ ) in Step 7 of Algorithm 2. This is because $\alpha$ and $\zeta$ are both equal to 0.5 based on Corollary 3.2 and 3.3. The $p$ calculated values represent possible values of $\delta$ . Then we utilize the $k$ -means algorithm with $k$ chosen via the silhouette score—a common technique for choosing the number of clusters. Note that the $k$ identified modes include the case for the null $\beta_j$ 's, where $\delta = 0$ , due to not knowing which variable is null and which is nonnull. Luckily, we can simply exclude the smallest mode (e.g., the closest one to zero), which corresponds to the mode for the null $\beta_j$ 's, resulting in $k - 1$ candidates. For every calculated pair of $\beta_j^+$ and $\beta_j^-$ , $\hat{\delta}_j$ is chosen to be the closest mode among the $k - 1$ candidates, compared to the value $\frac{\beta_j^+ + \beta_j^-}{2}$ . With Algorithm 2, we can perform feature selection without accessing the true $\delta$ information. In Section 4, all results about the proposed G²M method are based on the estimation algorithm (Algorithm 2). + +# Algorithm 2 Estimation algorithm + +1: Input: Design matrix $X$ , response $y$ , nominal FDR level $q$ and the number of modes for $\delta$ : $k \leq p$ . +2: Output: Algorithm 1(X, $y,\{\hat{\delta}_j\}_{j = 1}^p$ $q)$ +3: for $j = 1$ to $p$ do +4: Sample $z_{j}\sim \mathcal{N}(\mathbf{0},\mathbf{I}_{n})$ +5: Calculate $c_{j} = \frac{\|P_{\perp - j}x_{j}\|}{\|P_{\perp - j}z_{j}\|}$ and form $x_{j}^{+} = x_{j} + c_{j}z_{j}, x_{j}^{-} = x_{j} - c_{j}z_{j}$ . +6: Obtain the ordinary least squares estimator of $\hat{\beta}_j^+$ and $\hat{\beta}_j^-$ : + +$$ +(\hat {\beta} _ {j} ^ {+}, \hat {\beta} _ {j} ^ {-}, \hat {\beta} _ {- j}) = \arg \min _ {\beta_ {j} ^ {+}, \beta_ {j} ^ {-}, \beta_ {- j}} \| y - X _ {- j} \beta_ {- j} - x _ {j} ^ {+} \beta_ {j} ^ {+} - x _ {j} ^ {-} \beta_ {j} ^ {-} \| _ {2} ^ {2}. +$$ + +7: Calculate the value $v_{j} = \frac{|\beta_{j}^{+} + \beta_{j}^{-}|}{2}$ . +8: end for +9: Run $k$ -means algorithm to find $k$ modes given $\{v_{j}\}_{j = 1}^{p}$ , represented as $\{m_l\}_{l = 1}^k$ . +0: for $j = 1$ to $p$ do +1: $\hat{\delta}_j = \min_{m_l}|v_j - m_l|,l\in (1,\ldots ,k).$ +2: end for + +# 4 Experiment + +In this section, we evaluate the proposed $\mathrm{G}^2\mathrm{M}$ approach using three types of datasets, following the experimental setup described in Shen et al. [34] given its comprehensibility. Specifically, we consider: 1. Fully synthetic datasets, where the design matrix $X$ and the response $y$ are generated from known distributions; 2. Semi-synthetic datasets, where the design matrix is extracted from real-world data, and the response is generated using a predefined model; 3. A real-world case study, where both $X$ and $y$ are fully unknown. The experimental details and corresponding results are summarized in Section 4.1, Section 4.2, and Section 4.3, respectively. In each section, we provide details on the benchmarking methods, the configuration of the dataset, and the results. All experiments were carried out on an IBM AC922 server with 2x 20 core IBM POWER9 CPU @ 2.4GHz. + +Table 1: FDR and power with Gaussian design matrix $X$ for $\rho = {0.6}$ (left) and $\rho = {0.7}$ (right). Bold entries indicate the highest power given controlled FDR at level 0.1. Blue for the second best, and Red for FDR> 0.1 . + +
Methodρ = 0.6ρ = 0.7
OLSRidgeLASSOOLSRidgeLASSO
FDRPowerFDRPowerFDRPowerFDRPowerFDRPowerFDRPower
CRT [8]0.000.130.070.690.070.770.000.020.130.510.110.56
Distilled-CRT [19]0.530.980.270.980.130.970.580.950.250.930.110.87
Gaussian Mirror [45]0.090.540.050.670.040.740.050.180.040.510.230.48
Gaussian Mirror† [45]0.100.610.050.710.040.760.130.490.060.640.040.61
Data Splitting [9]0.000.000.060.660.040.500.000.000.060.560.050.41
Data Splitting† [9]0.720.070.080.710.170.700.730.070.070.660.250.70
HRT [40]0.010.230.000.210.000.220.010.100.000.100.000.10
Powerful Knockoff [37]0.000.000.050.480.060.690.010.000.040.290.050.55
Powerful Knockoff† [37]0.750.080.080.610.060.730.800.070.070.500.050.63
G2M (ours)0.080.650.050.820.050.820.080.360.030.670.030.68
G2M† (ours)0.080.720.040.800.060.840.100.550.030.670.040.70
+ +# 4.1 Synthetic Data + +To thoroughly evaluate performance, we adopt the experimental framework outlined in Shen et al. [34] to generate diverse synthetic datasets defined by $(X,y)$ . In this setup, $X\in \mathbb{R}^p$ represents dependent variables drawn from predefined distributions, and $y\in \mathbb{R}$ serves as the response variable. We model $y$ using a linear response framework: $y\sim \mathcal{N}(X^T\beta ,1)$ . The true underlying $\beta$ is a $p$ -dimensional vector with entries independently sampled from the distribution $\frac{p}{15\cdot\sqrt{n}}\cdot$ Rademacher(0.5). This choice differs from the commonly used scaling factor $\frac{p}{\sqrt{n}}\cdot$ Rademacher(0.5) [38, 22], enabling a more challenging evaluation configuration. + +In addition to the Copula and Gaussian mixture models considered in Shen et al. [34] that quantifies the nonlinear design matrices, we include the Gaussian setting, exploring various correlation levels among the features. Since methods such as knockoffs [8] and the conditional randomization test (CRT) [8] have closed-form solutions in Gaussian settings, incorporating this setup offers valuable insights into the consistency between empirical results and theoretical guarantees. Due to space limit, we defer the introduction of these settings in Appendix D. + +Benchmarking Methods & Settings4: Given the linear and nonlinear nature of the design matrix, we utilize two sets of benchmarking methods. The first set comprises methods that are theoretically proven to control the false discovery rate (FDR), including knockoff [8], conditional randomization test (CRT) [8], data splitting [9], Gaussian mirror [45], distilled-CRT [19], HRT [40], and powerful knockoff [37]. In addition, we also consider another rule to choose $\tau_q$ (Eq. (4)) according to Ren and Barber [29], to improve the power for certain applicable methods and mark them with “†”: + +$$ +\tau_ {q} = \min _ {t > 0} \left\{t: \frac {1 + | \{j : w _ {j} \leq - t \} |}{\max (1 , | \{j : w _ {j} \geq t \} |)} \leq q \quad \text {o r} \sum_ {j \in [ p ]} \mathbb {1} \{j: w _ {j} \geq t \} < \frac {1}{q} \right\}, \tag {4} +$$ + +where "1" is the indicator function. Essentially, the two equations differ only when $\min_{t > 0} \{ t : \frac{1 + |\{j : w_j \leq -t\}|}{\max(1, |\{j : w_j \geq t\}|)} \leq q \} > \min_{t > 0} \{ t : \sum_{j \in [p]} \mathbb{1} \{ j : w_j \geq t \} < \frac{1}{q} \}$ . This means when $q$ is small or + +the fraction of non-nulls is small, then the first term usually finds a larger $\tau_{q}$ , in favor of the FDR control. In this case, the power is pretty low, as when $\tau_{q}$ is large, there are only limited non-nulls selected. Instead of choosing $\tau_{q}$ with the first term (e.g., Eq. (1)), the second term chooses a smaller $\tau_{q}$ , which results in higher power. + +The second set focuses on deep learning-based methods designed to handle nonlinear design matrices, such as Deep Knockoff [30] $^{5}$ , DDLK [38] $^{6}$ , KnockoffGAN [14] $^{7}$ , sRMMD [22] $^{8}$ , and DeepDRK [34] $^{14}$ . + +Table 2: FDR and power with Copula and Gaussian Mixture design matrix $X$ . Bold entries indicate the highest power with controlled FDR at level 0.1. Blue for the second best, and Red for $\mathrm{{FDR}} > {0.1}$ . + +
MethodCitationGaussian MixtureClayton & Exp.Clayton & GammaJoe & Exp.Joe & Gamma
FDRPowerFDRPowerFDRPowerFDRPowerFDRPower
CRT[8]0.000.210.080.690.050.910.080.450.070.87
Distilled-CRT[19]0.170.220.060.430.060.360.050.260.040.26
Gaussian Mirror[45]0.050.830.070.540.070.890.060.340.080.86
Gaussian Mirror†[45]0.040.830.070.700.080.920.090.520.090.84
Data Splitting[9]0.070.720.070.520.070.790.060.280.080.76
Data Splitting†[9]0.090.760.090.620.080.810.100.490.090.78
HRT[40]0.000.180.010.150.010.500.010.090.020.50
Powerful Knockoff[37]0.080.620.030.170.050.380.040.110.070.39
Powerful Knockoff†[37]0.090.640.120.400.060.510.140.390.070.53
Deep Knockoff[30]0.741.000.290.880.250.950.400.860.260.94
DDLK[38]0.790.990.130.300.270.660.040.000.320.59
KnockoffGAN[14]0.440.990.070.350.090.700.050.170.090.60
sRMMD[22]0.721.000.290.880.240.940.310.780.260.93
DeepDRK[34]0.100.830.070.350.080.780.100.420.090.70
G2M (ours)-0.070.860.090.580.100.940.100.320.100.89
G2M† (ours)-0.060.920.060.750.090.950.100.610.100.91
+ +For the Gaussian setup, we focus solely on the first set of methods, as they align with the theoretical guarantees for Gaussian designs. We use three different fitting methods to generate the estimation coefficient $\hat{\beta}_j$ 's: ordinary least square (OLS), ridge regression, and LASSO, to consider common adaptation of these methods in the feature selection setup. In contrast, for the Copula and Gaussian mixture setups, we evaluate both sets of methods and consider ridge regression to be the fitting method as empirically it produces the best performance according to Shen et al. [34]. Results are presented for FDR and power with $(n,p) = (200,100)$ , at the FDR nominal level 0.1. All reported values are averaged over 100 independent repetitions. + +Results: In Table 1, we present results with Gaussian data in two different correlation settings (e.g., $\rho = 0.6, 0.7$ ). In both cases, the proposed $\mathrm{G}^2\mathrm{M}$ outperforms other benchmarking methods for achieving the highest power while controlling the FDR under the nominal level of 0.1. In addition, we notice that the power of the proposed $\mathrm{G}^2\mathrm{M}$ is always greater than that of the Gaussian mirror and data splitting methods, indicating the consistency between the empirical results and the theoretical reasoning. Surprisingly, we discover that $\mathrm{G}^2\mathrm{M}$ performs consistently across different fitting methods other than OLS (e.g., LASSO and ridge regression) despite the fact that the theory was developed in the least squares sense. This suggests a possible wide application of the proposed $\mathrm{G}^2\mathrm{M}$ method. + +In Table 2, on the other hand, we consider results with nonlinear design matrix $X$ with ridge regression model being the fitting method. Similar to the Gaussian setting, $\mathrm{G}^2\mathrm{M}$ achieves the highest power on all datasets while controlling the FDR under the nominal 0.1 level. The proposed method even outperforms those deep-learning-based methods, suggesting its power in wide applications. Importantly, this observation is consistent with the theoretical justification as the proof of the $\mathrm{G}^2\mathrm{M}$ method does not depend on the distribution of the design matrix $X$ . + +To further demonstrate the performance of our method, we investigate high dimensional settings by reducing the number of samples relative to the 100 dimensional input data. In addition, we consider the impact of noise to the power and FDR during feature selection. The comparison is deferred to + +Appendix E due to space limit. The results, however, suggest that our proposed method outperforms other benchmarking methods in most cases, indicating a new state-of-the-art. + +# 4.2 Semi-synthetic Data + +In this section, we conduct a semi-synthetic study by extracting the design matrix $X$ from two real-world datasets. + +Single-cell RNA Sequencing: The first dataset consists of single-cell RNA sequencing (scRNA-seq) data obtained from $10 \times$ Genomics $^{9}$ . Each entry of $X \in \mathbb{R}^{n \times p}$ represents the observed expression levels of $p$ genes across $n$ cells. For additional background on this dataset, we refer readers to Hansen et al. [13] and Agarwal et al. [1]. Following the preprocessing pipeline described in Hansen et al. [13], we obtain the final dataset $X$ with $n = 10,000$ and $p = 100^{10}$ . The response $y$ is formulated according to a nonlinear function of $X$ . Details including the form of $y$ and the data preparation about $X$ are included in Appendix F due to space limit. + +Inflammatory Bowel Disease (IBD): The second dataset, publicly available from the Metabolomics Workbench 11, originates from a real-world study titled "Longitudinal + +Table 3: FDR and power with two semi-synthetic datasets: RNA and IBD. Bold entries indicate the case with the highest power given controlled FDR at level 0.1, Blue for the second best, and Red for $\mathrm{FDR} > 0.1$ + +
MethodCitationSemi-RNASemi-IBD
FDRPowerFDRPower
CRT[8]0.390.890.150.56
Distilled-CRT[19]0.270.870.110.75
Gaussian Mirror[45]0.610.970.040.26
Gaussian Mirror†[45]0.060.580.080.55
Data Splitting[9]0.210.860.160.73
Data Splitting†[9]0.140.640.160.77
HRT[40]0.380.870.050.42
Powerful Knockoff[37]0.500.940.070.21
Powerful Knockoff†[37]0.140.430.140.51
Deep Knockoff[30]0.000.140.270.55
DDLK[38]0.140.810.090.26
KnockoffGAN[14]0.000.000.100.25
sRMMD[22]0.000.000.240.44
DeepDRK[34]0.080.730.100.25
G2M (ours)-0.000.430.070.45
G2M† (ours)-0.040.660.100.61
+ +Metabolomics of the Human Microbiome in Inflammatory Bowel Disease (IBD)" [20]. The objective of this study is to identify significant metabolites associated with two forms of inflammatory bowel disease: ulcerative colitis (UC) and Crohn's disease (CD). We use the C18 Reverse-Phase Negative Mode dataset, which comprises 546 samples and 91 metabolites. + +To handle missing values, we preprocess the dataset by removing metabolites with more than $20\%$ missing data, resulting in a final set of 80 metabolites. The data matrix is normalized entry-wise to have zero mean and unit variance, following a log transform and imputation of missing values using the $k$ -nearest neighbor algorithm, as described in Masud et al. [22]. The response $y$ uses the same linear model described in Section 4.1, which has lower signal strength compared to the experiment in Shen et al. [34]. + +In both semi-synthetic settings, we consider ridge regression as the fitting method as empirically it performs the best according to Shen et al. [34]. All reported values are averaged over 100 independent repetitions. The FDR nominal level is set to 0.1. + +Results: Since we consider the nonlinear design matrix $X$ , methods from both the deep-learning-based and non-deep-learning-based (see Section 4.1) are considered in this experiment. Results are presented in Table 3. We find that almost all non-deep-learning- and framework-based methods fail to control the FDR with the RNA data, and part of the methods fail on the IBD data. In comparison, the proposed $\mathrm{G}^2\mathrm{M}$ method is the only one that can successfully control the FDR under the nominal level while achieving the second-best in power (the best-performing DeepDRK achieves a higher power of 0.73 in this case). In the IBD setting, on the other hand, $\mathrm{G}^2\mathrm{M}$ beats all non-deep-learning-based methods and deep-learning-based methods for producing the highest power with the controlled FDR. This evidence suggests a stable application of $\mathrm{G}^2\mathrm{M}$ on real-world data. + +Table 4: Feature selection results with IBD real dataset that consider true $X$ and $y$ .: RNA and IBD. FDR level is specified as 0.2 . Since there is no ground truth on the features, we report "number of referenced metabolites/number of identified" in place of FDR or power. + +
ModelG2M (ours)†CRTDistilled-CRTGaussian Mirror†Data Splitting†HRTPowerful Knockoff
Referenced / Identified18/2212/2117/2618/2517/232/54/6
ModelDeepDRKDeep KnockoffsRMMDKnockoffGANDDLK
Referenced / Identified19/2315/205/512/1417/25
+ +# 4.3 Real Case Study + +We further conduct two case studies using real data (i.e., IBD dataset [20] and breast cancer dataset [43]) for both the design matrix $X$ and the response variable $y$ to qualitatively evaluate the selection performance of the proposed method. We defer the breast cancer dataset analysis in Appendix E.4. + +In the IBD analysis, the response variable $y$ is categorical, where $y = 1$ indicates that a sample is associated with UC/CD, and $y = 0$ otherwise. The covariates $X$ are identical to those used in the second semi-synthetic setup described in Section 4.2. Since ground truth is not available for this dataset, we evaluate the results by identifying evidence of IBD-associated metabolites from existing literature, following the curation in Shen et al. [34], which draws upon the following sources: 1. Metabolites explicitly documented as being associated with IBD, UC, or CD in the PubChem database $^{12}$ ; 2. Metabolites reported in peer-reviewed publications; 3. Metabolites mentioned in preprints. For convenience, we reproduce the nominal metabolite table from Shen et al. [34] in Appendix G.1. + +We use the DeepPINK [21] model as the fitting method to obtain the feature coefficients in consideration of the nonlinearity between the input $X$ and the response $y$ in this real setting. The FDR nominal level is set to 0.2. + +Results: We present feature selection results for the real IBD data in Table 4. Among the considered methods, clearly the $\mathrm{G}^2\mathrm{M}$ method performs on par with the DeepDRK method and identifies more reported metabolites while keeping a lower number of total selections compared to other benchmarking methods. This reveals the potential of the $\mathrm{G}^2\mathrm{M}$ method for real-world applications. For completeness, we include a full list of names for the identified metabolites by each method in Appendix G.2. + +# 5 Conclusion + +In this paper, we first identify a limitation of the existing mirror statistics in the data splitting paper—the strong unit variance assumption. We then proposed a variance-dependent Gaussian mirror method— $\mathrm{G}^2\mathrm{M}$ —and show both theoretically and empirically the performance of $\mathrm{G}^2\mathrm{M}$ compared to popular FDR-feature controlled frameworks and deep-learning-based knockoff methods with synthetic, semi-synthetic, and real datasets. The results demonstrate that the $\mathrm{G}^2\mathrm{M}$ method effectively controls the FDR while achieving the highest power in most cases and delivering comparable performance in others. These findings suggest the potential for broad adoption of the $\mathrm{G}^2\mathrm{M}$ method in real-world applications. + +Limitations and broader impacts: one limitation of the work is the normal distribution assumption on the fitting coefficients (e.g., $\beta_j^+$ and $\beta_j^-$ ). This is tied to the proposed UMP test statistics in Lemma 3.5 and the fundamental result in Dai et al. [9]. A possible future work would be to generalize this part beyond the normality assumption in light of existing work on the generalized linear model setting: e.g., Dai et al. [10]. Nonetheless, based on experimental results with synthetic, semi-synthetic or real case study, we demonstrate that despite having this limitation in the theoretical formulation, our model still outperformed the existing state-of-the-art methods, suggesting the importance of the work and its potential use cases in biological data where dimensionality and FDR are crucial aspects. + +# References + +[1] Divyansh Agarwal, Jingshu Wang, and Nancy R. Zhang. Data denoising and post-denoising corrections in single cell RNA sequencing. Statistical Science, 35(1):112-128, 2020. +[2] Ashwin N. Ananthakrishnan, Chengwei Luo, Vijay Yajnik, Hamed Khalili, John J. Garber, Betsy W. Stevens, Thomas Cleland, and Ramnik J. Xavier. Gut microbiome function predicts response to anti-integrin biologic therapy in inflammatory bowel diseases. Cell host & microbe, 21(5):603-610, 2017. +[3] Armin Askari, Quentin Rebjock, Alexandre d'Aspremont, and Laurent El Ghaoui. Fanok: Knockoffs in linear time. SIAM Journal on Mathematics of Data Science, 3(3):833-853, 2021. +[4] Rina Foygel Barber and Emmanuel J. Candès. Controlling the false discovery rate via knockoffs. The Annals of Statistics, 43(5):2055-2085, 2015. +[5] Cristina Bauset, Laura Gisbert-Ferrandiz, and Jesus Cosin-Roger. Metabolomics as a promising resource identifying potential biomarkers for inflammatory bowel disease. Journal of Clinical Medicine, 10(4):622, 2021. +[6] Shoaib Bin Masud, Conor Jenkins, Erika Hussey, Seth Elkin-Frankston, Phillip Mach, Elizabeth Dhummakupt, and Shuchin Aeron. Utilizing machine learning with knockoff filtering to extract significant metabolites in Crohn's disease with a publicly available untargeted metabolomics dataset. Plos one, 16(7):e0255240, 2021. +[7] P. A. Blaker, M. Arenas-Hernandez, M. A. Smith, E. A. Shobowale-Bakre, L. Fairbanks, P. M. Irving, J. D. Sanderson, and A. M. Marinaki. Mechanism of allopurinol induced TPMT inhibition. Biochemical pharmacology, 86(4):539-547, 2013. +[8] Emmanuel Candès, Yingying Fan, Lucas Janson, and Jinchi Lv. Panning for gold: 'Model-X' knockoffs for high dimensional controlled variable selection. Journal of the Royal Statistical Society: Series B (Statistical Methodology), 80(3):551-577, 2018. +[9] Chenguang Dai, Buyu Lin, Xin Xing, and Jun S Liu. False discovery rate control via data splitting. Journal of the American Statistical Association, 118(544):2503-2520, 2023. +[10] Chenguang Dai, Buyu Lin, Xin Xing, and Jun S Liu. A scale-free approach for false discovery rate control in generalized linear models. Journal of the American Statistical Association, 118 (543):1551-1565, 2023. +[11] DJ Fretland, DL Widomski, S Levin, and TS Gaginella. Colonic inflammation in the rabbit induced by phorbol-12-myristate-13-acetate. Inflammation, 14(2):143-150, 1990. +[12] Yeheng Ge, Sijia Zhang, and Xiao Zhang. False discovery rate control for high-dimensional cox model with uneven data splitting. Journal of Statistical Computation and Simulation, 94(7): 1462-1493, 2024. +[13] Derek Hansen, Brian Manzo, and Jeffrey Regier. Normalizing flows for knockoff-free controlled feature selection. In Advances in Neural Information Processing Systems, volume 35, pages 16125-16137, 2022. +[14] James Jordon, Jinsung Yoon, and Mihaela van der Schaar. KnockoffGAN: Generating knockoffs for feature selection using generative adversarial networks. In ICLR, 2018. +[15] Hon Wai Koon. A novel orally active metabolite reverses Crohn's disease-associated intestinal fibrosis. Inflammatory Bowel Diseases, 28(Supplement_1):S61-S62, 2022. +[16] Aonghus Lavelle and Harry Sokol. Gut microbiota-derived metabolites as key actors in inflammatory bowel disease. Nature reviews Gastroenterology & hepatology, 17(4):223-237, 2020. +[17] Thomas Lee, Thomas Clavel, Kirill Smirnov, Annemarie Schmidt, Ilias Lagkouvardos, Alesia Walker, Marianna Lucio, Bernhard Michalke, Philippe Schmitt-Kopplin, Richard Fedorak, et al. Oral versus intravenous iron replacement therapy distinctly alters the gut microbiota and metabolome in patients with IBD. Gut, 66(5):863-871, 2017. + +[18] Jundong Li, Kewei Cheng, Suhang Wang, Fred Morstatter, Robert P Trevino, Jiliang Tang, and Huan Liu. Feature selection: A data perspective. ACM computing surveys (CSUR), 50(6):1-45, 2017. +[19] Molei Liu, Eugene Katsevich, Lucas Janson, and Aaditya Ramdas. Fast and powerful conditional randomization testing via distillation. Biometrika, 109(2):277-293, 2022. +[20] Jason Lloyd-Price, Cesar Arze, Ashwin N. Ananthakrishnan, Melanie Schirmer, Julian Avila-Pacheco, Tiffany W. Poon, Elizabeth Andrews, Nadim J. Ajami, Kevin S. Bonham, Colin J. Brislawn, et al. Multi-omics of the gut microbial ecosystem in inflammatory bowel diseases. Nature, 569(7758):655–662, 2019. +[21] Yang Lu, Yingying Fan, Jinchi Lv, and William Stafford Noble. DeepPINK: reproducible feature selection in deep neural networks. In NeurIPS, volume 31, 2018. +[22] Shoaib Bin Masud, Matthew Werenski, James M Murphy, and Shuchin Aeron. Multivariate soft rank via entropy-regularized optimal transport: Sample efficiency and generative modeling. Journal of Machine Learning Research, 24(160):1-65, 2023. +[23] Rishi S. Mehta, Zachary L. Taylor, Lisa J. Martin, Michael J. Rosen, and Laura B. Ramsey. *SLCO1B1*15 allele is associated with methotrexate-induced nausea in pediatric patients with inflammatory bowel disease. *Clinical and translational science*, 15(1):63-69, 2022. +[24] Itta M. Minderhoud, Bas Oldenburg, Marguerite E. I. Schipper, Jose J. M. Ter Linde, and Melvin Samsom. Serotonin synthesis and uptake in symptomatic patients with Crohn's disease in remission. Clinical Gastroenterology and Hepatology, 5(6):714-720, 2007. +[25] Rajagopalan Lakshmi Narasimhan, Allison A. Throm, Jesvin Joy Koshy, Keith Metelo Raul Saldanha, Harikrishnan Chandranpillai, Rahul Deva Lal, Mausam Kumravat, Ajaya Kumar K. M., Aneesh Batra, Fei Zhong, et al. Inferring intestinal mucosal immune cell associated microbiome species and microbiota-derived metabolites in inflammatory bowel disease. bioRxiv, 2020. +[26] Binh T Nguyen, Bertrand Thirion, and Sylvain Arlot. A conditional randomization test for sparse logistic regression in high-dimension. Advances in Neural Information Processing Systems, 35:13691-13703, 2022. +[27] Andrea Nuzzo, Som Dutta Saha, Ellen Berg, Channa Jayawickreme, Joel Tocker, and James R. Brown. Expanding the drug discovery space with predicted metabolite-target interactions. Communications biology, 4(1):1-11, 2021. +[28] Xiaofa Qin. Etiology of inflammatory bowel disease: a unified hypothesis. World journal of gastroenterology: WJG, 18(15):1708, 2012. +[29] Zhimei Ren and Rina Foygel Barber. Derandomised knockoffs: leveraging e-values for false discovery rate control. Journal of the Royal Statistical Society Series B: Statistical Methodology, 86(1):122-154, 2024. +[30] Yaniv Romano, Matteo Sesia, and Emmanuel Candès. Deep knockoffs. Journal of the American Statistical Association, 115(532):1861-1872, 2020. +[31] Sameh Saber, Rania M Khalil, Walied S Abdo, Doaa Nassif, and Eman El-Ahwany. Olmesartan ameliorates chemically-induced ulcerative colitis in rats via modulating $\mathrm{NF}\kappa \mathrm{B}$ and Nrf-2/HO-1 signaling crosstalk. Toxicology and applied pharmacology, 364:120-132, 2019. +[32] Thorsten Schmidt. Coping with copulas. Copulas-From theory to application in finance, 3: 1-34, 2007. +[33] Elizabeth A. Scoville, Margaret M. Allaman, Caroline T. Brown, Amy K. Motley, Sara N. Horst, Christopher S. Williams, Tatsuki Koyama, Zhiguo Zhao, Dawn W. Adams, Dawn B. Beaulieu, et al. Alterations in lipid, amino acid, and energy metabolism distinguish Crohn's disease from ulcerative colitis and control subjects by serum metabolomic profiling. Metabolomics, 14(1): 1-12, 2018. + +[34] Hongyu Shen, Yici Yan, and Zhizhen Zhao. DeepDRK: Deep Dependency Regularized Knockoff for Feature Selection. In Advances in Neural Information Processing Systems, 2024. +[35] Johan D. Soderholm, Hans Oman, Lars Blomquist, Joggem Veen, Tuulikki Lindmark, and Gunnar Olaison. Reversible increase in tight junction permeability to macromolecules in rat ileal mucosa in vitro by sodium caprate, a constituent of milk fat. *Digestive diseases and sciences*, 43(7):1547–1552, 1998. +[36] Johan D. Söderholm, Gunnar Olaison, K. H. Peterson, L. E. Franzen, T. Lindmark, Mikael Wirén, Christer Tagesson, and Rune Sjödahl. Augmented increase in tight junction permeability by luminal stimuli in the non-inflamed ileum of Crohn's disease. Gut, 50(3):307-313, 2002. +[37] Asher Spector and Lucas Janson. Powerful knockoffs via minimizing reconstructability. The Annals of Statistics, 50(1):252-276, 2022. +[38] Mukund Sudarshan, Wesley Tansey, and Rajesh Ranganath. Deep direct likelihood knockoffs. In NeurIPS, volume 33, 2020. +[39] Karsten Suhre, So-Youn Shin, Ann-Kristin Petersen, Robert P. Mohney, David Meredith, Brigitte Wägele, Elisabeth Altmaier, Panos Deloukas, Jeanette Erdmann, Elin Grundberg, et al. Human metabolic individuality in biomedical and pharmaceutical research. Nature, 477(7362):54-60, 2011. +[40] Wesley Tansey, Victor Veitch, Haoran Zhang, Raul Rabadan, and David M Blei. The holdout randomization test for feature selection in black box models. Journal of Computational and Graphical Statistics, 31(1):151-162, 2022. +[41] Kan Uchiyama, Shunichi Odahara, Makoto Nakamura, Shigeo Koido, Kiyohiko Katahira, Hiromi Shiraishi, Toshifumi Ohkusa, Kiyotaka Fujise, and Hisao Tajiri. The fatty acid profile of the erythrocyte membrane in initial-onset inflammatory bowel disease patients. Digestive diseases and sciences, 58(5):1235-1243, 2013. +[42] Victor Uko, Suraj Thangada, and Kadakkal Radhakrishnan. Liver disorders in inflammatory bowel disease. Gastroenterology research and practice, 2012(1):642923, 2012. +[43] W. Wolberg, O. Mangasarian, N. Street, and W. Street. Breast cancer wisconsin (diagnostic) [dataset]. UCI Machine Learning Repository, 1993. URL https://archive.ics.uci.edu/ dataset/17/breast+cancer+wisconsin+diagnostic. +[44] Xin Xing, Yu Gui, Chenguang Dai, and Jun S Liu. Neural gaussian mirror for controlled feature selection in neural networks. arXiv preprint arXiv:2010.06175, 2020. +[45] Xin Xing, Zhigen Zhao, and Jun S Liu. Controlling false discovery rate using gaussian mirrors. Journal of the American Statistical Association, 118(541):222-241, 2023. +[46] Yanjie Zhong, Todd Kuffner, and Soumendra Lahiri. Conditional randomization rank test. arXiv preprint arXiv:2112.00258, 2021. + +# Appendix: $\mathbf{G}^2\mathbf{M}$ : A Generalized Gaussian Mirror Method to boost feature selection power + +This appendix is structured as follows: in Appendix 2 we provide details on related work and the connection with our method; in Appendix A we show the empirical observation of the variance of Gaussian mirror statistics; in Appendix B we provide proofs for the theorems and lemmas in the main paper; in Appendix E we provide additional experimental results; in Appendix F we detail the RNA data preparation for experiment in Section 4.2; in Appendix G we provide additional information for the IBD study. + +# A Empirical Evidence on Variance Differences in Gaussian Mirror Statistics + +This section provides empirical results about the variance (e.g. standard deviation) of the Gaussian mirror coefficients $\beta_{j}^{+}$ and $\beta_{j}^{-}$ given different design matrices. To start, we randomly generate 10000 samples of the standard deviations of $\beta_{j}^{+}$ and $\beta_{j}^{-}$ for both null and nonnull cases, respectively. Specifically, each sample is generated by first sampling a design matrix and then performing the Gaussian mirror perturbation based on a randomly sampled Gaussian vector $z_{j}$ . Both the design matrix $X$ and the sampled vector $z_{j}$ are used to calculate the corresponding $c_{j} = \frac{\|P_{\perp - j}x_{j}\|}{\|P_{\perp - j}z_{j}\|}$ a component in the Gassuin mirror. The feature $x_{j}$ is uniformly sampled on the index $j$ given either null or nonnull. With the information of $x_{j}, c_{j}$ and $z_{j}$ , we are able to calculate the standard deviation of $\beta_{j}^{+}$ and $\beta_{j}^{-}$ according to Corollary 3.4 and Eq. (2). + +We present the histograms of the results with Gaussian design matrix in Figure 1 and 2, Gaussian mixture design matrix in Figure 3 and 4, and IBD design matrix in Figure 5 and 6. The description on the Gaussian, Gaussian mixture and IBD design matrices can be found in Section 4. Clearly, the standard deviations for both null and nonnull coefficients are distributions that are not concentrated at 1, indicating the unrealistic assumption about the unit variance outlined in Dai et al. [9]. + +![](images/abccfe43df29ce719dda1dff26df2fae80f21d1bfcb7ee20e21704b8801be64f.jpg) +Figure 1: Histogram of the standard deviation of $\beta_j^+$ and $\beta_j^-$ for null variables over 10000 samples. The design matrix $X$ is based on Gaussian distributions. + +![](images/1378a5a90cd5644cbfa4566cff48de053da9cc8b9ed04dd512810cf33db8fce1.jpg) + +![](images/e8cdea18abe2a9898853d2fc2c92b0433e4bd4e01f5db5ba4c46c3f8e3d14db0.jpg) +Figure 2: Histogram of the standard deviation of $\beta_{j}^{+}$ and $\beta_{j}^{-}$ for nonnull variables over 10000 samples. The design matrix $X$ is based on Gaussian distributions. + +![](images/2901bb18da69f7117ca5b0d43809f50e083874ed12ecb3b6b7feb29f9526d72e.jpg) + +![](images/3ec5ccf0b0e6e27cd1516d5cb98238e951dad90a6c17147dc404c3717efc5fa8.jpg) +Figure 3: Histogram of the standard deviation of $\beta_{j}^{+}$ and $\beta_{j}^{-}$ for null variables over 10000 samples. The design matrix $X$ is based on Gaussian mixture distributions. + +![](images/4ce6e714c2f1e5e635fbcb1c55a3a4b3b210b5ef39790423cdd7123f16d67f3c.jpg) + +![](images/5d94468166b89f4966342b7fe4733c24ba50b0f9814010d4d2139adca65f3b3b.jpg) +Figure 4: Histogram of the standard deviation of $\beta_{j}^{+}$ and $\beta_{j}^{-}$ for nonnull variables over 10000 samples. The design matrix $X$ is based on Gaussian mixture distributions. + +![](images/5494f6847141ebf127a002d9e61d9e56e10364823d03e168458455db556c1278.jpg) + +![](images/268adf60f94556f0aac45cad77a802a3e47d3ac503a75e72a9906785ffbe07a9.jpg) +Figure 5: Histogram of the standard deviation of $\beta_j^+$ and $\beta_j^-$ for null variables over 10000 samples. The design matrix $X$ is based on the IBD dataset. + +![](images/79e5579526c53b1190ad1971d60ff0aa6c6f2933a191f43e0c08f76e602ae9ed.jpg) + +![](images/6a186dc0fdfb2e93c7991988d96ba07de2b3183de7063c9ccc8e167345002510.jpg) +Figure 6: Histogram of the standard deviation of $\beta_{j}^{+}$ and $\beta_{j}^{-}$ for nonnull variables over 10000 samples. The design matrix $X$ is based on the IBD dataset. + +![](images/e56c7dfb968672a9adfbcf923d5c113692d7c3cedcd7504f083d7144d2065ba8.jpg) + +# B Proof + +# B.1 Proposition 3.1 + +To show proposition 3.1, we first consider two models. The original model that produces the response $y$ and the new model that considers the inclusion of $x_{j}^{+}$ and $x_{j}^{-}$ in place of the original $x_{j}$ . In the following, we have the original model: + +$$ +y = X \beta + \epsilon . +$$ + +After modification of the design matrix, we have the new design matrix: + +$$ +X _ {\text {n e w}} = \left[ X _ {- j}, x _ {j} ^ {+}, x _ {j} ^ {-} \right], +$$ + +and the new coefficient vector: + +$$ +\beta_ {\mathrm {n e w}} = \left[ \begin{array}{c} \beta_ {- j} ^ {\mathrm {n e w}} \\ \beta_ {j} ^ {+} \\ \beta_ {- j} ^ {-}. \end{array} \right] +$$ + +To perform the least square fit with the new design matrix and the original response variable $y$ , we can expand normal equations $X_{\mathrm{new}}^{\top}X_{\mathrm{new}} / \beta_{\mathrm{new}} = X_{\mathrm{new}}^{\top}y$ : + +$$ +\left[ \begin{array}{c c c} X _ {- j} ^ {\top} X _ {- j} & X _ {- j} ^ {\top} x _ {j} ^ {+} & X _ {- j} ^ {\top} x _ {j} ^ {-} \\ (x _ {j} ^ {+}) ^ {\top} X _ {- j} & (x _ {j} ^ {+}) ^ {\top} x _ {j} ^ {+} & (x _ {j} ^ {+}) ^ {\top} x _ {j} ^ {-} \\ (x _ {j} ^ {-}) ^ {\top} X _ {- j} & (x _ {j} ^ {-}) ^ {\top} x _ {j} ^ {+} & (x _ {j} ^ {-}) ^ {\top} x _ {j} ^ {-} \end{array} \right] \left[ \begin{array}{c} \beta_ {- j} ^ {\text {n e w}} \\ \beta_ {j} ^ {+} \\ \beta_ {j} ^ {-} \end{array} \right] = \left[ \begin{array}{c} X _ {- j} ^ {\top} y \\ (x _ {j} ^ {+}) ^ {\top} y \\ (x _ {j} ^ {-}) ^ {\top} y \end{array} \right]. +$$ + +We use $\beta_{-j}^{\mathrm{new}}$ to distinguish from the original $\beta_{-j}$ that represents the coefficients in the original linear model with $j$ -th entry removed. Expanding the normal equation we have: + +$$ +\begin{array}{l} \left(X _ {- j} ^ {\top} x _ {j} + c _ {j} X _ {- j} ^ {\top} z _ {j}\right) \cdot \beta_ {j} ^ {+} + \left(X _ {- j} ^ {\top} x _ {j} - d _ {j} X _ {- j} ^ {\top} q _ {j}\right) \cdot \beta_ {j} ^ {-} = X _ {- j} ^ {\top} X _ {- j} \cdot \left(\beta_ {- j} - \beta_ {- j} ^ {\text {n e w}}\right) + X _ {- j} ^ {\top} x _ {j} \cdot \beta_ {j} + X _ {- j} ^ {\top} \epsilon , \\ (x _ {j} ^ {+ \top} X _ {- j}) \cdot \beta_ {- j} ^ {\text {n e w}} + (x _ {j} ^ {+ \top} x _ {j} ^ {+}) \cdot \beta_ {j} ^ {+} + (x _ {j} ^ {+ \top} x _ {j} ^ {-}) \cdot \beta_ {j} ^ {-} = x _ {j} ^ {+ \top} X _ {- j} \cdot \beta_ {j} + x _ {j} ^ {+ \top} x _ {j} \cdot \beta_ {j} + x _ {j} ^ {+ \top} \epsilon , \\ (x _ {j} ^ {- \top} X _ {- j}) \cdot \beta_ {- j} ^ {\text {n e w}} + (x _ {j} ^ {- \top} x _ {j} ^ {+}) \cdot \beta_ {j} ^ {+} + (x _ {j} ^ {- \top} x _ {j} ^ {-}) \cdot \beta_ {j} ^ {-} = x _ {j} ^ {- \top} X _ {- j} \cdot \beta_ {j} + x _ {j} ^ {- \top} x _ {j} \cdot \beta_ {j} + x _ {j} ^ {- \top} \epsilon . \\ \end{array} +$$ + +Since we have three equations and three unknowns: $\beta_{-j}^{\mathrm{new}}$ , $\beta_j^+$ and $\beta_j^-$ , we cancel out $\beta_{-j}^{\mathrm{new}}$ , leaving two functions that are the functions of $\beta_j^+$ and $\beta_j^-$ . Eventually, we obtain the final form that represents $\beta_j^+$ and $\beta_j^-$ as a linear function of the true beta $a_{j}$ and the noise $\epsilon$ : + +$$ +\beta_ {j} ^ {+} = \alpha \cdot \beta_ {j} + \gamma^ {\top} \cdot \epsilon , +$$ + +$$ +\beta_ {j} ^ {-} = \zeta \cdot \beta_ {j} + \eta^ {\top} \cdot \epsilon , +$$ + +where + +$$ +\alpha = \frac {N _ {j ^ {-}} \cdot F - M _ {j ^ {-}} \cdot H}{L}, \quad \gamma = \frac {N _ {j ^ {-}} \cdot G - M _ {j ^ {-}} \cdot K}{L}, +$$ + +$$ +\zeta = \frac {- N _ {j +} \cdot F + M _ {j +} \cdot H}{L}, \quad \eta = \frac {- N _ {j +} \cdot G + M _ {j +} \cdot K}{L}. +$$ + +All the involved variables are presented below: + +$$ +L = M _ {j +} \cdot N _ {j -} - M _ {j -} \cdot N _ {j +} +$$ + +$$ +M _ {j ^ {+}} = \left(x _ {j} ^ {\top} x _ {j} + 2 c z _ {j} ^ {\top} x _ {j} + c _ {j} ^ {2} z _ {j} ^ {\top} z _ {j}\right) - \left(x _ {j} ^ {\top} X _ {- j} + c _ {j} z _ {j} ^ {\top} X _ {- j}\right) A ^ {- 1} B, +$$ + +$$ +M _ {j -} = \left(x _ {j} ^ {\top} x _ {j} - d _ {j} x _ {j} ^ {\top} q _ {j} + c _ {j} z _ {j} ^ {\top} x _ {j} - c _ {j} d _ {j} z _ {j} ^ {\top} q _ {j}\right) - \left(x _ {j} ^ {\top} X _ {- j} + c _ {j} z _ {j} ^ {\top} X _ {- j}\right) A ^ {- 1} C, +$$ + +$$ +N _ {j ^ {+}} = \left(x _ {j} ^ {\top} x _ {j} + c _ {j} z _ {j} ^ {\top} x _ {j} - d _ {j} q _ {j} ^ {\top} x _ {j} - c _ {j} d _ {j} q _ {j} ^ {\top} z _ {j}\right) - \left(x _ {j} ^ {\top} X _ {- j} - d _ {j} q _ {j} ^ {\top} X _ {- j}\right) A ^ {- 1} B, +$$ + +$$ +N _ {j ^ {-}} = \left(x _ {j} ^ {\top} x _ {j} - 2 d q _ {j} ^ {\top} x _ {j} + d _ {j} ^ {2} q _ {j} ^ {\top} q _ {j}\right) - \left(x _ {j} ^ {\top} X _ {- j} - d _ {j} q _ {j} ^ {\top} X _ {- j}\right) A ^ {- 1} C, +$$ + +$$ +F = \left(x _ {j} + c _ {j} z _ {j}\right) ^ {\top} x _ {j} - \left(x _ {j} ^ {\top} X _ {- j} + c _ {j} z _ {j} ^ {\top} X _ {- j}\right) A ^ {- 1} D, +$$ + +$$ +G = \left(x _ {j} + c _ {j} z _ {j}\right) ^ {\top} \left(X _ {- j} - c _ {j} z _ {j} ^ {\top} X _ {- j}\right) A ^ {- 1} X _ {- j} ^ {\top}, +$$ + +$$ +H = \left(x _ {j} - d _ {j} q _ {j}\right) ^ {\top} x _ {j} - \left(x _ {j} ^ {\top} X _ {- j} - d _ {j} q _ {j} ^ {\top} X _ {- j}\right) A ^ {- 1} D, +$$ + +$$ +K = \left(x _ {j} - d _ {j} q _ {j}\right) ^ {\top} \left(X _ {- j} - d _ {j} q _ {j} ^ {\top} X _ {- j}\right) A ^ {- 1} X _ {- j} ^ {\top}, +$$ + +$$ +\begin{array}{l} A = X _ {- j} ^ {\top} X _ {- j}, \\ B = X _ {- j} ^ {\top} x _ {j} + c _ {j} X _ {- j} ^ {\top} z _ {j}, \\ C = X _ {- j} ^ {\top} x _ {j} - d _ {j} X _ {- j} ^ {\top} q _ {j}, \\ D = X _ {- j} ^ {\top} x _ {j}, \\ E = X _ {- j} ^ {\top} \epsilon . \\ \end{array} +$$ + +# B.2 Corollary 3.2 + +To prove, we only need to show $L = M_{j^{+}} \cdot N_{j^{-}} - M_{j^{-}} \cdot N_{j^{+}}$ and $N_{j^{-}} \cdot F - M_{j^{-}} \cdot H + -N_{j^{+}} \cdot F + M_{j^{+}} \cdot H$ are the same. We first expand + +$$ +\begin{array}{l} M _ {j ^ {+}} N _ {j ^ {-}} - M _ {j ^ {-}} N _ {j ^ {+}} = \\ - 2 c d \left(x _ {j} ^ {\top} x _ {j}\right) \left(q _ {j} ^ {\top} x _ {j}\right) - 2 c d \left(z _ {j} ^ {\top} x _ {j}\right) \left(q _ {j} ^ {\top} x _ {j}\right) - 2 d ^ {2} \left(x _ {j} ^ {\top} q _ {j}\right) ^ {2} + d _ {j} ^ {2} \left(q _ {j} ^ {\top} q _ {j}\right) \left(x _ {j} ^ {\top} x _ {j}\right) \\ + c _ {j} ^ {2} \left(z _ {j} ^ {\top} z _ {j}\right) \left(x _ {j} ^ {\top} x _ {j}\right) - c _ {j} ^ {2} \left(z _ {j} ^ {\top} x _ {j}\right) ^ {2} - c _ {j} ^ {2} d _ {j} \left(z _ {j} ^ {\top} z _ {j}\right) \left(x _ {j} ^ {\top} q _ {j}\right) + 2 c d \left(z _ {j} ^ {\top} q _ {j}\right) \left(x _ {j} ^ {\top} x _ {j}\right) \\ + d _ {j} ^ {2} c _ {j} (q _ {j} ^ {\top} q _ {j}) (z _ {j} ^ {\top} x _ {j}) + c _ {j} ^ {2} d _ {j} (z _ {j} ^ {\top} q _ {j}) (z _ {j} ^ {\top} x _ {j}) - c d ^ {2} (z _ {j} ^ {\top} q _ {j}) (x _ {j} ^ {\top} q _ {j}) \\ + \left[ \text {T e r m s} A ^ {- 1} \text {a n d h i g h e r o r d e r s} \right]. \\ \end{array} +$$ + +Similarly, we expand + +$$ +N _ {j ^ {-}} \cdot F - M _ {j ^ {-}} \cdot H + - N _ {j ^ {+}} \cdot F + M _ {j ^ {+}} \cdot H = +$$ + +$$ +(M _ {j ^ {+}} - M _ {j ^ {-}}) H + (N _ {j ^ {-}} - N _ {j ^ {+}}) F = +$$ + +$$ +\begin{array}{l} c _ {j} (z _ {j} ^ {\top} x _ {j}) (x _ {j} ^ {\top} x _ {j}) - c d (z _ {j} ^ {\top} x _ {j}) (x _ {j} ^ {\top} q _ {j}) + c _ {j} ^ {2} (z _ {j} ^ {\top} z _ {j}) (x _ {j} ^ {\top} x _ {j}) - c _ {j} ^ {2} d _ {j} (z _ {j} ^ {\top} z _ {j}) (x _ {j} ^ {\top} q _ {j}) \\ + d _ {j} \left(x _ {j} ^ {\top} q _ {j}\right) \left(x _ {j} ^ {\top} x _ {j}\right) - d _ {j} ^ {2} \left(x _ {j} ^ {\top} q _ {j}\right) ^ {2} + c d \left(z _ {j} ^ {\top} q _ {j}\right) \left(x _ {j} ^ {\top} x _ {j}\right) - c d ^ {2} \left(z _ {j} ^ {\top} q _ {j}\right) \left(x _ {j} ^ {\top} q _ {j}\right) \\ - d _ {j} \left(q _ {j} ^ {\top} x _ {j}\right) \left(x _ {j} ^ {\top} x _ {j}\right) - d c \left(q _ {j} ^ {\top} x _ {j}\right) \left(z _ {j} ^ {\top} x _ {j}\right) + c d \left(z _ {j} ^ {\top} q _ {j}\right) \left(x _ {j} ^ {\top} x _ {j}\right) + d _ {j} ^ {2} \left(q _ {j} ^ {\top} q _ {j}\right) \left(x _ {j} ^ {\top} x _ {j}\right) \\ + d _ {j} ^ {2} c _ {j} (q _ {j} ^ {\top} q _ {j}) (z _ {j} ^ {\top} x _ {j}) - c _ {j} ^ {2} (z _ {j} ^ {\top} x _ {j}) ^ {2} - c _ {j} (z _ {j} ^ {\top} x _ {j}) (x _ {j} ^ {\top} x _ {j}) + c _ {j} ^ {2} d _ {j} (z _ {j} ^ {\top} q _ {j}) (z _ {j} ^ {\top} x _ {j}) \\ \end{array} +$$ + +[Terms involving $A^{-1}$ and higher orders]. + +We can find matches for every term, proving the equality. + +# B.3 Corollary 3.3 + +The proof is similar to the one in Corollary 3.2, we expand terms and show the match. In this proof, we need to show $N_{j-} \cdot F - M_{j-} \cdot H = -N_{j+} \cdot F + M_{j+} \cdot H$ . We expand both sides and compare, which proves the corollary. In the following, we have: + +$$ +\begin{array}{l} - N _ {j +} \cdot F + M _ {j +} \cdot H = \\ c _ {j} ^ {2} \bigg [ (z _ {j} ^ {\top} z _ {j}) (x _ {j} ^ {\top} x _ {j}) - (z _ {j} ^ {\top} x _ {j}) ^ {2} \bigg ] \\ - c d \left(q _ {j} ^ {\top} x _ {j}\right) \left(z _ {j} ^ {\top} x _ {j}\right) + c d \left(q _ {j} ^ {\top} z _ {j}\right) \left(x _ {j} ^ {\top} x _ {j}\right) \\ + c _ {j} ^ {2} d _ {j} \left(q _ {j} ^ {\top} z _ {j}\right) \left(z _ {j} ^ {\top} x _ {j}\right) - c _ {j} ^ {2} d _ {j} \left(z _ {j} ^ {\top} z _ {j}\right) \left(q _ {j} ^ {\top} x _ {j}\right) \\ + \left[ \text {C o m p l e x} X _ {- j}, A ^ {- 1} \right], \\ \end{array} +$$ + +$$ +N _ {j ^ {-}} \cdot F - M _ {j ^ {-}} \cdot H = +$$ + +$$ +\begin{array}{l} \left[ - c d \left(q _ {j} ^ {\top} x _ {j}\right) \left(z _ {j} ^ {\top} x _ {j}\right) + c d \left(z _ {j} ^ {\top} q _ {j}\right) \left(x _ {j} ^ {\top} x _ {j}\right) \right] \\ + d _ {j} ^ {2} \bigg [ (q _ {j} ^ {\top} q _ {j}) (x _ {j} ^ {\top} x _ {j}) - (q _ {j} ^ {\top} x _ {j}) ^ {2} \bigg ] \\ + c d ^ {2} \left[ (q _ {j} ^ {\top} q _ {j}) (z _ {j} ^ {\top} x _ {j}) - (q _ {j} ^ {\top} x _ {j}) (z _ {j} ^ {\top} q _ {j}) \right] \\ + \left[ \text {C o m p l e x t e r m s i n v o l i n g} X _ {- j}, A ^ {- 1} \right]. \\ \end{array} +$$ + +After considering $c_{j} = d_{j}$ and $z_{j} = q_{j}$ , we can verify the two functions are equal, completing the proof. + +Remark B.1. Note that the Gaussian mirror work [45] only presents the $\beta_j^+ = \beta_j^- = 0.5$ without providing the formal proof. Combining Corollary 3.2 and Corollary 3.3, we can provide formal evidence for this statement. + +# B.4 Lemma 3.5 + +This proof is inspired by [9] for a more general setting. First, we identify the general form of the statistics following the proof in [9]. We then provide explicitly the form of $f$ (e.g., in Eq. (3)) according to the Neyman-Pearson lemma. The proof of the latter is omitted in [9]. We hope this proof can serve as a complement, which provides insights for people who are interested in any future extension on the optimal test statistics in this more general setting. + +To start with, let $Z_{1}, Z_{2} \sim N(\omega, \sigma_{1}^{2}), N(\omega, \sigma_{2}^{2})$ , respectively. $Z_{3}, Z_{4} \sim N(0, \sigma_{3}^{2})$ and $N(0, \sigma_{4}^{2})$ , respectively13. And $\omega \sim \delta \cdot \mathrm{Rademacher}(0.5)$ , $\delta > 0$ . All variables are independent. Following [9], we assume that the designated FDR control level $q \in (0, 1)$ satisfies $\frac{rq}{1 - q} < 1$ , otherwise selecting all features would maximize the power and also achieve asymptotic FDR control. Let $f_{\mathrm{opt}}(u, v)$ be the optimal choice, and let $S_{\mathrm{opt}}$ be the optimal selection result that achieves asymptotic FDR control. By the law of large numbers, we have: + +$$ +\lim _ {p \rightarrow \infty} \frac {\# \{j : j \in S _ {0} , j \in S _ {\text {o p t}} \}}{\# \{j : j \in S _ {\text {o p t}} \}} = \frac {P (j \in S _ {\text {o p t}} | j \in S _ {0})}{P (j \in S _ {\text {o p t}} | j \in S _ {0}) + r P (j \in S _ {\text {o p t}} | j \in S _ {1})} \leq q, \tag {5} +$$ + +in which the numerator is the type-I error. More precisely: + +$$ +P (j \in S _ {\mathrm {o p t}} | j \in S _ {1}) = P (\operatorname {s i g n} (Z _ {1} Z _ {2}) f _ {\mathrm {o p t}} (| Z _ {1} |, | Z _ {2} |) > t _ {\mathrm {o p t}}). +$$ + +$$ +P (j \in S _ {\text {o p t}} | j \in S _ {0}) = P (\text {s i g n} (Z _ {3} Z _ {4}) f _ {\text {o p t}} (| Z _ {3} |, | Z _ {4} |) > t _ {\text {o p t}}), +$$ + +Here, $t_{\mathrm{opt}} > 0$ is the cutoff that maximizes the power $P(j \in S_{\mathrm{opt}} | j \in S_1)$ , under the constraint that Eq. (5) holds. + +In practice, we consider testing whether the covariate $X_{j}$ is a null feature, with the significance level $\alpha$ specified as: + +$$ +\alpha = \frac {r q}{1 - q} P (j \in S _ {\mathrm {o p t}} | j \in S _ {1}) < 1. +$$ + +Given two realizations $\beta_j^+$ and $\beta_j^-$ , we consider their random variables independently following the normal distributions described by $Z_1$ , $Z_2$ if $X_j$ is a nonnull feature, and by $Z_3$ , $Z_4$ otherwise. According to Eq. (5), the test which rejects the null hypothesis (i.e., $j \in S_{\mathrm{opt}}$ ) if: + +$$ +\operatorname {s i g n} \left(\beta_ {j} ^ {+} \beta_ {j} ^ {-}\right) f _ {\text {o p t}} \left(\left| \beta_ {j} ^ {+} \right|, \left| \beta_ {j} ^ {-} \right|\right) > t _ {\text {o p t}} \tag {6} +$$ + +achieves the significance level $\alpha$ . + +We consider the following rejection rule with a certain form of $f$ that will be revealed later in the proof: + +$$ +\operatorname {s i g n} \left(\beta_ {j} ^ {+} \beta_ {j} ^ {-}\right) f \left(\left| \beta_ {j} ^ {+} \right|, \left| \beta_ {j} ^ {-} \right|\right) > t _ {\text {l i k}}, +$$ + +in which $t_{\mathrm{lik}} > 0$ satisfies: + +$$ +P (f (| Z _ {3} |, | Z _ {4} |) > t _ {\text {l i k}} \mid \operatorname {s i g n} (Z _ {3}) = \operatorname {s i g n} (Z _ {4})) = 2 \alpha . +$$ + +Let $S_{\mathrm{lik}}$ be the corresponding selection set. We first show that this rejection rule controls the type-I error below $\alpha$ . Indeed: + +$$ +P (j \in S _ {\mathrm {l i k}} | j \in S _ {0}) = \frac {1}{2} P (f (| \beta_ {j} ^ {+} |, | \beta_ {j} ^ {-} |) > t _ {\mathrm {l i k}} | j \in S _ {0}, \mathrm {s i g n} (\beta_ {j} ^ {+}) = \mathrm {s i g n} (\beta_ {j} ^ {-})) = \alpha . +$$ + +In terms of power, we have: + +$$ +\begin{array}{l} P (j \in S _ {\text {l i k}} | j \in S _ {1}) = p _ {w} P (f (| \beta_ {j} ^ {+} |, | \beta_ {j} ^ {-} |) > t _ {\text {l i k}} | j \in S _ {1}, \operatorname {s i g n} (\beta_ {j} ^ {+}) = \operatorname {s i g n} (\beta_ {j} ^ {-})) \\ \geq p _ {w} P \left(f _ {\text {o p t}} \left(\left| \beta_ {j} ^ {+} \right|, \left| \beta_ {j} ^ {-} \right|\right) > t _ {\text {o p t}} | j \in S _ {1}, \operatorname {s i g n} \left(\beta_ {j} ^ {+}\right) = \operatorname {s i g n} \left(\beta_ {j} ^ {-}\right)\right) = P (j \in S _ {\text {o p t}} | j \in S _ {1}), \\ \end{array} +$$ + +where $p_w = P\left(\mathrm{sign}(\beta_j^{(1)}) = \mathrm{sign}(\beta_j^{(2)}) \mid j \in S_1\right)$ . + +In the following, we show the proof of the inequality and the exact form of the function $f$ via the Neyman-Pearson lemma. + +Since we consider a function $f$ that takes the realizations as inputs (e.g., Eq. (6)), we first need to study the distribution of $|\beta_j^+|, |\beta_j^-|$ conditioning on the equal sign constraint: $\mathrm{sign}(\beta_j^+) = \mathrm{sign}(\beta_j^-)$ . + +We first define the event $E$ to represent $\mathrm{sign}(\beta_j^+) = \mathrm{sign}(\beta_j^-)$ . Then the events would have two cases: 1. both terms are positive; 2. both terms are negative. Under $H_0$ , since $\beta_j^+$ and $\beta_j^-$ are independent and symmetric around zero. Therefore: + +$$ +P _ {E | H _ {0}} = P _ {H _ {0}} (\beta_ {j} ^ {+} > 0, \beta_ {j} ^ {-} > 0) + P _ {H _ {0}} (\beta_ {j} ^ {+} < 0, \beta_ {j} ^ {-} < 0) = 0. 2 5 + 0. 2 5 = 0. 5. +$$ + +Under $H_{1}$ , on the other hand, we have: + +$$ +P _ {E \mid H _ {1}} = 0. 5 \cdot P _ {E \mid w = \delta} + 0. 5 \cdot P _ {E \mid w = - \delta}, +$$ + +where + +$$ +P _ {E | w = \delta} = P _ {E | w = - \delta} = \Phi \left(\frac {\delta}{\sigma_ {1}}\right) \Phi \left(\frac {\delta}{\sigma_ {2}}\right) + \left[ 1 - \Phi \left(\frac {\delta}{\sigma_ {1}}\right) \right] \left[ 1 - \Phi \left(\frac {\delta}{\sigma_ {2}}\right) \right], +$$ + +with $\Phi$ being the cumulative density function of the standard normal distribution. Combining the two equations above, we obtain: + +$$ +P _ {E \mid H _ {1}} = \Phi \left(\frac {\delta}{\sigma_ {1}}\right) \Phi \left(\frac {\delta}{\sigma_ {2}}\right) + \left[ 1 - \Phi \left(\frac {\delta}{\sigma_ {1}}\right) \right] \left[ 1 - \Phi \left(\frac {\delta}{\sigma_ {2}}\right) \right]. +$$ + +After characterizing the probability of the equal sign event, we consider two conditional density functions under $H_0$ and $H_1$ , respectively. Specifically under $H_0$ : + +$$ +f _ {| \beta_ {j} ^ {+} |, | \beta_ {j} ^ {-} | | E, H _ {0}} (a, b) = \frac {P _ {H _ {0}} (\beta_ {j} ^ {+} > 0 , \beta_ {j} ^ {-} > 0) \cdot f _ {H _ {0}} (a , b) + P _ {H _ {0}} (\beta_ {j} ^ {+} < 0 , \beta_ {j} ^ {-} < 0) \cdot f _ {H _ {0}} (- a , - b)}{P _ {E | H _ {0}}}, +$$ + +where $f_{H_0}(\beta_j^+, \beta_j^-) = f_{\beta_j^+}(\beta_j^+; 0, \sigma_3^2) \cdot f_{\beta_j^-}(\beta_j^-; 0, \sigma_4^2)$ . + +Similarly under $H_{1}$ , we have: + +$$ +f _ {| \beta_ {j} ^ {+} |, | \beta_ {j} ^ {-} | | E, H _ {1}} (a, b) = \frac {P _ {+ +} \cdot f _ {+ +} (a , b) + P _ {- -} \cdot f _ {- -} (a , b)}{P _ {E | H _ {1}}}, +$$ + +where: + +$$ +f _ {+ +} (a, b) = f _ {\beta_ {j} ^ {+}} (a; \delta , \sigma_ {1} ^ {2}) \cdot f _ {\beta_ {j} ^ {-}} (b; \delta , \sigma_ {2} ^ {2}), +$$ + +$$ +f _ {- -} (a, b) = f _ {\beta_ {j} ^ {+}} (a; - \delta , \sigma_ {1} ^ {2}) \cdot f _ {\beta_ {j} ^ {-}} (b; - \delta , \sigma_ {2} ^ {2}), +$$ + +$$ +P _ {+ +} = P _ {- -} = \frac {1}{2} \left[ \Phi \left(\frac {\delta}{\sigma_ {1}}\right) \Phi \left(\frac {\delta}{\sigma_ {2}}\right) + \Phi \left(- \frac {\delta}{\sigma_ {1}}\right) \Phi \left(- \frac {\delta}{\sigma_ {2}}\right) \right]. +$$ + +Overall, the likelihood ratio between $H_{1}$ and $H_{0}$ can be represented as: + +$$ +\Lambda (a, b) = \frac {f _ {| \beta_ {j} ^ {+} | , | \beta_ {j} ^ {-} | | E , H _ {1}} (a , b)}{f _ {| \beta_ {j} ^ {+} | , | \beta_ {j} ^ {-} | | E , H _ {0}} (a , b)} +$$ + +$$ += U \cdot \left[ P _ {+} \exp (- S _ {-}) + P _ {-} \exp (- S _ {+}) \right] +$$ + +where + +$$ +U = \left(\frac {1}{P _ {E | H _ {1}}}\right) \left(\frac {\sigma_ {3} \sigma_ {4}}{\sigma_ {1} \sigma_ {2}}\right), +$$ + +$$ +S _ {-} = \frac {(a - \delta) ^ {2}}{2 \sigma_ {1} ^ {2}} + \frac {(b - \delta) ^ {2}}{2 \sigma_ {2} ^ {2}} - \left[ \frac {a ^ {2}}{2 \sigma_ {3} ^ {2}} + \frac {b ^ {2}}{2 \sigma_ {4} ^ {2}} \right], +$$ + +$$ +S _ {+} = \frac {(a + \delta) ^ {2}}{2 \sigma_ {1} ^ {2}} + \frac {(b + \delta) ^ {2}}{2 \sigma_ {2} ^ {2}} - \left[ \frac {a ^ {2}}{2 \sigma_ {3} ^ {2}} + \frac {b ^ {2}}{2 \sigma_ {4} ^ {2}} \right], +$$ + +$$ +P _ {+} = \Phi \left(\frac {\delta}{\sigma_ {1}}\right) \Phi \left(\frac {\delta}{\sigma_ {2}}\right), +$$ + +$$ +P _ {-} = \left[ 1 - \Phi \left(\frac {\delta}{\sigma_ {1}}\right) \right] \left[ 1 - \Phi \left(\frac {\delta}{\sigma_ {2}}\right) \right]. +$$ + +In addition, we let $\sigma_{1} = \sigma_{3} = \sigma_{a}$ and $\sigma_{2} = \sigma_{4} = \sigma_{b}$ , leading to: + +$$ +U = \left(\frac {1}{P _ {E | H _ {1}}}\right), +$$ + +$$ +S _ {-} = \frac {\delta (\delta - 2 a)}{2 \sigma_ {a} ^ {2}} + \frac {\delta (\delta - 2 b)}{2 \sigma_ {b} ^ {2}}, +$$ + +$$ +S _ {+} = \frac {\delta (\delta + 2 a)}{2 \sigma_ {a} ^ {2}} + \frac {\delta (\delta + 2 b)}{2 \sigma_ {b} ^ {2}}, +$$ + +$$ +P _ {+} = \Phi \left(\frac {\delta}{\sigma_ {a}}\right) \Phi \left(\frac {\delta}{\sigma_ {b}}\right), +$$ + +$$ +P _ {-} = \left[ 1 - \Phi \left(\frac {\delta}{\sigma_ {a}}\right) \right] \left[ 1 - \Phi \left(\frac {\delta}{\sigma_ {b}}\right) \right]. +$$ + +Letting $\Lambda(a, b)$ be $f(a, b)$ completes the proof, as $\Lambda(a, b)$ is the UMP test statistic according to the Neyman-Pearson lemma. + +# B.5 Theorem 3.6 + +To show that the proposed test statistic $\mathrm{sign}(ab)f(a,b)$ (Eq. (3)) is better than the Gaussian mirror [45] and data splitting [9] counterparts, we prove with two parts. In the first part, we show that the mirror statistics can vary for different $j$ 's. Secondly, we show that the proposed test statistic (Eq. (3)) is UMP given $j$ . + +# B.5.1 Part I + +Differing from existing works like Gaussian mirror or data splitting that considers a generic distribution for all $\beta_{j}$ under $H_0$ and $H_{1}$ , we interpret the FDR under $H_{0}$ from another angle that considers the changes in distribution to $\beta_{j}$ across different $j$ 's. Namely, given an arbitrary mirror test statistic $\mathrm{sign}(\beta_j^+ \beta_j^-) f_j(\beta_j^+, \beta_j^-)$ that depends on the index $j$ , and a general index-agnostic counterpart $\mathrm{sign}(\beta_j^+ \beta_j^-) f(\beta_j^+, \beta_j^-)$ , the FDR can be defined as: + +$$ +P \left(\operatorname {s i g n} \left(\beta^ {+} \beta^ {-}\right) f \left(\beta^ {+}, \beta^ {-}\right) < t _ {q} \mid H _ {0}\right) = \sum_ {j \in [ 1, \dots , p ]} P (j) P \left(\operatorname {s i g n} \left(\beta_ {j} ^ {+} \beta_ {j} ^ {-}\right) f _ {j} \left(\beta_ {j} ^ {+}, \beta_ {j} ^ {-}\right) < t _ {q} ^ {j} \mid H _ {0}, j\right), \tag {7} +$$ + +where $P(j)$ refers to the probability of the presence of the index $j$ . In practice, we do not need to know the specification of this distribution as the proof relies only on the symmetry of the conditional distribution: $P(\mathrm{sign}(\beta_j^+ \beta_j^-) f_j(\beta_j^+, \beta_j^-)|H_0, j)$ . Apparently $\sum_{j \in [1, \dots, p]} P(j) = 1$ holds. $t_q$ is a threshold chosen such that the FDR can be controlled by the nominal level $q$ . $t_q^j$ adds the dependence on the index $j$ . In previous works, it is common to assume identical $\beta_j$ in $H_0$ and $H_1$ to have the same distribution across all $j$ 's, leading to the fact that $\mathrm{sign}(\beta_j^+ \beta_j^-) f(\beta_j^+, \beta_j^-) < t_q$ is the same as $\mathrm{sign}(\beta^+ \beta^-) f(\beta^+, \beta^-) < t_q$ . However, if one considers the selection rule for these approaches (e.g., Gaussian mirror/data splitting), it is clear that such a universal assumption need not to hold. This is because to control the FDR, the only necessary property is the symmetry of the null distribution. According to Eq. 7, we realize that even in the case where $\mathrm{sign}(\beta_j^+ \beta_j^-) f(\beta_j^+, \beta_j^-)$ varies across $j$ 's, the overall FDR can still be controlled if all of these distributions are symmetric about zero. The only difference is that for each $\mathrm{sign}(\beta_j^+ \beta_j^-) f(\beta_j^+, \beta_j^-)$ , the threshold $t_q^j$ is different given different distributions. And the proof of higher power is straightforward given this distributional flexibility. + +# B.5.2 Part II + +According to Lemma 3.5, we already showed that the proposed test statistic $\mathrm{sign}(\beta_j^+ \beta_j^-) f(\beta_j^+, \beta_j^-)$ in Eq. (3) is UMP given the setting considered in this paper. This means that given any arbitrary $t_q^j$ for some FDR level (not necessarily $q$ as $q$ refers to the general selection rule across all $\beta_j$ 's), we can achieve the highest power compared to the test statistics in Gaussian mirror and data splitting. This leads to the conclusion that the overall power is the highest, hence completing the proof. + +Remark B.2. This proof simply reveals a fact that under the controlled nominal FDR level $q$ , to find a better test statistic, we only need to maximize the power of individual test statistic for every $\beta_{j}$ (or, equivalently, $X_{j}$ ), rather than a general form of test statistic across all $j \in [1, \dots, p]$ . + +To the best of our knowledge, we are the first to provide this reasoning in the proof and we hope this brings insights into developing the UMP test statistics, which is beyond the scope of this paper. + +# C Complexity Discussion between Benchmarking Methods + +The complexity of the Gaussian Mirror method is $\mathcal{O}(np^3 + p^4)$ (for $n > p$ ). Essentially it runs $p$ ordinary least square fit (OLS), each of which has $\mathcal{O}(np^2 + p^3)$ complexity ( $n > p$ ). The computational complexity of Algorithm 2 is $\mathcal{O}(np^3 + p^4 + pki)$ (for $n > p$ ), where the $\mathcal{O}(pki)$ part is introduced by the k-means (following Lloyd's algorithm). $k$ is the number of clusters and $i$ stands for the number of iterations until convergence. In comparison, the data splitting runs two OLS, resulting in $\mathcal{O}(np^2 + p^3)$ complexity. The knockoff framework, according to Askari et al. [3], requires at least $\mathcal{O}(p^{3.5})$ to solve for the knockoff variable in a semi-definite programming setting. Later on, it requires one OLS for $2p$ dimension, resulting in $\mathcal{O}(p^{3.5} + 4np^2 + 8p^3)$ complexity. Other deep-learning-based knockoff variables require additional deep learning model fitting to obtain the knockoff statistics, making the complexity analysis hard given the choice of optimizer and the model architecture, hence ignored. According to Zhong et al. [46], the computational complexity for CRT, assuming $p = n$ is $\mathcal{O}(p^3 \log^2 p)$ . We did not find any rigorous complexity analysis about HRT and dCRT, however, based on the results reported in Table 2 of the paper [26], we believe the complexity is between model-X knockoff and CRT. + +# D Design Matrix Setup for Synthetic Experiment + +This section presents the model applied to the design matrix $X$ in the synthetic dataset setup in Section 4.1. + +Gaussian: We replicate the multivariate normal benchmark described in Romano et al. [30]. Specifically, we sample $x \sim \mathcal{N}(0, \Sigma)$ , where $\Sigma$ is a $d$ -dimensional covariance matrix with entries $\Sigma_{i,j} = \rho^{|i - j|}$ . This autoregressive Gaussian data structure captures strong correlations between adjacent features, with diminishing correlations as the distance between features increases. For our experiments, we consider $\rho \in \{0.5, 0.6, 0.7\}$ to provide additional context on how the change of $\rho$ affects the feature selection performance, in addition to the chosen 0.6 from Romano et al. [30]. Due to limited space in the main paper, we present results for $\rho = 0.5$ in Appendix E.1. + +Gaussian mixture: We utilize a Gaussian mixture model for $X$ , represented as $X \sim \sum_{k=1}^{3} \pi_k \mathcal{N}(\mu_k, \Sigma_k)$ , where the proportions of the three Gaussian components are given by $(\pi_1, \pi_2, \pi_3) = (0.4, 0.2, 0.4)$ . The mean vectors $\mu_k \in \mathbb{R}^p$ are defined as $\mu_k = \mathbf{1}_p \cdot 20 \cdot (k-1)$ , where $\mathbf{1}_p$ is a $p$ -dimensional vector of ones. The covariance matrices $\Sigma_k \in \mathbb{R}^{p \times p}$ have entries $(i,j)$ computed as $\rho_k^{|i-j|}$ , where $\rho_k = \rho_{\mathrm{base}}^{k-0.1}$ and $\rho_{\mathrm{base}} = 0.6$ . Both Gaussian and Gaussian mixture implementation can be found in [34] + +Copula Models: To evaluate data structures with more complex correlation, we incorporate copula models [32]. Specifically, we consider two copula families: Clayton and Joe, each parameterized with a consistent copula parameter of 2. Marginal distributions are selected as either a uniform distribution (via identity transformation) or an exponential distribution with a rate of 1. These copulas are implemented using the PyCop library $^{15}$ . Essentially, copulas are statistical tools designed to model and simulate complex dependencies among random variables, independently from the shapes of their marginal distributions. Unlike traditional multivariate models that assume linear or Gaussian relationships, copulas allow us to construct datasets where variables exhibit non-linear or asymmetric dependencies, better reflecting patterns seen in real-world data. In our study, we use two widely-studied copula families: the Clayton copula, which models strong lower-tail dependence (i.e., variables tend to move together when their values are low), and the Joe copula, which captures strong upper-tail dependence (i.e., variables tend to move together when their values are high). By + +specifying a copula parameter of 2, we control the strength of these dependencies in a consistent way across scenarios. For each simulated dataset, the individual variable distributions (margins) are chosen to be either uniform (via identity transformation) or exponential (rate=1), allowing us to assess the robustness of our methods under different data distributions. This setup enables a comprehensive evaluation of how our proposed methods perform under diverse and realistic correlation structures. + +# E Additional Experiment + +This section extends the experiments in the main paper to cover e.g., the high dimensional setting and various noise levels, to further demonstrate the performance of the proposed method. In particular, we consider synthetic experiments and focus on the ridge regression model for the fitting coefficients (see Sec. 4.1) for details). + +# E.1 Additional Gaussian Synthetic Dataset Result + +This section presents benchmarking results for the Gaussian setting complementing Table 1 with rho = 0.5. Results are presented in Table 5. + +Table 5: FDR and power with Gaussian design matrix $X$ for $\rho = {0.5}$ . Bold entries indicate the case with the highest power given controlled FDR level: 0.1. Blue entries indicate the second best. Red entries indicate $\mathrm{{FDR}} > {0.1}$ . + +
MethodOLSRidgeLASSO
FDRPowerFDRPowerFDRPower
CRT [8]0.380.930.270.950.170.99
Distilled-CRT [19]0.540.990.340.990.150.98
Data Splitting [9]0.000.000.130.820.080.70
Data Splitting† [9]0.060.750.130.820.130.76
Gaussian Mirror [45]0.100.790.050.770.070.84
Gaussian Mirror† [45]0.100.820.050.800.070.87
HRT [40]0.000.340.000.360.010.36
Powerful Knockoff [37]0.000.000.070.490.050.77
Powerful Knockoff† [37]0.060.800.090.600.050.80
G2M (ours)0.100.880.070.900.070.94
G2M† (ours)0.100.880.070.900.070.94
+ +# E.2 High Dimensional Setting + +We extend Table 1 and Table 2 with two additional low sample-size settings (i.e., $n / p = 30 / 100$ and $90 / 100$ ), considering ridge regression for its better performance compared to LASSO. + +Notably, given that we already consider a low signal strength setting in both synthetic and semi-synthetic experiments, following the experimental setup in Shen et al. [34] (see Sec. 4 in the main paper), to balance the reduced sample size with signal strength, we apply a multiplier to boost the signal strength. This is a common practice in statistical analysis. Specifically, we use the multiplier $\sqrt{200} / \sqrt{p}$ , where $p = 30$ or 100, to ensure a fair comparison between the new cases and the $n/p = 200/100$ case in Sec. 4. + +Additionally, for Data Splitting, Gaussian Mirror, and the proposed $G^2 M$ , we consider the non-dagger versions, as we observed that the dagger versions cannot properly control FDR when $p > n$ . The results for the Gaussian setting and the non-Gaussian setting are included in Table 6 and 7, respectively. + +From Table 6, we observe that the proposed $\mathrm{G}^2\mathrm{M}$ outperforms all other methods by achieving the highest power with controlled FDR. In addition, Table 7 suggests that $\mathrm{G}^2\mathrm{M}$ is generally a good feature selection algorithm for having the best or the second best power with controlled FDR. More importantly, in this case, $\mathrm{G}^2\mathrm{M}$ never exceeds the FDR nominal level of 0.1, suggesting its soundness in the proposed theoretical framework. + +Table 6: Extension of Table 1 with high-dimensional settings: FDR / power under two different $n/p$ ratios and two correlation strengths. Red entries indicate FDR failures ( $>0.1$ ), Bold entries indicate the highest power among methods controlling FDR $\leq 0.1$ , and Blue entries indicate the second highest. + +
Methodn/p = 90/100n/p = 30/100
ρ = 0.6ρ = 0.7ρ = 0.6ρ = 0.7
CRT [8]0.00 / 0.150.00 / 0.090.00 / 0.000.00 / 0.01
Distilled-CRT [19]0.59 / 0.950.59 / 0.920.04 / 0.060.09 / 0.09
Gaussian Mirror [45]0.11 / 0.550.08 / 0.540.01 / 0.030.00 / 0.00
Data Splitting [9]0.08 / 0.530.07 / 0.520.02 / 0.050.06 / 0.10
HRT [40]0.00 / 0.080.00 / 0.030.00 / 0.000.00 / 0.00
Powerful Knockoff [37]0.02 / 0.110.03 / 0.130.00 / 0.000.00 / 0.00
G2M (ours)0.06 / 0.580.08 / 0.550.09 / 0.130.05 / 0.11
+ +# E.3 Noise Level Variation + +To study robustness to noise, we follow Shen et al. [34] and vary the signal strength by scaling $\beta$ as $\frac{p}{10\sqrt{n}}$ , $\frac{p}{15\sqrt{n}}$ (already in the paper), and $\frac{p}{20\sqrt{n}}$ . Below, we report results for the additional $\frac{p}{10\sqrt{n}}$ and $\frac{p}{20\sqrt{n}}$ settings. All other experimental parameters remain the same. Similarly, ridge regression is considered for all methods. And results for the Gaussian setting and the non-Gaussian setting are included in Table 8 and 9, respectively. + +From Table 8, we observe that $\mathrm{G}^2\mathrm{M}$ beats other methods by having the highest power with controlled FDR—a case that is similarly revealed in Table 6. Besides, Table 9 also suggests a good performance of the proposed $\mathrm{G}^2\mathrm{M}$ for having 8 out of 10 best performing results (i.e., consider the highest power with controlled FDR). This indicate the outperformance of the proposed $\mathrm{G}^2\mathrm{M}$ in both Gaussian and non-Gaussian settings. + +# E.4 Another Real Case Study—Breast Cancer Dataset + +We conduct an additional real-world case study using the Breast Cancer Wisconsin (Diagnostic) dataset [43], which consists of 569 patient samples. Each sample is labeled with a binary diagnostic outcome (malignant vs. benign), accompanied by 30 quantitative features extracted from digitized images of fine-needle aspirate (FNA) of breast mass tissue. Following the same evaluation protocol as our previous case study, we reviewed the biomedical literature to identify features that have been consistently reported as highly indicative of malignant tissue. These literature-supported features serve as the ground truth for our analysis. Based on this reference set, we report in Table X the number of "referenced vs. identified" features selected by each competing method. In total, 22 features are recognized as clinically relevant. Results are in Table 10. + +# F Preparation of the RNA Data + +We first normalize the raw data $X$ to value range [0, 1] and then standardize it to have zero mean and unit variance. $Y$ is synthesized according to $X$ . We consider the response $Y$ is generated following the expression: + +$$ +k \in [ m / 4 ] +$$ + +$$ +\varphi_ {k} ^ {(1)}, \varphi_ {k} ^ {(2)} \sim \mathcal {N} (1, 1) +$$ + +$$ +\varphi_ {k} ^ {(3)}, \varphi_ {k} ^ {(4)}, \varphi_ {k} ^ {(5)} \sim \mathcal {N} (2, 1) \quad m / 4 \tag {8} +$$ + +$$ +\begin{array}{l} Y \mid X = \epsilon + \sum_ {k = 1} ^ {m / 4} \varphi_ {k} ^ {(1)} X _ {4 k - 3} + \varphi_ {k} ^ {(3)} X _ {4 k - 2} \\ + \varphi_ {k} ^ {(4)} \tanh \left(\varphi_ {k} ^ {(2)} X _ {4 k - 1} + \varphi_ {k} ^ {(5)} X _ {4 k}\right), \\ \end{array} +$$ + +Table 7: Extension of Table 2 with high-dimensional settings: FDR and power for each dataset under two sample-size regimes. Red entries indicate $\mathrm{FDR} > 0.1$ , Bold for the highest power among methods controlling $\mathrm{FDR} \leq 0.1$ , and Blue for the second highest. + +
DatasetMethodFDR / Power
n/p = 90/100n/p = 30/100
GaussianMixtureAR1CRT [8]0.00 / 0.000.00 / 0.00
Distilled-CRT [19]0.10 / 0.130.00 / 0.00
Gaussian Mirror [45]0.06 / 0.520.01 / 0.01
Data Splitting [9]0.05 / 0.320.06 / 0.07
HRT [40]0.00 / 0.010.00 / 0.01
Deep Knockoff [30]0.77 / 1.000.79 / 0.97
DDLK [38]0.25 / 0.210.00 / 0.00
KnockoffGAN [14]0.02 / 0.100.00 / 0.01
sRMMD [22]0.75 / 1.000.79 / 0.97
DeepDRK [34]0.01 / 0.200.00 / 0.00
G2M (ours)0.04 / 0.390.04 / 0.04
Copula: Clayton & Exp.CRT [8]0.00 / 0.000.00 / 0.00
Distilled-CRT [19]0.12 / 0.210.08 / 0.04
Gaussian Mirror [45]0.03 / 0.100.11 / 0.14
Data Splitting [9]0.07 / 0.200.05 / 0.02
HRT [40]0.00 / 0.000.00 / 0.00
Deep Knockoff [30]0.47 / 0.880.77 / 0.84
DDLK [38]0.00 / 0.000.00 / 0.00
KnockoffGAN [14]0.00 / 0.000.00 / 0.00
sRMMD [22]0.42 / 0.810.75 / 0.81
DeepDRK [34]0.08 / 0.150.17 / 0.35
G2M (ours)0.05 / 0.200.02 / 0.03
Copula: Clayton & GammaCRT [8]0.00 / 0.000.00 / 0.00
Distilled-CRT [19]0.13 / 0.190.07 / 0.07
Gaussian Mirror [45]0.02 / 0.180.12 / 0.18
Data Splitting [9]0.07 / 0.280.04 / 0.02
HRT [40]0.00 / 0.000.00 / 0.00
Deep Knockoff [30]0.50 / 0.940.78 / 0.89
DDLK [38]0.00 / 0.000.00 / 0.00
KnockoffGAN [14]0.00 / 0.030.00 / 0.00
sRMMD [22]0.52 / 0.950.78 / 0.92
DeepDRK [34]0.01 / 0.080.12 / 0.13
G2M (ours)0.03 / 0.180.10 / 0.10
Copula: Joe & Exp.CRT [8]0.00 / 0.000.00 / 0.00
Distilled-CRT [19]0.03 / 0.090.00 / 0.00
Gaussian Mirror [45]0.05 / 0.280.01 / 0.02
Data Splitting [9]0.06 / 0.190.02 / 0.01
HRT [40]0.00 / 0.000.00 / 0.00
Deep Knockoff [30]0.53 / 0.900.77 / 0.88
DDLK [38]0.00 / 0.000.00 / 0.00
KnockoffGAN [14]0.00 / 0.000.00 / 0.00
sRMMD [22]0.42 / 0.790.75 / 0.83
DeepDRK [34]0.09 / 0.210.08 / 0.00
G2M (ours)0.06 / 0.270.02 / 0.01
Copula: Joe & GammaCRT [8]0.00 / 0.000.00 / 0.00
Distilled-CRT [19]0.04 / 0.070.00 / 0.00
Gaussian Mirror [45]0.09 / 0.520.01 / 0.02
Data Splitting [9]0.07 / 0.280.02 / 0.01
HRT [40]0.00 / 0.000.00 / 0.00
Deep Knockoff [30]0.50 / 0.970.00 / 0.00
DDLK [38]0.01 / 0.000.00 / 0.00
KnockoffGAN [14]0.06 / 0.270.03 / 0.04
sRMMD [22]0.48 / 0.960.79 / 0.95
DeepDRK [34]0.05 / 0.300.03 / 0.05
G2M (ours)0.09 / 0.620.05 / 0.05
+ +Table 8: Extension of Table 1 with different noise levels: FDR / power under two additional different signal strength parameters $\frac{p}{10\sqrt{n}}$ and $\frac{p}{20\sqrt{n}}$ . Red entries indicate FDR>0.1; among those with FDR≤0.1, Bold entries indicate the highest power and Blue entries indicate the second highest. + +
Methodp10√np20√n
ρ = 0.6ρ = 0.7ρ = 0.6ρ = 0.7
CRT [8]0.04 / 0.960.05 / 0.880.07 / 0.420.09 / 0.25
Distilled-CRT [19]0.40 / 1.000.40 / 1.000.26 / 0.890.21 / 0.65
Gaussian Mirror [45]0.09 / 1.000.06 / 0.880.07 / 0.520.06 / 0.40
Data Splitting [9]0.11 / 0.810.07 / 0.770.03 / 0.560.08 / 0.45
HRT [40]0.01 / 0.710.02 / 0.460.02 / 0.130.00 / 0.05
Powerful Knockoff [37]0.08 / 0.780.02 / 0.530.05 / 0.200.02 / 0.06
G2M (ours)0.07 / 1.000.05 / 0.930.02 / 0.580.02 / 0.45
+ +where $\epsilon$ follows the standard normal distribution and the $m = 20$ for 20 covariates that are sampled uniformly. + +# G Supplementary Material for the Case Study + +# G.1 Nominal Metabolite List + +In Table 11, we include the list of the nominal metabolites curated by [34] for the IBD dataset. + +# G.2 Additional Results for the IBD Study + +Here we provide the supplementary information for the experimental results described in Section 4.3. In Table 12 and 13, we provide the list of identified metabolites by each of the considered models. This table provides additional information for Table 4 in the main paper which only includes metabolite counts due to limited space. + +Table 9: Extension of Table 2 with different noise levels: FDR and power under two signal strength regimes $\frac{p}{10\sqrt{n}}$ and $\frac{p}{20\sqrt{n}}$ . Red entries indicate FDR>0.1; Bold for the highest power among methods controlling FDR ≤ 0.1; and Blue for the second highest. + +
DatasetMethodp10√n(FDR / Power)p20√n(FDR / Power)
Gaussian MixtureCRT [8]0.31 / 0.400.30 / 0.40
Gaussian Mirror [45]0.06 / 1.000.07 / 0.62
Data Splitting [9]0.09 / 0.940.06 / 0.64
HRT [40]0.01 / 0.390.02 / 0.09
Deep Knockoff [30]0.56 / 0.990.76 / 1.00
DDLK [38]0.79 / 1.000.72 / 0.91
KnockoffGAN [14]0.21 / 1.000.52 / 0.98
sRMMD [22]0.60 / 1.000.77 / 1.00
DeepDRK [34]0.06 / 0.940.15 / 0.70
G2M (ours)0.06 / 1.000.05 / 0.78
Copula: Clayton & Exp.CRT [8]0.01 / 0.260.02 / 0.12
Gaussian Mirror [45]0.05 / 0.950.08 / 0.46
Data Splitting [9]0.08 / 0.860.13 / 0.46
HRT [40]0.02 / 0.580.00 / 0.06
Deep Knockoff [30]0.09 / 0.920.38 / 0.82
DDLK [38]0.49 / 0.880.02 / 0.05
KnockoffGAN [14]0.06 / 0.560.05 / 0.12
sRMMD [22]0.10 / 0.920.37 / 0.81
DeepDRK [34]0.07 / 0.850.17 / 0.51
G2M (ours)0.04 / 0.970.12 / 0.59
Copula: Clayton & GammaCRT [8]0.00 / 0.190.00 / 0.04
Gaussian Mirror [45]0.09 / 1.000.05 / 0.69
Data Splitting [9]0.10 / 0.960.09 / 0.72
HRT [40]0.02 / 0.920.02 / 0.29
Deep Knockoff [30]0.07 / 0.970.37 / 0.94
DDLK [38]0.47 / 0.950.14 / 0.30
KnockoffGAN [14]0.08 / 0.910.09 / 0.43
sRMMD [22]0.06 / 0.970.35 / 0.94
DeepDRK [34]0.06 / 0.920.12 / 0.67
G2M (ours)0.07 / 1.000.10 / 0.86
Copula: Joe & ExponentialCRT [8]0.11 / 0.570.08 / 0.23
Gaussian Mirror [45]0.05 / 0.760.15 / 0.35
Data Splitting [9]0.09 / 0.780.20 / 0.36
HRT [40]0.01 / 0.390.00 / 0.02
Deep Knockoff [30]0.18 / 0.910.48 / 0.79
DDLK [38]0.25 / 0.570.02 / 0.02
KnockoffGAN [14]0.05 / 0.370.03 / 0.06
sRMMD [22]0.16 / 0.880.40 / 0.68
DeepDRK [34]0.08 / 0.740.20 / 0.39
G2M (ours)0.08 / 0.910.19 / 0.43
Copula: Joe & GammaCRT [8]0.11 / 0.560.10 / 0.33
Gaussian Mirror [45]0.05 / 0.980.08 / 0.56
Data Splitting [9]0.07 / 0.920.10 / 0.58
HRT [40]0.01 / 0.910.01 / 0.16
Deep Knockoff [30]0.06 / 0.960.35 / 0.91
DDLK [38]0.48 / 0.950.09 / 0.17
KnockoffGAN [14]0.05 / 0.770.07 / 0.30
sRMMD [22]0.06 / 0.950.38 / 0.92
DeepDRK [34]0.05 / 0.860.12 / 0.61
G2M (ours)0.09 / 1.000.08 / 0.68
+ +Table 10: Feature selection performance on the Breast Cancer Wisconsin (Diagnostic) dataset [43]. Since the true support is unknown, we report "number of literature-referenced features / number of identified features" instead of FDR or power. + +
ModelG2M (ours)†CRTDistilled-CRTGaussian Mirror†Data Splitting†HRTPowerful Knockoff
Referenced / Identified18/215/612/1518/2611/152/421/26
ModelDeepDRKDeep KnockoffsRMMDKnockoffGANDDLK
Referenced / Identified15/2011/1513/1817/228/13
+ +Table 11: IBD-associated metabolites that are supported by the literature. This table includes all 47 referenced metabolites for the IBD case study. Each metabolite is supported by one of the three sources: PubChem, peer-reviewed publications, or preprints. For PubChem case, we report the PubChem reference ID (CID), and for the other two cases, we report the publication references. + +
Reference TypeMetaboliteSourceMetaboliteSource
PubChempalmitateCID: 985taurocholateCID: 6675
cholateCID: 221493p-hydroxyphenylacetateCID: 127
linoleateCID: 5280450deoxycholateCID: 222528
taurochenodeoxycholateCID: 387316
Publications12.13-diHOME[6]dodecanedioiate[6]
arachidonate[6]eicosatrienoate[6, 5]
eicosadienoate[6]docosapentaenoate[6, 5]
tauolithocholate[6]salicylate[6]
saccharin[6]1.2.3.4-tetrahydro-beta-caroline-1.3-dicarboxylate[6]
oleate[5]arachidate[5]
glycocholate[5]chenodeoxycholate[5]
phenyllactate[22, 16]glycolithocholate[5]
urobilin[22, 28]caproate[22, 17]
hydrocinnamate[22, 15]myristate[22, 11]
adrenate[22, 20]olmesartan[22, 31]
tetradecanedioate[39, 23]hexadecanedioate[39, 23]
oxypurinol[7]porphobilinogen[24]
caprate[35, 36]undecanedionate[17, 42]
stearate[2, 5]oleanate[27]
glycochenodeoxycholate[33]sebacate[17]
nervonic acid[41]lithocholate[5]
Preprintsalpha-muricholate[25]tauro-alpha-muricholate/tauro-beta-muricholate[25]
17-methylstearate[25]myristoleate[25]
taurodeoxycholate[25]ketodeoxycholate[25]
+ +Table 12: A list of literature-supported metabolites out of a total of 80 candidates. “*” indicates the important metabolites marked by the corresponding algorithms. + +
MetaboliteG2M†CRTDistilled-CRTGaussian Mirror†Data Splitting†HRTPowerful Knockoff
12.13-diHOME**
9.10-diHOME
caproate****
hydrocinnamate**
mandelate*
3-hydroxyoctanoate*
caprate*
indoleacetate
3-hydroxydecanoate
dodecanoate*
undecanedionate*
myristoleate*
myristate
dodecanedioate
pentadecanoate**
hydroxymyristate
palmitoleate*
palmitate*
tetradecanedioate****
10-heptadecenoate
2-hydroxyhexadecanoate*
alpha-linolenate*
linoleate***
oleate**
stearate***
hexadecanedioate**
10-nonadecenoate
nonadecanoate****
17-methylstearate**
eicosapentaenoate
arachidonate*****
eicosatrienoate*
eicosadienoate***
eicosanoate**
arachidate*
phytanate
docosahexaenoate**
docosapentaenoate***
adrenate****
13-docosenoate*
eicosanedioate****
oleanate*
masilinate*
lithocholate*
chenodeoxycholate**
deoxycholate
hyodeoxycholate/ursodeoxycholate*
ketodeoxycholate*
alpha-muricholate
cholate**
glycolithocholate
glycochenodeoxycholate**
glycodeoxycholate*
glycoursodeoxycholate****
glycocholate**
taurolithocholate*
taurochenodeoxycholate****
taurodeoxycholate**
tauro-alpha-muricholate/tauro-beta-muricholate*
taurocholate***
salicylate*****
saccharin*
azelate
sebacate
carboxybuprofen
olmesartan
1.2.3.4-tetrahydro-beta-carboline-1.3-dicarboxylate*
4-hydroxystyrene*****
acetytyrosine
alpha-CEHC*
carnosol**
oxypurinol
palmitoylethanolamide**
phenyllactate*****
p-hydroxyphenylacetate*
porphobilinogen**
urobilin****
nervonic acid***
oxymetazoline**
+ +Table 13: A continued list of literature-supported metabolites out of a total of 80 candidates to Table 12. \* indicates the important metabolites marked by the corresponding algorithms. + +
MetaboliteDeepDRKDeep KnockoffsRMMDKnockoffGANDDLK
12.13-diHOME*
9.10-diHOME
caproate****
hydrocinnamate
mandelate
3-hydroxyoctanoate
caprate
indoleacetate*
3-hydroxydecanoate
dodecanoate*
undecanedionate**
myristoleate
myristate
dodecanedioate*
pentadecanoate
hydroxymyristate
palmitoleate
palmitate*
tetradecanedioate*
10-heptadecenoate
2-hydroxyhexadecanoate
alpha-linolenate*
linoleate
oleate
stearate*
hexadecanedioate***
10-nonadecenoate
nonadecanoate
17-methylstearate***
eicosapentaenoate***
arachidonate****
eicosatrienoate***
eicosadienoate****
eicosenoate
arachidate*
phytanate
docosahexaenoate***
docosapentaenoate****
adrenate*****
13-docosenoate
eicosanedioate**
oleanate
masilinate
lithocholate*
chenodeoxycholate
deoxycholate***
hyodeoxycholate/ursodeoxycholate
ketodeoxycholate*
alpha-muricholate*
cholate*
glycolithocholate*
glycochenodeoxycholate
glycodeoxycholate
glycoursodeoxycholate
glycocholate
taurolithocholate*
taurochenodeoxycholate
taurodeoxycholate
taurohyodeoxycholate/tauroursodeoxycholate
tauro-alpha-muricholate/tauro-beta-muricholate**
taurocholate
salicylate****
saccharin*
azelate*
sebacate**
carboxyibuprofen
olmesartan
1.2.3.4-tetrahydro-beta-caroline-1.3-dicarboxylate**
4-hydroxystyrene***
acetytyrosine
alpha-CEHC
carnosol*
oxypurinol
palmitoylenolamide
phenyllactate****
p-hydroxyphenylacetate***
porphobilinogen*
urobilin****
nervonic acid
oxymetazoline***
+ +# NeurIPS Paper Checklist + +# 1. Claims + +Question: Do the main claims made in the abstract and introduction accurately reflect the paper's contributions and scope? + +Answer: [Yes] + +Justification: The abstract and introduction clearly stated the objective on proposing a new method in FDR-controlled feature selection regime. + +Guidelines: + +- The answer NA means that the abstract and introduction do not include the claims made in the paper. +- The abstract and/or introduction should clearly state the claims made, including the contributions made in the paper and important assumptions and limitations. A No or NA answer to this question will not be perceived well by the reviewers. +- The claims made should match theoretical and experimental results, and reflect how much the results can be expected to generalize to other settings. +- It is fine to include aspirational goals as motivation as long as it is clear that these goals are not attained by the paper. + +# 2. Limitations + +Question: Does the paper discuss the limitations of the work performed by the authors? + +Answer: [Yes] + +Justification: The limitation is essentially described in the form of assumptions in the theoretical parts in the method section. We have also discussed the limitation in the conclusion section. + +Guidelines: + +- The answer NA means that the paper has no limitation while the answer No means that the paper has limitations, but those are not discussed in the paper. +- The authors are encouraged to create a separate "Limitations" section in their paper. +- The paper should point out any strong assumptions and how robust the results are to violations of these assumptions (e.g., independence assumptions, noiseless settings, model well-specification, asymptotic approximations only holding locally). The authors should reflect on how these assumptions might be violated in practice and what the implications would be. +- The authors should reflect on the scope of the claims made, e.g., if the approach was only tested on a few datasets or with a few runs. In general, empirical results often depend on implicit assumptions, which should be articulated. +- The authors should reflect on the factors that influence the performance of the approach. For example, a facial recognition algorithm may perform poorly when image resolution is low or images are taken in low lighting. Or a speech-to-text system might not be used reliably to provide closed captions for online lectures because it fails to handle technical jargon. +- The authors should discuss the computational efficiency of the proposed algorithms and how they scale with dataset size. +- If applicable, the authors should discuss possible limitations of their approach to address problems of privacy and fairness. +- While the authors might fear that complete honesty about limitations might be used by reviewers as grounds for rejection, a worse outcome might be that reviewers discover limitations that aren't acknowledged in the paper. The authors should use their best judgment and recognize that individual actions in favor of transparency play an important role in developing norms that preserve the integrity of the community. Reviewers will be specifically instructed to not penalize honesty concerning limitations. + +# 3. Theory assumptions and proofs + +Question: For each theoretical result, does the paper provide the full set of assumptions and a complete (and correct) proof? + +# Answer: [Yes] + +Justification: Please refer to the method section for the theoretical results and the appendix for the proofs. + +# Guidelines: + +- The answer NA means that the paper does not include theoretical results. +- All the theorems, formulas, and proofs in the paper should be numbered and cross-referenced. +- All assumptions should be clearly stated or referenced in the statement of any theorems. +- The proofs can either appear in the main paper or the supplemental material, but if they appear in the supplemental material, the authors are encouraged to provide a short proof sketch to provide intuition. +- Inversely, any informal proof provided in the core of the paper should be complemented by formal proofs provided in appendix or supplemental material. +- Theorems and Lemmas that the proof relies upon should be properly referenced. + +# 4. Experimental result reproducibility + +Question: Does the paper fully disclose all the information needed to reproduce the main experimental results of the paper to the extent that it affects the main claims and/or conclusions of the paper (regardless of whether the code and data are provided or not)? + +# Answer: [Yes] + +Justification: The paper presents two algorithms that should provide full details of implementation. + +# Guidelines: + +- The answer NA means that the paper does not include experiments. +- If the paper includes experiments, a No answer to this question will not be perceived well by the reviewers: Making the paper reproducible is important, regardless of whether the code and data are provided or not. +- If the contribution is a dataset and/or model, the authors should describe the steps taken to make their results reproducible or verifiable. +- Depending on the contribution, reproducibility can be accomplished in various ways. For example, if the contribution is a novel architecture, describing the architecture fully might suffice, or if the contribution is a specific model and empirical evaluation, it may be necessary to either make it possible for others to replicate the model with the same dataset, or provide access to the model. In general, releasing code and data is often one good way to accomplish this, but reproducibility can also be provided via detailed instructions for how to replicate the results, access to a hosted model (e.g., in the case of a large language model), releasing of a model checkpoint, or other means that are appropriate to the research performed. +- While NeurIPS does not require releasing code, the conference does require all submissions to provide some reasonable avenue for reproducibility, which may depend on the nature of the contribution. For example + +(a) If the contribution is primarily a new algorithm, the paper should make it clear how to reproduce that algorithm. +(b) If the contribution is primarily a new model architecture, the paper should describe the architecture clearly and fully. +(c) If the contribution is a new model (e.g., a large language model), then there should either be a way to access this model for reproducing the results or a way to reproduce the model (e.g., with an open-source dataset or instructions for how to construct the dataset). +(d) We recognize that reproducibility may be tricky in some cases, in which case authors are welcome to describe the particular way they provide for reproducibility. In the case of closed-source models, it may be that access to the model is limited in some way (e.g., to registered users), but it should be possible for other researchers to have some path to reproducing or verifying the results. + +# 5. Open access to data and code + +Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material? + +Answer: [Yes] + +Justification: https://github.com/skyve2012/G2M. + +Guidelines: + +- The answer NA means that paper does not include experiments requiring code. +- Please see the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details. +- While we encourage the release of code and data, we understand that this might not be possible, so "No" is an acceptable answer. Papers cannot be rejected simply for not including code, unless this is central to the contribution (e.g., for a new open-source benchmark). +- The instructions should contain the exact command and environment needed to run to reproduce the results. See the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details. +- The authors should provide instructions on data access and preparation, including how to access the raw data, preprocessed data, intermediate data, and generated data, etc. +- The authors should provide scripts to reproduce all experimental results for the new proposed method and baselines. If only a subset of experiments are reproducible, they should state which ones are omitted from the script and why. +- At submission time, to preserve anonymity, the authors should release anonymized versions (if applicable). +- Providing as much information as possible in supplemental material (appended to the paper) is recommended, but including URLs to data and code is permitted. + +# 6. Experimental setting/details + +Question: Does the paper specify all the training and test details (e.g., data splits, hyperparameters, how they were chosen, type of optimizer, etc.) necessary to understand the results? + +Answer: [Yes] + +Justification: Such details are introduced and discussed in the experiment section. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The experimental setting should be presented in the core of the paper to a level of detail that is necessary to appreciate the results and make sense of them. +- The full details can be provided either with the code, in appendix, or as supplemental material. + +# 7. Experiment statistical significance + +Question: Does the paper report error bars suitably and correctly defined or other appropriate information about the statistical significance of the experiments? + +Answer: [Yes] + +Justification: As per feature selection, we follow the convention and report FDR and power (as empirical estimation of the means), with independent and repeated experiments. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The authors should answer "Yes" if the results are accompanied by error bars, confidence intervals, or statistical significance tests, at least for the experiments that support the main claims of the paper. +- The factors of variability that the error bars are capturing should be clearly stated (for example, train/test split, initialization, random drawing of some parameter, or overall run with given experimental conditions). + +- The method for calculating the error bars should be explained (closed form formula, call to a library function, bootstrap, etc.) +- The assumptions made should be given (e.g., Normally distributed errors). +- It should be clear whether the error bar is the standard deviation or the standard error of the mean. +- It is OK to report 1-sigma error bars, but one should state it. The authors should preferably report a 2-sigma error bar than state that they have a $96\%$ CI, if the hypothesis of Normality of errors is not verified. +- For asymmetric distributions, the authors should be careful not to show in tables or figures symmetric error bars that would yield results that are out of range (e.g. negative error rates). +- If error bars are reported in tables or plots, The authors should explain in the text how they were calculated and reference the corresponding figures or tables in the text. + +# 8. Experiments compute resources + +Question: For each experiment, does the paper provide sufficient information on the computer resources (type of compute workers, memory, time of execution) needed to reproduce the experiments? + +Answer: [Yes] + +Justification: The configuration is specified in Section 4. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The paper should indicate the type of compute workers CPU or GPU, internal cluster, or cloud provider, including relevant memory and storage. +- The paper should provide the amount of compute required for each of the individual experimental runs as well as estimate the total compute. +- The paper should disclose whether the full research project required more compute than the experiments reported in the paper (e.g., preliminary or failed experiments that didn't make it into the paper). + +# 9. Code of ethics + +Question: Does the research conducted in the paper conform, in every respect, with the NeurIPS Code of Ethics https://neurips.cc/public/EthicsGuidelines? + +Answer: [Yes] + +Justification: The paper conforms with the NeurIPS code of ethics. + +Guidelines: + +- The answer NA means that the authors have not reviewed the NeurIPS Code of Ethics. +- If the authors answer No, they should explain the special circumstances that require a deviation from the Code of Ethics. +- The authors should make sure to preserve anonymity (e.g., if there is a special consideration due to laws or regulations in their jurisdiction). + +# 10. Broader impacts + +Question: Does the paper discuss both potential positive societal impacts and negative societal impacts of the work performed? + +Answer: [Yes] + +Justification: The social impact and the importance of the work is discussed in the introduction section. + +Guidelines: + +- The answer NA means that there is no societal impact of the work performed. +- If the authors answer NA or No, they should explain why their work has no societal impact or why the paper does not address societal impact. + +- Examples of negative societal impacts include potential malicious or unintended uses (e.g., disinformation, generating fake profiles, surveillance), fairness considerations (e.g., deployment of technologies that could make decisions that unfairly impact specific groups), privacy considerations, and security considerations. +- The conference expects that many papers will be foundational research and not tied to particular applications, let alone deployments. However, if there is a direct path to any negative applications, the authors should point it out. For example, it is legitimate to point out that an improvement in the quality of generative models could be used to generate deepfakes for disinformation. On the other hand, it is not needed to point out that a generic algorithm for optimizing neural networks could enable people to train models that generate Deepfakes faster. +- The authors should consider possible harms that could arise when the technology is being used as intended and functioning correctly, harms that could arise when the technology is being used as intended but gives incorrect results, and harms following from (intentional or unintentional) misuse of the technology. +- If there are negative societal impacts, the authors could also discuss possible mitigation strategies (e.g., gated release of models, providing defenses in addition to attacks, mechanisms for monitoring misuse, mechanisms to monitor how a system learns from feedback over time, improving the efficiency and accessibility of ML). + +# 11. Safeguards + +Question: Does the paper describe safeguards that have been put in place for responsible release of data or models that have a high risk for misuse (e.g., pretrained language models, image generators, or scraped datasets)? + +Answer: [NA] + +Justification: N/A. + +Guidelines: + +- The answer NA means that the paper poses no such risks. +- Released models that have a high risk for misuse or dual-use should be released with necessary safeguards to allow for controlled use of the model, for example by requiring that users adhere to usage guidelines or restrictions to access the model or implementing safety filters. +- Datasets that have been scraped from the Internet could pose safety risks. The authors should describe how they avoided releasing unsafe images. +- We recognize that providing effective safeguards is challenging, and many papers do not require this, but we encourage authors to take this into account and make a best faith effort. + +# 12. Licenses for existing assets + +Question: Are the creators or original owners of assets (e.g., code, data, models), used in the paper, properly credited and are the license and terms of use explicitly mentioned and properly respected? + +Answer: [Yes] + +Justification: Existing works/codes are all properly cited. + +Guidelines: + +- The answer NA means that the paper does not use existing assets. +- The authors should cite the original paper that produced the code package or dataset. +- The authors should state which version of the asset is used and, if possible, include a URI. +- The name of the license (e.g., CC-BY 4.0) should be included for each asset. +- For scraped data from a particular source (e.g., website), the copyright and terms of service of that source should be provided. +- If assets are released, the license, copyright information, and terms of use in the package should be provided. For popular datasets, paperswithcode.com/datasets has curated licenses for some datasets. Their licensing guide can help determine the license of a dataset. + +- For existing datasets that are re-packaged, both the original license and the license of the derived asset (if it has changed) should be provided. +- If this information is not available online, the authors are encouraged to reach out to the asset's creators. + +# 13. New assets + +Question: Are new assets introduced in the paper well documented and is the documentation provided alongside the assets? + +Answer: [Yes] + +Justification: The assets are theoretical results and codes. The latter will be released upon acceptance. + +Guidelines: + +- The answer NA means that the paper does not release new assets. +- Researchers should communicate the details of the dataset/code/model as part of their submissions via structured templates. This includes details about training, license, limitations, etc. +- The paper should discuss whether and how consent was obtained from people whose asset is used. +- At submission time, remember to anonymize your assets (if applicable). You can either create an anonymized URL or include an anonymized zip file. + +# 14. Crowdsourcing and research with human subjects + +Question: For crowdsourcing experiments and research with human subjects, does the paper include the full text of instructions given to participants and screenshots, if applicable, as well as details about compensation (if any)? + +Answer: [NA] + +Justification: No such assets involved. + +Guidelines: + +- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. +- Including this information in the supplemental material is fine, but if the main contribution of the paper involves human subjects, then as much detail as possible should be included in the main paper. +- According to the NeurIPS Code of Ethics, workers involved in data collection, curation, or other labor should be paid at least the minimum wage in the country of the data collector. + +# 15. Institutional review board (IRB) approvals or equivalent for research with human subjects + +Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals (or an equivalent approval/review based on the requirements of your country or institution) were obtained? + +Answer: [NA] + +Justification: No such assets involved. + +Guidelines: + +- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. +- Depending on the country in which research is conducted, IRB approval (or equivalent) may be required for any human subjects research. If you obtained IRB approval, you should clearly state this in the paper. +- We recognize that the procedures for this may vary significantly between institutions and locations, and we expect authors to adhere to the NeurIPS Code of Ethics and the guidelines for their institution. + +- For initial submissions, do not include any information that would break anonymity (if applicable), such as the institution conducting the review. + +# 16. Declaration of LLM usage + +Question: Does the paper describe the usage of LLMs if it is an important, original, or non-standard component of the core methods in this research? Note that if the LLM is used only for writing, editing, or formatting purposes and does not impact the core methodology, scientific rigorousness, or originality of the research, declaration is not required. + +Answer: [NA] + +Justification: LLM is only used for grammatical correction and text polishing levels. + +Guidelines: + +- The answer NA means that the core method development in this research does not involve LLMs as any important, original, or non-standard components. +- Please refer to our LLM policy (https://neurips.cc/Conferences/2025/LLM) for what should or should not be described. \ No newline at end of file diff --git a/NeurIPS/2025/$_text{G}^2_text{M}$_ A Generalized Gaussian Mirror Method to Boost Feature Selection Power/images.zip b/NeurIPS/2025/$_text{G}^2_text{M}$_ A Generalized Gaussian Mirror Method to Boost Feature Selection Power/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..10f834d663e40de2589bfcce86adc510ce0f50c4 --- /dev/null +++ b/NeurIPS/2025/$_text{G}^2_text{M}$_ A Generalized Gaussian Mirror Method to Boost Feature Selection Power/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ae0497e5c7f22d4f81e22bece1955ac160158bd6522983bb92379753d6dde215 +size 1904823 diff --git a/NeurIPS/2025/$_text{G}^2_text{M}$_ A Generalized Gaussian Mirror Method to Boost Feature Selection Power/layout.json b/NeurIPS/2025/$_text{G}^2_text{M}$_ A Generalized Gaussian Mirror Method to Boost Feature Selection Power/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..5373a93223d0d369d9e74f059015d9a0db6da9b2 --- /dev/null +++ b/NeurIPS/2025/$_text{G}^2_text{M}$_ A Generalized Gaussian Mirror Method to Boost Feature Selection Power/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e3690deb6392b2bbd616a0121e03a4b17662914ad60f8c86cfa0d67ee3cee6b5 +size 1434387 diff --git a/NeurIPS/2025/$_text{S}^2$Q-VDiT_ Accurate Quantized Video Diffusion Transformer with Salient Data and Sparse Token Distillation/591034a3-721c-487d-aba5-5e442543ae0b_content_list.json b/NeurIPS/2025/$_text{S}^2$Q-VDiT_ Accurate Quantized Video Diffusion Transformer with Salient Data and Sparse Token Distillation/591034a3-721c-487d-aba5-5e442543ae0b_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..ed8feb06298c9a8ec47a310c120b358c164f658e --- /dev/null +++ b/NeurIPS/2025/$_text{S}^2$Q-VDiT_ Accurate Quantized Video Diffusion Transformer with Salient Data and Sparse Token Distillation/591034a3-721c-487d-aba5-5e442543ae0b_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:64aa02becb10531fccf946fed60854d1c134546bb7c226a3310812f7e613e4af +size 162775 diff --git a/NeurIPS/2025/$_text{S}^2$Q-VDiT_ Accurate Quantized Video Diffusion Transformer with Salient Data and Sparse Token Distillation/591034a3-721c-487d-aba5-5e442543ae0b_model.json b/NeurIPS/2025/$_text{S}^2$Q-VDiT_ Accurate Quantized Video Diffusion Transformer with Salient Data and Sparse Token Distillation/591034a3-721c-487d-aba5-5e442543ae0b_model.json new file mode 100644 index 0000000000000000000000000000000000000000..d5e6a236f17289d8c8a7a4ece01538cd2061a1b1 --- /dev/null +++ b/NeurIPS/2025/$_text{S}^2$Q-VDiT_ Accurate Quantized Video Diffusion Transformer with Salient Data and Sparse Token Distillation/591034a3-721c-487d-aba5-5e442543ae0b_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4e94efd9d1df1dab42151d96fdbbd0983080b0cd33ad04d6df9c12087a9aefcf +size 215163 diff --git a/NeurIPS/2025/$_text{S}^2$Q-VDiT_ Accurate Quantized Video Diffusion Transformer with Salient Data and Sparse Token Distillation/591034a3-721c-487d-aba5-5e442543ae0b_origin.pdf b/NeurIPS/2025/$_text{S}^2$Q-VDiT_ Accurate Quantized Video Diffusion Transformer with Salient Data and Sparse Token Distillation/591034a3-721c-487d-aba5-5e442543ae0b_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..4570f5bc5375e0136a0d7d7bac5d114fbab43b12 --- /dev/null +++ b/NeurIPS/2025/$_text{S}^2$Q-VDiT_ Accurate Quantized Video Diffusion Transformer with Salient Data and Sparse Token Distillation/591034a3-721c-487d-aba5-5e442543ae0b_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fba936514efb0fa2216e04c15f55065460c44c9b583c909863a2ed8c4af67592 +size 39961321 diff --git a/NeurIPS/2025/$_text{S}^2$Q-VDiT_ Accurate Quantized Video Diffusion Transformer with Salient Data and Sparse Token Distillation/full.md b/NeurIPS/2025/$_text{S}^2$Q-VDiT_ Accurate Quantized Video Diffusion Transformer with Salient Data and Sparse Token Distillation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..fd8acf10c53adcad5f42f61a6c12e8c1956cdb29 --- /dev/null +++ b/NeurIPS/2025/$_text{S}^2$Q-VDiT_ Accurate Quantized Video Diffusion Transformer with Salient Data and Sparse Token Distillation/full.md @@ -0,0 +1,797 @@ +# $\mathbf{S}^{2}\mathbf{Q}$ -VDiT: Accurate Quantized Video Diffusion Transformer with Salient Data and Sparse Token Distillation + +Weilun Feng $^{1,2}$ , Haotong Qin $^{3}$ , Chuanguang Yang $^{1}$ , Xiangqi Li $^{1,2}$ , Han Yang $^{1}$ , Yuqi Li $^{1}$ , Zhulin An $^{1}$ , Libo Huang $^{1}$ , Michele Magno $^{3}$ , Yongjun Xu $^{1}$ $^{1}$ State Key Laboratory of AI Safety, Institute of Computing Technology, Chinese Academy of Sciences + $^{2}$ University of Chinese Academy of Sciences ETH Zürich +{fengweilun24s, yangchuanguang, lixiangqi24s, anzhulin, xyj}@ict.ac.cn +{haotong.qin, michele.magno}@pbl.ee.ethz.ch, {yuqili010602, www.huanglibo}@gmail.com + +![](images/e80732e5509769ae8549cb32e7d5521aa09e844ab1f10bb622b08d6f4a25e361.jpg) +Figure 1: We present $\mathrm{S}^2\mathrm{Q}$ -VDiT, a post-training quantization method for video diffusion transformers. We quantize HunyuanVideo [24] to 4-bit weights and 6-bit activations without compromising visual quality. $\mathrm{S}^2\mathrm{Q}$ -VDiT can further achieve $3.9\times$ model compression and $1.3\times$ inference acceleration. + +# Abstract + +Diffusion transformers have emerged as the mainstream paradigm for video generation models. However, the use of up to billions of parameters incurs significant computational costs. Quantization offers a promising solution by reducing memory usage and accelerating inference. Nonetheless, we observe that the joint modeling of spatial and temporal information in video diffusion models (V-DMs) leads to extremely long token sequences, which introduces high calibration variance and learning challenges. To address these issues, we propose $\mathbf{S}^2\mathbf{Q}$ -VDiT, a posttraining quantization framework for V-DMs that leverages Salient data and Sparse token distillation. During the calibration phase, we identify that quantization performance is highly sensitive to the choice of calibration data. To mitigate this, we introduce Hessian-aware Salient Data Selection, which constructs high-quality calibration datasets by considering both diffusion and quantization characteristics unique to V-DMs. To tackle the learning challenges, we further analyze the sparse attention patterns inherent in V-DMs. Based on this observation, we propose + +Attention-guided Sparse Token Distillation, which exploits token-wise attention distributions to emphasize tokens that are more influential to the model's output. Under W4A6 quantization, $S^2 Q$ -VDiT achieves lossless performance while delivering $3.9 \times$ model compression and $1.3 \times$ inference acceleration. Code will be available at https://github.com/wlfeng0509/s2q-vdit. + +# 1 Introduction + +In recent years, diffusion transformer [39] has emerged as a powerful generative paradigm, demonstrating remarkable performance across diverse domains such as image synthesis [6, 26, 9, 57], audio generation [15], and increasingly, video generation [37, 35]. Among these, video diffusion models (V-DMs) [58, 24] represent a new frontier by extending the spatial generative capabilities of image diffusion models (I-DMs) into the spatial-temporal domain, enabling high-quality video synthesis from textual prompts. + +However, the transition from image to video generation introduces substantial computational challenges, primarily due to the exponential growth in token count introduced by the temporal dimension [35, 58, 24]. These memory and compute demands become particularly severe in large-scale video generation models [35, 58, 24], which contain up to billions of parameters, where each input consists of thousands or even tens of thousands of tokens. To enable efficient deployment of such models in resource-constrained environments, post-training quantization (PTQ) [32, 52, 20, 5] has become a widely adopted approach. PTQ compresses the pre-trained models into low-bit representations without modifying the model weights, relying only on a small dataset to calibrate quantization parameters with only hours on a single GPU [51, 28]. + +While PTQ has proven effective for I-DMs [30, 45, 54], directly applying it to V-DMs leads to substantial performance degradation [2, 62]. Prior works [2, 54, 62] have sought to improve V-DMs' quantization performance primarily from the perspective of quantizer design. In this paper, we delve deeper into the PTQ challenges specific to V-DMs, focusing on calibration data and optimization methods. + +We identify that the long token sequences characteristic of V-DMs significantly constrain the number of calibration samples (e.g., thousands for I-DMs vs. only dozens for V-DMs under equal computational budgets). Under such limited budgets, quantization performance becomes highly sensitive to the selection of calibration samples. Existing methods [54, 2, 62] typically employ random or uniform sampling strategies, which work reasonably well for I-DMs but fail to generalize well to only dozens of data for V-DMs. Moreover, we observe that V-DMs exhibit sparse attention patterns across all tokens. Current PTQ optimization frameworks [54, 30] treat all tokens equally during loss alignment between full-precision and quantized models. However, this uniform treatment is suboptimal for long token sequences, where only a small subset of tokens significantly impacts the final output. These observations highlight two fundamental challenges in PTQ for V-DMs: (1) the absence of a principled method for selecting calibration samples, and (2) the inefficiency of uniform token treatment during optimization, despite the varying importance of tokens. + +To address these limitations, we propose $\mathbf{S}^2\mathbf{Q}$ -VDiT, a post-training quantization framework tailored for V-DMs, built upon Salient data selection and Sparse token distillation. An overview of the proposed framework is illustrated in Fig. 2. First, we introduce Hessian-aware Salient Data Selection, which constructs calibration datasets by jointly assessing diffusion informativeness and quantization sensitivity. We define a unified metric to quantify sample's saliency to the denoising process and its sensitivity to quantization perturbations. Second, we present Attention-guided Sparse Token Distillation, a technique that leverages the inherent sparsity of spatial-temporal attention in V-DMs. Rather than treating all tokens equally during optimization, we reweight quantization losses based on token-wise attention distribution, allowing the model to focus more on the impactful representations. + +Our main contribution can be summarized as follows: + +- We empirically identify that V-DMs suffer from high calibration data variance in quantization performance. We propose Hessian-aware Salient Data Selection, which jointly considers diffusion informativeness and quantization sensitivity to construct effective calibration datasets. + +![](images/65212797b376acf5dd799a94a2df5b072738c2e593b485e8d0c48e29d7dda1e4.jpg) +Hessian-aware Salient Data Selection + +![](images/2cbd35e8869ca34e2a4e10f2b6b03075f74c699f876083a168e7796fe707ab49.jpg) +Attention-guided Sparse Token Distillation +Figure 2: Overview of $\mathrm{S}^2\mathrm{Q}$ -VDiT. The framework includes Hessian-aware Salient Data Selection (SDS) for constructing calibration dataset and Attention-guided Sparse Token Distillation (STD) for block-wise optimization. + +- We introduce Attention-guided Sparse Token Distillation, a method that leverages the inherent sparsity in spatial-temporal attention of V-DMs. We reweight the quantization loss of different tokens by measuring token-wise attention distribution. This enables the model to focus more on the impactful representations during optimization. +- Extensive experiments on large-scale video diffusion transformers with 2B to 13B parameters demonstrate that our $\mathbf{S}^2\mathbf{Q}$ -VDiT consistently outperforms existing PTQ baselines, achieving state-of-the-art performance under all quantization settings. + +# 2 Related Works + +Diffusion models[46, 17] have demonstrated strong generative capabilities in video generation tasks. However, up to billions of parameters [35, 49, 58, 24] pose major challenges for deployment in resource-constrained environments. Quantization has emerged as a widely adopted solution for model compression and acceleration [40, 14, 3, 21, 25]. A growing body of work has explored post-training quantization (PTQ) for diffusion models, particularly focusing on U-Net-based architectures [30, 45, 16, 18, 29, 63]. For the Diffusion Transformer architecture specifically, recent methods [54, 2, 10] have made further explorations from the perspective of data distribution and architecture characteristics on quantization behavior. To address performance degradation at ultra-low bits, several quantization-aware training approaches have been proposed [64, 31, 36, 65, 8, 11]. While effective, these methods typically require extensive training time and large-scale datasets, making them less practical in many scenarios. + +Despite these advances, most existing quantization research remains focused on image diffusion models (I-DMs), with limited exploration of video diffusion models (V-DMs). ViDiT-Q [62] and Q-DiT [2] have made the first explorations on the quantization of V-DMs. Q-DiT [2] introduces automatic quantization granularity allocation for fine-grained quantizer selection. ViDiT-Q [62] proposes static-dynamic quantization strategy to enhance quantization accuracy. While these approaches improve performance from different perspectives, they primarily focus on quantization granularity and quantizer design. In this paper, we tackle V-DM quantization from a new angle—calibration data quality and optimization strategy. Our method achieves lossless performance on various large-scale video diffusion transformers from 2B to 13B. + +# 3 Methods + +# 3.1 Preliminary + +Video Diffusion Transformer. Diffusion transformers [39] predict the target using the representation of multiple tokens $X \in \mathbb{R}^{n \times d}$ where $n$ and $d$ represent the number of tokens and feature dimension, respectively. For image diffusion models (I-DMs) [42, 26], $n = s$ accounts for spatial + +tokens. But for video diffusion models (V-DMs) [58, 24, 37], $n = s \times t$ incorporates the temporal dimension $t$ . This significantly increases the token count per sample (e.g., $t = 49$ for a 6-second video at 8 FPS), resulting in heightened memory consumption and greater optimization complexity. + +Post-training Quantization. Quantization maps the model weights and activation to low-bit integers for acceleration and memory saving. For a float vector $x$ , the symmetry quantization process can be formulated as: + +$$ +x _ {\text {i n t}} = \operatorname {c l a m p} \left(\operatorname {r o u n d} \left[ \frac {x}{\Delta} \right], - 2 ^ {N - 1}, 2 ^ {N - 1} - 1\right), \Delta = \frac {\operatorname* {m a x} (a b s (x))}{2 ^ {N - 1} - 1} \tag {1} +$$ + +where $N$ is the quantized bit, round $(\cdot)$ is the round operation, and clamp $(\cdot)$ constrains the value within integer range $[-2^{N-1}, 2^{N-1}-1]$ . Among quantization methods, post-training quantization (PTQ) is a more efficient method that only calibrates quantization parameters using a small calibration dataset $D_{\mathrm{calib}}$ without altering model weights. According to common practice [1, 55, 62], the quantization loss is expressed as: + +$$ +\mathcal {L} _ {\text {q u a n t}} = \mathbb {E} _ {X \sim D _ {\text {c a l i b}}} [ \| \theta^ {f} (x) - \theta^ {q} (x) \| ^ {2} ], \tag {2} +$$ + +where $\theta^f$ and $\theta^q$ denote the full-precision and quantized model parameters, respectively. $D_{\mathrm{calib}} \in \mathbb{R}^{N \times n \times d}$ where $N$ denotes the sample number in $D_{\mathrm{calib}}$ . Due to the limitations in computing resources and long token sequences in V-DMs, the calibration sample size $N$ is smaller than that in I-DMs, leading to higher variance in data representation. This variance is further exacerbated by the diverse text prompts and different denoising timesteps present in the diffusion models. + +# 3.2 Hessian-aware Salient Data Selection + +![](images/c53dd04ba9bcd59199660f29d898b73f6e3da3bd55c6fae50ae828beacbaf5d5.jpg) +Figure 3: Visualization of different calibration data on CogVideoX-2B. We compare our proposed method with All Timesteps from One Prompt (ATOP), All Timesteps from Five Prompts (ATFP), and Random Timesteps from Five Prompts (RTFP). Our method has better generation quality. + +![](images/5964d8943824dc98df3261bf3dce2ab4480408bef04185977ed45c8e624ce5fe.jpg) + +![](images/79db0457ff2683d569be874b003187dfd496919b3b4dc3bf433f47bd01952c7a.jpg) + +![](images/112603560a6fc184acfd3178ebe8d8c6b495a57f18661773768df3538e7a7b10.jpg) + +![](images/93ebbccbd19a552f3e5075a79aeb3179bfb35c2a46872b95e10bc29f88f42741.jpg) + +Observation 1. Calibration sample selection methods result in high variance of the quantized model performance. + +In line with we discussed in Sec. 3.1, we observed that under constrained calibrated data size, different samples have significant differences in the final model performance as shown in Fig. 3 and Fig. 6a. However, the sample selection method for V-DMs post-training quantization has not been thoroughly explored. Therefore, we hope to evaluate the importance of different data for V-DMs. To address this issue, we propose evaluating sample utility along two dimensions that naturally exist in the quantization of diffusion models: contribution to the diffusion process and sensitivity to quantization. + +Prior work on timestep distillation [43, 44] and caching [33, 22] indicates that skipping certain consecutive timesteps has limited impact on output quality, suggesting varying information content across different timesteps. Based on this insight, we measure the salient information of timestep $t$ for the whole denoising diffusion process by calculating the contribution of two consecutive timesteps latent representation. Given all candidate data among all the diffusion timesteps $[x_{1}, x_{2}, \dots, x_{T}]$ where $T$ is the total denoising timesteps defined in the pretrained models. We define the diffusion salience as: + +$$ +C _ {\text {d i f f}} = \frac {\left| \left| x _ {t} - x _ {t - 1} \right| \right| ^ {2}}{\left| \left| x _ {t} \right| \right| ^ {2}}, \tag {3} +$$ + +where $x_{t}$ stands for the denoised feature of timestep $t$ . A higher $C_{\mathrm{diff}}$ value denotes more informative denoising steps, while a lower $C_{\mathrm{diff}}$ value indicates that the contained information largely overlaps with the previous timestep. $C_{\mathrm{diff}}$ naturally measures the saliency of different timesteps during the diffusion denoising process. By focusing on the salient data, we can better approximate the distribution of the entire diffusion process and achieve better performance. + +We then consider the quantization of weight $W$ and its quantized version $\hat{W} = W + \Delta$ , the quantization loss that jointly considers the input $X$ can be approximated using a Taylor expansion: + +$$ +\begin{array}{l} \mathbb {E} [ \| X W ^ {\top} - X \hat {W} ^ {\top} \| ^ {2} ] = \mathbb {E} [ \| X W ^ {\top} - X (W + \Delta) ^ {\top} \| ^ {2} ] \\ \approx \Delta \mathrm {g} ^ {X} + \frac {1}{2} \Delta \mathrm {H} ^ {X} \Delta^ {\top}, \tag {4} \\ \end{array} +$$ + +where $\mathrm{g}^X$ is the gradient and $\mathrm{H}^X$ is the Hessian matrix. Using $\mathrm{g}^X = 0$ for a well-trained model provided in [32, 59] and $\mathrm{H}^X = \mathbb{E}[2X^\top X]$ provided in [13], Eq. (4) can be further simplified to: + +$$ +\mathbb {E} [ \| X W ^ {\top} - X \hat {W} ^ {\top} \| ^ {2} ] \approx \mathbb {E} [ \Delta (X ^ {\top} X) \Delta^ {\top} ], \tag {5} +$$ + +where Hessian matrix $X^{\top}X$ is given by Levenberg-Marquardt approximation [12, 38]. The Hessian matrix represents the inherent perturbation ability of sample $X$ to the quantization process, which leads us to define quantization salience as: + +$$ +C _ {\text {q u a n t}} = \left\| x _ {t} ^ {\top} x _ {t} \right\| _ {2}, \tag {6} +$$ + +where a larger $C_{\mathrm{quant}}$ denotes that $x_{t}$ is more sensitive to the quantization process due to the property of the Hessian matrix [13, 12, 59]. By focusing on the quantization-sensitive samples, we can further relieve the bridge between the original data distribution and quantization operations, making the quantized model more robust and perform better. + +To jointly emphasize diffusion informativeness and quantization sensitivity, we apply min-max normalization over the candidate calibration pool $\mathcal{D}_{\mathrm{calib}}$ : + +$$ +\bar {C} _ {\text {d i f f}} \left(x _ {t}\right) = \frac {C _ {\text {d i f f}} \left(x _ {t}\right) - C _ {\text {d i f f}} ^ {\min }}{C _ {\text {d i f f}} ^ {\max } - C _ {\text {d i f f}} ^ {\min}}, \bar {C} _ {\text {q u a n t}} \left(x _ {t}\right) = \frac {C _ {\text {q u a n t}} \left(x _ {t}\right) - C _ {\text {q u a n t}} ^ {\min }}{C _ {\text {q u a n t}} ^ {\max } - C _ {\text {q u a n t}} ^ {\min}}, \tag {7} +$$ + +where $C_{\mathrm{diff}}^{\mathrm{min}}, C_{\mathrm{diff}}^{\mathrm{max}}, C_{\mathrm{quant}}^{\mathrm{min}}$ , and $C_{\mathrm{quant}}^{\mathrm{max}}$ denote the minimum value and maximum value of all $C_{\mathrm{diff}}(\cdot)$ and $C_{\mathrm{quant}}(\cdot)$ respectively. The unified salience score is then defined as the product: + +$$ +C _ {\text {s a m p l e}} \left(x _ {t}\right) = \bar {C} _ {\text {d i f f}} \left(x _ {t}\right) \cdot \bar {C} _ {\text {q u a n t}} \left(x _ {t}\right) \leq \left(\frac {\bar {C} _ {\text {d i f f}} \left(x _ {t}\right) + \bar {C} _ {\text {q u a n t}} \left(x _ {t}\right)}{2}\right) ^ {2}, \tag {8} +$$ + +by the Arithmetic-Geometric Mean inequality [67] which ensuring the score is maximized only when both normalized metrics are high. This mutual-salience product metric inherently penalizes samples that are only strong on one dimension, aligns with mutual-information objectives, and yields a more strong, robust calibration set. + +# 3.3 Attention-guided Sparse Token Distillation + +![](images/78707d265ba58ec661275392ed1f119b26b86b9514e2a060d005138c50c81a0f.jpg) +(a) Attention heatmaps. + +![](images/5db72d5cd05c120c07e576d5d17d7c6ef658c9214a725835504bb903e3a0f877.jpg) +(b) Token-wise attention distribution. +Figure 4: Visualization of sparse attention patterns in CogVideoX-2B block-10. In (4a), fewer columns have significantly higher weights. In (4b), only $10\%$ of tokens have larger attention weights. + +Observation 2. The fully spatial-temporal attention in V-DMs exhibits certain sparse patterns, suggesting that only subsets of tokens notably impact the model output. + +Table 1: Performance of 4-bit weight and 6-bit activation quantization on text-to-video generation under VBench evaluation benchmark suite. We evaluate on Imaging Quality (IQ), Aesthetic Quality (AQ), Motion Smoothness (MS), Dynamic Degree (DD), Background Consistency (BC), Subject Consistency (SuC), Scene Consistency (ScC), and Overall Consistency (OC). Higher $(\uparrow)$ metrics represent better performance. Bold: the best result. + +
ModelMethodIQAQMSDDBCSuCScCOC
CogVideoX-2BFP58.6955.2597.9550.0096.4094.3033.7925.91
Q-DiT48.6347.6398.0819.4495.3092.1523.8424.00
PTQ4DiT42.9145.4998.485.5695.6592.8517.8821.15
SmoothQuant44.6044.3398.229.7295.6292.0418.6021.20
Quarot51.8948.4897.4931.9495.6193.0122.9723.57
ViDiT-Q51.9448.0697.4733.3395.5492.8722.1723.69
S²Q-VDiT55.4953.7498.1040.2896.0594.1632.7025.19
CogVideoX-5BFP61.8058.8897.6172.2295.5694.6345.2826.46
Q-DiT49.9450.1897.0343.0695.5291.5829.6524.49
PTQ4DiT43.5442.7097.774.1796.7093.3210.9321.75
SmoothQuant39.5036.9297.886.9496.3992.2823.1118.19
Quarot43.9544.8197.3331.9496.5892.2720.9322.34
ViDiT-Q48.8750.5197.6637.5096.2593.6027.7623.57
S²Q-VDiT60.7556.9097.4658.3396.7694.2446.6626.30
HunyuanVideoFP62.3062.4999.0056.9498.0895.3033.3626.85
Q-DiT50.2348.4098.9540.2897.1494.0318.4614.41
PTQ4DiT48.3150.1398.2619.4497.9594.3720.1919.85
SmoothQuant47.5556.0398.7727.7897.3394.5723.6925.47
Quarot52.3158.5099.1337.5097.9895.3125.5126.01
ViDiT-Q52.2158.3899.1241.6798.0295.2023.6926.15
S²Q-VDiT58.8359.6299.2048.6198.1595.5733.6526.91
+ +Let $x \in \mathbb{R}^{n \times d}$ be the token embeddings, we can express Eq. (2) in the summation form as follows: + +$$ +\mathcal {L} _ {\text {q u a n t}} = \frac {1}{n} \sum_ {j = 1} ^ {n} \left| \left| \theta^ {f} \left(x _ {j,:}\right) - \theta^ {q} \left(x _ {j,:}\right) \right| \right| ^ {2}, \tag {9} +$$ + +where $x_{j,:}$ refers to the $j_{th}$ token in the video diffusion transformer. This loss function assumes that each token contributes equally to the overall error between the quantized and full-precision models. However, due to the spatial-temporal modeling objectives, V-DMs typically require large-scale pretraining to achieve full convergence [37, 58, 24, 56]. + +In the post-training quantization (PTQ) stage, only a small dataset is used to calibrate the quantization parameters, which naturally limits the model's ability to optimize from all tokens. Nevertheless, attention maps derived from V-DMs reveal that only subsets of tokens significantly influence the final output (see Fig. 4 and Appendix Sec. H). This observation aligns with prior studies on attention in V-DMs [60, 4, 61, 66], which have shown that pruning irrelevant tokens has a negligible impact on generation quality. These findings motivate a strategy that focuses learning more intensely on salient tokens while relaxing constraints on less impactful ones. Thereby enabling better convergence and improved performance even with limited calibration data. + +To improve alignment between quantized and full-precision outputs, we reweight each token's contribution in the loss function based on its influence on the block output. Formally, we modify Eq. (9) to: + +$$ +\mathcal {L} _ {\text {q u a n t}} = \frac {1}{n} \sum_ {j = 1} ^ {n} \lambda_ {j} \left| \left| \theta^ {f} \left(x _ {j,:}\right) - \theta^ {q} \left(x _ {j,:}\right) \right| \right| ^ {2}, \tag {10} +$$ + +where $\lambda_{j}$ denotes the weighting factor corresponding to token $x_{j,}$ . Leveraging the attention mechanism within each transformer block of V-DMs, we can obtain the complete multi-head attention + +Table 2: Performance of both 4-bit weight and activation quantization on text-to-video generation under VBenchmark evaluation benchmark suite. + +
ModelMethodIQAQMSDDBCSuCScCOC
CogVideoX-2BFP58.6955.2597.9550.0096.4094.3033.7925.91
Q-DiT26.2627.6699.14098.0996.521.168.43
PTQ4DiT20.6628.5099.30097.6195.332.1111.11
SmoothQuant29.7628.3198.95097.6294.655.319.74
QuaRot43.2239.5997.5413.8996.1892.3512.2119.57
ViDiT-Q45.5642.0397.5712.596.0892.4311.9119.61
S²Q-VDiT53.7152.3198.0936.1196.1593.9934.2324.90
CogVideoX-5BFP61.8058.8897.6172.2295.5694.6345.2826.46
Q-DiT40.8033.0095.7136.1198.2696.990.221.91
PTQ4DiT41.4828.6396.38097.2995.0907.37
SmoothQuant40.3029.9995.761.3996.5496.020.446.51
QuaRot29.4135.3697.7715.2897.2392.718.3615.31
ViDiT-Q31.9536.7197.0915.2896.3793.0110.8516.91
S²Q-VDiT58.7655.3597.1847.2296.2593.6936.5626.02
+ +map $A \in \mathbb{R}^{H \times n \times n}$ where $H$ is the number of attention heads. $A$ naturally represents the importance matrix of different tokens within each block, and $A_{h,i,j}$ denotes the attention weight $j_{th}$ token receives from the $i_{th}$ token in $h_{th}$ attention head. We use the attention map $A$ to compute $\lambda_j$ using: + +$$ +S _ {j} = \sum_ {h, i} A _ {h, i, j}, \lambda_ {j} = \frac {S _ {j} - \min (S)}{\max (S) - \min (S)} \left(\lambda_ {\max } - \lambda_ {\min }\right) + \lambda_ {\min }, \tag {11} +$$ + +where $\min(S)$ and $\max(S)$ denote the minimum and maximum values in all $S$ respectively. The hyperparameters $\lambda_{\min}$ and $\lambda_{\max}$ define the normalization range for token importance. Ultimately, $\lambda_j$ quantifies each token's salience and helps guide the optimization process to prioritize alignment for tokens that exert greater influence. + +# 4 Experiments + +# 4.1 Experimental and Evaluation Settings + +Quantization Scheme. We employ uniform per-channel weight quantization and dynamic per-token activation quantization with channel-wise scale and rotation matrix same as prior works [2, 1, 62]. We use symmetry quantization for both weight and activation for better hardware acceleration and memory saving. We follow the block-wise post-training strategy used in prior works [30, 54, 2]. More implementation details and model settings can be seen in Appendix Sec. A. + +Evaluation Settings. We conduct text-to-video experiment on different scale SOTA models CogVideoX-2B, CogVideoX-5B [58] and HunyuanVideo-13B [24] for better evaluation. We evaluate the performance of the quantized model using the VBenchmark [19], which provides a comprehensive evaluation on video generation performance. Same as the prior works [2, 62], we select 8 major evaluation dimensions from VBenchmark to ensure a thorough assessment. We also present more experiments on EvalCrafter [34] benchmark in Appendix Sec. D. As current works [2, 62] have achieved almost lossless performance at high bits (e.g., 6-8 bits), we evaluated the performance at more challenging and unexplored low-bit W4A6 and W4A4 settings. + +Compared Methods. Consist with prior works [2, 62], we compare $S^2 Q$ -VDiT with current PTQ baseline methods. For diffusion baseline, we compare with Q-DiT [2], PTQ4DiT [54], and ViDiT-Q [62]. We further compare with strong LLM baseline, SmoothQuant [55] and QuaRot [1]. + +![](images/b865459c9355e935fba15094276c77a72483dac776158628de72d4f10cba2cef.jpg) +Prompt: A panda standing on a surfboard in the ocean in sunset. +(a) CogVideoX-5B. + +![](images/104cb55e8c603d5fbae3ea29556881d087beaa469cb249c9d4f27499ac3e4a60.jpg) + +![](images/9e4e84e462b04eb5a0f7362ee01168bd9d2d4d2bb1836f177e1b53b846e378c6.jpg) + +![](images/67b989e5126b1be12694c29f2b571999ca65e5dfc05dfbb391219e38dc13bdd9.jpg) +Prompt: A robot DJ is playing the turntable, in heavy raining futuristic toyo rooftop cyberpunk night, sci-fi, fantasy. +(b) HunyuanVideo-13B. +Figure 5: Visual comparison on different models under W4A6 quantization setting. + +# 4.2 Quantitative Comparison + +We present text-to-video experiment under VBenchmark evaluation benchmark suite in Tab. 1 and Tab. 2. W4A6 Quantization. In Tab. 1, we focus on relatively higher bit quantization setting of W4A6 (4-bit weight and 6-bit activation). In three different scale current V-DMs CogVideoX-2B, CogVideoX-5B, and HunyuanVideo-13B, our method outperforms all current quantization methods by a notable margin. Our $S^2Q$ -VDiT achieves almost lossless performance across all eight selected dimensions. For CogVideoX-5B, $S^2Q$ -VDiT even outperforms FP model with 46.66 scene consistency while other methods achieved the highest score of 29.65. + +W4A4 Quantization. In Tab. 2, we further explored the quantization performance of V-DMs under extremely low bit W4A4 settings. It is worth noting that this is currently the first exploration under 4-bit activation quantization. In this extremely low bit setting, $\mathrm{S}^2\mathrm{Q}$ -VDiT can still maintain $95\%$ of the model's performance while other methods showed significant performance degradation. Although some methods are particularly high in metrics such as SuC and BC, this is due to their almost collapsed generation quality. ViDiT-Q [62] pointed out that these metrics are particularly high on extremely collapsed methods, and maintaining performance closer to FP is better. ForCogVideoX-2B, our method achieves even lossless scene consistency of 34.23 while other methods achieved the highest score of 12.21 with almost a three times improvement. + +# 4.3 Visual comparison + +We present visual comparisons on different models under W4A6 in Fig. 5. Compared with the current SOTA methods QuaRot [1] and ViDiT-Q [62], $S^2$ Q-VDiT has significant improvements in image quality and dynamic degree, and is lessloss compared to FP models. For CogVideoX-5B, QuaRot can hardly generate clear images; ViDiT-Q lacks the ability in color richness and image details; $S^2$ Q-VDiT is significantly better in color, detail, and video dynamics. For HunyuanVideo, although all methods have not significantly reduced image clarity, the semantic information of QuaRot has severely declined; the generated characters and background details of ViDiT-Q are also insufficient. $S^2$ Q-VDiT maintains high quality in the details and colors of both the background and characters, and ensures the dynamic level of the video at different frame. The consistent and significant improvement on three different scales V-DMs also demonstrates the generalization and effectiveness of our method. We provide more visual comparison in Appendix Sec. I. + +![](images/c450d4757d94a3e564e8eb83c4f1f4219cbf3b8e68843a82df829027a795f744.jpg) +(a) Ablation study on SDS. +Figure 6: Ablation study of proposed methods on W4A4 CogVideoX-2B. + +![](images/1fcb080dcdef2c288e974a2f4938768aa24f0ab0039d089887a1d283006d5bb9.jpg) +(b) Ablation study on STD. + +Table 3: Ablation study on calibration data size. + +
MethodData SizeCalibration Time (Hour)Imaging QualityAesthetic QualityOverall Consistency
FP--58.6155.2525.91
S²Q-VDiT201.6453.5653.0724.69
S²Q-VDiT402.8855.4953.7425.19
S²Q-VDiT805.5655.5253.6425.21
+ +# 4.4 Ablation Study + +In Fig. 6, we present ablation studies on Hessian-aware Salient Data Selection (SDS) and Attention-guided Sparse Token Distillation (STD). To verify the effectiveness of these techniques, we conducted integration experiments with existing PTQ methods in Appendix Sec. E. + +Ablation on SDS. We study different calibration data selection methods with our proposed SDS and shown the results in Fig. 6a. We compare three different straightforward methods, including All Timesteps from One Prompt (ATOP), All Timesteps from Five Prompts (ATFP), and Random Timesteps from Five Prompts (RTFP). We selected 40 samples for all methods for fair comparison. We also present the visual comparison in Fig. 3. Our proposed SDS outperforms all other methods in terms of both visual and metric effects while other methods can not maintain high generation quality. We conducted more ablation experiments on the random seeds and decoupled the two saliences used in SDS. We present the experimental results in Appendix Sec. B. We further conduct an ablation study on calibration data size in CogVideoX-2B under W4A6 setting and present the results in Tab. 3. It can be seen that the calibration time increases almost linearly with the increase of data size. The performance of 40 data is significantly better than that of 20 data, but the performance improvement of 80 data is minor. Therefore, in the trade-off of performance and calibration time, we choose to use 40 data as the unified experimental settings. + +Ablation on STD. In Fig. 6b, we compare our proposed STD with no sparse distillation (w/o STD). It can be seen that compared with no STD, all distillation strategies can improve model performance. We also compare different hyperparameters used in Eq (11). We set $\lambda_{max} = 1$ as default and investigate different $\lambda_{min}$ selections which control the relaxation degree on less impactful tokens. It can be seen that all different $\lambda_{min}$ can improve quantization performance which proves the robustness of STD. We select $\lambda_{min} = 0.5$ in the main experiments for the most balanced performance improvement. We provide more visualization of the sparse patterns in Appendix Sec. H. + +# 4.5 Efficiency Study + +We study the deployment efficiency of different-scale video diffusion transformers in Tab. 5. We used the CUDA implementation provided in [62, 47] for deployment and conducted all experiments + +on a single NVIDIA A800 GPU. For Inference Memory and Latency, we use a batch size of 1 in Tab 5. Compared with baseline method PTQ4DiT [54], our method brings significant performance improvement with almost no extra inference burden. Compared with FP model, our method can bring $3.94 \times$ model memory saving, $1.56 \times$ inference memory saving, and $1.28 \times$ inference acceleration on CogVideoX-5B. In Appendix Sec. F, we conducted more experiments on deployment efficiency. + +# 4.6 Calibration Resource Cost + +Table 4: Calibration cost on W4A4 CogVideoX-2B. + +
MethodGPU Memory (GB)GPU Time (hour)Imaging QualityAesthetic Quality
FP--58.6155.25
Q-DiT29.852.6926.2627.66
PTQ4DiT33.302.2520.6628.50
S²Q-VDiT35.682.8853.7152.31
+ +We reported on the calibration resource consumption of our $\mathrm{S}^2\mathrm{Q}$ -VDiT compared with existing baseline methods Q-DiT [2] and PTQ4DiT [54] in Tab. 4. Compared with existing methods, $\mathrm{S}^2\mathrm{Q}$ -VDiT only increases 2GB memory consumption and 0.2h calibration time, but improves Imaging Quality from 26.26 to 53.71, significantly enhancing the quantization performance. This proves the high efficiency and performance of $\mathrm{S}^2\mathrm{Q}$ -VDiT. We further reported more detailed calibration resource consumption of each proposed component in Appendix Sec. G. + +Table 5: Efficiency study on different W4A6 models. + +
ModelMethodModel Storage (GB)Inference Memory (GB)Latency (s)Imaging QualityAesthetic Quality
CogVideoX-5BFP10.37515.801259.261.8058.88
PTQ4DiT2.63310.139203.143.5442.70
S²Q-VDiT2.63310.145203.260.7556.90
HunyuanVideoFP23.88129.260191.362.3062.49
PTQ4DiT6.49413.703175.148.3150.13
S²Q-VDiT6.49413.713175.258.8359.62
+ +# 5 Conclusion + +In this paper, we propose $S^2 Q$ -VDiT, a post-training quantization framework for V-DMs using Salient data and Sparse token distillation. To address the sensitivity to calibration data, we propose Hessian-aware Salient Data Selection to construct high-quality datasets from the perspectives of diffusion and quantization. To address the learning challenge brought by long token sequences, we propose Attention-guided Sparse Token Distillation, which utilizes the natural sparse attention in V-DMs to allocate more loss weights to important tokens. Extensive experiments have shown that $S^2 Q$ -VDiT outperforms all existing methods on different scales of V-DMs. + +# Acknowledgements + +This work is partially supported by the National Natural Science Foundation of China under Grant Number 62476264 and 62406312, the Postdoctoral Fellowship Program and China Postdoctoral Science Foundation under Grant Number BX20240385 (China National Postdoctoral Program for Innovative Talents), the Beijing Natural Science Foundation under Grant Number 4244098, the Science Foundation of the Chinese Academy of Sciences, and Swiss National Science Foundation (SNSF) project 200021E_219943 Neuromorphic Attention Models for Event Data (NAMED). + +# References + +[1] Saleh Ashkboos, Amirkeivan Mohtashami, Maximilian L Croci, Bo Li, Pashmina Cameron, Martin Jaggi, Dan Alistarh, Torsten Hoefler, and James Hensman. Quarot: Outlier-free 4-bit inference in rotated llms. arXiv preprint arXiv:2404.00456, 2024. +[2] Lei Chen, Yuan Meng, Chen Tang, Xinzhu Ma, Jingyan Jiang, Xin Wang, Zhi Wang, and Wenwu Zhu. Q-dit: Accurate post-training quantization for diffusion transformers. arXiv preprint arXiv:2406.17343, 2024. +[3] Krishna Teja Chitty-Venkata, Sparsh Mittal, Murali Emani, Venkatram Vishwanath, and Arun K Somani. A survey of techniques for optimizing transformer inference. Journal of Systems Architecture, page 102990, 2023. +[4] Hangliang Ding, Dacheng Li, Runlong Su, Peiyuan Zhang, Zhijie Deng, Ion Stoica, and Hao Zhang. Efficient-vdlt: Efficient video diffusion transformers with attention tile. arXiv preprint arXiv:2502.06155, 2025. +[5] Yifu Ding, Weilun Feng, Chuyan Chen, Jinyang Guo, and Xianglong Liu. Reg-ptq: Regression-specialized post-training quantization for fully quantized object detector. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16174-16184, 2024. +[6] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In *Forty-first international conference on machine learning*, 2024. +[7] Yuming Fang, Hanwei Zhu, Yan Zeng, Kede Ma, and Zhou Wang. Perceptual quality assessment of smartphone photography. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3677-3686, 2020. +[8] Weilun Feng, Haotong Qin, Chuanguang Yang, Zhulin An, Libo Huang, Boyu Diao, Fei Wang, Renshuai Tao, Yongjun Xu, and Michele Magno. Mpq-dm: Mixed precision quantization for extremely low bit diffusion models. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, pages 16595-16603, 2025. +[9] Weilun Feng, Chuanguang Yang, Zhulin An, Libo Huang, Boyu Diao, Fei Wang, and Yongjun Xu. Relational diffusion distillation for efficient image generation. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 205–213, 2024. +[10] Weilun Feng, Chuanguang Yang, Haotong Qin, Xiangqi Li, Yu Wang, Zhulin An, Libo Huang, Boyu Diao, Zixiang Zhao, Yongjun Xu, et al. Q-vdit: Towards accurate quantization and distillation of video-generation diffusion transformers. arXiv preprint arXiv:2505.22167, 2025. +[11] Weilun Feng, Chuanguang Yang, Haotong Qin, Yuqi Li, Xiangqi Li, Zhulin An, Libo Huang, Boyu Diao, Fuzhen Zhuang, Michele Magno, et al. Mpq-dmv2: Flexible residual mixed precision quantization for low-bit diffusion models with temporal distillation. arXiv preprint arXiv:2507.04290, 2025. +[12] Elias Frantar and Dan Alistarh. Optimal brain compression: A framework for accurate posttraining quantization and pruning. Advances in Neural Information Processing Systems, 35:4475-4488, 2022. +[13] Elias Frantar, Saleh Ashkboos, Torsten Hoefler, and Dan Alistarh. Gptq: Accurate post-training quantization for generative pre-trained transformers. arXiv preprint arXiv:2210.17323, 2022. +[14] Amir Gholami, Sehoon Kim, Zhen Dong, Zhewei Yao, Michael W Mahoney, and Kurt Keutzer. A survey of quantization methods for efficient neural network inference. In Low-Power Computer Vision, pages 291-326. Chapman and Hall/CRC, 2022. +[15] Jiarui Hai, Yong Xu, Hao Zhang, Chenxing Li, Helin Wang, Mounya Elhilali, and Dong Yu. Ezaudio: Enhancing text-to-audio generation with efficient diffusion transformer. arXiv preprint arXiv:2409.10819, 2024. + +[16] Yefei He, Luping Liu, Jing Liu, Weijia Wu, Hong Zhou, and Bohan Zhuang. Ptqd: Accurate post-training quantization for diffusion models. Advances in Neural Information Processing Systems, 36, 2024. +[17] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. +[18] Yushi Huang, Ruihao Gong, Jing Liu, Tianlong Chen, and Xianglong Liu. Tfmq-dm: Temporal feature maintenance quantization for diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7362-7371, 2024. +[19] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, et al. Vbench: Comprehensive benchmark suite for video generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21807-21818, 2024. +[20] Itay Hubara, Yury Nahshan, Yair Hanani, Ron Banner, and Daniel Soudry. Improving post training neural quantization: Layer-wise calibration and integer programming. arXiv preprint arXiv:2006.10518, 2020. +[21] Benoit Jacob, Skirmantas Kligys, Bo Chen, Menglong Zhu, Matthew Tang, Andrew Howard, Hartwig Adam, and Dmitry Kalenichenko. Quantization and training of neural networks for efficient integer-arithmetic-only inference. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2704-2713, 2018. +[22] Kumara Kahatapitiya, Haozhe Liu, Sen He, Ding Liu, Menglin Jia, Chenyang Zhang, Michael S Ryoo, and Tian Xie. Adaptive caching for faster video generation with diffusion transformers. arXiv preprint arXiv:2411.02397, 2024. +[23] Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. Musiq: Multi-scale image quality transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5148-5157, 2021. +[24] Weijie Kong, Qi Tian, Zijian Zhang, Rox Min, Zuozhuo Dai, Jin Zhou, Jiangfeng Xiong, Xin Li, Bo Wu, Jianwei Zhang, et al. Hunyuanvideo: A systematic framework for large video generative models. arXiv preprint arXiv:2412.03603, 2024. +[25] Raghuraman Krishnamoorthi. Quantizing deep convolutional networks for efficient inference: A whitepaper. arxiv 2018. arXiv preprint arXiv:1806.08342, 1806. +[26] Black Forest Labs. Flux. https://github.com/black-forest-labs/flux, 2024. +[27] LAION-AI. Aesthetic-predictor, 2022. Accessed: 2022-04-16. +[28] Jiedong Lang, Zhehao Guo, and Shuyu Huang. A comprehensive study on quantization techniques for large language models. In 2024 4th International Conference on Artificial Intelligence, Robotics, and Communication (ICAIRC), pages 224-231. IEEE, 2024. +[29] Muyang Li, Yujun Lin, Zhekai Zhang, Tianle Cai, Xiuyu Li, Junxian Guo, Enze Xie, Chenlin Meng, Jun-Yan Zhu, and Song Han. Svdqunat: Absorbing outliers by low-rank components for 4-bit diffusion models. arXiv preprint arXiv:2411.05007, 2024. +[30] Xiuyu Li, Yijiang Liu, Long Lian, Huanrui Yang, Zhen Dong, Daniel Kang, Shanghang Zhang, and Kurt Keutzer. Q-diffusion: Quantizing diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17535-17545, 2023. +[31] Yanjing Li, Sheng Xu, Xianbin Cao, Xiao Sun, and Baochang Zhang. Q-dm: An efficient low-bit quantized diffusion model. Advances in Neural Information Processing Systems, 36, 2024. +[32] Yuhang Li, Ruihao Gong, Xu Tan, Yang Yang, Peng Hu, Qi Zhang, Fengwei Yu, Wei Wang, and Shi Gu. Brecq: Pushing the limit of post-training quantization by block reconstruction. arXiv preprint arXiv:2102.05426, 2021. + +[33] Feng Liu, Shiwei Zhang, Xiaofeng Wang, Yujie Wei, Haonan Qiu, Yuzhong Zhao, Yingya Zhang, Qixiang Ye, and Fang Wan. Timestep embedding tells: It's time to cache for video diffusion model. arXiv preprint arXiv:2411.19108, 2024. +[34] Yaofang Liu, Xiaodong Cun, Xuebo Liu, Xintao Wang, Yong Zhang, Haoxin Chen, Yang Liu, Tieyong Zeng, Raymond Chan, and Ying Shan. Evalcrafter: Benchmarking and evaluating large video generation models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22139-22149, 2024. +[35] Yixin Liu, Kai Zhang, Yuan Li, Zhiling Yan, Chujie Gao, Ruoxi Chen, Zhengqing Yuan, Yue Huang, Hanchi Sun, Jianfeng Gao, et al. Sora: A review on background, technology, limitations, and opportunities of large vision models. arXiv preprint arXiv:2402.17177, 2024. +[36] Xudong Lu, Aojun Zhou, Ziyi Lin, Qi Liu, Yuhui Xu, Renrui Zhang, Yafei Wen, Shuai Ren, Peng Gao, Junchi Yan, et al. Terdit: Ternary diffusion models with transformers. arXiv preprint arXiv:2405.14854, 2024. +[37] Xin Ma, Yaohui Wang, Gengyun Jia, Xinyuan Chen, Ziwei Liu, Yuan-Fang Li, Cunjian Chen, and Yu Qiao. Latte: Latent diffusion transformer for video generation. arXiv preprint arXiv:2401.03048, 2024. +[38] Donald W Marquardt. An algorithm for least-squares estimation of nonlinear parameters. Journal of the society for Industrial and Applied Mathematics, 11(2):431-441, 1963. +[39] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. +[40] Ratko Pilipović, Patricio Bulić, and Vladimir Risojević. Compression of convolutional neural networks: A short survey. In 2018 17th International Symposium INFOTEH-JAHORINA (INFOTEH), pages 1-6. IEEE, 2018. +[41] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. +[42] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. +[43] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. arXiv preprint arXiv:2202.00512, 2022. +[44] Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. In European Conference on Computer Vision, pages 87-103. Springer, 2024. +[45] Yuzhang Shang, Zhihang Yuan, Bin Xie, Bingzhe Wu, and Yan Yan. Post-training quantization on diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1972-1981, 2023. +[46] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. +[47] Yuxuan Sun, Ruikang Liu, Haoli Bai, Han Bao, Kang Zhao, Yuening Li, Jiaxin Hu, Xianzhi Yu, Lu Hou, Chun Yuan, et al. Flatquant: Flatness matters for llm quantization. arXiv preprint arXiv:2410.09426, 2024. +[48] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16, pages 402-419. Springer, 2020. +[49] Team Wan, Ang Wang, Baole Ai, Bin Wen, Chaojie Mao, Chen-Wei Xie, Di Chen, Feiwu Yu, Haiming Zhao, Jianxiao Yang, et al. Wan: Open and advanced large-scale video generative models. arXiv preprint arXiv:2503.20314, 2025. + +[50] Yi Wang, Yinan He, Yizhuo Li, Kunchang Li, Jiashuo Yu, Xin Ma, Xinhao Li, Guo Chen, Xinyuan Chen, Yaohui Wang, et al. Internvid: A large-scale video-text dataset for multimodal understanding and generation. arXiv preprint arXiv:2307.06942, 2023. +[51] Lu Wei, Zhong Ma, Chaojie Yang, and Qin Yao. Advances in the neural network quantization: A comprehensive review. Applied Sciences, 14(17):7445, 2024. +[52] Xiuying Wei, Ruihao Gong, Yuhang Li, Xianglong Liu, and Fengwei Yu. Qdrop: Randomly dropping quantization for extremely low-bit post-training quantization. arXiv preprint arXiv:2203.05740, 2022. +[53] Haoning Wu, Erli Zhang, Liang Liao, Chaofeng Chen, Jingwen Hou, Annan Wang, Wenxiu Sun, Qiong Yan, and Weisi Lin. Exploring video quality assessment on user generated contents from aesthetic and technical perspectives. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20144-20154, 2023. +[54] Junyi Wu, Haoxuan Wang, Yuzhang Shang, Mubarak Shah, and Yan Yan. Ptq4dit: Post-training quantization for diffusion transformers. arXiv preprint arXiv:2405.16005, 2024. +[55] Guangxuan Xiao, Ji Lin, Mickael Seznec, Hao Wu, Julien Demouth, and Song Han. Smoothquant: Accurate and efficient post-training quantization for large language models. In International Conference on Machine Learning, pages 38087-38099. PMLR, 2023. +[56] Chuanguang Yang, Helong Zhou, Zhulin An, Xue Jiang, Yongjun Xu, and Qian Zhang. Cross-image relational knowledge distillation for semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12319-12328, 2022. +[57] Han Yang, Chuanguang Yang, Qiuli Wang, Zhulin An, Weilun Feng, Libo Huang, and Yongjun Xu. Multi-party collaborative attention control for image customization. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 7942-7951, 2025. +[58] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024. +[59] Zhihang Yuan, Chenhao Xue, Yiqi Chen, Qiang Wu, and Guangyu Sun. Ptq4vit: Post-training quantization for vision transformers with twin uniform quantization. In European conference on computer vision, pages 191–207. Springer, 2022. +[60] Jintao Zhang, Chengdong Xiang, Haofeng Huang, Jia Wei, Haocheng Xi, Jun Zhu, and Jianfei Chen. Spargeattn: Accurate sparse attention accelerating any model inference. arXiv preprint arXiv:2502.18137, 2025. +[61] Peiyuan Zhang, Yongqi Chen, Runlong Su, Hangliang Ding, Ion Stoica, Zhenghong Liu, and Hao Zhang. Fast video generation with sliding tile attention. arXiv preprint arXiv:2502.04507, 2025. +[62] Tianchen Zhao, Tongcheng Fang, Enshu Liu, Wan Rui, Widyadewi Soedarmadji, Shiyao Li, Zinan Lin, Guohao Dai, Shengen Yan, Huazhong Yang, et al. Vidit-q: Efficient and accurate quantization of diffusion transformers for image and video generation. arXiv preprint arXiv:2406.02540, 2024. +[63] Tianchen Zhao, Xuefei Ning, Tongcheng Fang, Enshu Liu, Guyue Huang, Zinan Lin, Shengen Yan, Guohao Dai, and Yu Wang. Mixdq: Memory-efficient few-step text-to-image diffusion models with metric-decoupled mixed precision quantization. In European Conference on Computer Vision, pages 285-302. Springer, 2025. +[64] Xingyu Zheng, Xianglong Liu, Yichen Bian, Xudong Ma, Yulun Zhang, Jiakai Wang, Jinyang Guo, and Haotong Qin. Bidm: Pushing the limit of quantization for diffusion models. arXiv preprint arXiv:2412.05926, 2024. + +[65] Xingyu Zheng, Haotong Qin, Xudong Ma, Mingyuan Zhang, Haojie Hao, Jiakai Wang, Zixiang Zhao, Jinyang Guo, and Xianglong Liu. Binarydm: Towards accurate binarization of diffusion model. arXiv preprint arXiv:2404.05662, 2024. +[66] Chang Zou, Xuyang Liu, Ting Liu, Siteng Huang, and Linfeng Zhang. Accelerating diffusion transformers with token-wise feature caching. In The Thirteenth International Conference on Learning Representations, 2025. +[67] Limin Zou and Youyi Jiang. Improved arithmetic-geometric mean inequality and its application. J. Math. Inequal, 9(1):107–111, 2015. + +# A Implementation Details + +In the main experiment, we use 10 random prompts for generating the candidate calibration samples. We finally selected 40 samples for post-training quantization for all methods. For our method, we use a channel-wise scale used in [55, 62, 54] and a rotation matrix used in [47] for linear quantization. We further use a learnable threshold for clipping the weight and activation min-max value as prior work [30, 18, 47]. We also use GPTQ weight quantizer [13] for our experiment, following prior work [2]. We conduct all the experiments on a single NVIDIA A800 GPU. + +For optimization, we train the diag-balancing scale, rotation-based matrix, and learnable clipping threshold following the layer-wise post-training quantization framework as prior works [30, 54]. We use 30 samples and train 15 epochs for each layer. We use AdamW optimizer and cosine learning rate scheduler. For the diag-balancing scale and rotation-based matrix, we use a learning rate of 5e-3. For the learnable clipping threshold, we use a learning rate of 5e-2. + +For deployment, we absorb all weight quantization parameters as prior works [54, 55, 62], which brings no extra burden. For activation quantization, we apply online dynamic quantization following [62, 1]. + +# B More Ablation on Hessian-aware Salient Data Selection + +Table 6: Performance of both 4-bit weight and activation quantization on CogVideoX-2B under three random seeds. + +
MethodImaging +QualityAesthetic +QualityMotion +Smooth.Dynamic +DegreeBG +Consist.Subject +Consist.Scene +Consist.Overall +Consist.
-58.6955.2597.9550.0096.4094.3033.7925.91
ATOS51.65±(1.76)49.79±(0.59)98.09±(0.16)29.17±(3.40)95.82±(0.35)93.24±(0.19)29.94±(1.35)24.31±(0.37)
ATDS50.63±(0.81)50.13±(0.25)98.05±(0.11)29.63±(2.62)95.94±(0.16)93.16±(0.41)30.98±(2.14)24.11±(0.27)
DTDS50.66±(1.04)50.33±(0.19)98.03±(0.14)31.48±(4.58)96.01±(0.16)93.07±(0.18)30.47±(1.77)24.75±(0.25)
DS52.73±(0.98)50.62±(0.81)98.15±(0.19)31.75±(2.73)96.06±(0.18)93.29±(0.15)31.38±(0.98)24.78±(0.22)
QS52.34±(0.85)51.17±(0.23)98.11±(0.12)32.01±(2.97)96.10±(0.17)93.57±(0.19)31.86±(0.90)24.79±(0.23)
SDS(Ours)52.95±(0.69)51.58±(0.11)98.16±(0.09)32.87±(2.36)96.13±(0.15)93.89±(0.17)32.75±(0.77)24.84±(0.26)
+ +In this section, we investigate the random seed influence on the quantization performance of different calibration datasets mentioned in Sec. 3.2 and Sec. 4.4. We compare our proposed Hessian-aware Salient Data Selection (SDS) with All Timesteps from One Prompt (ATOP), All Timesteps from Five Prompts (ATFP), and Random Timesteps from Five Prompts (RTFP) using three different random seeds. We further decoupled SDS into Diffusion Salience (DS) in Eq. (3) and Quantization Salience (QS) in Eq. (6) and reported the performance. We present the average results and variance in Tab. 6. + +Other straightforward sampling methods have lower average performance and larger variances, proving the influence of random seeds in these random sampling methods. Using our proposed diffusion salience (DS) or quantization salience (QS) can all improve the performance and reduce the impact of random seeds. Only using DS and QS can improve Scene Consistency to over 31 with variances less than 1, while other random sampling methods cannot achieve. By jointly considering two saliences, Hessian-aware Salient Data Selection (SDS) can achieve the best quantization performance with minimal impact from randomness. SDS achieved an average Imaging Quality of 52.95 with only 0.69 variance, while the random sampling only achieved the best average of 51.65 with 1.67 variance. + +# C Detailed Description of Selected Evaluation Metrics + +# C.1 VBenchmark + +For VBenchmark [19] benchmark, we follow the previous work ViDiT-Q [62], which selects 8 dimensions from three key aspects in video-generation task. + +Frame-wise Quality: In this aspect, we assess the quality of each individual frame without taking temporal quality into concern. + +- Imaging Quality assesses distortion (e.g., over-exposure, noise) presented in the generated frames using the MUSIQ [23] image quality predictor trained on the SPAQ [7] dataset. +- Aesthetic Quality evaluates the artistic and beauty value perceived by humans towards each video frame using the LAION aesthetic predictor [27]. + +Temporal Quality: In this aspect, we assess the cross-frame temporal consistency and dynamics. + +- Dynamic Degree evaluates the degree of dynamics (i.e., whether it contains large motions) generated by each model. +- Motion Smoothness evaluates whether the motion in the generated video is smooth, and follows the physical law of the real world. +- Subject Consistency assesses whether the subject's appearance remains consistent through out the whole video. +- Background Consistency evaluate the temporal consistency of the background scenes by calculating CLIP [41] feature similarity across frames. + +Semantics: In this aspect, we evaluate the video's adherence to the text prompt given by the user. consistency. + +- Scene evaluates whether the synthesized video is consistent with the intended scene described by the text prompt. +- Overall Consistency further use overall video-text consistency computed by ViCLIP [50] on general text prompts as an aiding metric to reflect both semantics and style consistency + +We use three different prompt sets provided by the official github repository of VBench [19] to generate videos. We generate one video for each prompt for evaluation same as ViDiT-Q [62]. + +- overall consistency.txt: includes 93 prompts, used to evaluate overall consistency, aesthetic quality, and imaging quality. +- subject consistency.txt: includes 72 prompts, used to evaluate subject consistency, dynamic degree, and motion smoothness. +- scene.txt: includes 86 prompts, used to evaluate scene and background consistency. + +# C.2 EvalCrafter Benchmark + +For EvalCrafter [34] benchmark, consistent with prior work ViDiT-Q [62], we select 5 low-level metrics to evaluate the generation performance. + +CLIPSIM and CLIP-Temp: CLIPSIM computes the image-text CLIP similarity for all frames in the generated videos, and we report the averaged results. This quantifies the similarity between input text prompts and generated videos. CLIP-Temp computes the CLIP similarity of each two consecutive frames of the generated videos and then gets the averages for each two frames. This quantifies the semantic consistency of generated videos. We use the CLIP-VIT-B/32 [50] model to compute CLIPSIM and CLIP-Temp. We use the implementation from EvalCrafter [34] to compute these two metrics. + +DOVER's VQA: VQA-Technical measures common distortions like noise, blur, and over-exposure. VQA-Aesthetic reflects aesthetic aspects such as the layout, the richness and harmony of colors, the photo-realism, naturalness, and artistic quality of the frames. We use the Dover [53] method to compute these two metrics. + +FLOW Score: Flow score was proposed in [34] to measure the general motion information of the video. We use RAFT [48] to extract the dense flows of the video in every two frames, and we calculate the average flow on these frames to obtain the average flow score of each generated video. + +We use the prompt sets provided by the official github repository of ViDiT-Q [62] to generate 10 videos for evaluation. We also attached the prompt sets in the supplementary material. + +# D Experiments on more metrics + +Following prior work [62], we evaluate different methods on EvalCrafter [34] benchmark for multi-aspects metrics evaluation. We select CLIPSIM, CLIP-Temp, DOVER [53] video quality assessment (VQA) metrics to evaluate the generation quality, and Flow-score to evaluate the temporal consistency. We conduct experiments on CogVideoX-2B, CogVideoX-5B, and HunyuanVideo-13B under W4A6 quantization setting. We present the evaluation results in Tab. 7. + +Table 7: Performance of 4-bit weight and 6-bit activation quantization on text-to-video generation under EvalCrafter benchmark. Higher (↑) metrics represent better performance. + +
ModelMethodCLIPSIMCLIP-TempVQA-AestheticVQA-TechnicalFLOW Score.
CogVideoX-2BFP0.18440.997876.6485.023.452
Q-DiT0.17870.997863.1567.372.331
PTQ4DiT0.17720.998558.7652.601.837
SmoothQuant0.17620.998155.1853.871.378
Quarot0.18080.997551.8356.792.867
ViDiT-Q0.18120.997653.0959.843.040
S²Q-VDiT0.18380.997970.5073.313.122
CogVideoX-5BFP0.18140.998278.8773.174.536
Q-DiT0.18350.997647.9646.722.967
PTQ4DiT0.17890.998422.9344.072.230
SmoothQuant0.17420.99763.0514.131.026
Quarot0.18050.998333.1043.673.040
ViDiT-Q0.17950.998042.0148.591.850
S²Q-VDiT0.18190.998773.4574.413.688
HunyuanVideoFP0.19100.998580.6663.511.674
Q-DiT0.18710.998756.4543.171.482
PTQ4DiT0.17860.997342.1733.691.089
SmoothQuant0.17820.99787.240.420.111
Quarot0.18730.997766.4952.810.899
ViDiT-Q0.18950.997866.2351.350.897
S²Q-VDiT0.19020.998577.8066.381.562
+ +It can be seen that under the EvalCrafter [34] benchmark, our $\mathrm{S}^2\mathrm{Q}$ -VDiT still achieved almost lossless performance and showed significant performance improvement compared to all comparison methods. Especially in terms of VQA-Technical metrics, our $\mathrm{S}^2\mathrm{Q}$ -VDiT even outperforms the full precision model on CogVideoX-5B and HunyuanVideo, while other methods show notable performance degradation. For CogVideoX-5B, $\mathrm{S}^2\mathrm{Q}$ -VDiT achieves 74.41 in VQA-Technical which outperforms the full precision model of 73.17, while current methods achieve the best of 48.59. + +# E Integration with Existing PTQ Methods + +The techniques that we proposed Hessian-aware Salient Data Selection (SDS) and Attention-guided Sparse Token Distillation (STD) can also be applied to existing block-wise optimization-based post-training quantization methods. To verify the generality of these two techniques, we combined them with the existing baseline method PTQ4DiT [54] and reported the performance improvement of these techniques on W4A6 CogVideoX-2B under VBenchmark [19] benchmark in Tab. 8. By using the calibration constructed by SDS, we further improved the performance of PTQ4DiT and increased Aesthetic Quality by 1.4. This demonstrates the improvement of SDS-constructed datasets under different optimization frameworks. From optimization perspective, we further improved the Aesthetic Quality to 47.27 by using sparse distillation STD. This also demonstrates the effectiveness and generalization of our attention-based optimization method. + +Table 8: Performance of 4-bit weight and 6-bit activation quantization on CogVideoX-2B under VBenchmark evaluation benchmark suite + +
MethodImaging QualityAesthetic QualityMotion Smooth.Dynamic DegreeBG Consist.Subject Consist.Scene Consist.Overall Consist.
FP58.6955.2597.9550.0096.4094.3033.7925.91
PTQ4DiT42.9145.4998.485.5695.6592.8517.8821.15
+SDS43.0646.8998.6411.1195.7993.3318.1022.27
+STD43.0847.2798.789.7295.9793.6819.0422.09
+ +# F More Experiments on Deployment Efficiency + +![](images/056d8c9509e3c23655610e9e3a186a95ef7bba10f265bd31b43deeb57d0b9d29.jpg) +(a) CogVideoX-2B. + +![](images/3d63971002dbc83adbc11798fe458b5b0a789b4b622c7cff696fcfc363405a3a.jpg) +(b) CogVideoX-5B. + +![](images/54e2c09cce12af01634f4b23736cb89cbe3d17379eec1ad35595c5d86b7e458f.jpg) +(c) HunyuanVideo-13B. +Figure 7: Deployment latency comparison under different batch size. + +We further expanded the experiments provided in Sec. 4.5. We compared the deployment efficiency of different models under different batch sizes in Fig. 7. Our $S^2$ Q-VDiT can bring consistent inference acceleration to different models under different batch sizes. Under the 50-step inference setting of CogVideoX-5B with a batch size of 4, our $S^2$ Q-VDiT can reduce the inference latency from 945.4s to 782.5s, achieving a significant acceleration of 163 seconds and outperforming the baseline method PTQ4DiT [54]. + +Table 9: Calibration cost about each component. + +
Hessian ApproximationAttention Computation
MethodConstruct Time (mins)Imaging QualityMethodCalibration Time (hours)Imaging Quality
CogVideoX-2B
FP-58.69FP-58.69
w/o Hessian7.70853.16w/o Attention2.8252.16
w Hessian7.71755.49w Attention2.8455.49
CogVideoX-5B
FP-61.80FP-61.80
w/o Hessian20.71958.91w/o Attention3.9758.23
w Hessian20.73460.75w Attention4.0060.75
HunyuanVideo-13B
FP-62.30FP-62.30
w/o Hessian19.50557.25w/o Attention5.7056.94
w Hessian19.50858.83w Attention5.7358.83
+ +# G More Detailed Calibration Resource Cost + +We reported the time increase caused by using the Hessian approximation when constructing the calibration dataset and the attention scores calculation across different scale video generation models in Tab. 9. + +It can be seen that the computational burden of using Hessian approximation is minor, but it can bring significant performance improvement. We use the Levenberg-Marquardt approximation [13] to calculate the Hessian approximation, which requires only one step matrix multiplication to obtain the approximate result, and is very efficient. + +Also, during the calibration process, we only need to use the Full-Precision model to conduct a single forward calculation of attention scores for all data in advance. When optimizing the quantization model, we can directly get the pre-computed attention scores by the data index, which brings minimal burden. + +# H More Visualization about Sparse Attention Pattern + +![](images/af0129dd96e235880afddc7ebb8033b17a9b3624a224658716d79b9880d50c45.jpg) +(a) Block-5. + +![](images/8e2de53a01588ccecd322e5ad268a23a5d74ae609626bd68489dc8ad4e6c1cd9.jpg) +(b) Block-12. + +![](images/bfaa4d8aaa6cbb40ee119da30dc7ddf5c4b703050ff2a07bc310a9001a089f5b.jpg) +(c) Block-13. + +![](images/3fcafcd4e59f30caf8d5d0d7b900c0f788d3d8644ffa831d0f2f3ca525fbfca0.jpg) +(d) Block-14. + +![](images/b25ff1ab9d139f6c1599b0b0d4e91b01352284b441fd20c8318c712b3c6ffe14.jpg) +(e) Block-15. + +![](images/000dfebb1a705817e83fecbfe6c741552c0bb1ade09dd1189d0f43efcc3e660f.jpg) +(f) Block-17. + +![](images/c283b1a7c019c2fa696995736718a5e4d5f4fe38ce8376cbb7c65136defcecd9.jpg) +(g) Block-18. +Figure 8: Visualization of attention heatmaps in CogVideoX-2B. + +![](images/637990258bf8ff85c645f700e2bceac8aa1b59162127e8abdba7f1c6ef7f22e2.jpg) +(h) Block-19. + +![](images/78cab21b1cef7aadb91d988fd2a44f8c110eefaaa91ad6ce786b742c0d636c0a.jpg) +(a) Block-5. + +![](images/f1a46dfa60e3fa0c2b799662589492a9218251a5cb796c798ee89d40dfc712bc.jpg) +(b) Block-12. + +![](images/04c9c7985e95ee141cff74f62a56e628ace68fce451840162572ebd412698803.jpg) +(c) Block-13. + +![](images/0e22b9520e9288eff5ff894720d86b97e90f5c21deb55e42c39487a7d5d01eda.jpg) +(d) Block-14. + +![](images/e24e4127c0359e751d1e75847c034e9cca7500b2240557f9062b5c095bc1b6c7.jpg) +(e) Block-15. +Figure 9: Visualization of token-wise attention distribution in CogVideoX-2B. + +![](images/e5c5dde7cbc996a719bf59838fa1f6a82e92bf32a6f6e6021db84e51563f099d.jpg) +(f) Block-17. + +![](images/3c8a53b9d190103adce39bbee6c12dd8fca6faba45dbc8983ae03c0c8f03480c.jpg) +(g) Block-18. + +![](images/32a0c120bddb01d7b9a1a2fe79d87d2e73e8bc81a94cdb0561f9e3d5fcb75a60.jpg) +(h) Block-19. + +We demonstrate the sparse attention patterns existing in V-DMs that we mentioned in Sec 3.3. We present more visualization results of different blocks of CogVideoX-2B in Fig. 8 and Fig. 9. There is a considerable degree of sparse attention patterns in the most layers of the model, and almost all $90\%$ tokens have significantly lower attention weights than the top $10\%$ tokens. This indicates that + +![](images/68ff816dbee70bb56fe6d86e376ba6f2791e0264d4d32ee901dbe9c6c0f977b2.jpg) +Figure 10: HunyuanVideo-13B results. Prompt: A cat wearing sunglasses on a beach. + +sparse attention is commonly present in V-DMs, and almost every layer only has a small portion of tokens that play an important role in the final output. This proves the universality of our observations in Sec. 3.3 and the effectiveness of our Attention-guided Sparse Token Distillation. + +# I More Visualization Results + +We present more visual comparison results on HunyuanVideo-13B [24], CogVideoX-5B, and CogVideoX-2B [58] under W4A6 quantization in the following figures. Compared with current methods SmoothQuant [55], Q-DiT [2], ViDiT-Q [62], our $\mathrm{S}^2\mathrm{Q}$ VDiT made notable visual improvement on different scale video diffusion models. This proves that our $\mathrm{S}^2\mathrm{Q}$ VDiT not only surpasses existing methods in terms of evaluation metrics but also shows significant improvement in visual effects, demonstrating the effectiveness of our $\mathrm{S}^2\mathrm{Q}$ VDiT. + +# J Limitations + +Although our $S^2\mathrm{Q}$ -VDiT outperforms existing methods, it cannot achieve completely lossless performance under the most difficult fully 4-bit quantization. We hope to optimize the quantization performance under low bit settings in the future. + +# K Broader Impacts + +Our quantized model may be used by people to generate false content, and we will require users to apply our model in legitimate and reasonable scenarios and label it as AI-generated. + +![](images/bdf5e5e126c8c036c470d7a0d2657671c4218c160a837fa5d668ab4a151a0f75.jpg) +Figure 11: HunyuanVideo-13B results. Prompt: A boat sailing leisurely along the Seine River with the Eiffel Tower in background. + +![](images/3e9e29bacb3305635011fe799c36576c7f798080b2a19bb4422493028fd53c5e.jpg) +Figure 12: HunyuanVideo-13B results. Prompt: A panda cooking in the kitchen. + +![](images/81c8ef2076561294c8f9c209b753c4298bd37a2b1e15530aabaf5fea2b72922d.jpg) +Figure 13: CogVideoX-5B results. Prompt: A beautiful coastal beach in spring, waves lapping on sand by Hokusai, in the style of Ukiyo. + +![](images/785c045c429243149b70e8fc1527c13aaf00065728f05ed5c6d8e084934ed451.jpg) +Figure 14:CogVideoX-5B results. Prompt: A modern art museum, with colorful paintings. + +![](images/baf1f6acd16d8bbb6b55644598cf8f943f54add9886e144e9c02739097ec6d5f.jpg) +Figure 15:CogVideoX-5B results. Prompt: Yoda playing guitar on the stage. + +![](images/d51b668e2cbbb873d7f6823a694e5b861954f89516910454b509b9416dc1d0a4.jpg) +Figure 16: CogVideoX-2B results. Prompt: Macro slo-mo. Slow motion cropped closeup of roasted coffee beans falling into an empty bowl. + +![](images/2230be29683f0ccc52e17c630b05688dba90333aa3b64469c70a6975102ceb2d.jpg) +Figure 17: CogVideoX-2B results. Prompt: A boat sailing leisurely along the Seine River with the Eiffel Tower in background by Vincent van Gogh. + +# NeurIPS Paper Checklist + +# 1. Claims + +Question: Do the main claims made in the abstract and introduction accurately reflect the paper's contributions and scope? + +Answer: [Yes] + +Justification: The claims reflect the paper's contributions and scope. + +Guidelines: + +- The answer NA means that the abstract and introduction do not include the claims made in the paper. +- The abstract and/or introduction should clearly state the claims made, including the contributions made in the paper and important assumptions and limitations. A No or NA answer to this question will not be perceived well by the reviewers. +- The claims made should match theoretical and experimental results, and reflect how much the results can be expected to generalize to other settings. +- It is fine to include aspirational goals as motivation as long as it is clear that these goals are not attained by the paper. + +# 2. Limitations + +Question: Does the paper discuss the limitations of the work performed by the authors? + +Answer: [Yes] + +Justification: We discuss the limitations of the work in Appendix Sec. J. + +Guidelines: + +- The answer NA means that the paper has no limitation while the answer No means that the paper has limitations, but those are not discussed in the paper. + +- The authors are encouraged to create a separate "Limitations" section in their paper. +- The paper should point out any strong assumptions and how robust the results are to violations of these assumptions (e.g., independence assumptions, noiseless settings, model well-specification, asymptotic approximations only holding locally). The authors should reflect on how these assumptions might be violated in practice and what the implications would be. +- The authors should reflect on the scope of the claims made, e.g., if the approach was only tested on a few datasets or with a few runs. In general, empirical results often depend on implicit assumptions, which should be articulated. +- The authors should reflect on the factors that influence the performance of the approach. For example, a facial recognition algorithm may perform poorly when image resolution is low or images are taken in low lighting. Or a speech-to-text system might not be used reliably to provide closed captions for online lectures because it fails to handle technical jargon. +- The authors should discuss the computational efficiency of the proposed algorithms and how they scale with dataset size. +- If applicable, the authors should discuss possible limitations of their approach to address problems of privacy and fairness. +- While the authors might fear that complete honesty about limitations might be used by reviewers as grounds for rejection, a worse outcome might be that reviewers discover limitations that aren't acknowledged in the paper. The authors should use their best judgment and recognize that individual actions in favor of transparency play an important role in developing norms that preserve the integrity of the community. Reviewers will be specifically instructed to not penalize honesty concerning limitations. + +# 3. Theory assumptions and proofs + +Question: For each theoretical result, does the paper provide the full set of assumptions and a complete (and correct) proof? + +Answer: [NA] + +Justification: The paper does not include theoretical results. + +Guidelines: + +- The answer NA means that the paper does not include theoretical results. +- All the theorems, formulas, and proofs in the paper should be numbered and cross-referenced. +- All assumptions should be clearly stated or referenced in the statement of any theorems. +- The proofs can either appear in the main paper or the supplemental material, but if they appear in the supplemental material, the authors are encouraged to provide a short proof sketch to provide intuition. +- Inversely, any informal proof provided in the core of the paper should be complemented by formal proofs provided in appendix or supplemental material. +- Theorems and Lemmas that the proof relies upon should be properly referenced. + +# 4. Experimental result reproducibility + +Question: Does the paper fully disclose all the information needed to reproduce the main experimental results of the paper to the extent that it affects the main claims and/or conclusions of the paper (regardless of whether the code and data are provided or not)? + +Answer: [Yes] + +Justification: We provide all the details in Sec. 4 and Appendix Sec. A. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- If the paper includes experiments, a No answer to this question will not be perceived well by the reviewers: Making the paper reproducible is important, regardless of whether the code and data are provided or not. +- If the contribution is a dataset and/or model, the authors should describe the steps taken to make their results reproducible or verifiable. + +- Depending on the contribution, reproducibility can be accomplished in various ways. For example, if the contribution is a novel architecture, describing the architecture fully might suffice, or if the contribution is a specific model and empirical evaluation, it may be necessary to either make it possible for others to replicate the model with the same dataset, or provide access to the model. In general, releasing code and data is often one good way to accomplish this, but reproducibility can also be provided via detailed instructions for how to replicate the results, access to a hosted model (e.g., in the case of a large language model), releasing of a model checkpoint, or other means that are appropriate to the research performed. + +- While NeurIPS does not require releasing code, the conference does require all submissions to provide some reasonable avenue for reproducibility, which may depend on the nature of the contribution. For example + +(a) If the contribution is primarily a new algorithm, the paper should make it clear how to reproduce that algorithm. +(b) If the contribution is primarily a new model architecture, the paper should describe the architecture clearly and fully. +(c) If the contribution is a new model (e.g., a large language model), then there should either be a way to access this model for reproducing the results or a way to reproduce the model (e.g., with an open-source dataset or instructions for how to construct the dataset). +(d) We recognize that reproducibility may be tricky in some cases, in which case authors are welcome to describe the particular way they provide for reproducibility. In the case of closed-source models, it may be that access to the model is limited in some way (e.g., to registered users), but it should be possible for other researchers to have some path to reproducing or verifying the results. + +# 5. Open access to data and code + +Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material? + +Answer: [Yes] + +Justification: We provide the code in supplemental material. + +Guidelines: + +- The answer NA means that paper does not include experiments requiring code. +- Please see the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details. +- While we encourage the release of code and data, we understand that this might not be possible, so "No" is an acceptable answer. Papers cannot be rejected simply for not including code, unless this is central to the contribution (e.g., for a new open-source benchmark). +- The instructions should contain the exact command and environment needed to run to reproduce the results. See the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details. +- The authors should provide instructions on data access and preparation, including how to access the raw data, preprocessed data, intermediate data, and generated data, etc. +- The authors should provide scripts to reproduce all experimental results for the new proposed method and baselines. If only a subset of experiments are reproducible, they should state which ones are omitted from the script and why. +- At submission time, to preserve anonymity, the authors should release anonymized versions (if applicable). +- Providing as much information as possible in supplemental material (appended to the paper) is recommended, but including URLs to data and code is permitted. + +# 6. Experimental setting/details + +Question: Does the paper specify all the training and test details (e.g., data splits, hyperparameters, how they were chosen, type of optimizer, etc.) necessary to understand the results? + +# Answer: [Yes] + +Justification: We provide the details in Appendix Sec. A. + +# Guidelines: + +- The answer NA means that the paper does not include experiments. +- The experimental setting should be presented in the core of the paper to a level of detail that is necessary to appreciate the results and make sense of them. +- The full details can be provided either with the code, in appendix, or as supplemental material. + +# 7. Experiment statistical significance + +Question: Does the paper report error bars suitably and correctly defined or other appropriate information about the statistical significance of the experiments? + +# Answer: [Yes] + +Justification: We report experiment about statistical significance in Appendix Sec. B. + +# Guidelines: + +- The answer NA means that the paper does not include experiments. +- The authors should answer "Yes" if the results are accompanied by error bars, confidence intervals, or statistical significance tests, at least for the experiments that support the main claims of the paper. +- The factors of variability that the error bars are capturing should be clearly stated (for example, train/test split, initialization, random drawing of some parameter, or overall run with given experimental conditions). +- The method for calculating the error bars should be explained (closed form formula, call to a library function, bootstrap, etc.) +- The assumptions made should be given (e.g., Normally distributed errors). +- It should be clear whether the error bar is the standard deviation or the standard error of the mean. +- It is OK to report 1-sigma error bars, but one should state it. The authors should preferably report a 2-sigma error bar than state that they have a $96\%$ CI, if the hypothesis of Normality of errors is not verified. +- For asymmetric distributions, the authors should be careful not to show in tables or figures symmetric error bars that would yield results that are out of range (e.g. negative error rates). +- If error bars are reported in tables or plots, The authors should explain in the text how they were calculated and reference the corresponding figures or tables in the text. + +# 8. Experiments compute resources + +Question: For each experiment, does the paper provide sufficient information on the computer resources (type of compute workers, memory, time of execution) needed to reproduce the experiments? + +# Answer: [Yes] + +Justification: We provide the compute resources in Sec. 4, Sec A, and Appendix Sec G. + +# Guidelines: + +- The answer NA means that the paper does not include experiments. +- The paper should indicate the type of compute workers CPU or GPU, internal cluster, or cloud provider, including relevant memory and storage. +- The paper should provide the amount of compute required for each of the individual experimental runs as well as estimate the total compute. +- The paper should disclose whether the full research project required more compute than the experiments reported in the paper (e.g., preliminary or failed experiments that didn't make it into the paper). + +# 9. Code of ethics + +Question: Does the research conducted in the paper conform, in every respect, with the NeurIPS Code of Ethics https://neurips.cc/public/EthicsGuidelines? + +Answer: [Yes] + +Justification: We conform with the NeurIPS Code of Ethics. + +Guidelines: + +- The answer NA means that the authors have not reviewed the NeurIPS Code of Ethics. +- If the authors answer No, they should explain the special circumstances that require a deviation from the Code of Ethics. +- The authors should make sure to preserve anonymity (e.g., if there is a special consideration due to laws or regulations in their jurisdiction). + +# 10. Broader impacts + +Question: Does the paper discuss both potential positive societal impacts and negative societal impacts of the work performed? + +Answer: [Yes] + +Justification: We discuss the broader impacts in Appendix Sec. K. + +Guidelines: + +- The answer NA means that there is no societal impact of the work performed. +- If the authors answer NA or No, they should explain why their work has no societal impact or why the paper does not address societal impact. +- Examples of negative societal impacts include potential malicious or unintended uses (e.g., disinformation, generating fake profiles, surveillance), fairness considerations (e.g., deployment of technologies that could make decisions that unfairly impact specific groups), privacy considerations, and security considerations. +- The conference expects that many papers will be foundational research and not tied to particular applications, let alone deployments. However, if there is a direct path to any negative applications, the authors should point it out. For example, it is legitimate to point out that an improvement in the quality of generative models could be used to generate deepfakes for disinformation. On the other hand, it is not needed to point out that a generic algorithm for optimizing neural networks could enable people to train models that generate Deepfakes faster. +- The authors should consider possible harms that could arise when the technology is being used as intended and functioning correctly, harms that could arise when the technology is being used as intended but gives incorrect results, and harms following from (intentional or unintentional) misuse of the technology. +- If there are negative societal impacts, the authors could also discuss possible mitigation strategies (e.g., gated release of models, providing defenses in addition to attacks, mechanisms for monitoring misuse, mechanisms to monitor how a system learns from feedback over time, improving the efficiency and accessibility of ML). + +# 11. Safeguards + +Question: Does the paper describe safeguards that have been put in place for responsible release of data or models that have a high risk for misuse (e.g., pretrained language models, image generators, or scraped datasets)? + +Answer: [NA] + +Justification: The paper poses no such risks. + +Guidelines: + +- The answer NA means that the paper poses no such risks. +- Released models that have a high risk for misuse or dual-use should be released with necessary safeguards to allow for controlled use of the model, for example by requiring that users adhere to usage guidelines or restrictions to access the model or implementing safety filters. +- Datasets that have been scraped from the Internet could pose safety risks. The authors should describe how they avoided releasing unsafe images. + +- We recognize that providing effective safeguards is challenging, and many papers do not require this, but we encourage authors to take this into account and make a best faith effort. + +# 12. Licenses for existing assets + +Question: Are the creators or original owners of assets (e.g., code, data, models), used in the paper, properly credited and are the license and terms of use explicitly mentioned and properly respected? + +Answer: [Yes] + +Justification: The models used in the paper are properly credited. + +Guidelines: + +- The answer NA means that the paper does not use existing assets. +- The authors should cite the original paper that produced the code package or dataset. +- The authors should state which version of the asset is used and, if possible, include a URL. +- The name of the license (e.g., CC-BY 4.0) should be included for each asset. +- For scraped data from a particular source (e.g., website), the copyright and terms of service of that source should be provided. +- If assets are released, the license, copyright information, and terms of use in the package should be provided. For popular datasets, paperswithcode.com/datasets has curated licenses for some datasets. Their licensing guide can help determine the license of a dataset. +- For existing datasets that are re-packaged, both the original license and the license of the derived asset (if it has changed) should be provided. +- If this information is not available online, the authors are encouraged to reach out to the asset's creators. + +# 13. New assets + +Question: Are new assets introduced in the paper well documented and is the documentation provided alongside the assets? + +Answer: [NA] + +Justification: We did not release new assets. + +Guidelines: + +- The answer NA means that the paper does not release new assets. +- Researchers should communicate the details of the dataset/code/model as part of their submissions via structured templates. This includes details about training, license, limitations, etc. +- The paper should discuss whether and how consent was obtained from people whose asset is used. +- At submission time, remember to anonymize your assets (if applicable). You can either create an anonymized URL or include an anonymized zip file. + +# 14. Crowdsourcing and research with human subjects + +Question: For crowdsourcing experiments and research with human subjects, does the paper include the full text of instructions given to participants and screenshots, if applicable, as well as details about compensation (if any)? + +Answer: [NA] + +Justification: The paper does not involve crowdsourcing nor research with human subjects + +Guidelines: + +- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. +- Including this information in the supplemental material is fine, but if the main contribution of the paper involves human subjects, then as much detail as possible should be included in the main paper. + +- According to the NeurIPS Code of Ethics, workers involved in data collection, curation, or other labor should be paid at least the minimum wage in the country of the data collector. + +# 15. Institutional review board (IRB) approvals or equivalent for research with human subjects + +Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals (or an equivalent approval/review based on the requirements of your country or institution) were obtained? + +Answer: [NA] + +Justification: The paper does not involve crowdsourcing nor research with human subjects. + +Guidelines: + +- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. +- Depending on the country in which research is conducted, IRB approval (or equivalent) may be required for any human subjects research. If you obtained IRB approval, you should clearly state this in the paper. +- We recognize that the procedures for this may vary significantly between institutions and locations, and we expect authors to adhere to the NeurIPS Code of Ethics and the guidelines for their institution. +- For initial submissions, do not include any information that would break anonymity (if applicable), such as the institution conducting the review. + +# 16. Declaration of LLM usage + +Question: Does the paper describe the usage of LLMs if it is an important, original, or non-standard component of the core methods in this research? Note that if the LLM is used only for writing, editing, or formatting purposes and does not impact the core methodology, scientific rigorousness, or originality of the research, declaration is not required. + +Answer: [NA] + +Justification: The core method development in this research does not involve LLMs as any important, original, or non-standard components. + +Guidelines: + +- The answer NA means that the core method development in this research does not involve LLMs as any important, original, or non-standard components. +- Please refer to our LLM policy (https://neurips.cc/Conferences/2025/LLM) for what should or should not be described. \ No newline at end of file diff --git a/NeurIPS/2025/$_text{S}^2$Q-VDiT_ Accurate Quantized Video Diffusion Transformer with Salient Data and Sparse Token Distillation/images.zip b/NeurIPS/2025/$_text{S}^2$Q-VDiT_ Accurate Quantized Video Diffusion Transformer with Salient Data and Sparse Token Distillation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..15df9942b83ee9112081ec067af447356ffe257f --- /dev/null +++ b/NeurIPS/2025/$_text{S}^2$Q-VDiT_ Accurate Quantized Video Diffusion Transformer with Salient Data and Sparse Token Distillation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:444c2793a5735f6fc21fd5793c6bb60ec384dbfc690658c58940308a497bda9d +size 2255984 diff --git a/NeurIPS/2025/$_text{S}^2$Q-VDiT_ Accurate Quantized Video Diffusion Transformer with Salient Data and Sparse Token Distillation/layout.json b/NeurIPS/2025/$_text{S}^2$Q-VDiT_ Accurate Quantized Video Diffusion Transformer with Salient Data and Sparse Token Distillation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..d29c32a64330d5e3f31bf40f69d7bc48f343ae1e --- /dev/null +++ b/NeurIPS/2025/$_text{S}^2$Q-VDiT_ Accurate Quantized Video Diffusion Transformer with Salient Data and Sparse Token Distillation/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d88de2a57dab4b60a44c0d905d0d9d86bb3d7e29927749933ecf014281d95420 +size 880966 diff --git a/NeurIPS/2025/$i$MIND_ Insightful Multi-subject Invariant Neural Decoding/73318b0a-5234-400e-a036-fa631a7129c3_content_list.json b/NeurIPS/2025/$i$MIND_ Insightful Multi-subject Invariant Neural Decoding/73318b0a-5234-400e-a036-fa631a7129c3_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..11a532f59b69f534a1686ffcc95ad0b9e685a253 --- /dev/null +++ b/NeurIPS/2025/$i$MIND_ Insightful Multi-subject Invariant Neural Decoding/73318b0a-5234-400e-a036-fa631a7129c3_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:50b7cbbe191868ebcece15ec970e21c1eebef7ded209782ba1f3ffe35692147d +size 197007 diff --git a/NeurIPS/2025/$i$MIND_ Insightful Multi-subject Invariant Neural Decoding/73318b0a-5234-400e-a036-fa631a7129c3_model.json b/NeurIPS/2025/$i$MIND_ Insightful Multi-subject Invariant Neural Decoding/73318b0a-5234-400e-a036-fa631a7129c3_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7cd8ec2f48f239c333e56aa30e91e5cbcae910c2 --- /dev/null +++ b/NeurIPS/2025/$i$MIND_ Insightful Multi-subject Invariant Neural Decoding/73318b0a-5234-400e-a036-fa631a7129c3_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:934492b54ef211b92eee4c5722bb5b2ab7487e28f08f906fc8f004f8d3a393b3 +size 245674 diff --git a/NeurIPS/2025/$i$MIND_ Insightful Multi-subject Invariant Neural Decoding/73318b0a-5234-400e-a036-fa631a7129c3_origin.pdf b/NeurIPS/2025/$i$MIND_ Insightful Multi-subject Invariant Neural Decoding/73318b0a-5234-400e-a036-fa631a7129c3_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..ebd6f801c0764c30c6e6218ed7d0ef02eacd999b --- /dev/null +++ b/NeurIPS/2025/$i$MIND_ Insightful Multi-subject Invariant Neural Decoding/73318b0a-5234-400e-a036-fa631a7129c3_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c473b0514ffe71f0db31d823258716eb737b7bde1947104c736231bd4c125889 +size 42026786 diff --git a/NeurIPS/2025/$i$MIND_ Insightful Multi-subject Invariant Neural Decoding/full.md b/NeurIPS/2025/$i$MIND_ Insightful Multi-subject Invariant Neural Decoding/full.md new file mode 100644 index 0000000000000000000000000000000000000000..51be1e8957d2b96c108f8c84316f533fdff8d0f7 --- /dev/null +++ b/NeurIPS/2025/$i$MIND_ Insightful Multi-subject Invariant Neural Decoding/full.md @@ -0,0 +1,929 @@ +# iMIND: Insightful Multi-subject Invariant Neural Decoding + +Zixiang Yin, Jiarui Li, Zhengming Ding* +Department of Computer Science, Tulane University {zyin, jli78, zding1}@tulane.edu +https://zachyin.com/imind + +# Abstract + +Decoding visual signals holds an appealing potential to unravel the complexities of cognition and perception. While recent reconstruction tasks leverage powerful generative models to produce high-fidelity images from neural recordings, they often pay limited attention to the underlying neural representations and rely heavily on pretrained priors. As a result, they provide little insight into how individual voxels encode and differentiate semantic content or how these representations vary across subjects. To mitigate this gap, we present an insightful Multi-subject Invariant Neural Decoding (iMIND) model, which employs a novel dual-decoding framework—both biometric and semantic decoding—to offer neural interpretability in a data-driven manner and deepen our understanding of brain-based visual functionalities. Our iMIND model operates through three core steps: establishing a shared neural representation space across subjects using a ViT-based masked autoencoder, disentangling neural features into complementary subject-specific and object-specific components, and performing dual decoding to support both biometric and semantic classification tasks. Experimental results demonstrate that iMIND achieves state-of-the-art decoding performance with minimal scalability limitations. Furthermore, iMIND empirically generates voxel-object activation fingerprints that reveal object-specific neural patterns and enable investigation of subject-specific variations in attention to identical stimuli. These findings provide a foundation for more interpretable and generalizable subject-invariant neural decoding, advancing our understanding of the voxel semantic selectivity as well as the neural vision processing dynamics. + +# 1 Introduction + +Deep learning models have recently been adopted in neuroscience as powerful tools for modeling brain activity, especially in the study of vision and cognition [6, 9, 19, 26]. A central goal in cognitive neuroscience is to understand how the brain transforms sensory input into meaningful representations that support recognition, memory, decision-making, and attention. Unlike behavioral annotations, neural signals provide a direct readout of these internal processes, revealing perceptual [3], emotional [32], and attentional dynamics [27] that cannot be fully captured by explicit labeling. Among available neural modalities, functional Magnetic Resonance Imaging (fMRI) [29] has been especially influential for its ability to non-invasively measure distributed patterns of cortical activity, enabling the study of how complex visual concepts are represented across brain regions. As such, fMRI not only offers a promising supervisory signal for aligning neural activity with computational models, but also serves as a critical tool for probing the neural basis of cognition and vision [16]. + +![](images/a167eba919ae9a8ded1240852eb9c9852b41323af5664523f7aa04fc9cb9a2f0.jpg) +Figure 1: Overview of our framework. iMIND involves biometric decoding (identifying individuals) and semantic decoding (classifying perceived objects). fMRI voxel signals are first flattened and encoded by a pretrained Masked Autoencoder (MAE) to generate latent neural tokens, which are then passed through a learnable subject-object disentanglement block via orthogonal basis transformation $\mathbf{B} = (\mathbf{B}_{\mathrm{subj}},\mathbf{B}_{\mathrm{obj}})$ . The resulting biometric neural tokens are directly pooled for subject classification, while the semantic neural tokens act as keys and values in a cross-attention module with the CLIP latent features of the corresponding visual stimuli as queries to reflect the semantic contents captured by each subject. The fused representation is then pooled for multi-label object classification. + +Most recent research has focused on reconstructing visual stimuli from brain activity by projecting neural representations into deep visual spaces and employing generative models such as GANs or diffusion models [5, 22, 35, 38, 41]. These approaches have yielded visually plausible, high-resolution images, suggesting that deep learning can serve as a bridge between brain activity and imagery. However, despite these remarkable achievements, we argue that reconstruction alone is fundamentally inadequate for understanding brain vision mechanism. Specifically, the following limitations exist: + +- Reconstruction relies heavily on pretrained generative priors, which often dominate the decoding process and may introduce model-specific biases that obscure genuine neural content. +- Brain recording may not encode all fine-grained details necessary for accurate reconstruction—especially across subjects—rendering pixel-level reconstruction ill-posed and often misleading. This ultimately shifts the focus from reconstruction to generation. + +Thus, we argue that reconstructing visual stimuli is neither a reliable nor interpretable strategy for decoding neural representations. A more effective alternative is to classify the subject's perceptual experience directly from fMRI data, enabling the identification of visual concepts embedded in neural responses [44]. This discriminative classification-based approach supports evaluation through standard metrics and allows researchers to disentangle both shared and individual-specific semantic components of brain activity—capabilities that reconstruction methods often fail to provide. + +Technically, current neural decoding solutions follow two main strategies: single-subject models [5, 22, 24, 25, 35, 38], which suffer from data scarcity, overfitting, and poor scalability due to the high cost of fMRI collection; and multi-subject models [4, 36, 41, 44] $^2$ , which face substantial inter-subject variability from anatomical and functional differences. This can lead to entangled neural representations, where subject-specific and object-related signals become mixed [10], degrading decoding performance and interpretability. This raises a key research question: + +"How can we develop a discriminative neural decoding framework that generalizes across individuals while preserving subject-specific semantic interpretations of visual stimuli?" + +To answer this, we present a novel approach to decoding brain activity using deep classification models, aiming to capture shared neural representations across individuals while preserving personalized interpretations shaped by diverse experiences and backgrounds. Toward this goal, we introduce the insightful Multi-subject Invariant Neural Decoding (iMIND) model-a dual-decoding + +framework (Figure 1) that supports both biometric decoding (identifying individuals) and semantic decoding (classifying perceived objects). By constructing voxel-object activation maps and examining subject-specific attentional patterns in response to complex, multi-object scenes, iMIND offers a principled path to disentangling individual and shared components of visual perception, ultimately advancing our understanding of brain-vision mechanisms. To sum up, our contributions are threefold: + +- We introduce a multi-subject dual-decoding framework that disentangles neural signals into subject-specific and object-specific components, enabling scalable and subject-aware brain analysis with state-of-the-art semantic and biometric decoding performance. +- We develop a visual-neural interaction module that identifies object-voxel activation patterns in a data-driven manner, revealing how different subjects encode object-level semantics in the brain. +- We perform a comprehensive analysis of attention-based variability across subjects viewing the same visual stimuli, providing insights into individualized neural responses under time-constrained conditions. + +# 2 Proposed Method + +# 2.1 Problem Formulation + +Consider a neural dataset containing brain activities of subjects $S$ in response to visual stimuli drawn from an image dataset with $M$ samples $\mathcal{D} = \{(\mathbf{X}_i,\mathbf{y}_i)\}_{i = 1}^M$ , where each visual stimulus $\mathbf{X}_i$ is a three-channel image with a resolution of $H\times W$ , paired with a non-exclusive ground-truth label $\mathbf{y}_i\in \mathbb{R}^C$ representing $C$ object categories. During neural recordings, all images $\mathbf{X}$ are viewed by a group of subjects $S$ . For each subject $s\in S$ viewing the $i$ -th image $\mathbf{X}_i$ , an fMRI voxel response $\mathbf{V}_{i,s}$ is recorded, capturing neural activity specific to the subject-image pair. Note that the voxel signal $\mathbf{V}_{i,s}$ is flattened as an 1D vector with various lengths ( $12,682\sim 17,907$ ) across subjects due to biometric differences in neural structure. Following [5], we apply wrap-around padding to achieve a uniform voxel length $L$ . Details on pre-processing procedures are provided in Appendix A.3.2. + +# 2.2 Framework Overview + +Our goal is to decouple arbitrary fMRI voxel signals into subject-invariant and subject-specific components from both semantic and biometric perspectives by developing a visual-neural model $\mathcal{M}$ . Formally, this can be expressed as a mapping: + +$$ +\mathcal {M}: \underbrace {\mathbb {R} ^ {H \times W \times 3}} _ {\text {i m a g e}} \times \underbrace {\mathbb {R} ^ {L}} _ {\text {v o x e l s}} \rightarrow \underbrace {\mathbb {R} ^ {C}} _ {\text {s e m a n t i c s}} \times \underbrace {\mathbb {R} ^ {| S |}} _ {\text {b i o m e t r i c s}}, \tag {1} +$$ + +where semantic decoding seeks to extract object-related neural representations that are consistent across subjects, whereas biometric decoding focuses on capturing subject-specific neural patterns that are independent of the visual stimuli. + +To address both tasks, we propose the insightful Multi-subject Invariant Neural Decoding model, abbreviated as iMIND. Building on the SC-MBM framework [5], our method constructs a $d$ -dimensional shared latent neural space $\mathcal{F}$ across subjects to reduce the noise and redundancy inherent in fMRI signals [40]. Specifically, we employ a Vision Transformer (ViT)-based encoder $\mathcal{E}:\mathbb{R}^L\to \mathbb{R}^{N\times d}$ that projects each input voxel signal $\mathbf{V}$ into a set of $N$ neural feature tokens, denoted as $\mathbf{F} = \{\mathbf{f}_j\}_{j = 1}^N$ , where each $\mathbf{f}_j$ is a $d$ -dimensional neural feature token. This yields $N$ neural representations per fMRI input, all embedded in the shared latent space $\mathcal{F}$ . The encoder $\mathcal{E}(\cdot)$ is pretrained using a masked autoencoding objective in a self-supervised manner, emphasizing voxel-wise reconstruction. + +Subsequently, we disentangle the learned features into object-specific and subject-specific components. The subject-specific features are used for subject classification (biometric decoding), while the object-specific features are aligned with frozen CLIP visual embeddings of the corresponding images via a cross-attention mechanism, enabling multi-label object classification (semantic decoding). + +# 2.3 Subject-Object Disentanglement + +The self-supervised pretraining reduces noise and redundancy in fMRI signals by encouraging the model to capture generalizable patterns. However, the resulting neural features $\mathcal{F}$ are not explicitly + +tailored for downstream decoding tasks such as biometric identification or semantic classification. Because the reconstruction objective treats all latent information as equally relevant, it often leads to entangled representations-mixing subject-specific and object-specific components. This entanglement limits the interpretability of the learned features and hinders task-specific performance, especially when distinguishing between individual variability and shared visual semantics is essential. + +Mathematically, this can be formalized by interpreting each neural feature token $\mathbf{f} = (f_1, f_2, \ldots, f_d) \in \mathbb{R}^d$ as the coordinate representation of a corresponding neural point $\mathbf{p}$ in the latent neural feature space $\mathcal{F}$ . By default, this representation is expressed with respect to the standard basis $\mathbf{E} = \{\mathbf{e}_k \in \mathbb{R}^d\}_{k=1}^d$ of $\mathbb{R}^d$ , denoted as $[\mathbf{p}]_{\mathbf{E}}$ : + +$$ +[ \mathbf {p} ] _ {\mathbf {E}} := \mathbf {f} = \sum_ {k = 1} ^ {d} f _ {k} \mathbf {e} _ {k} \in \mathbb {R} ^ {d}. \tag {2} +$$ + +From this perspective, the entanglement of subject-specific and object-specific information within $\mathbf{f}$ can be attributed to the default use of $\mathbf{E}$ as the basis for spanning the neural feature space $\mathcal{F}$ . As $\mathbf{E}$ is implicitly determined by the self-supervised task in the first training stage, this choice is beyond direct control, resulting in an inevitable blending of both subject and object information within the feature representation $\mathbf{f}$ for any neural point $\mathbf{p} \in \mathcal{F}$ . + +To enable effective downstream biometric and semantic decoding, we propose a solution from the perspective of feature disentanglement [8, 14, 39, 43]. We begin with assuming that the subject-specific and object-specific information within a neural point $\mathbf{p}$ are linearly entangled in the current neural feature representation $\mathbf{f}$ under the basis $\mathbf{E}$ . While this may not fully characterize the complexities of neural dynamics, it serves as a simplified approximation to provide a meaningful step toward understanding the interplay between subject-specific and object-specific information. More crucially, this assumption applies at the latent feature level, not at the original fMRI signal level. At this level, the assumption is reasonable, as it aligns with the principles of deep classification tasks, where linear MLP classifiers rely on deep neural networks to transform inputs into linearly separable features for accurate classification. Based on this assumption, we resolve the linear entanglement by re-representing $\mathbf{p}$ with respect to a new basis $\mathbf{B}$ . Specifically, we seek a learnable basis $\mathbf{B} = (\mathbf{B}_{\mathrm{subj}}, \mathbf{B}_{\mathrm{obj}})$ of $\mathbb{R}^d$ space that perfectly separates the subject-specific and object-specific features. Mathematically, this re-representation would allow $\mathbf{z}$ , the representation of the same neural point $\mathbf{p}$ with respect to the new basis $\mathbf{B}$ , to be distinctly split into subject-specific $\mathbf{z}_{\mathrm{subj}}$ and object-specific $\mathbf{z}_{\mathrm{obj}}$ components: + +$$ +\mathbf {z} = [ \mathbf {p} ] _ {\mathbf {B}} = \left([ \mathbf {p} ] _ {\mathbf {B} _ {\text {s u b j}}}, [ \mathbf {p} ] _ {\mathbf {B} _ {\text {o b j}}}\right) = \left(\mathbf {z} _ {\text {s u b j}}, \mathbf {z} _ {\text {o b j}}\right) \in \mathbb {R} ^ {d}. \tag {3} +$$ + +According to the mathematical property of bases, the separation of $\mathbf{z}_{\mathrm{subj}}$ and $\mathbf{z}_{\mathrm{obj}}$ is guaranteed as long as $\mathbf{B}$ forms an orthonormal basis of $\mathbb{R}^d$ , i.e., $\mathbf{BB}^\top = \mathbf{I}_d$ , where $\mathbf{I}_d$ is the identity matrix of rank $d$ . With this orthonormality condition satisfied, the transformation from the original representation of $\mathbf{f}$ to the new representation $\mathbf{z}$ of the same neural point $\mathbf{p}$ can be derived as follows: + +$$ +\mathbf {z} = [ \mathbf {p} ] _ {\mathbf {B}} = \mathbf {B} ^ {- 1} [ \mathbf {p} ] _ {\mathbf {E}} = \mathbf {B} ^ {\top} [ \mathbf {p} ] _ {\mathbf {E}} = \mathbf {B} ^ {\top} \mathbf {f} \in \mathbb {R} ^ {d}. \tag {4} +$$ + +Combined with Eq. (3), we finally arrive at our subject-object disentanglement formulation: + +$$ +\mathbf {z} _ {\text {s u b j}} = \mathbf {B} _ {\text {s u b j}} ^ {\top} \mathbf {f} \in \mathbb {R} ^ {d _ {\text {s u b j}}} \quad \text {a n d} \quad \mathbf {z} _ {\text {o b j}} = \mathbf {B} _ {\text {o b j}} ^ {\top} \mathbf {f} \in \mathbb {R} ^ {d _ {\text {o b j}}}. \tag {5} +$$ + +From a feature transformation perspective, we realize subject-object disentanglement by decomposing the original neural space $\mathcal{F}$ into two complementary (orthogonal) subspaces: the subject-specific neural subspace $\mathcal{F}_{\mathrm{subj}}$ and the object-specific neural subspace $\mathcal{F}_{\mathrm{obj}}$ . This disentanglement is achieved by learning two orthonormal sets, $\mathbf{B}_{\mathrm{subj}} \in \mathbb{R}^{d \times d_{\mathrm{subj}}}$ and $\mathbf{B}_{\mathrm{obj}} \in \mathbb{R}^{d \times d_{\mathrm{obj}}}$ , which span $\mathcal{F}_{\mathrm{subj}}$ and $\mathcal{F}_{\mathrm{obj}}$ respectively. In our framework, $d_{\mathrm{obj}}$ is treated as a user-defined value, with $d_{\mathrm{subj}} \coloneqq d - d_{\mathrm{obj}}$ to complete the basis. A formal and theoretic proof is provided in Appendix A.2. + +The complementary relationship, $\mathcal{F} = \mathcal{F}_{\mathrm{subj}} \oplus \mathcal{F}_{\mathrm{obj}}$ , guarantees a clear separation of subject and object information in the transformed neural representation $\mathbf{z} = (\mathbf{z}_{\mathrm{subj}}, \mathbf{z}_{\mathrm{obj}}) \in \mathbb{R}^d$ for each neural point $\mathbf{p} \in \mathcal{F}$ , establishing a foundation for the subsequent biometric and semantic decoding tasks. + +# 2.4 Biometric & Semantic Decoding + +In this section, we describe our approach to decoding fMRI signals biometrically and semantically. Note that an fMRI signal $\mathbf{V}$ is represented by $\mathbf{F}$ as a set of $N$ neural tokens $\{\mathbf{f}_j\}_{j=1}^N$ within the + +latent neural space $\mathcal{F}$ . Based on Eq. (5), the subject-specific feature map $\mathbf{Z}_{\mathrm{subj}} \in \mathbb{R}^{N \times d_{\mathrm{subj}}}$ and object-specific feature map $\mathbf{Z}_{\mathrm{obj}} \in \mathbb{R}^{N \times d_{\mathrm{obj}}}$ are generated from $\mathbf{F} \in \mathbb{R}^{N \times d}$ as follows: + +$$ +\mathbf {Z} _ {\text {s u b j}} = \mathbf {F B} _ {\text {s u b j}} \quad \text {a n d} \quad \mathbf {Z} _ {\text {o b j}} = \mathbf {F B} _ {\text {o b j}}. \tag {6} +$$ + +# 2.4.1 Biometric Decoding + +The biometric neural decoding is driven by a supervised multi-subject classification task. We apply a Global Average Pooling (GAP) operator, $\mathcal{G}_{\mathrm{subj}}: \mathbb{R}^{N \times d_{\mathrm{subj}}} \to \mathbb{R}^{d_{\mathrm{subj}}}$ , to the subject-specific neural feature map $\mathbf{Z}_{\mathrm{subj}}$ to build a subject class token: + +$$ +\mathbf {z} _ {\text {s u b j}} ^ {\text {c l s}} = \mathcal {G} _ {\text {s u b j}} \left(\mathbf {Z} _ {\text {s u b j}}\right). \tag {7} +$$ + +Finally, a linear multi-subject classifier $\mathcal{C}_{\mathrm{subj}}: \mathbb{R}^{d_{\mathrm{subj}}} \to \mathbb{R}^S$ is applied for the final biometric prediction: + +$$ +\hat {\mathbf {y}} _ {\text {s u b j}} = \mathcal {C} _ {\text {s u b j}} \left(\mathbf {z} _ {\text {s u b j}} ^ {\text {c l s}}\right). \tag {8} +$$ + +# 2.4.2 Semantic Decoding + +To establish a feature-level connection between the fMRI voxel signals $\mathbf{V}$ and the semantic content of visual stimulus $\mathbf{X}$ , we utilize the vision feature map $\mathbf{F}_{\mathbf{x}} \in \mathbb{R}^{N_{\mathbf{x}} \times d_{\mathbf{x}}}$ extracted from the last layer of a frozen CLIP visual encoder [33]. In contrast to most neural vision approaches that project fMRI features into the CLIP visual space [5, 38], our method takes the opposite strategy by treating the CLIP vision features as queries to extract corresponding neural object features from $\mathbf{Z}_{\mathrm{obj}}$ . + +This design is motivated by the complementary roles of CLIP and fMRI signals. CLIP features encode subject-invariant and stimulus-driven semantics with rich spatial and conceptual structure, effectively serving as a pseudo-ground-truth reference for object existence. In contrast, fMRI captures subject-specific neural responses that reveal how different individuals attend to these semantic components. To fuse these two modalities, we applied a multi-head cross-attention extractor $\mathcal{A}:\mathbb{R}^{N_{\mathbf{x}}\times d_{\mathbf{x}}}\times \mathbb{R}^{N\times d_{\mathrm{obj}}}\times \mathbb{R}^{N\times d_{\mathrm{obj}}}\to \mathbb{R}^{N_{\mathbf{x}}\times d_{\mathrm{obj}}}$ , defined as follows: + +$$ +\mathbf {Z} _ {\mathrm {o b j}} ^ {\mathbf {F} _ {\mathbf {x}}} = \mathcal {A} \left(\text {Q u e r y} = \mathbf {F} _ {\mathbf {x}}, \text {K e y} = \mathbf {Z} _ {\mathrm {o b j}}, \text {V a l u e} = \mathbf {Z} _ {\mathrm {o b j}}\right). \tag {9} +$$ + +Here, the CLIP embeddings act as queries, while the fMRI-derived $\mathbf{Z}_{\mathrm{obj}}$ serves as keys and values. This configuration ensures that the fused representation remains fundamentally neural in nature: fMRI signals determine what semantic components are prioritized, while CLIP provides the structured semantic reference frame. The resulting cross-attention has two key effects: + +- Semantic Prioritization - CLIP embeddings query the fMRI features, and attention weights highlight which parts of the CLIP semantic space align with neural activations. This allows the model to anchor predictions in semantically grounded content; +- Subject-Specific Modulation - The fMRI responses modulate CLIP-driven semantics in a subject-dependent manner, enabling the model to capture how different individuals selectively emphasize different attributes of visual stimuli that share similar semantic contents. + +In this way, CLIP will not dominate or overwrite the neural signal, but rather provides a semantically structured scaffold. The fMRI features dynamically shape which aspects of that structure are emphasized, yielding a bi-directional synergy. Our semantic decoding thus remains primarily rooted in the fMRI modality, with CLIP assisting in refining object-specific neural features for final multi-label object classification. Similar to what we have done in the biometric decoding in the previous section, a global feature operator $\mathcal{G}_{\mathrm{obj}}:\mathbb{R}^{N_{\mathbf{x}}\times d_{\mathrm{obj}}}\to \mathbb{R}^{d_{\mathrm{obj}}}$ transforms $\mathbf{Z}_{\mathrm{obj}}^{\mathbf{F}_{\mathbf{x}}}$ into an object class token: + +$$ +\mathbf {z} _ {\mathrm {o b j}} ^ {\mathrm {c l s}} = \mathcal {G} _ {\mathrm {o b j}} \left(\mathbf {Z} _ {\mathrm {o b j}} ^ {\mathbf {F} _ {\mathbf {X}}}\right). \tag {10} +$$ + +Following that a multi-label object classifier $\mathcal{C}_{\mathrm{obj}}: \mathbb{R}^{d_{\mathrm{obj}}} \to \mathbb{R}^C$ is applied for final semantic prediction: + +$$ +\hat {\mathbf {y}} _ {\mathrm {o b j}} = \mathcal {C} _ {\mathrm {o b j}} \left(\mathbf {z} _ {\mathrm {o b j}} ^ {\mathrm {c l s}}\right). \tag {11} +$$ + +# 2.5 Model Training + +The training process of our model consists of two stages. In the first stage, we follow the approach of SC-MBM [5] to pre-train a ViT-based masked autoencoder for fMRI data, constructing a latent neural space via minimizing reconstruction error with a Mean-Square Error (MSE) loss. + +In the second stage, we retain only the fMRI encoder $\mathcal{E}(\cdot)$ from the first stage and optimize it with all other parameters in the proposed architecture. Three loss functions guide this stage. First, for biometric decoding, we introduce a subject classification loss $\mathcal{L}_{\mathrm{subj}}$ , which computes the cross-entropy $\mathcal{H}_{\mathrm{CE}}$ with softmax activation against the one-hot label $\mathbf{y}_{\mathrm{subj}}$ from the ground-truth subject index $\mathbf{y}_{\mathrm{subj}}$ + +$$ +\mathcal {L} _ {\mathrm {s u b j}} := \mathcal {H} _ {\mathrm {C E}} \left(\operatorname {s o f t m a x} \left(\hat {\mathbf {y}} _ {\mathrm {s u b j}}\right), \mathbf {y} _ {\mathrm {s u b j}}\right). \tag {12} +$$ + +For multi-label semantic decoding, we employ an object loss $\mathcal{L}_{\mathrm{obj}}$ which is a binary cross-entropy function $\mathcal{H}_{\mathrm{BCE}}$ with sigmoid activation $\sigma (\cdot)$ : + +$$ +\mathcal {L} _ {\mathrm {o b j}} := \mathcal {H} _ {\mathrm {B C E}} \left(\sigma \left(\hat {\mathbf {y}} _ {\mathrm {o b j}}\right), \mathbf {y} _ {\mathrm {o b j}}\right). \tag {13} +$$ + +Finally, we impose an orthonormal constraint on the learnable basis concatenation $\mathbf{B} = (\mathbf{B}_{\mathrm{subj}},\mathbf{B}_{\mathrm{obj}})$ , as defined in Eq. (6). This orthonormal loss $\mathcal{L}_{\mathrm{orth}}$ ensures the perfect separation of subject-specific and object-specific features by minimizing $\mathcal{L}_{\mathrm{orth}} := \| \mathbf{BB}^{\top} - \mathbf{I}_d\|_{\mathbf{F}}^2$ , where $\| \cdot \|_{\mathbf{F}}$ indicates the Frobenius matrix norm, and $\mathbf{I}_d \in \mathbb{R}^{d\times d}$ is the identity matrix of rank $d$ . + +In summary, in the second training stage, the total objective $\mathcal{L}$ is formulated as: + +$$ +\mathcal {L} = \mathcal {L} _ {\text {s u b j}} + \mathcal {L} _ {\text {o b j}} + \lambda \mathcal {L} _ {\text {o r t h}}, \tag {14} +$$ + +where $\lambda$ serves as a trade-off hyperparameter to balance the orthonormal constraint against the subject and object classification losses, which are considered equally important and share the same scale. + +# 3 Experiments + +# 3.1 Experimental Setup + +Dataset. We evaluate our iMIND framework using the Natural Scenes Dataset (NSD) [2], a comprehensive, publicly available fMRI dataset capturing brain responses from 8 human subjects viewing natural scenes from MS-COCO [23]. Each subject passively viewed a set of 10,000 images for 3s, each repeated three times; 1,000 of these images were shared across all subjects, while the remaining 9,000 were unique to each individual, with no overlap between subjects. Due to incomplete sessions and data availability restrictions, not all trials are accessible for every subject, resulting in a total of 213,000 trials across all participants before pre-processing. In line with previous NSD studies [17, 35, 36, 38], we used standardized train/test splits and averaged fMRI activations over repetitions for each image within each subject. This pre-processing yielded 69,566 training samples and 7,674 test samples, allowing us to train a single multi-subject model across all 8 subjects. Additional details on NSD data, fMRI pre-processing, and iMIND implementation can be found in Appendix A.3.3. + +# 3.2 Neural Decoding Performance + +Semantic Decoding. For the semantic decoding task, we evaluate and compare with other models using three standard metrics for multi-label classification: mean Average Precision (mAP), the area under the receiver operating characteristic curve (AUC), and Hamming distance (Hamming). Table 1 categorizes models based on their ability to process multi-subject fMRI signals simultaneously or on a per-subject basis, as well as the modalities used for object classification. While our iMIND model is designed to process both image and fMRI modalities, it can also be adapted as a single-modality model by simply removing the cross-attention mechanism in Eq. (9) and using $\mathbf{Z}_{\mathrm{obj}}$ in Eq. (6) directly for semantic decoding. Experimental results indicate that iMIND achieves superior performance across all three metrics, establishing a new state-of-the-art for semantic decoding in both single-modality and multi-modality settings. + +Table 1: Semantic decoding performance on the NSD dataset. + +
Model TypeMethodsModalitiesmAP ↑AUC ↑Hamming ↓
Single-subjectMLP [44]*fMRI.258.854.033
ViT [44]*fMRI.238.815.032
Multi-subjectMLP [44]*fMRI.150.767.039
ViT [44]*fMRI.156.755.038
EMB [4]fMRI.220.825.035
CLIP-MUSED [44]fMRI+Image+Text.258.877.030
iMIND (Ours)fMRI.309.913.027
fMRI+Image.784.984.012
+ +* directly sourced from [44] as benchmarks due to the limited research on semantic neural decoding + +![](images/4f05f61ea93e7c1d5a88a54b657150b46d5d83d18e13348743bdb67b0ea6b64c.jpg) +Figure 2: Average activations of tokens in $\mathbf{Z}_{\mathrm{obj}}$ for the object Chair. + +Biometric Decoding. To the best of our knowledge, no existing neural decoding models support subject classification using the fMRI modality. To provide a comprehensive and fair evaluation, we established baseline models following the exact preprocessing steps from our proposed method and conducted both supervised and unsupervised biometric decoding, as presented in Table 2. Top-1 accuracy (ACC) and Matthews Correlation Coefficient (MCC) [28] are used as metrics. The poor performance of naive subject classification methods underscores the com + +plexity of neural data across subjects, whereas the near-perfect classification achieved by iMIND highlights the effectiveness and necessity of our subject-object disentanglement approach. This demonstrates that within our framework, biometric fMRI features are highly discriminative across subjects and our linearity assumption is reasonable. By facilitating the extraction of task-relevant features, our disentanglement method further enhances downstream semantic and biometric decoding. Details on how we build biometric decoding baselines are provided in Appendix A.4. + +Table 2: Biometric decoding performance. + +
MethodSettingACCMCC
K-MeansEuclidean.181.068
Cosine.232.126
MLPPlain.283.181
L2 norm.377.290
L2 + ReLU.573.526
iMIND-.999.999
+ +# 3.3 Subject-Invariant Decoding + +The primary motivation for the subject-object disentanglement design, introduced in Section 2.3, is to decompose the entangled neural feature $\mathbf{F}$ into subject-specific and object-oriented components $\mathbf{Z}_{\mathrm{subj}}$ and $\mathbf{Z}_{\mathrm{obj}}$ for a better biometric and semantic decoding. This approach expects object-wise token contributions in $\mathbf{Z}_{\mathrm{obj}}$ to remain consistent across subjects. Using the object chair as an example, we visualize 280 tokens' activations averaged across all correct predictions by subject in Figure 2. It turns out that at the feature level, our method successfully achieves subject-invariant decoding, as token activations display high similarity with only negligible subject-level variations. This outcome demonstrates our model's effectiveness in extracting object-specific information from complex fMRI data, offering a robust framework for multi-subject fMRI decoding. + +# 3.4 Visual-Neural Relationship + +We empirically investigate the relationship between brain activities and semantic objects in visual stimuli, leveraging both GradCAM [37] and Attention Roll-out [1]. + +Subject-wise 1D Activation Pattern. Taking subj01 as an example, we calculate voxel-wise activations within low-level visual regions of interest (ROIs) (V1-V4) and a wider high-level visual ROI in response to three objects: person, horse, and chair. This calculation takes the median activation across all true positive samples predicted by our model on the test set. As shown in Figure 3, the + +![](images/775f3798a2f8a5e4d94dc6aa59ab85b703d0fca0aced561e51e92c42f1517834.jpg) +Figure 3: 1D Object-Voxel activations by brain vision ROIs for subj01. + +![](images/36d70697ed5c8a64e7a77eb0d23b5d691d0e3a5e5cda08a820b3d0afcb2d2f50.jpg) +Figure 4: 3D Object-Voxel activations of person, car, bird, and horse for subj01, subj03, and subj05. + +y-axis represents median activation, while the x-axis represents voxel IDs, providing a clear overview of the active voxels across both low-level and high-level visual ROIs when subj01 recognizes each object. The unique activation patterns, with specific voxels responding to each object, indicate that brain voxels function differently from image pixels. In images, object locations are believed spatially random, so evaluating pixel-wise activation does not yield consistent spatial activation patterns. In contrast, fMRI voxels exhibit specialized roles in processing visual information, suggesting that brain voxels are organized by functional responsibility with a degree of spatial invariance—especially in high-level visual ROIs. This conclusion aligns with the existing studies from neuroscience [20, 21]. + +Cross-subject 3D Activation Pattern. We present 3D brain activation patterns in Figure 4 for four objects-person, car, bird, horse-visualized across three subjects: subj01, subj03, and subj05. Light yellow regions denote non-visual areas that were excluded from the dataset, resulting in uniformly absent signals in these regions. Based on the visualization, the following observations are made: + +- Consistency across subjects: the objects bird and horse show a broad similarity across subjects, particularly in the region of the higher visual cortex and predominantly in the left hemisphere. This consistency suggests that certain high-level features associated with animals may be processed in similar ways across individuals, reflecting stable visual processing pathways in the brain. +- Object sensitivity: the activation intensity for object person appears stronger and more concentrated, indicating that the brain may allocate increased neural resources or "attention" to socially relevant stimuli (people), compared to less socially significant objects like bird. This result is supported by neuroscience research [15, 42]. +- Representational flexibility: while general patterns are shared across subjects, the intensity and spatial distribution of activation vary slightly for certain objects, such as car. These variations may reflect individual differences in brain anatomy or prior experiences that influence object representation and visual information processing. This flexibility of the brain's adaptability to personal needs and experiences is known as neural plasticity [7, 11, 30]. + +# 3.5 Variations in Subject Attention + +A key contribution of iMIND is its use of subject-invariant CLIP visual features to explore how different subjects focus on distinct objects when receiving the same stimulus. Figure 5 illustrates this attention variation: the first column displays the original visual stimulus with six ground-truth + +![](images/37e494c117b1853e81f9e1739e253ca287fbd0b86e96a32cb18b836052cbfa5b.jpg) + +![](images/18ad1dd59cf577e42f92333bd9f779aa7d0c8f0a1a942f0bbc0fdcefb77e2b49.jpg) +Figure 5: Variations in subjects' attention to different objects. The leftmost image shows the visual stimulus, labeled with six objects: person, dining table, cup, fork, hot dog, and chair. Four of them are selected for visualization. Plots in the second column represent the shared attention across all subjects, and the remaining eight columns show the residual, subject-specific attention alongside predicted probabilities to compare recognition confidence and priority. + +![](images/166f725fe9812cf67aaf1d2546423ee054532f5ec836a0403c660ccdb90a8391.jpg) + +![](images/e3509f7cd96a3b93bb4162c4e07cfab4de889a8510736b085bdf0adee8e8d77a.jpg) + +![](images/589940a6330dba9d75c4be02476b3ce7924b144ebbf270500bd5e7105d641142.jpg) + +annotations, the second column represents the shared attention map common across all subjects, and the remaining eight columns show the residual, subject-specific attention patterns. Four object-specific attention maps are visualized on rows with predicted logits to compare recognition confidence. Considering images are shown for only 3s [2], patterns of attention and object recognition offer even more intriguing insights into rapid, automatic processes of visual information and neural encoding. + +Temporal Constraints on Attention Allocation: Within a brief 3-second viewing window, the brain must rapidly parse and prioritize elements of a complex scene. Notably, despite occupying only a modest portion of the image, the object chair consistently receives high attention across all subjects. This suggests that chairs are processed in an early, feed-forward manner-likely due to their high salience and distinctive visual features that enable rapid recognition under time constraints. To validate this, we analyzed prediction confidence for chair in the training set across subjects, grouped + +![](images/55fa2be1d235f61944bc0fafb743a449e8b52e7e54de322bd4f628db5c19b570.jpg) +Figure 6: Recognition of chair by object size. + +by object size (Figure 6). The results confirm that chairs of sufficient size are reliably identified by all participants. This immediate and confident response highlights the efficiency of the visual system in detecting familiar, contextually relevant objects with minimal cognitive effort. + +Subject-specific Focus Under Time Constraints: Under these brief viewing conditions, subject-specific differences in attention to objects like cup, fork, and hot dog become especially revealing. Variation in attention to cup, particularly with subj01 and subj03 achieving recognition confidence (predicted probabilities $>0.5$ ) by focusing more precisely on its location, suggests that these individuals may possess faster or more selective attentional strategies. Such patterns point to individual differences in visual processing speed, attentional control, or + +![](images/86219a2150fa0030181d4e6261e6f87e8581b6cc84df74c2cb55691f9d92c13a.jpg) +Figure 7: Sensitivity to cup across subjects. + +perceptual expertise that influence object prioritization under time constraints. These findings are consistent with training results (Figure 7), where subj01 and subj03 show the highest sensitivity to cups, as measured by weighted MCC. Interestingly, although none of the subjects successfully recog- + +size fork or hot dog in Figure 5, all but subj02 allocate more attention to fork than hot dog, suggesting a subtle yet consistent attentional bias that reflects object familiarity or contextual relevance. + +To the best of our knowledge, the proposed iMIND is the first model to capture subtle variations in how quickly and differently individuals allocate attention within a constrained time frame, demonstrating the model's robustness in simulating real-world neural processes. The model's ability to account for both shared and individual-specific attention patterns in response to brief stimulus exposure can inform the development of neural decoding approaches that better reflect human variability, especially in time-sensitive applications like real-time scene analysis or autonomous driving. Complete details for all visualized figures are provided in the Appendix. + +In sum, the fact that subjects allocate attention differently within just a few seconds underscores the efficiency of neural mechanisms in prioritizing objects and the role of individual cognitive differences. This rapid, nuanced attention mapping highlights how our iMIND framework captures the interplay of shared and individual neural patterns, bridging cognitive neuroscience with computational modeling to decode visual attention in real-world scenarios. + +Table 3: Ablation on loss functions. + +
ID\( \mathcal{L}_{\text{subj}} \)\( \mathcal{L}_{\text{orth}} \)λmAP (%)
Full✓✓.178.36
1✓-7.15 (↓)
2✓.1-11.17 (↓)
3-11.07 (↓)
4✓✓.01-0.94 (↓)
5✓✓1-2.95 (↓)
6✓✓10-0.84 (↓)
+ +# 3.6 Ablation Studies + +Loss Functions. Our novel designs—subject-object disentanglement and the dual-decoding framework—are considered two key factors in achieving SOTA semantic decoding performance. To evaluate their effectiveness and necessity, we test combinations of the two loss functions, $\mathcal{L}_{\mathrm{subj}}$ and $\mathcal{L}_{\mathrm{orth}}$ , along with a trade-off hyperparameter $\lambda$ . Table 3 confirms that both $\mathcal{L}_{\mathrm{orth}}$ for subject-object disentanglement and the dual-decoding design are crucial for achieving high semantic performance. Moreover, the trade-off parameter seems to have a minimal effect on the overall results. The optimal model utilizes all three loss functions with a trade-off parameter of $\lambda = 0.1$ . + +Model Variants. We investigate the impact of two key hyperparameters on the performance of semantic decoding: $d_{\mathrm{obj}}$ , the dimension of the neural object space for subject-object disentanglement in Section 2.3, and $h$ , the number of heads in multi-head cross-attention module in Eq. (9). According to Table 4, increasing the number of heads does not necessarily lead to performance gain, as it may result in potential overfitting. In addition, we found that performance degradation remains minimal as long as there is sufficient feature space allocated for the neural object information. Ultimately, the optimal object classification performance, in terms of mAP, is achieved with $h = 4$ and $d_{\mathrm{obj}} = 700$ . + +Table 4: Ablation on heads and ${d}_{\text{obj }}$ . + +
IDHead(s)\(d_{\text{obj}}\)mAP (%)
Full470078.36
11700-1.95 (↓)
22700-1.12 (↓)
36700-1.35 (↓)
48700-9.18 (↓)
54100-4.01 (↓)
64200-2.11 (↓)
74300-1.04 (↓)
84400-1.38 (↓)
94500-0.89 (↓)
104600-1.14 (↓)
+ +# 4 Conclusion + +In this paper, we introduce an innovative multi-subject dual-decoding framework that decomposes latent fMRI representations into distinct subject-specific and object-specific components using a robust basis transformation. This approach enables precise biometric decoding through individualized neural features, while shared object-oriented features facilitate subject-invariant semantic decoding by querying with CLIP-derived visual representations. Our framework not only establishes a new benchmark for semantic decoding accuracy but also reveals variations in attentional focus across subjects when viewing identical visual stimuli. Additionally, we construct object-specific activation patterns at the voxel level, offering data-driven insights into the brain's visual processing mechanisms. + +In future work, we aim to leverage large-scale fMRI datasets to develop more robust and informative pretrained models for extracting latent neural features. Additionally, we plan to collaborate with brain scientists to deepen our understanding of how specific voxel patterns in fMRI data relate to semantic object representations. This domain knowledge will help bridge the gap between visual features and neural signals, further enhancing the interpretability and accuracy of brain-based decoding models. + +# References + +[1] Samira Abnar and Willem Zuidema. Quantifying attention flow in transformers. In Dan Jurafsky, Joyce Chai, Natalie Schluter, and Joel Tetreault, editors, Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4190-4197, Online, July 2020. Association for Computational Linguistics. +[2] Emily J Allen, Ghislain St-Yves, Yihan Wu, Jesse L Breedlove, Jacob S Prince, Logan T Dowdle, Matthias Nau, Brad Caron, Franco Pestilli, Ian Charest, et al. A massive 7t fmri dataset to bridge cognitive neuroscience and artificial intelligence. Nature neuroscience, 25(1):116-126, 2022. +[3] Antonello Baldassarre, Christopher M Lewis, Giorgia Committeri, Abraham Z Snyder, Gian Luca Romani, and Maurizio Corbetta. Individual variability in functional connectivity predicts performance of a perceptual task. Proceedings of the National Academy of Sciences, 109(9):3516-3521, 2012. +[4] Omar Chehab, Alexandre Defossez, Loiseau Jean-Christophe, Alexandre Gramfort, and JeanRemi King. Deep recurrent encoder: an end-to-end network to model magnetoencephalography at scale. Neurons, Behavior, Data Analysis, and Theory, 2022. +[5] Zijiao Chen, Jiaxin Qing, Tiange Xiang, Wan Lin Yue, and Juan Helen Zhou. Seeing beyond the brain: Conditional diffusion model with sparse masked modeling for vision decoding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22710-22720, 2023. +[6] Davide Cortinovis, Marius V Peelen, and Stefania Bracci. Tool representations in human visual cortex. Journal of Cognitive Neuroscience, 37(3):515-531, 2025. +[7] Hans P Op de Beeck, Ineke Pillet, and J Brendan Ritchie. Factors determining where category-selective areas emerge in visual cortex. Trends in cognitive sciences, 23(9):784-797, 2019. +[8] Andrea Dittadi, Frederik Trauble, Francesco Locatello, Manuel Wüthrich, Vaibhav Agrawal, Ole Winther, Stefan Bauer, and Bernhard Schölkopf. On the transfer of disentangled representations in realistic settings. In 9th International Conference on Learning Representations, 2021. +[9] Changde Du, Kaicheng Fu, Bincheng Wen, Yi Sun, Jie Peng, Wei Wei, Ying Gao, Shengpei Wang, Chuncheng Zhang, Jinpeng Li, et al. Human-like object concept representations emerge naturally in multimodal large language models. Nature Machine Intelligence, pages 1-16, 2025. +[10] Julien Dubois and Ralph Adolphs. Building a science of individual differences from fmri. Trends in cognitive sciences, 20(6):425-443, 2016. +[11] Scott L Fairhall and Alfonso Caramazza. Brain regions that represent amodal conceptual knowledge. Journal of Neuroscience, 33(25):10552-10558, 2013. +[12] Bruce Fischl. Freesurfer. Neuroimage, 62(2):774-781, 2012. +[13] Bruce Fischl, Martin I Sereno, Roger BH Tootell, and Anders M Dale. High-resolution intersubject averaging and a coordinate system for the cortical surface. Human brain mapping, 8(4):272-284, 1999. +[14] Marco Fumero, Florian Wenzel, Luca Zancato, Alessandro Achille, Emanuele Rodolà, Stefano Soatto, Bernhard Schölkopf, and Francesco Locatello. Leveraging sparse and shared feature activations for disentangled representation learning. Advances in Neural Information Processing Systems, 36:27682-27698, 2023. +[15] Michelle R Greene and Aude Oliva. The briefest of glances: The time course of natural scene understanding. Psychological science, 20(4):464-472, 2009. +[16] Kalanit Grill-Spector and Rafael Malach. The human visual cortex. Annu. Rev. Neurosci., 27(1):649-677, 2004. + +[17] Zijin Gu, Keith Jamison, Amy Kuceyeski, and Mert R. Sabuncu. Decoding natural image stimuli from fmri data with a surface-based convolutional network. In Ipek Oguz, Jack Noble, Xiaoxiao Li, Martin Styner, Christian Baumgartner, Mirabela Rusu, Tobias Heinmann, Despina Kontos, Bennett Landman, and Benoit Dawant, editors, Medical Imaging with Deep Learning, volume 227 of Proceedings of Machine Learning Research, pages 107-118. PMLR, 10-12 Jul 2024. +[18] Zijin Gu, Keith Wakefield Jamison, Meenakshi Khosla, Emily J Allen, Yihan Wu, Ghislain St-Yves, Thomas Naselaris, Kendrick Kay, Mert R Sabuncu, and Amy Kuceyeski. Neurogen: activation optimized image synthesis for discovery neuroscience. NeuroImage, 247:118812, 2022. +[19] Martin N Hebart, Charles Y Zheng, Francisco Pereira, and Chris I Baker. Revealing the multidimensional mental representations of natural objects underlying human similarity judgements. Nature human behaviour, 4(11):1173-1185, 2020. +[20] Klaus Hoenig, Eun-Jin Sim, Viktor Bochev, Bärbel Herrnberger, and Markus Kiefer. Conceptual flexibility in the human brain: dynamic recruitment of semantic maps from visual, motor, and motion-related areas. Journal of Cognitive Neuroscience, 20(10):1799-1814, 2008. +[21] Alexander G Huth, Shinji Nishimoto, An T Vu, and Jack L Gallant. A continuous semantic space describes the representation of thousands of object and action categories across the human brain. Neuron, 76(6):1210-1224, 2012. +[22] Sikun Lin, Thomas Sprague, and Ambuj K Singh. Mind reader: Reconstructing complex images from brain activities. Advances in Neural Information Processing Systems, 35:29624-29636, 2022. +[23] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dóllár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. +[24] Andrew Luo, Maggie Henderson, Leila Wehbe, and Michael Tarr. Brain diffusion for visual exploration: Cortical discovery using large scale generative models. Advances in Neural Information Processing Systems, 36, 2024. +[25] Andrew F Luo, Margaret M Henderson, Michael J Tarr, and Leila Wehbe. *Brainscuba: Fine-grained natural language captions of visual cortex selectivity.* arXiv preprint arXiv:2310.04420, 2023. +[26] Florian P Mahner, Lukas Mutenthaler, Umut Güçlü, and Martin N Hebart. Dimensions underlying the representational alignment of deep neural networks with humans. Nature Machine Intelligence, 7(6):848-859, 2025. +[27] George R Mangun, Michael H Buonocore, Massimo Girelli, and Amishi P Jha. Erp and fmri measures of visual spatial selective attention. Human brain mapping, 6(5-6):383-389, 1998. +[28] Brian W Matthews. Comparison of the predicted and observed secondary structure of t4 phage lysozyme. Biochimica et Biophysica Acta (BBA)-Protein Structure, 405(2):442-451, 1975. +[29] Paul M Matthews and Peter Jezzard. Functional magnetic resonance imaging. Journal of Neurology, Neurosurgery & Psychiatry, 75(1):6-12, 2004. +[30] Arne May. Experience-dependent structural plasticity in the adult human brain. Trends in cognitive sciences, 15(10):475-482, 2011. +[31] John Mazziotta, Arthur Toga, Alan Evans, Peter Fox, Jack Lancaster, Karl Zilles, Roger Woods, Tomas Paus, Gregory Simpson, Bruce Pike, et al. A four-dimensional probabilistic atlas of the human brain. Journal of the American Medical Informatics Association, 8(5):401-430, 2001. +[32] K Luan Phan, Tor Wager, Stephan F Taylor, and Israel Liberzon. Functional neuroanatomy of emotion: a meta-analysis of emotion activation studies in pet and fmri. Neuroimage, 16(2):331-348, 2002. + +[33] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. +[34] N Apurva Ratan Murty, Pouya Bashivan, Alex Abate, James J DiCarlo, and Nancy Kanwisher. Computational models of category-selective brain regions enable high-throughput tests of selectivity. Nature communications, 12(1):5540, 2021. +[35] Paul Scotti, Atmadeep Banerjee, Jimmie Goode, Stepan Shabalin, Alex Nguyen, Aidan Dempster, Nathalie Verlinde, Elad Yundler, David Weisberg, Kenneth Norman, et al. Reconstructing the mind's eye: fmri-to-image with contrastive learning and diffusion priors. Advances in Neural Information Processing Systems, 36, 2024. +[36] Paul Steven Scotti, Mihir Tripathy, Cesar Torrico, Reese Kneeland, Tong Chen, Ashutosh Narang, Charan Santhirasegaran, Jonathan Xu, Thomas Naselaris, Kenneth A Norman, et al. Mindeye2: Shared-subject models enable fmri-to-image with 1 hour of data. In Forty-first International Conference on Machine Learning, 2024. +[37] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision, pages 618-626, 2017. +[38] Yu Takagi and Shinji Nishimoto. High-resolution image reconstruction with latent diffusion models from human brain activity. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14453-14463, 2023. +[39] Frederik Trauble, Elliot Creager, Niki Kilbertus, Francesco Locatello, Andrea Dittadi, Anirudh Goyal, Bernhard Scholkopf, and Stefan Bauer. On disentangled representations learned from correlated data. In International conference on machine learning, pages 10401-10412. PMLR, 2021. +[40] Kamil Uğurbil, Junqian Xu, Edward J Auerbach, Steen Moeller, An T Vu, Julio M Duarte-Carvajalino, Christophe Lenglet, Xiaoping Wu, Sebastian Schmitter, Pierre Francois Van de Moortele, et al. Pushing spatial and temporal resolution for functional and diffusion mri in the human connectome project. Neuroimage, 80:80-104, 2013. +[41] Shizun Wang, Songhua Liu, Zhenxiong Tan, and Xinchao Wang. Mindbridge: A cross-subject brain decoding framework. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11333–11342, 2024. +[42] Galit Yovel and Nancy Kanwisher. The neural basis of the behavioral face-inversion effect. Current biology, 15(24):2256-2262, 2005. +[43] Qi Zhang, Yifei Wang, and Yisen Wang. Identifiable contrastive learning with automatic feature importance discovery. Advances in Neural Information Processing Systems, 36, 2024. +[44] Qiongyi Zhou, Changde Du, Shengpei Wang, and Huiguang He. Clip-mused: Clip-guided multi-subject visual neural information semantic decoding. In The Twelfth International Conference on Learning Representations, 2024. + +# NeurIPS Paper Checklist + +# 1. Claims + +Question: Do the main claims made in the abstract and introduction accurately reflect the paper's contributions and scope? + +Answer: [Yes] + +Justification: we clearly state our contributions that match claims made in the abstract and introduction. we also did theoretic proof and conduct extensively experiments to validate our claims in the result section in the main paper and in Appendix + +Guidelines: + +- The answer NA means that the abstract and introduction do not include the claims made in the paper. +- The abstract and/or introduction should clearly state the claims made, including the contributions made in the paper and important assumptions and limitations. A No or NA answer to this question will not be perceived well by the reviewers. +- The claims made should match theoretical and experimental results, and reflect how much the results can be expected to generalize to other settings. +- It is fine to include aspirational goals as motivation as long as it is clear that these goals are not attained by the paper. + +# 2. Limitations + +Question: Does the paper discuss the limitations of the work performed by the authors? + +Answer: [Yes] + +Justification: Due to page limit, we don't include a limitation section in the main paper. Instead, we point out the limitation of the proposed method in Appendix A.6. + +Guidelines: + +- The answer NA means that the paper has no limitation while the answer No means that the paper has limitations, but those are not discussed in the paper. +- The authors are encouraged to create a separate "Limitations" section in their paper. +- The paper should point out any strong assumptions and how robust the results are to violations of these assumptions (e.g., independence assumptions, noiseless settings, model well-specification, asymptotic approximations only holding locally). The authors should reflect on how these assumptions might be violated in practice and what the implications would be. +- The authors should reflect on the scope of the claims made, e.g., if the approach was only tested on a few datasets or with a few runs. In general, empirical results often depend on implicit assumptions, which should be articulated. +- The authors should reflect on the factors that influence the performance of the approach. For example, a facial recognition algorithm may perform poorly when image resolution is low or images are taken in low lighting. Or a speech-to-text system might not be used reliably to provide closed captions for online lectures because it fails to handle technical jargon. +- The authors should discuss the computational efficiency of the proposed algorithms and how they scale with dataset size. +- If applicable, the authors should discuss possible limitations of their approach to address problems of privacy and fairness. +- While the authors might fear that complete honesty about limitations might be used by reviewers as grounds for rejection, a worse outcome might be that reviewers discover limitations that aren't acknowledged in the paper. The authors should use their best judgment and recognize that individual actions in favor of transparency play an important role in developing norms that preserve the integrity of the community. Reviewers will be specifically instructed to not penalize honesty concerning limitations. + +# 3. Theory assumptions and proofs + +Question: For each theoretical result, does the paper provide the full set of assumptions and a complete (and correct) proof? + +# Answer: [NA] + +Justification: We don't make any theoretic contributions in the paper, but we do rely on the assumption that the subject and object information is linear entangled in the latent fMRI space. + +# Guidelines: + +- The answer NA means that the paper does not include theoretical results. +- All the theorems, formulas, and proofs in the paper should be numbered and cross-referenced. +- All assumptions should be clearly stated or referenced in the statement of any theorems. +- The proofs can either appear in the main paper or the supplemental material, but if they appear in the supplemental material, the authors are encouraged to provide a short proof sketch to provide intuition. +- Inversely, any informal proof provided in the core of the paper should be complemented by formal proofs provided in appendix or supplemental material. +- Theorems and Lemmas that the proof relies upon should be properly referenced. + +# 4. Experimental result reproducibility + +Question: Does the paper fully disclose all the information needed to reproduce the main experimental results of the paper to the extent that it affects the main claims and/or conclusions of the paper (regardless of whether the code and data are provided or not)? + +# Answer: [Yes] + +Justification: All details necessary for reproducing the experimental results are fully described in either the main paper or appendix. The code will be released upon acceptance. + +# Guidelines: + +- The answer NA means that the paper does not include experiments. +- If the paper includes experiments, a No answer to this question will not be perceived well by the reviewers: Making the paper reproducible is important, regardless of whether the code and data are provided or not. +- If the contribution is a dataset and/or model, the authors should describe the steps taken to make their results reproducible or verifiable. +- Depending on the contribution, reproducibility can be accomplished in various ways. For example, if the contribution is a novel architecture, describing the architecture fully might suffice, or if the contribution is a specific model and empirical evaluation, it may be necessary to either make it possible for others to replicate the model with the same dataset, or provide access to the model. In general, releasing code and data is often one good way to accomplish this, but reproducibility can also be provided via detailed instructions for how to replicate the results, access to a hosted model (e.g., in the case of a large language model), releasing of a model checkpoint, or other means that are appropriate to the research performed. +- While NeurIPS does not require releasing code, the conference does require all submissions to provide some reasonable avenue for reproducibility, which may depend on the nature of the contribution. For example +(a) If the contribution is primarily a new algorithm, the paper should make it clear how to reproduce that algorithm. +(b) If the contribution is primarily a new model architecture, the paper should describe the architecture clearly and fully. +(c) If the contribution is a new model (e.g., a large language model), then there should either be a way to access this model for reproducing the results or a way to reproduce the model (e.g., with an open-source dataset or instructions for how to construct the dataset). +(d) We recognize that reproducibility may be tricky in some cases, in which case authors are welcome to describe the particular way they provide for reproducibility. In the case of closed-source models, it may be that access to the model is limited in some way (e.g., to registered users), but it should be possible for other researchers to have some path to reproducing or verifying the results. + +# 5. Open access to data and code + +Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material? + +Answer: [Yes] + +Justification: In appendix, we describe the used dataset as well as the version of publicly available pretrained deep models we rely on in details. + +Guidelines: + +- The answer NA means that paper does not include experiments requiring code. +- Please see the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details. +- While we encourage the release of code and data, we understand that this might not be possible, so "No" is an acceptable answer. Papers cannot be rejected simply for not including code, unless this is central to the contribution (e.g., for a new open-source benchmark). +- The instructions should contain the exact command and environment needed to run to reproduce the results. See the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details. +- The authors should provide instructions on data access and preparation, including how to access the raw data, preprocessed data, intermediate data, and generated data, etc. +- The authors should provide scripts to reproduce all experimental results for the new proposed method and baselines. If only a subset of experiments are reproducible, they should state which ones are omitted from the script and why. +- At submission time, to preserve anonymity, the authors should release anonymized versions (if applicable). +- Providing as much information as possible in supplemental material (appended to the paper) is recommended, but including URLs to data and code is permitted. + +# 6. Experimental setting/details + +Question: Does the paper specify all the training and test details (e.g., data splits, hyperparameters, how they were chosen, type of optimizer, etc.) necessary to understand the results? + +Answer: [Yes] + +Justification: Please refer to the implementation details in appeix for all the details regarding hyperparameter choice, preprocessing, etc. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The experimental setting should be presented in the core of the paper to a level of detail that is necessary to appreciate the results and make sense of them. +- The full details can be provided either with the code, in appendix, or as supplemental material. + +# 7. Experiment statistical significance + +Question: Does the paper report error bars suitably and correctly defined or other appropriate information about the statistical significance of the experiments? + +Answer: [No] + +Justification: We don't do statistical significance check simply because our model outperforms existing models by a large margin that don't need a statistical test to prove it. Also, the model is trained and tested on a large dataset. It would be time-consuming to do such test. + +Guidelines: + +- The answer NA means that the paper does not include experiments. + +- The authors should answer "Yes" if the results are accompanied by error bars, confidence intervals, or statistical significance tests, at least for the experiments that support the main claims of the paper. +- The factors of variability that the error bars are capturing should be clearly stated (for example, train/test split, initialization, random drawing of some parameter, or overall run with given experimental conditions). +- The method for calculating the error bars should be explained (closed form formula, call to a library function, bootstrap, etc.) +- The assumptions made should be given (e.g., Normally distributed errors). +- It should be clear whether the error bar is the standard deviation or the standard error of the mean. +- It is OK to report 1-sigma error bars, but one should state it. The authors should preferably report a 2-sigma error bar than state that they have a $96\%$ CI, if the hypothesis of Normality of errors is not verified. +- For asymmetric distributions, the authors should be careful not to show in tables or figures symmetric error bars that would yield results that are out of range (e.g. negative error rates). +- If error bars are reported in tables or plots, The authors should explain in the text how they were calculated and reference the corresponding figures or tables in the text. + +# 8. Experiments compute resources + +Question: For each experiment, does the paper provide sufficient information on the computer resources (type of compute workers, memory, time of execution) needed to reproduce the experiments? + +Answer: [Yes] + +Justification: We list the computing resources needed for the proposed method, whose parameters are publicly available. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The paper should indicate the type of compute workers CPU or GPU, internal cluster, or cloud provider, including relevant memory and storage. +- The paper should provide the amount of compute required for each of the individual experimental runs as well as estimate the total compute. +- The paper should disclose whether the full research project required more compute than the experiments reported in the paper (e.g., preliminary or failed experiments that didn't make it into the paper). + +# 9. Code of ethics + +Question: Does the research conducted in the paper conform, in every respect, with the NeurIPS Code of Ethics https://neurips.cc/public/EthicsGuidelines? + +Answer: [Yes] + +Justification: We closely follow the code of ethics by NeurIPS. + +Guidelines: + +- The answer NA means that the authors have not reviewed the NeurIPS Code of Ethics. +- If the authors answer No, they should explain the special circumstances that require a deviation from the Code of Ethics. +- The authors should make sure to preserve anonymity (e.g., if there is a special consideration due to laws or regulations in their jurisdiction). + +# 10. Broader impacts + +Question: Does the paper discuss both potential positive societal impacts and negative societal impacts of the work performed? + +Answer: [NA] + +Justification: The paper is an exploring work and far away from application. So, it may not have any societal impacts for now. + +# Guidelines: + +- The answer NA means that there is no societal impact of the work performed. +- If the authors answer NA or No, they should explain why their work has no societal impact or why the paper does not address societal impact. +- Examples of negative societal impacts include potential malicious or unintended uses (e.g., disinformation, generating fake profiles, surveillance), fairness considerations (e.g., deployment of technologies that could make decisions that unfairly impact specific groups), privacy considerations, and security considerations. +- The conference expects that many papers will be foundational research and not tied to particular applications, let alone deployments. However, if there is a direct path to any negative applications, the authors should point it out. For example, it is legitimate to point out that an improvement in the quality of generative models could be used to generate deepfakes for disinformation. On the other hand, it is not needed to point out that a generic algorithm for optimizing neural networks could enable people to train models that generate Deepfakes faster. +- The authors should consider possible harms that could arise when the technology is being used as intended and functioning correctly, harms that could arise when the technology is being used as intended but gives incorrect results, and harms following from (intentional or unintentional) misuse of the technology. +- If there are negative societal impacts, the authors could also discuss possible mitigation strategies (e.g., gated release of models, providing defenses in addition to attacks, mechanisms for monitoring misuse, mechanisms to monitor how a system learns from feedback over time, improving the efficiency and accessibility of ML). + +# 11. Safeguards + +Question: Does the paper describe safeguards that have been put in place for responsible release of data or models that have a high risk for misuse (e.g., pretrained language models, image generators, or scraped datasets)? + +Answer: [No] + +Justification: we don't think the reuse of our data or models will have a high risk for misuse. + +# Guidelines: + +- The answer NA means that the paper poses no such risks. +- Released models that have a high risk for misuse or dual-use should be released with necessary safeguards to allow for controlled use of the model, for example by requiring that users adhere to usage guidelines or restrictions to access the model or implementing safety filters. +- Datasets that have been scraped from the Internet could pose safety risks. The authors should describe how they avoided releasing unsafe images. +- We recognize that providing effective safeguards is challenging, and many papers do not require this, but we encourage authors to take this into account and make a best faith effort. + +# 12. Licenses for existing assets + +Question: Are the creators or original owners of assets (e.g., code, data, models), used in the paper, properly credited and are the license and terms of use explicitly mentioned and properly respected? + +Answer: [Yes] + +Justification: The data and the pretrained deep models are fully described. We give credits to the original creator in the main paper. + +# Guidelines: + +- The answer NA means that the paper does not use existing assets. +- The authors should cite the original paper that produced the code package or dataset. +- The authors should state which version of the asset is used and, if possible, include a URL. +- The name of the license (e.g., CC-BY 4.0) should be included for each asset. + +- For scraped data from a particular source (e.g., website), the copyright and terms of service of that source should be provided. +- If assets are released, the license, copyright information, and terms of use in the package should be provided. For popular datasets, paperswithcode.com/datasets has curated licenses for some datasets. Their licensing guide can help determine the license of a dataset. +- For existing datasets that are re-packaged, both the original license and the license of the derived asset (if it has changed) should be provided. +- If this information is not available online, the authors are encouraged to reach out to the asset's creators. + +# 13. New assets + +Question: Are new assets introduced in the paper well documented and is the documentation provided alongside the assets? + +Answer: [Yes] + +Justification: We will release our code upon acceptance, which include new assets such as trained model weights. + +Guidelines: + +- The answer NA means that the paper does not release new assets. +- Researchers should communicate the details of the dataset/code/model as part of their submissions via structured templates. This includes details about training, license, limitations, etc. +- The paper should discuss whether and how consent was obtained from people whose asset is used. +- At submission time, remember to anonymize your assets (if applicable). You can either create an anonymized URL or include an anonymized zip file. + +# 14. Crowdsourcing and research with human subjects + +Question: For crowdsourcing experiments and research with human subjects, does the paper include the full text of instructions given to participants and screenshots, if applicable, as well as details about compensation (if any)? + +Answer: [No] + +Justification: Our research relies on human neural signals data, but it has no potential risks for incurred participants. + +Guidelines: + +- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. +- Including this information in the supplemental material is fine, but if the main contribution of the paper involves human subjects, then as much detail as possible should be included in the main paper. +- According to the NeurIPS Code of Ethics, workers involved in data collection, curation, or other labor should be paid at least the minimum wage in the country of the data collector. + +# 15. Institutional review board (IRB) approvals or equivalent for research with human subjects + +Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals (or an equivalent approval/review based on the requirements of your country or institution) were obtained? + +Answer: [No] + +Justification: Our research relies on human neural signals data, but it has no potential risks for incurred participants. + +Guidelines: + +- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. +- Depending on the country in which research is conducted, IRB approval (or equivalent) may be required for any human subjects research. If you obtained IRB approval, you should clearly state this in the paper. +- We recognize that the procedures for this may vary significantly between institutions and locations, and we expect authors to adhere to the NeurIPS Code of Ethics and the guidelines for their institution. +- For initial submissions, do not include any information that would break anonymity (if applicable), such as the institution conducting the review. + +# 16. Declaration of LLM usage + +Question: Does the paper describe the usage of LLMs if it is an important, original, or non-standard component of the core methods in this research? Note that if the LLM is used only for writing, editing, or formatting purposes and does not impact the core methodology, scientific rigorousness, or originality of the research, declaration is not required. + +Answer: [NA] + +Justification: the core method development in this research does not involve LLMs as any important, original, or non-standard components. + +Guidelines: + +- The answer NA means that the core method development in this research does not involve LLMs as any important, original, or non-standard components. +- Please refer to our LLM policy (https://neurips.cc/Conferences/2025/LLM) for what should or should not be described. + +# A Technical Appendices and Supplementary Material + +# A.1 Related Work + +# A.1.1 Single-Subject v.s. Multi-Subject Models + +The application of deep learning to neural data has initially centered on single-subject models tailored to individual participants. BrainDiVE [24] adopts a generative approach, synthesizing images predicted to activate specific regions of the human visual cortex. Moving beyond visual modalities, Mind Reader [22] incorporates textual information to reconstruct complex images containing multiple objects from brain activities. Extending this multimodal approach, BrainSCUBA [25] takes advantage of contrastive vision-language models and large-language models to generate voxel-wise captions, eliminating the need for human-annotated voxel-caption data. While single-subject models have achieved notable success, they face inherent limitations. These models require large amounts of subject-specific data to train robust models, which is challenging given the high costs and effort involved in collecting fMRI data. Furthermore, they are prone to overfitting, exhibit poor generalizability across individuals, and struggle with scalability when applied to larger datasets or diverse populations. + +To overcome these challenges, multi-subject models aim to unify data across participants, enabling shared representation learning. However, this approach introduces significant complexity due to inter-subject variability, which arises from static anatomical differences and dynamic functional responses. Various methods have been proposed to overcome these obstacles. [4] employs subject embeddings and recurrent architectures to account for inter-trial and inter-subject variability, outperforming many single-subject models in predicting MEG time series. MindBridge [41] introduces a biologically inspired aggregation function and a cyclic fMRI reconstruction mechanism to achieve subject-invariant representation learning. MindEye2 [36] aligns spatial patterns of fMRI activity to a shared latent space using subject-specific ridge regression, improving out-of-subject generalization with limited training data and achieving state-of-the-art results in image retrieval and reconstruction. More recently, CLIP-MUSED [44] introduced learnable subject-specific tokens to facilitate the aggregation of multi-subject data without a linear increase in model parameters. This approach integrates representational similarity analysis (RSA) to guide token representation learning based on the topological relationships of visual stimuli in the latent visual space. + +These advancements demonstrate the potential of multi-subject models to surpass the limitations of single-subject approaches, providing more generalizable and scalable solutions for neural decoding tasks. However, to the best of our knowledge, existing multi-subject neural decoding models predominantly adopt what we term a suppressive strategy for handling inter-subject variability. This approach aims to minimize subject-specific differences during learning, progressively refining features to become more task-relevant as the model deepens. In these frameworks, subject-specific information is often treated as noise or an obstacle to effective decoding. In contrast, our iMIND framework proposes an instructive strategy. Rather than suppressing subject-specific differences, our model embraces this variability by explicitly disentangling subject-specific features from task-relevant ones. By doing so, iMIND not only preserves individual-specific neural representations but also leverages them positively to enhance both subject-specific and shared task-related decoding. This dual-decoding approach enables iMIND to achieve superior performance while offering insights into both individual neural patterns and shared semantic representations. + +# A.1.2 Vision-Neural Interactions + +Decoding visual information from neural signals is an inherently multi-modal task, involving the alignment and interaction of at least two modalities: images and neural signals (eg., fMRI). Broadly speaking, approaches for vision-neural modality alignment can be categorized into two branches based on the direction of projection between visual and neural spaces. + +The first branch projects neural signals into a pre-trained latent visual space. This approach is exemplified by works such as [35], which maps flattened spatial patterns of fMRI activity across 3D cortical tissue cubes into the image embedding space of a pre-trained CLIP model. Similarly, [38] predicts latent representations of presented images from fMRI signals within the early visual cortex. Other notable works in this branch include [5, 22, 36, 41], which leverage pre-trained visual generative models, such as GANs [34] and diffusion models, for reconstruction tasks. These models capitalize + +on large-scale visual datasets and avoid re-training resource-intensive generative architectures, which would otherwise be infeasible given the scarcity of paired neural-visual data. The second branch adopts the opposite approach by projecting latent visual image features into the neural space. This method is particularly useful for generating synthetic stimuli that activate specific brain regions, enabling the study of feature preferences in different areas of the brain. Classic examples include [18, 24, 25], which investigate neural activation patterns in response to synthetic stimuli derived from visual features. + +Our iMIND model takes a fundamentally different approach to vision-neural modality interaction. Rather than projecting between modalities, we use CLIP-derived vision features as queries to extract corresponding neural object features directly from neural representations. This design choice is motivated by several factors. First, as a semantic neural decoding framework, iMIND does not rely on resource-intensive generative models. Second, direct projections between modalities often result in significant information loss and modality gaps that require careful handling. Most importantly, our approach is expected to enable the investigation of subject-specific attention variations when viewing the same visual stimuli. By treating the CLIP vision features as pseudo-ground-truths for object presence, we leverage their subject-invariant properties as an anchor to explore how neural responses to specific objects differ across subjects. This design uniquely aligns with the goals of understanding inter-subject variability in neural decoding. + +# A.2 Theoretic Validation for Basis Transformation + +In this section, we present the basis transformation in linear algebra and establish the relationship between coordinates in different bases. The derivation ensures clarity in transitioning from one basis to another, essential for interpreting subject-specific neural space $\mathcal{F}_{\mathrm{subj}}$ and object-specific neural space $\mathcal{F}_{\mathrm{obj}}$ mentioned at the end of Section 2.3 in the main paper. We begin with the following formal claim: + +Claim Let $\mathcal{V}$ be a vector space of cardinality $d$ over $\mathbb{R}$ , with a standard basis $\mathbf{E} = \{\mathbf{e}_1, \mathbf{e}_2, \dots, \mathbf{e}_d\}$ of $\mathbb{R}^d$ and an arbitrary basis $\mathbf{B} = \{\mathbf{b}_1, \mathbf{b}_2, \dots, \mathbf{b}_d\}$ of $\mathbb{R}^d$ . For any vector $\mathbf{v} \in \mathcal{V}$ , if its coordinate with respect to the standard basis $\mathbf{E}$ is given by $[\mathbf{v}]_{\mathbf{E}} \in \mathbb{R}^d$ , then its coordinate with respect to the basis $\mathbf{B}$ can be derived as: + +$$ +[ \mathbf {v} ] _ {\mathbf {B}} = \mathbf {P} _ {\mathbf {E} \rightarrow \mathbf {B}} \cdot [ \mathbf {v} ] _ {\mathbf {E}}, \tag {15} +$$ + +where $\mathbf{P}_{\mathbf{E}\rightarrow \mathbf{B}}\in \mathbb{R}^{d\times d}$ is the change-of-basis matrix from $\mathbf{E}$ to $\mathbf{B}$ , defined as: + +$$ +\mathbf {P} _ {\mathbf {E} \rightarrow \mathbf {B}} := \left[\begin{array}{c c c c}|&|&\dots&|\\\mathbf {b} _ {1}&\mathbf {b} _ {2}&\dots&\mathbf {b} _ {d}\\|&|&\dots&|\end{array}\right] ^ {- 1}. \tag {16} +$$ + +Proof Since $\mathbf{v} \in \mathcal{V}$ and $\mathbf{E}$ forms a basis for the vector space $\mathcal{V}$ , $\mathbf{v}$ can be written as a linear combination of all basis vectors from $\mathbf{E}$ : + +$$ +\mathbf {v} = \sum_ {i = 1} ^ {d} a _ {i} \mathbf {e} _ {i}, \tag {17} +$$ + +where $a_{i}\in \mathbb{R}$ are scalars. In this case, the coordinate of $\mathbf{v}$ with respect to $\mathbf{E}$ is: + +$$ +[ \mathbf {v} ] _ {\mathbf {E}} = \left(a _ {1}, a _ {2}, \dots , a _ {d}\right) ^ {\top} \in \mathbb {R} ^ {d}. \tag {18} +$$ + +Similarly, because $\mathbf{B}$ also forms a basis for $\nu$ , we can express $\mathbf{v}$ as: + +$$ +\mathbf {v} = \sum_ {i = 1} ^ {d} w _ {i} \mathbf {b} _ {i}, \tag {19} +$$ + +where $w_{i}\in \mathbb{R}$ are scalars. The coordinate of $\mathbf{v}$ with respect to $\mathbf{B}$ is: + +$$ +[ \mathbf {v} ] _ {\mathbf {B}} = \left(w _ {1}, w _ {2}, \dots , w _ {d}\right) ^ {\top} \in \mathbb {R} ^ {d}. \tag {20} +$$ + +Since each basis vector $\mathbf{e}_i$ within $\mathbf{E}$ is also an element of the vector space $\mathcal{V}$ , it can also be written as a linear combination of all basis vectors from $\mathbf{B}$ : + +$$ +\mathbf {e} _ {i} = \sum_ {j = 1} ^ {d} p _ {j i} \mathbf {b} _ {j}. \tag {21} +$$ + +Writing the equation above in matrix form, we obtain: + +$$ +\left[ \begin{array}{c c c c} | & | & \dots & | \\ \mathbf {e} _ {1} & \mathbf {e} _ {2} & \dots & \mathbf {e} _ {d} \\ | & | & \dots & | \end{array} \right] = \left[ \begin{array}{c c c c} p _ {1 1} & p _ {1 2} & \dots & p _ {1 d} \\ p _ {2 1} & \ddots & & p _ {2 d} \\ \vdots & & \ddots & \vdots \\ p _ {d 1} & p _ {d 2} & \dots & p _ {d d} \end{array} \right] \left[ \begin{array}{c c c c} | & | & \dots & | \\ \mathbf {b} _ {1} & \mathbf {b} _ {2} & \dots & \mathbf {b} _ {d} \\ | & | & \dots & | \end{array} \right]. \tag {22} +$$ + +Denoting the middle matrix as $\mathbf{P}$ and solving it, we get: + +$$ +\mathbf {P} := \left[ \begin{array}{c c c c} p _ {1 1} & p _ {1 2} & \dots & p _ {1 d} \\ p _ {2 1} & \ddots & & p _ {2 d} \\ \vdots & & \ddots & \vdots \\ p _ {d 1} & p _ {d 2} & \dots & p _ {d d} \end{array} \right] = \left[ \begin{array}{c c c c} | & | & \dots & | \\ \mathbf {b} _ {1} & \mathbf {b} _ {2} & \dots & \mathbf {b} _ {d} \\ | & | & \dots & | \end{array} \right] ^ {- 1}. \tag {23} +$$ + +Next, let's plug each $\mathbf{e}_i$ into Eq. (17) using the formulation of Eq. (21): + +$$ +\mathbf {v} = \sum_ {i = 1} ^ {d} a _ {i} \mathbf {e} _ {i} = \sum_ {i = 1} ^ {d} \left(a _ {i} \sum_ {j = 1} ^ {d} p _ {j i} \mathbf {b} _ {j}\right) = \sum_ {j = 1} ^ {d} \left(\sum_ {i = 1} ^ {d} a _ {i} p _ {j i}\right) \mathbf {b} _ {j}. \tag {24} +$$ + +Combining with Eq. (19), we have: + +$$ +\sum_ {j = 1} ^ {d} w _ {j} \mathbf {b} _ {j} = \sum_ {j = 1} ^ {d} \left(\sum_ {i = 1} ^ {d} a _ {i} p _ {j i}\right) \mathbf {b} _ {j}. \tag {25} +$$ + +Move everything to the left hand: + +$$ +\sum_ {j = 1} ^ {d} \left[ w _ {j} - \left(\sum_ {i = 1} ^ {d} a _ {i} p _ {j i}\right) \right] \mathbf {b} _ {j} = 0. \tag {26} +$$ + +According to the claim, $\mathbf{B} = \{\mathbf{b}_1,\mathbf{b}_2,\dots ,\mathbf{b}_d\}$ is a basis of $\mathbf{R}^d$ . Therefore, Eq. (26) holds if and only if: + +$$ +w _ {j} - \sum_ {i = 1} ^ {d} a _ {i} p _ {j i} = 0 \quad \text {f o r} j = 1, 2 \dots , d. \tag {27} +$$ + +Equivalently, we have: + +$$ +\left[ \begin{array}{c} w _ {1} \\ w _ {2} \\ \vdots \\ w _ {d} \end{array} \right] = \sum_ {i = 1} ^ {d} a _ {i} \left[ \begin{array}{c} p _ {1 i} \\ p _ {2 i} \\ \vdots \\ p _ {d i} \end{array} \right] = \left[ \begin{array}{c c c c} p _ {1 1} & p _ {1 2} & \dots & p _ {1 d} \\ p _ {2 1} & \ddots & & p _ {2 d} \\ \vdots & & \ddots & \vdots \\ p _ {d 1} & p _ {d 2} & \dots & p _ {d d} \end{array} \right] \left[ \begin{array}{c} a _ {1} \\ a _ {2} \\ \vdots \\ a _ {d} \end{array} \right]. \tag {28} +$$ + +Using the coordinates expression defined in Eq. (18) and Eq. (20) along with Eq. (23), we obtain the following equation: + +$$ +[ \mathbf {v} ] _ {\mathbf {B}} = \mathbf {P} \cdot [ \mathbf {v} ] _ {\mathbf {E}} \quad \text {w h e r e} \mathbf {P} = \left[ \begin{array}{c c c c} | & | & \dots & | \\ \mathbf {b} _ {1} & \mathbf {b} _ {2} & \dots & \mathbf {b} _ {d} \\ | & | & \dots & | \end{array} \right] ^ {- 1}. \tag {29} +$$ + +Finally, we complete the proof of the claim. + +In our iMIND model, subject-object disentanglement is achieved through a basis transformation described above, where the new basis $\mathbf{B}$ is treated as learnable parameters optimized by loss propagation. We further enforce the orthonormality constraint on $\mathbf{B}$ , as this ensures that any subspace spanned by a subset of $\mathbf{B}$ is orthogonal (complementary) to its counterparts in the original feature space. Specifically, in our model, this constraint guarantees that the subject-specific neural space $\mathcal{F}_{\mathrm{subj}}$ , spanned by $\mathbf{B}_{\mathrm{subj}}$ , and the object-specific neural space $\mathcal{F}_{\mathrm{obj}}$ , spanned by $\bar{\mathbf{B}}_{\mathrm{obj}}$ , are complementary and non-overlapping. Consequently, this orthogonal decomposition yields a perfect separation of subject-specific and object-specific neural features within the latent neural representations. + +# A.3 NSD dataset, Pre-processing, and Implementation + +# A.3.1 NSD Dataset + +The Natural Scenes Dataset (NSD) [2] is a groundbreaking resource in cognitive neuroscience and artificial intelligence, designed to capture extensive, high-resolution fMRI data during natural scene perception. It includes whole-brain fMRI measurements of eight human participants at 7T field strength, with a spatial resolution of $1.8\mathrm{mm}$ . Participants viewed a total of 70,566 natural scene images, with 10,000 unique images per subject (9,000 unique to each participant and 1,000 shared across all participants). Images were sourced from the richly annotated Microsoft COCO dataset [23], ensuring ecological relevance and diversity. The experiment was conducted over 30-40 sessions per participant, taking a rapid event-related design with continuous recognition tasks to guarantee engagement and probe both short- and long-term memory processes. During neural recording, participants were tasked with identifying objects in the image, with each visual stimulus presented for only three seconds per trial. This design makes the NSD particularly well-suited for investigating the mechanisms of rapid attention and visual recognition in human vision. Advanced pre-processing completed by the authors, including denoising and voxel-specific hemodynamic response modeling, yielded high-quality single-trial beta estimates with exceptional signal-to-noise ratios. Complementing the functional data, NSD includes extensive anatomical scans, resting-state data, and behavioral performance measures, enabling multi-faceted investigations of vision and memory. This dataset, with its unparalleled scale and quality, serves as a valuable benchmark for developing and testing machine learning models that aim to decode brain activity and simulate neural representations of natural scenes. In addition, the NSD dataset supports multiple widely used neuroimaging atlases to facilitate data analysis and integration with existing frameworks. Functional data are provided in both native cortical surface space and standard volumetric spaces, including fsaverage [13] and MNI152 [31], enabling compatibility with tools like FreeSurfer [12] and FMRIB Software Library. Additionally, the dataset includes manually defined regions of interest (ROIs) for retinotopic mapping and category-selective areas, such as the early visual cortex and higher-order regions in the ventral visual stream, which is the atlas that we used in our iMIND model. These comprehensive atlases allow researchers to seamlessly apply NSD data to diverse analytic pipelines and cross-study comparisons. + +# A.3.2 Preprocessing + +Our preprocessing pipeline begins with splitting the dataset into training and testing sets. Due to incomplete sessions and data availability constraints, not all trials are accessible for every subject, resulting in a total of 213,000 trials across all participants. Among these, neural recordings corresponding to 1,000 images viewed by all subjects are allocated to the testing set, comprising a total of 21,118 test trials. The remaining neural recordings, corresponding to images viewed exclusively by individual subjects, are included in the training set, resulting in 191,882 training trials. Both training and testing trials are standardized voxel-wise using the mean and standard deviation calculated from the training set. Since each image is presented to a subject three times, we average the fMRI responses across repetitions for each image within each subject. This results in 69,566 training samples and 7,674 testing samples, allowing us to train and evaluate a single multi-subject model across all eight subjects. For each sample, we use the nsdgeneral atlas provided by the NSD dataset to extract visual voxel signals as a 1D vector. However, the number of visual voxels varies between subjects due to anatomical differences, with the voxel length $L_{s}$ ranging from 12,682 to 17,907 across subjects. To unify the input length, we apply a padding strategy inspired by Mind-Vis [5], which conducts wrap-around padding. This approach avoids issues arising from truncation or constant padding to the maximum voxel length. Additionally, since our fMRI encoder is based on a Vision Transformer (ViT), which requires input lengths divisible by the user-defined patch size (64 in our model), we adjust the uniform voxel length $L$ accordingly. The final voxel length across subjects is set to $L = 17,920$ ensuring compatibility with the model while maintaining consistency across participants. + +# A.3.3 Implementation Details + +As described in Section 2 of the main paper, our proposed architecture is trained in two stages. The first stage involves pre-training a ViT-based masked autoencoder, similar to SC-MBM [5], using a self-supervised fMRI reconstruction task. In this stage, we choose a patch size of 64 voxels with a masking ratio of 0.75. The encoder has a hidden dimension of 768 and consists of 12 layers + +Table 5: Generalizability to unseen subjects + +
IDTrained onTested onmAP (%)
M7subj01-07subj01-07.7904
subj01-07subj08.7842
M8 (Full)subj01-08subj01-07.7842
subj01-08subj08.7909
+ +of 6-head self-attention, while the decoder has a hidden dimension of 512 and 8 layers of 8-head self-attention. In the second stage, we discard the decoder and inherit only the encoder from the first stage, which outputs pre-trained feature $\mathbf{F} \in \mathbb{R}^{N \times d}$ with $N = 280$ and $d = 768$ . We set the object neural space dimension $d_{\mathrm{obj}} = 700$ and choose a 4-head cross-attention module for fMRI-vision feature interactions. The CLIP visual encoder we used is clip-vit-base-patch16 released by OpenAI, which remains frozen at all stages of the proposed framework. A trade-off parameter $\lambda$ of 0.1 is set by default to enforce the orthonormal constraint of the learnable basis $\mathbf{B}$ . A detailed investigation is provided in Section 3.6. During this stage, all parameters are optimized end-to-end for subject and object classification tasks. For either stage, we train the model for 100 epochs, including 10 warm-up epochs. The learning rate is initialized at $7.5 \times 10^{-4}$ and terminated at zero adjusted dynamically by a cosine scheduler. The batch size is set to 200. Optimization is performed via the AdamW optimizer with a weight decay of 0.05. All experiments are conducted on two Nvidia RTX 6000 Ada GPUs, with the first stage taking approximately 1.5 hours and the second stage around 2 hours to complete. + +# A.4 Subject Classification Baselines + +To the best of our knowledge, no existing models support subject classification. To provide a comprehensive evaluation, we established baseline models using the exact fMRI preprocessing steps in Appendix A.3.2 and conducted both supervised and unsupervised biometric decoding. All methods are trained and tested on identical data splits and fMRI voxel sets as iMIND, ensuring a fair comparison on the same held-out unseen test set. + +For supervised learning, we employ a single linear layer trained in two ways: + +- Linear Regression: We minimize the mean squared error (MSE) between the input (padded fMRI voxel signals) and the target (one-hot subject IDs), using the ordinary least squares closed-form solution; +- Classification: We train an identical architecture with cross-entropy loss, treating subject identification as a standard classification task. + +For unsupervised learning, we evaluate K-Means clustering with two distance metrics: + +Euclidean (L2) distance; +- Cosine similarity. + +Since the number of subjects is known (8), we set the number of clusters to 8 as well. To measure performance, we compute accuracy and MCC by optimally aligning the learned clusters with ground-truth subject IDs. + +# A.5 Subject Generalizability + +For completeness, we conducted an additional experiment to assess subject generalizability within NSD as shown in Table 5. In our original setup (denoted as M8), iMIND was trained on data from all 8 subjects. For this experiment, we introduced a variant (M7), where iMIND was trained using only the first 7 subjects and tested on the held-out data of subj08. M7/M8 achieved an overall mAP of .7904/.7842 on the first 7 subjects and .7842/.7909 on subj08. These results demonstrate that our proposed method exhibits a reasonable and strong generalizability in neural signal semantic decoding, particularly for unseen subjects. + +# A.6 Limitations + +While our method achieves state-of-the-art performance in both semantic and biometric decoding tasks, several limitations remain unresolved. + +First, the current approach to neural feature extraction may not be optimal. Although functional, the pretrained neural reconstruction stage produces fMRI reconstructions—both numerically and visually—that underperform compared to the ground-truth voxel signals. This suggests that the masked autoencoder (MAE) backbone may not be the most effective architecture for this task, warranting further exploration. + +Second, flattening voxel inputs discards crucial spatial relationships among neighboring voxels, despite neuroscientific evidence that proximal voxels exhibit functional coupling in visual processing. Future work could explore advanced architectures—such as 3D SwinTransformers, which are explicitly designed for volumetric fMRI data and have demonstrated efficacy in neurological disease diagnosis—to better preserve spatial hierarchies and improve feature learning. + +Third, we acknowledge that the brain's functional dynamics are fundamentally non-linear and complex. Our assumption that subject-specific and object-specific components are linearly entangled at the latent feature level is a simplifying inductive bias introduced to enable interpretable, computationally tractable disentanglement. We agree it fully captures the richness of brain representations; rather, it serves as a first-order approximation that enables clear factorization of subject identity and semantic content from fMRI signals. If the linearity assumption does not fully hold, we expect the following potential implications: + +- If subject-object interactions are fundamentally non-linear, the disentangled object representation $\mathbf{Z}_{\mathrm{obj}}$ may still retain residual subject-specific information, potentially introducing subject bias in semantic decoding and diminishing our model's generalizability for unseen subjects. +- Conversely, enforcing strict linear disentanglement may suppress relevant non-linear object features in $\mathbf{Z}_{\mathrm{obj}}$ , potentially smoothing out sharp voxel-object modulations or degrading decoding performance for fine-grained categories. + +Last, our analysis of visual mechanisms relies on post hoc interpretation methods (Grad-CAM and Rollout), which provide only approximate explanations of model behavior. A more principled approach would involve explainable-by-design architectures for fMRI feature extraction, which we leave for future work. + +# A.7 Broader Impacts + +Our work represents a pioneering step toward decoding the brain's visual processing mechanisms, with far-reaching implications for both neuroscience and artificial intelligence. By modeling how the brain transforms visual signals into neural activity and high-level semantics, we aim to uncover the fine-grained functional organization of visual regions—such as those specialized for distinguishing closely related objects (e.g., dogs vs. cats). + +This understanding could enable breakthroughs in brain-computer interfaces (BCIs), where precise neural decoding could restore or augment vision for impaired individuals. Conversely, it also raises ethical considerations: the same principles could theoretically be used to manipulate neural signals, artificially inducing semantic perceptions (e.g., generating "fake" visual concepts in the brain). Such capabilities would necessitate rigorous ethical frameworks to prevent misuse while maximizing societal benefit. + +Further, our computational approach bridges AI and neuroscience, offering interpretable models that could inspire more biologically plausible machine vision systems. By aligning artificial and biological vision, we may accelerate progress in both fields—from improving AI's robustness to advancing treatments for neurological disorders. + +# A.8 Technical Details on Object-Voxel Visualization + +In our experiments, we empirically investigate the relationship between brain activities and semantic objects in visual stimuli. In this section, we detail how voxel contributions to object recognition are + +visualized in our framework. Given an input fMRI voxel signal $\mathbf{V} \in \mathbb{R}^L$ , we obtain its object-specific neural feature map $\mathbf{Z}_{\mathrm{obj}} \in \mathbb{R}^{N \times d_{\mathrm{obj}}}$ before cross-attention module in our model. Using GradCAM [37], we are able to build an activation map $\mathbf{t} \in \mathbb{R}^N$ , which quantifies the contribution of each neural token $\mathbf{z}_{\mathrm{obj}} \in \mathbb{R}^{d_{\mathrm{obj}}}$ in object-specific neural feature map $\mathbf{Z}_{\mathrm{obj}} \in \mathbb{R}^{N \times d_{\mathrm{obj}}}$ to the correct recognition of the object of interest. + +Unlike CNN-based models, which can simply resize $t$ to match the size of the input $\mathbf{V}$ because of their spatial invariance nature (a one-to-one correspondence between input patches and latent feature vectors), our ViT-based neural encoder in iMIND model lacks such spatial invariance properties by default. For instance, the first neural token in $\mathbf{Z}_{\mathrm{obj}}$ does not necessarily correspond to the first voxel patch of $\mathbf{V}$ , making direct resizing infeasible. + +To address this, we leverage Attention Roll-out [1] to approximate how information flows from the input voxels $\mathbf{V}$ to the neural tokens in feature map $\mathbf{Z}_{\mathrm{obj}}$ . Specifically, this information flow in our ViT-based encoder can be measured as follows: + +$$ +\mathbf {A} ^ {l} = \mathbf {A} ^ {(l)} \mathbf {A} ^ {(l - 1)} \dots \mathbf {A} ^ {(2)} \mathbf {A} ^ {(1)}, \tag {30} +$$ + +where $\mathbf{A}^{(k)}\in \mathbb{R}^{N\times N}$ represents the attention weights in the $k$ -th self attention layer of the ViT encoder. Based on the mathematical property of the self-attention mechanism, the element $a_{ij}$ of the attention weights $\mathbf{A}$ at every transformer block defines how much attention flows from the token $j$ in the previous layer to the token $i$ in the next layer. Therefore, each element $a_{ij}^l$ within $\mathbf{A}^l$ defined in Eq. (30) quantifies the degree of information flow from the $j$ -th voxel patch of input fMRI signal $\mathbf{V}\in \mathbb{R}^{L}$ to the $i$ -th token in the feature map $\mathbf{Z}_{\mathrm{obj}}$ . + +Next, we combine the GradCAM-based token contributions $\mathbf{t} \in \mathbb{R}^N$ with the cumulative information flow $\mathbf{A}^l$ to derive the voxel-level activation measurement $\mathbf{T}$ : + +$$ +\mathbf {T} = \mathbf {t} \cdot \mathbf {A} ^ {l} \in \mathbb {R} ^ {N}. \tag {31} +$$ + +Here, $\mathbf{T}$ measures the contribution of each voxel path immediately after embedding the original 1D voxel signal $\mathbf{V}$ of length $L$ into $N$ patches. Since $\mathbf{T}$ is now positionally aligned with the input voxel patches, it can be safely upsampled from size $N$ to $L$ to obtain a voxel-level activation map for each fMRI sample. Unfortunately, this upsampled activation map is partially synthetic because wrap-around padding was applied during preprocessing to achieve a uniform, model-compatible voxel length $L$ across subjects. This padding introduces artificial "fake" voxels. The advantage of the wrap-around padding strategy is that it allows us to trace the origins of these fake voxels. To restore the true voxel activation map, we retain the activations of the real voxels, while for the fake voxels, we trace their origins and assign their values as the maximum of the original and artificial activations. This approach ensures that the restored activation map accurately represents the contributions of real voxels while mitigating the impact of synthetic padding. + +Finally, this process is repeated across all samples containing the object of interest, enabling us to investigate voxel semantic selectivity, as illustrated in Figure 3 of the main paper. To achieve the 3D activation like Figure 4 in the main paper, we can just map each flattened voxel back to 3D brain space using the provided nsdgeneral atlas. This methodology allows us to map neural activations back to their voxel-level origins, providing insights into the relationship between neural representations and object recognition. + +# A.9 Technical Details on Figure 6 + +We first computed the pixel occupation ratio for all training images containing the object chair. This ratio was derived by dividing the number of chair pixels (using MS-COCO annotations for masking) by the total image resolution. Since the raw pixel ratios exhibited a highly skewed distribution, we applied a log transformation to approximate a normal-like distribution as shown in Figure A.9. + +Next, we calculated the mean $\mu$ and standard deviation $\sigma$ of the log-transformed ratios. To partition the chairs into size-based categories, we defined three intervals: + +- Small chairs: $(-\infty, \mu - 0.5\sigma)$ +Medium chairs: $(\mu -0.5\sigma ,\mu +0.5\sigma)$ +Large chairs: $(\mu +0.5\sigma ,0)$ + +![](images/b2167f901021652094b9ebab126f53ecd0459571663202dac1e549782a6ea4a8.jpg) +Figure 8: Chair pixel ratio distribution in training set + +![](images/1766dc2ececf5d1b760e46df988f20d944a2115167be2df9a65f6dd632114564.jpg) + +Finally, for each subject, we generated a boxplot of the predicted probabilities for the chair class, stratified by these size groups. + +# A.10 Technical Details on Figure 7 + +To analyze differences in subjects' sensitivity to the object cup, we first computed the pixel occupation ratio for all training images containing cup and calculated the baseline MCC subject by subject as the pre-adjusted sensitivity measure. Since each subject viewed distinct images during training, we accounted for distribution shifts in both the size and frequency of cup appearances across subjects. To ensure a fair comparison, we adjusted the MCC by normalizing it with the subject-wise average pixel ratio. The resulting weighted MCC is visualized in Figure 7. + +# A.11 More Analytical and Visual Results + +# A.11.1 Voxel Sensitivity + +To examine voxel sensitivities, we analyzed the mean (x-axis) and standard deviation (y-axis) of voxel activations across five brain ROIS for objects bench and chair, as presented in Figure A.11.1. The results indicate a quadratic relationship in voxel sensitivity across these regions, allowing us to classify voxels into three distinct groups based on their mean activation and variability (standard deviation). Each group reflects a unique role in semantic decoding within visual regions: + +- Bystanders – This group, characterized by the lowest mean and standard deviation, consists of voxels that consistently contribute minimal information to the semantic decoding of visual stimuli. These voxels are either not responsible for distinguishing specific objects (bench and chair) or likely located in regions less involved in object discrimination, and instead providing generalized but stable responses across diverse stimuli. +- Discriminators – This higher mean and the highest standard deviation group includes voxels that show selective, highly variable responses, playing a key role in differentiating between features and supporting object-specific sensitivity. These voxels likely drive the flexibility needed for nuanced and accurate decoding of semantic information in visual stimuli. +- Supporters - The highest mean, low standard deviation voxels, characterized by strong, consistent activation, likely represent core object features and provide a stable foundation for robust and invariant decoding across subjects. They likely provide stable, foundational support for correctly classifying objects across different conditions. + +These findings suggest that voxel sensitivity patterns vary across the visual hierarchy, with each group contributing distinct vision information-processing roles in object recognition in the brain. + +# A.11.2 Single-subject 1D Activation Pattern + +Similar to Figure 3 in the main paper, we provide more visualization results on 1D Object-Voxel activation below: + +![](images/431340dc461bc3cdc6d16e2b91fbce2e64deb70dd3bb7576c04e8f9b2f81b3e0.jpg) +Figure 9: Object-Voxel activations (std v.s. mean) by vision ROIs for subj01 + +![](images/9b396ead8225798dfa3fe1249d14b24100678d3be1122800caaf3e9ff37628ff.jpg) +Figure 10: 1D Object-Voxel activations by brain vision ROIs for subj01. + +![](images/6e35ba877bc149c70d205ba014fd955d7de4995dbbe5f05681dd2aa430511a53.jpg) +Figure 11: 1D Object-Voxel activations by brain vision ROIs for subj02. + +![](images/8eaafeb6ec3e98f2c15462b183cd4ad29e4f6fb4abef3befdc97fe0336adc8b3.jpg) +Figure 12: 1D Object-Voxel activations by brain vision ROIs for subj03. + +![](images/4979e6bd3053c5c8b21f75ef46531d4fdd06f2fca2b0d96ff1c24de170c829bc.jpg) +Figure 13: 1D Object-Voxel activations by brain vision ROIs for subj04. + +![](images/4a98c572ad4747aa49f13fad4b13cea7f2d213fc2fce073d084911abfeaa06a2.jpg) +Figure 14: 1D Object-Voxel activations by brain vision ROIs for subj05. + +![](images/9d7e7fc1d785f4a7ddac20dd70c4605ca587890565236b639feb743ca57fd6d9.jpg) +Figure 15: 1D Object-Voxel activations by brain vision ROIs for subj06. + +![](images/12408e842dc6512af722b135f78fbb07d9d70f3f11c344c1b3d19c5add4e761a.jpg) +Figure 16: 1D Object-Voxel activations by brain vision ROIs for subj07. + +![](images/54d85426c013d10d97f593a0ed10933cb70fb278e175b46fa0ace914dba0ff37.jpg) +Figure 17: 1D Object-Voxel activations by brain vision ROIs for subj08. + +# A.11.3 Cross-subject 3D Activation Pattern + +Similar to Figure 4 in the main paper, we provide more visualization results on 3D Object-Voxel activation below: + +![](images/9a89e8749e429b41e435a1a09e472bc58b6feb393dd0ecb1ef395f1ba252dd4b.jpg) +Figure 18: 3D Object-Voxel activations of person, bicycle, car, and motorcycle for subj01, subj03, subj04, and subj05. + +![](images/94961cdf458067f19b37082a3949c8dbfca59c5f4f08201547d9dbcfa254a562.jpg) +Figure 19: 3D Object-Voxel activations of airplane, train, truck, and boat for subj01, subj03, subj04, and subj05. + +![](images/59576f9c90210b6e8432606346ad0d54b58ec7710670559aba2b57de3ad33520.jpg) +Figure 20: 3D Object-Voxel activations of boat, traffic light, fire hydrant, and stop sign for subj01, subj03, subj04, and subj05. + +![](images/ccd6f47265f525f5b5692c53f148c42c382a917cc5eb5aa06c5600c2e1e8c3fd.jpg) +Figure 21: 3D Object-Voxel activations of bench, bird, cat, and dog for subj01, subj03, subj04, and subj05. + +![](images/e342b1bd18506b545d47f785a327347c740b9745863cb488653fc13771d913d6.jpg) +Figure 22: Variations in subjects' attention to different objects. Four objects: person, cup, chair, and dining table are selected for visualization. + +# A.11.4 Variations in Subject Attention + +Similar to Figure 4 in the main paper, we provide more visualization results on variations in subjects' attention to different objects in the same image. The leftmost image shows the visual stimulus. Plots in the second column represent the shared attention across all subjects, and the remaining eight columns show the residual, subject-specific attention alongside predicted probabilities to compare recognition confidence and priority. + +![](images/a54e328f752dbcfd69979de5bc4844f2679ab5d37f61e1b639a4b27c4c9443e6.jpg) + +![](images/4e067e986af02eb3182cd22367ab73ccd334b9fa6f18ba2dc1181947b731c5cc.jpg) + +![](images/4510c8d5d3f9c822967ceaf0a56d183cf26686bf695003ca36db7bcc210a33f1.jpg) + +![](images/30539ff381be91a126c7624e2e1d521e029e9dc63ca1e1ae80ffbd67ac9eb151.jpg) + +![](images/304354d361522db9b8b80c63f317cd4e93e1f83a01607533308e9c1568f78774.jpg) + +![](images/3028dc4b2b06434381811410dc0b498532169d36350b38b1cee651042dab38a0.jpg) + +![](images/c5a279e1a5f5c1d54d19d82a07890fb081e654ba6acff44952560c93f881c3e3.jpg) + +![](images/f78d523a7458ae1ddd71b7ff32e4f4247f6684ade652c72c4c669cb9dbf14482.jpg) + +![](images/fba7000d8a16449646856b1bbf2ce7d4afa215b9edee6ea52fc005e04d57a0e4.jpg) + +![](images/4a300e93328cdb41f521ccee16a232e80e48feef950bcc45ff87053cb7fe7a44.jpg) + +![](images/f7f73ab39c01b132536c17d7e72205531269eb31816b8dcffda08abec0a7ba1a.jpg) + +![](images/9a155bfe257c00eac66618b3890af46f2e581c60494b6737a57180abcd8494e3.jpg) + +![](images/bae039647c59744b46fb56c7098e7565be678d461b1e108a648157f25abb4e91.jpg) +Figure 23: Variations in subjects' attention to different objects. Four objects: cup, fork, knife, and dining table are selected for visualization. + +![](images/f81a1f008eb50760a65378f72a368b2b75a491c5f522f5b1528107f341c39326.jpg) + +![](images/2148daf106728fcf06500c6c004acf9a97b1988e8347e499d13b619bf3a5902c.jpg) + +![](images/8b11be7e2c5008039fa90a718758fd518cad45609f2e3ad52b6a7ea9db1a7378.jpg) + +![](images/4230b743a806da050d3fb000771436b7443159da0dd43d25aabad0f603f4b32f.jpg) + +![](images/2321cfb58fb1d2c8991502312fa6bdac74a89d62480438bc592eea3fd2ff70ab.jpg) + +![](images/06dbeeb0c89698cec872ff6294c8646d0f2e7fe268fe9e4befdde487ae079484.jpg) +Figure 24: Variations in subjects' attention to different objects. Three objects: person, car, and truck are selected for visualization. + +![](images/b972ea4a052e70f50e7cebeed6c69c29d1f78ced48c45b73f193b3460718e79d.jpg) +Figure 25: Variations in subjects' attention to different objects. Three objects: person, sports ball, and tennis racket are selected for visualization. \ No newline at end of file diff --git a/NeurIPS/2025/$i$MIND_ Insightful Multi-subject Invariant Neural Decoding/images.zip b/NeurIPS/2025/$i$MIND_ Insightful Multi-subject Invariant Neural Decoding/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..8a2810e1f458b4c63d7602d166309670e7d72ae7 --- /dev/null +++ b/NeurIPS/2025/$i$MIND_ Insightful Multi-subject Invariant Neural Decoding/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:959dc37f980cb450779e32c1d268973bfa5f9d7bf093c409c25451b50133048a +size 3095509 diff --git a/NeurIPS/2025/$i$MIND_ Insightful Multi-subject Invariant Neural Decoding/layout.json b/NeurIPS/2025/$i$MIND_ Insightful Multi-subject Invariant Neural Decoding/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..6727ffbebee9969b5d0de5597136f88161df90dd --- /dev/null +++ b/NeurIPS/2025/$i$MIND_ Insightful Multi-subject Invariant Neural Decoding/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:82ab56bc17b86e0638d7b8b7188000dfabbf75ef0faa8ad7ac4f160a546e318d +size 1042772 diff --git a/NeurIPS/2025/1000 Layer Networks for Self-Supervised RL_ Scaling Depth Can Enable New Goal-Reaching Capabilities/4a8d9df5-287c-441c-8654-78be23752307_content_list.json b/NeurIPS/2025/1000 Layer Networks for Self-Supervised RL_ Scaling Depth Can Enable New Goal-Reaching Capabilities/4a8d9df5-287c-441c-8654-78be23752307_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..2ef9405fc92ef8377c94c8c23d5fda5af0afb09c --- /dev/null +++ b/NeurIPS/2025/1000 Layer Networks for Self-Supervised RL_ Scaling Depth Can Enable New Goal-Reaching Capabilities/4a8d9df5-287c-441c-8654-78be23752307_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:67afaac15c3b90e135b45650e0ca276e19965cec5ec542cdda4f4f9bfdeeefd4 +size 164834 diff --git a/NeurIPS/2025/1000 Layer Networks for Self-Supervised RL_ Scaling Depth Can Enable New Goal-Reaching Capabilities/4a8d9df5-287c-441c-8654-78be23752307_model.json b/NeurIPS/2025/1000 Layer Networks for Self-Supervised RL_ Scaling Depth Can Enable New Goal-Reaching Capabilities/4a8d9df5-287c-441c-8654-78be23752307_model.json new file mode 100644 index 0000000000000000000000000000000000000000..a57eac600a3784bbc007734b5779f12dc850ba64 --- /dev/null +++ b/NeurIPS/2025/1000 Layer Networks for Self-Supervised RL_ Scaling Depth Can Enable New Goal-Reaching Capabilities/4a8d9df5-287c-441c-8654-78be23752307_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4a852a85e6ecc2358ba7089e00fc61ea69af7bc05b718e58f038ccdf4955828f +size 206763 diff --git a/NeurIPS/2025/1000 Layer Networks for Self-Supervised RL_ Scaling Depth Can Enable New Goal-Reaching Capabilities/4a8d9df5-287c-441c-8654-78be23752307_origin.pdf b/NeurIPS/2025/1000 Layer Networks for Self-Supervised RL_ Scaling Depth Can Enable New Goal-Reaching Capabilities/4a8d9df5-287c-441c-8654-78be23752307_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..008e434cc960c91263492601dbf8ec208a8edafa --- /dev/null +++ b/NeurIPS/2025/1000 Layer Networks for Self-Supervised RL_ Scaling Depth Can Enable New Goal-Reaching Capabilities/4a8d9df5-287c-441c-8654-78be23752307_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:504fc52ac9fa74457f9de82d143b39c762deaee8b5d8eb790406b4a3efea1342 +size 4312939 diff --git a/NeurIPS/2025/1000 Layer Networks for Self-Supervised RL_ Scaling Depth Can Enable New Goal-Reaching Capabilities/full.md b/NeurIPS/2025/1000 Layer Networks for Self-Supervised RL_ Scaling Depth Can Enable New Goal-Reaching Capabilities/full.md new file mode 100644 index 0000000000000000000000000000000000000000..b3d0471f6a0e87de2ad9598ed522a28367d12dc6 --- /dev/null +++ b/NeurIPS/2025/1000 Layer Networks for Self-Supervised RL_ Scaling Depth Can Enable New Goal-Reaching Capabilities/full.md @@ -0,0 +1,786 @@ +# 1000 Layer Networks for Self-Supervised RL: Scaling Depth Can Enable New Goal-Reaching Capabilities + +Kevin Wang + +Princeton University + +kw6487@princeton.edu + +Ishaan Javali + +Princeton University + +ijavali@princeton.edu + +Michał Bortkiewicz + +Warsaw University of Technology + +michal.bortkiewicz.dokt@pw.edu.pl + +Tomasz Trzciński + +Warsaw University of Technology, + +Tooploox, IDEAS Research Institute + +Benjamin Eysenbach + +Princeton University + +eysenbach@princeton.edu + +# Abstract + +Scaling up self-supervised learning has driven breakthroughs in language and vision, yet comparable progress has remained elusive in reinforcement learning (RL). In this paper, we study building blocks for self-supervised RL that unlock substantial improvements in scalability, with network depth serving as a critical factor. Whereas most RL papers in recent years have relied on shallow architectures (around 2 - 5 layers), we demonstrate that increasing the depth up to 1024 layers can significantly boost performance. Our experiments are conducted in an unsupervised goal-conditioned setting, where no demonstrations or rewards are provided, so an agent must explore (from scratch) and learn how to maximize the likelihood of reaching commanded goals. Evaluated on simulated locomotion and manipulation tasks, our approach increases performance on the self-supervised contrastive RL algorithm by $2 \times -50 \times$ , outperforming other goal-conditioned baselines. Increasing the model depth not only increases success rates but also qualitatively changes the behaviors learned. The project webpage and code can be found here: https://wang-kevin3290.github.io/scaling-crl/. + +# 1 Introduction + +While scaling model size has been an effective recipe in many areas of machine learning, its role and impact in reinforcement learning (RL) remain unclear. The typical model size for state-based RL tasks is between 2 to 5 layers (Raffin et al., 2021; Huang et al., 2022). In contrast, it is not uncommon to use very deep networks in other domain areas; Llama 3 (Dubey et al., 2024) and Stable Diffusion 3 (Esser et al., 2024) have hundreds of layers. In fields such as vision (Radford et al., 2021; Zhai et al., 2021; Dehghani et al., 2023) and language (Srivastava et al., 2023), models often only acquire the ability to solve certain tasks once they are larger than a critical scale. In the RL setting, many researchers have searched for similar emergent phenomena (Srivastava et al., 2023), but these papers typically report only small marginal benefits and typically only on tasks where small models already achieve some degree of success (Nauman et al., 2024b; Lee et al., 2024; Farebrother et al., 2024). A key open question in RL today is whether it is possible to achieve similar jumps in performance by scaling RL networks. + +![](images/8b6d70bfe50bfa266f3b46590be162e35e0925588f97eb4eb80ba8aa9ee9f240.jpg) + +![](images/2f13fc85a7133990a216e3c4210820663f4f68d4646797dbf8e8ae4f31419b99.jpg) + +![](images/5603eb5508ad076b5f668ae567dc3b008de582a8f6dd0c951aaf85c1246a9cec.jpg) + +![](images/db2639db5c38779e24abe7e4c23da44cc20457aac8a7ab3455a008db2c70d176.jpg) + +![](images/6fa117cacfa0ca26a66b10bba61309d85a812279298819c901f7f105bf0d9c9c.jpg) + +![](images/e21c5fdb92bcf65843fb4d3cffbc9dd4ed8d1fd4f3d93d5a0fddcedbf1f56c89.jpg) +Figure 1: Scaling network depth yields performance gains across a suite of locomotion, navigation, and manipulation tasks, ranging from doubling performance to $50 \times$ improvements on Humanoid-based tasks. Notably, rather than scaling smoothly, performance often jumps at specific critical depths (e.g., 8 layers on Ant Big Maze, 64 on Humanoid U-Maze), which correspond to the emergence of qualitatively distinct policies (see Section 4). + +![](images/96f794a0c415b90cab83b2ed5357a6d771ae9773b9f29eba0e539dce713f8a78.jpg) + +![](images/06fd295488124f1bf86c8b0e1e382d03e4ab425c0e298aaf341723de47145355.jpg) + +![](images/e47e2e7ddfad95ff4e362a767406c4edfe224784b9251d6f78534304e91e0186.jpg) + +![](images/c8b27a055fde43961292b2f97a91b90a64480f8f5394968fd88fe70a1f28c0c9.jpg) + +![](images/7f24fd41864d7df09ae2aaeb0eaaf472871092cd6e908d1cb4baebcba6a4ae13.jpg) + +At first glance, it makes sense why training very large RL networks should be difficult: the RL problem provides very few bits of feedback (e.g., only a sparse reward after a long sequence of observations), so the ratio of feedback to parameters is very small. The conventional wisdom (LeCun, 2016), reflected in many recent models (Radford, 2018; Chen et al., 2020; Goyal et al., 2019), has been that large AI systems must be trained primarily in a self-supervised fashion and that RL should only be used to finetune these models. Indeed, many of the recent breakthroughs in other fields have been primarily achieved with self-supervised methods, whether in computer vision (Caron et al., 2021; Radford et al., 2021; Liu et al., 2024), NLP (Srivastava et al., 2023), or multimodal learning (Zong et al., 2024). Thus, if we hope to scale reinforcement learning methods, self-supervision will likely be a key ingredient. + +In this paper, we will study building blocks for scaling reinforcement learning. Our first step is to rethink the conventional wisdom above: "reinforcement learning" and "self-supervised learning" are not diametric learning rules, but rather can be married together into self-supervised RL systems that explore and learn policies without reference to a reward function or demonstrations (Eysenbach et al., 2021, 2022; Lee et al., 2022). In this work, we use one of the simplest self-supervised RL algorithms, contrastive RL (CRL) (Eysenbach et al., 2022). The second step is to recognize the importance of increasing data availability. We will do this by building on recent GPU-accelerated RL frameworks (Makoviychuk et al., 2021; Rutherford et al., 2023; Rudin et al., 2022; Bortkiewicz et al., 2024). The third step is to increase network depth, using networks that are up to $100 \times$ deeper than those typically found in prior work. Stabilizing the training of such networks will require incorporating architectural techniques from prior work, including residual connections (He et al., 2015), layer normalization (Ba et al., 2016), and Swish activation (Ramachandran et al., 2018). Our experiments will also study the relative importance of batch size and network width. + +The primary contribution of this work is to show that a method that integrates these building blocks into a single RL approach exhibits strong scalability: + +- Empirical Scalability: We observe a significant performance increase, more than $20 \times$ in half of the environments and outperforming other standard goal-conditioned baselines. These performance gains correspond to qualitatively distinct policies that emerge as the scale increases. +- Scaling Depth in Network Architecture: While many prior RL works have primarily focused on increasing network width, they often report limited or even negative returns + +when expanding depth (Lee et al., 2024; Nauman et al., 2024b). In contrast, our approach unlocks the ability to scale along the axis of depth, yielding performance improvements that surpass those from scaling width alone (see Sec. 4). + +- Empirical Analysis: We conduct an extensive analysis of the key components in our scaling approach, uncovering critical factors and offering new insights. + +We anticipate that future research may build on this foundation by uncovering additional building blocks. + +# 2 Related Work + +Natural Language Processing (NLP) and Computer Vision (CV) have recently converged in adopting similar architectures (i.e. transformers) and shared learning paradigms (i.e self-supervised learning), which together have enabled transformative capabilities of large-scale models (Vaswani et al., 2017; Srivastava et al., 2023; Zhai et al., 2021; Dehghani et al., 2023; Wei et al., 2022). In contrast, achieving similar advancements in reinforcement learning (RL) remains challenging. Several studies have explored the obstacles to scaling large RL models, including parameter underutilization (Obando-Ceron et al., 2024), plasticity and capacity loss (Lyle et al., 2024, 2022), data sparsity (Andrychowicz et al., 2017; LeCun, 2016), and training instabilities (Ota et al., 2021; Henderson et al., 2018; Van Hasselt et al., 2018; Nauman et al., 2024a). As a result, current efforts to scale RL models are largely restricted to specific problem domains, such as imitation learning (Tuyls et al., 2024), multi-agent games (Neumann and Gros, 2022), language-guided RL (Driess et al., 2023; Ahn et al., 2022), and discrete action spaces (Obando-Ceron et al., 2024; Schwarzer et al., 2023). + +Recent approaches suggest several promising directions, including new architectural paradigms (Obando-Ceron et al., 2024), distributed training approaches (Ota et al., 2021; Espeholt et al., 2018), distributional RL (Kumar et al., 2023), and distillation (Team et al., 2023). Compared to these approaches, our method makes a simple extension to an existing self-supervised RL algorithm. The most recent works in this vein include Lee et al. (2024) and Nauman et al. (2024b), which leverage residual connections to facilitate the training of wider networks. These efforts primarily focus on network width, noting limited gains from additional depth, thus both works use architectures with only four MLP layers. In our method, we find that scaling width indeed improves performance (Section 4.4); however, our approach also enables scaling along depth, proving to be more powerful than width alone. + +One notable effort to train deeper networks is described by Farebrother et al. (2024), who cast value-based RL into a classification problem by discretizing the TD objective into a categorical cross-entropy loss. This approach draws on the conjecture that classification-based methods can be more robust and stable and thus may exhibit better scaling properties than their regressive counterparts (Torgo and Gama, 1996; Farebrother et al., 2024). The CRL algorithm that we use effectively uses a cross-entropy loss as well (Eysenbach et al., 2022). Its InfoNCE objective is a generalization of the cross-entropy loss, thereby performing RL tasks by effectively classifying whether current states and actions belong to the same or different trajectory that leads toward a goal state. In this vein, our work serves as a second piece of evidence that classification, much like cross-entropy's role in the scaling success in NLP, could be a potential building block in RL. + +# 3 Preliminaries + +This section introduces notation and definitions for goal-conditioned RL and contrastive RL. Our focus is on online RL, where a replay buffer stores the most recent trajectories, and the critic is trained in a self-supervised manner. + +Goal-Conditioned Reinforcement Learning We define a goal-conditioned MDP as tuple $\mathcal{M}_g = (\mathcal{S},\mathcal{A},p_0,p,p_g,r_g,\gamma)$ , where the agent interacts with the environment to reach arbitrary goals (Kaelbling, 1993; Andrychowicz et al., 2017; Blier et al., 2021). At every time step $t$ , the agent observes state $s_t\in S$ and performs a corresponding action $a_{t}\in \mathcal{A}$ . The agent starts interaction in states sampled from $p_0(s_0)$ , and the interaction dynamics are defined by the transition probability distribution $p(s_{t + 1}\mid s_t,a_t)$ . Goals $g\in \mathcal{G}$ are defined in a goal space $\mathcal{G}$ , which is related to $\mathcal{S}$ via a mapping $f:\mathcal{S}\to \mathcal{G}$ . For example, $\mathcal{G}$ may correspond to a subset of state dimensions. The prior distribution + +over goals is defined by $p_{g}(g)$ . The reward function is defined as the probability density of reaching the goal in the next time step $r_{g}(s_{t},a_{t})\triangleq (1 - \gamma)p(s_{t + 1} = g\mid s_{t},a_{t})$ , with discount factor $\gamma$ . + +In this setting, the goal-conditioned policy $\pi(a \mid s, g)$ receives both the current observation of the environment as well as a goal. We define the discounted state visitation distribution as $p_{\gamma}^{\pi(\cdot \mid \cdot, g)}(s) \triangleq (1 - \gamma) \sum_{t=0}^{\infty} \gamma^{t} p_{t}^{\pi(\cdot \mid \cdot, g)}(s)$ , where $p_{t}^{\pi}(s)$ is the probability that policy $\pi$ visits $s$ after exactly $t$ steps, when conditioned with $g$ . This last expression is precisely the $Q$ -function of the policy $\pi(\cdot \mid \cdot, g)$ for the reward $r_{g} \colon Q_{g}^{\pi}(s, a) \triangleq p_{\gamma}^{\pi(\cdot \mid \cdot, g)}(g \mid s, a)$ . The objective is to maximize the expected reward: + +$$ +\max _ {\pi} \mathbb {E} _ {p _ {0} (s _ {0}), p _ {g} (g), \pi (\cdot | \cdot , g)} \left[ \sum_ {t = 0} ^ {\infty} \gamma^ {t} r _ {g} \left(s _ {t}, a _ {t}\right) \right]. \tag {1} +$$ + +Contrastive Reinforcement Learning. Our experiments will use the contrastive RL algorithm (Eysenbach et al., 2022) to solve goal-conditioned problems. Contrastive RL is an actor-critic method; we will use $f_{\phi, \psi}(s, a, g)$ to denote the critic and $\pi_{\theta}(a \mid s, g)$ to denote the policy. The critic is parametrized with two neural networks that return state, action pair embedding $\phi(s, a)$ and goal embedding $\psi(g)$ . The critic's output is defined as the $l^2$ -norm between these embeddings: $f_{\phi, \psi}(s, a, g) = \| \phi(s, a) - \psi(g) \|_2$ . The critic is trained with the InfoNCE objective (Sohn, 2016) as in previous works (Eysenbach et al., 2022, 2021; Zheng et al., 2023, 2024; Myers et al., 2024; Bortkiewicz et al., 2024). Training is conducted on batches $\mathcal{B}$ , where $s_i, a_i, g_i$ represent the state, action, and goal (future state) sampled from the same trajectory, while $g_j$ represents a goal sampled from a different, random trajectory. The objective function is defined as: + +$$ +\min _ {\phi , \psi} \mathbb {E} _ {\mathcal {B}} \left[ - \sum_ {i = 1} ^ {| \mathcal {B} |} \log \left(\frac {e ^ {f _ {\phi , \psi} \left(s _ {i} , a _ {i} , g _ {i}\right)}}{\sum_ {j = 1} ^ {K} e ^ {f _ {\phi , \psi} \left(s _ {i} , a _ {i} , g _ {j}\right)}}\right) \right]. +$$ + +The policy $\pi_{\theta}(a\mid s,g)$ is trained to maximize the critic: + +$$ +\max_{\pi_{\theta}}\mathbb{E}_{\substack{p_{0}(s_{0}),p(s_{t + 1}|s_{t},a_{t}),\\ p_{g}(g),\pi_{\theta}(a|s,g)}}\left[f_{\phi ,\psi}(s,a,g)\right]. +$$ + +Residual Connections We incorporate residual connections (He et al., 2015) into our architecture, following their successful use in RL (Farebrother et al., 2024; Lee et al., 2024; Nauman et al., 2024b). A residual block transforms a given representation $\mathbf{h}_i$ by adding a learned residual function $F_{i}(\mathbf{h}_{i})$ to the original representation. Mathematically, this is expressed as: + +$$ +\mathbf {h} _ {i + 1} = \mathbf {h} _ {i} + F _ {i} \left(\mathbf {h} _ {i}\right) +$$ + +where $\mathbf{h}_{i + 1}$ is the output representation, $\mathbf{h}_i$ is the input representation, and $F_{i}(\mathbf{h}_{i})$ is a transformation learned through the network (e.g., using one or more layers). The addition ensures that the network learns modifications to the input rather than entirely new transformations, helping to preserve useful features from earlier layers. Residual con + +![](images/5f4a6b3971d486a087ec2675951571bedebf16fb56f825f59b05c9b34895afa2.jpg) +Figure 2: Architecture. Our approach integrates residual connections into both the actor and critic networks of the Contrastive RL algorithm. The depth of this residual architecture is defined as the total number of Dense layers across the residual blocks, which, with our residual block size of 4, equates to $4N$ . + +nections improve gradient propagation by introducing shortcut paths (He et al., 2016; Veit et al., 2016), enabling more effective training of deep models. + +# 4 Experiments + +# 4.1 Experimental Setup + +**Environments.** All RL experiments use the JaxGCRL codebase (Bortkiewicz et al., 2024), which facilitates fast online GCRL experiments based on Brax (Freeman et al., 2021) and MJX (Todorov + +et al., 2012) environments. The specific environments used are a range of locomotion, navigation, and robotic manipulation tasks, for details see Appendix B. We use a sparse reward setting, with $r = 1$ only when the agent is in the goal proximity. For evaluation, we measure the number of time steps (out of 1000) that the agent is near the goal. When reporting an algorithm's performance as a single number, we compute the average score over the last five epochs of training. + +Architectural Components We employ residual connections from the ResNet architecture (He et al., 2015), with each residual block consisting of four repeated units of a Dense layer, a Layer Normalization (Ba et al., 2016) layer, and Swish activation (Ramachandran et al., 2018). We apply the residual connections immediately following the final activation of the residual block, as shown in Figure 2. In this paper, we define the depth of the network as the total number of Dense layers across all residual blocks in the architecture. In all experiments, the depth refers to the configuration of the actor network and both critic encoder networks, which are scaled jointly, except for the ablation experiment in Section 4.4. + +# 4.2 Scaling Depth in Contrastive RL + +We start by studying how increasing network depth can increase performance. Both the JaxGCRL benchmark and relevant prior work (Lee et al., 2024; Nauman et al., 2024b; Zheng et al., 2024) use MLPs with a depth of 4, and as such we adopt it as our baseline. In contrast, we will study networks of depth 8, 16, 32, and 64. The results in Figure 1 demonstrate that deeper networks achieve significant performance improvements across a diverse range of locomotion, navigation, and manipulation tasks. Compared to the 4-layer models typical in prior work, deeper networks achieve $2 - 5 \times$ gains in robotic manipulation tasks, over $20 \times$ gains in long-horizon maze tasks such as Ant U4-Maze and Ant U5-Maze, and over $50 \times$ gains in humanoid-based tasks. The full table of performance increases up to depth 64 is provided in Table 1. + +In Figure 12, we present results the same 10 environments, but compared against SAC, SAC+HER, TD3+HER, GCBC, and GCSL. Scaling CRL leads to substantial performance improvements, outperforming all other baselines in 8 out of 10 tasks. The only exception is SAC on the Humanoid Maze environments, where it exhibits greater sample efficiency early on; however, scaled CRL eventually reaches comparable performance. These results highlight that scaling the depth of the CRL algorithm enables state-of-the-art performance in goal-conditioned reinforcement learning. + +# 4.3 Emergent Policies Through Depth + +A closer examination of the results from the performance curves in Figure 1 reveals a notable pattern: instead of a gradual improvement in performance as depth increases, there are pronounced jumps that occur once a critical depth threshold is reached (also shown in Figure 5). The critical depths vary by environment, ranging from 8 layers (e.g. Ant Big Maze) to 64 layers in the Humanoid U-Maze task, with further jumps occurring even at depths of 1024 layers (see the Testing Limits section, Section 4.4). + +Prompted by this observation, we visualized the learned policies at various depths and found qualitatively distinct skills and behaviors exhibited. This is particularly pronounced in the humanoid-based tasks, as illustrated in Figure 3. Networks with a depth of 4 exhibit rudimentary policies where the agent either falls or throws itself toward the target. Only at a critical depth + +of 16 does the agent develop the ability to walk upright into the goal. In the Humanoid U-Maze environment, networks of depth 64 struggle to navigate around the intermediary wall, collapsing on the ground. Remarkably at a depth of 256, the agent learns unique behaviors on Humanoid U-Maze. These behaviors include folding forward into a leveraged position to propel itself over walls and + +![](images/701d892a5239ccfb9506d491f5e7bb2099ec5a4e6252998a00ea23ff64bede3f.jpg) +Figure 3: Increasing depth results in new capabilities: Row 1: A depth-4 agent collapses and throws itself toward the goal. Row 2: A depth-16 agent walks upright. Row 3: A depth-64 agent struggles and falls. Row 4: A depth-256 agent vaults the wall acrobatically. + +![](images/ef97b398075c57486ab06de471206ff2d05ae47770fd7a42699554c3b5e23b65.jpg) +Figure 5: Critical depth and residual connections. Incrementally increasing depth results in marginal performance gains (left). However, once a critical threshold is reached, performance improves dramatically (right) for networks with residual connections. + +![](images/e9704e0ef79a526ab246089b8a586b1789c694c4467846dd54a12f5a818c0a9c.jpg) +Figure 6: Actor vs. Critic. In Arm Push Easy, scaling the critic is more effective; in Ant Big Maze, the actor matters more. For Humanoid, scaling both is necessary. These results suggest that actor and critic scaling can complement each other for CRL. + +shifting into a seated posture over the intermediary obstacle to worm its way toward the goal (one of these policies is illustrated in the fourth row of Figure 3). To the best of our knowledge, this is the first goal-conditioned approach to document such behaviors on the humanoid environment. + +# 4.4 What Matters for CRL Scaling + +Width vs. Depth Past literature has shown that scaling network width can be effective (Lee et al., 2024; Nauman et al., 2024b). In Figure 4, we find that scaling width is also helpful in our experiments: wider networks consistently outperform narrower networks (depth held constant at 4). However, depth seems to be a more effective axis for scaling: simply doubling the depth to 8 (width held constant at 256) outperforms the widest networks in all three environments. The advantage of depth scaling is most pronounced in the Humanoid environment (observation dimension 268), followed by Ant Big Maze (dimension 29) and Arm Push Easy (dimension 17), suggesting that the comparative benefit may increase with higher observation dimensionality. + +Note additionally that the parameter count scales linearly with width but quadratically with depth. For comparison, a network with 4 MLP layers + +and 2048 hidden units has roughly $35\mathrm{M}$ parameters, while one with a depth of 32 and 256 hidden units has only around 2M. Therefore, when operating under a fixed FLOP compute budget or specific memory constraints, depth scaling may be a more computationally efficient approach to improving network performance. + +Scaling the Actor vs. Critic Networks To investigate the role of scaling in the actor and critic networks, Figure 6 presents the final performance for various combinations of actor and critic depths across three environments. Prior work (Nauman et al., 2024b; Lee et al., 2024) focuses on scaling the critic network, finding that scaling the actor degrades performance. In contrast, while we do find that scaling the critic is more impactful in two of the three environments (Humanoid, Arm Push Easy), our method benefits from scaling the actor network jointly, with one environment (Ant Big Maze) demonstrating actor scaling to be more impactful. Thus, our method suggests that scaling both the actor and critic networks can play a complementary role in enhancing performance. + +Deep Networks Unlock Batch Size Scaling Scaling batch size has been well-established in other areas of machine learning (Chen et al., 2022; Zhang et al., 2024). However, this approach has not translated as effectively to reinforcement learning (RL), and prior work has even reported negative impacts on value-based RL (Obando-Ceron et al., 2023). Indeed, in our experiments, + +![](images/ca912580887f9d1e8f6001538b6887abf878d6a917e53749d9165fe0826da491.jpg) +Figure 4: Scaling network width vs. depth. Here, we reflect findings from previous works (Lee et al., 2024; Nauman et al., 2024b) which suggest that increasing network width can enhance performance. However, in contrast to prior work, our method is able to scale depth, yielding more impactful performance gains. For instance, in the Humanoid environment, raising the width to 2048 (depth=4) fails to match the performance achieved by simply doubling the depth to 8 (width=256). The comparative advantage of scaling depth is more pronounced as the observational dimensionality increases. + +eters, while one with a depth of 32 and 256 hidden voting under a fixed FLOP compute budget or specific are computationally efficient approach to improving + +![](images/d089124a4059d3d2e22682f2c9f776bc41ede1a030ba39d02e4482209c5b14dd.jpg) +Figure 7: Deeper networks unlock batch size scaling. We find that as depth increases from 4 to 64 in Humanoid, larger networks can effectively leverage batch size scaling to achieve further improvements. + +![](images/0830455da50529ff50d90a02e0573a39977572c35a45c43663bade3ceea6c2e8.jpg) + +![](images/4223380e9585423e46f0bf20188b90454ad6eac3b3d284669ae2704884e53817.jpg) + +![](images/d20629ad3c844c484299e3b30024cea98a5c26d67cb1cc9300278b72e0b3402f.jpg) + +![](images/2a860af1d67aa25c1a8c867b5fbac24046789b366cb96caefa2b6e7a753d6823.jpg) + +simply increasing the batch size for the original CRL networks yields only marginal differences in performance (Figure 7, top left). + +At first glance, this might seem counterintuitive: since reinforcement learning typically involves fewer informational bits per piece of training data (LeCun, 2016), one might expect higher variance in batch loss or gradients, suggesting the need for larger batch sizes to compensate. At the same time, this possibility hinges on whether the model in question can actually make use of a bigger batch size—in domains of ML where scaling has been successful, larger batch sizes usually bring the most benefit when coupled with sufficiently large models (Zhang et al., 2024; Chen et al., 2022). One hypothesis is that the small models traditionally used in RL may obscure the underlying benefits of larger batch size. + +To test this hypothesis, we study the effect of increasing the batch size for networks of varying depths. As shown in Figure 7, scaling the batch size becomes effective as network depth grows. This finding offers evidence that by scaling network capacity, we may simultaneously unlock the benefits of larger batch size, potentially making it an important component in the broader pursuit of scaling self-supervised RL. + +Training Contrastive RL with 1000+ Layers We next study whether further increasing depth beyond 64 layers further improves performance. We use the Humanoid maze tasks as these are both the most challenging environments in the benchmark and also seem to benefit from the deepest scaling. The results, shown in Figure 12, indicate that performance continues to substantially improve as network depth reaches 256 and 1024 layers in the Humanoid U-Maze environment. While we were unable to scale beyond 1024 layers due to computational constraints, we expect to see continued improvements with even greater depths, especially on the most challenging tasks. + +![](images/93d9021d02f10a5b54949f0ddfb6f6ef29ec218ce7c89ece721b4316060e19c4.jpg) + +![](images/54d6c98f0dc7b7c16b6b243a2259800b17c1ba963f9fabb6c242a9910858a0f6.jpg) + +![](images/03f3934d33f3321e16c568a9d12ee757b1dd780e65f0d86bab8bdb5c732e6f89.jpg) + +![](images/498f3f6f3081658ea9b286744b741c465be38770618dc8e0a6eedd5eeee4a8c8.jpg) +Figure 8: We disentangle the effects of exploration and expressivity on depth scaling by training three networks in parallel: a "collector," plus one deep and one shallow learner that train only from the collector's shared replay buffer. In all three environments, when using a deep collector (i.e. good data coverage), the deep learner outperforms the shallow learner, indicating that expressivity is crucial when controlling for good exploration. With a shallow collector (poor exploration), even the deep learner cannot overcome the limitations of insufficient data coverage. As such, the benefits of depth scaling arise from a combination of improved exploration and increased expressivity working jointly. + +# 4.5 Why Scaling Happens + +Depth Enhances Contrastive Representations The long-horizon setting has been a long-standing challenge in RL particularly in unsupervised goal-conditioned settings where there is no auxiliary reward feedback (Gupta et al., 2019). The family of U-Maze environments requires a global understanding of the maze layout for effective navigation. We consider a variant of the Ant U-Maze environment, the U4-maze, in which the agent must initially move in the direction opposite the goal to loop around and ultimately reach it. As shown in Figure 9, we observe a qualitative difference in the behavior of the shallow network (depth 4) compared to the deep network (depth 64). The visualized Q-values computed from the critic encoder representations reveal that the depth 4 network seemingly relies on Euclidean distance to the goal as a proxy for the Q value, even when a wall obstructs the direct path. In contrast, the depth 64 critic network learns richer representations, enabling it to effectively capture the topology of the maze as visualized by the trail of high Q values along the inner edge. These findings suggest that increasing network depth leads to richer learned representations, + +enabling deeper networks to better capture environment topology and achieve more comprehensive state-space coverage in a self-supervised manner. + +![](images/4ba2f275d4db91ccab249c6454fc3897ed5e1ab245f6b523d4bf49ae5f0cad7c.jpg) +Figure 9: Deeper Q-functions are qualitatively different. In the U4-Maze, the start and goal positions are indicated by the $\odot$ and $\mathbf{G}$ symbols respectively, and the visualized Q values are computed via the $L_{2}$ distance in the learned representation space, i.e., $Q(s,a,g) = \| \phi (s,a) - \psi (g)\| _2$ . The shallow depth 4 network (left) naively relies on Euclidean proximity, showing high Q values near the start despite a maze wall. In contrast, the depth 64 network (right) clusters high Q values at the goal, gradually tapering along the interior. + +Depth Enhances Exploration and Expressivity in a Synergized Way Our earlier results suggested that deeper networks achieve greater state-action coverage. To better understand why scaling works, we sought to determine to whether improved data alone explains the benefits of scaling, or whether it acts in conjunction with other factors. Thus, we designed an experiment in Figure 8 in which we train three networks in parallel: one network, the "collector," interacts with the environment and writes all experience to a shared replay buffer. Alongside it, two additional "learners", one deep and one shallow, train concurrently. Crucially, these two learners never collect their own data; they train only from the collector's buffer. This design holds the data distribution constant while varying the model's capacity, so any performance gap between the deep and shallow learners must come from expressivity rather than exploration. When the collector is deep (e.g., depth 32), across all three environments the deep learner substantially outperforms the shallow one across all three environments, indicating that the expressivity of the deep networks is critical. On the other hand, we repeat the experiment with shallow collectors (e.g., depth 4), which explores less effectively and therefore populates the buffer with low-coverage experience. Here, both the deep and shallow learners struggle and achieve similarly poor performance, which indicates that the deep network's additional capacity does not overcome the limitations of insufficient data coverage. As such, scaling depth enhances exploration and expressivity in a synergized way: stronger learning capacity drives more extensive exploration, and strong data coverage is essential to fully realize the power of stronger learning capacity. Both aspects jointly contribute to improved performance. + +Deep Networks Learn to Allocate Greater Representational Capacity to States Near the Goal In Figure 10 we take a successful trajectory in the Humanoid environment and visualize the embeddings of state-action encoder along this trajectory for both deep vs. shallow networks. While the shallow network (Depth 4) tends to cluster near-goal states tightly together, the deep network produces more "spread out" representations. This distinction is important: in a self-supervised setting, we want our representations to separate states that matter—particularly future or goal-relevant states—from random ones. As such, we want to allocate more representational capacity to such critical regions. This suggests that deep networks may learn to allocate representational capacity more effectively to state regions that matter most for the downstream task. + +![](images/c2bd040060058e03ae99a5da3656b88e9b45c88f92e5e23f6d093a48664acc00.jpg) +Successful Trajectory Path in Humanoid Env + +![](images/133fec0c43a418167ce315216ccfaf292ce805c42cfa9a9d2d069a49fee1cd81.jpg) +Trajectory in Embedding Space (Depth 4) + +![](images/f68960329e357b2e301b2fe64c9a3127d7797a53baf21050dc286d25b38fb1f1.jpg) +Trajectory in Embedding Space (Depth 64) +Figure 10: We visualize state-action embeddings from shallow (depth 4) and deep (depth 64) networks along a successful trajectory in the Humanoid task. Near the goal, embeddings from the deep network expand across a curved surface, while those from the shallow network form a tight cluster. This suggests that deeper networks may devote greater representational capacity to regions of the state space that are more frequently visited and play a more critical role in successful task completion. + +![](images/064cae44897ca9dba940701fcc3a75b4891bc56c1175543c5469f802b936f986.jpg) + +# Deeper Networks Enable Partial Experience + +Stitching Another key challenge in reinforcement learning is learning policies that can generalize to tasks unseen during training. To evaluate this setting, we designed a modified version of the Ant U-Maze environment. As shown in Figure 11 (top right), the original JaxGCRL benchmark assesses the agent's performance on the three farthest goal positions located on the opposite side of the wall. However, instead of training on all possible subgoals (a superset of the evaluation state-goal pairs), we modified the setup to train on start-goal pairs that are at most 3 units apart, ensuring that none of the evaluation pairs ever appear in the training set. Figure 11 demonstrates that depth 4 networks show limited generalization, solving only the easiest goal (4 units away from the start). Depth 16 networks achieve moderate success, while depth 64 networks excel, sometimes solving the most challenging goal position. These results suggest that the increasing network depth results in some degree of stitching, combining $\leq 3$ -unit pairs to navigate the 6-unit span of the U-Maze. + +The (CRL) Algorithm is Key In Appendix A, we show that scaled CRL outperforms other baseline goal-conditioned algorithms and advance the SOTA for goal-conditioned RL. We observe that for temporal difference methods + +![](images/f1f2d019cbd5925795ac76990a74b021edbc8f9b8b096ee736fa97e99e7226cf.jpg) + +![](images/63259bec9267556a0b9eec65671bbb6113c1636e3b24efebc5fc27ae10382b53.jpg) + +![](images/7fbfead07d2a0343f27c19808fc28429dbed6f12cfbc2ef6cb43a05fe8f3f262.jpg) +Figure 11: Deeper networks exhibit improved generalization. (Top left) We modify the training setup of the Ant U-Maze environment such that start-goal pairs are separated by $\leq 3$ units. This design guarantees that no evaluation pairs (Top right) were encountered during training, testing the ability for combinatorial generalization via stitching. (Bottom) Generalization ability improves as network depth grows from 4 to 16 to 64 layers. + +(SAC, SAC+HER, TD3+HER), the performance saturates for networks of depth 4, and there is either zero or negative performance gains from deeper networks. This is in line with previous research showing that these methods benefit mainly from width (Lee et al., 2024; Nauman et al., 2024b). These results suggest that the self-supervised CRL algorithm is critical. + +We also experiment with scaling more self-supervised algorithms, namely Goal-Conditioned Behavioral Cloning (GCBC) and Goal-Conditioned Supervised Learning (GCSL). While these methods yield zero success in certain environments, they show some utility in arm manipulation tasks. Interestingly, even a very simple self-supervised algorithm like GCBC benefits from increased depth. This + +points to a promising direction for future work of further investigating other self-supervised methods to uncover potentially different or complementary recipes for scaling self-supervised RL. + +Finally, recent work has augmented goal-conditioned RL with quasimetric architectures, leveraging the fact that temporal distances satisfy a triangle inequality-based invariance. In Appendix A, we also investigate whether the depth scaling effect persists when applied to these quasimetric networks. + +# 4.6 Does Depth Scaling Improve Offline Contrastive RL? + +In preliminary experiments, we evaluated depth scaling in the offline goal-conditioned setting using OGBench (Park et al., 2024). We found little evidence that increasing the network depth of CRL improves performance in this offline setting. To further investigate this, we conducted ablations: (1) scaling critic depth while holding the actor at 4 or 8 layers, and (2) applying cold initialization to the final layers of the critic encoders (Zheng et al., 2024). In all cases, baseline depth 4 networks often had the highest success. A key direction for future work is to see if our method can be adapted to enable scaling in the offline setting. + +# 5 Conclusion + +Arguably, much of the success of vision and language models today is due to the emergent capabilities they exhibit from scale (Srivastava et al., 2023), leading to many systems reducing the RL problem to a vision or language problem. + +A critical question for large AI models is: where does the data come from? Unlike supervised learning paradigms, RL methods inherently address this by jointly optimizing both the model and the data collection process through exploration. Ultimately, determining effective ways of building RL systems that demonstrate emergent capabilities may be important for transforming the field into one that trains its own large models. We believe that our work is a step towards these systems. By integrating key components for scaling up RL into a single approach, we show that model performance consistently improves as scale increases in complex tasks. In addition, deep models exhibit qualitatively better behaviors which might be interpreted as implicitly acquired skills necessary to reach the goal. + +Limitations. The primary limitations of our results are that scaling network depth comes at the cost of compute. An important direction for future work is to study how distributed training might be used to leverage even more compute, and how techniques such as pruning and distillation might be used to decrease the computational costs. + +Impact Statement This paper presents work whose goal is to advance the field of Machine Learning. There are many potential societal consequences of our work, none which we feel must be specifically highlighted here. + +Acknowledgments. We gratefully acknowledge Galen Collier, Director of Researcher Engagement, for support through the NSF grant, as well as the staff at Princeton Research Computing for their invaluable assistance. We also thank Colin Lu for his discussions and contributions to this work. This research was also partially supported by the National Science Centre, Poland (grant no. 2023/51/D/ST6/01609), and the Warsaw University of Technology through the Excellence Initiative: Research University (IDUB) program. Finally, we would also like to thank Jens Tuyls and Harshit Sikchi for providing helpful commends and feedback on the manuscript. + +![](images/1089a432ac57e32fb6151d0f2d6be79116c405c25d20c4f45c8d47c219948a9e.jpg) +Figure 12: Testing the limits of scale. We extend the results from Figure 1 by scaling networks even further on the challenging Humanoid maze environments. We observe continued performance improvements with network depths of 256 and 1024 layers on Humanoid U-Maze. Note that for the 1024-layer networks, we observed the actor loss exploding at the onset of training, so we maintained the actor depth at 512 while using 1024-layer networks only for the two critic encoders. + +![](images/671c24af599f59d28fe70bed48c4ccdcee70f8c14684e4f0b29ed068f92781aa.jpg) + +# References + +Ahn, M., Brohan, A., Brown, N., Chebotar, Y., Cortes, O., David, B., Finn, C., Gopalakrishnan, K., Hausman, K., Herzog, A., Ho, D., Hsu, J., Ibarz, J., Ichter, B., Irpan, A., Jang, E., Ruano, R. J., Jeffrey, K., Jesmonth, S., Joshi, N., Julian, R. C., Kalashnikov, D., Kuang, Y., Lee, K.-H., Levine, S., Lu, Y., Luu, L., Parada, C., Pastor, P., Quiambao, J., Rao, K., Rettinghouse, J., Reyes, D., Sermanet, P., Sievers, N., Tan, C., Toshev, A., Vanhoucke, V., Xia, F., Xiao, T., Xu, P., Xu, S., and Yan, M. (2022). Do as i can, not as i say: Grounding language in robotic affordances. Conference on Robot Learning. +Andrychowicz, M., Wolski, F., Ray, A., Schneider, J., Fong, R., Welinder, P., McGrew, B., Tobin, J., Pieter Abbeel, O., and Zaremba, W. (2017). Hindsight Experience Replay. In Neural Information Processing Systems, volume 30. +Ba, J. L., Kiros, J. R., and Hinton, G. E. (2016). Layer normalization. arXiv preprint arXiv: 1607.06450. +Blier, L., Tallec, C., and Ollivier, Y. (2021). Learning Successor States and Goal-Dependent Values: A Mathematical Viewpoint. +Bortkiewicz, M., Pałucki, W., Myers, V., Dziarmaga, T., Arczewski, T., Kuciński, L., and Eysenbach, B. (2024). Accelerating goal-conditioned rl algorithms and research. arXiv preprint arXiv:2408.11052. +Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., and Joulin, A. (2021). Emerging properties in self-supervised vision transformers. arXiv preprint arXiv: 2104.14294. +Chang, B., Meng, L., Haber, E., Tung, F., and Begert, D. (2018). Multi-level residual networks from dynamical systems view. In 6th International Conference on Learning Representations, ICLR 2018, Vancouver, BC, Canada, April 30 - May 3, 2018, Conference Track Proceedings. OpenReview.net. +Chen, C., Zhang, J., Xu, Y., Chen, L., Duan, J., Chen, Y., Tran, S. D., Zeng, B., and Chilimbi, T. (2022). Why do we need large batchesizes in contrastive learning? a gradient-bias perspective. In Oh, A. H., Agarwal, A., Belgrave, D., and Cho, K., editors, Advances in Neural Information Processing Systems. +Chen, T., Kornblith, S., Swersky, K., Norouzi, M., and Hinton, G. E. (2020). Big self-supervised models are strong semi-supervised learners. Advances in neural information processing systems, 33:22243-22255. +Dehghani, M., Djolonga, J., Mustafa, B., Padlewski, P., Heek, J., Gilmer, J., Steiner, A., Caron, M., Geirhos, R., Alabdulmohsin, I. M., Jenatton, R., Beyer, L., Tschannen, M., Arnab, A., Wang, X., Riquelme, C., Minderer, M., Puigcerver, J., Evci, U., Kumar, M., van Steenkiste, S., Elsayed, G. F., Mahendran, A., Yu, F., Oliver, A., Huot, F., Bastings, J., Collier, M., Gritsenko, A., Birodkar, V., Vasconcelos, C., Tay, Y., Mensink, T., Kolesnikov, A., Paveti'c, F., Tran, D., Kipf, T., Luvcic, M., Zhai, X., Keysers, D., Harmsen, J., and Houlsby, N. (2023). Scaling vision transformers to 22 billion parameters. International Conference on Machine Learning. +Driess, D., Xia, F., Sajjadi, M. S. M., Lynch, C., Chowdhery, A., Ichter, B., Wahid, A., Tompson, J., Vuong, Q., Yu, T., Huang, W., Chebotar, Y., Sermanet, P., Duckworth, D., Levine, S., Vanhoucke, V., Hausman, K., Toussaint, M., Greff, K., Zeng, A., Mordatch, I., and Florence, P. R. (2023). Palm-e: An embodied multimodal language model. International Conference on Machine Learning. +Dubey, A., Jauhri, A., Pandey, A., Kadian, A., Al-Dahle, A., Letman, A., Mathur, A., Schelten, A., Yang, A., Fan, A., et al. (2024). The llama 3 herd of models. arXiv preprint arXiv:2407.21783. +Espeholt, L., Soyer, H., Munos, R., Simonyan, K., Mnih, V., Ward, T., Doron, Y., Firoiu, V., Harley, T., Dunning, I., et al. (2018). Impala: Scalable distributed deep-rl with importance weighted actor-learner architectures. In International conference on machine learning, pages 1407-1416. PMLR. +Esser, P., Kulal, S., Blattmann, A., Entezari, R., Müller, J., Saini, H., Levi, Y., Lorenz, D., Sauer, A., Boesel, F., et al. (2024). Scaling rectified flow transformers for high-resolution image synthesis. In *Forty-first International Conference on Machine Learning*. +Eysenbach, B., Salakhutdinov, R., and Levine, S. (2021). C-Learning: Learning to Achieve Goals via Recursive Classification. In International Conference on Learning Representations. arXiv. +Eysenbach, B., Zhang, T., Levine, S., and Salakhutdinov, R. R. (2022). Contrastive learning as goal-conditioned reinforcement learning. Advances in Neural Information Processing Systems, 35:35603-35620. +Farebrother, J., Orbay, J., Vuong, Q., Taiga, A. A., Chebotar, Y., Xiao, T., Irpan, A., Levine, S., Castro, P. S., Faust, A., Kumar, A., and Agarwal, R. (2024). Stop Regressing: Training Value Functions via Classification for Scalable Deep RL. + +Freeman, C. D., Frey, E., Raichuk, A., Girgin, S., Mordatch, I., and Bachem, O. (2021). Brax - a Differentiable Physics Engine for Large Scale Rigid Body Simulation. In NeurIPS Datasets and Benchmarks. arXiv. +Goyal, P., Mahajan, D., Gupta, A., and Misra, I. (2019). Scaling and benchmarking self-supervised visual representation learning. In Proceedings of the IEEE/cvf International Conference on computer vision, pages 6391-6400. +Gupta, A., Kumar, V., Lynch, C., Levine, S., and Hausman, K. (2019). Relay policy learning: Solving long-horizon tasks via imitation and reinforcement learning. Conference on Robot Learning. +He, K., Zhang, X., Ren, S., and Sun, J. (2015). Deep residual learning for image recognition. Computer Vision and Pattern Recognition. +He, K., Zhang, X., Ren, S., and Sun, J. (2016). Identity Mappings in Deep Residual Networks, pages 630-645. Springer International Publishing. +Henderson, P., Islam, R., Bachman, P., Pineau, J., Precup, D., and Meger, D. (2018). Deep reinforcement learning that matters. In Proceedings of the AAAI conference on artificial intelligence, volume 32. +Huang, S., Dossa, R. F. J., Ye, C., Braga, J., Chakraborty, D., Mehta, K., and Araujo, J. G. (2022). Cleanrl: High-quality single-file implementations of deep reinforcement learning algorithms. Journal of Machine Learning Research, 23(274):1-18. +Kaelbling, L. P. (1993). Learning to achieve goals. In *IJCAI*, volume 2, pages 1094–8. CiteSeer. +Kumar, A., Agarwal, R., Geng, X., Tucker, G., and Levine, S. (2023). Offline q-learning on diverse multi-task data both scales and generalizes. In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023. OpenReview.net. +LeCun, Y. (2016). Predictive learning. Invited talk at the 30th Conference on Neural Information Processing Systems (NIPS). Barcelona, Spain. +Lee, H., Hwang, D., Kim, D., Kim, H., Tai, J. J., Subramanian, K., Wurman, P. R., Choo, J., Stone, P., and Seno, T. (2024). SimBa: Simplicity Bias for Scaling Up Parameters in Deep Reinforcement Learning. +Lee, K.-H., Nachum, O., Yang, M., Lee, L., Freeman, D., Xu, W., Guadarrama, S., Fischer, I., Jang, E., Michalewski, H., and Mordatch, I. (2022). Multi-Game Decision Transformers. +Liu, B., Feng, Y., Liu, Q., and Stone, P. (2023). Metric Residual Networks for Sample Efficient Goal-Conditioned Reinforcement Learning. +Liu, H., Li, C., Wu, Q., and Lee, Y. J. (2024). Visual instruction tuning. Advances in neural information processing systems, 36. +Lyle, C., Rowland, M., and Dabney, W. (2022). Understanding and preventing capacity loss in reinforcement learning. arXiv preprint arXiv:2204.09560. +Lyle, C., Zheng, Z., Khetarpal, K., van Hasselt, H., Pascanu, R., Martens, J., and Dabney, W. (2024). Disentangling the causes of plasticity loss in neural networks. arXiv preprint arXiv:2402.18762. +Makoviychuk, V., Wawrzyniak, L., Guo, Y., Lu, M., Storey, K., Macklin, M., Hoeller, D., Rudin, N., Allshire, A., Handa, A., et al. (2021). Isaac gym: High performancegpu-based physics simulation for robot learning. arXiv preprint arXiv:2108.10470. +Myers, V., Zheng, C., Dragan, A., Levine, S., and Eysenbach, B. (2024). Learning temporal distances: Contrastive successor features can provide a metric structure for decision-making. International Conference on Machine Learning. +Nauman, M., Bortkiewicz, M., Milos, P., Trzcinski, T., Ostaszewski, M., and Cygan, M. (2024a). Overestimation, overfitting, and plasticity in actor-critic: the bitter lesson of reinforcement learning. In *Forty-first International Conference on Machine Learning*, ICML 2024, Vienna, Austria, July 21-27, 2024. OpenReview.net. +Nauman, M., Ostaszewski, M., Jankowski, K., Miłos, P., and Cygan, M. (2024b). Bigger, Regularized, Optimistic: Scaling for compute and sample-efficient continuous control. +Neumann, O. and Gros, C. (2022). Scaling laws for a multi-agent reinforcement learning model. arXiv preprint arXiv:2210.00849. +Obando-Ceron, J., Bellemare, M. G., and Castro, P. S. (2023). Small batch deep reinforcement learning. Neural Information Processing Systems. Published at NeurIPS 2023. + +Obando-Ceron, J., Sokar, G., Willi, T., Lyle, C., Farebrother, J., Foerster, J. N., Dziugaite, G., Precup, D., and Castro, P. S. (2024). Mixtures of experts unlock parameter scaling for deep rl. International Conference on Machine Learning. +Ota, K., Jha, D. K., and Kanezaki, A. (2021). Training larger networks for deep reinforcement learning. arXiv preprint arXiv:2102.07920. +Park, S., Frans, K., Eysenbach, B., and Levine, S. (2024). Ogbench: Benchmarking offline goal-conditioned rl. arXiv preprint arXiv: 2410.20092. +Radford, A. (2018). Improving language understanding by generative pre-training. +Radford, A., Kim, J. W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., and Sutskever, I. (2021). Learning transferable visual models from natural language supervision. International Conference on Machine Learning. +Raffin, A., Hill, A., Gleave, A., Kanervisto, A., Ernestus, M., and Dormann, N. (2021). Stable-baselines3: Reliable reinforcement learning implementations. Journal of Machine Learning Research, 22(268):1-8. +Ramachandran, P., Zoph, B., and Le, Q. V. (2018). Searching for activation functions. In 6th International Conference on Learning Representations, ICLR 2018, Vancouver, BC, Canada, April 30 - May 3, 2018, Workshop Track Proceedings. OpenReview.net. +Rudin, N., Hoeller, D., Reist, P., and Hutter, M. (2022). Learning to walk in minutes using massively parallel deep reinforcement learning. In Conference on Robot Learning, pages 91-100. PMLR. +Rutherford, A., Ellis, B., Gallici, M., Cook, J., Lupu, A., Ingvarsson, G., Willi, T., Khan, A., de Witt, C. S., Souly, A., et al. (2023). Jaxmarl: Multi-agent rl environments and algorithms in jax. arXiv preprint arXiv:2311.10090. +Schwarzer, M., Obando-Ceron, J. S., Courville, A. C., Bellemare, M. G., Agarwal, R., and Castro, P. S. (2023). Bigger, better, faster: Human-level atari with human-level efficiency. In Krause, A., Brunskill, E., Cho, K., Engelhardt, B., Sabato, S., and Scarlett, J., editors, International Conference on Machine Learning, ICML 2023, 23-29 July 2023, Honolulu, Hawaii, USA, volume 202 of Proceedings of Machine Learning Research, pages 30365-30380. PMLR. +Sohn, K. (2016). Improved Deep Metric Learning With Multi-Class N-Pair Loss Objective. In Neural Information Processing Systems, volume 29. Curran Associates, Inc. +Srivastava, A., Rastogi, A., Rao, A., et al. (2023). Beyond the imitation game: Quantifying and extrapolating the capabilities of language models. Trans. Mach. Learn. Res. +Team, A. A., Bauer, J., Baumli, K., Baveja, S., Behbahani, F., Bhoopchand, A., Bradley-Schmieg, N., Chang, M., Clay, N., Collister, A., et al. (2023). Human-timescale adaptation in an open-ended task space. arXiv preprint arXiv:2301.07608. +Todorov, E., Erez, T., and Tassa, Y. (2012). Mujoco: A Physics Engine for Model-Based Control. In IEEE/RSJ International Conference on Intelligent Robots and Systems, pages 5026-5033. IEEE, IEEE. +Torgo, L. and Gama, J. (1996). Regression by classification. In Advances in Artificial Intelligence: 13th Brazilian Symposium on Artificial Intelligence, SBIA'96 Curitiba, Brazil, October 23-25, 1996 Proceedings 13, pages 51-60. Springer. +Tuyls, J., Madeka, D., Torkkola, K., Foster, D., Narasimhan, K., and Kakade, S. (2024). Scaling Laws for Imitation Learning in Single-Agent Games. +Van Hasselt, H., Doron, Y., Strub, F., Hessel, M., Sonnerat, N., and Modayil, J. (2018). Deep reinforcement learning and the deadly triad. arXiv preprint arXiv:1812.02648. +Vaswani, A., Shazeer, N. M., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, L., and Polosukhin, I. (2017). Attention is all you need. nips. +Veit, A., Wilber, M., and Belongie, S. (2016). Residual networks behave like ensembles of relatively shallow networks. arXiv preprint arXiv: 1605.06431. +Wang, T., Torralba, A., Isola, P., and Zhang, A. (2023a). Optimal goal-reaching reinforcement learning via quasimetric learning. + +Wang, T., Torralba, A., Isola, P., and Zhang, A. (2023b). Optimal goal-reaching reinforcement learning via quasimetric learning. In Krause, A., Brunskill, E., Cho, K., Engelhardt, B., Sabato, S., and Scarlett, J., editors, International Conference on Machine Learning, ICML 2023, 23-29 July 2023, Honolulu, Hawaii, USA, volume 202 of Proceedings of Machine Learning Research, pages 36411-36430. PMLR. +Wei, J., Tay, Y., Bommasani, R., Raffel, C., Zoph, B., Borgeaud, S., Yogatama, D., Bosma, M., Zhou, D., Metzler, D., Chi, E. H., Hashimoto, T., Vinyals, O., Liang, P., Dean, J., and Fedus, W. (2022). Emergent abilities of large language models. Trans. Mach. Learn. Res. +Zhai, X., Kolesnikov, A., Houlsby, N., and Beyer, L. (2021). Scaling vision transformers. Computer Vision and Pattern Recognition. +Zhang, H., Morwani, D., Vyas, N., Wu, J., Zou, D., Ghai, U., Foster, D., and Kakade, S. (2024). How does critical batch size scale in pre-training? arXiv preprint arXiv: 2410.21676. +Zheng, C., Eysenbach, B., Walke, H., Yin, P., Fang, K., Salakhutdinov, R., and Levine, S. (2024). Stabilizing Contrastive RL: Techniques for Offline Goal Reaching. In International Conference on Learning Representations. arXiv. +Zheng, C., Salakhutdinov, R., and Eysenbach, B. (2023). Contrastive Difference Predictive Coding. In Twelfth International Conference on Learning Representations. arXiv. +Zong, Y., Aodha, O. M., and Hospedales, T. (2024). Self-supervised multimodal learning: A survey. + +# A Additional Experiments + +# A.1 Scaled CRL Outperforms All Other Baselines on 8 out of 10 Environments + +![](images/730bd078e073f82e0afc72edfe007bbf905c5ff4a3c9bbb17bc1e231027dabcd.jpg) + +![](images/92b4b4dd6cc5cb9ae652725811d49c200f71bb103ff45683cdf306fa189becf4.jpg) + +![](images/5626038346aff0b231cf08e379c30b04195618e8844594ab5d8e7ab93d258d54.jpg) + +![](images/717faa98da86c75af1e132ea28cb18fc6cd09ef6207bcf71594db5f6cafc42be.jpg) + +![](images/10de8a92b83061a3bbb3daaa246604221b44a22a04ed92bc7917aed706e0aab0.jpg) + +![](images/c39850f767db2200a5efbaa9a7409da9f169122a54148f62209df6dc4f9ccd6c.jpg) + +![](images/dcb3c394d95f998c3ca94f8c798ec68de3c6ef449f21629a409730e694488b94.jpg) + +![](images/a56ee6ea2c1ba803074c50e470b55d7e72d34b4a5b41e65c4022394c5e1871ef.jpg) + +![](images/5cc03295b480b2fffcbd1ef09560b179d6f41d2d148dc30e504730376ceecd0d.jpg) + +![](images/69acfead8a21a555087ad96ad7168846526846cf2f272b9a3d1edad2bef7c374.jpg) + +![](images/f7d10f01112b08059ef3e70f652871a7c4d44c970666de21ecb1eef39bb36b79.jpg) +Figure 12: Scaled CRL (Ours) outperforms baselines CRL (original), SAC, SAC+HER, TD3+HER, GCSL, and GCBC in 8 out 10 environments. + +In Figure 1, we demonstrated that increasing the depth of the CRL algorithm leads to significant performance improvements over the original CRL (see also Table 1). Here, we show that these gains translate to state-of-the-art results in online goal-conditioned RL, with Scaled CRL outperforming both standard TD-based methods such as SAC, SAC+HER, and TD3+HER, as well as self-supervised imitation-based approaches like GCBC and GCSL. + +# A.2 The CRL Algorithm is Key: Depth Scaling is Not Effective on Other Baselines + +Next, we investigate whether increasing network depth in the baseline algorithms yields similar performance improvements as observed in CRL. We find that SAC, SAC+HER, and TD3+HER do not benefit from depths beyond four layers, which is consistent with prior findings (Lee et al., 2024; Nauman et al., 2024b). Additionally, GCSL and GCBC fail to achieve any meaningful performance on the Humanoid and Ant Big Maze tasks. Interestingly, we do observe one exception, as GCBC exhibits improved performance with increased depth in the Arm Push Easy environment. + +Table 1: Increasing network depth (depth $D = 4 \rightarrow {64}$ ) increases performance on CRL (Figure 1). Scaling depth exhibits the greatest benefits on tasks with the largest observation dimension (Dim). + +
TaskDimD=4D=64Imprv.
Arm Binpick Hard38 ±4219 ±155.7×
Arm Push Easy17308 ±33762 ±302.5×
Arm Push Hard171 ±11410 ±132.4×
Ant U4-Maze11.4 ±4.1286 ±3625×
Ant U5-Maze290.97 ±0.761 ±1863×
Ant Big Maze61 ±20441 ±257.3×
Ant Hardest Maze215 ±8387 ±211.8×
Humanoid12.6 ±1.3649 ±1952×
Humanoid U-Maze2683.2 ±1.2159 ±3350×
Humanoid Big Maze0.06 ±0.0459 ±211051×
+ +![](images/ffc3088b66f0b9f1ea4005a46a1150f559426c2e59b908aca27211eb266c445c.jpg) + +![](images/fd2ff44be34ecf3d74b0de975defdf82f8de8e4863060e980122d5d7b1c522b6.jpg) + +![](images/91a96e98fa35859b13b79c5b3f887754b4076efe1b0a158c7405f47d466ba96b.jpg) + +![](images/b4e6ebb0d9243d9112e93ced18543f95a10d285889f36c8ef7b1c888da4a8275.jpg) + +![](images/2ef7d485f6f0183e500a1357ee91daff2187fcf40a7c3808651c9a900e6cd95d.jpg) + +![](images/5f291d52f6bc83ee9d3fc85dcd3041b4f1fd33c290279d83931f9ca07254c10c.jpg) + +![](images/77a33f0f2a577985758a50f764b45456d9661417bb2f42007a7207cb4900d7cf.jpg) + +![](images/08639185cef7a515607224824ff62fed516b53463e5ba031aa4f55615815ab03.jpg) + +![](images/1f1061f51b305cb129308726d79a18fae1489c49f92c116244d3823d000c8d6f.jpg) + +![](images/e78849b80ba0f1e1c07b880e4f931355fefe6288386adec39938b2d5ab4f891e.jpg) + +![](images/5930254e5700b75ae34b6d8eceee07e48198112e5290f62fbfa0bbda4d0b3aa3.jpg) + +![](images/7212d8f7cea86ac9e1ea1ff8f89244148e56c9a2919d662a1746b8d3789b752c.jpg) + +![](images/aa3babc7dad7d6cd819f9d09851cc7d00bf09d334de95780163d97cb4d9a43ee.jpg) + +![](images/c018f4961757f0f584dd83baffd97f135b20e1c28606310ae6941e0002be978e.jpg) + +![](images/3582caab873fb979c5456607a84ff67fe744710c36cfb3e297fd09c61e2ed424.jpg) + +![](images/01f2fceac3a46fac59a8cf1d032a977b4a868fffa14956a730508d0851c51c44.jpg) +Figure 13: Depth scaling yields limited gains for SAC, SAC+HER, TD3+HER, GCSL, and GCBC. + +![](images/d4f07b04d7626920c01ba3cb9dc3c224ac3cfecdca5b3ff88e4a09ccd030bb26.jpg) + +![](images/295cbc38794a2cc8382873a69c3a5576bb3c9aa73cfd7a5aa70b2aa4762efc42.jpg) + +# A.3 Additional Scaling Experiments: Offline GCBC, BC, and QRL + +We further investigate several additional scaling experiments. As shown in Figure 14, our approach successfully scales with depth in the offline GCBC setting on the antmaze-medium-stitch task from OGBench. We find that our the combination of layer normalization, residual connections, and Swish activations is critical, suggesting that our architectural choices may be applied to unlock depth scaling in other algorithms and settings. We also attempt to scale depth for behavioral cloning and the QRL (Wang et al., 2023a) algorithm—in both of these cases, however, we observe negative results. + +![](images/be38a439855f1beb09b5904f549e5925eafadea61df57235ed34627022314729.jpg) +Figure 14: Our approach successfully scales depth in offline GCBC on antmaze-medium-stitch (OGBench). In contrast, scaling depth for BC (antmaze-giant-navigate, expert SAC data) and for both online (FetchPush) and offline QRL (pointmaze-giant-stitch, OGBench) yield negative results. + +![](images/001fb349a93760545ae79ffefe97e98ea4ef790899cfd33082bb6799f4f08b6c.jpg) +Depth + +![](images/5ab56d730e05524ad8c1d04051ea1efc4cd28de3acefebc22b88271f53fb379f.jpg) + +![](images/f0ee878c0e553800f68e7859a7e5704c49eea6ec5f8541ec9d75e53b14b9f95c.jpg) + +# A.4 Can Depth Scaling also be Effective for Quasimetric Architectures? + +Prior work (Wang et al., 2023b; Liu et al., 2023) has found that temporal distances satisfy an important invariance property, suggesting the use of quasimetric architectures when learning temporal distances. Our next experiment tests whether changing the architecture affects the scaling properties of self-supervised RL. Specifically, we use the CMD-1 algorithm (Myers et al., 2024), which employs a backward NCE loss with MRN representations. The results indicate that scaling benefits are not limited to a single neural network parametrization. However, MRN's poor performance on the Ant U5-Maze task suggests further innovation is needed for consistent scaling with quasimetric models. + +![](images/461d7d9d447b32970c2d4812f3a7b2dbcc00b4b63445059d391e954be6c82099.jpg) +Figure 15: Performance of depth scaling on CRL augmented with quasimetric architectures (CMD-1). + +# A.5 Additional Architectural Ablations: Layer Norm and Swish Activation + +We conduct ablation experiments to validate the architectural choices of layer norm and swish activation. Figure 16 shows that removing layer normalization performs significantly worse. Additionally, scaling with ReLU significantly hampers scalability. These results, along with Figure 5 show that all of our architectural components—residual connections, layer norm, and swish activations—are jointly essential to unlocking the full performance of depth scaling. + +![](images/4f0afa3c93eb9cab976eb565dd8daecdfa5fb3ea8ebad983c235ae5653e0cdab.jpg) +Figure 16: (Left) Layer Norm is essential for scaling depth. (Right) Scaling with ReLU activations leads to worse performance compared to Swish activations. + +# A.6 Can We Integrate Novel Architectural Innovations from the Emerging RL Scaling Literature? + +Recently, Simba-v2 proposed a new architecture for scalable RL. Its key innovation is the replacement of layer normalization with hyperspherical normalization, which projects network weights onto the unit-norm hypersphere after each gradient update. As shown, the same depth-scaling trends hold when adding hyperspherical normalization to our architecture, and it further improves the sample efficiency of depth scaling. This demonstrates that our method can naturally incorporate new architectural innovations emerging in the RL scaling literature. + +Table 2: Integrating hyperspherical normalization in our architecture enhances the sample efficiency of depth scaling. + +
Steps to reach ≥200 successSteps to reach ≥400 successSteps to reach ≥600 success
Depth41632Depth41632Depth41632
With-5042With-6248With-7767
Without-6454Without-7564Without--77
+ +# A.7 Residuals Norms in Deep Networks + +Prior work has noted decreasing residual activation norms in deeper layers (Chang et al., 2018). We investigate whether this pattern also holds in our setting. For the critic, the trend is generally evident, especially in very deep architectures (e.g., depth 256). The effect is not as pronounced in the actor. + +![](images/e0f72721430a116fce3aefcb60d0a04a1c90e8eb1feb08bd7e479401c4e33a95.jpg) +Average Residual Magnitudes (L2 Norm) + +![](images/44df65accbd9b1d02a14dfc6e1a925fa2cb15380bad5e67a58fc43dfce74d772.jpg) + +![](images/275e9f53dfec89312b23be85ed5a9f499a95dedcce12cd36a7ccbbc6d3abbe26.jpg) + +![](images/37f567c8c0513b45ded73a0f3cec29eb9bb841cee1e530ed456ccc0766c3de88.jpg) +Figure 17: L2 norms of residual activations in networks with depths of 32, 64, 128, and 256. + +# A.8 Scaling Depth for Offline Goal-conditioned RL + +![](images/706725daf7ac042a4ec4662893537bcfa4a6d2edd5494df61bc4db85fe0909fc.jpg) +Figure 18: To evaluate the scalability of our method in the offline setting, we scaled model depth on OGBench (Park et al., 2024). In two out of three environments, performance drastically declined as depth scaled from 4 to 64, while a slight improvement was seen on antimaze-medium-stitch-v0. Successfully adapting our method to scale offline GCRL is an important direction for future work. + +# B Experimental Details + +# B.1 Environment Setup and Hyperparameters + +![](images/8e14b25a1a5efd6141724a7054694c68ee689f6f04da31c782256ff51307da2b.jpg) +Figure 19: The scaling results of this paper are demonstrated on the JaxGCRL benchmark, showing that they replicate across a diverse range of locomotion, navigation, and manipulation tasks. These tasks are set in the online goal-conditioned setting where there are no auxiliary rewards or demonstrations. Figure taken from (Bortkiewicz et al., 2024). + +Our experiments use the JaxGCRL suite of GPU-accelerated environments, visualized in Figure 19, and a contrastive RL algorithm with hyperparameters reported in Table 7. In particular, we use 10 environments, namely: ant/big_maze, ant_hardest_maze, arm_binpick-hard, arm.push_easy, arm.push-hard, humanoid, humanoidBIG_maze, humanoid_u_maze, ant_u4_maze, ant_u5_maze. + +# B.2 Python Environment Differences + +In all plots presented in the paper, we used MJX 3.2.6 and Brax 0.10.1 to ensure a fair and consistent comparison. During development, we noticed discrepancies in physics behavior between the environment versions we employed (the CleanRL version of JaxGCRL) and the version recommended in a more recent commit of JaxGCRL (Bortkiewicz et al., 2024). Upon examination, the performance differences (shown in Figure 20) stem from a difference in versions in the MJX and Brax packages. Nonetheless, in both sets of MJX and Brax versions, performance scales monotonically with depth. + +![](images/c5fe01c60cc53c18272483270de572d1c295fb9fe149e45737bf09ce4f14acce.jpg) +Figure 20: Scaling behavior for humanoid in two different python environments: MJX=3.2.3, Brax=0.10.5 and MJX=3.2.6, Brax=0.10.1 (ours) version of JaxGCRL. Scaling depth improves the performance significantly for both versions. In the environment we used, training requires fewer environment steps to reach a marginally better performance than in other Python environment. + +# B.3 Wall-clock Time of Our Approach + +We report the wall-clock time of our approach in Table 3. The table shows results for depths of 4, 8, 16, 32, and 64 across all ten environments, and for the Humanoid U-Maze environment, scaling up to 1024 layers. Overall, wall-clock time increases approximately linearly with depth beyond a certain point. + +Table 3: Wall-clock time (in hours) for Depth 4, 8, 16, 32, and 64 across all 10 environments. + +
EnvironmentDepth 4Depth 8Depth 16Depth 32Depth 64
Humanoid1.48 ± 0.002.13 ± 0.013.40 ± 0.015.92 ± 0.0110.99 ± 0.01
Ant Big Maze2.12 ± 0.002.77 ± 0.004.04 ± 0.016.57 ± 0.0211.66 ± 0.03
Ant U4-Maze1.98 ± 0.272.54 ± 0.013.81 ± 0.016.35 ± 0.0111.43 ± 0.03
Ant U5-Maze9.46 ± 1.7510.99 ± 0.0216.09 ± 0.0131.49 ± 0.3446.40 ± 0.12
Ant Hardest Maze5.11 ± 0.006.39 ± 0.008.94 ± 0.0113.97 ± 0.0123.96 ± 0.06
Arm Push Easy9.97 ± 1.0311.02 ± 1.2912.20 ± 1.4314.94 ± 1.9619.52 ± 1.97
Arm Push Hard9.74 ± 1.0510.55 ± 1.2011.98 ± 1.4914.40 ± 1.6418.53 ± 0.06
Arm Binpick Hard18.41 ± 2.1617.48 ± 1.8819.47 ± 0.0521.91 ± 1.9329.64 ± 6.10
Humanoid U-Maze8.72 ± 0.0111.29 ± 0.0116.36 ± 0.0326.48 ± 0.0546.74 ± 0.04
Humanoid Big Maze12.45 ± 0.0215.02 ± 0.0120.34 ± 0.0130.61 ± 0.0550.33 ± 0.05
+ +Table 4: Total wall-clock time (in hours) for training from Depth 4 up to Depth 1024 in the Humanoid U-Maze environment. + +
DepthTime (h)
43.23 ± 0.001
84.19 ± 0.003
166.07 ± 0.003
329.83 ± 0.006
6417.33 ± 0.003
12832.67 ± 0.124
25673.83 ± 2.364
512120.88 ± 2.177
1024134.15 ± 0.081
+ +# B.4 Wall-clock Time: Comparison to Baselines + +Since the baselines use standard sized networks, naturally our scaled approach incurs higher raw wall-clock time per environment step (Table 5). However, a more practical metric is the time required to reach a given performance level. As shown in Table 6, our approach outperforms the strongest baseline, SAC, in 7 of 10 environments while requiring less wall-clock time. + +Table 5: Wall-clock training time comparison of our method vs. baselines across all 10 environments. + +
EnvironmentScaled CRLSACSAC+HERTD3GCSLGCBC
Humanoid11.0 ± 0.00.5 ± 0.00.6 ± 0.00.8 ± 0.00.4 ± 0.00.6 ± 0.0
Ant Big Maze11.7 ± 0.01.6 ± 0.01.6 ± 0.01.7 ± 0.01.5 ± 0.31.4 ± 0.1
Ant U4-Maze11.4 ± 0.01.2 ± 0.01.3 ± 0.01.3 ± 0.00.7 ± 0.01.1 ± 0.1
Ant U5-Maze46.4 ± 0.15.7 ± 0.06.1 ± 0.06.2 ± 0.02.8 ± 0.15.6 ± 0.5
Ant Hardest Maze24.0 ± 0.04.3 ± 0.04.5 ± 0.05.0 ± 0.02.1 ± 0.64.4 ± 0.5
Arm Push Easy19.5 ± 0.68.3 ± 0.08.5 ± 0.08.4 ± 0.06.4 ± 0.18.3 ± 0.3
Arm Push Hard18.5 ± 0.08.5 ± 0.08.6 ± 0.08.3 ± 0.15.2 ± 0.37.4 ± 0.5
Arm Binpick Hard29.6 ± 1.320.7 ± 0.120.7 ± 0.018.4 ± 0.38.0 ± 0.916.2 ± 0.4
Humanoid U-Maze46.7 ± 0.03.0 ± 0.03.5 ± 0.05.4 ± 0.03.1 ± 0.17.2 ± 0.8
Humanoid Big Maze50.3 ± 0.08.6 ± 0.09.3 ± 0.07.5 ± 1.15.1 ± 0.011.4 ± 1.9
+ +Table 6: Wall-clock time (in hours) for our approach to surpass SAC's final performance. As shown, our approach surpasses SAC performance in less wall-clock time in 7 out of 10 environments. The N/A* entries are because in those environments, scaled CRL doesn't outperform SAC. + +
EnvironmentSACScaled CRL (Depth 64)
Humanoid0.466.37
Ant Big Maze1.550.00
Ant U4-Maze1.160.00
Ant U5-Maze5.730.00
Ant Hardest Maze4.330.45
Arm Push Easy8.321.91
Arm Push Hard8.506.65
Arm Binpick Hard20.704.43
Humanoid U-Maze3.04N/A*
Humanoid Big Maze8.55N/A*
+ +Table 7: Hyperparameters + +
HyperparameterValue
num_timesteps100M-400M (varying across tasks)
update-to-data (UTD) ratio1:40
max_replay_size10,000
min_replay_size1,000
episode_length1,000
discounting0.99
num_envs512
batch_size512
policy_lr3e-4
critic_lr3e-4
contrastive_loss_functionInfoNCE
energy_functionL2
logsumexp_penalty0.1
Network depthdepends on the experiment
Network widthdepends on the experiment
representation dimension64
+ +# NeurIPS Paper Checklist + +# 1. Claims + +Question: Do the main claims made in the abstract and introduction accurately reflect the paper's contributions and scope? + +Answer: [Yes] + +Justification: The abstract contains 3 main claims: (1) Depth scaled to 1024 layers; (2) Performance increases 2-50x on CRL and outperforms other goal-conditioned baselines. (3) These performance gains leads to qualitatively new learned behaviors. Each of these claims are clearly substantiated in the main text in Section 4. + +Guidelines: + +- The answer NA means that the abstract and introduction do not include the claims made in the paper. +- The abstract and/or introduction should clearly state the claims made, including the contributions made in the paper and important assumptions and limitations. A No or NA answer to this question will not be perceived well by the reviewers. +- The claims made should match theoretical and experimental results, and reflect how much the results can be expected to generalize to other settings. +- It is fine to include aspirational goals as motivation as long as it is clear that these goals are not attained by the paper. + +# 2. Limitations + +Question: Does the paper discuss the limitations of the work performed by the authors? + +Answer: [Yes] + +Justification: We included a Limitations section that describes the main limitation of our paper, which is latency of deep networks. We also multiple times in the paper demarcated where our research can be extended by future work. + +Guidelines: + +- The answer NA means that the paper has no limitation while the answer No means that the paper has limitations, but those are not discussed in the paper. +- The authors are encouraged to create a separate "Limitations" section in their paper. +- The paper should point out any strong assumptions and how robust the results are to violations of these assumptions (e.g., independence assumptions, noiseless settings, model well-specification, asymptotic approximations only holding locally). The authors should reflect on how these assumptions might be violated in practice and what the implications would be. +- The authors should reflect on the scope of the claims made, e.g., if the approach was only tested on a few datasets or with a few runs. In general, empirical results often depend on implicit assumptions, which should be articulated. +- The authors should reflect on the factors that influence the performance of the approach. For example, a facial recognition algorithm may perform poorly when image resolution is low or images are taken in low lighting. Or a speech-to-text system might not be used reliably to provide closed captions for online lectures because it fails to handle technical jargon. +- The authors should discuss the computational efficiency of the proposed algorithms and how they scale with dataset size. +- If applicable, the authors should discuss possible limitations of their approach to address problems of privacy and fairness. +- While the authors might fear that complete honesty about limitations might be used by reviewers as grounds for rejection, a worse outcome might be that reviewers discover limitations that aren't acknowledged in the paper. The authors should use their best judgment and recognize that individual actions in favor of transparency play an important role in developing norms that preserve the integrity of the community. Reviewers will be specifically instructed to not penalize honesty concerning limitations. + +# 3. Theory assumptions and proofs + +Question: For each theoretical result, does the paper provide the full set of assumptions and a complete (and correct) proof? + +Answer: [NA] + +Justification: This is an empirical paper. As such, no theoretical results that require assumptions or proofs. + +Guidelines: + +- The answer NA means that the paper does not include theoretical results. +- All the theorems, formulas, and proofs in the paper should be numbered and cross-referenced. +- All assumptions should be clearly stated or referenced in the statement of any theorems. +- The proofs can either appear in the main paper or the supplemental material, but if they appear in the supplemental material, the authors are encouraged to provide a short proof sketch to provide intuition. +- Inversely, any informal proof provided in the core of the paper should be complemented by formal proofs provided in appendix or supplemental material. +- Theorems and Lemmas that the proof relies upon should be properly referenced. + +# 4. Experimental result reproducibility + +Question: Does the paper fully disclose all the information needed to reproduce the main experimental results of the paper to the extent that it affects the main claims and/or conclusions of the paper (regardless of whether the code and data are provided or not)? + +Answer: [Yes] + +Justification: Yes, documentation for reproducing the experiments is included alongside the anonymous code. + +Guidelines: + +- The answer NA means that the paper does not include experiments. + +- If the paper includes experiments, a No answer to this question will not be perceived well by the reviewers: Making the paper reproducible is important, regardless of whether the code and data are provided or not. + +- If the contribution is a dataset and/or model, the authors should describe the steps taken to make their results reproducible or verifiable. + +- Depending on the contribution, reproducibility can be accomplished in various ways. For example, if the contribution is a novel architecture, describing the architecture fully might suffice, or if the contribution is a specific model and empirical evaluation, it may be necessary to either make it possible for others to replicate the model with the same dataset, or provide access to the model. In general, releasing code and data is often one good way to accomplish this, but reproducibility can also be provided via detailed instructions for how to replicate the results, access to a hosted model (e.g., in the case of a large language model), releasing of a model checkpoint, or other means that are appropriate to the research performed. + +- While NeurIPS does not require releasing code, the conference does require all submissions to provide some reasonable avenue for reproducibility, which may depend on the nature of the contribution. For example + +(a) If the contribution is primarily a new algorithm, the paper should make it clear how to reproduce that algorithm. +(b) If the contribution is primarily a new model architecture, the paper should describe the architecture clearly and fully. +(c) If the contribution is a new model (e.g., a large language model), then there should either be a way to access this model for reproducing the results or a way to reproduce the model (e.g., with an open-source dataset or instructions for how to construct the dataset). +(d) We recognize that reproducibility may be tricky in some cases, in which case authors are welcome to describe the particular way they provide for reproducibility. In the case of closed-source models, it may be that access to the model is limited in some way (e.g., to registered users), but it should be possible for other researchers to have some path to reproducing or verifying the results. + +# 5. Open access to data and code + +Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material? + +Answer: [Yes] + +Justification: See link to anonymous code in Abstract. + +Guidelines: + +- The answer NA means that paper does not include experiments requiring code. +- Please see the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details. +- While we encourage the release of code and data, we understand that this might not be possible, so "No" is an acceptable answer. Papers cannot be rejected simply for not including code, unless this is central to the contribution (e.g., for a new open-source benchmark). +- The instructions should contain the exact command and environment needed to run to reproduce the results. See the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details. +- The authors should provide instructions on data access and preparation, including how to access the raw data, preprocessed data, intermediate data, and generated data, etc. +- The authors should provide scripts to reproduce all experimental results for the new proposed method and baselines. If only a subset of experiments are reproducible, they should state which ones are omitted from the script and why. +- At submission time, to preserve anonymity, the authors should release anonymized versions (if applicable). +- Providing as much information as possible in supplemental material (appended to the paper) is recommended, but including URLs to data and code is permitted. + +# 6. Experimental setting/details + +Question: Does the paper specify all the training and test details (e.g., data splits, hyperparameters, how they were chosen, type of optimizer, etc.) necessary to understand the results? + +Answer: [Yes] + +Justification: See Experiments section and Appendix on Experimental Details + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The experimental setting should be presented in the core of the paper to a level of detail that is necessary to appreciate the results and make sense of them. +- The full details can be provided either with the code, in appendix, or as supplemental material. + +# 7. Experiment statistical significance + +Question: Does the paper report error bars suitably and correctly defined or other appropriate information about the statistical significance of the experiments? + +Answer: [Yes] + +Justification: Error bars in figures depict one standard error across random seeds. We used 5 seeds in Figure 1. For other figures in the main text, we could only run 3 seeds because of computational constraints. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The authors should answer "Yes" if the results are accompanied by error bars, confidence intervals, or statistical significance tests, at least for the experiments that support the main claims of the paper. + +- The factors of variability that the error bars are capturing should be clearly stated (for example, train/test split, initialization, random drawing of some parameter, or overall run with given experimental conditions). +- The method for calculating the error bars should be explained (closed form formula, call to a library function, bootstrap, etc.) +- The assumptions made should be given (e.g., Normally distributed errors). +- It should be clear whether the error bar is the standard deviation or the standard error of the mean. +- It is OK to report 1-sigma error bars, but one should state it. The authors should preferably report a 2-sigma error bar than state that they have a $96\%$ CI, if the hypothesis of Normality of errors is not verified. +- For asymmetric distributions, the authors should be careful not to show in tables or figures symmetric error bars that would yield results that are out of range (e.g. negative error rates). +- If error bars are reported in tables or plots, The authors should explain in the text how they were calculated and reference the corresponding figures or tables in the text. + +# 8. Experiments compute resources + +Question: For each experiment, does the paper provide sufficient information on the computer resources (type of compute workers, memory, time of execution) needed to reproduce the experiments? + +Answer: [Yes] + +Justification: Compute resources are detailed in the appendix. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The paper should indicate the type of compute workers CPU or GPU, internal cluster, or cloud provider, including relevant memory and storage. +- The paper should provide the amount of compute required for each of the individual experimental runs as well as estimate the total compute. +- The paper should disclose whether the full research project required more compute than the experiments reported in the paper (e.g., preliminary or failed experiments that didn't make it into the paper). + +# 9. Code of ethics + +Question: Does the research conducted in the paper conform, in every respect, with the NeurIPS Code of Ethics https://neurips.cc/public/EthicsGuidelines? + +Answer: [Yes] + +Justification: No known violations of the Code of Ethics. + +Guidelines: + +- The answer NA means that the authors have not reviewed the NeurIPS Code of Ethics. +- If the authors answer No, they should explain the special circumstances that require a deviation from the Code of Ethics. +- The authors should make sure to preserve anonymity (e.g., if there is a special consideration due to laws or regulations in their jurisdiction). + +# 10. Broader impacts + +Question: Does the paper discuss both potential positive societal impacts and negative societal impacts of the work performed? + +Answer: [Yes] + +Justification: The Conclusion notes that there are no immediately societal impacts of the work. + +Guidelines: + +- The answer NA means that there is no societal impact of the work performed. + +- If the authors answer NA or No, they should explain why their work has no societal impact or why the paper does not address societal impact. +- Examples of negative societal impacts include potential malicious or unintended uses (e.g., disinformation, generating fake profiles, surveillance), fairness considerations (e.g., deployment of technologies that could make decisions that unfairly impact specific groups), privacy considerations, and security considerations. +- The conference expects that many papers will be foundational research and not tied to particular applications, let alone deployments. However, if there is a direct path to any negative applications, the authors should point it out. For example, it is legitimate to point out that an improvement in the quality of generative models could be used to generate deepfakes for disinformation. On the other hand, it is not needed to point out that a generic algorithm for optimizing neural networks could enable people to train models that generate Deepfakes faster. +- The authors should consider possible harms that could arise when the technology is being used as intended and functioning correctly, harms that could arise when the technology is being used as intended but gives incorrect results, and harms following from (intentional or unintentional) misuse of the technology. +- If there are negative societal impacts, the authors could also discuss possible mitigation strategies (e.g., gated release of models, providing defenses in addition to attacks, mechanisms for monitoring misuse, mechanisms to monitor how a system learns from feedback over time, improving the efficiency and accessibility of ML). + +# 11. Safeguards + +Question: Does the paper describe safeguards that have been put in place for responsible release of data or models that have a high risk for misuse (e.g., pretrained language models, image generators, or scraped datasets)? + +Answer: [NA] + +Justification: No immediate impact to high-risk applications. + +Guidelines: + +- The answer NA means that the paper poses no such risks. +- Released models that have a high risk for misuse or dual-use should be released with necessary safeguards to allow for controlled use of the model, for example by requiring that users adhere to usage guidelines or restrictions to access the model or implementing safety filters. +- Datasets that have been scraped from the Internet could pose safety risks. The authors should describe how they avoided releasing unsafe images. +- We recognize that providing effective safeguards is challenging, and many papers do not require this, but we encourage authors to take this into account and make a best faith effort. + +# 12. Licenses for existing assets + +Question: Are the creators or original owners of assets (e.g., code, data, models), used in the paper, properly credited and are the license and terms of use explicitly mentioned and properly respected? + +Answer: [NA] + +Justification: Benchmarks used are appropriately cited in the main text. + +Guidelines: + +- The answer NA means that the paper does not use existing assets. +- The authors should cite the original paper that produced the code package or dataset. +- The authors should state which version of the asset is used and, if possible, include a URI. +- The name of the license (e.g., CC-BY 4.0) should be included for each asset. +- For scraped data from a particular source (e.g., website), the copyright and terms of service of that source should be provided. + +- If assets are released, the license, copyright information, and terms of use in the package should be provided. For popular datasets, paperswithcode.com/datasets has curated licenses for some datasets. Their licensing guide can help determine the license of a dataset. +- For existing datasets that are re-packaged, both the original license and the license of the derived asset (if it has changed) should be provided. +- If this information is not available online, the authors are encouraged to reach out to the asset's creators. + +# 13. New assets + +Question: Are new assets introduced in the paper well documented and is the documentation provided alongside the assets? + +Answer: [NA] + +Justification: Datasets and benchmark used are all from prior work and appropriately cited. Guidelines: + +- The answer NA means that the paper does not release new assets. +- Researchers should communicate the details of the dataset/code/model as part of their submissions via structured templates. This includes details about training, license, limitations, etc. +- The paper should discuss whether and how consent was obtained from people whose asset is used. +- At submission time, remember to anonymize your assets (if applicable). You can either create an anonymized URL or include an anonymized zip file. + +# 14. Crowdsourcing and research with human subjects + +Question: For crowdsourcing experiments and research with human subjects, does the paper include the full text of instructions given to participants and screenshots, if applicable, as well as details about compensation (if any)? + +Answer: [NA] + +Justification: No crowdsourcing experiments. + +Guidelines: + +- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. +- Including this information in the supplemental material is fine, but if the main contribution of the paper involves human subjects, then as much detail as possible should be included in the main paper. +- According to the NeurIPS Code of Ethics, workers involved in data collection, curation, or other labor should be paid at least the minimum wage in the country of the data collector. + +# 15. Institutional review board (IRB) approvals or equivalent for research with human subjects + +Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals (or an equivalent approval/review based on the requirements of your country or institution) were obtained? + +Answer: [NA] + +Justification: No human subject experiments + +Guidelines: + +- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. +- Depending on the country in which research is conducted, IRB approval (or equivalent) may be required for any human subjects research. If you obtained IRB approval, you should clearly state this in the paper. + +We recognize that the procedures for this may vary significantly between institutions and locations, and we expect authors to adhere to the NeurIPS Code of Ethics and the guidelines for their institution. +- For initial submissions, do not include any information that would break anonymity (if applicable), such as the institution conducting the review. + +# 16. Declaration of LLM usage + +Question: Does the paper describe the usage of LLMs if it is an important, original, or non-standard component of the core methods in this research? Note that if the LLM is used only for writing, editing, or formatting purposes and does not impact the core methodology, scientific rigorousness, or originality of the research, declaration is not required. + +Answer: [NA] + +Justification: LLMs were not used in writing the paper, and were only used for occasional code debugging. + +Guidelines: + +- The answer NA means that the core method development in this research does not involve LLMs as any important, original, or non-standard components. +- Please refer to our LLM policy (https://neurips.cc/Conferences/2025/LLM) for what should or should not be described. \ No newline at end of file diff --git a/NeurIPS/2025/1000 Layer Networks for Self-Supervised RL_ Scaling Depth Can Enable New Goal-Reaching Capabilities/images.zip b/NeurIPS/2025/1000 Layer Networks for Self-Supervised RL_ Scaling Depth Can Enable New Goal-Reaching Capabilities/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..44e2e03e618354c68f936693d96cfa00a7fced7e --- /dev/null +++ b/NeurIPS/2025/1000 Layer Networks for Self-Supervised RL_ Scaling Depth Can Enable New Goal-Reaching Capabilities/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:aaa30a08d5c04cca666a5694bcbdafee49f75331b6f7eaa0c25343db46e98c15 +size 1458132 diff --git a/NeurIPS/2025/1000 Layer Networks for Self-Supervised RL_ Scaling Depth Can Enable New Goal-Reaching Capabilities/layout.json b/NeurIPS/2025/1000 Layer Networks for Self-Supervised RL_ Scaling Depth Can Enable New Goal-Reaching Capabilities/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b5cf451ef1df14940b60e231acf1180c4786e89e --- /dev/null +++ b/NeurIPS/2025/1000 Layer Networks for Self-Supervised RL_ Scaling Depth Can Enable New Goal-Reaching Capabilities/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:85c7250365a716f25faf728a570aeb05d42026e1cecefb5716386eb0d8734bdf +size 816747 diff --git a/NeurIPS/2025/1000+ FPS 4D Gaussian Splatting for Dynamic Scene Rendering/b10427d1-bee4-4ee0-a8a7-46eb2a4b8e6a_content_list.json b/NeurIPS/2025/1000+ FPS 4D Gaussian Splatting for Dynamic Scene Rendering/b10427d1-bee4-4ee0-a8a7-46eb2a4b8e6a_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..706a7c56134c28ca55bf0d2f97cddb4ff7a4e2a8 --- /dev/null +++ b/NeurIPS/2025/1000+ FPS 4D Gaussian Splatting for Dynamic Scene Rendering/b10427d1-bee4-4ee0-a8a7-46eb2a4b8e6a_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a0c17532f90ea2210aa8e56b69ae75bb7cdbb55b845e32122091898d980ca262 +size 121997 diff --git a/NeurIPS/2025/1000+ FPS 4D Gaussian Splatting for Dynamic Scene Rendering/b10427d1-bee4-4ee0-a8a7-46eb2a4b8e6a_model.json b/NeurIPS/2025/1000+ FPS 4D Gaussian Splatting for Dynamic Scene Rendering/b10427d1-bee4-4ee0-a8a7-46eb2a4b8e6a_model.json new file mode 100644 index 0000000000000000000000000000000000000000..a61f67e7ff403236b08c386fb212adab5117be49 --- /dev/null +++ b/NeurIPS/2025/1000+ FPS 4D Gaussian Splatting for Dynamic Scene Rendering/b10427d1-bee4-4ee0-a8a7-46eb2a4b8e6a_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1aeefca0a8d970fea3e167ae05985e5e10bfa84275cba7330f2b644364e44fcd +size 161505 diff --git a/NeurIPS/2025/1000+ FPS 4D Gaussian Splatting for Dynamic Scene Rendering/b10427d1-bee4-4ee0-a8a7-46eb2a4b8e6a_origin.pdf b/NeurIPS/2025/1000+ FPS 4D Gaussian Splatting for Dynamic Scene Rendering/b10427d1-bee4-4ee0-a8a7-46eb2a4b8e6a_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..0f7ea9dad9f7e243a05926eaca2d08c78f3069a5 --- /dev/null +++ b/NeurIPS/2025/1000+ FPS 4D Gaussian Splatting for Dynamic Scene Rendering/b10427d1-bee4-4ee0-a8a7-46eb2a4b8e6a_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1be91691d19bcd6fcc14615f8672947540c758ac4aeb197b67f18f8d9c5c18de +size 13959299 diff --git a/NeurIPS/2025/1000+ FPS 4D Gaussian Splatting for Dynamic Scene Rendering/full.md b/NeurIPS/2025/1000+ FPS 4D Gaussian Splatting for Dynamic Scene Rendering/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d6af972ded72119287eab0908bd59cac517b7770 --- /dev/null +++ b/NeurIPS/2025/1000+ FPS 4D Gaussian Splatting for Dynamic Scene Rendering/full.md @@ -0,0 +1,618 @@ +# 1000+ FPS 4D Gaussian Splating for Dynamic Scene Rendering + +Yuheng Yuan $^{1,*}$ Qiuhong Shen $^{1,*}$ Xingyi Yang $^{2,1}$ Xinchao Wang $^{1,\dagger}$ + +$^{1}$ National University of Singapore $^{2}$ The Hong Kong Polytechnic University + +{yuhengyuan,qiuhong.shen}@u.nus.edu，xingyi.yang@polyu.edu.hk，xinchao@nus.edu.sgProject page:https://4dgs-1k.github.io/ + +![](images/d7da3e01919a899f1f1fe155031be96fed437ed9c7d9b621dd4a3c341152120d.jpg) +Figure 1: Compressibility and Rendering Speed. We introduce 4DGS-1K, a novel compact representation with high rendering speed. In contrast to 4D Gaussian Splatting (4DGS) [1], we can achieve rasterization at $1000+$ FPS while maintaining comparable photorealistic quality with only $2\%$ of the original storage size. The right figure is the result tested on the N3V [2] datasets, where the radius of the dot corresponds to the storage size. + +![](images/df972e730f95fdee97959ea45405be10c339ce251edb116fd531c067c1394de8.jpg) + +![](images/44b34462cb9aca22914038f602a15a90a6f836f99f503293eda0163e9a4cd0d0.jpg) + +# Abstract + +4D Gaussian Splitting (4DGS) has recently gained considerable attention as a method for reconstructing dynamic scenes. Despite achieving superior quality, 4DGS typically requires substantial storage and suffers from slow rendering speed. In this work, we delve into these issues and identify two key sources of temporal redundancy. (Q1) Short-Lifespan Gaussians: 4DGS uses a large portion of Gaussians with short temporal span to represent scene dynamics, leading to an excessive number of Gaussians. (Q2) Inactive Gaussians: When rendering, only a small subset of Gaussians contributes to each frame. Despite this, all Gaussians are processed during rasterization, resulting in redundant computation overhead. To address these redundancies, we present 4DGS-1K, which runs at over 1000 FPS on modern GPUs. For Q1, we introduce the Spatial-Temporal Variation Score, a new pruning criterion that effectively removes short-lifespan Gaussians while encouraging 4DGS to capture scene dynamics using Gaussians with longer temporal spans. For Q2, we store a mask for active Gaussians across consecutive frames, significantly reducing redundant computations. Compared to vanilla 4DGS, our method achieves a $41 \times$ reduction in storage and $9 \times$ faster rasterization on complex dynamic scenes, while maintaining comparable visual quality. + +# 1 Introduction + +Novel view synthesis for dynamic scenes allows for the creation of realistic representations of 4D environments, which is essential in fields like computer vision, virtual reality, and augmented reality. Traditionally, this area has been led by neural radiance fields (NeRF) [2-6], which model opacity and + +color over time to depict dynamic scenes. While effective, these NeRF-based methods come with high training and rendering costs, limiting their practicality, especially in real-time applications and on devices with limited resources. + +Recently, point-based representations like 4D Gaussian Splatting (4DGS) [1] have emerged as strong alternatives. 4DGS models a dynamic scene using a set of 4D Gaussian primitives, each with a 4-dimensional mean and a $4 \times 4$ covariance matrix. At any given timestamp, a 4D Gaussian is decomposed into a set of conditional 3D Gaussians and a marginal 1D Gaussian, the latter controlling the opacity at that moment. This mechanism allows 4DGS to effectively capture both static and dynamic features of a scene, enabling high-fidelity dynamic scene reconstruction. + +However, representing dynamic scenes with 4DGS is both storage-intensive and slow. Specifically, 4DGS often requires millions of Gaussians, leading to significant storage demands (averaging 2GB for each scene on the N3V [2] dataset) and suboptimal rendering speed. In comparison, mainstream deformation field methods [7] require only about 90MB for the same dataset. Therefore, reducing the storage size of 4DGS [1] and improving rendering speed are essential for efficiently representing complex dynamic scenes. + +We look into the cause of such an explosive number of Gaussian and place a specific emphasis on two key issues. (Q1) A large portion of Gaussians exhibit a short temporal span. In empirical experiments, 4DGS tends to favor "flicking" Gaussians to fit complex dynamic scenes, which just influence a short portion of the temporal domain. This necessitates that 4DGS relies on a large number of Gaussians to reconstruct a high-fidelity scene. As a result, substantial storage is needed to record the attributes of these Gaussians. (Q2) Inactive Gaussians lead to redundant computation. During rendering, 4DGS needs to process all Gaussians. However, only a very small portion of Gaussians are active at that moment. Therefore, most of the computation time is spent on inactive Gaussians. This phenomenon greatly hampers the rendering speed. In this paper, we introduce 4DGS-1K, a framework that significantly reduces the number of Gaussians to minimize storage requirements and speedup rendering while maintaining high-quality reconstruction. To address these issues, 4DGS-1K introduces a two-step pruning approach: + +- Pruning Short-Lifespan Gaussians. We propose a novel pruning criterion called the spatial-temporal variation score, which evaluates the temporal impact of each Gaussian. Gaussians with minimal influence are identified and pruned, resulting in a more compact scene representation with fewer Gaussians with short temporal span. +- Filtering Inactive Gaussians. To further reduce redundant computations during rendering, we use a key-frame temporal filter that selects the Gaussians needed for each frame. On top of this, we share the masks for adjacent frames. This is based on our observation that Gaussians active in adjacent frames often overlap significantly. + +Besides, the pruning in step 1 enhances the masking process in step 2. By pruning Gaussians, we increase the temporal influence of each Gaussian, which allows us to select sparser key frames and further reduce storage requirements. + +We have extensively tested our proposed model on various dynamic scene datasets including real and synthetic scenes. As shown in Figure 1, 4DGS-1K reduces storage costs by $41\times$ on the Neural 3D Video datasets [2] while maintaining equivalent scene representation quality. Crucially, it enables real-time rasterization speeds exceeding 1,000 FPS. These advancements collectively position 4DGS-1K as a practical solution for high-fidelity dynamic scene modeling without compromising efficiency. + +In summary, our contributions are three-fold: + +- We delve into the temporal redundancy of 4D Gaussian Splatting, and explain the main reason for the storage pressure and suboptimal rendering speed. +- We introduce 4DGS-1K, a compact and memory-efficient framework to address these issues. It consists of two key components, a spatial-temporal variation score-based pruning strategy and a temporal filter. +- Extensive experiments demonstrate that 4DGS-1K not only achieves a substantial storage reduction of approximately $41 \times$ but also accelerates rasterization to $1000+$ FPS while maintaining high-quality reconstruction. + +# 2 Related Work + +# 2.1 Novel view synthesis for static scenes + +Recently, neural radiance fields(NeRF) [3] have achieved encouraging results in novel view synthesis. NeRF [3] represents the scene by mapping 3D coordinates and view dependency to color and opacity. Since NeRF [3] requires sampling each ray by querying the MLP for hundreds of points, this significantly limits the training and rendering speed. Subsequent studies [8-15] have attempted to speed up the rendering by introducing specialized designs. However, these designs also constrain the widespread application of these models. In contrast, 3D Gaussian Splatting(3DGS) [16] has gained significant attention, which utilizes anisotropic 3D Gaussians to represent scenes. It achieves high-quality results with intricate details, while maintaining real-time rendering performance. + +# 2.2 Novel view synthesis for dynamic scenes + +Dynamic NVs pose new challenges due to the temporal variations in the input images. Previous NeRF-based dynamic scene representation methods [2, 4-6, 17-22] handle dynamic scenes by learning a mapping from spatiotemporal coordinates to color and density. Unfortunately, these NeRF-based models are constrained in their applications due to low rendering speeds. Recently, 3D Gaussians Splatting [16] has emerged as a novel explicit representation, with many studies [7, 23-27] attempting to model the dynamic scenes based on it. 4D Gaussian Splatting(4DGS) [1] is one of the representatives. It utilizes a set of 4D Gaussian primitives. However, 4DGS often requires a huge redundant number of Gaussians for dynamic scenes. These Gaussians lead to tremendous storage and suboptimal rendering speed. To this end, we focus on analyzing the temporal redundancy of 4DGS [1] in hopes of developing a novel framework to achieve lower storage requirements and higher rendering speeds. + +# 2.3 Gaussian Splitting Compression + +3D Gaussian-based large-scale scene reconstruction typically requires millions of Gaussians, resulting in the requirement of up to several gigabytes of storage. Therefore, subsequent studies have attempted to tackle these issues. Specifically, Compgs [28] and Compact3D [29] employ vector quantization to store Gaussians within codebooks. Concurrently, inspired by model pruning, some studies [30-35] have proposed criterion to prune Gaussians by a specified ratio. However, compared to 3DGS [16], 4DGS [1] introduces an extra temporal dimension to enable dynamic representation. Previous 3DGS-based methods may therefore be unsuitable for 4DGS. Consequently, we first identify a key limitation leading to this problem, referred as temporal redundancy. Furthermore, we propose a novel pruning criterion leveraging spatial-temporal variation, and a temporal filter to achieve more efficient storage requirements and higher rendering speed. + +# 3 Preliminary of 4D Gaussian Splitting + +Our framework builds on 4D Gaussian Splitting (4DGS) [1], which reconstructs dynamic scenes by optimizing a collection of anisotropic 4D Gaussian primitives. For each Gaussian, it is characterized by a 4D mean $\mu = (\mu_x,\mu_y,\mu_z,\mu_t)\in \mathbb{R}^4$ coupled with a covariance matrix $\Sigma \in \mathbb{R}^{4\times 4}$ . + +By treating time and space dimensions equally, the 4D covariance matrix $\Sigma$ can be decomposed into a scaling matrix $S_{4D} = (s_x,s_y,s_z,s_t)\in \mathbb{R}^4$ and a rotation matrix $R_{4D}\in \mathbb{R}^{4\times 4}$ . $R_{4D}$ is represented by a pair of left quaternion $q_{l}\in \mathbb{R}^{4}$ and right quaternion $q_{r}\in \mathbb{R}^{4}$ . + +During rendering, each 4D Gaussian is decomposed into a conditional 3D Gaussian and a 1D Gaussian at a specific time $t$ . Moreover, the conditional 3D Gaussian can be derived from the properties of the multivariate Gaussian with: + +$$ +\mu_ {x y z \mid t} = \mu_ {1: 3} + \Sigma_ {1: 3, 4} \Sigma_ {4, 4} ^ {- 1} (t - \mu_ {t}) +$$ + +$$ +\Sigma_ {x y z | t} = \Sigma_ {1: 3, 1: 3} - \Sigma_ {1: 3, 4} \Sigma_ {4, 4} ^ {- 1} \Sigma_ {4, 1: 3} \tag {1} +$$ + +Here, $\mu_{1:3} \in \mathbb{R}^3$ and $\Sigma_{1:3,1:3} \in \mathbb{R}^{3 \times 3}$ denote the spatial mean and covariance, while $\mu_t$ and $\Sigma_{4,4}$ are scalars representing the temporal components. To perform rasterization, given a pixel under view $\mathcal{I}$ + +![](images/c01b20d6775b1d9f5043919e68a5d7ddcf0f83515724b268943038eb53cf2421.jpg) +(a) + +![](images/ab501d240515e0b77c414c9f8ca2cb0443d47029c68fbdfb6f1f271a5046e4b5.jpg) +(b) +Figure 2: Temporal redundancy Study. (a) The $\Sigma_{t}$ distribution of 4DGS. The red line shows the result of vanilla 4DGS. The other two lines represent our model has effectively reduced the number of transient Gaussians with small $\Sigma_{t}$ . (b) The active ratio during rendering at different timestamps. It demonstrates that most of the computation time is spent on inactive Gaussians in vanilla 4DGS. However, 4DGS-1K can significantly reduce the occurrence of inactive Gaussians during rendering to avoid unnecessary computations. (c) This figure shows the IoU between the set of active Gaussians in the first frame and frame t. It proves that active Gaussians tend to overlap significantly across adjacent frames. + +![](images/8685f8ae44bcb1cb192a5a81dff59e6bb0865519eb2615e7fa4c59a26cd373f0.jpg) +(c) + +and timestamp $t$ , its color $\mathcal{I}(u, v, t)$ can be computed by blending visible Gaussians that are sorted by their depth: + +$$ +\mathcal {I} (u, v, t) = \sum_ {i} ^ {N} c _ {i} (d) \alpha_ {i} \prod_ {j = 1} ^ {i - 1} (1 - \alpha_ {j}) \tag {2} +$$ + +with + +$$ +\alpha_ {i} = p _ {i} (t) p _ {t} (u, v | t) \sigma_ {i} +$$ + +$$ +p _ {i} (t) \sim \mathcal {N} \left(t; \mu_ {t}, \Sigma_ {4, 4}\right) \tag {3} +$$ + +where $c_{i}(d)$ is the color of each Gaussian, and $\alpha_{i}$ is given by evaluating a 2D Gaussian with covariance $\Sigma_{2D}$ multiplied with a learned per-point opacity $\sigma_{i}$ and temporal Gaussian distribution $p_{i}(t)$ . In the following discussion, we denote $\Sigma_{4,4}$ as $\Sigma_{t}$ for simplicity. + +Temporal Redundancy. Despite achieving high quality, 4DGS requires a huge number of Gaussians to model dynamic scenes. We identify a key limitation leading to this problem: 4DGS represents scenes through temporally independent Gaussians that lack explicit correlation across time. This means that, even static objects are redundantly represented by hundreds of Gaussians, which inconsistently appear or vanish across timesteps. We refer to this phenomenon as temporal redundancy. As a result, scenes end up needing more Gaussians than they should, leading to excessive storage demands and suboptimal rendering speeds. In Section 4, we analyze the root causes of this issue and propose a set of solutions to reduce the count of Gaussians. + +# 4 Methodology + +Our goal is to compress 4DGS by reducing the number of Gaussians while preserving rendering quality. To achieve this, we first analyze the redundancies present in 4DGS, as detailed in Section 4.1. Building on this analysis, we introduce 4DGS-1K in Section 4.2, which incorporates a set of compression techniques designed for 4DGS. 4DGS-1K enables rendering speeds of over 1,000 FPS on modern GPUs. + +# 4.1 Understanding Redundancy in 4DGS + +This section investigates why 4DGS requires an excessive number of Gaussians to represent dynamic scenes. In particular, we identify two key factors. First, 4DGS models object motion using a large number of transient Gaussians that inconsistently appear and disappear across timesteps, leading to redundant temporal representations. Second, for each frame, only a small fraction of Gaussians actually contribute to the rendering. We discuss those problems below. + +Massive Short-Lifespan Gaussians. We observe that 4DGS tends to store numerous Gaussians that flicker in time. We refer to these as Short-Lifespan Gaussians. To investigate this property, + +![](images/754dfb6a45ff6a4576d73ee3ee6f5eaf5cd1d85a61d4085d8987baaef805544f.jpg) +(a) Transient Gaussian Pruning + +![](images/26791847cf2d45431ecdd901cc943cf72257565110b5042df5bb48efda4c404b.jpg) +(b) Temporal Filter +Figure 3: Overview of 4DGS-1K. (a) We first calculate the spatial-temporal variation score for each 4D Gaussian on training views, to prune Gaussians with short lifespan (The Red Gaussian). (b) The temporal filter is introduced to filter out inactive Gaussians before the rendering process to alleviate suboptimal rendering speed. At a given timestamp $t$ , the set of Gaussians participating in rendering is derived from the two adjacent key-frames, $t_0$ and $t_{0 + \Delta_t}$ . + +we analyze the Gaussians' opacity, which controls visibility. Intuitively, Short-Lifespan Gaussians exhibit an opacity pattern that rapidly increases and then suddenly decreases. In 4DGS, this behavior is typically reflected in the time variance parameter $\Sigma_{t}$ —small $\Sigma_{t}$ values indicate a short lifespan. + +Observations. Specifically, we plot the distribution of $\Sigma_t$ for all Gaussians in the Sear Steak scene. As shown in Figure 2a, most of Gaussians has small $\Sigma_t$ values (e.g. $70\%$ have $\Sigma_t < 0.25$ ). + +Therefore, in 4DGS, nearly all Gaussians have a short lifespan. This property leads to high storage needs and slower rendering. + +Inactive Gaussians. Another finding is that, during the forward rendering, actually, only a small fraction of Gaussians are contributing. Interestingly, active ones tend to overlap significantly across adjacent frames. To quantify this, we introduce two metrics: (1) Active ratio. This ratio is defined as the proportion of the total number of active Gaussians across all views at any moment relative to the total number of Gaussians. (2) Activation Intersection-over-Union (IoU). This is computed as IoU between the set of active Gaussians in the first frame and in frame $t$ . + +Observations. Again, we plot the two metrics from Sear Steak scene. As shown in Figure 2b, nearly $85\%$ of Gaussians are inactive at each frame, even though all Gaussians are processed during rendering. Moreover, Figure 2c demonstrates that the active Gaussians remain quite consistent over time, with an IoU above $80\%$ over a 20-frame window. + +The inactive gaussians bring a significant issue in 4DGS, because each 4D Gaussian must be decomposed into a 3D Gaussian and a 1D Gaussian before rasterization (see eq. (1)). Therefore, a large portion of computational resources is wasted on inactive Gaussians. + +In summary, redundancy in 4DGS comes from massive Short-Lifespan Gaussians and inactive Gaussians. These insights motivate our compression strategy to eliminate redundant computations while preserving rendering quality. + +# 4.2 4DGS-1K for Fast Dynamic Scene Rendering + +Building on the analysis above, we introduce 4DGS-1K, a suite of compression techniques specifically designed for 4DGS to eliminate redundant Gaussians. As shown in Figure 3, this process involves two key steps. First, we identify and globally prune unimportant Gaussians with low Spatial-Temporal Variation Score in Section 4.2.1. Second, we apply local pruning using a temporal filter to inactive Gaussians that are not needed at each timestep in Section 4.2.2. + +# 4.2.1 Pruning with Spatial-Temporal Variation Score + +We first prune unimportant 4D Gaussians to improve efficiency. Like 3DGS, we remove those that have a low impact on rendered pixels. Besides, we additionally remove short-lifespan Gaus- + +sians—those that persist only briefly over time. To achieve this, we introduce a novel spatial-temporal variation score as the pruning criterion for 4DGS. It is composed of two parts, spatial score that measures the Gaussians contributions to the pixels in rendering, and temporal score considering the lifespan of Gaussians. + +Spatial score. Inspired by the previous method [30, 31] and $\alpha$ -blending in 3DGS [16], we define the spatial score by aggregating the ray contribution of Gaussian $g_{i}$ along all rays $r$ across all input images at a given timestamp. It can accurately capture the contribution of each Gaussian to one pixel. Consequently, the spatial contribution score $S^S$ is obtained by traversing all pixels: + +$$ +\mathcal {S} _ {i} ^ {S} = \sum_ {k = 1} ^ {N H W} \alpha_ {i} \prod_ {j = 1} ^ {i - 1} \left(1 - \alpha_ {j}\right) \tag {4} +$$ + +where $\mathbf{N}$ denotes the number of training views, $\mathrm{H},\mathrm{W}$ denote the height and width of the images, and $\alpha_{i}\prod_{j = 1}^{i - 1}(1 - \alpha_{j})$ reflects the contribution of $i^{th}$ Gaussian to the final color of all pixels according to the alpha composition in eq. (2). + +Temporal score. It is expected to assign a higher temporal score to Gaussians with a longer lifespan. To quantify this, we compute the second derivative of temporal opacity function $p_i(t)$ defined in eq. (3). The second derivative $p_i^{(2)}(t)$ is computed as + +$$ +p _ {i} ^ {(2)} (t) = \left(\frac {\left(t - \mu_ {t}\right) ^ {2}}{\Sigma_ {t} ^ {2}} - \frac {1}{\Sigma_ {t}}\right) p _ {i} (t) \tag {5} +$$ + +Intuitively, large second derivative magnitude corresponds to unstable, short-lived Gaussians, while low second derivative indicates smooth, persistent ones. + +Moreover, since the second derivative spans the real number domain $\mathbb{R}$ , we apply tanh function to map it to the interval $(0,1)$ . Consequently, the score for opacity variation, $S_{i}^{TV}$ , of each Gaussian $g_{i,t}$ is expressed as: + +$$ +\mathcal {S} _ {i} ^ {T V} = \sum_ {t = 0} ^ {T} \frac {1}{0 . 5 \cdot \operatorname {t a n h} \left(\left| p _ {i} ^ {(2)} (t) \right|\right) + 0 . 5}. \tag {6} +$$ + +In addition to the opacity range rate, the volume of 4D Gaussians is necessary to be considered, as described in eq. (1). The volume should be normalized following the method in [30], denoted as $\gamma(S^{4D}) = \text{Norm}(V(S^{4D}))$ . Therefore, the final temporal score $S_{i}^{T} = S_{i}^{TV}\gamma(S_{i}^{4D})$ + +Finally, by aggregating both spatial and temporal score, the spatial-temporal variation score $S_{i}$ can be written as: + +$$ +\mathcal {S} _ {i} = \sum_ {t = 0} ^ {T} \mathcal {S} _ {i} ^ {T} \mathcal {S} _ {i} ^ {S} \tag {7} +$$ + +Pruning. All 4D Gaussians are ranked based on their spatial-temporal variation score $S_{i}$ , and Gaussians with lower scores are pruned to reduce the storage burden of 4DGS [1]. The remaining Gaussians are optimized over a set number of iterations to compensate for minor losses resulting from pruning. + +# 4.2.2 Fast rendering with temporal filtering + +Our analysis reveals that inactive Gaussians induces unnecessary computations in 4DGS, significantly slowing down rendering. To address this issue, we introduce a temporal filter that dynamically selects active Gaussians. We observed that active Gaussians in adjacent frames overlap considerably (as detailed in Section 4.1), which allows us to share their corresponding masks across a window of frames. + +Key-frame based Temporal Filtering. Based on this observation, we design a key-frame based temporal filtering for active Gaussians. We select sparse key-frames at even intervals and share their masks with surrounding frames. + +Specifically, we select a list of key-frame timestamps $\{t_i\}_{i=0}^T$ , where $T$ depends on the chosen interval $\Delta_t$ . For each $t_i$ , we render the images from all training views at current timestamp and calculate the visibility list $\{m_{i,j}\}_{j=1}^N$ , where $m_{i,j}$ is the visibility mask obtained by eq. (2) from the $j^{th}$ training + +Table 1: Quantitative comparisons on the Neural 3D Video Dataset. + +
MethodPSNR↑SSIM↑LPIPS↓Storage(MB)↓FPS↑Raster FPS↑#Gauss↓
Neural Volume1[4]22.80-0.295----
DyNeRF1[2]29.58-0.083280.015--
StreamRF[18]28.26--531010.90--
HyperReel[5]31.100.9270.0963602.00--
K-Planes[6]31.63-0.0183110.30--
4K4D[36]21.29--2519290--
Dynamic 3DGS[37]30.670.9300.0992764460--
4D Gaussian[7]31.150.9400.0499030--
E-D3DGS[26]31.310.9450.0373574--
Swift4D[38]32.23-0.043120125--
Grid4D[39]31.49--146116--
STG[40]32.050.9460.044200140--
4D-RotorGS[41]31.620.9400.140-277--
MEGA[42]31.49-0.0562577--
Compact3D[29]31.690.9450.05415186--
4DGS[1]32.01-0.055-114--
4DGS2[1]31.910.9460.0522085901183333160
Ours31.880.9460.0524188051092666632
Ours-PP31.870.9440.053508051092666632
+ +1 The metrics of the model are tested without "coffee martini" and the resolution is set to $1024 \times 768$ . +2 The retrained model from the official implementation. + +Table 2: Quantitative comparisons on the D-NeRF Dataset. + +
MethodPSNR↑SSIM↑LPIPS↓Storage(MB)↓FPS↑Raster FPS↑#Gauss↓
DNeRF[19]29.670.950.08-0.1--
TiNeuVox[43]32.670.970.04-1.6--
K-Planes[6]31.070.970.02-1.2--
4D Gaussian[7]32.990.970.0518104--
Deformable3DGS[23]40.430.990.012770-131428
SC-GS[44]40.65--28126--
Grid4D[39]39.91--93166--
4D-RotorGS[41]34.260.970.031121257--
4DGS[1]34.090.980.02----
4DGS1[1]32.990.970.032783761232445076
Ours33.340.970.03421462248266460
Ours-PP33.370.970.0371462248266460
+ +1 The retrained model from the official implementation. + +viewpoint at timestamp $t_i$ and $N$ is the number of training views at current timestamp. The final set of active Gaussian masks is given by $\left\{\bigcup_{j=1}^{N} m_{i,j}\right\}_{i=0}^{T}$ . + +Filter based Rendering. To render the images from any viewpoint at a given timestamp $t_{test}$ , we consider its two nearest key-frames, denoted as $t_l$ and $t_r$ . Then, we perform rasterization while only considering the Gaussians marked by mask $\left\{\bigcup_{j=1}^{N} m_{i,j}\right\}_{i=l,r}$ . This method explicitly filters out inactive Gaussians to speed up rendering. + +Note that using long intervals may overlook some Gaussians, reducing rendering quality. Therefore, we fine-tune Gaussians recorded by the masks to compensate for losses. + +# 5 Experiment + +# 5.1 Experimental Settings + +Datasets. We utilize two dynamic scene datasets to demonstrate the effectiveness of our method: (1) Neural 3D Video Dataset (N3V) [2]. This dataset consists of six dynamic scenes, and the resolution is $2704 \times 2028$ . For a fair comparison, we align with previous work [1, 40] by conducting evaluations at a half-resolution of 300 frames. (2) D-NeRF Dataset [19]. This dataset is a monocular video dataset comprising eight videos of synthetic scenes. We choose standard test views that originate from novel camera positions not encountered during the training process. + +Evaluation Metrics. To evaluate the quality of rendering dynamic scenes, we employ several commonly used image quality assessment metrics: Peak Signal-to-Noise Ratio (PSNR), Structural + +![](images/a18b9956322a2c02c41121e9d67531e89dbf246cdf7756c7c1aad71b5d3498dd.jpg) + +![](images/cc9f764867bbce90a59abe0073a0fe9082faf051c2460776b2ef1989f31407d5.jpg) + +![](images/ce68cf0094a6b90c454734f4e852eebf300507edcc8e8e1875aa01699d16b653.jpg) + +![](images/99e3572a8a9c15ce78fe6ae72f9e98c1da50ca096164134a0e2df108d7a7b374.jpg) + +![](images/8532c4bf5897410caacdf2eb87b72f0bcfd9f43ae0be93bc4b8156060467af52.jpg) + +![](images/40ac77f10bcbf78be97acb1d262432e7c251133e8806ccb3eeba5dae28f5834d.jpg) + +![](images/297f630aaca124c167b8302f8eabf6381f71f14a01065747ccf3e7d0feec6da8.jpg) +(a) Results on Sear Steak Scene. + +![](images/f7481c792cf6f7d54f318399678cfcde82dd099f3b1b7baeb52e60c97357b4f7.jpg) + +![](images/9ffe9f81ad96656f1a30329a3f66919e5d3033708628f62a1cfe0ee7abef59da.jpg) + +![](images/1295c55e7497f076910fac7f9f644adb72c16a70cf470fbe693f571e2c60b3d7.jpg) + +![](images/0f16e907d14697fedab9fd727b3ca7ca67f65fc6b86e7236492417beda0b885c.jpg) + +![](images/7a3db6c9e8409646a0663439408137993b8dee046ce3baa17f83909ad7b76965.jpg) + +![](images/0bd678ba72a4ad673a313daeb97525b83b9d51356e254eb83b041ed17aad4823.jpg) + +![](images/12babc436717259b6ce823ea0cbabade2863a45f09b02e6fa532807b52a475cd.jpg) + +![](images/7edff9f508f2a2ca896f926afd7a84e85d43c9d96457bfbf7c76263274cfd98b.jpg) + +![](images/3da27150b83e2a877d270eec48b43dd577d16ac7699cd42be0de807a68d080d6.jpg) + +![](images/a7c0a266f29deebdfd8eaec1ba81310b5364d7214e00ef019bd011a0b4af4ff7.jpg) +Figure 4: Qualitative comparisons of 4DGS and our method. + +![](images/212cea1e903c4d4907493655cda30a72bfa177bfdd82d6bd25266e7107246e1e.jpg) +(b) Results on Trex Scene. + +![](images/0d42203e5bdaaa10dc448bc09515668b989c1e7c92087402f208f6e1e371d8f7.jpg) + +![](images/2d10f076b87e2f81bc7388f6c6c36edea412c33639284694d354f4d1d8671263.jpg) + +![](images/24e336f7a0d464e783f93d43f460b1927359d9c4979dd512974c1d1e837b9572.jpg) + +Table 3: Ablation study of per-component contribution. + +
IDMethod\DatasetPSNR↑SSIM↑LPIPS↓Storage(MB)↓FPS↑Raster FPS↑#Gauss↓
FilterPruningPP
avanilla 4DGS131.910.94580.05182085901183333160
b✓1,231.510.94460.053920912425613333160
c✓229.560.93540.060520913005613333160
d✓31.920.94620.0513417312600666632
e✓✓31.880.94570.05244188051092666632
f✓2✓31.630.94520.05244187891080666632
g✓✓✓31.870.94440.0532508051092666632
+ +1 The result with environment map. 2 The result without finetuning. + +Similarity Index Measure (SSIM), and Learned Perceptual Image Patch Similarity (LPIPS) [45]. Following the previous work, LPIPS [45] is computed using AlexNet [46] and VGGNet [47] on the N3V dataset and the D-NeRF dataset, respectively. Moreover, we report the number of Gaussians and storage. To demonstrate the improvement in rendering speed, we report two types of FPS: (1) FPS. It considers the entire rendering function. Due to interference from other operations, it can't effectively demonstrate the acceleration achieved by our method. (2) Raster FPS. It only considers the rasterization, the most computationally intensive component during rendering. + +Baselines. Our primary baseline for comparison is 4DGS [1], which serves as the foundation of our model. Moreover, we compare 4DGS-1K with two concurrent works on 4D compression, MEGA [42] and Compact3D [29]. Certainly, we conduct comparisons with 4D-RotorGS [41] which is another form of representation for 4D Gaussian Splating with the capability for real-time rendering speed and high-fidelity rendering results. In addition, we also compare our work against NeRF-based methods, like Neural Volume [4], DyNeRF [2], StreamRF [18], HyperReel [5], DNeRF [19], K-Planes [6] and 4K4D [36]. Furthermore, other recent competitive Gaussian-based methods are also considered in our comparison, including Dynamic 3DGS [37], STG [40], 4DGAaussian [7], E-D3DGS [26], Swift4D [38], Grid4D [39] and SC-GS [44]. + +Implementation Details. Our method is tested in a single RTX 3090 GPU. We train our model following the experiment setting in 4DGS [1]. After training, we perform the pruning and filtering strategy. Then, we fine-tune 4DGS-1K for 5,000 iterations while disabling additional clone/split operations. For pruning strategy, the pruning ratio is set to $80\%$ on the N3V Dataset, and $85\%$ on the D-NeRF Dataset. For the temporal filtering, we set the interval $\Delta_t$ between key-frames to 20 frames + +on the N3V Dataset. Considering the varying capture speeds on the D-NeRF dataset, we select 6 key-frames rather than a specific frame interval. Additionally, to further compress the storage of 4DGS [1], we implement post-processing techniques in our model, denoted as Ours-PP. It includes vector quantization [28] on SH of Gaussians and compressing the mask of filter into bits. + +Note that we don't apply environment maps implemented by 4DGS on Coffee Martini and Flame Salmon scenes, which significantly affects the rendering speed. Subsequent results indicate that removing it for 4DGS-1K does not significantly degrade the rendering quality. + +# 5.2 Results and Comparisons + +Comparisons on real-world dataset. Table 1 presents a quantitative evaluation on the N3V dataset. 4DGS-1K achieves rendering quality comparable to the current baseline. Compared to 4DGS [1], we achieve a $41\times$ compression and $9\times$ faster in rendering speed at the cost of a $0.04dB$ reduction in PSNR. In addition, compared to MEGA [42] and Compact3D [29], two concurrent works on 4D compression, the rendering speeds are $10\times$ and $4\times$ faster respectively while maintaining a comparable storage requirement and high quality reconstruction. Moreover, the FPS of 4DGS-1K far exceeds the current state-of-the-art levels. It is nearly twice as fast as the current fastest model, Dynamic 3DGS [37] while requiring only $1\%$ of the storage size. Additionally, 4DGS-1K achieves better visual quality than that of Dynamic 3DGS [37], with an increase of about $1.2dB$ in PSNR. Compared to the storage-efficient model, E-D3DGS [26] and DyNeRF [2] we achieve an increase of over $0.5dB$ in PSNR and fast rendering speed. Figure 4 offers qualitative comparisons for the Sear Steak, demonstrating that our results contain more vivid details. + +Comparisons on synthetic dataset. In our experiments, we benchmarked 4DGS-1K against several baselines using the monocular synthetic dataset introduced by D-NeRF [19]. The result is shown in Table 2. Compared to 4DGS [1], our method achieves up to $40\times$ compression and $4\times$ faster rendering speed. It is worth noting that the rendering quality of our model even surpasses that of the original 4DGS, with an increase of about $0.38dB$ in PSNR. Furthermore, our approach exhibits higher rendering quality and smaller storage overhead compared to most Gaussian-based methods. We provide qualitative results in Figure 4 for a more visual assessment. + +# 5.3 Ablation Study + +To evaluate the contribution of each component, we conducted ablation experiments on the N3V dataset [2]. More ablations are provided in the supplement(See appendix B). + +Pruning. As shown in Table 3, our pruning strategy achieves $5 \times$ compression ratio and $5 \times$ faster rasterization speed while slightly improving rendering quality. As shown in Figure 2a, our pruning strategy also reduces the presence of Gaussians with short lifespan. As such, 4DGS-1k processes far fewer unnecessary Gaussians (See Figure 2b) during rendering. Moreover, as shown in Figure 2c, the pruning process expands the range of adjacent frames. It allows larger intervals for the temporal filter. + +Temporal Filtering. As illustrated in Table 3, the results of b and c are obtained by directly applying the filter without fine-tuning. It proves that this component can enhance the rendering speed of 4DGS. However, as mentioned in Section 4.1, the 4DGS contains a huge number of short lifespan Gaussians. It results in some Gaussians being overlooked in the filter, causing a slight decrease in rendering quality. However, through pruning, most Gaussians are ensured to have long lifespan, making them visible even at large intervals. Therefore, it alleviates the issue of Gaussians being overlooked (See f). Furthermore, appropriate fine-tuning allows the Gaussians in the active Gaussians list to relearn the scene features to compensate for the loss incurred by the temporal filter (See e and f). + +# 6 Conclusion + +In this paper, we present 4DGS-1K, a compact and memory-efficient dynamic scene representation capable of running at over 1000 FPS on modern GPUs. We introduce a novel pruning criterion called the spatial-temporal variation score, which eliminates a significant number of redundant Gaussian points in 4DGS, drastically reducing storage requirements. Additionally, we propose a temporal filter that selectively activates only a subset of Gaussians during each frame's rendering. This approach enables our rendering speed to far surpass that of existing baselines. Compared to vanilla 4DGS, + +4DGS-1K achieves a $41\times$ reduction in storage and $9\times$ faster rasterization speed while maintaining high-quality reconstruction. + +# Acknowledgement + +This project is supported by the National Research Foundation, Singapore, under its Medium Sized Center for Advanced Robotics Technology Innovation. + +# References + +[1] Zeyu Yang, Hongye Yang, Zijie Pan, and Li Zhang. Real-time photorealistic dynamic scene representation and rendering with 4d gaussian splatting. arXiv preprint arXiv:2310.10642, 2023. +[2] Tianye Li, Mira Slavcheva, Michael Zollhoefer, Simon Green, Christoph Lassner, Changil Kim, Tanner Schmidt, Steven Lovegrove, Michael Goesele, Richard Newcombe, et al. Neural 3d video synthesis from multi-view video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5521-5531, 2022. +[3] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65 (1):99-106, 2021. +[4] Stephen Lombardi, Tomas Simon, Jason Saragih, Gabriel Schwartz, Andreas Lehrmann, and Yaser Sheikh. Neural volumes: Learning dynamic renderable volumes from images. arXiv preprint arXiv:1906.07751, 2019. +[5] Benjamin Attal, Jia-Bin Huang, Christian Richardt, Michael Zollhoefer, Johannes Kopf, Matthew O'Toole, and Changil Kim. Hyperreel: High-fidelity 6-dof video with ray-conditioned sampling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16610-16620, 2023. +[6] Sara Fridovich-Keil, Giacomo Meanti, Frederik Rahbæk Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12479–12488, 2023. +[7] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20310-20320, 2024. +[8] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European conference on computer vision, pages 333-350. Springer, 2022. +[9] Katja Schwarz, Axel Sauer, Michael Niemeyer, Yiyi Liao, and Andreas Geiger. Voxgraf: Fast 3d-aware image synthesis with sparse voxel grids. Advances in Neural Information Processing Systems, 35:33999-34011, 2022. +[10] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5459-5469, 2022. +[11] Liao Wang, Jiakai Zhang, Xinhang Liu, Fuqiang Zhao, Yanshun Zhang, Yingliang Zhang, Minye Wu, Jingyi Yu, and Lan Xu. Fourier plenoc trees for dynamic radiance field rendering in real-time. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13524-13534, 2022. +[12] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5501–5510, 2022. +[13] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM transactions on graphics (TOG), 41(4):1-15, 2022. +[14] Christian Reiser, Songyou Peng, Yiyi Liao, and Andreas Geiger. Kilonerf: Speeding up neural radiance fields with thousands of tiny mlp's. In Proceedings of the IEEE/CVF international conference on computer vision, pages 14335-14345, 2021. +[15] Huan Wang, Jian Ren, Zeng Huang, Kyle Olszewski, Mengei Chai, Yun Fu, and Sergey Tulyakov. R2l: Distilling neural radiance field to neural light field for efficient novel view synthesis. In European Conference on Computer Vision, pages 612-629. Springer, 2022. + +[16] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. +[17] Ben Mildenhall, Pratul P Srinivasan, Rodrigo Ortiz-Cayon, Nima Khademi Kalantari, Ravi Ramamoorthi, Ren Ng, and Abhishek Kar. Local light field fusion: Practical view synthesis with prescriptive sampling guidelines. ACM Transactions on Graphics (ToG), 38(4):1-14, 2019. +[18] Lingzhi Li, Zhen Shen, Zhongshu Wang, Li Shen, and Ping Tan. Streaming radiance fields for 3d video synthesis. Advances in Neural Information Processing Systems, 35:13485-13498, 2022. +[19] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10318–10327, 2021. +[20] Ang Cao and Justin Johnson. Hexplane: A fast representation for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 130-141, 2023. +[21] Liangchen Song, Anpei Chen, Zhong Li, Zhang Chen, Lele Chen, Junsong Yuan, Yi Xu, and Andreas Geiger. Nerfplayer: A streamable dynamic scene representation with decomposed neural radiance fields. IEEE Transactions on Visualization and Computer Graphics, 29(5):2732-2742, 2023. +[22] Feng Wang, Sinan Tan, Xinghang Li, Zeyue Tian, Yafei Song, and Huaping Liu. Mixed neural voxels for fast multi-view video synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19706-19716, 2023. +[23] Ziyi Yang, Xinyu Gao, Wen Zhou, Shaohui Jiao, Yuqing Zhang, and Xiaogang Jin. Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20331-20341, 2024. +[24] Zhiyang Guo, Wengang Zhou, Li Li, Min Wang, and Houqiang Li. Motion-aware 3d gaussian splatting for efficient dynamic scene reconstruction. arXiv preprint arXiv:2403.11447, 2024. +[25] Zhicheng Lu, Xiang Guo, Le Hui, Tianrui Chen, Min Yang, Xiao Tang, Feng Zhu, and Yuchao Dai. 3d geometry-aware deformable gaussian splatting for dynamic view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8900-8910, 2024. +[26] Jeongmin Bae, Seoha Kim, Youngsik Yun, Hahyun Lee, Gun Bang, and Youngjung Uh. Per-gaussian embedding-based deformation for deformable 3d gaussian splatting. arXiv preprint arXiv:2404.03613, 2024. +[27] Devikalyan Das, Christopher Wewer, Raza Yunus, Eddy Ilg, and Jan Eric Lenssen. Neural parametric gaussians for monocular non-rigid object reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10715-10725, 2024. +[28] KL Navaneet, Kossar Pourahmadi Meibodi, Soroush Abbasi Koohpayegani, and Hamed Pirsiavash. Compgs: Smaller and faster gaussian splatting with vector quantization. In European Conference on Computer Vision, 2024. +[29] Joo Chan Lee, Daniel Rho, Xiangyu Sun, Jong Hwan Ko, and Eunbyung Park. Compact 3d gaussian representation for radiance field. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21719-21728, 2024. +[30] Zhiwen Fan, Kevin Wang, Kairun Wen, Zehao Zhu, Dejia Xu, and Zhangyang Wang. Lightgaussian: Unbounded 3d gaussian compression with $15\mathrm{x}$ reduction and $200+$ fps. arXiv preprint arXiv:2311.17245, 2023. +[31] Guangchi Fang and Bing Wang. Mini-splatting: Representing scenes with a constrained number of gaussians. arXiv preprint arXiv:2403.14166, 2024. +[32] Michael Niemeyer, Fabian Manhardt, Marie-Julie Rakotosaona, Michael Oechsle, Daniel Duckworth, Rama Gosula, Keisuke Tateno, John Bates, Dominik Kaeser, and Federico Tombari. Radsvat: Radiance field-informed gaussian splatting for robust real-time rendering with $900+$ fps. arXiv preprint arXiv:2403.13806, 2024. +[33] Muhammad Salman Ali, Maryam Qamar, Sung-Ho Bae, and Enzo Tartaglione. Trimming the fat: Efficient compression of 3d gaussian splats through pruning. arXiv preprint arXiv:2406.18214, 2024. +[34] Panagiotis Papantonakis, Georgios Kopanas, Bernhard Kerbl, Alexandre Lanvin, and George Drettakis. Reducing the memory footprint of 3d gaussian splattering. Proceedings of the ACM on Computer Graphics and Interactive Techniques, 7(1):1-17, 2024. + +[35] Wenkai Liu, Tao Guan, Bin Zhu, Lili Ju, Zikai Song, Dan Li, Yuesong Wang, and Wei Yang. Efficientgs: Streamlining gaussian splatting for large-scale high-resolution scene representation. arXiv preprint arXiv:2404.12777, 2024. +[36] Zhen Xu, Sida Peng, Haotong Lin, Guangzhao He, Jiaming Sun, Yujun Shen, Hujun Bao, and Xiaowei Zhou. 4k4d: Real-time 4d view synthesis at 4k resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 20029-20040, 2024. +[37] Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. arXiv preprint arXiv:2308.09713, 2023. +[38] Jiahao Wu, Rui Peng, Zhiyan Wang, Lu Xiao, Luyang Tang, Jinbo Yan, Kaiqiang Xiong, and Ronggang Wang. Swift4d: Adaptive divide-and-conquer gaussian splattering for compact and efficient reconstruction of dynamic scene. arXiv preprint arXiv:2503.12307, 2025. +[39] Jiawei Xu, Zexin Fan, Jian Yang, and Jin Xie. Grid4d: 4d decomposed hash encoding for high-fidelity dynamic gaussian splatting. Advances in Neural Information Processing Systems, 37:123787-123811, 2024. +[40] Zhan Li, Zhang Chen, Zhong Li, and Yi Xu. Spacetime gaussian feature splatting for real-time dynamic view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8508-8520, 2024. +[41] Yuanxing Duan, Fangyin Wei, Qiyu Dai, Yuhang He, Wenzheng Chen, and Baoquan Chen. 4d-rotor gaussian splatting: towards efficient novel view synthesis for dynamic scenes. In ACM SIGGRAPH 2024 Conference Papers, pages 1–11, 2024. +[42] Xinjie Zhang, Zhening Liu, Yifan Zhang, Xingtong Ge, Dailan He, Tongda Xu, Yan Wang, Zehong Lin, Shuicheng Yan, and Jun Zhang. Mega: Memory-efficient 4d gaussian splatting for dynamic scenes. arXiv preprint arXiv:2410.13613, 2024. +[43] Jiemin Fang, Taoran Yi, Xinggang Wang, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Matthias Nießner, and Qi Tian. Fast dynamic radiance fields with time-aware neural voxels. In SIGGRAPH Asia 2022 Conference Papers, pages 1-9, 2022. +[44] Yi-Hua Huang, Yang-Tian Sun, Ziyi Yang, Xiaoyang Lyu, Yan-Pei Cao, and Xiaojuan Qi. Sc-gs: Sparse-controlled gaussian splatting for editable dynamic scenes. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4220-4230, 2024. +[45] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. +[46] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. Advances in neural information processing systems, 25, 2012. +[47] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. + +# NeurIPS Paper Checklist + +# 1. Claims + +Question: Do the main claims made in the abstract and introduction accurately reflect the paper's contributions and scope? + +Answer: [Yes] + +Justification: The main claims made in the abstract and introduction accurately reflect the paper's contributions and scope. + +Guidelines: + +- The answer NA means that the abstract and introduction do not include the claims made in the paper. +- The abstract and/or introduction should clearly state the claims made, including the contributions made in the paper and important assumptions and limitations. A No or NA answer to this question will not be perceived well by the reviewers. +- The claims made should match theoretical and experimental results, and reflect how much the results can be expected to generalize to other settings. +- It is fine to include aspirational goals as motivation as long as it is clear that these goals are not attained by the paper. + +# 2. Limitations + +Question: Does the paper discuss the limitations of the work performed by the authors? + +Answer: [Yes] + +Justification: We discuss it in supplementary material. + +Guidelines: + +- The answer NA means that the paper has no limitation while the answer No means that the paper has limitations, but those are not discussed in the paper. +- The authors are encouraged to create a separate "Limitations" section in their paper. +- The paper should point out any strong assumptions and how robust the results are to violations of these assumptions (e.g., independence assumptions, noiseless settings, model well-specification, asymptotic approximations only holding locally). The authors should reflect on how these assumptions might be violated in practice and what the implications would be. +- The authors should reflect on the scope of the claims made, e.g., if the approach was only tested on a few datasets or with a few runs. In general, empirical results often depend on implicit assumptions, which should be articulated. +- The authors should reflect on the factors that influence the performance of the approach. For example, a facial recognition algorithm may perform poorly when image resolution is low or images are taken in low lighting. Or a speech-to-text system might not be used reliably to provide closed captions for online lectures because it fails to handle technical jargon. +- The authors should discuss the computational efficiency of the proposed algorithms and how they scale with dataset size. +- If applicable, the authors should discuss possible limitations of their approach to address problems of privacy and fairness. +- While the authors might fear that complete honesty about limitations might be used by reviewers as grounds for rejection, a worse outcome might be that reviewers discover limitations that aren't acknowledged in the paper. The authors should use their best judgment and recognize that individual actions in favor of transparency play an important role in developing norms that preserve the integrity of the community. Reviewers will be specifically instructed to not penalize honesty concerning limitations. + +# 3. Theory assumptions and proofs + +Question: For each theoretical result, does the paper provide the full set of assumptions and a complete (and correct) proof? + +Answer: [NA] + +Justification: This paper does not involve theoretical result. + +# Guidelines: + +- The answer NA means that the paper does not include theoretical results. +- All the theorems, formulas, and proofs in the paper should be numbered and cross-referenced. +- All assumptions should be clearly stated or referenced in the statement of any theorems. +- The proofs can either appear in the main paper or the supplemental material, but if they appear in the supplemental material, the authors are encouraged to provide a short proof sketch to provide intuition. +- Inversely, any informal proof provided in the core of the paper should be complemented by formal proofs provided in appendix or supplemental material. +- Theorems and Lemmas that the proof relies upon should be properly referenced. + +# 4. Experimental result reproducibility + +Question: Does the paper fully disclose all the information needed to reproduce the main experimental results of the paper to the extent that it affects the main claims and/or conclusions of the paper (regardless of whether the code and data are provided or not)? + +# Answer: [Yes] + +Justification: The paper discloses all the information needed to reproduce the main experimental results of the paper in Section 5. + +# Guidelines: + +- The answer NA means that the paper does not include experiments. +- If the paper includes experiments, a No answer to this question will not be perceived well by the reviewers: Making the paper reproducible is important, regardless of whether the code and data are provided or not. +- If the contribution is a dataset and/or model, the authors should describe the steps taken to make their results reproducible or verifiable. +- Depending on the contribution, reproducibility can be accomplished in various ways. For example, if the contribution is a novel architecture, describing the architecture fully might suffice, or if the contribution is a specific model and empirical evaluation, it may be necessary to either make it possible for others to replicate the model with the same dataset, or provide access to the model. In general, releasing code and data is often one good way to accomplish this, but reproducibility can also be provided via detailed instructions for how to replicate the results, access to a hosted model (e.g., in the case of a large language model), releasing of a model checkpoint, or other means that are appropriate to the research performed. +- While NeurIPS does not require releasing code, the conference does require all submissions to provide some reasonable avenue for reproducibility, which may depend on the nature of the contribution. For example +(a) If the contribution is primarily a new algorithm, the paper should make it clear how to reproduce that algorithm. +(b) If the contribution is primarily a new model architecture, the paper should describe the architecture clearly and fully. +(c) If the contribution is a new model (e.g., a large language model), then there should either be a way to access this model for reproducing the results or a way to reproduce the model (e.g., with an open-source dataset or instructions for how to construct the dataset). +(d) We recognize that reproducibility may be tricky in some cases, in which case authors are welcome to describe the particular way they provide for reproducibility. In the case of closed-source models, it may be that access to the model is limited in some way (e.g., to registered users), but it should be possible for other researchers to have some path to reproducing or verifying the results. + +# 5. Open access to data and code + +Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material? + +# Answer: [No] + +Justification: The code will be open-sourced upon acceptance. + +# Guidelines: + +- The answer NA means that paper does not include experiments requiring code. +- Please see the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details. +- While we encourage the release of code and data, we understand that this might not be possible, so "No" is an acceptable answer. Papers cannot be rejected simply for not including code, unless this is central to the contribution (e.g., for a new open-source benchmark). +- The instructions should contain the exact command and environment needed to run to reproduce the results. See the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details. +- The authors should provide instructions on data access and preparation, including how to access the raw data, preprocessed data, intermediate data, and generated data, etc. +- The authors should provide scripts to reproduce all experimental results for the new proposed method and baselines. If only a subset of experiments are reproducible, they should state which ones are omitted from the script and why. +- At submission time, to preserve anonymity, the authors should release anonymized versions (if applicable). +- Providing as much information as possible in supplemental material (appended to the paper) is recommended, but including URLs to data and code is permitted. + +# 6. Experimental setting/details + +Question: Does the paper specify all the training and test details (e.g., data splits, hyperparameters, how they were chosen, type of optimizer, etc.) necessary to understand the results? + +# Answer: [Yes] + +Justification: The detailed experiment settings are listed in Section 5. + +# Guidelines: + +- The answer NA means that the paper does not include experiments. +- The experimental setting should be presented in the core of the paper to a level of detail that is necessary to appreciate the results and make sense of them. +- The full details can be provided either with the code, in appendix, or as supplemental material. + +# 7. Experiment statistical significance + +Question: Does the paper report error bars suitably and correctly defined or other appropriate information about the statistical significance of the experiments? + +# Answer: [NA] + +Justification: This paper follows the existing work about 4DGS and lists the essential detailed quantitative results in Section 5. None of them report error bars. + +# Guidelines: + +- The answer NA means that the paper does not include experiments. +- The authors should answer "Yes" if the results are accompanied by error bars, confidence intervals, or statistical significance tests, at least for the experiments that support the main claims of the paper. +- The factors of variability that the error bars are capturing should be clearly stated (for example, train/test split, initialization, random drawing of some parameter, or overall run with given experimental conditions). +- The method for calculating the error bars should be explained (closed form formula, call to a library function, bootstrap, etc.) +- The assumptions made should be given (e.g., Normally distributed errors). + +- It should be clear whether the error bar is the standard deviation or the standard error of the mean. +- It is OK to report 1-sigma error bars, but one should state it. The authors should preferably report a 2-sigma error bar than state that they have a $96\%$ CI, if the hypothesis of Normality of errors is not verified. +- For asymmetric distributions, the authors should be careful not to show in tables or figures symmetric error bars that would yield results that are out of range (e.g. negative error rates). +- If error bars are reported in tables or plots, The authors should explain in the text how they were calculated and reference the corresponding figures or tables in the text. + +# 8. Experiments compute resources + +Question: For each experiment, does the paper provide sufficient information on the computer resources (type of compute workers, memory, time of execution) needed to reproduce the experiments? + +Answer: [Yes] + +Justification: The paper provides sufficient information on the computer resources in Section 5 and supplemental material. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The paper should indicate the type of compute workers CPU or GPU, internal cluster, or cloud provider, including relevant memory and storage. +- The paper should provide the amount of compute required for each of the individual experimental runs as well as estimate the total compute. +- The paper should disclose whether the full research project required more compute than the experiments reported in the paper (e.g., preliminary or failed experiments that didn't make it into the paper). + +supplemental material. + +# 9. Code of ethics + +Question: Does the research conducted in the paper conform, in every respect, with the NeurIPS Code of Ethics https://neurips.cc/public/EthicsGuidelines? + +Answer: [Yes] + +Justification: The research conducted in this paper conform the NeurIPS Code of Ethics. + +Guidelines: + +- The answer NA means that the authors have not reviewed the NeurIPS Code of Ethics. +- If the authors answer No, they should explain the special circumstances that require a deviation from the Code of Ethics. +- The authors should make sure to preserve anonymity (e.g., if there is a special consideration due to laws or regulations in their jurisdiction). + +# 10. Broader impacts + +Question: Does the paper discuss both potential positive societal impacts and negative societal impacts of the work performed? + +Answer: [Yes] + +Justification: : This paper discusses the potential societal impacts in Section 6 and supplemental material. + +Guidelines: + +- The answer NA means that there is no societal impact of the work performed. +- If the authors answer NA or No, they should explain why their work has no societal impact or why the paper does not address societal impact. +- Examples of negative societal impacts include potential malicious or unintended uses (e.g., disinformation, generating fake profiles, surveillance), fairness considerations (e.g., deployment of technologies that could make decisions that unfairly impact specific groups), privacy considerations, and security considerations. + +- The conference expects that many papers will be foundational research and not tied to particular applications, let alone deployments. However, if there is a direct path to any negative applications, the authors should point it out. For example, it is legitimate to point out that an improvement in the quality of generative models could be used to generate deepfakes for disinformation. On the other hand, it is not needed to point out that a generic algorithm for optimizing neural networks could enable people to train models that generate Deepfakes faster. +- The authors should consider possible harms that could arise when the technology is being used as intended and functioning correctly, harms that could arise when the technology is being used as intended but gives incorrect results, and harms following from (intentional or unintentional) misuse of the technology. +- If there are negative societal impacts, the authors could also discuss possible mitigation strategies (e.g., gated release of models, providing defenses in addition to attacks, mechanisms for monitoring misuse, mechanisms to monitor how a system learns from feedback over time, improving the efficiency and accessibility of ML). + +# 11. Safeguards + +Question: Does the paper describe safeguards that have been put in place for responsible release of data or models that have a high risk for misuse (e.g., pretrained language models, image generators, or scraped datasets)? + +Answer: [NA] + +Justification: This paper does not pose such risk. + +Guidelines: + +- The answer NA means that the paper poses no such risks. +- Released models that have a high risk for misuse or dual-use should be released with necessary safeguards to allow for controlled use of the model, for example by requiring that users adhere to usage guidelines or restrictions to access the model or implementing safety filters. +- Datasets that have been scraped from the Internet could pose safety risks. The authors should describe how they avoided releasing unsafe images. +- We recognize that providing effective safeguards is challenging, and many papers do not require this, but we encourage authors to take this into account and make a best faith effort. + +# 12. Licenses for existing assets + +Question: Are the creators or original owners of assets (e.g., code, data, models), used in the paper, properly credited and are the license and terms of use explicitly mentioned and properly respected? + +Answer: [Yes] + +Justification: This paper follows the applicable licenses and terms of usage. + +Guidelines: + +- The answer NA means that the paper does not use existing assets. +- The authors should cite the original paper that produced the code package or dataset. +- The authors should state which version of the asset is used and, if possible, include a URL. +- The name of the license (e.g., CC-BY 4.0) should be included for each asset. +- For scraped data from a particular source (e.g., website), the copyright and terms of service of that source should be provided. +- If assets are released, the license, copyright information, and terms of use in the package should be provided. For popular datasets, paperswithcode.com/datasets has curated licenses for some datasets. Their licensing guide can help determine the license of a dataset. +- For existing datasets that are re-packaged, both the original license and the license of the derived asset (if it has changed) should be provided. + +- If this information is not available online, the authors are encouraged to reach out to the asset's creators. + +# 13. New assets + +Question: Are new assets introduced in the paper well documented and is the documentation provided alongside the assets? + +Answer: [NA] + +Justification: The paper does not release new assets. + +Guidelines: + +- The answer NA means that the paper does not release new assets. +- Researchers should communicate the details of the dataset/code/model as part of their submissions via structured templates. This includes details about training, license, limitations, etc. +- The paper should discuss whether and how consent was obtained from people whose asset is used. +- At submission time, remember to anonymize your assets (if applicable). You can either create an anonymized URL or include an anonymized zip file. + +# 14. Crowdsourcing and research with human subjects + +Question: For crowdsourcing experiments and research with human subjects, does the paper include the full text of instructions given to participants and screenshots, if applicable, as well as details about compensation (if any)? + +Answer: [NA] + +Justification: This paper does not involve crowdsourcing nor research with human subjects. + +Guidelines: + +- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. +- Including this information in the supplemental material is fine, but if the main contribution of the paper involves human subjects, then as much detail as possible should be included in the main paper. +- According to the NeurIPS Code of Ethics, workers involved in data collection, curation, or other labor should be paid at least the minimum wage in the country of the data collector. + +# 15. Institutional review board (IRB) approvals or equivalent for research with human subjects + +Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals (or an equivalent approval/review based on the requirements of your country or institution) were obtained? + +Answer: [NA] + +Justification: The paper does not involve crowdsourcing nor research with human subjects. + +Guidelines: + +- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. +- Depending on the country in which research is conducted, IRB approval (or equivalent) may be required for any human subjects research. If you obtained IRB approval, you should clearly state this in the paper. +- We recognize that the procedures for this may vary significantly between institutions and locations, and we expect authors to adhere to the NeurIPS Code of Ethics and the guidelines for their institution. +- For initial submissions, do not include any information that would break anonymity (if applicable), such as the institution conducting the review. + +# 16. Declaration of LLM usage + +Question: Does the paper describe the usage of LLMs if it is an important, original, or non-standard component of the core methods in this research? Note that if the LLM is used only for writing, editing, or formatting purposes and does not impact the core methodology, scientific rigorousness, or originality of the research, declaration is not required. + +# Answer: [NA] + +Justification: The core method development in this research does not involve LLMs. + +# Guidelines: + +- The answer NA means that the core method development in this research does not involve LLMs as any important, original, or non-standard components. +- Please refer to our LLM policy (https://neurips.cc/Conferences/2025/LLM) for what should or should not be described. \ No newline at end of file diff --git a/NeurIPS/2025/1000+ FPS 4D Gaussian Splatting for Dynamic Scene Rendering/images.zip b/NeurIPS/2025/1000+ FPS 4D Gaussian Splatting for Dynamic Scene Rendering/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..2919a0b847d8b8a69fb6901e625fa8147c6544f1 --- /dev/null +++ b/NeurIPS/2025/1000+ FPS 4D Gaussian Splatting for Dynamic Scene Rendering/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:afb1d7fc2320faa552bcb8fc89f169079e69269fafe77d9d31f84f395c12c656 +size 409186 diff --git a/NeurIPS/2025/1000+ FPS 4D Gaussian Splatting for Dynamic Scene Rendering/layout.json b/NeurIPS/2025/1000+ FPS 4D Gaussian Splatting for Dynamic Scene Rendering/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..cf86ea9242d20741ce7e3d6ddf5684190b3c3060 --- /dev/null +++ b/NeurIPS/2025/1000+ FPS 4D Gaussian Splatting for Dynamic Scene Rendering/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0b4d0ba8b34b6ef2ce9a14be2c18d02c0c9df75aa3a7023289f3dc36f9dacf07 +size 673299 diff --git a/NeurIPS/2025/3BASiL_ An Algorithmic Framework for Sparse plus Low-Rank Compression of LLMs/2af2a6c4-d824-4f9d-a26b-4132df89fa21_content_list.json b/NeurIPS/2025/3BASiL_ An Algorithmic Framework for Sparse plus Low-Rank Compression of LLMs/2af2a6c4-d824-4f9d-a26b-4132df89fa21_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..f1a1c9d139f08c4502d1ef0522e5bca00c34ba2d --- /dev/null +++ b/NeurIPS/2025/3BASiL_ An Algorithmic Framework for Sparse plus Low-Rank Compression of LLMs/2af2a6c4-d824-4f9d-a26b-4132df89fa21_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b49cb730f91939a8ffbf0798b32833383b076c1ec585b6b82ee590f562d930da +size 233477 diff --git a/NeurIPS/2025/3BASiL_ An Algorithmic Framework for Sparse plus Low-Rank Compression of LLMs/2af2a6c4-d824-4f9d-a26b-4132df89fa21_model.json b/NeurIPS/2025/3BASiL_ An Algorithmic Framework for Sparse plus Low-Rank Compression of LLMs/2af2a6c4-d824-4f9d-a26b-4132df89fa21_model.json new file mode 100644 index 0000000000000000000000000000000000000000..5aedd4145641f8d7d9c5444c70f1fb13869d411e --- /dev/null +++ b/NeurIPS/2025/3BASiL_ An Algorithmic Framework for Sparse plus Low-Rank Compression of LLMs/2af2a6c4-d824-4f9d-a26b-4132df89fa21_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dbbe0d75f4a0c417e2f78e2eeb21d3a697dd8032b3b01547e5ba76d546166efc +size 280002 diff --git a/NeurIPS/2025/3BASiL_ An Algorithmic Framework for Sparse plus Low-Rank Compression of LLMs/2af2a6c4-d824-4f9d-a26b-4132df89fa21_origin.pdf b/NeurIPS/2025/3BASiL_ An Algorithmic Framework for Sparse plus Low-Rank Compression of LLMs/2af2a6c4-d824-4f9d-a26b-4132df89fa21_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..e7c3204d082bb556caf78552022125e854b9f492 --- /dev/null +++ b/NeurIPS/2025/3BASiL_ An Algorithmic Framework for Sparse plus Low-Rank Compression of LLMs/2af2a6c4-d824-4f9d-a26b-4132df89fa21_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:672cecfe38467589f1b88c856b9c3d57d11d2ec341c0a28273db134d756549ae +size 3221024 diff --git a/NeurIPS/2025/3BASiL_ An Algorithmic Framework for Sparse plus Low-Rank Compression of LLMs/full.md b/NeurIPS/2025/3BASiL_ An Algorithmic Framework for Sparse plus Low-Rank Compression of LLMs/full.md new file mode 100644 index 0000000000000000000000000000000000000000..9b8140e33cc64938b0f7345951fab1cacd235ad0 --- /dev/null +++ b/NeurIPS/2025/3BASiL_ An Algorithmic Framework for Sparse plus Low-Rank Compression of LLMs/full.md @@ -0,0 +1,850 @@ +# Mehdi Makni + +Operations Research Center Massachusetts Institute of Technology mmakni@mit.edu + +# Xiang Meng + +Operations Research Center +Massachusetts Institute of Technology +mengx@mit.edu + +# Rahul Mazumder + +Operations Research Center +Massachusetts Institute of Technology +rahulmaz@mit.edu + +# Abstract + +Sparse plus Low-Rank $(\mathbf{S} + \mathbf{LR})$ decomposition of Large Language Models (LLMs) has emerged as a promising direction in model compression, aiming to decompose pre-trained model weights into a sum of sparse and low-rank matrices $\mathbf{W} \approx \mathbf{S} + \mathbf{LR}$ . Despite recent progress, existing methods often suffer from substantial performance degradation compared to dense models. In this work, we introduce 3BASiL-TM, an efficient one-shot post-training method for $(\mathbf{S} + \mathbf{LR})$ decomposition of LLMs that addresses this gap. Our approach first introduces a novel 3-Block Alternating Direction Method of Multipliers (ADMM) method, termed 3BASiL, to minimize the layer-wise reconstruction error with convergence guarantees. We then design an efficient transformer-matching (TM) refinement step that jointly optimizes the sparse and low-rank components across transformer layers. This step minimizes a novel memory-efficient loss that aligns outputs at the transformer level. Notably, the TM procedure is universal as it can enhance any $(\mathbf{S} + \mathbf{LR})$ decomposition, including pure sparsity. Our numerical experiments show that 3BASiL-TM reduces the WikiText2 perplexity gap relative to dense LLaMA-8B model by over $30\%$ under a (2:4 Sparse + 64 LR) configuration, compared to prior methods. Moreover, our method achieves over $2.5\mathrm{x}$ faster compression runtime on an A100 GPU compared to SOTA $(\mathbf{S} + \mathbf{LR})$ method. Our code is available at https://github.com/mazumder-lab/3BASiL. + +# 1 Introduction + +Large Language Models (LLMs) have demonstrated exceptional performance across diverse tasks including complex reasoning [Xu et al., 2025], text generation [Achiam et al., 2023], mathematical problem-solving [Romera-Paredes et al., 2024], and code synthesis [Roziere et al., 2023]. However, state-of-the-art LLMs [Achiam et al., 2023, Dubey et al., 2024, Google, 2023] with billions of parameters face substantial deployment challenges due to their computational and memory requirements. These constraints substantially limit real-time applications and deployment on resource-constrained devices. Consequently, model compression techniques have emerged as an essential research direction to increase LLM accessibility while preserving their accuracy and functionality. + +Established methods for model compression primarily include neural network pruning [LeCun et al., 1989, Hassibi and Stork, 1992, Han et al., 2015b] and quantization [Han et al., 2015a, 2016]. For LLMs, recent research has focused on one-shot post-training compression methods [Frantar and + +Alistarh, 2023, Dettmers et al., 2023, Lin et al., 2024, Frantar et al., 2022, Behdin et al., 2023, Meng et al., 2024a,b] that compress model weights using a minimal calibration dataset without expensive retraining. These approaches have become particularly attractive as they enable efficient compression of modern LLMs even on a single commodity GPU. + +An exciting line of research in one-shot compression studies the task of decomposing pretrained weight matrices $\mathbf{W}$ into a compressed backbone component (e.g., sparse or quantized) and a low-rank component: $\mathbf{W} \approx \mathcal{C}(\mathbf{W}) + \mathbf{L}\mathbf{R}$ . This LoRA-aware formulation effectively integrates with Low-Rank Adaptation (LoRA) methods [Hu et al., 2022], allowing efficient downstream adaptation by freezing $\mathcal{C}(\mathbf{W})$ and fine-tuning only the low-rank components, which serve as a smart initialization to LoRA. Guo et al. [2024], Li et al. [2024] demonstrate that this approach outperforms the sequential approach of first compressing the model $\mathbf{W} \approx \mathcal{C}(\mathbf{W})$ (under similar backbone components constraints) followed by standard LoRA fine-tuning (Noise & Zero adapter). + +One-shot Quantized plus Low-Rank decomposition methods for LLMs [Guo et al., 2024, Li et al., 2024, Saha et al., 2024] have demonstrated exceptional efficiency. These works decompose pretrained LLM matrices into low-rank components and memory-efficient quantized backbones, enabling aggressive quantization while preserving model performance. + +Parallel advances in Sparse plus Low-Rank $(\mathbf{S} + \mathbf{LR})$ decomposition [Zhang and Papyan, 2025, Makni et al., 2025] combine the strengths of pruning and matrix factorization. OATS [Zhang and Papyan, 2025] pioneered post-training $(\mathbf{S} + \mathbf{LR})$ compression for LLMs, demonstrating its viability as an alternative to unstructured pruning at the same compression rates and achieving CPU acceleration via DeepSparse [NeuralMagic, 2021]. HASSLE-Free [Makni et al., 2025] established a unified $(\mathbf{S} + \mathbf{LR})$ framework showing an underlying connection between OATS and various LLM pruning methods—they focus on N:M sparse plus low-rank decompositions for GPU acceleration using specialized CUDA kernels from [Mozaffari et al., 2024]. Despite their promise, existing $(\mathbf{S} + \mathbf{LR})$ algorithms for LLMs rely exclusively on alternating minimization approaches. Due to the complexity of the underlying optimization problem, these procedures have limited convergence guarantees and may perform poorly in joint optimization of the sparse and low-rank components. Indeed, our empirical evidence in Figure 5a and Figure 5b suggests that our proposed algorithm more effectively optimizes the sparse and low-rank parts compared to an alternating minimization approach. + +In this paper, we propose 3BASiL1, an elegant 3-block ADMM approach tailored for $(\mathbf{S} + \mathbf{LR})$ decomposition of LLMs. Unlike prior approaches that separate pruning and low-rank fitting steps, 3BASiL explicitly models their interaction through simultaneous optimization under a unified objective with provable convergence guarantees. We formulate the weight decomposition problem with explicit sparsity and rank constraints, decomposing it into three variable sets—sparse component, low-rank component, and original weights—optimized within an iterative ADMM framework. This approach precisely enforces sparsity pattern and rank constraints at each iteration via closed-form proximal updates while minimizing reconstruction error with respect to the original model weights. + +Additionally, we propose a (memory-efficient) transformer-matching (TM) procedure that refines sparse and low-rank components by aligning transformer block outputs with the dense model. In contrast to prior $(\mathbf{S} + \mathbf{LR})$ methods, which allow only low-rank components refinement via LoRA after layerwise compression, TM enables joint refinement of both sparse and low-rank components at the transformer level. This procedure is compatible with any existing $(\mathbf{S} + \mathbf{LR})$ method and can be applied prior to LoRA fine-tuning with minimal computational overhead, providing a more effective initialization for downstream adaptation. + +Our contributions can be summarized as follows: + +1. 3-Block ADMM We introduce 3BASiL, a novel 3-Block Alternating Direction Method of Multipliers (ADMM) algorithm specifically designed for Sparse plus Low-Rank $(\mathbf{S} + \mathbf{LR})$ decomposition of Language Models. Our method explicitly captures interactions between sparse and low-rank components within a unified optimization framework, while providing theoretical convergence guarantees as well. Moreover, 3BASiL offers remarkable computational advantages, achieving over $7\mathrm{x}$ speedup compared to the strong HASSLE-free-ALPS baseline, when compressing a Llama3.2-3B model on an L40 48GB GPU. + +![](images/f9b95fcd2553378fa7763f17550200eafea68c7c7fea45b7cf890603a98eee06.jpg) +Figure 1: Overview of the proposed 3BASiL framework. (Left) For each layer in a Transformer, we employ multi-Block ADMM to efficiently decompose weights into high-quality Sparse plus Low-Rank components by minimizing the layer reconstruction objective. (Right) At the Transformer level, we apply gradient-based optimization to jointly refine all sparse and low-rank components across layers to match the original transformer's output, with the resulting low-rank components serving as smart initialization for subsequent LoRA fine-tuning. + +2. Transformer matching and Universality We introduce TM, a novel (memory-efficient) refinement procedure that jointly optimizes sparse and low-rank components across transformer layers. This approach significantly improves sparse component quality with minimal computational cost by directly leveraging transformer-level outputs, addressing a major limitation in current $(\mathbf{S} + \mathbf{L}\mathbf{R})$ methods. Crucially, our TM procedure is universally applicable and can enhance any existing $(\mathbf{S} + \mathbf{L}\mathbf{R})$ decomposition method, including purely sparse compression, providing superior initialization for subsequent LoRA fine-tuning. +3. Empirical Validation and State-of-the-Art Results We introduce 3BASiL-TM as a new state-of-the-art method for $(\mathbf{S} + \mathbf{LR})$ one-shot decomposition of Large Language Models. It significantly improves LLM evaluation benchmarks including perplexity of different datasets and various zero-shot tasks. Specifically, our numerical experiments show that 3BASiL-TM reduces the WikiText2 perplexity gap to dense model by over $30\%$ compared to prior methods for a Llama-8B model under a (2:4 Sparse + 64 LR) configuration. It also provides significant compression runtime speedups compared to other $(\mathbf{S} + \mathbf{LR})$ decomposition techniques for LLMs. + +# 2 Highly effective Sparse plus Low-Rank decomposition via ADMM + +# 2.1 Problem formulation + +We compress the layers of an LLM sequentially, one at a time by minimizing the reconstruction error between the outputs of pre-trained weights and compressed ones on a set of given input activations. Formally, let $\widehat{\mathbf{W}}$ represent the pre-trained weight matrix of a given layer, and $\mathbf{X}$ denote the input activations (i.e., output of previous layers) on a set of $N$ calibration samples. In our setting, the goal of layer-wise reconstruction is to find a $(\mathbf{S} + \mathbf{L}\mathbf{R})$ decomposition that minimizes the $\ell_2$ error between the outputs of the original and decomposed weights—this can be formulated as follows: + +$$ +\min _ {\mathbf {S}, \mathbf {L}} \frac {1}{2} \left\| \mathbf {X} \widehat {\mathbf {W}} - \mathbf {X} (\mathbf {S} + \mathbf {L}) \right\| _ {F} ^ {2} + \frac {\lambda}{2} \left\| \widehat {\mathbf {W}} - (\mathbf {S} + \mathbf {L}) \right\| _ {F} ^ {2} \quad \text {s . t .} \quad \mathbf {S} \in \mathcal {S}, \quad \operatorname {r a n k} (\mathbf {L}) \leq r. \tag {1} +$$ + +Above $\| \cdot \| _F$ denotes the Frobenius norm, $S$ denotes the set of matrices satisfying a specified sparsity constraint (e.g., unstructured sparsity with given sparsity level or N:M sparsity); S and L denote the sparse and low-rank components, respectively. Parameter $\lambda >0$ encourages the decomposed weights to remain close to the pre-trained ones. + +# 2.2 A multi-Block ADMM approach for layer-wise reconstruction + +The primary challenge in optimizing problem (1) lies in the joint minimization of $\mathbf{S}$ and $\mathbf{L}$ under two complex constraints—sparsity and low-rank. To address this, we employ the Alternating Direction Method of Multipliers (ADMM), which enables separate updates of $\mathbf{S}$ and $\mathbf{L}$ at each iteration while + +maintaining their interdependence through a Lagrangian multiplier. This approach preserves the power of joint optimization while making the problem tractable. Our 3-block ADMM introduces an auxiliary variable $\mathbf{D}$ as a copy of the sparse component $\mathbf{S}$ , reformulating problem (1) as: + +$$ +\min _ {\mathbf {S}, \mathbf {D}, \mathbf {L}} \frac {1}{2} \left\| \mathbf {X} \widehat {\mathbf {W}} - \mathbf {X} (\mathbf {S} + \mathbf {L}) \right\| _ {F} ^ {2} + \frac {\lambda}{2} \left\| \widehat {\mathbf {W}} - (\mathbf {S} + \mathbf {L}) \right\| _ {F} ^ {2} + \mathbb {I} _ {S} (\mathbf {D}) \tag {2} +$$ + +$$ +\begin{array}{l l} \text {s . t .} & \mathbf {S} = \mathbf {D}, \quad \text {r a n k} (\mathbf {L}) \leq r. \end{array} +$$ + +where $\mathbb{I}_S(\mathbf{D})$ is an indicator function that equals to infinity when $\mathbf{D} \notin S$ and zero otherwise. The augmented Lagrangian function with dual variable $\mathbf{V}$ and a quadratic penalty parameter $\rho > 0$ reads: + +$$ +\mathcal {L} _ {\rho} (\mathbf {S}, \mathbf {L}, \mathbf {D}, \mathbf {V}) = \frac {1}{2} \left\| \mathbf {X} \widehat {\mathbf {W}} - \mathbf {X} (\mathbf {S} + \mathbf {L}) \right\| _ {F} ^ {2} + \frac {\lambda}{2} \left\| \widehat {\mathbf {W}} - (\mathbf {S} + \mathbf {L}) \right\| _ {F} ^ {2} + \mathbb {I} _ {S} (\mathbf {D}) + \frac {\rho}{2} \left\| \mathbf {S} - \mathbf {D} + \frac {\mathbf {V}}{\rho} \right\| _ {F} ^ {2}. +$$ + +The method proceeds by minimizing the augmented Lagrangian with respect to three variables sequentially: the sparse component $\mathbf{S}$ , the low-rank component $\mathbf{L}$ , and sparse component's constrained copy $\mathbf{D}$ , followed by a dual update (in variable $\mathbf{V}$ ). This sequential optimization over three variable blocks gives the method its name: 3-Block ADMM. At iteration $t$ , the updates are: + +$$ +\mathbf {S} ^ {(t + 1)} = \arg \min _ {\mathbf {S}} \mathcal {L} _ {\rho} (\mathbf {S}, \mathbf {L} ^ {(t)}, \mathbf {D} ^ {(t)}, \mathbf {V} ^ {(t)}) \qquad \mathbf {L} ^ {(t + 1)} = \arg \min _ {\mathbf {L}} \mathcal {L} _ {\rho} (\mathbf {S} ^ {(t + 1)}, \mathbf {L}, \mathbf {D} ^ {(t)}, \mathbf {V} ^ {(t)}) +$$ + +$$ +\mathbf {D} ^ {(t + 1)} = \arg \min _ {\mathbf {D}} \mathcal {L} _ {\rho} (\mathbf {S} ^ {(t + 1)}, \mathbf {L} ^ {(t + 1)}, \mathbf {D}, \mathbf {V} ^ {(t)}) \quad \mathbf {V} ^ {(t + 1)} = \mathbf {V} ^ {(t)} + \rho (\mathbf {S} ^ {(t + 1)} - \mathbf {D} ^ {(t + 1)}). +$$ + +Below, we derive the updates. For notational simplicity, we denote $\mathbf{H} = \mathbf{X}^{\top}\mathbf{X} + \lambda \mathbf{I}$ + +S-block update Since $\mathcal{L}_{\rho}(\mathbf{S},\mathbf{L}^{(t)},\mathbf{D}^{(t)},\mathbf{V}^{(t)})$ is a quadratic function of $\mathbf{S}$ , we obtain the closed-form solution by setting the gradient to zero: + +$$ +\mathbf {S} ^ {(t + 1)} = \left(\mathbf {H} + \rho \mathbf {I}\right) ^ {- 1} \left(\mathbf {H} \left(\widehat {\mathbf {W}} - \mathbf {L} ^ {(t)}\right) - \mathbf {V} ^ {(t)} + \rho \mathbf {D} ^ {(t)}\right). \tag {3} +$$ + +L-block update Note that the L-optimization subproblem can be reformulated as minimizing $\| \mathbf{H}^{1 / 2}(\widehat{\mathbf{W}} -\mathbf{S}^{(t + 1)} - \mathbf{L})\| _F^2$ subject to the rank constraint. When $\mathbf{H}$ is full-rank (satisfied for any $\lambda >0$ ), this problem has the closed-form solution (see Section 5 for a discussion about rank-reduced regression results): + +$$ +\mathbf {L} ^ {(t + 1)} = \mathbf {H} ^ {- 1 / 2} P _ {r} \left(\mathbf {H} ^ {1 / 2} \left(\widehat {\mathbf {W}} - \mathbf {S} ^ {(t + 1)}\right)\right), \tag {4} +$$ + +where $P_r$ denotes the best rank- $r$ approximation, which can be computed via SVD.2 + +D-block update The optimization over $\mathbf{D}$ involves projecting $\mathbf{S}^{(t + 1)} + \mathbf{V}^{(t)} / \rho$ onto the sparsity constraint set $S$ , which corresponds to magnitude-based pruning of $(\mathbf{S}^{(t + 1)} + \mathbf{V}^{(t)} / \rho)$ —we sort $[(S^{(t + 1)} + V^{(t)} / \rho)_{ij}]^2$ and retain only the largest values. For unstructured pruning, we keep a predetermined fraction of the largest values; for N:M structured sparsity, we retain N largest values out of every M consecutive weights. + +In practice, we employ an iteration-dependent penalty parameter $\rho_{t}$ , giving the following updates: + +$$ +\mathbf {S} ^ {(t + 1)} = (\mathbf {H} + \rho_ {t} \mathbf {I}) ^ {- 1} \left(\mathbf {H} (\widehat {\mathbf {W}} - \mathbf {L} ^ {(t)}) - \mathbf {V} ^ {(t)} + \rho_ {t} \mathbf {D} ^ {(t)}\right) \mathbf {L} ^ {(t + 1)} = \mathbf {H} ^ {- 1 / 2} P _ {r} (\mathbf {H} ^ {1 / 2} (\widehat {\mathbf {W}} - \mathbf {S} ^ {(t + 1)})) +$$ + +$$ +\mathbf {D} ^ {(t + 1)} = P _ {\mathcal {S}} \left(\mathbf {S} ^ {(t + 1)} + \mathbf {V} ^ {(t)} / \rho_ {t}\right) \quad \mathbf {V} ^ {(t + 1)} = \mathbf {V} ^ {(t)} + \rho_ {t} \left(\mathbf {S} ^ {(t + 1)} - \mathbf {D} ^ {(t + 1)}\right). \tag {5} +$$ + +Computational complexity We implement several tricks to reduce the computational cost in the S and L-update steps, which constitute the major computational cost of 3-Block ADMM algorithm. For the S-update step, we adopt the approach of Meng et al. [2024a] by pre-computing (once) and storing the eigenvalue decomposition $\mathbf{H} = \mathbf{U}\pmb{\Sigma}\mathbf{U}^{\top}$ . This allows us to efficiently calculate the matrix inverse $(\mathbf{H} + \rho \mathbf{I})^{-1} = \mathbf{U}(\pmb{\Sigma} + \rho \mathbf{I})^{-1}\mathbf{U}^{\top}$ for varying values of $\rho$ across iterations. For an efficient L-update step, we store the matrices $\mathbf{H}^{-1/2} = \mathbf{U}\pmb{\Sigma}^{-1/2}$ , and $\mathbf{H}^{1/2} = \pmb{\Sigma}^{1/2}\mathbf{U}^{\top}$ and employ a randomized-SVD procedure [Halko et al., 2011] for numerical efficiency. In the context of LLMs, the weight matrices scale with the transformer's hidden dimension $N$ . Our algorithm's per iteration time complexity comprises: five matrix-matrix multiplications with complexity $O(N^3)$ , a Randomized-SVD operation with complexity $O(N^2r)$ to enforce rank constraint (using constant oversampling and power iterations as in [Halko et al., 2011]), and a projection onto $S$ requiring at most $O(N^2\log(N))$ for sorting and thresholding operations—across the entire matrix for unstructured sparsity or within blocks for semi-structured sparsity. The overall time complexity is $O(N^3)$ . + +# 2.3 Convergence of ADMM + +Despite its appeal and usage, the convergence properties of 3-Block ADMM remain theoretically challenging. Chen et al. [2016] demonstrated that without additional conditions, the algorithm may fail to converge, while later works [Lin et al., 2015, Wang et al., 2018] established various sufficient conditions for convergence. + +We observe that our proposed 3-block ADMM approach can be reformulated as a standard 2-block ADMM by treating $(\mathbf{L},\mathbf{D})$ as a single variable block. This reformulation is valid because the Lagrangian is separable with respect to $\mathbf{L}$ and $\mathbf{D}$ , meaning their joint minimization yields equivalent updates to sequential optimization (although 3-blocks remain the "natural" way to conceptualize the updates). While Meng et al. [2024a] established convergence guarantees for ADMM applied to layerwise pruning, their analysis addresses a different problem formulation than ours. Specifically, they apply ADMM solely to unstructured pruning, whereas our approach extends to $(\mathbf{S} + \mathbf{LR})$ decomposition. Our framework includes a low-rank component with relatively complex updates in each iteration, which introduces additional mathematical challenges in convergence analysis that prevent direct application of the results in Meng et al. [2024a]. + +To address this gap, we establish the following novel convergence guarantee that ensures the decomposition converges as long as we choose penalty parameter $\rho_{t}$ that increases sufficiently rapidly (refer to Appendix A for a complete proof). + +Theorem 1. Let $\{\mathbf{S}^{(t)}\}_{t = 0}^{\infty}$ and $\{\mathbf{L}^{(t)}\}_{t = 0}^{\infty}$ be the sequence generated according to update rule (5). Suppose the penalty parameter $\rho_{t}$ chosen at iteration $t$ is non-decreasing and satisfies $\sum_{t = 0}^{\infty}1 / \rho_t < \infty$ . Then for any $t\geq 1$ : + +$$ +\max \left\{\| \mathbf {S} ^ {(t + 1)} - \mathbf {S} ^ {(t)} \| _ {F}, \| \mathbf {L} ^ {(t + 1)} - \mathbf {L} ^ {(t)} \| _ {F} \right\} \leq C / \rho_ {t - 1}, \tag {6} +$$ + +where $C$ is a constant depending on $\mathbf{X}$ , $\widehat{\mathbf{W}}$ , $\lambda$ , $\rho_0$ , and $\sum_{t=0}^{\infty} 1 / \rho_t$ . In particular, there exists a matrix $\bar{\mathbf{W}}$ such that $\mathbf{S}^{(t)} + \mathbf{L}^{(t)} \rightarrow \bar{\mathbf{W}}$ as $t \to \infty$ . + +# 3 Transformer-level matching + +After layer-wise pruning, LoRA can directly refine the low-rank components in the $(\mathbf{S} + \mathbf{LR})$ decomposition for task adaptation. However, the sparse components are not well-optimized by this process, as they are determined solely via layer-wise objectives. These layer-wise objectives are imperfect proxies for the true end-to-end loss function. On the other hand, fully optimizing the sparse components using the true end-to-end loss is computationally expensive and requires a full backpropagation on the entire network. To address this limitation, we introduce an efficient transformer-matching refinement step that leverages transformer-level information to enhance the sparse components. This procedure is efficient because it requires comparable CUDA memory and runtime to the compression algorithms themselves. + +Our transformer-matching procedure jointly optimizes all sparse and low-rank components across layers within a transformer block to better match the original transformer's output. It acts as an intermediate loss function between layer-wise proxies and the true end-to-end loss. This approach can enhance any $(\mathbf{S} + \mathbf{LR})$ decomposition, including pruning (where $\mathbf{LR} = \mathbf{0}$ ). Figure 2 illustrates the performance gains obtained after applying TM to state-of-the-art one-shot $(\mathbf{S} + \mathbf{LR})$ decomposition algorithms. In Table 3, we show results of applying transformer-matching to pruning algorithms with pure sparsity constraints like WandA [Sun et al., 2024], SparseGPT [Frantar and Alistarh, 2023], and ALPS [Meng et al., 2024a] highlighted in dark red. + +Formally, for each transformer block $T_{i}$ with $L$ layers, after obtaining sparse and low-rank components $\{\mathbf{S}^{(i,\ell)},\mathbf{L}^{(i,\ell)}\}_{\ell = 1}^{L}$ through layer-wise pruning, we denote the support of sparse components as $S^{(i,\ell)} = \mathrm{Supp}(\mathbf{S}^{(i,\ell)})$ . Let $\mathbf{X}_i$ represent the outputs from the previously compressed transformer block $T_{i - 1}$ . We then refine these components using a transformer-level reconstruction loss: + +$$ +\min _ {\left\{\mathbf {S} ^ {(i, \ell)}, \mathbf {L} ^ {(i, \ell)} \right\} _ {\ell = 1} ^ {L}} \left\| T _ {i} \left(\mathbf {X} _ {i}; \left\{\mathbf {W} ^ {(i, \ell)} \right\} _ {\ell = 1} ^ {L}\right) - T _ {i} \left(\mathbf {X} _ {i}; \left\{\mathbf {S} ^ {(i, \ell)} + \mathbf {L} ^ {(i, \ell)} \right\} _ {\ell = 1} ^ {L}\right) \right\| _ {F} ^ {2}, \tag {7} +$$ + +$$ +\begin{array}{l} \text {s . t .} \quad \operatorname {S u p p} (\mathbf {S} ^ {(i, \ell)}) \subset \mathcal {S} ^ {(i, \ell)}, \quad \operatorname {r a n k} (\mathbf {L} ^ {(i, \ell)}) \leq r ^ {(i, \ell)} \end{array} +$$ + +where this constraint optimizes the weights of the decomposed components. Due to the non-linear activations between layers, we use gradient-based optimization methods such as Adam. Nonetheless, this optimization remains computationally efficient as it is performed using iteratively chunks + +of the small calibration dataset used for compression. Additionally, the forward/backward passes are limited to only one transformer block. The transformer-matching approach offers two key advantages. First, it creates a more accurate proxy of the original loss function by directly minimizing the discrepancy between the original and compressed transformer outputs, resulting in higher-performance pruned models. Second, it reduces accumulated errors—introduced in layer-wise pruning where input activations $\mathbf{X}$ are computed from outputs of previously pruned layers—by ensuring that activations fed into subsequent layers more faithfully match those of the dense model: + +$$ +T _ {i} \left(\mathbf {X} _ {i}; \left\{\mathbf {S} ^ {(i, \ell)} + \mathbf {L} ^ {(i, \ell)} \right\} _ {\ell = 1} ^ {L}\right) = \mathbf {X} _ {i + 1} \approx \mathbf {X} _ {i + 1} ^ {\text {(o r a c l e)}} = T _ {i} \left(\mathbf {X} _ {i}; \left\{\mathbf {W} ^ {(i, \ell)} \right\} _ {\ell = 1} ^ {L}\right), \tag {8} +$$ + +therefore providing better activation statistics for compression on subsequent transformers compared to layer-wise reconstruction which only matches weight matrices layer by layer. + +After transformer-matching, the refined sparse components $\mathbf{S}^{(i,\ell)}$ remain fixed during downstream fine-tuning, while the low-rank components $\mathbf{L}^{(i,\ell)}$ serve as smart initializations for efficient LoRA adaptation to specific tasks. + +![](images/9a0972059aa438aefaf3902c5c676f2d2a9b13c67431564f64916bd9ad0b5a1b.jpg) +Transformer-Matching Universality + +![](images/e5ee9819da983290a1a025542a09e445dbcd6abc6d9e1f63d143f23ef8deb814.jpg) +Figure 2: Our transformer matching (TM) procedure improves any one-shot $(\mathbf{S} + \mathbf{L}\mathbf{R})$ decomposition method (see baselines in Section 4) with a small computational overhead. Circled markers represent standard $(\mathbf{S} + \mathbf{L}\mathbf{R})$ methods, while filled markers indicate their TM-enhanced versions. Black arrows illustrate performance gains due to TM. The compression runtimes are reported in hours. Llama3-8B models were run on a A100 GPU, while Llama3.2-3B were run on a L40 GPU. Our proposal 3BASiL-TM, remains significantly faster: (left) over $2\times$ speedup on an A100 80GB for the Llama3-8B model decomposed to (2:4+64LR) configuration, and (right) over $3\times$ speedup on an L40 48GB for the Llama3.2-3B model decomposed to (4:8+64LR) configuration (both compared to Hf-ALPS). + +# 4 Experimental results + +# 4.1 Experimental setup + +Models and LLM Evaluation Protocol To rigorously assess the effectiveness of our proposed approach 3BASiL and transformer-matching (TM) procedure, we conducted extensive experiments on the Llama-3 and Llama-3.2 model families [Dubey et al., 2024] and scaled results in one experiment to a OPT-30B [Zhang et al., 2022] model, hence covering architectures with number of parameters ranging from 1B to 30B. Following the widely adopted setup introduced by Frantar and Alistarh [2023] for one-shot compression, we select the calibration set consisting of 128 randomly sampled text segments (2048 tokens each) from the C4 [Raffel et al., 2020] train dataset's first shard. This calibration set is shared across all evaluated compression methods to ensure consistency. + +We adopt two evaluation criteria: (1) perplexity as a foundational measure of language modeling quality, and (2) zero-shot task performance to assess practical downstream capabilities post-compression. Perplexity is measured using three standard benchmarks: WikiText2 [Merit et al., 2017], Penn Treebank [Marcus et al., 1994], and C4 validation samples, computed using Hugging-Face's full-stride perplexity protocol [Per, 2022]. For zero-shot evaluation, we utilize the LM Harness framework [Gao et al.] on a diverse suite of eight zero-shot tasks: PIQA [Bisk et al., 2020], ARC-Easy/Challenge [Clark et al., 2018], HellaSwag [Zellers et al., 2019], Winogrande [Sakaguchi et al., 2021], RTE [Poliak, 2020], OpenbookQA [Banerjee et al., 2019], and BoolQ [Clark et al., 2019]. We report individual scores for each benchmark and the average across all tasks. + +For perplexity, $(\downarrow)$ lower values are preferred. For zero-shot tasks, $(\uparrow)$ higher values are preferred. + +Baselines Our main baselines are OATS [Zhang and Papyan, 2025], HASSLE-free-SparseGPT (Hf-SparseGPT) and HASSLE-free-ALPS (Hf-ALPS)—the latter two use pruning approaches + +SparseGPT [Frantar and Alistarh, 2023] and ALPS [Meng et al., 2024a], respectively, in the sparsification step of the alternating minimization algorithm proposed by Makni et al. [2025]. + +For all these baselines, we follow the original configuration and perform 80 steps of alternating minimization. For HASSLE-free methods, we propose an improved implementation that replaces their original parameterization of $\mathbf{L} = \mathbf{U}\mathbf{V}^{\top}$ and gradient-based optimization with the closed-form solution provided in Equation (4). This modification leads to improved compression runtime and better downstream LLM evaluation metrics—see Table 6. Under this improved implementation, the method EoRA [Liu et al., 2024], which applies the update in Equation (4) once after one round of compression, reduces to HASSLE-free (alternating minimization approach) with a number of iterations equal to one. EoRA is the fastest $(\mathbf{S} + \mathbf{L}\mathbf{R})$ method but underperforms HASSLE-free which uses more alternating minimization steps (default=80), and hence there is a large gap compared to our approach on most model/configuration settings. We show some results of EoRA in Table 6. + +More details on the implementation of 3BASiL, TM and the baselines (with improved implementation) are provided in Appendix B. + +# 4.2 Numerical results + +Our evaluation focuses primarily on $(\mathrm{N}:\mathbf{M} + \mathbf{L}\mathbf{R})$ decompositions, which enable efficient GPU acceleration via specialized CUDA kernels [Mozaffari et al., 2024, Makni et al., 2025]. We evaluate both one-shot compression performance and downstream LoRA fine-tuning capabilities. Additionally, we demonstrate the generality of our approach through experiments with unstructured sparsity and integration with sparsity allocation methods. The downstream LoRA experiments have been motivated by recent studies [Li et al., 2024, Guo et al., 2024, Saha et al., 2024] suggesting that decompositions of the form $\mathcal{C}(\mathbf{W}) + \mathbf{L}\mathbf{R}$ are LoRA-aware: i.e. low-rank components obtained from compression can act as smart initialization to improve downstream LoRA fine-tuning. Further numerical experiments where we ablate on TM and LoRA fine-tuning for $(\mathbf{S} + \mathbf{L}\mathbf{R})$ methods can be found in Appendix C. + +One-shot (Sparse + LR) results We compare 3BASiL to prior $(\mathbf{S} + \mathbf{LR})$ decomposition methods in the one-shot compression setting—i.e., without fine-tuning. Table 1 reports results for LLaMA3.2 family under various $(\mathrm{N}:\mathrm{M} + 64\mathrm{LR})$ configurations. Table 2 and Figure 3 show results for similar configurations for the LLaMA3-8B model. 3BASiL reduces perplexity by up to $8\%$ compared to previous SOTA (due to better layer-wise reconstruction—see Figure 5a and Figure 5b), with the TM step yielding further dramatic improvements of up to $40\%$ perplexity reduction. + +We also compare $(\mathbf{S} + \mathbf{LR})$ decompositions with semi-structured pure pruning methods under a fixed compression ratio $\rho = 50\%$ . Results in Table 3 show that 3BASiL-TM achieves the best compression-performance trade-off under $(3:8 + \mathrm{LR})$ configurations among different $(\mathbf{S} + \mathbf{LR})$ methods. Additionally, we expand our $(\mathbf{S} + \mathbf{LR})$ experiments to include a $(2:4 + 112)$ configuration for OPT-30B model [Zhang et al., 2022]. This configuration uses a $1.56\%$ Low-Rank Adapter (hidden size 7168). Under this configuration, Mozaffari et al. [2024] report a $1.53\mathrm{x}$ speedup as well as a $0.63\mathrm{x}$ memory reduction compared to dense model. Results are reported in Table 4. + +For unstructured sparsity configurations, we benchmark 3BASiL against prior $(S + LR)$ methods on a "less aggressive" $(50\% +128)$ compression for both Llama3.2-1B and Llama3-8B models. Table 5 shows that our proposed method maintains its advantage even in this near-lossless configuration regime. We further evaluate 3BASiL under high sparsity ratios with (Unstructured $+64$ ) configurations and demonstrate how our method integrates with the sparsity allocation method OWL [Yin et al., 2024] for the Llama3-8B model—see Table 13 in Appendix B. + +These results highlight the effectiveness and flexibility of our method 3BASiL. + +LoRA fine-tuning after one-shot compression After applying $(\mathbf{S} + \mathbf{LR})$ decomposition, the resulting low-rank components can serve as initialization for LoRA fine-tuning on downstream tasks to recover lost performance. We conducted limited LoRA fine-tuning on $10\%$ of the first C4 training dataset shard (approximately 15 million tokens), with detailed hyperparameters in Appendix B. Figure 4a demonstrates that LFT-3BASiL-TM significantly reduces the C4 perplexity of $(\mathbf{S} + \mathbf{LR})$ decompositions, particularly under aggressive compression regimes like $2:8 + 64\mathrm{LR}$ . Moreover, while LoRA fine-tuning can recover a large portion of the performance lost due to compression, an advanced one-shot decomposition approach retains its advantage post fine-tuning. For instance, LFT-3BASiL-TM still outperforms competing decomposition methods after LoRA fine-tuning of $2:8 + 64\mathrm{LR}$ configurations, achieving approximately $8\%$ lower perplexity. + +![](images/1c53d638984dbde23529f2d6782dad036646e2022c4356c4e33040085ac06ea8.jpg) +Figure 3: One-shot C4 perplexity analysis of Llama3-8B under different (N:M + 64LR) configurations. + +Table 1: Perplexity of Llama-3.2 family + +
MethodConfigLlama-3.2-1BLlama-3.2-3B
C4 ↓WT2 ↓PTB ↓C4 ↓WT2 ↓PTB ↓
OATS2:8+64LR640.86605.20779.86531.47494.31674.71
Hf-SparseGPT162.45134.21170.12106.07106.17151.92
Hf-ALPS107.1494.71124.1769.9665.34108.68
3BAsiL97.5086.59100.3573.0072.26110.10
3BAsiL-TM55.2449.7469.4945.3542.3868.29
OATS3:8+64LR125.9192.13115.8065.0847.2781.29
Hf-SparseGPT43.5034.1851.1634.6626.6039.76
Hf-ALPS37.8029.0043.6027.9422.7734.59
3BAsiL34.8126.9641.5526.3520.6631.77
3BAsiL-TM26.2620.7532.0920.8917.1825.31
OATS4:8+64LR28.0619.6932.9019.2513.4021.67
Hf-SparseGPT22.2415.9027.3517.0912.3019.19
Hf-ALPS20.7114.9024.7516.0411.5118.17
3BAsiL20.0414.2624.2715.6510.9717.39
3BAsiL-TM18.6613.1922.4614.8910.2916.52
OATS2:4+64LR41.8028.4545.3625.1817.4128.60
Hf-SparseGPT27.2519.4532.6320.3815.0323.23
Hf-ALPS23.9017.6628.9618.4513.7920.50
3BAsiL23.1617.2727.7717.8913.1220.10
3BAsiL-TM20.4615.2324.6016.3711.7918.34
Dense-14.019.7517.5911.337.8113.53
+ +Table 2: One-shot (N:M Sparse + LR) decomposition performance for Meta-Llama-3-8B. + +
MethodConfigC4 ↓WT2 ↓PTB ↓PIQA ↑HS ↑ARC-E ↑ARC-C ↑WG ↑RTE ↑OQA ↑BoolQ ↑Avg ↑
OATS3:8+64LR58.8840.7667.3563.7139.4842.6824.3253.9152.7128.4063.9846.15
Hassle-free-SparseGPT29.3221.4632.0668.6651.9950.9730.3863.8553.0732.0071.3152.78
Hassle-free-ALPS23.9318.2026.3170.6256.5454.4230.1264.7255.2332.8071.9654.55
3BASiL23.0718.0324.8471.0656.9657.7032.5966.6954.5133.0066.7054.90
3BASiL-TM18.1114.2620.4774.0561.8560.7334.7365.9854.5134.8076.9157.94
OATS4:8+64LR16.3810.8817.2375.8467.6067.0941.2170.8860.2938.2073.6161.84
Hassle-free-SparseGPT14.659.8815.2177.0969.9569.3241.8171.2756.3240.6079.3963.22
Hassle-free-ALPS14.049.4414.4576.8271.1971.0444.4572.7756.6840.2078.1363.91
3BASiL13.749.2114.2476.8872.0570.1644.8072.1461.0141.4080.8964.92
3BASiL-TM13.028.6413.7078.2472.5973.1147.3571.9863.1842.4080.4966.17
OATS2:4+64LR21.5914.7623.4172.7460.7060.8634.8165.5157.7635.2068.3256.99
Hassle-free-SparseGPT17.7712.3818.7174.8165.0466.1638.5770.0954.8738.4077.7160.71
Hassle-free-ALPS16.1511.3816.7175.1967.1064.4438.9169.5359.9339.4078.3861.61
3BASiL15.7611.2316.2576.5067.6167.2140.1070.2464.2638.2078.2962.80
3BASiL-TM14.349.7814.8877.4869.5867.2140.5371.2761.3739.8079.5163.34
Meta-Llama-3-8B Dense-9.446.1411.1880.7979.1777.6953.3372.8569.6845.0081.4469.99
+ +
MethodConfigC4↓WT2↓PTB↓PIQA ↑ARC-E ↑ARC-C ↑
Wanda2:438.2126.8947.1367.6349.3729.01
Wanda-TM15.9111.0317.6075.1463.9740.19
SparseGPT22.6516.2225.1571.1656.4832.59
SparseGPT-TM15.3010.8316.7776.2865.2840.36
ALPS19.6214.5021.7373.7860.0635.84
ALPS-TM14.9610.6516.3576.8865.0339.85
Wanda4:822.7015.5826.6272.0358.6336.09
Wanda-TM13.999.2815.0977.4868.2242.75
SparseGPT17.5912.2918.4875.6863.3839.59
SparseGPT-TM13.689.2814.5178.0770.2943.94
ALPS16.0611.1716.6076.1266.2540.87
ALPS-TM13.599.1514.1877.5869.5743.94
OATS2:8+LR21.0314.5424.1573.6759.6837.12
Hf-SparseGPT20.0515.0322.0174.0560.5236.18
Hf-ALPS17.8913.0719.1174.5465.5339.08
3BASI-L15.2010.6415.8076.7170.0843.52
3BASI-L-TM13.819.5014.7477.1573.3644.54
OATS3:8+LR16.8711.4318.5375.2465.9139.85
Hf-SparseGPT16.1611.3616.7175.7967.5541.04
Hf-ALPS14.8510.2015.4277.1569.4043.64
3BASI-L13.739.2914.6278.4571.4243.43
3BASI-L-TM13.018.6913.7477.8075.0047.44
Llama3-8B Dense-9.446.1411.1880.7977.6953.33
+ +Table 3: One-shot (N:M Sparse + LR) decomposition performance of Llama3-8B model. The compression ratio (percentage of nonzero parameters retained) is fixed to be $\rho = 0.5$ . For Perplexity, $(\downarrow)$ lower values are preferred. For zero-shot tasks, $(\uparrow)$ higher values are preferred. Bolded values correspond to the overall best compression scheme that satisfies $\rho = 0.5$ . Underlined values correspond to the best pure pruning algorithm for the same compression. This shows the universality of transformer-matching to pure sparsity constraints. + +
MethodC4 ↓WT2 ↓PTB ↓Time (hrs) ↓
OATS-1011.7510.4814.655.81
Hf-SparseGPT11.5810.1714.395.97
Hf-ALPS-1011.5610.0514.334.33
3BASiL11.5310.0414.264.20
Dense11.449.5614.04-
+ +Table 4: One-shot (2:4 + 112) decomposition of OPT-30B model. This configuration results in efficient inference. We limit the compression runtime to 6 A100 GPU hours. 3BASiL-TM largely exceeds this period. We limit the alternating minimization steps of Hf-ALPS and 0ATS to 10 to fit within the time constraint. + +Table 5: One-shot (50% + 128) decomposition for Llama3.2-1B and Meta-Llama-3-8B models. + +
MethodConfigC4↓WT2↓PTB↓PIQA↑HS↑ARC-E↑ARC-C↑WG↑RTE↑OQA↑BoolQ↑Avg↑
OATS50%+12817.9912.1621.4071.7157.9657.2833.7959.9852.7133.0063.9453.80
Hf-SparseGPT17.2511.9920.8772.9159.0456.8233.1158.8857.7635.0057.0353.82
Hf-ALPS16.8111.6620.1272.8059.9257.6233.1158.6455.9635.2059.6654.11
3BASIL16.1711.1620.0073.8360.4258.0434.4760.3853.7936.8058.2054.49
3BASIL-TM15.7810.8719.3373.2360.6659.2634.5661.0159.2136.6064.1356.08
Llama-3.2-1B Dense-14.019.7517.5974.5963.6660.4836.2660.6956.6837.2063.9856.69
OATS50%+12812.257.7812.9278.4075.3273.9949.1573.8058.8441.8079.4266.34
Hf-SparseGPT11.987.7712.8579.1175.8875.0049.4073.3263.1843.8078.3267.25
Hf-ALPS12.097.9912.8678.7876.2976.5251.1973.0960.6540.2081.6267.29
3BASIL11.517.4712.3679.5476.6974.7548.7272.6967.8743.0080.2467.94
3BASIL-TM11.277.3012.2679.6576.0775.8447.7871.9870.4044.2080.7068.33
Meta-Llama-3-8B Dense-9.446.1411.1880.7979.1777.6953.3372.8569.6845.0081.4469.99
+ +Figure 4: C4 perplexity performance of Llama3-8B & Llama3.2-1B before/after LoRA fine-tuning. +![](images/be923a572ace536ae43e490d87b2720104e4ab994bb51e56ce93f509286cb784.jpg) +(a) C4 ppl of Llama3-8B model under different $(\mathbf{S} + \mathbf{LR})$ configurations after LoRA. + +![](images/78bd7d769d0c8c208e8a762803a6d3c280489587d7c045c615cd67c061d990e6.jpg) +(b) C4 perplexity gap to dense model (Llama3.2-1B) under $(50\% + 128\mathrm{LR})$ configuration. + +# 5 Related Work + +One-shot Sparse/Quantized plus Low-Rank compression The seminal works of Yu et al. [2017] proposed a compression technique for a neural network using sparsity plus low-rank constraints. However, the authors study small-scale vision models. In addition, they consider a compression that needs to be repeated over multiple rounds (decomposing selected layers and followed by a retraining process). Our focus is different; we are interested in compressing at LLM-scale in one-shot (no expensive retraining). Recent methods in LLM compression have focused on effectively combining low-rank decomposition with quantization or sparsity. EoRA [Liu et al., 2024] has been proposed as a method to compensate for the loss produced by a general-purpose compressed weight $\mathcal{C}(\mathbf{W})$ using a low-rank component, it does the low-rank fitting step once post the initial weight compression, which could include combinations of sparsity and quantization. LoftQ [Li et al., 2024] jointly optimizes quantization and LoRA initialization by solving $\min_{\mathbf{Q},\mathbf{L}}\| \mathbf{W} - (\mathbf{Q} + \mathbf{L})\|_F$ , where $\mathbf{W}$ represents the original weights, $\mathbf{Q}$ the quantized component, and $\mathbf{L}$ the low-rank component. LQ-LoRA [Guo et al., 2024] extends this by incorporating Fisher information weighting, approximately solving $\min_{\mathbf{Q},\mathbf{L}}\| \mathbf{F} \odot (\mathbf{W} - (\mathbf{Q} + \mathbf{L}))\|_F$ . CALDERA [Saha et al., 2024] further considers the layer-wise reconstruction error, optimizing $\min_{\mathbf{Q},\mathbf{L}}\| \mathbf{X}\mathbf{W} - \mathbf{X}(\mathbf{Q} + \mathbf{L})\|_F$ to maintain the outputs of individual layers rather than mere weight approximation. From the $(\mathbf{S} + \mathbf{LR})$ perspective, OATS [Zhang and Papyan, 2025] proposes an outlier-aware alternating minimization, effectively reducing to solving $\min_{\mathbf{S},\mathbf{L}}\| \mathbf{D}\mathbf{W} - \mathbf{D}(\mathbf{S} + \mathbf{L})\|_F$ with $\mathbf{D} = \mathrm{diag}(\mathbf{X}^T\mathbf{X})$ , as noted by Makni et al. [2025]. HASSLE-free [Makni et al., 2025] directly tackles layer-wise reconstruction error $\min_{\mathbf{S},\mathbf{L}}\| \mathbf{X}\mathbf{W} - \mathbf{X}(\mathbf{S} + \mathbf{L})\|_F$ using alternating minimization. While methods such as OATS and HASSLE-free separately optimize sparse and low-rank components, our proposed approach, 3BASiL, distinctly utilizes a unified optimization framework via a 3-block ADMM formulation, jointly optimizing sparse and low-rank components simultaneously. + +Sparse plus Low-Rank structures in transformers Beyond model compression, sparse plus low-rank structures have a strong presence in the context of LLMs. LoRAPrune [Zhang et al., 2024] is a purse sparsification method, which prunes a model (iteratively) by designing a memory-efficient LoRA-guided (low-rank structure) pruning criterion. In contrast, LoSA (low-rank Sparse Adapta + +tion) [Huang et al., 2025] jointly applies LoRA fine-tuning and pruning in a unified framework to obtain a fine-tuned sparse-only (as opposed to $(\mathbf{S} + \mathbf{LR}))$ model, by dynamically sparsifying the LoRA weights and adjusting their rank. SLTrain [Han et al., 2024] addresses $(\mathbf{S} + \mathbf{LR})$ from a training perspective. It pre-trains an LLM using a fixed random sparse mask plus trainable low-rank factors (similar to LoRA), achieving comparable accuracy to dense training with far fewer parameters. SLTrain demonstrates the benefits of $(\mathbf{S} + \mathbf{LR})$ structure for pre-training but it doesn't solve the post-hoc decomposition problem of a dense model. There are connections between our transformer-matching step and SLTrain as they both train sparse (fixed support) and low-rank components, but they minimize different loss functions and serve different purposes. + +ADMM approaches to compress networks The Alternating Direction Method of Multipliers (ADMM) [Boyd et al., 2011, Davis and Yin, 2016] is an effective optimization technique for problems with coupled variables that has been successfully applied to neural network compression. Ye et al. [2018] introduced ADMM-based progressive weight pruning that optimizes the original loss function under sparsity constraints, which Ye et al. [2019] extended to preserve adversarial robustness during compression. In contrast, recent methods have scaled ADMM to LLMs through layerwise reconstruction: Boža [2024] employed ADMM to solve a convex problem recovering optimal weights on a fixed support of the weight matrix, while Meng et al. [2024a] utilized ADMM for a non-convex problem that jointly optimizes both support and weights. Our proposed method differs from these prior works as we explore a 3-block ADMM in model compression that simultaneously optimizes $(\mathbf{S} + \mathbf{LR})$ components with theoretical convergence guarantees. + +Exact Low-Rank updates for layer-wise compression The problem of exact low-rank updates found in Equation (4) has original roots from classical reduced-rank regression methods [Izenman, 1975, Reinsel and Velu, 1998], which provide closed-form solutions for optimally approximating linear regression models under rank constraints. Recent work, including CALDERA [Saha et al., 2024] and the low-rank correction method by Scetbon and Hensman [2024], applies these closed-form updates to compress large language models into $\mathbf{W} \approx \mathbf{Q} + \mathbf{L}\mathbf{R}$ . We also use these exact low-rank updates by integrating them directly in Equation (4) within our ADMM framework for $(\mathbf{S} + \mathbf{L}\mathbf{R})$ decomposition. + +# 6 Conclusion and limitations + +We present 3BASiL as a highly-efficient $(\mathbf{S} + \mathbf{LR})$ decomposition algorithm with theoretical convergence guarantees. It provides high-quality solutions to the layer-wise decomposition problem presented in Equation (1) in terms of objective minimization (Figure 5a and Figure 5b) compared to competing $(\mathbf{S} + \mathbf{LR})$ decomposition methods. We further refine these decomposed weights with our novel (memory-efficient) transformer matching step TM that can enhance any $(\mathbf{S} + \mathbf{LR})$ decomposition. This shows that one route for optimal compression results (in the context of $\mathcal{C}(\mathbf{W}) + \mathbf{LR}$ ) is to unfold the LLM compression into 3 minimization steps: (i) [layer-wise reconstruction] this is the loss that has been considered in many SOTA pruning/quantization algorithms [Frantar and Alistarh, 2023, Meng et al., 2024a, Saha et al., 2024, Frantar et al., 2022, Meng et al., 2024b], (ii) [transformer-matching] this is an intermediate loss function (to be optimized in a memory-efficient manner) which is a more reliable approximation to the true loss function than simple layer-wise reconstruction, and (iii) [LoRA fine-tuning] plugs the obtained low-rank components as smart initialization for LoRA to minimize the true LLM loss function. We believe that our 3-block ADMM approach and TM can generalize to quantization or quantized-sparse constraints. We leave these explorations for future works. While we have shown how to integrate sparsity allocation mechanisms like OWL to our framework, it remains to explore dedicated methods that can algorithmically allocate different sparsity/rank configurations to different layers to further improve efficiency-utility-computations tradeoffs. + +# Acknowledgements + +This research is supported in part by grants from the Office of Naval Research (N000142512504, N000142212665). We acknowledge the MIT Engaging cluster for providing HPC resources that have contributed to the research results reported within this paper. Additionally, we thank Google for providing us with Google Cloud Credits. We thank Shibal Ibrahim, Ryan Lucas, and Gabriel Afriat for their helpful discussions. + +# References + +Perplexity of fixed-length models, 2022. URL https://huggingface.co/docs/transformers/perplexity. +Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. +Pratyay Banerjee, Kuntal Kumar Pal, Arindam Mitra, and Chitta Baral. Careful selection of knowledge to solve open book question answering. arXiv preprint arXiv:1907.10738, 2019. +Kayhan Behdin, Ayan Acharya, Sathiya Keerthi Aman Gupta, and Rahul Mazumder. Quantease: Optimization-based quantization for language models—an efficient and intuitive algorithm. stat, 1050:5, 2023. +Yonatan Bisk, Rowan Zellers, Jianfeng Gao, Yejin Choi, et al. Piqa: Reasoning about physical commonsense in natural language. In Proceedings of the AAAI conference on artificial intelligence, volume 34, pages 7432-7439, 2020. +Stephen Boyd, Neal Parikh, Eric Chu, Borja Peleato, Jonathan Eckstein, et al. Distributed optimization and statistical learning via the alternating direction method of multipliers. Foundations and Trends® in Machine learning, 3(1):1-122, 2011. +Vladimír Boža. Fast and optimal weight update for pruned large language models. arXiv preprint arXiv:2401.02938, 2024. +Caihua Chen, Bingsheng He, Yinyu Ye, and Xiaoming Yuan. The direct extension of admn for multi-block convex minimization problems is not necessarily convergent. Mathematical Programming, 155(1):57-79, 2016. +Christopher Clark, Kenton Lee, Ming-Wei Chang, Tom Kwiatkowski, Michael Collins, and Kristina Toutanova. *BoolQ: Exploring the surprising difficulty of natural yes/no questions.* In Jill Burstein, Christy Doran, and Thamar Solorio, editors, *Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies*, Volume 1 (Long and Short Papers), pages 2924–2936, Minneapolis, Minnesota, June 2019. Association for Computational Linguistics. doi: 10.18653/v1/N19-1300. URL https://aclanthology.org/N19-1300/. +Peter Clark, Isaac Cowhey, Oren Etzioni, Tushar Khot, Ashish Sabharwal, Carissa Schoenick, and Oyvind Tafjord. Think you have solved question answering? try arc, the ai2 reasoning challenge. arXiv preprint arXiv:1803.05457, 2018. +Damek Davis and Wotao Yin. Convergence rate analysis of several splitting schemes. Splitting methods in communication, imaging, science, and engineering, pages 115-163, 2016. +Tim Dettmers, Ruslan Svirschevski, Vage Egiazarian, Denis Kuznedev, Elias Frantar, Saleh Ashkoos, Alexander Borzunov, Torsten Hoefler, and Dan Alistarh. Spqr: A sparse-quantized representation for near-lossless llm weight compression. arXiv preprint arXiv:2306.03078, 2023. +Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. +Elias Frantar and Dan Alistarh. Sparsegpt: Massive language models can be accurately pruned in one-shot. In International Conference on Machine Learning, pages 10323-10337. PMLR, 2023. +Elias Frantar, Saleh Ashkboos, Torsten Hoefler, and Dan Alistarh. Gptq: Accurate post-training quantization for generative pre-trained transformers. arXiv preprint arXiv:2210.17323, 2022. +L Gao, J Tow, B Abbasi, S Biderman, S Black, A DiPofi, C Foster, L Golding, J Hsu, A Le Noac'h, et al. A framework for few-shot language model evaluation, 12 2023. URL https://zenodo.org/records/10256836, 7. + +Gemini Team Google. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. +Han Guo, Philip Greengard, Eric Xing, and Yoon Kim. LQ-loRA: Low-rank plus quantized matrix decomposition for efficient language model finetuning. In The Twelfth International Conference on Learning Representations, 2024. URL https://openreview.net/forum?id= xw29Vv0MmU. +Nathan Halko, Per-Gunnar Martinsson, and Joel A Tropp. Finding structure with randomness: Probabilistic algorithms for constructing approximate matrix decompositions. SIAM review, 53 (2):217-288, 2011. +Andi Han, Jiaxiang Li, Wei Huang, Mingyi Hong, Akiko Takeda, Pratik Jawanpuria, and Bamdev Mishra. SLTrain: a sparse plus low rank approach for parameter and memory efficient pretraining. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. URL https://openreview.net/forum?id=MXze4H7opg. +Song Han, Huizi Mao, and William J Dally. Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding. arXiv preprint arXiv:1510.00149, 2015a. +Song Han, Jeff Pool, John Tran, and William Dally. Learning both weights and connections for efficient neural network. Advances in neural information processing systems, 28, 2015b. +Song Han, Xingyu Liu, Huizi Mao, Jing Pu, Ardavan Pedram, Mark A Horowitz, and William J Dally. Eie: Efficient inference engine on compressed deep neural network. ACM SIGARCH Computer Architecture News, 44(3):243-254, 2016. +Babak Hassibi and David Stork. Second order derivatives for network pruning: Optimal brain surgeon. Advances in neural information processing systems, 5, 1992. +Edward J Hu, yelong shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuzhhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2022. URL https://openreview.net/forum?id=nZeVKeeFYf9. +Weizhong Huang, Yuxin Zhang, Xiawu Zheng, Liuyang, Jing Lin, Yiwu Yao, and Rongrong Ji. Dynamic low-rank sparse adaptation for large language models. In The Thirteenth International Conference on Learning Representations, 2025. URL https://openreview.net/forum?id= oXh0939Zzq. +Alan Julian Izenman. Reduced-rank regression for the multivariate linear model. Journal of multivariate analysis, 5(2):248-264, 1975. +Yann LeCun, John Denker, and Sara Solla. Optimal brain damage. Advances in neural information processing systems, 2, 1989. +Yixiao Li, Yifan Yu, Chen Liang, Nikos Karampatziakis, Pengcheng He, Weizhu Chen, and Tuo Zhao. Loftq: LoRA-fine-tuning-aware quantization for large language models. In The Twelfth International Conference on Learning Representations, 2024. URL https://openreview.net/forum?id=LzPWwPAdY4. +Ji Lin, Jiaming Tang, Haotian Tang, Shang Yang, Wei-Ming Chen, Wei-Chen Wang, Guangxuan Xiao, Xingyu Dang, Chuang Gan, and Song Han. Awq: Activation-aware weight quantization for on-device llm compression and acceleration. Proceedings of Machine Learning and Systems, 6: 87-100, 2024. +Tian-Yi Lin, Shi-Qian Ma, and Shu-Zhong Zhang. On the sublinear convergence rate of multi-block admm. Journal of the Operations Research Society of China, 3:251-274, 2015. +Shih-Yang Liu, Huck Yang, Chein-Yi Wang, Nai Chit Fung, Hongxu Yin, Charbel Sakr, Saurav Muralidharan, Kwang-Ting Cheng, Jan Kautz, Yu-Chiang Frank Wang, Pavlo Molchanov, and Min-Hung Chen. Eora: Training-free compensation for compressed llm with eigenspace low-rank approximation. CoRR, abs/2410.21271, 2024. URL https://doi.org/10.48550/arXiv.2410.21271. + +Mehdi Makni, Kayhan Behdin, Zheng Xu, Natalia Ponomareva, and Rahul Mazumder. A unified framework for sparse plus low-rank matrix decomposition for LLMs. In The Second Conference on Parsimony and Learning (Proceedings Track), 2025. URL https://openreview.net/forum?id=hyN75SAJTI. +Mitchell Marcus, Grace Kim, Mary Ann Marcinkiewicz, Robert MacIntyre, Ann Bies, Mark Ferguson, Karen Katz, and Britta Schasberger. The penn treebank: Annotating predicate argument structure. In Proceedings of the Workshop on Human Language Technology, HLT '94, page 114-119, USA, 1994. Association for Computational Linguistics. ISBN 1558603573. doi: 10.3115/1075812.1075835. URL https://doi.org/10.3115/1075812.1075835. +Xiang Meng, Kayhan Behdin, Haoyue Wang, and Rahul Mazumder. Alps: Improved optimization for highly sparse one-shot pruning for large language models. In Annual Conference on Neural Information Processing Systems, NeurIPS, 2024a. +Xiang Meng, Shibal Ibrahim, Kayhan Behdin, Hussein Hazimeh, Natalia Ponomareva, and Rahul Mazumder. Osscar: One-shot structured pruning in vision and language models with combinatorial optimization. In *Forty-first International Conference on Machine Learning*, 2024b. +Stephen Merity, Caiming Xiong, James Bradbury, and Richard Socher. Pointer sentinel mixture models. In International Conference on Learning Representations, 2017. URL https://openreview.net/forum?id=Byj72udxe. +Mohammad Mozaffari, Amir Yazdanbakhsh, Zhao Zhang, and Maryam Mehri Dehnavi. Slope: Double-pruned sparse plus lazy low-rank adapter pretraining of llms. arXiv preprint arXiv:2405.16325, 2024. +NeuralMagic. Deepsparse: A cpu runtime for sparse inference of neural networks., 2021. URL https://github.com/neuralmagic/deepsparse. +Adam Paszke, Sam Gross, Soumith Chintala, Gregory Chanan, Edward Yang, Zachary DeVito, Zeming Lin, Alban Desmaison, Luca Antiga, and Adam Lerer. Automatic differentiation in pytorch. 2017. +Adam Poliak. A survey on recognizing textual entailment as an nlp evaluation. arXiv preprint arXiv:2010.03061, 2020. +Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of Machine Learning Research, 21(140):1-67, 2020. URL http://jmlr.org/papers/v21/20-074.html. +Gregory C Reinsel and Raja P Velu. Multivariate reduced-rank regression. Springer, 1998. +Bernardino Romera-Paredes, Mohammadamin Barekatain, Alexander Novikov, Matej Balog, M. Pawan Kumar, Emilien Dupont, Francisco J. R. Ruiz, Jordan S. Ellenberg, Pengming Wang, Omar Fawzi, Pushmeet Kohli, and Alhussein Fawzi. Mathematical discoveries from program search with large language models. Nat., 625(7995):468-475, January 2024. URL https://doi.org/10.1038/s41586-023-06924-6. +Baptiste Roziere, Jonas Gehring, Fabian Gloeckle, Sten Sootla, Itai Gat, Xiaqing Ellen Tan, Yossi Adi, Jingyu Liu, Romain Sauvestre, Tal Remez, et al. Code llama: Open foundation models for code. arXiv preprint arXiv:2308.12950, 2023. +Rajarshi Saha, Naomi Sagan, Varun Srivastava, Andrea Goldsmith, and Mert Pilanci. Compressing large language models using low rank and low precision decomposition. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. URL https://openreview.net/forum?id=1kx30pcqSZ. +Keisuke Sakaguchi, Ronan Le Bras, Chandra Bhagavatula, and Yejin Choi. Winogrande: An adversarial winograd schema challenge at scale. Communications of the ACM, 64(9):99-106, 2021. +Meyer Sctbon and James Hensman. Low-rank correction for quantized llms. arXiv preprint arXiv:2412.07902, 2024. + +Mingjie Sun, Zhuang Liu, Anna Bair, and J Zico Kolter. A simple and effective pruning approach for large language models. In The Twelfth International Conference on Learning Representations, 2024. URL https://openreview.net/forum?id=PxoFut3dWW. +Fenghui Wang, Wenfei Cao, and Zongben Xu. Convergence of multi-block bregman admm for nonconvex composite problems. Science China Information Sciences, 61:1-12, 2018. +Fengli Xu, Qianyue Hao, Zefang Zong, Jingwei Wang, Yunke Zhang, Jingyi Wang, Xiaochong Lan, Jiahui Gong, Tianjian Ouyang, Fanjin Meng, et al. Towards large reasoning models: A survey of reinforced reasoning with large language models. arXiv preprint arXiv:2501.09686, 2025. +Shaokai Ye, Tianyun Zhang, Kaiqi Zhang, Jiayu Li, Kaidi Xu, Yunfei Yang, Fuxun Yu, Jian Tang, Makan Fardad, Sijia Liu, et al. Progressive weight pruning of deep neural networks using admm. arXiv preprint arXiv:1810.07378, 2018. +Shaokai Ye, Kaidi Xu, Sijia Liu, Hao Cheng, Jan-Henrik Lambrechts, Huan Zhang, Aojun Zhou, Kaisheng Ma, Yanzhi Wang, and Xue Lin. Adversarial robustness vs. model compression, or both? In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 111-120, 2019. +Lu Yin, You Wu, Zhenyu Zhang, Cheng-Yu Hsieh, Yaqing Wang, Yiling Jia, Mykola Pechenizkiy, Yi Liang, Zhangyang Wang, and Shiwei Liu. Outlier weighed layerwise sparsity (OWL): A missing secret sauce for pruning LLMs to high sparsity, 2024. URL https://openreview.net/forum?id=p0Bvr1PxFd. +Xiyu Yu, Tongliang Liu, Xinchao Wang, and Dacheng Tao. On compressing deep models by low rank and sparse decomposition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7370-7379, 2017. +Rowan Zellers, Ari Holtzman, Yonatan Bisk, Ali Farhadi, and Yejin Choi. HellaSwag: Can a machine really finish your sentence? In Anna Korhonen, David Traum, and Lluis Márquez, editors, Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 4791-4800, Florence, Italy, July 2019. Association for Computational Linguistics. doi: 10.18653/v1/P19-1472. URL https://aclanthology.org/P19-1472/. +Mingyang Zhang, Hao Chen, Chunhua Shen, Zhen Yang, Linlin Ou, Xinyi Yu, and Bohan Zhuang. LoRAPrune: Pruning meets low-rank parameter-efficient fine-tuning, 2024. URL https://openreview.net/forum?id=9KVT1e1qf7. +Stephen Zhang and Vardan Papyan. OATS: Outlier-aware pruning through sparse and low rank decomposition. In The Thirteenth International Conference on Learning Representations, 2025. URL https://openreview.net/forum?id=DLDuVbxORA. +Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, et al. Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068, 2022. + +# Appendix + +# A Proofs of Theorem 1 + +Proof. For conciseness, throughout the proof, we denote $\mathbf{H} = \mathbf{X}^{\top}\mathbf{X} + \lambda \mathbf{I}$ and $\mathbf{G} = (\mathbf{X}^{\top}\mathbf{X} + \lambda \mathbf{I})\widehat{\mathbf{W}}$ . We denote $C_F$ as a large constant such that + +$$ +\max \{1, \| \mathbf {H} ^ {- 1 / 2} \| _ {2}, \| \mathbf {H} \| _ {2}, \| \mathbf {G} \| _ {F} \} \leq C _ {F}. \tag {9} +$$ + +To establish the theorem, we first present the following three lemmas. + +Lemma A.1. Let $\{\mathbf{D}^{(t)}\}_{t = 0}^{\infty}$ , $\{\mathbf{L}^{(t)}\}_{t = 0}^{\infty}$ and $\{\mathbf{V}^{(t)}\}_{t = 0}^{\infty}$ be the sequence generated according to update rule (5). Then for any $t \geq 1$ , it holds + +$$ +\left\| \mathbf {L} ^ {(t)} \right\| _ {F} \leq C _ {F} ^ {3} \left(1 + \left\| \mathbf {D} ^ {(t)} \right\| _ {F} + \frac {\left\| \mathbf {V} ^ {(t)} \right\| _ {F}}{\rho_ {t - 1}} + \frac {\left\| \mathbf {V} ^ {(t - 1)} \right\| _ {F}}{\rho_ {t - 1}}\right). \tag {10} +$$ + +Lemma A.2. Let $\{\mathbf{D}^{(t)}\}_{t = 0}^{\infty}$ and $\{\mathbf{V}^{(t)}\}_{t = 0}^{\infty}$ be the sequence generated according to update rule (5). Then for any $t\geq 1$ it holds + +$$ +\left\| \mathbf {V} ^ {(t + 1)} \right\| _ {F} \leq \left(C _ {F} + C _ {F} ^ {4}\right) \left(1 + \left\| \mathbf {D} ^ {(t)} \right\| _ {F} + \frac {\left\| \mathbf {V} ^ {(t)} \right\| _ {F}}{\rho_ {t - 1}} + \frac {\left\| \mathbf {V} ^ {(t - 1)} \right\| _ {F}}{\rho_ {t - 1}}\right). \tag {11} +$$ + +and + +$$ +\left\| \mathbf {D} ^ {(t + 1)} - \mathbf {D} ^ {(t)} \right\| _ {F} \leq \frac {2 C _ {F} + 2 C _ {F} ^ {4}}{\rho_ {t}} \left(1 + \left\| \mathbf {D} ^ {(t)} \right\| _ {F} + \frac {\left\| \mathbf {V} ^ {(t)} \right\| _ {F}}{\rho_ {t - 1}} + \frac {\left\| \mathbf {V} ^ {(t - 1)} \right\| _ {F}}{\rho_ {t - 1}}\right). \tag {12} +$$ + +Lemma A.3. Let $\{\mathbf{D}^{(t)}\}_{t=0}^{\infty}$ and $\{\mathbf{V}^{(t)}\}_{t=0}^{\infty}$ be the sequence generated according to update rule (5). Then for any $t \geq 1$ , it holds + +$$ +\begin{array}{l} \| \mathbf {D} ^ {(t)} \| _ {F} + \frac {\| \mathbf {V} ^ {(t)} \| _ {F}}{\rho_ {t - 1}} + \frac {\| \mathbf {V} ^ {(t - 1)} \| _ {F}}{\rho_ {t - 1}} \\ \leq \exp \left(3 \left(C _ {F} + C _ {F} ^ {4}\right) \sum_ {s = 1} ^ {t - 1} \frac {1}{\rho_ {s - 1}}\right) \cdot \left(\| \mathbf {D} ^ {(1)} \| _ {F} + \frac {\| \mathbf {V} ^ {(1)} \| _ {F}}{\rho_ {0}} + \frac {\| \mathbf {V} ^ {(0)} \| _ {F}}{\rho_ {0}} + \sum_ {s = 1} ^ {t - 1} \frac {3 \left(C _ {F} + C _ {F} ^ {4}\right)}{\rho_ {s - 1}}\right) \tag {13} \\ \end{array} +$$ + +Returning to the proof of the main theorem, define + +$$ +\begin{array}{l} C _ {A} = 2 (C _ {F} + C _ {F} ^ {4}) \left[ 1 + \exp \left(3 (C _ {F} + C _ {F} ^ {4}) \sum_ {s = 1} ^ {\infty} \frac {1}{\rho_ {s - 1}}\right) \cdot \right. \\ \left. \left(\| \mathbf {D} ^ {(1)} \| _ {F} + \frac {\| \mathbf {V} ^ {(1)} \| _ {F}}{\rho_ {0}} + \frac {\| \mathbf {V} ^ {(0)} \| _ {F}}{\rho_ {0}} + \sum_ {s = 1} ^ {\infty} \frac {3 (C _ {F} + C _ {F} ^ {4})}{\rho_ {s - 1}}\right) \right]. \\ \end{array} +$$ + +It follows from the update rules (5) that $C_A$ is a constant depending on $\mathbf{X}$ , $\widehat{\mathbf{W}}$ , $\lambda$ , $\rho_0$ , and $\sum_{t=0}^{\infty} 1 / \rho_t$ . + +Lemma A.2 together with Lemma A.3 yields + +$$ +\left\| \mathbf {D} ^ {(t + 1)} - \mathbf {D} ^ {(t)} \right\| _ {F} \leq \frac {2 C _ {F} + 2 C _ {F} ^ {4}}{\rho_ {t}} \left(1 + \left\| \mathbf {D} ^ {(t)} \right\| _ {F} + \frac {\left\| \mathbf {V} ^ {(t)} \right\| _ {F}}{\rho_ {t - 1}} + \frac {\left\| \mathbf {V} ^ {(t - 1)} \right\| _ {F}}{\rho_ {t - 1}}\right) \leq \frac {C _ {A}}{\rho_ {t}}. \tag {15} +$$ + +and + +$$ +\left\| \mathbf {V} ^ {(t + 1)} \right\| _ {F} \leq \left(C _ {F} + 2 C _ {F} ^ {4}\right) \left(1 + \left\| \mathbf {D} ^ {(t)} \right\| _ {F} + \frac {\left\| \mathbf {V} ^ {(t)} \right\| _ {F}}{\rho_ {t - 1}} + \frac {\left\| \mathbf {V} ^ {(t - 1)} \right\| _ {F}}{\rho_ {t - 1}}\right) \leq \frac {C _ {A}}{2}. \tag {16} +$$ + +It then follows from $\mathbf{V}$ -update rule and triangle inequality that + +$$ +\begin{array}{l} \| \mathbf {S} ^ {(t + 1)} - \mathbf {S} ^ {(t)} \| _ {F} \leq \| \mathbf {S} ^ {(t + 1)} - \mathbf {D} ^ {(t + 1)} \| _ {F} + \| \mathbf {D} ^ {(t + 1)} - \mathbf {D} ^ {(t)} \| _ {F} + \| \mathbf {S} ^ {(t)} - \mathbf {D} ^ {(t)} \| _ {F} \\ \leq \frac {\| \mathbf {V} ^ {(t + 1)} \| _ {F} + \| \mathbf {V} ^ {(t)} \| _ {F}}{\rho_ {t}} + \| \mathbf {D} ^ {(t + 1)} - \mathbf {D} ^ {(t)} \| _ {F} + \frac {\| \mathbf {V} ^ {(t)} \| _ {F} + \| \mathbf {V} ^ {(t - 1)} \| _ {F}}{\rho_ {t - 1}} \\ \leq \frac {3 C _ {A}}{\rho_ {t - 1}}. \tag {17} \\ \end{array} +$$ + +According to L-update rule, we have + +$$ +\begin{array}{l} \| \mathbf {L} ^ {(t + 1)} - \mathbf {L} ^ {(t)} \| _ {F} = \left\| \mathbf {H} ^ {- 1 / 2} P _ {r} (\mathbf {H} ^ {1 / 2} (\widehat {\mathbf {W}} - \mathbf {S} ^ {(t + 1)})) - \mathbf {H} ^ {- 1 / 2} P _ {r} (\mathbf {H} ^ {1 / 2} (\widehat {\mathbf {W}} - \mathbf {S} ^ {(t)})) \right\| _ {F} \\ \leq \| \mathbf {H} ^ {- 1 / 2} \| _ {2} \left\| P _ {r} \left(\mathbf {H} ^ {1 / 2} \left(\widehat {\mathbf {W}} - \mathbf {S} ^ {(t + 1)}\right)\right) - P _ {r} \left(\mathbf {H} ^ {1 / 2} \left(\widehat {\mathbf {W}} - \mathbf {S} ^ {(t)}\right)\right) \right\| _ {F} \\ \leq C _ {F} \| \mathbf {H} ^ {1 / 2} \| _ {2} \| \mathbf {S} ^ {(t + 1)} - \mathbf {S} ^ {(t)} \| _ {F} \tag {18} \\ \leq C _ {F} ^ {2} \| \mathbf {S} ^ {(t + 1)} - \mathbf {S} ^ {(t)} \| _ {F} \\ \leq \frac {3 C _ {F} ^ {2} C _ {A}}{\rho_ {t - 1}}. \\ \end{array} +$$ + +Therefore, with constant $C = 3C_F^2 C_A$ , we obtain + +$$ +\max \left\{\| \mathbf {S} ^ {(t + 1)} - \mathbf {S} ^ {(t)} \| _ {F}, \| \mathbf {L} ^ {(t + 1)} - \mathbf {L} ^ {(t)} \| _ {F} \right\} \leq \frac {C}{\rho_ {t - 1}}. \tag {19} +$$ + +Since $\sum_{s=0}^{\infty} 1 / \rho_s < \infty$ , both $\{\mathbf{S}^{(t)}\}_{t=0}^{\infty}$ and $\{\mathbf{L}^{(t)}\}_{t=0}^{\infty}$ are Cauchy sequences. Therefore, there exist matrices $\bar{\mathbf{S}}$ and $\bar{\mathbf{L}}$ such that $\mathbf{S}^{(t)} \to \bar{\mathbf{S}}$ and $\mathbf{L}^{(t)} \to \bar{\mathbf{L}}$ as $t \to \infty$ . Setting $\bar{\mathbf{W}} = \bar{\mathbf{S}} + \bar{\mathbf{L}}$ , we conclude that $\mathbf{S}^{(t)} + \mathbf{L}^{(t)} \to \bar{\mathbf{W}}$ as $t \to \infty$ . + +# A.1 Proof of Lemma A.1 + +Proof. The $\mathbf{L}$ -update rule in (5), together with (9) yields + +$$ +\begin{array}{l} \| \mathbf {L} ^ {(t)} \| _ {F} = \left\| \mathbf {H} ^ {- 1 / 2} P _ {r} \left(\mathbf {H} ^ {1 / 2} \left(\widehat {\mathbf {W}} - \mathbf {S} ^ {(t)}\right)\right) \right\| _ {F} \\ \leq \| \mathbf {H} ^ {- 1 / 2} \| _ {2} \left\| P _ {r} \left(\mathbf {H} ^ {1 / 2} \left(\widehat {\mathbf {W}} - \mathbf {S} ^ {(t)}\right)\right) \right\| _ {F} \\ \leq C _ {F} \left\| \mathbf {H} ^ {1 / 2} \left(\widehat {\mathbf {W}} - \mathbf {S} ^ {(t)}\right) \right\| _ {F} \tag {20} \\ \leq C _ {F} \left\| \mathbf {H} ^ {1 / 2} \right\| _ {2} \| \widehat {\mathbf {W}} \| _ {F} + C _ {F} \left\| \mathbf {H} ^ {1 / 2} \right\| _ {2} \left\| \mathbf {S} ^ {(t)} \right\| _ {F} \\ \leq C _ {F} ^ {2} \| \widehat {\mathbf {W}} \| _ {F} + C _ {F} ^ {2} \| \mathbf {S} ^ {(t)} \| _ {F}, \\ \end{array} +$$ + +where the second inequality follows from the non-expansiveness of rank-r projection operator $P_r$ in Frobenius norm. It then follows from the V-update rule in (5) that + +$$ +\begin{array}{l} \| \mathbf {L} ^ {(t)} \| _ {F} \leq C _ {F} ^ {2} \| \widehat {\mathbf {W}} \| _ {F} + C _ {F} ^ {2} \| \mathbf {S} ^ {(t)} \| _ {F} \\ = C _ {F} ^ {2} \| \widehat {\mathbf {W}} \| _ {F} + C _ {F} ^ {2} \left\| \mathbf {D} ^ {(t)} + \frac {\mathbf {V} ^ {(t)} - \mathbf {V} ^ {(t - 1)}}{\rho_ {t - 1}} \right\| _ {F} \tag {21} \\ \leq C _ {F} ^ {3} \left(1 + \| \mathbf {D} ^ {(t)} \| _ {F} + \frac {\| \mathbf {V} ^ {(t)} \| _ {F}}{\rho_ {t - 1}} + \frac {\| \mathbf {V} ^ {(t - 1)} \| _ {F}}{\rho_ {t - 1}}\right). \\ \end{array} +$$ + +□ + +# A.2 Proof of Lemma A.2 + +Proof. According to the S-update rule in (5), it holds + +$$ +\begin{array}{l} \mathbf {S} ^ {(t + 1)} - \mathbf {D} ^ {(t)} + \frac {\mathbf {V} ^ {(t)}}{\rho_ {t}} = (\mathbf {H} + \rho_ {t} \mathbf {I}) ^ {- 1} (\mathbf {G} - \mathbf {H L} ^ {(t)} - \mathbf {V} ^ {(t)} + \rho_ {t} \mathbf {D} ^ {(t)}) - \mathbf {D} ^ {(t)} + \frac {\mathbf {V} ^ {(t)}}{\rho_ {t}} \\ = \left(\left(\mathbf {H} + \rho_ {t} \mathbf {I}\right) ^ {- 1} \rho_ {t} - \mathbf {I}\right) \mathbf {D} ^ {(t)} + \left(\mathbf {H} + \rho_ {t} \mathbf {I}\right) ^ {- 1} \left(\mathbf {G} - \mathbf {H L} ^ {(t)} - \mathbf {V} ^ {(t)}\right) + \frac {\mathbf {V} ^ {(t)}}{\rho_ {t}} \\ = - \frac {1}{\rho_ {t}} \left(\mathbf {I} + \frac {\mathbf {H}}{\rho_ {t}}\right) ^ {- 1} \mathbf {H D} ^ {(t)} + \frac {1}{\rho_ {t}} \left(\mathbf {I} + \frac {\mathbf {H}}{\rho_ {t}}\right) ^ {- 1} \left(\mathbf {G} - \mathbf {H L} ^ {(t)} - \mathbf {V} ^ {(t)}\right) + \frac {\mathbf {V} ^ {(t)}}{\rho_ {t}} \\ = \frac {1}{\rho_ {t}} \left(\mathbf {I} + \frac {\mathbf {H}}{\rho_ {t}}\right) ^ {- 1} \left(\mathbf {G} - \mathbf {H L} ^ {(t)} - \mathbf {H D} ^ {(t)}\right) + \frac {1}{\rho_ {t}} \left[ \mathbf {I} - \left(\mathbf {I} + \frac {\mathbf {H}}{\rho_ {t}}\right) ^ {- 1} \right] \mathbf {V} ^ {(t)} \\ = \frac {1}{\rho_ {t}} \left(\mathbf {I} + \frac {\mathbf {H}}{\rho_ {t}}\right) ^ {- 1} \left(\mathbf {G} - \mathbf {H L} ^ {(t)} - \mathbf {H D} ^ {(t)} + \frac {\mathbf {H V} ^ {(t)}}{\rho_ {t}}\right) \tag {22} \\ \end{array} +$$ + +Therefore, we obtain + +$$ +\begin{array}{l} \left\| \mathbf {S} ^ {(t + 1)} - \mathbf {D} ^ {(t)} + \frac {\mathbf {V} ^ {(t)}}{\rho_ {t}} \right\| _ {F} \leq \frac {1}{\rho_ {t}} \left\| \left(\mathbf {I} + \frac {\mathbf {H}}{\rho_ {t}}\right) ^ {- 1} \right\| _ {2} \left\| \mathbf {G} - \mathbf {H L} ^ {(t)} - \mathbf {H D} ^ {(t)} + \frac {\mathbf {H V} ^ {(t)}}{\rho_ {t}} \right\| _ {F} \\ \leq \frac {1}{\rho_ {t}} \left\| \mathbf {G} - \mathbf {H L} ^ {(t)} - \mathbf {H D} ^ {(t)} + \frac {\mathbf {H V} ^ {(t)}}{\rho_ {t}} \right\| _ {F} \tag {23} \\ \leq \frac {1}{\rho_ {t}} \left(\| \mathbf {G} - \mathbf {H L} ^ {(t)} - \mathbf {H D} ^ {(t)} \| _ {F} + \frac {\| \mathbf {H V} ^ {(t)} \| _ {F}}{\rho_ {t}}\right). \\ \end{array} +$$ + +Denote $\tilde{\mathcal{I}} := \{(i,j) \in [N_{in}] \times [N_{out}] \mid \mathbf{D}_{ij}^{(t)} = 0\}$ . It follows from the D-update rule and the definition of the projection operator that + +$$ +\begin{array}{l} \left\| \mathbf{D}^{(t + 1)} - \mathbf{S}^{(t + 1)} - \frac{\mathbf{V}^{(t)}}{\rho_{t}}\right\|_{F}^{2} = \min_{\substack{\mathcal{I}\subseteq [N_{in}]\times [N_{out}]\\ |\mathcal{I}| = N_{in}N_{out} - k}}\sum_{(i,j)\in \mathcal{I}}\left(\mathbf{S}^{(t + 1)} + \frac{\mathbf{V}^{(t)}}{\rho_{t}}\right)_{i,j}^{2} \\ \leq \sum_ {(i, j) \in \tilde {\mathcal {I}}} \left(\mathbf {S} ^ {(t + 1)} + \frac {\mathbf {V} ^ {(t)}}{\rho_ {t}}\right) _ {i, j} ^ {2} = \sum_ {(i, j) \in \tilde {\mathcal {I}}} \left(\mathbf {S} ^ {(t + 1)} - \mathbf {D} ^ {(t)} + \frac {\mathbf {V} ^ {(t)}}{\rho_ {t}}\right) _ {i, j} ^ {2} \tag {24} \\ \leq \left\| \mathbf {S} ^ {(t + 1)} - \mathbf {D} ^ {(t)} + \frac {\mathbf {V} ^ {(t)}}{\rho_ {t}} \right\| _ {F} ^ {2} \\ \end{array} +$$ + +Together with (23), we get + +$$ +\left\| \mathbf {D} ^ {(t + 1)} - \mathbf {S} ^ {(t + 1)} - \frac {\mathbf {V} ^ {(t)}}{\rho_ {t}} \right\| _ {F} \leq \frac {1}{\rho_ {t}} \left(\| \mathbf {G} - \mathbf {H L} ^ {(t)} - \mathbf {H D} ^ {(t)} \| _ {F} + \frac {\| \mathbf {H V} ^ {(t)} \| _ {F}}{\rho_ {t}}\right). \tag {25} +$$ + +It then follows from the V-update rule that + +$$ +\frac {\left\| \mathbf {V} ^ {(t + 1)} \right\| _ {F}}{\rho_ {t}} = \left\| \mathbf {D} ^ {(t + 1)} - \mathbf {S} ^ {(t + 1)} - \frac {\mathbf {V} ^ {(t)}}{\rho_ {t}} \right\| _ {F} \leq \frac {1}{\rho_ {t}} \left(\left\| \mathbf {G} - \mathbf {H L} ^ {(t)} - \mathbf {H D} ^ {(t)} \right\| _ {F} + \frac {\left\| \mathbf {H V} ^ {(t)} \right\| _ {F}}{\rho_ {t}}\right) \tag {26} +$$ + +According to Lemma A.1 and the monotonicity of $\{\rho_t\}_{t=0}^{\infty}$ , it holds + +$$ +\begin{array}{l} \| \mathbf {G} - \mathbf {H L} ^ {(t)} - \mathbf {H D} ^ {(t)} \| _ {F} + \frac {\| \mathbf {H V} ^ {(t)} \| _ {F}}{\rho_ {t}} \leq \| \mathbf {G} \| _ {F} + \| \mathbf {H} \| _ {2} \| \mathbf {L} ^ {(t)} \| _ {F} + \| \mathbf {H} \| _ {2} \| \mathbf {D} ^ {(t)} \| _ {F} + \frac {\| \mathbf {H} \| _ {2} \| \mathbf {V} ^ {(t)} \| _ {F}}{\rho_ {t}} \\ \leq C _ {F} \left(1 + \| \mathbf {D} ^ {(t)} \| _ {F} + \frac {\| \mathbf {V} ^ {(t)} \| _ {F}}{\rho_ {t - 1}}\right) + C _ {F} \| \mathbf {L} ^ {(t)} \| _ {F} \\ \leq \left(C _ {F} + C _ {F} ^ {4}\right) \left(1 + \| \mathbf {D} ^ {(t)} \| _ {F} + \frac {\| \mathbf {V} ^ {(t)} \| _ {F}}{\rho_ {t - 1}} + \frac {\| \mathbf {V} ^ {(t - 1)} \| _ {F}}{\rho_ {t - 1}}\right). \tag {27} \\ \end{array} +$$ + +Together with inequality (26), this establishes the first inequality of the lemma. Furthermore, by summing up (23) and (25) and applying the triangle inequality, we verify the second inequality. $\square$ + +# A.3 Proof of Lemma A.3 + +Proof. It follows from Lemma A.2 that + +$$ +\frac {\left\| \mathbf {V} ^ {(t + 1)} \right\| _ {F}}{\rho_ {t}} \leq \frac {C _ {F} + C _ {F} ^ {4}}{\rho_ {t}} \left(1 + \left\| \mathbf {D} ^ {(t)} \right\| _ {F} + \frac {\left\| \mathbf {V} ^ {(t)} \right\| _ {F}}{\rho_ {t - 1}} + \frac {\left\| \mathbf {V} ^ {(t - 1)} \right\| _ {F}}{\rho_ {t - 1}}\right) \tag {28} +$$ + +and + +$$ +\begin{array}{l} \| \mathbf {D} ^ {(t + 1)} \| _ {F} \leq \| \mathbf {D} ^ {(t)} \| _ {F} + \| \mathbf {D} ^ {(t + 1)} - \mathbf {D} ^ {(t)} \| _ {F} \\ \leq \left\| \mathbf {D} ^ {(t)} \right\| _ {F} + \frac {2 C _ {F} + 2 C _ {F} ^ {4}}{\rho_ {t}} \left(1 + \left\| \mathbf {D} ^ {(t)} \right\| _ {F} + \frac {\left\| \mathbf {V} ^ {(t)} \right\| _ {F}}{\rho_ {t - 1}} + \frac {\left\| \mathbf {V} ^ {(t - 1)} \right\| _ {F}}{\rho_ {t - 1}}\right). \tag {29} \\ \end{array} +$$ + +Summing up these two inequalities yields + +$$ +\begin{array}{l} \| \mathbf {D} ^ {(t + 1)} \| _ {F} + \frac {\| \mathbf {V} ^ {(t + 1)} \| _ {F}}{\rho_ {t}} + \frac {\| \mathbf {V} ^ {(t)} \| _ {F}}{\rho_ {t}} \leq \| \mathbf {D} ^ {(t + 1)} \| _ {F} + \frac {\| \mathbf {V} ^ {(t + 1)} \| _ {F}}{\rho_ {t}} + \frac {\| \mathbf {V} ^ {(t)} \| _ {F}}{\rho_ {t - 1}} \\ \leq \frac {3 C _ {F} + 3 C _ {F} ^ {4}}{\rho_ {t}} \left(1 + \| \mathbf {D} ^ {(t)} \| _ {F} + \frac {\| \mathbf {V} ^ {(t)} \| _ {F}}{\rho_ {t - 1}} + \frac {\| \mathbf {V} ^ {(t - 1)} \| _ {F}}{\rho_ {t - 1}}\right) + \| \mathbf {D} ^ {(t)} \| _ {F} + \frac {\| \mathbf {V} ^ {(t)} \| _ {F}}{\rho_ {t - 1}} \tag {30} \\ \leq \left(1 + \frac {3 C _ {F} + 3 C _ {F} ^ {4}}{\rho_ {t - 1}}\right) \left(\| \mathbf {D} ^ {(t)} \| _ {F} + \frac {\| \mathbf {V} ^ {(t)} \| _ {F}}{\rho_ {t - 1}} + \frac {\| \mathbf {V} ^ {(t - 1)} \| _ {F}}{\rho_ {t - 1}}\right) + \frac {3 C _ {F} + 3 C _ {F} ^ {4}}{\rho_ {t - 1}}, \\ \end{array} +$$ + +Denote $a_{t} \coloneqq \| \mathbf{D}^{(t)}\|_{F} + \| \mathbf{V}^{(t)}\|_{F} / \rho_{t - 1} + \| \mathbf{V}^{(t - 1)}\|_{F} / \rho_{t - 1}$ , then the above inequality can be rewritten as + +$$ +a _ {t + 1} \leq \left(1 + \frac {3 C _ {F} + 3 C _ {F} ^ {4}}{\rho_ {t - 1}}\right) a _ {t} + \frac {3 C _ {F} + 3 C _ {F} ^ {4}}{\rho_ {t - 1}} \tag {31} +$$ + +Therefore, + +$$ +\begin{array}{l} \frac {a _ {t + 1}}{\prod_ {s = 1} ^ {t} (1 + 3 (C _ {F} + C _ {F} ^ {4}) / \rho_ {s - 1})} \leq \frac {a _ {t}}{\prod_ {s = 1} ^ {t - 1} (1 + 3 (C _ {F} + C _ {F} ^ {4}) / \rho_ {s - 1})} + \frac {3 (C _ {F} + C _ {F} ^ {4})}{\rho_ {t - 1} \prod_ {s = 0} ^ {t} (1 + 3 (C _ {F} + C _ {F} ^ {4}) / \rho_ {s - 1})} \\ \leq \frac {a _ {t}}{\prod_ {s = 1} ^ {t - 1} \left(1 + 3 \left(C _ {F} + C _ {F} ^ {4}\right) / \rho_ {s - 1}\right)} + \frac {3 \left(C _ {F} + C _ {F} ^ {4}\right)}{\rho_ {t - 1}} \tag {32} \\ \end{array} +$$ + +It then follows from telescoping that + +$$ +\frac {a _ {t}}{\prod_ {s = 1} ^ {t - 1} \left(1 + 3 \left(C _ {F} + C _ {F} ^ {4}\right) / \rho_ {s - 1}\right)} \leq a _ {1} + \sum_ {s = 1} ^ {t - 1} \frac {3 \left(C _ {F} + C _ {F} ^ {4}\right)}{\rho_ {s - 1}} \tag {33} +$$ + +Note that + +$$ +\prod_ {s = 1} ^ {t - 1} \left(1 + 3 \left(C _ {F} + C _ {F} ^ {4}\right) / \rho_ {s - 1}\right) \leq \exp \left(3 \left(C _ {F} + C _ {F} ^ {4}\right) \sum_ {s = 1} ^ {t - 1} \frac {1}{\rho_ {s - 1}}\right), \tag {34} +$$ + +recalling the definition of $a_{t}$ completes the proof. + +# B Additional Experimental Details + +Computing environments All experiments were conducted on a computing cluster. Unless otherwise specified, we utilized an Intel Xeon Gold 6248 machine with 16 CPU cores and a single NVIDIA L40 48GB / A100 80GB / H100 80GB GPU. When runtime compression results are reported, all experiments have been run on the same node (including GPU) configuration. All language models and pruning methods were implemented using the PyTorch library Paszke et al. [2017]. + +Implementation Details of 3BASiL We use $\mathbf{H}' = \mathbf{H} + 0.005\mathrm{diag}(\mathbf{X}^\top \mathbf{X}) + 0.005\mathrm{Tr}(\mathbf{X}^\top \mathbf{X})\mathbf{I}$ . + +In practice, we employ an iteration-dependent penalty parameter $\rho_{t}$ , giving the following updates at iteration $t$ : + +$$ +\mathbf {S} ^ {(t + 1)} = (\mathbf {H} + \rho_ {t} \mathbf {I}) ^ {- 1} \left(\mathbf {H} \left(\widehat {\mathbf {W}} - \mathbf {L} ^ {(t)}\right) - \mathbf {V} ^ {(t)} + \rho_ {t} \mathbf {D} ^ {(t)}\right) \mathbf {L} ^ {(t + 1)} = \mathbf {H} ^ {- 1 / 2} P _ {r} \left(\mathbf {H} ^ {1 / 2} \left(\widehat {\mathbf {W}} - \mathbf {S} ^ {(t + 1)}\right)\right) +$$ + +$$ +\mathbf {D} ^ {(t + 1)} = P _ {\mathcal {S}} \left(\mathbf {S} ^ {(t + 1)} + \mathbf {V} ^ {(t)} / \rho_ {t}\right) \quad \mathbf {V} ^ {(t + 1)} = \mathbf {V} ^ {(t)} + \rho_ {t} \left(\mathbf {S} ^ {(t + 1)} - \mathbf {D} ^ {(t + 1)}\right). \tag {35} +$$ + +The initial $\rho_0 = 0.1$ . The $\rho$ -update for ADMM depends on the support change similar to what was proposed by Meng et al. [2024a]. The $(\mathbf{S} + \mathbf{LR})$ decomposition is more "sensitive" to increasing $\rho$ aggressively compared to pure pruning in the works of Meng et al. [2024a]. We use the following $\rho$ update rules. We update $\rho$ every 10 iteration based on a step function that depends on the current value of $\rho_t$ and $s_t \coloneqq |\operatorname{Supp}(\mathbf{D}^{(t)})\Delta \operatorname{Supp}(\mathbf{D}^{(t - 10)})|$ , which represents the number of elements in the symmetric difference between $\operatorname{Supp}(\mathbf{D}^{(t)})$ and $\operatorname{Supp}(\mathbf{D}^{(t - 10)})$ . Specifically, we set + +$$ +\rho_ {t + 1} = \left\{ \begin{array}{l l} 1. 1 \rho_ {t} & \text {i f} s _ {t} \geq 0. 1 k, \\ 1. 0 5 \rho_ {t} & \text {i f} s _ {t} \geq 0. 0 0 5 k, \\ 1. 0 2 \rho_ {t} & \text {i f} s _ {t} \geq 0. 5. \end{array} \right. \tag {36} +$$ + +It is worth noting that the algorithm can converge significantly faster if we set these parameters to the ones proposed by Meng et al. [2024a] (ADMM for pruning) but the solution quality can be slightly compromised. + +Implementation Details of transformer matching (TM) Given a transformer $T_{i}$ with input activations $\mathbf{X}_{\mathrm{i}}$ (obtained from the outputs of the previously compressed transformer block $T_{i-1}$ , we start by creating a copy of $T_{i}$ , termed $T_{i}^{\mathrm{ori}}$ ). We then compress $T_{i}$ layers using an $(\mathbf{S} + \mathbf{LR})$ method. We now replace dense layers with LoRA layers that contain new linear sparse layers and low-rank components $\mathbf{A}, \mathbf{B}$ . We set all parameters in transformer block $T_{i}$ to be trainable and minimize using Adam the loss $\| T_{i}(\mathbf{X}_{\mathrm{i}}) - T_{i}^{(ori)}(\mathbf{X}_{\mathrm{i}})\|_{F}^{2}$ . The input activations fed into subsequent transformer blocks are $T_{i}^{(TM)}(\mathbf{X}_{\mathrm{i}})$ , where $T_{i}^{(TM)}$ is the transformer block after $(\mathbf{S} + \mathbf{LR})$ decomposition and TM refinement steps. + +For TM step, we employ the Adam optimizer with PyTorch's default hyperparameters. We use 20 epochs (on the 128 calibration data points selected for compression). The batch size used is 8. The learning rate is $2e^{-5}$ using a Cosine Annealing Scheduler with $\eta_{\mathrm{min}} = 4e^{-6}$ . + +Baseline Implementation Details Below are the implementation specifications for: + +- OATS: We adopt the official implementation from Zhang and Papyan [2025] (accessible via GitHub) and apply the default hyperparameters and 80 alternating minimization steps. +- HASSLE-free-SparseGPT: We adopt the official implementation from Makni et al. [2025] (accessible via GitHub) and provide an improved implementation that uses the closed-form solution Equation (4) for the low-rank fitting step. We apply the default hyperparameters and 80 alternating minimization steps. +- HASSLE-free-ALPS: We adopt the official implementation from Makni et al. [2025] (accessible via GitHub) and provide an improved implementation that uses the closed-form solution Equation (4) for the low-rank fitting step. We apply the default hyperparameters and 80 alternating minimization steps. + +In Table 3, we use the values reported in Makni et al. [2025]. For all other reported values, instead of minimizing $\| \mathbf{X}(\mathbf{W} - \mathbf{M})\|_F$ s.t. $\mathrm{rank}(\mathbf{M}) \leq r$ by reparameterizing $\mathbf{M} = \mathbf{U}\mathbf{V}^\top$ and optimizing + +
ModelAlgorithmPerplexity (↓)Zero-shot (↑)
C4WT2PTBPIQAHSARC-EARC-CWGRTEOQABoolQAvg
Llama3-8BHf-SparseGPT-original18.0612.6618.6674.8664.7763.8537.3769.2256.6836.4076.1259.91
Hf-SparseGPT-ours17.7712.3818.7174.8165.0466.1638.5770.0954.8738.4077.7160.71
EoRA-SparseGPT21.8915.6923.9172.2558.8757.7034.5666.3054.8733.8073.5856.49
Hf-ALPS-original16.7611.8317.7675.0866.3763.6437.5469.6964.6237.2077.8961.50
Hf-ALPS-ours16.1511.3816.7175.1967.1064.4438.9169.5359.9339.4078.3861.61
EoRA-ALPS18.6913.6120.5573.9962.2161.0737.2068.5957.4036.0074.1658.83
+ +Table 6: Comparison of (original paper), our reproduced results (with improved implementation) and EoRA (reduces to HASSLE-free with alternating minimization steps set to 1) for Llama3-8B under the configuration $(2:4 + 64)$ . Perplexity (lower is better), Zero-shot accuracy (higher is better). + +with gradient-descent on $\mathbf{U}$ and $\mathbf{V}$ as proposed by the authors, we use our improved implementation of HASSLE-free with closed-form solution Equation (4). This results in significant speedup improvements. A slight improvement in LLM evaluation benchmarks is also sometimes observed using the improved implementation. This is expected because gradient-descent on $\mathbf{U}$ and $\mathbf{V}$ approximately solves the reduced-rank regression problem, whereas the closed-form solution is an optimal solution. + +Table 6 shows an extract of the differences between the implementation of HASSLE-free proposed in Makni et al. [2025] and ours (using closed-form solution for low-rank update). Moreover, the original paper reports a compression runtime (of a Llama3-8B under a 2:4+64LR configuration) of 20.13 hours using a single A100 80GB GPU, whereas we report a compression runtime (for the same setup) of 15.71 hours in Figure 2 (using a single A100 80GB GPU) thanks to the efficiency of the closed-form solution. It is worth noting that 3BASiL and 3BASiL-TM are still over 7 times and 3 times, respectively, faster than HASSLE-free-ALPS, even when using the improved implementation for HASSLE-free. + +LoRA Finetuning Details We follow a similar LoRA fine-tuning pipeline to the one introduced in Guo et al. [2024]. For LoRA fine-tuning use a learning rate of 2e-5 and a batch size of size 64 per step. The block size used is 1024 tokens per batch. The effective batch size is obtained by using a physical batch size of 2 on GPU with 32 gradient accumulation steps before each weight update. Training is conducted on $10\%$ of the first shard of the C4 training dataset, which contains over 15 million tokens. We employ the Adam optimizer with PyTorch's default hyperparameters. A cosine learning rate scheduler is used, with a warm-up ratio of 0.03 and no weight decay applied. + +Layer-wise reconstruction error of 3BASiL Figure 5a and Figure 5b show the objective of Equation (1) attained by 3BASiL and other $(\mathbf{S} + \mathbf{LR})$ methods for the first transformer block of a 2:4+64LR decomposition of a Llama3-8B model for Attention and MLP layers, respectively. + +![](images/ae4577e4d7e178e69c3fd37bf1aa9f3ef89901c352a3a2bdbd941a24627682d1.jpg) +(a) True loss for attention layers (linear scale). + +![](images/2d98e4f2cfdd064f2b6bea43b3bd913ddfab0287ea752f6fd8d2bfa9964fbaf9.jpg) +(b) True loss for MLP layers (log scale). +Figure 5: Comparison of true loss values introduced in Equation (1) across different $(\mathbf{S} + \mathbf{LR})$ methods. Lower values indicate better optimization quality. 3BASiL consistently outperforms other methods, particularly for attention layers. + +# C Additional Experimental Results + +We provide additional performance results considered in Section 4. We compare different $(\mathbf{S} + \mathbf{L}\mathbf{R})$ algorithms and their TM-enhanced versions (apply TM as an add-on to the decomposition algorithm). In that case, we add the suffix -TM to the algorithm. We mark algorithms with TM in gray. We also study the results of the $(\mathbf{S} + \mathbf{L}\mathbf{R})$ decomposition after LoRA fine-tuning as described in Appendix B. In that case, we add the prefix LFT- to the algorithm. + +Example: LFT-0ATS-TM denotes the results of $(\mathbf{S} + \mathbf{LR})$ decomposition after (i) using OATS to obtain sparse and low-rank components, (ii) refine these decomposed components with TM and (iii) LoRA fine-tunes the model by using the low-rank components from the $(\mathbf{S} + \mathbf{LR})$ decomposition as a smart initialization. + +
MethodConfigC4↓WT2↓PTB↓PIQA ↑HS↑ARC-E ↑ARC-C ↑WG ↑RTE ↑OQA ↑BoolQ ↑Avg
OATS2:8+64LR640.86605.20779.8652.0127.6728.6623.1249.9652.7125.0037.8337.12
OATS-TM116.2999.92126.0655.5528.7831.4420.5651.3852.7126.8046.7939.25
Hackle-free-SparseGPT162.45134.21170.1254.1928.2831.4022.0149.1752.7126.8057.2540.23
Hackle-free-SparseGPT-TM74.5067.5988.0557.2430.0333.1221.3352.3352.7126.2060.3441.66
Hackle-free-ALPS107.1494.71124.1755.3929.9832.0720.8252.7253.0727.2049.2740.06
Hackle-free-ALPS-TM58.3052.3073.9758.8132.2134.9321.5952.8052.7126.2062.2042.68
3BASI-L97.5086.59100.3556.9130.4932.3721.0853.7553.0724.4061.7441.73
3BASI-L-TM55.2449.7469.4958.8132.8035.1422.7853.0453.0726.6062.1443.05
OATS3:8+64LR125.9192.13115.8057.1332.1435.9822.6151.4651.9926.8062.5142.58
OATS-TM36.3227.6941.9863.1738.4743.6923.2152.8852.7129.8062.0845.75
Hackle-free-SparseGPT43.5034.1851.1661.8638.2441.2924.6653.7552.7129.2062.0845.47
Hackle-free-SparseGPT-TM30.4824.1837.2865.1841.4045.4125.6855.9652.7130.4062.2647.38
Hackle-free-ALPS37.8029.0043.6064.0441.4742.3425.8554.2254.5130.4062.0246.86
Hackle-free-ALPS-TM27.3421.4234.2366.3844.4845.6225.5155.8853.4330.4062.2347.99
3BASI-L34.8126.9641.5564.9142.3445.2427.4756.1253.0731.2062.6047.87
3BASI-L-TM26.2620.7532.0966.4345.4747.4727.0557.7752.7132.0062.2048.89
OATS4:8+64LR28.0619.6932.9067.3049.0748.4827.6556.2052.7130.2062.4549.26
OATS-TM20.6514.5924.5069.3751.3755.0531.4057.2254.5130.6062.8151.54
Hackle-free-SparseGPT22.2415.9027.3570.5152.0850.7629.4457.3858.1234.2062.6351.89
Hackle-free-SparseGPT-TM19.6313.8524.1770.2453.1255.0531.1456.8354.8733.8063.1852.28
Hackle-free-ALPS20.7114.9024.7569.5953.2753.0729.7857.7753.7933.6063.3351.78
Hackle-free-ALPS-TM19.0713.7023.0671.1154.8255.3031.8358.6452.7131.6062.4252.30
3BASI-L20.0414.2624.2770.6254.5555.7230.7260.0655.2334.0063.0652.99
3BASI-L-TM18.6613.1922.4672.0955.4855.6031.4859.1253.0734.0063.4653.04
OATS2:4+64LR41.8028.4545.3663.2841.8947.0126.3753.5151.2628.4063.0946.85
OATS-TM23.8917.0427.9968.2347.7651.0127.7355.6456.3232.6062.3250.20
Hackle-free-SparseGPT27.2519.4532.6367.6347.7045.9626.9655.8852.7130.4062.1448.67
Hackle-free-SparseGPT-TM22.1716.4126.6769.1049.9050.2927.9956.5957.0433.4062.4250.84
Hackle-free-ALPS23.9017.6628.9669.1549.6249.6628.1657.7755.2332.0063.0650.58
Hackle-free-ALPS-TM20.9315.3525.1570.4651.1451.0928.2458.3357.0434.6063.5551.81
3BASI-L23.1617.2727.7769.8051.7451.3527.8258.7254.8733.4062.8451.32
3BASI-L-TM20.4615.2324.6070.1852.9152.0630.1258.9652.7133.4062.3951.59
Llama-3.2-1B Dense-14.019.7517.5974.5963.6660.4836.2660.6956.6837.2063.9856.69
+ +Table 7: One-shot (N:M Sparse + LR) decomposition performance for Llama-3.2-1B. For Perplexity, $(\downarrow)$ lower values are better. For zero-shot tasks, $(\uparrow)$ higher values are better. + +
MethodConfigC4 ↓WT2 ↓PTB ↓PIQA ↑HS ↑ARC-E ↑ARC-C ↑WG ↑RTE ↑OQA ↑BoolQ ↑Avg
OATS2:8+64LR531.47494.31674.7152.5027.3328.1623.2949.5752.7126.6039.6037.47
OATS-TM100.8787.20120.9856.6429.0130.1320.6550.9952.7126.0062.1141.03
Hassle-free-SparseGPT106.07106.17151.9254.6229.6729.9221.6750.2852.7126.6061.9340.93
Hassle-free-SparseGPT-TM61.5056.0290.3758.9832.5333.7522.1051.7852.7126.4062.1142.55
Hassle-free-ALPS69.9665.34108.6857.3432.5933.5920.8250.6752.7127.0062.2642.12
Hassle-free-ALPS-TM46.1244.0361.2561.2636.3636.8323.1252.8052.7125.0062.5143.82
3BASI L73.0072.26110.1057.2932.6234.0121.4251.1452.7126.8062.2042.27
3BASI-TM45.3542.3868.2961.1036.9038.1722.7853.1253.0726.0062.6644.23
OATS3:8+64LR65.0847.2781.2961.7537.8042.1723.8952.8852.7127.2062.7545.14
OATS-TM27.0920.9430.2167.6847.2651.2628.4157.4652.7129.2064.6549.83
Hassle-free-SparseGPT34.6626.6039.7665.9446.1947.7726.8858.9653.0729.6065.0249.18
Hassle-free-SparseGPT-TM23.6919.5427.4569.7051.6452.7829.3560.2255.9630.6062.7251.62
Hassle-free-ALPS27.9422.7734.5969.1550.1853.3229.0161.4852.7132.0063.5851.43
Hassle-free-ALPS-TM21.5217.8026.4271.0054.4557.3730.8059.7556.3233.4066.0253.64
3BASI L26.3520.6631.7768.6651.4452.1029.9561.2554.1531.2068.3252.13
3BASI-TM20.8917.1825.3171.8255.3555.2232.0860.8554.1533.4065.4153.53
OATS4:8+64LR19.2513.4021.6772.4761.1060.8235.1566.3057.4034.0073.2457.56
OATS-TM15.9211.0017.8274.2163.4665.6137.8066.1464.6236.8072.2660.11
Hassle-free-SparseGPT17.0912.3019.1973.8363.0164.2336.5265.5958.1237.8072.0858.90
Hassle-free-SparseGPT-TM15.4210.8417.2374.6565.2264.9438.1465.1958.8440.4069.6959.63
Hassle-free-ALPS16.0411.5118.1774.5464.6363.7636.9566.3859.5736.8072.0859.34
Hassle-free-ALPS-TM15.0710.5416.8475.1965.9366.8440.0267.4060.6540.0074.0161.25
3BASI L15.6510.9717.3975.6865.8766.4639.4267.2559.9338.8073.5260.87
3BASI-TM14.8910.2916.5275.7966.4667.0538.8266.0659.9339.2072.3260.70
OATS2:4+64LR25.1817.4128.6070.8954.7657.7432.7661.1753.0732.8070.4054.20
OATS-TM18.0812.8520.2272.4259.4662.7935.2462.2758.8434.6070.2156.98
Hassle-free-SparseGPT20.3815.0323.2371.5558.6259.9332.9463.8557.4033.6069.9455.98
Hassle-free-SparseGPT-TM17.2412.6619.4173.7861.5061.6634.6463.6958.4837.2068.6257.45
Hassle-free-ALPS18.4513.7920.5073.7860.8263.3035.4964.5657.4035.8072.7857.99
Hassle-free-ALPS-TM16.6012.1118.5974.2162.9763.0537.5466.1458.8436.0070.5858.67
3BASI L17.8913.1220.1073.3461.9962.5035.0766.4661.7339.6071.8059.06
3BASI-TM16.3711.7918.3473.7863.3863.0538.0564.2559.5736.8071.1358.75
Llama-3.2-3B Dense-11.337.8113.5377.4873.6171.6345.9969.8554.5143.0073.3963.68
+ +Table 8: One-shot (N:M Sparse + LR) decomposition performance for Llama-3.2-3B. For Perplexity, $(\downarrow)$ lower values are better. For zero-shot tasks, $(\uparrow)$ higher values are better. + +
MethodConfigC4↓WT2↓PTB↓PIQA↑HS↑ARC-E↑ARC-C↑WG↑RTE↑OQA↑BoolQ↑Avg
OATS2:8+64LR424.55431.81590.8851.6328.1127.6123.7249.7252.7127.0038.2637.34
73.4264.21100.5958.4331.2332.2020.2251.3852.7126.0062.2941.81
88.3996.61109.7154.9530.7731.4820.6550.7552.7126.2061.5941.14
46.2442.7571.6661.4836.7636.7423.1253.5152.7127.6063.1544.38
60.1656.0377.1157.9434.9234.6421.6754.6252.7127.8056.1242.55
36.5034.3150.1464.9141.1840.2824.4056.9952.7128.2059.6046.03
56.9953.8372.4859.2535.3435.9821.5954.0652.7127.4064.6243.87
36.1633.5152.8763.6041.4739.8124.3258.4152.7126.8063.6746.35
OATS3:8+64LR58.8840.7667.3563.7139.4842.6824.3253.9152.7128.4063.9846.15
22.6717.1724.4671.2254.2954.8831.4863.2254.1532.8071.3854.18
29.3221.4632.0668.6651.9950.9730.3863.8553.0732.0071.3152.78
19.9715.5122.2872.8558.7258.1233.9665.6761.3732.0074.9257.20
23.9318.2026.3170.6256.5454.4230.1264.7255.2332.8071.9654.55
18.3814.5220.1574.4861.5659.0133.2867.4857.4035.2075.1157.94
23.0718.0324.8471.0656.9657.7032.5966.6954.5133.0066.7054.90
18.1114.2620.4774.0561.8560.7334.7365.9854.5134.8076.9157.94
OATS4:8+64LR16.3810.8817.2375.8467.6067.0941.2170.8860.2938.2073.6161.84
13.779.0614.3776.8270.1570.1643.5270.8865.7040.2077.8964.41
14.659.8815.2177.0969.9569.3241.8171.2756.3240.6079.3963.22
13.408.9014.1177.5871.4273.2343.6070.4064.9841.4079.3965.25
14.049.4414.4576.8271.1971.0444.4572.7756.6840.2078.1363.91
13.218.7113.8578.5672.5472.8145.7371.4365.3441.4079.8865.96
13.749.2114.2476.8872.0570.1644.8072.1461.0141.4080.8964.92
13.028.6413.7078.2472.5973.1147.3571.9863.1842.4080.4966.17
OATS2:4+64LR21.5914.7623.4172.7460.7060.8634.8165.5157.7635.2068.3256.99
15.4910.6116.1176.0165.6667.0040.6168.5956.6836.6075.6960.86
17.7712.3818.7174.8165.0466.1638.5770.0954.8738.4077.7160.71
14.9510.2815.9776.8868.1867.2141.8169.4664.9838.2078.8163.19
16.1511.3816.7175.1967.1064.4438.9169.5359.9339.4078.3861.61
14.4510.0015.2377.0969.3267.2640.1070.1760.6538.8075.7562.39
15.7611.2316.2576.5067.6167.2140.1070.2464.2638.2078.2962.80
14.349.7814.8877.4869.5867.2140.5371.2761.3739.8079.5163.34
Meta-Llama-3-8B Dense-9.446.1411.1880.7979.1777.6953.3372.8569.6845.0081.4469.99
+ +Table 9: One-shot (N:M Sparse + LR) decomposition performance for Meta-Llama-3-8B. For Perplexity, $(\downarrow)$ lower values are better. For zero-shot tasks, $(\uparrow)$ higher values are better. + +
MethodConfigC4↓WT2↓PTB↓PIQA↑HS↑ARC-E↑ARC-C↑WG↑RTE↑OQA↑BoolQ↑Avg
LFT-OATS2:8+64LR71.4163.5895.0556.5829.2230.7222.1053.8352.7123.0061.1041.16
LFT-OATS-TM53.0147.6868.9859.0931.0736.1521.8452.4952.7124.6062.1442.51
LFT-Hasle-free-SparseGPT51.1947.0963.8759.0331.1635.1422.1850.8352.7126.6060.7042.29
LFT-Hasle-free-SparseGPT-TM42.8539.4655.5360.5533.3735.2722.1850.3652.7128.0062.1443.07
LFT-Hasle-free-ALPS44.1541.0257.4559.3633.2234.6822.1851.9352.7126.6060.9542.70
LFT-Hasle-free-ALPS-TM37.5435.5050.1560.2835.8037.5023.6355.0952.7127.2062.1444.29
LFT-3BASIL36.0933.4245.1360.5535.5237.7123.2153.1253.0727.0062.0244.02
LFT-3BASIL-TM36.0334.1347.1160.0736.1637.2924.6655.4153.0729.4062.3244.80
LFT-OATS3:8+64LR32.2025.4741.7462.3539.7144.3625.6852.3353.7929.2062.1446.20
LFT-OATS-TM27.0521.1434.8365.5143.3746.2526.6254.9352.7129.6061.9947.62
LFT-Hasle-free-SparseGPT27.1721.9634.5565.7244.0944.9926.7955.1752.7130.6061.5947.71
LFT-Hasle-free-SparseGPT-TM24.7420.3832.3667.0345.7148.3626.6255.0152.7130.2062.0248.46
LFT-Hasle-free-ALPS25.2020.0332.2665.6146.1945.1228.1654.8555.9632.0062.6348.81
LFT-Hasle-free-ALPS-TM23.4218.8130.3268.2347.8546.5927.2255.2554.5131.2062.2349.14
LFT-3BASIL22.9718.2329.7467.3648.3048.3629.5255.7254.8731.8062.7249.83
LFT-3BASIL-TM22.7318.2929.9468.0149.2248.8629.3556.8352.7133.2063.2150.17
LFT-OATS4:8+64LR28.0619.6932.9067.3049.0748.4827.6556.2052.7130.2062.4549.26
LFT-OATS-TM18.9513.7723.6570.7354.2556.3632.5158.8054.8733.6062.7252.98
LFT-Hasle-free-SparseGPT19.3814.1524.7471.7153.7853.5830.8056.3557.7633.2060.6752.23
LFT-Hasle-free-SparseGPT-TM18.4513.3523.6071.5555.2156.0232.9456.4353.0736.0063.4353.08
LFT-Hasle-free-ALPS18.7713.8223.6371.0655.5955.0130.2956.9953.7933.8062.9152.43
LFT-Hasle-free-ALPS-TM18.1113.2822.7571.9356.9756.9932.0059.4354.5133.8062.6653.54
LFT-3BASIL17.8812.9922.5672.7456.9156.8631.6660.6259.5736.4062.3954.64
LFT-3BASIL-TM17.7512.8222.3072.5256.8156.4432.9459.5151.6235.2063.6153.58
LFT-OATS2:4+64LR23.5517.5229.6067.0848.0449.2827.9055.3354.1531.6062.5749.49
LFT-OATS-TM21.0015.4025.7769.8050.9252.1529.0155.9655.2333.0062.6351.09
LFT-Hasle-free-SparseGPT21.5615.9926.8169.5351.0348.6528.6756.0452.3531.4062.1449.98
LFT-Hasle-free-SparseGPT-TM20.1615.1625.2371.1153.4052.8629.6957.2256.6835.0062.1752.27
LFT-Hasle-free-ALPS20.3815.5725.4771.1652.5353.4129.6956.9954.8733.6062.8451.89
LFT-Hasle-free-ALPS-TM19.4214.5124.3171.1654.0153.1628.9258.0154.1534.2060.4351.75
LFT-3BASIL19.2514.6024.5671.7654.7253.1629.0157.7753.4334.0063.1552.12
LFT-3BASIL-TM19.0714.3723.8871.4455.1254.0429.9558.1753.4333.6062.6952.30
Llama-3.2-1B Dense-14.019.7517.5974.5963.6660.4836.2660.6956.6837.2063.9856.69
+ +Table 10: (N:M Sparse + LR) decomposition performance for Llama-3.2-1B after LoRa Fine-Tuning (LFT). For Perplexity, $(\downarrow)$ lower values are better. For zero-shot tasks, $(\uparrow)$ higher values are better. + +
MethodConfigC4↓WT2↓PTB↓PIQA↑HS↑ARC-E↑ARC-C↑WG↑RTE↑OQA↑BoolQ↑Avg
LFT-OATS2:8+64LR48.5344.5162.2158.2732.8436.9522.7852.9652.7125.6061.4442.94
LFT-OATS-TM100.8787.20120.9856.6429.0130.1320.6550.9952.7126.0062.1141.03
LFT-Hassle-free-SparseGPT35.9133.4446.6461.4837.3336.7023.4651.5452.3526.2062.1743.90
LFT-Hassle-free-SparseGPT-TM31.4630.7742.1262.7940.1938.7623.6354.7853.0729.0058.3245.07
LFT-Hassle-free-ALPS31.4329.5044.3963.3840.6539.7323.4655.7252.7128.4046.0643.76
LFT-Hassle-free-ALPS-TM28.2226.8136.2365.0243.6940.7823.4655.7252.7128.4060.4646.28
LFT-3BASIL30.5828.2741.3063.1141.2541.7524.2355.0952.7128.4063.8846.30
LFT-3BASIL-TM27.4826.1838.3565.3444.3743.1826.5457.4653.4329.2061.0447.57
LFT-OATS3:8+64LR21.9816.4427.2369.5952.2652.7829.4457.2259.2131.0064.0751.95
27.0920.9430.2167.6847.2651.2628.4157.4652.7129.2064.6549.83
LFT-OATS-TM20.0115.7224.8770.4656.2554.5930.7260.2258.1233.2063.6453.40
LFT-Hassle-free-SparseGPT18.7315.1823.2571.0058.4954.2132.5161.7259.2135.4053.0053.19
LFT-Hassle-free-SparseGPT-TM18.8315.5422.9671.6058.2757.1532.5163.2254.8733.8066.6454.76
LFT-Hassle-free-ALPS18.0414.6823.0872.9160.2660.2334.2261.0962.8236.0058.9055.80
LFT-3BASIL18.4014.6122.9072.2059.7057.8332.1762.6755.2335.2068.4755.43
LFT-3BASIL-TM17.6914.3822.4272.9661.2157.3234.0461.8058.4834.8055.8154.55
LFT-OATS4:8+64LR15.5010.7717.9475.6365.4765.2438.6565.9857.4037.8065.6958.98
15.9211.0017.8274.2163.4665.6137.8066.1464.6236.8072.2660.11
LFT-OATS-TM14.9710.6217.4875.2466.8765.8739.2566.2258.1238.6072.1160.29
LFT-Hassle-free-SparseGPT-TM14.4710.2516.9275.6867.7966.3739.7666.5455.9642.8067.8660.35
LFT-Hassle-free-ALPS14.6010.3316.7375.7967.4264.2737.6365.8262.8238.2055.7858.47
LFT-Hassle-free-ALPS-TM14.3010.1116.5075.6368.1866.7541.2167.8859.2140.0069.3961.03
LFT-3BASIL14.3810.1016.5676.7768.0667.2640.7067.4859.5739.6068.3560.97
LFT-3BASIL-TM14.159.8916.2977.2668.4466.2039.5966.6962.8239.8072.2961.64
LFT-OATS2:4+64LR25.1812.0820.2673.9961.5862.8836.0963.4661.7336.6062.6357.37
18.0812.8520.2272.4259.4662.7935.2462.2758.8434.6070.2156.98
LFT-OATS-TM16.3611.7919.4373.9463.7163.5135.0764.0151.6236.8068.9957.21
LFT-Hassle-free-SparseGPT-TM15.6511.3718.4175.1465.6562.8836.3564.3356.3238.6066.8558.27
LFT-Hassle-free-ALPS15.8111.5718.4474.5964.9164.0636.5264.5654.5137.8068.6958.20
LFT-Hassle-free-ALPS-TM15.3711.2217.9074.7066.4564.5638.4866.2255.6038.6066.0958.84
LFT-3BASIL15.5211.2317.8775.0866.4563.0536.6066.0664.9839.8069.9460.25
LFT-3BASIL-TM15.1910.9617.5974.4866.4663.0139.0863.0661.3738.4069.2759.39
Llama-3.2-3B Dense-11.337.8113.5377.4873.6171.6345.9969.8554.5143.0073.3963.68
+ +Table 11: (N:M Sparse + LR) decomposition performance for Llama-3.2-3B after LoRa Fine-Tuning (LFT). For Perplexity, $(\downarrow)$ lower values are better. For zero-shot tasks, $(\uparrow)$ higher values are better. + +
MethodConfigC4↓WT2↓PTB↓PIQA↑HS↑ARC-E↑ARC-C↑WG↑RTE↑OQA↑BoolQ↑Avg
LFT-OATS2:8+64LR37.4631.4949.7162.2436.8838.7224.2351.9352.7127.2062.0244.49
LFT-OATS-TM28.2323.8636.2565.6143.8142.0925.2652.6452.7130.6064.1347.11
LFT-Hassle-free-SparseGPT28.8024.4733.9462.3543.3040.2824.9154.8553.0729.4064.4646.58
LFT-Hassle-free-SparseGPT-TM24.8922.1431.6566.0048.4943.7726.3758.2552.7129.8066.2748.96
LFT-Hassle-free-ALPS25.3121.9731.9966.3848.4243.7326.6259.4353.4329.6047.1646.85
LFT-Hassle-free-ALPS-TM22.8520.2328.0067.9552.1846.8427.9960.5453.4331.2061.4450.20
LFT-3BASIL24.5121.4330.0566.8149.6344.0226.3760.0655.6030.6068.1050.15
LFT-3BASIL-TM22.4520.0029.0068.1252.9745.9226.9660.6953.4332.2070.3451.33
LFT-OATS3:8+64LR17.8712.6520.5972.3661.2057.4535.5863.9353.0735.0070.5856.15
LFT-OATS-TM16.1811.4318.3473.6165.4360.6537.4666.5457.4036.8075.7259.20
LFT-Hassle-free-SparseGPT16.6512.0718.8473.9464.7758.0037.2066.6961.0135.0069.7958.30
LFT-Hassle-free-SparseGPT-TM15.6811.5118.0875.6866.5062.0437.5467.2570.4035.2076.7961.42
LFT-Hassle-free-ALPS15.9211.7717.8974.5966.8660.3536.4368.6765.3437.2074.7160.52
LFT-Hassle-free-ALPS-TM15.2911.4617.7475.9068.3461.6637.0368.9862.4536.2069.3359.99
LFT-3BASIL15.6411.7917.8574.7067.4763.5538.6567.4855.6038.2073.2759.87
LFT-3BASIL-TM15.1111.3817.2775.4168.1563.1739.0868.0362.0937.6077.3461.36
LFT-OATS4:8+64LR13.098.6714.6477.8072.3170.9245.3970.7260.6540.4075.6064.22
LFT-Hassle-free-SparseGPT12.738.5714.0378.1873.4570.4543.8671.3562.4541.4078.8164.99
LFT-Hassle-free-SparseGPT-TM12.388.3213.7178.7874.2974.1245.6570.4866.7940.8079.2466.27
LFT-Hassle-free-ALPS12.559.4414.4576.8271.1971.0444.4572.7756.6840.2078.1363.91
LFT-Hassle-free-ALPS-TM12.318.2913.4378.8474.9273.8648.2971.1968.5940.8079.5167.00
LFT-3BASIL12.398.3313.4378.5174.8571.0045.5672.4561.7343.4077.9265.68
LFT-3BASIL-TM12.178.2513.4179.3374.4073.2348.7271.9061.7341.6078.9066.23
LFT-OATS2:4+64LR14.349.6716.2477.1569.5465.6640.7068.2766.4338.8073.7362.54
LFT-OATS-TM13.469.0914.8177.6971.2569.9944.2869.7758.8440.6078.0763.81
LFT-Hassle-free-SparseGPT13.839.5015.3577.0471.6268.8641.7270.0161.3739.4076.0263.25
LFT-Hassle-free-SparseGPT-TM13.359.1514.8479.0072.5767.8043.6070.0163.9040.4075.6964.12
LFT-Hassle-free-ALPS13.499.2914.6177.4872.4067.0942.9270.5667.1539.8077.1364.32
LFT-Hassle-free-ALPS-TM13.209.1214.3478.2472.5167.3841.8970.8862.4540.4075.6363.67
LFT-3BASIL13.359.2514.5477.3772.9069.1943.5271.9867.1541.2076.2464.94
LFT-3BASIL-TM13.079.0014.1778.7373.1768.1441.8971.9859.5739.2076.4263.64
Meta-Llama-3-8B Dense-9.446.1411.1880.7979.1777.6953.3372.8569.6845.0081.4469.99
+ +Table 12: (N:M Sparse + LR) decomposition performance for Meta-Llama-3-8B after LoRa Fine-Tuning (LFT). For Perplexity, $(\downarrow)$ lower values are better. For zero-shot tasks, $(\uparrow)$ higher values are better. + +
MethodConfigC4↓WT2↓PTB↓PIQA↑HS↑ARC-E↑ARC-C↑WG↑RTE↑OQA↑BoolQ↑Avg
3BASIL+OWL70% + 6425.3120.5928.3371.2755.5654.2530.8064.5653.4333.2072.9154.50
23.2119.5427.0971.7157.9056.6933.5367.5654.1533.4077.8656.60
3BASIL-TM19.5516.4421.6474.1659.7658.0433.2866.3858.1235.2073.1557.26
3BASIL-TM+OWL19.5216.2221.3773.5659.9258.8831.4064.3356.3235.4070.8056.33
3BASIL80% + 6462.8561.0879.4959.5235.0735.4022.2754.2252.7127.0060.9543.39
50.5158.1679.0961.7039.4237.6323.9859.3555.6028.0068.2346.74
3BASIL-TM36.5139.3257.9465.0742.0539.9425.0058.8052.7126.0064.5946.77
3BASIL-TM+OWL36.3238.1956.0564.9641.9241.3725.4358.5652.7128.2063.4647.08
Meta-Llama-3-BB Dense-9.446.1411.1880.7979.1777.6953.3372.8569.6845.0081.4469.99
+ +Table 13: Impact of OWL on 3BASiL for (Unstructured + 64) decompositions of Meta-Llama-3-8B. + +# NeurIPS Paper Checklist + +# 1. Claims + +Question: Do the main claims made in the abstract and introduction accurately reflect the paper's contributions and scope? + +Answer: [Yes] + +Justification: We conclude the introduction with a paragraph that explicitly outlines the paper's contributions and scope. + +Guidelines: + +- The answer NA means that the abstract and introduction do not include the claims made in the paper. +- The abstract and/or introduction should clearly state the claims made, including the contributions made in the paper and important assumptions and limitations. A No or NA answer to this question will not be perceived well by the reviewers. +- The claims made should match theoretical and experimental results, and reflect how much the results can be expected to generalize to other settings. +- It is fine to include aspirational goals as motivation as long as it is clear that these goals are not attained by the paper. + +# 2. Limitations + +Question: Does the paper discuss the limitations of the work performed by the authors? + +Answer: [Yes] + +Justification: We discuss the limitation of our work at the end of Section ?? + +Guidelines: + +- The answer NA means that the paper has no limitation while the answer No means that the paper has limitations, but those are not discussed in the paper. +- The authors are encouraged to create a separate "Limitations" section in their paper. +- The paper should point out any strong assumptions and how robust the results are to violations of these assumptions (e.g., independence assumptions, noiseless settings, model well-specification, asymptotic approximations only holding locally). The authors should reflect on how these assumptions might be violated in practice and what the implications would be. +- The authors should reflect on the scope of the claims made, e.g., if the approach was only tested on a few datasets or with a few runs. In general, empirical results often depend on implicit assumptions, which should be articulated. +- The authors should reflect on the factors that influence the performance of the approach. For example, a facial recognition algorithm may perform poorly when image resolution is low or images are taken in low lighting. Or a speech-to-text system might not be used reliably to provide closed captions for online lectures because it fails to handle technical jargon. +- The authors should discuss the computational efficiency of the proposed algorithms and how they scale with dataset size. +- If applicable, the authors should discuss possible limitations of their approach to address problems of privacy and fairness. +- While the authors might fear that complete honesty about limitations might be used by reviewers as grounds for rejection, a worse outcome might be that reviewers discover limitations that aren't acknowledged in the paper. The authors should use their best judgment and recognize that individual actions in favor of transparency play an important role in developing norms that preserve the integrity of the community. Reviewers will be specifically instructed to not penalize honesty concerning limitations. + +# 3. Theory assumptions and proofs + +Question: For each theoretical result, does the paper provide the full set of assumptions and a complete (and correct) proof? + +Answer: [Yes] + +Justification: We clearly state all assumptions in Theorem 1 and provide a rigorous proof in Appendix A. + +# Guidelines: + +- The answer NA means that the paper does not include theoretical results. +- All the theorems, formulas, and proofs in the paper should be numbered and cross-referenced. +- All assumptions should be clearly stated or referenced in the statement of any theorems. +- The proofs can either appear in the main paper or the supplemental material, but if they appear in the supplemental material, the authors are encouraged to provide a short proof sketch to provide intuition. +- Inversely, any informal proof provided in the core of the paper should be complemented by formal proofs provided in appendix or supplemental material. +- Theorems and Lemmas that the proof relies upon should be properly referenced. + +# 4. Experimental result reproducibility + +Question: Does the paper fully disclose all the information needed to reproduce the main experimental results of the paper to the extent that it affects the main claims and/or conclusions of the paper (regardless of whether the code and data are provided or not)? + +Answer: [Yes] + +Justification: We present a detailed description of the proposed 3-Block ADMM algorithm, including update rules and computational procedures, in Section 2, and describe the Transformer matching procedure in Section 3. Additional implementation details necessary for reproducing our results are provided in Appendix B. + +# Guidelines: + +- The answer NA means that the paper does not include experiments. +- If the paper includes experiments, a No answer to this question will not be perceived well by the reviewers: Making the paper reproducible is important, regardless of whether the code and data are provided or not. +- If the contribution is a dataset and/or model, the authors should describe the steps taken to make their results reproducible or verifiable. +- Depending on the contribution, reproducibility can be accomplished in various ways. For example, if the contribution is a novel architecture, describing the architecture fully might suffice, or if the contribution is a specific model and empirical evaluation, it may be necessary to either make it possible for others to replicate the model with the same dataset, or provide access to the model. In general, releasing code and data is often one good way to accomplish this, but reproducibility can also be provided via detailed instructions for how to replicate the results, access to a hosted model (e.g., in the case of a large language model), releasing of a model checkpoint, or other means that are appropriate to the research performed. +- While NeurIPS does not require releasing code, the conference does require all submissions to provide some reasonable avenue for reproducibility, which may depend on the nature of the contribution. For example +(a) If the contribution is primarily a new algorithm, the paper should make it clear how to reproduce that algorithm. +(b) If the contribution is primarily a new model architecture, the paper should describe the architecture clearly and fully. +(c) If the contribution is a new model (e.g., a large language model), then there should either be a way to access this model for reproducing the results or a way to reproduce the model (e.g., with an open-source dataset or instructions for how to construct the dataset). +(d) We recognize that reproducibility may be tricky in some cases, in which case authors are welcome to describe the particular way they provide for reproducibility. In the case of closed-source models, it may be that access to the model is limited in some way (e.g., to registered users), but it should be possible for other researchers to have some path to reproducing or verifying the results. + +# 5. Open access to data and code + +Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material? + +Answer: [Yes] + +Justification: We will release the codes if the paper is accepted. + +Guidelines: + +- The answer NA means that paper does not include experiments requiring code. +- Please see the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details. +- While we encourage the release of code and data, we understand that this might not be possible, so "No" is an acceptable answer. Papers cannot be rejected simply for not including code, unless this is central to the contribution (e.g., for a new open-source benchmark). +- The instructions should contain the exact command and environment needed to run to reproduce the results. See the NeurIPS code and data submission guidelines (https: //nips.cc/public/guides/CodeSubmissionPolicy) for more details. +- The authors should provide instructions on data access and preparation, including how to access the raw data, preprocessed data, intermediate data, and generated data, etc. +- The authors should provide scripts to reproduce all experimental results for the new proposed method and baselines. If only a subset of experiments are reproducible, they should state which ones are omitted from the script and why. +- At submission time, to preserve anonymity, the authors should release anonymized versions (if applicable). +- Providing as much information as possible in supplemental material (appended to the paper) is recommended, but including URLs to data and code is permitted. + +# 6. Experimental setting/details + +Question: Does the paper specify all the training and test details (e.g., data splits, hyperparameters, how they were chosen, type of optimizer, etc.) necessary to understand the results? + +Answer: [Yes] + +Justification: We provide detailed training and evaluation settings for both our proposed pipeline and the baseline methods in Appendix B. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The experimental setting should be presented in the core of the paper to a level of detail that is necessary to appreciate the results and make sense of them. +- The full details can be provided either with the code, in appendix, or as supplemental material. + +# 7. Experiment statistical significance + +Question: Does the paper report error bars suitably and correctly defined or other appropriate information about the statistical significance of the experiments? + +Answer: [No] + +Justification: While we aimed to provide rigorous evaluation, we were constrained by computational resources and thus could not include statistical significance measures. We have, however, ensured consistent settings and fair comparisons across all baselines. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The authors should answer "Yes" if the results are accompanied by error bars, confidence intervals, or statistical significance tests, at least for the experiments that support the main claims of the paper. + +- The factors of variability that the error bars are capturing should be clearly stated (for example, train/test split, initialization, random drawing of some parameter, or overall run with given experimental conditions). +- The method for calculating the error bars should be explained (closed form formula, call to a library function, bootstrap, etc.) +- The assumptions made should be given (e.g., Normally distributed errors). +- It should be clear whether the error bar is the standard deviation or the standard error of the mean. +It is OK to report 1-sigma error bars, but one should state it. The authors should preferably report a 2-sigma error bar than state that they have a $96\%$ CI, if the hypothesis of Normality of errors is not verified. +- For asymmetric distributions, the authors should be careful not to show in tables or figures symmetric error bars that would yield results that are out of range (e.g. negative error rates). +- If error bars are reported in tables or plots, The authors should explain in the text how they were calculated and reference the corresponding figures or tables in the text. + +# 8. Experiments compute resources + +Question: For each experiment, does the paper provide sufficient information on the computer resources (type of compute workers, memory, time of execution) needed to reproduce the experiments? + +Answer: [Yes] + +Justification: We provide details of the computational resources used for our experiments in Appendix B. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The paper should indicate the type of compute workers CPU or GPU, internal cluster, or cloud provider, including relevant memory and storage. +- The paper should provide the amount of compute required for each of the individual experimental runs as well as estimate the total compute. +- The paper should disclose whether the full research project required more compute than the experiments reported in the paper (e.g., preliminary or failed experiments that didn't make it into the paper). + +# 9. Code of ethics + +Question: Does the research conducted in the paper conform, in every respect, with the NeurIPS Code of Ethics https://neurips.cc/public/EthicsGuidelines? + +Answer: [Yes] + +Justification: We have carefully reviewed the NeurIPS Code of Ethics and confirm that all research presented in this paper adheres to its principles. + +Guidelines: + +The answer NA means that the authors have not reviewed the NeurIPS Code of Ethics. +- If the authors answer No, they should explain the special circumstances that require a deviation from the Code of Ethics. +- The authors should make sure to preserve anonymity (e.g., if there is a special consideration due to laws or regulations in their jurisdiction). + +# 10. Broader impacts + +Question: Does the paper discuss both potential positive societal impacts and negative societal impacts of the work performed? + +Answer: [NA] + +Justification: To the best of our knowledge, our work has no societal impact. + +Guidelines: + +- The answer NA means that there is no societal impact of the work performed. + +- If the authors answer NA or No, they should explain why their work has no societal impact or why the paper does not address societal impact. +- Examples of negative societal impacts include potential malicious or unintended uses (e.g., disinformation, generating fake profiles, surveillance), fairness considerations (e.g., deployment of technologies that could make decisions that unfairly impact specific groups), privacy considerations, and security considerations. +- The conference expects that many papers will be foundational research and not tied to particular applications, let alone deployments. However, if there is a direct path to any negative applications, the authors should point it out. For example, it is legitimate to point out that an improvement in the quality of generative models could be used to generate deepfakes for disinformation. On the other hand, it is not needed to point out that a generic algorithm for optimizing neural networks could enable people to train models that generate Deepfakes faster. +- The authors should consider possible harms that could arise when the technology is being used as intended and functioning correctly, harms that could arise when the technology is being used as intended but gives incorrect results, and harms following from (intentional or unintentional) misuse of the technology. +- If there are negative societal impacts, the authors could also discuss possible mitigation strategies (e.g., gated release of models, providing defenses in addition to attacks, mechanisms for monitoring misuse, mechanisms to monitor how a system learns from feedback over time, improving the efficiency and accessibility of ML). + +# 11. Safeguards + +Question: Does the paper describe safeguards that have been put in place for responsible release of data or models that have a high risk for misuse (e.g., pretrained language models, image generators, or scraped datasets)? + +Answer: [NA] + +Justification: To the best of our knowledge, our work poses no such risks. + +Guidelines: + +- The answer NA means that the paper poses no such risks. +- Released models that have a high risk for misuse or dual-use should be released with necessary safeguards to allow for controlled use of the model, for example by requiring that users adhere to usage guidelines or restrictions to access the model or implementing safety filters. +- Datasets that have been scraped from the Internet could pose safety risks. The authors should describe how they avoided releasing unsafe images. +- We recognize that providing effective safeguards is challenging, and many papers do not require this, but we encourage authors to take this into account and make a best faith effort. + +# 12. Licenses for existing assets + +Question: Are the creators or original owners of assets (e.g., code, data, models), used in the paper, properly credited and are the license and terms of use explicitly mentioned and properly respected? + +Answer: [Yes] + +Justification: At the start of Section 4, we reference all datasets and models involved in our experiments. The sources of the code used are listed in Appendix B. + +Guidelines: + +- The answer NA means that the paper does not use existing assets. +- The authors should cite the original paper that produced the code package or dataset. +- The authors should state which version of the asset is used and, if possible, include a URI. +- The name of the license (e.g., CC-BY 4.0) should be included for each asset. +- For scraped data from a particular source (e.g., website), the copyright and terms of service of that source should be provided. + +- If assets are released, the license, copyright information, and terms of use in the package should be provided. For popular datasets, paperswithcode.com/datasets has curated licenses for some datasets. Their licensing guide can help determine the license of a dataset. +- For existing datasets that are re-packaged, both the original license and the license of the derived asset (if it has changed) should be provided. +- If this information is not available online, the authors are encouraged to reach out to the asset's creators. + +# 13. New assets + +Question: Are new assets introduced in the paper well documented and is the documentation provided alongside the assets? + +Answer: [NA] + +Justification: Our paper does not release new assets. + +Guidelines: + +- The answer NA means that the paper does not release new assets. +- Researchers should communicate the details of the dataset/code/model as part of their submissions via structured templates. This includes details about training, license, limitations, etc. +- The paper should discuss whether and how consent was obtained from people whose asset is used. +- At submission time, remember to anonymize your assets (if applicable). You can either create an anonymized URL or include an anonymized zip file. + +# 14. Crowdsourcing and research with human subjects + +Question: For crowdsourcing experiments and research with human subjects, does the paper include the full text of instructions given to participants and screenshots, if applicable, as well as details about compensation (if any)? + +Answer: [NA] + +Justification: Our paper does not involve crowdsourcing nor research with human subjects. + +Guidelines: + +- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. +- Including this information in the supplemental material is fine, but if the main contribution of the paper involves human subjects, then as much detail as possible should be included in the main paper. +- According to the NeurIPS Code of Ethics, workers involved in data collection, curation, or other labor should be paid at least the minimum wage in the country of the data collector. + +# 15. Institutional review board (IRB) approvals or equivalent for research with human subjects + +Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals (or an equivalent approval/review based on the requirements of your country or institution) were obtained? + +Answer: [NA] + +Justification: Our paper does not involve crowdsourcing nor research with human subjects. + +Guidelines: + +- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. +- Depending on the country in which research is conducted, IRB approval (or equivalent) may be required for any human subjects research. If you obtained IRB approval, you should clearly state this in the paper. + +- We recognize that the procedures for this may vary significantly between institutions and locations, and we expect authors to adhere to the NeurIPS Code of Ethics and the guidelines for their institution. +- For initial submissions, do not include any information that would break anonymity (if applicable), such as the institution conducting the review. + +# 16. Declaration of LLM usage + +Question: Does the paper describe the usage of LLMs if it is an important, original, or non-standard component of the core methods in this research? Note that if the LLM is used only for writing, editing, or formatting purposes and does not impact the core methodology, scientific rigorousness, or originality of the research, declaration is not required. + +Answer: [NA] + +Justification: Although our work focuses on pruning LLMs, the core methods proposed do not involve LLMs as important, original, or non-standard components of the algorithm itself. + +Guidelines: + +- The answer NA means that the core method development in this research does not involve LLMs as any important, original, or non-standard components. +- Please refer to our LLM policy (https://neurips.cc/Conferences/2025/LLM) for what should or should not be described. \ No newline at end of file diff --git a/NeurIPS/2025/3BASiL_ An Algorithmic Framework for Sparse plus Low-Rank Compression of LLMs/images.zip b/NeurIPS/2025/3BASiL_ An Algorithmic Framework for Sparse plus Low-Rank Compression of LLMs/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..77f886abc7f8c13ada8abacbefbad91d0f56f968 --- /dev/null +++ b/NeurIPS/2025/3BASiL_ An Algorithmic Framework for Sparse plus Low-Rank Compression of LLMs/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9cc0b7c4f0fecd97d9a66f808c4f4db15c25a9e86e4daf5aac6cc4dced5b9918 +size 2202544 diff --git a/NeurIPS/2025/3BASiL_ An Algorithmic Framework for Sparse plus Low-Rank Compression of LLMs/layout.json b/NeurIPS/2025/3BASiL_ An Algorithmic Framework for Sparse plus Low-Rank Compression of LLMs/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b69e0a79eded1ed0c7f54eac54712f2c3d9da465 --- /dev/null +++ b/NeurIPS/2025/3BASiL_ An Algorithmic Framework for Sparse plus Low-Rank Compression of LLMs/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7c218beb2b11d2005b1b386d198dbe164f4668d97b291f2a0a9bf33f6d79f40c +size 1021916 diff --git a/NeurIPS/2025/3D Equivariant Visuomotor Policy Learning via Spherical Projection/20f2c39d-2687-48ce-8af6-bc79fc25f92b_content_list.json b/NeurIPS/2025/3D Equivariant Visuomotor Policy Learning via Spherical Projection/20f2c39d-2687-48ce-8af6-bc79fc25f92b_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..35dcceb23ed9e76ab0458ab8ea5d5f909c2d422d --- /dev/null +++ b/NeurIPS/2025/3D Equivariant Visuomotor Policy Learning via Spherical Projection/20f2c39d-2687-48ce-8af6-bc79fc25f92b_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c279e62b70c3b799236b2ebf392ee64cc268466fef9cc36bc772a4bc9203c617 +size 192474 diff --git a/NeurIPS/2025/3D Equivariant Visuomotor Policy Learning via Spherical Projection/20f2c39d-2687-48ce-8af6-bc79fc25f92b_model.json b/NeurIPS/2025/3D Equivariant Visuomotor Policy Learning via Spherical Projection/20f2c39d-2687-48ce-8af6-bc79fc25f92b_model.json new file mode 100644 index 0000000000000000000000000000000000000000..4bede5b6c5d177df5886de9bb719276cc9e56e1e --- /dev/null +++ b/NeurIPS/2025/3D Equivariant Visuomotor Policy Learning via Spherical Projection/20f2c39d-2687-48ce-8af6-bc79fc25f92b_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:103175d9ca927c6984af2439d9bbd1372a547de15a411a89121580ca86c6d98f +size 247750 diff --git a/NeurIPS/2025/3D Equivariant Visuomotor Policy Learning via Spherical Projection/20f2c39d-2687-48ce-8af6-bc79fc25f92b_origin.pdf b/NeurIPS/2025/3D Equivariant Visuomotor Policy Learning via Spherical Projection/20f2c39d-2687-48ce-8af6-bc79fc25f92b_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..a3f2643a8b6449889d032db157a51d5c0e885cfe --- /dev/null +++ b/NeurIPS/2025/3D Equivariant Visuomotor Policy Learning via Spherical Projection/20f2c39d-2687-48ce-8af6-bc79fc25f92b_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fd7ca908b1fd64466657c1f87878e8d5b148f034bf0c76675cc8f4c6b9df824e +size 12577097 diff --git a/NeurIPS/2025/3D Equivariant Visuomotor Policy Learning via Spherical Projection/full.md b/NeurIPS/2025/3D Equivariant Visuomotor Policy Learning via Spherical Projection/full.md new file mode 100644 index 0000000000000000000000000000000000000000..da85d3a0be53b40c00e3706b12984ea6047b60fa --- /dev/null +++ b/NeurIPS/2025/3D Equivariant Visuomotor Policy Learning via Spherical Projection/full.md @@ -0,0 +1,855 @@ +# 3D Equivariant Visuomotor Policy Learning via Spherical Projection + +Boce Hu $^{1}$ Dian Wang $^{2,2}$ David Klee $^{1}$ Heng Tian $^{1}$ Xupeng Zhu $^{1}$ Haojie Huang $^{1}$ Robert Platt $^{†,1}$ Robin Walters $^{†,1}$ + +$^{1}$ Northeastern University $^{2}$ Stanford University $^{\dagger}$ Equal Advising + +https://isp-3d.github.io/ + +# Abstract + +Equivariant models have recently been shown to improve the data efficiency of diffusion policy by a significant margin. However, prior work that explored this direction focused primarily on point cloud inputs generated by multiple cameras fixed in the workspace. This type of point cloud input is not compatible with the now-common setting where the primary input modality is an eye-in-hand RGB camera like a GoPro. This paper closes this gap by incorporating into the diffusion policy model a process that projects features from the 2D RGB camera image onto a sphere. This enables us to reason about symmetries in $\mathrm{SO}(3)$ without explicitly reconstructing a point cloud. We perform extensive experiments in both simulation and the real world that demonstrate that our method consistently outperforms strong baselines in terms of both performance and sample efficiency. Our work, Image-to-Sphere Policy (ISP), is the first SO(3)-equivariant policy learning framework for robotic manipulation that works using only monocular RGB inputs. + +# 1 Introduction + +The eye-in-hand configuration, where the primary perception modality is a camera mounted near the wrist of the robot, is an important setting for robotic policy learning. This setup avoids the need for carefully calibrated external camera systems, is easier to integrate onto mobile robot platforms, and provides fine-grained visual details in the regions where the end-effector interacts with the environment. Moreover, it is used in a growing number of large robot datasets [27, 28, 42, 29, 2, 46]. + +Despite recent advances in equivariant learning [66, 63], there remains a lack of effective network architectures for leveraging equivariant invariant neural networks improve data efficiency and domain symmetries directly into the model [34], performance of diffusion policy [1, 70]. However, perform best with point cloud data captured from RGB data, current equivariant policies are unable to the problem and underperform the point cloud question: can SO(3)-equivariance be achieved + +![](images/d532b8dda8ffe8d770ca0661301ce3b391258aca998c0dd9f3bba494f39a228e.jpg) +Figure 1: We propose the first SO(3)-equivariant policy learning framework based on a single eye-in-hand RGB image, where the predicted action sequence transforms equivariantly under the same group action $g \in \mathrm{SO}(3)$ applied to the whole scene. + +structure in this setting using only RGB input. Equivalent generalization by incorporating prior knowledge of [60, 41], and have recently been shown to enhance the filter, existing equivariant diffusion policy frameworks from multiple depth cameras [58]. When used with a flexible to fully leverage the SO(3) structure present in our version significantly [65]. This naturally raises the question directly from monocular RGB images to support the + +data-efficient visuomotor learning? Such a capability should also have the potential to serve as a modular, plug-and-play component that generalizes seamlessly to richer sensing setups. + +This paper addresses this challenge by introducing a novel diffusion policy framework that incorporates SO(3)-equivariance into eye-in-hand visuomotor learning. Our method first projects features extracted from 2D RGB observations onto a sphere and then rotates the resulting spherical signal to compensate for camera motion. This yields a stable, SO(3)-equivariant representation that is well-suited for downstream equivariant architectures. Unlike prior work that relies on segmented point clouds [58, 70] or calibrated multi-camera systems [65], our approach maintains equivariance throughout the entire policy and supports robust, sample-efficient closed-loop control directly from raw eye-in-hand inputs. To the best of our knowledge, this is the first framework to learn SO(3)-equivariant visuomotor policies from monocular RGB observations in eye-in-hand settings. + +# Our key contributions are summarized as follows: + +- We introduce Image-to-Sphere Policy (ISP), the first SO(3)-equivariant policy learning framework that uses spherical projection from 2D RGB inputs to model 3D symmetries. +- We theoretically prove that our method achieves global SO(3)-equivariance and local SO(2)-invariance, facilitating policy learning. +- We validate our method through extensive experiments, achieving an average success rate improvement of $11.6\%$ over twelve simulation tasks and $42.5\%$ across four real-world tasks. + +# 2 Related Work + +Eye-in-hand Policy Learning Eye-in-hand policy learning [22, 26, 69, 67, 11, 43] has become a flexible and scalable alternative to traditional systems that rely on multiple fixed, externally mounted cameras with precise calibration [65, 25, 58, 71, 5]. By mounting cameras on the robot's wrist, these methods simplify deployment, avoid explicit calibration, and ease demonstration collection [6, 36, 52, 23]. However, the constantly shifting viewpoint introduces challenges like partial observability, which motivates the use of closed-loop policies that can handle local, viewpoint-dependent observations [72, 23, 4, 47, 73]. Recent work has explored transformer-based [59] or diffusion-based [15] architectures for eye-in-hand-control [74, 11], showing promising results across diverse manipulation tasks. Despite this progress, existing approaches often require large-scale demonstration data [35, 37], and often overlook symmetry structures inherently present in observations. Our method addresses this gap by introducing equivariant representations that encode geometric structure for eye-in-hand settings. + +Closed-loop Visuomotor Policy Learning Early approaches to closed-loop visuomotor policy learning relied on reinforcement learning and CNN-based policies to map visual inputs to single-step actions [31, 72, 27]. While effective in simple tasks, these methods were sample-inefficient and struggled to capture multimodal behaviors, as each action was predicted independently without considering temporal context. To address this, subsequent work introduced temporal modeling into behavior cloning frameworks, such as BCRNN [39] and BeT [53], to improve sequential consistency and planning horizons. Building on this direction, recent advances have adopted generative policy models [5, 44, 71, 65], which model multi-step action sequences as a denoising process conditioned on observations. These approaches offer stronger expressiveness and improved multimodal behavior modeling. ISP extends this line of work by further integrating structural inductive biases, which enable more generalizable closed-loop control in complex manipulation settings. + +Equivariance in Robotic Manipulation Equivariant and invariant representations have been shown to improve performance and sample efficiency [13, 54, 10, 56, 33, 18, 14, 32, 7]. Prior work has incorporated equivariant architectures for open-loop pick-and-place tasks [76, 64, 17, 19, 9, 45, 61, 21], showing strong performance with fewer demonstrations. Recently, equivariance has been extended to closed-loop diffusion policies [65, 70, 58]. EquiDiff [65] employs an SO(2)-equivariant architecture to enhance Diffusion Policy [5]. EquiBot [70] adopts an SIM(3)-equivariant structure, and ET-SEED [58] performs trajectory-level SE(3)-equivariant diffusion, both leveraging segmented point cloud inputs to model spatial symmetries, thereby improving policy generalization. However, these approaches typically rely on multi-camera setups with fixed viewpoints or preprocessed 3D inputs. These constraints reduce their practicality in eye-in-hand settings, where the continuously shifting viewpoint and monocular RGB input violate the assumptions of existing equivariant models. To fill this gap, ISP models symmetry in the eye-in-hand RGB setting, preserves SO(3)-equivariance, and can integrate with other frameworks to enhance their effectiveness without additional preprocessing. + +# 3 Background + +# 3.1 Representations of SO(3) + +A group representation encodes symmetry by mapping elements of a group to linear transformations. In this work, we focus on the special orthogonal group $\mathrm{SO}(3)$ of 3D rotations. A representation of $\mathrm{SO}(3)$ is a homomorphism $\rho : \mathrm{SO}(3) \to \mathrm{GL}(V)$ , where $V$ is a finite-dimensional vector space and $\mathrm{GL}(V)$ denotes the group of invertible linear transformations on $V$ . We highlight three commonly used representations in robotics and geometric deep learning: + +- Degree-0 trivial representation $\rho_0$ : Maps every $g \in \mathrm{SO}(3)$ to the identity transformation on $\mathbb{R}$ . This is used for rotation-invariant quantities, such as scalar sensor readings or gripper states. +- Degree-1 standard representation $\rho_{1}$ : Maps $g\in \mathrm{SO}(3)$ to a $3\times 3$ rotation matrix acting on $v\in \mathbb{R}^3$ via $\rho_1(g)v = gv$ , capturing vector features like positions and directions. +- Higher degree irreducible representations $\rho_{\ell}$ : For $\ell \in \mathbb{N}$ , the representation $\rho_{\ell} \colon \mathrm{SO}(3) \to \mathrm{GL}(\mathbb{R}^{2\ell + 1})$ is given by the Wigner $D$ -matrix of degree $\ell$ . It is used to describe higher degree features, such as relative poses, and is often used for latent features in equivariant neural networks. + +# 3.2 Spherical Harmonics and Fourier Coefficients + +Spherical harmonics $Y_{\ell}^{m}:\mathbb{S}^{2}\to \mathbb{R}$ form an orthonormal basis for square-integrable functions on the 2-sphere and realize the irreducible representations of SO(3). Any spherical function $\Phi :\mathbb{S}^2\to \mathbb{R}^d$ can thus be expanded as: + +$$ +\Phi (\theta , \phi) = \sum_ {\ell = 0} ^ {\infty} \sum_ {m = - \ell} ^ {\ell} c _ {\ell} ^ {m} Y _ {\ell} ^ {m} (\theta , \phi), \tag {1} +$$ + +where $c_{\ell}^{m}$ are the corresponding Fourier coefficients. The mapping $\Phi \mapsto \{c_{\ell}^{m}\}$ is known as the Spherical Fourier Transform. Under a rotation $R \in \mathrm{SO}(3)$ , each coefficient vector $c_{\ell} \in \mathbb{R}^{2\ell + 1}$ transforms linearly via the representation $\rho_{\ell}$ : + +$$ +c _ {\ell} ^ {\prime} = \rho_ {\ell} (R) \cdot c _ {\ell}. \tag {2} +$$ + +This enables efficient rotation-equivariant operations on spherical signals in the spectral domain. + +# 3.3 Diffusion Policy + +Diffusion-based policy learning [24, 48] is a class of imitation learning methods that model distributions over action trajectories using denoising diffusion probabilistic models (DDPMs) [15]. These methods iteratively denoise sequences of noisy actions, conditioned on observations, to recover expert-like behavior. Formally, given an observation $\mathcal{O}$ and diffusion timestep $k$ , the policy predicts a noise estimate $\epsilon^k$ from a corrupted action sequence $\mathbf{a}^k = \mathbf{a}^0 + \epsilon^k$ using a denoiser network $\Gamma$ . The model is trained to minimize the denoising objective $\mathcal{L}_{\mathrm{diff}} = \mathbb{E}_{\mathbf{a}^0, k, \epsilon^k} \left[ \| \Gamma(\mathcal{O}, \mathbf{a}^k, k) - \epsilon^k \|^2 \right]$ . At test time, the policy generates actions by iteratively denoising a randomly initialized sequence from Gaussian noise. Recent extensions [65] incorporate symmetry priors by designing the denoiser to be equivariant with respect to a transformation group $G$ . Specifically, for compact groups such as SO(3), the denoiser $\Gamma$ is required to satisfy the equivariance constraint: + +$$ +\Gamma (g \cdot \mathcal {O}, g \cdot \mathbf {a} ^ {k}, k) = g \cdot \Gamma (\mathcal {O}, \mathbf {a} ^ {k}, k) \quad \forall g \in G. \tag {3} +$$ + +This formulation ensures that the denoising process respects the symmetry of the environment. + +# 3.4 Problem formulation + +We study closed-loop robotic visuomotor policy learning through behavior cloning, where a policy is trained to imitate expert demonstrations. Given an observation sequence $\mathcal{O} = \{o_{t - k + 1},\dots,o_t\}$ at timestep $t$ , the learned policy predicts an action chunk $A = \{\mathbf{a}_{t + 1},\dots,\mathbf{a}_{t + n}\}$ , where $k$ and $n$ are the observation length and prediction horizon, respectively. Each observation $o = (I,P)$ consists of an RGB image $I$ from the wrist camera, proprioceptive input $P$ describing the end-effector pose. Prior work has shown that higher performance is achieved when using absolute action representations, i.e., actions expressed in the world frame [5, 65]. Following this, we represent each predicted action + +![](images/78a46b6b3e31cc2afbdf51251dfccb0102be8462fbe58c2af7efcf8d013496ea.jpg) +Figure 2: Overview of Image-to-Sphere Policy (ISP) (a) An SO(3)-equivariant observation encoder extracts features from the RGB input, projects them onto the sphere, and applies an equivariance correction using the gripper orientation $R_{x}$ to account for the camera's dynamic viewpoint (red arrow). The corrected spherical signal $\Phi_{\mathrm{corr}}(x)$ is then processed by spherical convolution layers to extract SO(3) signals. Proprioceptive inputs are embedded via equivariant linear layers. Both image and proprioceptive features are represented as a set of Fourier coefficients $c_{\ell}$ on SO(3) and fused (yellow block). (b) The encoded spherical signals are transformed back to the spatial domain via inverse Fourier transform, sampling finite group elements as the conditioning vector for SO(3)-equivariant denoising. The noisy action sequence is processed in the same way, through equivariant linear layers and projected onto the same group elements. + +$a_{t} \in \mathbb{R}^{10}$ as the absolute end-effector pose including a position in $\mathbb{R}^3$ , an orientation represented as a 6D rotation vector in $\mathbb{R}^6$ (see [75]), and a gripper open state in $\mathbb{R}^1$ . As noted by [65], the absolute action parametrization has a symmetry: 3D transformations of the world frame result in the same 3D transformations to the action. We formalize the equivariance properties of ISP in Section 4.1. + +# 4 Method + +Figure 2 illustrates an overview of our proposed method, which consists of two key components, an SO(3)-equivariant observation encoder (Figure 2a) and an SO(3)-equivariant diffusion module (Figure 2b). The observation encoder uses spherical projection to map image-extracted features onto a hemisphere and applies spherical convolutions to ensure SO(3)-equivariance, producing the conditioning vector for the diffusion process. The diffusion module is designed as an SO(3)-equivariant function of the conditioning vectors and noisy inputs. As a result, the entire policy is end-to-end SO(3)-equivariant. In the following subsections, we first describe our observation encoder, which extracts SO(3)-equivariant features from 2D images, and then our equivariant diffusion module. + +# 4.1 SO(3) Equivariant Observation Encoder + +This section describes how we construct an SO(3)-equivariant observation encoder that maps 2D images and robot proprioception into a 3D feature representation. The observation $x \in X$ consists of two parts, an eye-in-hand RGB image $I$ , that captures visual information, and proprioceptive data, $P \in \mathbb{R}^7$ , including the end-effector's 6D pose (position and orientation) and gripper state. Both these signals need to be represented in a way that encodes equivariance. Representing $P$ is relatively easy. Following [65, 70, 51], we embed end-effector pose in SO(3) using the standard representation and gripper state using the trivial representation (Section 3.1). In contrast, encoding the 2D image input $I$ into SO(3)-equivariant features is harder because changes in the pose of the wrist-mounted camera induce out-of-plane viewpoint variations that are hard to model. We address this by projecting a standard 2D encoding of the image onto the sphere, as described below and first proposed in the context of object pose estimation [30, 16]. This enables us to reason about SO(3) action using its irreducible representations encoded as Wigner D-matrices (Section 3.1). + +![](images/a65de86087ee0970784641e3587d09af575de72e083bcf76065d19fbd9179342.jpg) +Figure 3: Illustration of Equivariance Correction. The left side shows two identical scenes under different global transformations. Since the wrist-mounted camera captures images in its local frame, the resulting images, and thus the projected spherical signals, remain identical across both scenes. By applying the gripper orientation $R$ as an equivariance correction, we align these spherical signals to a common world frame, ensuring their equivariant transformation under global scene rotations. + +Image Encoder Our image encoder is detailed in Figure 2a. First, we encode the input image $I$ from the observation $x$ using a standard SO(2)-equivariant image encoder $\lambda$ . Next, the resulting feature map $\lambda(I)$ is mapped onto the sphere using a learnable orthographic projection (see Appendix A for details). This converts a "flat" image into a spherical signal $\Phi(x): \mathbb{S}^2 \to \mathbb{R}^d$ that is easier to manipulate in SO(3). We represent this spherical signal in the spectral domain as truncated Fourier coefficients calculated using the spherical Fourier transform (Equation 1). + +Equivalence Correction At this point, the image encoding has been projected onto the sphere and represented using spherical harmonics. However, there is a problem. Since global 3D transformations of the world transform the camera and objects equally, the observed image and the projected signal $\Phi(x)$ would be invariant. This introduces a mismatch in that $\Phi(x)$ stays constant while the world and actions rotate, thereby breaking global SO(3)-equivariance. We accommodate this by rotating the spherical signal by an amount corresponding to the SO(3) orientation of the gripper. We call this the equivariance correction factor, and it is illustrated in Figure 3. On the left of Figure 3, we see two scenes that are the same except for an SO(3) rotation of $g$ . The eye-in-hand camera image (of the banana) is the same in both situations, even though the scene is rotated. This results in the identical projected spherical signal. It is only by applying the equivariance correction factor to the two respective signals ( $R_1$ and $R_2$ ) that we recover the camera pose in the spherical signal. This ensures that the spherical signals produced in different camera poses are represented in a consistent global frame. We analyze this approach below. + +Definition 1 (Equivalence Correction). Let $G$ be a group acting on the input space $X$ and output space $Y$ . For a function $f: X \to Y$ , an Equivalence-Correction map is any $\mathcal{C}: X \to G$ satisfying $\mathcal{C}(g \cdot x) f(g \cdot x) = g \mathcal{C}(x) f(x)$ for all $g \in G$ , and $x \in X$ . The corrected function $f_{corr}(x) = \mathcal{C}(x) f(x)$ is therefore $G$ -equivariant. + +Notice that Definition 1 implies $f(x)$ and $f(gx)$ are in the same orbit. Equivalence Correction is similar to a canonicalization map $c: X \to G$ , where $f_{\mathrm{cano}} = c(x)f(c(x)^{-1}x$ transforms the input to a canonical frame, then transforms the output back to the original frame. When $f(x) = f(gx)$ , Equivalence Correction is a special case of canonicalization where $f(c(x)^{-1}x) = f(x)$ is invariant, so it only transforms the output to restore equivariance without altering the input. + +We now show that Definition 1 is satisfied when the correction map is chosen to be the end-effector rotation. Let $x \in X$ denote the robot observation at a given timestep, which includes an eye-in-hand RGB image $I$ and the corresponding camera (end-effector) pose $R_x \in \mathrm{SO}(3)$ in the world frame. Let $\Phi(x)$ denote the spherical signal derived from the image $I$ , expressed in the camera frame, and let $\rho$ be a representation of SO(3) acting on $\Phi(x)$ . + +Proposition 1 (Equivariance Correction via End-Effector Pose). The map $\mathcal{C} \colon (I, R_x) \mapsto R_x$ , which assigns each camera image to its corresponding camera pose $R_x \in \mathrm{SO}(3)$ is an equivariance correction. The corrected signal $\Phi_{corr}(x) = \rho(\mathcal{C}(x))\Phi(x) = \rho(R_x)\Phi(x)$ is in a world-aligned frame. Thus, the mapping $\Phi_{corr}$ is $\mathrm{SO}(3)$ -equivariant: for any global rotation $g \in \mathrm{SO}(3)$ , we have $\Phi_{corr}(g \cdot x) = \rho(g)\Phi_{corr}(x)$ . + +Proof. Let $g \in \mathrm{SO}(3)$ be a global rotation applied simultaneously to the scene and the camera. Since the image is recorded in the camera frame, the spherical signal is unaffected, i.e. $\Phi(g \cdot x) = \Phi(x)$ , while the camera pose updates as $R_x \mapsto R_{gx} = gR_x$ . For the corrected signal, we therefore obtain + +$$ +\Phi_ {\text {c o r r}} (g x) = \rho (R _ {g x}) \Phi (g x) = \rho (g) \rho (R _ {x}) \Phi (x) = \rho (g) \Phi_ {\text {c o r r}} (x), \tag {4} +$$ + +where the second equality follows from the homomorphism property $\rho(gR_x) = \rho(g)\rho(R_x)$ of the representation $\rho$ . Hence the map $\Phi_{\mathrm{corr}}$ is $\mathrm{SO}(3)$ -equivariant. + +A concrete realization of $\rho$ with the spherical-harmonic coefficients and Wigner $D$ -matrices is given in Appendix B, where the proposition reduces to the rotation of coefficient vectors in Eq. 2. + +Camera-rotation invariance Our model also enforces an additional symmetry, rotations of the camera around its optical axis while the object remains stationary. These rotations form an SO(2) subgroup. Such rotations transform both the image and the camera pose, but their effects cancel out in the corrected world-frame signal. We now formalize the invariance of the corrected world-frame signal under such transformations. + +Proposition 2 (Invariance to SO(2) Rotation of the Eye-in-hand Camera). Let $g \in \mathrm{SO}(2)$ be a rotation about the camera's optical axis. Then, under the transformation $(I, R_x) \mapsto (g \cdot I, R_x g^{-1})$ , the corrected signal defined in Proposition 1 remains invariant: $\Phi_{corr}(g \cdot x) = \Phi_{corr}(x)$ . + +Proof. Assume the image encoder $\lambda$ is SO(2)-equivariant, i.e., $\lambda(g \cdot I) = g \cdot \lambda(I)$ for all $g \in \mathrm{SO}(2)$ . Because spherical projection and spherical Fourier transform preserve equivariance, the spherical signal satisfies $\Phi(g \cdot x) = g \cdot \Phi(x)$ . Meanwhile, the camera pose transforms as $R_x \mapsto R_x g^{-1}$ , since applying an SO(2) rotation $g$ in the camera frame corresponds to right-multiplying its world-frame orientation $R_x$ by $g^{-1}$ (i.e., rotating the camera relative to itself). The corrected signal is: + +$$ +\Phi_ {\text {c o r r}} (g \cdot x) = \rho \left(R _ {x} g ^ {- 1}\right) \Phi (g \cdot x) = \rho \left(R _ {x} g ^ {- 1}\right) \rho (g) \Phi (x) = \rho \left(R _ {x}\right) \Phi (x) = \Phi_ {\text {c o r r}} (x). \tag {5} +$$ + +Thus, the corrected signal is invariant under any $\mathrm{SO}(2)$ rotations of the eye-in-hand camera. + +![](images/a6d3739aa33b5cce9fd8888d6b09b3653b2a553346a56ce551f88fdf1a4f6b42.jpg) + +By combining Propositions 1 and 2, we obtain a two-level symmetry in the encoder: the features are globally SO(3)-equivariant and locally SO(2)-invariant to rotations of the camera around its optical axis. These properties are inherently preserved without requiring additional constraints. As shown in Section 5, encoding these properties into the network leads to empirically improved performance. + +# 4.2 SO(3) Equivariant Diffusion + +As described in Section 3.3, we enforce end-to-end SO(3)-equivariance by requiring the denoising network $\Gamma$ to satisfy: $\Gamma(g \cdot \mathcal{O}, g \cdot \mathbf{a}^k, k) = g \cdot \Gamma(\mathcal{O}, \mathbf{a}^k, k)$ for all $g \in \mathrm{SO}(3)$ . To achieve this, we extend the 2D denoising model from EquiDiff [65] to 3D. EquiDiff applies a shared 1D temporal U-Net [49] independently to each group element in $C_n \subset \mathrm{SO}(2)$ . This element-wise weight sharing guarantees that the same parameters act on every group element, resulting in a noise embedding in the regular representation. To generalize to 3D, we approximate the continuous symmetry group SO(3) with a finite subgroup and perform sampling accordingly. Denote $H \subset \mathrm{SO}(3)$ the subgroup that the diffusion process is equivariant to (e.g., the icosahedral group $I_{60}$ ). Denote $S \subset \mathrm{SO}(3)$ a set that is closed under $H$ , i.e., $HS = S$ . Intuitively, $S$ could be viewed as copies of the rotations in $H$ , each with different offset angles to capture a denser discrete signal. Given a signal $\Psi : \mathrm{SO}(3) \to \mathbb{R}^d$ , we first sample $\Psi(S) = \{\Psi(s_i) : s_i \in S\}$ and then evaluate the U-Net pointwise on each sample $\Gamma(\Psi(S)) = \{\Gamma(\Psi(s_i)) : s_i \in S\}$ , where both the input and output can be treated as copies of the regular representations of $H$ . Since $g \in H$ permutes the order of $\Psi(S)$ and $\Gamma(\Psi(S))$ identically, the entire process is $H$ -equivariant. Because the spherical convolution layers output a signal on SO(3), we can flexibly choose any finite group $H$ and sampling set $S$ for discretization. In our implementation, we use both $C_8 \subset \mathrm{SO}(2)$ and $I_{60} \subset \mathrm{SO}(3)$ as choices of $H$ . We refer readers to Appendix C for further details. + +# 4.3 End-to-End Symmetry Analysis + +In this section, we analyze the equivariant properties of our method. First, due to the SO(3)-equivariant encoder (Proposition 1) and the SO(3)-equivariant diffusion model (Section 4.2), our policy has end-to-end symmetry to global scene SO(3) rotations. This significantly improves its sample efficiency and generalizability to world coordinate frame changes. + +![](images/809f4ac1b598c526de0d6d254b0ea21ae1f0439af7cb6592b462fd4dc53097af.jpg) +(1) Square D2 + +![](images/067a11154d559bbe275190d5f8e1511cae7078dd8ae7b731a466e183d64fc11f.jpg) +(2) Hammer Cleanup D1 +Figure 5: A subset of experimental environments from MimicGen. Left: external view of the task. Right: eye-in-hand observation used in the experiments. The full set of tasks is shown in Appendix D. + +![](images/fe5aa2c52c9f201bf72caaf7d30ea8c1a767b2ebf5708f52417b99c0ba87d5d8.jpg) +(3) Coffee Preparation D1 + +![](images/cc3ff41433a69b9c6f5111aa5ac806fab6aa97470967ad77e33331eb8eb67b9f.jpg) +(4) Stack Three D1 + +![](images/ccd3220657afbb16a55fc683f0948dc508349713a330ed1c42b1b8df8c67ea37.jpg) + +The benefit of SO(2) camera-rotation invariance (Proposition 2) is subtle. Under a rotation of the gripper with respect to the workspace, no a priori constraint can be placed on how the action trajectory should transform. However, our diffusion model receives a representation from the observation encoder that is equivariant to this rotation because it is constructed from invariant features (from the spherical signals) and equivariant features (from the end-effector rotation), thus providing a structured geometric bias. Figure 4 illustrates the benefit of this design. In both states (a) and (b), the gripper (triangle) aims to reach the same goal pose (star), but in (b) it + +![](images/8c4cde6b0be3213e495b4d1eec03481814f74d0dcec299eff1f85893cebd2957.jpg) +(a) +Figure 4: Illustration of translation invariance and rotation equivariance-to-invariance transition. + +![](images/7842ea00cf519d9be637a84797d12a8005b47dba9c338808700c33c00283877f.jpg) +(b) + +is rotated by $90^{\circ}$ around its optical axis. Translationally, the action in (b) should remain invariant (red dots), while rotationally, it should gradually transition from equivariant (yellow) to invariant (green) behavior. The equivariant component in the representation ensures that the model can correctly handle the initial $90^{\circ}$ rotation through its symmetry, while the invariant component provides stability and goal alignment. Together, this representation offers a geometric inductive bias for learning such trajectories, whereas non-equivariant models must infer these patterns purely from data. The advantage is empirically validated in Section 5. + +# 5 Experiments + +# 5.1 Simulation + +Experiment Setting We evaluate ISP on twelve robotic manipulation tasks from the MimicGen benchmark [40], which is widely used in previous work on closed-loop policy learning [8, 65]. A representative subset of these simulation tasks is shown in Figure 5 (see Appendix D for a full description of all twelve MimicGen tasks). Policies are trained and evaluated exclusively using eye-in-hand RGB observations (right image in each subfigure of Figure 5). To capture sufficient scene context, we enlarge the camera's field of view (FOV) to approximate a typical fisheye camera setup and re-generate the enlarged FOV observations using the original Mimicgen demonstrations for our method and baselines. For each task, we train three independent models with different random seeds (0, 1, and 2) for each of the 100- and 200-demonstration settings. The models are evaluated 60 times throughout training using 50 fixed rollouts per evaluation. We report the average of the best success rates from the three runs. Task and training details are provided in Appendix D and Appendix E. + +Baselines Our experiments aim to validate the benefits of explicitly modeling equivariance in eye-in-hand visuomotor policies. We evaluate two versions of ISP with different symmetry levels, an SO(3)-equivariant version and an SO(2)-equivariant variant, which is symmetric only about rotations in the plane of the table. Although the SO(3) version has more symmetry, the SO(2) version is more lightweight, which may be preferable in some settings. We compare against three strong baselines: (1) Diffusion Policy [5]: A diffusion-based policy without any equivariance, serving as the primary reference. (2) EquiDiff (modified) [65]: Designed for fixed-camera settings, it achieves SO(2) equivariance via an equivariant image encoder and an equivariant temporal U-Net. For eye-in-hand control, we replace its image encoder with a standard ResNet [12], so only proprioception and denoising remain equivariant. (3) ACT [74]: A transformer-based behavior cloning method. To ensure a fair comparison, all experiments in the following sections, including ablations and method variants, consistently apply SO(2) data augmentation during training by rotating the end-effector pose in both proprioception and actions, equivalent to jointly rotating the gripper and scene. + +Table 1: Success rates (%) on MimicGen tasks with 100 and 200 demonstrations, averaged over 3 seeds. We report both overall mean and per-task performance. The best result is highlighted in bold, and the second best is underlined. Full results with standard deviations are in Appendix F. + +
MethodMeanStack D1Stack Three D1Square D2Threading D0Three Pc. D0Hammer Cl. D1
100200100200100200100200100200100200100200
ISP-SO(3)65.2 (+11.6)75.0 (+10.5)9910070883551909271796673
ISP-SO(2)65.0 (+11.4)73.1 (+8.6)9810075883251858775807173
DiffPo53.664.1919643771225778773735963
EquiDiff53.064.596996180919899274795974
ACT23.040.945771237310365328503563
Mug Cl. D1Coffee D2Kitchen D1Pick Place D0Coffee Prep. D1Nut Asse. D0
Method100200100200100200100200100200100200
ISP-SO(3)545964697579426641617582
ISP-SO(2)566159636572466147567484
DiffPo496153556171364837525162
EquiDiff516247615567284627394056
ACT2537213521519148163749
+ +Results Table 1 reports the maximum success rates across all methods and configurations. In terms of performance, ISP-SO(3) achieves the best results in 21 out of 24 task settings, consistently outperforming the baselines. The remaining three settings show only marginal differences (within $1 - 2\%$ ), all within the standard error margins. Similarly, ISP-SO(2) outperforms baselines in 20 settings, which further validates the effectiveness of our design. With only 100 demonstrations, our model exceeds the best-performing baseline by an average of $11.6\%$ . With 200 demonstrations, the advantage remains similar at $10.5\%$ . Importantly, our model trained with 100 demonstrations surpasses all baselines trained with 200 demonstrations and additional data augmentation, clearly demonstrating superior data efficiency. These results collectively highlight that the explicit modeling of equivariance is the key factor driving both the improved performance and enhanced sample efficiency of our method. Appendix F provides the full experimental results with standard deviations across three random seeds. + +Ablation Study To assess the contribution of each component of our method, we conduct an ablation study on four representative tasks with 100 demonstrations: Stack Three D1, Square D2, Coffee D2, and Nut assembly D0. We evaluate the following variants of ISP-SO(3), each corresponding to a core module in our design: (1) Sphere: With or without the spherical projection and spherical convolutions for + +Table 2: Ablation study results. A red cross indicates that the corresponding module is removed in that variant. + +
SphereEquiEncEquiUSta.Cof.Nut.Squ.Mean
X✓✓63.361.359.023.351.8 (-9.2)
✓X✓66.057.361.332.054.2 (-6.8)
✓✓X68.758.758.032.054.3 (-6.7)
✓✓✓70.064.075.334.761.0
+ +extracting SO(3)-equivariant features from images. (2) EquiEnc: With or without the proposed equivariant image encoder that captures SO(2)-invariant features (Proposition 2). (3) EquiU: With or without an equivariant temporal denoising U-Net in the diffusion module. The results are summarized in Table 2. Removing the spherical projection leads to the largest performance drop of $9.2\%$ , highlighting its critical role in capturing symmetries, despite the use of data augmentation. Disabling the equivariant image encoder and the equivariant U-Net results in drops of $6.8\%$ and $6.7\%$ , respectively. These results demonstrate that all three components: spherical lifting, invariant encoding, and equivariant denoising, are essential for the overall effectiveness of our method. In addition, we further investigate the role of Equivariance Correction (Proposition 1) by comparing delta and absolute control strategies in Appendix G. + +The Benefits of Pretraining While our method already benefits from explicit equivariance, we further explore whether incorporating a pretrained image encoder can provide additional performance gains. Intuitively, pretraining can introduce stronger geometric priors and yield higher-quality visual features, especially beneficial in data-limited regimes. To evaluate this effect, we conduct experiments on the MimicGen using 100 demonstrations and the same evaluation protocol described above. We compare the ISP-SO(2) with two variants: Pretraining, which initializes the image encoder with an + +Table 3: Success rates (%) on MimicGen tasks with 100 demonstrations, comparing pretrained and scratch initialization of the equivariant image encoder. Results are averaged over three seeds. Values in parentheses indicate the performance difference between the two settings. + +
MethodMeanStack D1Stack Three D1Square D2Threading D0Three Pc. D0Hammer Cl. D1
ISP-SO(2) (Pretraining)72.1(+7.1)98.0 (=)81.3 (+6.6)56.0 (+24.0)91.3 (+6.6)76.7 (+2.0)72.7 (+2.0)
ISP-SO(2) (Scratch)65.098.074.732.084.774.770.7
Mug Cl. D1Coffee D2Kitchen D1Pick Place D0Coffee Pre. D1Nut Assembly D0
ISP-SO(2) (Pretraining)54.0 (-2.0)66.7 (+8.0)64.0 (-0.7)56.3 (+10.6)63.3 (+16.6)85.0 (+11.3)
ISP-SO(2) (Scratch)56.058.764.745.746.773.7
+ +![](images/40e71539364b9da5235e012a14c7d5c40f273a64355617951d489c8c647e4679.jpg) +(1) Box-Pipe Disassembly + +![](images/118a895b4a9fcd1cf81cb68f2d31996d59119d3c07e2f872ba6bb07dab772a09.jpg) +(2) U-Pipe Disassembly +Figure 6: Real-world environments for evaluation. A GoPro camera is mounted on the robot's wrist to capture eye-in-hand observations. In each subfigure, the left image shows the initial state, while the right image shows the goal state. See Appendix H for detailed task descriptions. + +![](images/ae514ca178a6ae3bcfb03a026fc1f94ddbf76bfa2a19792da64dad25c25de295.jpg) +(3) 3D-Pipe Disassembly + +![](images/dc17b5f53aaa013d14e419cf02f35d524077bf775292378eb0018840ef3930f5.jpg) +(4) Grocery Bag Retrieval + +ImageNet-1k [50]-pretrained equivariant ResNet-18, and Scratch, which trains the entire model from random initialization. Table 3 reports the maximum evaluation success rates. + +Results show that ISP-SO(2) (Pretraining) surpasses ISP-SO(2) (Scratch) by $7.1\%$ , indicating consistently improved final performance across most tasks. Moreover, the pretrained version with only 100 demonstrations achieves comparable performance to training from scratch with 200 demonstrations, further highlighting its data efficiency. These findings demonstrate the effectiveness of pretraining in providing richer and more stable visuomotor representations. Although the performance gains are marginal or absent in a few tasks, this suggests that naive pretraining may not always align perfectly with the downstream visuomotor learning objective. Developing pretraining strategies that are tailored to equivariant visuomotor policy representations is, therefore, a promising future direction. + +# 5.2 Real World + +Physical Setups Our real robot experiments use a Universal Robot UR5 equipped with a Robotiq-85 Gripper and custom-designed soft fingers. A GoPro camera is mounted on the wrist, following prior setups [6, 11, 35]. Demonstrations are collected via the Gello teleoperation interface [68], with observations and actions recorded at $5\mathrm{Hz}$ . Following [5, 65], we employ DDIM [55] to reduce the number of denoising steps to 16. Figure 6 illustrates the four real-world manipulation tasks. The first three tasks involve pipe disassembly, each focusing on different challenges in closed-loop control: background-object segmentation (Box-Pipe), long-horizon control (U-Pipe), and handling complex 3D geometries (3D-Pipe). The fourth task involves retrieving objects from a deformable grocery bag, for which wrist-mounted camera observations are the only reliable source of visual information due to severe occlusions and limited external visibility. We compare ISP-SO(3) against the Diffusion Policy [5]. Further details on the physical setup, task visualization, goal specification, and practical guidelines for data collection are provided in Appendix H and Appendix I. + +Results Table 4 reports success rates over 20 trials per task. Our method consistently outperforms the Diffusion Policy [5] baseline, with significant improvements on Box-Pipe (80% vs. 10%) and 3D-Pipe (75% vs. 15%). The former benefits from more precise visual representations that distinguish the gray pipe from the gray box background, while the latter showcases the advantage of SO(3)-equivariant features for reasoning over complex 3D geometries. The U-Pipe task also shows a notable gain (85% vs. 65%), demonstrating the sustained and stable performance of our equivariant method in the long-horizon task. On the Grocery Bag task, which heavily relies on eye-in-hand perception, our method achieves a 95% success rate. This shows its high stability and robustness. These results confirm the effectiveness of our equivariant design in addressing diverse manipulation challenges in the real world. See Appendix J for a detailed failure analysis. We further evaluate the computational efficiency of ISP in real-world settings, with a comprehensive discussion provided in Appendix K. In addition, we discuss potential limitations and practical considerations of equivariance in Appendix L. + +![](images/3ffa15f146f6498909f435f1d63ca9d8fd8ba9e690d0217e2b84a1b34cf4c8cc.jpg) +(a) Lighting Change + +![](images/10dd9382d9583c63cc7559eb8be5b8e172a77323e13ea9639a78627e6e3751b3.jpg) +(b) Background Clutter +Figure 7: Real-world perturbation scenarios used to evaluate the robustness and generalization of our method on the Box-Pipe Disassembly task. + +![](images/6d801c48a6fe86e19a4d16cb9cd4f0ded5948be057281ffe4f5b214da0dcbae2.jpg) +(c) Partial Camera Occlusion + +Robustness to Real-World Perturbations To further evaluate the robustness and generalization ability of our policy, we conducted additional real-world experiments on the Box-Pipe Disassembly task under various domain shifts. First, we altered the lighting conditions by introducing a strong white point light + +Table 4: Real-world task performance over 20 trials. The number of demonstrations used for training each task is shown in the second row. + +
Box-PipeU-Pipe3D-PipeGrocery Bag
# Demos65656560
ISP-SO(3)80%(16/20)85%(17/20)75%(15/20)95%(19/20)
DiffPo [5]10%(2/20)65%(13/20)15%(3/20)75%(15/20)
+ +source near the workspace, which substantially changed the shadows and color temperature of the scene (Figure 7a). Second, we perturbed the background by placing multiple household objects on the table to create clutter (Figure 7b). Finally, to test robustness against partial occlusion, we repeatedly and briefly blocked the eye-in-hand camera by rapidly waving different objects (e.g., a toy golf club and a flower) in front of it during policy rollout (Figure 7c). Using the same initial states and 20 rollouts per condition as in the previous real-world experiments, ISP-SO(2) achieves success rates of $85\%$ under lighting changes, $75\%$ with background clutter, and $85\%$ under partial camera occlusion. For reference, the performance of ISP-SO(3) without perturbations is $80\%$ . These results demonstrate that the proposed method generalizes well to real-world disturbances and maintains strong task performance under challenging visual conditions. + +# 6 Conclusion + +In this paper, we propose Image-to-Sphere Policy (ISP), the first SO(3)-equivariant policy learning framework for eye-in-hand visuomotor control using only monocular RGB inputs. By lifting 2D image features onto the sphere and introducing an equivariance correction mechanism to compensate for dynamic camera viewpoints, our method achieves global SO(3)-equivariance and local SO(2)-invariance without relying on depth sensors or multi-camera setups. This design enables robust and sample-efficient policy learning in dynamic, real-world settings. Extensive experiments in both simulation and real-world tasks demonstrate that ISP consistently outperforms strong baselines, achieving higher success rates with fewer demonstrations. Our work provides a general and effective algorithmic solution that is both deployable and scalable for eye-in-hand visuomotor learning. + +**Limitations** Our method has several limitations for future investigation. First, we only consider a single wrist-mounted RGB camera. While this view provides fine-grained local information, it lacks the global scene context that an agent-view camera could offer. Effectively combining these complementary perspectives remains an important challenge. Second, our approach models rotational equivariance but does not address translational equivariance. This limits the model's ability to generalize to object translations within the scene. Extending the equivariance correction to handle camera translations is a promising direction for future work. Third, the use of equivariant networks increases training time. Although inference remains efficient, reducing training overhead through more lightweight architectures would further enhance practicality. Fourth, our current method focuses on single-arm manipulation. Extending the framework to bimanual systems, where coordination between two arms is required, is a natural next step. Finally, our method does not yet leverage vision-language models. Integrating high-level semantic understanding through vision language models could further improve generalization and task understanding in more diverse environments. + +# Acknowledgments + +We would like to thank Rachel Lim and Andrew Cole for their assistance with the real-world experiments as well as all members of the Helping Hands Lab for their valuable discussions and feedback on the manuscript. This work was supported in part by NSF grants 2107256, 2134178, 2314182, 2409351, 2442658, and NASA grant 80NSSC19K1474. + +# References + +[1] Johann Brehmer, Joey Bose, Pim De Haan, and Taco Cohen. EDGI: Equivariant Diffusion for Planning with Embodied Agents. arXiv preprint arXiv:2303.12410, 2023. +[2] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Joseph Dabis, Chelsea Finn, Keerthana Gopalakrishnan, Karol Hausman, Alex Herzog, Jasmine Hsu, et al. Rt-1: Robotics transformer for real-world control at scale. arXiv preprint arXiv:2212.06817, 2022. +[3] Gabriele Cesa, Leon Lang, and Maurice Weiler. A program to build E(N)-equivariant steerable CNNs. In International Conference on Learning Representations, 2022. URL https://openreview.net/forum?id=WE4qe9xlnQw. +[4] Ricson Cheng, Arpit Agarwal, and Katerina Fragkiadaki. Reinforcement learning of active vision for manipulating objects under occlusions. In Conference on robot learning, pages 422-431. PMLR, 2018. +[5] Cheng Chi, Zhenjia Xu, Siyuan Feng, Eric Cousineau, Yilun Du, Benjamin Burchfiel, Russ Tedrake, and Shuran Song. Diffusion policy: Visuomotor policy learning via action diffusion. The International Journal of Robotics Research, page 02783649241273668, 2023. +[6] Cheng Chi, Zhenjia Xu, Chuer Pan, Eric Cousineau, Benjamin Burchfiel, Siyuan Feng, Russ Tedrake, and Shuran Song. Universal manipulation interface: In-the-wild robot teaching without in-the-wild robots. arXiv preprint arXiv:2402.10329, 2024. +[7] Zihao Dong, Alan Papalia, Leonard Jung, Alenna Spiro, Philip R Osteen, Christa S Robison, and Michael Everett. Learning smooth state-dependent traversability from dense point clouds. arXiv preprint arXiv:2506.04362, 2025. +[8] Niklas Funk, Julien Urain, Joao Carvalho, Vignesh Prasad, Georgia Chalvatzaki, and Jan Peters. Actionflow: Equivariant, accurate, and efficient policies with spatially symmetric flow matching. arXiv preprint arXiv:2409.04576, 2024. +[9] Chongkai Gao, Zhengrong Xue, Shuying Deng, Tianhai Liang, Siqi Yang, Lin Shao, and Huazhe Xu. Riemann: Near real-time se (3)-equivariant robot manipulation without point cloud segmentation. arXiv preprint arXiv:2403.19460, 2024. +[10] Jiaqi Guan, Wesley Wei Qian, Xingang Peng, Yufeng Su, Jian Peng, and Jianzhu Ma. 3D Equivariant Diffusion for Target-Aware Molecule Generation and Affinity Prediction. In The Eleventh International Conference on Learning Representations, 2023. +[11] Huy Ha, Yihuai Gao, Zipeng Fu, Jie Tan, and Shuran Song. Umi on legs: Making manipulation policies mobile with manipulation-centric whole-body controllers. arXiv preprint arXiv:2407.10353, 2024. +[12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2016. +[13] Lingshen He, Yuxuan Chen, Zhengyang Shen, Yiming Dong, Yisen Wang, and Zhouchen Lin. Efficient equivariant network. In Advances in Neural Information Processing Systems, 2021. +[14] Lingshen He, Yuxuan Chen, Zhengyang Shen, Yibo Yang, and Zhouchen Lin. Neural epdos: Spatially adaptive equivariant partial differential operator based networks. In The international conference on learning representations, 2022. + +[15] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. +[16] Owen Howell, David Klee, Ondrej Biza, Linfeng Zhao, and Robin Walters. Equivariant single view pose prediction via induced and restriction representations. Advances in Neural Information Processing Systems, 36:47251-47263, 2023. +[17] Boce Hu, Xupeng Zhu, Dian Wang, Zihao Dong, Haojie Huang, Chenghao Wang, Robin Walters, and Robert Platt. Orbitgrasp: $SE(3)$ -equivariant grasp learning. arXiv preprint arXiv:2407.03531, 2024. +[18] Boce Hu, Heng Tian, Dian Wang, Haojie Huang, Xupeng Zhu, Robin Walters, and Robert Platt. Push-grasp policy learning using equivariant models and grasp score optimization. IEEE Robotics and Automation Letters, 10(11):11180-11187, 2025. doi: 10.1109/LRA.2025.3606392. +[19] Haojie Huang, Dian Wang, Robin Walters, and Robert Platt. Equivariant transporter network. arXiv preprint arXiv:2202.09400, 2022. +[20] Haojie Huang, Owen Howell, Dian Wang, Xupeng Zhu, Robin Walters, and Robert Platt. Fourier transporter: Bi-equivariant robotic manipulation in 3d. arXiv preprint arXiv:2401.12046, 2024. +[21] Haojie Huang, Dian Wang, Arsh Tangri, Robin Walters, and Robert Platt. Leveraging symmetries in pick and place. The International Journal of Robotics Research, 43(4):550-571, 2024. +[22] Physical Intelligence, Kevin Black, Noah Brown, James Darpinian, Karan Dhabalia, Danny Driess, Adnan Esmail, Michael Equi, Chelsea Finn, Niccolo Fusai, et al. $\pi_{0.5}$ : A vision-language-action model with open-world generalization. arXiv preprint arXiv:2504.16054, 2025. +[23] Rishabh Jangir, Nicklas Hansen, Sambaran Ghosal, Mohit Jain, and Xiaolong Wang. Look closer: Bridging egocentric and third-person views with transformers for robotic manipulation. IEEE Robotics and Automation Letters, 7(2):3046-3053, 2022. +[24] Michael Janner, Yilun Du, Joshua Tenenbaum, and Sergey Levine. Planning with Diffusion for Flexible Behavior Synthesis. In International Conference on Machine Learning, pages 9902-9915. PMLR, 2022. +[25] Mingxi Jia, Dian Wang, Guanang Su, David Klee, Xupeng Zhu, Robin Walters, and Robert Platt. Seil: Simulation-augmented equivariant imitation learning. arXiv preprint arXiv:2211.00194, 2022. +[26] Yunfan Jiang, Ruohan Zhang, Josiah Wong, Chen Wang, Yanjie Ze, Hang Yin, Cem Gokmen, Shuran Song, Jiajun Wu, and Li Fei-Fei. Behavior robot suite: Streamlining real-world whole-body manipulation for everyday household activities. arXiv preprint arXiv:2503.05652, 2025. +[27] Dmitry Kalashnikov, Alex Irpan, Peter Pastor, Julian Ibarz, Alexander Herzog, Eric Jang, Deirdre Quillen, Ethan Holly, Mrinal Kalakrishnan, Vincent Vanhoucke, et al. Scalable deep reinforcement learning for vision-based robotic manipulation. In Conference on robot learning, pages 651-673. PMLR, 2018. +[28] Dmitry Kalashnikov, Jacob Varley, Yevgen Chebotar, Benjamin Swanson, Rico Jonschkowski, Chelsea Finn, Sergey Levine, and Karol Hausman. Mt-opt: Continuous multi-task robotic reinforcement learning at scale. arXiv preprint arXiv:2104.08212, 2021. +[29] Alexander Khazatsky, Karl Pertsch, Suraj Nair, Ashwin Balakrishna, Sudeep Dasari, Siddharth Karamcheti, Soroush Nasiriany, Mohan Kumar Srirama, Lawrence Yunliang Chen, Kirsty Ellis, et al. Droid: A large-scale in-the-wild robot manipulation dataset. arXiv preprint arXiv:2403.12945, 2024. +[30] David M Klee, Ondrej Biza, Robert Platt, and Robin Walters. Image to sphere: Learning equivariant features for efficient pose prediction. arXiv preprint arXiv:2302.13926, 2023. + +[31] Sergey Levine, Chelsea Finn, Trevor Darrell, and Pieter Abbeel. End-to-end training of deep visuomotor policies. Journal of Machine Learning Research, 17(39):1-40, 2016. +[32] Yikang Li, Yeqing Qiu, Yuxuan Chen, Lingshen He, and Zhouchen Lin. Affine equivariant networks based on differential invariants. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5546-5556, 2024. +[33] Yikang Li, Yeqing Qiu, Yuxuan Chen, and Zhouchen Lin. Affine steerable equivariant layer for canonicalization of neural networks. In The international conference on learning representations, 2025. +[34] Yi-Lun Liao and Tess Smidt. Equformer: Equivariant graph attention transformer for 3d atomistic graphs. In The Eleventh International Conference on Learning Representations, 2023. URL https://openreview.net/forum?id=KwmPfARgOTD. +[35] Fanqi Lin, Yingdong Hu, Pingyue Sheng, Chuan Wen, Jiacheng You, and Yang Gao. Data scaling laws in imitation learning for robotic manipulation. arXiv preprint arXiv:2410.18647, 2024. +[36] Kehui Liu, Chuyue Guan, Zhongjie Jia, Ziniu Wu, Xin Liu, Tianyu Wang, Shuai Liang, Pengan Chen, Pingrui Zhang, Haoming Song, et al. Fastumi: A scalable and hardware-independent universal manipulation interface with dataset. arXiv e-prints, pages arXiv-2409, 2024. +[37] Songming Liu, Lingxuan Wu, Bangguo Li, Hengkai Tan, Huayu Chen, Zhengyi Wang, Ke Xu, Hang Su, and Jun Zhu. Rdt-1b: a diffusion foundation model for bimanual manipulation. arXiv preprint arXiv:2410.07864, 2024. +[38] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. +[39] Ajay Mandlekar, Danfei Xu, Josiah Wong, Soroush Nasiriany, Chen Wang, Rohun Kulkarni, Li Fei-Fei, Silvio Savarese, Yuke Zhu, and Roberto Martín-Martín. What matters in learning from offline human demonstrations for robot manipulation. arXiv preprint arXiv:2108.03298, 2021. +[40] Ajay Mandlekar, Soroush Nasiriany, Bowen Wen, Iretiayo Akinola, Yashraj Narang, Linxi Fan, Yuke Zhu, and Dieter Fox. Mimicgen: A data generation system for scalable robot learning using human demonstrations. arXiv preprint arXiv:2310.17596, 2023. +[41] Arnab Kumar Mondal, Pratheeksha Nair, and Kaleem Siddiqi. Group equivariant deep reinforcement learning. arXiv preprint arXiv:2007.03437, 2020. +[42] Abby O'Neill, Abdul Rehman, Abhiram Maddukuri, Abhishek Gupta, Abhishek Padalkar, Abraham Lee, Acorn Pooley, Agrim Gupta, Ajay Mandlekar, Ajinkya Jain, et al. Open x-embodiment: Robotic learning datasets and rt-x models: Open x-embodiment collaboration 0. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 6892–6903. IEEE, 2024. +[43] Karl Pertsch, Kyle Stachowicz, Brian Ichter, Danny Driess, Suraj Nair, Quan Vuong, Oier Mees, Chelsea Finn, and Sergey Levine. Fast: Efficient action tokenization for vision-language-action models. arXiv preprint arXiv:2501.09747, 2025. +[44] Aaditya Prasad, Kevin Lin, Jimmy Wu, Linqi Zhou, and Jeannette Bohg. Consistency policy: Accelerated visuomotor policies via consistency distillation. arXiv preprint arXiv:2405.07503, 2024. +[45] Yu Qi, Yuanchen Ju, Tianming Wei, Chi Chu, Lawson LS Wong, and Huazhe Xu. Two by two: Learning multi-task pairwise objects assembly for generalizable robot manipulation. CVPR 2025, 2025. +[46] Yu Qi, Haibo Zhao, Ziyu Guo, Siyuan Ma, Ziyan Chen, Yaokun Han, Renrui Zhang, Zitiantao Lin, Shiji Xin, Yijian Huang, et al. Bear: Benchmarking and enhancing multimodal language models for atomic embodied capabilities. arXiv preprint arXiv:2510.08759, 2025. + +[47] Allen Z Ren, Justin Lidard, Lars L Ankile, Anthony Simeonov, Pulkit Agrawal, Anirudha Majumdar, Benjamin Burchfiel, Hongkai Dai, and Max Simchowitz. Diffusion policy policy optimization. arXiv preprint arXiv:2409.00588, 2024. +[48] Moritz Reuss, Maximilian Li, Xiaogang Jia, and Rudolf Lioutikov. Goal conditioned imitation learning using score-based diffusion policies. In Robotics: Science and Systems, 2023. +[49] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention, pages 234-241. Springer, 2015. +[50] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115:211-252, 2015. +[51] Hyunwoo Ryu, Jiwoo Kim, Hyunseok An, Junwoo Chang, Joohwan Seo, Taehan Kim, Yubin Kim, Chaewon Hwang, Jondeun Choi, and Roberto Horowitz. Diffusion-edfs: Bi-equivariant denoising generative modeling on se (3) for visual robotic manipulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18007-18018, 2024. +[52] Mingyo Seo, H. Andy Park, Shenli Yuan, Yuke Zhu, , and Luis Sentis. Legato: Cross-embodiment imitation using a grasping tool. IEEE Robotics and Automation Letters (RA-L), 2025. +[53] Nur Muhammad Shafiullah, Zichen Cui, Ariuntuya Arty Altanzaya, and Lerrel Pinto. Behavior transformers: Cloning $k$ modes with one stone. Advances in neural information processing systems, 35:22955-22968, 2022. +[54] Anthony Simeonov, Yilun Du, Andrea Tagliasacchi, Joshua B Tenenbaum, Alberto Rodriguez, Pulkit Agrawal, and Vincent Sitzmann. Neural descriptor fields: Se (3)-equivariant object representations for manipulation. In 2022 International Conference on Robotics and Automation (ICRA), pages 6394-6400. IEEE, 2022. +[55] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. +[56] Sangli Teng, William Clark, Anthony Bloch, Ram Vasudevan, and Maani Ghaffari. Lie algebraic cost function design for control on lie groups. In 2022 IEEE 61st Conference on Decision and Control (CDC), pages 1867-1874. IEEE, 2022. +[57] Nathaniel Thomas, Tess Smidt, Steven Kearnes, Lusann Yang, Li Li, Kai Kohlhoff, and Patrick Riley. Tensor field networks: Rotation-and translation-equivariant neural networks for 3d point clouds. arXiv preprint arXiv:1802.08219, 2018. +[58] Chenrui Tie, Yue Chen, Ruihai Wu, Boxuan Dong, Zeyi Li, Chongkai Gao, and Hao Dong. Et-seed: Efficient trajectory-level se (3) equivariant diffusion policy. arXiv preprint arXiv:2411.03990, 2024. +[59] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. +[60] Robin Walters, Jinxi Li, and Rose Yu. Trajectory prediction using equivariant continuous convolution. arXiv preprint arXiv:2010.11344, 2020. +[61] Dian Wang, Mingxi Jia, Xupeng Zhu, Robin Walters, and Robert Platt. On-robot learning with equivariant models. arXiv preprint arXiv:2203.04923, 2022. +[62] Dian Wang, Jung Yeon Park, Neel Sortur, Lawson LS Wong, Robin Walters, and Robert Platt. The surprising effectiveness of equivariant models in domains with latent symmetry. arXiv preprint arXiv:2211.09231, 2022. + +[63] Dian Wang, Robin Walters, and Robert Platt. so(2)-equivariant reinforcement learning. arXiv preprint arXiv:2203.04439, 2022. +[64] Dian Wang, Robin Walters, Xupeng Zhu, and Robert Platt. Equivariant $q$ learning in spatial action spaces. In Conference on Robot Learning, pages 1713-1723. PMLR, 2022. +[65] Dian Wang, Stephen Hart, David Surovik, Tarik Kelestemur, Haojie Huang, Haibo Zhao, Mark Yeatman, Jiuguang Wang, Robin Walters, and Robert Platt. Equivariant diffusion policy. arXiv preprint arXiv:2407.01812, 2024. +[66] Maurice Weiler and Gabriele Cesa. General e (2)-equivariant steerable cnns. Advances in neural information processing systems, 32, 2019. +[67] Jimmy Wu, Rika Antonova, Adam Kan, Marion Lepert, Andy Zeng, Shuran Song, Jeannette Bohg, Szymon Rusinkiewicz, and Thomas Funkhouser. Tidybot: Personalized robot assistance with large language models. Autonomous Robots, 47(8):1087-1102, 2023. +[68] Philipp Wu, Yide Shentu, Zhongke Yi, Xingyu Lin, and Pieter Abbeel. Gello: A general, low-cost, and intuitive teleoperation framework for robot manipulators. In 2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 12156-12163. IEEE, 2024. +[69] Xiaomeng Xu, Dominik Bauer, and Shuran Song. Robopanoptes: The all-seeing robot with whole-body dexterity. arXiv preprint arXiv:2501.05420, 2025. +[70] Jingyun Yang, Zi-ang Cao, Congyue Deng, Rika Antonova, Shuran Song, and Jeannette Bohg. Equibot: Sim (3)-equivariant diffusion policy for generalizable and data efficient learning. arXiv preprint arXiv:2407.01479, 2024. +[71] Yanjie Ze, Gu Zhang, Kangning Zhang, Chenyuan Hu, Muhan Wang, and Huazhe Xu. 3d diffusion policy: Generalizable visuomotor policy learning via simple 3d representations. arXiv preprint arXiv:2403.03954, 2024. +[72] Tianhao Zhang, Zoe McCarthy, Owen Jow, Dennis Lee, Xi Chen, Ken Goldberg, and Pieter Abbeel. Deep imitation learning for complex manipulation tasks from virtual reality teleoperation. In 2018 IEEE international conference on robotics and automation (ICRA), pages 5628-5635. IEEE, 2018. +[73] Haibo Zhao, Dian Wang, Yizhe Zhu, Xupeng Zhu, Owen Howell, Linfeng Zhao, Yaoyao Qian, Robin Walters, and Robert Platt. Hierarchical equivariant policy via frame transfer. In Forty-second International Conference on Machine Learning, 2025. URL https://openreview.net/forum?id=nAv5ketrHq. +[74] Tony Z Zhao, Vikash Kumar, Sergey Levine, and Chelsea Finn. Learning fine-grained bimanual manipulation with low-cost hardware. arXiv preprint arXiv:2304.13705, 2023. +[75] Yi Zhou, Connelly Barnes, Jingwan Lu, Jimei Yang, and Hao Li. On the Continuity of Rotation Representations in Neural Networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5745-5753, 2019. +[76] Xupeng Zhu, Dian Wang, Ondrej Biza, Guanang Su, Robin Walters, and Robert Platt. Sample efficient grasp learning using equivariant models. arXiv preprint arXiv:2202.09468, 2022. + +# A Orthographic Projection Details + +The orthographic projection in our method follows the approach of [30], which lifts 2D feature maps onto the unit sphere in the camera frame through a signal remapping operation. Unlike traditional geometric projections that rely on explicit camera calibration or depth information, our projection is entirely learned and does not depend on a predefined 3D center or physical camera parameters. + +In practice, the spherical signal is sampled on a HEALPix grid, which provides an equal-area, hierarchical discretization of the sphere. For each point on the sphere, we apply a learnable weighted aggregation over the entire 2D feature map to compute its corresponding signal value. This design allows the network to flexibly determine how spatial features are mapped onto the sphere, rather than relying on fixed projection kernels. This is particularly useful for wide FOV images, where classical orthographic lifting can introduce distortions near image boundaries. By learning the mapping, the model can implicitly compensate for such distortions. However, this also means that robustness to changes in camera intrinsics (e.g., different FOVs or lens distortions) is not explicitly enforced. A promising future direction is to train the projection module under diverse intrinsic settings to support the model to learn a more general and transferable projection function. + +# B Spectral Realization of the Equivariance Correction + +In this section, we provide a concrete spectral realization of the equivariance correction introduced in Proposition 1, using the spherical-harmonic coefficients and Wigner $D$ -matrices. + +Proof. Let $x$ be the observation with camera pose $R_{x} \in \mathrm{SO}(3)$ and let $c_{\ell}(x) \in \mathbb{R}^{2\ell + 1}$ denotes spherical harmonic coefficients. Under a global rotation $g \in \mathrm{SO}(3)$ applied to both the scene and the camera, the camera pose transforms as $R_{x} \mapsto R_{gx} = gR_{x}$ . Since the signal $\Phi(x)$ is expressed in the local (camera) frame, the spherical coefficients remain unchanged under the global transformation, so $c_{\ell}(gx) = c_{\ell}(x)$ . Applying Equation 2 with the updated camera pose, the corrected coefficients at $qx$ are: + +$$ +c _ {\ell , \operatorname {c o r r}} (g x) = D ^ {\ell} \left(R _ {g x}\right) c _ {\ell} (g x) = D ^ {\ell} \left(g R _ {x}\right) c _ {\ell} (x). \tag {6} +$$ + +Since the Wigner $D$ -matrices $D^{\ell}$ form a group representation of $\mathrm{SO}(3)$ , they satisfy the homomorphism property: $D^{\ell}(gR_{x}) = D^{\ell}(g)D^{\ell}(R_{x})$ . Substituting this, we obtain: + +$$ +c _ {\ell , \operatorname {c o r r}} (g x) = D ^ {\ell} (g) D ^ {\ell} \left(R _ {x}\right) c _ {\ell} (x) \tag {7} +$$ + +Recognizing that $c_{\ell, \mathrm{corr}}(x) = D^{\ell}(R_x)c_{\ell}(x)$ by Proposition 1, we conclude: + +$$ +c _ {\ell , \operatorname {c o r r}} (g x) = D ^ {\ell} (g) c _ {\ell , \operatorname {c o r r}} (x) \tag {8} +$$ + +Thus, the corrected coefficients $c_{\ell,\mathrm{corr}}(x)$ transform equivariantly under the group action $g \in \mathrm{SO}(3)$ . + +This result shows that equivariance correction can be implemented spectrally by left-multiplying the spherical harmonic coefficients with Wigner $D$ -matrices according to the camera orientation. This aligns the signal, originally expressed in the camera frame, to a common world frame for consistent and equivariant downstream processing across varying viewpoints. + +# C Implementation of Our Policy + +Our model consists of an SO(3)-equivariant observation encoder followed by an SO(3)-equivariant diffusion module, both implemented using escnn [3] and e3nn [57]. + +Given an observation $x \in X$ , the SO(2)-equivariant image encoder $\lambda$ first maps the RGB image $I$ into a regular representation, which is then mapped to a trivial representation $\lambda(I) \in \mathbb{R}^{n \times h \times w}$ , where $n, h,$ and $w$ denote the number of channels, height, and width, respectively. These 2D features are lifted to the sphere via orthographic projection, producing a signal $\Phi(x)$ on $\mathbf{S}^2$ . To account for varying viewpoints, we use the gripper orientation $R_x$ as an equivariance correction factor to align the spherical signal into a common reference frame. In our setup, the wrist-mounted camera is rigidly attached to the gripper, so the gripper orientation provides a fixed proxy for the camera pose. This + +approximation is sufficient for aligning the image features with the proprioceptive signals, and any minor misalignments can be further handled by the equivariant convolution layers. The corrected signal is then processed by a sequence of $\mathrm{S}^2\rightarrow \mathrm{SO}(3)$ and $\mathrm{SO}(3)\to \mathrm{SO}(3)$ spherical convolution layers to generate the signal $\Psi (x)$ on $\mathrm{SO}(3)$ . The proprioceptive state is encoded using the irreps $\rho_0$ and $\rho_{1}$ , and passed through $\mathrm{SO}(3)$ -equivariant linear layers to yield Fourier coefficients of the same type as $\Psi (x)$ . These are then concatenated with the image signal to form the global conditioning vector $e_{\mathbf{o}}\in \mathbb{R}^{u\times d_{\mathbf{o}}}$ , where $u$ is the number of channels and $d_{\mathbf{o}}$ is the feature dimension. Similarly, the noisy action chunk $\mathbf{a}^k$ is embedded into $e_{\mathbf{a}}\in \mathbb{R}^{u\times d_{\mathbf{a}}\times n}$ , where $d_{\mathbf{a}}$ denotes the number of action feature channels and $n$ the number of time steps. An inverse FFT is applied to sample both $e_{\mathbf{o}}$ and $e_{\mathbf{a}}$ onto discrete subgroups, either the icosahedral group $I_{60}\subset \mathrm{SO}(3)$ or the cyclic group $C_8\subset \mathrm{SO}(2)$ , producing $e_{\mathbf{o}}\in \mathbb{R}^{p\times d_{\mathbf{o}}}$ , $e_{\mathbf{a}}\in \mathbb{R}^{p\times d_{\mathbf{a}}\times n}$ , where $p = 60$ or 8 is the number of group elements. For each group element $g\in I_{60}$ or $g\in C_8$ , a shared SO(3)- or SO(2)-equivariant 1-D temporal U-Net processes the action sequence $e_{\mathbf{a}}^g$ , conditioned on the observation $e_{\mathbf{o}}^g$ and diffusion step $k$ . This design follows the point-wise equivariant processing strategy proposed in [65], ensuring equivariance across group elements. Finally, an equivariant decoder maps the denoised representation to the noise estimate $\epsilon^k$ . + +# D Simulation Settings + +![](images/64b4d81fb751cebf9c6e84c1e0393efcb428c7d14badcc332065713a056d0f73.jpg) +(1) Stack D1 + +![](images/8612dbc3b5e77f38f2b0d2f20a998a6f6165658e88cc00871b53150ab7f52149.jpg) +(2) Stack Three D1 + +![](images/ec3437ba4da0f63f7d61395d1ab697d7bd1b563726e9ce461a223f467e553753.jpg) +(3) Square D2 + +![](images/012e8dc95dc1bcb928b9bc19f88bdee0e03f1a48ae4abc8af5de5e2ec1b2d03c.jpg) +(4) Threading D0 + +![](images/ff9793526ec90a9d7fea33470667cb100364354ff75093e71c344f2a673e0b9a.jpg) +(5) Three Pc. Assembly D0 + +![](images/0ddd66f228c56fb433e5cd900603b78b7b946cd13841fea796ee0b9acca894f3.jpg) +(6) Hammer Cleanup D1 + +![](images/3834b1fd323ae2b60b20fa1bbef3834da9be730f0bf8f3ff357003fff9cbe152.jpg) +(7) Coffee D2 + +![](images/7fdb5b2add1e86aeffa4cac78e9a85852d55a60cea01663271607ee828cf9526.jpg) +(8) Mug Cleanup D1 + +![](images/91635e433e1deeee19705ddc31714c66aad16f3eef2965ad495ba89a8ee0eab9.jpg) +(9) Kitchen D1 + +![](images/2b7d59b9e15aada5d935dd998f585832277bfa273e3808c1820453fd82fdefac.jpg) +(10) Pick Place D0 + +![](images/20285e7e4f04a2a523c0f866126f57518e0f6d86ad80852d0fd3ca84974e3b9b.jpg) +(11) Coffee Preparation D1 +Figure 8: The twelve simulation tasks from the MimicGen [40] simulator. In each subfigure, the left image shows the task scene, while the right image shows the corresponding eye-in-hand view. + +![](images/86c3e671e1df3073c0fd76b5197c75fbda5b3c79fb33a42d29cb8a13d15cc8ad.jpg) +(12) Nut Assembly D0 + +Figure 8 illustrates the twelve tasks in the MimicGen simulation. In each subfigure, the left image shows the full environment scene from the agent's view, while the right image is the eye-in-hand RGB observation used by the model. Following prior work [40, 65], we set the resolution of the eye-in-hand image to $3 \times 84 \times 84$ and adopt the same maximum episode length. To enable the wrist-mounted camera to capture more contextual information, we increase its FOV from 75 to 130 degrees, similar to that of a typical wide-angle camera. + +# E Training Details + +For the simulation experiments, we follow the hyperparameter settings from prior work [65, 5]. In detail, we use an observation window of two history steps for ISP-SO(3) and one step for ISP-SO(2). In both cases, the denoising network outputs a sequence of 16 action steps, which are used for optimization during training, while only the first 8 steps are executed during evaluation. During training, input images are randomly cropped to a resolution of $76 \times 76$ , while a center crop is applied at evaluation time. We train all models using the AdamW [38] optimizer with Exponential Moving Average, and adopt the DDPM [15] framework with 100 denoising steps for both training and evaluation. For all baselines, we retain their original hyperparameter settings for evaluation and only adjust the number of training steps to ensure consistency across methods. All methods are trained on the same dataset and evaluated using three random seeds. + +For the real-world experiments, we use the same hyperparameters as in the simulation, except that we replace DDPM with DDIM [55] for both training and evaluation, and reduce the number of denoising steps to 16 at evaluation time. However, we find that using a resolution of $76 \times 76$ is insufficient for fine-grained manipulation in the real world, as the extremely wide FOV from the GoPro camera causes each pixel to correspond to a relatively large spatial region in the original setting. To address this, we increase the input resolution to $224 \times 224$ . Specifically, starting from the original $720 \times 720$ RGB image captured using a GoPro with the Max Lens Mod, we apply a center crop of size $480 \times 480$ , followed by resizing to $224 \times 224$ . In addition, we apply standard data augmentations, including random cropping, rotation, and color jitter, to improve the robustness of both our method and the baselines. + +All models are trained on single GPUs using compute clusters and workstations equipped with multiple high-performance consumer-grade GPUs. + +# F Full Simulation Experiment Results with Standard Deviations + +Table 5 presents the same results as Table 1, with standard deviations included. + +Table 5: Maximum success rates (%) on MimicGen tasks with 100 and 200 demonstrations across different methods, averaged over three random seeds. The $\pm$ indicates standard deviation. + +
MethodStack D1Stack Three D1Square D2Threading D0
100200100200100200100200
ISP-SO(3)99.3 ± 1.2100.0 ± 0.070.0 ± 2.088.0 ± 2.034.7 ± 4.251.3 ± 2.390.0 ± 2.092.0 ± 0.0
ISP-SO(2)98.0 ± 2.0100.0 ± 0.074.7 ± 7.688.0 ± 2.032.0 ± 0.050.7 ± 5.084.7 ± 1.287.3 ± 3.1
DiffPo90.7 ± 4.296.0 ± 2.043.3 ± 4.276.7 ± 4.212.0 ± 2.025.3 ± 3.177.3 ± 10.386.7 ± 7.0
EquiDiff96.0 ± 0.098.7 ± 1.261.3 ± 5.080.0 ± 2.08.7 ± 1.219.3 ± 1.288.7 ± 5.892.0 ± 2.0
ACT45.3 ± 7.677.3 ± 2.312.0 ± 2.036.7 ± 9.92.7 ± 1.210.0 ± 2.036.0 ± 6.953.3 ± 6.1
Three Pc. Assembly D0Hammer Cleanup D1Mug Cleanup D1Coffee D2
Method100200100200100200100200
ISP-SO(3)70.7 ± 1.279.3 ± 1.266.0 ± 0.073.3 ± 1.254.0 ± 8.758.7 ± 2.364.0 ± 0.068.7 ± 3.1
ISP-SO(2)74.7 ± 1.280.0 ± 2.070.7 ± 1.273.3 ± 2.356.0 ± 2.060.7 ± 1.258.7 ± 3.163.3 ± 4.2
DiffPo72.7 ± 3.173.3 ± 2.358.7 ± 7.663.3 ± 11.749.3 ± 8.361.0 ± 1.753.3 ± 3.154.7 ± 4.2
EquiDiff74.0 ± 5.378.7 ± 1.259.3 ± 4.274.0 ± 2.050.7 ± 2.362.0 ± 0.047.3 ± 3.161.3 ± 2.3
ACT28.0 ± 4.050.0 ± 5.334.7 ± 2.362.7 ± 5.824.7 ± 3.137.3 ± 5.820.7 ± 3.134.7 ± 2.3
Kitchen D1Pick Place D0Coffee Preparation D1Nut Assembly D0
Method100200100200100200100200
ISP-SO(3)75.3 ± 3.179.3 ± 4.242.0 ± 4.465.7 ± 5.540.7 ± 2.361.3 ± 2.375.3 ± 2.582.0 ± 7.5
ISP-SO(2)64.7 ± 2.372.0 ± 2.045.7 ± 8.061.0 ± 5.646.7 ± 4.656.0 ± 0.073.7 ± 7.684.3 ± 1.5
DiffPo60.7 ± 8.170.7 ± 3.136.3 ± 2.147.7 ± 1.537.3 ± 1.252.0 ± 5.351.3 ± 3.862.3 ± 1.5
EquiDiff55.3 ± 1.266.7 ± 2.327.7 ± 2.946.3 ± 3.527.3 ± 1.238.7 ± 2.340.0 ± 4.056.3 ± 3.1
ACT21.3 ± 1.250.7 ± 3.18.7 ± 1.513.7 ± 2.57.3 ± 2.316.0 ± 2.036.7 ± 1.249.0 ± 2.0
+ +# G Invariance via Delta Control vs. Equivariance via Rotation Correction + +One of the core components of our method is the rotation correction step, which aligns the spherical signals to a common reference frame to preserve SO(3)-equivariance throughout the policy pipeline. A natural alternative is to remove this step and instead express actions in the moving gripper frame, referred to as delta actions in [6], which can also be interpreted as a sequence of incremental transforms. This formulation leads to an SE(3)-invariant system, as both perception and action are expressed relative to the gripper's local frame. This raises an important question: Is rotation correction necessary if delta actions can achieve similar symmetry properties through invariance? + +Empirical Evidence To investigate this, we conducted additional experiments comparing absolute and delta action control on two MimicGen tasks: Square D2 and Nut Assembly D0. Specifically, we evaluated (a) a variant of our method without rotation correction that uses delta control, and (b) the original Diffusion Policy with delta control. Table 6 summarizes the results with 100 demonstrations. + +Table 6: Comparison of absolute and delta control on two MimicGen tasks with 100 demonstrations. Values in parentheses indicate performance differences relative to ISP-SO(2) (Absolute). + +
MethodSquare D2Nut Assembly D0
ISP-SO(2) (Absolute)3274
ISP-SO(2) (Delta, No Rotation Correction)22 (-10)57 (-17)
DiffPo (Absolute)12 (-20)51 (-23)
DiffPo (Delta)14 (-18)22 (-52)
+ +![](images/e9dd7645402d0e3b917aca8abe44b768492173ef416f2e0fb76c07d82d0e2cb2.jpg) +Figure 9: Real-world experimental setup. We use a UR5 robot equipped with a Robotiq-85 gripper and custom-designed soft fingers. A GoPro camera is mounted on the wrist to capture visual observations. Demonstrations are collected using the Gello teleoperation interface (bottom right). + +We observe that absolute action consistently outperforms delta action. Similar trends have been reported in prior work, such as EquiDiff [65] and the Diffusion Policy [5], where delta or velocity control often results in inferior performance. + +Generalization Perspective Although gripper-relative control guarantees invariance under single-camera setups, it does not generalize seamlessly to multi-camera or hybrid sensing configurations, where additional viewpoints can break this invariance assumption. In contrast, aligning both observations and actions to a shared world frame establishes a consistent global reference across all sensors. This property supports more flexible sensor integration and improved generalization in complex environments with multiple or moving viewpoints. + +# H Details of the Real-World Experiment + +Figure 9 shows our real-world experimental setup. Demonstrations are collected using the Gello teleoperation interface [68]. While the robot is teleoperated in joint space, we record end-effector actions, including position, rotation, and gripper state. Visual observations and actions are recorded synchronously at each timestep. + +Figure 10 illustrates the initial state distributions for each task. In Box-Pipe Disassembly, two pipes with different colors are connected to a junction box, where one pipe shares the same color as the box may confuse the policy. The orientations of the pipes are randomized. In U-Pipe Disassembly, four pipe fittings are arranged in a U-shape and initialized with random rotations. In 3D-Pipe Disassembly, two pipes are connected with independently randomized 3D orientations. In Grocery Bag Retrieval, a toy banana is randomly placed inside a deformable plastic bag. The robot must reach into the bag, identify and retrieve the banana, and place it into a transparent container with minor positional + +![](images/7ed1040f9ed0589a5f5dff2305493ec3c4ec0addb83c7b9158ffe14aa5e40fce.jpg) +(a) Box-Pipe Disassembly + +![](images/974ea6d7f76a333520255e60d1ea33adf0b0fa991130fe8366725b931d0efe47.jpg) +(b) U-Pipe Disassembly + +![](images/536ae6df021a43209105f71353f2cc1153120137ace441f06907b8731b9c342a.jpg) +(c) 3D-Pipe Disassembly +Figure 10: Distribution of random initial states used in the real-world experiments. + +![](images/bdc49154f8d9242d6256c33bff08108b31eee717c5e49d48d105a188ba35da04.jpg) +(d) Grocery Bag Retrieval + +variation. All subfigures in Figure 10 show averaged visualizations across multiple randomized initializations. + +We visualize one episode for each task in Figure 11. These tasks emphasize different aspects. The pipe disassembly tasks require precise, closed-loop control to smoothly extract the pipes. This makes them particularly challenging for open-loop policies. The Grocery Bag Retrieval task highlights the importance of the eye-in-hand camera, as the target object is difficult to perceive and localize using only external views. + +# I Practical Guidelines for Data Collection + +Before starting data collection on the real robot, it is critical to establish a predefined task execution strategy to ensure motion simplicity, efficiency, and cross-operator consistency. Such a strategy typically involves defining consistent action sequences, execution ranges, and task progression patterns, helping to avoid ambiguous or poorly structured scenarios that may lead to robotic indecision or an undesirably large amount of multimodal behavior during training. + +Based on our experience, during data collection, demonstrations should: + +1. Uniformly cover as many task-relevant initial states as possible. +2. Maintain a consistent end-effector speed within and between trajectories without interruption. +3. Avoid unnecessary stops, pauses, or other irregular motion patterns. +4. Synchronize sensing and control to minimize latency-induced artifacts. +5. Regularly verify alignment between the robot and sensors to prevent drift and maintain data consistency. + +After data collection, all trajectories should be automatically or visually inspected to detect potential issues. In particular, segments exhibiting robotic hesitation or stalling, most commonly near the beginning and end of each demonstration, as well as episodes containing negative or low-quality behavior, should be identified and removed. Consistent inspection and pruning of low-quality data can significantly improve the stability and performance of policy learning. + +These practical steps help ensure that the collected data are clean, diverse, and informative, which can ultimately enhance the robustness and generalization of learned visuomotor policies. + +# J Real World Experiment Failure Analysis + +In the Box-Pipe Disassembly task, one of the primary failure cases arises from the inability to distinguish between the gray junction box and the gray pipe. For the original Diffusion Policy, the policy consistently misidentifies the box as the pipe to be disassembled, which triggers the robot's emergency stop. While our method occasionally encounters the same issue, the failure rate is significantly lower. This suggests that our method is more data-efficient and better at learning robust visual distinctions from limited demonstrations. + +Table 7: Comparison of training and inference efficiency on a single RTX 4090 GPU. + +
MethodTraining Speed (rel.)Inference Time (ms)Real-Time Capable
DiffPo1×68Yes
ISP-SO(2)2.6× slower63Yes
ISP-SO(3)5.4× slower75Yes
+ +In the U-Pipe Disassembly task, a common failure mode for our method and the baseline occurs when pulling the red pipe inadvertently causes a connected pipe to be extracted as well. In such cases, the robot grasps both pipes simultaneously. We consider this a partial success. However, the baseline additionally suffers from incorrect orientation predictions under certain initial states, leading to more frequent failures. + +In the 3D-Pipe Disassembly task, our method occasionally fails to identify the correct grasp orientation. In contrast, the baseline struggles consistently with this issue and rarely completes the task successfully. One major contributing factor is the multimodality of the task. During data collection, it is difficult to maintain consistency in demonstration strategies because pipes can be grasped in multiple orientations. Nevertheless, by incorporating 3D symmetries, our method is more robust to such variations and generalizes better across diverse configurations. + +In the Grocery Bag Retrieval task, failure cases primarily result from unsuccessful grasp attempts or inaccuracies during the placement phase. The deformable nature of the bag and the partial occlusion of the banana present additional challenges, especially under limited visual feedback. + +# K Computational Efficiency Analysis + +In this section, we provide quantitative comparisons of the computational efficiency of our method during both training and inference. Our results are measured on a single RTX 4090 GPU. + +Training Efficiency Compared to the original Diffusion Policy [5], ISP-SO(2) is approximately $2.6 \times$ slower, and ISP-SO(3) is approximately $5.4 \times$ slower during training. This increase is expected due to the added computational complexity of the equivariant layers. Nevertheless, the training speed remains practical for large-scale policy learning. + +Inference Efficiency Despite the higher training cost, our method maintains high efficiency during inference, making it well-suited for real-time deployment. Table 7 summarizes the average inference time of each method in the real-world settings with 16-step DDIM sampling [55]. All methods exhibit comparable inference speeds. The SO(2) variant is slightly faster than the baseline, primarily due to its lighter-weight diffusion U-Net and the use of a smaller history observation window. Although the SO(3) variant is marginally slower, its inference time ( $\sim$ 75 ms) remains close to DiffPo ( $\sim$ 68 ms), well within real-time control requirements (e.g., 10 Hz). + +# L On Limitations and Practical Considerations of Equivariance + +Interaction with Real-World Asymmetries Equivariance may face challenges in manipulation scenarios where asymmetries in the physical world are important. A representative example is tasks involving asymmetric robot kinematics, such as left-right differences in reachable workspace. Although equivariance allows the model to generalize across rotated scenes, joint limits are not preserved under rotation, which may lead to infeasible or suboptimal actions. Another example is manipulation involving heavy objects, where gravity breaks rotational symmetry in practice. An object that is easy to manipulate in one orientation may become unstable or infeasible to lift when rotated. Despite these challenges, prior work has shown that equivariant models can remain robust in the presence of symmetry-breaking factors such as visual appearance, camera pose, and shadows [62]. These asymmetries are already encoded in the input, allowing the model to learn appropriate behaviors without violating the equivariant structure. While cases where equivariance leads to performance degradation are relatively uncommon, they do highlight scenarios where symmetry-breaking mechanisms may be beneficial. A promising future direction is to augment + +equivariant architectures with non-equivariant components or task-specific inductive biases (e.g., gravity-aware priors or joint-limit encodings) to better capture real-world asymmetries. + +Discretization of Continuous Symmetry Groups In our framework, equivariance is enforced by sampling finite subgroups of $\mathrm{SO}(3)$ (e.g., $I_{60}$ or $C_8$ ) and applying a shared U-Net across the sampled group elements. This approximation may, in principle, introduce discrepancies for rotations outside the sampled set. Empirically, however, no significant performance degradation was observed. In fact, discrete subgroups often lead to superior performance compared to continuous irreducible representations, consistent with prior findings in equivariant learning [3, 66]. While this approach reduces the theoretical degree of symmetry, it provides a more scalable and expressive modeling strategy by avoiding the computational overhead and activation function constraints associated with continuous irreps. Similar strategies have also demonstrated strong empirical effectiveness in other robotics applications [65, 20]. To further mitigate potential limitations and improve generalization, we apply random $\mathrm{SO}(2)$ rotations as a data augmentation strategy to the end-effector pose in both proprioceptive inputs and actions during training. Additionally, subgroup sampling is not restricted to a single set: multiple sets can be employed in practice to increase angular coverage when necessary. Finally, while equivariant architectures introduce structural inductive biases, they do not inherently limit the model's ability to generalize beyond the sampled rotations. With sufficient data diversity and augmentation, the network is able to interpolate smoothly across $\mathrm{SO}(3)$ , thereby alleviating the potential impact of discretization. + +# M Broader Impact + +This work has several potential social impacts, both positive and negative. On the positive side, our proposed method enables more data-efficient and generalizable robot policy learning in 3D environments. This can facilitate the development of more robust and capable household robots, particularly in settings where labeled demonstrations are limited. Moreover, by leveraging geometric symmetries and closed-loop visuomotor control from wrist-mounted cameras, our method could lower the barrier for deploying autonomous robots in unstructured real-world environments, thereby expanding accessibility and utility. + +However, as with many data-driven learning methods, our approach inherits limitations tied to the quality and intent of the training data. Since the robot policy is learned entirely through imitation, any unsafe, biased, or suboptimal behavior demonstrated during data collection may be reflected in the final policy. Furthermore, the increased autonomy enabled by our method underscores the importance of safety monitoring and responsible deployment, especially in applications involving human interaction. + +![](images/99a25d578c16fa220c645728651ef0de01cbfa3027bd1680875c37a153935b66.jpg) + +![](images/88722c255783f31751310fb1afa2eb47da2a2814144df85d218a7f882506d1f4.jpg) + +![](images/142155492890877bdd0cf115eb2a458a0afddbc838cde42c670f39d5c44c37e4.jpg) + +![](images/353fc3604010ec1eb97bd95d15e051b7441928981f39267a285c99c9f3f20d35.jpg) + +![](images/3602f5779327639734838b4af82ee561f49fdae52ada193e6588f254c1e66ac5.jpg) +(a) Box-Pipe Disassembly + +![](images/36de5eea3ecfb861df794af59b51421249e63227851af90be318714fe4c4e115.jpg) + +![](images/700dde744c3b4e15bf145e80ca86f89f32779767c7b1c9bf336d3a422ed6f894.jpg) + +![](images/6e970eb03253a11747f46759fe809cc84e94e728eb62bd0091ae2820eee4ec9f.jpg) + +![](images/3f1fd32303ec1d48f204a834b64eada91083c922c5d3e4a2fc721d8cc734cfa6.jpg) + +![](images/569235219fc909b9fcefe171c499b0dd8723359e0bfafb3aa2c5ab5cfe0e8e51.jpg) + +![](images/bf3f444970e508702f7c5d42ec4d7c2628ee816bb6e2c4c6372d9e6b9fedde3c.jpg) +(b) U-pipe Disassembly + +![](images/90ee98da5daf9a44745696d3dfd280af076ba815b67f3a132b8af06dd987d425.jpg) + +![](images/dc27d6853ff2ce3a4887060a2dbd30d5da2178b7db5f60be0228a707d23cb488.jpg) + +![](images/f199f9b2182a46a7a4188c684b7cc1cdecf2238885d5da692ed24b4829c0e44c.jpg) + +![](images/c1b8b42ac9f54f0568c357d318b93d03869aa51fe65ea9b094747d8ee78a3fa5.jpg) + +![](images/b6b843f0a870a8bd69f0d85ad0022be83adf7f710512ab1be8024a71fe771861.jpg) + +![](images/609f535c5cf8d5eb3d556b71dab395415246c903c83e2b7be37901e3d9000e30.jpg) + +![](images/2ed0dd3b5b8dd3e24336d05b5c1359a80f5d606bcbdab33e175efef9f4755a1e.jpg) + +![](images/ef94feeba8404a5fc8349dd3d9418ec24a33b8abd9a2a57b72ac8d8a0da687fa.jpg) + +![](images/b69303d8819b92995dfcbd74a693e83912db43392e1be36cf575988361208b7f.jpg) +(c) 3D-Pipe Disassembly + +![](images/ebcb7d4ce81567ba75c131db85e3941036ba1bbd25a900b45732b772fd876074.jpg) + +![](images/44621cf3dc01679e66068cdbd37c8b352d272da918a6df9bf700e2d4e20e26f5.jpg) +Figure 11: Visualization of one episode for each task. Each subfigure illustrates a key action step in the trajectory. + +![](images/90c79c94a77a7937428b63b111b360c387bfc62a5d9186ce1e0f3a1a055e85f1.jpg) +(d) Grocery Bag Retrieval + +![](images/a36b5caf426b7e44f25c528b6cd7da94e2a54797f0ea1f8ecce901b46316f629.jpg) + +# NeurIPS Paper Checklist + +The checklist is designed to encourage best practices for responsible machine learning research, addressing issues of reproducibility, transparency, research ethics, and societal impact. Do not remove the checklist: The papers not including the checklist will be desk rejected. The checklist should follow the references and follow the (optional) supplemental material. The checklist does NOT count towards the page limit. + +Please read the checklist guidelines carefully for information on how to answer these questions. For each question in the checklist: + +- You should answer [Yes], [No], or [NA]. +- [NA] means either that the question is Not Applicable for that particular paper or the relevant information is Not Available. +- Please provide a short (1–2 sentence) justification right after your answer (even for NA). + +The checklist answers are an integral part of your paper submission. They are visible to the reviewers, area chairs, senior area chairs, and ethics reviewers. You will be asked to also include it (after eventual revisions) with the final version of your paper, and its final version will be published with the paper. + +The reviewers of your paper will be asked to use the checklist as one of the factors in their evaluation. While "[Yes]" is generally preferable to "[No]", it is perfectly acceptable to answer "[No]" provided a proper justification is given (e.g., "error bars are not reported because it would be too computationally expensive" or "we were unable to find the license for the dataset we used"). In general, answering "[No]" or "[NA]" is not grounds for rejection. While the questions are phrased in a binary way, we acknowledge that the true answer is often more nuanced, so please just use your best judgment and write a justification to elaborate. All supporting evidence can appear either in the main paper or the supplemental material, provided in appendix. If you answer [Yes] to a question, in the justification please point to the section(s) where related material for the question can be found. + +IMPORTANT, please: + +- Delete this instruction block, but keep the section heading "NeurIPS Paper Checklist", +- Keep the checklist subsection headings, questions/answers and guidelines below. +- Do not modify the questions and only use the provided macros for your answers. + +# 1. Claims + +Question: Do the main claims made in the abstract and introduction accurately reflect the paper's contributions and scope? + +Answer: [Yes] + +Justification: The abstract and introduction clearly state the core claims of the paper, including our key contributions and the assumptions. These claims are well-supported by both the theoretical analysis and experimental results presented. + +Guidelines: + +- The answer NA means that the abstract and introduction do not include the claims made in the paper. +- The abstract and/or introduction should clearly state the claims made, including the contributions made in the paper and important assumptions and limitations. A No or NA answer to this question will not be perceived well by the reviewers. +- The claims made should match theoretical and experimental results, and reflect how much the results can be expected to generalize to other settings. +- It is fine to include aspirational goals as motivation as long as it is clear that these goals are not attained by the paper. + +# 2. Limitations + +Question: Does the paper discuss the limitations of the work performed by the authors? + +Answer: [Yes] + +Justification: The paper includes a discussion of the limitations of our method. Specifically, we clarify the computational efficiency issues introduced by using equivariant networks, and discuss limitations related to the task settings (e.g., single-arm manipulation and single-view observations). We also outline potential directions for future research to address these limitations and extend the applicability of our approach. + +# Guidelines: + +- The answer NA means that the paper has no limitation while the answer No means that the paper has limitations, but those are not discussed in the paper. +- The authors are encouraged to create a separate "Limitations" section in their paper. +- The paper should point out any strong assumptions and how robust the results are to violations of these assumptions (e.g., independence assumptions, noiseless settings, model well-specification, asymptotic approximations only holding locally). The authors should reflect on how these assumptions might be violated in practice and what the implications would be. +- The authors should reflect on the scope of the claims made, e.g., if the approach was only tested on a few datasets or with a few runs. In general, empirical results often depend on implicit assumptions, which should be articulated. +- The authors should reflect on the factors that influence the performance of the approach. For example, a facial recognition algorithm may perform poorly when image resolution is low or images are taken in low lighting. Or a speech-to-text system might not be used reliably to provide closed captions for online lectures because it fails to handle technical jargon. +- The authors should discuss the computational efficiency of the proposed algorithms and how they scale with dataset size. +- If applicable, the authors should discuss possible limitations of their approach to address problems of privacy and fairness. +- While the authors might fear that complete honesty about limitations might be used by reviewers as grounds for rejection, a worse outcome might be that reviewers discover limitations that aren't acknowledged in the paper. The authors should use their best judgment and recognize that individual actions in favor of transparency play an important role in developing norms that preserve the integrity of the community. Reviewers will be specifically instructed to not penalize honesty concerning limitations. + +# 3. Theory assumptions and proofs + +Question: For each theoretical result, does the paper provide the full set of assumptions and a complete (and correct) proof? + +Answer: [Yes] + +Justification: All definitions, propositions, and assumptions are clearly stated, numbered, and cross-referenced. + +# Guidelines: + +- The answer NA means that the paper does not include theoretical results. +- All the theorems, formulas, and proofs in the paper should be numbered and cross-referenced. +- All assumptions should be clearly stated or referenced in the statement of any theorems. +- The proofs can either appear in the main paper or the supplemental material, but if they appear in the supplemental material, the authors are encouraged to provide a short proof sketch to provide intuition. +- Inversely, any informal proof provided in the core of the paper should be complemented by formal proofs provided in appendix or supplemental material. +- Theorems and Lemmas that the proof relies upon should be properly referenced. + +# 4. Experimental result reproducibility + +Question: Does the paper fully disclose all the information needed to reproduce the main experimental results of the paper to the extent that it affects the main claims and/or conclusions of the paper (regardless of whether the code and data are provided or not)? + +# Answer: [Yes] + +Justification: We will submit the code as supplementary material and release the code implementation for all experiments in this paper. Detailed descriptions of our method are also provided in the appendix. + +# Guidelines: + +- The answer NA means that the paper does not include experiments. +- If the paper includes experiments, a No answer to this question will not be perceived well by the reviewers: Making the paper reproducible is important, regardless of whether the code and data are provided or not. +- If the contribution is a dataset and/or model, the authors should describe the steps taken to make their results reproducible or verifiable. +- Depending on the contribution, reproducibility can be accomplished in various ways. For example, if the contribution is a novel architecture, describing the architecture fully might suffice, or if the contribution is a specific model and empirical evaluation, it may be necessary to either make it possible for others to replicate the model with the same dataset, or provide access to the model. In general, releasing code and data is often one good way to accomplish this, but reproducibility can also be provided via detailed instructions for how to replicate the results, access to a hosted model (e.g., in the case of a large language model), releasing of a model checkpoint, or other means that are appropriate to the research performed. +- While NeurIPS does not require releasing code, the conference does require all submissions to provide some reasonable avenue for reproducibility, which may depend on the nature of the contribution. For example +(a) If the contribution is primarily a new algorithm, the paper should make it clear how to reproduce that algorithm. +(b) If the contribution is primarily a new model architecture, the paper should describe the architecture clearly and fully. +(c) If the contribution is a new model (e.g., a large language model), then there should either be a way to access this model for reproducing the results or a way to reproduce the model (e.g., with an open-source dataset or instructions for how to construct the dataset). +(d) We recognize that reproducibility may be tricky in some cases, in which case authors are welcome to describe the particular way they provide for reproducibility. In the case of closed-source models, it may be that access to the model is limited in some way (e.g., to registered users), but it should be possible for other researchers to have some path to reproducing or verifying the results. + +# 5. Open access to data and code + +Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material? + +# Answer: [Yes] + +Justification: We include the complete code for data generation and all models in the supplementary material, ensuring full reproducibility of our results. A public GitHub repository will be provided with the final version of the paper. + +# Guidelines: + +- The answer NA means that paper does not include experiments requiring code. +- Please see the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details. +- While we encourage the release of code and data, we understand that this might not be possible, so "No" is an acceptable answer. Papers cannot be rejected simply for not including code, unless this is central to the contribution (e.g., for a new open-source benchmark). +- The instructions should contain the exact command and environment needed to run to reproduce the results. See the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details. + +- The authors should provide instructions on data access and preparation, including how to access the raw data, preprocessed data, intermediate data, and generated data, etc. +- The authors should provide scripts to reproduce all experimental results for the new proposed method and baselines. If only a subset of experiments are reproducible, they should state which ones are omitted from the script and why. +- At submission time, to preserve anonymity, the authors should release anonymized versions (if applicable). +- Providing as much information as possible in supplemental material (appended to the paper) is recommended, but including URLs to data and code is permitted. + +# 6. Experimental setting/details + +Question: Does the paper specify all the training and test details (e.g., data splits, hyperparameters, how they were chosen, type of optimizer, etc.) necessary to understand the results? + +Answer: [Yes] + +Justification: In the Appendix, we provide the training details of our method, along with the code as well. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The experimental setting should be presented in the core of the paper to a level of detail that is necessary to appreciate the results and make sense of them. +- The full details can be provided either with the code, in appendix, or as supplemental material. + +# 7. Experiment statistical significance + +Question: Does the paper report error bars suitably and correctly defined or other appropriate information about the statistical significance of the experiments? + +Answer: [Yes] + +Justification: We report the full experimental results with standard errors computed over multiple random seeds. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The authors should answer "Yes" if the results are accompanied by error bars, confidence intervals, or statistical significance tests, at least for the experiments that support the main claims of the paper. +- The factors of variability that the error bars are capturing should be clearly stated (for example, train/test split, initialization, random drawing of some parameter, or overall run with given experimental conditions). +- The method for calculating the error bars should be explained (closed form formula call to a library function, bootstrap, etc.) +- The assumptions made should be given (e.g., Normally distributed errors). +- It should be clear whether the error bar is the standard deviation or the standard error of the mean. +- It is OK to report 1-sigma error bars, but one should state it. The authors should preferably report a 2-sigma error bar than state that they have a $96\%$ CI, if the hypothesis of Normality of errors is not verified. +- For asymmetric distributions, the authors should be careful not to show in tables or figures symmetric error bars that would yield results that are out of range (e.g. negative error rates). +- If error bars are reported in tables or plots, The authors should explain in the text how they were calculated and reference the corresponding figures or tables in the text. + +# 8. Experiments compute resources + +Question: For each experiment, does the paper provide sufficient information on the computer resources (type of compute workers, memory, time of execution) needed to reproduce the experiments? + +# Answer: [Yes] + +Justification: We provide the details of the compute resources in the appendix. + +# Guidelines: + +- The answer NA means that the paper does not include experiments. +- The paper should indicate the type of compute workers CPU or GPU, internal cluster, or cloud provider, including relevant memory and storage. +- The paper should provide the amount of compute required for each of the individual experimental runs as well as estimate the total compute. +- The paper should disclose whether the full research project required more compute than the experiments reported in the paper (e.g., preliminary or failed experiments that didn't make it into the paper). + +# 9. Code of ethics + +Question: Does the research conducted in the paper conform, in every respect, with the NeurIPS Code of Ethics https://neurips.cc/public/EthicsGuidelines? + +# Answer: [Yes] + +Justification: Our work conforms, in every respect, with the NeurIPS Code of Ethics. + +# Guidelines: + +- The answer NA means that the authors have not reviewed the NeurIPS Code of Ethics. +- If the authors answer No, they should explain the special circumstances that require a deviation from the Code of Ethics. +- The authors should make sure to preserve anonymity (e.g., if there is a special consideration due to laws or regulations in their jurisdiction). + +# 10. Broader impacts + +Question: Does the paper discuss both potential positive societal impacts and negative societal impacts of the work performed? + +# Answer: [Yes] + +Justification: We discuss the broader impacts of our work in the appendix, which consists of both potential positive and negative societal implications. + +# Guidelines: + +- The answer NA means that there is no societal impact of the work performed. +- If the authors answer NA or No, they should explain why their work has no societal impact or why the paper does not address societal impact. +- Examples of negative societal impacts include potential malicious or unintended uses (e.g., disinformation, generating fake profiles, surveillance), fairness considerations (e.g., deployment of technologies that could make decisions that unfairly impact specific groups), privacy considerations, and security considerations. +- The conference expects that many papers will be foundational research and not tied to particular applications, let alone deployments. However, if there is a direct path to any negative applications, the authors should point it out. For example, it is legitimate to point out that an improvement in the quality of generative models could be used to generate deepfakes for disinformation. On the other hand, it is not needed to point out that a generic algorithm for optimizing neural networks could enable people to train models that generate Deepfakes faster. +- The authors should consider possible harms that could arise when the technology is being used as intended and functioning correctly, harms that could arise when the technology is being used as intended but gives incorrect results, and harms following from (intentional or unintentional) misuse of the technology. +- If there are negative societal impacts, the authors could also discuss possible mitigation strategies (e.g., gated release of models, providing defenses in addition to attacks, mechanisms for monitoring misuse, mechanisms to monitor how a system learns from feedback over time, improving the efficiency and accessibility of ML). + +# 11. Safeguards + +Question: Does the paper describe safeguards that have been put in place for responsible release of data or models that have a high risk for misuse (e.g., pretrained language models, image generators, or scraped datasets)? + +Answer: [NA] + +Justification: The dataset used in this work is publicly available, and our method does not pose a high risk of misuse. + +Guidelines: + +- The answer NA means that the paper poses no such risks. +- Released models that have a high risk for misuse or dual-use should be released with necessary safeguards to allow for controlled use of the model, for example by requiring that users adhere to usage guidelines or restrictions to access the model or implementing safety filters. +- Datasets that have been scraped from the Internet could pose safety risks. The authors should describe how they avoided releasing unsafe images. +- We recognize that providing effective safeguards is challenging, and many papers do not require this, but we encourage authors to take this into account and make a best faith effort. + +# 12. Licenses for existing assets + +Question: Are the creators or original owners of assets (e.g., code, data, models), used in the paper, properly credited and are the license and terms of use explicitly mentioned and properly respected? + +Answer: [Yes] + +Justification: We have cited all relevant prior works, datasets, and software packages used in this paper. License, copyright information, and terms of use will be provided in our GitHub repository. + +Guidelines: + +- The answer NA means that the paper does not use existing assets. +- The authors should cite the original paper that produced the code package or dataset. +- The authors should state which version of the asset is used and, if possible, include a URL. +- The name of the license (e.g., CC-BY 4.0) should be included for each asset. +- For scraped data from a particular source (e.g., website), the copyright and terms of service of that source should be provided. +- If assets are released, the license, copyright information, and terms of use in the package should be provided. For popular datasets, paperswithcode.com/datasets has curated licenses for some datasets. Their licensing guide can help determine the license of a dataset. +- For existing datasets that are re-packaged, both the original license and the license of the derived asset (if it has changed) should be provided. +- If this information is not available online, the authors are encouraged to reach out to the asset's creators. + +# 13. New assets + +Question: Are new assets introduced in the paper well documented and is the documentation provided alongside the assets? + +Answer: [Yes] + +Justification: The related docs will be included in the supplementary material. + +Guidelines: + +- The answer NA means that the paper does not release new assets. +- Researchers should communicate the details of the dataset/code/model as part of their submissions via structured templates. This includes details about training, license, limitations, etc. + +- The paper should discuss whether and how consent was obtained from people whose asset is used. +- At submission time, remember to anonymize your assets (if applicable). You can either create an anonymized URL or include an anonymized zip file. + +# 14. Crowdsourcing and research with human subjects + +Question: For crowdsourcing experiments and research with human subjects, does the paper include the full text of instructions given to participants and screenshots, if applicable, as well as details about compensation (if any)? + +Answer: [NA] + +Justification: This work does not involve crowdsourcing nor research with human subjects. + +Guidelines: + +- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. +- Including this information in the supplemental material is fine, but if the main contribution of the paper involves human subjects, then as much detail as possible should be included in the main paper. +- According to the NeurIPS Code of Ethics, workers involved in data collection, curation or other labor should be paid at least the minimum wage in the country of the data collector. + +# 15. Institutional review board (IRB) approvals or equivalent for research with human subjects + +Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals (or an equivalent approval/review based on the requirements of your country or institution) were obtained? + +Answer: [NA] + +Justification: This work does not involve crowdsourcing nor research with human subjects. + +Guidelines: + +- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. +- Depending on the country in which research is conducted, IRB approval (or equivalent) may be required for any human subjects research. If you obtained IRB approval, you should clearly state this in the paper. +- We recognize that the procedures for this may vary significantly between institutions and locations, and we expect authors to adhere to the NeurIPS Code of Ethics and the guidelines for their institution. +- For initial submissions, do not include any information that would break anonymity (if applicable), such as the institution conducting the review. + +# 16. Declaration of LLM usage + +Question: Does the paper describe the usage of LLMs if it is an important, original, or non-standard component of the core methods in this research? Note that if the LLM is used only for writing, editing, or formatting purposes and does not impact the core methodology, scientific rigorousness, or originality of the research, declaration is not required. + +Answer: [NA] + +Justification: The core method development in this research does not involve LLMs as any important, original, or non-standard components. + +Guidelines: + +- The answer NA means that the core method development in this research does not involve LLMs as any important, original, or non-standard components. +- Please refer to our LLM policy (https://neurips.cc/Conferences/2025/LLM) for what should or should not be described. \ No newline at end of file diff --git a/NeurIPS/2025/3D Equivariant Visuomotor Policy Learning via Spherical Projection/images.zip b/NeurIPS/2025/3D Equivariant Visuomotor Policy Learning via Spherical Projection/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..0a58d6546466beaec021dd4dc618782780218a69 --- /dev/null +++ b/NeurIPS/2025/3D Equivariant Visuomotor Policy Learning via Spherical Projection/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2832d3dd063d861a30823c0f3dc29a87f8e111ae425f6be8bd063468553f55c9 +size 1136145 diff --git a/NeurIPS/2025/3D Equivariant Visuomotor Policy Learning via Spherical Projection/layout.json b/NeurIPS/2025/3D Equivariant Visuomotor Policy Learning via Spherical Projection/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a7a66016b964caa825606f1c6aeeda5d6934a079 --- /dev/null +++ b/NeurIPS/2025/3D Equivariant Visuomotor Policy Learning via Spherical Projection/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0d07b97db6f54fff6e074e5df15c779fd8c7ab89047b2e93c1014341a5b05234 +size 1077550 diff --git a/NeurIPS/2025/3D Gaussian Flats_ Hybrid 2D_3D Photometric Scene Reconstruction/bfec7c41-0681-41c8-8491-f468a3e77d73_content_list.json b/NeurIPS/2025/3D Gaussian Flats_ Hybrid 2D_3D Photometric Scene Reconstruction/bfec7c41-0681-41c8-8491-f468a3e77d73_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..6dcb570f522664ab2942e69189dcb3337250cfeb --- /dev/null +++ b/NeurIPS/2025/3D Gaussian Flats_ Hybrid 2D_3D Photometric Scene Reconstruction/bfec7c41-0681-41c8-8491-f468a3e77d73_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7b0286c5aaf1a6516249b7dac357597494b4a714644fcb4dbe7b1a2777a381db +size 132205 diff --git a/NeurIPS/2025/3D Gaussian Flats_ Hybrid 2D_3D Photometric Scene Reconstruction/bfec7c41-0681-41c8-8491-f468a3e77d73_model.json b/NeurIPS/2025/3D Gaussian Flats_ Hybrid 2D_3D Photometric Scene Reconstruction/bfec7c41-0681-41c8-8491-f468a3e77d73_model.json new file mode 100644 index 0000000000000000000000000000000000000000..2ab4a1591122fd8adb41e0d121fc239befa1b0c9 --- /dev/null +++ b/NeurIPS/2025/3D Gaussian Flats_ Hybrid 2D_3D Photometric Scene Reconstruction/bfec7c41-0681-41c8-8491-f468a3e77d73_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:133b7ad374dc6ae11ae11415500aa567c648e0c0eea38c65b4d6fdc4ac000438 +size 169572 diff --git a/NeurIPS/2025/3D Gaussian Flats_ Hybrid 2D_3D Photometric Scene Reconstruction/bfec7c41-0681-41c8-8491-f468a3e77d73_origin.pdf b/NeurIPS/2025/3D Gaussian Flats_ Hybrid 2D_3D Photometric Scene Reconstruction/bfec7c41-0681-41c8-8491-f468a3e77d73_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..7611c79144c545d023669e96bf3d2ef9b29a8163 --- /dev/null +++ b/NeurIPS/2025/3D Gaussian Flats_ Hybrid 2D_3D Photometric Scene Reconstruction/bfec7c41-0681-41c8-8491-f468a3e77d73_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a2cb282626c4d59d52976d54fdd60c96beca38aa72dcfb412a392c295e1b99b4 +size 40052412 diff --git a/NeurIPS/2025/3D Gaussian Flats_ Hybrid 2D_3D Photometric Scene Reconstruction/full.md b/NeurIPS/2025/3D Gaussian Flats_ Hybrid 2D_3D Photometric Scene Reconstruction/full.md new file mode 100644 index 0000000000000000000000000000000000000000..2651182b03aac307067db451736f1d3928a158cb --- /dev/null +++ b/NeurIPS/2025/3D Gaussian Flats_ Hybrid 2D_3D Photometric Scene Reconstruction/full.md @@ -0,0 +1,619 @@ +# 3D Gaussian Flats: Hybrid 2D/3D Photometric Scene Reconstruction + +# Maria Taktasheva + +Simon Fraser University +maria_taktasheva@sfu.ca + +# Lily Goli* + +University of Toronto + +# Alessandro Fiorini + +University of Bologna + +# Zhen Li + +Simon Fraser University + +# Daniel Rebain + +University of British Columbia + +# Andrea Tagliasacchi* + +Simon Fraser University + +University of Toronto + +![](images/7a1503cbf7aba2e1499c093300306dd848a0dc7ce46ea98103a59dec183a71f3.jpg) +Faux render of 3D Gaussian Flats +Figure 1: Teaser - We introduce 3D Gaussian Flats, a hybrid representation of 2D Gaussians on semantically distinct planar surfaces and 3D Gaussians elsewhere (left). Our method achieves a photorealistic quality on par with fully 3D approaches, while improving geometry over surface reconstruction methods (right) e.g. no visible hole in the middle of the 'garden' scene from Mip-NeRF360 [1]. + +![](images/405f69e31a994698a042adb41ab39c98cfa8dd850063e7acf0fdf6b8decf86a9.jpg) +Ground Truth + +![](images/ffb5f1e65e0f2895bac8b0a1fa0a07d834c918719ed7000ad4a80a8419488b23.jpg) +Ours + +![](images/502cc33e72a40c20ddfffa9381e59354461bed6830472f3b627d0a6b0057dd6c.jpg) +Ours + +![](images/017935bc1c83b88cda04617897541b374711d07210fdc35fb9375492ae3bf033.jpg) +3DGS-MCMC + +![](images/7efd56dc26c4118d69ac8c4f975cdec5372e3c8291e3e67a50da44902dc7b8c5.jpg) +2DGS +3DGS +MCMC + +![](images/5e5b919d6a6c481c606bf78d94ffb9bdb7378843dd61a3b4066faff44ef0f6a1.jpg) +2DGS + +# Abstract + +Recent advances in radiance fields and novel view synthesis enable creation of realistic digital twins from photographs. However, current methods struggle with flat, texture-less surfaces, creating uneven and semi-transparent reconstructions, due to an ill-conditioned photometric reconstruction objective. Surface reconstruction methods solve this issue but sacrifice visual quality. We propose a novel hybrid 2D/3D representation that jointly optimizes constrained planar (2D) Gaussians for modeling flat surfaces and freeform (3D) Gaussians for the rest of the scene. Our end-to-end approach dynamically detects and refines planar regions, improving both visual fidelity and geometric accuracy. It achieves state-of-the-art depth estimation on ScanNet++ and ScanNetv2, and excels at mesh extraction without overfitting to a specific camera model, showing its effectiveness in producing high-quality reconstruction of indoor scenes. + +# 1 Introduction + +Recent advances in radiance fields and novel view synthesis have enabled the creation of realistic digital twins from collections of real-world photographs [2, 3]. These techniques allow for high-fidelity 3D reconstructions that capture intricate details of real-world scenes, making them invaluable for applications in virtual reality, gaming, cultural heritage preservation, and scientific visualization. + +However, when optimizing for novel view synthesis on flat and texture-less surfaces (e.g. walls, ceilings, tables that are prevalent in indoor scenes), current methods struggle in producing a faithful 3D reconstruction as the problem is photometrically under-constrained [4]. Specifically, modern novel view synthesis frameworks like [5, 6], which are optimized via volume rendering, model flat surfaces with low densities, resulting in non-opaque representations of solid surfaces; see the surface of the table in Figure 1 as an example. Conversely, surface reconstruction methods that assume solid, flat surfaces avoid this limitation [7]. However, they compromise visual quality in favor of a more parsimonious 3D reconstruction; see figure 1. Our core research question is whether these seemingly conflicting objectives could be achieved simultaneously. + +Some approaches have attempted to answer this questions by first creating a full 3D representation, and then - post-training - detecting planar surfaces to enable 3D planar reconstruction [8, 9]. However, these methods do not leverage planar assumptions during the optimization of the scene representation itself, limiting their effectiveness. Others enforce planar assumptions during training through various regularizer losses [10]. However, these losses can be hard to tune, as they are only suitable for the portion of the scene that is solid and flat, hindering the reconstruction whenever these assumptions are violated. + +In contrast to these methods, we propose to look at the problem in an end-to-end fashion, conjoining the process of photometric to the one of planar surface reconstruction. To achieve this, we introduce a hybrid 2D/3D representation, where flat surfaces are modeled with 2D Gaussian splats [7] that are confined to 2D planes, while the remaining of the scene is modeled with a classical, and more expressive, 3DGS model [6]. By jointly optimizing planar (2D) and freeform (3D) Gaussians, our approach enables better fitting of the final representation to planar surfaces within the scene. During photometric optimization, our method dynamically detects planar regions, and adaptively grows their extent, resulting in reconstruction that retains high visual quality (as measured by PSNR) compared to a classical 3DGS scene, while simultaneously achieving superior geometric accuracy (as measured by depth error). + +Our evaluations demonstrate that this hybrid representation achieves state-of-the-art depth estimation results on challenging indoor datasets including the new ScanNet++ dataset which was designed for dense reconstruction tasks using NeRF-based approaches, and the legacy ScanNetv2 dataset with sparser camera views. Our method delivers crisp reconstructed surfaces, while maintaining competitive visual quality compared to fully 3D representations. Beyond novel view synthesis, our approach has application in mesh extraction for planar surfaces, producing high-quality meshes and accurate mesh segmentation results across diverse capture setups (DSLR and iPhone captures), without the overfitting issues that negatively affect previous methods trained on specific camera models. + +# 2 Related Work + +Modern neural scene reconstruction methods aim to generate high-quality 3D representations from 2D images for applications like novel view synthesis [5, 6]. Despite significant progress, volumetric approaches struggle to accurately reconstruct planar surfaces [11], while surface reconstruction methods fail to recover volumetric effects [12]. Finding an approach that accurately reconstructs planar geometry without compromising the quality of the surrounding scene geometry and appearance is a key challenge. + +Representations for differentiable rendering Neural Radiance Field (NeRF) [5] pioneered scene reconstruction with a 3D neural representation optimized through differentiable volumetric rendering. 3D Gaussian Splatting (3DGS) [6] overcame NeRF slow training/rendering speed by representing scenes as efficiently rasterizable 3D Gaussians, dramatically accelerating rendering while maintaining quality. The impressive speed-quality balance of 3DGS quickly established it as a standard approach, + +with recent advancements such as 3DGS-MCMC [13] further enhancing its accessibility by eliminating the dependency on SfM initialization. Despite these innovations, volumetric representations still struggle with clean geometry reconstruction in flat and textureless surfaces common in indoor environments, hindering applications like mesh extraction. Our method addresses these challenges through a hybrid 2D/3D Gaussian representation that achieves superior geometric reconstruction while preserving rendering quality. + +Surface representations and planar constraints While NeRF [5] and 3DGS [6] employ fully volumetric representations, alternative approaches such as [11, 14] model scenes as solid surfaces. This philosophy inspired SuGaR [15], to use a regularization term that encourages the Gaussians to align with the surface of the scene, and later 2DGS [7], which uses 2D Gaussian primitives to reconstruct surfaces outperforming prior surface reconstruction methods [11, 14, 15]. Recent work [16] uses 2D Gaussians as in 2DGS, with multi-view depth and normal regularization to improve surface quality, while RaDe-GS [17] enables depth and normal rasterization for 3D Gaussians to support similar regularization. Other works introduced more explicit primitives, including planes [18, 19], estimizable geometry through learnable opacity maps [20], and soup of planes for dynamic reconstruction [21]. While these methods excel at representing flat surfaces with clean geometry, they typically sacrifice rendering quality and struggle to model phenomena that are better explained by volumetric effects, rather than surfaces. Some methods enforce planar constraints only as regularization losses, such as Guo et al. [22] that uses Manhattan world assumptions on semantically segmented regions and Chen et al. [23] that enforces plane normal consistency in textureless regions. Although helpful, regularizers can be difficult to tune. Our approach instead explicitly detects and optimizes planes within scene reconstruction, avoiding such issues. + +3D plane detection and reconstruction Another research direction detects planar surfaces in an initial 3D reconstruction and fits planes only to detected regions, extending single-image plane detection [24, 25] to multi-view settings. PlanarNeRF [26] adds a plane-predicting MLP branch to NeRF, supervised via ground truth labels or plane detection consistency across frames, but prevents plane MLP gradients from affecting the geometry prediction branch. PlanarRecon [8] reconstructs a sparse feature volume, which is decoded into plane features and clustered. AirPlanes [9] and NeuralPlane [27] build 3D-consistent plane embeddings per 3D point, emphasizing semantic priors for accurate detection. While we also use semantic knowledge, our method jointly detects and optimizes planes alongside scene reconstruction, allowing geometry to benefit from planar constraints. Further, unlike these methods, our approach yields full scene reconstructions suitable for novel view synthesis, vs. a coarse surface reconstruction. + +Hybrid representations Recent hybrid 2D-3D approaches have explored planar surface representation. Kim and Lim [28] integrate meshes into 3DGS for indoor scenes, using SAM [29] to detect planar surfaces and represent them with meshes while employing 3D Gaussians for other objects. Zanjani et al. [30] combine SAM segmentation with normal estimation to lift 2D plane descriptors to 3D, clustering the planar Gaussians using a tree structure. In contrast, our method offers a simpler solution by representing the scene with a mixture of 2D and 3D Gaussians. This design remains fully compatible with the 3DGS rendering pipeline, eliminating the need for complex hybrid mesh handling, or hierarchical tree structures. + +# 3 Method + +Given $N$ posed images $\{I_c\}$ and $M$ planar surfaces $\{P_p\}$ , each specified by binary image masks $\{\mathcal{M}_{p,c}\}$ , we aim to reconstruct a hybrid novel view synthesis method that combines a classical 3DGS model with a 2D piecewise planar representation of the scene. Our goal is to reconstruct the scene so that the planar surfaces are accurately recovered and compactly represented by 2D Gaussian primitives, while the rest of the scene is modeled with 3D Gaussians, with the key objective of avoiding the artifacts that can typically be seen when using 3D primitives to model planar surfaces; see Figure 1. + +![](images/3bea045433d677876a4ae6c331fef3a99563bb3de61ea646a5e001b24fcd44fb.jpg) +Figure 2: Overview - Training of our model is split into two parts: warm-up, in which 3D Gaussians are trained as in [6] using a photometric loss; and planar training, in which 3D Gaussians and planar Gaussians are trained along with the parameters of the planes to which planar Gaussians are locked. Planar training is performed in alternating phases, with Gaussian parameters frozen while plane parameters are optimized, and vice versa. Legend: learnable (warm up), learnable (Gaussian phase), learnable (plane phase). + +# 3.1 Hybrid representation + +Our representation consists of $M$ planes $\mathcal{P} = \{P_p\}$ , each characterized by its 3D origin and normal $(\mathbf{o}_p, \mathbf{n}_p)$ . The geometry of each such plane $P_p$ is represented through a set of 2D Gaussians $\mathcal{G} = \{\mathbf{g}_k\}_{k=1}^{K_k}$ such that, + +$$ +\mathbf {g} _ {k} = \mathcal {N} \left(\mu_ {k}, \Sigma_ {k}\right), \quad \mu_ {k} \in P _ {k}, \quad \Sigma_ {k} \in \mathbb {R} ^ {2 \times 2}. \tag {1} +$$ + +Here, $\mu_{k}$ is the center of the $k$ -th Gaussian on the plane $P_{p}$ , and $\Sigma_{k}$ is the 2D covariance matrix, parametrized with a 2D in-plane rotation $\mathbf{R}_{k}$ and a 2D diagonal scale matrix $\mathbf{S}_{k}$ . The plane-to-world transformation matrix is defined as $\mathbf{T}_{\mathrm{pw}} = \mathrm{hom}(\mathbf{R},\mathbf{o})$ , where $\mathbf{R}$ is any rotation matrix that satisfies $\hat{z} = \mathbf{Rn}$ with $\hat{z}$ being the unit vector along the z-axis in the world frame. Thus, the degrees of freedom of planar Gaussians can be mapped to world coordinates through the rigid transformation: + +$$ +\bar {\mu} _ {k} = \mathbf {T} _ {\mathrm {p w}} [ \mu_ {k}; 0; 1 ], \quad \bar {\Sigma} _ {k} = \mathbf {T} _ {\mathrm {p w}} \operatorname {d i a g} (\Sigma_ {k}, 1, 1) \mathbf {T} _ {\mathrm {p w}} ^ {\top} \tag {2} +$$ + +yielding a standard 3D Gaussian primitive representation suitable for rendering. The remaining scene geometry is represented by unconstrained 3D Gaussians $\bar{\mathcal{G}} = \{\bar{\mathbf{g}}_k\}_{k=1}^{\bar{K}}$ : + +$$ +\bar {\mathbf {g}} _ {k} = \mathcal {N} (\bar {\mu} _ {k}, \bar {\Sigma} _ {k}), \quad \mu_ {k} \in \mathbb {R} ^ {3}, \quad \Sigma_ {k} \in \mathbb {R} ^ {3 \times 3} \tag {3} +$$ + +All Gaussians have view-dependent colors $\mathbf{c}$ represented as Spherical Harmonics, and opacity $\alpha$ as in vanilla 3DGS. To reconstruct the scene with our hybrid representation, we need to optimize the degrees of freedom of planes $\mathcal{P}$ , 2D planar Gaussians $\mathcal{G}$ , and 3D freeform Gaussians $\bar{\mathcal{G}}$ . We begin our optimization with a warm-up stage using only 3D Gaussians (for N=3500 iterations). After that, we begin our planar reconstruction where in each round of optimization we: (i) dynamically initialize plane parameters by robustly fitting planes to the current representation (section 3.2); (ii) alternate between optimizing plane and Gaussian parameters (section 3.2); (iii) densify our representation through a (slightly modified) MCMC densification, due to the challenges of optimizing compact-support functions (section 3.4). + +# 3.2 Plane initialization + +For compactness of notation, let us drop our indices, and consider the binary mask $\mathcal{M} \gets \mathcal{M}_{c,p}$ for the $p$ -th planar surface in the $c$ -th view, and denote with $\pi$ the function that projects a 3D point to the coordinate frame of the $n$ -th image. We start by selecting all the Gaussians (i) whose mean projects into the mask, (ii) that are sufficiently opaque, and (iii) that lie within a shell of the expected ray termination of the $n$ -th image: + +$$ +\tilde {\mathcal {G}} = \left\{\bar {\mathbf {g}} _ {k} \mid \pi (\bar {\mu} _ {k}) \in \mathcal {M}, \alpha_ {k} > \kappa , | D (\pi (\bar {\mu} _ {k})) - d _ {k} | < \delta \right\}, \tag {4} +$$ + +where the thresholds $\alpha_{\mathrm{th}} = 0.1$ and $d_{\mathrm{th}} = 0.05$ are hyper-parameters that control this selection process, and where $D$ is the expected ray termination map (i.e. depth map), and $d_{k}$ is the depth of the + +![](images/370daa86e6e23fde972eecb675e474793fe2aefd433fb840310df5869901ef66.jpg) +Figure 3: Planar Relocation - A freeform Gaussian (teal) gets relocated to the plane to become a planar Gaussian (brown), when both its distance to the plane $(d_{\perp})$ and along $(d_{\parallel})$ the plane are small. + +Gaussians. We then extract a candidate plane $P$ by RANSAC optimization on the point cloud that samples the Gaussians: + +$$ +P, \mathcal {I} = \operatorname {R A N S A C} \left(\left\{x \sim \bar {\mathbf {g}} \mid \bar {\mathbf {g}} \in \tilde {\mathcal {G}} \right\}, \epsilon\right) \tag {5} +$$ + +where we accept $P$ as a viable plane candidate only whenever the mean inlier residual is lower than $\epsilon$ . The set $\mathcal{I}$ includes the indexes of Gaussians in $\tilde{\mathcal{G}}$ that are inliers of the RANSAC process. We further discard planes that are too small with set $\mathcal{I}$ having a smaller size than 100. Once a plane corresponding to $\mathcal{M}$ has been accepted, all the semantic masks for that plane $p$ are excluded from subsequent RANSAC runs. The plane initialization process is repeated for remaining masks, after each completed round of plane and Gaussian optimization, as described in Section 3.3. + +Snapping We then remove the discovered inliers from the set of 3D Gaussians $\bar{\mathcal{G}}\gets \bar{\mathcal{G}}\setminus \mathcal{I}$ , and add them to our set of 2D Gaussians $\mathcal{G}\gets \mathcal{G}\cup \mathcal{I}$ . During the latter operation, we clip 3D Gaussians to 2D to become planar by transforming to the local plane coordinates, and set the third component of their means and scales to zero. Further, only rotation about the $z$ -axis in local plane coordinates is preserved + +Active set update If the accepted plane $\mathcal{P}_i$ has an angular distance below a threshold to an already existing plane, while its origin $\mathbf{o}_i$ also has a small Euclidean distance to the closest Gaussian center on that plane, we merge the two planes. Otherwise, the plane is added as a new plane to the active set of planes $\mathcal{P}$ . In merging, we assign the new plane's Gaussians to the previously found one. This allows our optimization to merge planar areas that have only been partially observed in any view. + +# 3.3 Optimization + +We optimize our representation by block-coordinate descent, starting each round of optimization by only optimizing the plane parameters for a fixed number of 10 iterations, and then freezing these, and optimizing the Gaussian parameters (both 2D and 3D) for another 100 iterations. This alternation in optimization is critical to avoid instability; see an ablation in figure 7. In the first optimization block, within each iteration, the parameters of the $p$ -th plane within the $c$ -th image are optimized by the loss: + +$$ +\underset {\mathbf {o} _ {p}, \mathbf {n} _ {p}} {\arg \min } = \underbrace {\left\| I _ {c} - \tilde {I} _ {c} \right\| _ {1}} _ {\mathcal {L} _ {\text {p h o t o}}} + \lambda_ {\text {m a s k}} \underbrace {\left\| \mathcal {M} _ {c , p} - \tilde {\mathcal {M}} _ {c , p} \right\| _ {1}} _ {\mathcal {L} _ {\text {m a s k}}}, \tag {6} +$$ + +where $\tilde{\mathcal{M}}$ is the predicted plane mask, obtained by rendering the mixture of Gaussians with binarized color (white for planar, and black for 3D), and alpha blended using the original Gaussian opacities during the rasterization. In the second optimization block, we optimize all Gaussian parameters jointly: + +$$ +\underset {\mathcal {G}, \bar {\mathcal {G}}} {\arg \min } \mathcal {L} _ {\text {p h o t o}} + \lambda_ {\text {m a s k}} \mathcal {L} _ {\text {m a s k}} + \lambda_ {\text {T V}} \mathcal {L} _ {\text {T V}} + \lambda_ {\text {s c a l e}} \mathcal {L} _ {\text {s c a l e}} + \lambda_ {\text {o p a c i t y}} \mathcal {L} _ {\text {o p a c i t y}}, \tag {7} +$$ + +where $\mathcal{L}_{\mathrm{TV}}$ is the total depth variation regularization from Niemeyer et al. [10], $\mathcal{L}_{\mathrm{scale}}$ is the scale regularizer and $\mathcal{L}_{\mathrm{opacity}}$ is the opacity regularizer from Kheradmand et al. [13] that vanishes the size of Gaussians that are unconstrained by the photometric loss. Note that planar Gaussians move rigidly during plane optimization (6), and move locally in the plane during Gaussian optimization (7), as only their 2D in-plane parameters are optimized. + +# 3.4 Planar relocation + +We follow 3DGS-MCMC [13] in our training dynamics. For densification of planes, we rely on relocating low-opacity Gaussians to locations of dense high-opacity Gaussians, as this allows + +![](images/4601dcea2809b0af7bf1a40cb83808fae60eabc0eb8d4c948c8d437cf5b523e0.jpg) +Figure 4: Novel View Synthesis - Quantitative and qualitative results show significant improvement in predicted depth compared to previous work, while maintaining comparable rendering quality to the full 3D representations. + +transferring between 3D and 2D/planar Gaussians. However, the number of Gaussians on planes, especially when the plane has weak texture, is usually low, leading to a slow densification rate for planes / planar Gaussians. To address this issue, whenever a freeform Gaussian projects into the current mask $\pi (\bar{\mu}_k)\in \mathcal{M}$ , and it is sufficiently close to the currently reconstruction, we stochastically re-locate it to the plane. To measure distance, we identify the 2D Gaussian with the smallest Euclidean distance to $\bar{\mu}_k$ , and measure its distance in the direction of the plane normal $d_{\perp}$ , and the one along the plane $d_{||}$ ; see Figure 3. We stochastically relocate this if both distances are sufficiently small, as expressed by the following Bernoulli distribution: + +$$ +p \sim \mathcal {B} (\beta), \beta = \left[ 1 - \Phi \left(\frac {d _ {\perp}}{\sigma_ {\perp}}\right) \right] \cdot \left[ 1 - \Phi \left(\frac {d _ {\parallel}}{\sigma_ {\parallel}}\right) \right], \tag {8} +$$ + +where $\Phi$ is the cumulative distribution function of a Gaussian, and $\sigma_{\perp}$ and $\sigma_{\parallel}$ are hyper-parameters that control the stochastic re-location. + +# 4 Results + +We validate our proposed method for scene reconstruction through the novel view synthesis task on common indoor scene datasets, assessing both rendered image and depth quality metrics (section 4.1). We then show an application of our method to mesh extraction for planar surfaces (section 4.2). Finally, we validate our design choices through an ablation study on different aspects of the method (section 4.3). We provide our implementation details in the supplementary material. + +![](images/7948438cec2fc0990c8a3e824813bffb9b2335d60ac9c09c0b9495cbc2eb2ce0.jpg) +Figure 5: Novel View Synthesis on ScanNetv2 - Our method outperforms baselines in image and depth quality on ScanNetv2 despite sparse camera views. + +
Metric3DGS-MCMC2DGSOurs
RMSE↓0.460.600.40
MAE↓0.370.440.31
AbsRel↓0.190.230.16
δ < 1.25 ↑0.610.630.70
δ < 1.25² ↑0.870.770.90
δ < 1.25³ ↑0.950.830.97
PSNR↑20.1821.4421.75
LPIPS↓0.290.300.27
SSIM↑0.830.850.86
# primitives500K809K500K
(% planar)(17.6%)
+ +# 4.1 Novel View Synthesis - Figures 4 and 5 + +We evaluate our hybrid representation's novel view synthesis on common indoor scene reconstruction benchmarks and provide comparisons with both state-of-the-art fully 3D representations and 2D surface reconstruction approaches. We show a significant improvement in the reconstructed surface geometry while maintaining high visual quality. + +Datasets We perform evaluations on common indoor scene benchmarks ScanNet++[31] and ScanNetv2[32], as they primarily feature indoor scenes with flat textureless surfaces suitable for the task at hand. ScanNet++ provides dense scenes with SfM camera poses and sparse point clouds, designed primarily for 3D reconstruction approaches that follow the NeRF [5] paradigm. Conversely, the legacy version of ScanNet i.e. ScanNetv2 offers sparser views without SfM information. Our method works with or without initial sparse point clouds, enabling reconstruction initialized with sparse SfM point cloud on ScanNet++ and experiments with randomly initialized point clouds on ScanNetv2. For ScanNet++, we use 11 training scenes with ground truth meshes for depth derivation, utilizing iPhone video streams, sampling every 10th frame for training at $2 \times$ downsampling and every 8th for testing. We chose the scenes that are diverse in their content and contain various planar surfaces. For ScanNet, we evaluate on 5 scenes with sufficient overlapping views of planar surfaces following the data preparation scheme of [27]. The 2D plane masks were generated using PlaneRecNet [25] and propagated through the image sequence with SAMv2 video processor [29]. + +Baselines We compare against SOTA reconstruction methods, both fully 3D representations and 2D surface reconstruction methods. For 3D representations, we compare with vanilla 3DGS [6], and 3DGS-MCMC [13] as it is more robust version to random initializations, and has higher rendering quality. Within photometric surface reconstruction methods, we compare to 2DGS [7] as a widely used state-of-the-art, as well as to PGSR [16] and RaDe-GS [17], which more recently report improved depth quality. All methods are trained for 30K iterations. + +Metrics We use the common image quality metrics PSNR, SSIM and LPIPS for evaluating the rendered RGB. Further, we choose depth as a strong indicator for the quality of the reconstructed surface geometry. We provide depth quality metrics by computing the rendered depth as the expected ray termination at each pixel. We report RMSE, MAE and average absolute error relative to ground truth depth (AbsRel). Additionally, we provide depth accuracy percentage at different error thresholds similar to [33]. The metrics are computed only on the defined portion of the ground-truth depths. We further report the total number of primitives in our model and the percentage that are planar (and thus can be represented more compactly). + +Analysis Quantitative and qualitative results across both datasets show significant improvement in depth accuracy compared to all baselines. Notably, our method achieves comparable image quality to SOTA 3D representations on dense ScanNet++ scenes while surpassing them in depth quality, evidenced by sharper geometry reconstruction in qualitative examples. The slight PSNR difference with 3D methods reflects a trade-off: our constrained geometry enforces correct structure, while unconstrained methods can inflate PSNR by fitting view-dependent effects with incorrect geometry. + +![](images/478bedcfbc45dd4ed2705f677a1e44189b688e2ccf8cc9e312706349ecff5f7d.jpg) +Figure 6: Mesh Extraction - Our method shows consistent results across iPhone and DSLR captures, while baselines typically overfit to one camera type. Qualitatively, our approach extracts complete meshes for most target planes with fewer inaccurate plane detections (shown in gray) compared to baselines. Target planes are shown with distinct colors on the ground truth. + +In the sparser ScanNetv2 scenes, our approach delivers superior performance in both depth and image quality, leveraging the planar prior of indoor environments to overcome the geometric ambiguity that challenges pure 3D methods in sparse captures. Our method also substantially outperforms 2DGS in both image fidelity and depth accuracy metrics. + +# 4.2 Mesh Extraction - figure 6 + +Our method enables mesh extraction from reconstructed 3D planar surfaces. For each plane, we un-project all 2D segmentation masks to 3D by computing ray-plane intersections, yielding a point cloud. This point cloud is downsampled using fixed-size voxels and rasterized onto plane coordinates to create an occupancy grid. We then use Marching Squares for contour extraction (We omit small contours with less than 100 points), followed by ear-clipping triangulation to produce the final mesh. We evaluate the quality of the retrieved mesh for the planar surfaces and compare our method to planar reconstruction methods. + +Datasets We use ScanNet++ to extract planar surface meshes. We show results both on the subset of this dataset captured by iPhone and also the DSLR subset, showing that our method can handle different camera models, while previous methods usually overfit to one modality. For ground truth, we follow the approach of Watson et al. [9] to obtain a ground truth planar mesh. We then only consider the subset of planes in the ground truth mesh that we have annotated segmentation masks for each scene. We provide details on selecting these planes in Appendix E. + +Baselines We compare against previous planar reconstruction methods AirPlanes [9] and PlanarRecon [8] that provide extracted planar mesh as output of their methods. We follow the same evaluation setting as in the original papers on the iPhone subset of the dataset. For DSLR images, we crop the images to the specified FoV in each baseline to match their training distribution. + +![](images/8300c5e60cc2f9c787f2b1e98b4c7b2163cd8162e6147520d9e59a2f8c2cf30f.jpg) +Figure 7: Ablation on design choices - Loss components and optimization strategy are critical, with simultaneous plane-Gaussian optimization causing significant drops. 2D Gaussian snapping greatly improves depth accuracy compared to regularization alternatives. Similarly, Gaussian relocation is essential. + +Metrics We report mesh accuracy metrics including accuracy, precision, recall, completeness and Chamfer distance as defined in Ye et al. [27]. We also provide mesh segmentation metrics that evaluate how well detected plane segments match ground truth segments following [9]. + +Analysis Our method outperforms the baselines on DSLR images subset of the dataset. Unlike previous methods that are trained on specific modalities (i.e. phone camera) and struggle to transfer to different camera models (i.e. DSLR camera), our approach maintains consistent mesh quality due to having zero-shot mesh extraction on test scenes through photometric reconstruction. Additionally, our method outperforms PlanarRecon on iPhone data, while having competitive performance to AirPlanes. Qualitative results reveal that both PlanarRecon and AirPlanes extract extraneous planes with numerous random small fragments, resulting in unsightly and impractical meshes. In contrast, our method produces clean planar surfaces, yielding a more coherent and usable reconstruction. + +# 4.3 Ablation - Figure 7 + +We ablate our design choices and additionally test our method's robustness to random point cloud initialization (in table 1). + +Loss design We ablate the effect of $\mathcal{L}_{mask}$ and $\mathcal{L}_{TV}$ . Although removing these losses reduces the image quality by some margin, it affects depth quality more significantly. Qualitative rendering shows that $\mathcal{L}_{mask}$ contributes significantly to detecting and growing 2D Gaussians. + +**Optimization design** Our method is based on optimizing Gaussians and plane parameters together in an alternating fashion. We show that fixing plane parameters with no optimization degrades our results both quantitatively and qualitatively. Simultaneous joint optimization of Gaussians and planes also affects the results negatively. In Figure 7, note how the floor plane gets stuck above the ground level, as revealed by its intersection with the bin. + +2D Gaussian design Using hybrid 2D/3D Gaussians is one of the main components of our design. Therefore, we ablate the necessity of having 2D Gaussians by disabling snapping as described in Section 3.2. This shows a significant drop in depth accuracy, which is also evident in qualitative results. As an alternative to snapping, we can regularize the smallest scale component in planar Gaussians. However, we find that this approach is difficult to tune and provides suboptimal results. Finally, we ablate our densification process with relocation of Gaussians to planes. Without relocation, planes are not fully detected, with the planar Gaussians comprising the plane maintaining low opacity. Furthermore, some of the Gaussians remain close to the plane while not being detected as belonging to that plane. + +# 5 Conclusions + +We introduce 3D Gaussian Flats, a hybrid 2D/3D Gaussian representation that accurately models planar surfaces without sacrificing rendering quality. Our method jointly optimizes 2D Gaussians constrained to planar surfaces alongside free-form Gaussians for the remaining scene. By leveraging semantic segmentation masks, we predict both a full 3D representation and semantically distinct planes for planar mesh extraction in indoor scenes. Our approach achieves state-of-the-art depth estimation on indoor scene benchmarks while maintaining high image quality. Additionally, our planar mesh extraction method generalizes across different camera models, overcoming domain gap limitations that typically cause previous methods to fail. + +**Limitations** Our reliance on initial 3DGS reconstruction often generates insufficient Gaussians in flat areas with no texture, although this potentially can be addressed via more adaptive densification strategies. Further, using a weak spherical harmonics appearance model still leads to building extra geometry to compensate for view-dependent effects, which a stronger appearance model would resolve. Additionally, we depend on 2D semantic masks from SAMv2 that may contain errors, but our method will naturally improve alongside advances in semantic segmentation. Finally, our RANSAC-based approach, while robust, introduces computational overhead that extends training time. We believe our hybrid representation opens exciting new avenues for research into more efficient approaches that balance geometric precision with visual fidelity. + +# References + +[1] Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. CVPR, 2022. URL https://github.com/google-research/multinerf.1 +[2] A. Tewari, J. Thies, B. Mildenhall, P. Srinivasan, E. Tretschk, W. Yifan, C. Lassner, V. Sitzmann, R. Martin-Brualla, S. Lombardi, T. Simon, C. Theobalt, M. Nießner, J. T. Barron, G. Wetzstein, M. Zollhöfer, and V. Golyanik. Advances in neural rendering. Computer Graphics Forum, 2022. 2 +[3] Guikun Chen and Wenguan Wang. A survey on 3d gaussian splatting. arXiv preprint arXiv:2401.03890, 2025. 2 +[4] Lily Goli, Cody Reading, Silvia Sellán, Alec Jacobson, and Andrea Tagliasacchi. Bayes' Rays: Uncertainty quantification in neural radiance fields. CVPR, 2024. URL https://github.com/BayesRays/BayesRays.2 +[5] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. URL https://github.com/bmild/nerf.2,3,7 +[6] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 2023. URL https://github.com/graphdeco-inria/gaussian-splatting.2,3,4,6,7 +[7] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically accurate radiance fields. In SIGGRAPH, 2024. URL https://github.com/hbb1/2d-gaussian-splatting. 2, 3, 6, 7, 1 +[8] Yiming Xie, Matheus Gadelha, Fengting Yang, Xiaowei Zhou, and Huaizu Jiang. Planarrecon: Real-time 3d plane detection and reconstruction from posed monocular videos. In CVPR, 2022. URL https://github.com/neu-vi/PlanarRecon.2,3,8,6 +[9] Jamie Watson, Filippo Aleotti, Mohamed Sayed, Zawar Qureshi, Oisin Mac Aodha, Gabriel Brostow, Michael Firman, and Sara Vicente. Airplanes: Accurate plane estimation via 3d-consistent embeddings. In CVPR, 2024. URL https://github.com/nianticlabs/airplanes.2,3,8,9,6 + +[10] Michael Niemeyer, Jonathan T. Barron, Ben Mildenhall, Mehdi S. M. Sajjadi, Andreas Geiger, and Noha Radwan. Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In CVPR, 2022. URL https://github.com/google-research/google-research/tree/master/regnerf.2,5,7 +[11] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. NeurIPS, 2021. URL https://github.com/Totoro97/NeuS.2,3 +[12] Zian Wang, Tianchang Shen, Merlin Nimier-David, Nicholas Sharp, Jun Gao, Alexander Keller, Sanja Fidler, Thomas Müller, and Zan Gojcic. Adaptive shells for efficient neural radiance field rendering. ACM TOG., 2023. 2 +[13] Shakiba Kheradmand, Daniel Rebain, Gopal Sharma, Weiwei Sun, Jeff Tseng, Hossam Isack, Abhishek Kar, Andrea Tagliasacchi, and Kwang Moo Yi. 3d gaussian splatting as markov chain monte carlo. In NeurIPS, 2024. URL https://github.com/ubc-vision/3dgs-mcmc. 3, 5, 6, 7, 1 +[14] Zhaoshuo Li, Thomas Müller, Alex Evans, Russell H Taylor, Mathias Unberath, Ming-Yu Liu, and Chen-Hsuan Lin. Neuralangelo: High-fidelity neural surface reconstruction. In CVPR, 2023. URL https://github.com/NVlabs/neuralangelo.3 +[15] Antoine Guédon and Vincent Lepetit. Sugar: Surface-aligned gaussian splatting for efficient 3d mesh reconstruction and high-quality mesh rendering. In CVPR, 2024. URL https://github.com/Aanttwo/SuGaR.3 +[16] Danpeng Chen, Hai Li, Weicai Ye, Yifan Wang, Weijian Xie, Shangjin Zhai, Nan Wang, Haomin Liu, Hujun Bao, and Guofeng Zhang. Pgsr: Planar-based gaussian splatting for efficient and high-fidelity surface reconstruction. IEEE Transactions on Visualization and Computer Graphics, 2024. URL https://github.com/zju3dv/PGSR.3, 6, 7 +[17] Baowen Zhang, Chuan Fang, Rakesh Shrestha, Yixun Liang, Xiaoxiao Long, and Ping Tan. Rade-gs: Rasterizing depth in gaussian splatting. arXiv preprint arXiv:2406.01467, 2024. URL https://github.com/BaowenZ/RaDe-GS.3,6,7 +[18] Zhi-Hao Lin, Wei-Chiu Ma, Hao-Yu Hsu, Yu-Chiang Frank Wang, and Shenlong Wang. Neurmips: Neural mixture of planar experts for view synthesis. In CVPR, 2022. URL https://github.com/chih-hao-lin/neurmips.3 +[19] Bin Tan, Rui Yu, Yujun Shen, and Nan Xue. Planarsplatting: Accurate planar surface reconstruction in 3 minutes. In CVPR, 2025. 3 +[20] David Svitov, Pietro Morerio, Lourdes Agapito, and Alessio Del Bue. Billboard splatting (bb-splat): Learnable textured primitives for novel view synthesis. arXiv preprint arXiv:2411.08508, 2024. URL https://github.com/david-svitov/BBSplat.3 +[21] Yao-Chih Lee, Zhoutong Zhang, Kevin Blackburn-Matzen, Simon Niklaus, Jianming Zhang, Jia-Bin Huang, and Feng Liu. Fast view synthesis of casual videos with soup-of-planes. In ECCV, 2024. 3 +[22] Haoyu Guo, Sida Peng, Haotong Lin, Qianqian Wang, Guofeng Zhang, Hujun Bao, and Xiaowei Zhou. Neural 3d scene reconstruction with the manhattan-world assumption. In CVPR, 2022. URL https://github.com/zju3dv/manhattan_sdf.3 +[23] Zheng Chen, Chen Wang, Yuan-Chen Guo, and Song-Hai Zhang. Structnerf: Neural radiance fields for indoor scenes with structural hints. IEEE TPAMI, 2023. 3 +[24] Chen Liu, Kihwan Kim, Jinwei Gu, Yasutaka Furukawa, and Jan Kautz. Planercnn: 3d plane detection and reconstruction from a single image. In CVPR, 2019. URL https://github.com/NVlabs/planercnn.3 +[25] Yaxu Xie, Fangwen Shu, Jason Rambach, Alain Pagani, and Didier Stricker. Planerecnet: Multi-task learning with cross-task consistency for piece-wise plane detection and reconstruction from a single rgb image. In BMVC, 2021. URL https://github.com/EryiXie/PlaneRecNet.3,7,4,6 + +[26] Zheng Chen, Qingan Yan, Huangying Zhan, Changjiang Cai, Xiangyu Xu, Yuzhong Huang, Weihan Wang, Ziyue Feng, Lantao Liu, and Yi Xu. Planarnerf: Online learning of planar primitives with neural radiance fields. arXiv preprint arXiv:2401.00871, 2023. 3 +[27] Hanqiao Ye, Yuzhou Liu, Yangdong Liu, and Shuhan Shen. Neuralplane: Structured 3d reconstruction in planar primitives with neural fields. In ICLR, 2025. URL https://github.com/3dv-casia/NeuralPlane. 3, 7, 9 +[28] Jiyeop Kim and Jongwoo Lim. Integrating meshes and 3d gaussians for indoor scene reconstruction with sam mask guidance. arXiv preprint arXiv:2407.16173, 2024. 3 +[29] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junting Pan, Kalyan Vasudev Alwala, Nicolas Carion, Chao-Yuan Wu, Ross Girshick, Piotr Dólar, and Christoph Feichtenhofer. Sam 2: Segment anything in images and videos. *ICLR*, 2025. URL https://github.com/facebookresearch/segment-anything. 3, 7, 4, 6 +[30] Farhad G Zanjani, Hong Cai, Hanno Ackermann, Leila Mirvakhabova, and Fatih Porikli. Planar gaussian splatting. In WACV, 2025. 3 +[31] Chandan Yeshwanth, Yueh-Cheng Liu, Matthias Nießner, and Angela Dai. Scannet++: A high-fidelity dataset of 3d indoor scenes. In ICCV, 2023. URL https://kaldir.vc.in.tum.de/scannetpp. Licensed under the ScanNet++ Terms of Use. 7, 17, 1, 3, 4 +[32] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In CVPR, 2017. URL http://www.scan-net.org. Licensed under the MIT License. 7, 17 +[33] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In CVPR, 2024. URL https://github.com/LiheYoung/Depth-Anything. 7 +[34] Thomas Schöps, Johannes L. Schonberger, Silvano Galliani, Torsten Sattler, Konrad Schindler, Marc Pollefeys, and Andreas Geiger. A multi-view stereo benchmark with high-resolution images and multi-camera videos. In Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 1 +[35] Zehao Yu, Torsten Sattler, and Andreas Geiger. Gaussian opacity fields: Efficient adaptive surface reconstruction in unbounded scenes. ACM Transactions on Graphics, 2024. 1 +[36] Matias Turkulainen, Xuqian Ren, Iaroslav Melekhov, Otto Seiskari, Esa Rahtu, and Juho Kannala. Dn-splatter: Depth and normal priors for gaussian splatting and meshing. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2025. 1 + +# NeurIPS Paper Checklist + +# 1. Claims + +Question: Do the main claims made in the abstract and introduction accurately reflect the paper's contributions and scope? + +Answer: [Yes] + +Justification: A discussion of current state of the literature can be found in section 2, proposed method is defined in section 3 and the discussion of results in section 4 support the claims made in the abstract and introduction. + +Guidelines: + +- The answer NA means that the abstract and introduction do not include the claims made in the paper. +- The abstract and/or introduction should clearly state the claims made, including the contributions made in the paper and important assumptions and limitations. A No or NA answer to this question will not be perceived well by the reviewers. +- The claims made should match theoretical and experimental results, and reflect how much the results can be expected to generalize to other settings. +- It is fine to include aspirational goals as motivation as long as it is clear that these goals are not attained by the paper. + +# 2. Limitations + +Question: Does the paper discuss the limitations of the work performed by the authors? + +Answer: [Yes] + +Justification: Limitations of the work are listed in section 5. + +Guidelines: + +- The answer NA means that the paper has no limitation while the answer No means that the paper has limitations, but those are not discussed in the paper. +- The authors are encouraged to create a separate "Limitations" section in their paper. +- The paper should point out any strong assumptions and how robust the results are to violations of these assumptions (e.g., independence assumptions, noiseless settings, model well-specification, asymptotic approximations only holding locally). The authors should reflect on how these assumptions might be violated in practice and what the implications would be. +- The authors should reflect on the scope of the claims made, e.g., if the approach was only tested on a few datasets or with a few runs. In general, empirical results often depend on implicit assumptions, which should be articulated. +- The authors should reflect on the factors that influence the performance of the approach. For example, a facial recognition algorithm may perform poorly when image resolution is low or images are taken in low lighting. Or a speech-to-text system might not be used reliably to provide closed captions for online lectures because it fails to handle technical jargon. +- The authors should discuss the computational efficiency of the proposed algorithms and how they scale with dataset size. +- If applicable, the authors should discuss possible limitations of their approach to address problems of privacy and fairness. +- While the authors might fear that complete honesty about limitations might be used by reviewers as grounds for rejection, a worse outcome might be that reviewers discover limitations that aren't acknowledged in the paper. The authors should use their best judgment and recognize that individual actions in favor of transparency play an important role in developing norms that preserve the integrity of the community. Reviewers will be specifically instructed to not penalize honesty concerning limitations. + +# 3. Theory assumptions and proofs + +Question: For each theoretical result, does the paper provide the full set of assumptions and a complete (and correct) proof? + +# Answer: [NA] + +Justification: The paper does not contain theoretical results. + +# Guidelines: + +- The answer NA means that the paper does not include theoretical results. +- All the theorems, formulas, and proofs in the paper should be numbered and cross-referenced. +- All assumptions should be clearly stated or referenced in the statement of any theorems. +- The proofs can either appear in the main paper or the supplemental material, but if they appear in the supplemental material, the authors are encouraged to provide a short proof sketch to provide intuition. +- Inversely, any informal proof provided in the core of the paper should be complemented by formal proofs provided in appendix or supplemental material. +- Theorems and Lemmas that the proof relies upon should be properly referenced. + +# 4. Experimental result reproducibility + +Question: Does the paper fully disclose all the information needed to reproduce the main experimental results of the paper to the extent that it affects the main claims and/or conclusions of the paper (regardless of whether the code and data are provided or not)? + +# Answer: [Yes] + +Justification: All experiments descriptions in the paper follow the same structure, discussing datasets, baselines, metrics and analysing the results with provided hyperparameter settings and data sampling methodology. In Appendix E, we additionally discuss data preparation procedures. + +# Guidelines: + +- The answer NA means that the paper does not include experiments. +- If the paper includes experiments, a No answer to this question will not be perceived well by the reviewers: Making the paper reproducible is important, regardless of whether the code and data are provided or not. +- If the contribution is a dataset and/or model, the authors should describe the steps taken to make their results reproducible or verifiable. +- Depending on the contribution, reproducibility can be accomplished in various ways. For example, if the contribution is a novel architecture, describing the architecture fully might suffice, or if the contribution is a specific model and empirical evaluation, it may be necessary to either make it possible for others to replicate the model with the same dataset, or provide access to the model. In general, releasing code and data is often one good way to accomplish this, but reproducibility can also be provided via detailed instructions for how to replicate the results, access to a hosted model (e.g., in the case of a large language model), releasing of a model checkpoint, or other means that are appropriate to the research performed. +- While NeurIPS does not require releasing code, the conference does require all submissions to provide some reasonable avenue for reproducibility, which may depend on the nature of the contribution. For example +(a) If the contribution is primarily a new algorithm, the paper should make it clear how to reproduce that algorithm. +(b) If the contribution is primarily a new model architecture, the paper should describe the architecture clearly and fully. +(c) If the contribution is a new model (e.g., a large language model), then there should either be a way to access this model for reproducing the results or a way to reproduce the model (e.g., with an open-source dataset or instructions for how to construct the dataset). +(d) We recognize that reproducibility may be tricky in some cases, in which case authors are welcome to describe the particular way they provide for reproducibility. In the case of closed-source models, it may be that access to the model is limited in some way (e.g., to registered users), but it should be possible for other researchers to have some path to reproducing or verifying the results. + +# 5. Open access to data and code + +Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material? + +Answer: [Yes] + +Justification: The code is provided with the supplementary material, the used datasets are public. + +Guidelines: + +- The answer NA means that paper does not include experiments requiring code. +- Please see the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details. +- While we encourage the release of code and data, we understand that this might not be possible, so "No" is an acceptable answer. Papers cannot be rejected simply for not including code, unless this is central to the contribution (e.g., for a new open-source benchmark). +- The instructions should contain the exact command and environment needed to run to reproduce the results. See the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details. +- The authors should provide instructions on data access and preparation, including how to access the raw data, preprocessed data, intermediate data, and generated data, etc. +- The authors should provide scripts to reproduce all experimental results for the new proposed method and baselines. If only a subset of experiments are reproducible, they should state which ones are omitted from the script and why. +- At submission time, to preserve anonymity, the authors should release anonymized versions (if applicable). +- Providing as much information as possible in supplemental material (appended to the paper) is recommended, but including URLs to data and code is permitted. + +# 6. Experimental setting/details + +Question: Does the paper specify all the training and test details (e.g., data splits, hyperparameters, how they were chosen, type of optimizer, etc.) necessary to understand the results? + +Answer: [Yes] + +Justification: The hyperparameters and data splits are provided in section 4 and Appendix F. Hyperparameter choice justification is discussed. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The experimental setting should be presented in the core of the paper to a level of detail that is necessary to appreciate the results and make sense of them. +- The full details can be provided either with the code, in appendix, or as supplemental material. + +# 7. Experiment statistical significance + +Question: Does the paper report error bars suitably and correctly defined or other appropriate information about the statistical significance of the experiments? + +Answer: [No] + +Justification: It is infeasible to report statistical significance for all compared methods due to the amount of compute required and available. The observed variance in the conducted experiments was negligible. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The authors should answer "Yes" if the results are accompanied by error bars, confidence intervals, or statistical significance tests, at least for the experiments that support the main claims of the paper. + +- The factors of variability that the error bars are capturing should be clearly stated (for example, train/test split, initialization, random drawing of some parameter, or overall run with given experimental conditions). +- The method for calculating the error bars should be explained (closed form formula call to a library function, bootstrap, etc.) +- The assumptions made should be given (e.g., Normally distributed errors). +- It should be clear whether the error bar is the standard deviation or the standard error of the mean. +- It is OK to report 1-sigma error bars, but one should state it. The authors should preferably report a 2-sigma error bar than state that they have a $96\%$ CI, if the hypothesis of Normality of errors is not verified. +- For asymmetric distributions, the authors should be careful not to show in tables or figures symmetric error bars that would yield results that are out of range (e.g. negative error rates). +- If error bars are reported in tables or plots, The authors should explain in the text how they were calculated and reference the corresponding figures or tables in the text. + +# 8. Experiments compute resources + +Question: For each experiment, does the paper provide sufficient information on the computer resources (type of compute workers, memory, time of execution) needed to reproduce the experiments? + +Answer: [Yes] + +Justification: Information is provided in Appendix F. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The paper should indicate the type of compute workers CPU or GPU, internal cluster or cloud provider, including relevant memory and storage. +- The paper should provide the amount of compute required for each of the individual experimental runs as well as estimate the total compute. +- The paper should disclose whether the full research project required more computer than the experiments reported in the paper (e.g., preliminary or failed experiments that didn't make it into the paper). + +# 9. Code of ethics + +Question: Does the research conducted in the paper conform, in every respect, with the NeurIPS Code of Ethics https://neurips.cc/public/EthicsGuidelines? + +Answer: [Yes] + +Justification: The authors reviewed and followed the NeurIPS Code of Ethics. In particular, used datasets are anonymized. + +Guidelines: + +The answer NA means that the authors have not reviewed the NeurIPS Code of Ethics +- If the authors answer No, they should explain the special circumstances that require a deviation from the Code of Ethics. +- The authors should make sure to preserve anonymity (e.g., if there is a special consideration due to laws or regulations in their jurisdiction). + +# 10. Broader impacts + +Question: Does the paper discuss both potential positive societal impacts and negative societal impacts of the work performed? + +Answer: [NA] + +Justification: The method improves the quality and efficiency in the tasks of Novel View Synthesis and planar mesh extraction – however, these improvements do not introduce conceptually new capabilities that would require revision of societal impact when compared to prior work. + +# Guidelines: + +- The answer NA means that there is no societal impact of the work performed. +- If the authors answer NA or No, they should explain why their work has no societal impact or why the paper does not address societal impact. +- Examples of negative societal impacts include potential malicious or unintended uses (e.g., disinformation, generating fake profiles, surveillance), fairness considerations (e.g., deployment of technologies that could make decisions that unfairly impact specific groups), privacy considerations, and security considerations. +- The conference expects that many papers will be foundational research and not tied to particular applications, let alone deployments. However, if there is a direct path to any negative applications, the authors should point it out. For example, it is legitimate to point out that an improvement in the quality of generative models could be used to generate deepfakes for disinformation. On the other hand, it is not needed to point out that a generic algorithm for optimizing neural networks could enable people to train models that generate Deepfakes faster. +- The authors should consider possible harms that could arise when the technology is being used as intended and functioning correctly, harms that could arise when the technology is being used as intended but gives incorrect results, and harms following from (intentional or unintentional) misuse of the technology. +- If there are negative societal impacts, the authors could also discuss possible mitigation strategies (e.g., gated release of models, providing defenses in addition to attacks, mechanisms for monitoring misuse, mechanisms to monitor how a system learns from feedback over time, improving the efficiency and accessibility of ML). + +# 11. Safeguards + +Question: Does the paper describe safeguards that have been put in place for responsible release of data or models that have a high risk for misuse (e.g., pretrained language models, image generators, or scraped datasets)? + +Answer: [NA] + +Justification: The paper poses no such risks. + +# Guidelines: + +- The answer NA means that the paper poses no such risks. +- Released models that have a high risk for misuse or dual-use should be released with necessary safeguards to allow for controlled use of the model, for example by requiring that users adhere to usage guidelines or restrictions to access the model or implementing safety filters. +- Datasets that have been scraped from the Internet could pose safety risks. The authors should describe how they avoided releasing unsafe images. +- We recognize that providing effective safeguards is challenging, and many papers do not require this, but we encourage authors to take this into account and make a best faith effort. + +# 12. Licenses for existing assets + +Question: Are the creators or original owners of assets (e.g., code, data, models), used in the paper, properly credited and are the license and terms of use explicitly mentioned and properly respected? + +Answer: [Yes] + +Justification: Used datasets are credited with licences ([31] and [32]), competing methods are cited after the method short name in every table, and the citation includes code access links crediting license. Used codebases are credited in Appendix E. + +# Guidelines: + +- The answer NA means that the paper does not use existing assets. +- The authors should cite the original paper that produced the code package or dataset. +- The authors should state which version of the asset is used and, if possible, include a URL. + +- The name of the license (e.g., CC-BY 4.0) should be included for each asset. +- For scraped data from a particular source (e.g., website), the copyright and terms of service of that source should be provided. +- If assets are released, the license, copyright information, and terms of use in the package should be provided. For popular datasets, paperswithcode.com/datasets has curated licenses for some datasets. Their licensing guide can help determine the license of a dataset. +- For existing datasets that are re-packaged, both the original license and the license of the derived asset (if it has changed) should be provided. +- If this information is not available online, the authors are encouraged to reach out to the asset's creators. + +# 13. New assets + +Question: Are new assets introduced in the paper well documented and is the documentation provided alongside the assets? + +Answer: [Yes] + +Justification: The code contains README file with all the information to reproduce the paper results. + +Guidelines: + +- The answer NA means that the paper does not release new assets. +- Researchers should communicate the details of the dataset/code/model as part of their submissions via structured templates. This includes details about training, license, limitations, etc. +- The paper should discuss whether and how consent was obtained from people whose asset is used. +- At submission time, remember to anonymize your assets (if applicable). You can either create an anonymized URL or include an anonymized zip file. + +# 14. Crowdsourcing and research with human subjects + +Question: For crowdsourcing experiments and research with human subjects, does the paper include the full text of instructions given to participants and screenshots, if applicable, as well as details about compensation (if any)? + +Answer: [NA] + +Justification: The paper does not involve crowdsourcing nor research with human subjects. + +Guidelines: + +- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. +- Including this information in the supplemental material is fine, but if the main contribution of the paper involves human subjects, then as much detail as possible should be included in the main paper. +- According to the NeurIPS Code of Ethics, workers involved in data collection, curation, or other labor should be paid at least the minimum wage in the country of the data collector. + +# 15. Institutional review board (IRB) approvals or equivalent for research with human subjects + +Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals (or an equivalent approval/review based on the requirements of your country or institution) were obtained? + +Answer: [NA] + +Justification: The paper does not involve crowdsourcing nor research with human subjects. + +Guidelines: + +- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. + +- We recognize that the procedures for this may vary significantly between institutions and locations, and we expect authors to adhere to the NeurIPS Code of Ethics and the guidelines for their institution. +- For initial submissions, do not include any information that would break anonymity (if applicable), such as the institution conducting the review. + +# 16. Declaration of LLM usage + +Question: Does the paper describe the usage of LLMs if it is an important, original, or non-standard component of the core methods in this research? Note that if the LLM is used only for writing, editing, or formatting purposes and does not impact the core methodology, scientific rigorousness, or originality of the research, declaration is not required. + +Answer: [NA] + +Justification: LLM policy is reflected in OpenReview submission. The paper does not involve LLMs as any important, original, or non-standard components. + +Guidelines: + +- The answer NA means that the core method development in this research does not involve LLMs as any important, original, or non-standard components. +- Please refer to our LLM policy (https://neurips.cc/Conferences/2025/LLM) for what should or should not be described. + +# A Full mesh extraction results - Figures 8 to 10 + +We evaluate our hybrid representation on the task of full mesh extraction using the method from [7], we do it in addition to the planar-only mesh extraction experiments presented in Section 4.2, concatenating the two meshes together and comparing them to common benchmarks from Section 4.1. + +Datasets We evaluate on ScanNet++ [31], a common indoor scene benchmark, as well as on subset of suitable indoor/outdoor scenes from ETH3D [34], which provides high quality mesh, and is more challenging because of sparse image supervision. + +Baselines For ScanNet++ we reuse the models trained on iPhone data stream and evaluated on the task of NVS in Section 4.1 to access mesh quality reconstruction. On ETH3D, in addition, we evaluate Gaussian Opacity Fields (GOF) [35], an extension of 2DGS for higher quality mesh reconstruction, and DNSplatter [36], a method supervising 3DGS with mono-depth + +To obtain the mesh, we use TSDF fusion with the median depth estimate for 3DGS, 2DGS, DNSplatter and ours, rather than the expected ray termination as in default settings (i.e., average depth). For PGSR we use their proposed unbiased depth computation, and for Gaussian Opacity Fields we extract the mesh using the level set of the Gaussians, hence the mesh is not colored. + +Metrics We use the same metrics as for meshing task in planar mesh experiments Section 4.2. We compute the F1-score at $5\mathrm{cm}$ threshold. For both of the datasets, we use every 8th image as a test image. + +Analysis We provide full mesh renders along with the metrics on ScanNet++ in Figure 8. For ETH3D, in addition to mesh renders in Figure 10, we provide rendered novel views from the test set in Figure 9. Note that captured planar surfaces are unbiased and outline well the structures of the scenes. Moreover, on in the sparse view setting on ETH3D dataset we achieve a notable rendering quality improvement. + +# B Additional ablations - Tables 1 and 2 + +Random initialization We analyze the effect of having sparse point cloud initialization versus random initialization in our method on 11 DSLR scenes from ScanNet++ [31]. for random initialization we do 5000 iterations in our warmup stage, as opposed to the usual 3500. We show that our method maintains the robustness to random initialization similar to 3DGS-MCMC [13], and despite a drop in number of planar Gaussians, it achieves comparable depth and image quality metrics to our method when initialized with SfM sparse point cloud. + +Table 1: Ablation on initialization - Our method is robust to random initialization and achieves comparable performance to when initialized with SfM point cloud. + +
MethodPSNR↑SSIM↑LPIPS↓RMSE↓MAE↓AbsRel↓#primitives(%planar)
3DGS-MCMC (SfM)23.380.870.240.410.240.261.13M
Ours (SfM)23.420.870.240.200.130.121.13M(31%)
Ours (Random)23.300.860.250.210.140.131.13M(21%)
+ +Full metrics set for ablation on design choices We provide the full set of metrics for ablation on design choices (described in section 4.3) in the table 2. + +![](images/2fe147556389cd44b515177b05d824a6770b20c847fbd4e5121a9e4744d2b6ab.jpg) +Figure 8: Full Mesh Extraction Results on ScanNet++ - Out method achieves competitive performance for surface reconstruction, while mainating the rendering quality. Checkered surfaces indicate different planes, planes are usually behind the TSDF-extracted mesh as they represent unbiased surfaces. Some of the meshes are shown from outside of the indoor scene to highlight the planar alignment. + +# C Additional video and 3D mesh results + +We provide video renderings of RGB and depth for our method compared to baselines in https://theialab.github.io/3dgs-flats. Video results best capture the significant enhancement of our approach over baselines in depth estimation and accurately modeling scene geometry. + +![](images/bbb80f3129b80804cbc61e8a67ecc3c4ec9b3f6f301745d46b67b7653f2c7b10.jpg) +Figure 9: Rendering Results on ETH3D Scenes - Our method outperforms the baselines in terms of rendering quality on this set of sparse view outdoor/indoor scenes, and the planar representation is crucial for achieving good novel view synthesis in sparse scenarios. + +
MethodElectroTerraceDelivery area
PSNR↑LPIPS↓SSIM↑PSNR↑LPIPS↓SSIM↑PSNR↑LPIPS↓SSIM↑
3DGS16.450.380.7220.770.270.7819.480.290.83
2DGS16.400.410.7220.820.290.7919.260.350.81
GOF17.340.360.7120.800.270.7519.400.330.79
PGSR------16.640.410.69
DNSplatter------19.560.240.77
Ours18.720.310.7522.570.220.8122.560.210.87
+ +Table 2: Ablation on design choices – Loss components and optimization strategy are critical, with simultaneous plane-Gaussian optimization causing significant drops. 2D Gaussian snapping greatly improves depth accuracy compared to regularization alternatives. Similarly, Gaussian relocation is essential. + +
Full modelPSNR↑ 26.83LPIPS↓ 0.27SSIM↑ 0.86RMSE↓ 0.25MAE↓ 0.18AbsRel↓ 0.09
Loss design:
w/o LTV23.240.340.820.340.240.13
w/o Lmask24.020.320.830.620.530.29
Optimization design:
w/o plane optimization21.080.370.800.540.430.24
simult. joint optimization19.520.380.790.400.320.18
2D Gaussian design:
w/o snapping25.530.310.840.380.310.17
reg. w/o snapping21.690.350.810.360.280.15
w/o relocation20.000.370.800.590.500.28
+ +# D Additional qualitative results – Figures 11 and 12 + +We provide more qualitative evidence for the performance of our method compared to 2DGS [7], 3DGS [6] and 3DGS-MCMC [13] baselines on the ScanNet++ [31] dataset in figure 11. The results show how baselines particularly struggle with reconstructing accurate geometry for the textureless areas while our method significantly improves upon these methods in depth estimation and keeps the visual quality of images. + +![](images/bd6d5b5dad160aea27bbaa0f83306d4d0c26b2631d6dd41d6dabb7b645cad87a.jpg) +Figure 10: Full Mesh Extraction Results on ETH3D Scenes - Our method outperforms the baselines. + +Further, we provide more visualization for our estimated planes on ScanNet++ [31] dataset, showcasing the perfect alignment of our planes with the detected planar surfaces in figure 12. + +# E Input planar masks + +2D semantic masks Our method relies on input consistent 2D segmentation masks of planar surfaces. To obtain these masks, we can either annotate each image collection manually or automate the process for larger scenes. To automatically obtain the 2D segmentation masks, we employ PlaneRecNet [25] and SAMv2 video segmentation model [29], to create an annotation pipeline. We first input images to PlaneRecNet to obtain 2D plane annotations that are not semantically consistent across the image collection. We set the plane probability threshold to 0.5. While this method works well on iPhone images, it produces fewer plane annotations for DSLR images, that are out of distribution for its network trained on iPhone data. We then input these unmatched masks as seed to SAMv2. In order to do that, we order image collections that are not already sampled from a video. We propagate masks from the initial frame in 16-frame chunks of the sequence to the next 15 frames, and match SAMv2's prediction with any subsequent 2D masks output from [25], using Hungarian matching with an IoU metric. Although largely effective, this process is prone to error accumulation through mask propagation. However, we assume resultant masks are semantically consistent across the image sequence. We provide sample segmentation of an input sequence in the supplementary video and on the website. + +![](images/e93c323897421e47e1c9cb4915acf586e7e358fa80f35b26b62aaa70321d2a7d.jpg) +Figure 11: Novel view synthesis and depth - Qualitative results on ScanNet++ iPhone dataset show our superior performance in both image quality and depth estimation in novel views. Note the limitation of the quality of Gaussian Splatting based methods for textureless surfaces, which makes the plane fitting procedure less robust. + +![](images/b8a3011ae559e73026c13a707ba225e7f70e485a8d83bab3d490999ccbd178fd.jpg) + +![](images/a426876e9a19668e325ebede1cbf816f7a0ee4913aa00acef958681e2b458a97.jpg) + +![](images/48facd09a9d8286d51b1c8af924394013d06f50fc1894b378e67a66a098a42a3.jpg) + +![](images/51f9e7099796fbcd21ff7959af1c6d9a093d820d4288c6d1c1377b2c58fb35cc.jpg) +Figure 12: We provide visualizations of our output planes on the rendered test views of ScanNet++ DSLR streams (top 2 rows) and iPhone stream (bottom 2 rows). Pink markings are due to the anonymization of the original ScanNet++ dataset. While some planar surfaces are missed due to lack of manual 2D planar mask annotation, the captured planes are reconstructed faithfully. + +![](images/73146067a32b82a59dd491b90452f730a4ab8d097f75e9dbd7603350ffd4b7a0.jpg) + +![](images/f40691df9e5c9b4308dc0bfadea0fecc14845b7b3ebbd0646707280dc5d07e01.jpg) + +![](images/2ea0bde79ff6eda0afec9f18c78c0006f9aa6b8443fb638e57c49a8db715ef90.jpg) + +![](images/b003bcd915e35c1a7366587a974e04129c29d7e92195ae240ce216ad64a679ea.jpg) + +![](images/0e3e095753983cdfa9fabd0e05f4b0525f2c42e90dc3baac21d418064b398227.jpg) + +![](images/c290cce50202fecd6503308cda8c2a02d782e8b93fda21a84d57ff40cb67b8fa.jpg) + +![](images/39942705b731ea470adf10002e92a84e1f7d6660d65d2f34d65969181b098b0a.jpg) + +![](images/ef17a18cb9fe4ec8791c817d854f542c6f70f45d10f680d38a9bce3c8a319979.jpg) + +Masked ground truth meshes For the planar mesh extraction task, we only consider planes with annotated segmentation masks from the ground truth mesh, as the 2D plane segmentation task is orthogonal to our method. To identify the relevant subset of planes, we unproject points from the ground truth depth maps that correspond to each plane according to its segmentation mask. We then fit a plane to each resulting point cloud using RANSAC and compile these fitted planes into a set $S$ . We match planes from the ground truth mesh to those in set S by applying two criteria: the normal cosine distance must be less than 0.99 to at least one plane in $S$ , and the distance between their closest points must be less than 0.1. Doing this allows for computational efficiency and increased robustness to missing or undersegmented planes in the input 2D annotations. + +Code We release our code3 publicly for reproducibility purposes and to facilitate future research in this area. We base our code on the 3DGS-MCMC paper [13] and additionally use SAMv2 [29], and PlaneRecNet [25] to generate masks. The baselines are evaluated using their official released code [7, 6, 16, 17, 13, 8, 9]. We further utilize AirPlanes [9] code to compute meshing metrics. + +# F Hyperparameters settings + +We use $\sigma_{\perp}$ and $\sigma_{\parallel}$ as hyper-parameters that control the stochastic re-location. These parameters are chosen depending on the metric scale of the dataset, and are defined in millimeters. For both datasets we used $\sigma_{\perp} = 0.01$ and $\sigma_{\parallel} = 0.3$ . We observe that setting $\lambda_{\mathrm{mask}} = 0.1$ , yields best results empirically. For regularizers, we use $\lambda_{\mathrm{TV}} = 0.1$ , $\lambda_{\mathrm{scale}} = 0.01$ and $\lambda_{\mathrm{opacity}} = 0.01$ following [10] and [13]. We use the same scheduling policy for learning plane origin and normal (rotation) as for the Gaussian means the vanilla 3DGS. All experiments were conducted on a single A6000 ADA GPU, with 46GB memory. The method runs for approximately 1 hour for a single ScanNet++/ScanNetV2 scene, which is comparable to PGSR [16], the second best method for geometric quality according to our experiments and $1.5\times$ longer than 3DGS-MCMC [13], the best method for Novel View Synthesis. The training time is increased due to the RANSAC overhead and block-coordinate descent optimization of planar parameters. Additionally, mesh extraction takes $\sim 3$ minutes and SAM mask propagation is on average 7 minutes long, depending on the scene type. We believe that the training time can be reduced in future work with addition of customized CUDA kernels. \ No newline at end of file diff --git a/NeurIPS/2025/3D Gaussian Flats_ Hybrid 2D_3D Photometric Scene Reconstruction/images.zip b/NeurIPS/2025/3D Gaussian Flats_ Hybrid 2D_3D Photometric Scene Reconstruction/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..8e13fc1e54286e41c4f5222c0a6b8d41e05a6f84 --- /dev/null +++ b/NeurIPS/2025/3D Gaussian Flats_ Hybrid 2D_3D Photometric Scene Reconstruction/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f027b178e2fe26314d8ba393ff1ac3a77429d08ddbafbd7015711865cf41e820 +size 1849089 diff --git a/NeurIPS/2025/3D Gaussian Flats_ Hybrid 2D_3D Photometric Scene Reconstruction/layout.json b/NeurIPS/2025/3D Gaussian Flats_ Hybrid 2D_3D Photometric Scene Reconstruction/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..11e55b2a293ac1d17ac5ff56e554fbd31ccd4f10 --- /dev/null +++ b/NeurIPS/2025/3D Gaussian Flats_ Hybrid 2D_3D Photometric Scene Reconstruction/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d98cccd3594fcf746dedd947d5037455fc15eb148c7584981fd78b34f7c427d6 +size 664511 diff --git a/NeurIPS/2025/3D Gaussian Splatting based Scene-independent Relocalization with Unidirectional and Bidirectional Feature Fusion/64fceee8-48f9-4972-bfcf-8de3c8fc2ec6_content_list.json b/NeurIPS/2025/3D Gaussian Splatting based Scene-independent Relocalization with Unidirectional and Bidirectional Feature Fusion/64fceee8-48f9-4972-bfcf-8de3c8fc2ec6_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..e27a48a9d1793d87fc1e39bfb89bfb92822ec1f6 --- /dev/null +++ b/NeurIPS/2025/3D Gaussian Splatting based Scene-independent Relocalization with Unidirectional and Bidirectional Feature Fusion/64fceee8-48f9-4972-bfcf-8de3c8fc2ec6_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b7c4ed0e285d4dfba7e9423fa09b2f1c0694163d10f71dab56e7ec64078947b1 +size 156307 diff --git a/NeurIPS/2025/3D Gaussian Splatting based Scene-independent Relocalization with Unidirectional and Bidirectional Feature Fusion/64fceee8-48f9-4972-bfcf-8de3c8fc2ec6_model.json b/NeurIPS/2025/3D Gaussian Splatting based Scene-independent Relocalization with Unidirectional and Bidirectional Feature Fusion/64fceee8-48f9-4972-bfcf-8de3c8fc2ec6_model.json new file mode 100644 index 0000000000000000000000000000000000000000..0b78626c0ae1476ceb0bc2db9129d463093900a4 --- /dev/null +++ b/NeurIPS/2025/3D Gaussian Splatting based Scene-independent Relocalization with Unidirectional and Bidirectional Feature Fusion/64fceee8-48f9-4972-bfcf-8de3c8fc2ec6_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c58fbdfa76cfd6c9315086077096d1125e4e320d2055db0c7fc11114b3b7e3fb +size 207517 diff --git a/NeurIPS/2025/3D Gaussian Splatting based Scene-independent Relocalization with Unidirectional and Bidirectional Feature Fusion/64fceee8-48f9-4972-bfcf-8de3c8fc2ec6_origin.pdf b/NeurIPS/2025/3D Gaussian Splatting based Scene-independent Relocalization with Unidirectional and Bidirectional Feature Fusion/64fceee8-48f9-4972-bfcf-8de3c8fc2ec6_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..9c645501d256cf93773ef6401ee5cdea1f81ea08 --- /dev/null +++ b/NeurIPS/2025/3D Gaussian Splatting based Scene-independent Relocalization with Unidirectional and Bidirectional Feature Fusion/64fceee8-48f9-4972-bfcf-8de3c8fc2ec6_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a60f221826db3bd766e05358b969ce2bb9a060927e72c582c68172edd86c4167 +size 13754186 diff --git a/NeurIPS/2025/3D Gaussian Splatting based Scene-independent Relocalization with Unidirectional and Bidirectional Feature Fusion/full.md b/NeurIPS/2025/3D Gaussian Splatting based Scene-independent Relocalization with Unidirectional and Bidirectional Feature Fusion/full.md new file mode 100644 index 0000000000000000000000000000000000000000..3433ea845a8c86887434130d8ae8b67f5d345ac8 --- /dev/null +++ b/NeurIPS/2025/3D Gaussian Splatting based Scene-independent Relocalization with Unidirectional and Bidirectional Feature Fusion/full.md @@ -0,0 +1,705 @@ +# 3D Gaussian Splating based Scene-independent Camera Relocalization with Unidirectional and Bidirectional Feature Fusion + +Junyi Wang1,2 + +junyiwang@sdu.edu.cn + +Yuze Wang + +wangyuze19980709@163.com + +Wantong Duan + +wantongd@buaa.edu.cn + +Meng Wang2 + +wangm05@buaa.edu.cn + +Yue Qi $^{2,3*}$ + +qy@buaa.edu.cn + +# Abstract + +Visual localization is a critical component across various domains. The recent emergence of novel scene representations, such as 3D Gaussian Splatting (3D GS), introduces new opportunities for advancing localization pipelines. In this paper, we propose a novel 3D GS-based framework for RGB based, scene-independent camera relocalization, with three main contributions. First, we design a two-stage pipeline with fully exploiting 3D GS. The pipeline consists of an initial stage, which utilizes 2D-3D correspondences between image pixels and 3D Gaussians, followed by pose refinement using the rendered image by 3D GS. Second, we introduce a 3D GS based Relocalization Network, termed GS-RelocNet, to establish correspondences for initial camera pose estimation. Additionally, we present a refinement network that further optimizes the camera pose. Third, we propose a unidirectional 2D-3D feature fusion module and a bidirectional image feature fusion module, integrated into GS-RelocNet and the refinement network, respectively, to enhance feature sharing across the two stages. Experimental results on public 7 Scenes, Cambridge Landmarks, TUM RGB-D and Bonn demonstrate state-of-the-art performance. Furthermore, the beneficial effects of the two feature fusion modules and pose refinement are also highlighted. In summary, we believe that the proposed framework can be a novel universal localization pipeline for further research. + +# 1 Introduction + +Visual localization is considered a fundamental research problem in computer vision and is applied across a variety of scenarios, including Augmented Reality (AR), Mixed Reality (MR), robotics, and autonomous driving Wang et al. (2024c); Jia et al. (2024); Zhu et al. (2024). The primary function of a localization algorithm is to estimate the 6-DoF (Degrees of Freedom) camera pose within a target environment. + +As current works cover, two main types of methods are investigated to achieve robust localization performance, including feature matching and geometry regression. Specially, feature matching methods employ either hand-crafted Liu et al. (2017) or learned features Sun et al. (2021) to establish pixel correspondences for localization, while geometry regression approaches train the deep network + +![](images/1c71929d9474762a821a40556dd448bcf4bd4e6dccecf711ebd8c94b004a3100.jpg) +Figure 1: (a) Localization pipeline. The framework predicts the initial pose by establishing 2D-3D correspondences between pixels and 3D Gaussians, followed by a pose refinement process to optimize pose by rendered views using 3D GS. (b) Localization performance comparison. Our method obtains the least localization error on the two datasets, while supporting scene-independent localization. + +![](images/19fc8bb9bbd62e1bb4adde073a591765fdaa5c1275db18dba147b937417c229c.jpg) +(a) + +![](images/3f2665145052f773d3bb53657600bfde81c2c66795b3ed4ee8219082cde888f3.jpg) + +![](images/d7c19af70b2a74bdb4940b79afe61251d46f6f88a0d1bf427a6088a36ed64c22.jpg) + +![](images/554e373016e85ee1a2e3f45f135572fa1b7b7dfd2940020fe648a50c099d503b.jpg) +(b) + +to solve the camera pose, by Absolute Pose Regression (APR) Chen et al. (2022) or Scene Coordinate Regression (SCR) Wang & Qi (2023b). However, due to the representation, these methods always ignore texture and illumination information, which limits their capacity to fully represent the scene. + +Recent neural and geometric 3D structures for Novel View Synthesis (NVS) have gained significant popularity for scene representat in Mildenhall et al. (2021); Kerbl et al. (2023); Hu et al. (2023). Specially, 3D Gaussian Splatting (3D GS) offers a well-balanced approach to training and rendering performance for NVS, presenting new opportunities for the localization pipeline. However, how to leverage 3D GS for robust and accurate localization remains a significant challenge. + +Current methods predominantly utilize 3D GS for pose refinement Keetha et al. (2024); Yan et al. (2024); Liu et al. (2025), which heavily relies on the accuracy of initial camera pose estimation. When the initial pose estimation is inaccurate or fails, the refinement process becomes ineffective. To address this, our motivation lies in the employment of 3D GS for both pose initialization and refinement, while achieving scene-independent relocalization to enhance robustness. + +As illustrated in Fig. 1(a), we introduce a novel relocalization framework that first establishes 2D-3D correspondences between image pixels and 3D Gaussians, followed by a refinement stage that predicts the relative pose between real and rendered views using 3D GS. A feature fusion module is incorporated in both stages to enhance correspondence regression. To the best of our knowledge, this is the first 3D GS based, scene-independent relocalization framework, offering a robust solution for challenging localization tasks. Specially, the term "scene-independent" indicates that our framework can achieve robust relocalization in a target scene without requiring scene-specific pre-training, in contrast to most "scene-dependent" methods that necessitate prior training on the target scene. Our contributions are summarized as follows. + +1. We propose a innovative framework for scene-independent camera relocalization. The framework comprises initial pose estimation by establishing correspondences between pixels of the input image and the scene expressed by the 3D Gaussians, and pose refinement by predicting the relative pose between the input image and rendered view using 3D GS. +2. We design GS-RelocNet for predicting the correspondences to obtain the initial camera pose. Within GS-RelocNet, we introduce a unidirectional feature fusion module to merge geometry and texture features for learning confidence scores between each pixel and 3D Gaussian. +3. We propose a pose refinement network based on 3D GS. In the refinement network, we present a bidirectional feature fusion module to combine features from rendered and real images. + +To validate the performance of the framework, we conduct experiments on 7 Scenes and Cambridge Landmarks. As illustrated in Fig. 1(b), the results demonstrate the state-of-the-art localization performance on the two datasets. Meanwhile, our framework can support scene-independent relocalization, denoting that it can perform relocalization in unseen scenes. + +# 2 Related Works + +# 2.1 Localization with Feature Matching + +Traditional hand-crafted feature matching based methods typically follow a pipeline consisting of feature extraction, matching, global map construction, and optimization Liu et al. (2017); Sattler et al. (2017). Moreover, semantic SLAM systems Yang & Scherer (2019); You et al. (2023); Lin et al. (2024b); Xi et al. (2025); Zhang et al. (2025) incorporate semantic information derived from learned features to enhance the robustness and accuracy of traditional hand-crafted feature based processes. Alternatively, learned feature matching methods aim to estimation pixel correspondences for pose estimation Wang et al. (2022, 2024d). Based on LoFTR Sun et al. (2021), Efficient LoFTR Wang et al. (2024d) performed the transformer with an aggregated attention mechanism using adaptive token selection for efficiency. + +# 2.2 Localization with Geometry Regression + +The geometry regression methods can be broadly categorized into two approaches, including APR and SCR methods. APR methods train the deep network to learn the relationship between 2D images and 6-DoF camera poses Chen et al. (2022, 2024b). While APR methods offer high computational efficiency, they are often limited by accuracy and generalization issues Liu et al. (2024b). Alternatively, SCR techniques calibrate the pose by using the Kabsch or Perspective-n-Point (PnP) algorithm to the known source and evaluated target coordinate, which achieve considerable localization performance Wang & Qi (2021, 2023b). Recent hot SCR researches are DUSt3R Wang et al. (2024b) and its subsequent extension works Leroy et al. (2024), By using large training samples, the methods achieve outstanding localization accuracy and generalization ability. These methods focus on geometry regression, bug can not fully exploit texture features. + +# 2.3 Localization with Neural Radiance Field (NeRF) and 3D GS + +Due to the outstanding NVS performance, NeRF is applied to the localization pipeline with iterative rendering and pose updates Germain et al. (2022); Moreau et al. (2023); Chen et al. (2024a); Wang et al. (2023); Xu et al. (2024). NeRFect Match Zhou et al. (2024b) explored the potential of NeRF's internal features in establishing precise 2D-3D matches for localization. With the shift in the NVS field from NeRF to 3D GS, STDLoc Huang et al. (2025) introduced a matching-oriented Gaussian sampling strategy and a scene-specific detector to achieve efficient and robust pose estimation. + +# 3 Method + +# 3.1 Overview + +Given a target image and a 3D scene model expressed through 3D GS, our method predicts the 6-DoF camera pose of the target image within the scene. The overall localization process is composed of two stages, both utilizing 3D GS in distinct ways. In the initial pose estimation stage, we establish the correspondences between each image pixel and 3D Gaussian, followed by a PnP algorithm with RANSAC to solve the initial camera pose. Based on the predicted pose, we proceed to the refinement stage, where we first render the view using 3D GS, and then predict the relative pose between the target and rendered images to perform pose optimization. + +The advantages of the proposed pipeline are twofold. First, we use 3D Gaussians to represent the scene in the initial pose estimation stage. Compared to the point cloud representation in SCR methods, 3D Gaussians retain more detailed geometric and texture information. Additionally, by establishing correspondences between image pixels and 3D Gaussians, GS-RelocNet enables scene-independent relocalization. Second, we employ the rendered view generated by 3D GS for pose refinement, which reduces the domain gap between the rendered and real views. Through this refinement process, the localization accuracy is further enhanced. + +# 3.2 Initial Pose Estimation by GS-RelocNet + +Pose estimation process. In the initial stage, we regress the confidence matrix between image pixels and selected Gaussians by GS-RelocNet. Based on the confidence threshold, we establish the correspondences between pixels and Gausaians, followed by a PnP method with RANSAC to solve the initial pose. + +![](images/a2ea3539876280931b24a12c033ab9f4a3e549b05390eb8907a4606ad3757b85.jpg) + +![](images/86d80eca40b0c868720f99a80f175f0c28ce04745d49ba5536323e013ff2ff32.jpg) +Figure 2: Architecture of GS-RelocNet. The GS-RelocNet framework integrates an RGB feature encoder, a 3D Gaussian feature encoder, an RGB descriptor decoder, a point descriptor decoder, and a confidence metric regression module. In the figure, arrows indicate the process flow, and numbers adjacent to each block denote the corresponding filter size. + +![](images/d30fc3e115e7a2e6110391851279e1f7e64fcff0bf5d57b1b43372036b0bae8f.jpg) + +Inputs and outputs of GS-RelocNet. The GS-RelocNet receives a monocular RGB image and a 3D model expressed by 3D Gaussians as inputs, and outputs the confidence scores that quantify the correspondence between pixels and 3D Gaussians. These confidence scores are then used to establish 2D-3D correspondences for camera pose prediction. + +Input processing. For 2D image processing, GS-RelocNet employs patch partition and embedding to segment the images into multiple parts. For the 3D Gaussians, the point cloud serialization and embedding are exploited to transform unstructured 3D Gaussians into a structured format. Specifically, the position, alpha, covariance matrix and spherical harmonic function of 3D Gaussian are processed independently, with the features from all four components concatenated. Additionally, we incorporate features from DINO V2 Oquab et al. (2023), along with a depth estimation head, as supplementary input to enhance geometry feature learning. Notably, the parameters of the DINO V2 model are kept frozen during this process. + +Architecture of GS-RelocNet. The detailed architecture of GS-RelocNet is presented in Fig. 2. It consists of a spatial image encoder, a 3D Gaussian encoder, an image descriptor decoder, a 3D Gaussian descriptor decoder, and a confidence matrix decoder. Within both encoders and decoders, we incorporate a unidirectional feature fusion module to facilitate effective feature sharing from the model to the RGB image. The image branch employs a Swin Transformer Liu et al. (2021) architecture, consisting of multiple Swin Blocks, while the point cloud branch utilizes Point Transformer V3 Wu et al. (2024) for 3D Gaussian feature learning. + +Unidirectional feature fusion module. Between consecutive image and 3D Gaussian learning blocks, we propose a unidirectional feature fusion module to combine 2D and 3D features, shown at the lower left of Fig. 2. The module takes image features and geometry features as inputs, and outputs the fused features of the two parts. Let the input RGB feature have dimensions $H_{u} \times W_{u} \times D_{u}$ , and the model feature have dimensions $N_{m} \times D_{m}$ . The whole fusion process is as follows. + +Step 1, alignment of features. The module aligns the image and 3D model features. Specifically, the RGB feature is reshaped to $N_{m} \times (H_{u} \times W_{u} \times D_{u} / N_{m})$ , followed by a 1D convolution to transform it to $N_{m} \times D_{m}$ , aligning with the model feature dimensions. + +Step 2, fusion with self-attention. In this step, the features from the image and 3D Gaussian are first added. Subsequently, multi-head self-attention is applied to added features and original model features respectively. Finally, the both features are again added to achieve feature fusion. + +Step 3, feature transformation. The combined feature is reshaped to $H_{u} \times W_{u} \times (N_{m} \times D_{m} / H_{u} / W_{u})$ . A Swin Transformer block is then applied to restore the feature dimensions to $H_{u} \times W_{u} \times D_{u}$ , matching the input RGB feature. The input RGB feature is added to this output to produce the merged feature. + +Step 4, iterative fusion. Steps 1 through 3 are iteratively applied. Specifically, the fused features from the current iteration are used as the RGB input for the next iteration. + +Through these steps, GS-RelocNet can perform one-way fusion of 2D image and 3D Gaussian features. Specifically, the model features can influence the image descriptor learning, but not the other way around. On one hand, this feature sharing enhances RGB feature descriptor learning. On the other hand, the independence of model descriptor learning allows for pre-prediction of model descriptors before the inference stage, significantly accelerating the overall process. In summary, we argue that this fusion mechanism contributes to the localization task, and its effectiveness will be validated in the subsequent experiments. + +Confidence matrix regression. After regressing the $N_{i}$ image descriptors and $N_{g}$ 3D Gaussian descriptors, GS-RelocNet performs regression of a confidence matrix with $N_{i} * N_{g}$ scores, where each score represents the confidence between its corresponding image and 3D Gaussian. The regression process proceeds as follows. First, for the image features, GS-RelocNet applies a positional encoding operation followed by a self-attention operation. Similarly, for the 3D Gaussian features, GS-RelocNet also applies a self-attention operation. Next, cross-attention is applied to the two sets of features. Finally, a dual softmax operation is employed to predict the final confidence matrix. + +Loss of GS-RelocNet. To train GS-RelocNet, we need to construct the ground truth confidence between each 3D Gaussian and image pixel. Given a 3D Gaussian with 2D covariance matrix $\Sigma$ under the current view, then the ground truth confidence is calculated as the following formula, + +$$ +C _ {g} = \frac {1}{2 \pi | \Sigma |} \exp \left[ - \frac {1}{2} (x - \mu) ^ {T} \Sigma^ {- 1} (x - \mu) \right], \tag {1} +$$ + +where $\mu$ denotes the 2D Gaussian center, and $x$ is the pixel position. + +# 4 Pose Refinement + +Refinement process. After processing with GS-RelocNet, we obtain the initial camera pose of the target image. Subsequently, we rendered the current view using 3D GS with the predicted camera pose. The refinement process is to predict the relative pose $(R_r^p,T_r^p)$ between the input real image and rendered view. Finally, the refined pose $(R_f^p,T_f^p)$ is obtained by the following formula. + +$$ +R _ {f} ^ {p} = R _ {r} ^ {p} * R _ {i} ^ {p}, \quad T _ {f} ^ {p} = R _ {r} ^ {p} * T _ {i} ^ {p} + T _ {r} ^ {p}. \tag {2} +$$ + +To predict $R_{r}^{p}, T_{r}^{p}$ , we propose a refinement network to regress residual coordinate map proposed in the work Wang & Qi (2023a) between the rendered and real views. + +Refinement network. As illustrated in Fig. 3, the network takes real and rendered views as inputs, and outputs the residual coordinate map, followed by a PnP method with RANSAC to predict the relative camera pose $(R_{r}^{p},T_{r}^{p})$ . The architecture of the refinement network consists of several Swin blocks to learn features of both real and rendered images. + +Bidirectional feature fusion. Similar to GS-RelocNet, we design a feature fusion module to combine the features of real and rendered images. The key distinction is that the feature fusion process is bidirectional, following these steps. First, the two sets of features are added together, and a self-attention operation is applied to fuse the real and rendered features. Second, a Swin block is used to further process and learn the fused features. Finally, the resulting features of both parts are obtained by adding the original features to the fused output. Notably, the three steps outlined above can be repeated multiple times to refine the feature fusion process. + +![](images/c28c6032e665af2c809b2da4286569787fd5c8010784ef6a482a143a657d09ae.jpg) +Figure 3: Architecture of the refinement network. The network takes real and rendered views as inputs, and outputs the residual coordinate map to predict the relative camera pose, which is composed of a real image feature encoder, a rendered image feature encoder and a bidirectional fusion module. + +![](images/74dd8eb7aa3a6c9236dbe57b339fda94249bfbf9eefb8298a18887bc75726174.jpg) + +Loss function of refinement network. For the loss of the residual coordinate $(Loss_{map})$ , we use $L_{1}$ loss to train the refinement network. In addition to the coordinate map loss, the refinement network also exploits the auxiliary loss to facilitate feature learning. Specifically, the pose loss is expressed using the following formula, + +$$ +\operatorname {L o s s} _ {a u x} = \exp (- T) \cdot \| \mathbf {T} ^ {p} - \mathbf {T} ^ {g} \| + T + \exp (- Q) \cdot \| \mathbf {Q} ^ {p} - \mathbf {Q} ^ {q} \| + Q, \tag {3} +$$ + +where $\mathbf{T}^p$ , $\mathbf{T}^g$ represent the prediction and ground truth of camera position, $\mathbf{Q}^p$ , $\mathbf{Q}^g$ mean the prediction and ground truth of camera orientation, and $T$ , $Q$ are variables learned by the refinement network to balance the three terms. + +Although the auxiliary loss directly outputs the 6-DoF relative camera pose, it is not used as the final result in our framework. This decision is based on the observation that direct learning based methods typically yield less accurate localization results Wang et al. (2020). Hence the refinement network uses the map to predict the camera pose, with the auxiliary loss as an additional guidance. + +$$ +L o s s = L o s s _ {m a p} + \alpha * L o s s _ {a u x}. \tag {4} +$$ + +# 4.1 Datasets and Implementation Details + +Datasets and train-test split. We conduct experiments on indoor 7 Scenes Shotton et al. (2013), TUM RGB-D Sturm et al. (2012), Bonn Palazzolo et al. (2019), ScanNet Dai et al. (2017), outdoor MegaDepth Li & Snavely (2018) and Cambridge Landmarks Kendall et al. (2015). The implementation is divided into scene-dependent and scene-independent settings. In scene-independent setting, we train GS-RelocNet on ScanNet and test it on 7 Scenes, TUM RGB-D and Bonn. For outdoor scene-independent setting, GS-RelocNet is trained on MegaDepth Li & Snavely (2018) and tested on Cambridge Landmarks. When performing the relocalization task on in a scene-dependent manner, GS-RelocNet is trained on 7 Scenes and Cambridge Landmarks respectively. + +Gaussian selection. In training and inference stages, 4096 Gaussians are selected by spatially uniform sampling. Specifically, we partition the 3D space into $S_{x} \times S_{y} \times S_{z}$ grids with each grid resolution of $0.1\mathrm{m}$ . Let $N_{t}$ denote the total number of Gaussians. For each grid cell containing $N_{g}$ Gaussians, we randomly sample $N_{g} * 4096 / N_{t}$ Gaussians. If the number of sampled Gaussians is less than 4096, we randomly duplicate some samples to reach the desired count for training. + +GS-RelocNet details. The RGB input of our framework is resized to $256 \times 256$ pixels, while 4096 3D Gaussians spatially uniform sampled from all 3D Gaussians, are processed by the Point Transformer network. In Fig. 2, the module setting and output filter size are indicated near the corresponding modules. Additionally, GS-RelocNet leverages an ADAM W optimizer with learning rates $2 \times 10^{-4}$ . + +Refinement network training details. After obtaining the initial poses, we use 3D GS to render the view. Both the real and rendered images are resized to $128 \times 128$ , while the output size of the relative structure is $64 \times 64$ . In Fig. 3, the module setting and output filter size are also annotated. The loss coefficient $\alpha$ in Eq. (4) is set to 0.3, the variables $T, Q$ are initially set to 0.0. + +Inference details. After training GS-RelocNet, we use the selected Gaussians and the test image to obtain the confidence map. Then we initially apply a fixed threshold (set to 0.7 in our experiments) to eliminate correspondences with similarity scores below this value. Subsequently, for each pixel + +associated with multiple Gaussian correspondences, the pixel's coordinate is computed as a weighted average of the selected Gaussians, weighted by their confidence values. Additionally, if the number of correspondences falls below 100, we re-run GS-RelocNet by selecting an alternative set of 4096 Gaussians from the grids that contain Gaussians with confidence higher than 0.7. After determining the pixel and its weighted Gaussian coordinate, we use PnP with RANSAC to solve the initial pose. + +Based on the initial pose, we render the view by trained 3D GS. Given the input image and rendered view, we predict the residual coordinate map by the refinement network, followed by a PnP method with RANSAC to solve the relative pose between them. Finally, the refined pose is obtained by Eq. 2. + +# 4.2 Static 7 Scenes + +Table 1: Median position (cm) and rotation $(^{\circ})$ errors on 7 Scenes. The sign $\ddagger$ means the result with SfM pseudo ground truth, while others leverage the original KinectFusion ground truth. In each scene, the red and blue marks represent the first and second. + +
MethodChessFireHeadsOfficePumpkinKitchenStairsMean
Scene-dependent
APRMS-Transformer Shavit et al. (2021) (ICCV 2021)11/6.3823/11.513/13.018/8.1417/8.4216/8.9229/10.318/9.51
DFNet Chen et al. (2022) (ECCV 2022)3/1.126/2.304/2.296/1.547/1.927/1.7412/2.636/1.93
Marepo Chen et al. (2024b) (CVPR 2024)2.1/1.242.3/1.391.8/2.032.8/1.263.5/1.484.2/1.715.6/1.673.2/1.54
MS-HyperPose Ferens & Keller (2025) (CVPR 2025)11/4.3423/9.7913/10.717/6.0516/5.2417/6.8627/6.0018/7.00
SCR\( ACE^‡ \) Brachmann et al. (2023) (CVPR 2023)0.55/0.180.83/0.330.53/0.331.0/0.291.1/0.220.77/0.212.89/0.811.1/0.34
DeViLoc Giang et al. (2024) (CVPR 2024)2/0.782/0.741/0.653/0.824/1.023/1.194/1.122.7/1.10
\( GLACE^‡ \) Wang et al. (2024a) (CVPR 2024)0.6/0.180.9/0.340.6/0.341.1/0.290.9/0.230.8/0.203.2/0.931.2/0.36
NeRFCROSSFIRE Moreau et al. (2023) (CVPR 2023)1/0.45/1.93/2.35/1.63/0.82/0.812/1.94/1.10
\( NeRFMatch^‡ \) Zhou et al. (2024a) (ECCV 2024)0.9/0.301.1/0.401.5/1.003/0.802.2/0.601.0/0.3010.1/1.702.8/0.70
PMNet Lin et al. (2024a) (ECCV 2024)4/1.7010/4.517/4.237/1.9614/3.3314/3.3616/3.6210/3.24
3D GSDFNet + GS-CPR‡ Liu et al. (2025) (ICLR 2025)0.7/0.200.9/0.320.6/0.361.2/0.321.3/0.310.9/0.252.2/0.611.1/0.34
\( ACE + GS-CPR‡ \) Liu et al. (2025) (ICLR 2025)0.5/0.150.6/0.250.4/0.280.9/0.261.0/0.230.7/0.171.4/0.420.8/0.25
\( STDLoc^‡ \) Huang et al. (2025) (CVPR 2025)0.46/0.150.57/0.240.45/0.260.86/0.240.93/0.210.63/0.191.42/0.410.76/0.24
\( Ours^‡ \) (No Refinement)0.44/0.170.61/0.240.39/0.300.89/0.240.95/0.280.60/0.221.36/0.390.75/0.26
\( Ours^‡ \)0.41/0.150.55/0.210.37/0.260.85/0.240.92/0.250.58/0.181.30/0.350.71/0.23
Scene-independent
Hand-craftedActive Search Sattler et al. (2016) (TPAMI)4/1.963/1.532/1.459/3.618/3.107/3.373/2.2251/2.46
RPRRelocNet Balntas et al. (2018) (ECCV 2018)21/10.9032/11.8015/13.4031/10.3040/10.9033/10.3033/11.4029.3/11.29
Relative PoseNet Laskar et al. (2017) (ICCV 2017)31/15.0040/19.0024/22.2038/14.1044/18.2041/16.5035/23.6036.1/18.37
SCRInLoc Taira et al. (2018) (CVPR 2018)3/1.053/1.072/1.163/1.055/1.554/1.319/2.474.1/1.38
Pixloc Sarlin et al. (2021) (CVPR 2021)2/0.802/0.731/0.823/0.824/1.213/1.205/1.302.9/0.98
Wang et al. Wang & Qi (2023a) (ISMAR 2023)2.4/0.972.0/0.991.6/1.272.4/1.013.7/1.202.8/1.143.1/1.222.6/1.11
DUSt3R Wang et al. (2024b) (CVPR 2024)3/0.964/1.021/1.004/1.045/1.264/1.3621/4.066/1.53
Reloc3R Dong et al. (2025) (CVPR 2025)3/0.994/1.132/1.235/0.887/1.145/1.2312/1.255.4/1.12
3D GSOurs (No Refinement)1.3/0.811.2/0.650.7/0.731.6/0.892.7/1.012.5/1.102.1/1.001.7/0.88
Ours1.0/0.721.0/0.640.6/0.701.4/0.822.2/0.932.0/1.021.9/0.921.4/0.82
+ +In Table 1, we provide the experimental localization results on 7 Scenes. with both scene-dependent and scene-independent categories using the median position and rotation errors. + +Scene-dependent method comparison. In scene-independent setting, our results gain best accuracy on both mean position and orientation metrics. Among 7 scenes, our method obtains the best position on all 7 scenes and orientation on 5 scenes, demonstrating the state-of-the-art performance under scene-dependent situation. Additionally, our method achieves $1.21\text{cm} /0.61^{\circ}$ with the original ground truth, which also demonstrates superior accuracy compared to other methods evaluated. + +Scene-independent method comparison. With scene-independent approaches, our method achieves the lowest mean position and orientation errors on all 7 scenes. Moreover, our method significantly reduces the position error from other best $2.6\mathrm{cm} / 0.98^{\circ}$ to $1.4\mathrm{cm} / 0.82^{\circ}(\downarrow 46.2\% /16.3\%)$ . Besides the localization accuracy, to our knowledge, our framework is the first scene-independent relocalization method with 3D GS. + +Result visualization. Fig. 4 provides visualization results of the camera trajectories with green poses denoting the ground truth and blue ones representing our results. The results show minimal differences between the predicted and ground truth poses, demonstrating the suitability of our method. + +# 4.3 Dynamic TUM RGB-D and Bonn + +Challenges on TUM RGB-D and Bonn. The TUM RGB-D test sequences involve two individuals walking around a table increase the complexity of localization. Similarly, the Bonn dataset features highly dynamic sequences, such as individuals manipulating boxes or interacting with balloons. + +![](images/cf129eec85e68262e46719dcce06c492ab215075e3f2b0ec6abb31c652da1121.jpg) +Figure 4: The visualization results of camera pose on 7 Scenes. In each scene, the green and blue poses denote the ground truth and prediction respectively. + +Table 2: RMSE of ATE [cm] results in four dynamic scenes of TUM RGB-D. In each scene, the red and blue marks represent the first and second respectively. + +
Methodfr3_walking_Mean
xyzstaticryphalf
Hand-craftedORB-SLAM2 Mur-Artal & Tardós (2017)45.99.365.832.838.5
DGM-VINS Song et al. (2023)3.61.37.13.33.8
Hand-crafted + SemanticsDynaSLAM Bescos et al. (2018)1.50.63.52.52.0
DS-SLAM Yu et al. (2018)2.50.844.43.012.7
LC-CRF SLAM Du et al. (2020)1.61.14.62.82.5
3D GSOurs1.10.42.22.01.4
+ +Root Mean Square Error (RMSE) of Absolute Trajectory Error (ATE) on dynamic TUM RGB-D. Table 2 presents the RMSE of ATE for four dynamic scenes, compared against ORB-SLAM2, DynaSLAM, DS-SLAM, LC-CRF SLAM, and DGM-VINS. Our approach consistently outperforms these state-of-the-art SLAM systems in the RMSE of ATE metric, demonstrating superior localization accuracy in dynamic environments. + +RMSE of ATE on dynamic Bonn. We evaluated our framework on the Bonn dataset across 20 test scenes, consistent with LC-CRF SLAM, and compared it with ReFusion Palazzolo et al. (2019), MaskFusion Runz et al. (2018), and LC-CRF SLAM Du et al. (2020). The mean RMSE of ATE results are $23.8\mathrm{cm}$ (ReFusion), $25.1\mathrm{cm}$ (MaskFusion), $6.8\mathrm{cm}$ (LC-CRF SLAM), and $4.3\mathrm{cm}$ (Ours). Our framework achieves the lowest mean RMSE of ATE, highlighting its exceptional localization accuracy in highly dynamic scenes. + +# 4.4 Cambridge Landmarks + +Table 3: Median localization results on Cambridge Landmarks compared with other methods. Units of position and orientation are centimeter (cm) and $^\circ$ . $\star$ means the result with a scene-independent setting. In each scene, the red and blue marks represent the first and second. + +
MethodCollegeHospitalShopStMaryMean
APRMS-Transformer Shavit et al. (2021) (ICCV 2021)85/1.45175/2.4388/3.20166/4.12129/2.80
DFNet Chen et al. (2022) (ECCV 2022)73/2.37200/2.9867/2.21137/4.02119/2.90
SCRInLoc Taira et al. (2018) (CVPR 2018)46/0.848/1.011/0.518/0.631/0.73
DSAC* Brachmann & Rother (2021) (TPAMI)15/0.321/0.45/0.313/0.414/0.35
ACE Brachmann et al. (2023) (CVPR 2023)29/0.3831/0.615/0.319/0.621/0.47
DUSt3R-224* Wang et al. (2024b) (CVPR 2024)20/0.3226/0.469/0.3811/0.3817/0.39
Reloc3R-224* Dong et al. (2025) (CVPR 2025)47/0.4187/0.6618/0.5341/0.7348/0.58
NeRFNeuMap Tang et al. (2023) (CVPR 2023)14/0.219/0.46/0.317/0.514/0.35
CROSSFIRE Moreau et al. (2023) (ICCV 2023)47/0.743/0.720/1.239/1.437/1.00
NeRFMatch Zhou et al. (2024b) (ECCV 2024)12.7/0.220.7/0.48.7/0.411.3/0.40.13/0.35
PMNet Lin et al. (2024a) (ECCV 2024)68/1.97103/1.3158/2.10133/3.7391/2.28
3D GaussianDFNet + GS-CPR Liu et al. (2025) (ICLR 2025)26/0.3448/0.7210/0.3627/0.6228/0.51
ACE + GS-CPR Liu et al. (2025) (ICLR 2025)25/0.2926/0.385/0.2313/0.4117/33
STDLoc Huang et al. (2025) (CVPR 2025)15/0.1711.9/0.213/0.134.7/0.149/0.16
Ours (No Refinement)11/0.1913/0.264/0.187/0.159/0.20
Ours*12/0.1813/0.255/0.197/0.209/0.21
Ours9/0.1510/0.193/0.155/0.137/0.16
+ +In Table 3, we provide the experimental localization results on Cambridge Landmarks in comparison with APR, SCR, NeRF and 3D GS methods. The results on Cambridge Landmarks demonstrate that our method gains the best performance on both mean position and orientation metrics. Compared + +with the recent PMNet Lin et al. (2024a), the position error is significantly reduced from 91cm to 7cm, further validating the effectiveness of our framework. + +Generalization on Cambridge Landmarks. To assess generalization, we trained GS-RelocNet on the MegaDepth dataset Li & Snavely (2018) and tested it on the Cambridge Landmarks dataset in a scene-independent setting (marked as $\dagger$ ). The mean pose error across four scenes is $9\mathrm{cm} / 0.21^{\circ}$ , surpassing the performance of DUS3R ( $17\mathrm{cm} / 0.39^{\circ}$ ) and Reloc3R ( $48\mathrm{cm} / 0.58^{\circ}$ ). These results underscore the robustness and generalization capability of GS-RelocNet in diverse, unseen environments. + +Discussion of pose refinement. In Tables 1 and 3, we also present the localization performance without pose refinement. Two notable observations emerge from the results. First, pose refinement leads to improvements in localization accuracy across all scenes, demonstrating the positive impact of the refinement process using 3D Gaussians. Second, even without pose refinement, our method remains competitive with other state-of-the-art approaches. On 7 Scenes, the scene-dependent result is comparable with STDLoc, while the scene-independent performance is obviously more accurate than others. On Cambridge, the results are also comparable to those of STDLoc. + +Discussion of running time. Our framework efficiency comprises initial pose estimation with confidence map regression and PnP, and pose refinement with view rendering, residual map regression and PnP. On average, it processes testing images at 65 ms (15.4 FPS) on an Nvidia 4090 GPU across 7 Scenes and Cambridge Landmarks. The average running times per frame are as follows, 39 ms for confidence regression, 4 ms for initial PnP with RANSAC, 9 ms for view rendering, 8 ms for residual coordinate regression, and 5 ms for PnP in pose refinement. This outperforms 3D GS based methods like ACE Brachmann et al. (2023) + GS-CPR Liu et al. (2025) (190ms, 5.3 FPS) and STDLoc Huang et al. (2025) (143ms, 7 FPS), highlighting our computational efficiency. + +Discussion of comparison with 3G GS based methods. From Tables 1 and 3, we can see that localization performance of 3D GS based STDLoc and ACE + GS-CPR is slightly less accurate than ours, but also is competitive. Compared to these methods, the additional superiorities of our framework lies in two aspects. First, our method supports scene-independent relocalization, while STDLoc and GS-CPR requests training before localization in the target scene. Second, our inference speed is faster than the other two methods, achieving more than twice speed. + +# 4.5 Detailed Studies + +Table 4: Scene-independent localization results on 7 Scenes with original ground truth and Cambridge Landmarks with different settings. + +
Fusion in encoderFusion in decoderFusion in Refinement7 ScenesCambridge Landmarks
S1×××2.1/1.1714/0.35
S2✓××1.9/1.1413/0.32
S3×✓×1.9/1.1112/0.36
S4××✓2.0/1.0313/0.35
S5×✓✓1.6/0.9711/0.28
S6✓×✓1.7/0.9210/0.26
S7✓✓×1.5/0.849/0.23
S8✓✓✓1.4/0.829/0.21
Iterations of fusion model in GS-RelocNet
S911.7/0.9212/0.29
S1021.4/0.829/0.21
S1141.4/0.8610/0.22
S1281.5/0.8711/0.24
+ +Discussion of outstanding performance. The results of the aforementioned experiments demonstrate state-of-the-art performance of our framework. We attribute our outstanding performance to three main factors. First, our framework leverages the full potential of 3D GS for both initial pose estimation and pose refinement. Second, GS-RelocNet is specifically designed to establish accurate correspondences between pixels and 3D Gaussians. Third, we propose a refinement network that predicts the relative pose between real and rendered images using a bidirectional feature fusion module. In the following sections, we present ablation experiments to demonstrate the effectiveness of these fusion modules. + +Ablation studies of fusion modules. In S1 - S8 of Table 4, we conduct ablation experiments of the three fusion modules, including the unidirectional feature fusion module in encoder, decoder of + +GS-RelocNet and the bidirectional fusion module in the refinement network. In S5, S6, and S7, each module is removed individually. With this setting, the two features are reshaped and concatenated directly. It is evident that accuracy across all three datasets decreases in comparison to setting S8. In S2, S3, and S4, two modules are removed individually, and the results are less accurate than those in settings S5, S6, and S7. In setting S1, where all three modules are removed, the accuracy decreases most significantly, further confirming the positive impact of the three fusion modules. + +Detailed studies of the iterations in the unidirectional feature fusion module. In S9 - S12 of Table 4, we explore the effect of the number of iterations in the unidirectional feature fusion module by setting the iteration values to 1, 2, 4, and 8, respectively. When the iteration is set to 1, the result demonstrates that the estimation performance is less accurate than the other 3 settings. Moreover, the results for iterations set to 2, 4, and 8 are comparable. This discrepancy can be explained by the fact that with only one iteration, the RGB and point cloud features are not sufficiently fused, resulting in less accurate estimates. In contrast, with 2, 4, or 8 iterations, the fusion between the two feature types is sufficiently accomplished for RGB based pose estimation, leading to comparable performance. + +Limitation discussion. A primary limitation of our framework is its reliance on a high-quality 3D GS model of the target scene. When the 3D GS model is of suboptimal quality, localization performance may degrade, leading to failures or significant errors. + +# 4.6 AR Application + +![](images/5082764182ff270d007691968adfc812e9abdec9c10f67bc03ff8152b05affbd.jpg) +Figure 5: AR effect on the Chess and Office scene of 7 Scenes. We render a virtual walking person and a bottle onto the Chess scene, and two virtual bottles and a static person onto the Office scene based on the predicted camera pose of our framework. + +To demonstrate the performance in real-world AR applications, we present the virtual-real fusion results for two scenes using the predicted pose by our framework, the Chess and Office scenes from the 7 Scenes dataset, as shown in Fig. 5. Specifically, we render a virtual walking person and a bottle onto the Chess scene, and two virtual bottles and a static person onto the Office scene. + +# 5 Conclusion + +In this paper, we propose a novel 3D Gaussian based camera relocalization framework, composed of of two stages, an initial pose estimation stage, which predicts 2D-3D correspondences between image pixels and 3D Gaussians, and a refinement stage, which estimates the residual pose between the target and rendered views. To estimate the 2D-3D correspondences, we introduce a descriptor matching network called GS-RelocNet. Within GS-RelocNet, we design a unidirectional feature fusion model to combine RGB features with Gaussian features. After obtaining the initial camera pose, we proceed with the pose refinement network. In this refinement network, we propose a bidirectional feature fusion model to merge the features from the real and rendered images. To validate the performance of our framework, we conduct experiments on both indoor 7 Scenes, TUM RGB-D, Bonn and outdoor Cambridge Landmarks datasets. The results demonstrate state-of-the-art localization accuracy on both datasets. Additionally, we provide detailed studies on the feature fusion modules and the refinement stage, further highlighting the effectiveness of our approach. + +In summary, this paper makes a significant contribution by introducing a scene-independent localization framework through the full utilization of 3D Gaussians. By leveraging 3D Gaussians in both the initial and refinement stages, our method is able to deliver more accurate localization results across a variety of scenes. We hope that the proposed framework could be a universal localization pipeline. + +# Acknowledgments + +The paper is supported by Shandong Provincial Natural Science Foundation (No. ZR2024QF215), Key Research and Development Program of Rizhao (No. 2024ZDYF010053), National Natural Science Foundation of China (No. 62072020) and the Open Project Program of State Key Laboratory of Virtual Reality Technology and Systems, Beihang University (No. VRLAB2024A**). + +The authors thank the Zhiyang Innovation Technology Co., Ltd. for computing power and data support. + +# References + +Balntas, V., Li, S., and Prisacariu, V. Relocnet: Continuous metric learning relocalisation using neural nets. In European Conference on Computer Vision, pp. 751-767, 2018. +Bescos, B., Fácil, J. M., Civera, J., and Neira, J. Dynaslam: Tracking, mapping, and inpainting in dynamic scenes. IEEE Robotics and Automation Letters, 3(4):4076-4083, 2018. +Brachmann, E. and Rother, C. Learning less is more-6d camera localization via 3d surface regression. In IEEE Conference on Computer Vision and Pattern Recognition, pp. 4654-4662, 2018. +Brachmann, E. and Rother, C. Visual camera re-localization from rgb and rgb-d images using dsac. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(9):5847-5865, 2021. +Brachmann, E., Cavallari, T., and Prisacariu, V. A. Accelerated coordinate encoding: Learning to relocalize in minutes using rgb and poses. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 5044-5053, 2023. +Chen, S., Li, X., Wang, Z., and Prisacariu, V. A. Dfnet: Enhance absolute pose regression with direct feature matching. In European Conference on Computer Vision, pp. 1-17. Springer, 2022. +Chen, S., Bhalgat, Y., Li, X., Bian, J.-W., Li, K., Wang, Z., and Prisacariu, V. A. Neural refinement for absolute pose regression with feature synthesis. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 20987-20996, 2024a. +Chen, S., Cavallari, T., Prisacariu, V. A., and Brachmann, E. Map-relative pose regression for visual relocalization. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 20665-20674, 2024b. +Dai, A., Chang, A. X., Savva, M., Halber, M., Funkhouser, T., and Nießner, M. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In IEEE Conference on Computer Vision and Pattern Recognition, pp. 5828-5839, 2017. +Dong, S., Wang, S., Liu, S., Cai, L., Fan, Q., Kannala, J., and Yang, Y. Reloc3r: Large-scale training of relative camera pose regression for generalizable, fast, and accurate visual localization. In IEEE Computer Vision and Pattern Recognition Conference, pp. 16739-16752, 2025. +Du, Z.-J., Huang, S.-S., Mu, T.-J., Zhao, Q., Martin, R. R., and Xu, K. Accurate dynamic slam using crf-based long-term consistency. IEEE Transactions on Visualization and Computer Graphics, 28(4):1745-1757, 2020. +Ferens, R. and Keller, Y. Hyperpose: Hypernetwork-infused camera pose localization and an extended cambridge landmarks dataset. arXiv preprint arXiv:2303.02610, 2025. +Germain, H., DeTone, D., Pascoe, G., Schmidt, T., Novotny, D., Newcombe, R., Sweeney, C., Szeliski, R., and Balntas, V. Feature query networks: Neural surface description for camera pose refinement. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 5071-5081, 2022. +Giang, K. T., Song, S., and Jo, S. Learning to produce semi-dense correspondences for visual localization. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 19468-19478, 2024. +Hu, J., Mao, M., Bao, H., Zhang, G., and Cui, Z. Cp-slam: Collaborative neural point-based slam system. Advances in Neural Information Processing Systems, 36:39429-39442, 2023. +Huang, Z., Yu, H., Shentu, Y., Yuan, J., and Zhang, G. From sparse to dense: Camera relocalization with scene-specific detector from feature gaussian splatting. In IEEE Computer Vision and Pattern Recognition Conference, pp. 27059-27069, 2025. + +Jia, P., Liu, Y., Li, X., Zhao, X., Wang, Y., Du, Y., Han, X., Wei, X., Wang, S., and Yin, D. G3: an effective and adaptive framework for worldwide geolocation using large multi-modality models. Advances in Neural Information Processing Systems, 37:53198-53221, 2024. +Keetha, N., Karhade, J., Jatavallabhula, K. M., Yang, G., Scherer, S., Ramanan, D., and Luiten, J. Splatam: Splat track & map 3d gaussians for dense rgb-d slam. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 21357-21366, 2024. +Kendall, A. and Cipolla, R. Geometric loss functions for camera pose regression with deep learning. In IEEE Conference on Computer Vision and Pattern Recognition, pp. 5974-5983, 2017. +Kendall, A., Grimes, M., and Cipolla, R. Posenet: A convolutional network for real-time 6-dof camera relocalization. In IEEE International Conference on Computer Vision, pp. 2938-2946, 2015. +Kerbl, B., Kopanas, G., Leimkuhler, T., and Drettakis, G. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42(4):139-1, 2023. +Kulhanek, J., Peng, S., Kukelova, Z., Pollefeys, M., and Sattler, T. Wildgaussians: 3d gaussian splatting in the wild. arXiv preprint arXiv:2407.08447, 2024. +Laskar, Z., Melekhov, I., Kalia, S., and Kannala, J. Camera relocalization by computing pairwise relative poses using convolutional neural network. In IEEE International Conference on Computer Vision, pp. 929-938, 2017. +Leroy, V., Cabon, Y., and Revaud, J. Grounding image matching in 3d with mast3r. In European Conference on Computer Vision, pp. 71-91, 2024. +Li, Z. and Snavely, N. Megadepth: Learning single-view depth prediction from internet photos. In IEEE Conference on Computer Vision and Pattern Recognition, pp. 2041-2050, 2018. +Lin, J., Gu, J., Wu, B., Fan, L., Chen, R., Liu, L., and Ye, J. Learning neural volumetric pose features for camera localization. In European Conference on Computer Vision, pp. 198-214. Springer, 2024a. +Lin, X., Ruan, J., Yang, Y., He, L., Guan, Y., and Zhang, H. Robust data association against detection deficiency for semantic slam. IEEE Transactions on Automation Science and Engineering, 21(1):868-880, 2024b. +Liu, C., Chen, S., Bhalgat, Y., Hu, S., Cheng, M., Wang, Z., Prisacariu, V. A., and Braud, T. Gsloc: Efficient camera pose refinement via 3d gaussian splatting. arXiv preprint arXiv:2408.11085, 2024a. +Liu, C., Chen, S., Zhao, Y., Huang, H., Prisacariu, V., and Braud, T. Hr-apr: Apr-agnostic framework with uncertainty estimation and hierarchical refinement for camera relocalisation. arXiv preprint arXiv:2402.14371, 2024b. +Liu, C., Chen, S., Bhalgat, Y. S., Hu, S., Cheng, M., Wang, Z., Prisacariu, V. A., and Braud, T. Gs-cpr: Efficient camera pose refinement via 3d gaussian splattering. In International Conference on Learning Representations, 2025. +Liu, L., Li, H., and Dai, Y. Efficient global 2d-3d matching for camera localization in a large-scale 3d map. In IEEE International Conference on Computer Vision, pp. 2372-2381, 2017. +Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., and Guo, B. Swin transformer: Hierarchical vision transformer using shifted windows. In IEEE/CVF International Conference on Computer Vision, pp. 10012-10022, 2021. +Mildenhall, B., Srinivasan, P. P., Tancik, M., Barron, J. T., Ramamoorthi, R., and Ng, R. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. +Moreau, A., Piasco, N., Bennehar, M., Tsishkou, D., Stanciulescu, B., and de La Fortelle, A. Crossfire: Camera relocalization on self-supervised features from an implicit representation. In IEEE International Conference on Computer Vision, pp. 252-262, 2023. +Mur-Artal, R. and Tardós, J. D. Orb-slam2: An open-source slam system for monocular, stereo, and rgb-d cameras. IEEE Transactions on Robotics, 33(5):1255-1262, 2017. +Oquab, M., Darcet, T., Moutakanni, T., Vo, H. V., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., Howes, R., Huang, P.-Y., Xu, H., Sharma, V., Li, S.-W., Galuba, W., Rabbat, M., Assran, M., Ballas, N., Synnaeve, G., Misra, I., Jegou, H., Mairal, J., Labatut, P., Joulin, A., and Bojanowski, P. Dinov2: Learning robust visual features without supervision, 2023. + +Palazzolo, E., Behley, J., Lottes, P., Giguere, P., and Stachniss, C. Refusion: 3d reconstruction in dynamic environments for rgb-d cameras exploiting residuals. In IEEE/RSJ International Conference on Intelligent Robots and Systems, pp. 7855-7862, 2019. +Runz, M., Buffier, M., and Agapito, L. Maskfusion: Real-time recognition, tracking and reconstruction of multiple moving objects. In IEEE International Symposium on Mixed and Augmented Reality, pp. 10-20. IEEE, 2018. +Sarlin, P.-E., Unagar, A., Larsson, M., Germain, H., Toft, C., Larsson, V., Pollefeys, M., Lepetit, V., Hammarstrand, L., Kahl, F., et al. Back to the feature: Learning robust camera localization from pixels to pose. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 3247-3257, 2021. +Sattler, T., Leibe, B., and Kobbelt, L. Efficient & effective prioritized matching for large-scale image-based localization. IEEE Transactions on Pattern Analysis and Machine Intelligence, 39(9):1744-1756, 2016. +Sattler, T., Leibe, B., and Kobbelt, L. Efficient & effective prioritized matching for large-scale image-based localization. IEEE Transactions on Pattern Analysis and Machine Intelligence, 39(9):1744-1756, 2017. +Shavit, Y., Ferens, R., and Keller, Y. Learning multi-scene absolute pose regression with transformers. In IEEE/CVF International Conference on Computer Vision, pp. 2733-2742, 2021. +Shotton, J., Glocker, B., Zach, C., Izadi, S., Criminisi, A., and Fitzgibbon, A. Scene coordinate regression forests for camera relocalization in rgb-d images. In IEEE Conference on Computer Vision and Pattern Recognition, pp. 2930-2937, 2013. +Song, B., Yuan, X., Ying, Z., Yang, B., Song, Y., and Zhou, F. Dgm-vins: Visual-inertial slam for complex dynamic environments with joint geometry feature extraction and multiple object tracking. IEEE Transactions on Instrumentation and Measurement, 72:1-11, 2023. +Sturm, J., Engelhard, N., Endres, F., Burgard, W., and Cremers, D. A benchmark for the evaluation of rgb-d slam systems. In IEEE/RSJ International Conference on Intelligent Robots and Systems, pp. 573-580, 2012. +Sun, J., Shen, Z., Wang, Y., Bao, H., and Zhou, X. Loftr: Detector-free local feature matching with transformers. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 8922-8931, 2021. +Taira, H., Okutomi, M., Sattler, T., Cimpoi, M., Pollefeys, M., Sivic, J., Pajdla, T., and Torii, A. Inloc: Indoor visual localization with dense matching and view synthesis. In IEEE Conference on Computer Vision and Pattern Recognition, pp. 7199-7209, 2018. +Tang, S., Tang, S., Tagliasacchi, A., Tan, P., and Furukawa, Y. Neumap: Neural coordinate mapping by auto-transdecoder for camera localization. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 929-939, 2023. +Valentin, J., Dai, A., Nießner, M., Kohli, P., Torr, P., Izadi, S., and Keskin, C. Learning to navigate the energy landscape. In International Conference on 3D Vision, pp. 323-332, 2016. +Wang, B., Chen, C., Lu, C. X., Zhao, P., Trigoni, N., and Markham, A. Atloc: Attention guided camera localization. In AAAI Conference on Artificial Intelligence, volume 34, pp. 10393-10401, 2020. +Wang, F., Jiang, X., Galliani, S., Vogel, C., and Pollefeys, M. Glace: Global local accelerated coordinate encoding. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 21562-21571, 2024a. +Wang, J. and Qi, Y. Camera relocalization using deep point cloud generation and hand-crafted feature refinement. In IEEE International Conference on Robotics and Automation, pp. 5891-5897, 2021. +Wang, J. and Qi, Y. Scene-independent localization by learning residual coordinate map with cascaded localizers. In IEEE International Symposium on Mixed and Augmented Reality, pp. 79-88, 2023a. +Wang, J. and Qi, Y. Simultaneous scene-independent camera localization and category-level object pose estimation via multi-level feature fusion. In IEEE Conference Virtual Reality and 3D User Interfaces, pp. 254-264. IEEE, 2023b. +Wang, Q., Zhang, J., Yang, K., Peng, K., and Stiefelhagen, R. Matchformer: Interleaving attention in transformers for feature matching. In Asian Conference on Computer Vision, pp. 2746-2762, 2022. +Wang, S., Leroy, V., Cabon, Y., Chidlovskii, B., and Revaud, J. Dust3r: Geometric 3d vision made easy. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 20697-20709, 2024b. +Wang, T., Sheng, H., Chen, R., Yang, D., Cui, Z., Wang, S., Cong, R., and Zhao, M. Light field depth estimation: A comprehensive survey from principles to future. High-Confidence Computing, 4(1):100187, 2024c. + +Wang, Y., Yan, Y., Shi, D., Zhu, W., Xia, J., Jeff, T., Jin, S., Gao, K., Li, X., and Yang, X. Nerf-ibvs: visual servo based on nerf for visual localization and navigation. Advances in Neural Information Processing Systems, 36: 8292-8304, 2023. +Wang, Y., He, X., Peng, S., Tan, D., and Zhou, X. Efficient loftr: Semi-dense local feature matching with sparse-like speed. arXiv preprint arXiv:2403.04765, 2024d. +Wu, X., Jiang, L., Wang, P.-S., Liu, Z., Liu, X., Qiao, Y., Ouyang, W., He, T., and Zhao, H. Point transformer v3: Simpler faster stronger. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 4840-4851, 2024. +Xi, J., Zhang, W., Xu, Z., Zhu, S., Tang, L., and Zhao, L. Three-dimensional dynamic gesture recognition method based on convolutional neural network. High-Confidence Computing, 5(1):100280, 2025. +Xu, Y., Jiang, H., Xiao, Z., Feng, J., and Zhang, L. Dg-slam: Robust dynamic gaussian splattering slam with hybrid pose optimization. Advances in Neural Information Processing Systems, 37:51577-51596, 2024. +Yan, C., Qu, D., Xu, D., Zhao, B., Wang, Z., Wang, D., and Li, X. Gs-slam: Dense visual slam with 3d gaussian splatting. In EEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 19595-19604, 2024. +Yang, S. and Scherer, S. Cubeslam: Monocular 3-d object slam. IEEE Transactions on Robotics, 35(4):925-938, 2019. +You, M., Luo, C., Zhou, H., and Zhu, S. Dynamic dense crf inference for video segmentation and semantic slam. Pattern Recognition, 133:109023, 2023. +Yu, C., Liu, Z., Liu, X.-J., Xie, F., Yang, Y., Wei, Q., and Fei, Q. Ds-slam: A semantic visual slam towards dynamic environments. In IEEE/RSJ International Conference on Intelligent Robots and Systems, pp. 1168-1174, 2018. +Zhang, D., Wang, C., Wang, W., Li, P., Qin, M., and Wang, H. Gaussian in the wild: 3d gaussian splatting for unconstrained image collections. In European Conference on Computer Vision, pp. 341-359, 2024. +Zhang, J., Peng, W., Xiao, A., Liu, T., Fu, J., Chen, J., and Yan, Z. Kans-detr: Enhancing detection transformer with kolmogorov-arnold networks for small object. High-Confidence Computing, pp. 100336, 2025. +Zhou, L., Luo, Z., Shen, T., Zhang, J., Zhen, M., Yao, Y., Fang, T., and Quan, L. Kfnet: Learning temporal camera relocalization using kalman filtering. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 4919-4928, 2020. +Zhou, Q., Maximov, M., Litany, O., and Leal-Taixe, L. The nefrect match: Exploring nef features for visual localization. In European Conference on Computer Vision, pp. 108-127, 2024a. +Zhou, Q., Maximov, M., Litany, O., and Leal-Taixe, L. The nefrect match: Exploring nef features for visual localization. In European Conference on Computer Vision, pp. 108-127. Springer, 2024b. +Zhu, J., Yan, S., Wang, L., Zhang, S., Liu, Y., and Zhang, M. Lod-loc: Aerial visual localization using lod 3d map with neural wireframe alignment. Advances in Neural Information Processing Systems, 37:119063-119098, 2024. + +# NeurIPS Paper Checklist + +# 1. Claims + +Question: Do the main claims made in the abstract and introduction accurately reflect the paper's contributions and scope? + +Answer: [Yes] + +Justification: [Yes] + +Guidelines: + +- The answer NA means that the abstract and introduction do not include the claims made in the paper. +- The abstract and/or introduction should clearly state the claims made, including the contributions made in the paper and important assumptions and limitations. A No or NA answer to this question will not be perceived well by the reviewers. +- The claims made should match theoretical and experimental results, and reflect how much the results can be expected to generalize to other settings. +- It is fine to include aspirational goals as motivation as long as it is clear that these goals are not attained by the paper. + +# 2. Limitations + +Question: Does the paper discuss the limitations of the work performed by the authors? + +Answer: [Yes] + +Justification: [Yes] + +Guidelines: + +- The answer NA means that the paper has no limitation while the answer No means that the paper has limitations, but those are not discussed in the paper. +- The authors are encouraged to create a separate "Limitations" section in their paper. +- The paper should point out any strong assumptions and how robust the results are to violations of these assumptions (e.g., independence assumptions, noiseless settings, model well-specification, asymptotic approximations only holding locally). The authors should reflect on how these assumptions might be violated in practice and what the implications would be. +- The authors should reflect on the scope of the claims made, e.g., if the approach was only tested on a few datasets or with a few runs. In general, empirical results often depend on implicit assumptions, which should be articulated. +- The authors should reflect on the factors that influence the performance of the approach. For example, a facial recognition algorithm may perform poorly when image resolution is low or images are taken in low lighting. Or a speech-to-text system might not be used reliably to provide closed captions for online lectures because it fails to handle technical jargon. +- The authors should discuss the computational efficiency of the proposed algorithms and how they scale with dataset size. +- If applicable, the authors should discuss possible limitations of their approach to address problems of privacy and fairness. +- While the authors might fear that complete honesty about limitations might be used by reviewers as grounds for rejection, a worse outcome might be that reviewers discover limitations that aren't acknowledged in the paper. The authors should use their best judgment and recognize that individual actions in favor of transparency play an important role in developing norms that preserve the integrity of the community. Reviewers will be specifically instructed to not penalize honesty concerning limitations. + +# 3. Theory assumptions and proofs + +Question: For each theoretical result, does the paper provide the full set of assumptions and a complete (and correct) proof? + +Answer: [NA] + +# Justification: [NA] + +# Guidelines: + +- The answer NA means that the paper does not include theoretical results. +- All the theorems, formulas, and proofs in the paper should be numbered and cross-referenced. +- All assumptions should be clearly stated or referenced in the statement of any theorems. +- The proofs can either appear in the main paper or the supplemental material, but if they appear in the supplemental material, the authors are encouraged to provide a short proof sketch to provide intuition. +- Inversely, any informal proof provided in the core of the paper should be complemented by formal proofs provided in appendix or supplemental material. +- Theorems and Lemmas that the proof relies upon should be properly referenced. + +# 4. Experimental result reproducibility + +Question: Does the paper fully disclose all the information needed to reproduce the main experimental results of the paper to the extent that it affects the main claims and/or conclusions of the paper (regardless of whether the code and data are provided or not)? + +# Answer: [Yes] + +Justification: [Yes] + +# Guidelines: + +- The answer NA means that the paper does not include experiments. +- If the paper includes experiments, a No answer to this question will not be perceived well by the reviewers: Making the paper reproducible is important, regardless of whether the code and data are provided or not. +- If the contribution is a dataset and/or model, the authors should describe the steps taken to make their results reproducible or verifiable. +- Depending on the contribution, reproducibility can be accomplished in various ways. For example, if the contribution is a novel architecture, describing the architecture fully might suffice, or if the contribution is a specific model and empirical evaluation, it may be necessary to either make it possible for others to replicate the model with the same dataset, or provide access to the model. In general, releasing code and data is often one good way to accomplish this, but reproducibility can also be provided via detailed instructions for how to replicate the results, access to a hosted model (e.g., in the case of a large language model), releasing of a model checkpoint, or other means that are appropriate to the research performed. +- While NeurIPS does not require releasing code, the conference does require all submissions to provide some reasonable avenue for reproducibility, which may depend on the nature of the contribution. For example +(a) If the contribution is primarily a new algorithm, the paper should make it clear how to reproduce that algorithm. +(b) If the contribution is primarily a new model architecture, the paper should describe the architecture clearly and fully. +(c) If the contribution is a new model (e.g., a large language model), then there should either be a way to access this model for reproducing the results or a way to reproduce the model (e.g., with an open-source dataset or instructions for how to construct the dataset). +(d) We recognize that reproducibility may be tricky in some cases, in which case authors are welcome to describe the particular way they provide for reproducibility. In the case of closed-source models, it may be that access to the model is limited in some way (e.g., to registered users), but it should be possible for other researchers to have some path to reproducing or verifying the results. + +# 5. Open access to data and code + +Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material? + +Answer: [Yes] + +Justification: [Yes] + +Guidelines: + +- The answer NA means that paper does not include experiments requiring code. +- Please see the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details. +- While we encourage the release of code and data, we understand that this might not be possible, so "No" is an acceptable answer. Papers cannot be rejected simply for not including code, unless this is central to the contribution (e.g., for a new open-source benchmark). +- The instructions should contain the exact command and environment needed to run to reproduce the results. See the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details. +- The authors should provide instructions on data access and preparation, including how to access the raw data, preprocessed data, intermediate data, and generated data, etc. +- The authors should provide scripts to reproduce all experimental results for the new proposed method and baselines. If only a subset of experiments are reproducible, they should state which ones are omitted from the script and why. +- At submission time, to preserve anonymity, the authors should release anonymized versions (if applicable). +- Providing as much information as possible in supplemental material (appended to the paper) is recommended, but including URLs to data and code is permitted. + +# 6. Experimental setting/details + +Question: Does the paper specify all the training and test details (e.g., data splits, hyperparameters, how they were chosen, type of optimizer, etc.) necessary to understand the results? + +Answer: [Yes] + +Justification: [Yes] + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The experimental setting should be presented in the core of the paper to a level of detail that is necessary to appreciate the results and make sense of them. +- The full details can be provided either with the code, in appendix, or as supplemental material. + +# 7. Experiment statistical significance + +Question: Does the paper report error bars suitably and correctly defined or other appropriate information about the statistical significance of the experiments? + +Answer: [Yes] + +Justification: [Yes] + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The authors should answer "Yes" if the results are accompanied by error bars, confidence intervals, or statistical significance tests, at least for the experiments that support the main claims of the paper. +- The factors of variability that the error bars are capturing should be clearly stated (for example, train/test split, initialization, random drawing of some parameter, or overall run with given experimental conditions). +- The method for calculating the error bars should be explained (closed form formula, call to a library function, bootstrap, etc.) +- The assumptions made should be given (e.g., Normally distributed errors). +- It should be clear whether the error bar is the standard deviation or the standard error of the mean. + +- It is OK to report 1-sigma error bars, but one should state it. The authors should preferably report a 2-sigma error bar than state that they have a $96\%$ CI, if the hypothesis of Normality of errors is not verified. +- For asymmetric distributions, the authors should be careful not to show in tables or figures symmetric error bars that would yield results that are out of range (e.g. negative error rates). +- If error bars are reported in tables or plots, The authors should explain in the text how they were calculated and reference the corresponding figures or tables in the text. + +# 8. Experiments compute resources + +Question: For each experiment, does the paper provide sufficient information on the computer resources (type of compute workers, memory, time of execution) needed to reproduce the experiments? + +Answer: [Yes] + +Justification: [Yes] + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The paper should indicate the type of compute workers CPU or GPU, internal cluster, or cloud provider, including relevant memory and storage. +- The paper should provide the amount of compute required for each of the individual experimental runs as well as estimate the total compute. +- The paper should disclose whether the full research project required more computer than the experiments reported in the paper (e.g., preliminary or failed experiments that didn't make it into the paper). + +# 9. Code of ethics + +Question: Does the research conducted in the paper conform, in every respect, with the NeurIPS Code of Ethics https://neurips.cc/public/EthicsGuidelines? + +Answer: [Yes] + +Justification: [Yes] + +Guidelines: + +- The answer NA means that the authors have not reviewed the NeurIPS Code of Ethics. +- If the authors answer No, they should explain the special circumstances that require a deviation from the Code of Ethics. +- The authors should make sure to preserve anonymity (e.g., if there is a special consideration due to laws or regulations in their jurisdiction). + +# 10. Broader impacts + +Question: Does the paper discuss both potential positive societal impacts and negative societal impacts of the work performed? + +Answer: [NA] + +Justification: [NA] + +Guidelines: + +- The answer NA means that there is no societal impact of the work performed. +- If the authors answer NA or No, they should explain why their work has no societal impact or why the paper does not address societal impact. +- Examples of negative societal impacts include potential malicious or unintended uses (e.g., disinformation, generating fake profiles, surveillance), fairness considerations (e.g., deployment of technologies that could make decisions that unfairly impact specific groups), privacy considerations, and security considerations. +- The conference expects that many papers will be foundational research and not tied to particular applications, let alone deployments. However, if there is a direct path to any negative applications, the authors should point it out. For example, it is legitimate to point out that an improvement in the quality of generative models could be used to + +generate deepfakes for disinformation. On the other hand, it is not needed to point out that a generic algorithm for optimizing neural networks could enable people to train models that generate Deepfakes faster. + +- The authors should consider possible harms that could arise when the technology is being used as intended and functioning correctly, harms that could arise when the technology is being used as intended but gives incorrect results, and harms following from (intentional or unintentional) misuse of the technology. +- If there are negative societal impacts, the authors could also discuss possible mitigation strategies (e.g., gated release of models, providing defenses in addition to attacks, mechanisms for monitoring misuse, mechanisms to monitor how a system learns from feedback over time, improving the efficiency and accessibility of ML). + +# 11. Safeguards + +Question: Does the paper describe safeguards that have been put in place for responsible release of data or models that have a high risk for misuse (e.g., pretrained language models, image generators, or scraped datasets)? + +Answer: [NA] + +Justification: [NA] + +Guidelines: + +- The answer NA means that the paper poses no such risks. +- Released models that have a high risk for misuse or dual-use should be released with necessary safeguards to allow for controlled use of the model, for example by requiring that users adhere to usage guidelines or restrictions to access the model or implementing safety filters. +- Datasets that have been scraped from the Internet could pose safety risks. The authors should describe how they avoided releasing unsafe images. +- We recognize that providing effective safeguards is challenging, and many papers do not require this, but we encourage authors to take this into account and make a best faith effort. + +# 12. Licenses for existing assets + +Question: Are the creators or original owners of assets (e.g., code, data, models), used in the paper, properly credited and are the license and terms of use explicitly mentioned and properly respected? + +Answer: [Yes] + +Justification: [Yes] + +Guidelines: + +- The answer NA means that the paper does not use existing assets. +- The authors should cite the original paper that produced the code package or dataset. +- The authors should state which version of the asset is used and, if possible, include a URL. +- The name of the license (e.g., CC-BY 4.0) should be included for each asset. +- For scraped data from a particular source (e.g., website), the copyright and terms of service of that source should be provided. +- If assets are released, the license, copyright information, and terms of use in the package should be provided. For popular datasets, paperswithcode.com/datasets has curated licenses for some datasets. Their licensing guide can help determine the license of a dataset. +- For existing datasets that are re-packaged, both the original license and the license of the derived asset (if it has changed) should be provided. +- If this information is not available online, the authors are encouraged to reach out to the asset's creators. + +# 13. New assets + +Question: Are new assets introduced in the paper well documented and is the documentation provided alongside the assets? + +Answer: [NA] + +Justification: [NA] + +Guidelines: + +- The answer NA means that the paper does not release new assets. +- Researchers should communicate the details of the dataset/code/model as part of their submissions via structured templates. This includes details about training, license, limitations, etc. +- The paper should discuss whether and how consent was obtained from people whose asset is used. +- At submission time, remember to anonymize your assets (if applicable). You can either create an anonymized URL or include an anonymized zip file. + +# 14. Crowdsourcing and research with human subjects + +Question: For crowdsourcing experiments and research with human subjects, does the paper include the full text of instructions given to participants and screenshots, if applicable, as well as details about compensation (if any)? + +Answer: [NA] + +Justification: [NA] + +Guidelines: + +- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. +- Including this information in the supplemental material is fine, but if the main contribution of the paper involves human subjects, then as much detail as possible should be included in the main paper. +- According to the NeurIPS Code of Ethics, workers involved in data collection, curation, or other labor should be paid at least the minimum wage in the country of the data collector. + +# 15. Institutional review board (IRB) approvals or equivalent for research with human subjects + +Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals (or an equivalent approval/review based on the requirements of your country or institution) were obtained? + +Answer: [NA] + +Justification: [NA] + +Guidelines: + +- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. +- Depending on the country in which research is conducted, IRB approval (or equivalent) may be required for any human subjects research. If you obtained IRB approval, you should clearly state this in the paper. +- We recognize that the procedures for this may vary significantly between institutions and locations, and we expect authors to adhere to the NeurIPS Code of Ethics and the guidelines for their institution. +- For initial submissions, do not include any information that would break anonymity (if applicable), such as the institution conducting the review. + +# 16. Declaration of LLM usage + +Question: Does the paper describe the usage of LLMs if it is an important, original, or non-standard component of the core methods in this research? Note that if the LLM is used only for writing, editing, or formatting purposes and does not impact the core methodology, scientific rigorousness, or originality of the research, declaration is not required. + +Answer: [NA] + +# Justification: [NA] + +# Guidelines: + +- The answer NA means that the core method development in this research does not involve LLMs as any important, original, or non-standard components. +- Please refer to our LLM policy (https://neurips.cc/Conferences/2025/LLM) for what should or should not be described. + +# A Technical Appendices and Supplementary Material + +# A.1 Residual Coordinate Map for Pose Refinement + +Given a pair of images, the residual coordinate map denotes the XYZ coordinate difference between the current and previous camera coordinate space Wang & Qi (2023a), which is used to predict the relative pose. In our paper, given the real input image and rendered view with initially predicted pose $R_{i}^{p}, T_{i}^{p}$ , we use the residual coordinate map to solve the relative pose $R_{r}^{p}, T_{r}^{p}$ between them. Then the refined pose is obtained by Eq. 2. + +The construction process of residual coordinate map is as follows. We substitute the depth information (Z-axis value) with the grayscale value $M_{d}$ denotes the XYZ coordinates under the camera space of the input real view, obtained by uniformly sampling from the grayscale image. For each point $\mathbf{p} \in M_d$ , we first transform it to the world space. Then, it is converted to the camera space of the rendered frame. Finally, the relative point is obtained by subtracting the original coordinate from the transformed one. In summary, the coordinate representation is defined by the following formula. + +$$ +M _ {r} = \left(R _ {r} ^ {p} - E\right) M _ {d} + T _ {r} ^ {p}, \tag {5} +$$ + +where $E$ denotes the identity matrix. Through regressing the coordinate map by the refinement network, $R_{r}^{p}, T_{r}^{p}$ can be predicted by the PnP method with RANSAC. + +# A.2 More Dataset Details + +7 Scenes Shotton et al. (2013). This dataset includes seven indoor scenes, each containing 2 to 10 sequences. It provides depth images, color frames, and ground truth poses. All scenes were recorded using a handheld Kinect RGB-D camera at a resolution of $640 \times 480$ and are divided into separate training and testing sets. The ground truth poses were obtained using the KinectFusion system. This dataset has recently been used as a benchmark in studies Kendall & Cipolla (2017); Brachmann & Rother (2018); Zhou et al. (2020), making it convenient for comparison with other methods. + +Cambridge Landmarks Kendall et al. (2015). This outdoor dataset includes several large outdoor environments. In this paper, we use five scenes (College, Hospital, Shop, and Church) to evaluate localization accuracy. Ground truth poses in each scene were calibrated using Visual SFM, and a sparse point cloud is also provided. + +ScanNet Dai et al. (2017). ScanNet contains over 1,500 scans, amounting to around 2.5 million views. The dataset was captured using a user-friendly and scalable RGB-D capture system. To evaluate scene-independent performance, we train GS-RelocNet on ScanNet and tested it on the 7 Scenes dataset. + +TUM RGB-D Sturm et al. (2012). The TUM RGB-D dataset is designed to benchmark visual odometry and SLAM systems. We select the four dynamic scenes (fr_walking_xyz, fr_walking-static, fr_walking_rpy, fr_walking_half) from TUM RGB-D to evaluate the localization performance in dynamic environments. + +Bonn Palazzolo et al. (2019). The Bonn dataset is tailored for dynamic localization, featuring highly dynamic sequences. We select 20 sequences from the dynamic subset (same as LC-CRF SLAM Du et al. (2020)), where individuals perform various tasks such as manipulating boxes or interacting with balloons, along with 2 static sequences. + +12 Scenes Valentin et al. (2016). The 12 Scenes dataset features 12 larger indoor environments, with volumes ranging from $14m^3$ to $79m^3$ . + +# A.3 More Implementation Details + +3D GS training details. To construct the 3D GS model, we first utilize COLMAP to generate an initial point cloud using ground truth poses. Subsequently, we employ the original 3D GS model with its default configuration settings. + +-iterations: 30000. Total number of training iterations. +-position_lr_init: 0.00016. The initial learning rate of the Gaussian position. +-position_lr_final: 0.0000016. The final learning rate of the Gaussian position. + +-position_lr_delay_mult: 0.01. The delay multiplier before the learning rate decay begins. +-position_lr_max_steps: 30000. The total number of steps for the learning rate decay. +-feature_lr: 0.0025. Learning rate of spherical harmonic function coefficients. +- opacity_lr: 0.05. Learning rate of opacity. +-scaling_lr: 0.005. Scaling learning rate. +-rotation_lr: 0.001. Learning rate of rotation. +-densify_from_iter: 500. Densification begins from which iteration. +-densify_until_iter: 15000. Densification ends at which iteration. +-densification_interval: 100. Perform densification and pruning checks every few iterations. +- opacity_prune_threshold: 0.005. Opacity pruning threshold. +-densify_grad_threshold: 0.0002. The gradient threshold for densifying the Gaussian sphere. + +PnP details. The PnP with RANSAC uses OpenCV implementation with following parameters. + +-iterationsCount: 100. The number of RANSAC iterations. +-reprojectionError: 8. Threshold for reprojection error. +-confidence: 0.99. Degree of confidence. +flags:SOLVEPNP.IterATIVE.PnP solver algorithm. + +Pose extension in the refinement network. In the refinement network training, small pose differences between rendered and real frames require high coordinate accuracy, which increases the learning difficulty. To address this, we extend the relative pose using fixed coefficients. The position is scaled directly, and the orientation is expanded through a transformation between the quaternion and Euler angles. Specifically, the extension coefficient is set to 8.0. + +Initial pose estimation by PnP. For the PnP solver with RANSAC, we adapt the traditional RANSAC framework by incorporating our predicted confidence scores. Conventionally, RANSAC determines the final result based on the number of inlier points. In our modified approach, we instead use the sum of the confidence values of the inlier points to make this determination, thereby improving the reliability of the pose estimation. + +# A.4 More Results on 7 Scenes + +Table 5: The percentage of localization error under $5cm$ , $5^{\circ}$ and $2cm$ , $2^{\circ}$ on indoor 7 Scenes compared with other methods. Specially, the sign $\ddagger$ means the result with SfM pseudo ground truth, while others leverage the original KinectFusion ground truth. The red and blue marks represent the first and second. + +
Method5cm, 5° (↑)2cm, 2° (↑)
APRDFNet Chen et al. (2022)43.18.4
Marepo Chen et al. (2024b)84.033.7
SCRDSAC*‡ Brachmann & Rother (2021)97.880.7
ACE‡ Brachmann et al. (2023)97.183.3
GLACE‡ Wang et al. (2024a)95.682.2
NeRFNeReS Chen et al. (2024a)78.345.9
HR-APR Liu et al. (2024b)76.440.2
NeRFMatch‡ Zhou et al. (2024b)78.4-
3D GaussianDFNet + GS-CPR‡ Liu et al. (2025) (Accepted by ICLR 2025)94.276.5
ACE + GS-CPR‡ Liu et al. (2025) (Accepted by ICLR 2025)100.093.1
STDLoc‡ Huang et al. (2025) (Accepted by CVPR 2025)99.190.9
DFNet + GS-CPR‡ Liu et al. (2024a)94.276.5
Ours‡99.894.9
+ +$5\mathrm{cm}, 5^{\circ}$ and $2\mathrm{cm}, 2^{\circ}$ metric on 7 Scenes. Besides the accuracy, the localization stability is also an important metric, always expressed by the percentage of position and orientation error under $5\mathrm{cm}, 5^{\circ}$ and $2\mathrm{cm}, 2^{\circ}$ . Table 5 presents a comparison of our results with those of DFNet Chen et al. (2022), + +Marepo Chen et al. (2024b), DSAC* Brachmann & Rother (2021), ACE Brachmann et al. (2023), GLACE Wang et al. (2024a), NeReS Chen et al. (2024a), HR-APR Liu et al. (2024b), NeRFMatch Zhou et al. (2024b) and GSLoc Liu et al. (2024a). + +Our approach achieves an accuracy of $99.8\%$ , slightly below the state-of-the-art performance of ACE Brachmann et al. (2023) + GS-CPR Liu et al. (2025) $(100\%)$ . However, our method outperforms all other competing approaches. For the more stringent $2cm, 2^{\circ}$ metric, our framework demonstrates at least a $1.7\%$ improvement in accuracy compared to the next-best method, underscoring its robustness in challenging indoor environments. In comparison to DFNet Chen et al. (2022) + GS-CPR Liu et al. (2025), our method consistently achieves higher accuracy across both metrics. Notably, while GS-CPR relies on accurate initial pose estimates, our approach excels independently, demonstrating superior generalization without requiring such priors. + +# A.5 Results on 12 Scenes + +Table 6: The percentage of localization error under $2cm,2^{\circ}$ on 12 Scenes compared with other methods. The red and blue marks represent the first and second. The results are reported in Liu et al. (2024a). + +
Method2cm, 2° (↑)
APRMarepo Chen et al. (2024b)50.4
SCRDSAC* Brachmann & Rother (2021)96.7
ACE Brachmann et al. (2023)97.2
GLACE Wang et al. (2024a)97.5
3D GaussianMarepo Chen et al. (2024b) + GS-CPR Liu et al. (2025)90.9
Ours98.7
+ +$2cm, 2^{\circ}$ metric on 12 Scenes. To further evaluate localization performance, we conduct experiments on the 12 Scenes dataset using the $2cm, 2^{\circ}$ metric. Table 6 presents the results in comparison with other approaches. Our method achieves the highest accuracy in localization for the $2cm, 2^{\circ}$ metric. Compared to the 3D Gaussian based refinement method, GS-CPR Liu et al. (2025), our approach demonstrates achieves a $7.6\%$ improvement. + +# A.6 More Detailed Studies of GS-RelocNet + +Table 7: Scene-independent localization results on 7 Scenes with original ground truth and Cambridge Landmarks with different settings. + +
GS-RelocNet +3D Gaussian Number7 ScenesCambridge Landmarks
S15122.4/1.1414/0.36
S210242.0/0.9412/0.29
S320481.7/0.8910/0.22
S440961.4/0.829/0.21
S581921.5/0.889/0.21
S6163841.9/0.9712/0.28
S7Using Point Cloud instead of 3D Gaussians2.5/1.4316/0.45
Refinement Network
Rendering Method
S83D model1.8/0.9212/0.28
S9NeRF1.6/0.8410/0.23
S103D GS1.4/0.829/0.21
+ +Discussion of 3D Gaussian number in GS-RelocNet. In S1 - S6 of Table 7, we experiment with different numbers of 3D Gaussians. On one hand, the results with 512, 1024, and 16384 Gaussians are inferior to those with 4096 Gaussians, indicating that fewer Gaussians are insufficient for learning the scene features, while too many Gaussians increase the learning difficulty. On the other hand, the results with 2048, 4096, and 8192 Gaussians are comparable, suggesting that these configurations are sufficient for adequately learning the scene. + +Subsequently, a common concern may lie in the large scene with 4096 Gaussians. To address this, our framework may adopt a coarse-to-fine strategy. In the coarse stage, we uniformly sample 3D + +Gaussians across the entire scene and select those with high confidence scores, iterating this process as needed. In the refinement stage, we focus on Gaussians in proximity to the high-confidence selections, establishing 2D-3D correspondences between image pixels and these 3D Gaussians, followed by PnP to estimate the camera pose. + +Discussion of 3D Gaussian sampling strategy. For the sampling strategy, we uniformly sample 3D Gaussians within grid cells to ensure a sufficient number of correspondences for robust pose estimation. Within each grid cell, Gaussians are randomly selected for training. Although we use 4096 Gaussians per training iteration, multiple iterations allow most Gaussians to be utilized. To evaluate the robustness of this strategy, we conducted experiments using random sampling instead of uniform grid based sampling. The results show mean pose errors of $0.76\mathrm{cm} / 0.26^{\circ}$ in the scene-dependent setting on the 7-Scenes dataset and $7\mathrm{cm} / 0.18^{\circ}$ on the Cambridge Landmarks dataset. These results indicate minimal performance degradation, demonstrating the robustness of our sampling strategy. + +Discussion of usage of 3D Gaussians. In S7 of Table 7, we replace 3D Gaussians with the point cloud of the scene. The results clearly show a significant decrease in localization accuracy compared to using 3D Gaussians. We believe this is due to the fact that 3D Gaussians retain more texture and illumination information, while point clouds only capture geometric details. + +Discussion of Spherical Harmonics (SH) in 3D GS. To evaluate the contribution of SH in 3D GS, we exclude SH and perform experiments on the 7 Scenes and Cambridge Landmarks datasets without using SH as the input of GS-RelocNet. The scene-independent results reveal a consistent increase in localization error. On 7 Scenes, the error rises from $1.4 / 0.82$ to $1.7 / 0.93$ ( $\uparrow 0.3 / 0.11$ ), and on Cambridge Landmarks, it increases from $9 / 0.21$ to $10 / 0.26$ ( $\uparrow 1 / 0.05$ ). This observed increase in error across both datasets confirms the effectiveness of SH in enhancing the performance of 3D GS for localization tasks. + +Discussion of DINO v2 features. In GS-RelocNet, we also DINO v2 features as an additional input. To validate its effect, we conduct experiments without DINO v2 features. The results are $0.73 / 0.25$ ( $\uparrow 0.02 / 0.02$ with scene-dependent setting), $1.4 / 0.84$ ( $\uparrow 0.0 / 0.02$ with scene-independent setting) on 7 Scenes and $10 / 0.24$ ( $\uparrow 1 / 0.03$ ) on Cambridge Landmarks with scene-independent setting. We can see that the additional feature plays a slightly positive role on localization improvement. In the scene-dependent setting on 7 Scenes, it is even negligible. As an explanation, this is because that ScanNet contains large samples for training, making GS-RelocNet can learn the additional features itself. + +# A.7 More Detailed Studies of Refinement + +Discussion of rendering method. In S8 - S10 of Table 7, we present results using rendered images from both the 3D model and the NeRF network. The results demonstrate that the 3D GS based refinement yields the most significant improvements. When using rendered images from the 3D model, the results are comparable to those without refinement, likely due to the domain gap between rendered and real views. By incorporating NeRF and 3D GS, the lighting conditions are also considered, reducing this gap. + +Discussion of cross-attention layers. To explore the potential of cross-attention layers, we conducted two ablation studies in the pose refinement stage. First, we integrated a cross-attention layer into the bidirectional feature fusion module of the refinement network. This yielded mean pose errors of $0.73\mathrm{cm} / 0.24^{\circ}$ on the 7 Scenes dataset and $7\mathrm{cm} / 0.17^{\circ}$ on the Cambridge Landmarks dataset in the scene-dependent setting, which are comparable to our original results. Second, we replaced the refinement network with Dust3R ViT variant, incorporating cross-attention. The scene-dependent errors are $0.76\mathrm{cm} / 0.27^{\circ}$ on 7 Scenes and $9\mathrm{cm} / 0.21^{\circ}$ on Cambridge Landmarks, indicating slightly reduced accuracy. We attribute the limited impact of cross-attention layers to the small input resolution of the refinement network $(128\times 128)$ , which constrains their effectiveness. + +Discussion of environment conditions. Our current implementation does not include specific adaptations for environmental changes, such as variations in lighting or weather. Upon reviewing other 3D GS based methods, including GS-CPR and STDLoc, we found no explicit mention of specialized designs addressing such conditions. To mitigate the impact of environmental variations, we propose exploring 3D GS variants designed for enhanced robustness to environmental changes, as demonstrated in recent works Zhang et al. (2024); Kulhanek et al. (2024). These approaches could be integrated into our framework to improve performance under diverse conditions. + +Discussion of RANSAC. To predict the initial pose, we use the PnP with RANSAC using predicted confidence. To validate this, we also make the experiments with the traditional RANSAC using inlier number. The scene-independent results are $1.6 / 0.86$ ( $\uparrow 0.1 / 0.04$ ) and $10 / 0.24$ ( $\uparrow 1 / 0.03$ ) on Cambridge Landmarks. The results show a slight accuracy improvement with confidence in RANSAC. \ No newline at end of file diff --git a/NeurIPS/2025/3D Gaussian Splatting based Scene-independent Relocalization with Unidirectional and Bidirectional Feature Fusion/images.zip b/NeurIPS/2025/3D Gaussian Splatting based Scene-independent Relocalization with Unidirectional and Bidirectional Feature Fusion/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..7e2940ce9496344dde0814a300d5b6232fc12ca7 --- /dev/null +++ b/NeurIPS/2025/3D Gaussian Splatting based Scene-independent Relocalization with Unidirectional and Bidirectional Feature Fusion/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2d7fa3dcb423ae58f1f631d17d9f593ee6e16a8f702827752370b0359b42181b +size 829478 diff --git a/NeurIPS/2025/3D Gaussian Splatting based Scene-independent Relocalization with Unidirectional and Bidirectional Feature Fusion/layout.json b/NeurIPS/2025/3D Gaussian Splatting based Scene-independent Relocalization with Unidirectional and Bidirectional Feature Fusion/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..bde9d8cfa979465b28a8992e41f355ef6755c389 --- /dev/null +++ b/NeurIPS/2025/3D Gaussian Splatting based Scene-independent Relocalization with Unidirectional and Bidirectional Feature Fusion/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a71daf4039e15ad789189ef8ef973cd134d44c33cc21c7e0e80bcaf4423e9d80 +size 783729 diff --git a/NeurIPS/2025/3D Human Pose Estimation with Muscles/63b1c657-eda7-47cd-9f63-6882cee2567e_content_list.json b/NeurIPS/2025/3D Human Pose Estimation with Muscles/63b1c657-eda7-47cd-9f63-6882cee2567e_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..fb76cc17aad606c8145259b0def47401d0959dd1 --- /dev/null +++ b/NeurIPS/2025/3D Human Pose Estimation with Muscles/63b1c657-eda7-47cd-9f63-6882cee2567e_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ba68cb74f833b88a863a0b19be4b0d9d6d73553fd3ffd4fde113321ee0c320b1 +size 165761 diff --git a/NeurIPS/2025/3D Human Pose Estimation with Muscles/63b1c657-eda7-47cd-9f63-6882cee2567e_model.json b/NeurIPS/2025/3D Human Pose Estimation with Muscles/63b1c657-eda7-47cd-9f63-6882cee2567e_model.json new file mode 100644 index 0000000000000000000000000000000000000000..ecd253bf5d71ada4811d0dee1e6d509a6d7c270a --- /dev/null +++ b/NeurIPS/2025/3D Human Pose Estimation with Muscles/63b1c657-eda7-47cd-9f63-6882cee2567e_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e1384bbe4fba739920b2a325aa041fdb4d2f8538b1131dcd0f424ed4db2d83bb +size 215534 diff --git a/NeurIPS/2025/3D Human Pose Estimation with Muscles/63b1c657-eda7-47cd-9f63-6882cee2567e_origin.pdf b/NeurIPS/2025/3D Human Pose Estimation with Muscles/63b1c657-eda7-47cd-9f63-6882cee2567e_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..a51e2482284119c1ac9da857f5c66913cd8adaf5 --- /dev/null +++ b/NeurIPS/2025/3D Human Pose Estimation with Muscles/63b1c657-eda7-47cd-9f63-6882cee2567e_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7e29da355b29eb8af400d7f990c3b17e8d3f4e4614e3dcdebbd5346cc9962137 +size 2635448 diff --git a/NeurIPS/2025/3D Human Pose Estimation with Muscles/full.md b/NeurIPS/2025/3D Human Pose Estimation with Muscles/full.md new file mode 100644 index 0000000000000000000000000000000000000000..bcf234a59101c29fbf0e19d3400d2ae741ef7e02 --- /dev/null +++ b/NeurIPS/2025/3D Human Pose Estimation with Muscles/full.md @@ -0,0 +1,833 @@ +# 3D Human Pose Estimation with Muscles + +Kevin Zhu AliAsghar MohammadiNasrabadi Alexander Wong John McPhee + +University of Waterloo + +{k79zhu, aa27moha, a28wong, mcphee}@uwaterloo.ca + +# Abstract + +We introduce MusclePose as an end-to-end learnable physics-infused 3D human pose estimator that incorporates muscle-dynamics modeling to infer human dynamics from monocular video. Current physics pose estimators aim to predict physically plausible poses by enforcing the underlying dynamics equations that govern motion. Since this is an underconstrained problem without force-annotated data, methods often estimate kinetics with external physics optimizers that may not be compatible with existing learning frameworks, or are too slow for real-time inference. While more recent methods use a regression-based approach to overcome these issues, the estimated kinetics can be seen as auxiliary predictions, and may not be physically plausible. To this end, we build on existing regression-based approaches, and aim to improve the biofidelity of kinetic inference with a multihypothesis approach — by inferring joint torques via Lagrange's equations and via muscle dynamics modeling with muscle torque generators. Furthermore, MusclePose predicts detailed human anthropometrics based on values from biomechanics studies, in contrast to existing physics pose estimators that construct their human models with shape primitives. We show that MusclePose is competitive with existing 3D pose estimators in positional accuracy, while also able to infer plausible human kinetics and muscle signals consistent with values from biomechanics studies, without requiring an external physics engine. + +# 1 Introduction + +3D human pose estimation (HPE) is a fundamental task in computer vision that involves the localization of 3D human joints from images, which allows the user to track human movement from videos, leading to a plethora of potential downstream applications. However, since many pose estimators are purely data-driven, the inferred motion is modeled implicitly, which may lead to physically impossible poses and movements. + +Physics-based human pose estimation (PHPE) methods aim to mitigate these artifacts by enforcing the underlying dynamics equations that govern the kinematic state $\mathcal{K} = \{\pmb {q},\dot{\pmb{q}},\ddot{\pmb{q}}\}$ , + +$$ +\mathfrak {M} (\boldsymbol {q}, \mathcal {A}) \cdot \ddot {\boldsymbol {q}} + \mathfrak {C} (\boldsymbol {q}, \dot {\boldsymbol {q}}, \mathcal {A}) = \boldsymbol {\tau} _ {q} + \mathfrak {F} \tag {1} +$$ + +where $\mathbf{q}$ are generalized coordinates that describe motion, often in terms of translational (e.g. 3D position of the root) and rotational (e.g. joint rotations) degrees of freedom (DoF). We denote "dot" (\*) as the time derivative and "double dot" (\*" as the 2nd time derivative of a variable. $\mathfrak{M}$ is the mass matrix and $\mathfrak{C}$ contains the Coriolis, centrifugal, and gravitational forces, for a human with anthropometric features $\mathcal{A}$ at a given state $\kappa$ . Here, we loosely lump together a human's dimensions, mass and inertia properties, and other intrinsic and mobility features using the anthropometrics term $\mathcal{A}$ . On the right hand side, $\tau_{q}$ describes the human joint torques generated by each DoF, and $\mathfrak{F}$ are external forces, both in the generalized space. + +In this paper, we deal with monocular pose estimation, where our only input source is a monocular video, without force sensors. When only one unknown external force $\mathfrak{F}$ is applied on the human, + +![](images/12c3a6f1551c2a2fd81b788dc3e74bb77310a6f3b66054fc8d9cfa7934c0901a.jpg) +Figure 1: Overall framework of MusclePose. + +we can solve for $\tau_{q}$ by enforcing the entries of $\tau_{q}$ that corresponds to the root link to be zero. However, when there are more than one external force applied to the human simultaneously at different locations, which is often the case (e.g. when both feet are in contact with the ground), Eq. (1) becomes underconstrained. + +Without large-scale force-annotated video datasets, many methods estimate the corresponding kinetics via optimization with an external physics engine [20, 19, 77, 41]. However, these physics engines are often either non-differentiable and cannot be trained end-to-end, or are too slow for real-time inference. Furthermore, as discussed in [85], these methods are often combined with reinforcement learning to reach a desired outcome, but the effects of changing inputs on the outputs are unknown. Since joint torques can be hard to agree on in the biomechanics community, as they are often computed from different models, with different assumptions and post-processing, a more flexible learning framework may be preferred. More recently, PHPE methods have begun regressing kinetics directly with neural networks [85, 37, 63]. While the regression-based approach improves the kinematic reliability of the predicted motion, the inferred kinetics can be seen as auxiliary predictions, which may not be directly constrained and may be physically implausible. Although these kinetic predictions are not the main focus of these pose estimators, they may still be of interest for downstream applications. In sports for example, in addition to kinematics, practitioners and researchers are often interested in analyzing the whole-body musculoskeletal dynamics of athletes. To do so, a multibody model of skeletal dynamics is commonly used in combination with an optimal control algorithm to generate predictive simulations of athlete movements [5, 28, 49]. However, these optimal control algorithms can take hours or days to produce results. + +To this end, we build on existing regression-based PHPE approaches, to infer human kinetics simultaneously with kinematics, without a physics engine, and propose MusclePose (Fig. 1) to improve the plausibility of the predicted kinetics. To mitigate the underconstrained problem of regressing kinetics, we use a multihypothesis approach, and compute torques via Lagrange's equations, and also via muscle dynamics modeling with muscle torque generators (MTGs) [51, 25]. + +To maintain fidelity when modeling human movement, classical muscle models often represent muscles as linear actuators, and capture the nonlinear dependence of muscle tension on muscle length and the rate of lengthening [66, 32] using various Hill-type muscle models [23]. However, incorporating detailed muscles requires solving the actuator redundancy problem [3] and computing complex and varying musculoskeletal geometries [60, 12]. To overcome these drawbacks, parametric MTG models were proposed to mimic the behavior of muscles crossing a given joint to directly approximate joint torque by modeling kinematic dependence on active torque generation and passive impedance (Eq. (11)). Essentially, MTGs infer net joint torques from a joint's kinematics and activation levels, which is what we ultimately want, as we are not interested in isolated muscle tensions or granular joint contact forces. And since MTGs consist of differentiable equations, we are able to incorporate them into our learning framework, and train our pose estimator end-to-end. + +Moreover, for computational efficiency, existing PHPE methods rely on human models with anthropometrics estimated from the predicted human dimensions, or use the intrinsics properties (e.g. inertia + +and mass properties) of primitive shapes (e.g. spheres and simple rods), as proxies. From, Eq. (1), we see that, even if the kinematics state $\mathcal{K}$ and external forces $\mathfrak{F}$ are accurate, but $\mathcal{A}$ is not, the inferred torques $\tau_{q}$ may not correspond to the actual human performing the motion. For example, existing pose estimators may infer the center of mass (CoM) of body parts by taking the mean of the predicted surface mesh, assuming constant density [85]. However, since the composition of bones, muscles, internal organs, etc. is different, the human body's density is not uniform [14]. For example, the CoM of the upper torso is slightly towards the left side [16], whereas taking the mean vertices will be in the center. As such, we further predict detailed anthropometrics for each human, and keep them close to values taken from biomechanics studies. + +In summary, we introduce MusclePose to comprehensively predict human kinematics, kinetics, muscle signals, and detailed anthropometrics from monocular video. Specifically, we want a pose estimator with (i) a flexible learning framework easily adaptable for different scenarios, (ii) a reasonable degree of biofidelity, (iii) inference speed and (iv) positional accuracy both on par with purely kinematic pose estimators. To satisfy (i) and (iii), MusclePose is regression-based, consists of customizable and swappable components, can be trained end-to-end, and does not require an external physics engine. For (ii), MusclePose is the first pose estimator to incorporate muscle dynamics modeling and predict detailed human anthropometrics. We demonstrate improvements in the inferred kinetics on actions including walking from the H36M dataset [27] and baseball pitching and golf swings from PennAction [84]. Also, the use of MTGs allows us to further assess human motion at a musculoskeletal level, and we show that our inferred muscle signals are comparative to those from biomechanics studies, as well as to EMG data of pertinent muscle groups. Lastly, for (iv), we evaluate our method on benchmark 3D HPE datasets, H36M [27] and 3DPWoc, [71], to show that MusclePose is kinematically competitive with state-of-the-art (SOTA) pose estimators. + +![](images/c60636756476372514476cc7fab559b7655ff9081c5d449c04f46287c3833470.jpg) +Figure 2: Examples of MTG curves for hip flexion. $\tau_{\text{passive}}$ models the passive torque [79] as a double exponential function. $\tau_{\omega}$ models the active-torque-angular-speed relationship [69, 67] as a piecewise function. $\tau_{\theta}$ models the active-torque-angle relationship [21, 33] as the non-negative portion of a polynomial. + +# 2 Related work + +Monocular 3D human pose estimation. Early deep learning 3D HPE approaches use convolutional neural networks to directly estimate human 3D keypoint positions from images, with intermediate values represented by 3D heatmaps [57], location maps [50], or 2D heatmaps with depth regression [87]. The more recent and popular approach lifts 2D keypoints to 3D, essentially forming a monocular sparse-depth estimation task. The lifting network can be fully-connected layers [47], temporal convolution networks [10, 58], graph convolution networks [74, 7, 11], or transformers [89, 39, 86]. Human pose and shape estimation (HPSE) refers to predicting a 3D surface mesh of humans. The popular model-based HPSE approach [38, 35, 31] predicts input parameters of a parametric human model, such as SMPL, which infers a 3D mesh from rotation and shape parameters. Non-parametric approaches [42, 52, 36] directly regress 3D coordinates of mesh vertices. Other methods combine both approaches, such as [70], which predicts a volumetric representation before fitting a SMPL model, or [45], which calibrates model-based mesh predictions with 3D keypoints. + +Physics based pose estimation. To achieve more physically plausible human motion, PHPE methods apply dynamic constraints to encourage contact and penalize motion jitter, ground penetration, and unbalanced postures. [77, 63, 61, 64] model contact forces between the foot and ground, [20, 19, 82] include contact points between the full body and ground, while [41] also models human interaction with stick-like hand tools. Optimization-based frameworks [20, 19, 77, 64, 41] simulate physically plausible human motion from a physics engine and minimize an objective function to keep the simulated motion close to the detections obtained from a kinematic pose estimator. These frameworks are also combined with reinforcement learning [82, 59]. Recently, to overcome the need of an external physics engine, regression-based frameworks [85, 37, 63] directly estimate human kinetics using neural networks. As discussed in Sec. 1, this is an underconstrained problem, for which we hope to mitigate, and further inject biofidelity. We incorporate MTGs to do so. + +Muscle torque generators (MTGs). Due to their simplicity, MTGs have been increasingly popular in multibody dynamics simulations as they reduce computational cost while maintaining a reasonable degree of biofidelity. Recently, MTGs have been incorporated to simulate human movement post hip and knee replacement surgeries [13], human interactions with exoskeletons [22, 26], and manual wheelchair propulsion [5]. For more dynamic movements, such as in sports, examples of MTG-driven simulations of athlete motor control include golf [49] and cycling [28]. + +# 3 MusclePose + +We propose MusclePose (Fig. 1) as a physics-based pose estimator to directly regress comprehensive human dynamics from a monocular video of length $T$ . We use a transformer encoder to refine initial pose estimates, and produce latent motion features $\phi^{\{1:T\}}$ , which are used as inputs for 5 customizable modules to infer human anthropometrics $A$ , kinematics $\mathcal{K}$ , external forces $\mathcal{F}$ , joint torques via Lagrange's equations $\tau_{q}$ , and joint torques via MTGs $\tau_{MTG}$ . We describe our prototype in the following subsections, where we extract muscle signals $\alpha^{\{1:T\}}$ and residual terms $\mathcal{E}$ , $\delta^{\{1:T\}}$ , $\mathcal{3}$ from $\phi^{\{1:T\}}$ as inputs for the 5 modules. All variables described in this section are sequences of length $T$ , except $\mathcal{E}$ , $\mathcal{3}$ and shape parameters $\beta$ , and we drop the superscript $(\{1:T\})$ . + +Compared to the most recent regression-based PHPE method, PhysPT [85], we do not use a transformer decoder to directly regress joint torques, and instead, regress the input parameters of MTG models. Our motivation stems from the popular HPSE approach that regresses SMPL parameters instead of the human mesh directly, which not only reduces the computation complexity but also geometrically constrains the predicted human, as SMPL infers the mesh via forward kinematics. In parallel, we use MTGs to avoid estimating complex musculoskeletal geometries and granular joint contact forces, while enforcing a constraint on the inferred torques from Lagrange's equations. + +# 3.1 Kinematics estimation + +We follow the common approach in PHPE and clean initial kinematic estimates $\{\hat{\pmb{\theta}},\hat{\pmb{\beta}},\hat{\mathbf{T}}\}$ generated by some existing kinematic pose estimator, to obtain the refined $\{\pmb {\theta},\pmb {\beta},\mathbf{T},\mathbf{c}\}$ as our prediction. Here, $\mathbf{T}\in \mathbb{R}^3$ represents 3D pelvis translation in the world frame, and c are binary contact labels. Rotation parameters $\pmb {\theta} = \{\pmb {\theta}_0,\dots,\pmb {\theta}_{23}\}$ represents local rotations of the 24 SMPL keypoints, relative to their parents in the SMPL kinematic tree, with $\pmb{\theta}_0$ being the pelvis orientation in the world frame. We follow prior work [34] and predict the 6D continuous rotation representation [88] for each $\pmb {\theta}_k\in \mathbb{R}^6$ . Shape parameter $\beta \in \mathbb{R}^{10}$ denotes the first 10 principal components of SMPL's shape space. Since our inputs lack the shape information that RGB images provide, we follow the hybrid approach in [89] to regress shape residuals that are combined with initial predictions. We also use the same approach and regress anthropometric residuals $\mathcal{E}$ later on in Sec. 3.2 and force residuals $\delta$ in Sec. 3.3. + +The parametric human model, SMPL, then uses a collection of linear functions to map these parameters to a triangulated mesh $\mathcal{V}$ of 6890 vertices that represents the surface of the human body, and 24 SMPL keypoint positions $\mathbf{P}$ : + +$$ +\{\mathcal {V}, \mathbf {P} \} = \operatorname {S M P L} (\boldsymbol {\theta}, \boldsymbol {\beta}) + \mathbf {T} \tag {2} +$$ + +We define the kinematic loss $\mathcal{L}_{kin}$ with weights $\lambda_{kin}$ as + +$$ +\mathcal {L} _ {\text {k i n}} = \boldsymbol {\lambda} _ {\text {k i n}} \cdot \left[ \begin{array}{l l l l l l} \mathcal {L} _ {p} & \mathcal {L} _ {v} & \mathcal {L} _ {\theta} & \mathcal {L} _ {\beta} & \mathcal {L} _ {\text {n o r m}} & \mathcal {L} _ {c} \end{array} \right] ^ {\intercal} \tag {3} +$$ + +where the first five losses are from [89] which penalize joint position, linear velocity, SMPL parameter prediction L1 errors, and minimize the L2 norms of the SMPL parameters; and $\mathcal{L}_c$ is the binary contact loss from [83]. + +To facilitate multibody dynamics modeling in the following sections, we convert the predicted coordinates to generalized coordinates, $\pmb{q} = [X_0, q_0, q_1, \dots, q_{N_k}]^\intercal \in \mathbb{R}^{N_{DoF}}$ , where $\pmb{X}_0 \in \mathbb{R}^3$ is the global root translation, and each $\pmb{q}_k$ describes the joint's rotational DoFs. Specifically, each $q_i \in \pmb{q}_k$ are ZXY Euler angles converted from the predicted $\theta_k$ , to match the International Society of Biomechanics (ISB) format, where a joint's local $z$ -direction corresponds to flexion/extension, $x$ for abduction/adduction and $y$ for internal/external rotation. We denote the predicted kinematics as $\mathcal{K} = \{\pmb{q}, \dot{\pmb{q}}, \ddot{\pmb{q}}\}$ , with the velocity and acceleration terms estimated via finite differences. + +# 3.2 Human estimation + +Human model. We assume a rigid multibody dynamics model of a human with $N_{k} = 18$ segments and $N_{DoF} = 47$ total degrees of freedom (DoF). The 3D positions of the 18 joints, each corresponding to a segment, are the 24 SMPL keypoint positions minus the 5 end-effectors and the spine3 SMPL keypoint. The wrists, elbows, scapulas, each contain 2 rotational DoFs, knees each with 1 rotational DoF, root with 3 rotational and 3 translational DoFs, and 3 rotational DoFs for each of the remaining joints, for a total of 47 DoFs. We selected this configuration as it aligns best with biomechanics studies with anthropometric measurements that we use in the remaining sections. + +Anthropometrics prediction. To predict the human's anthropometrics $\mathcal{A} = \cup_{k}\{m_{k},I_{0,k},CoM_{k}\}$ , specifically the mass $m_{k}$ , inertia tensor at zero rotation $I_{0,k}$ , and CoM of all segments, we scale literature values $\bar{\mathcal{A}}$ from [16] based on the predicted human shapes $\beta$ and add the predicted offsets $\mathcal{E}$ + +$$ +\mathcal {A} = s _ {\beta} \bar {\mathcal {A}} + \mathcal {E} \tag {4} +$$ + +The scaling term $s_{\beta}$ is computed from the predicted $\beta$ , with details in the supplementary material. + +# 3.3 Kinetics estimation + +Ground reaction forces and moments (GRFM) prediction. Let $\mathcal{F}_k = [\mathbf{F}_k, \mathbf{M}_k]^{\top}$ be the GRFM applied on the CoM of each segment $k$ . We infer $\mathcal{F} = \sum_{k} \mathcal{F}_k$ from our previous predictions and our regressed force residuals $\delta$ , + +$$ +\mathcal {F} = \operatorname {G R F M} \operatorname {m o d e l} (\mathcal {K}, \mathcal {A}, \delta) \tag {5} +$$ + +Since we trained our model on the AMASS dataset [46] and feet-ground contact labels from RoHM [83], we assume feet-ground contact only for simplicity, as with many PHPE methods [37, 20, 82]. Omitting subscript $k$ , let $\mathbf{F} = [F_{X}, F_{Y}, F_{Z}]^{\intercal}$ be the force in world cartesian coordinates where $Y$ is the vertical direction, and let $\mathbf{z} = [z_{x}, z_{y}, z_{z}]^{\intercal}$ be the center of pressure (CoP) in the foot's local coordinates where $x$ is along the length of the foot (i.e. $\mathbf{M} = R_{ankle}^{0} \mathbf{z} \times \mathbf{F}$ where $R_{ankle}^{0}$ is the ankle's world orientation). From the regressed residuals $\delta_{\{Y,l\}} \subset \delta$ and the kinematics of each foot $\mathcal{K}_{foot}$ , we estimate the vertical force applied on the foot scaled by bodyweight $F_{Y}^{W} = F_{Y} / W$ , and the CoP along the foot scaled by foot length $z_{x}^{l} = z_{x} / l_{foot}$ , + +$$ +\left\{F _ {Y} ^ {W}, z _ {x} ^ {l} \right\} = \eta \mathcal {K} _ {\text {f o o t}} + \delta_ {\{Y, l \}} \tag {6} +$$ + +where linear coefficients $\eta$ were fitted on the forceplate data in [72]. The remaining $\delta$ terms are scaling factors between -1 and 1 to ensure the values in the other directions are physically possible (i.e. $F_{X}^{2} + F_{Z}^{2}\leq \delta_{\mu}^{2}F_{Y}^{2}$ and $\pmb{z}$ is within the foot's dimensions) + +$$ +F _ {X} = \delta_ {X} \delta_ {\mu} F _ {Y}, \quad F _ {Z} = \delta_ {Z} \sqrt {\delta_ {\mu} ^ {2} F _ {Y} ^ {2} - F _ {X} ^ {2}} \tag {7} +$$ + +$$ +z _ {y} = - \left| \delta_ {h} l _ {h} \right|, \quad z _ {z} = \delta_ {s} \left(l _ {w} / 2\right) \tag {8} +$$ + +where $l_w, l_h$ are the foot's width and height, respectively. Additional details of our GRFM model can be found in the supplementary material. + +Inverse dynamics via Lagrange's. From here, we can analytically compute the mass matrix $\mathfrak{M}$ , Coriolis term $\mathfrak{C}$ , external forces in the generalized space $\mathfrak{F}$ , and infer joint torques in the generalized space $\pmb{\tau}_{q}$ from the equations of motion: + +$$ +\tau_ {q} = \operatorname {L a g r a n g e} ^ {\prime} \mathrm {s} (\mathcal {K}, \mathcal {A}, \mathcal {F}) = \mathfrak {M} \ddot {\boldsymbol {q}} + \mathfrak {C} - \mathfrak {F} \tag {9} +$$ + +We include the calculations of these terms in the supplementary material. We define a residual force loss $\mathcal{L}_{res}$ to minimize the resulting forces and torques at the root, which correspond to the first 6 entries of $\pmb{\tau}_q$ + +$$ +\mathcal {L} _ {r e s} = \left| \tau_ {q [: 6 ]} \right| \tag {10} +$$ + +Inverse dynamics via MTGs. Simultaneously, we use parametric MTG models [51] to infer joint torques $\tau_{MTG}$ from the predicted kinematics $\mathcal{K}$ and muscle activations $\alpha$ . Specifically, this kinematic dependence is separated into active torque generation $\tau_{active}$ and passive impedance $\tau_{passive}$ . For each joint rotational DoF $q \in q_{[6:]}$ with angular velocity $\dot{q}$ , let muscle signal $\alpha \in [0,1]$ represent the joint's corresponding activation level for this DoF, we compute the corresponding torque as + +$$ +\tau_ {M T G} = \operatorname {M T G} (\mathcal {K}, \mathcal {A}, \alpha , \mathfrak {z}) = \tau_ {\text {a c t i v e}} + \tau_ {\text {p a s s i v e}} \tag {11} +$$ + +The active torque is further broken down into + +$$ +\tau_ {\text {a c t i v e}} = \alpha \cdot \tau_ {\omega} (\dot {q}) \cdot \tau_ {\theta} (q) \cdot \tau_ {0} (\mathcal {A}, 3) \tag {12} +$$ + +where $\tau_{\omega}(\dot{q};\gamma_{\omega})$ models the active-torque-angular-speed relationship [69, 67] and $\tau_{\theta}(q;\gamma_{\theta})$ models the active-torque-angle relationship [21, 33], as shown in Fig. 2. These relationships are parameterized by the $\gamma$ coefficients, which are unique for each joint's DoF and direction, and are identified via dynamometry. This joint-dependent parameterization preserves physiological realism (e.g., hip flexion and knee extension should exhibit different peak torque and passive stiffness profiles), unlike uniform torque models that assume identical properties across the body. For this paper, we use the set of $\gamma$ values summarized in [54, 53]. + +$\tau_0(\mathcal{A}; \gamma_i, \gamma_e)$ is the peak isokinetic torque that controls peak MTG output at zero joint velocity, which can be measured with a dynamometer. $\tau_0$ is estimated in [54, 53] as a linear approximation of the human's intrinsic, scaled by certain external factors such as the human's fitness or activity level. Since these external factors (and some intrinsic properties) are not readily known, we take the mean effects $\gamma_i, \gamma_e$ from [54, 53], and add regressed offsets 3. We compute $\tau_0$ as + +$$ +\tau_ {0} = \left(\gamma_ {i} \mathcal {A} + 3 _ {i}\right) \left(\gamma_ {e} + 3 _ {e}\right) \tag {13} +$$ + +Furthermore, to account for stability, we assume each joint is driven by a pair of agonist-antagonist MTGs — a flexor (+) and an extensor (-), that corresponds to the movement direction. Hence, for each joint rotational DoF, we regress 2 muscle signals $\{\alpha^{flex},\alpha^{ext}\}$ , and the active torque becomes: + +$$ +\tau_ {\text {a c t i v e}} = \alpha^ {\text {f l e x}} \tau_ {\omega} ^ {\text {f l e x}} \tau_ {\theta} ^ {\text {f l e x}} \tau_ {0} ^ {\text {f l e x}} + \alpha^ {\text {e x t}} \tau_ {\omega} ^ {\text {e x t}} \tau_ {\theta} ^ {\text {e x t}} \tau_ {0} ^ {\text {e x t}} \tag {14} +$$ + +$\tau_{passive}(q; \gamma_p)$ is the passive torque [1] of a joint that arises when the surrounding muscles, tendons, and ligaments are strained and intensifies near anatomical joint limits [1, 81]. The joint's viscous damping and nonlinear stiffness are parameterized by $\gamma_p$ , which encourages the joint to move within its range of motion, as a large restoring torque is produced otherwise, as shown in Fig. 2. Equations to compute $\tau_\omega, \tau_\theta, \tau_{passive}$ can be found in the supplementary material. + +We define the torque loss $\mathcal{L}_{\tau}$ as the absolute difference between the two sets of predicted joint torques, and another regularizing term $\mathcal{L}_{\epsilon}$ for all regressed residuals: + +$$ +\mathcal {L} _ {\tau} = \left| \boldsymbol {\tau} _ {q [ 6: ]} - \boldsymbol {\tau} _ {M T G} \right| \tag {15} +$$ + +$$ +\mathcal {L} _ {\epsilon} = \left| \left| \mathcal {E} \right| \right| _ {2} + \left| \left| \delta \right| \right| _ {2} + \left| \left| 3 \right| \right| _ {2} \tag {16} +$$ + +Finally, we have dynamic loss with weights $\lambda_{dyn}$ + +$$ +\mathcal {L} _ {\text {d y n}} = \boldsymbol {\lambda} _ {\text {d y n}} \cdot \left[ \begin{array}{l l l} \mathcal {L} _ {\tau} & \mathcal {L} _ {\text {r e s}} & \mathcal {L} _ {\epsilon} \end{array} \right] ^ {\intercal} \tag {17} +$$ + +# 4 Experiments + +# 4.1 Implementation and datasets + +For training, we used the AMASS dataset [46], with feet-ground contact labels from [83]. As such, we trained and evaluated on sequences with feet-ground contact only (denoted $\dagger$ ), which is also the case for many PHPE experiments [37, 20, 82]. We trained MusclePose end-to-end with a sequence input length of 16 frames, using total loss $\mathcal{L}_{total} = \mathcal{L}_{kin} + \mathcal{L}_{dyn}$ for 25 epochs, using the AdamW optimizer [44] with a weight decay of $10^{-4}$ and an initial learning rate of $10^{-4}$ that decreases by $20\%$ every 5 epochs. Following common curriculum learning [2] practices, we split the training into two phases — for the first 20 epochs, we trained using the ground truth as input, followed by 5 epochs using the model's predictions as inputs. + +For evaluation, we assessed positional accuracy on the inference results of the H36M test set [27] and object-occlusion subset of 3DPW (3DPWoc) [71]. As with training, we removed input sequences containing non-feet-ground contact, the sitting and sitting down actions in H36M, and courtyard laceshoe, flat guitar, outdoors climbing, outdoors freestyle, outdoors parcours, downtown stairs in 3DPWoc. We further assessed kinetic biofidelity from 3 actions — walking from H36M, and baseball pitch and golf swing from the PennAction dataset (PA) [84]. As with existing large-scale human video datasets, since neither datasets include force-annotations, we compared our inference results with existing biomechanics studies of these movements, and commented on overall trends and plausibility. We selected walking because human gait is heavily studied in biomechanics [72, 18, 76, 8, 29, 75], and is a relatively consistent and cyclic movement. We included the latter two actions to evaluate faster movements, for which we were able to find published lab measurements [55, 80, 62]. During inference, to promote a closer comparison with the SOTA regression-based physics pose estimator PhysPT [85], we used the same kinematic estimator, CLIFF [40], to extract initial kinematic estimates, and the global trajectory predictor in [85] to extract initial root DoFs. The rationale for using CLIFF in [85] is that it produces competitive positional accuracy but lacks in physical plausibility. + +# 4.2 Positional accuracy + +We followed standard evaluation protocol and reported the mean per-joint positional error (MJE) and procrustes-aligned MJE (PJE) in Tab. 1, for the 14 LSP [30] keypoints in millimetres. MJE is the root-aligned mean Euclidean distance in millimeters between the predicted and ground truth 3D keypoints. PJE is the MJE after aligning the predicted pose with the ground truth in translation, rotation, and scale using the Procrustes method. For 3DPWoc, since the data is captured using a moving camera with unknown extrinsics, and our method predicts the global root DoFs directly, we reported PJE only, with additional ablations in Tab. 3 to show consistency of results. + +We see that MusclePose outperforms other PHPE methods on H36M but is slightly worse than PhysPT on 3DPWoc. This, along with the overall worse positional accuracy of PHPE compared to purely kinematic methods, could be due to a kinematics-kinetics trade-off, as our method produced a lower residual force (Tab. 4). Specifically, the root DoFs in 3DPW may be harder to estimate due to the moving camera, leading to higher residual forces, which the model may try to reduce (lower kinetic error) by estimating a set of local joint kinematics slightly different from the original motion (higher kinematic error). + +Table 1: Positional accuracy on H36M and 3DPWoc. "opti" denotes kinetics obtained from an external physics optimizer, and "repr" denotes regressed by a neural network. $\dagger$ denotes sequences with feet-ground contact only. + +
KineticsH36M3DPWoc PJE↓
MJE↓PJE↓
kinematicHybrIK [38]-55.433.6-
HybrIK [38]-†56.4†36.7-
CLIFF [40]-52.236.8-
CLIFF [40]-†46.5†32.4†24.0
PHPESimPoE [82]opti.†56.7†41.6-
DiffPhy [20]opti.†81.7†55.6-
D&D [37]repr.†52.5†35.5-
PhysPT [85]repr.†50.6†35.5†25.9
MusclePose(ours)repr.†48.4†33.5†27.6
+ +# 4.3 Biofidelity + +Since most PHPE methods do not report kinetics results, we mainly compared ours with values we reproduced from PhysPT. For even comparison, for both pose estimators, we computed joint torques in the generalized space $\pmb{\tau}_{q}$ via Lagrange's equations (9) from the predicted motion, anthropometrics, and GRFM. Unlike the biomechanics studies, we did not apply additional signal post-processing or smoothing to $\pmb{\tau}_{q}$ . Hence, we see more noise and spikes in the pose estimators' results, which could also be amplified by the low frame rate of PennAction. + +Qualitative. Since the different biomechanics studies computed torques differently from different datasets, we comment on general trends and evaluate qualitatively. In Fig. 3, we plotted median torques scaled by predicted body weight, with a $25 - 75\%$ quantile band, for select joints of the 3 actions, for which we found reference values. Overall, compared to PhysPT (gray), we see that MusclePose (ours, purple) more closely follows the trends and magnitudes of the reference values (greens and yellow). + +![](images/2f60b53c00f8895bae2f16c5867ff7cb0b213dabe6dba0f4912f308b35c25585.jpg) +Figure 3: Median predicted joint torques scaled by bodyweight, with a $25 - 75\%$ quantile band, for gait cycles, the downswing phase of golf drives, and the arm acceleration phase of baseball pitches, compared to values from biomechanics studies. + +The 2 leftmost columns of Fig. 3 correspond to flexion/extension of the hip, knee, and ankle torques for walking, scaled such that the toe-off occurs at $60\%$ of the gait cycle, and compared to reference torques from [73, 18, 75]. We see that MusclePose produced more reasonable trends overall, whereas PhysPT produced torques with low magnitudes but with higher extreme values. For hip flexion, our results resembled more of the yellow curve that was computed from a wearable system in [73], where the authors attributed their errors to a lack of shear force measurement, leading to more noticeable errors in the hips than more distal joints due to the increase in moment arms. Since the subjects in H36M walk in a small circle with slower and varying speeds, a discrepancy can arise in the generated shear force, leading to the discrepancy in hip torque, as the subjects in the biomechanics studies walk in a straight line at a consistent pace. This could also contribute to the low magnitude of ankle torque, where different timings (toe-off/heel-off/touch-down) could affect ankle power generation, as explained in [6]. + +The middle 2 columns of Fig. 3 correspond to lumbar torques during the downswing phase of golf drives, scaled such that the maximum lumbar rotation occurs at $2/3$ of the motion, and compared to reference torques from [55]. While we see noticeable spikes from both pose estimators, the extreme values for lumbar lateral bending and axial rotation were much higher for PhysPT. + +The 2 rightmost columns of Fig. 3 correspond to shoulder lateral/medial rotation torque during the arm acceleration phase of baseball pitches, scaled such that the maximum moment occurs at $80\%$ of the motion, and compared to reference torques from professional pitchers (darker green) in [62], as well as amateurs (lighter green) in [80]. Although the peak of our $75\%$ quantile slightly exceeds the shoulder medial rotation limit reported in [62], our band was able to cover values from both skill groups, whereas PhysPT's band was below the amateurs, even though the pitchers in PA range from teenage amateurs to adult professionals. + +Quantitative. In the bottom right of Fig. 3, we reported the mean residual forces $\{F_{res},\tau_{res}\}$ , mean out of range joint torques $\tau_{OOR}$ , and the median value of the sum of GRFs in the direction opposite of gravity $\mathrm{GRF}_v$ . Residual $F_{res}$ and $\tau_{res}$ were computed as the mean L2 norms of the entries of $\tau_q$ that correspond to the translational and rotational DoFs of the root. $\tau_{OOR}$ was computed as the mean absolute amount outside of joint torque limits (red and blue values in Fig. 3) reported in [1, 43, 62]. These values are further scaled by predicted body weight and denoted $(^{/W})$ . In the last column, we also reported the mean predicted segment mass $m^{/M}$ as a percentage of body mass $M$ . + +Overall, we see that MusclePose inferred more reasonable kinetics, indicated by the lower residual forces, less extreme joint torques, and a median $\mathrm{GRF}_v$ closer to body weight. Similar to its joint torques, PhysPT's $\mathrm{GRF}_v$ values were overall lower in magnitude, with occasional spikes. While MusclePose's maximum $\mathrm{GRF}_v$ were very large for the golf swing and baseball pitch, our $99\%$ quantiles were more comparable to the maximum values of about 1.3 times body weight for golf reported in [55], and about 2.3 for pitching in [9]. + +![](images/f336f1d4267e5edcd5f489b4cd3755a268bc3269cf855986019d5eaea33984fe.jpg) +Figure 4: Mean predicted muscle activations (dashed) for gait cycles compared to EMG data and predictions of pertinent muscles from literature. Min-max scaling applied to all values. + +# 4.4 Ablations and additional evaluation + +We reported ablations results in Tab. 2. Row 2 includes results without custom anthropometrics, and instead computed directly from the SMPL mesh, assuming constant density. Row 3 includes results without MTGs. Row 4 includes results using kinematic-based muscle activations, with details in the next paragraph. We see that, overall, MusclePose has better positional accuracy, along with lower residual forces $(\mathbf{F}_{res} = (F_{res} + \tau_{res}) / 2)$ , and median $\mathrm{GRF}_v$ closest to bodyweight. + +Table 2: Ablations results. + +
\( ^\dagger \)H36M\( ^\dagger \)3DPWocH36M - WalkPA - Golf swingPA - Pitch
MJE↓\( F_{res}^{\prime W} \)↓PJE↓\( F_{res}^{\prime W} \)↓\( F_{res}^{\prime W} \)↓med GRFvW\( F_{res}^{\prime W} \)↓med GRFvW\( F_{res}^{\prime W} \)↓med GRFvW
MusclePose48.40.0827.60.250.080.980.560.980.680.99
w/o A49.50.2627.90.350.220.740.610.640.670.39
w/o \( {\tau }_{MTG} \)49.70.1427.20.300.150.940.660.880.740.90
with \( \alpha \)50.90.1628.40.260.130.930.590.970.690.90
+ +Furthermore, due to the lack of extrinsic information in 3DPWoc, we reported results from using different kinematic estimators in Tab. 3 to show consistency. + +Table 3: ${}^{ \dagger }$ 3DPWoc results with different kinematic estimators. + +
Ours+CLIFFCLIFF [40]Ours+WHAMWHAM [65]Ours+CoMotionCoMotion [56]
PJE↓27.624.028.522.732.630.7
FWres↓0.3-0.3-0.2-
+ +Muscle activations. Since we were not able to find public video datasets with corresponding MTG activation signals to directly compare with, we followed the evaluation procedure in [29] to assess general trends, and overlayed our mean muscle activation predictions (black, dashed) for gait cycles from H36M with EMG data of pertinent muscles from other gait studies [72, 76, 8, 29] in rows 1 and 3 of Fig. 4, with min-max scaling applied to all values. We also included the predicted activations (dotted) from [29]; however, they use a different muscle model. While muscle activations can be seen as surrogate representations of EMGs, the two are not exactly the same. Hence, mismatches in timing and magnitude will exist, and peaks and valleys may be further amplified by the min-max scaling applied. In general, raw EMG values can vary widely due to electrode placement. + +To mimic methods that regress joint torques as a linear combination of joint kinematics, we experimented with estimating the muscle activation as a "kinematic effort term" (denoted $\hat{\alpha}$ ), specifically as a joint's angular velocity relative to its limit plus an additionally regressed offset term: + +$$ +\hat {\alpha} ^ {d} = \dot {q} ^ {d} / \dot {q} _ {\text {m a x}} ^ {d} + \mathfrak {Z} _ {\alpha} ^ {d}, \quad d \in \{f l e x, e x t \} \tag {18} +$$ + +We plotted $\hat{\alpha}$ results (gray, dashed) in rows 2 and 4 of Fig. 4. In comparison, MusclePose (rows 1 and 3) seems to better follow literature trends overall, such as having a more noticeable hitch (or second peak) for hip adduction, external rotation, knee flexion, etc. Furthermore, for the $\hat{\alpha}$ case, ankle + +plantarflexion seems to be deactivated during the middle of the gait cycle, when it should peak. Row 4 of Tab. 2 also shows that MusclePose quantitatively outperforms the $\hat{\alpha}$ case. + +Kinematic plausibility. In addition to the joint positional errors in Sec. 4.2, metrics such as acceleration loss (ACC), foot skating (FS), and ground penetration (GP) were introduced to further evaluate kinematic plausibility. We computed these values for H36M and 3DPWoc in Tab. 4, where ACC is the mean L2 norm in mm/frame $^2$ between the predicted and ground truth keypoint accelerations to access jitter. We also included mean torque variation (MTV) as the mean absolute change in joint torques over consecutive frames (in Newton\*metres/frame) to assess torque continuity. FS is the average displacement in mm of vertices in contact with the ground in consecutive frames. GP is the average vertical distance to the ground in mm of vertices below the ground. + +Table 4: Plausibility metrics. ${}^{P}$ indicates Procrustes aligned. (%f) indicates % of frames. + +
Pos. MJE↓Kinematic plausibilityFloat (%f)Kinetic plausibilityGRFw/20q99 maxGRFw/20q99 max
ACC↓FSGPHmin > {1,10,20} mmFres↓MTVmed
H36MCLIFF46.526.3----------
PhysPT50.613.734.76.8{59.031.68.5}0.45.3{0.42.410.0}{7.220.860.0}
MusclePose48.412.937.226.0{8.03.01.3}0.12.5{1.01.23.0}{3.03.05.2}
3DPWocCLIFF24.0P13.8P----------
PhysPT25.9P3.0P7.811.2{82.973.957.2}0.927.0{0.51.23.9}{5.311.852.4}
MusclePose27.6P4.3P12.830.8{6.04.73.7}0.312.1{1.01.64.3}{5.35.36.3}
+ +While we see an improvement in jitter from both physics pose estimators, as indicated by a lower ACC compared to CLIFF, a lower FS or GP may not be strictly better. For one, foot sliding may occur naturally. And in terms of the latter, while human bodies deform under pressure and contact, the SMPL mesh does not model this deformation, and will instead penetrate the object it is in contact with [68]. During walking for example, minimizing GP while assuming this rigidity may restrict natural ankle rotation, potentially leading to the smaller ankle torques compared to literature values in Fig. 3. On the other hand, we should also check for floating. We reported the percentage of frames $(\% \mathrm{f})$ when the minimum vertex height $\mathcal{H}_{min}$ is above certain thresholds (1, 10, 20mm). Since we removed the non-feet-ground contact sequences, there are very minimal frames where "floating" occurs. We see that while PhysPT has less GP, it also includes more floating. + +Ground reaction force. We can also characterize floating as when $\mathrm{GRF}_v$ is small. As such, we also reported the percentage of frames when $\mathrm{GRF}_v$ is below certain thresholds (1%, 10%, 50% of body weight) in Tab. 4. The results are consistent with $\mathcal{H}_{\text{min}}$ , both indicating less floating (lower %f) for MusclePose. In Fig. 5, we plotted the predicted median vertical GRF of a foot, divided by body weight, for gait cycles in H36M. Compared to PhysPT (gray), we see that MusclePose's predictions (purple) are closer to literature values (greens) from [18, 75]. + +![](images/194f4012e265ae4f6a762481afd6408ab4aace3cf50c52237fe00159cadc52ba.jpg) +Figure 5: Median vertical GRF divided by body weight, of a foot for gait cycles in H36M, with a $25 - 75\%$ quantile band. + +# 5 Conclusion + +In conclusion, we introduced MusclePose as the first PHPE method to simultaneously predict human kinematics, kinetics, muscle signals, and detailed anthropometrics from monocular video. In Sec. 4.3 and 4.4, we showed how the additions of muscle-dynamics modeling and detailed anthropometrics predictions improve the kinetic plausibility of regression-based PHPE, while being competitive with purely-kinematic pose estimators in positional accuracy in Sec. 4.2. Our framework consists of customizable components, does not require an external physics engine, and can be trained end-to-end. + +Acknowledgements. We acknowledge financial support from the Canada Research Chairs Program, Canadian Sports Institute Ontario, and a Mitacs grant. + +# References + +[1] D. E. Anderson, M. L. Madigan, and M. A. Nussbaum. Maximum voluntary joint torque as a function of joint angle and angular velocity: Model development and application to the lower limb. Journal of Biomechanics, 40(14):3105-3113, 2007. +[2] Y. Bengio, J. Louradour, R. Collobert, and J. Weston. Curriculum learning. In Proceedings of the 26th Annual International Conference on Machine Learning, ICML '09, page 41-48, 2009. +[3] N. A. Bernshtein. The co-ordination and regulation of movements. 1967. +[4] C. Brown, W. McNally, and J. McPhee. Optimal control of joint torques using direct collocation to maximize ball carry distance in a golf swing. Multibody System Dynamics, 50(3), 2020. +[5] C. Brown and J. J. McPhee. Predictive forward dynamic simulation of manual wheelchair propulsion on a rolling dynamometer. Journal of biomechanical engineering, 2020. +[6] A. Buchmann, S. Wenzler, L. Welte, and D. Renjewski. The effect of including a mobile arch, toe joint, and joint coupling on predictive neuromuscular simulations of human walking. Scientific Reports, 2024. +[7] Y. Cai, L. Ge, J. Liu, J. Cai, T. J. Cham, J. Yuan, and N. M. Thalmann. Exploiting spatial-temporal relationships for 3D pose estimation via graph convolutional networks. In Proceedings of the IEEE International Conference on Computer Vision, volume 2019-October, 2019. +[8] T. Castermans, M. Duvinage, G. Cheron, and T. Dutoit. Towards effective non-invasive brain-computer interfaces dedicated to gait rehabilitation systems. *Brain Sciences*, 4(1):1-48, 2014. +[9] S.-W. Chen, W.-T. Tang, J.-T. Kung, T.-Y. Hung, W.-H. Lin, Y.-L. Chen, and D. J. Burgee. Comparison of ground reaction force among stride types in baseball pitching. Sports Biomechanics, 0(0):1-14, 2024. +[10] Y. Cheng, B. Yang, B. Wang, and R. T. Tan. 3D human pose estimation using spatio-temporal networks with explicit occlusion training. In AAAI 2020 - 34th AAAI Conference on Artificial Intelligence, 2020. +[11] H. Ci, C. Wang, X. Ma, and Y. Wang. Optimizing network structure for 3D human pose estimation. In Proceedings of the IEEE International Conference on Computer Vision, volume 2019-October, 2019. +[12] D. J. Cleather and A. M. J. Bull. Lower-extremity musculoskeletal geometry affects the calculation of patellofemoral forces in vertical jumping and weightlifting. Proceedings of the Institution of Mechanical Engineers, Part H: Journal of Engineering in Medicine, 224:1073 - 1083, 2010. +[13] B. Danaei and J. J. McPhee. Model-based acetabular cup orientation optimization based on minimizing the risk of edge-loading and implant impingement following total hip arthroplasty. Journal of biomechanical engineering, 2022. +[14] R. Drillis, R. Contini, and M. Bluestein. Body segment parameters. Artificial limbs, 8(1):44-66, 1964. +[15] G. A. Dudley, R. T. Harris, M. R. Duvoisin, B. M. Hather, and P. Buchanan. Effect of voluntary vs. artificial activation on the relationship of muscle torque to speed. Journal of Applied Physiology, 69(6), 1990. +[16] R. Dumas, L. Chéze, and J. P. Verriest. Adjustments to mcconville et al. and young et al. body segment inertial parameters. Journal of biomechanics, 40 3:543-53, 2007. +[17] R. Featherstone. Rigid Body Dynamics Algorithms. 2008. +[18] C. A. Fukuchi, R. K. Fukuchi, and M. Duarte. A public dataset of overground and treadmill walking kinematics and kinetics in healthy individuals. *PeerJ*, 6, 2018. +[19] E. Gartner, M. Andriluka, H. Xu, and C. Sminchisescu. Trajectory Optimization for Physics-Based Reconstruction of 3d Human Pose from Monocular Video. In Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition, volume 2022-June, 2022. +[20] E. Gartner, M. Andriluka, E. Coumans, and C. Sminchisescu. Differentiable dynamics for articulated 3d human motion reconstruction. In Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 2022. +[21] D. Haering, C. Pontonnier, N. Bideau, G. Nicolas, and G. Dumont. Using Torque-Angle and Torque-Velocity Models to Characterize Elbow Mechanical Function: Modeling and Applied Aspects. Journal of Biomechanical Engineering, 141(8), 2019. + +[22] N. Haraguchi, A. Nasr, K. A. Inkol, K. Hase, and J. McPhee. Human and passive lower-limb exoskeleton interaction analysis: Computational study with dynamics simulation using nonlinear model predictive control. 2023 62nd Annual Conference of the Society of Instrument and Control Engineers (SICE), pages 844-849, 2023. +[23] A. V. Hill. The heat of shortening and the dynamic constants of muscle. Proceedings of The Royal Society B: Biological Sciences, 126:136-195, 1938. +[24] P. D. Hoang, R. B. Gorman, G. Todd, S. C. Gandevia, and R. D. Herbert. A new method for measuring passive length-tension properties of human gastrocnemius muscle in vivo. Journal of Biomechanics, 38(6):1333-1341, 2005. +[25] K. A. Inkol, C. Brown, W. McNally, C. Jansen, and J. McPhee. Muscle torque generators in multibody dynamic simulations of optimal sports performance. Multibody System Dynamics, 50(4), 2020. +[26] K. A. Inkol and J. J. McPhee. Using dynamic simulations to estimate the feasible stability region of feet-in-place balance recovery for lower-limb exoskeleton users. 2022 9th IEEE RAS/EMBS International Conference for Biomedical Robotics and Biomechatronics (BioRob), pages 1-6, 2022. +[27] C. Ionescu, D. Papava, V. Olaru, and C. Sminchisescu. Human3.6M: Large scale datasets and predictive methods for 3D human sensing in natural environments. IEEE Transactions on Pattern Analysis and Machine Intelligence, 36(7), 2014. +[28] C. Jansen and J. J. McPhee. Predictive dynamic simulation of olympic track cycling standing start using direct collocation optimal control. Multibody System Dynamics, 49:53-70, 2020. +[29] C. T. John, F. C. Anderson, J. S. Higginson, and S. L. D. and. Stabilisation of walking by intrinsic muscle properties revealed in a three-dimensional muscle-driven simulation. Computer Methods in Biomechanics and Biomedical Engineering, 16(4):451-462, 2013. +[30] S. Johnson and M. Everingham. Clustered pose and nonlinear appearance models for human pose estimation. In British Machine Vision Conference, 2010. +[31] A. Kanazawa, M. J. Black, D. W. Jacobs, and J. Malik. End-to-End Recovery of Human Shape and Pose. In Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 2018. +[32] B. Katz. The relation between force and speed in muscular contraction. The Journal of Physiology, 96, 1939. +[33] M. A. King, C. Wilson, and M. R. Yeadon. Evaluation of a torque-driven model of jumping for height. Journal of Applied Biomechanics, 22(4), 2006. +[34] M. Kocabas, N. Athanasiou, and M. J. Black. Vibe: Video inference for human body pose and shape estimation. In Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 2020. +[35] N. Kolotouros, G. Pavlakos, M. J. Black, and K. Daniilidis. Learning to reconstruct 3d human pose and shape via model-fitting in the loop. In Proceedings of the International Conference on Computer Vision, 2019. +[36] N. Kolotouros, G. Pavlakos, and K. Daniilidis. Convolutional mesh regression for single-image human shape reconstruction. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4496-4505, 2019. +[37] J. Li, S. Bian, C. Xu, G. Liu, G. Yu, and C. Lu. D &D: Learning Human Dynamics from Dynamic Camera. In ECCV, 2022. +[38] J. Li, C. Xu, Z. Chen, S. Bian, L. Yang, and C. Lu. Hybrid analytical-neural inverse kinematics solution for 3d human pose and shape estimation. In Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 2021. +[39] W. Li, H. Liu, H. Tang, P. Wang, and L. Van Gool. MHFormer: Multi-Hypothesis Transformer for 3D Human Pose Estimation. In Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition, volume 2022-June, 2022. +[40] Z. Li, J. Liu, Z. Zhang, S. Xu, and Y. Yan. Cliff: Carrying location information in full frames into human pose and shape estimation. In European Conference on Computer Vision, 2022. + +[41] Z. Li, J. Sedlar, J. Carpentier, I. Laptev, N. Mansard, and J. Sivic. Estimating 3D motion and forces of person-object interactions from monocular video. In Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition, volume 2019-June, 2019. +[42] K. Lin, L. Wang, and Z. Liu. Mesh graphormer. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 12919-12928, 2021. +[43] D. M. Lindsay and J. F. Horton. Trunk rotation strength and endurance in healthy normals and elite male golfers with and without low back pain. North American journal of sports physical therapy : NAJSPT, 1 2:80-9, 2006. +[44] I. Loshchilov and F. Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2017. +[45] T. Luan, Y. Wang, J. Zhang, Z. Wang, Z. Zhou, and Y. Qiao. Pc-hmr: Pose calibration for 3d human mesh recovery from 2d images/videos. In AAAI Conference on Artificial Intelligence, 2021. +[46] N. Mahmood, N. Ghorbani, N. F. Troje, G. Pons-Moll, and M. Black. AMASS: Archive of motion capture as surface shapes. In Proceedings of the IEEE International Conference on Computer Vision, volume 2019-October, 2019. +[47] J. Martinez, R. Hossain, J. Romero, and J. J. Little. A Simple Yet Effective Baseline for 3d Human Pose Estimation. In Proceedings of the IEEE International Conference on Computer Vision, volume 2017-October, 2017. +[48] W. McNally and J. McPhee. Dynamic Optimization of the Golf Swing Using a Six Degree-of-Freedom Biomechanical Model. 2018. +[49] W. J. McNally and J. J. McPhee. Dynamic optimization of the golf swing using a six degree-of-freedom biomechanical model. 2018. +[50] D. Mehta, H. Rhodin, D. Casas, P. Fua, O. Sotnychenko, W. Xu, and C. Theobalt. Monocular 3D human pose estimation in the wild using improved CNN supervision. In Proceedings - 2017 International Conference on 3D Vision, 3DV 2017, 2018. +[51] M. Millard, T. K. Uchida, A. Seth, and S. L. Delp. Flexing computational muscle: modeling and simulation of musculotendon dynamics. Journal of biomechanical engineering, 135 2:021005, 2013. +[52] G. Moon and K. M. Lee. I21-meshnet: Image-to-lixel prediction network for accurate 3d human pose and mesh estimation from a single rgb image. ArXiv, abs/2008.03713, 2020. +[53] A. Nasr, A. Hashemi, and J. McPhee. Scalable musculoskeletal model for dynamic simulations of upper body movement. Computer Methods in Biomechanics and Biomedical Engineering, 2023. +[54] A. Nasr and J. McPhee. Scalable musculoskeletal model for dynamic simulations of lower body movement. Computer methods in biomechanics and biomedical engineering, pages 1-27, 2024. +[55] S. Nesbit. Development of a full-body biomechanical model of the golf swing. International Journal of Modelling and Simulation, 27(4):392-404, 2007. +[56] A. Newell, P. Hu, L. Lipson, S. R. Richter, and V. Koltun. Comotion: Concurrent multi-person 3d motion. In ICLR, 2025. +[57] G. Pavlakos, X. Zhou, K. G. Derpanis, and K. Daniilidis. Coarse-to-fine volumetric prediction for single-image 3D human pose. In Proceedings - 30th IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2017, volume 2017-January, 2017. +[58] D. Pavllo, C. Feichtenhofer, D. Grangier, and M. Auli. 3D human pose estimation in video with temporal convolutions and semi-supervised training. In Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition, volume 2019-June, 2019. +[59] X. B. Peng, P. Abbeel, S. Levine, and M. Van De Panne. DeepMimic: Example-guided deep reinforcement learning of physics-based character skills. ACM Transactions on Graphics, 37(4), 2018. +[60] E. D. Pieri, M. Lund, A. Gopalakrishnan, K. P. Rasmussen, D. E. Lunn, and S. J. Ferguson. Refining muscle geometry and wrapping in the tlem 2 model for improved hip contact force prediction. PLoS ONE, 13, 2018. + +[61] D. Rempe, L. J. Guibas, A. Hertzmann, B. Russell, R. Villegas, and J. Yang. Contact and Human Dynamics from Monocular Video. In 19th ACM SIGGRAPH / Eurographics Symposium on Computer Animation 2020, SCA 2020 - Showcases, 2020. +[62] M. Sabick, M. Torry, Y.-K. Kim, and R. Hawkins. Humeral torque in professional baseball pitchers. The American journal of sports medicine, 32:892-8, 07 2004. +[63] S. Shimada, V. Golyanik, W. Xu, P. Pérez, and C. Theobalt. Neural monocular 3D human motion capture with physical awareness. ACM Transactions on Graphics, 40(4), 2021. +[64] S. Shimada, V. Golyanik, W. Xu, and C. Theobalt. Phys Cap. ACM Transactions on Graphics, 39(6), 2020. +[65] S. Shin, J. Kim, E. Halilaj, and M. J. Black. Wham: Reconstructing world-grounded humans with accurate 3d motion. In CVPR, 2024. +[66] T. Siebert, C. Rode, W. Herzog, O. Till, and R. Blickhan. Nonlinearities make a difference: comparison of two common hill-type models with real muscle. Biological Cybernetics, 98:133-143, 2008. +[67] E. J. Sprigings. Simulation of the force enhancement phenomenon in muscle. Computers in Biology and Medicine, 16(6), 1986. +[68] S. Tripathi, L. Müller, C.-H. P. Huang, O. Taheri, M. J. Black, and D. Tzionas. 3d human pose estimation via intuitive physics. ArXiv, abs/2303.18246, 2023. +[69] A. J. van Soest and M. F. Bobbert. The contribution of muscle properties in the control of explosive movements. Biological Cybernetics, 69(3), 1993. +[70] G. Varol, D. Ceylan, B. C. Russell, J. Yang, E. Yumer, I. Laptev, and C. Schmid. Bodynet: Volumetric inference of 3d human body shapes. ArXiv, abs/1804.04875, 2018. +[71] T. von Marcard, R. Henschel, M. Black, B. Rosenhahn, and G. Pons-Moll. Recovering accurate 3d human pose in the wild using imus and a moving camera. In European Conference on Computer Vision (ECCV), sep 2018. +[72] H. Wang, A. Basu, G. Durandau, and M. Sartori. Comprehensive Kinetic and EMG Dataset of Daily Locomotion with 6 types of Sensors, May 2022. +[73] H. Wang, A. Basu, G. Durandau, and M. Sartori. A wearable real-time kinetic measurement sensor setup for human locomotion. Wearable Technologies, Feb. 2023. +[74] J. Wang, S. Yan, Y. Xiong, and D. Lin. Motion Guided 3D Pose Estimation from Videos. In Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), volume 12358 LNCS, 2020. +[75] D. A. Winter. Biomechanics and motor control of human movement. John Wiley & Sons, Hoboken, NJ, USA, 4th edition, 2009. +[76] A. R. Wu, F. Dzeladini, T. J. H. Brug, F. Tamburella, N. L. Tagliamonte, E. H. F. van Asseldonk, H. van der Kooij, and A. J. Ijspeert. An adaptive neuromuscular controller for assistive lower-limb exoskeletons: A preliminary study on subjects with spinal cord injury. Frontiers in Neurorobotics, Volume 11 - 2017, 2017. +[77] K. Xie, T. Wang, U. Iqbal, Y. Guo, S. Fidler, and F. Shkurti. Physics-based Human Motion Estimation and Synthesis from Videos. In Proceedings of the IEEE International Conference on Computer Vision, 2021. +[78] G. T. Yamaguchi. Dynamic Modeling of Musculoskeletal Motion. 2001. +[79] G. T. Yamaguchi. Dynamic modeling of musculoskeletal motion: A vectorized approach for biomechanical analysis in three dimensions. Springer, Boston, MA, USA, 1 edition, 2006. +[80] K. Yoichi, E. Sato, and T. Yamaji. Biomechanical analysis of the pitching characteristics of adult amateur baseball pitchers throwing standard and lightweight balls. Journal of Physical Therapy Science, 32:816-822, 12 2020. +[81] Y. S. Yoon and J. M. Mansour. The passive elastic moment at the hip. Journal of Biomechanics, 15(12):905-910, 1982. +[82] Y. Yuan, S.-E. Wei, T. Simon, K. Kitani, and J. Saragih. Simpoe: Simulated character control for 3d human pose estimation. In Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 2021. + +[83] S. Zhang, B. L. Bhatnagar, Y. Xu, A. Winkler, P. Kadlecek, S. Tang, and F. Bogo. Rohm: Robust human motion reconstruction via diffusion. In CVPR, 2024. +[84] W. Zhang, M. Zhu, and K. G. Derpanis. From actemes to action: A strongly-supervised representation for detailed action understanding. In 2013 IEEE International Conference on Computer Vision, pages 2248-2255, 2013. +[85] Y. Zhang, J. O. Kephart, Z. Cui, and Q. Ji. Physpt: Physics-aware pretrained transformer for estimating human dynamics from monocular videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2305-2317, June 2024. +[86] C. Zheng, S. Zhu, M. Mendieta, T. Yang, C. Chen, and Z. Ding. 3D Human Pose Estimation with Spatial and Temporal Transformers. In Proceedings of the IEEE International Conference on Computer Vision, 2021. +[87] X. Zhou, Q. Huang, X. Sun, X. Xue, and Y. Wei. Towards 3D Human Pose Estimation in the Wild: A Weakly-Supervised Approach. In Proceedings of the IEEE International Conference on Computer Vision, volume 2017-October, 2017. +[88] Y. Zhou, C. Barnes, J. Lu, J. Yang, and H. Li. On the continuity of rotation representations in neural networks. In Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 2019. +[89] W. Zhu, X. Ma, Z. Liu, L. Liu, W. Wu, and Y. Wang. Learning human motion representations: A unified perspective. In Proceedings of the International Conference on Computer Vision, 2023. + +# NeurIPS Paper Checklist + +# 1. Claims + +Question: Do the main claims made in the abstract and introduction accurately reflect the paper's contributions and scope? + +Answer: [Yes] + +Justification: our contributions are summarized in the last paragraph of Sec. 1, with corresponding results in Sec. 4. + +Guidelines: + +- The answer NA means that the abstract and introduction do not include the claims made in the paper. +- The abstract and/or introduction should clearly state the claims made, including the contributions made in the paper and important assumptions and limitations. A No or NA answer to this question will not be perceived well by the reviewers. +- The claims made should match theoretical and experimental results, and reflect how much the results can be expected to generalize to other settings. +- It is fine to include aspirational goals as motivation as long as it is clear that these goals are not attained by the paper. + +# 2. Limitations + +Question: Does the paper discuss the limitations of the work performed by the authors? + +Answer: [Yes] + +Justification: limitations are discussed in the supplementary material. + +Guidelines: + +- The answer NA means that the paper has no limitation while the answer No means that the paper has limitations, but those are not discussed in the paper. +- The authors are encouraged to create a separate "Limitations" section in their paper. +- The paper should point out any strong assumptions and how robust the results are to violations of these assumptions (e.g., independence assumptions, noiseless settings, model well-specification, asymptotic approximations only holding locally). The authors should reflect on how these assumptions might be violated in practice and what the implications would be. +- The authors should reflect on the scope of the claims made, e.g., if the approach was only tested on a few datasets or with a few runs. In general, empirical results often depend on implicit assumptions, which should be articulated. +- The authors should reflect on the factors that influence the performance of the approach. For example, a facial recognition algorithm may perform poorly when image resolution is low or images are taken in low lighting. Or a speech-to-text system might not be used reliably to provide closed captions for online lectures because it fails to handle technical jargon. +- The authors should discuss the computational efficiency of the proposed algorithms and how they scale with dataset size. +- If applicable, the authors should discuss possible limitations of their approach to address problems of privacy and fairness. +- While the authors might fear that complete honesty about limitations might be used by reviewers as grounds for rejection, a worse outcome might be that reviewers discover limitations that aren't acknowledged in the paper. The authors should use their best judgment and recognize that individual actions in favor of transparency play an important role in developing norms that preserve the integrity of the community. Reviewers will be specifically instructed to not penalize honesty concerning limitations. + +# 3. Theory assumptions and proofs + +Question: For each theoretical result, does the paper provide the full set of assumptions and a complete (and correct) proof? + +Answer: [Yes] + +Justification: all formulas used can be found in Sec. 3, with details in the supplementary material. + +# Guidelines: + +- The answer NA means that the paper does not include theoretical results. +- All the theorems, formulas, and proofs in the paper should be numbered and cross-referenced. +- All assumptions should be clearly stated or referenced in the statement of any theorems. +- The proofs can either appear in the main paper or the supplemental material, but if they appear in the supplemental material, the authors are encouraged to provide a short proof sketch to provide intuition. +- Inversely, any informal proof provided in the core of the paper should be complemented by formal proofs provided in appendix or supplemental material. +- Theorems and Lemmas that the proof relies upon should be properly referenced. + +# 4. Experimental result reproducibility + +Question: Does the paper fully disclose all the information needed to reproduce the main experimental results of the paper to the extent that it affects the main claims and/or conclusions of the paper (regardless of whether the code and data are provided or not)? + +# Answer: [Yes] + +Justification: all models implemented can be found in Sec. 3, with details, such as model coefficients or hyperparameters, listed in the supplementary material. Experiment implementation information, for both training and evaluation, can be found in Sec. 4.1, with additional details in the supplementary material. + +# Guidelines: + +- The answer NA means that the paper does not include experiments. + +- If the paper includes experiments, a No answer to this question will not be perceived well by the reviewers: Making the paper reproducible is important, regardless of whether the code and data are provided or not. + +- If the contribution is a dataset and/or model, the authors should describe the steps taken to make their results reproducible or verifiable. + +- Depending on the contribution, reproducibility can be accomplished in various ways. For example, if the contribution is a novel architecture, describing the architecture fully might suffice, or if the contribution is a specific model and empirical evaluation, it may be necessary to either make it possible for others to replicate the model with the same dataset, or provide access to the model. In general, releasing code and data is often one good way to accomplish this, but reproducibility can also be provided via detailed instructions for how to replicate the results, access to a hosted model (e.g., in the case of a large language model), releasing of a model checkpoint, or other means that are appropriate to the research performed. + +- While NeurIPS does not require releasing code, the conference does require all submissions to provide some reasonable avenue for reproducibility, which may depend on the nature of the contribution. For example + +(a) If the contribution is primarily a new algorithm, the paper should make it clear how to reproduce that algorithm. +(b) If the contribution is primarily a new model architecture, the paper should describe the architecture clearly and fully. +(c) If the contribution is a new model (e.g., a large language model), then there should either be a way to access this model for reproducing the results or a way to reproduce the model (e.g., with an open-source dataset or instructions for how to construct the dataset). +(d) We recognize that reproducibility may be tricky in some cases, in which case authors are welcome to describe the particular way they provide for reproducibility. In the case of closed-source models, it may be that access to the model is limited in some way (e.g., to registered users), but it should be possible for other researchers to have some path to reproducing or verifying the results. + +# 5. Open access to data and code + +Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material? + +Answer: [No] + +Justification: while our code is not yet ready for open source, we included in the supplementary material, all necessary details required to reproduce our model and results, including implementation details of all novel components, as well as all the publicly available code repositories we borrowed from to implement and train our model, and to produce and evaluate our results. + +Guidelines: + +- The answer NA means that paper does not include experiments requiring code. +- Please see the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details. +- While we encourage the release of code and data, we understand that this might not be possible, so "No" is an acceptable answer. Papers cannot be rejected simply for not including code, unless this is central to the contribution (e.g., for a new open-source benchmark). +- The instructions should contain the exact command and environment needed to run to reproduce the results. See the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details. +- The authors should provide instructions on data access and preparation, including how to access the raw data, preprocessed data, intermediate data, and generated data, etc. +- The authors should provide scripts to reproduce all experimental results for the new proposed method and baselines. If only a subset of experiments are reproducible, they should state which ones are omitted from the script and why. +- At submission time, to preserve anonymity, the authors should release anonymized versions (if applicable). +- Providing as much information as possible in supplemental material (appended to the paper) is recommended, but including URLs to data and code is permitted. + +# 6. Experimental setting/details + +Question: Does the paper specify all the training and test details (e.g., data splits, hyperparameters, how they were chosen, type of optimizer, etc.) necessary to understand the results? + +Answer: [Yes] + +Justification: training and test information can be found in Sec. 4.1, with additional details in the supplementary material. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The experimental setting should be presented in the core of the paper to a level of detail that is necessary to appreciate the results and make sense of them. +- The full details can be provided either with the code, in appendix, or as supplemental material. + +# 7. Experiment statistical significance + +Question: Does the paper report error bars suitably and correctly defined or other appropriate information about the statistical significance of the experiments? + +Answer: [Yes] + +Justification: we include quantile bands in Fig. 3. + +Guidelines: + +- The answer NA means that the paper does not include experiments. + +- The authors should answer "Yes" if the results are accompanied by error bars, confidence intervals, or statistical significance tests, at least for the experiments that support the main claims of the paper. +- The factors of variability that the error bars are capturing should be clearly stated (for example, train/test split, initialization, random drawing of some parameter, or overall run with given experimental conditions). +- The method for calculating the error bars should be explained (closed form formula, call to a library function, bootstrap, etc.) +- The assumptions made should be given (e.g., Normally distributed errors). +- It should be clear whether the error bar is the standard deviation or the standard error of the mean. +- It is OK to report 1-sigma error bars, but one should state it. The authors should preferably report a 2-sigma error bar than state that they have a $96\%$ CI, if the hypothesis of Normality of errors is not verified. +- For asymmetric distributions, the authors should be careful not to show in tables or figures symmetric error bars that would yield results that are out of range (e.g. negative error rates). +- If error bars are reported in tables or plots, The authors should explain in the text how they were calculated and reference the corresponding figures or tables in the text. + +# 8. Experiments compute resources + +Question: For each experiment, does the paper provide sufficient information on the computer resources (type of compute workers, memory, time of execution) needed to reproduce the experiments? + +Answer: [Yes] + +Justification: computing information is included in the supplementary material. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The paper should indicate the type of compute workers CPU or GPU, internal cluster, or cloud provider, including relevant memory and storage. +- The paper should provide the amount of compute required for each of the individual experimental runs as well as estimate the total compute. +- The paper should disclose whether the full research project required more compute than the experiments reported in the paper (e.g., preliminary or failed experiments that didn't make it into the paper). + +# 9. Code of ethics + +Question: Does the research conducted in the paper conform, in every respect, with the NeurIPS Code of Ethics https://neurips.cc/public/EthicsGuidelines? + +Answer: [Yes] + +Justification: the research conducted in the paper conform with the NeurIPS Code of Ethics. + +Guidelines: + +- The answer NA means that the authors have not reviewed the NeurIPS Code of Ethics. +- If the authors answer No, they should explain the special circumstances that require a deviation from the Code of Ethics. +- The authors should make sure to preserve anonymity (e.g., if there is a special consideration due to laws or regulations in their jurisdiction). + +# 10. Broader impacts + +Question: Does the paper discuss both potential positive societal impacts and negative societal impacts of the work performed? + +Answer: [Yes] + +Justification: societal impacts are discussed in the supplementary material. + +Guidelines: + +- The answer NA means that there is no societal impact of the work performed. +- If the authors answer NA or No, they should explain why their work has no societal impact or why the paper does not address societal impact. +- Examples of negative societal impacts include potential malicious or unintended uses (e.g., disinformation, generating fake profiles, surveillance), fairness considerations (e.g., deployment of technologies that could make decisions that unfairly impact specific groups), privacy considerations, and security considerations. +- The conference expects that many papers will be foundational research and not tied to particular applications, let alone deployments. However, if there is a direct path to any negative applications, the authors should point it out. For example, it is legitimate to point out that an improvement in the quality of generative models could be used to generate deepfakes for disinformation. On the other hand, it is not needed to point out that a generic algorithm for optimizing neural networks could enable people to train models that generate Deepfakes faster. +- The authors should consider possible harms that could arise when the technology is being used as intended and functioning correctly, harms that could arise when the technology is being used as intended but gives incorrect results, and harms following from (intentional or unintentional) misuse of the technology. +- If there are negative societal impacts, the authors could also discuss possible mitigation strategies (e.g., gated release of models, providing defenses in addition to attacks, mechanisms for monitoring misuse, mechanisms to monitor how a system learns from feedback over time, improving the efficiency and accessibility of ML). + +# 11. Safeguards + +Question: Does the paper describe safeguards that have been put in place for responsible release of data or models that have a high risk for misuse (e.g., pretrained language models, image generators, or scraped datasets)? + +Answer: [NA] + +Justification: the paper poses no such risk. + +Guidelines: + +- The answer NA means that the paper poses no such risks. +- Released models that have a high risk for misuse or dual-use should be released with necessary safeguards to allow for controlled use of the model, for example by requiring that users adhere to usage guidelines or restrictions to access the model or implementing safety filters. +- Datasets that have been scraped from the Internet could pose safety risks. The authors should describe how they avoided releasing unsafe images. +- We recognize that providing effective safeguards is challenging, and many papers do not require this, but we encourage authors to take this into account and make a best faith effort. + +# 12. Licenses for existing assets + +Question: Are the creators or original owners of assets (e.g., code, data, models), used in the paper, properly credited and are the license and terms of use explicitly mentioned and properly respected? + +Answer: [Yes] + +Justification: assets used in this paper are cited in the main text, and the license and terms are respected and mentioned in the supplementary material. + +Guidelines: + +- The answer NA means that the paper does not use existing assets. +- The authors should cite the original paper that produced the code package or dataset. +- The authors should state which version of the asset is used and, if possible, include a URL. +- The name of the license (e.g., CC-BY 4.0) should be included for each asset. + +- For scraped data from a particular source (e.g., website), the copyright and terms of service of that source should be provided. +- If assets are released, the license, copyright information, and terms of use in the package should be provided. For popular datasets, paperswithcode.com/datasets has curated licenses for some datasets. Their licensing guide can help determine the license of a dataset. +- For existing datasets that are re-packaged, both the original license and the license of the derived asset (if it has changed) should be provided. +- If this information is not available online, the authors are encouraged to reach out to the asset's creators. + +# 13. New assets + +Question: Are new assets introduced in the paper well documented and is the documentation provided alongside the assets? + +Answer: [Yes] + +Justification: the paper introduces MusclePose as a novel human pose estimation framework, with details on implementation and application. + +Guidelines: + +- The answer NA means that the paper does not release new assets. +- Researchers should communicate the details of the dataset/code/model as part of their submissions via structured templates. This includes details about training, license, limitations, etc. +- The paper should discuss whether and how consent was obtained from people whose asset is used. +- At submission time, remember to anonymize your assets (if applicable). You can either create an anonymized URL or include an anonymized zip file. + +# 14. Crowdsourcing and research with human subjects + +Question: For crowdsourcing experiments and research with human subjects, does the paper include the full text of instructions given to participants and screenshots, if applicable, as well as details about compensation (if any)? + +Answer: [NA] + +Justification: this paper does not involve crowdsourcing, and uses human data from publicly available datasets. + +Guidelines: + +- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. +- Including this information in the supplemental material is fine, but if the main contribution of the paper involves human subjects, then as much detail as possible should be included in the main paper. +- According to the NeurIPS Code of Ethics, workers involved in data collection, curation, or other labor should be paid at least the minimum wage in the country of the data collector. + +# 15. Institutional review board (IRB) approvals or equivalent for research with human subjects + +Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals (or an equivalent approval/review based on the requirements of your country or institution) were obtained? + +Answer: [NA] + +Justification: this paper does not involve crowdsourcing, and uses human data from publicly available datasets. + +Guidelines: + +- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. +- Depending on the country in which research is conducted, IRB approval (or equivalent) may be required for any human subjects research. If you obtained IRB approval, you should clearly state this in the paper. +- We recognize that the procedures for this may vary significantly between institutions and locations, and we expect authors to adhere to the NeurIPS Code of Ethics and the guidelines for their institution. +- For initial submissions, do not include any information that would break anonymity (if applicable), such as the institution conducting the review. + +# 16. Declaration of LLM usage + +Question: Does the paper describe the usage of LLMs if it is an important, original, or non-standard component of the core methods in this research? Note that if the LLM is used only for writing, editing, or formatting purposes and does not impact the core methodology, scientific rigorousness, or originality of the research, declaration is not required. + +Answer: [NA] + +Justification: LLMs were not used for research development nor any part of this paper. + +Guidelines: + +- The answer NA means that the core method development in this research does not involve LLMs as any important, original, or non-standard components. +- Please refer to our LLM policy (https://neurips.cc/Conferences/2025/LLM) for what should or should not be described. + +# Supplementary material for 3D Human Pose Estimation with Muscles + +# A Technical appendix + +# A.1 Human model + +We assume a rigid multibody dynamics model of a human with $N_{k} = 18$ joints - pelvis, lumbar joint, thoracic joint, neck, scapulas, shoulders, elbows, wrists, hips, knees, and ankles. The pelvis is set as the root, with 3 rotational and 3 translational degrees of freedom (DoFs). The scapulas have 2 DoFs, corresponding to depression/elevation and protraction/retraction. The elbows have 2 DoFs, corresponding to flexion/extension and forearm pronation/supination. The wrists have 2 DoFs, corresponding to flexion/extension and ulnar/radial deviation. The knees have 1 DoF, corresponding to flexion/extension. All remaining joints have 3 DoFs, for a total of 47 DoFs. We selected this configuration as it aligns best with existing biomechanics models that we implemented, such as for anthropometrics estimation [16] and MTGs [51, 25]. + +Anthropometrics estimation. We predict the human's anthropometrics by combining our regressed residuals $\mathcal{E}$ with initial estimates based on literature values $\bar{\mathcal{A}}$ scaled by our predicted human dimensions. Specifically, we want to predict $\mathcal{A} = \cup_{k}\{m_{k},I_{0,k},\chi_{k}\}$ , where $m_{k}$ is the mass of segment $k$ , with $I_{0,k}$ as its inertia tensor at zero rotation with scaling matrix $\Lambda_{k}$ , and $\chi_{k}$ as its local CoM position relative to its segment length. For the remainder of the section, we assume relevant units to be in seconds, radians, meters, kilograms, Newtons. + +From the predicted $\beta$ and Eq. (2), we can compute the human's volume and all segment lengths $L_{k}$ . We further compute the human's initial bodymass estimate $\hat{M}$ as its volume multiplied by a constant density of $985~kg / m^2$ . For segment $k$ , let $s_{L,k} = L_k / H$ be its segment length relative to height, and $s_{m,k} = m_k / M$ be its mass relative to bodymass. Let "bar" ( $-$ ) denote the human values measured by Dumas et al. in [16]. We set our initial estimates as $\bar{\mathcal{A}}$ scaled by $s_{L,k} / \bar{s}_{L,k}$ : + +$$ +\{M, s _ {m, k} ^ {\prime}, \Lambda_ {k}, \chi_ {k} \} = \{\hat {M}, \bar {s} _ {m, k} \frac {s _ {L , k}}{\bar {s} _ {L , k}}, \bar {\Lambda} _ {k}, \bar {\chi} _ {k} \} + \mathcal {E} \tag {19} +$$ + +$$ +m _ {k} = s _ {m, k} M, \text {w h e r e} s _ {m, k} = \frac {s _ {m , k} ^ {\prime}}{\sum_ {j} s _ {m , j} ^ {\prime}} \tag {20} +$$ + +$$ +I _ {0, k} = m _ {k} L _ {k} ^ {2} \Lambda_ {k} \tag {21} +$$ + +Lastly, we compute the body weight $W$ of the human in Newtons, with $g = 9.8m / s^2$ as + +$$ +W = g \sum_ {k} m _ {k} \tag {22} +$$ + +# A.2 GRFM Model + +Let $\mathcal{F} = [\mathbf{F},\mathbf{M}]^{\intercal}$ be the ground reaction forces and moments (GRFM) applied at the CoM of a foot in global cartesian coordinates. Let $\mathbf{F} = [F_X,F_Y,F_Z]^{\intercal}$ where $Y$ is the vertical direction, and $\mathbf{z} = [z_x,z_y,z_z]^{\intercal}$ be the center of pressure (CoP) in the foot's local coordinates where $x$ is along the length of the foot, such that + +$$ +\mathbf {M} = R _ {\text {a n k l e}} ^ {0} \mathbf {z} \times \mathbf {F} \tag {23} +$$ + +where $R_{k}$ is joint $k$ 's local rotation matrix, and $R_{k}^{0} = R_{p(k)}^{0}R_{k}$ describes the chain of rotational transformations from the world frame to its local frame. We use lower case $x,y,z$ to denote the foot's local coordinates, and its dimensions $\{l_l,l_w,l_h\}$ as shown in Fig. 6. + +We predict the force in the vertical direction scaled by body weight $F_{Y}^{W} = F_{Y} / W$ , and CoP along the foot scaled by foot length $z_{x}^{l} = z_{x} / l_{l}$ , from initial estimates based on the foot's kinematics $\Psi$ and linear coefficients $\eta$ . Furthermore, let $\mu$ be the coefficient of friction, initialized at 0.8. With our regressed residuals $\delta$ , and binary contact $\mathbf{c}$ , we infer: + +$$ +\left\{F _ {Y} ^ {W}, z _ {x} ^ {l}, \mu \right\} = \left\{\eta_ {F Y} \Psi , \eta_ {z x} \Psi , 0. 8 \right\} + \delta_ {\{Y, l, \mu \}} \tag {24} +$$ + +$$ +F _ {Y} = F _ {Y} ^ {W} \cdot \text {b o d y w e i g h t} \tag {25} +$$ + +$$ +z _ {x} = z _ {x} ^ {l} \cdot l _ {l} \tag {26} +$$ + +Specifically, $\Psi = [1, P_{ankle,Y}, P_{oppAnkle,Y}, \dot{P}_{ankle,Y}, \ddot{P}_{ankle,Y}, q_{ankle,z}, \dot{q}_{ankle,z}, \ddot{q}_{ankle,z},]$ includes the ankle's linear kinematics in the direction opposite of gravity, and angular kinematics corresponding to plantar/dorsiflexion. Linear coefficients were fitted on the forceplate data in [72], with $\eta_{FY} = [0.3116, 3.1785, -2.2963, 0.4151, 0.0088, 0.3374, -0.1206, -0.0089]$ and $\eta_{zx} = [0.68996, -3.1508, 0.5925, 0.21997, 0.0035, 0.18502, -0.03311, -0.00212]$ . + +The remaining $\delta$ terms are scaling factors between -1 and 1 to ensure the values in the other directions are physically possible (i.e. $F_{X}^{2} + F_{Z}^{2}\leq \mu^{2}F_{Y}^{2}$ and $\pmb{z}$ is within the foot's dimensions) + +$$ +F _ {X} = \delta_ {X} \mu F _ {Y}, \quad F _ {Z} = \delta_ {Z} \sqrt {\mu^ {2} F _ {Y} ^ {2} - F _ {X} ^ {2}} \tag {27} +$$ + +$$ +z _ {y} = - \left| \delta_ {h} l _ {h} \right|, \quad z _ {z} = \delta_ {s} \left(l _ {w} / 2\right) \tag {28} +$$ + +![](images/3c745305a5b3e22211b8795ff101c0b563473786e4f618307c5ace5619d89c20.jpg) +Figure 6: Foot local coordinate system and dimensions, with length $l_{l} = l_{f} + l_{b}$ , width $l_{w} = 2l_{s}$ and CoM height $l_{h}$ . + +# A.3 Muscle torque generators + +We compute MTG torque $\tau_{MTG}$ using the equations below that are parameterized by the $\gamma$ coefficients that can be found in the tables of [54, 53]. For each joint rotational DoF $q \in q_{[6:]}$ , with angular velocity $\dot{q}$ , let muscle signal $\alpha \in [0,1]$ represent the joint's corresponding activation level for this DoF, we separate its $\tau_{MTG}$ into active torque generation $\tau_{active}$ and passive impedance $\tau_{passive}$ + +$$ +\tau_ {M T G} = \tau_ {\text {a c t i v e}} + \tau_ {\text {p a s s i v e}} \tag {29} +$$ + +We compute the active torque as + +$$ +\tau_ {\text {a c t i v e}} = \alpha \tau_ {\omega} \tau_ {\theta} \tau_ {0} \tag {30} +$$ + +where $\tau_{\omega}$ models the active-torque-angular-speed relationship [69, 67] and is paramerized as a piecewise function with coefficients $\gamma_{1:3}$ . + +$$ +\tau_ {\omega} (\dot {q}) = \mathbb {1} _ {\dot {q} < 0} \left(\frac {(1 - \gamma_ {1}) | \omega_ {\max } | - (\gamma_ {2} + 1) \gamma_ {1} \gamma_ {3} \dot {q}}{(1 - \gamma_ {1}) | \omega_ {\max } | + (\gamma_ {2} + 1) \gamma_ {1} \dot {q}} + \mathbb {1} _ {\dot {q} \geq 0} \left(\frac {| \omega_ {\max } | - \dot {q}}{| \omega_ {\max } | + \gamma_ {2} \dot {q}}\right) \right. \tag {31} +$$ + +The peak velocity $\omega_{max}$ for each joint we use the values from [54, 53]. The coefficient $\gamma_{1}$ is the ratio of the maximum eccentric isokinetic torque over the maximum isometric torque [69, 15], $\gamma_{2}$ is the slope of the eccentric and concentric functions when the angular velocity is zero [69], and $\gamma_{3}$ is a shape factor that influences the curvature of the hyperbola in the torque-velocity concentric relationship [4]. + +$\tau_{\theta}$ models the active-torque-angle relationship [21, 33] and is represented by the non-negative portion of a polynomial (32) with coefficients $\gamma_{4:6}$ + +$$ +\tau_ {\theta} (q) = \left(\gamma_ {4} + \gamma_ {5} q + \gamma_ {6} q ^ {2}\right) _ {+} \tag {32} +$$ + +$\tau_0$ is the peak isokinetic torque that controls peak MTG output at zero joint velocity, which can be measured via dynamometry. + +$\tau_{passive}$ is the passive torque [1] of a joint that arises when the surrounding muscles, tendons, and ligaments are strained and intensifies near anatomical joint limits [1, 24, 81]. A joint's viscous damping and nonlinear stiffness are commonly described by a double exponential function [79] + +$$ +\tau_ {\text {p a s s i v e}} = \gamma_ {1 0 e} ^ {- \gamma_ {1 1} (\boldsymbol {q} - \boldsymbol {q} _ {\min })} - \gamma_ {1 2} e ^ {\gamma_ {1 3} (\boldsymbol {q} - \boldsymbol {q} _ {\max })} - \gamma_ {1 4} \omega \tag {33} +$$ + +where $\gamma_{10 - 14}$ are passive coefficients from [48] and $\gamma_{11}$ is the rotational damping linear coefficient [78] to reflect viscoelasticity. This encourages the joint to move within its range of motion (RoM), as a large restoring torque is produced otherwise. + +# A.4 Inverse dynamics + +We compute $\tau_{q}$ using Lagrange's equations derived from d'Alembert's Principle of virtual work + +$$ +\boldsymbol {\tau} _ {q} = \mathfrak {M} \ddot {\boldsymbol {q}} + \mathfrak {C} - \mathfrak {F} \tag {34} +$$ + +We can write the terms on the right hand side as: + +$$ +\mathfrak {M} = \sum_ {k} J _ {k} ^ {\intercal} \mathcal {M} _ {k} J _ {k}, \tag {35} +$$ + +$$ +\mathfrak {C} = \sum_ {k} \left(J _ {k} ^ {\intercal} \mathcal {M} _ {k} \dot {J} _ {k} + J _ {k} ^ {\intercal} \left[ \begin{array}{c c} 0 & 0 \\ 0 & [ J _ {\Omega , k} \dot {q} ] _ {s} \end{array} \right] \mathcal {M} _ {k} J _ {k}\right) \dot {\boldsymbol {q}} \tag {36} +$$ + +$$ +\mathfrak {F} = J _ {L F o o t} ^ {\intercal} \mathcal {F} _ {L F o o t} + J _ {R F o o t} ^ {\intercal} \mathcal {F} _ {R F o o t} \tag {37} +$$ + +where $\mathbf{I}_3$ is the identity matrix, $([\cdot ]_s)$ denotes the skew-symmetric form, and + +$$ +\mathcal {M} _ {k} = \left[ \begin{array}{c c} m _ {k} \mathbf {I} _ {3} & 0 \\ 0 & R _ {k} ^ {0} I _ {0, k} \left(R _ {k} ^ {0}\right) \tau \end{array} \right] \tag {38} +$$ + +To deal with potential energy, we offset the root acceleration in the direction of gravity by $-9.8\mathrm{m / s}^2$ . Jacobian matrix $J$ is the mapping from the generalized space to the global Cartesian coordinates, such that for linear and angular velocities $\mathbf{V}_k,\Omega_k$ in global Cartesian coordinates, we have: + +$$ +J _ {k} \dot {\boldsymbol {q}} = \left[ \begin{array}{l} J _ {V, k} \\ J _ {\Omega , k} \end{array} \right] \dot {\boldsymbol {q}} = \left[ \begin{array}{l} V _ {k} \\ \boldsymbol {\Omega} _ {k} \end{array} \right] \tag {39} +$$ + +$J$ can be computed analytically using a recursive algorithm such as in [17]. For segment $k$ , we define its parent segment $p(k)$ as its neighboring segment that is closer to the root. Other than the root, each segment has one and only one parent. We define $k$ 's children $ch(k)$ as its neighboring segments further away from the root. Let $\boldsymbol{r}_{a \to b}$ denote the 3D displacement from point $a$ to $b$ . For segment $k$ , with linear velocity $\boldsymbol{V}_k$ at its CoM and linear velocity $\boldsymbol{V}_k^{joint}$ at its corresponding joint, we have + +$$ +\boldsymbol {V} _ {k} ^ {j o i n t} = \boldsymbol {V} _ {p (k)} + \boldsymbol {\Omega} _ {p (k)} \times \boldsymbol {r} _ {p (k) \rightarrow k ^ {j o i n t}} \Rightarrow J _ {V, k} ^ {j o i n t} = J _ {V, p (k)} - \left[ \boldsymbol {r} _ {p (k) \rightarrow k ^ {j o i n t}} \right] J _ {\Omega , p (k)} \tag {40} +$$ + +and the velocity at the CoM of segment $k$ becomes: + +$$ +\boldsymbol {V} _ {k} = \boldsymbol {V} _ {k} ^ {\text {j o i n t}} + \boldsymbol {\Omega} _ {k} \times \boldsymbol {r} _ {k ^ {\text {j o i n t}} \rightarrow k} \Rightarrow J _ {V, k} = J _ {V, p (k)} - \left[ \boldsymbol {r} _ {p (k) \rightarrow k ^ {\text {j o i n t}}} \right] J _ {\Omega , p (k)} - \left[ \boldsymbol {r} _ {k ^ {\text {j o i n t}} \rightarrow k} \right] J _ {\Omega , k} \tag {41} +$$ + +From (41), we can compute the time derivative recursively as: + +$$ +\dot {J} _ {V, k} = \dot {J} _ {V, p (k)} - \left[ \boldsymbol {r} _ {p (k) \rightarrow k ^ {\text {j o i n t}}} \right] \dot {J} _ {\Omega , p (k)} - \left[ \boldsymbol {r} _ {k ^ {\text {j o i n t}} \rightarrow k} \right] \dot {J} _ {\Omega , k} \tag {42} +$$ + +The global angular velocity of $k$ in skew symmetric form is: + +$$ +\begin{array}{l} [ \boldsymbol {\Omega} _ {k} ] = \dot {R} _ {k} ^ {0} \left(R _ {k} ^ {0}\right) ^ {\intercal} = \left(R _ {p (k)} ^ {0} \cdot R _ {k}\right) \left(R _ {p (k)} ^ {0} R _ {k}\right) ^ {\intercal} (43) \\ = \dots = \dot {R} _ {p (k)} ^ {0} \left(R _ {p (k)} ^ {0}\right) ^ {\intercal} + R _ {p (k)} ^ {0} \left(\dot {R} _ {k} R _ {k} ^ {\intercal}\right) \left(R _ {p (k)} ^ {0}\right) ^ {\intercal} (44) \\ = \left[ \boldsymbol {\Omega} _ {p (k)} \right] + R _ {p (k)} ^ {0} \left[ \boldsymbol {\omega} _ {k} \right] \left(R _ {p (k)} ^ {0}\right) ^ {\intercal} \quad (\because [ \boldsymbol {\omega} _ {k} ] = \dot {R} _ {k} R _ {k} ^ {\intercal}) (45) \\ \Rightarrow \boldsymbol {\Omega} _ {k} = \boldsymbol {\Omega} _ {p (k)} + R _ {p (k)} ^ {0} \boldsymbol {\omega} _ {k} \quad (\because [ A \boldsymbol {b} ] = A [ \boldsymbol {b} ] A ^ {\intercal}) (46) \\ \end{array} +$$ + +To avoid confusion of notation, we also write joint $k$ 's rotation $\pmb{\theta}_k \triangleq \pmb{q}_k$ , i.e. we have generalized coordinates $\pmb{q} = \begin{bmatrix} \pmb{X}_0 \\ \pmb{\theta} \end{bmatrix} \in \mathbb{R}^{N_{DoF} \times 1}$ where $\pmb{X}_0$ is the global root translation, $\pmb{\theta}_0$ is the global root rotation, $\pmb{\theta}_k$ describes the local rotation of segment $k$ relative to its parent $p(k)$ , and $\pmb{\theta}^{\intercal} = [\pmb{\theta}_0^{\intercal} \pmb{\theta}_1^{\intercal} \dots \pmb{\theta}_{N_k}^{\intercal}]$ . Let $J_{\omega,k}$ be the local Jacobian such that $\omega_k = J_{\omega,k} \dot{\pmb{\theta}}_k$ . We can compute $\Omega_k$ recursively: + +$$ +\begin{array}{l} \boldsymbol {\Omega} _ {k} = \boldsymbol {\Omega} _ {p (k)} + R _ {p (k)} ^ {0} J _ {\omega , k} \dot {\boldsymbol {\theta}} _ {k} (47) \\ = 0 + J _ {\omega , 0} \dot {\boldsymbol {\theta}} _ {0} + \dots + R _ {p (p (k))} ^ {0} J _ {\omega , p (k)} \dot {\boldsymbol {\theta}} _ {p (k)} + R _ {p (k)} ^ {0} J _ {\omega , k} \dot {\boldsymbol {\theta}} _ {k} (48) \\ \triangleq J _ {\Omega , k} \dot {\boldsymbol {q}} (49) \\ \end{array} +$$ + +Let $\mathcal{P}_k$ denote the set of all ancestors of $k$ and itself $(k\in \mathcal{P}_k)$ , we split $J_{\Omega ,k}$ into $N_{k} + 1$ blocks of size $3\times 3$ : + +$$ +J _ {\Omega , k} = \left[ \begin{array}{l l l l l} 0 _ {3 \times 3} & J _ {\omega , 0} & \mathbb {1} _ {1 \in \mathcal {P} _ {k}} R _ {p (1)} ^ {0} J _ {\omega , 1} & \dots & \mathbb {1} _ {N _ {k} \in \mathcal {P} _ {k}} R _ {p (K)} ^ {0} J _ {\omega , N _ {k}} \end{array} \right] \in \mathbb {R} ^ {3 \times N _ {D o F}} \tag {50} +$$ + +For segment $k$ , if we represent rotation $\theta_{k} = [\theta_{k,1}\quad \theta_{k,2}\quad \theta_{k,3}]\stackrel {\Delta}{=}[\alpha \quad \beta \quad \gamma ]\in \mathbb{R}^{3}$ using 3 Euler angles, removing subscript $k$ for notation simplicity, we have + +$$ +[ \boldsymbol {\omega} ] = \dot {R} R ^ {\intercal} = \sum_ {i} \frac {\partial R}{\partial \theta_ {i}} R ^ {\intercal} \dot {\theta} _ {i} = \frac {\partial R}{\partial \alpha} R ^ {\intercal} \dot {\alpha} + \frac {\partial R}{\partial \beta} R ^ {\intercal} \dot {\beta} + \frac {\partial R}{\partial \gamma} R ^ {\intercal} \dot {\gamma} \tag {51} +$$ + +and it remains to compute $\mathbf{J}$ 's s.t. + +$$ +J _ {\omega} \triangleq \left[ \begin{array}{l l l} \mathbf {J} _ {1} & \mathbf {J} _ {2} & \mathbf {J} _ {3} \end{array} \right], \quad \text {w h e r e} \left[ \mathbf {J} _ {1} \right] = \frac {\partial R}{\partial \alpha} R ^ {\intercal}, \quad \left[ \mathbf {J} _ {2} \right] = \frac {\partial R}{\partial \beta} R ^ {\intercal}, \quad \left[ \mathbf {J} _ {3} \right] = \frac {\partial R}{\partial \gamma} R ^ {\intercal} \tag {52} +$$ + +$$ +\text {w h i c h s a t i s f i e s} [ \boldsymbol {\omega} ] = \sum_ {i} [ \mathbf {J} _ {i} ] \dot {\theta} _ {i} \quad \text {a n d} \boldsymbol {\omega} = J _ {\omega} \dot {\boldsymbol {\theta}} \tag {53} +$$ + +Finally, let segment $l \in \mathcal{P}_k$ be an ancestor of $k$ and denote $\{J_{\Omega,k}\}_{l} \triangleq R_{p(l)}^{0} J_{\omega,l}$ as the $(l+2)$ -th $3 \times 3$ block in $J_{\Omega,k}$ from (50), we can compute its time derivative as: + +$$ +\left\{\dot {J} _ {\Omega , k} \right\} _ {l} = \dot {R} _ {p (l)} ^ {0} J _ {\omega , l} + R _ {p (l)} ^ {0} \dot {J} _ {\omega , l}, \quad \text {w i t h} \dot {J} _ {\omega , l} = \sum_ {l, i} \frac {\partial J _ {\omega , l}}{\partial \theta_ {l , i}} \dot {\theta} _ {l, i} \tag {54} +$$ + +following (52), it remains to compute $\mathbf{J}$ 's s.t. + +$$ +\dot {J} _ {\omega , l} \triangleq \left[ \mathbf {j} _ {l, 1} \quad \mathbf {j} _ {l, 2} \quad \mathbf {j} _ {l, 3} \right], \quad \text {w h e r e} \mathbf {j} _ {l, j} = \sum_ {l, i} \frac {\partial \mathbf {J} _ {l , j}}{\partial \theta_ {l , i}} \dot {\theta} _ {l, i} \tag {55} +$$ + +# A.5 Neural network + +We trained a transformer encoder consisting of 8 layers with a latent dimension of 256, using total loss $\mathcal{L}_{total}$ with weights $\lambda_{kin} = [0.5,10,1000,1,20,1000]$ and $\lambda_{dyn} = [100,100,20]$ . We trained on AMASS with a sequence input length of 16 frames, after removing sequences containing non-feet-ground contact, with contact labels from [83], for 25 epochs. We used the AdamW optimizer [44] with a weight decay of $10^{-4}$ and an initial learning rate of $10^{-4}$ that decreases by $20\%$ every 5 epochs. The entire process can be trained in about 12 hours on a single Titan Xp GPU. \ No newline at end of file diff --git a/NeurIPS/2025/3D Human Pose Estimation with Muscles/images.zip b/NeurIPS/2025/3D Human Pose Estimation with Muscles/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..48894177ac426b0cb8783df9f31b310571c2d140 --- /dev/null +++ b/NeurIPS/2025/3D Human Pose Estimation with Muscles/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:44682e6ebd3b5da8b6421ebf8b3954d56477e44a3cb2b14e1d45fd85e8b504be +size 652262 diff --git a/NeurIPS/2025/3D Human Pose Estimation with Muscles/layout.json b/NeurIPS/2025/3D Human Pose Estimation with Muscles/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..5db55e92009b328a0c8b4f8b611f47a1d62cbc35 --- /dev/null +++ b/NeurIPS/2025/3D Human Pose Estimation with Muscles/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7507b177a9c6e3083084d9eb5ddece1892ef7f5273a941097193ea5ed0737362 +size 954753 diff --git a/NeurIPS/2025/3D Interaction Geometric Pre-training for Molecular Relational Learning/e803c955-9e56-414d-b048-fe9bf7d42510_content_list.json b/NeurIPS/2025/3D Interaction Geometric Pre-training for Molecular Relational Learning/e803c955-9e56-414d-b048-fe9bf7d42510_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..fbbf7d04054e9b698af6e9dd8ff7aa9bc5600b99 --- /dev/null +++ b/NeurIPS/2025/3D Interaction Geometric Pre-training for Molecular Relational Learning/e803c955-9e56-414d-b048-fe9bf7d42510_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b1b58d19bafce7b588c7c8e5e634e6a70ec8251f196e09bb3cf47380c7b8161e +size 188632 diff --git a/NeurIPS/2025/3D Interaction Geometric Pre-training for Molecular Relational Learning/e803c955-9e56-414d-b048-fe9bf7d42510_model.json b/NeurIPS/2025/3D Interaction Geometric Pre-training for Molecular Relational Learning/e803c955-9e56-414d-b048-fe9bf7d42510_model.json new file mode 100644 index 0000000000000000000000000000000000000000..08c1ce6d240ac5bbcaba4e4e665164a284ffb93c --- /dev/null +++ b/NeurIPS/2025/3D Interaction Geometric Pre-training for Molecular Relational Learning/e803c955-9e56-414d-b048-fe9bf7d42510_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a0a778f40b78f4d540f848d3d0f7f2705a70bd3e0afdca12c5b80c6caa6c0103 +size 240458 diff --git a/NeurIPS/2025/3D Interaction Geometric Pre-training for Molecular Relational Learning/e803c955-9e56-414d-b048-fe9bf7d42510_origin.pdf b/NeurIPS/2025/3D Interaction Geometric Pre-training for Molecular Relational Learning/e803c955-9e56-414d-b048-fe9bf7d42510_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..6519f8d2afcd7fe63989f2993365cde7a052611c --- /dev/null +++ b/NeurIPS/2025/3D Interaction Geometric Pre-training for Molecular Relational Learning/e803c955-9e56-414d-b048-fe9bf7d42510_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9a519f1bd30b090b2513743a6243fa930abb9fb4bc1bd1d6a1e2ea6c623d97f6 +size 1592245 diff --git a/NeurIPS/2025/3D Interaction Geometric Pre-training for Molecular Relational Learning/full.md b/NeurIPS/2025/3D Interaction Geometric Pre-training for Molecular Relational Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..fcdc04b2e04a1381cf841e3b5b9fa58e71e58e32 --- /dev/null +++ b/NeurIPS/2025/3D Interaction Geometric Pre-training for Molecular Relational Learning/full.md @@ -0,0 +1,799 @@ +# 3D Interaction Geometric Pre-training for Molecular Relational Learning + +mkyeong Lee $^{1}$ , Yunhak Oh $^{1}$ , Heewoong Noh $^{1}$ , Gyoung S. Na $^{1,2}$ , Inkai Xu $^{3}$ , Hanchen Wang $^{3,4}$ , Tianfan Fu $^{5}$ , Chanyoung Park $^{1*}$ + +$^{1}$ KAIST $^{2}$ KRICT $^{3}$ Stanford University $^{4}$ Genentech $^{5}$ Nanjing University + +# Abstract + +Molecular Relational Learning (MRL) is a rapidly growing field that focuses on understanding the interaction dynamics between molecules, which is crucial for applications ranging from catalyst engineering to drug discovery. Despite recent progress, earlier MRL approaches are limited to using only the 2D topological structure of molecules, as obtaining the 3D interaction geometry remains prohibitively expensive. This paper introduces a novel 3D geometric pre-training strategy for MRL (3DMRL) that incorporates a 3D virtual interaction environment, overcoming the limitations of costly traditional quantum mechanical calculation methods. With the constructed 3D virtual interaction environment, 3DMRL trains 2D MRL model to learn the global and local 3D geometric information of molecular interaction. Extensive experiments on various tasks using real-world datasets, including out-of-distribution and extrapolation scenarios, demonstrate the effectiveness of 3DMRL, showing up to a $24.93\%$ improvement in performance across 40 tasks. Our code is publicly available at https://github.com/Namkyeong/3DMRL. + +# 1 Introduction + +Molecular relational learning (MRL) focuses on understanding the interaction dynamics between molecules and has gained significant attention from researchers thanks to its diverse applications [21, 32]. Despite recent advancements in MRL, previous works tend to ignore molecules' 3D geometric information and instead focus solely on their 2D topological structures. However, in molecular science, the 3D geometric information of molecules (Figure 1 (a)) is crucial for understanding and predicting molecular behavior across various contexts, ranging from + +physical properties [1] to biological functions [10]. This is particularly important in MRL, as geometric information plays a key role in molecular interactions by determining how molecules recognize, interact, and bind with one another in their interaction environment [35]. In traditional molecular dynamics simulations, explicit solvent models, which directly consider the detailed environment of molecular interaction, have demonstrated superior performance compared to implicit solvent models, which simplify the solvent as a continuous medium, highlighting the significance of modeling the complex geometries of interaction environments [47]. + +However, acquiring stereochemical structures of molecules is often very costly, resulting in limited availability of such 3D geometric information for downstream tasks [24]. Consequently, in the + +![](images/6cfaba64278bed1156d0eb6f9452edd373e125bb2497b54a903ea69d054d83db.jpg) +(a) Single Molecule + +![](images/4804d1a879a51f771dbd0f60c1551780c5f01cae6c5e80e0d490cf31b0d713a3.jpg) +(b) Molecular Interaction Environment +Figure 1: 3D geometry of (a) an individual molecule and (b) the molecular interaction environment. + +domain of molecular property prediction (MPP), there has been substantial progress in injecting 3D geometric information to 2D molecular graph encoders during the pre-training phase, while utilizing only the 2D molecular graph encoder for downstream tasks [36, 25]. In contrast, compared to the MPP, pre-training strategies for MRL have been surprisingly underexplored, primarily due to the following two distinct challenges in modeling complex molecular interaction environments. + +Firstly, interactions between molecules occur through complex geometry as they are chaotically distributed in space as shown in Figure 1 (b). Therefore, it is essential to consider not only each molecule's independent geometry but also their relative positions and orientations in space. This requirement further complicates the acquisition of geometric information, making it more challenging to obtain detailed 3D geometry of molecular interaction environments. Consequently, it is essential to model an interaction environment that can simulate molecular interactions based solely on the 3D geometry of the individual molecules. + +Secondly, even after constructing the interaction environment, how to inject the geometry between molecules during interactions are not trivial. More specifically, while the global geometry of the interaction environment is essential for understanding overall interactions and system stability, the local geometry is also critical for examining localized interactions and precise molecular behaviors. Therefore, developing pre-training strategies that effectively capture the complementary global and local geometries between molecules and their interaction environment is essential. + +To address these challenges, we introduce a novel 3D geometric pre-training strategy that is applicable to various MRL models by incorporating the 3D geometry of the interaction environment for molecules (3DMRL). Specifically, instead of relying on costly traditional quantum mechanical calculation methods to obtain interaction environments, we first propose a virtual interaction environment involving multiple molecules designed to simulate real molecular interactions. Then, during the pre-training stage, a 2D MRL model is trained to produce representations that are globally aligned with those of the 3D virtual interaction environment via contrastive learning. Additionally, the 2D MRL model is trained to predict the localized relative geometry between molecules within this virtual interaction environment, allowing the model to effectively learn fine-grained atom-level interactions between molecules. These two pre-training strategies enable the 2D MRL model to be pre-trained to understand the nature of molecular interactions, facilitating positive transfer to a wide range of downstream MRL tasks. In this paper, we make the following contributions: + +- Rather than relying on costly traditional quantum mechanical calculation methods to obtain interaction geometry, we propose a virtual interaction geometry made up of multiple molecules to mimic the molecular interaction environment observed in real-world conditions. +- We propose pre-training strategies that allow the 2D MRL model to learn representations of the 3D interaction environment, capturing both its global and local geometries. +- We conduct extensive experiments across various MRL models pre-trained with 3DMRL on a range of MRL tasks, including out-of-distribution and extrapolation scenarios. These experiments demonstrate improvements of up to $24.93\%$ compared to MRL methods trained from scratch, underscoring the versatility of 3DMRL (Section 5). + +To the best of our knowledge, this is the first paper proposing pre-training strategies specifically designed for molecular relational learning. + +# 2 Related Works + +Molecular Relational Learning. Molecular Relational Learning (MRL) focuses on understanding the interaction dynamics between paired molecules. Delfos [23] employs recurrent neural networks combined with attention mechanisms to predict solvation-free energy, a key factor influencing the solubility of chemical substances, using SMILES string as input. Similarly, CIGIN [32] utilizes message-passing neural networks [11] along with a cross-attention mechanism to capture atomic representations for solvation-free energy prediction. In a different context, Joung et al. [17] use graph convolutional networks [18] to generate representations of chromophores and solvents, which are then used to predict various optical and photophysical properties of chromophores, essential for developing new materials with vibrant colors. Meanwhile, MHCADDI [4] introduces a co-attentive message passing network [38] designed for predicting drug-drug interactions (DDI), which aggregates information from all atoms within a pair of molecules, not just within individual molecules. Recently, + +CGIB [21] and CMRL [22] have introduced a comprehensive framework for MRL tasks, such as predicting solvation-free energy, chromophore-solute interactions, and drug-drug interactions. These models achieve this by identifying core functional groups involved in molecular interactions using information bottleneck and causal theory, respectively. However, prior studies have largely ignored molecules' 3D geometric information despite its well-established importance in comprehending various molecular properties. + +3D Pre-training for Molecular Property Prediction. Recently, the molecular science community has shown increasing interest in pre-training machine learning models with unlabeled data, primarily due to the scarcity of labeled data for downstream tasks [22, 37, 44]. A promising approach in this area leverages molecules' inherent nature, which can be effectively represented as both 2D topological graphs and 3D geometric graphs. For instance, 3D Infomax [36] aims to enhance mutual information between 2D and 3D molecular representations using contrastive learning. GraphMVP [24] extends this concept by introducing a generative pre-training framework alongside contrastive learning. More recently, Noisy Nodes [46] and MoleculeSDE [25] have introduced methods to learn the 3D geometric distribution of molecules using a denoising framework, thereby uncovering the connection between the score function and the force field of molecules. Although the 3D structure of molecules has been effectively leveraged in pre-training for predicting single molecular properties, it remains surprisingly underexplored in the context of molecular relational learning (MRL). We provide more detailed explanations with the figure in Appendix A. + +# 3 Preliminaries + +# 3.1 Problem Statement + +Notations. Given a molecule $g$ , we first consider a 2D molecular graph, denoted as $g_{2\mathrm{D}} = (\mathbf{X}, \mathbf{A})$ where $\mathbf{X} \in \mathbb{R}^{N \times F}$ represents the atom attribute matrix, and $\mathbf{A} \in \mathbb{R}^{N \times N}$ is the adjacency matrix, with $\mathbf{A}_{ij} = 1$ if a covalent bond exists between atoms $i$ and $j$ . Additionally, we define a 3D conformer as $g_{3\mathrm{D}} = (\mathbf{X}, \mathbf{R})$ , where $\mathbf{R} \in \mathbb{R}^{N \times 3}$ is the matrix of 3D coordinates, each row representing the spatial position of an individual atom. + +Task Description. Given a 2D molecular graph pair $(g_{\mathrm{2D}}^{1}, g_{\mathrm{2D}}^{2})$ and 3D conformer pair $(g_{\mathrm{3D}}^{1}, g_{\mathrm{3D}}^{2})$ , our goal is to pre-train the 2D molecular encoders $f_{\mathrm{2D}}^{1}$ and $f_{\mathrm{2D}}^{2}$ simultaneously with the virtual interaction geometry $g_{\mathrm{vr}}$ , derived from the 3D conformer pair. Then, the pre-trained 2D molecular encoders $f_{\mathrm{2D}}^{1}$ and $f_{\mathrm{2D}}^{2}$ are utilized for various MRL downstream tasks. + +# 3.2 2D MRL Model Architecture + +In this paper, we mainly focus on 1) the construction of virtual interaction geometry, and 2) pretraining strategies for MRL. Therefore, we employ existing model architectures for 2D MRL, i.e., CIGIN [32], which provides a straightforward yet effective framework for MRL as depicted in Figure 2 (a). Specifically, for each pair of 2D molecular graphs, denoted as $g_{\mathrm{2D}}^{1}$ and $g_{\mathrm{2D}}^{2}$ , the graph neural networks (GNNs)-based molecular encoders $f_{\mathrm{2D}}^{1}$ and $f_{\mathrm{2D}}^{2}$ initially produce an atom embedding matrix for each molecule, formulated as: + +$$ +\mathbf {E} ^ {1} = f _ {2 \mathrm {D}} ^ {1} \left(g _ {2 \mathrm {D}} ^ {1}\right), \quad \mathbf {E} ^ {2} = f _ {2 \mathrm {D}} ^ {2} \left(g _ {2 \mathrm {D}} ^ {2}\right), \tag {1} +$$ + +where $\mathbf{E}^1\in \mathbb{R}^{N^1\times d}$ and $\mathbf{E}^2\in \mathbb{R}^{N^2\times d}$ are the atom embedding matrices for $g_{2\mathrm{D}}^{1}$ and $g_{2\mathrm{D}}^{2}$ , containing $N^1$ and $N^2$ atoms, respectively. Next, we capture the interactions between nodes in $g_{2\mathrm{D}}^{1}$ and $g_{2\mathrm{D}}^{2}$ using an interaction matrix $\mathbf{I}\in \mathbb{R}^{N^1\times N^2}$ , defined by $\mathbf{I}_{ij} = \mathrm{sim}(\mathbf{E}_i^1,\mathbf{E}_j^2)$ , where $\mathrm{sim}(\cdot ,\cdot)$ represents the cosine similarity measure. Subsequently, we derive new embedding matrices $\tilde{\mathbf{E}}^1\in \mathbb{R}^{N^1\times d}$ and $\tilde{\mathbf{E}}^2\in \mathbb{R}^{N^2\times d}$ for each graph, reflecting their respective interactions. This is computed using $\tilde{\mathbf{E}}^1 = \mathbf{I}\cdot \mathbf{E}^2$ and $\tilde{\mathbf{E}}^2 = \mathbf{I}^\top \cdot \mathbf{E}^1$ , where $\cdot$ denotes matrix multiplication. Here, $\tilde{\mathbf{E}}^1$ represents the node embeddings of $g_{2\mathrm{D}}^{1}$ that incorporates the interaction information with nodes in $g_{2\mathrm{D}}^{2}$ , and similarly for $\tilde{\mathbf{E}}^2$ . To obtain the final node embeddings, we concatenate the original and interaction-based embeddings for each graph, resulting in $\mathbf{H}^1 = (\mathbf{E}^1 ||\tilde{\mathbf{E}}^1)\in \mathbb{R}^{N^1\times 2d}$ and $\mathbf{H}^2 = (\mathbf{E}^2 ||\tilde{\mathbf{E}}^2)\in \mathbb{R}^{N^2\times 2d}$ . Finally, we apply the Set2Set function [40] to compute the graph-level embeddings $\mathbf{z}_{2\mathrm{D}}^{1}$ and $\mathbf{z}_{2\mathrm{D}}^{2}$ for graph $g_{2\mathrm{D}}^{1}$ and $g_{2\mathrm{D}}^{2}$ , respectively. + +![](images/602729008a4d7000d9fb2323f8b04025e6d162ff8a06dbf83559447b1128be63.jpg) +Figure 2: Overall Framework: (a) 2D MRL model architecture (Section 3.2). (b) Virtual interaction geometry construction (Section 4.1). (c) $SE(3)$ -Invariant Global Geometry Learning (Section 4.2.1). (d) $SE(3)$ -Equivariant Local Relative Geometry Learning (Section 4.2.2). + +![](images/2676705921c6620e5b564a0407dc9925f7a8b73bf7f27523a1f178fd16e794f0.jpg) + +![](images/96a05e440e9aaaceaa35ac11f6ebd01de12bea06f9ded4daa6e222cbc28bfe46.jpg) + +# 4 Methodology + +In this section, we introduce our method, named 3DMRL, a novel pre-training framework for MRL utilizing 3D geometry information. Specifically, in Section 4.1, we introduce how to construct the virtual interaction geometry that can be utilized instead of expensive calculation of real interaction geometry of molecules. Then, in Section 4.2, we present two complementary geometric pre-training strategies for the 2D MRL model to acquire representations aligned with the constructed virtual interaction geometry in both global and local perspectives. The overall framework is depicted in Figure 2, and the pseudocode for the entire framework is provided in Appendix F. + +# 4.1 Virtual Interaction Geometry Construction + +While the 3D geometry of molecules plays a significant role in predicting molecular properties, acquiring this information involves a trade-off between cost and accuracy. For example, RDKit's ETKDG algorithm [20] is fast but less accurate. In contrast, the widely adopted metadynamics method, CREST [12], achieves a more balanced compromise between speed and accuracy, yet still requires around 6 hours to process a drug-like molecule. This challenge is even more pronounced in molecular interaction systems, which necessitates not just the geometry of individual molecules but also the relative spatial arrangements between multiple molecules [6]. Moreover, an appropriate initial geometry of a molecular interaction system is highly dependent on individual molecules in the system [27]. For this reason, a data-agnostic process for generating the initial molecular geometry is crucial for flexible and robust representation learning on molecular interaction systems. + +Drawing inspiration from the explicit solvent models used in traditional molecular dynamics simulations [9], we propose a one-to-many geometric configuration that involves a relatively larger molecule $g_{\mathrm{3D}}^{1}$ , determined based on its radius, surrounded by multiple smaller molecules $g_{\mathrm{3D}}^{2}$ as shown in Figure 2 (b) [28]. Specifically, for a given conformer pair $(g_{\mathrm{3D}}^{1} = (\mathbf{X}^{1}, \mathbf{R}^{1}), g_{\mathrm{3D}}^{2} = (\mathbf{X}^{2}, \mathbf{R}^{2}))$ , we create an environment by arranging the smaller molecules $(g_{\mathrm{3D}}^{2,1}, \ldots, g_{\mathrm{3D}}^{2,i}, \ldots, g_{\mathrm{3D}}^{2,n})$ around a centrally placed larger molecule $g_{\mathrm{3D}}^{1}$ as follows: + +[Step 1] Select Target Atoms in the Larger Molecule. We start by randomly selecting $n$ atoms from the larger molecule $g_{\mathrm{3D}}^{1}$ that are not part of any aromatic ring. This choice is based on the fact that aromatic rings are more stable and less likely to engage in chemical reactions. + +[Step 2] Positioning the Smaller Molecules. Each smaller molecule in $(g_{3\mathrm{D}}^{2,1},\ldots ,g_{3\mathrm{D}}^{2,i},\ldots ,g_{3\mathrm{D}}^{2,n})$ is then placed close to one of the $n$ selected atoms in the larger molecule $g_{3\mathrm{D}}^{1}$ . This positioning + +is achieved by transiting and rotating the original 3D coordinates $\mathbf{R}^2$ of the smaller molecule $g_{\mathrm{3D}}^2$ , following the method widely employed in computational chemistry [19]. + +- [Step 2-1] Determine Transition Direction. For flexible and robust molecular relational learning, we follow a widely used strategy that samples initial geometries from parameterized stochastic processes [42]. Specifically, we generate a normalized random Gaussian noise vector $\varepsilon$ (with a norm of 1), which will be used to set the direction for the transition. We then scale this direction vector $\varepsilon$ by the radius of the smaller molecule, $r^2$ , to establish the transition distance. +- [Step 2-2] Transit and Rotate to the New Position. The new 3D coordinates for each smaller molecule are determined using the formula $\mathbf{R}^{2,i} = \mathbf{R}^2 +\varepsilon_i*r^2 +\mathbf{R}_i^1$ , where $\mathbf{R}_i^1\in \mathbb{R}^3$ represents the 3D position of the $i$ -th selected atom in the larger molecule $g_{3\mathrm{D}}^{1}$ . This operation is performed through broadcasting, meaning $\mathbf{R}_i^1$ and $\varepsilon_{i}$ are added to each row of $\mathbf{R}^2$ . Additionally, we apply a random rotation matrix to rotate the small molecule after its transition. This transition and rotation operations ensure that each smaller molecule is positioned close to its corresponding selected atom on the larger molecule, simulating a realistic interaction environment. + +[Step 3] Constructing Virtual Interaction Geometry. After positioning each smaller molecule $g_{\mathrm{3D}}^{2,i}$ near the $i$ -th selected atom in the larger molecule $g_{\mathrm{3D}}^{1}$ , we compile all the 3D coordinates to form a unified virtual environment $g_{\mathrm{vr}}$ . This process involves combining the coordinate matrix $\mathbf{R}^{1}$ of the larger molecule $g_{\mathrm{3D}}^{1}$ , with the transited coordinates $(\mathbf{R}^{2,1}, \ldots, \mathbf{R}^{2,i}, \ldots, \mathbf{R}^{2,n})$ of the smaller molecules $(g_{\mathrm{3D}}^{2,1}, \ldots, g_{\mathrm{3D}}^{2,i}, \ldots, g_{\mathrm{3D}}^{2,n})$ , resulting in $\mathbf{R}_{\mathrm{vr}} = (\mathbf{R}^{1}\|\mathbf{R}^{2,1}\| \ldots\|\mathbf{R}^{2,i}\| \ldots\|\mathbf{R}^{2,n}) \in \mathbb{R}^{(N^{1} + n\cdot N^{2}) \times 3}$ . Additionally, it involves concatenating all the atom attribute matrices to form $\mathbf{X}_{\mathrm{vr}} = (\mathbf{X}^{1}\|\mathbf{X}^{2}\| \ldots\|\mathbf{X}^{2}) \in \mathbb{R}^{(N^{1} + n\cdot N^{2}) \times F}$ , thereby defining the virtual interaction geometry as $g_{\mathrm{vr}} = (\mathbf{X}_{\mathrm{vr}}, \mathbf{R}_{\mathrm{vr}})$ . Note that multiple small molecules share the same attribute matrix $\mathbf{X}^{2}$ , since we use the atom attribute irrelevant to the atomic coordinates. + +Note that such randomized configurations of interaction environment is a well-established strategy in molecular simulations. For instance, protein-ligand docking protocols (e.g., Rosetta) often initialize ligands in random orientations relative to the protein before searching for binding modes. Similarly, Monte Carlo insertion methods like Widom's test-particle approach randomly insert solvent molecules to explore solute-solvent configurations without bias. Moreover, while we construct the virtual interaction geometry (Step 1 to Step 3) at each epoch during the pre-training phase, the virtual environment can be generated in real time because transition and rotation are matrix operations. Therefore, we argue that our approach allows efficient sampling over a wide range of distances and orientations while remaining physically sound: in the limit of sufficient sampling, no unphysical configuration is favored, and the process mimics the early stages of solvation when solvent molecules approach from arbitrary directions. In Section 5 and Appendix E.4, we analyze the environment in various aspects, justifying the proposed approach for constructing a virtual interaction environment. + +# 4.2 Pre-training Strategies + +Once the virtual interaction geometry is established, we pre-train the 2D MRL model using two complementary geometry learning strategies: $SE(3)$ -invariant global geometry learning (Section 4.2.1) and $SE(3)$ -equivariant local relative geometry learning (Section 4.2.2). + +# 4.2.1 SE(3)-Invariant Global Geometry Learning + +Given a paired 2D molecular graphs $(g_{\mathrm{2D}}^{1}, g_{\mathrm{2D}}^{2})$ and its corresponding 3D virtual interaction geometry $g_{\mathrm{vr}}$ , we first encode them with a 2D MRL model, and a geometric deep learning model, respectively. For 2D molecular graphs, we compute the molecule-level representations, $\mathbf{z}_{\mathrm{2D}}^{1}$ and $\mathbf{z}_{\mathrm{2D}}^{2}$ , for each molecule $g_{\mathrm{2D}}^{1}$ and $g_{\mathrm{2D}}^{2}$ , respectively, as outlined in the Section 3.2. Following this, we derive the 2D interaction representation $\mathbf{z}_{\mathrm{2D}}$ , by concatenating these two representations, i.e., $\mathbf{z}_{\mathrm{2D}} = (\mathbf{z}_{\mathrm{2D}}^{1}||\mathbf{z}_{\mathrm{2D}}^{2})$ . On the other hand, to encode the 3D virtual interaction geometry $g_{\mathrm{vr}} = (\mathbf{X}_{\mathrm{vr}}, \mathbf{R}_{\mathrm{vr}})$ , we use geometric GNNs $f_{\mathrm{3D}}$ that output $SE(3)$ invariant [7] representations $\mathbf{z}_{\mathrm{3D}}$ given the coordinates of atoms $\mathbf{R}_{\mathrm{vr}}$ in virtual interaction geometry [34], i.e., $\mathbf{z}_{\mathrm{3D}} = f_{\mathrm{3D}}(\mathbf{R}_{\mathrm{vr}})$ . Then, as shown in Figure 2 (c), we align the 2D interaction representation $\mathbf{z}_{\mathrm{2D}}$ and the 3D geometry representation $\mathbf{z}_{\mathrm{3D}}$ via Normalized + +temperature-scaled cross entropy loss [3] as follows: + +$$ +\mathcal {L} _ {\mathrm {g l o b}} = - \frac {1}{N _ {\mathrm {b a t c h}}} \sum_ {i = 1} ^ {N _ {\mathrm {b a t c h}}} \left[ \log \frac {e ^ {\operatorname {s i m} \left(\mathbf {z} _ {2 \mathrm {D} , i} , \mathbf {z} _ {3 \mathrm {D} , i}\right) / \tau}}{\sum_ {k = 1} ^ {N _ {\mathrm {b a t c h}}} e ^ {\operatorname {s i m} \left(\mathbf {z} _ {2 \mathrm {D} , i} , \mathbf {z} _ {3 \mathrm {D} , k}\right) / \tau}} + \log \frac {e ^ {\operatorname {s i m} \left(\mathbf {z} _ {3 \mathrm {D} , i} , \mathbf {z} _ {2 \mathrm {D} , i}\right) / \tau}}{\sum_ {k = 1} ^ {N _ {\mathrm {b a t c h}}} e ^ {\operatorname {s i m} \left(\mathbf {z} _ {3 \mathrm {D} , i} , \mathbf {z} _ {2 \mathrm {D} , k}\right) / \tau}} \right]. +$$ + +where $\operatorname{sim}(\cdot, \cdot)$ represents cosine similarity, $\tau$ denotes the temperature hyperparameter, and $N_{\mathrm{batch}}$ refers to the number of pairs within a batch. By training the 2D MRL model to output interaction representations that align with the 3D interaction geometry, the model effectively learns the overall global geometry of molecular interactions during the pre-training phase. + +# 4.2.2 $SE(3)$ -Equivariant Local Relative Geometry Learning + +Beyond the overall global geometry of interaction, it is essential to learn about the intermolecular local relative geometry between molecules during molecular interactions, as their localized relative geometry governs how molecules interact in various environments. To achieve this, we propose pre-training the 2D MRL model to learn local relative geometry by predicting the 3D geometry of the paired molecule, specifically by training a smaller 2D molecule encoder to predict the geometry of the larger molecule. However, predicting relative geometry from a 2D representation is challenging because the prediction must adhere to the physical properties of the molecule, specifically being equivariant to rotations and transitions in 3D Euclidean space, also known as $SE(3)$ -equivariance [7]. To address this, we propose predicting the relative geometry between molecules by utilizing local frame [5], which allows for flexible conversion between invariant and equivariant features. + +More specifically, given the position $\mathbf{R}^{2,i}$ of the $i$ -th small molecule $g_{3\mathrm{D}}^{2,i}$ in the constructed virtual interaction geometry, we first define an orthogonal local frame $\mathcal{F}_{k,l}$ between atoms $k$ and $l$ within molecule $g_{3\mathrm{D}}^{2,i}$ as follows: + +$$ +\mathcal {F} _ {k, l} = \left(\frac {\mathbf {r} _ {k} - \mathbf {r} _ {l}}{\| \mathbf {r} _ {k} - \mathbf {r} _ {l} \|}, \frac {\mathbf {r} _ {k} \times \mathbf {r} _ {l}}{\| \mathbf {r} _ {k} \times \mathbf {r} _ {l} \|}, \frac {\mathbf {r} _ {k} - \mathbf {r} _ {l}}{\| \mathbf {r} _ {k} - \mathbf {r} _ {l} \|} \times \frac {\mathbf {r} _ {k} \times \mathbf {r} _ {l}}{\| \mathbf {r} _ {k} \times \mathbf {r} _ {l} \|}\right), \tag {2} +$$ + +where $\mathbf{r}_k\in \mathbb{R}^3$ and $\mathbf{r}_l\in \mathbb{R}^3$ indicate the position of atoms $k$ and $l$ in constructed virtual interaction geometry, respectively. For simplicity, please note that we will omit the molecule index $i$ in the notation from here. With the established local frame, we derive the invariant 3D feature for the edge between atoms $k$ and $l$ by projecting their coordinates into the local frame, i.e., $\mathbf{e}_{3\mathrm{D}}^{k,l} = \mathrm{Projection}_{\mathcal{F}_{k,l}}(\mathbf{r}_k,\mathbf{r}_l)\in \mathbb{R}^d$ . Additionally, we obtain the 2D invariant edge feature between atoms $k$ and $l$ by concatenating the respective features from the 2D molecular graph, i.e., $\mathbf{e}_{2\mathrm{D}}^{k,l} = \mathrm{MLP}(\mathbf{H}_k^2 ||\mathbf{H}_l^2)\in \mathbb{R}^d$ . Now that we have both invariant 2D and 3D features, we can derive the final invariant edge feature $\mathbf{e}^{k,l}$ by combining these invariant edge features as follows: + +$$ +\mathbf {e} _ {k, l} = \mathbf {e} _ {\mathrm {2 D}} ^ {k, l} + \mathbf {e} _ {\mathrm {3 D}} ^ {k, l}. \tag {3} +$$ + +We define the edge feature set $\mathcal{E}$ , which includes $\mathbf{e}_{k,l}$ for every possible pair of atoms. + +With the invariant final edge feature set $\mathcal{E}$ , we can further process the small molecule information through GNNs to predict the geometry of the larger molecule. To achieve this, we first obtain the atom features specific to the $i$ -th small molecule by concatenating the $i$ -th atom representation of the larger molecule (to which the $i$ -th small molecule is assigned) with each atom representation of the small molecule, i.e., $\tilde{\mathbf{X}} = (\mathbf{H}^2 ||\mathbf{H}_i^1)\in \mathbb{R}^{N^2\times 4d}$ using broadcasting. This approach allows the model to learn a more precise geometry by incorporating the features of the assigned atom in the larger molecule. Next, with the edge feature set $\mathcal{E}$ and the atom feature $\tilde{\mathbf{X}}$ , we derive the final edge representation $\mathbf{h}_{k,l}$ through multiple GNN layers, represented as $\mathbf{h}_{k,l} = \mathrm{GNN}(\tilde{\mathbf{X}},\mathcal{E})$ . Finally, we determine the relative geometry $\hat{f}_k$ between the atom $k$ of the small molecule and the central larger molecule by combining the final invariant edge representation $\mathbf{h}_{k,l}$ with our $SE(3)$ -equivariant frame $\mathcal{F}_{k,l}$ as follows: + +$$ +\hat {f} _ {k} = \sum_ {l} \mathbf {h} _ {k, l} \odot \mathcal {F} _ {k, l}, \tag {4} +$$ + +where $\odot$ indicates element-wise product. This approach guarantees our predicted relative geometry $\hat{f}_k$ to be $SE(3)$ -equivariant. Then, we calculate the relative geometry prediction loss as follows: + +$$ +\mathcal {L} _ {\text {l o c a l}} = \frac {1}{n \cdot N ^ {2}} \sum_ {i = 1} ^ {n} \sum_ {k = 1} ^ {N ^ {2}} \left| \left| f _ {k} ^ {i} - \hat {f} _ {k} ^ {i} \right| \right| _ {2} ^ {2}, \tag {5} +$$ + +where $f_{k}^{i}$ represents the ground truth relative geometry between the larger molecule and the $k$ -th atom of the $i$ -th small molecule. We define the relative geometry $f_{k}^{i}$ as the direction between the $k$ -th atom of the $i$ -th small molecule and the $i$ -th atom of the larger molecule to which the small molecule is attached, i.e., $f_{k}^{i} = (\mathbf{R}_{k}^{2,i} - \mathbf{R}_{i}^{1}) / ||\mathbf{R}_{k}^{2,i} - \mathbf{R}_{i}^{1}||_{2}$ . Note that $\mathcal{L}_{\mathrm{local}}$ is calculated for every molecule pair in the batch, although we have omitted this notation for simplicity. + +Finally, we pre-train the 2D MRL model by jointly optimizing two proposed losses, i.e., $SE(3)$ -invariant global geometry loss and $SE(3)$ -equivariant local relative geometry loss, as follows: + +$$ +\mathcal {L} _ {\text {p r e - t r a i n}} = \mathcal {L} _ {\text {g l o b}} + \alpha \cdot \mathcal {L} _ {\text {l o c a l}}, \tag {6} +$$ + +where $\alpha$ is a hyperparameter that determines the trade-off between the global geometry loss and the local geometry loss. After task-agnostic pre-training, the 2D molecular encoders $f_{\mathrm{2D}}^{1}$ and $f_{\mathrm{2D}}^{2}$ are fine-tuned for specific downstream tasks where access to 3D geometric information is limited. + +# 4.3 Discussion + +While we define local relative geometry for learning fine-grained interactions between molecules, we can view local relative geometry as an interaction force between molecules. This provides a physically motivated supervision signal rooted in classical intermolecular forces, many of which are central and act along the internuclear axis. For example, van der Waals interactions (described by the Lennard-Jones potential) exhibit repulsive or attractive forces directed along this axis. At short distances, repulsion dominates and aligns directly outward between nuclei. + +This supervision scheme serves as a central-force approximation, consistent with classical force fields, and offers a lightweight surrogate for full force labels, which would require costly quantum chemistry or MD simulations. Notably, SchNet[34] demonstrated that even approximate force signals improve learning of molecular interactions. Our direction-based supervision enables the model to learn geometric features like hydrogen bond alignment or steric repulsion trajectories in an SE(3)-equivariant manner. + +Since solvent atoms are placed near specific solute atoms, the dominant interaction direction aligns with the interatomic vector, making it a reasonable proxy for the net force axis. Thus, the unit direction vector serves as a pseudo-force label, conveying the primary interaction axis and encouraging the model to encode directionality of intermolecular interactions. + +# 5 Experiments + +# 5.1 Experimental Setup + +Downstream Tasks. Following a prior study [21], we employ ten datasets to comprehensively evaluate the performance of 3DMRL on two tasks: 1) molecular interaction prediction, and 2) drug-drug interaction (DDI) prediction. For the molecular interaction prediction task, we utilize the Chromophore dataset [16], which pertains to three optical properties of chromophores, along with five other datasets related to the solvation free energy of solutes: MNSol [26], FreeSolv [29], CompSol [30], Abraham [13], and CombiSolv [39]. In the Chromophore dataset, we focus on the maximum absorption wavelength (Absorption), maximum emission wavelength (Emission), and excited state lifetime (Lifetime) properties. For the DDI prediction task, we use two datasets: ZhangDDI [48] and ChChMiner [49], both of which contain labeled DDI data. We provide further details on pre-training and downstream task datasets in Appendix B.1 and B.2, respectively. + +Baseline methods. We validate the effectiveness of 3DMRL by using it to enhance various recent state-of-the-art molecular relational learning methods, including MPNN [11], AttentiveFP [43], CIGIN [32], CGIB [21], and $\mathbf{CGIB}_{\mathrm{Cont}}$ [21]. Additionally, we compare our proposed pre-training framework, 3DMRL, with recent molecular pre-training approaches that aim to learn 3D structure of individual molecules, such as 3D Infomax [36], GraphMVP [24], and MoleculeSDE [25]. It is important to note that these approaches involve pre-training a single encoder for molecular property prediction (MPP Pre-training in Table 11), whereas our work is pioneering in training two separate encoders simultaneously during pre-training for molecular relational learning (MRL Pre-training in Table 11). For the baseline methods, we use the original authors' code and conduct the experiments in the same environment as 3DMRL to ensure a fair comparison. Moreover, while we choose to mainly + +Table 1: Performance improvement in molecular interaction tasks across different models with our proposed pre-training strategy (RMSE) $(\downarrow)$ . We conduct 15 independent runs for each model and report their mean along with the standard deviation (in parentheses). Colors indicate the performance improvement compared to the models trained from scratch. + +
ModelChromophoreMNSolFreeSolvCompSolAbrahamCombiSolv
AbsorptionEmissionLifetime
MPNN22.00 (0.30)26.34 (0.41)0.789 (0.021)0.643 (0.005)1.127 (0.110)0.420 (0.018)0.640 (0.008)0.614 (0.031)
+ 3DMRL19.96 (0.12)25.21 (0.31)0.753 (0.018)0.609 (0.008)1.068 (0.087)0.377 (0.020)0.550 (0.051)0.599 (0.025)
Improvement9.27%4.29%4.56%5.28%5.24%10.24%14.06%2.44%
AttentiveFP22.86 (0.30)28.70 (0.23)0.871 (0.010)0.570 (0.021)1.019 (0.070)0.350 (0.008)0.426 (0.042)0.471 (0.028)
+ 3DMRL22.80 (0.61)28.54 (1.97)0.784 (0.013)0.562 (0.031)0.901 (0.059)0.271 (0.009)0.378 (0.027)0.448 (0.011)
Improvement0.26%0.55%9.99%1.40%11.57%22.57%11.26%4.88%
CIGIN19.66 (0.69)25.84 (0.23)0.821 (0.017)0.582 (0.022)0.958 (0.116)0.369 (0.018)0.421 (0.018)0.464 (0.002)
+ 3DMRL18.00 (0.17)24.21 (0.09)0.729 (0.014)0.528 (0.019)0.839 (0.105)0.277 (0.006)0.371 (0.031)0.435 (0.006)
Improvement8.44%6.30%11.20%9.28%12.42%24.93%11.87%6.25%
CGIB18.37 (0.35)24.52 (0.25)0.808 (0.015)0.562 (0.008)0.876 (0.037)0.321 (0.002)0.404 (0.037)0.448 (0.008)
+ 3DMRL17.93 (0.35)23.92 (0.29)0.733 (0.009)0.538 (0.020)0.842 (0.078)0.274 (0.002)0.370 (0.027)0.442 (0.015)
Improvement2.40%5.90%9.28%4.27%3.88%14.64%8.42%1.33%
CGIBCont18.59 (0.24)24.68 (0.49)0.803 (0.019)0.561 (0.012)0.897 (0.098)0.333 (0.005)0.404 (0.039)0.452 (0.015)
+ 3DMRL17.90 (0.17)**23.94 (0.24)0.720 (0.020)0.524 (0.018)*0.863 (0.075)0.284 (0.007)0.372 (0.021)0.441 (0.022)
Improvement3.71%3.00%10.33%6.59%3.79%14.71%7.92%2.43%
+ +compare 2D encoder pre-training approach, we also compare 3D encoder pre-training approaches [15, 31, 8] in Appendix E.5. We provide more details on the compared methods in Appendix C. + +Evaluation protocol. Following Pathak et al. [32], for the molecular interaction prediction task, we evaluate the models under a 5-fold cross-validation scheme. The dataset is randomly split into 5 subsets and one of the subsets is used as the test set, while the remaining subsets are used to train the model. A subset of the test set is selected as the validation set for hyperparameter selection and early stopping. We repeat 5-fold cross-validation three times (i.e., 15 runs in total) and report the accuracy and standard deviation of the repeats. For the DDI prediction task [21], we conduct experiments on two different out-of-distribution scenarios, namely molecule split and scaffold split. For the molecule split, the performance is evaluated when the models are presented with new molecules not included in the training dataset. In the scaffold split setting [14], just like in the molecule split, molecules corresponding to scaffolds that were not seen during training will be used for testing. For both splits, we repeat 5 independent experiments with different random seeds on split data, and report the accuracy and the standard deviation of the repeats. In both scenarios, we split the data into training, validation, and test sets with a ratio of $60/20/20\%$ . We provide details on evaluation protocol, model implementation, and model training in Section D. + +# 5.2 Experimental Results + +We begin by comparing each model architecture trained from scratch with the same architecture pre-trained using our proposed strategy, referred to as +3DMRL in Table 1. We have the following observations: 1) 3DMRL obtains consistent improvements over the base graph neural networks in all 40 tasks (across various datasets and neural architectures), achieving up to $24.93\%$ relative reduction in RMSE. While the paper is written based on CIGIN for better understanding in Section 3.2, we could observe performance improvements not only in CIGIN but also in various other model architectures, demonstrating the versatility of proposed pre-training strategies. We further demonstrate how our pre-training strategies are adopted to various model architectures in Appendix C. + +Additionally, we compare our pre-training strategies with recent molecular pre-training approaches proposed for molecular property prediction (MPP) of a single molecule. Table 11 (a) and (b) show the results for the molecular interaction prediction task, and the drug-drug interaction (DDI) task, respectively. As these approaches are originally designed for single molecules, we first pre-train the GNNs using each strategy, then incorporate the pre-trained GNNs into the CIGIN architecture and fine-tune them for various MRL downstream tasks. We have the following observations: 2) Although MPP pre-training methods have demonstrated success in molecular property prediction in prior studies, they did not yield satisfactory results in molecular relational learning tasks and, in some cases, even resulted in negative transfer. This highlights the need for creating specialized pre-training strategies tailored to MRL tasks. We further demonstrate the MPP pre-training strategy with a large-scale dataset still performs worse than 3DMRL in Appendix E.1. 3) On the other hand, pre-training + +Table 2: Performance of CIGIN model on (a) molecular interaction tasks using different pre-training strategies (RMSE) $(\downarrow)$ and (b) out-of-distribution DDI tasks using different pre-training strategies (AUROC) $(\uparrow)$ . For each dataset, we highlight the best method in bold. + +
Strategy(a) Molecular Interaction Tasks (RMSE ↓)(b) Drug-Drug Interaction Task (AUROC ↑)
ChromophoreMNSolFreeSolvCompSolAbrahamCombiSolv(c) Molecule Split(d) Scaffold Split
AbsorptionEmissionLifetimeZhangDDIChChMinerZhangDDIChChMiner
No Pre-training19.66 (0.69)25.84 (0.23)0.821 (0.017)0.567 (0.014)0.884 (0.074)0.331 (0.029)0.412 (0.028)0.458 (0.002)71.75 (0.76)76.21 (1.19)70.96 (1.40)75.81 (0.79)
MPP (molecular property prediction) Pre-training
3D Infomax18.71 (0.61)24.59 (0.22)0.790 (0.022)0.585 (0.015)0.873 (0.103)0.321 (0.041)0.426 (0.036)0.464 (0.004)71.01 (2.19)76.05 (1.30)70.90 (1.63)74.87 (1.08)
GraphMVP18.40 (0.62)24.73 (0.14)0.797 (0.022)0.561 (0.025)1.010 (0.115)0.301 (0.025)0.418 (0.020)0.437 (0.015)71.82 (1.44)76.42 (1.68)71.73 (0.95)76.13 (1.01)
MoleculeSDE18.56 (0.24)24.91 (0.10)0.836 (0.040)0.564 (0.018)0.971 (0.122)0.308 (0.024)0.426 (0.028)0.454 (0.012)70.07 (0.58)76.37 (1.14)69.46 (1.55)76.03 (1.13)
MRL (molecular relational learning) Pre-training
3DMRL18.00 (0.17)24.21 (0.09)0.729 (0.014)0.528 (0.019)0.839 (0.105)0.277 (0.006)0.371 (0.031)0.435 (0.006)74.00 (0.72)78.93 (0.59)74.85 (1.58)78.56 (1.03)
+ +with 3DMRL consistently delivers significant performance improvements across downstream tasks. This validates the effectiveness of our approach, as it successfully integrates scientific knowledge into the pre-training strategy, enhancing the model's overall performance. 4) Additionally, for the DDI task in Table 11 (b), we observed that the performance improvement is more pronounced in challenging scenarios ((d) Scaffold split) compared to less difficult ones ((c) Molecule split). This highlights the enhanced generalization ability of 3DMRL in out-of-distribution scenarios, demonstrating its potential for drug discovery applications where robust generalization across unknown molecules is essential. We explore the extrapolation capability of 3DMRL in Appendix E.2. + +# 5.3 Model Analysis + +Ablation Studies. To further understand our model, we conduct an ablation study to investigate the impact of two key components on the final performance. Specifically, as shown in Equation 6, the objective function contains two terms: (i) global geometry loss and (ii) intermolecular local geometry loss; we curate two variants that involve only (i) (denoted only glob.) and only (ii) (denoted only local) in Figure 3. As shown in Figure 3, learning the global geometry plays a particularly critical role. Removing it from 3DMRL results in a significant performance drop, even falling below MPP pre-training strategies such as 3D Infomax and GraphMVP. This is because the global + +![](images/ad0c2f234bc4e7948a8f36fe26a2723e7f4b75aac32eb9774537ec6464fea9d0.jpg) +Figure 3: Ablation studies. + +geometry loss allows the model to capture the overall interaction geometry at the molecular level, while the local geometry loss focuses on learning more fine-grained, atom-level interactions. However, combining both losses, as in 3DMRL, yields the best results, demonstrating the importance of leveraging the strengths of both levels of granularity. We provide further detailed results of ablation studies in Appendix E.3. + +Molecule Collision Analysis. While the virtual environment is designed to carefully mimic the nature of molecular interactions, as discussed in Section 4.1, molecule collisions can still occur within the environment. To first examine how molecule collisions affect model performance, we created a "No Radius" model that does not take the radius into account during pre-training. Looking at the atomic overlap in Table 3, we observed that 3DMRL, which utilizes radius in + +Table 3: Model performance in various 3D interaction environments with reduced collision. + +
Atomic OverlapTime (min/epoch)Performance
AbsorptionEmissionLifetime
No Radius73.84%5.3018.6825.370.745
Fixed Direction19.28 %5.3018.2624.240.734
Twice Radius10.28 %5.3218.2324.250.730
Regenerate0.0%23.0118.2023.860.727
3DMRL25.12%5.3218.0024.210.729
+ +formation, significantly reduces the overlap ratio between molecules compared to the "No Radius" configuration. Moreover, we found that in the "No Radius" case, where the atomic overlap ratio is very high, the performance was much lower than that of the 3DMRL. To further investigate whether further reducing atomic overlap would be helpful, we experimented with several additional configurations. The "Fixed direction" configuration was designed to prevent overlap caused by random direction placement by positioning the solvent along the direction from the origin to the target atom. "Twice radius" refers to multiplying the radius in the equation in line 170. These methods reduce atomic overlap by decreasing randomness and increasing the distance between molecules, respectively; however, in terms of performance, they were either similar to or worse than 3DMRL. + +The experimental results showed that both methods were able to reduce atomic overlap, but in terms of performance, they were either similar to or worse than 3DMRL. Lastly, we introduce "Regenerate," which regenerates the 3D virtual environment every time a collision occurs between any molecules in a virtual environment. Although the collision between molecules can certainly be avoided in this case, this approach incurs high computational complexity. In Table 3, we observe that the performance gain of "Regenerate" is minimal, despite its significantly higher computational requirements. Based on these results, we argue that 3DMRL strikes an appropriate balance between computation and performance. + +Sensitivity analysis on $n$ . Moreover, we conduct a sensitivity analysis to explore the empirical effect of the number of target atoms $n$ , which determines the number of small molecules in a virtual interaction geometry. In Figure 4 (a), we observe that the model performs the best when using five small molecules to construct the virtual interaction geometry. More specifically, using too few small molecules ( $n = 2$ ) results in poorer performance, as it fails to adequately simulate real-world interaction environments. On the other hand, the model performance also declines as the number of small molecules increases, likely due to the 3D geometry encoder overfitting to the small + +![](images/22b46c6113994d098d84626db966b4ce6dc19b5fbccdaabea0e68ab13ad124ad.jpg) +Figure 4: Sensitivity analysis. + +molecules with an excessive count. Furthermore, we observe that as the number of target atoms increases, more extensive computational resources are required to encode the 3D interaction geometry during pre-training. Hence, selecting an appropriate number of target atoms is crucial for both model performance and computational efficiency. We provide further analyses on virtual interaction geometry and other tasks in Appendix E.4. + +Sensitivity analysis on $\alpha$ . We conduct sensitivity analysis on $\alpha$ , which controls the weight of local geometry loss, in Equation 6. In Figure 4 (b), the model's performance declines as $\alpha$ increases from 0.1, primarily because it overly emphasizes atom-level interactions between the molecules instead of considering the overall interaction geometry. Conversely, we also notice a drop in performance when local geometry loss is not utilized ( $\alpha = 0.0$ ), as this causes the model to lose ability in learning fine-grained atom-level interactions. It is important to note that while we set $n = 5$ and $\alpha = 0.1$ during pre-training, models pre-trained with varying $n$ and $\alpha$ consistently outperform those trained from scratch, demonstrating 3DMRL's robustness. + +# 6 Conclusion + +In this work, we propose 3DMRL, a novel pre-training framework that effectively integrates 3D geometric information into MRL. By constructing a virtual interaction geometry and utilizing local and global geometry prediction, our approach effectively incorporates complex 3D interaction geometry information into 2D MRL models. Experimental results demonstrate that 3DMRL significantly enhances the performance of 2D MRL models across various downstream tasks and neural architectures, validating the importance of incorporating 3D geometric data. + +For future work, we intend to develop and train a virtual interaction geometry generator capable of mimicking MD trajectories of molecular interactions. We will then substitute this generator for the purely random generation method currently used in Section 4.1, providing a more physically informed signal.. Furthermore, we plan to broaden this research in drug-target binding affinity prediction, a core task in drug discovery that involves complex protein structures as the larger molecule. + +# Acknowledgement + +This work was supported by the Institute of Information & Communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (RS-2025-02304967, AI Star Fellowship (KAIST)). Additionally, this research received funding from the National Research Foundation of Korea (NRF) through two separate grants: RS-2024-00335098 (funded by the Korea government (MSIT)) and RS-2022-NR068758 (funded by the Ministry of Science and ICT). + +# References + +[1] Atkins, P. W., De Paula, J., and Keeler, J. Atkins' physical chemistry. Oxford university press, 2023. +[2] Axelrod, S. and Gomez-Bombarelli, R. Geom, energy-annotated molecular conformations for property prediction and molecular generation. Scientific Data, 9(1):185, 2022. +[3] Chen, T., Kornblith, S., Norouzi, M., and Hinton, G. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pp. 1597-1607. PMLR, 2020. +[4] Deac, A., Huang, Y.-H., Velickovic, P., Lio, P., and Tang, J. Drug-drug adverse effect prediction with graph co-attention. arXiv preprint arXiv:1905.00534, 2019. +[5] Du, W., Zhang, H., Du, Y., Meng, Q., Chen, W., Zheng, N., Shao, B., and Liu, T.-Y. Se (3) equivariant graph neural networks with complete local frames. In International Conference on Machine Learning, pp. 5583-5608. PMLR, 2022. +[6] Durrant, J. D. and McCammon, J. A. Molecular dynamics simulations and drug discovery. BMC biology, 9:1-9, 2011. +[7] Duval, A., Mathis, S. V., Joshi, C. K., Schmidt, V., Miret, S., Malliaros, F. D., Cohen, T., Lio, P., Bengio, Y., and Bronstein, M. A hitchhiker's guide to geometric gnu ns for 3d atomic systems. arXiv preprint arXiv:2312.07511, 2023. +[8] Feng, S., Ni, Y., Lan, Y., Ma, Z.-M., and Ma, W.-Y. Fractional denoising for 3d molecular pre-training. In International Conference on Machine Learning, pp. 9938-9961. PMLR, 2023. +[9] Frenkel, D. and Smit, B. Understanding molecular simulation: from algorithms to applications. Elsevier, 2023. +[10] Fu, Y., Lu, Y., Wang, Y., Zhang, B., Zhang, Z., Yu, G., Liu, C., Clarke, R., Herrington, D. M., and Wang, Y. Ddn3. 0: Determining significant rewiring of biological network structure with differential dependency networks. Bioinformatics, pp. btae376, 2024. +[11] Gilmer, J., Schoenholz, S. S., Riley, P. F., Vinyals, O., and Dahl, G. E. Neural message passing for quantum chemistry. In International conference on machine learning, pp. 1263-1272. PMLR, 2017. +[12] Grimme, S. Exploration of chemical compound, conformer, and reaction space with meta-dynamics simulations based on tight-binding quantum chemical calculations. Journal of chemical theory and computation, 15(5):2847-2862, 2019. +[13] Grubbs, L. M., Saifullah, M., Nohelli, E., Ye, S., Achi, S. S., Acree Jr, W. E., and Abraham, M. H. Mathematical correlations for describing solute transfer into functionalized alkane solvents containing hydroxyl, ether, ester or ketone solvents. Fluid phase equilibria, 298(1): 48-53, 2010. +[14] Huang, K., Fu, T., Gao, W., Zhao, Y., Roohani, Y., Leskovec, J., Coley, C. W., Xiao, C., Sun, J., and Zitnik, M. Therapeutics data commons: Machine learning datasets and tasks for drug discovery and development. arXiv preprint arXiv:2102.09548, 2021. +[15] Jiao, R., Han, J., Huang, W., Rong, Y., and Liu, Y. Energy-motivated equivariant pretraining for 3d molecular graphs. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 37, pp. 8096-8104, 2023. +[16] Joung, J. F., Han, M., Jeong, M., and Park, S. Experimental database of optical properties of organic compounds. Scientific data, 7(1):1-6, 2020. +[17] Joung, J. F., Han, M., Hwang, J., Jeong, M., Choi, D. H., and Park, S. Deep learning optical spectroscopy based on experimental database: potential applications to molecular design. JACS Au, 1(4):427-438, 2021. + +[18] Kipf, T. N. and Welling, M. Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907, 2016. +[19] Kuroshima, D., Kilgour, M., Tuckerman, M. E., and Rogal, J. Machine learning classification of local environments in molecular crystals. Journal of chemical theory and computation, 20 (14):6197-6206, 2024. +[20] Landrum, G. Rdkit documentation. Release, 1(1-79):4, 2013. +[21] Lee, N., Hyun, D., Na, G. S., Kim, S., Lee, J., and Park, C. Conditional graph information bottleneck for molecular relational learning. In International Conference on Machine Learning, pp. 18852-18871. PMLR, 2023. +[22] Lee, N., Yoon, K., Na, G. S., Kim, S., and Park, C. Shift-robust molecular relational learning with causal substructure. In Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pp. 1200–1212, 2023. +[23] Lim, H. and Jung, Y. Delfos: deep learning model for prediction of solvation free energies in generic organic solvents. Chemical science, 10(36):8306-8315, 2019. +[24] Liu, S., Wang, H., Liu, W., Lasenby, J., Guo, H., and Tang, J. Pre-training molecular graph representation with 3d geometry. arXiv preprint arXiv:2110.07728, 2021. +[25] Liu, S., Du, W., Ma, Z.-M., Guo, H., and Tang, J. A group symmetric stochastic differential equation model for molecule multi-modal pretraining. In International Conference on Machine Learning, pp. 21497-21526. PMLR, 2023. +[26] Marenich, A. V., Kelly, C. P., Thompson, J. D., Hawkins, G. D., Chambers, C. C., Giesen, D. J., Winget, P., Cramer, C. J., and Truhlar, D. G. Minnesota solvation database (mnsol) version 2012. 2020. +[27] Martínez, L., Andrade, R., Birgin, E. G., and Martínez, J. M. Packmol: A package for building initial configurations for molecular dynamics simulations. Journal of computational chemistry, 30(13):2157-2164, 2009. +[28] Megyes, T., Balint, S., Peter, E., Grósz, T., Bakó, I., Krienke, H., and Bellissent-Funel, M.-C. Solution structure of nano3 in water: Diffraction and molecular dynamics simulation study. The Journal of Physical Chemistry B, 113(13):4054-4064, 2009. +[29] Mobley, D. L. and Guthrie, J. P. Freesolv: a database of experimental and calculated hydration free energies, with input files. Journal of computer-aided molecular design, 28(7):711-720, 2014. +[30] Moine, E., Privat, R., Sirjean, B., and Jaubert, J.-N. Estimation of solvation quantities from experimental thermodynamic data: Development of the comprehensive compsol databank for pure and mixed solutes. Journal of Physical and Chemical Reference Data, 46(3):033102, 2017. +[31] Ni, Y., Feng, S., Ma, W.-Y., Ma, Z.-M., and Lan, Y. Sliced denoising: A physics-informed molecular pre-training method. arXiv preprint arXiv:2311.02124, 2023. +[32] Pathak, Y., Laghuvarapu, S., Mehta, S., and Priyakumar, U. D. Chemically interpretable graph interaction network for prediction of pharmacokinetic properties of drug-like molecules. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pp. 873-880, 2020. +[33] Ryu, J. Y., Kim, H. U., and Lee, S. Y. Deep learning improves prediction of drug–drug and drug–food interactions. Proceedings of the national academy of sciences, 115(18):E4304-E4311, 2018. +[34] Schütt, K., Kindermans, P.-J., Sauceda Felix, H. E., Chmiela, S., Tkatchenko, A., and Müller, K.-R. Schnet: A continuous-filter convolutional neural network for modeling quantum interactions. Advances in neural information processing systems, 30, 2017. +[35] Silverman, R. B. and Holladay, M. W. The organic chemistry of drug design and drug action. Academic press, 2014. + +[36] Stärk, H., Beaini, D., Corso, G., Tossou, P., Dallago, C., Gunnemann, S., and Lio, P. 3d infomax improves gnns for molecular property prediction. In International Conference on Machine Learning, pp. 20479-20502. PMLR, 2022. +[37] Velez-Arce, A., Huang, K., Li, M. M., Lin, X., Gao, W., Fu, T., Kellis, M., Pentelute, B. L., and Zitnik, M. Tdc-2: Multimodal foundation for therapeutic science. bioRxiv, pp. 2024-06, 2024. +[38] Velicković, P., Cucurull, G., Casanova, A., Romero, A., Lio, P., and Bengio, Y. Graph attention networks. arXiv preprint arXiv:1710.10903, 2017. +[39] Vermeire, F. H. and Green, W. H. Transfer learning for solvation free energies: From quantum chemistry to experiments. Chemical Engineering Journal, 418:129307, 2021. +[40] Vinyals, O., Bengio, S., and Kudlur, M. Order matters: Sequence to sequence for sets. arXiv preprint arXiv:1511.06391, 2015. +[41] Wang, Y., Min, Y., Chen, X., and Wu, J. Multi-view graph contrastive representation learning for drug-drug interaction prediction. In Proceedings of the Web Conference 2021, pp. 2921-2933, 2021. +[42] Wu, F. and Li, S. Z. Diffmd: a geometric diffusion model for molecular dynamics simulations. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 37, pp. 5321-5329, 2023. +[43] Xiong, Z., Wang, D., Liu, X., Zhong, F., Wan, X., Li, X., Li, Z., Luo, X., Chen, K., Jiang, H., et al. Pushing the boundaries of molecular representation for drug discovery with the graph attention mechanism. Journal of medicinal chemistry, 63(16):8749-8760, 2019. +[44] Xu, B., Lu, Y., Li, C., Yue, L., Wang, X., Hao, N., Fu, T., and Chen, J. Smiles-mamba: Chemical mamba foundation models for drug admet prediction. arXiv preprint arXiv:2408.05696, 2024. +[45] Xu, K., Hu, W., Leskovec, J., and Jegelka, S. How powerful are graph neural networks? arXiv preprint arXiv:1810.00826, 2018. +[46] Zaidi, S., Schaarschmidt, M., Martens, J., Kim, H., Teh, Y. W., Sanchez-Gonzalez, A., Battaglia, P., Pascanu, R., and Godwin, J. Pre-training via denoising for molecular property prediction. arXiv preprint arXiv:2206.00133, 2022. +[47] Zhang, J., Zhang, H., Wu, T., Wang, Q., and Van Der Spoel, D. Comparison of implicit and explicit solvent models for the calculation of solvation free energy in organic solvents. Journal of chemical theory and computation, 13(3):1034-1043, 2017. +[48] Zhang, W., Chen, Y., Liu, F., Luo, F., Tian, G., and Li, X. Predicting potential drug-drug interactions by integrating chemical, biological, phenotypic and network data. BMC bioinformatics, 18(1):1-12, 2017. +[49] Zitnik, M., Sosic, R., and Leskovec, J. Biosnap datasets: Stanford biomedical network dataset collection. Note: http://snap.stanford.edu/biodataCitedby, 5(1), 2018. + +# NeurIPS Paper Checklist + +The checklist is designed to encourage best practices for responsible machine learning research, addressing issues of reproducibility, transparency, research ethics, and societal impact. Do not remove the checklist: The papers not including the checklist will be desk rejected. The checklist should follow the references and follow the (optional) supplemental material. The checklist does NOT count towards the page limit. + +Please read the checklist guidelines carefully for information on how to answer these questions. For each question in the checklist: + +- You should answer [Yes], [No], or [NA]. +- [NA] means either that the question is Not Applicable for that particular paper or the relevant information is Not Available. +- Please provide a short (1–2 sentence) justification right after your answer (even for NA). + +The checklist answers are an integral part of your paper submission. They are visible to the reviewers, area chairs, senior area chairs, and ethics reviewers. You will be asked to also include it (after eventual revisions) with the final version of your paper, and its final version will be published with the paper. + +The reviewers of your paper will be asked to use the checklist as one of the factors in their evaluation. While "[Yes]" is generally preferable to "[No]", it is perfectly acceptable to answer "[No]" provided a proper justification is given (e.g., "error bars are not reported because it would be too computationally expensive" or "we were unable to find the license for the dataset we used"). In general, answering "[No]" or "[NA]" is not grounds for rejection. While the questions are phrased in a binary way, we acknowledge that the true answer is often more nuanced, so please just use your best judgment and write a justification to elaborate. All supporting evidence can appear either in the main paper or the supplemental material, provided in appendix. If you answer [Yes] to a question, in the justification please point to the section(s) where related material for the question can be found. + +IMPORTANT, please: + +- Delete this instruction block, but keep the section heading "NeurIPS Paper Checklist", +- Keep the checklist subsection headings, questions/answers and guidelines below. +- Do not modify the questions and only use the provided macros for your answers. + +# 1. Claims + +Question: Do the main claims made in the abstract and introduction accurately reflect the paper's contributions and scope? + +Answer: [Yes] + +Justification: Yes, abstract and introduction accurately reflect the paper's contributions and scope. + +Guidelines: + +- The answer NA means that the abstract and introduction do not include the claims made in the paper. +- The abstract and/or introduction should clearly state the claims made, including the contributions made in the paper and important assumptions and limitations. A No or NA answer to this question will not be perceived well by the reviewers. +- The claims made should match theoretical and experimental results, and reflect how much the results can be expected to generalize to other settings. +- It is fine to include aspirational goals as motivation as long as it is clear that these goals are not attained by the paper. + +# 2. Limitations + +Question: Does the paper discuss the limitations of the work performed by the authors? + +Answer: [Yes] + +Justification: In Section 5.3, we have discussed that our approach generates the configuration with collapsed molecules. + +# Guidelines: + +- The answer NA means that the paper has no limitation while the answer No means that the paper has limitations, but those are not discussed in the paper. +- The authors are encouraged to create a separate "Limitations" section in their paper. +- The paper should point out any strong assumptions and how robust the results are to violations of these assumptions (e.g., independence assumptions, noiseless settings, model well-specification, asymptotic approximations only holding locally). The authors should reflect on how these assumptions might be violated in practice and what the implications would be. +- The authors should reflect on the scope of the claims made, e.g., if the approach was only tested on a few datasets or with a few runs. In general, empirical results often depend on implicit assumptions, which should be articulated. +- The authors should reflect on the factors that influence the performance of the approach. For example, a facial recognition algorithm may perform poorly when image resolution is low or images are taken in low lighting. Or a speech-to-text system might not be used reliably to provide closed captions for online lectures because it fails to handle technical jargon. +- The authors should discuss the computational efficiency of the proposed algorithms and how they scale with dataset size. +- If applicable, the authors should discuss possible limitations of their approach to address problems of privacy and fairness. +- While the authors might fear that complete honesty about limitations might be used by reviewers as grounds for rejection, a worse outcome might be that reviewers discover limitations that aren't acknowledged in the paper. The authors should use their best judgment and recognize that individual actions in favor of transparency play an important role in developing norms that preserve the integrity of the community. Reviewers will be specifically instructed to not penalize honesty concerning limitations. + +# 3. Theory assumptions and proofs + +Question: For each theoretical result, does the paper provide the full set of assumptions and a complete (and correct) proof? + +Answer: [NA] + +Justification: The paper does not include theoretical results. + +# Guidelines: + +- The answer NA means that the paper does not include theoretical results. +- All the theorems, formulas, and proofs in the paper should be numbered and cross-referenced. +- All assumptions should be clearly stated or referenced in the statement of any theorems. +- The proofs can either appear in the main paper or the supplemental material, but if they appear in the supplemental material, the authors are encouraged to provide a short proof sketch to provide intuition. +- Inversely, any informal proof provided in the core of the paper should be complemented by formal proofs provided in appendix or supplemental material. +- Theorems and Lemmas that the proof relies upon should be properly referenced. + +# 4. Experimental result reproducibility + +Question: Does the paper fully disclose all the information needed to reproduce the main experimental results of the paper to the extent that it affects the main claims and/or conclusions of the paper (regardless of whether the code and data are provided or not)? + +Answer: [Yes] + +Justification: We provide the source code in the external URL along with the implementation details in the paper. + +Guidelines: + +- The answer NA means that the paper does not include experiments. + +- If the paper includes experiments, a No answer to this question will not be perceived well by the reviewers: Making the paper reproducible is important, regardless of whether the code and data are provided or not. + +- If the contribution is a dataset and/or model, the authors should describe the steps taken to make their results reproducible or verifiable. + +- Depending on the contribution, reproducibility can be accomplished in various ways. For example, if the contribution is a novel architecture, describing the architecture fully might suffice, or if the contribution is a specific model and empirical evaluation, it may be necessary to either make it possible for others to replicate the model with the same dataset, or provide access to the model. In general, releasing code and data is often one good way to accomplish this, but reproducibility can also be provided via detailed instructions for how to replicate the results, access to a hosted model (e.g., in the case of a large language model), releasing of a model checkpoint, or other means that are appropriate to the research performed. + +- While NeurIPS does not require releasing code, the conference does require all submissions to provide some reasonable avenue for reproducibility, which may depend on the nature of the contribution. For example + +(a) If the contribution is primarily a new algorithm, the paper should make it clear how to reproduce that algorithm. +(b) If the contribution is primarily a new model architecture, the paper should describe the architecture clearly and fully. +(c) If the contribution is a new model (e.g., a large language model), then there should either be a way to access this model for reproducing the results or a way to reproduce the model (e.g., with an open-source dataset or instructions for how to construct the dataset). +(d) We recognize that reproducibility may be tricky in some cases, in which case authors are welcome to describe the particular way they provide for reproducibility. In the case of closed-source models, it may be that access to the model is limited in some way (e.g., to registered users), but it should be possible for other researchers to have some path to reproducing or verifying the results. + +# 5. Open access to data and code + +Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material? + +Answer: [Yes] + +Justification: We provide access to the code. + +Guidelines: + +- The answer NA means that paper does not include experiments requiring code. +- Please see the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details. +- While we encourage the release of code and data, we understand that this might not be possible, so "No" is an acceptable answer. Papers cannot be rejected simply for not including code, unless this is central to the contribution (e.g., for a new open-source benchmark). +- The instructions should contain the exact command and environment needed to run to reproduce the results. See the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details. +- The authors should provide instructions on data access and preparation, including how to access the raw data, preprocessed data, intermediate data, and generated data, etc. +- The authors should provide scripts to reproduce all experimental results for the new proposed method and baselines. If only a subset of experiments are reproducible, they should state which ones are omitted from the script and why. +- At submission time, to preserve anonymity, the authors should release anonymized versions (if applicable). + +- Providing as much information as possible in supplemental material (appended to the paper) is recommended, but including URLs to data and code is permitted. + +# 6. Experimental setting/details + +Question: Does the paper specify all the training and test details (e.g., data splits, hyperparameters, how they were chosen, type of optimizer, etc.) necessary to understand the results? + +Answer: [Yes] + +Justification: We have provided a detailed evaluation protocol and hyperparameters for model training in Appendix D. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The experimental setting should be presented in the core of the paper to a level of detail that is necessary to appreciate the results and make sense of them. +- The full details can be provided either with the code, in appendix, or as supplemental material. + +# 7. Experiment statistical significance + +Question: Does the paper report error bars suitably and correctly defined or other appropriate information about the statistical significance of the experiments? + +Answer: [Yes] + +Justification: We provide statistical error for each table. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The authors should answer "Yes" if the results are accompanied by error bars, confidence intervals, or statistical significance tests, at least for the experiments that support the main claims of the paper. +- The factors of variability that the error bars are capturing should be clearly stated (for example, train/test split, initialization, random drawing of some parameter, or overall run with given experimental conditions). +- The method for calculating the error bars should be explained (closed form formula, call to a library function, bootstrap, etc.) +- The assumptions made should be given (e.g., Normally distributed errors). +- It should be clear whether the error bar is the standard deviation or the standard error of the mean. +- It is OK to report 1-sigma error bars, but one should state it. The authors should preferably report a 2-sigma error bar than state that they have a $96\%$ CI, if the hypothesis of Normality of errors is not verified. +- For asymmetric distributions, the authors should be careful not to show in tables or figures symmetric error bars that would yield results that are out of range (e.g. negative error rates). +- If error bars are reported in tables or plots, The authors should explain in the text how they were calculated and reference the corresponding figures or tables in the text. + +# 8. Experiments compute resources + +Question: For each experiment, does the paper provide sufficient information on the computer resources (type of compute workers, memory, time of execution) needed to reproduce the experiments? + +Answer: [Yes] + +Justification: We have provided computing resources in Appendix D. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The paper should indicate the type of compute workers CPU or GPU, internal cluster, or cloud provider, including relevant memory and storage. + +- The paper should provide the amount of compute required for each of the individual experimental runs as well as estimate the total compute. +- The paper should disclose whether the full research project required more computer than the experiments reported in the paper (e.g., preliminary or failed experiments that didn't make it into the paper). + +# 9. Code of ethics + +Question: Does the research conducted in the paper conform, in every respect, with the NeurIPS Code of Ethics https://neurips.cc/public/EthicsGuidelines? + +Answer: [Yes] + +Justification: We follow the NeurIPS code of Ethics. + +Guidelines: + +- The answer NA means that the authors have not reviewed the NeurIPS Code of Ethics. +- If the authors answer No, they should explain the special circumstances that require a deviation from the Code of Ethics. +- The authors should make sure to preserve anonymity (e.g., if there is a special consideration due to laws or regulations in their jurisdiction). + +# 10. Broader impacts + +Question: Does the paper discuss both potential positive societal impacts and negative societal impacts of the work performed? + +Answer: [NA] + +Justification: Since our model is about science application, we believe there are no potential societal impacts of the work. + +Guidelines: + +- The answer NA means that there is no societal impact of the work performed. +- If the authors answer NA or No, they should explain why their work has no societal impact or why the paper does not address societal impact. +- Examples of negative societal impacts include potential malicious or unintended uses (e.g., disinformation, generating fake profiles, surveillance), fairness considerations (e.g., deployment of technologies that could make decisions that unfairly impact specific groups), privacy considerations, and security considerations. +- The conference expects that many papers will be foundational research and not tied to particular applications, let alone deployments. However, if there is a direct path to any negative applications, the authors should point it out. For example, it is legitimate to point out that an improvement in the quality of generative models could be used to generate deepfakes for disinformation. On the other hand, it is not needed to point out that a generic algorithm for optimizing neural networks could enable people to train models that generate Deepfakes faster. +- The authors should consider possible harms that could arise when the technology is being used as intended and functioning correctly, harms that could arise when the technology is being used as intended but gives incorrect results, and harms following from (intentional or unintentional) misuse of the technology. +- If there are negative societal impacts, the authors could also discuss possible mitigation strategies (e.g., gated release of models, providing defenses in addition to attacks, mechanisms for monitoring misuse, mechanisms to monitor how a system learns from feedback over time, improving the efficiency and accessibility of ML). + +# 11. Safeguards + +Question: Does the paper describe safeguards that have been put in place for responsible release of data or models that have a high risk for misuse (e.g., pretrained language models, image generators, or scraped datasets)? + +Answer: [NA] + +Justification: This paper poses no such risks. + +Guidelines: + +- The answer NA means that the paper poses no such risks. +- Released models that have a high risk for misuse or dual-use should be released with necessary safeguards to allow for controlled use of the model, for example by requiring that users adhere to usage guidelines or restrictions to access the model or implementing safety filters. +- Datasets that have been scraped from the Internet could pose safety risks. The authors should describe how they avoided releasing unsafe images. +- We recognize that providing effective safeguards is challenging, and many papers do not require this, but we encourage authors to take this into account and make a best faith effort. + +# 12. Licenses for existing assets + +Question: Are the creators or original owners of assets (e.g., code, data, models), used in the paper, properly credited and are the license and terms of use explicitly mentioned and properly respected? + +Answer: [Yes] + +Justification: We have adequately cited the relevant works and data sources. + +# Guidelines: + +- The answer NA means that the paper does not use existing assets. +- The authors should cite the original paper that produced the code package or dataset. +- The authors should state which version of the asset is used and, if possible, include a URL. +- The name of the license (e.g., CC-BY 4.0) should be included for each asset. +- For scraped data from a particular source (e.g., website), the copyright and terms of service of that source should be provided. +- If assets are released, the license, copyright information, and terms of use in the package should be provided. For popular datasets, paperswithcode.com/datasets has curated licenses for some datasets. Their licensing guide can help determine the license of a dataset. +- For existing datasets that are re-packaged, both the original license and the license of the derived asset (if it has changed) should be provided. +- If this information is not available online, the authors are encouraged to reach out to the asset's creators. + +# 13. New assets + +Question: Are new assets introduced in the paper well documented and is the documentation provided alongside the assets? + +Answer: [Yes] + +Justification: Our source code is well documented in an external URL, and also describes the implementation details in the Appendix. + +# Guidelines: + +- The answer NA means that the paper does not release new assets. +- Researchers should communicate the details of the dataset/code/model as part of their submissions via structured templates. This includes details about training, license, limitations, etc. +- The paper should discuss whether and how consent was obtained from people whose asset is used. +- At submission time, remember to anonymize your assets (if applicable). You can either create an anonymized URL or include an anonymized zip file. + +# 14. Crowdsourcing and research with human subjects + +Question: For crowdsourcing experiments and research with human subjects, does the paper include the full text of instructions given to participants and screenshots, if applicable, as well as details about compensation (if any)? + +Answer: [NA] + +Justification: This paper does not involve crowd sourcing nor research with human subjects. + +# Guidelines: + +- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. +- Including this information in the supplemental material is fine, but if the main contribution of the paper involves human subjects, then as much detail as possible should be included in the main paper. +- According to the NeurIPS Code of Ethics, workers involved in data collection, curation or other labor should be paid at least the minimum wage in the country of the data collector. + +# 15. Institutional review board (IRB) approvals or equivalent for research with human subjects + +Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals (or an equivalent approval/review based on the requirements of your country or institution) were obtained? + +Answer: [NA] + +Justification: This paper does not involve crowd sourcing nor research with human subjects. + +# Guidelines: + +- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. +- Depending on the country in which research is conducted, IRB approval (or equivalent) may be required for any human subjects research. If you obtained IRB approval, you should clearly state this in the paper. +- We recognize that the procedures for this may vary significantly between institutions and locations, and we expect authors to adhere to the NeurIPS Code of Ethics and the guidelines for their institution. +- For initial submissions, do not include any information that would break anonymity (if applicable), such as the institution conducting the review. + +# 16. Declaration of LLM usage + +Question: Does the paper describe the usage of LLMs if it is an important, original, or non-standard component of the core methods in this research? Note that if the LLM is used only for writing, editing, or formatting purposes and does not impact the core methodology, scientific rigorousness, or originality of the research, declaration is not required. + +Answer: [NA] + +Justification: We only use LLMs for editing. + +# Guidelines: + +- The answer NA means that the core method development in this research does not involve LLMs as any important, original, or non-standard components. +- Please refer to our LLM policy (https://neurips.cc/Conferences/2025/LLM) for what should or should not be described. + +# Supplementary Material for 3D Interaction Geometric Pre-training for Molecular Relational Learning + +A Molecular Relational Learning 22 +B Datasets 22 +B.1 Pre-Training Datasets 22 +B.2 Downstream Task Datasets 23 +CBaselines Setup 24 +D Implementation Details 25 +D.1 Evaluation Protocol 25 +D.2 Model architecture 26 +D.3 Model training 26 +E Additional Experimental Results 26 +E.1 Molecular Property Prediction Pre-training with Large-Scale Datasets 26 +E.2 Extrapolation in Molecular Interaction Task 26 +E.3 Ablation Studies 28 +E.4 Further Virtual Interaction Environment Analysis 28 +E.5 3D Encoder Pre-training Approaches 29 + +F Pseudocode 30 + +# A Molecular Relational Learning + +In this section, we provide further clarification on molecular relational learning by contrasting it with conventional molecular property prediction tasks. As illustrated in Figure 5 (a), conventional molecular property prediction focuses on learning the properties of a single molecule. Models like GraphMVP, 3D Infomax, and MoleculeSDE utilize the 3D information of individual molecules during pre-training to improve performance in downstream tasks aimed at predicting single molecular properties. + +In contrast, as shown in Figure 5 (b), molecular relational learning focuses on learning the properties of molecules after their interactions. Our pre-training approach trains both encoders simultaneously to learn 3D information from the virtual environment $g_{vr}$ . What sets our approach apart from traditional molecular pretraining is its specific tailoring to our Molecular Relational Learning strategy. This allows the two encoders to learn how paired molecules interact in 3D space, which is essential for various downstream tasks in Molecular Relational Learning. + +![](images/25220107eb1e2b36e9f62409370c11e2a81d7630456cbd160413cf12151ed1e0.jpg) +(a) Molecule Property Prediction + +![](images/304fff636d7a5ef5d4a3e10b4f93ab04029233b54cd44dcd5809d7fbf5a8a038.jpg) +(b) Molecular Relational Learning +Figure 5: Difference between the conventional pre-training strategy for (a) molecular property prediction and our (b) molecular relational learning. + +# B Datasets + +# B.1 Pre-Training Datasets + +We utilize three distinct datasets, i.e., Chromophore, CombiSolv, and DDI, to pre-train 3DMRL for each downstream task as described in Section 5. Specifically, we use the Chromophore dataset for downstream tasks involving the optical properties of chromophores, the CombiSolv dataset for tasks related to the solvation free energy of solutes, and the DDI dataset, which we created for the drug-drug interaction task. + +- The Chromophore dataset [16] consists of 20,236 combinations derived from 6,815 chromophores and 1,336 solvents, provided in SMILES string format. For pre-training, we initially convert chromophores and solvents into their respective 3D structures via rdkit, resulting in 6,524 3D structures for chromophores and 1,255 for solvents. These 6,524 unique chromophores are then randomly paired with the 1,255 solvents to generate a sufficient number of pairs. Out of the possible 8,187,620 chromophore-solvent combinations, we randomly sample $1\%$ , which corresponds to 81,876 pairs, for pre-training. +- The CombiSolv dataset [39] contains 10,145 combinations derived from 1,368 solutes and 291 solvents, provided in SMILES string format. Similar to our approach with the Chromophore dataset, we first convert solutes and solvents into their corresponding 3D structures, yielding + +1,368 3D structures for solutes and 290 for solvents. From the potential random combinations, we select 79,344 solute-solvent pairs, representing $20\%$ of all possible pairs. + +- For the DDI dataset, we compile drug-drug pairs from the ZhangDDI [48], ChChMiner [49], and DeepDDI [33] datasets. From a total of 235,547 positive pairs, we randomly sample $40\%$ (i.e., 94,218 pairs) for use as the pre-training dataset. While chromophores and solutes act as the larger molecule $g^{1}$ in molecular interaction tasks, in the DDI dataset, we designate the drug with the larger radius as the larger molecule. + +# B.2 Downstream Task Datasets + +Molecular Interaction Prediction. For the molecular interaction prediction task, we transform the SMILES strings into graph structures using the CIGIN implementation available on GitHub [32]. Regarding the datasets related to solvation free energies, such as MNSol, FreeSolv, CompSol, Abraham, and CombiSolv, we utilize SMILES-based datasets from previous studies [39]. Following previous work [21], we specifically filter the data to include only solvation free energies measured at temperatures of $298\mathrm{K}$ $(\pm 2)$ and exclude any data involving ionic liquids and ionic solutes [39]. + +- The Chromophore dataset [16] consists of 20,236 combinations derived from 6,815 chromophores and 1,336 solvents, provided in SMILES string format. This dataset includes optical properties sourced from scientific publications, with unreliable experimental results being excluded after thorough examination of absorption and emission spectra. In our work, we assess model performance by predicting three key properties: maximum absorption wavelength (Absorption), maximum emission wavelength (Emission), and excited state lifetime (Lifetime), which are crucial for designing chromophores for specific applications. To ensure the integrity of each dataset, we remove any NaN values that were not reported in the original publications. Additionally, following previous work [21], for the Lifetime data, we apply log normalization to the target values to mitigate skewness in the dataset, thereby enhancing training stability. +- The MNSol dataset [26] features 3,037 experimentally measured free energies of solvation or transfer for 790 distinct solutes and 92 solvents. For our study, we focus on 2,275 pairs comprising 372 unique solutes and 86 solvents, in alignment with prior research [39]. +- The FreeSolv dataset [29] offers 643 hydration free energy values, both experimental and calculated, for small molecules in water. In our research, we utilize 560 experimental measurements, consistent with the dataset selection criteria from previous studies [39]. +- The CompSol dataset [30] has been designed to illustrate the impact of hydrogen-bonding association effects on solvation energies. For our study, we analyze 3,548 solute-solvent pairs, encompassing 442 distinct solutes and 259 solvents, in accordance with prior research parameters [39]. +- The Abraham dataset [13], curated by the Abraham research group at University College London, provides extensive data on solvation. For this study, we focus on 6,091 solute-solvent combinations, comprising 1,038 distinct solutes and 122 solvents, as outlined in previous research [39]. +- The CombiSolv dataset [39] integrates the data from MNSol, FreeSolv, CompSol, and Abraham, encompassing a total of 10,145 solute-solvent combinations. This dataset features 1,368 unique solutes and 291 distinct solvents. + +Drug-Drug Interaction (DDI) Prediction. In the drug-drug interaction prediction task, we utilize the positive drug pairs provided in the MIRACLE GitHub repository3, which excludes data instances that cannot be represented as graphs from SMILES strings. To create negative samples, we generate a corresponding set by sampling from the complement of the positive drug pairs. This approach is applied to both datasets. Additionally, for the classification task, we adhere to the graph conversion process outlined by MIRACLE [41]. + +- The ZhangDDI dataset [48] includes data on 548 drugs and 48,548 pairwise interactions, along with various types of similarity information pertaining to these drug pairs. + +Table 4: Statistics of datasets. $\mathcal{G}^1$ and $\mathcal{G}^2$ are defined in Section 5.1. + +
TaskDatasetG1G2# G1# G2# Pairs
Molecular InteractionChromophore4AbsorptionChromophoreSolvent6,41672517,276
EmissionChromophoreSolvent6,4121,02118,141
LifetimeChromophoreSolvent2,7552476,960
MNSol5SoluteSolvent372862,275
FreeSolv6SoluteSolvent5601560
CompSol7SoluteSolvent4422593,548
Abraham8SoluteSolvent1,0381226,091
CombiSolv9SoluteSolvent1,49532610,145
Drug-Drug InteractionZhangDDI10Small-molecule DrugSmall-molecule Drug54454440,255
ChChMiner11Small-molecule DrugSmall-molecule Drug94994921,082
+ +- The ChChMiner dataset [49] comprises 1,322 drugs and 48,514 annotated DDIs, sourced from drug labels and scientific literature. + +Despite the ChChMiner dataset containing a significantly higher number of drug instances compared to the ZhangDDI dataset, the number of labeled DDIs is nearly equivalent. This suggests that the ChChMiner dataset exhibits a much sparser network of relationships between drugs. + +# C Baselines Setup + +To validate the effectiveness of 3DMRL, we primarily evaluate molecular relational learning model architectures trained from scratch for downstream tasks, as well as the same models that are first pre-trained with 3DMRL and then fine-tuned for various downstream tasks. We include the following molecular relational learning model architectures: + +- MPNN (Message Passing Neural Networks) [11] was originally proposed to predict the various chemical properties of a single molecule. For molecular relational learning tasks, we independently encode each molecule in a pair using MPNN and then concatenate their representations. To apply 3DMRL for MPNN, we first obtain the atom representation matrices $\mathbf{E}^1$ and $\mathbf{E}^2$ using $f_{\mathrm{2D}}^{1}$ and $f_{\mathrm{2D}}^{1}$ , which are MPNNs. Then, we directly use $\mathbf{E}^1$ and $\mathbf{E}^2$ instead of the $\mathbf{H}^1$ and $\mathbf{H}^2$ , which considers the interaction between two molecules in Section 3.2. That is, we obtain graph-level embeddings $\mathbf{z}_{\mathrm{2D}}^{1}$ and $\mathbf{z}_{\mathrm{2D}}^{2}$ via $\mathbf{E}^1$ and $\mathbf{E}^2$ with Set2set readout function. Following contrastive learning is done with $\mathbf{z}_{\mathrm{2D}}^{1}$ and $\mathbf{z}_{\mathrm{2D}}^{2}$ , and the edge representations $\mathbf{e}_{\mathrm{2D}}^{k,l}$ and initial atom representations for relative geometry $\hat{\mathbf{X}}$ is obtained through $\mathbf{E}^1$ and $\mathbf{E}^2$ . One can simply alternate $\mathbf{H}^1$ and $\mathbf{H}^2$ in Section 4 to $\mathbf{E}^1$ and $\mathbf{E}^2$ . +- AttentiveFP [43] was also initially proposed to predict various chemical properties of individual molecules by employing a graph attention mechanism to gather more information from relevant molecular datasets. For molecular relational learning tasks, we independently encode each molecule in a pair using MPNN and then concatenate their representations. + +More specifically, AttentiveFP first obtain atom representation matrices $\mathbf{H}^1$ and $\mathbf{H}^2$ using $f_{\mathrm{2D}}^{1}$ and $f_{\mathrm{2D}}^{1}$ , which consist of GAT and GRU layers. Then, the model obtain initial molecule representation $\tilde{\mathbf{z}}_{\mathrm{2D}}^{1}$ and $\tilde{\mathbf{z}}_{\mathrm{2D}}^{2}$ which are further enhanced by considering other molecules in a batch through GAT layers. After passing multiple GAT layers, the model obtain final molecule representations $\tilde{\mathbf{z}}_{\mathrm{2D}}^{1}$ and $\tilde{\mathbf{z}}_{\mathrm{2D}}^{2}$ . In our framework, contrastive learning is done with $\mathbf{z}_{\mathrm{2D}}^{1}$ and $\mathbf{z}_{\mathrm{2D}}^{2}$ , + +and the edge representations $\mathbf{e}_{2\mathrm{D}}^{k,l}$ and initial atom representations for relative geometry $\hat{\mathbf{X}}$ is obtained through $\mathbf{H}^1$ and $\mathbf{H}^2$ . + +- CIGIN (Chemically Interpretable Graph Interaction Network) [32] proposes to model the interaction between the molecules through a dot product between atoms in paired molecules. By doing so, they successfully predict the solubility of drug molecules. We provide detailed descriptions on how to apply 3DMRL for CIGIN in Section 4. +- CGIB (Conditional Graph Information Bottleneck) and $\mathbf{CGIB}_{\mathrm{cont}}$ (Conditional Graph Information Bottleneck with Contrastive Learning)[21] aim to enhance generalization in molecular relational learning by identifying the core substructure of molecules during chemical reactions, based on the information bottleneck theory. While CIGIN is limited to predicting drug solubility, CGIB and $\mathbf{CGIB}_{\mathrm{cont}}$ extend molecular relational learning to predict the optical properties of chromophores in various solvents, molecule solubility in various solvents, and drug-drug interactions. + +CGIB and $\mathbf{CGIB}_{\mathrm{cont}}$ model architectures are highly similar to CIGIN, but they have another branch named compress module, which aims to inject noise to the atoms that are not important during the model. Specifically, they obtain $\mathbf{T}^1$ that is node representation matrix with noise, and obtain $\mathbf{z}_{\mathcal{G}_{\mathrm{CIB}}^1}$ from the noise injected matrix along with $\mathbf{z}_{\mathcal{G}^1}$ and $\mathbf{z}_{\mathcal{G}^2}$ which are obtained from $\mathbf{H}^1$ and $\mathbf{H}^2$ , respectively. To apply 3DMRL for CGIB, we pre-train the model without noise injection module, thereby using $\mathbf{H}^1$ , $\mathbf{H}^2$ , $\mathbf{z}_{\mathcal{G}^1}$ , and $\mathbf{z}_{\mathcal{G}^2}$ in CGIB as $\mathbf{H}^1$ , $\mathbf{H}^2$ , $\mathbf{z}_{\mathrm{2D}}^1$ , and $\mathbf{z}_{\mathrm{2D}}^2$ in Section 4. After pre-training staget, all the modules including noise injection module is trained for the downstream tasks. + +In addition to the model architectures, we also compare the recent state-of-the-art molecular pretraining methods based on CIGIN architecture. Since molecular pre-training methods are specifically designed for a single molecule, we pre-train each molecule encoder in CIGIN architecture and adopted the pre-trained weights for molecular relational learning downstream tasks. In Section 5, we include following molecular pre-training approaches: + +- No pre-training does not involve pertaining process and fine-tune the model using labeled data +- 3D Infomax [36] increase the mutual information between 2D and 3D molecular representations using contrastive learning +- GraphMVP [24] incorporates a generative pre-training framework in addition to contrastive learning +- MoleculeSDE [25] designs a denoising framework to capture the 3D geometric distribution of molecules, thereby revealing the relationship between the score function and the molecular force field. + +To apply these approaches for MRL, we first pre-train the each encoder $f_{2\mathrm{D}}^{1}$ and $f_{2\mathrm{D}}^{2}$ in Section 3.2 with the above approaches. Then, the pre-trained encoders $f_{2\mathrm{D}}^{1}$ and $f_{2\mathrm{D}}^{2}$ are utilized to output the representations $\mathbf{E}^{1}$ and $\mathbf{E}^{2}$ , following the remaining pipeline of the model outlined in Section 3.2. That is, each molecule encoder $f_{2\mathrm{D}}^{1}$ and $f_{2\mathrm{D}}^{2}$ implicitly possesses knowledge about the 3D structure of individual molecules, but not the complex interaction geometry between multiple molecules. + +# D Implementation Details + +# D.1 Evaluation Protocol + +Following Pathak et al. [32], for the molecular interaction prediction task, we evaluate the models under a 5-fold cross-validation scheme. The dataset is randomly split into 5 subsets and one of the subsets is used as the test set, while the remaining subsets are used to train the model. A subset of the test set is selected as the validation set for hyperparameter selection and early stopping. We repeat 5-fold cross-validation three times (i.e., 15 runs in total) and report the accuracy and standard deviation of the repeats. For the DDI prediction task [21], we conduct experiments on two different out-of-distribution scenarios, namely molecule split and scaffold split. For the molecule split, the performance is evaluated when the models are presented with new molecules not included in the training dataset. Specifically, let $\mathbb{G}$ denote the total set of molecules in the dataset. Given $\mathbb{G}$ , we split $\mathbb{G}$ into $\mathbb{G}_{\mathrm{old}}$ and $\mathbb{G}_{\mathrm{new}}$ , so that $\mathbb{G}_{\mathrm{old}}$ contains the set of molecules that have been seen in the training phase, and $\mathbb{G}_{\mathrm{new}}$ contains the set of molecules that have not been seen in the training phase. + +Then, the new split of dataset consists of $\mathcal{D}_{\mathrm{train}} = \{(\mathcal{G}^1,\mathcal{G}^2)\in \mathcal{D}|\mathcal{G}^1\in \mathbb{G}_{\mathrm{old}}\wedge \mathcal{G}^2\in \mathbb{G}_{\mathrm{old}}\}$ and $\mathcal{D}_{\mathrm{test}} = \{(\mathcal{G}^1,\mathcal{G}^2)\in \mathcal{D}|\mathcal{(G}^1\in \mathbb{G}_{\mathrm{new}}\wedge \mathcal{G}^2\in \mathbb{G}_{\mathrm{new}})\vee (\mathcal{G}^1\in \mathbb{G}_{\mathrm{new}}\wedge \mathcal{G}^2\in \mathbb{G}_{\mathrm{old}})\vee (\mathcal{G}^1\in \mathbb{G}_{\mathrm{old}}\wedge \mathcal{G}^2\in$ $\mathbb{G}_{\mathrm{new}})\}$ . We use a subset of $\mathcal{D}_{\mathrm{test}}$ as the validation set in inductive setting. In the scaffold split setting [14], just like in the molecule split, molecules corresponding to scaffolds that were not seen during training will be used for testing. For both splits, we repeat 5 independent experiments with different random seeds on split data, and report the accuracy and the standard deviation of the repeats. In both scenarios, we split the data into training, validation, and test sets with a ratio of $60 / 20 / 20\%$ . + +# D.2 Model architecture + +For the 2D MRL model, following a previous work [32], we use 3-layer MPNNs [11] as our backbone molecule encoder to learn the representation of solute and solvent for the molecular interaction prediction, while we use a GIN [45] to encode both drugs for the drug-drug interaction prediction task [21]. We utilize a hidden dimension of 56 for molecular interaction tasks and 300 for drug-drug interaction tasks, employing the ReLU activation function for both. For the 3D virtual environment encoder $f_{\mathrm{3D}}$ , we utilize SchNet [34], which guarantees an SE(3)-invariant representation of the environment. For both molecular interaction and drug-drug interaction tasks, we configure SchNet with 128 hidden channels, 128 filters, 6 interaction layers, and a cutoff distance of 5.0. + +# D.3 Model training + +For model optimization during Pre-training stage, we employ the Adam optimizer with an initial learning rate of 0.0005 for the chromophore task, 0.0001 for the solvation free energy task, and 0.0005 for the DDI tasks. The model is optimized over 100 epochs during pre-training. + +In the downstream tasks, the learning rate was reduced by a factor of $10^{-1}$ after 20 epochs of no improvement in model performance in validation set, following the approach in a previous work [32], with the initial learning rate of 0.005 for the chromophore task, 0.001 for the solvation free energy task, and 0.0005 for the DDI tasks. + +Computational resources. We perform all pre-training on a 40GB NVIDIA A6000 GPU, whereas all downstream tasks are executed on a 24GB NVIDIA GeForce RTX 3090 GPU. + +Software configuration. Our model is implemented using Python 3.7, PyTorch 1.9.1, RDKit 2020.09.1, and Pytorch-geometric 2.0.3. + +# E Additional Experimental Results + +# E.1 Molecular Property Prediction Pre-training with Large-Scale Datasets + +Although MPP pre-training approaches demonstrate unsatisfactory performance in Section 5, a positive aspect is their ability to leverage large-scale datasets containing both 2D and 3D molecular information. Consequently, we further explore whether utilizing a large-scale pre-training dataset can enhance MPP pre-training strategies in MRL tasks. To do so, we pre-train the encoders with each strategy with randomly sampled 50K molecules in GEOM dataset [2], which consists of 2D topological information and 3D geometric information, following the previous work [24]. In Table 5, we observe that a large-scale pre-training dataset does not consistently result in performance improvements for MRL downstream tasks and can still cause negative transfer in various tasks. On the other hand, we note that MoleculeSTM benefits the most from the large-scale dataset among the strategies, likely due to the complexity of its denoising framework, which necessitates a large-scale dataset to learn the data distribution effectively. Nevertheless, it still exhibits negative transfer in the FreeSolve dataset and performs worse than 3DMRL, highlighting the need for a pre-training strategy specifically tailored to molecular relational learning. + +# E.2 Extrapolation in Molecular Interaction Task + +The model's generalization ability in out-of-distribution (OOD) datasets is crucial for its application in real-world scientific discovery processes. To this end, we further conduct experiments on molecular interaction tasks by assuming out-of-distribution scenarios, as shown in Table 6. Specifically, we split the dataset based on molecular structure, i.e., molecule split and scaffold split, similar to the approach + +Table 5: Performance comparison of CIGIN model on molecular interaction tasks using different pre-training strategies and pre-training dataset (RMSE) $(\downarrow)$ . The blue color signifies a positive transfer between the pre-training task and the downstream task, whereas the orange color denotes a negative transfer between the pre-training task and the downstream task. Pre-training Dataset indicates the pre-training datasets used during pre-training. + +
StrategyPre-training DatasetChromophoreMNSolFreeSolvCompSolAbrahamCombiSolv
AbsorptionEmissionLifetime
No Pre-training-19.66 (0.69)25.84 (0.23)0.821 (0.017)0.567 (0.014)0.884 (0.074)0.331 (0.029)0.412 (0.028)0.458 (0.002)
MPP (molecular property prediction) Pre-training
3D InfomaxMRL18.71 (0.61)24.59 (0.22)0.790 (0.022)0.585 (0.015)0.873 (0.103)0.321 (0.041)0.426 (0.036)0.464 (0.004)
GEOM18.82 (0.24)25.14 (0.18)0.795 (0.021)0.589 (0.027)0.899 (0.080)0.319 (0.019)0.418 (0.023)0.466 (0.017)
GraphMVPMRL18.40 (0.62)24.73 (0.14)0.797 (0.022)0.561 (0.025)1.010 (0.115)0.301 (0.025)0.418 (0.020)0.437 (0.015)
GEOM18.85 (0.74)24.87 (0.54)0.784 (0.014)0.551 (0.013)0.900 (0.059)0.325 (0.007)0.410 (0.036)0.437 (0.007)
MoleculeSDEMRL18.56 (0.24)24.91 (0.10)0.836 (0.040)0.564 (0.018)0.971 (0.122)0.308 (0.024)0.426 (0.028)0.454 (0.012)
GEOM18.72 (0.16)24.77 (0.48)0.773 (0.023)0.560 (0.086)0.909 (0.142)0.290 (0.008)0.399 (0.034)0.449 (0.007)
MRL (molecular relational learning) Pre-training
3DMRLMRL18.00 (0.17)24.21 (0.09)0.729 (0.014)0.528 (0.019)0.839 (0.105)0.277 (0.006)0.371 (0.031)0.435 (0.006)
+ +used in the DDI task in Section 5. It is important to note that this scenario is significantly more challenging than the out-of-distribution DDI task in Section 5 because it involves a regression task, which can also be viewed as an extrapolation task. As shown in Table 6, we observe that pre-training approaches generally benefit model performance in extrapolation tasks, with the exception of one case, namely 3D Infomax for the Lifetime dataset. Among the pre-training approaches, 3DMRL performs the best, underscoring the extrapolation capability of 3DMRL. + +Table 6: Performance comparison of the CIGIN model on extrapolation in molecular interaction tasks using different pre-training strategies (RMSE) $(\downarrow)$ . + +
StrategyMolecule SplitScaffold Split
AbsorptionEmissionLifetimeAbsorptionEmissionLifetime
No Pre-training27.51 (0.74)37.04 (1.07)1.205 (0.033)59.55 (1.35)60.11 (1.98)1.221 (0.033)
MPP (molecular property prediction) Pre-training
3D Infomax27.38 (1.19)36.98 (1.24)1.257 (0.050)58.34 (1.89)58.67 (1.00)1.207 (0.041)
GraphMVP26.93 (1.89)36.51 (0.92)1.201 (0.034)59.27 (1.57)57.67 (1.14)1.199 (0.024)
MoleculeSDE27.26 (1.19)36.48 (1.12)1.135 (0.077)57.75 (0.74)58.74 (1.02)1.214 (0.010)
MRL (molecular relational learning) Pre-training
3DMRL25.01 (1.51)34.66 (0.89)1.033 (0.027)57.58 (1.62)57.53 (1.13)1.178 (0.010)
+ +# E.3 Ablation Studies + +We provide further ablation studies on molecular interaction task and drug-drug interaction task in Table 7 and 8, respectively. + +Table 7: Further results from ablation studies on molecular interaction tasks. + +
StrategyChromophoreMNSolFreeSolvCompSolAbrahamCombiSolv
AbsorptionEmissionLifetime
Only Glob.18.30 (0.16)24.70 (0.16)0.739 (0.015)0.531 (0.022)0.874 (0.060)0.301 (0.018)0.376 (0.029)0.458 (0.014)
Only Local19.34 (0.50)24.80 (0.05)0.804 (0.011)0.587 (0.019)1.184 (0.173)0.330 (0.028)0.391 (0.020)0.466 (0.021)
3DMRL18.00 (0.17)24.21 (0.09)0.729 (0.014)0.528 (0.019)0.839 (0.105)0.277 (0.006)0.371 (0.031)0.435 (0.006)
+ +Table 8: Further results from ablation studies on drug-drug interaction tasks. + +
Strategy(a) Molecule Split(b) Scaffold Split
ZhangDDIChChMinerZhangDDIChChMiner
Only Glob.73.09 (0.83)77.68 (0.55)73.18 (0.59)76.79 (1.13)
Only Local73.45 (1.29)75.93 (1.14)73.41 (2.28)74.29 (1.79)
3DMRL74.00 (0.72)78.93 (0.59)74.85 (1.58)78.56 (1.03)
+ +# E.4 Further Virtual Interaction Environment Analysis + +Sensitivity Analysis on DDI Datasets. In Table 9, we provide sensitivity analysis results in drug-drug interaction tasks. + +Table 9: Sensitivity analysis on $n$ in drug-drug interaction tasks. + +
Molecule SplitScaffold Split
ZhangDDIChChMinerZhangDDIChChMiner
n = 273.7777.1574.7677.01
n = 574.0078.9374.8578.56
n = 1073.9679.1274.3677.76
n = 2073.9478.7574.0377.64
+ +Further Environment Analysis. While we propose assigning a single small molecule to each target atom in Section 4.1, we also investigate the impact of varying the number of assigned small molecules per atom in the larger molecule. As illustrated in Figure 6, we observe a decline in model performance as the number of small molecules per atom increases, given a fixed number of target atoms $n$ . This suggests that modeling interactions between multiple small molecules and a single atom in a larger molecule can degrade model performance. This is consistent with the scientific understanding that, although hydrogen bonding can occasionally allow multiple molecules to interact with a single atom simultaneously, steric and electronic hindrances frequently impede such interactions. Thus, we contend that our proposed virtual interaction geometry appropriately reflects the real-world physics in molecular interactions. + +![](images/1cea21882ed8d7a2a0c5ad0ce1227b0c03053050ed084bac5ce74d5a0b5e84c7.jpg) +Figure 6: Further environment analysis results. + +![](images/3daae4b85597494b888faca5c482b844f61824a91e467397a85c1fd8a57a0235.jpg) + +![](images/112221119e16e56e19e7bdf5876b449fe1d81c75203600ce45d5c376894dff83.jpg) + +Number of Larger Molecules. In Section 4.1, we initially constructed the virtual geometry in a one-to-many manner (one larger molecule and many smaller molecules) to effectively mimic the + +explicit solvent model in traditional MD simulations. However, in this section, we explore the many-to-many configurations between larger molecules and smaller molecules. In Table 10, we observe that the best performance was achieved when there was only one larger molecule. However, since the performance differences were not significant, we can conclude that our model is robust across various configurations. + +Table 10: Model performance on different number of larger molecules. + +
# Larger moleculesAbsorptionEmissionLifetime
1 (Ours)18.0024.210.729
218.2824.430.738
318.3724.350.749
+ +# E.5 3D Encoder Pre-training Approaches + +Since the core concept of our paper is to inject 3D information into a 2D encoder, we choose baseline approaches that pre-train 2D molecular encoder with 3D information. In this section, we compare the approaches that pre-train 3D molecular encoder with 3D information. To do so, since the elaborately calculated 3D structure of the molecules is not available for our datasets, we first calculate the 3D structure of the molecules in the dataset using RDKit ETKDG algorithm. However, some of the molecules in the dataset were not able to obtain 3D structures through RDKit ETKDG algorithm. We excluded these molecules from the experiment, and the results are shown below. + +Before Conversion: Absorption - 17,276 pairs, Emission - 18,141 pairs, and Lifetime - 6,960 pairs. + +After Conversion: Absorption - 16,756 pairs, Emission - 17,525 pairs, and Lifetime - 6,740 pairs. + +Table 11: Performance of various 3D encoder pre-training strategies in RMSE (↓). Note that these results are not directly comparable since some of the molecules in the dataset were not able to obtain 3D structures through RDKit ETKDG algorithm. + +
AbsorptionEmissionLifetime
3D-EMGP18.6224.060.753
SliDe21.9628.870.859
Frad19.5828.430.781
3DMRL18.0024.210.729
+ +# F Pseudocode + +In this section, we provide pseudocode of 3DMRL in Algorithm 1. + +Algorithm 1 Overall framework of 3DMRL. +1: Input: 2D molecular topology graphs \(g_{\mathrm{2D}}^{1}, g_{\mathrm{2D}}^{2}\) 3D molecular geometric graphs \(g_{\mathrm{3D}}^{1}, g_{\mathrm{3D}}^{2}\) 2D graph encoders \(f_{\mathrm{2D}}^{1}, f_{\mathrm{2D}}^{2}\) 3D Virtual Interaction Geometry Encoder \(f_{\mathrm{3D}}\) +2: Pre-Training Stage: +3: For epoch in epochs: \(\mathbf{z}_{\mathrm{2D}}^{1}, \mathbf{z}_{\mathrm{2D}}^{2}, \mathbf{H}^{1}, \mathbf{H}^{2} = 2\mathrm{D}\) MRL ENCODER (\(g_{\mathrm{2D}}^{1}, g_{\mathrm{2D}}^{2}\)) +5: \(\mathbf{z}_{\mathrm{2D}} = (\mathbf{z}_{\mathrm{2D}}^{1}||\mathbf{z}_{\mathrm{2D}}^{2})\) +6: \(g_{\mathrm{vr}} = \mathrm{VIRTUAL~INTERACTION~GEOMETRY~CONSTRUCTION~}(g_{\mathrm{3D}}^{1}, g_{\mathrm{3D}}^{2})\) +7: \(\mathbf{z}_{\mathrm{3D}} = f_{\mathrm{3D}}(g_{\mathrm{vr}})\) /\*Virtual Geometry Encoding via SchNet \*/ +8: \(\mathcal{L}_{\mathrm{glob.}} = \mathrm{SE}(3)\) INVARIANT GLOBAL GEOMETRY LEARNING (\(\mathbf{z}_{\mathrm{2D}}, \mathbf{z}_{\mathrm{3D}}\)) +9: \(\mathcal{L}_{\mathrm{local.}} = \frac{1}{n}\sum_{i=1}^{n}\mathrm{SE}(3)\) EQUIVARIANT LOCAL RELATIVE GEOMETRY LEARNING (\(g_{\mathrm{3D}}^{2,i}, \mathbf{H}^{1}, \mathbf{H}^{2}\)) +10: \(\mathcal{L}_{\mathrm{pre - train.}} = \mathcal{L}_{\mathrm{glob.}} + \alpha \cdot \mathcal{L}_{\mathrm{local}}\) +11: Update \(f_{\mathrm{2D}}^{1}, f_{\mathrm{2D}}^{2}\), and \(f_{\mathrm{3D}}\) +12: Function 2D MRL ENCODER (\(g_{\mathrm{2D}}^{1}, g_{\mathrm{2D}}^{2}\)) +13: \(\mathbf{E}^{1} = f_{\mathrm{2D}}^{1}(g_{\mathrm{2D}}^{1}), \quad \mathbf{E}^{1} = f_{\mathrm{2D}}^{2}(g_{\mathrm{2D}}^{2})\) +14: \(\mathbf{I}_{ij} = \sin(\mathbf{E}_i, \mathbf{E}_j^2)\) where sim(,) is cosine similarity +15: \(\tilde{\mathbf{E}}^{1} = \mathbf{I} \cdot \tilde{\mathbf{E}}^{2}, \quad \tilde{\mathbf{E}}^{2} = \mathbf{I}^{\top} \cdot \mathbf{E}^{1}\) +16: \(\mathbf{H}^{1} = (\mathbf{E}^{1}||\tilde{\mathbf{E}}^{1}), \quad \mathbf{H}^{2} = (\mathbf{E}^{2}||\tilde{\mathbf{E}}^{2})\) +17: \(\mathbf{z}_{\mathrm{2D}}^{1} = \operatorname{Set2set}(\mathbf{H}^{1}), \quad \mathbf{z}_{\mathrm{2D}}^{2} = \operatorname{Set2set}(\mathbf{H}^{2})\) +18: return \(\mathbf{z}_{\mathrm{2D}}^{1}, \mathbf{z}_{\mathrm{2D}}^{2}, \mathbf{H}^{1}, \mathbf{H}^{2}\) +19: Function VIRTUAL INTERACTION GEOMETRY CONSTRUCTION (\(g_{\mathrm{3D}}^{1}, g_{\mathrm{3D}}^{2}\)) +20: Randomly select \(n\) atoms in larger molecule \(g_{\mathrm{3D}}^{1}\) +21: Copy small molecule \(g_{\mathrm{3D}}^{2}\) to \(n\) small molecules \(g_{\mathrm{3D}}^{2,1}, \ldots, g_{\mathrm{3D}}^{2,i}, \ldots, g_{\mathrm{3D}}^{2,n}\) +22: Generate a normalized random Gaussian noise vector \(\varepsilon\) +23: Create new 3D coordinates for each smaller molecule \(g_{\mathrm{3D}}^{2,i}\) +\(\mathbf{R}^{2,i} = \mathbf{R}^2 + \varepsilon_i * r^2 + \mathbf{R}_i^1\) /\* Broadcasting operation \*/ +24: Create virtual interaction geometry \(g_{\mathrm{vr}}\) +\(\mathbf{R}_{\mathrm{vr}} = (\mathbf{R}^1||\mathbf{R}^{2,1}||...||\mathbf{R}^{2,i}||...||\mathbf{R}^{2,n})\) +\(\mathbf{X}_{\mathrm{vr}} = (\mathbf{X}^1||\mathbf{X}^2||...||\mathbf{X}^2)\) +\(g_{\mathrm{vr}} = (\mathbf{X}_{\mathrm{vr}}, \mathbf{R}_{\mathrm{vr}})\) +25: return \(g_{\mathrm{vr}}\) +26: Function SE(3) INVARIANT GLOBAL GEOMETRY LEARNING (\(\mathbf{z}_{\mathrm{2D}}, \mathbf{z}_{\mathrm{3D}}\) ) +27: return \(\mathcal{L}_{\mathrm{glob.}} = -\frac{1}{N_{\mathrm{batch}}} \sum_{i=1}^{N_{\mathrm{batch}}} \left[ \log \frac{\textit{e}^{\sin(\mathbf{z}_{\mathrm{2D}}, i, z_{\mathrm{3D}}, i)/\tau}}{\sum_{k=1}^{N_{\mathrm{batch}}} \textit{e}^{\sin(\mathbf{z}_{\mathrm{2D}}, i, z_{\mathrm{3D}}, k)/\tau}} + \log \frac{\textit{e}^{\sin(\mathbf{z}_{\mathrm{3D}}, i, z_{\mathrm{2D}}, i)/\tau}}{\sum_{k=1}^{N_{\mathrm{batch}}} \textit{e}^{\sin(\mathbf{z}_{\mathrm{3D}}, i, z_{\mathrm{2D}}, k)/\tau}} \right]\) +28: Function SE(3) EQUIVARIANT LOCAL RELATIVE GEOMETRY LEARNING (\(g_{\mathrm{3D}}^{2,i}, \mathbf{H}^{1}, \mathbf{H}^{2}\)) +29: For all edges (\(k,l\) in \(g_{\mathrm{3D}}^{2,i}\): +\[ F_k, l = \left( \frac{\mathbf{k}_r - \mathbf{k}_l}{||\mathbf{k}_r - \mathbf{k}_l||}, \frac{\mathbf{k}_r\times\mathbf{k}_l}{||\mathbf{k}_r\times\mathbf{k}_l||}, \frac{\mathbf{k}_r - \mathbf{k}_l}{||\mathbf{k}_r - \mathbf{k}_l||} \times \frac{\mathbf{k}_r\times\mathbf{k}_l}{||\mathbf{k}_r\times\mathbf{k}_l||} \right), /*\) Construct Orthogonal Frame +where \(r_k \in \mathbb { R } ^ { 3 }\) indicates the position of atoms \(k\) +\[ e^{\boldsymbol {k}, l} = P r o j c t i o n _ {\boldsymbol {F}, l} (\boldsymbol {r}_k, r_l) / * S e t (3)- I n v a r i a n t F e l d u s e +\[ e^{\boldsymbol {k}, l} = M L P (H _ { k } ^ { 2 } | | H _ { l } ^ { 2 } ) / * S e t (3)- I n v a r i a n t F e l d u s e +\[ e ^ {k , l } = e ^ {k , l } + e ^ {k , l } . / * S e t (3)- I n v a r i a n t F e l d u s e +\[ X = (H _ { l } ^ { 2 } | | H _ { l } ^ { 1 } ) / * S e t (3)- I n v a r i a n t F e l d u s e +\[ h _ { k , l } = G N N (X , E ) , w h e r e E i d e n s i d e s a l l e d g e s i n g _ { 3 D } ^ { 2 , i } / * S e t (3)- I n v a r i a n t F e l d u s e +\[ f _ { k } = \sum _ { l } h _ { k , l } \odot F _ { k , l } / * S e t (3)- I n v a r i a n t F e l d u s e +] +return L _ { c o l a r } = 1 / N ^ { 2 } ∑ _ { k = 1 } ^ { N ^ { 2 } } \| f _ { k } ^ { i } - f _ { k } ^ { i } \| _ { 2 } ^ { 2 } \ No newline at end of file diff --git a/NeurIPS/2025/3D Interaction Geometric Pre-training for Molecular Relational Learning/images.zip b/NeurIPS/2025/3D Interaction Geometric Pre-training for Molecular Relational Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..08cde7bcfcaca08120b363ae5a661ac5d111e311 --- /dev/null +++ b/NeurIPS/2025/3D Interaction Geometric Pre-training for Molecular Relational Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a6bd33f1d6d952b2a0622709e4eaa25409107891e1331a462f5222b144bc8fb8 +size 701086 diff --git a/NeurIPS/2025/3D Interaction Geometric Pre-training for Molecular Relational Learning/layout.json b/NeurIPS/2025/3D Interaction Geometric Pre-training for Molecular Relational Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..8edea0888de30d83669dcd4d8f91c5e9ad0538bc --- /dev/null +++ b/NeurIPS/2025/3D Interaction Geometric Pre-training for Molecular Relational Learning/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:733721a0c05e582e18e999a1514fb1fee57f86eeab41ae2b8eb8c4e6eeb13560 +size 1004206 diff --git a/NeurIPS/2025/3D Visual Illusion Depth Estimation/0ec266a2-d4ad-43f5-acb5-3c432ce7212a_content_list.json b/NeurIPS/2025/3D Visual Illusion Depth Estimation/0ec266a2-d4ad-43f5-acb5-3c432ce7212a_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..647f42b2edb735764657ca48ac0c6c3de3726df7 --- /dev/null +++ b/NeurIPS/2025/3D Visual Illusion Depth Estimation/0ec266a2-d4ad-43f5-acb5-3c432ce7212a_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5c048804fa98ecb74c8a36819f1e18f45c8b7985fa80ebdf76cff8a04dd74417 +size 254969 diff --git a/NeurIPS/2025/3D Visual Illusion Depth Estimation/0ec266a2-d4ad-43f5-acb5-3c432ce7212a_model.json b/NeurIPS/2025/3D Visual Illusion Depth Estimation/0ec266a2-d4ad-43f5-acb5-3c432ce7212a_model.json new file mode 100644 index 0000000000000000000000000000000000000000..539c002da08fa08f818a1dc80ace06c1cd443464 --- /dev/null +++ b/NeurIPS/2025/3D Visual Illusion Depth Estimation/0ec266a2-d4ad-43f5-acb5-3c432ce7212a_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d0a15f5cdb872280c93a07026995f0aaf7eb1d80b15543ffc6c1d4909eeab3c4 +size 323582 diff --git a/NeurIPS/2025/3D Visual Illusion Depth Estimation/0ec266a2-d4ad-43f5-acb5-3c432ce7212a_origin.pdf b/NeurIPS/2025/3D Visual Illusion Depth Estimation/0ec266a2-d4ad-43f5-acb5-3c432ce7212a_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..8ca879ebd43c84e804db94d1ead185e510abd512 --- /dev/null +++ b/NeurIPS/2025/3D Visual Illusion Depth Estimation/0ec266a2-d4ad-43f5-acb5-3c432ce7212a_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:990021398f74bbdd2db04b85ee029174897d605993d71bde2e6554baa29f0ada +size 14075398 diff --git a/NeurIPS/2025/3D Visual Illusion Depth Estimation/full.md b/NeurIPS/2025/3D Visual Illusion Depth Estimation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..c33fe2b49052769c2f1f8c0c0b2e5c2c2febeadb --- /dev/null +++ b/NeurIPS/2025/3D Visual Illusion Depth Estimation/full.md @@ -0,0 +1,1526 @@ +# 3D Visual Illusion Depth Estimation + +Chengtang Yao $^{1,2}$ , Zhidan Liu $^{1,2}$ , Jiaxi Zeng $^{1,2}$ , Lidong Yu $^{3,4}$ , Yuwei Wu $^{1,2}$ , Yunde Jia $^{2,1*}$ + +1Beijing Key Laboratory of Intelligent Information Technology, fComputer Science & Technology, Beijing Institute of Technology, China + +2Guangdong Provincial Key Laboratory of Machine Perception and Intelligent Computing, Shenzhen MSU-BIT University, Shenzhen, China + +$^{3}$ NVIDIA, $^{4}$ NEOLIX + +{zdliu, wuyuwei, jiayunde}@bit.edu.cn + +{yao.c.t.adam, yvlidong, jiaxizeng.jx}@gmail.com + +![](images/1a985fd5df46e45b1eff7f4d0791c3c209366e0f6a4d9dff5ca748c75c57c0cb.jpg) + +HuggingFace Dataset + +![](images/82ec6d9da1045c25a53dfb7709b655a38a43297fb1e883a07e356d06d31c56d1.jpg) + +GitHub Project + +![](images/193900ee26c69ac77ae0e07c7853fb6513ce3b92d414a328510a736850e5d384.jpg) + +HuggingFace Model + +![](images/9a572d56551cc83841dc30a0ac2e8ea2072e16827d566f31dd1f173a9a6824b7.jpg) + +# Abstract + +3D visual illusion is a perceptual phenomenon where a two-dimensional plane is manipulated to simulate three-dimensional spatial relationships, making a flat artwork or object look three-dimensional in the human visual system. In this paper, we reveal that the machine visual system is also seriously fooled by 3D visual illusions, including monocular and binocular depth estimation. In order to explore and analyze the impact of 3D visual illusion on depth estimation, we collect a large dataset containing almost 3k scenes and 200k images to train and evaluate SOTA monocular and binocular depth estimation methods. We also propose a 3D visual illusion depth estimation framework that uses common sense from the vision language model to adaptively fuse depth from binocular disparity and monocular depth. Experiments show that SOTA monocular, binocular, and multi-view depth estimation approaches are all fooled by various 3D visual illusions, while our method achieves SOTA performance. + +# 1 Introduction + +Depth estimation aims to recover the 3D geometry of a scene from a single image or an image sequence. It is a long-standing and challenging vision problem, with extensive research in monocular depth estimation [46, 47, 3, 17], stereo matching [27, 42, 5, 13], and multi-view reconstruction [41, 39]. These works have achieved impressive performance in typical, well-structured scenes, approaching human-level perception. However, beyond typical scenes, there exist many 3D visual illusion scenes that make a flat artwork or object look three-dimensional, as illustrated in Figure 1. These 3D visual illusions mislead the depth perception and seriously affect the downstream applications, causing safety-critical risks in AR/VR and robotics. + +In this paper, we present a 3D-Visual-Illusion dataset to investigate the impact of 3D visual illusions on depth estimation. The dataset includes five types of illusions: inpainting illusion (e.g., inpainting on walls or floors), picture illusion (e.g., image printed/drawn on a paper), replay illusion (e.g., videos replayed on different screens), holography illusion, and mirror illusion (e.g., specular and transparent surfaces). It comprises nearly 3,000 scenes and 200,000 images, covering various environments from small objects to large scenes and from indoor to outdoor settings. We construct the dataset from both virtual and real-world data. Virtual data is generated using two separate pipelines: one based on web-sourced videos and the other on generative models. Real-world data is captured using a stereo camera and a solid-state LiDAR depth sensor. + +![](images/ec4474d90ebed4e87b4ccb62311a6888284c22314e48e9aa06a1d864a260fbd4.jpg) +Figure 1: The visualization of 3D visual illusions. + +The evaluation results on the 3D-Visual-Illusion dataset reveal distinct failure modes for different SOTA depth estimation models. Monocular methods, which rely on the mapping from texture cues to 3D geometry, are easily misled by illusion patterns such as printed images or screen content. In contrast, stereo methods depend on pixel correspondences and fail on transparent or reflective surfaces like glass and mirrors, where conflicting signals distort the matching process. Notably, stereo and monocular methods exhibit complementary strengths, often succeeding where the other fails. Stereo methods succeed on texture-rich illusions, while monocular models can recover mirror geometry through learned priors. Yet, each alone is insufficient to handle the full spectrum of 3D visual illusions, and this complementarity motivates us to seek a unified framework that leverages the strengths of both. + +Inspired by the strong generalization capability of vision-language models (VLMs) on mirror illusions (see the supplementary materials for details), we propose a VLM-driven monocular-stereo fusion framework to fuse stereo and monocular priors. The model leverages commonsense knowledge from VLMs to assess the reliability of monocular and binocular depth across different regions, enabling more effective depth fusion. Our model consists of two components: a dual-branch prediction network and a VLM-based fusion network. The dual-branch network takes a rectified image pair as input and simultaneously predicts monocular depth and binocular disparity. The VLM-based fusion network employs a pre-trained vision model to extract features from the left RGB image. These features are mapped into a shared embedding space using a large language model conditioned on a language prompt. The embedding features are then used to generate a confidence map via flow matching. The confidence map is used to align the affine-invariant monocular depth to metric scale, which is then fused with the binocular disparity to produce the final depth map. Experiments on our dataset and the Booster dataset demonstrate that our method achieves SOTA performance under a wide range of 3D visual illusions. + +# 2 Related Work + +# 2.1 Stere Matching + +Stereo matching is a pixel-wise labeling task that relies on dense correspondence between a pair of images. The SOTA methods are either GRU-based iterative methods or Transformer-based methods. The former methods predict the disparity update and iteratively approximate the GT value in a GRU framework [27, 24, 16, 48, 5, 10]. They have achieved great performance in both benchmark and zero-shot generalization testing. The latter methods use a Transformer to learn matching and predict the disparity map [25, 12, 43, 44, 41, 23, 39]. The Transformer-based methods achieve superior performance by learning from large-scale data. In this paper, we collect a comprehensive, large-scale dataset to thoroughly investigate and evaluate the impact of 3D visual illusions on matching methods. + +We reveal that stereo matching methods are highly susceptible to various illusions. Our method leverages common sense from VLM to detect mirror illusions and help rectify these illusions. + +# 2.2 Monocular Depth Estimation + +Monocular depth estimation is a pixel-wise regression task based on a single image. Recent deep learning methods leverage diffusion models [17, 8, 50] or Transformers [33, 2, 47, 45] to extract depth-related features, in both supervised and self-supervised settings [9, 11, 19]. Despite their impressive generalization performance across diverse scenes, these methods fundamentally rely on monocular cues, which are susceptible to 3D visual illusions, much like the human visual system. In this work, we introduce a large-scale benchmark to evaluate state-of-the-art monocular depth models under such illusions. Our results show that existing methods are consistently misled by these challenging patterns. To address this, we propose leveraging matching-based depth cues as complementary information to enhance monocular depth estimation. + +# 2.3 Large Vision-Language Model + +The large vision-language model (VLM) injects common sense from billions of textual data to support vision understanding and generation [49]. It presents great power in various tasks, like visual question answering, image generation, and navigation. The methods of these tasks mainly adapt a pre-trained VLM to specific datasets to preserve the generalization ability, while promoting the understanding of specific tasks. To further facilitate the training of VLM in downstream tasks, a lot of methods explore different finetuning strategies, like Prompt [4], Adapter [14], LoRA [15], and LST [35]. In this paper, inspired by the strong detection ability of VLM on mirror illusions, we use VLM to predict the confidence of the disparity map to recover the metric version of monocular depth. The common sense from large VLM is beneficial for the confidence estimation in various complex scenes. + +# 3 3D Visual Illusion Dataset + +We construct the 3D-Visual-Illusion dataset to investigate the challenges posed by 3D visual illusions in depth estimation. The dataset comprises nearly 3,000 scenarios, with over 200,000 frames for training and 617 frames for testing. It includes images of various resolutions, up to a maximum of $1080 \times 1920$ , and spans a wide range of scenes, from indoor environments and small objects to large-scale street views. The dataset covers five types of illusions: inpainting illusion (e.g., inpainting on a wall/floor), picture illusion (e.g., picture printed/drawn on a paper), replay illusion (e.g., video replayed on a different screen), holography illusion, and mirror illusion (e.g., specular or transparent surfaces). Data is collected from both virtual and real-world sources. Details of the construction process for the virtual and real subsets are provided in the following sections. + +# 3.1 Virtual Data + +We collect a large amount of video data from websites and text-to-video generative models. We take the videos as left image sequences and generate disparity maps and right images. + +Video Collection We adopt two distinct data collection strategies for the first four types of illusions and for mirror illusions. For inpainting, picture, replay, and holography illusions, we crawl 5,226 web videos (over 52M frames) using keyword-based search. We then apply a vision-language model, Qwen2-VL-72B [1, 40], to automatically filter out irrelevant frames, reducing the dataset to 4,519 videos (1.4M frames). Further manual filtering removes blurry or occluded frames, resulting in 1,384 high-quality videos with 236K frames. Mirror illusions are difficult to collect from the web due to the rarity of mirror-related keywords and high-quality videos. To address this, we generate videos using SOTA generative models, including Sora [30], Kling [21], and HunyuanVideo [20]. Prompts are initially created with ChatGPT and manually refined. Videos violating physical plausibility are discarded. In total, we collect 234 high-quality mirror illusion videos comprising 2,382 frames. + +Depth Generation After collecting videos from both web sources and generative models, we generate depth for each frame using the pipelines illustrated in Figure 2 and 3. Different pipelines are used for the two data sources for the following reasons: (1) Web-sourced videos typically involve fixed cameras, providing limited viewpoints and making accurate scene reconstruction difficult. (2) Generative videos are used primarily for mirror illusions, which require modeling the geometry of the + +![](images/1b17a5a960dcb3e639839059af36f41159bd4e554bbc9c830680eb360a46cab9.jpg) +Figure 2: The data generation pipeline for web-sourced data. + +![](images/63f7d2ca04d707132339d5c6a12800a9fb4df06594c87c0c72966652eec7ba74.jpg) +Figure 3: The data generation pipeline for videos from generative models. + +reflected (mirror) world to generate right-view images. However, monocular depth estimation may ambiguously predict either the mirror surface or the reflected scene, leading to inconsistent results. + +For web videos, we use the pre-trained DepthAnything V2 [46] to predict inverse depth, which is treated as disparity under unknown camera parameters. However, in regions affected by 3D visual illusions, the predicted disparity is often severely inaccurate. To correct this, we introduce a neighboring support region, assuming it lies on the same plane as the illusion region, and use it as a reference for disparity correction. Segmentation masks for both illusion and support regions are obtained using SAM2 [34]. Since the automatic mode struggles to detect illusions accurately, we manually annotate all frames using SAM2's click mode, removing redundant or imperceptible cases. + +After obtaining the mask of illusion and support regions, we fit a plane using the points within a support region. In the standard 3D camera coordinate system $(X,Y,Z)$ , the general form of a plane with parameters $(\alpha,\beta,\gamma,\delta)$ is $\alpha \cdot x + \beta \cdot y + \gamma \cdot z + \delta = 0$ . Given the relationship between image coordinate $(u,v)$ , disparity $d$ , and $(X,Y,Z)$ : + +$$ +(u, v, d) = \frac {1}{z} (x, y, B) \cdot \left(f _ {x}, f _ {y}, \frac {f _ {x} + f _ {y}}{2}\right) + \left(c _ {x}, c _ {y}, 0\right), +$$ + +the planar structures in 3D space $(X,Y,Z)$ remain planar in the disparity space $(u,v,d)$ with plane parameters $(\alpha, \beta, \delta, \gamma)$ : $\alpha \cdot u + \beta \cdot v + \delta \cdot d + \gamma = 0$ . This property is crucial for our generation, as it allows plane fitting directly in disparity space under unknown camera intrinsics and baseline, avoiding the need to convert disparity into depth. Given a set of $N$ points from the support region, $\{(u_i, v_i, d_i)\}_{i=1}^N$ , the goal of plane fitting becomes a least squares fitting problem over parameters $(\alpha, \beta, \delta, \gamma)$ . To mitigate the impact of noise during fitting, we adopt RANSAC [7] for robust plane estimation (see supplemental materials for details). The fitted plane is then used to rectify disparity values within the illusion regions. After obtaining the rectified disparity map, we apply an additional denoising step to ensure smooth transitions along the boundaries between support regions and their surroundings. + +For videos from generative models, we reconstruct the entire scene using InstantSplat [6], which first estimates geometry via DUSt3R [41] and synthesizes novel views using Gaussian Splatting (GS) [18]. + +We extract the disparity map from the geometry derived from DUSt3R and refine it through a series of post-processing steps, including semantic segmentation, RANSAC-based plane fitting, and disparity denoising. + +Right Image Generation The right-view images for generative-model videos are directly rendered using Gaussian Splatting (GS). For web-sourced videos, right views are generated by warping the left images using monocular disparity. Due to scale ambiguity, we estimate an optimal scale factor $s$ via binary search, terminating when most warped pixels fall within the valid image width: + +$$ +\tilde {s} = \arg \min _ {s} \left| \frac {\sum_ {x} \mathbb {1} (0 \leq u - s \cdot d _ {(u , v)} < W)}{N} - \tau \right|, \tag {1} +$$ + +where $\mathbb{1}(\cdot)$ is the indicator function, $d$ is the disparity, $(u,v)$ are pixel coordinates, $W$ is the image width, $N$ is the number of pixels, and $\tau$ is the target ratio of valid pixels. We use the scaled disparity $\tilde{d} = \tilde{s}\cdot d$ to warp the left image accordingly. In cases where multiple source pixels are warped to the same target location due to occlusions, we retain the one with the largest disparity to maintain consistency: + +$$ +I _ {r} (u ^ {\prime}, v) = I _ {l} (u ^ {*}, v), +$$ + +$$ +u ^ {*} = \arg \max _ {u} \left\{d _ {(u, v)} \mid u - \tilde {d} _ {(u, v)} = u ^ {\prime} \right\}. \tag {2} +$$ + +To address holes after warping, we apply an image inpainting method [36] to produce visually complete right-view images. The full generation algorithm is detailed in the supplemental materials. + +# 3.2 Real-world Data + +In addition to virtual data, we collect real-world data comprising 72 scenes and 617 frames. The setup includes a stereo camera (ZED Mini) and a LiDAR-based depth sensor (Realsense L515), with details provided in the supplemental materials. To ensure accurate alignment, the two sensors are rigidly mounted, and their relative pose is calibrated using a checkerboard. The L515 depth map is then warped to the ZED left camera frame based on the calibration to construct the ground-truth depth. + +Due to the lower resolution of the L515, direct pixel-wise warping to the higher-resolution ZED image will result in sparse depth maps and incorrect projections, particularly in occluded regions where background depths may overwrite foreground pixels. To address this, we first densify the L515 point cloud by upsampling its depth map $\mathcal{Z}_L$ via nearest-neighbor interpolation and proportionally scaling its intrinsic matrix $K_{L}$ . + +After densifying the point cloud, image coordinates from the L515, $(U_L,V_L)$ , are first projected into the 3D camera coordinate $(X_L,Y_L,Z_L)$ using the L515 depth map. These 3D points are then transformed to the ZED left camera's coordinate system $(X_Z,Y_Z,Z_Z)$ . Finally, they are projected onto the ZED left image coordinates $(U_Z,V_Z)$ : + +$$ +\left[ X _ {L}, Y _ {L}, Z _ {L} \right] = Z _ {L} \cdot K _ {L} ^ {- 1} \cdot \left[ U _ {L}, V _ {L}, 1 \right], +$$ + +$$ +\left[ X _ {Z}, Y _ {Z}, Z _ {Z} \right] = R \cdot \left[ X _ {L}, Y _ {L}, Z _ {L} \right] + T, \tag {3} +$$ + +$$ +[ U _ {Z}, V _ {Z}, 1 ] = K _ {Z} \cdot [ X _ {Z} / Z _ {Z}, Y _ {Z} / Z _ {Z}, 1 ]. +$$ + +Here, $R \in \mathbb{R}^{3 \times 3}$ and $T \in \mathbb{R}^{3 \times 1}$ denote the rotation and translation matrices between the L515 and the ZED cameras, both obtained via calibration. $K_{L} \in \mathbb{R}^{3 \times 3}$ and $K_{Z} \in \mathbb{R}^{3 \times 3}$ represent the intrinsic matrices of the L515 and the ZED left camera, respectively. + +In the projected coordinates $(U_Z, V_Z)$ , multiple 3D points $P_m$ may map to the same pixel due to slanted surfaces or occlusions. To resolve this, we apply Z-buffering to retain the point with the minimum depth for the ZED depth map $\mathcal{Z}_Z$ : + +$$ +\mathcal {Z} _ {Z} (u _ {Z}, v _ {Z}) = z _ {Z} ^ {*}, +$$ + +$$ +z _ {Z} ^ {*} = \min _ {\left(u _ {Z} ^ {\prime}, v _ {Z} ^ {\prime}, z _ {Z} ^ {\prime}\right) \in P _ {m}} \left\{Z _ {Z} ^ {\prime} \mid \left(u _ {Z} ^ {\prime}, v _ {Z} ^ {\prime}\right) = \left(u _ {Z}, v _ {Z}\right) \right\}. \tag {4} +$$ + +Although upsampling greatly densifies the point cloud, projecting from a lower-resolution to a higher-resolution space may still introduce small holes. To address this, we apply connected component + +![](images/5b2009c061598fcf52d5610d7153ebfeee31214fcc4b3485057a863924db891d.jpg) +Figure 4: The pipeline of VLM-driven binocular and monocular disparity fusion model. + +analysis to identify missing regions and fill them using image inpainting [38]. The inpainted depth values are further smoothed to ensure seamless transitions with surrounding areas. + +Furthermore, to identify unreliable depth values, the depth map projected onto the ZED image is reprojected back to the L515 coordinate: + +$$ +\begin{array}{l} \left[ X _ {Z \rightarrow L}, Y _ {Z \rightarrow L}, Z _ {Z \rightarrow L} \right] = R ^ {- 1} \cdot \left(Z _ {Z} \cdot K _ {Z} ^ {- 1} \cdot \left[ U _ {Z}, V _ {Z}, 1 \right] - T\right), \\ [ U _ {Z \rightarrow L}, V _ {Z \rightarrow L}, 1 ] = \frac {1}{Z _ {Z \rightarrow L}} \cdot K _ {L} [ X _ {Z \rightarrow L}, Y _ {Z \rightarrow L}, Z _ {Z \rightarrow L} ]. \tag {5} \\ \end{array} +$$ + +The reprojected pixels that correspond to invalid depth values or exhibit large depth differences with their ZED counterparts are marked as unreliable: + +$$ +\mathcal {Z} _ {Z} \left(u _ {Z}, v _ {Z}\right) = \left\{\begin{array}{l l}0&\text {i f} \mathcal {Z} _ {L} \left(u _ {Z \rightarrow L}, v _ {Z \rightarrow L}\right) = = 0 o r\\&\left| z _ {Z \rightarrow L} - \mathcal {Z} _ {L} \left(u _ {Z \rightarrow L}, v _ {Z \rightarrow L}\right)\right| > \epsilon ,\\\mathcal {Z} _ {z} \left(u _ {z}, v _ {z}\right)&\text {o t h e r w i s e .}\end{array}\right. \tag {6} +$$ + +Here, $\epsilon$ is a manually defined threshold. To further refine depth quality, median filtering is applied to suppress noise and remove outlier points from the point cloud. Finally, the depth map $\mathcal{Z}_Z$ is converted into disparity map $D$ : + +$$ +D = B \cdot F / \mathcal {Z} _ {Z}, \tag {7} +$$ + +where $B$ denotes the baseline between the ZED's stereo cameras, and $F$ is the focal length of the ZED camera. The entire algorithm is detailed in the supplemental materials. + +# 4 VLM-Driven Monocular-Stereo Fusion Model + +As illustrated in Figure 4, our model first uses a dual-branch prediction network to predict the binocular disparity and the monocular depth. Then, a VLM-based fusion network is used to produce the final disparity map by fusing the binocular disparity and the monocular depth. + +# 4.1 Dual-Branch Prediction Network + +The dual-branch prediction network comprises a binocular disparity estimation branch and a monocular depth estimation branch. The stereo branch adopts an iterative optimization framework, extracting features from rectified image pairs and constructing a cost volume via dot product. Starting from an initial disparity of zero, a GRU-based module iteratively refines the disparity map. The monocular branch takes the left image as input and uses the frozen DepthAnything V2 [46] to predict affine-invariant inverse depth, which is disparity under unknown camera parameters. We also extract features from frozen DepthAnything V2, followed by learnable convolutions. These adapted features serve as left-view context to guide disparity refinement in the stereo branch. + +# 4.2 VLM-Based Fusion Network + +As shown in Figure 4, the VLM-based fusion network contains three main stages: the VLM prediction stage, the confidence map generation stage, and the global fusion stage. + +VLM Prediction Stage In this stage, we utilize the pretrained QwenVL2-7B model [40, 1]. The visual prompt comprises the left image, binocular disparity map, and monocular disparity map. Since the textures that mislead monocular depth estimation are often too complex and diverse to be explicitly described, the language prompt is instead designed from the perspective of materials that typically confuse stereo matching. By leveraging the general reasoning capabilities of the vision-language model, we extract embedding features that help assess the relative reliability of monocular and binocular depth cues under language guidance. + +Confidence Map Generation Stage This stage aims to transform the embedding features back into image space to generate a confidence map. Inspired by the flow-matching framework Flux [26, 31], we learn a guided path flow from Gaussian noise to a complex confidence distribution: + +$$ +y _ {k _ {i + 1}} = y _ {k _ {i}} + \Delta K \cdot v _ {k _ {i}} \left(y _ {k _ {i}}, c _ {e}\right), \tag {8} +$$ + +where $k$ is uniformly sampled in the interval [0, 1], $\Delta K = 1 / K$ , and $K$ is the total number of steps. $y_{0}$ is sampled from a prior Gaussian distribution, $y_{k_i}$ is the intermediate state at $i$ -th sampling, $c_e$ is the conditional embedding, and $v_{t_k}$ denotes the predicted velocity field at time $t_k$ under conditions $(y_{t_k}, c_e)$ . + +Specifically, the language prompts are first mapped into the embedding space. The image and text embeddings are concatenated to form $c_{e}$ , which is combined with the intermediate state $y_{t_k}$ and passed through a stack of Transformers. Cross-attention is used to inject conditional information, and the velocity field $v_{t_k}$ is predicted. After multiple iterations via Equation 8, the final state $y_{1}$ is reshaped into a 2D format and decoded into image space using a variational autoencoder. The resulting feature is then concatenated with the cost volume and passed through convolution layers to predict the confidence map $I_{c}$ : + +$$ +I _ {c} = \sigma \left(\mathcal {F} _ {c} \left(\left[ \mathcal {G} _ {c} \left(y _ {1} ^ {\prime}\right), V (D _ {s}) \right]\right)\right), \tag {9} +$$ + +where $\sigma$ is the sigmoid function, $\mathcal{F}_c$ denotes convolution, $\mathcal{G}_c$ is the VAE decoder, $y_1'$ is the reshaped $y_1$ , and $V(D_s)$ is the cost volume sampled around binocular disparity $D_s$ . + +Global Fusion Stage Global fusion first aligns monocular disparity $D_{M}$ to the absolute/metric disparity space and then fuses the aligned monocular disparity $\tilde{D}_{m}$ with binocular disparity $D_{s}$ . The alignment is achieved by affine transformation parameters $s_{m}, t_{m}$ : + +$$ +\tilde {D} _ {m} = s _ {m} \cdot D _ {m} + t _ {m}, +$$ + +$$ +s _ {m}, t _ {m} = \arg \min _ {s _ {m}, t _ {m}} \sum_ {(u, v)} \left(s _ {m} \cdot D _ {m} (u, v) + t _ {m} - D _ {s} (u, v)\right) ^ {2}. \tag {10} +$$ + +$s_m$ and $t_m$ are learned through convolutions on the concatenation of the monocular disparity $D_m$ and the binocular disparity $D_s$ . Since $s_m$ and $t_m$ on low-confident regions are unreliable, we further refine the parameters by pooling $s_m$ and $t_m$ of the high-confident neighbors. After acquiring refined parameters $s_m$ and $t_m$ , we compute the aligned monocular disparity $\tilde{D}_m$ using Equation 10. $\tilde{D}_m$ , $D_s$ , and $I_c$ are then concatenated and passed through convolutions and upsampling layers to generate the final high-resolution disparity map. + +# 5 Experiment + +We first pre-train the models on the SceneFlow dataset [28], and then fine-tune them on the virtual 3D-Visual-Illusion data. The fine-tuned models are evaluated on both the real-world 3D-Visual-Illusion data and the Booster training set [32]. We compare our method with monocular depth estimation approaches (DepthAnything V2 [46], Marigold [17], Metric3D [47], and DepthPro [3]), multi-view foundation models (Dust3R [41] and VGGT [39]), and stereo matching methods (RAFT-Stereo [27], Selective-IGEV [42], and MochaStereo [5]). For additional details on training procedures, loss functions, evaluation metrics, and prompt designs, please refer to the supplementary material. + +# 5.1 Influence of 3D Visual Illusions + +Table 1 presents a comprehensive comparison of the real-world data in the 3D-Visual-Illusion dataset. The real-world data is mainly composed of inpainting, picture, and replay illusions. The results of all compared methods are obtained from official code and weights, where the stereo methods use model weights pretrained on the SceneFlow dataset. + +Table 1: Evaluation results on illusion regions for real-world data of the 3D-Visual-Illusion dataset. align means alignment using globally shared affine parameters computed from ground truth. + +
MethodFinetuneDisparity SpaceDepth Space
EPE ↓bad2 ↓bad3 ↓bad5 ↓AbsRel ↓RMSE ↓δ1 ↑
DA V2 [46]×5.8161.4543.1830.570.140.1592.86
Metric3D [47]×12.4694.1191.1482.050.340.2948.97
DA V2 metric [46]×16.2492.5387.4375.250.520.3948.75
DepthPro [3]×12.2687.0880.6062.430.280.2565.92
Marigold [17]×21.1665.6759.6753.190.450.3763.65
DA V2 metric [46] + align×5.2356.8245.5028.890.170.1593.70
Metric3D [47] + align×5.7066.2650.9240.430.170.1794.80
DepthPro [3] + align×4.3644.9834.9824.700.090.1093.83
Dust3R [41]×6.7452.8945.3136.610.250.2287.09
VGGT [39]×6.1653.3244.8937.200.130.1278.46
RAFT-Stereo [27]×1.6224.3213.202.970.040.0699.18
Selective-RAFT [42]×1.5823.4612.652.570.030.0799.60
Selective-IGEV [42]×1.6724.0613.112.990.040.1099.26
MochaStereo [5]×1.7525.4914.113.540.040.1198.76
StereoAnything [13]×2.4129.0016.156.540.110.3296.23
ours✓1.7726.7215.733.600.030.0899.60
+ +Table 2: Zero-shot generalization on the balanced set of Booster dataset with quarter resolution. All: all regions, Trans: transparent regions, NonTrans: nontransparent regions. + +
MethodFinetuneAllTransNonTrans
EPE ↓bad2 ↓bad3 ↓bad5 ↓EPE ↓bad2 ↓bad3 ↓bad5 ↓EPE ↓bad2 ↓bad3 ↓bad5 ↓
DA V2 [46]×3.1648.8336.9821.227.1977.9869.0250.912.9147.2535.0919.25
Metric3D [47]×35.5599.7099.3297.7341.5599.3798.9397.9134.8999.7199.3297.59
DA V2 metric [46]×21.5594.2891.3784.2128.4293.0490.7286.7820.9494.2591.2283.84
DepthPro [3]×24.4492.9890.2384.2525.6592.9888.9083.0524.1492.9190.1684.19
Marigold [17]×5.9957.9047.1332.638.4676.3365.9051.525.7256.7945.8731.26
DA V2 metric + align×5.7162.7048.9432.1812.7277.2468.4654.705.4562.0548.1731.36
Metric3D + align×3.0943.0529.6516.858.7276.8764.6847.622.7641.2827.9115.22
DepthPro + align×4.0253.7640.3023.816.0264.2555.1940.393.9653.2539.6123.12
Dust3R [41]×3.7048.5734.1619.538.6973.4064.0350.183.3447.2132.5617.93
VGGT [39]×3.7034.0523.4414.5810.7872.2265.1255.833.3232.2721.3412.24
RAFT-Stereo [27]×4.0817.6114.8712.179.5567.8459.4347.463.1313.1010.708.63
Selective-RAFT [42]×4.0519.4816.6413.5710.0870.0261.7949.642.9214.9412.389.93
Selective-IGEV [42]×4.5219.2316.5113.849.2267.0058.9947.213.5214.6912.2810.20
MochaStereo [5]×3.7916.7714.2411.779.1866.6458.1045.782.8212.2510.118.30
StereoAnything [13]×4.3624.1319.2014.5010.5473.4863.5349.373.2920.0915.3711.08
ours✓2.4313.849.986.917.3256.7747.8336.451.7610.066.544.08
+ +In Table 1, we can observe that monocular depth estimation methods struggle notably with inpainting, replay, and picture illusions, leading to large errors in both disparity and depth spaces. Even when ground-truth alignment is used to convert relative depth to absolute scale, their performance remains significantly worse than stereo methods and ours. As for recent foundation models, such as Dust3R [41] and VGGT [39], they exhibit a strong monocular bias in illusion-rich scenes. In comparison with stereo matching methods depending on explicit correspondence, our method achieves comparable results, indicating that it preserves strong matching constraints. Qualitative results in Figure 5 further illustrate the serious influence of + +Table 3: Results of Marigold with/without finetuning on 3D-Visual-Illusion dataset. + +
RegionFinetuneEPE↓bad2↓bad3↓AbsRel↓δ1↑
Illusion×21.1665.6759.670.4563.65
Illusion✓13.6774.8255.200.2871.04
Non-illusion×7.6149.1839.560.1879.76
Non-illusion✓7.1055.6344.090.1677.44
+ +Table 4: Results of stereo methods on Booster dataset with/without finetuning on 3D-Visual-Illusion data. + +
MethodFinetuneTransNonTrans
EPE↓bad2↓bad3↓EPE↓bad2↓bad3↓
RAFT-Stereo [27]×9.5567.8459.433.2313.1310.75
RAFT-Stereo [27]Sparse15.3680.3472.347.1227.4724.01
RAFT-Stereo [27]Dense9.2474.1060.6717.3922.4818.96
Selective-IGEV [42]×9.5066.8558.903.6014.7412.34
Selective-IGEV [42]Sparse9.4264.0654.215.9714.6312.46
Selective-IGEV [42]Dense10.3969.6559.405.3219.0015.64
+ +Table 5: Zero-shot generalization on Middlebury dataset at half resolution. Evaluation is conducted in metric disparity space over the entire image, without restricting maximum disparity. + +
MetricSelective-RAFTSelective-IGEVMochaStereoStereoAnythingRAFT-StereoOurs
EPE ↓2.342.592.662.891.921.50
Bad-2 ↓12.0411.7910.1811.9312.6011.79
+ +![](images/fd4af111c1e46620abd93a1b70f433fc8b772bd8895f8fc704301d34034c09b8.jpg) + +![](images/bc098226d5fdb4ff31fd6a3f78a0865e51396f9a588d6b172f0b6a99c57e0961.jpg) +Right Image +Figure 5: Visualization on our dataset. + +![](images/b34c4427bb98ed8b3ce5033192259b8633aca2056ac7a8847f2ee648101ba841.jpg) + +![](images/d7f565f4905e01648bcc52c0742fe1c4ebd67fa3923fbf05a36a7f3cb964704b.jpg) +Ours + +![](images/ba422ec35ca81623cc4bfc24a0d27e5a56a1ae602a5e5d51cdffd1d4a5bcdece.jpg) + +![](images/f86484323c1d5d63593c0cc62afdc43be3007c46a8650bc5b3f49322e548209c.jpg) +Dust3R + +![](images/6820065e919f0079b6874d41dc55caf9be76a44edeba5784d29c9071e6111cc5.jpg) + +![](images/ceda005b17225807238362d64dd9f442ff45bbcaf22f0547a890073fb275b584.jpg) +VGGT + +![](images/e1e40d795df080669d4f0cd2868742845bc88e76057138d43d700b7cf0628774.jpg) + +![](images/4ccfcc7f842a1c3c605f5542d23cbf6d37f531adfa058bdbf43860c453710cfd.jpg) +Metric3D + +![](images/20702b11afc1659439a8303dd9e957389733c7c541df732b2ebfdfe3dfb1ca1a.jpg) +Left Image + +![](images/a2241acfa5a40c6be9c232c6c8b458f5808e46c62d4dbf47972ea7959d245af8.jpg) +GT + +![](images/a5ca343554661b0039f57573159f7706c5c9b7401bf4cfe4136a0acc16ba3649.jpg) +Mocha + +![](images/f0ce96fea19091eafb97d95a09ff731055ddb2642b87659f956f5c8ad3b352e5.jpg) +RAFTStereo + +![](images/704dac191c86e396ad7b7ff697c2538357af50637b2a182282be3bc0afa55423.jpg) +Right Image +Figure 6: Visualization on Booster dataset. + +![](images/49aa75e8ff318ec5c1202a3e5e8fd2a2919c918163519b9f4cf12d33107bb7f5.jpg) +Ours + +![](images/fe2e449851d650f28f8cee84dd8ed5d1c8025fe89560817254c31be2db240689.jpg) +Selective-RAFT + +![](images/518921c1e5bd2d1a6a9e88a5d2e594092a37564d7b4f610b8e5d95d95434a569.jpg) +Selective-IGEV + +3d visual illusions on SOTA depth estimation methods. More visualization, please refer to our supplemental materials. + +The Booster dataset [32] includes many objects with specular and transparent surfaces, making it well-suited for evaluating generalization to mirror illusions. As shown in Table 2, monocular methods achieve the best performance, especially DepthPro. In contrast, binocular methods are easily fooled by mirrors and specular objects. Qualitative results in Figure 6 further show that mirror illusions pose a serious challenge to SOTA binocular methods. See supplemental materials for more visualizations. + +Here, we also compare classical binocular and monocular methods with/without finetuning on 3D-Visual-Illusion data. Table 3 shows Marigold's performance on illusion and non-Illusion regions of real-world 3D-Visual-Illusion data. We omit DepthAnything V2 results, as its official code fails to converge on the virtual data. This is because its official implementation trains on metric depth, which is sensitive to scale variations, while our virtual data varies significantly in scale. In contrast, the official code of Marigold supports training on affine-invariant depth, ensuring stable learning. Table 3 shows that fine-tuning improves most metrics on illusion regions, suggesting Marigold adjusts its predictions toward planar surfaces, but at the cost of degraded performance on non-Illusion regions. This supports our hypothesis that monocular depth models rely on fixed mappings from texture cues (e.g., shape, shadow, perspective, defocus), making it difficult to distinguish illusion textures from real object textures. The worsened bad2 metric in illusion regions indicates incomplete overfitting, while the limited EPE gain in non-Illusion regions likely stems from correcting a few extreme outliers rather than consistent improvement. + +Table 4 presents the performance of SOTA stereo models with and without finetuning on the 3D-Visual-Illusion dataset. The Sparse and Dense represent different augmentation strategies used during training. The results show that all stereo models achieve only limited improvement on transparent regions after finetuning, while suffering a significant performance drop on non-transparent regions. This indicates that standard stereo architectures cannot effectively learn from such data. When finetuning on our virtual illusion data, different illusion types are inherently conflicting: mirror illusions rely on spatial context (i.e., monocular priors) for accurate depth estimation, whereas inpainting, picture, replay, and holography illusions deliberately mislead models by distorting these priors. Thus, features learned from mirror illusions are compromised when the model is trained on other illusion types, leading to conflicting learning of monocular priors. Moreover, since we assume a flat plane to rectify disparity during the generation of virtual illusion data, the finetuned stereo models tend to produce overly flat disparities. This results in a slight improvement on transparent glass regions but severe degradation on other non-transparent and non-planar objects. + +In addition to illusion scenes, we also present the performance of our model in the Middlebury dataset. We compare our model with several SOTA stereo-based approaches using metric disparity space over the entire image, without restricting the maximum disparity range. Table 5 shows that our model does not degrade in performance on these mundane scenes. On the contrary, it achieves improvements, particularly in terms of EPE. + +# 5.2 Ablation Study + +We conduct ablation studies on the Booster dataset to evaluate the contribution of each module and various fusion strategies. As shown in Table 6, the baseline stereo-only model performs poorly, indicating that binocular cues alone are insufficient in regions with challenging materials. Introducing + +Table 6: Ablation study on the Booster dataset. MF: Monocular Feature, PF: Post Fusion, APF: Adaptive Post Fusion, SF: Stereo Fusion, VLM: Vision-Language Model. + +
MFPFAPFSFVLMEPE ↓bad2 ↓bad3 ↓bad4 ↓bad5 ↓bad6 ↓bad7 ↓
15.1180.3872.3566.0661.3257.0452.97
✓8.3669.8961.0153.5047.4742.1637.43
✓✓9.2568.4659.0351.4845.8640.2935.60
✓✓9.5972.7761.9552.9046.3140.2835.12
✓✓10.4081.9467.5757.8250.1744.8139.69
✓✓7.3256.7747.8341.4836.4532.2828.75
+ +monocular depth through simple post fusion (PF, where fusion is guided by confidence generated from image features) significantly improves generalization. Incorporating monocular features into the stereo branch $(\mathrm{MF} + \mathrm{PF})$ further improves performance on the bad metrics, although it slightly degrades the EPE error rate, which indicates better overall geometry but more severe outlier shifts. Adaptive post fusion (APF) employs two independent GRUs to iteratively update the monocular and binocular disparity during fusion, where each other's disparity is used as update guidance. Although this strategy can bring some improvements, it may introduce noise due to inconsistent updates between the two branches. SF uses a single GRU to fuse binocular disparity with fixed monocular disparity, which makes performance worse, highlighting the risks of naively reusing uncertain priors. Finally, our full VLM-based fusion approach achieves the best performance, improving the bad2 metric by over 10 points. This demonstrates the strong reasoning capability of the VLM in handling visually ambiguous regions and its effectiveness in guiding reliable depth fusion. We also evaluate the VLM's confidence by comparing the predicted confidence maps with the disparity error maps, as defined in Equation 14. This analysis is conducted on the Booster dataset under a zero-shot generalization setting. The results show that the error rate of our confidence estimation is approximately $20\%$ , demonstrating a strong generalization ability, even in previously unseen illusory scenes. + +# 5.3 Discussion + +Illusion Effect on Different Depth Paradigms: (1) Monocular estimation relies on texture-based cues (e.g., shape, perspective, shadow, defocus) learned from RGB image. When these cues are artificially simulated on flat surfaces ( inpainting, pictures, replays, holograms), the model is easily misled, producing incorrect depths. Mirror illusions, however, can often be resolved through scene-level context learned from large-scale data. (2) Stereo estimation instead depends on pixel-wise correspondence. In mirror scenes, reflections overlap with real surfaces, causing ambiguous matches and depth errors. For inpainting, pictures, replays, and holography, stereo matching remains effective due to cross-view texture consistency. + +Limitations and Future Work: (1) The virtual data generation pipeline relies on manual semantic segmentation, which is labor-intensive and time-consuming. Given the lack of 3D visual illusion data, and the fact that existing detectors, segmentation models, and VLMs are often fooled, even humans will be fooled in complex cases, manual collection remains a practical step at this stage. Developing an automatic pipeline is an important future direction. (2) The real-world subset currently covers only a limited range of illusions ( inpainting, picture, and replay). We plan to extend it to broader types and more diverse real-world scenes. (3) The VLM-driven fusion is effective but computationally expensive. Designing lighter, more efficient fusion methods is worth further exploration. (4) Our study focuses on pure illusions without compositing. Future challenges include combinations of multiple illusions, semantically ambiguous objects, and entirely novel illusion types. + +# 6 Conclusion + +In this paper, we introduce the 3D-Visual-Illusion dataset, a large-scale benchmark for evaluating the depth estimation models under 3D visual illusions. The dataset covers diverse illusion types and scene categories, including indoor and outdoor settings, as well as both virtual and real-world data. Our experiments show that state-of-the-art models are easily fooled by various illusions, each exhibiting distinct failure modes. Monocular methods act as generative models, mapping texture cues to 3D geometry, and thus can be deceived by carefully simulated textures. In contrast, stereo methods serve as discriminative models that rely on pixel-wise correspondence, which breaks down when multiple objects project onto the same pixels. Finally, we introduce a VLM-driven monocular-stereo fusion model, which leverages commonsense reasoning from a vision-language model to assess cue reliability and achieve more robust depth estimation. + +Acknowledgment This work was supported by the Shenzhen Science and Technology Program under Grant No. JCYJ20241202130548062, the Natural Science Foundation of Shenzhen under Grant No. JCYJ20230807142703006, the Natural Science Foundation of China (NSFC) under Grants No. 62172041 and No. 62176021. + +# References + +[1] Bai Jinze, Bai Shuai, Yang Shusheng, Wang Shijie, Tan Sinan, Wang Peng, Lin Junyang, Zhou Chang, Zhou Jingren. Qwen-VL: A Versatile Vision-Language Model for Understanding, Localization, Text Reading, and Beyond // arXiv preprint arXiv:2308.12966. 2023. +[2] Bhat Shariq Farooq, Alhashim Ibraheem, Wonka Peter. Adabins: Depth estimation using adaptive bins // Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). 2021. 4009-4018. +[3] Bochkovskii Aleksei, Delaunoy Amael, Germain Hugo, Santos Marcel, Zhou Yichao, Richter Stephan R., Koltun Vladlen. Depth Pro: Sharp Monocular Metric Depth in Less Than a Second // International Conference on Learning Representations (ICLR). 2025. +[4] Brown Tom, Mann Benjamin, Ryder Nick, Subbiah Melanie, Kaplan Jared D, Dhariwal Prafulla, Neelakantan Arvind, Shyam Pranav, Sastry Girish, Askell Amanda, others. Language models are few-shot learners // Advances in Neural Information Processing Systems (NeurIPS). 2020. 33. 1877-1901. +[5] Chen Ziyang, Long Wei, Yao He, Zhang Yongjun, Wang Bingshu, Qin Yongbin, Wu Jia. MoCha-Stereo: Motif Channel Attention Network for Stereo Matching // Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). 2024. 27768-27777. +[6] Fan Zhiwen, Cong Wenyan, Wen Kairun, Wang Kevin, Zhang Jian, Ding Xinghao, Xu Danfei, Ivanovic Boris, Pavone Marco, Pavlakos Georgios, others. Instantsplat: Unbounded sparse-view pose-free gaussian splatting in 40 seconds // arXiv preprint arXiv:2403.20309. 2024. 2. 4. +[7] Fischler Martin A, Bolles Robert C. Random sample consensus: A paradigm for model fitting with applications to image analysis and automated cartography // Communications of the Acm. 1981. 24. 381-395. +[8] Fu Xiao, Yin Wei, Hu Mu, Wang Kaixuan, Ma Yuexin, Tan Ping, Shen Shaojie, Lin Dahua, Long Xiaoxiao. Geowizard: Unleashing the diffusion priors for 3d geometry estimation from a single image // Proceedings of the European Conference on Computer Vision (ECCV). 2025. 241-258. +[9] Godard Clément, Mac Aodha Oisin, Firman Michael, Brostow Gabriel J. Digging into self-supervised monocular depth estimation // Proceedings of the IEEE International Conference on Computer Vision (ICCV). 2019. 3828-3838. +[10] Guan Tongfan, Wang Chen, Liu Yun-Hui. Neural Markov Random Field for Stereo Matching // Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). 2024. 5459-5469. +[11] Guizilini Vitor, Ambrus Rares, Pillai Sudeep, Raventos Allan, Gaidon Adrien. 3d packing for self-supervised monocular depth estimation // Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). 2020. 2485-2494. +[12] Guo Weiyu, Li Zhaoshuo, Yang Yongkui, Wang Zheng, Taylor Russell H, Unberath Mathias, Yuille Alan, Li Yingwei. Context-enhanced stereo transformer // Proceedings of the European Conference on Computer Vision (ECCV). 2022. 263-279. +[13] Guo Xianda, Zhang Chenming, Zhang Youmin, Nie Dujun, Wang Ruilin, Zheng Wenzhao, Poggi Matteo, Chen Long. Stereo Anything: Unifying Stereo Matching with Large-Scale Mixed Data // arXiv preprint arXiv:2411.14053. 2024. +[14] Houlsby Neil, Giurgiu Andrei, Jastrzebski Stanislaw, Morrone Bruna, De Laroussilhe Quentin, Gesmundo Andrea, Attariyan Mona, Gelly Sylvain. Parameter-efficient transfer learning for NLP // Proceedings of the International Conference on Machine Learning (ICML). 2019. 2790-2799. +[15] Hu Edward J, Wallis Phillip, Allen-Zhu Zeyuan, Li Yuanzhi, Wang Shean, Wang Lu, Chen Weizhu, others. LoRA: Low-Rank Adaptation of Large Language Models // Proceedings of the International Conference on Learning Representations (ICLR). 2022. + +[16] Jing Junpeng, Li Jiankun, Xiong Pengfei, Liu Jiangyu, Liu Shuaicheng, Guo Yichen, Deng Xin, Xu Mai, Jiang Lai, Sigal Leonid. Uncertainty Guided Adaptive Warping for Robust and Efficient Stereo Matching // Proceedings of the IEEE International Conference on Computer Vision (ICCV). 2023. 3318-3327. +[17] Ke Bingxin, Obukhov Anton, Huang Shengyu, Metzger Nando, Daudt Rodrigo Caye, Schindler Konrad. Repurposing diffusion-based image generators for monocular depth estimation // Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). 2024. 9492-9502. +[18] Kerbl Bernhard, Kopanas Georgios, Leimkuhler Thomas, Drettakis George. 3d gaussian splatting for real-time radiance field rendering. // ACM Trans. Graph. 2023. 42. 139-1. +[19] Klingner Marvin, Termöhlen Jan-Aike, Mikolajczyk Jonas, Fingscheidt Tim. Self-supervised monocular depth estimation: Solving the dynamic object problem by semantic guidance // Proceedings of the European Conference on Computer Vision (ECCV). 2020. 582-600. +[20] Kong Weijie, Tian Qi, Zhang Zijian, Min Rox, Dai Zuozhuo, Zhou Jin, Xiong Jiangfeng, Li Xin, Wu Bo, Zhang Jianwei, others. Hunyuanvideo: A systematic framework for large video generative models // arXiv preprint arXiv:2412.03603. 2024. +[21] Kuaishou. Kling: AI Video Generation Tool. 2024. Accessed: 2025-03-08. +[22] Labs Black Forest. FLUX. 2024. +[23] Leroy Vincent, Cabon Yohann, Revaud Jérôme. Grounding image matching in 3d with mast3r // European Conference on Computer Vision (ECCV). 2024. 71-91. +[24] Li Jiankun, Wang Peisen, Xiong Pengfei, Cai Tao, Yan Ziwei, Yang Lei, Liu Jiangyu, Fan Haoqiang, Liu Shuaicheng. Practical stereo matching via cascaded recurrent network with adaptive correlation // Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). 2022. 16263-16272. +[25] Li Zhaoshuo, Liu Xingtong, Drenkow Nathan, Ding Andy, Creighton Francis X, Taylor Russell H, Unberath Mathias. Revisiting stereo depth estimation from a sequence-to-sequence perspective with transformers // Proceedings of the IEEE International Conference on Computer Vision (ICCV). 2021. 6197-6206. +[26] Lipman Yaron, Chen Ricky TQ, Ben-Hamu Heli, Nickel Maximilian, Le Matt. Flow Matching for Generative Modeling // Proceedings of the International Conference on Learning Representations (ICLR). 2023. +[27] Lipson Lahav, Teed Zachary, Deng Jia. Raft-stereo: Multilevel recurrent field transforms for stereo matching // International Conference on 3D Vision. 2021. 218-227. +[28] Mayer Nikolaus, Ilg Eddy, Hausser Philip, Fischer Philipp, Cremers Daniel, Dosovitskiy Alexey, Brox Thomas. A large dataset to train convolutional networks for disparity, optical flow, and scene flow estimation // Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). 2016. 4040-4048. +[29] Mousavian Arsalan, Anguelov Dragomir, Flynn John, Kosecka Jana. 3D Bounding Box Estimation Using Deep Learning and Geometry // arXiv preprint arXiv:1612.00496. 2017. +[30] OpenAI. Sora: AI Video Generation Model. 2024. Accessed: 2025-03-08. +[31] Peebles William, Xie Saining. Scalable diffusion models with transformers // Proceedings of the IEEE International Conference on Computer Vision (ICCV). 2023. 4195-4205. +[32] Ramirez Pierluigi Zama, Tosi Fabio, Poggi Matteo, Salti Samuele, Mattoccia Stefano, Di Stefano Luigi. Open Challenges in Deep Stereo: the Booster Dataset // Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). 2022. 21168-21178. +[33] Ranfil René, Bochkovskiy Alexey, Koltun Vladlen. Vision transformers for dense prediction // Proceedings of the IEEE International Conference on Computer Vision (ICCV). 2021. 12179-12188. +[34] Ravi Nikhila, Gabeur Valentin, Hu Yuan-Ting, Hu Ronghang, Ryali Chaitanya, Ma Tengyu, Khedr Haitham, Rädle Roman, Rolland Chloe, Gustafson Laura, Mintun Eric, Pan Junting, Alwala Kalyan Vasudev, Carion Nicolas, Wu Chao-Yuan, Girshick Ross, Dollar Piotr, Feichtenhofer Christoph. SAM 2: Segment Anything in Images and Videos // arXiv preprint arXiv:2408.00714. 2024. +[35] Sung Yi-Lin, Cho Jaemin, Bansal Mohit. Lst: Ladder side-tuning for parameter and memory efficient transfer learning // Advances in Neural Information Processing Systems (NeurIPS). 2022. 35. 12991-13005. + +[36] Suvorov Roman, Logacheva Elizaveta, Mashikhin Anton, Remizova Anastasia, Ashukha Arsenii, Silvestrov Aleksei, Kong Naejin, Goka Harshith, Park Kiwoong, Lempitsky Victor. Resolution-robust Large Mask Inpainting with Fourier Convolutions // arXiv preprint arXiv:2109.07161. 2021. +[37] Team Pallets. Flask: A lightweight WSGI web application framework. 2025. Accessed: 2025-03-08. +[38] Telea Alexandru. An image inpainting technique based on the fast marching method // Journal of Graphics Tools. 2004. 9. 23-34. +[39] Wang Jianyuan, Chen Minghao, Karaev Nikita, Vedaldi Andrea, Rupprecht Christian, Novotny David. VGGT: Visual Geometry Grounded Transformer // Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). 2025. +[40] Wang Peng, Bai Shuai, Tan Sinan, Wang Shijie, Fan Zhihao, Bai Jinze, Chen Keqin, Liu Xuejing, Wang Jialin, Ge Wenbin, Fan Yang, Dang Kai, Du Mengfei, Ren Xuancheng, Men Rui, Liu Dayiheng, Zhou Chang, Zhou Jingren, Lin Junyang. Qwen2-VL: Enhancing Vision-Language Model's Perception of the World at Any Resolution // arXiv preprint arXiv:2409.12191. 2024. +[41] Wang Shuzhe, Leroy Vincent, Cabon Yohann, Chidlovskii Boris, Revaud Jerome. Dust3r: Geometric 3d vision made easy // Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). 2024. 20697-20709. +[42] Wang Xianqi, Xu Gangwei, Jia Hao, Yang Xin. Selective-stereo: Adaptive frequency information selection for stereo matching // Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). 2024. 19701-19710. +[43] Weinzaepfel Philippe, Leroy Vincent, Lucas Thomas, Brégier Romain, Cabon Yohann, Arora Vaibhav, Antsfeld Leonid, Chidlovskii Boris, Csurka Gabriela, Revaud Jérôme. Croco: Self-supervised pre-training for 3d vision tasks by cross-view completion // Advances in Neural Information Processing Systems (NeurIPS). 2022. 3502-3516. +[44] Weinzaepfel Philippe, Lucas Thomas, Leroy Vincent, Cabon Yohann, Arora Vaibhav, Brégier Romain, Csurka Gabriela, Antsfeld Leonid, Chidlovskii Boris, Revaud Jérôme. CroCo v2: Improved cross-view completion pre-training for stereo matching and optical flow // Proceedings of the IEEE International Conference on Computer Vision (ICCV). 2023. 17969-17980. +[45] Yang Lihe, Kang Bingyi, Huang Zilong, Xu Xiaogang, Feng Jiashi, Zhao Hengshuang. Depth anything: Unleashing the power of large-scale unlabeled data // Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). 2024. 10371-10381. +[46] Yang Lihe, Kang Bingyi, Huang Zilong, Zhao Zhen, Xu Xiaogang, Feng Jiashi, Zhao Hengshuang. Depth Anything V2 // arXiv preprint arXiv:2406.09414. 2024. +[47] Yin Wei, Zhang Chi, Chen Hao, Cai Zhipeng, Yu Gang, Wang Kaixuan, Chen Xiaozhi, Shen Chunhua. Metric3d: Towards zero-shot metric 3d prediction from a single image // Proceedings of the IEEE International Conference on Computer Vision (ICCV). 2023. 9043-9053. +[48] Zeng Jiaxi, Yao Chengtang, Yu Lidong, Wu Yuwei, Jia Yunde. Parameterized cost volume for stereo matching // Proceedings of the IEEE International Conference on Computer Vision (ICCV). 2023. 18347-18357. +[49] Zhang Jingyi, Huang Jiaxing, Jin Sheng, Lu Shijian. Vision-language models for vision tasks: A survey // IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI). 2024. 5625-5644. +[50] Zhao Wenliang, Rao Yongming, Liu Zuyan, Liu Benlin, Zhou Jie, Lu Jiwen. Unleashing text-to-image diffusion models for visual perception // Proceedings of the IEEE International Conference on Computer Vision (ICCV). 2023. 5729-5739. + +# Appendix + +# A 3D Visual Illusion Dataset + +# A.1 Video Collection + +We collect a large amount of videos from web-source data and generative models, covering five types of illusions: inpainting illusion (e.g., inpainting on a wall/floor), picture illusion (e.g., picture printed/drawn on a paper), replay illusion (e.g., video replayed on different screens), holography illusion, and mirror illusion (e.g., specular or transparent surfaces), as shown in Figure 7. When collecting videos from generative models, we observe four important considerations in the design of text prompts. (1) Level of Detail in Prompts: Overly fine-grained control in prompts often leads to physically unrealistic results, such as requiring the object in the mirror to maintain the same pose as its real-world counterpart, enforcing perfect mirror symmetry, or specifying excessive positional details. Instead, less detailed scene descriptions tend to produce more physically accurate and realistic results. (2) Challenges with Dynamic Objects: Generating videos with dynamic objects proves particularly difficult. The virtual image in the mirror and the real-world objects often exhibit motion inconsistencies. As a result, we focus primarily on static scenes or those with only slight object movement. (3) Layout Complexity: Complex scene layouts frequently lead to mismatches between the mirror world and the real world, causing spatial inconsistencies. (4) Camera Motion: To ensure a stable and realistic scene, the camera is required to pan slowly. Excessive camera movement may result in abrupt rotations or scene transitions, disrupting the illusion. + +![](images/8a89d42bbfa3ed059f03cfede5551bc2fc72046b04014a72a16208b3eb854295.jpg) +(a) Inpainting Illusion + +![](images/400bca50ca726d75479de165a1f89bdeee81830ddc5c5574cef96f689d2b2422.jpg) + +![](images/fa6fe5291dd21bfeca1c4729785e4f2d4b057d3aea96aa36232eda3bb03c6b41.jpg) + +![](images/6e70ea76a49ccf3db16df07f458e8c05375a440bdb4af2fa3565edcd1fe898c2.jpg) +(b) Picture Illusion + +![](images/7f615fd3a04eed33b48767939e6d94c1b3dae1eb3242693e20514d448059788b.jpg) + +![](images/ca1897a616fca3aeeaa71ce0bc5c6538d28d3ecbc40fae79a3a54381e8779871.jpg) + +![](images/e8e34013de4cbac6892669704d6d6ba3160624dfcdab0e7d181e76343f602f95.jpg) +(c) Replay Illusion + +![](images/3cb51392762219861e51b0c1a484c38e254147192add5a00db91583ac035f91f.jpg) + +![](images/163d1d1a55083b6d8c3bcc9d3476a447a6d733af08f50a69a1f13d9bab251910.jpg) + +![](images/ee159e5a31916fc2a7d5f5acb7fa77ad880c8a6bce42d8d4a47da5d971dc3554.jpg) +(d) Holography Illusion + +![](images/8e5ad0822151f1b65c5cfc4d27aacab009ed92a1a5d3bbda6b43141449d04187.jpg) + +![](images/e99ffa134a44f1a681181c763f9600f25c48375dba2b15555707231f46bd6843.jpg) + +![](images/931ac69ae2230c938d11945380f3818f5670f9ac21385a4b4fd78acc607e67a8.jpg) +Figure 7: The visualization of 3D visual illusions. + +![](images/885e924a87cc9e5719ca3943cfca69afc7702d2ac65987331e6e1a2a20d53063.jpg) +(e) Mirror Illusion + +![](images/285ae510e3d8e09f15a46c9714ab0a552576c999b59dbf58c3660539c1460f2f.jpg) + +![](images/0b1a8b594808c248fdfa3a36f4f90a5097a868b3293544fc2d4564d237296923.jpg) + +# A.2 Depth Geneation + +To mitigate the impact of noise in plane fitting, we adopt RANSAC for robust plane estimation: + +$$ +\min _ {\alpha , \beta , \delta , \gamma} \sum_ {i = 1} ^ {N} \left(\alpha \cdot u _ {i} + \beta \cdot v _ {i} + \delta \cdot d _ {i} + \gamma\right) ^ {2}, \tag {11} +$$ + +$$ +s u b j e c t \quad t o \quad \alpha^ {2} + \beta^ {2} + \delta^ {2} = 1. +$$ + +$(\alpha, \beta, \delta, \gamma)$ are the plane parameters, $(u, v)$ is image plane coordinate and $d$ is disparity. As illustrated in Algorithm 1,, we randomly sample three points to define a candidate plane at each iteration of RANSAC. The plane normal $(\alpha, \beta, \delta)$ is computed as the cross product of vectors formed by these three points, and the offset $\gamma$ is derived by substituting one point into $\alpha \cdot u + \beta \cdot v + \delta \cdot d + \gamma = 0$ . We then compute the distance from each point in the support region to the candidate plane to determine the inliers. After all iterations, the candidate plane with the largest number of inliers is selected, which are taken as the best inlier set. The plane parameters $(\alpha, \beta, \delta, \gamma)$ are then estimated by computing the + +eigenvector corresponding to the smallest eigenvalue of the covariance matrix constructed from the best inlier set. We also present visualizations of the rendered and rectified depth images in Figure 8 and 9. After applying plane fitting for rectification, the resulting depth map becomes smoother and more geometrically accurate. + +Algorithm 1 RANSAC Plane Fitting +Require: Point set $\mathbf{P} = \{(u_i,v_i,d_i)\}_{i = 1}^N\in \mathbb{R}^{N\times 3}$ , inlier threshold $\tau_{d}$ , sub-sample size per iteration $M$ max iterations $T_{p}$ +Ensure: Optimal plane parameters $\pi^{*} = [\alpha ,\beta ,\delta ,\gamma ]$ +1: Initialize: best_score $= 0$ best(plane $= 0$ best_inliers $= \emptyset$ +2: for $t = 1$ to $T_{p}$ do +3: Randomly sample $M$ sets of 3-point tuples: $Q = \{(u_i^0,v_i^0,d_i^0),(u_i^1,v_i^1,d_i^1),(u_i^2,v_i^2,d_i^2)\} _i^M\in \mathbb{R}^{M\times 3\times 3}$ +4: for $b = 1$ to $M$ do +5: $\mathbf{v}_{10} = (u_i^1,v_i^1,d_i^1) - (u_i^0,v_i^0,d_i^0),\mathbf{v}_{20} = (u_i^2,v_i^2,d_i^2) - (u_i^0,v_i^0,d_i^0)$ +6: $\mathbf{n}_b = \mathbf{v}_{10}\times \mathbf{v}_{20}$ {Normal via cross product} +7: $d_b = -\mathbf{n}_b^\top [u_i^1,v_i^1,d_i^1]$ +8: $\pi_t[b] = [\mathbf{n}_b^\top ,d_b]$ +9: end for +10: $\mathbf{D}_t = \pi_t[\mathbf{P},\mathbf{1}]^\top /\| \pi_t[:,0:3]\| _2$ {Batch distance computation} +11: $\mathbf{M}_t = \| \mathbf{D}_t\| < \tau_d$ +12: $\mathbf{c}_t = \mathrm{sum}(\mathbf{M}_t,\mathrm{dim} = 1)$ +13: $k = \arg \max c_{t},c_{\mathrm{max}} = c_{t}[k]$ +14: if $c_{\mathrm{max}} >$ best_score then +15: best_score $= c_{\mathrm{max}}$ +16: best(plane $= \pi_{t}[k]$ +17: best_inliers $= \mathbf{M}_t[k]$ +18: end if +19: end for +20: Refinement via Eigen Decomposition +21: PInliers $=$ P[best_inliers] +22: $\mathbf{S} = [\mathbf{P}_{\mathrm{inliers}},\mathbf{1}]^\top [\mathbf{P}_{\mathrm{inliers}},\mathbf{1}]$ +23: $(\mathbf{W},\mathbf{V}) = \mathrm{eigh}(\mathbf{S})$ +24: $\pi^{*} = \mathbf{V}(:,0]$ +25: return $\pi^{*}$ + +# A.3 Right Image Geneation + +The right-view images for generative-model videos are directly rendered using Gaussian Splatting (GS). For web-sourced videos, right views are generated by warping the left images using monocular disparity. As shown in Algorithm 2, we generate a right-view image $\hat{I}_R$ from a given left-view image $\hat{I}_L$ and disparity map $D$ . It begins by estimating an appropriate disparity scaling factor $s$ via binary search, ensuring that a sufficient proportion of the projected pixels fall within valid image bounds. Using the computed $s$ , pixel coordinates are mapped from the left to the right view, with invalid coordinates filtered out. An initial right-view image is synthesized by transferring valid pixel values based on the mapping. Finally, image inpainting is applied to fill missing regions, resulting in the completed right-view image $\hat{I}_R$ . The algorithm outputs both $\hat{I}_R$ and the estimated scaling factor $s$ . We also present the visualization of the initial warped image and the inpainted image in Figure refig: vis web source. The inpainting process effectively fills in the missing regions, resulting in a more complete and visually coherent right-view image. + +# A.4 Real-world Data + +# A.4.1 Camera System + +We collect real-world data using a stereo camera (ZED Mini) and a LiDAR-based depth sensor (Realsense L515). The sensors are rigidly mounted and calibrated using a checkerboard to ensure + +![](images/f4ce75e7580ff33a45792dfdb367ec2c43422497b362561872d234c3aa7bf169.jpg) +Left Image + +![](images/f6a5072fc277b1702cdb78f057dd3c23528664ad5504bfac75bcd4abec83b565.jpg) + +![](images/b1a6af5f1c651b5132abeee31fd652fb5aca2215bbe75c4442cf31c134ef8d09.jpg) + +![](images/df91f34acdfa362f8ce830031b2c443d4627a2d6c9f75644dce748c1fbbc5495.jpg) + +![](images/14474228d0e96fe9cc28382f7540822be165a150c6ed3b102ecd1dc8244df5a1.jpg) +Figure 8: The visualization of results for video from generative models. + +![](images/3c08a3362c4448268324258d905c0fdae04c088a519abde5f52b59102929ecb0.jpg) +Right Image + +![](images/60eaf54f5f36646b38b98dec906c2320a62b1b466bc9a9998736a4ae71d81e32.jpg) + +![](images/89a02d3133d9960f667e80689a249f3a34ca4f409084f3aa29d4c3d99b53aabb.jpg) + +![](images/1fc381c1d7b844ec6aba5f65dd1bc74d352fb957d024225ede8dc8be39e80671.jpg) + +![](images/b1074def98831b2150bceca301fbd3d38e9be2fc294a92bbf50cbeef829a5d69.jpg) + +![](images/02a3e9af48852577c1e3a3dfdcb41a31bf1c62d2fcb43efa28c476a11b620232.jpg) +Rendered Depth + +![](images/d582526192e49f9168a2d410ac2c78c8fabce41df41cb0637b06da9be96908ad.jpg) + +![](images/87749ffc2a0bb875685e5daaed842e42d00e6ac78c4346399fb7632d84c33947.jpg) + +![](images/d6e99957552f39eba39c80c3cf393ea07b0a4da77622f0921cc11c07fc742dae.jpg) + +![](images/47bf1f2bfb63269edf823ac4d71a13e8547e8e4c4f545f450591c9e31b2cd678.jpg) + +![](images/00eee363fff9477d84ff21fe796f66b13bfc7729dbb5243bc4309c53393d0020.jpg) +Rectified Depth + +![](images/df8fd14be48bac2425ad6610326d8bca6a053f46b2ae061814e09f8a11f657e7.jpg) + +![](images/b7355646b53baa4916abfc4d84479e06ff4bdd6fb5985ad1dc133f86967dd878.jpg) + +![](images/4bb1ba084172d9d9759478c2831a3f07f1cccf97a1dae57a92d03fdd731bcd13.jpg) + +![](images/5f179e03eefef0fd9a05229b1b80a11f0e9ae5e22185d2ffff735e3883ff5d1e.jpg) + +![](images/57731c481ae7512d3d9178f89651ebc80605b78dddebaff32dcdf6d34cd62ae3.jpg) +RGB Image + +![](images/5226f4d0af39742a6f2edadcb4d8e99410076feb6a0ddc9860847c36eb054563.jpg) + +![](images/246574fce318a22c9fb044d0e9a8f723afb18a70a389d24390def3f6e63914ce.jpg) + +![](images/94347aeaa60cadce0c81d48c840d99d39560ebbaa77218fe62f7ecd45669545c.jpg) +Figure 9: The visualization of results for web-source video. + +![](images/02532afe9bc3a1f1ec07305ab8c31e177f7c9c96f02e2b93042d2dbcf14b5767.jpg) +Depth Estimation + +![](images/eff69ecaafca276ad9533289ff89140d191d8a252c64e5e739e81cb5c66dd4f3.jpg) + +![](images/87921e7c3065e0ea7ba310b8e91f91b8633f9f97a5c80aa0c64766ffd87b7341.jpg) + +![](images/e71c1e99a4a26f4ef6edf3d333f363d10051f66a49b32733e471456aa61352d7.jpg) + +![](images/652a764253af6a844b6762eff994a2b90af4a0ac26ee34aee45e80152a333069.jpg) +Rectified Depth + +![](images/43caebecea25f3ea99b0586c8d70b5df1e2f4e2d5d1f300a53f4cf64cbd24522.jpg) + +![](images/b3e0f86f651a5c5cc19cb86a2625d6e6daf5e09e614bb4f15b66e737f7d50703.jpg) + +![](images/01d1bf1c8acd49007656b7d566d51c97edaa05684dfa0c96195eefd93262ead1.jpg) + +![](images/13024a65e834fd30f0bc47b133e994bea0773bc5adf999b759db75a316e85b79.jpg) +Warping + +![](images/30c041af59e0d71fed086c653a763c2198600293a15fbee00a65cf7d9c087f70.jpg) + +![](images/cf01e64df5aae5bffbcd0a249058e43a1dd2251d75192e171902809b7bf8b077.jpg) + +![](images/36bdc958d1dd96d82b0c864814f2bb72936450860efef14fc498175567ef3d6d.jpg) + +![](images/96f027e1ad92177c0ca2f3f4e1e49641587dba5430bef170b99770bbae2efadb.jpg) +Inpainting + +![](images/22e4410ef1f664d14f2337d7b7a9d0f49a079ee87c58b949dc947c8621a988a4.jpg) + +![](images/d6b033e94b5f44ff2a7e881994168ec884acede5e4d6cdfd9558822daa28100d.jpg) + +![](images/d89c33709f600a77cdfdf3ecb82a0431d0b83b5d05261b2209e4b2a46c74a9a8.jpg) + +accurate alignment, as shown in Figure 10. The ZED Mini captures RGB images, while the L515 provides depth maps. The intrinsic and extrinsic parameters of both cameras are obtained through calibration. The calibration process involves capturing multiple images of the checkerboard pattern from different angles and distances, allowing for accurate estimation of the camera parameters. + +# A.4.2 Depth Map Projection + +The L515 depth map is warped to the ZED left camera to construct the ground-truth depth. As shown in Algorithm 3, the process begins by upsampling the depth map and scaling the intrinsic matrix accordingly. 3D points are then computed and transformed from the L515 frame to the ZED frame using calibrated extrinsics, followed by projection onto the ZED image plane. The resulting depth values are splatted to the ZED image grid, and missing regions are filled using inpainting and + +Algorithm 2 Right Image Generation +```txt +Require: Left image $I_L \in \mathbb{R}^{H \times W \times 3}$ , disparity map $D \in \mathbb{R}^{H \times W}$ , valid pixel threshold $\theta = 0.9$ , maximum iterations $T_g$ +``` + +Ensure: Synthesized right image $\hat{I}_R$ , scaling factor $s$ + +1: Step 1: Compute Scaling Factor +2: Initialize: $l = 0$ , $r = W / (4 \cdot \max(D))$ , $\epsilon = 10^{-6}$ , $t = 0$ +3: while $|r - l| > \epsilon$ and $t < T_g$ do +4: $t = t + 1$ +5: $s = (l + r) / 2$ +6: Coordinate projection: $U' = U - s \cdot D$ +7: Compute valid pixel ratio: $\eta = \frac{1}{HW}\sum \mathbb{I}(U^{\prime}\in [0,W))$ +8: if $\eta \geq \theta$ then +9: $l = s$ +0: else +11: $r = s$ +12: end if +13: end while +14: Final scaling factor: $s = (l + r) / 2$ +15: Step 2: Image Coordinate Mapping +16: Generate coordinate grid: $(u,v) = \mathsf{MESHGRID}(0:W - 1,0:H - 1)$ +17: Compute projected coordinates: $u' = u - s \cdot D(u, v)$ +18: Quantize coordinates: $\hat{u}^{\prime} = \{\lfloor u^{\prime}\rfloor ,\lceil u^{\prime}\rceil \}$ +19: Filter invalid coordinates: $\{\hat{u}^{\prime}\mid \hat{u}^{\prime}\geq 0$ and $\hat{u}^{\prime} < W\}$ +20: Step 3: Right View Image Synthesis +21: Initialize: $I_R = \mathbf{0}^{H \times W \times 3}$ +22: for each pixel $(u,v)$ do +23: Generate initial right-view image $I_R$ : $I_R(u', v) = I_L(u^*, v)$ , $u^* = \arg \max_u \{ d_{(u,v)} \mid u - s \cdot D_{(u,v)} = u' \}$ . +24: end for +25: Step 4: Image Completion Perform inpainting on $I_R$ to fill invalid regions and obtain the final right-view image $\hat{I}_R$ +26: return $\hat{I}_R, s$ + +![](images/fc4aa7d2ea1a89351ebecd10699042c61285f9086709ed2eb9f60c0d597559b4.jpg) +Figure 10: Camera System and Calibration Visualization + +![](images/46439986bd61a3ac113622f4b5d7f2327c623909f719bb7800cd8cb49e951d08.jpg) +L515 RGB Camera +ZED Left Camera + +![](images/36441b42c4cf2739feb3f055df930e9094b307f7ea779bfdc7eef7fb9a585c13.jpg) +ZED Right Camera +ZED Left Camera + +guided filtering. To ensure consistency, a backward reprojection step verifies each pixel's validity by comparing it with the original L515 depth. Finally, noise is suppressed using median filtering, and valid depth values are converted into disparities based on the ZED stereo baseline and focal length. + +Algorithm 3 Depth Map Reprojection +Require: Depth map from L515 camera $\mathcal{Z}_L\in \mathbb{R}^{H_L\times W_L}$ , RGB image from ZED left camera $I_{Z}\in \mathbb{R}^{H^{\prime}\times W^{\prime}\times 3}$ , intrinsic matrix of L515 $K_{L}\in \mathbb{R}^{3\times 3}$ , rotation matrix $R\in \mathbb{R}^{3\times 3}$ and translation matrix $T\in \mathbb{R}^{3\times 1}$ from L515 to ZED-left, upsampling factor $s = 3$ , ZED stereo baseline $B$ , and ZED focal length $F$ +Ensure: Disparity map of ZED left camera $D\in \mathbb{R}^{H^{\prime}\times W^{\prime}}$ +1: Step 1: Depth Upsampling +2: $\tilde{\mathcal{Z}}_{\mathrm{L}} = \mathrm{resize}(\mathcal{Z}_{\mathrm{L}},\mathrm{scale} = s,\mathrm{interp} = \mathrm{NEAREST})$ +3: $\tilde{K}_{\mathrm{L}} = s\cdot K_{\mathrm{L}}$ +4: Step 2: Coordinate Transformation +5: for each pixel $(u_{L},v_{L})$ in $\tilde{\mathcal{Z}}_{\mathrm{L}}$ do +6: $[x_Z,y_Z,z_Z] = R\cdot \tilde{\mathcal{Z}}_L\cdot \tilde{K}_L^{-1}\cdot [u_L,v_L,1]^T +T$ +7: $[u_Z,v_Z,1] = K_Z\cdot [x_Z / z_Z,y_Z / z_Z,1]$ +8: end for +9: Step 3: Depth Projection +10: Initialize $\mathcal{Z}_{\mathrm{Z}} = \infty^{H^{\prime}\times W^{\prime}}$ +11: for each projected point $(u_{Z}^{i},v_{Z}^{i},z_{Z}^{i})$ do +12: $(u_{1},v_{1}) = (\lfloor u_{Z}^{i}\rfloor ,\lfloor v_{Z}^{i}\rfloor)$ $(u_{2},v_{2}) = (\lceil u_{Z}^{i}\rceil ,\lceil v_{Z}^{i}\rceil)$ +13: for $(u,v)\in \{(u_1,v_1),(u_1,v_2),(u_2,v_1),(u_2,v_2)\}$ do +14: if $(u,v)$ is within image bounds then +15: $\mathcal{Z}_{\mathrm{Z}}(v,u) = \min (\mathcal{Z}_{\mathrm{Z}}(v,u),z_{Z}^{i})$ +16: end if +17: end for +18: end for +19: Step 4: Hole Filling +20: $\mathcal{M}_{\mathrm{invalid}} = (\mathcal{Z}_{\mathrm{Z}} == \infty)$ $\mathcal{Z}_{\mathrm{Z}} = \mathcal{Z}_{\mathrm{Z}}\odot \neg \mathcal{M}_{\mathrm{invalid}} + \mathbf{0}\odot \mathcal{M}_{\mathrm{invalid}}$ +21: $\mathcal{M}_{\mathrm{small}} = \mathrm{connectedComponents}(\mathcal{M}_{\mathrm{invalid}},\mathrm{area\_th} = 100)$ +22: $\mathcal{Z}_{\mathrm{repair\_small}} = \mathrm{inpaint}(\mathcal{Z}_{\mathrm{Z}},\mathcal{M}_{\mathrm{small}})$ $\mathcal{Z}_{\mathrm{repair\_all}} = \mathrm{inpaint}(\mathcal{Z}_{\mathrm{Z}},\mathcal{M}_{\mathrm{invalid}})$ +23: $\mathcal{Z}_{\mathrm{repair\_all}} = \mathrm{guidedFilter}(I_Z,\mathcal{Z}_{\mathrm{repair\_all}},\mathrm{radius} = 5,\epsilon = 1e - 3)$ +24: $\mathcal{Z}_{\mathrm{repair}} = \mathcal{Z}_{\mathrm{Z}}\odot \neg \mathcal{M}_{\mathrm{invalid}} + \mathcal{Z}_{\mathrm{repair\_all}}\odot \neg (\mathcal{Z}_{\mathrm{repair\_small}} == 0)\odot \mathcal{M}_{\mathrm{invalid}}$ +25: Step 5: Backward Reprojection for Invalid Region Detection +26: for each pixel $(u_Z,v_Z)$ in $\mathcal{Z}_{\mathrm{repair}}$ do +27: $[x_{Z\to L},y_{Z\to L},z_{Z\to L}] = R^{-1}\cdot (z_Z\cdot K_Z^{-1}\cdot [u_Z,v_Z,1] - T)$ +28: $[u_{Z\to L},v_{Z\to L},1] = \frac {1}{z _ { Z } \to L } . K _ { L } [ x _ { Z } \to L , y _ { Z } \to L , z _ { Z } \to L ]$ +29: if $\mathcal{Z}_L(v_{Z\to L},u_{Z\to L}) == 0$ or $\| \mathcal{Z}_L(v_{Z\to L},u_{Z\to L}) - z_{Z\to L}\| >\tau$ then +30: $\mathcal{Z}_{\mathrm{repair}}(u_Z,v_Z) = 0$ +31: end if +32: end for +33: Step 6: Noise Suppression +34: $\mathcal{Z}_{\mathrm{smaoath}} = \mathrm{medianFilter}(\mathcal{Z}_{\mathrm{repair}},\mathrm{size} = 3)$ +35: Mnoise $= |\mathcal{Z}_{\mathrm{repair}} - \mathcal{Z}_{\mathrm{smaoath}}| > 0.03$ +36: Zfinal $= \mathcal{Z}_{\mathrm{repair}}\odot \neg \mathcal{M}_{\mathrm{noise}} + 0$ C Mnoise +37: Step 7: Disparity Computation +38: $D = B\cdot F / Z_{\mathrm{final}}$ +39: return D + +# B VLM-Driven Monocular-Stereo Fusion Model + +Inspired by the strong reasoning ability of vision-language models (VLMs), as shown in Figure 11. We propose a VLM-driven monocular-stereo fusion model that leverages commonsense knowledge to assess the reliability of monocular and stereo depth in different regions, enabling effective depth fusion. + +![](images/f305e44042c4a4810f9197dcbe7ebf840752385447ee4f4633d57112c4065456.jpg) +Is the middle area of the picture a mirror or a distant scene? + +![](images/cdbf04cfc31557506b58aadaffd4cf25e80a7abdc15f3c8439f7f622729859f7.jpg) + +![](images/cad209f2c99266bc98e4affac3b5643a9820ca89da42d609a82e619bfad32fb7.jpg) +Figure 11: Visualization of commonsense knowledge from VLMs. The left image shows a scene with a mirror. The right image depicts a scene with a transparent object, where the background is visible through the object. + +The middle area of the image is a mirror, not a distant scene. Here are a few clues that confirm this: + +1. The identical robes: The man walking appears to be wearing the exact same robe as the man whose back is facing the camera. +2. Mirored motion: The positioning and posture of the reflected figure match precisely what you would expect if he were mirroring the man in the foreground. +3. Matching details: Elements in the background - like the furniture, wall decor, and the carpet - align symmetrically across the center line, which supports it being a reflection. + +So yes, you're looking at a mirror with a reflection of the man and the room behind him. + +![](images/de4097c87d1655500698bea14765eabfa7b28c1b62212b1f45a288822f3c538c.jpg) + +The middle area of the picture is a mirror. This can be determined by the reflection of the person in the foreground, who is facing the mirror and whose back is visible to the viewer. The reflection shows the same person from the front, indicating that they are looking into a mirror. Additionally, the symmetry and the way the room is reflected further confirm that it is a mirror rather than a distant scene. + +![](images/321ecf90d42c42817165b0bc1ee06c702ca4ac3a9e5c09d63c4d2df252ac003e.jpg) +Is the middle area of the picture a mirror or a distant scene? + +![](images/37785b25378e278d3f0e5c50cbf7e4045ea0332e018125cbd79f9df520bd7da7.jpg) + +![](images/69b238873f12a4cd72dee9a2c543efae8caf319bdb3c5085751329d9e5d8c0cd.jpg) + +The middle area of the picture is glass, specifically a large glass window or door. Here's how you can tell: + +1. There is a clear reflection of a plant and interior elements (like a chair and railing), which wouldn't appear like that in a distant scene. +2. The reflections are slightly transparent, allowing some visibility of the outdoors (grass, mountains, and sky) through the glass. +3. The angle and lighting strongly suggest this is a reflective surface, typical of glass windows or sliding doors. + +So yes, you're looking at a glass surface with reflections of the interior space. + +![](images/af87084ef759944324ef4499da2e28bcab82d175670e158364cc3688103ba372.jpg) + +The middle area of the picture is a glass window. It reflects the surrounding environment, including trees and possibly distant scenery, but it also shows the interior of a building through its transparency. The reflections on the glass create an illusion that might make it appear like a distant scene, but it is indeed a glass surface reflecting and transmitting light from both sides. + +# B.1 Loss + +The supervised loss function consists of two main components: one $(\mathcal{L}_d)$ for the disparity maps and the other $(\mathcal{L}_c)$ for the confidence map: + +$$ +\mathcal {L} = \mathcal {L} _ {d} + w \mathcal {L} _ {c}, \tag {12} +$$ + +where $w$ is a manually set weighting factor for balancing the confidence map loss. + +For disparity supervision, we use the $L_{1}$ loss to supervise each iteratively updated disparity $D_{s}^{t}$ , the aligned monocular disparity $\tilde{D}_{m}$ , and the final predicted disparity $D_{f}$ . The loss function is defined as: + +$$ +\begin{array}{l} \mathcal {L} _ {d} = \sum_ {t = 1} ^ {T} \gamma_ {d} ^ {T + 2 - t} \left| \left| D _ {s} ^ {t} - D _ {G} \right| \right| _ {1} \tag {13} \\ + \gamma_ {d} | | \tilde {D} _ {m} - D _ {G} | | _ {1} + | | D _ {f} - D _ {G} | | _ {1}. \\ \end{array} +$$ + +Here, $D_G$ denotes the ground-truth disparity, and $\gamma_d$ is a weighting coefficient to balance contributions from intermediate predictions. + +For confidence map supervision, we adopt the Focal Loss, where the ground-truth for confidence is derived based on the disparity difference between the final stereo prediction $D_{s}^{T}$ and the ground-truth + +$D_G$ : + +$$ +\mathcal {L} _ {c} = \frac {1}{N} \sum_ {i} \alpha_ {c} \cdot \left(1 - e ^ {- \mathcal {L} _ {b} (i)}\right) ^ {\gamma_ {c}} \cdot \mathcal {L} _ {b} (i), +$$ + +$$ +\mathcal {L} _ {b} = - \bar {I} _ {c} \log I _ {c} - (1 - \bar {I} _ {c}) \log (1 - I _ {c}), \tag {14} +$$ + +$$ +\bar {I} _ {c} = \mathbb {I} (\text {I n t e r p o l a t e} (| D _ {G} - D _ {s} ^ {T} |, \text {s c a l e} = \frac {1}{4}) < \frac {5}{4}), +$$ + +In this formulation, $\alpha_{c}$ and $\gamma_{c}$ are the hyperparameters of the Focal Loss, $\mathbb{I}$ is the indicator function, and Interpolate denotes the downsampling operation that resizes the supervision signal to $\frac{1}{4}$ of the original resolution, matching the resolution used in intermediate network outputs. + +# C Experiments + +# C.1 Evaluation Metric + +We evaluate model performance in both disparity space and depth space. In disparity space, two commonly used metrics are adopted. (1) End-Point Error (EPE): $EPE = \frac{1}{N} \sum_{i} |r_i - \bar{r}_i|$ , where $r$ and $\bar{r}$ denote the predicted and ground-truth disparity values, respectively. EPE measures the average absolute disparity error in pixels. (2) Bad- $x$ Error: bad- $x = \frac{1}{N} \sum_{i} \mathbb{I}(|r_i - \bar{r}_i| > x)$ , which indicates the percentage of pixels where the disparity error exceeds $x$ pixels. This metric is especially useful for evaluating the robustness of the model under boundary conditions. In depth space, four standard evaluation metrics are employed. (1) Absolute Relative Error (AbsRel): AbsRel = $\frac{1}{N} \sum_{i} \frac{|r_i - \tilde{r}_i|}{\tilde{r}_i}$ , which evaluates the relative difference between predictions and ground-truth values, normalized to mitigate the impact of scale and unit differences, making it suitable for datasets with diverse depth ranges. (2) Root Mean Squared Error (RMS): RMS = $\sqrt{\frac{1}{N} \sum_{i} (r_i - \bar{r}_i)^2}$ . (3) Log10 Error: $log10 = \frac{1}{N} \sum_{i} |\log_{10}(r_i) - \log_{10}(\bar{r}_i)|$ . (4) Threshold Accuracy ( $\delta_1$ ): $\delta_1 = \frac{1}{N} \sum_{i} \mathbb{I}(\max(\frac{y_i}{\bar{y}_i}, \frac{\tilde{y}_i}{y_i}) < 1.25)$ , which measures the proportion of pixels for which the predicted depth falls within a certain ratio (e.g., 1.25) of the ground truth. The model is initially trained on the SceneFlow dataset, then fine-tuned on the 3D-Visual-Illusion training set, and finally evaluated on the 3D-Illusion test set and the Booster training set. + +# C.2 Implementation Details + +For dataset construction, we use Qwen2-VL-72B [1, 40] to perform initial screening, reducing the dataset from 5226 videos with 52M frames to 4,519 videos with 1.4M frames. We then built a Flask[37]-based web tool to manually reduce data to 1384 videos with 236k frames. We further developed a more convenient flask-based web app for SAM2 [34] to acquire the semantic mask of illusion and support regions. During the semantic segmentation, we delete frames with redundant content and illusions imperceptible to humans, reducing data from 236k frames to 176,530 frames. Later, we rectify the depth values of illusion regions from the reference of support regions, which is further used to generate the right images for web-source data. Besides web-source data, We also use large generative models to generate 234 videos with 2382 frames. The video generation is achieved via Sora [30] and Kling [21], and a small part of the data is generated from HunyuanVideo [20]. We then use InstantSplat [6], DUSt3R [41], and GS [18] to generate the right images and depth map, followed by similar depth post-processing. + +As for our VLM-driven monocular-stereo fusion network, we benefit from the vision and language foundation model and use Depthanything V2 [45, 46] as a pre-trained monocular model, QwenV12-7B [40, 1] as pre-trained VLM, and FLUX [22] as a diffusion model. We use Lora [15] to fine-tune the last layer of QwenV12-7B and the Q$$V projection layer of FLUX on $4 \times$ H100 with a batch size of 6 on each GPU. The entire training takes almost 20 days. + +# C.3 Prompts + +In dataset construction, we design a prompt to prefilter bad frames using Qwen2-VL-72B [1, 40] due to the large amount of videos collected from the Internet. The prompts are as follows: + +Reply to me in the format of a string concatenating 'yes' or 'no' with '.'. Each 'yes or 'no' is an answer to each following question. Does this image feature any flat artistic creation of landscapes where the surface of the creation is flat and has no ups and downs? Does this image contain any areas with perspective illusions? Does this image contain any optical-illusion graffiti or artwork? Does this image contain any transparent or high-reflective areas? Does this image show a display screen playing 3D objects or scenes? Does the image contain areas that make you mistake them for 3D objects? Does this image contain excessive watermarks or captions that seriously affect its quality? Does this image contain small watermarks or captions in the corners? Is this image too blurry? Are most regions of the artistic creation covered by a single/two hands? Is this image a software interface? Is only the figure of the artist clear, but the others are blurry, like artwork, screen, or areas that make you mistake them for 3D objects? + +We use the answer from Qwen2-VL-72B to filter out bad frames. We reduce the data from 5,226 videos and 52 million frames to 4,519 videos and 1.4 million frames. + +In addition to web-sourced data, we also use videos produced by generative models, resulting in 234 videos comprising a total of 2,382 frames. The primary generative models used are Sora and Kling, with a small portion of the data sourced from HunyuanVideo [20]. The initial prompts were generated using ChatGPT, with the prompt used for generation as follows: + +Please provide 100 unique and detailed bilingual (Chinese and English) prompts, each with an index number, for generating text-to-video scenes that include mirror reflections. The prompts must meet the following requirements: 1. Specify the mirror type and describe the entities in the scene, the overall layout, and their spatial relationship to the mirror. 2. Include a diverse range of mirror types: dressing mirrors, vanity mirrors, full-length mirrors, bathroom mirrors, car rearview mirrors, polished stainless steel, etc. 3. Ensure varied scene distributions: residential settings, commercial spaces, and public areas. 4. The combination of mirror type and scene context must be reasonable (e.g., polished stainless steel is appropriate in a kitchen but not in a study). 5. Entity configuration: some scenes should include people in front of the mirror (e.g., a woman combing her hair or a customer trying on clothes), some should feature objects (e.g., plants, cosmetics, books), and others should show only the mirror reflecting surfaces like walls. 6. Each prompt must describe the physical correspondence between the real object and its reflection. 7. Avoid overly complex layouts in individual scenes. 8. Ensure a balance of richly textured and minimally textured elements within the same scene. 9. All objects in the scene must remain static, with only slow camera panning; descriptions implying motion (e.g., "a moving car") are inappropriate. 10. Descriptions should be as precise and detailed as possible. + +The generated prompts were subsequently refined to avoid producing low-quality video outputs, as pointed in Section A.1. Below are some examples of the prompts: + +Generate a video showing a cozy, modern living room. A single minimalist-designed mirror is mounted on the wall, with clearly defined edges and realistic reflections. The scene combines intricate furniture textures with a monochromatic background, and the camera pans slowly. + +Generate a video set in a creative art space. A uniquely shaped mirror hangs on the wall, featuring accurate reflections and distinct boundaries. The scene includes complex graffiti textures and smooth surfaces, with slow camera panning. + +A static and art-deco inspired living room with a framed mirror above a tufted velvet sofa, reflecting physical laws accurately, geometric patterns, sleek metal finishes, and glamorous lighting. Realistic, glamorous lighting, retro. + +A static and rustic farmhouse dining area with a reclaimed wood-framed mirror on a weathered brick wall, highlighting a crisp realistic reflection, a sturdy wooden table, vintage chairs, and warm pendant lighting. Realistic, warm lighting, rustic. + +Our VLM-driven monocular-stereo fusion framework employs Depthanything V2 [45, 46] as the pre-trained monocular network, QwenV12-7B [40, 1] as the pre-trained visual-language network, and FLUX [26, 31] as the flow matching network. The language prompt for the pre-trained visual-language network is: + +Are there any transparent or reflective objects? Like mirror, glass, window, showcase, and so on? If true, reply to me with the list of corner coordinates of each object in the format of (x1,y1,x2,y2,x3,y3,x4,y4) in the image. If false, reply with an empty list of corners. + +The language prompt for the pre-trained flow matching network is: + +Using the provided features extracted by QwenVL2, generate a binary segmentation mask for the image. Highlight all transparent or reflective objects (e.g., mirrors, glass, windows, showcases) in white (255), while marking all other regions in black (0). + +# C.4 Computational Cost + +The detailed inference metrics, including runtime and memory consumption, are presented in Table 7. All experiments were conducted on a single NVIDIA H100 GPU with an input resolution of $1920 \times 1080$ . The majority of the computational cost arises from the VLM part. + +
ModelMemory UsageInference Time (per iteration)
RAFT-Stereo5610 MB0.87 s/it
DepthAnything V23584 MB0.18 s/it
Ours53959 MB4.77 s/it
+ +Table 7: Comparison of memory usage and inference time across models. + +# C.5 Visualization + +We present more visualization on the 3D-Visual-Illusion dataset and Booster dataset in Figure 12, 13, and 14. The results demonstrate that our method can effectively handle various types of visual illusions. The depth maps generated by our model exhibit high fidelity and accuracy, even in challenging scenarios with complex visual illusions. The depth maps from VGGT and Dust3R mirror the significance of fusing monocular priors and multi-view matching. + +We also present the visualization of 3D detection on the real data of the 3D-Visual-Illusion dataset in Figure 15. We obtain the results from YOLO3D [29], and the results show that 3D visual illusions can seriously affect the performance of 3D detection. We believe that the 3D visual illusion will become more and more important as the vision foundation models become more and more powerful, especially when used in downstream applications, like 3D detection, occupancy and planning. + +![](images/c693ebd781dd8ee75346ea2d0cbfd7988b03d32825c4aa0542e9f6bc82069a47.jpg) + +![](images/5436fba1cbdc98b9d7237ed01cfb1fc868c644a930381dc6245a47c9450eebfc.jpg) +DA v2 DepthPro + +![](images/fa3bfe31065e58043a21404d0785f7278a60cc79da80a659e82a39e65e0965e9.jpg) + +![](images/88956b9b8372c34d4603b1f24b960f506b1b8230092a74961bfa1393662225b8.jpg) +Moge + +![](images/41ba282504a66eef84ab9023a0d25735663960e2b09b0e56cdcc639d15d092de.jpg) +Right Image + +![](images/f5625ad519584cdf184b509a9146042a8387054064721d822668ab286631e38a.jpg) +Dust3R +VGGT + +![](images/3c30a79dfd252d95cc69391cb174217596980dd728c523b7de69c975b7209fe6.jpg) +Metric3d + +![](images/f7196eafcd6a0c895c83f998714ad324c9b340d89a9d76b623084453b73dfa7b.jpg) +Ours + +![](images/dfe22998596f0bf900eb42ae9fe6ccbbbb7c05661b907abef309d4e870a75e57.jpg) +Left Image + +![](images/aac2c1c670b72f675a7c35bd06b33a5124e678c094793f6054036101e4c0aa17.jpg) +DA v2 +DepthPro + +![](images/550b04cc182b1d3135c6e34a5517fff6dcb488dd1e364cd138eb47b42bd24484.jpg) +Marigold + +![](images/f3f74387f2cfca5084446b4f047df8264714d2f256baa4bbf0059cfb7335b707.jpg) +Moge + +![](images/46a5dfe8b46dca829fff11548c296cbd343cb70b984e585d5a735dd3680758ce.jpg) +Right Image + +![](images/4ea848cc1f657fadf4b56af6693c7eeb5df23bf0a9b120f380912704a5b07b03.jpg) +Dust3R +VGGT + +![](images/0550bc7f20992ec45e567ffa1dfb54706537702f5231231fc02cbde57110aa51.jpg) +Metric3d + +![](images/9b002927ed889b2e89d975c660c6ebb94b1837bfff417405d0f7c1a88fd4de68.jpg) +Ours + +![](images/82dcea173cc08b52b1cca702ddea362d91ed3396742e00a486a0233303a74728.jpg) +Left Image + +![](images/828c8d4f0f85779c00b310300d2394b86d4197ea06002216b84367ccc9d23a92.jpg) +DA v2 +DepthPro + +![](images/ab578e9461bf3073a261996a3771cd9dffd57d2df5ddbb7cbb4a9c78c411b798.jpg) +Marigold + +![](images/f90a8947cd363a466d98adccaa367d022d2bf437763baff428610f057c8587b5.jpg) +Moge + +![](images/aad463dd762f422f867deff4deee2bb93460f87f6cd012a58e44d4aa26690bcf.jpg) +Right Image + +![](images/7d6c8d94f8bde59a04a6be443c529674e5273e55a242837be3808a6f75948381.jpg) +Dust3R + +![](images/13e41c7878dcbc3f51bb6100a850cebfb141994a36b2e404f9913903aeb572d7.jpg) +VGGT + +![](images/b8d9841d9ecae2d71c4e33c44dcb5f891bfa01012b31732601331d8e321e5bac.jpg) +Metric3d + +![](images/1118d2f312b1d598670b17d8faa66f9edde1efcc8e88dfec8cd635bdad9e3eca.jpg) +Ours + +![](images/3ca98f8e98eb18362ecd65db4cd3c1f439e7b7be05d4eb73168dd6536c9d84d0.jpg) +Left Image + +![](images/f0ec5ede5114a9b38e3ab4bece2521cf8fed87245b387f096629afb74979def1.jpg) +DA v2 + +![](images/209affbb21e8f371e449ba9d6e6ce4b1a21481292da1790c0fddb8d5e5b69451.jpg) +DepthPro + +![](images/96360560dd4a97185420c5225daceefb32952e350d0dfd6c2af2bb057cde5f1b.jpg) +Marigold + +![](images/e6d945d17f89cfed86b7a7ae919357e4c38d8df636461b9c3257d3cb6719c5c1.jpg) +Moge + +![](images/0cdbe111b0a53bed8e9615729e98e6a642df4a7a6a6ad2fff9e44b21ddec5ced.jpg) +Right Image + +![](images/0078c0401c535edb600684f74473517b39c391ab1cfe19dca5541d3ab7482fc9.jpg) +Dust3R VGGT + +![](images/81989ee7ef043f4007220743ab78138afaaa0200dbfb88e8e1cc836979265e6e.jpg) + +VGGT + +![](images/7467a18dca72a635c8e3cee6bb526af15b219f8cc56972d098b7801279843d6b.jpg) +Metric3d + +![](images/7713f9182eb539f660840994a27a70c5336af9d7e088fa6d756557e9919fc3c2.jpg) +Ours + +![](images/7825b1eeac908c4fd768f6761746c1fc1cea7708149971e5c1242c5cada5e08a.jpg) +Left Image + +![](images/e23cc9188459f05fe49aef70c9c10eb70d86b3f5bd7eeefed95500e95989b617.jpg) +DA v2 + +![](images/b991d9291c4daeaa6dece4fb82b09d0514bef95090448b24da4235fd8c99c609.jpg) +DepthPro + +![](images/911f9323632e2083e378b1610308593b5fdda77c94891ab30a5d62c8c0c367fe.jpg) +Marigold + +![](images/26aa40c4ebfe310dcb1ec9061694fdd5f74ce132f53be5e92fa8a0216664a6f1.jpg) +Moge + +![](images/ced0622283da0064b92f4b06bba741b20f1a8e38d1c1c52b07189e36b5668a62.jpg) +Right Image + +![](images/6bfcc8ea5d1be9513ca4fa9ec67a2cc30bd398c4fefde1c4c16ca63152316810.jpg) +Dust3R VGGT + +![](images/d6298d42af180d5bbacd6e8bb51b7376153d42022856b2bb609e2b95d142c904.jpg) + +VGGT + +![](images/97415925745b615bb420038156a4ae767f52b8107cff7c85f3fa480ad68a68f5.jpg) +Metric3d + +![](images/91647d3800f5ddab425ffd139ebc4bb40db4f11baea5929823c542491bde2423.jpg) +Ours + +![](images/ad0b13fced20d073066d9c1d75e394c9bda61c293c43b8cfc78fd1d1d11dfb69.jpg) +Left Image + +![](images/bde8ea8dc6fe8ffd90910f211ec7ce24fdb3b8fefd0e586b127dec1c5088455d.jpg) +DA v2 + +![](images/5453fe5e5b3ec84652eb64036398323ab2dab1f095f5d3884662fbf9da255d60.jpg) +DepthPro + +![](images/d064441e94b4ac87dbfc0891fc6e3b5c17ee1514260c883484e87508b3f66d01.jpg) +Marigold + +![](images/bec0f3b38bccb2114fb4ed6c35fabd77a2c154271b8dd3f23ce776a61fb8304c.jpg) +Moge + +![](images/3f2f7883956d44279bf2bea5ad6003d7c4f612905a1a4a96e6142fc7ab3bc3f9.jpg) +Right Image + +![](images/ac39bac9688309ca7895a2417cb7d27b29136df0ef8239b822047d775f76953f.jpg) +Dust3R + +![](images/4f8e8f9933818725bfcb6e17585b02ee8c8f0b02752fefed201cd8324d607385.jpg) +VGGT + +![](images/4351399d801ab89671919ea8e8fc19f4ca9ca20de789638652d9a535a4ee2141.jpg) +Metric3d + +![](images/7cd476a7d4e70f87c5db0ce478af359cc12754b94ac26e3453dea56b1cd61ab1.jpg) +Ours + +![](images/bfcaf38874188078984a7bfbe19c5f161c28112f1226c15b6dd920a41534b686.jpg) +Left Image + +![](images/21c2bc35a9f9fac98c429a9ef503ac74c315e66a792c8c4264ba17602c6a6338.jpg) +DAv2 + +![](images/ce6d84856b873e28f1540a2cdd5088a18becfaaa89ca42a716ad75a5600cdd3f.jpg) +DepthPro + +![](images/e5b5a4d6a25983a50a961e94f027862f3d99d1f8ad3f73e7e206cbb6f42782d8.jpg) +Marigold + +![](images/2e27cd72f9ab719677c42f4d1c4e2a5b8fd6dbe8549d9c1576f147573e0de50c.jpg) +Moge + +![](images/fd219638f108a0d21c2892e2a810c4c1ebfe4725adef96e4d2cb266f7d4e8c78.jpg) +Right Image + +![](images/01abae4ac57fabe395bf80dc11a5a825801bf532a307badaa9557c95211056fc.jpg) +Dust3R + +![](images/aebb170e7058792a3977f5440e19548149c41c86007218b2d0f7926b3586d6be.jpg) +VGGT + +![](images/ff677339f1314825114e074c0d8635576c04026840c7919bc05062d9c7db1e5b.jpg) +Metric3d + +![](images/e6da792a30bdcff0f2c62d853d3fe23d4bfe3768c401abe92925fb1c9be8c92a.jpg) +Ours + +![](images/3f805c9070df5a3b1c809332576616cc847972944a4ebac7ef8d9b2716becb01.jpg) +Left Image + +![](images/c10060af44f58da511b1399e04aace8cb18ef045dec32c4e1e3c1679c1221e93.jpg) +DA v2 + +![](images/cccbb130b2ac82160f3a951ef30c382de9e8fce59ecdf365033a17cf7475cace.jpg) +DepthPro + +![](images/c43e10d7ab6dad883e6e15408039f65b90e572436137db655908dba0bc15f908.jpg) +Marigold + +![](images/8ab76ca9593bd98333549e4c39c2df6e6437ca1e53ccb27d2e1178130cb2faa6.jpg) +Moge + +![](images/dc9d02594d4a63b7907e553e9325a813b9985ed7cacf069e8af1a35cef183483.jpg) +Right Image + +![](images/18d697985d4df6755f33c7fa0a5c06acb2625815968a4c6334106ec0e4c6c5ba.jpg) +Dust3R + +![](images/8565ca7fc7b61d63d41c34df90aa180d8be897c6d4aa27199af650a6998fd936.jpg) +VGGT + +![](images/74cf23c0ff339decb120a69ffde65acd79e91fdedd75bccdac2bf9fd2168ab49.jpg) +Metric3d + +![](images/11d98714f0366a45f1829534f952b90c503a2f81065a37478032e05cd1ad2465.jpg) +Ours + +![](images/8a27d3dd451c23c30c8d910cfb44c1a9aed02d331cdc3c90fb48c1adb3e29fdd.jpg) +Left Image + +![](images/1255a73990c5d35ae28e6c0f8383765c78697df08a47e2bae1e94c452d97e5e2.jpg) +DA v2 + +![](images/5dbc7553fa40baace4dc1c4bc14020d2612aba6c67bdc139afa662bebe36d99f.jpg) +DepthPro + +![](images/1a0f4fcc9a049e27b7c3a60546be258d79ebc7f5b7a23d021e83f3da2f248dca.jpg) +Marigold + +![](images/af38228259710d5acad1be5e851a10f51c14ec37765b863a7c35ebe48a07a39c.jpg) + +![](images/992825d0ae16f32b06515cd6f0e26609439a711f015330763b937804749a7e2e.jpg) +Right Image +Figure 12: The visualization of results on virtual data of the 3D-Visual-Illusion dataset. + +![](images/f1e2dcc4c0c1f6b91ad41df6737c5cfc05f35933ee57110a1c7d73d06b8d4113.jpg) +Dust3R + +![](images/6429ccfbd2d0356234c218e3229277af6ac3fdf7c8ca793863b61af0964d7bac.jpg) +VGGT + +![](images/5ea896f3aae3d2d69873dbd20548dab86caa1e16ade78c96471029465563e775.jpg) +Metric3d + +![](images/49c5ecec4ddda9bb9d7f495ee49bd4b3ad21b1fa98e857103ea8e00ec1016306.jpg) +Moge +Ours + +![](images/2695df9284b329e4a7e7a8b736de72f6fa8c37ad93a6c6902f31a9b4e6a4705e.jpg) + +![](images/e83b35e666e95c3bb9f3f1eee3f23dc38740defbcd16cf7f0643ecef1daac3e3.jpg) +GT + +![](images/08b2cb567cfbd800717940c724ea48366c7b071e3d5d6221940d5f38eb38bb34.jpg) + +![](images/f686641345c57183790961c5c0b5809a7f954f0de29da74f72b74a260ac1d8e5.jpg) + +![](images/1de4876d5cd8237cb0af6ee118a65610777a245648b0f6321303ba1b1f761277.jpg) + +![](images/800807cdbe3b5e1a2f2e9f4b9774b7a489bb79e92c02919ef95cc910d782951e.jpg) +Left Image +Right Image + +![](images/d8b6279c15aafb5ebf9dc68313910b37aba58196895dfaf27fd810b79d057613.jpg) +Ours + +![](images/73b07067e636130b0cf71fa7cb0c618f7b409a977739570d2258c633cc8b6483.jpg) +DepthAnything +Dust3R + +![](images/24efb211289157f5efdc373aa03085f20eb71b3e0609e38d81bc9d4752a75595.jpg) +Marigold +VGGT + +![](images/5474cbf8030cf19f0c90edc58ea7fe8b5517a660a582899b1bd05925799a03eb.jpg) +DepthPro +Metric3D + +![](images/444a629bfc9d2770bdef87c50f788b6421003d899cd9403a727b56ade2fae51d.jpg) +Left Image + +![](images/2598a849f90a40e94ce6e2aab435d6abd0c7909be53c1f1a6cad3ec33c39742a.jpg) +GT + +![](images/752676a38c29d34cc19f869bd8fb9de01ee8b67b1bd3e8cbece4ad70554084bb.jpg) +DepthAnything + +![](images/2da65a6c0c03e6f186d89d410f7a3121ecfa7e13c2b6989d18188e4bd1841425.jpg) +Marigold + +![](images/87180c22a9165f6720cb3bf885355269e7e77d7c24c1100adab7a6039bb30c08.jpg) +DepthPro + +![](images/e8d3e643bf6b8066342050e94da6a6b52004d0b3b4472b1ed742c1f15dbfd151.jpg) +Right Image + +![](images/d82111205f23ef31c14462b7a277343683715e0905dbf3666f9526cd5798acb2.jpg) +Ours + +![](images/cacb6f015df7b3f4c4c2344589042fd1531457047463306127c9c484fdf9304a.jpg) +Dust3R + +![](images/d978adfb70cf77f46e673d9c3f6621b55331620af2ea1772b1b654dcbd472012.jpg) +VGGT + +![](images/40dd587c584e0b5211f5ffaf17059051922c4dbb68af97bec2560a4b34a44c43.jpg) +Metric3D + +![](images/670624a593c9e23844b9764496437a4dc3602b17a2b3f55968d5d0052928aa4c.jpg) + +![](images/832c4b6d1a337f3add51a0ed37e85a6d85530fb2ec86298b818ca94c5a2da5a5.jpg) +GT + +![](images/c9fbc2ca4a206a69f8a80d1228f1b68954d6a4553c65177abb32a7a634de90bf.jpg) + +![](images/e87dbe85b067132859df120af65629a3741746007d9ae1c51a75f2386f8b58e8.jpg) + +![](images/66c730839313e26c72740ab269335b92f6bc96f7587130beaa4be413496c2d4c.jpg) + +![](images/ab723e793e4aaaa869a495d109a2df3744b7e711710be6f85e1a9cb38a3a2d07.jpg) +Left Image +Right Image +Figure 13: The visualization of results on real data of the 3D-Visual-Illusion dataset. + +![](images/9c1c8ff0493b70b4ec22e683065e3733bd34693f72f73c2e4b243d2a61193953.jpg) +Ours + +![](images/7a724cd1a7fc04dab975d9eb5091207985c43190e99ffe518f2f10399ec8dd01.jpg) +DepthAnything +Dust3R + +![](images/8da232e2cf7d95c2f9cc186faa4c500ddae8687e3a43721d79db389974e82225.jpg) +Marigold +VGGT + +![](images/8f9842f6d07be9fd61a9c964b869730e42656ba7d34f91ef50317bfef0f029e2.jpg) +DepthPro +Metric3D + +![](images/a738f0a098ea912f6ec6ab4c5f29bb42286c4a32647fa3f76ba73f7235a4619f.jpg) +Left Image + +![](images/be6c0f30f32bb4fe7f831c32ef5c35bd6a237840dd4e4646e53b73b8afbd2657.jpg) +GT + +![](images/37ac8af485daa517b8205e38aba8a42de5ea67995d71f94def0b9625c51d6d2f.jpg) +Mocha + +![](images/20afe927732d7a09d356bb61adbdfdf48f2bf9b67cc4d76d9f5f06ca8854f7a8.jpg) +RAFTStereo + +![](images/29d7461a7bb5e103ca780228923bf0bafda5549f22ad1a14a7bb2390ace8b996.jpg) +Right Image + +![](images/824df7dd7685742edf69055f76125aefa8604aeb0e0024bccb067b1e9ffd1b3a.jpg) +Ours + +![](images/8aba645a47fa671eab1c47cb8e05369f15272dca0fd2ab27e3db0dc60a2eac73.jpg) +Selective-RAFT + +![](images/f04e4ff29e496c9a85d08fc27fffa993beae626479af2993094dfce6d619e1ff.jpg) +Selective-IGEV + +![](images/de8f6cdeba913fc3b83eccb7412b0479e2a64539a12001452d495957b2104c96.jpg) +Left Image + +![](images/acde2e549ecba3ca2c1daef929f90067311b18f9e32fe69c164b35e83d310220.jpg) +GT + +![](images/acd6b020c597e81abd6e23faf65282453203a7be2ea87b4713352cf30e637edc.jpg) +Mocha-Stereo + +![](images/9648e9350447daf1242bfe64040bd26c28eb928dc3204c3807004a7076cd15fb.jpg) + +![](images/d0d3c35ab48ae7d269b788d69acaee0b45ffbc68090e4f2301a27845755393bb.jpg) +Right Image + +![](images/c8f5c28ea2ef70356340fcd85292d71d7dcc81af647080c886b13fd32dffb587.jpg) +Ours + +![](images/857c762b6e5511b511df19bb5fafb05cd89c1da0159a6845952c41ab97e1617c.jpg) +Selective-RAFT + +![](images/a91e50efd8f81bb58be20cfb25e77d786ea7354bcc49fb13ebad2d7cd766252b.jpg) +RAFTStereo +Selective-IGEV + +![](images/a9587516415bece249a42a6551e64392a641786c0a3a218e9aae14b951de048c.jpg) +Left Image + +![](images/8e5fe53726f63856674f1b6bfade88af09960f2a0fcad1905d7fef85144c6f33.jpg) +GT + +![](images/5a2402cea4de9509da0116f80f90a84e360074aa20139ae2c9820a56aad0acd2.jpg) +Mocha-Stereo + +![](images/26a15736d372bf33161dc9779740c5d4a2b95c9b51fd9483a866f9a3143dff39.jpg) +RAFTStereo + +![](images/141cdb0c20d2ea2369d83109393252a75a791850d9a5bf156ba0cdac2f536681.jpg) +Right Image +Figure 14: The visualization of results on the Booster dataset. + +![](images/5fd572b4c44708624ec77953eb73a5e08b1550cf0b2b827aa66446d57ef4d3e2.jpg) +Ours + +![](images/db3235deea4bbc7c576d4c37768cbb1b08b15081c96cca8e4caf6d79a819f1a5.jpg) +Selective-RAFT + +![](images/3c0ca6857206d0ba35165a40d63f6d43deb7e0ad4e221287049f077bd51cff1f.jpg) +Selective-IGEV + +![](images/d4ed04b0376033caee1c0cc34e4e900c573027183605c58c6e3f91c57ae381bf.jpg) + +![](images/f888c340e72b62b097884202ca091b805b7231023b7f3c4b035070db0aa0131d.jpg) + +![](images/35f54c2deb80035ffb22aa2bd275ba4f9f47479f5190368b1c63bae9dc5c1283.jpg) + +![](images/3e3aa9cd9eadd7e492d2cc9d93a9551e7834907fa00c28a8825f2851af7009d4.jpg) + +![](images/86884557f2a42009ee420b86df78d6acdaba6eda6c59cecb35fba90e798d5ffb.jpg) + +![](images/c4c181b788c6ad8c8d7ef2e05352dbec5ce393a77e6d5593ffb884fc689f58fd.jpg) + +![](images/7254483e481942e86731f84ba107c19d17e20cb774639f7326f375c75717ac08.jpg) + +![](images/2b7960015eeb9e13a01fbc9f3c244ae8639a481d8c2062cf954103b70e033998.jpg) + +![](images/606b9f2c8ea25d0e49e1e87058f950ad1dd997059244b6cf8b76f75e1058cfa5.jpg) + +![](images/f1f5f881139f3a865dcadf533791b97bd6b65e53a557600464e88531723909ce.jpg) + +![](images/05c8cf39b31f6b556d5b979d90274407f1b289b06383459bc1e387df89416738.jpg) + +![](images/2de48f9749ac9a1d2743ebadf7e5781ce0b5c74c579ace8eb3c92e03c2b35e12.jpg) + +![](images/fca41624bb1a9033934b25ddbb23541170fbca984938859832743305bbc11cc9.jpg) +Figure 15: The visualization of 3D detection on real data of 3D-Visual-Illusion dataset. + +![](images/9803a350e7d3ec8e63ace085140f0c26ad246ad4ca6d1dc03faf82385119f24d.jpg) + +![](images/6347c908f00bbbb27fe7522398b81dd1e4582e26f4e5bf84c3cbcc8f6f0730f5.jpg) + +# NeurIPS Paper Checklist + +# 1. Claims + +Question: Do the main claims made in the abstract and introduction accurately reflect the paper's contributions and scope? + +Answer: [Yes] + +Justification: The abstract and introduction contain our contributions and scope, including the 3D-Visual-Illusion dataset and the VLM-driven monocular-stereo fusion model. + +Guidelines: + +- The answer NA means that the abstract and introduction do not include the claims made in the paper. +- The abstract and/or introduction should clearly state the claims made, including the contributions made in the paper and important assumptions and limitations. A No or NA answer to this question will not be perceived well by the reviewers. +- The claims made should match theoretical and experimental results, and reflect how much the results can be expected to generalize to other settings. +- It is fine to include aspirational goals as motivation as long as it is clear that these goals are not attained by the paper. + +# 2. Limitations + +Question: Does the paper discuss the limitations of the work performed by the authors? + +Answer: [Yes] + +Justification: We have discussed the limitations in the Conclusion section. + +Guidelines: + +- The answer NA means that the paper has no limitation while the answer No means that the paper has limitations, but those are not discussed in the paper. +- The authors are encouraged to create a separate "Limitations" section in their paper. +- The paper should point out any strong assumptions and how robust the results are to violations of these assumptions (e.g., independence assumptions, noiseless settings, model well-specification, asymptotic approximations only holding locally). The authors should reflect on how these assumptions might be violated in practice and what the implications would be. +- The authors should reflect on the scope of the claims made, e.g., if the approach was only tested on a few datasets or with a few runs. In general, empirical results often depend on implicit assumptions, which should be articulated. +- The authors should reflect on the factors that influence the performance of the approach. For example, a facial recognition algorithm may perform poorly when image resolution is low or images are taken in low lighting. Or a speech-to-text system might not be used reliably to provide closed captions for online lectures because it fails to handle technical jargon. +- The authors should discuss the computational efficiency of the proposed algorithms and how they scale with dataset size. +- If applicable, the authors should discuss possible limitations of their approach to address problems of privacy and fairness. +- While the authors might fear that complete honesty about limitations might be used by reviewers as grounds for rejection, a worse outcome might be that reviewers discover limitations that aren't acknowledged in the paper. The authors should use their best judgment and recognize that individual actions in favor of transparency play an important role in developing norms that preserve the integrity of the community. Reviewers will be specifically instructed to not penalize honesty concerning limitations. + +# 3. Theory assumptions and proofs + +Question: For each theoretical result, does the paper provide the full set of assumptions and a complete (and correct) proof? + +Answer: [NA] + +Justification: Our work is not related to theorems. + +# Guidelines: + +- The answer NA means that the paper does not include theoretical results. +- All the theorems, formulas, and proofs in the paper should be numbered and cross-referenced. +- All assumptions should be clearly stated or referenced in the statement of any theorems. +- The proofs can either appear in the main paper or the supplemental material, but if they appear in the supplemental material, the authors are encouraged to provide a short proof sketch to provide intuition. +- Inversely, any informal proof provided in the core of the paper should be complemented by formal proofs provided in appendix or supplemental material. +- Theorems and Lemmas that the proof relies upon should be properly referenced. + +# 4. Experimental result reproducibility + +Question: Does the paper fully disclose all the information needed to reproduce the main experimental results of the paper to the extent that it affects the main claims and/or conclusions of the paper (regardless of whether the code and data are provided or not)? + +Answer: [Yes] + +Justification: We provide the implementation details including hyperparameter settings, baseline selection and evaluation details. + +# Guidelines: + +- The answer NA means that the paper does not include experiments. +- If the paper includes experiments, a No answer to this question will not be perceived well by the reviewers: Making the paper reproducible is important, regardless of whether the code and data are provided or not. +- If the contribution is a dataset and/or model, the authors should describe the steps taken to make their results reproducible or verifiable. +- Depending on the contribution, reproducibility can be accomplished in various ways. For example, if the contribution is a novel architecture, describing the architecture fully might suffice, or if the contribution is a specific model and empirical evaluation, it may be necessary to either make it possible for others to replicate the model with the same dataset, or provide access to the model. In general, releasing code and data is often one good way to accomplish this, but reproducibility can also be provided via detailed instructions for how to replicate the results, access to a hosted model (e.g., in the case of a large language model), releasing of a model checkpoint, or other means that are appropriate to the research performed. +- While NeurIPS does not require releasing code, the conference does require all submissions to provide some reasonable avenue for reproducibility, which may depend on the nature of the contribution. For example +(a) If the contribution is primarily a new algorithm, the paper should make it clear how to reproduce that algorithm. +(b) If the contribution is primarily a new model architecture, the paper should describe the architecture clearly and fully. +(c) If the contribution is a new model (e.g., a large language model), then there should either be a way to access this model for reproducing the results or a way to reproduce the model (e.g., with an open-source dataset or instructions for how to construct the dataset). +(d) We recognize that reproducibility may be tricky in some cases, in which case authors are welcome to describe the particular way they provide for reproducibility. In the case of closed-source models, it may be that access to the model is limited in some way (e.g., to registered users), but it should be possible for other researchers to have some path to reproducing or verifying the results. + +# 5. Open access to data and code + +Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material? + +# Answer: [No] + +Justification: We do not provide open access to the data and code at this time, but can publish part of them at the rebuttal stage if the reviewers need it. The complete data and code will be published after the paper is accepted. + +# Guidelines: + +- The answer NA means that paper does not include experiments requiring code. +- Please see the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details. +- While we encourage the release of code and data, we understand that this might not be possible, so "No" is an acceptable answer. Papers cannot be rejected simply for not including code, unless this is central to the contribution (e.g., for a new open-source benchmark). +- The instructions should contain the exact command and environment needed to run to reproduce the results. See the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details. +- The authors should provide instructions on data access and preparation, including how to access the raw data, preprocessed data, intermediate data, and generated data, etc. +- The authors should provide scripts to reproduce all experimental results for the new proposed method and baselines. If only a subset of experiments are reproducible, they should state which ones are omitted from the script and why. +- At submission time, to preserve anonymity, the authors should release anonymized versions (if applicable). +- Providing as much information as possible in supplemental material (appended to the paper) is recommended, but including URLs to data and code is permitted. + +# 6. Experimental setting/details + +Question: Does the paper specify all the training and test details (e.g., data splits, hyperparameters, how they were chosen, type of optimizer, etc.) necessary to understand the results? + +# Answer: [Yes] + +Justification: We provide the implementation details including hyperparameter settings, baseline selection and evaluation details + +# Guidelines: + +- The answer NA means that the paper does not include experiments. +- The experimental setting should be presented in the core of the paper to a level of detail that is necessary to appreciate the results and make sense of them. +- The full details can be provided either with the code, in appendix, or as supplemental material. + +# 7. Experiment statistical significance + +Question: Does the paper report error bars suitably and correctly defined or other appropriate information about the statistical significance of the experiments? + +# Answer: [No] + +Justification: We follow existing work in the areas we work in and do not provide statistical significance for fair comparisons. + +# Guidelines: + +- The answer NA means that the paper does not include experiments. +- The authors should answer "Yes" if the results are accompanied by error bars, confidence intervals, or statistical significance tests, at least for the experiments that support the main claims of the paper. +- The factors of variability that the error bars are capturing should be clearly stated (for example, train/test split, initialization, random drawing of some parameter, or overall run with given experimental conditions). + +- The method for calculating the error bars should be explained (closed form formula, call to a library function, bootstrap, etc.) +- The assumptions made should be given (e.g., Normally distributed errors). +- It should be clear whether the error bar is the standard deviation or the standard error of the mean. +- It is OK to report 1-sigma error bars, but one should state it. The authors should preferably report a 2-sigma error bar than state that they have a $96\%$ CI, if the hypothesis of Normality of errors is not verified. +- For asymmetric distributions, the authors should be careful not to show in tables or figures symmetric error bars that would yield results that are out of range (e.g. negative error rates). +- If error bars are reported in tables or plots, The authors should explain in the text how they were calculated and reference the corresponding figures or tables in the text. + +# 8. Experiments compute resources + +Question: For each experiment, does the paper provide sufficient information on the computer resources (type of compute workers, memory, time of execution) needed to reproduce the experiments? + +Answer: [Yes] + +Justification: We provide the computer resources for reproducing the experiments. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The paper should indicate the type of compute workers CPU or GPU, internal cluster, or cloud provider, including relevant memory and storage. +- The paper should provide the amount of compute required for each of the individual experimental runs as well as estimate the total compute. +- The paper should disclose whether the full research project required more compute than the experiments reported in the paper (e.g., preliminary or failed experiments that didn't make it into the paper). + +# 9. Code of ethics + +Question: Does the research conducted in the paper conform, in every respect, with the NeurIPS Code of Ethics https://neurips.cc/public/EthicsGuidelines? + +Answer: [Yes] + +Justification: Our work conforms with the NeurIPS Code of Ethics. + +Guidelines: + +- The answer NA means that the authors have not reviewed the NeurIPS Code of Ethics. +- If the authors answer No, they should explain the special circumstances that require a deviation from the Code of Ethics. +- The authors should make sure to preserve anonymity (e.g., if there is a special consideration due to laws or regulations in their jurisdiction). + +# 10. Broader impacts + +Question: Does the paper discuss both potential positive societal impacts and negative societal impacts of the work performed? + +Answer: [NA] + +Justification: There is no societal impact of our work performed. + +Guidelines: + +- The answer NA means that there is no societal impact of the work performed. +- If the authors answer NA or No, they should explain why their work has no societal impact or why the paper does not address societal impact. +- Examples of negative societal impacts include potential malicious or unintended uses (e.g., disinformation, generating fake profiles, surveillance), fairness considerations (e.g., deployment of technologies that could make decisions that unfairly impact specific groups), privacy considerations, and security considerations. + +- The conference expects that many papers will be foundational research and not tied to particular applications, let alone deployments. However, if there is a direct path to any negative applications, the authors should point it out. For example, it is legitimate to point out that an improvement in the quality of generative models could be used to generate deepfakes for disinformation. On the other hand, it is not needed to point out that a generic algorithm for optimizing neural networks could enable people to train models that generate Deepfakes faster. +- The authors should consider possible harms that could arise when the technology is being used as intended and functioning correctly, harms that could arise when the technology is being used as intended but gives incorrect results, and harms following from (intentional or unintentional) misuse of the technology. +- If there are negative societal impacts, the authors could also discuss possible mitigation strategies (e.g., gated release of models, providing defenses in addition to attacks, mechanisms for monitoring misuse, mechanisms to monitor how a system learns from feedback over time, improving the efficiency and accessibility of ML). + +# 11. Safeguards + +Question: Does the paper describe safeguards that have been put in place for responsible release of data or models that have a high risk for misuse (e.g., pretrained language models, image generators, or scraped datasets)? + +Answer: [Yes] + +Justification: We use search engines to access Internet data, and search engines have their own methods to avoid security safety risks. Moreover, samples in the test set we curated have been reviewed case by case. + +Guidelines: + +- The answer NA means that the paper poses no such risks. +- Released models that have a high risk for misuse or dual-use should be released with necessary safeguards to allow for controlled use of the model, for example by requiring that users adhere to usage guidelines or restrictions to access the model or implementing safety filters. +- Datasets that have been scraped from the Internet could pose safety risks. The authors should describe how they avoided releasing unsafe images. +- We recognize that providing effective safeguards is challenging, and many papers do not require this, but we encourage authors to take this into account and make a best faith effort. + +# 12. Licenses for existing assets + +Question: Are the creators or original owners of assets (e.g., code, data, models), used in the paper, properly credited and are the license and terms of use explicitly mentioned and properly respected? + +Answer: [Yes] + +Justification: We've cited the original paper of the code and model we used. + +Guidelines: + +- The answer NA means that the paper does not use existing assets. +- The authors should cite the original paper that produced the code package or dataset. +- The authors should state which version of the asset is used and, if possible, include a URL. +- The name of the license (e.g., CC-BY 4.0) should be included for each asset. +- For scraped data from a particular source (e.g., website), the copyright and terms of service of that source should be provided. +- If assets are released, the license, copyright information, and terms of use in the package should be provided. For popular datasets, paperswithcode.com/datasets has curated licenses for some datasets. Their licensing guide can help determine the license of a dataset. + +- For existing datasets that are re-packaged, both the original license and the license of the derived asset (if it has changed) should be provided. +- If this information is not available online, the authors are encouraged to reach out to the asset's creators. + +# 13. New assets + +Question: Are new assets introduced in the paper well documented and is the documentation provided alongside the assets? + +Answer: [No] + +Justification: We will provide open access to part of the new assets at the rebuttal stage if the reviewers need it. The complete assets will be published after the paper is accepted. + +Guidelines: + +- The answer NA means that the paper does not release new assets. +- Researchers should communicate the details of the dataset/code/model as part of their submissions via structured templates. This includes details about training, license, limitations, etc. +- The paper should discuss whether and how consent was obtained from people whose asset is used. +- At submission time, remember to anonymize your assets (if applicable). You can either create an anonymized URL or include an anonymized zip file. + +# 14. Crowdsourcing and research with human subjects + +Question: For crowdsourcing experiments and research with human subjects, does the paper include the full text of instructions given to participants and screenshots, if applicable, as well as details about compensation (if any)? + +Answer: [NA] + +Justification: The dataset are built by co-authors. + +Guidelines: + +- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. +- Including this information in the supplemental material is fine, but if the main contribution of the paper involves human subjects, then as much detail as possible should be included in the main paper. +- According to the NeurIPS Code of Ethics, workers involved in data collection, curation, or other labor should be paid at least the minimum wage in the country of the data collector. + +# 15. Institutional review board (IRB) approvals or equivalent for research with human subjects + +Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals (or an equivalent approval/review based on the requirements of your country or institution) were obtained? + +Answer: [NA] + +Justification: The dataset are built by co-authors. + +Guidelines: + +- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. +- Depending on the country in which research is conducted, IRB approval (or equivalent) may be required for any human subjects research. If you obtained IRB approval, you should clearly state this in the paper. +- We recognize that the procedures for this may vary significantly between institutions and locations, and we expect authors to adhere to the NeurIPS Code of Ethics and the guidelines for their institution. + +- For initial submissions, do not include any information that would break anonymity (if applicable), such as the institution conducting the review. + +# 16. Declaration of LLM usage + +Question: Does the paper describe the usage of LLMs if it is an important, original, or non-standard component of the core methods in this research? Note that if the LLM is used only for writing, editing, or formatting purposes and does not impact the core methodology, scientific rigorousness, or originality of the research, declaration is not required. + +Answer: [Yes] + +Justification: We describe the use of LLMs in the Sections 3.1 and 4.2. + +Guidelines: + +- The answer NA means that the core method development in this research does not involve LLMs as any important, original, or non-standard components. +- Please refer to our LLM policy (https://neurips.cc/Conferences/2025/LLM) for what should or should not be described. \ No newline at end of file diff --git a/NeurIPS/2025/3D Visual Illusion Depth Estimation/images.zip b/NeurIPS/2025/3D Visual Illusion Depth Estimation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..83d2c80f83657bc2be5fc5a950ac63e471df920e --- /dev/null +++ b/NeurIPS/2025/3D Visual Illusion Depth Estimation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6499c74e97753f2423ef69c19add7df3d3888bf778d926324f57564c24a7edb2 +size 1765350 diff --git a/NeurIPS/2025/3D Visual Illusion Depth Estimation/layout.json b/NeurIPS/2025/3D Visual Illusion Depth Estimation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..fe5aa280639a6e08ce06710e4c2e3ba541c613f6 --- /dev/null +++ b/NeurIPS/2025/3D Visual Illusion Depth Estimation/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0093c1f22be7aa426864f50332515574d07f0e33f42dec1fa0625c4e1e1d1f50 +size 1635936 diff --git a/NeurIPS/2025/3D-Agent_ A Tri-Modal Multi-Agent Responsive Framework for Comprehensive 3D Object Annotation/b7ddcf35-7b87-4e57-a1da-f034ad401c0a_content_list.json b/NeurIPS/2025/3D-Agent_ A Tri-Modal Multi-Agent Responsive Framework for Comprehensive 3D Object Annotation/b7ddcf35-7b87-4e57-a1da-f034ad401c0a_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..c5b729bc01fa5bcfa88dec401c93a6666d9039a6 --- /dev/null +++ b/NeurIPS/2025/3D-Agent_ A Tri-Modal Multi-Agent Responsive Framework for Comprehensive 3D Object Annotation/b7ddcf35-7b87-4e57-a1da-f034ad401c0a_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2fa5b1cef48a46db88169183d2389e8d16fe74d9420fa220e0a567aefdc4de61 +size 187183 diff --git a/NeurIPS/2025/3D-Agent_ A Tri-Modal Multi-Agent Responsive Framework for Comprehensive 3D Object Annotation/b7ddcf35-7b87-4e57-a1da-f034ad401c0a_model.json b/NeurIPS/2025/3D-Agent_ A Tri-Modal Multi-Agent Responsive Framework for Comprehensive 3D Object Annotation/b7ddcf35-7b87-4e57-a1da-f034ad401c0a_model.json new file mode 100644 index 0000000000000000000000000000000000000000..4e86587cdd85780b9dc396df46b6089fd02403b0 --- /dev/null +++ b/NeurIPS/2025/3D-Agent_ A Tri-Modal Multi-Agent Responsive Framework for Comprehensive 3D Object Annotation/b7ddcf35-7b87-4e57-a1da-f034ad401c0a_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5129725699300765800e9f945c5754c38cf4d74fe39bfedc7e255a2887ab6a4b +size 231480 diff --git a/NeurIPS/2025/3D-Agent_ A Tri-Modal Multi-Agent Responsive Framework for Comprehensive 3D Object Annotation/b7ddcf35-7b87-4e57-a1da-f034ad401c0a_origin.pdf b/NeurIPS/2025/3D-Agent_ A Tri-Modal Multi-Agent Responsive Framework for Comprehensive 3D Object Annotation/b7ddcf35-7b87-4e57-a1da-f034ad401c0a_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..27a03e75043a91ab3d208ca575f547db7f160f8c --- /dev/null +++ b/NeurIPS/2025/3D-Agent_ A Tri-Modal Multi-Agent Responsive Framework for Comprehensive 3D Object Annotation/b7ddcf35-7b87-4e57-a1da-f034ad401c0a_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ed01c2a9d57383723102b819e0f9fafa91d36b3bd9e7409600a450eae54f477c +size 6877630 diff --git a/NeurIPS/2025/3D-Agent_ A Tri-Modal Multi-Agent Responsive Framework for Comprehensive 3D Object Annotation/full.md b/NeurIPS/2025/3D-Agent_ A Tri-Modal Multi-Agent Responsive Framework for Comprehensive 3D Object Annotation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..de7c7bc636df363dd1b76a6ddf9c5c832b232e62 --- /dev/null +++ b/NeurIPS/2025/3D-Agent_ A Tri-Modal Multi-Agent Responsive Framework for Comprehensive 3D Object Annotation/full.md @@ -0,0 +1,779 @@ +# Tri-MARF: A Tri-Modal Multi-Agent Responsive Framework for Comprehensive 3D Object Annotation + +Jusheng Zhang $^{1}$ , Yijia Fan $^{1}$ , Zimo Wen $^{2}$ , Jian Wang $^{3}$ , Keze Wang $^{1,\dagger}$ + +$^{1}$ Sun Yat-sen University + $^{2}$ Shanghai Jiao Tong University + $^{3}$ Snap Inc. +† Corresponding author: kezewang@gmail.com + +# Abstract + +Driven by the applications in autonomous driving, robotics, and augmented reality, 3D object annotation is a critical task compared to 2D annotation, such as spatial complexity, occlusion, and viewpoint inconsistency. The existing methods relying on single models often struggle with these issues. In this paper, we introduce Tri-MARF, a novel framework that integrates tri-modal inputs (i.e., 2D multi-view images, text descriptions, and 3D point clouds) with multi-agent collaboration to enhance the 3D annotation process. Our Tri-MARF consists of three specialized agents: a vision-language model agent that generates multi-view descriptions, an information aggregation agent that selects optimal descriptions, and a gating agent that aligns text descriptions with 3D geometries for more refined captioning. Extensive experiments on the Objaverse-LVIS, Objaverse-XL, and ABO datasets demonstrate the superiority of our Tri-MARF, which achieves a CLIPScore of 88.7 (compared to 78.6-82.4 for other SOTA methods), retrieval accuracy of 45.2/43.8 (ViLT R@5), and an impressive throughput of 12,000 objects per hour on a single NVIDIA A100 GPU. + +# 1 Introduction + +3D object annotation is crucial in computer vision, providing semantic labels for 3D data across autonomous driving [30, 5, 28, 24, 17, 40, 80, 61], robotics, and AR applications. The existing methods primarily use single large-scale models without leveraging multi-agent collaboration, creating challenges with complex scenes [26, 17, 20, 75, 74, 52, 72]. Unlike 2D annotation, 3D annotation faces unique difficulties: increased spatial relationship complexity, occlusion issues [13, 1, 32, 71], and viewpoint variations affecting cross-view consistency [45]. Traditional 3D annotation methods face serious problems with multi-view data: single models struggle to handle viewpoint differences, geometric complexity, and semantic consistency simultaneously. When objects are partially occluded or only visible from specific angles, existing methods often generate incomplete or inconsistent annotations. Existing approaches, which typically rely on vision-language models, often encounter hallucination problems and description inconsistencies [77, 58, 68]. The existing methods struggle to maintain perspective consistency when using multi-view information and often overlook inherent geometric information by relying solely on 2D images [42, 14]. + +After performing deep and comprehensive analysis of these challenges, we find that it is vital to overcome the inherent difficulty of a single decision-making system by optimizing multiple competing objectives simultaneously, i.e., accuracy, completeness, consistency [64, 46, 41], and efficiency [18]. In complex 3D annotation tasks, a single model often struggles to balance these goals effectively [62], akin to a solitary expert lacking proficiency across all domains. These formidable challenges inspire a key question: how can we design a system that collaborates like a team of human experts, + +![](images/dc712329ece8b1c0e52e7d9310428b2295907e1b4db910e4bd6bf26999ff40f8.jpg) +Figure 1: Comparison example of our Tri-MARF captions with previous SOTA methods. Our Tri-MARF not only accurately recognizes the specific names of objects, but also provides rich and correct details. Some keywords in the annotations are shown in red, and the specific names of the objects are shown in orange. Please note that only our Tri-MARF can mark them out. + +![](images/92540af4ade5a93d583f0c1bdc387e14fac4e10415d8cd1f9fcea5807eac8296.jpg) + +![](images/3180a68f8f145f128e6aca2fe2a89db634455239e7989c73fcc3b27b3b86c3c5.jpg) + +![](images/dbb8469069fb27db314b61c2d11fd8919d3ccd4919ab3e9de6ac6fc3de370f00.jpg) + +![](images/4af360a99f33c1ee964cd0a372ac98415d3a492795906b427254eddf449a8442.jpg) + +![](images/22c6807199fdfdb1e151aaf83b7cef37b521c856c5025f831918683de43eaab4.jpg) + +Cap3d + +![](images/bcb96ca1592e98de9246bc62df6525fb2f71a3cdac8e70b139ec345fef924437.jpg) + +A silver-gray streamlined sports car. + +A black dragon, the whole is in pixel style. + +golden state warriors nba 2k17 + +Scoreagg + +![](images/49ba55c9b84be62c1de526a3cbc0f0689947100f346b97ed0e22336dd1418b03.jpg) + +The streamlined sports car has a silver-gray metal body, a black lower spoiler, cool scissor doors on the side and dark windows, and a low and wide rear, which highlights the futuristic and sporty feel as a whole. + +A black mechanical dragon made of metal, with sharp claws and purple eyes. The overall design is full of futuristic technology and may be a figure. + +The image shows a basketball player in a blue and yellow jersey, preparing to dribble the ball. The jersey has the words "Golden State Warriors" and the number 30 printed on it, indicating that he is a member of the Golden State Warriors. + +Tri-MARF + +This sleek, modern sports car, specifically a Lamborghini, rendered in high-gloss metal and carbon fiber, featuring sharp angles, a low profile, and large black alloy wheels with red brake calipers. Designed for high-speed performance, it is showcased against a plain white background, likely for promotional or luxury display purposes. + +This 3D model is a model of an Ender Dragon in Minecraft with a central rectangular structure and symmetrical wing-like fins. Its industrial design includes three vertical columns capped with square elements, suggesting a decorative or structural utility. Set against a light background, it appears intended for exhibition or thematic displays. + +This digital 3D model depicts Stephen Curry of the Golden State Warriors in action. He wears a blue and yellow Warriors uniform with the number 30, white knee pads, and turquoise shoes, holding an orange basketball with black lines. The artwork uses digital textures to mimic fabric, rubber, and skin, set against a neutral background for promotional or entertainment use. + +addressing the task from multiple specialized perspectives [49, 55, 65]. Motivated by this, multi-agent systems offer a natural framework by decomposing complex tasks into specialized sub-tasks, allowing distinct agents to leverage their respective strengths. For 3D object annotation, this implies deploying dedicated agents to handle geometric feature recognition, semantic understanding, and cross-view consistency separately. However, a central challenge in multi-agent systems lies in effectively coordinating the decisions of diverse agents [31, 67], particularly under uncertainty and conflicting information. To address this coordination problem, reinforcement learning emerges as an ideal solution. By dynamically learning optimal policies, reinforcement learning techniques [43, 44] enable the system to continuously refine and optimize collaborative decision-making, surpassing the limitations of predefined rules. Integrating multi-agent systems with reinforcement learning offers numerous advantages [50, 35, 57, 69], including robustness, adaptability, and enhanced performance in tackling complex problems [36, 78]. However, this introduces new difficulty, e.g., designing appropriate reward signals to evaluate annotation quality and embedding reinforcement learning seamlessly into the workflow. + +To address this issue, we propose a novel annotation framework called Tri-MARF (Tri-Modal Multi-Agent Response Framework). The core idea is to adopt a multi-stage pipeline where specialized agents handle each task, with reinforcement learning incorporated to enhance decision-making, particularly in the critical text aggregation phase. As shown in Figure 2, our Tri-MARF consists of four stages to progressively refine and integrate annotation information: Data Preparation Stage: 3D objects are sourced from datasets like Objaverse [12, 11], generating multi-view 2D images and point cloud features to capture structural details; Initial VLM Annotation Stage: A vision-language model (VLM) agent generates preliminary descriptions for each viewpoint. To ensure accuracy, we employ multi-round Q&A with Qwen2.5 [39]. For each view, five candidate responses are produced, and a RoBERT [23] model clusters these using DBSCAN to yield the text description for that perspective; Reinforcement Learning-Based Information Aggregation Stage: We introduce an agent based on a multi-armed bandit [48] with an upper confidence bound algorithm to aggregate candidate descriptions from different views into a coherent, high-confidence global description. This agent models multi-view annotation as a multi-armed bandit problem, where each description candidate serves as an "arm" and the system dynamically learns to select optimal descriptions through exploration-exploitation balance. Unlike static rules or simple voting, our agent adapts to different object types and viewpoint scenarios. This agent learns to balance visual consistency, geometric accuracy, and semantic richness through reward functions. This agent is trained using a composite reward function incorporating VLM confidence scores and CLIP [40] similarity between images and generated captions to dynamically balance exploration and exploitation and optimizing cross-view description consistency through continuous learning; Gating Stage: Cosine similarity between aggregated text and 3D point cloud is computed via an encoder [53], with a threshold determining + +![](images/d2832fecd447ea2ac87b51a2ee180d28f55e7fe6743fe591460f46ad177fe527.jpg) +Figure 2: The illustration of our Tri-MARF for 3D object annotation, featuring a collaborative multi-agent mechanism. The process starts with Agent 1 (VLM Annotation Agent), which uses a visual language model (e.g., Qwen2.5-VL-72B-Instruct) to generate 5 text descriptions for each view of a 3D object from six standard viewpoints (front, back, left, right, top, bottom). These descriptions are then processed by Agent 2 (Information Aggregation Agent), which uses RoBERTa+DBSCAN for semantic embedding clustering, CLIP for visual-text alignment, and integrates a multi-armed bandit (MAB) model to optimize description selection and balance exploration and exploitation to obtain the final captions. Agent 3 (Point Cloud Gating Agent) uses threshold control to align text and 3D point clouds, further reducing the wrong results produced by VLM annotations. Please note that our point cloud is a pre-rendered asset. + +further annotation needs. Adaptive agent outperforms rule-based methods by dynamically adapting to scenes and learning from errors to optimize 3D annotation quality (Figure 1). + +This mainpaper introduces Tri-MARF, a novel multi-stage annotation framework that integrates specialized agents to tackle inconsistencies in 3D object annotation. By leveraging multi-agent collaboration with reinforcement learning, Tri-MARF enhances decision-making and ensures annotation consistency, offering a robust and adaptive solution. Extensive experiments on benchmarks like Objaverse-LVIS, Objverse-XL, and ABO demonstrate Tri-MARF's superior performance, surpassing existing methods in annotation accuracy, description consistency, and linguistic quality. To date, it has annotated approximately 2 million 3D models. + +# 2 Related Works + +Neural 3D Object Annotation has evolved from manual labeling to automated approaches. Chang et al. [7] established ShapeNet, Mo et al. [29] developed PartNet, Yi et al. [63] focused on semantic segmentation, and Savva et al. [38, 60] contributed SHREC benchmarks. Recent vision-language models transformed this landscape: ULIP [58] bridged 3D point clouds and text, PointCLIP [77] adapted CLIP for point clouds, while [45] introduced cross-modal embeddings and Zeng et al. [51] explored prompt engineering. Cap3D [26] pioneered synthetic-to-real transfer but struggled with cross-view consistency. 3D-LLM [17] attempts to solve this via specialized objectives, while [32] developed a complex scene understanding method. + +Multi-Agent Systems for Visual Understanding decompose complex visual tasks into manageable subtasks to enhance efficiency and accuracy in visual understanding. Maes et al. [27] established the foundational concept of collaborating agents, Zhong et al. [79] demonstrated that specialists outperform monolithic models, and Deng et al. [13] introduced multi-agent approaches to 3D scene understanding. Deng et al. [47] and Zhong et al. [79] showed improved robustness through viewpoint integration, while Aghasian et al. [1] developed hierarchical protocols for agent collaboration, and Chafii et al. [4] implemented emergent communication between agents. For scene comprehension, Wei et al. [54] proposed 3D scene graph generation frameworks, Johnson et al. [19]'s DenseCap system demonstrates the importance of specialized roles for image annotation, and Cai et al. [3] and Liu et al. [22] explored dynamic agent routing to enhance system adaptability. Unlike these fixed-protocol systems, our Tri-MARF introduces a tri-modal approach with multi-agent collaboration to optimize collaboration, adaptively weighting information across different modalities to maintain high performance across varying scene complexities. + +Reinforcement Learning for Decision Making in Vision Systems has employed reinforcement learning to optimize viewpoint selection in 3D environments [30], complemented by advances in exploration strategies that improve training efficiency for visual perception tasks [66]. Recently, multi-armed bandit algorithms [48], including upper confidence bound strategies that balance exploration and exploitation [2], have been well researched. These approaches have proven effective for content selection in visual applications [81] and have been extended to contextual settings where visual representations inform decision-making [8]. Multi-modal fusion research has advanced through adaptive weighting mechanisms based on reinforcement learning principles [16], particularly in vision-language tasks where modality importance varies contextually [15, 70]. Recent advances [73, 76] demonstrate the ongoing evolution of these approaches for complex decision-making tasks. In contrast, our Tri-MARF implements a simple yet effective architecture with specialized agents for 3D annotation tasks, demonstrating superior adaptability across diverse object categories and viewpoint conditions. + +# 3 Methodology + +Our Tri-MARF framework addresses key challenges in 3D annotation through three specialized agents in the following four-stage annotation pipeline. Formally, suppose $V = \{ \text{front}, \text{back}, \text{left}, \text{right}, \text{top}, \text{bottom} \}$ represent the standardized viewpoints. Data Preparation Stage generates six corresponding images $\{ I_v : v \in V \}$ for each object, which are encoded and passed to the vision-language model without manual feature engineering. For each viewpoint $v$ , the Initial VLM Annotation Stage produces: $\\( D_v = \{ C_{v,i} \}_{i=1}^M, \$ a \) a set of $M$ candidate descriptions (Tri-MARF employs $M = 5$ with temperature-controlled sampling) each with an associated confidence score. Information Aggregation Stage transforms these texts into BERT embedding space for semantic clustering, while CLIP evaluates each description's visual alignment with image $I_v$ . This dual-evaluation process produces scored response pairs $(C_{v,i}, s_{v,i})$ , where $s_{v,i}$ represents a composite score of semantic distinctiveness and vision-text correlation. In Information Aggregation Stage, we frames candidate descriptions as arms in a multi-armed bandit problem, selecting one description $\hat{C}_v$ per view through UCB (Upper Confidence Bound) exploration-exploitation balancing. The information aggregation agent continuously updates reward estimates to favor higher-quality descriptions as it learns. Finally, we fuse the optimized view-specific descriptions $\hat{C}_v : v \in V$ into a coherent global annotation, with weighted emphasis on informative perspectives. Gating Stage input the description obtained in the previous step into the pretrained text encoder for encoding. At the same time, we input the 3D point cloud of the model obtained in the preparation stage into the pre-trained point cloud encoder for encoding. Then, we calculate the cosine similarity between the encoded text and the point cloud to determine whether the cosine similarity is greater than the empirical threshold $\alpha = 0.557$ (Please refer to Supp. 11.6). If it is greater than this threshold, the corresponding sample is retained; if it is less than this threshold, the sample is marked as a questionable sample for manual annotation. + +# 3.1 Initial VLM Annotation Stage + +For each view image $I_{v}$ , Tri-MARF employs a sophisticated vision-language model agent that uses an innovative multi-turn prompting strategy. Unlike conventional single-prompt approaches, we implement a structured dialogue system with Qwen2.5-VL-72B-Instruct[39] that mirrors expert visual analysis. Our prompting protocol unfolds in three strategic phases: Viewpoint-aware identification. We orient the model to recognize its viewing perspective (e.g., "This is the front view. What object do you see and what is its specific name?"), ensuring attention focuses on viewpoint-specific diagnostic cues. Systematic attribute elicitation. We use targeted follow-up prompts to elicit key attributes such as color, material, and structural components, guaranteeing sufficient feature coverage even under complex viewpoints. Contextual integration. The extracted observations are integrated into consistent, coherent descriptions that preserve viewpoint alignment and emphasize distinguishing characteristics. This transforms annotation quality by decomposing complex visual reasoning into manageable sub-tasks, yielding significantly more detailed and accurate descriptions than conventional single-prompt methods. To maximize semantic coverage and reduce annotation bias, we empirically introduce stochastic diversity sampling at temperature = 0.7 with $M = 5$ descriptions per view, generating alternative interpretations that capture different object aspects. Each description $C_{v,i}$ retains token-level log-probabilities, enabling sophisticated confidence assessment. + +Confidence Score Computation. We introduce a novel probabilistic confidence metric for each description, addressing the critical challenge of uncertainty quantification in 3D annotation. The confidence score $\mathrm{Conf}(C)$ quantifies semantic reliability through average token log-likelihood, providing a principled measure of model certainty. For tokens $t_1, t_2, \ldots, t_N$ with conditional probabilities $P(t_i \mid \text{context})$ , we compute: + +$$ +\operatorname {C o n f} (C) = \frac {1}{N} \sum_ {i = 1} ^ {N} | \log P \left(t _ {i} \mid \text {c o n t e x t u p t o} t _ {i}\right) |. \tag {1} +$$ + +This formulation captures the model's internal uncertainty during generation—lower $\operatorname{Conf}(C)$ values indicate higher confidence (higher token probabilities), while elevated scores signal potential unreliability (perhaps from rare descriptors or uncertain attributions). This confidence metric serves dual purposes in our reinforcement learning pipeline: flagging potentially hallucinated content for rejection, and informing bandit-based selection between semantically similar candidate descriptions. This probabilistic approach to confidence assessment represents a significant advancement over deterministic methods, enabling more reliable annotation in ambiguous situations. + +Importance of Multi-View Inputs. Our Tri-MARF innovatively processes six standard views of each 3D object, addressing the fundamental challenge of viewpoint inconsistency that plagues single-view methods. Unlike traditional approaches that rely on limited perspectives, our Tri-MARF captures comprehensive spatial relationships through front, back, left, right, top, and bottom views. This deliberate design tackles the inherent difficulty of 3D annotation—objects often conceal critical features from any single viewpoint. For instance, a vehicle's diagnostic features distribute across multiple angles: brand identifiers on the front, distinctive lighting arrays at the rear, profile silhouettes from sides, and functional components from top/bottom perspectives. Our Tri-MARF approach systematically mitigates occlusion problems by exploiting complementary information across perspectives, creating an integrated understanding impossible with conventional methods. This redundancy provides crucial resilience: when noise or occlusion compromises one viewpoint, alternative angles maintain annotation integrity. Furthermore, our Tri-MARF supports cross-view verification, confirming the existence of features across multiple viewpoints and reducing the inconsistency problem prevalent in standard VLM methods. This comprehensive spatial coverage forms the foundation for Tri-MARF's exceptional accuracy and completeness. + +# 3.2 Information Aggregation Stage + +After obtaining multiple description candidates per view, our Tri-MARF employs the aggregation agent to perform semantic clustering to eliminate redundancy and then relevance weighting to evaluate each description $C_{v,1}, \ldots, C_{v,M}$ . These steps transform a raw list of $M$ descriptions into a smaller set of unique, scored responses, setting the stage for the final selection. To identify when different generated sentences are essentially saying the same thing, we project each candidate description into a high-dimensional semantic space using a pre-trained language model (BERT). Let $C_{v,i}$ and $C_{v,j}$ be two candidate descriptions for the same view $v$ . We compute their embeddings $E_{v,i} = \mathrm{BERT}(C_{v,i})$ , $E_{v,j} = \mathrm{BERT}(C_{v,j})$ as fixed-length vectors. The semantic similarity between the two descriptions is cosine similarity of their embedding vectors: + +$$ +S _ {i j} = \cos \left(E _ {v, i}, E _ {v, j}\right) = \frac {E _ {v , i} \cdot E _ {v , j}}{\| E _ {v , i} \| \| E _ {v , j} \|.} \tag {2} +$$ + +performing this step, Tri-MARF condenses the candidate descriptions, removing duplicative entries and preparing a canonical description for each distinct idea. From each cluster, we select a representative description. Ideally, all descriptions in one cluster are paraphrases, so any could serve as the cluster's exemplar. Tri-MARF chooses the highest-scoring description in each cluster as the canonical representative. At this stage, we have not yet defined the score that comes next with the CLIP weighting, but once scores are assigned, we compute: $C^{(k)}$ canonical := arg max $C_{v,i} \in \mathcal{C}ksv, i$ , which produces a set of unique descriptions for the view one per cluster. These are the candidate descriptions for competing in the final selection. + +# 3.2.1 Relevance Weighting + +Next, our Tri-MARF evaluates how well each candidate description is grounded in the actual image using CLIP[40]. CLIP provides a function that maps an image $I$ and a text $T$ into a shared feature space such that related image-text pairs have high cosine similarity. We obtain an image embedding + +$\mathbf{I}_v = f_{\mathrm{CLIP}}^{\mathrm{img}}(I_v)$ for the view $v$ and a text embedding $\mathbf{T}_{v,i} = f_{\mathrm{CLIP}}^{\mathrm{text}}(C_{v,i})$ . The relevance of description $C_{v,i}$ to image $I_{v}$ is: $\cos \theta_{v,i}; = ;\frac{\mathbf{I}_v\cdot\mathbf{T}_v,i}{|\mathbf{I}_v|;|\mathbf{T}_v,i|!}$ . We convert the raw similarity into a probabilistic weight via a softmax over the $M$ descriptions of that view: + +$$ +w _ {v, i} = \frac {\exp (\cos \theta_ {v , i})}{\sum_ {k = 1} ^ {M} \exp (\cos \theta_ {v , k})}. \tag {3} +$$ + +The candidate descriptions that align better with the image are assigned a higher weight. Then, our Tri-MARF combines the semantic clustering information and the CLIP visual alignment into a single score for each description. Let $S_{\mathrm{conf},i}$ denote a confidence score for candidate description $i$ , and let $w_{i}$ be the CLIP-based weight after softmax normalization. The final weighted score is: $s_i := (1 - \alpha) \cdot S_{\mathrm{conf},i} + \alpha \cdot w_i$ , where $\alpha \in [0,1]$ controls the balance between text similarity and image-text similarity. In Tri-MARF, a smaller $\alpha$ prioritizes text-based confidence, while a larger $\alpha$ emphasizes visual-semantic alignment, combining signals for responses with both textual relevance and visual correctness. + +# 3.2.2 Multi-Armed Bandit-Based Response Aggregation + +Even after clustering and scoring, there may be multiple plausible descriptions for a given view. Rather than arbitrarily picking the top one, the information aggregation agent of our Tri-MARF uses a multi-armed bandit (MAB) model to adaptively select the best description over time, especially when feedback signals are available. Define the set of arms $\mathcal{A} = a_1, a_2, \ldots, a_K$ corresponding to the $K$ canonical descriptions for the current view. When Tri-MARF chooses arm $a_k$ (i.e., uses description $C_{\text{canonical}}^{(k)}$ as the annotation), it receives a reward $r_k$ that reflects the annotation quality. Over many trials, the goal is to maximize: $\max_{\pi}; \mathbb{E}\left[\sum_{t=1}^{T} r_{a_t}\right]$ . We assume each arm $a_k$ has an underlying expected reward $\mu_k$ . The challenge is the classic exploration-exploitation trade-off: the algorithm should try different arms to learn their rewards but also exploit the best one found so far. Tri-MARF employs the UCB1 variant, which calculates for each arm $a$ : $a_t; =$ ; $\arg \max_{a \in \mathcal{A}} \left(\hat{r}_a; +; c, \sqrt{\frac{2\ln t}{n_a}}\right)$ , where $\hat{r}_a$ is the empirical mean reward, $n_a$ is how many times arm $a$ has been chosen, $t$ is the current round, and $c$ is an exploration weight. This rule formalizes "optimism in the face of uncertainty," ensuring that arms with high potential or insufficient exploration are tried sufficiently. When a reward $r_a$ is observed, the empirical mean is updated by $\hat{r}_a \gets \frac{(n_a - 1), \hat{r}_a + r_a}{n_a}$ . Over time, Tri-MARF converges to favoring the arm with the highest true reward. We chose UCB due to its simplicity, strong regret bounds, and easy interpretability. + +Cross-View Processing and Global Description Synthesis. Once the final descriptions of the individual views are selected, we ensure consistency and fuse them into a global 3D object annotation.Front/Back View Prioritization. Front and back views receive priority, assuming they carry critical identifying information about category and appearance. We assign higher weight $w_{FB}$ to these descriptions: $\mathrm{Priority}(C_{FB}) = w_{FB} \cdot \mathrm{Score}(C_{FB})$ . In practice, front/back descriptions identify the object while other views (side, top, bottom) provide supplementary details. Core Description Ex + +![](images/12f6bf75552f5750a1ea052d7755a085d27670e47f45c1eea8d3be53a55b464f.jpg) +Figure 3: Detailed demonstration of the gating agent of our Tri-MARF. The pre-trained Uni3d encoder is used to handle point cloud and text matching on the open domain. + +traction. From the front/back combined description $C_{FB}$ , Tri-MARF extracts only the first sentence as the core identification sentence: $S_{\mathrm{core}} = \mathrm{First\_Sentence}(C_{FB})$ . For other views, a compiled description $C_{\mathrm{other}}$ is formed by selecting the best or longest candidate from side/top/bottom views. Global Description Assembly. The final global description is then: $C_{\mathrm{global}} = S_{\mathrm{core}} + C_{\mathrm{other}}$ . A scoring formula like Scoreglobal = $\frac{\mathrm{Score}(CFB) + \mathrm{Score}(C_{\mathrm{other}})}{2}$ evaluates how well the merged description aligns with both identity and detailed attributes. + +Table 1: The comparison of caption quality and efficiency for 3D object annotation. The highest value of each metric is in bold. The two values of ViLT R@5 (e.g. 45.2/43.8) represent the retrieval performance of Image-to-Text (I2T) and Text-to-Image (T2I) respectively. + +
MethodObjaverse-LVIS (1k)Objaverse-XL (5k)ABO (6.4k)Speed (objects/luor)
A/B ScoreCLIPScoreViLT R@5A/B ScoreCLIPScoreViLT R@5A/B ScoreCLIPScoreViLT R@5
Human Annotation2.382.440.0 / 38.52.981.037.0 / 35.52.978.933.8 / 32.50.12k
Our Tri-MARF-88.745.2 / 43.8-86.140.5 / 38.9-82.337.1 / 35.612k
Cap3D3.378.635.2 / 33.43.576.432.1 / 30.53.574.828.9 / 27.38k
ScoreAgg3.980.137.8 / 36.03.778.534.5 / 33.04.276.231.2 / 30.09k
3D-LLM3.277.434.9 / 33.33.475.631.8 / 30.33.373.028.4 / 26.96.5k
PointCLIP2.065.322.4 / 20.82.363.119.5 / 18.02.260.717.2 / 15.75k
ULIP-23.075.233.1 / 31.53.273.829.7 / 28.23.171.426.5 / 25.07k
GPT4Point1.862.918.7 / 17.12.060.516.3 / 14.81.958.214.6 / 13.14k
Metadata1.565.220.1 / 18.7---2.161.516.3 / 15.0-
+ +# 3.3 Gating Stage + +To mitigate limitations of traditional 2D image annotation in discriminating geometric properties, we introduce a similarity gating agent based on point cloud-text alignment, as shown in Figure 3. We employ pre-trained encoders $\mathbf{E}_p$ (3D point cloud) and $\mathbf{E}_t$ (text) to extract geometric and semantic features respectively, both in $\mathbb{R}^d$ dimension. Cross-modal matching is quantified using: Cosine Similarity $= \frac{\mathbf{E}_p\cdot\mathbf{E}_t}{|\mathbf{E}_p|_2|\mathbf{E}_t|_2}$ where $\cdot$ denotes dot product and $|\cdot |_{2}$ represents L2 norm. Based on validation grid search, we set dynamic threshold $\alpha = 0.577$ as confidence criterion. When similarity falls below threshold, the gating agent triggers dual-check: critical category samples undergo manual review while redundant samples are filtered out. This geometric-semantic consistency gating effectively suppresses annotation hallucination in visual language models and better leverages intrinsic 3D object information. + +# 4 Experiments + +We rigorously evaluate our Tri-MARF across four experiments to validate its 3D understanding capabilities: (1) Caption quality is assessed on Objaverse-LVIS [12], Objaverse-XL [11], and ABO [9] via A/B testing against human annotations and automated metrics (CLIPScore, ViLT [21] retrieval). (2) Type inference accuracy on Objaverse-LVIS is compared against CAP3D [25], ScoreAgg [20], and Human Annotation using GPT-4o [33] scoring and human validation. (3) The effect of selecting different numbers of viewpoints on the annotation quality is used to justify why we choose 6 viewpoints instead of 8 viewpoints in the previous work [80]. (4) We also conducted annotation experiments on clean 3D point cloud datasets and real-world noisy datasets, demonstrating that our Tri-MARF has high generalization performance. Please refer to Supp. 14 for more detailed experimental settings. + +# 4.1 3D Captioning Test + +Experimental Setup. We evaluate the caption quality of our Tri-MARF for 3D object annotation on Objaverse-LVIS (1k sampled), Objaverse-XL (5k sampled), and ABO (6.4k objects). The captions of our Tri-MARF are compared with those from Cap3D, ScoreAgg, ULIP-2 [59], PointCLIP [77], 3D-LLM [17], GPT4Point [37], human annotations, and metadata in terms of quality and efficiency. Random sampling ensures representativeness. Quality is measured via A/B testing (1-5 scale), CLIPScore, and ViLT retrieval. Note that the speed here is estimated according to the overall rate. Baseline models follow official configurations. Tri-MARF's detailed settings are in Supp. 6. We also explored the model's ability to understand scenes (See Supp. 9) Experimental Results and Analyses. As shown in Table 1, our Tri-MARF achieves state-of-the-art performance across all semantic alignment metrics while maintaining the highest annotation throughput (12k objects/hour) on a single NVIDIA A100 GPU. By design, Tri-MARF serves as the reference baseline for human preference evaluation (A/B scores), implicitly outperforming all methods through pairwise comparisons. Tri-MARF dominates CLIPScore (88.7 vs. 78.6-82.4 on Objverse-LVIS) and ViLT R@5 (45.2/43.8 vs. 35.2-40.0), demonstrating superior cross-modal alignment. Notably, Tri-MARF-generated captions surpass human annotations in semantic precision (CLIPScore +6.3 on ABO) while avoiding human annotators' preference bias. Please refer to Supp. 12 for more details. Tri-MARF excels in 3D object captioning, with higher CLIPScore and ViLT retrieval accuracy showing effective feature capture by the multi-agent approach. High A/B test scores confirm the reinforcement learning-based aggregated + +Table 2: Cross-dataset generalization experimental results comparison. The highest values are highlighted in bold. + +
MethodShapeNet-CoreScanNetModelNet40
CLIP↑ViLT R@5↑GPT-4↑CLIP↑ViLT R@5↑GPT-4↑CLIP↑ViLT R@5↑GPT-4↑
Human Annotation81.737.8 / 36.04.279.534.8 / 33.24.380.236.0 / 34.54.0
Our Tri-MARF83.238.6 / 36.84.380.335.2 / 33.74.081.536.7 / 35.24.2
Cap3D76.533.1 / 31.53.673.229.8 / 28.13.274.331.2 / 29.83.4
ScoreAgg79.135.4 / 33.93.975.632.1 / 30.43.577.233.8 / 32.33.7
3D-LLM75.832.5 / 30.93.572.529.1 / 27.63.173.630.7 / 29.23.3
PointCLIP63.421.7 / 20.22.360.819.3 / 17.92.162.120.5 / 19.12.2
ULIP-273.731.4 / 29.83.370.627.8 / 26.32.972.329.4 / 28.03.1
GPT4Point61.219.5 / 18.12.158.717.4 / 16.01.960.318.9 / 17.52.0
+ +agent aligns with human description habits. Faster runtime underscores the lightweight MAB strategy, ensuring efficiency for large-scale 3D dataset annotation. + +# 4.2 Type Annotation Experiment + +To evaluate the performance of various classification methods on the Objaverse-LVIS dataset, we compare Tri-MARF and ScoreAgg against CAP3D and manual annotation. The comparison focuses on two primary evaluation metrics: (1) the accuracy of string matching between predicted and ground-truth labels, and (2) the semantic accuracy assessed by GPT-4o after comparing model-generated subtitles with standard answers to account for potential synonym mismatches. This experimental design ensures a comprehensive evaluation of both syntactic and semantic alignment. We also compare caption results with a single-agent baseline using the same input and architecture(See in 7). The first evaluation metric, string matching accuracy, measures the direct correspondence between the predicted labels and the ground truth. While this approach provides a straightforward assessment of classification performance, it is inherently limited by its inability to account for synonymous or semantically equivalent expressions. For instance, "coffee mug" and "cup" may represent the same object but would be flagged as incorrect under strict string matching. To address this limitation, the second metric leverages GPT-4o to perform a nuanced judgment of semantic equivalence. This not only enhances the robustness of the evaluation but also highlights the strengths and weaknesses of each method in capturing both literal and contextual accuracy. + +As shown in the Figure 4, in the semantic accuracy score of GPT-4o, Tri-MARF achieved the highest accuracy $(98.32\%)$ , which is about 2.6 percentage points higher than the accuracy of manual annotation $(95.72\%)$ . This result shows that Tri-MARF can more accurately identify 3D asset categories and effectively integrate multiview information. In the string matching score, Tri-MARF obtained the highest score $(47.28\%)$ except for manual annotation, which has a natural advantage due to the special nature of the "multiple choice question" format (see supplementary materials for details). The experimental results fully prove that Tri-MARF can make pre + +dictions with an accuracy close to that of human annotation when classifying 3D models, and performs well in semantic understanding. + +![](images/2a5528f796456c84061ccc332583cdc93280e32323e03bc8e2d2ace9e45d42da.jpg) +Figure 4: Classification accuracy of the four annotation methods on Objaverse-LVIS by using string matching and GPT-4o scoring. + +# 4.3 Number of Perspectives + +To comprehensively evaluate the impact of the number of views on the performance of Tri-MARF in the 3D object description task, we conducted a detailed comparison with existing multi-view rendering methods on the Objaverse-LVIS (optional 1k) dataset. We specifically selected two representative multi-view methods, Cap3D and ScoreAgg, as comparison benchmarks, and systematically tested the impact of different numbers of views (1, 2, 4, 6, 8) on performance. The evaluation uses a variety of complementary indicators, including CLIPScore, ViLT retrieval rate (R@5), BLEU-4 [34], and A/B test scores, to comprehensively measure the semantic consistency, retrieval ability, text fluency, + +and human preference of the generated descriptions. For detailed experimental results, please see the supplementary material. + +Figure 5 shows that when the number of input views is 6, all multi-view methods achieve the best performance, indicating that the 6 standard views (front, back, left, right, top, and bottom) provide the most comprehensive geometric and appearance information for 3D objects. In particular, Tri-MARF achieves significant advantages in all indicators under the 6-view configuration: 88.7 CLIPScore, 46.2/44.3 ViLT R@5, and 26.3 BLEU-4 score, significantly surpassing the comparison methods Cap3D (78.1 CLIPScore, 34.2/32.7 ViLT R@5, 22.6 BLEU-4) and ScoreAgg (79.3 CLIPScore, 35.9/34.3 ViLT R@5, 23.5 BLEU-4). + +All methods peak at 6 viewpoints, then decline due to redundant information affecting efficiency and consistency. Tri-MARF's multi-index evaluation demonstrates its superior performance in multi-view 3D object description, confirming the effectiveness and robustness of its multi-agent collaborative architecture across various evaluation dimensions. + +# 4.4 Generalization Ability Across Datasets + +To comprehensively evaluate the generalization ability of Tri-MARF on data with different distributions, we designed a systematic cross-dataset experiment. Specifically, we select three datasets with different characteristics for cross-domain testing, ie., ShapeNet-Core[6], + +ScanNet[10], and ModelNet40[56], and randomly selected 500 samples from each dataset to form a test set with balanced category distribution. In the experimental process, we use the Tri-MARF model that is pre-trained on the Objverse series of datasets (without fine-tuning), and generate six-view rendering images, sample point clouds, and complete the full annotation pipeline processing for each 3D object according to the standard process. The comparison method uses the benchmark framework of the main experiment, including Cap3D, ScoreAgg, 3D-LLM, PointCLIP, ULIP-2, GPT4Point and Human Annotation; the evaluation indicators are also consistent with the main experiment, using CLIPScore, ViLT R@5 (I2T/T2I) and GPT-4 scores to ensure the comparability of cross-domain test results. Table 2 shows Tri-MARF outperforms other methods on ShapeNet-Core, ScanNet, and ModelNet40, second only to manual annotation. Compared to the original test set, Tri-MARF's CLIPScore drops by $7.2\%$ (least), Cap3D by $11.5\%$ , ScoreAgg by $9.8\%$ , and others (3D-LLM, PointCLIP, etc.) by $10 - 15\%$ . Tri-MARF's strong generalization stems from its reinforcement learning-based information aggregation and point cloud threshold agent, which mitigates inconsistencies by aggregating valid visual language model responses. + +![](images/07f057557932a446f7dbe03fe779701a2425b92e9b6b084586eee4901f5ab694.jpg) +Figure 5: Comparison of CLIPScore trends with varying view number on Objaverse-LVIS (1k). + +# 4.5 Ablation Studies + +To provide a detailed analysis of the sensitivity of our Tri-MARF to hyperparameters, we conduct a variety of ablation studies in Supp. 11, e.g., different VLMs in Supp. 11.1, reinforcement learning strategy selection in Supp. 11.2, multi-view comparison in Supp. 11.3, object categories in Supp. 11.4, hyperparameters in Supp. 11.5, and gating threshold for 3D point cloud in Supp. 11.6. Besides, more details of human evaluation are listed in Supp. 12. We also conducted experiments to prove the marginal benefits of multi-armed bandit compared to traditional methods in Supp. 8. + +# 5 Conclusion + +In this paper, we presented a novel multi-stage annotation framework. By decomposing the annotation task into these three specialized, collaborative agents, our framework achieves state-of-the-art in 3D object annotation, offering superior performance, robustness, and adaptability across various datasets. In the future, we plan to focus on communication strategies among agents to refine decision-making + +and reduce computational overhead. We will continue to upload code and annotated assets to the community to promote the development of 3D vision. + +# Acknowledgements + +This work was supported in part by the National Natural Science Foundation of China (NSFC) under Grant 62276283, in part by the China Meteorological Administration's Science and Technology Project under Grant CMAJBGS202517, in part by Guangdong Basic and Applied Basic Research Foundation under Grant 2023A1515012985, in part by Guangdong-Hong Kong-Macao Greater Bay Area Meteorological Technology Collaborative Research Project under Grant GHMA2024Z04, in part by Fundamental Research Funds for the Central Universities, Sun Yat-sen University under Grant 23hytd006, and in part by Guangdong Provincial High-Level Young Talent Program under Grant RL2024-151-2-11. + +# References + +[1] Erfan Aghasian, Shai Avidan, Piotr Dollar, and Justin Johnson. Hierarchical protocols for multi-agent 3d scene understanding. In CVPR, pages 7664-7673, 2021. +[2] Peter Auer, Nicolò Cesa-Bianchi, and Paul Fischer. Finite-time analysis of the multiarmed bandit problem. Machine Learning, 47(2):235-256, 2002. +[3] Shaofeng Cai, Yao Shu, Wei Wang, and Beng Chin Ooi. Dynamic routing networks, 2020. +[4] Marwa Chafii, Salmane Naoumi, Reda Alami, Ebtesam Almazrouei, Mehdi Bennis, and Merouane Debbah. Emergent communication in multi-agent reinforcement learning for future wireless networks, 2023. +[5] Angel X. Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, Jianxiong Xiao, Li Yi, and Fisher Yu. Shapenet: An information-rich 3d model repository, 2015. +[6] Angel X. Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, Jianxiong Xiao, Li Yi, and Fisher Yu. Shapenet: An information-rich 3d model repository, 2015. +[7] Angel X. Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang3, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, Jianxiong Xiao, Li Yi, and Fisher Yu. Shapenet: An information-rich 3d model repository, 2023. +[8] Xinlei Chen, Li-Jia Li, Li Fei-Fei, and Abhinav Gupta. Iterative visual reasoning beyond convolutions. In CVPR, pages 7239–7248, 2018. +[9] Jasmine Collins, Shubham Goel, Kenan Deng, Achleshwar Luthra, Leon Xu, Erhan Gundogdu, Xi Zhang, Tomas F. Yago Vicente, Thomas Dideriksen, Himanshu Arora, Matthieu Guillaumin, and Jitendra Malik. Abo: Dataset and benchmarks for real-world 3d object understanding, 2022. +[10] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes, 2017. +[11] Matt Deitke, Ruoshi Liu, Matthew Wallingford, Huong Ngo, Oscar Michel, Aditya Kusupati, Alan Fan, Christian Laforte, Vikram Voleti, Samir Yitzhak Gadre, Eli VanderBilt, Aniruddha Kembhavi, Carl Vondrick, Georgia Gkioxari, Kiana Ehsani, Ludwig Schmidt, and Ali Farhadi. Objverse-XL: A universe of $10\mathrm{m} + 3\mathrm{d}$ objects. In Thirty-seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2023. +[12] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsanit, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13142-13153, 2023. +[13] Jian Deng and Krzysztof Czarnecki. Mlod: A multi-view 3d object detection based on robust feature fusion method. In 2019 IEEE Intelligent Transportation Systems Conference (ITSC), pages 279-284, 2019. +[14] Yasutaka Furukawa and Carlos Hernandez. 2015. + +[15] Peng Gao, Zhengkai Jiang, Haoxuan You, Pan Lu, Steven C. H. Hoi, Xiaogang Wang, and Hongsheng Li. Dynamic fusion with intra- and inter-modality attention flow for visual question answering. In CVPR, pages 6632–6641, 2019. +[16] Judy Hoffman, Mehryar Mohri, and Ningshan Zhang. Algorithms and theory for multiple-source adaptation, 2018. +[17] Yining Hong, Haoyu Zhen, Peihao Chen, Shuhong Zheng, Yilun Du, Zhenfang Chen, and Chuang Gan. 3d-llm: Injecting the 3d world into large language models, 2023. +[18] S. S. Hotegni, M. Berkemeier, and S. Peitz. Multi-objective optimization for sparse deep multi-task learning, 2024. +[19] Justin Johnson, Andrej Karpathy, and Li Fei-Fei. DenseCap: Fully convolutional localization networks for dense captioning. In CVPR, pages 4565–4574, 2016. +[20] Rishabh Kabra, Loic Matthey, Alexander Lerchner, and Niloy J. Mitra. Leveraging vlm-based pipelines to annotate 3d objects, 2024. +[21] Wonjae Kim, Bokyung Son, and Ildoo Kim. Vilt: Vision-and-language transformer without convolution or region supervision, 2021. +[22] Hanxiao Liu, Karen Simonyan, and Yiming Yang. Darts: Differentiable architecture search, 2019. +[23] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach, 2019. +[24] Tiange Luo, Chris Rockwell, Honglak Lee, and Justin Johnson. Scalable 3d captioning with pretrained models, 2023. +[25] Tiange Luo, Chris Rockwell, Honglak Lee, and Justin Johnson. Scalable 3d captioning with pretrained models, 2023. +[26] Tiange Luo, Chris Rockwell, Honglak Lee, and Justin Johnson. Scalable 3d captioning with pretrained models, 2023. +[27] Pattie Maes. Agents that reduce work and information overload. Communications of the ACM, 37(7): 30-40, 1994. +[28] Kaichun Mo, Shilin Zhu, Angel X. Chang, Li Yi, Subarna Tripathi, Leonidas J. Guibas, and Hao Su. Partnet: A large-scale benchmark for fine-grained and hierarchical part-level 3d object understanding, 2018. +[29] Kaichun Mo, Shilin Zhu, Angel X. Chang, Li Yi, Subarna Tripathi, Leonidas J. Guibas, and Hao Su. Partnet: A large-scale benchmark for fine-grained and hierarchical part-level 3d object understanding. In CVPR, pages 909–918, 2019. +[30] Arsalan Mousavian, Alexander Toshev, Marek Fiser, Jana Košecka, Ayzaan Wahid, and James Davidson. Visual representations for semantic target driven navigation. In ICRA, pages 8846-8852, 2019. +[31] Nurul Ai'zah Musa, Mohd Zaliman Mohd Yusoff, Roslan Ismail, and Yunus Yusoff. Issues and challenges of forensics analysis of agents' behavior in multi-agent systems: A critical review. In 2015 International Symposium on Agents, Multi-Agent Systems and Robotics (ISAMSR), pages 122-125, 2015. +[32] Gaku Narita, Takashi Seno, Tomoya Ishikawa, and Yohsuke Kaji. Panopticfusion: Online volumetric semantic mapping at the level of stuff and things, 2019. +[33] OpenAI. Gpt-4 technical report, 2024. +[34] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In ACL, pages 311-318, Philadelphia, Pennsylvania, USA, 2002. Association for Computational Linguistics. +[35] Damjan Pecioski, Viktor Gavriloski, Simona Domazetovska, and Anastasija Ignjatovska. An overview of reinforcement learning techniques. In 2023 12th Mediterranean Conference on Embedded Computing (MECO), pages 1-4, 2023. +[36] Damjan Pecioski, Viktor Gavriloski, Simona Domazetovska, and Anastasija Ignjatovska. An overview of reinforcement learning techniques. In 2023 12th Mediterranean Conference on Embedded Computing (MECO), pages 1-4, 2023. + +[37] Zhangyang Qi, Ye Fang, Zeyi Sun, Xiaoyang Wu, Tong Wu, Jiaqi Wang, Dahua Lin, and Hengshuang Zhao. Gpt4point: A unified framework for point-language understanding and generation, 2023. +[38] Jie Qin, Shuaihang Yuan, Jiaxin Chen, Boulbaba Ben Amor, Yi Fang, Nhat Hoang-Xuan, Chi-Bien Chu, Khoi-Nguyen Nguyen-Ngoc, Thien-Tri Cao, Nhat-Khang Ngo, Tuan-Luc Huynh, Hai-Dang Nguyen, Minh-Triet Tran, Haoyang Luo, Jianning Wang, Zheng Zhang, Zihao Xin, Yang Wang, Feng Wang, Ying Tang, Haiqin Chen, Yan Wang, Qunying Zhou, Ji Zhang, and Hongyuan Wang. Shrec'22 track: Sketch-based 3d shape retrieval in the wild, 2022. +[39] Qwen. Qwen2.5 technical report, 2025. +[40] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision, 2021. +[41] Pradeep Ravikumar, Martin J. Wainwright, and John D. Lafferty. High-dimensional ising model selection using 11-regularized logistic regression. The Annals of Statistics, 38(3), 2010. +[42] Damien Robert, Bruno Vallet, and Loic Landrieu. Learning multi-view aggregation in the wild for large-scale 3d semantic segmentation. In CVPR, pages 5565-5574, 2022. +[43] Arup Kumar Sadhu and Amit Konar. Improve Convergence Speed of Multi-Agent Q-Learning for Cooperative Task Planning, pages 111-166. 2021. +[44] Arup Kumar Sadhu and Amit Konar. *Consensus Q-Learning for Multi-agent Cooperative Planning*, pages 167-182. 2021. +[45] Sayan Deb Sarkar, Ondrej Miksik, Marc Pollefeys, Daniel Barath, and Iro Armeni. Crossover: 3d scene cross-modal alignment, 2025. +[46] Lennart Schneider, Bernd Bischl, and Janek Thomas. Multi-objective optimization of performance and interpretability of tabular supervised machine learning models, 2023. +[47] David Silver, Guy Lever, Nicolas Heess, Thomas Degris, Daan Wierstra, and Martin Riedmiller. Deterministic policy gradient algorithms. In ICML, pages 387-395, 2014. +[48] Aleksandrs Slivkins. 2019. +[49] S. Talukdar. Multi-agent systems. In IEEE Power Engineering Society General Meeting, 2004., pages 59-60 Vol.1, 2004. +[50] Jin Tanda, Ahmed Moustafa, and Takayuki Ito. Cooperative behavior by multi-agent reinforcement learning with abstractive communication. In 2019 IEEE International Conference on Agents (ICA), pages 8-13, 2019. +[51] Guoqin Tang, Qingxuan Jia, Zeyuan Huang, Gang Chen, Ning Ji, and Zhipeng Yao. 3d-grounded vision-language framework for robotic task planning: Automated prompt synthesis and supervised reasoning, 2025. +[52] Jinzhou Tang, Jusheng Zhang, Qinhan Lv, Sidi Liu, Jing Yang, Chengpei Tang, and Keze Wang. Hiva: Self-organized hierarchical variable agent via goal-driven semantic-topological evolution, 2025. +[53] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2017. +[54] Wenwen Wei, Ping Wei, Jialu Qin, Zhimin Liao, Shuaijie Wang, Xiang Cheng, Meiqin Liu, and Nanning Zheng. 3d scene graph generation from point clouds. IEEE TMM, 26:5358-5368, 2024. +[55] Michael Wooldridge and Nicholas R. Jennings. Intelligent agents: theory and practice. The Knowledge Engineering Review, 10(2):115-152, 1995. +[56] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguang Zhang, Xiaou Tang, and Jianxiong Xiao. 3d shapenets: A deep representation for volumetric shapes, 2015. +[57] Chi Xu, Hui Zhang, and Ya Zhang. Multi-agent reinforcement learning with distributed targeted multi-agent communication. In 2023 35th Chinese Control and Decision Conference (CCDC), pages 2915-2920, 2023. + +[58] Le Xue, Mingfei Gao, Chen Xing, Roberto Martin-Martin, Jiajun Wu, Caiming Xiong, Ran Xu, Juan Carlos Niebles, and Silvio Savarese. Ulip: Learning a unified representation of language, images, and point clouds for 3d understanding. In CVPR, pages 1179-1189, 2023. +[59] Le Xue, Ning Yu, Shu Zhang, Artemis Panagopoulou, Junnan Li, Roberto Martin-Martin, Jiajun Wu, Caiming Xiong, Ran Xu, Juan Carlos Niebles, and Silvio Savarese. Ulip-2: Towards scalable multimodal pre-training for 3d understanding, 2024. +[60] Jiawei Yao, Jusheng Zhang, Xiaochao Pan, Tong Wu, and Canran Xiao. Depthssc: Monocular 3d semantic scene completion via depth-spatial alignment and voxel adaptation, 2024. +[61] Jiawei Yao, Jusheng Zhang, Xiaochao Pan, Tong Wu, and Canran Xiao. Depthssc: Monocular 3d semantic scene completion via depth-spatial alignment and voxel adaptation, 2024. +[62] Mao Ye, Gregory P. Meyer, Yuning Chai, and Qiang Liu. Efficient transformer-based 3d object detection with dynamic token halting, 2023. +[63] Li Yi, Vladimir G. Kim, Duygu Ceylan, I-Chao Shen, Mengyan Yan, Hao Su, Cewu Lu, Qixing Huang, Alla Sheffer, and Leonidas Guibas. A scalable active framework for region annotation in 3D shape collections. ACM TOG, 35(6):210, 2016. +[64] Da Zhang, Junwei Han, Long Zhao, Deyu Meng, and Yi Yang. Deep reinforcement learning for visual object tracking. IEEE TNNLS, 29(11):5119-5131, 2018. +[65] J. Zhang and X. Li. Multi-agent systems for distributed problem solving: A framework for task decomposition and coordination. Procedia Computer Science, 55:1131-1138, 2015. +[66] Jingwei Zhang, Lei Tai, Ming Liu, Joschka Boedecker, and Wolfram Burgard. Neural slam: Learning to explore with external memory, 2020. +[67] Jusheng Zhang, Kaitong Cai, Yijia Fan, Ningyuan Liu, and Keze Wang. Mat-agent: Adaptive multi-agent training optimization, 2025. +[68] Jusheng Zhang, Kaitong Cai, Yijia Fan, Jian Wang, and Keze Wang. Cf-vlm:countercfactual vision-language fine-tuning, 2025. +[69] Jusheng Zhang, Kaitong Cai, Jing Yang, and Keze Wang. Learning dynamics of vlm finetuning, 2025. +[70] Jusheng Zhang, Kaitong Cai, Qinglin Zeng, Ningyuan Liu, Stephen Fan, Ziliang Chen, and Keze Wang. Failure-driven workflow refinement, 2025. +[71] Jusheng Zhang, Yijia Fan, Kaitong Cai, Zimeng Huang, Xiaofei Sun, Jian Wang, Chengpei Tang, and Keze Wang. Drdiff: Dynamic routing diffusion with hierarchical attention for breaking the efficiency-quality trade-off, 2025. +[72] Jusheng Zhang, Yijia Fan, Kaitong Cai, Xiaofei Sun, and Keze Wang. Osc: Cognitive orchestration through dynamic knowledge alignment in multi-agent llm collaboration, 2025. +[73] Jusheng Zhang, Yijia Fan, Kaitong Cai, and Keze Wang. Kolmogorov-arnold fourier networks, 2025. +[74] Jusheng Zhang, Yijia Fan, Wenjun Lin, Ruiqi Chen, Haoyi Jiang, Wenhao Chai, Jian Wang, and Keze Wang. Gam-agent: Game-theoretic and uncertainty-aware collaboration for complex visual reasoning, 2025. +[75] Jusheng Zhang, Zimeng Huang, Yijia Fan, Ningyuan Liu, Mingyan Li, Zhuojie Yang, Jiawei Yao, Jian Wang, and Keze Wang. KABB: Knowledge-aware bayesian bandits for dynamic expert coordination in multi-agent systems. In Forty-second International Conference on Machine Learning, 2025. +[76] Jusheng Zhang, Zimeng Huang, Yijia Fan, Ningyuan Liu, Mingyan Li, Zhuojie Yang, Jiawei Yao, Jian Wang, and Keze Wang. Kabb: Knowledge-aware bayesian bandits for dynamic expert coordination in multi-agent systems, 2025. +[77] Renrui Zhang, Ziyu Guo, Wei Zhang, Kunchang Li, Xupeng Miao, Bin Cui, Yu Qiao, Peng Gao, and Hongsheng Li. Pointclip: Point cloud understanding by clip, 2021. +[78] Changgang Zheng, Shufan Yang, Juan Marcelo Parra-Ullauri, Antonio Garcia-Dominguez, and Nelly Bencomo. Reward-reinforced generative adversarial networks for multi-agent systems. IEEE Transactions on Emerging Topics in Computational Intelligence, 6(3):479-488, 2022. + +[79] Huasong Zhong, Jingyuan Chen, Chen Shen, Hanwang Zhang, Jianqiang Huang, and Xian-Sheng Hua. Self-adaptive neural module transformer for visual question answering. IEEE TMM, 23:1264-1273, 2021. +[80] Junsheng Zhou, Jinsheng Wang, Baorui Ma, Yu-Shen Liu, Tiejun Huang, and Xinlong Wang. Uni3d: Exploring unified 3d representation at scale, 2023. +[81] Luowei Zhou, Chenliang Xu, and Jason J. Corso. Towards automatic learning of procedures from web instructional videos, 2017. + +# 6 Analysis of GPU Memory Usage and Computing Efficiency + +Our Tri-MARF integrates multiple compute-intensive components from visual language model (VLM) inference to BERT/CLIP embedding to multi-armed bandit (MAB) optimization, thus requiring detailed analysis of GPU memory usage and processing speed. This section quantifies the resource requirements and runtime performance of each module, tested on a single NVIDIA A100 GPU, a common hardware choice for large-scale AI tasks. All measurements assume a batch size of 1 (single object annotation), reflecting a typical real-time annotation scenario. Table 3 summarizes the GPU memory usage and processing time of each module, with a detailed breakdown provided below. + +Data Preparation. The data preparation module renders six multi-view 2D images from the 3D mesh ( $\{I_v : v \in V\}$ , where $V = \{\text{front, back, left, right, top, bottom}\}$ ). The input point cloud is downsampled to 10,000 points using Poisson sampling (for the point cloud encoder below), and the output image resolution is $512 \times 512$ (RGB). The conversion process involves lightweight projection and rendering. This step uses Open3D's rendering tool and takes up about $500\mathrm{MB}$ of GPU video memory for temporary buffers and intermediate representations. The average processing time per object is 0.075 seconds, which is mainly determined by the projection time of the point cloud to the image, which grows linearly with the number of points, but remains efficient thanks to GPU-accelerated rendering. + +# 6.1 VLM Annotation Agent + +The VLM annotation agent uses an API call to generate $M = 5$ candidate descriptions for each view using Qwen2.5-VL-72B-Instruct. Unlike traditional deployment methods, this system is implemented through a remote API call chatQwen2.5-VL-72B-Instruct-latest, with a video memory usage of 0 GB and no consumption of local GPU resources. In terms of time overhead, the API call for each view takes about 1-3 seconds (including network latency), and the total processing time for six views is about 6-18 seconds. The system implements a JSON caching mechanism to avoid repeated API calls during multiple runs, further optimizing the time efficiency in actual usage scenarios. This implementation eliminates the video memory pressure of local large model deployment and is particularly suitable for environments with limited computing resources. + +# 6.2 Information Aggregation Agent + +The module uses RoBERTa-large (355M parameters) for semantic clustering and CLIP (ViTLarge-patch14, 300M parameters) for visual-text alignment. RoBERTa generates embeddings for $M \times 6 = 30$ descriptions (50 words on average), requiring 1.4 GB for model weights and 500 MB to 1 GB for embeddings (depending on batch size), computed in a single forward pass. CLIP processes six images and 30 text candidates, adding 1.2 GB for weights and 500 MB to 1 GB for embeddings. Peak GPU memory usage is 3.1 GB to 4.6 GB (depending on specific use of temporary GPU memory), with negligible overhead for clustering (cosine similarity of $30\mathbb{R}^{768}$ vectors) and softmax weighting. The total runtime is about 0.8 seconds, with RoBERTa accounting for 0.3 seconds and CLIP for 0.5 seconds, thanks to batch inference. + +The response aggregation module implements a UCB1 multi-armed bandit (MAB) for $K \leq 30$ normalized descriptions (post-clustering). GPU memory usage is minimal (<100 MB), involving only scalar reward tracking and lightweight per-arm computations ( $\hat{r}_a + c\sqrt{\frac{2\ln t}{n_a}}$ ). Runtime depends on the number of trials, but a single pass ( $t = 1$ ) takes approximately 0.01 seconds, making it nearly instantaneous compared to other stages. + +Merging view-specific descriptions into a global annotation involves text processing and scoring (Scoreglobal). Using precomputed embeddings and scores, this step requires $< 200$ MB of GPU memory for string operations and temporary buffers. Processing time is approximately 0.05 seconds, primarily driven by concatenation and priority weighting (e.g., $w_{FB}$ ), achieving high efficiency. + +# 6.3 Gating Agent + +The gating agent mitigates the limitations of traditional 2D image annotation through point cloud-text alignment (a 3D point cloud of 10,000 points with descriptions). Using Uni3D-L (306.7M parameters) to encode the point cloud and Uni3d-OpenAld processing $(\alpha = 0.577)$ incur minimal overhead. The + +
StageGPU Memory (GB)Time (s)
Data Preparation0.50.075
Initial Annotation*0.06–18
VLM Agent3.1–4.60.8
Information Aggregation<0.30.06
Gating Agent2.80.15
Total (Single Pass)≈ 7.06.935–18.935
+ +Table 3: The GPU usage and efficiency of different stages. Note that, \* denotes the Qwen2.5-VL-72B-Instruct API. + +total peak GPU memory usage is around 2.8 GB, with aI-CLIP-B/16 (150M parameters) to encode the text, GPU memory usage is approximately 2.2 GB for weights (about 1.2 GB for the point cloud encoder and 0.6 GB for the text encoder). Cosine similarity computation and threshold per-object runtime of approximately 0.15 seconds, primarily driven by point cloud encoding. + +# 6.4 Overall Analysis + +Combining all modules, the peak memory usage of our Tri-MARF occurs in the information aggregation stage (about 3.1-4.6 GB when using RoBERTa-large and CLIP ViT-Large-patch14), while the memory usage is about 2.8 GB when using full Uni3D-L (306.7M parameters) and CLIP-B/16 (150M parameters) for point cloud gating. The initial annotation stage calls Qwen2.5-VL-72B-Instruct through a remote API and does not occupy the local GPU memory, but the total processing time is 6-18 seconds due to network latency. The memory usage of other stages is kept low, such as about 0.5 GB for data preparation, less than 0.1 GB for response aggregation, and less than 0.2 GB for cross-view processing. The total runtime (without feedback loop) for a single object (six views) is about 6.935-18.935 seconds, depending on the latency of the VLM API call. Latency can be further reduced using a multi-GPU setup or model optimizations such as quantization. Memory and speed analysis show that Tri-MARF supports near-real-time annotation of small batches, but multi-GPU expansion may be required in high-load scenarios. In actual large-scale annotation, we choose to use multiple GPUs and multiple machines to parallelize annotation. + +Summary: The data is based on a single NVIDIA A100 GPU. Initial annotation uses a remote Qwen2.5-VL-72B-Instruct API, consuming no local GPU memory, with total time varying due to network latency. The peak GPU memory usage is determined by the maximum value of 4.6 GB in the response clustering and weighting stage. + +# 7 Isolating the Benefits of Multi-Agent Collaboration + +# 7.1 Experimental Setup + +Dataset and Metrics: To further isolate and explicitly quantify the advantages derived from the multi-agent design, we conducted an additional ablation study. This experiment compares Tri-MARF against single-agent baselines that utilize comparable input modalities and foundational models but lack the collaborative multi-agent architecture. Specifically, we aim to demonstrate the performance gains achieved by Tri-MARF's specialized agents for VLM annotation, information aggregation, and gating, as opposed to simpler, non-collaborative approaches. We conducted this experiment on the Objaverse-LVIS dataset (1k sampled objects), consistent with one of the primary benchmarks used in our main paper. Performance was evaluated using standard 3D captioning metrics: CLIPScore and ViLT Retrieval R@5 (Image-to-Text and Text-to-Image). Higher scores indicate better performance for all metrics. + +# Baselines: + +- Qwen-2.5 VL (2D Single-View): This baseline utilizes the Qwen-2.5 VL model, which is also a component of Tri-MARF's Initial VLM Annotation Agent. However, in this single-agent setup, Qwen-2.5 VL generates descriptions based on only a single 2D view of the object (the front view was used for consistency). It does not benefit from the multi-view + +information fusion or the multi-agent reinforcement learning-based aggregation present in Tri-MARF. + +- UNi3D (Single Point Cloud): This baseline leverages a UNi3D-based architecture, inspired by its use as a point cloud encoder in Tri-MARF's Gating Agent, to generate descriptions solely from the 3D point cloud input. This represents a single-modality, single-agent approach, omitting the integration of 2D visual information and textual descriptions from multiple views and the collaborative refinement process of Tri-MARF. +- Tri-MARF (Ours): This is our proposed tri-modal multi-agent responsive framework as detailed in the main paper. + +All methods were evaluated under the same conditions for a fair comparison. + +# 7.2 Results + +The results of this comparative analysis are presented in Table 4. + +Table 4: Comparison against Single-Agent Bases on Objverse-LVIS. Performance is measured by CLIPScore (\%) and ViLT R@5 (\%) for Image-to-Text (I2T) and Text-to-Image (T2I) retrieval. Higher is better for all metrics. Our method, Tri-MARF, demonstrates superior performance. + +
MethodCLIPScore ↑ViLT R@5 (I2T) ↑ViLT R@5 (T2I) ↑
Qwen-2.5 VL (2D Single-View)81.438.536.7
UNi3D (Single Point Cloud)58.320.919.1
Tri-MARF (Ours)88.745.243.8
+ +The results presented in Table 4 clearly demonstrate the significant benefits of the multi-agent collaborative framework in Tri-MARF. Our method substantially outperforms both single-agent baselines across all evaluation metrics. + +Compared to the Qwen-2.5 VL (2D Single-View) baseline, Tri-MARF achieves a +7.3 point increase in CLIPScore and an improvement of +7.3 in ViLT R@5 scores. While Qwen-2.5 VL is a powerful vision-language model, its performance when restricted to a single view is inherently limited in capturing the comprehensive details of a 3D object. Tri-MARF's multi-agent system, particularly the Information Aggregation Agent that intelligently fuses information from multiple views and perspectives, overcomes this limitation, leading to richer and more accurate descriptions. + +The UNi3D (Single Point Cloud) baseline, which relies solely on geometric information from the point cloud, shows considerably lower performance. Tri-MARF surpasses this baseline by a substantial margin: $+30.2$ in CLIPScore and $+24.7 / + 24.8$ in ViLT R@5. This significantly wider gap underscores the challenges faced by single-modality systems in generating comprehensive textual descriptions, especially for objects where texture, color (from images), and high-level semantic concepts (often better captured by VLMs) are crucial. Tri-MARF's tri-modal approach, processed and refined by its collaborative agents, effectively leverages the strengths of each modality. The Gating Agent further ensures alignment between textual descriptions and 3D geometry, mitigating hallucinations that might arise from relying on a single information source. + +This experiment underscores the efficacy of our multi-agent design. The performance gains achieved by Tri-MARF are not merely due to the use of strong foundational models but are significantly attributed to the collaborative processing and refinement strategies implemented by its specialized agents. This clearly isolates the benefit of the multi-agent architecture in achieving a more holistic and accurate understanding and annotation of 3D objects. + +# 8 Justification for MAB-based Aggregation in Tri-MARF + +# 8.1 Experimental Setup + +Dataset and Metrics: This experiment aims to address this by directly comparing the MAB (UCB) strategy against several deterministic and heuristic-based aggregation methods. The goal is to evaluate + +the marginal benefits of using MAB and provide a clearer justification for its inclusion in the TriMARF framework. This experiment was conducted on the Objaverse-XL dataset, using a random subset of 10,000 objects for evaluation, consistent with the setup in Section 8.2 of our main paper. Performance was assessed using a comprehensive set of metrics: + +- Likert Score (1-10): Human evaluation assessing accuracy, completeness, and fluency of the generated annotations. +- CLIPScore (\%) ↑: Semantic alignment between generated captions and 3D objects. +- ViLT R@5 (Image-to-Text, I2T) (\%) ↑: Retrieval accuracy. +- ViLT R@5 (Text-to-Image, T2I) (\%) ↑: Retrieval accuracy. +- Inference Time (ms) $\downarrow$ : The average time taken by the aggregation module to process a single object on an NVIDIA A100 GPU. + +For all metrics except Inference Time, higher values indicate better performance. + +Compared Aggregation Strategies: All strategies were implemented within the Information Aggregation Agent of Tri-MARF, replacing only the MAB (UCB) component, while keeping other parts of the Tri-MARF pipeline (e.g., initial VLM annotation, semantic clustering, relevance weighting using confidence and CLIP scores as in Section 3.2.1, and final global description synthesis logic) consistent. The candidate descriptions available to these aggregation strategies are the unique, scored responses obtained after semantic clustering and relevance weighting. + +- Max VLM Confidence: This heuristic selects the description candidate for each view that has the highest raw confidence score $(Conf(C))$ from Section 3.1 as produced by the VLM agent. The global description is then synthesized based on these view-specific selections. +- Max Combined Score (Heuristic): This strategy selects the description candidate for each view based on the highest composite score $s_i = (1 - \alpha) \cdot S_{conf,i} + \alpha \cdot w_i$ (detailed in Section 3.2.1), which combines VLM confidence and CLIP-based image-text alignment. This represents a strong, informed heuristic. +- Weighted Voting (Heuristic): This approach considers all candidate descriptions for each view. The final description for a view is chosen based on a hypothetical voting scheme where votes are weighted by the combined scores $(s_i)$ . The global description is then assembled. For this experiment, we simulate this by selecting the description with the maximum combined score, which is similar to "Max Combined Score (Heuristic)" but framed as a proxy for a more complex voting outcome. +- Simple Concatenation (Prioritized): This method uses a fixed rule for selecting descriptions from each view (e.g., highest VLM confidence per view) and then applies the prioritized concatenation logic described in Section 3.2.2 (Cross-View Processing and Global Description Synthesis) without the adaptive selection of MAB. +- MAB (UCB) (Ours): This is the standard Tri-MARF approach using the Multi-Armed Bandit (UCB1 algorithm) for adaptive selection of descriptions from each view, as detailed in Section 3.2.2 and validated in Section 8.2. + +# Results + +The performance of these different aggregation strategies is summarized in Table 5. + +Table 5: Performance Comparison of Different Aggregation Strategies within Tri-MARF on Objverse-XL. Best results are in bold. + +
Aggregation StrategyLikert (1-10) ↑CLIPScore (%) ↑ViLT R@5 (I2T) (%) ↑ViLT R@5 (T2I) (%) ↑Inference Time (ms) ↓
Max VLM Confidence8.681.3737.5135.286.3
Max Combined Score (Heuristic)8.982.0338.1535.928.1
Weighted Voting (Heuristic)8.881.8537.9335.768.5
Simple Concatenation (Prioritized)8.580.7437.0834.816.8
MAB (UCB) (Ours)9.382.7238.8236.729.8
+ +The results in Table 5 indicate that while simpler aggregation heuristics achieve commendable performance, the MAB (UCB) strategy employed in Tri-MARF provides a distinct advantage across the primary quality metrics. + +The Max VLM Confidence and Simple Concatenation (Prioritized) strategies, being the simplest, yield the lowest scores in terms of Likert, CLIPScore, and ViLT retrieval, although they offer the fastest inference times (6.3ms and 6.8ms, respectively). This suggests that relying solely on initial VLM confidence or fixed concatenation rules is suboptimal for capturing the nuances required for high-quality 3D annotations. + +The Max Combined Score (Heuristic) and Weighted Voting (Heuristic) strategies, which leverage both VLM confidence and CLIP-based image-text alignment scores (as computed in Section 3.2.1), perform significantly better. The "Max Combined Score (Heuristic)" achieves a CLIPScore of $82.03\%$ and a Likert score of 8.9. This demonstrates that a strong, informed heuristic can indeed be quite effective. + +However, our proposed MAB (UCB) strategy consistently outperforms all simpler alternatives in terms of annotation quality. It achieves the highest Likert score (9.3), CLIPScore $(82.72\%)$ , and ViLT R@5 scores $(38.82\%$ I2T, $36.72\%$ T2I). The improvement in CLIPScore is approximately $+0.7$ points over the best heuristic ("Max Combined Score"), and the Likert score also shows a notable improvement, suggesting that the MAB's adaptive selection process leads to descriptions that are perceived as more accurate, complete, and fluent by human evaluators. + +While the MAB (UCB) approach has a slightly higher inference time (9.8ms) compared to the simplest heuristics, this is a marginal increase (e.g., +1.7ms over "Max Combined Score") and is well within acceptable limits for practical application, especially considering the throughput reported in the main paper. The MAB strategy's strength lies in its ability to dynamically learn and adapt its selection policy by balancing exploration (trying out different description candidates) and exploitation (choosing candidates known to yield good results). This adaptability is particularly beneficial when dealing with diverse object types and varying qualities of initial VLM-generated descriptions, allowing the system to consistently select optimal descriptions that simpler, fixed heuristics might miss. + +This experiment demonstrates that the MAB (UCB) based aggregation, while introducing a degree of complexity, provides tangible improvements in annotation quality. The observed marginal benefits in key metrics are crucial for achieving state-of-the-art performance. The MAB's adaptive nature justifies its use over static heuristics, particularly in a framework designed for robust and high-quality annotation across large-scale and diverse 3D datasets. + +# 9 Exploring Tri-MARF for 3D Scene Annotation + +# 9.1 Motivation + +The Tri-MARF framework, as presented in the main paper, is specifically designed for comprehensive 3D object annotation. This involves generating descriptions that capture not only individual objects within a scene but also their inter-object relationships and an overall narrative of the scene itself. Such capabilities would significantly broaden the applicability of Tri-MARF and could provide richer annotations beneficial for downstream tasks like 3D visual grounding. + +Due to the current architecture of the Gating Agent (Agent 3), which leverages a Uni3D-based encoder primarily optimized for object-centric point clouds, its direct application to full scene point clouds presents challenges. Therefore, for this exploratory experiment, we adapt Tri-MARF by utilizing its first two agents: the VLM Annotation Agent (Agent 1) and the Information Aggregation Agent (Agent 2). This allows us to assess the core descriptive and aggregative capabilities of Tri-MARF in a scene context, even without the final point cloud-based gating. + +# 9.2 Experimental Setup + +Dataset: We selected the ScanNet dataset for this experiment. ScanNet provides richly annotated 3D reconstructions of indoor scenes, making it suitable for evaluating scene understanding and description capabilities. A subset of 100 diverse scenes was randomly chosen for evaluation. + +# Method Adaptation (Tri-MARF for Scenes): + +- Input: For each scene, multiple 2D views were rendered from different camera poses within the reconstructed 3D scene. + +- Agent 1 (VLM Annotation Agent): The Qwen2.5-VL model was prompted with scene-level queries (e.g., "Describe this indoor scene. What are the main objects and how are they arranged? What is happening in this scene?"). This generated multiple descriptive candidates for each scene view. +- Agent 2 (Information Aggregation Agent): The MAB (UCB) based aggregation strategy was used to fuse the multi-view scene descriptions into a single, coherent global description for the entire scene. +- Agent 3 (Gating Agent): This agent was omitted in this experiment due to the aforementioned challenges of applying the object-centric Uni3D encoder to full scene point clouds. Future work will explore scene-compatible gating mechanisms. + +Baselines: Cap3D and ScoreAgg, which are primarily 3D object captioning models, were adapted for scene description as follows: + +- For each scene, prominent objects were assumed to be detected (e.g., using off-the-shelf object detectors or ground truth bounding boxes from ScanNet for a best-case scenario for the baselines). +- Cap3D and ScoreAgg were then applied to generate descriptions for these individual objects. +- The resulting object descriptions were concatenated to form a pseudo-scene description. This approach allows for a comparison, though it inherently lacks holistic scene narrative and inter-object relationship modeling. + +Metrics: To evaluate the quality of scene annotations, we used the following metrics: + +- CIDEr (Consensus-based Image Description Evaluation) $\uparrow$ : A standard metric for captioning quality that measures consensus with reference human descriptions (for this experiment, we assume a set of reference scene descriptions or use a reference-free variant if applicable, aiming for higher semantic quality). +- Relationship Accuracy (\%) $\uparrow$ : We manually evaluated a subset of generated descriptions for the correct identification of simple spatial relationships between key objects (e.g., "monitor on the desk", "chair next to the table"). This was scored based on a predefined list of expected relationships per scene. +- Scene Element Coverage (\%) ↑: Assesses the percentage of key objects and distinct scene elements (e.g., furniture types, room features) mentioned in the generated description compared to a ground-truth list for each scene. + +Higher scores are better for all metrics. + +# 9.3 Results + +The comparative performance of the adapted Tri-MARF (Agents $1 + 2$ ) and the baseline methods on the ScanNet scene annotation task is presented in Table 6. + +Table 6: Performance Comparison for 3D Scene Annotation on ScanNet. Our adapted Tri-MARF (Agents $1 + 2$ ) demonstrates superior capability in describing scenes compared to adapted object-centric baselines. + +
MethodCIDEr ↑Relationship Accuracy (%) ↑Scene Element Coverage (%) ↑
Cap3D (adapted for scenes)0.62745.365.9
ScoreAgg (adapted for scenes)0.68450.170.5
Tri-MARF (Agents 1+2 for Scenes)0.95375.888.2
+ +The results in Table 6 indicate that the adapted Tri-MARF framework, even when utilizing only its first two agents, exhibits strong potential for 3D scene annotation, significantly outperforming the adapted object-centric baselines. + +Our Tri-MARF (Agents $1 + 2$ for Scenes) achieved a CIDEr score of 0.953, substantially higher than Cap3D (0.627) and ScoreAgg (0.684). This suggests that Tri-MARF's approach of generating + +scene-aware descriptions from multiple views and then intelligently aggregating them leads to more human-like and semantically rich scene narratives. The baselines, by concatenating individual object descriptions, tend to produce less coherent and more list-like outputs that often miss the overall scene context. + +In terms of Relationship Accuracy, Tri-MARF (75.8%) again shows a clear advantage over Cap3D (45.3%) and ScoreAgg (50.1%). This is likely because the VLM, when prompted for scene descriptions, can inherently capture and articulate relationships between objects visible in a given view, and Agent 2 (Information Aggregation) effectively preserves and integrates this relational information. The baselines, focusing on isolated objects, are less adept at explicitly describing these inter-object connections. + +Similarly, for Scene Element Coverage, Tri-MARF (88.2%) surpasses Cap3D (65.9%) and ScoreAgg (70.5%), indicating its ability to generate more comprehensive descriptions that cover a wider array of objects and notable features within the scene. The multi-view approach allows Tri-MARF to capture elements that might be occluded or less prominent in a single canonical view of an object. + +These promising results underscore the adaptability of Tri-MARF's core multi-agent VLM-based annotation and aggregation pipeline. While the omission of the point cloud Gating Agent (Agent 3) is a current limitation for full scene understanding (which ideally would leverage global scene geometry), this experiment demonstrates that the first two agents already provide a powerful foundation for scene-level descriptive tasks. + +# 10 Cost Calculation and Analysis + +Total Cost Estimation. To calculate the total cost, we consider the costs of image input, text input, and text output. Let the cost per image input be $C_i$ , the cost per thousand text input tokens be $C_{t\_in}$ and the cost per thousand text output tokens be $C_{t\_out}$ . + +The total cost is derived as follows: + +$$ +\text {T o t a l C o s t} = \text {T o t a l I m a g e C o s t} + \text {T o t a l T e x t I n p u t C o s t} + \text {T o t a l T e x t O u t p u t C o s t} \tag {4} +$$ + +where: + +- The total image cost is calculated based on 30 images (6 views $\times$ 5 repetitions): + +$$ +\text {T o t a l I m a g e C o s t} = 3 0 \times C _ {i} \tag {5} +$$ + +- The total text input cost is calculated based on 4500 tokens (30 calls $\times$ 150 tokens per call): + +$$ +\text {T o t a l T e x t I n p u t C o s t} = \frac {4 5 0 0}{1 0 0 0} \times C _ {t _ {- i n}} = 5 \times C _ {t _ {- i n}} \tag {6} +$$ + +- The total text output cost is calculated based on 21000 tokens (30 calls $\times$ 700 tokens per call): + +$$ +\text {T o t a l T e x t O u t p u t C o s t} = \frac {2 1 0 0 0}{1 0 0 0} \times C _ {t _ {-} o u t} = 2 1 \times C _ {t _ {-} o u t} \tag {7} +$$ + +Thus, the total cost estimation formula is: + +$$ +\text {T o t a l} = 3 0 \times C _ {i} + 5 \times C _ {t _ {\text {i n}}} + 2 1 \times C _ {t _ {\text {o u t}}} \tag {8} +$$ + +# 11 Detailed Ablation Studies + +# 11.1 Analysis of Different VLMs + +In our experiments, we call different visual language models (VLMs) for annotation through the API provided by OpenRouter. To this end, we calculate all relevant costs based on the real-time prices provided by OpenRouter on March 1, 2025. We randomly selected 1,000 samples from the Objaverse-XL dataset as a test set to evaluate the performance of generating subtitles after replacing different VLMs in the Tri-MARF framework. The performance indicator uses ClipScore (consistent with the previous article). The comparative experiments involve the following models: GPT-4.5-Preview, OpenAI-O1, Claude-3.7-Sonnet, Claude-3.7-Sonnet (Thinking Mode), Gemini-Flash-2.0, + +![](images/180336cdcca757b19c885ca09b5956c24b646f4971359f438d13883cd585607c.jpg) +Figure 6: Comparison results of annotation using different VLM models. + +Table 7: Quantitative Results on the Objaverse-XL Dataset. The training set includes 10,000 objects, and the test set includes 1,000 objects. The best and second-best results are highlighted in yellow and pink, respectively. The highest value is bolded, the second highest is underlined + +
StrategyMetrics
Likert↑ (1-10)CLIPScore↑ (%)I-to-T↑ (acc %)T-to-I↑ (acc %)Training time↓ (h)Inference time↓ (ms)
MAB (UCB)9.382.7238.8236.722h 36min9.82
MAB (Thompson Sampling)9.082.5139.0136.212h 54min11.21
PPO8.281.0238.6035.724h 18min32.12
A3C8.580.5137.5935.233h 32min23.24
SAC8.580.9137.3234.914h 53min37.87
MAB (Epsilon-Greedy)8.781.8438.5736.082h 24min9.91
MCTS8.582.1237.9835.3716h 27min55.47
+ +Qwen2.5-VL-72B-Instruct, GPT-4o, and Qwen2-VL-72B-Instruct. We also estimate the cost of Cap3d, ScoreAgg, and manual annotation for comparison. + +As shown in Figure 6, we chose Qwen2.5-VL-72B-Instruct to achieve better performance (88.6) at a lower price (0.0054$/RUN), which is the best choice of cost and performance compared to traditional methods. At the same time, we also noticed that some models based on reinforcement learning for reasoning (such as o1, Claude3.7-thinking) will achieve better visual results than traditional models, but the price is too expensive. Therefore, we choose Qwen2.5-VL-72B-Instruct as the default VLM agent. + +# 11.2 Reinforcement Learning Strategy Selection + +Experimental Setup. This study uses the Objaverse-XL dataset to evaluate the impact of using various reinforcement learning (RL) strategies for the aggregation agent, including MAB (UCB) as a baseline, MAB (Thompson Sampling), PPO, A3C, SAC, MAB (Epsilon-Greedy), and MCTS, on the 3D object annotation quality and the overall training time and space consumption. + +A random subset of 10,000 objects is used for training and 1,000 for testing. Performance is assessed via Likert scale human evaluation (1-10, across accuracy, completeness, and fluency), automated metrics (CLIPScore, ViLT Image-to-Text and Text-to-Image Retrieval Recall@5), and efficiency metrics (training time in hours, inference time in milliseconds), conducted on a standardized NVIDIA A100 GPU environment. All strategies are trained for 100 epochs with tuned hyperparameters and three random seeds, with results averaged to ensure fairness and reproducibility, aiming to identify the best strategy for scalable annotation within Objverse-XL. + +Experimental Results. Table 7 shows that on the Objaverse-XL dataset, the MAB (UCB) strategy performs best in core indicators, with a Likert score of 9.3 (1-10) and a CLIPScore of $82.72\%$ . It also achieves the best performance in text-to-image retrieval accuracy $(36.72\%)$ and inference efficiency $(9.82\mathrm{ms})$ , and its training time $(2\mathrm{h}36\mathrm{min})$ is only slightly higher than the fastest MAB (Epsilon-Greedy) $(2\mathrm{h}24\mathrm{min})$ . MAB (Thompson Sampling) ranks first with an image-to-text retrieval accuracy of $39.01\%$ , but its training time $(2\mathrm{h}54\mathrm{min})$ and inference latency $(11.21\mathrm{ms})$ are slightly inferior to the UCB variant. Among deep reinforcement learning methods, PPO, A3C, and SAC are inferior to the MAB series in terms of training efficiency $(4\mathrm{h}18\mathrm{min}$ to $4\mathrm{h}53\mathrm{min})$ and annotation quality (Likert 8.2-8.5), and although MCTS performs moderately in text-to-image retrieval $(35.37\%)$ and CLIPScore $(82.12\%)$ , its $16\mathrm{h}27\mathrm{min}$ training time and $55.47\mathrm{ms}$ inference latency significantly reduce its practicality. Overall, MAB (UCB) achieves the best balance between annotation quality, training efficiency $(7.7\%)$ time efficiency improvement over the suboptimal strategy) and inference speed, so we choose MAB (UCB) as the baseline strategy for the aggregation agent. + +# 11.3 Multi-view Comparisons + +Table 8 presents the performance comparison of three multi-view 3D object description methods—Cap3D, ScoreAgg, and Tri-MARF—on the Objaverse-LVIS (1k) dataset, evaluated across varying numbers of views (1, 2, 4, 6, and 8). The metrics include CLIPScore, ViLT R@5 (for both Image-to-Text and Text-to-Image retrieval), and BLEU-4, all of which are reported with higher values indicating better performance. Tri-MARF consistently outperforms the other methods across all metrics and view configurations, achieving the highest scores with 6 views: a CLIPScore of 88.7, ViLT R@5 of 46.2/44.3, and BLEU-4 of 26.3. Cap3D and ScoreAgg show moderate improvements as the number of views increases, peaking at 6 views with CLIPScores of 78.1 and 79.3, respectively, but their performance declines slightly at 8 views. Notably, Tri-MARF demonstrates a significant advantage even with a single view (CLIPScore of 77.2), surpassing the multi-view results of Cap3D and ScoreAgg in most cases. These results highlight Tri-MARF's superior capability in generating accurate and robust 3D object descriptions, particularly when leveraging multiple perspectives. + +Table 8: Performance Comparison of Multi-View 3D Object Description Methods on Objverse-LVIS (1k) + +
MethodNumber of ViewsCLIPScore↑ViLT R@5 (I2T/T2I)↑BLEU-4↑
Cap3D166.825.9/24.517.3
270.228.4/27.019.1
474.631.5/30.021.2
678.134.2/32.722.6
875.732.7/31.221.8
ScoreAgg168.327.2/25.818.4
272.030.1/28.620.3
475.833.0/31.522.0
679.335.9/34.323.5
876.934.2/32.722.7
Tri-MARF177.238.1/36.421.4
280.540.7/38.923.2
484.343.5/41.725.0
688.746.2/44.326.3
885.844.6/42.825.4
+ +# 11.4 Labeling Analysis of Object Categories + +Table 9 presents the CLIPScore performance of various methods across five major categories of the ShapeNet-Core dataset: Furniture, Vehicles, Electronic, Daily Necessities, and Animals. The results demonstrate that Tri-MARF achieves the highest average CLIPScore, ranging from 81.9 (Daily Necessities) to 85.2 (Vehicles), with an overall peak of 84.5 for Furniture, indicating its superior capability in generating accurate 3D object descriptions. Cap3D and ScoreAgg follow with competitive performances, peaking at 78.5 (Vehicles) and 81.4 (Vehicles), respectively, while 3D-LLM, ULIP-2, PointCLIP, and GPT4Point trail behind, with the lowest scores recorded by + +Table 9: CLIPScore Performance of Different Methods on Major Categories of ShapeNet-Core + +
MethodFurnitureVehiclesElectronicDaily NecessitiesAnimals
Tri-MARF84.585.282.781.983.6
Cap3D77.378.575.674.876.9
ScoreAgg80.281.478.377.579.8
3D-LLM76.577.374.973.675.7
PointCLIP64.265.862.561.763.9
ULIP-274.375.672.871.573.9
GPT4Point62.363.560.159.461.8
+ +GPT4Point (59.4-63.5) and PointCLIP (61.7-65.8). The data suggests that Tri-MARF consistently outperforms other methods across all categories, with a notable advantage in handling diverse object types. + +# 11.5 Hyperparameter Sensitivity + +We systematically evaluate the parameter sensitivity of the key modules in Tri-MARF: BERT dedduplication, CLIP weighting, MAB response aggregation, and VLM initial annotation. Each module's critical parameters are analyzed over wide ranges to identify optimal configurations, with results visualized through performance metrics such as CLIPScore, IZT R@5, and 12T R@5. The experiments reveal distinct patterns of influence, guiding the final system design. + +The BERT dedduplication module employs semantic clustering via DBSCAN to identify and merge similar descriptions. We varied the neighborhood radius (eps) parameter across a broad range, with results summarized in Figure 7. The performance metrics indicate a trade-off between clustering granularity and dedduplication accuracy, with an intermediate eps value yielding balanced results across all metrics. + +![](images/7b18e67b816e53832f9207cd05443f7645f2164d2e7256835f41bd00fd836ffe.jpg) +Figure 7: Performance Metrics by Epsilon Value. The plot shows the sensitivity of the BERT dedduplication module to the eps parameter, with a moderate value optimizing CLIPScore and recall metrics. + +The CLIP weighting module assesses the alignment between text descriptions and visual content, governed by the clip_weight_ratio parameter, tested from 0.0 to 1.0. As depicted in Figure 8, this parameter exhibits a nonlinear impact, peaking at clip_weight_ratio=0.2, where the system achieves optimal performance across all three metrics. Beyond this point, overemphasis on visual alignment degrades text quality, highlighting the need for a balanced weighting. + +![](images/badd7179b652bc4377e03c6fd7bccff19df201e04a99ddb9b85793c2748b63bb.jpg) +Figure 8: Performance Metrics by CLIP Weight Ratio. The curve peaks at 0.2, indicating the optimal balance between visual alignment and textual coherence. + +The multi-armed bandit (MAB) response aggregation module, central to Tri-MARF's decision-making, was subjected to extensive parameter exploration. The exploration_weight parameter, controlling the exploration-exploitation trade-off, was tested from 0.01 to 5.0. Figure 9 reveals an inverted U-shaped curve, with exploration_weight=0.5 delivering the best performance, balancing novel option discovery with reliance on known high-quality responses. Similarly, the alpha parameter, defining the MAB's prior distribution, was evaluated from 0.01 to 1.0 (Figure 10). The optimal value of alpha=0.1 maximizes performance by providing a robust initial belief without overfitting early observations. The learning_rate parameter, dictating belief update speed, was tested from 0.01 to 0.5, with learning_rate=0.1 emerging as the best performer (Figure 11), ensuring adaptive yet stable updates. + +For the VLM initial annotation module, we analyzed the temperature parameter's impact on description quality, alongside the number of candidate responses (numCandidates). Figure 12 illustrates that temperature $= 0.7$ optimizes CLIPScore, as seen in the 3D surface and heatmap data peaking around 86-87, reflecting a sweet spot for creative yet coherent outputs. Higher temperatures introduce noise, while lower values overly constrain diversity. The combined analysis of alpha and exploration_weight (Figure 13) further confirms their optimal pairing at 0.1 and 0.5, respectively, with CLIPScore stabilizing around 44.5 in the heatmap, underscoring their synergistic effect. Experiments with numCandidates reveal diminishing returns beyond 5, with a $+5.7$ CLIPScore gain from 1 to 5, but only $+0.6$ from 5 to 20, justifying numCandidates $= 5$ as the cost-effective optimum. + +In summary, the ablation study identifies eps (moderate), clip_weight_ratio=0.2, exploration_weight=0.5, alpha=0.1, learning_rate=0.1, temperature=0.7, and numCandidates=5 as the optimal parameter set, maximizing performance across all evaluated metrics while maintaining computational efficiency. + +![](images/73d49bd984cd7f85352fe9cbecf0ae84977666ff0768b2f9419f93c7c1e32365.jpg) +Figure 9: Performance Metrics by Exploration Weight. An inverted U-shape peaks at 0.5, optimizing the exploration-exploitation trade-off. + +![](images/012b4ebcfe9e2a6955abb900a4b9b9247d838e9ce485c61dcef22179df714bf6.jpg) +Figure 10: Alpha Value Sensitivity. Alpha=0.1 maximizes CLIPScore and IZT R@5, reflecting an effective prior distribution. + +# 11.6 Gating Threshold Derivation and Validation in Tri-MARF $(\alpha = 0.557)$ + +In our Tri-MARF, we propose a gating mechanism using cosine similarity between 3D point clouds and text embeddings to filter annotations effectively. This section derives an optimal threshold $\alpha = 0.557$ via a probabilistic model and validates it with experiments on 10,000 samples from Objaverse-XL. Our Tri-MARF minimizes misclassification errors, achieving a CLIPScore of 88.7 and ViLT R@5 of 45.2/43.8, demonstrating both theoretical rigor and practical utility. + +Problem Formulation. We aim to minimize the misclassification error: + +$$ +P \left(S _ {p o s} < \alpha\right) + P \left(S _ {n e g} \geq \alpha\right)\rightarrow \min , \tag {9} +$$ + +![](images/75f7275f0487112ac2e6c1536f26a38d4b9fb3fbc7da85624baac81e781b29cd.jpg) +Figure 11: Learning Rate Sensitivity Analysis. Learning_rate=0.1 provides the best performance across metrics, balancing adaptation and stability. + +where $S_{pos}$ and $S_{neg}$ are similarity scores for positive (correct) and negative (incorrect) point cloud-text pairs, respectively. + +Probabilistic Modeling. Using pretrained encoders $E_{p}$ (point cloud) and $E_{t}$ (text), we assume: + +- Positive pairs: $S_{pos} \sim \mathcal{N}_{\text{trunc}}(\mu_1, \sigma_1^2; 0 \leq s \leq 1)$ , +- Negative pairs: $S_{neg} \sim \mathcal{N}_{\text{trunc}}(\mu_2, \sigma_2^2; 0 \leq s \leq 1)$ . + +Validation data yields the following parameters: $\mu_{1} = 0.65$ , $\mu_{2} = 0.35$ , $\sigma_{1} = 0.1$ , $\sigma_{2} = 0.15$ . + +Optimal Threshold. The optimal $\alpha$ satisfies: + +$$ +f _ {p o s} (\alpha) = f _ {n e g} (\alpha). \tag {10} +$$ + +Substituting Gaussian PDFs: + +$$ +\frac {1}{\sigma_ {1} \sqrt {2 \pi}} e ^ {- \frac {\left(\alpha - \mu_ {1}\right) ^ {2}}{2 \sigma_ {1} ^ {2}}} = \frac {1}{\sigma_ {2} \sqrt {2 \pi}} e ^ {- \frac {\left(\alpha - \mu_ {2}\right) ^ {2}}{2 \sigma_ {2} ^ {2}}}. \tag {11} +$$ + +Taking the natural logarithm: + +$$ +\ln \left(\frac {1}{\sigma_ {1}}\right) - \frac {\left(\alpha - \mu_ {1}\right) ^ {2}}{2 \sigma_ {1} ^ {2}} = \ln \left(\frac {1}{\sigma_ {2}}\right) - \frac {\left(\alpha - \mu_ {2}\right) ^ {2}}{2 \sigma_ {2} ^ {2}}. \tag {12} +$$ + +Rearranging: + +$$ +\frac {(\alpha - \mu_ {2}) ^ {2}}{\sigma_ {2} ^ {2}} - \frac {(\alpha - \mu_ {1}) ^ {2}}{\sigma_ {1} ^ {2}} = 2 \ln \left(\frac {\sigma_ {2}}{\sigma_ {1}}\right). \tag {13} +$$ + +Expanding into a quadratic form: + +$$ +\alpha^ {2} \left(\frac {1}{\sigma_ {2} ^ {2}} - \frac {1}{\sigma_ {1} ^ {2}}\right) + 2 \alpha \left(\frac {\mu_ {1}}{\sigma_ {1} ^ {2}} - \frac {\mu_ {2}}{\sigma_ {2} ^ {2}}\right) + \left(\frac {\mu_ {2} ^ {2}}{\sigma_ {2} ^ {2}} - \frac {\mu_ {1} ^ {2}}{\sigma_ {1} ^ {2}} - 2 \ln \left(\frac {\sigma_ {2}}{\sigma_ {1}}\right)\right) = 0. \tag {14} +$$ + +Define $A\alpha^2 + B\alpha + C = 0$ , where: + +$A = \frac{1}{0.15^2} -\frac{1}{0.1^2} = 44.44 - 100 = -55.56,$ +- $B = 2\left(\frac{0.65}{0.1^2} - \frac{0.35}{0.15^2}\right) = 2(65 - 15.56) = 98.89,$ +- $C = \frac{0.35^2}{0.15^2} - \frac{0.65^2}{0.1^2} - 2\ln \left(\frac{0.15}{0.1}\right) = 5.4444 - 42.25 - 2(0.4055) = -37.6166.$ + +![](images/b2415c7a2f1b034997729f8313a6c5629534a51e1990bd9ce9b3a972cf54374e.jpg) +CLIPScore 3D Surface + +![](images/b547e402e9535a6b277466c7ca7b209ca883af42dabdd9abefd25b5660e014a4.jpg) +Impact of Temperature on VLM Performance +CLIPScore Heatmap + +I2TR@5 Heatmap +Figure 12: Impact of Temperature on VLM Performance. The 3D surface and heatmap peak at temperature=0.7, with CLIPScore reaching 86-87. +![](images/f1cf2c867faca54eedf67267fd8f01ee92f392f0270399b3fcbe11ee3a6af8c3.jpg) +Temperature parameter has a decisive impact on the quality of descriptions generated by VLM. A temperature of 0.7 achieves optimal performance across all three metrics, showing an inverted U-shaped relationship. + +![](images/56ca7499e2ffecd13550739e7471c5d9af17885914b121ff550cb9ddfd9aa064.jpg) +T21 R@5 Heatmap + +Solving: + +$$ +\alpha = \frac {- B \pm \sqrt {B ^ {2} - 4 A C}}{2 A}, \tag {15} +$$ + +$$ +\Delta = 9 8. 8 9 ^ {2} - 4 (- 5 5. 5 6) (- 3 7. 6 1 6 6) = 1 4 2 5. 4 6 2 5, \tag {16} +$$ + +$$ +\sqrt {\Delta} \approx 3 7. 7 5, +$$ + +$$ +\alpha_ {1} \approx 0. 5 5 7, \quad \alpha_ {2} \approx 1. 2 2 4. \tag {17} +$$ + +Since $\alpha_{2} > 1$ is invalid, we select $\alpha = 0.557$ . + +Experimental Validation. We use 10,000 point cloud-text pairs from Objaverse-XL, with 5,000 positive and 5,000 negative pairs (randomly mismatched). Cosine similarities are computed via $E_{p}$ (PointNet++-based) and $E_{t}$ (BERT-based) on an NVIDIA RTX 3090 using PyTorch. + +Distribution Verification We fit truncated Gaussians to $S_{pos}$ and $S_{neg}$ , estimating parameters and performing KS tests. The results are shown in Figure 14, indicating that the estimated parameters closely match our theoretical assumptions, with high KS p-values confirming consistency. + +Threshold Optimization. We compute FNR, FPR, and total error for $\alpha \in [0.4, 0.7]$ , with results detailed in Figure 15. The AUC from ROC analysis is 0.91, and $\alpha = 0.557$ achieves the lowest total error of 0.25, outperforming other thresholds. + +System Performance. We evaluate Tri-MARF performance across $\alpha$ values, focusing on CLIPScore and ViLT R@5, as shown in Figure 16. At $\alpha = 0.557$ , the system achieves a CLIPScore of 88.7 + +![](images/825905919509e3c1e5035a8f674dfbe35991d7394378b00c66137fbef5c6c852.jpg) + +![](images/503c19b93f08ccaeaea212e046462367a83b97f401bbc977498c47e42ef465d7.jpg) + +![](images/ca31fe280fc2f27d3bcb9fc82d66f88d2a33c8f147772e6835f74777722fe527.jpg) +Figure 13: Combined Parameter Analysis: Alpha and Exploration Weight. The surface plot and heatmap confirm alpha=0.1 and exploration_weight=0.5 as the optimal configuration, with CLIPScore around 44.5. + +![](images/cec8c0b7e9d2bdee7ff5c3c183cb7fa83f44a8943e5d675f707e804898f24093.jpg) + +![](images/b232c0198e1f9cd0aa0469f99d90a07a438bcb7990bbec450a41e502baad0a45.jpg) +Figure 14: Distribution Parameters: Theoretical vs. Estimated with KS Test p-values + +and ViLT R@5 of 45.2/43.8. Sensitivity analysis around $\alpha = 0.557 \pm 0.02$ , presented in Figure 17, shows fluctuations below $2\%$ , confirming robustness. + +KL Divergence. We calculate $D_{KL}(P_{pos} \| P_{neg})$ to assess discriminative power, with results in Figure 18. The peak value of 2.30 at $\alpha = 0.557$ supports its optimality. + +Robustness. We test $\alpha = 0.557$ under varied distributions, as shown in Figure 19. Performance remains strong, with CLIPScore dropping only slightly to 87.5 under a more overlapping distribution, aided by Tri-MARF's multi-agent design. + +![](images/c35d5624857acdd02704745bff223d4de6c420a1ce9ff26115102a8fb9a0e271.jpg) +Figure 15: Error Rates Across $\alpha$ + +![](images/ff5ca320882baf73af924b5c44b7395ab07d474204036df1f0b6ad37e7085b8e.jpg) +Figure 16: System Performance Across $\alpha$ + +Baseline Comparison. We compare $\alpha = 0.557$ against baselines in Figure 20. It consistently outperforms alternatives, achieving a CLIPScore of 88.7 versus 86.3 for $\alpha = 0.5$ and 85.7 for no gating. + +Theoretically, $\alpha = 0.557$ balances partially overlapping distributions ( $\mu_1 - \mu_2 = 0.3 < 2\sqrt{\sigma_1^2 + \sigma_2^2} \approx 0.36$ ). Experiments, as detailed in Figures 14 to 20, confirm its efficacy, with minor deviations (e.g., $\alpha = 0.551$ in exact computation) resolved through practical tuning. The architecture of Tri-MARF enhances the robustness of annotation. We derive and validate $\alpha = 0.557$ as an optimal gating threshold in Tri-MARF, supported by rigorous theory and comprehensive experiments across Figures 14 to 20. Future work may explore adaptive thresholds for varying distributions. + +![](images/3e876a87ac53a681299996d65acd029b7d2e6339d3e6c5350c50c096767d7916.jpg) +Sensitivity Analysis at $\alpha = 0.557\pm 0.02$ + +![](images/767190e640f74dfcc84e435b3c9f6f957ae033eb45f7e4ecb3a6d0482de0f91c.jpg) + +Figure 17: Sensitivity Analysis at $\alpha = 0.557\pm 0.02$ +![](images/4ee29897c3eff3fe6b4fc9a45196713492fb6e93feeeabadd7ef0f7f5aabcfd1.jpg) +Sensitivity analysis reveals high stability across a range 0.537-0.577, with performance fluctuations $< 0.7\%$ . The optimal threshold $\alpha = 0.557$ consistently delivers the best results across all evaluated metrics. + +![](images/cb9e3f31a94b1b4d73ef32dd6794619d9e2d4fa956fa6a17543d971c85396340.jpg) + +![](images/a61c1a136afb52be6f428e0355515e7cbd39cbdcd461db4131024c94b1fb13ce.jpg) +KL Divergence (DKL) Across Different α Values +Figure 18: KL Divergence Across $\alpha$ + +# 12 Details of Human Evaluation + +Human evaluations are conducted to validate Tri-MARF's performance in caption quality assessment, type annotation validation, and reinforcement learning strategy selection. All annotators were hired through a crowdsourcing platform and required basic English proficiency and at least one year of experience in image or text annotation. Below are the details of each experiment's methodology, participant recruitment, and evaluation protocols. We obtained local Institutional review board (IRB) approvals before conducting the experiment. + +![](images/d813b2fdcc817f6292d68dc7d7ad30765c4d61efdc4ff84bb52bf857f5d66083.jpg) +Model Robustness Comparison Across Different Parameter Distributions +Figure 19: Robustness Across Distributions + +![](images/c064193ab43bd36db67cb42c816c6bfee5692e427d644a5c072fd420b5492117.jpg) +Figure 20: Baseline Comparison + +# 12.1 Human Evaluation in 3D Captioning Test + +Compare Tri-MARF generated captions against baselines (e.g., Cap3D, ScoreAgg, Human Annotation) via A/B testing. Five annotators were recruited from the crowdsourcing platform. + +Two hundred objects were randomly sampled from each dataset—Objaverse-LVIS (1k), Objverse-XL (5k), and ABO (6.4k)—totaling 600 objects. Annotators evaluated pairs of captions (Tri-MARF vs. a baseline, randomly ordered) on a 1-5 Likert scale for accuracy (object description match), completeness (key feature coverage), and linguistic quality (clarity and grammar). Each annotator assessed 40 pairs per dataset (120 pairs total), with tasks evenly distributed. Scores were averaged across annotators and objects, with Tri-MARF as the reference baseline. The task was completed in five days, with each annotator working 5 hours per day. + +# 12.2 Human Verification in Type Annotation Experiments + +To verify the semantic accuracy of object type classification by Tri-MARF and baselines, as well as automated metrics. Three annotators were hired from a crowdsourcing platform. + +300 objects were randomly selected from Objaverse-LVIS. Annotators received 3D models and renderings of their 6 viewpoints. They observed the 3D objects and selected the category (e.g., "mug" vs. "cup") in a six-choice question, simply selecting the most appropriate option. Each object was reviewed by two annotators, and the third annotator resolved disagreements by majority voting. The + +results established a human annotation baseline. The task was completed in three days, with each annotator working for 2 hours. + +# 12.3 Human Evaluation of Reinforcement Learning Strategies + +Assess annotation quality of RL strategies (e.g., MAB UCB, PPO, MCTS) using a Likert scale. Four annotators were recruited from the crowdsourcing platform. + +Twenty-five annotations per RL strategy (7 strategies, 175 total) were sampled from the Objaverse-XL test set (1,000 objects). Annotators rated each annotation on a 1-10 Likert scale for accuracy (description correctness), completeness (detail inclusion), and fluency (readability). Each annotator evaluated 43-44 annotations, with strategy origins blinded. Scores were averaged to yield the final Likert score. The task was completed in four days, with each annotator working 2.5 hours. + +# 12.4 General Protocol and Quality Control + +- Training: Annotators completed a 15-minute online training module via the platform, using sample objects and annotations to understand criteria. +- Quality Control: Inter-rater reliability was tracked with Cohen's Kappa, achieving 0.76 (substantial agreement). Ratings differing by more than 2 points were reviewed by a platform supervisor, with $5\%$ of responses rechecked for consistency. +Compensation: Annotators were paid $15/hour. At the same time, ensure that no personnel are replaced during the marking period + +# 13 Robustness Evaluation Under Occlusion + +In this section, we investigate the robustness of the Tri-MARF framework when 3D objects are partially occluded, a common challenge in real-world scenarios such as autonomous driving and robotics. To evaluate this, we randomly selected 500 objects from the Objaverse-XL dataset and introduced artificial occlusion by overlaying random black planes on their 3D assets, simulating varying degrees of obstruction. Both the unoccluded and occluded 3D models were processed using our Tri-MARF, with experimental parameters consistent with the main experiments, including the use of Qwen2.5-VL-72B-Instruct for initial annotation, RoBERTa+DBSCAN for clustering, MAB (UCB) with exploration_weight = 0.5, alpha = 0.1, and learning_rate = 0.1 for aggregation, and a point cloud gating threshold alpha = 0.557. The CLIPScore was recorded to compare the quality of generated captions under occluded versus unoccluded conditions. + +The results indicate that Tri-MARF maintains robust performance under occlusion. For unoccluded objects, the average CLIPScore was 86.1, aligning with the main experiment (Table 1). For occluded objects, the CLIPScore dropped to an average of 82.3 (a $4.2\%$ decrease), with variations depending on occlusion severity. This suggests that the multi-agent collaboration, particularly the reinforcement learning-based aggregation and point cloud gating, effectively mitigates the impact of missing visual data by leveraging complementary views and geometric consistency. The figure 21 shows one of our test examples and the output results + +The slight degradation in CLIPScore highlights the challenge of occlusion but demonstrates Tri-MARF's ability to infer missing features, supported by the VLM's multi-turn prompting and the MAB's dynamic selection. Tri-MARF's robustness is evident, suggesting its generalization to occluded scenarios. + +# 14 Additional Details of all the experiments + +This section outlines the comparison models, datasets, and evaluation metrics utilized to evaluate the performance of Tri-MARF in 3D object annotation tasks. These components are selected to provide a robust and comprehensive assessment of our proposed method against existing approaches. + +![](images/74e4a1b66ac391a54bf759c0e111eb49fa0e20f5a426decba8a4dc43f042361c.jpg) +Figure 21: Occlusion experiment demonstration: The object is likely a chair, viewed from behind, with a 3D model showing an upward perspective. It combines wood and metal, featuring a rounded backrest, vertical supports, and a circular base suggesting a swivel or rocking mechanism. The smooth, polished surface indicates it's well-maintained or new. Designed for comfort, it could be a rocking chair suited for indoor use in living rooms, bedrooms, or similar settings. + +# 14.1 Comparison Models + +- Cap3D: Cap3D is a leading model for 3D object captioning that uses multi-view rendering to produce descriptions, serving as a baseline to compare against Tri-MARF's multi-agent collaborative framework. It excels in generating captions from multiple perspectives but lacks the reinforcement learning and point cloud processing capabilities that enhance Tri-MARF's robustness and accuracy. +- ScoreAgg: This model improves captioning accuracy by aggregating scores from multiple views, though it falls short of Tri-MARF's performance due to its inability to handle noisy data effectively. It provides a useful benchmark for evaluating the benefits of Tri-MARF's advanced aggregation strategy. +- ULIP-2: ULIP-2 integrates language and 3D point clouds for enhanced understanding but relies on single-view processing, limiting its generalization compared to Tri-MARF's multi-view, multi-agent approach. It highlights the advantage of our method in achieving superior cross-modal alignment. +- PointCLIP: PointCLIP employs CLIP for feature extraction from point clouds, yet its simplistic aggregation struggles with complex 3D structures, unlike Tri-MARF's sophisticated framework. It serves to demonstrate Tri-MARF's improvement in handling intricate object details. +- 3D-LLM: Combining large language models with 3D data, 3D-LLM offers high-quality captions but is computationally heavy, contrasting with Tri-MARF's efficient, lightweight design. This comparison underscores our method's balance of quality and speed. +- GPT4Point: GPT4Point merges point cloud data with GPT-4 for captioning, but its high latency and weaker cross-modal alignment make it less competitive than Tri-MARF. It illustrates the efficiency gains from our reinforcement learning-based aggregation. +- Human Annotation: Human annotations provide a gold-standard reference for caption quality, though they are slow and costly compared to Tri-MARF's automated, high-throughput approach. Tri-MARF aims to rival or exceed this standard efficiently. + +- Metadata: Dataset metadata offers a basic benchmark for annotation, often lacking the semantic depth Tri-MARF achieves with its contextually rich descriptions. It helps quantify our method's improvement over rudimentary annotations. + +# 14.2 Datasets + +- Objaverse-LVIS: Objaverse-LVIS is a large-scale dataset with richly annotated 3D objects across diverse categories, ideal for testing Tri-MARF's caption quality and type inference accuracy. It challenges models with its variety, ensuring robust evaluation of generalization. +- Objaverse-XL: An expanded version of Objaverse, Objverse-XL includes a vast array of 3D objects, with a 5k-object subset used to assess Tri-MARF's scalability and performance on large-scale data. Its breadth tests the model's ability to handle extensive datasets efficiently. +- ABO: Focused on furniture and household items, ABO's 6.4k real-world objects evaluate Tri-MARF's precision in annotating detailed, specific 3D models. It provides a practical testbed for real-world application scenarios. +- ShapeNet-Core: Containing 51,300 synthetic 3D models across 55 categories, ShapeNet-Core is used to test Tri-MARF's adaptability to different data distributions in cross-dataset experiments. Its structured nature contrasts with noisier real-world datasets. +- ScanNet: ScanNet's 1,513 scanned point clouds of indoor scenes introduce noise and incompleteness, assessing Tri-MARF's robustness in real-world conditions. It challenges the model to perform reliably despite imperfect data. +- ModelNet40: With 12,311 CAD models across 40 categories, ModelNet40 tests Tri-MARF on clean, well-structured 3D data, evaluating performance consistency. Its standardized format complements the diversity of other datasets. + +# 14.3 Evaluation Metrics + +- A/B Testing: Human evaluators score captions on a 1-5 scale to gauge quality and preference, offering a subjective measure of Tri-MARF's alignment with human expectations. It directly assesses user satisfaction with generated annotations. +- CLIPScore: CLIPScore measures semantic alignment between captions and 3D objects using text-image embedding similarity, providing an automated metric for Tri-MARF's accuracy. It ensures objective evaluation of cross-modal consistency. +- ViLT Retrieval (R@5): This metric evaluates Tri-MARF's retrieval accuracy (recall at rank 5) for image-to-text and text-to-image tasks, testing its ability to match queries with correct annotations. It highlights the model's retrieval effectiveness. +- GPT-4o Scoring: Used for type inference, GPT-4o compares predicted labels to ground truth, accounting for synonyms to assess Tri-MARF's semantic accuracy. It offers a nuanced evaluation beyond strict string matching. +- String Matching Accuracy: This metric calculates exact matches between predicted and ground-truth labels, providing a simple yet strict measure of Tri-MARF's type inference precision. It may undervalue semantically correct but lexically different terms. +- BLEU-4: BLEU-4 assesses caption fluency and grammatical correctness by comparing n-gram overlap with reference texts, used here to evaluate Tri-MARF's viewpoint experiment outcomes. It ensures the generated text is linguistically sound. \ No newline at end of file diff --git a/NeurIPS/2025/3D-Agent_ A Tri-Modal Multi-Agent Responsive Framework for Comprehensive 3D Object Annotation/images.zip b/NeurIPS/2025/3D-Agent_ A Tri-Modal Multi-Agent Responsive Framework for Comprehensive 3D Object Annotation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..5df8d688d1ebce4872c44f7e22f70960967d5176 --- /dev/null +++ b/NeurIPS/2025/3D-Agent_ A Tri-Modal Multi-Agent Responsive Framework for Comprehensive 3D Object Annotation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:29448983544d4765da01b9143daaf566b53ee57e6a1a9d85032ec6c49c7524c3 +size 1189225 diff --git a/NeurIPS/2025/3D-Agent_ A Tri-Modal Multi-Agent Responsive Framework for Comprehensive 3D Object Annotation/layout.json b/NeurIPS/2025/3D-Agent_ A Tri-Modal Multi-Agent Responsive Framework for Comprehensive 3D Object Annotation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..8a131b081258158d549011d5829157478deabed2 --- /dev/null +++ b/NeurIPS/2025/3D-Agent_ A Tri-Modal Multi-Agent Responsive Framework for Comprehensive 3D Object Annotation/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cf0abe0662873ae56bdb3747c7ba8bc49ef22b451cc7155461dcac495e23eb81 +size 916717 diff --git a/NeurIPS/2025/3D-GSRD_ 3D Molecular Graph Auto-Encoder with Selective Re-mask Decoding/07703aac-b0e9-44a7-9c91-88716b6109c0_content_list.json b/NeurIPS/2025/3D-GSRD_ 3D Molecular Graph Auto-Encoder with Selective Re-mask Decoding/07703aac-b0e9-44a7-9c91-88716b6109c0_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..51e2133461aa1b2d1ef9c88af196efe777d1b1a2 --- /dev/null +++ b/NeurIPS/2025/3D-GSRD_ 3D Molecular Graph Auto-Encoder with Selective Re-mask Decoding/07703aac-b0e9-44a7-9c91-88716b6109c0_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1cc544eb84f04ebf0efeafbd7eaa69909ab7f1cd3a4770b5b084497168ffc99d +size 161417 diff --git a/NeurIPS/2025/3D-GSRD_ 3D Molecular Graph Auto-Encoder with Selective Re-mask Decoding/07703aac-b0e9-44a7-9c91-88716b6109c0_model.json b/NeurIPS/2025/3D-GSRD_ 3D Molecular Graph Auto-Encoder with Selective Re-mask Decoding/07703aac-b0e9-44a7-9c91-88716b6109c0_model.json new file mode 100644 index 0000000000000000000000000000000000000000..5ca3709761cd3d85d91a5923ca50fe308d9b77f5 --- /dev/null +++ b/NeurIPS/2025/3D-GSRD_ 3D Molecular Graph Auto-Encoder with Selective Re-mask Decoding/07703aac-b0e9-44a7-9c91-88716b6109c0_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f8f83aa22c9ad4af1426d534a2227338914c204076f965d56a9330b50739b663 +size 210841 diff --git a/NeurIPS/2025/3D-GSRD_ 3D Molecular Graph Auto-Encoder with Selective Re-mask Decoding/07703aac-b0e9-44a7-9c91-88716b6109c0_origin.pdf b/NeurIPS/2025/3D-GSRD_ 3D Molecular Graph Auto-Encoder with Selective Re-mask Decoding/07703aac-b0e9-44a7-9c91-88716b6109c0_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..3e8624f813c1d1123ecc593ac3bce0ae0ed87a38 --- /dev/null +++ b/NeurIPS/2025/3D-GSRD_ 3D Molecular Graph Auto-Encoder with Selective Re-mask Decoding/07703aac-b0e9-44a7-9c91-88716b6109c0_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:88d614729fb092b8b7cbc8dc13d718a225978f8ac396861aa3abde4c9ced30a6 +size 1916155 diff --git a/NeurIPS/2025/3D-GSRD_ 3D Molecular Graph Auto-Encoder with Selective Re-mask Decoding/full.md b/NeurIPS/2025/3D-GSRD_ 3D Molecular Graph Auto-Encoder with Selective Re-mask Decoding/full.md new file mode 100644 index 0000000000000000000000000000000000000000..488a370a5b0635cb1bdea04f5c7d8243c2843ec7 --- /dev/null +++ b/NeurIPS/2025/3D-GSRD_ 3D Molecular Graph Auto-Encoder with Selective Re-mask Decoding/full.md @@ -0,0 +1,762 @@ +# 3D-GSRD: 3D Molecular Graph Auto-Encoder with Selective Re-mask Decoding + +Chang Wu $^{1*}$ , Zhiyuan Liu $^{2*}$ , Wen Shu $^{3}$ , Liang Wang $^{4}$ , Yanchen Luo $^{1}$ , Wenqiang Lei $^{3}$ , Yatao Bian $^{2}$ , Junfeng Fang $^{2\dagger}$ , Xiang Wang $^{1\dagger}$ + +1 University of Science and Technology of China 2 National University of Singapore 3 Sichuan University 4 Institute of Automation, Chinese Academy of Sciences + +wuchang0124@mail.ustc.edu.cn, acharkq@gmail.com, shuwen@stu.scu.edu.cn, liang.wang@cripac.ia.ac.cn, luoyanchen@mail.ustc.edu.cn, wenqianglei@scu.edu.cn, ybian@nus.edu.sg, fangjf1997@gmail.com, xiangwang1223@gmail.com + +* Equal contribution. † Corresponding author. + +# Abstract + +Masked graph modeling (MGM) is a promising approach for molecular representation learning (MRL). However, extending the success of re-mask decoding from 2D to 3D MGM is non-trivial, primarily due to two conflicting challenges: avoiding 2D structure leakage to the decoder, while still providing sufficient 2D context for reconstructing re-masked atoms. To address these challenges, we propose 3D-GSRD: a 3D Molecular Graph Auto-Encoder with Selective Re-mask Decoding. The core innovation of 3D-GSRD lies in its Selective Re-mask Decoding (SRD), which re-masks only 3D-relevant information from encoder representations while preserving the 2D graph structures. This SRD is synergistically integrated with a 3D Relational-Transformer (3D-ReTrans) encoder alongside a structure-independent decoder. We analyze that SRD, combined with the structure-independent decoder, enhances the encoder's role in MRL. Extensive experiments show that 3D-GSRD achieves strong downstream performance, setting a new state-of-the-art on 7 out of 8 targets in the widely used MD17 molecular property prediction benchmark. The code is released at https://github.com/WuChang0124/3D-GSRD. + +# 1 Introduction + +Molecular representation learning (MRL) [1-3] is fundamental to a wide range of downstream tasks, including de novo drug design [4], molecular dynamics simulation [5], and molecular property prediction [6, 7]. Given the abundance of unlabeled molecular data in this field, self-supervised pretraining has emerged as a key strategy for learning effective molecular representations. Previous works have primarily focused on 1D molecular strings [8, 9] and 2D molecular graphs [10-12], achieving promising results. However, they often neglect critical 3D structural information, which is crucial for capturing molecular properties such as the highest occupied molecular orbital, molecular dynamics, and energy functions [13]. This limitation has led to a growing interest in incorporating 3D molecular coordinates into pretraining frameworks. + +Masked graph modeling (MGM) has emerged as a leading paradigm for 3D molecular pretraining, aiming to learn data distributions by reconstructing randomly masked graph features [14, 15]. As illustrated in Figure 1, its 3D variant typically consists of three key components: (1) 3D graph masking, which perturbs the original 3D molecular graph by randomly masking features such as 3D coordinates, atom types, and chemical bonds [14, 16]; (2) a 3D graph encoder, which processes the + +![](images/7033c50cac3facada45a55912c18e2d205e39141abc01e3262f7bb5629a7fb78.jpg) +Figure 1: Illustration of 3D MGM with re-mask decoding and 2D structure leakage. + +![](images/a3373478b20c5a9774c6516cc573a5111410ad779c91c7e7b151d6f546dfe776.jpg) +Figure 2: Reconstruction loss across two pretraining settings. We compare settings using frozen encoders (3D-ReTrans) and structure-independent (Transformer) versus structure-dependent (3D-ReTrans) decoders. + +masked graph to generate molecular representations; and (3) a 3D graph decoder, which reconstructs the masked features from the encoded representations. After pretraining via reconstruction, the 3D graph encoder is finetuned on downstream tasks to enhance performance. + +However, a long-standing challenge in MGM, mirroring similar issues in masked image modeling [17], is the misalignment between the reconstruction objective and representation learning [18]. Specifically, minimizing the reconstruction error often leads models to focus on low-level graph features, such as atom types and 3D coordinates, rather than learning high-level graph semantics required for property prediction [19]. To mitigate this, re-mask decoding is introduced for MGM pretraining on 2D molecular graphs [15, 20, 18]. This method re-masks the encoder's representation of previously masked atoms before feeding them to the decoder (Figure 1). In this way, the encoder is prevented from reconstructing the masked atoms directly and is encouraged to focus on generating high-quality representations of the unmasked graph regions, which the decoder then uses for reconstruction. This approach shifts the encoder's focus from graph reconstruction to MRL and yields substantial downstream improvements for 2D MGM. + +To adapt re-mask decoding for 3D MGM, we identify two seemingly contradictory challenges: + +- Leaking 2D structure to decoder weakens encoder's MRL capability. The decoder should rely solely on the encoder representations to reconstruct masked features. Exposing the decoder directly to 2D molecular structures (e.g., chemical bond connections) diminishes the encoder's role in MRL, as the decoder can recover masked features using the provided 2D molecular structures, even with poor encoder representations. For example, Figure 2 shows that a frozen randomly-initialized encoder with a trainable structure-dependent decoder can achieve relatively low reconstruction loss when predicting masked atomic coordinates. Consequently, the encoder focuses less on capturing structures, leading to suboptimal MRL performance. However, existing re-mask decoding methods [15, 18, 20] typically use structure-dependent decoders like graph neural networks [21], which exacerbate this issue. +- Structure-independent decoding can prevent structure leakage, but hinders reconstruction of re-masked atoms. A naive solution to prevent structure leakage is to use a structure-independent decoder, which consumes no 2D structure input. However, this approach fails to account for the relative positions and contextual relationships of re-masked atoms within the 2D graph, making it challenging to distinguish between re-masked atoms during reconstruction. To address this, we propose leveraging the 3D graph encoder to generate 2D structural contexts for re-masked atoms. This ensures that the encoder is effectively trained for structural representation while preventing structure leakage beyond the encoder's representations. + +To address the challenges above, we introduce 3D Molecular Graph Auto-Encoder with Selective Re-mask Decoding (3D-GSRD), a 3D MGM framework with three key elements: (1) the Selective Remask Decoding (SRD) that re-masks only 3D-relevant information from the encoder representations while preserving its 2D structural context; (2) a structure-independent decoder that derives all structural information exclusively from the encoder; and (3) 3D Relational-Transformer (termed as 3D-ReTrans) as 3D graph encoder, effectively integrating 3D molecular features (e.g., atomic coordinates) and 2D features (e.g., bonding connections) for MRL. + +Specifically, SRD ensures the preservation of 2D graph structures during re-masking by reintroducing 2D information through a 2D graph Position Encoder (2D-PE). Crucially, this 2D-PE is trained via distillation from the 3D graph encoder's representations, ensuring its information is fully contained within the 3D encoder, as demonstrated in Section 5. The distillation process also allows the 2D-PE's structural encoding capability to improve alongside the 3D encoder's advancements during pretraining. To complement SRD, we employ a structure-independent Transformer [22] decoder that derives 2D graph structure exclusively from the 3D encoder and the 2D-PE. Together with SRD, our decoder provides the re-masked atoms with rich 2D graph contexts that are distilled from the 3D graph encoder, while preventing 2D structure leakage. + +Encoding 3D molecules is challenging due to their multi-modal nature (e.g., discrete atom types vs. continuous coordinates) and multi-granular structure (e.g., atom-wise vs. pair-wise features). Prior works like PaiNN [23] and TorchMD-NET [24] address this by using equivariant architectures that separately process scalar features (e.g., atom types, distances) and vector features (e.g., directional geometry). Building on these insights, we introduce 3D-ReTrans as our 3D graph encoder, extending the Relational-Transformer [25] to incorporate both scalar and vector features while maintaining its scalability and flexibility to process both atom-wise and pair-wise features. Specifically, we introduce a tailored attention mechanism that incorporates pairwise distances and interactions directly into attention weights, along with a 3D Update Layer that jointly updates scalar and vector features. This design yields strong performance on MRL and serves as a robust backbone for 3D MGM pretraining. + +Finally, we include in-depth analysis to showcase the inner mechanism of SRD and our structure-independent decoder, demonstrating our key claims of shifting the encoder's focus to MRL while preventing structure leakage in the decoder. Based on these revealed advantages in MGM pretraining, 3D-GSRD demonstrates superior performance when being fine-tuned on downstream datasets, achieving new state-of-the-art on 7 out of 8 molecules for MD17 [26]. + +# 2 Related Work + +Molecular pretraining has emerged as a fundamental approach for molecular representation learning [27-29], critical for various downstream tasks, such as molecule property prediction. + +3D Molecular Denoising and Masked Graph Modeling. Recent advances in 3D molecular pretraining have focused on 3D structure learning through coordinate denoising and masking. For example, [30-32] introduce noise to atomic coordinates and then reconstruct them. SubGDiff [2] adds distinct Gaussian noise to different substructures of 3D molecular conformation and performs denoising via a diffusion process. MolSpectra [33] uses the energy spectra to enhance 3D molecular representation learning during denoising. As for masking, Uni-Mol [14] and Uni-Mol2 [16] employ masked coordinates prediction as one of the self-supervised tasks, while other works like [1] focus on masking and predicting bond lengths and angles. + +Other 3D Molecular Pretraining Methods. EPT [34] proposes a multi-domain 3D pretraining approach by combining atom-level features for small molecules and residue-level features in proteins. 3D PGT [35] designs three generative pretraining tasks, including predicting bond length, bond angle, and dihedral angle, and introduces an adaptive fusion strategy for these tasks, using total energy as a surrogate metric to optimize their combination weight. GraphMVP [36] and 3D Infomax [7] use contrastive learning to transfer knowledge from the 3D encoder into the 2D graph encoder. + +Graph Position Encoding and Structure Encoding. Position Encoding (PE) encodes the spatial position of a given node within a graph [37]. Some methods [38, 39] use adjacency, Laplacian, or distance matrices to represent PE. Other approaches like [40-42] leverage shortest paths, heat kernels, or Green's function to compute pair-wise distance, capturing the distance and directional relationships between nodes. Currently, in MOL-AE [19], SMILES strings are used to provide PE to the decoder as an identifier. Structure Encoding (SE) encodes the structural information of graphs and subgraphs. Common methods include node degree [43], Laplacian matrices [38], and Boolean indicators that specify whether two nodes belong to the same substructure [44]. Unlike these methods, we propose using a 2D graph position encoder distilled from a 3D graph encoder to produce SE, which provides rich and effective context information. + +2D Molecular Graph Pretraining. Previous methods primarily focus on leveraging 2D molecular graphs to learn molecular representation. A popular technique is masked graph modeling [15, 8, 10], + +![](images/da22d3d4b36f5dc0aee3502bd4e61867caad8bfeed97b2ade6c06ad5438e56ef.jpg) +Figure 3: Overview of 3D-GSRD. It contains three key elements: (1) a 3D-ReTrans encoder; (2) the SRD that re-masks only 3D-relevant information from the encoder representations while preserving its 2D structure information via 2D-from-3D distillation; (3) a structure-independent decoder. + +typically comprising three key components [18]: graph tokenizer [11, 3], graph masking [45, 46], and graph autoencoder [47-49]. Another prominent line of work is contrastive learning [50, 12, 51], which aims to pull positive pairs and push negative pairs apart in the representation space. Notably, methods such as GeomGCL [6], GraphMVP [36], and 3D Infomax [7] incorporate 3D molecular conformations as auxiliary information to enhance 2D graph representations via contrastive objectives. While effective for 2D molecular pretraining, these methods overlook 3D features, which are crucial for molecular representation learning. Moreover, directly extending these methods to 3D molecular pretraining is non-trivial due to the increased complexity and spatial nature of 3D molecular data. + +Our method is similar to GraphMVP [36] in leveraging both 2D and 3D molecular graphs but differs in objective and design. While GraphMVP transfers 3D information into a 2D encoder to enhance 2D graph representations, our method distills 3D representation into 2D-PE, ensuring the 2D-PE's embedding is fully contained within the 3D encoder to avoid structure leakage in decoding. Additionally, GraphMVP aligns the 2D and 3D views of the same molecule and contrasts views of different molecules using contrastive losses. Our method instead uses a cosine similarity loss that encourages 2D-PE to generate structural encodings closely aligned with the 3D encoder. This offers a simpler and more efficient framework for 2D structure-informed decoding without structure leakage. + +# 3 Preliminary: 3D Masked Graph Modeling + +Notations. A 3D molecular graph is represented as $G = (\mathbf{x}, \mathbf{a}, \mathbf{e})$ , where $\mathbf{x} \in \mathbb{R}^{N \times 3}$ denotes the 3D atomic coordinates, $\mathbf{a} \in \mathbb{R}^{N \times *}$ represents the atom types, and $\mathbf{e} \in \mathbb{R}^{N \times N \times *}$ captures the atomic pair features, such as inter-atomic distances and bonds. $N$ is the number of atoms. + +Graph Masking. Given a molecular graph $G$ , the 3D coordinates $\{\mathbf{x}_i \in \mathbb{R}^3 | i \in \mathcal{V}_m\}$ of a randomly selected subset of atoms $\mathcal{V}_m$ are masked. For each masked atom $i \in \mathcal{V}_m$ , its original coordinates $\mathbf{x}_i$ is replaced by a learnable special token $\mathbf{m}_x \in \mathbb{R}^3$ . The coordinate matrix $\mathbf{x}$ after masking is denoted as $\tilde{\mathbf{x}}$ , and the masked graph is denoted as $\tilde{G} = (\tilde{\mathbf{x}}, \mathbf{a}, \mathbf{e})$ . Some prior works [19] instead remove all information corresponding to the masked atoms, including their 3D coordinates, atom types, and pairwise features, which results in a masked graph $\tilde{G} = (\tilde{\mathbf{x}}, \tilde{\mathbf{a}}, \tilde{\mathbf{e}})$ . In this work, we adopt this latter masking strategy, fully excluding the masked atoms from the input graph. + +3D Graph Auto-Encoder and 2D Structure Leakage. The 3D graph auto-encoder comprises a graph encoder $\phi_e(\cdot)$ and a graph decoder $\phi_d(\cdot)$ . The encoder processes the masked graph $\tilde{G}$ to produce graph representations $\mathbf{h} = \phi_e(\tilde{G}) \in \mathbb{R}^{N \times *}$ . The decoder then predicts the masked coordinates $\{\hat{\mathbf{x}}_i | i \in \mathcal{V}_m\}$ using $\phi_d(\mathbf{h})$ , and optionally incorporating the pair features $\phi_d(\mathbf{h}, \mathbf{e})$ . However, using pair features $\mathbf{e}$ introduces 2D structure leakage, as the decoder relies on additional information beyond the encoder's representation $\mathbf{h}$ . This weakens the encoder's role in MRL, because the decoder can leverage pair features to compensate for any deficiencies in the encoder's representations. Despite this drawback, such leakage is common in previous MGM works [15, 20, 18] that utilize Graph Neural Networks [21] as decoders. In contrast, methods that avoid 2D structure leakage [14, 16] mostly use weak decoders, such as MLPs, which can lead to suboptimal MGM pretraining. + +Re-mask Decoding. Before passing $\mathbf{h}$ into the decoder, re-mask decoding replaces the representations of the previously masked atoms $\mathcal{V}_m$ with a learnable token $\mathbf{m}_h$ , preventing the encoder from directly + +![](images/85851c1354ebff720afb769dcc9afe07f788da8df251e0864fa1ca84f9887acd.jpg) +(a) 3D-ReTrans + +![](images/3de67c38de44f6e70e8d248628365ef96ee88c1b84d97cbbb024d76df0493017.jpg) +(b) 3D Relational-Attention +Figure 4: Illustration of 3D-ReTrans. (a) 3D-ReTrans is constructed by stacking multiple 3D Relational-Attention and 3D Update Layers. (b) 3D Relational-Attention that processes both atom-wise and pair-wise features. (c) 3D Update Layer that includes a residual connection. + +![](images/81f438fea3ca617b5877513dc1d05affd2d547dd64cbcbef539812369ada7951.jpg) +(c) 3D Update Layer + +predicting the masked coordinates. This encourages the encoder to focus on learning meaningful representations for the unmasked graph regions. The re-masked representation $\tilde{\mathbf{h}}$ is defined as: + +$$ +\tilde {\mathbf {h}} _ {i} = \operatorname {r e - m a s k} \left(\mathbf {h} _ {i}\right) = \left\{ \begin{array}{l l} \mathbf {m} _ {h}, & \forall i \in \mathcal {V} _ {m}, \\ \mathbf {h} _ {i}, & \text {o t h e r w i s e .} \end{array} \right. \tag {1} +$$ + +MGM Loss. The pretraining objective minimizes the mean squared error between the ground truth coordinates $\{\mathbf{x}_i | i \in \mathcal{V}_m\}$ of the masked atoms and the decoder's predicted coordinates $\{\hat{\mathbf{x}}_i | i \in \mathcal{V}_m\}$ : + +$$ +\mathcal {L} _ {\mathrm {M G M}} = \sum_ {i \in \mathcal {V} _ {m}} \| \hat {\mathbf {x}} _ {i} - \mathbf {x} _ {i} \| ^ {2}. \tag {2} +$$ + +# 4 Methodology: 3D-GSRD + +In this section, we present our method 3D Molecular Graph Auto-Encoder with Selective Re-mask Decoding (3D-GSRD) (Figure 3). Below, we start by elaborating on the Selective Re-mask Decoding and the pretraining objective of 3D-GSRD. We then describe the encoder of 3D-ReTrans. + +# 4.1 SRD: Selective Re-mask Decoding + +Here we introduce SRD to improve 3D MGM. Re-mask decoding is proposed to address the mismatch between the reconstruction objective of 2D MGM and MRL [15, 18]. Our SRD extends this approach to 3D MGM while overcoming issues such as 2D structure leakage and providing 2D contexts to re-masked atoms. + +Re-mask Decoding with 2D Graph Position Encoder. As Figure 3 shows, given the encoder representation $\mathbf{h} = \phi_e(\tilde{\mathbf{x}},\tilde{\mathbf{a}},\tilde{\mathbf{e}})$ of $\tilde{G}$ , SRD can be defined as: + +$$ +\operatorname {S R D} (\mathbf {h}, \tilde {G}) = \operatorname {r e - m a s k} (\mathbf {h}) + \operatorname {s t o p - g r a d} \left(\phi_ {2 d} (\mathbf {a}, \mathbf {e})\right), \tag {3} +$$ + +where $\mathrm{SRD}(\mathbf{h},\tilde{G})$ is directly fed into the decoder for masked prediction; re-mask $(\cdot)$ is the standard re-mask; and stop-grad $(\cdot)$ stops the gradient flow to the 2D graph position encoder $\phi_{2d}$ , which generates $\tilde{G}$ 's 2D representation $\phi_{2d}(\mathbf{a},\mathbf{e})\in \mathbb{R}^{N\times *}$ . + +Building a 2D Graph Position Encoder without Structure Leakage via 2D-from-3D Distillation. The 2D graph position encoder $\phi_{2d}(\mathbf{a},\mathbf{e})$ is the key component of SRD. For unmasked atoms, $\phi_{2d}$ conveys the same information as the 3D graph encoder $\phi_e$ , preventing any information leakage beyond what $\phi_e$ has captured. For re-masked atoms, $\phi_{2d}$ offers the necessary 2D contexts that would have been available from $\phi_e$ without re-masking, enabling the decoder to distinguish the relative positions of re-masked atoms. To prevent information leakage, $\phi_{2d}$ is trained exclusively through knowledge distillation from the 3D encoder, without any gradient updates from the MGM loss. This is enforced by the stop-grad $(\cdot)$ , which blocks the gradient flow from the MGM loss into $\phi_{2d}$ . Further, the knowledge distillation loss for $\phi_{2d}$ can be written as: + +$$ +\mathcal {L} _ {\text {d i s t i l l}} = - \sum_ {i \notin \mathcal {V} _ {m}} \cos \left(\phi_ {2 d} (\mathbf {a}, \mathbf {e}) _ {i}, \operatorname {s t o p - g r a d} \left(\phi_ {e} (\mathbf {x}, \mathbf {a}, \mathbf {e}) _ {i}\right)\right), \tag {4} +$$ + +where $\cos (\cdot ,\cdot)$ denotes cosine similarity. The loss applies only to unmasked atoms for the consistency of $\phi_{2\mathrm{d}}$ 's training objective. We employ stop-grad(·) to prevent updating the 3D graph encoder $\phi_e$ allowing it to focus on MRL. + +While the 2D graph position encoder can be any graph encoder, we implement it as a 2D-ReTrans, a simplified version of the 3D-ReTrans that excludes 3D coordinates and distance inputs. Our experiments demonstrate the effectiveness of this design. + +# 4.2 Pretraining 3D-GSRD + +3D Graph Auto-Encoder. We employ the 3D-ReTrans as encoder and employ SRD with a structure-independent decoder of transformer [22]. In this way, we avoid structure leakage to the decoder beyond the encoder's information and provide strong 2D structural contexts for the re-masked atoms. + +Pretraining Loss. We combine the MGM loss and 2D-from-3D distillation loss for pretraining: + +$$ +\mathcal {L} _ {\text {p r e t r a i n}} = \mathcal {L} _ {\mathrm {M G M}} + \mathcal {L} _ {\text {d i s t i l l}}. \tag {5} +$$ + +# 4.3 3D Relational-Transformer + +Encoding 3D molecular graphs $G = (\mathbf{x}, \mathbf{a}, \mathbf{e})$ presents significant challenges of processing 3D coordinates $\mathbf{x}$ while preserving 3D equivariance and integrating the pairwise features $\mathbf{e}$ , whose shape $(N, N, *)$ differs from the atomic features $(N, *)$ . Prior works have primarily focused on ensuring 3D equivariance while processing the 3D coordinates $\mathbf{x}$ [24, 52], but have paid less attention to effectively incorporating pair features $\mathbf{e}$ . Most existing methods [14, 52, 53] incorporate pairwise representations of scalar values in self-attention layers [22], limiting their ability to capture the high-dimensional nature of inter-atomic interactions. While TorchMD-NET [24] models distances as high-dimensional pairwise features, extending it to include chemical bonds remains challenging. + +To address the challenges, we propose 3D-ReTrans as our encoder, leveraging the Relational-Transformer's [25] scalability and flexibility to incorporate pair features, while enabling it to process 3D coordinates. Draw inspiration from prior works [24, 23, 54], a core design is to explicitly separate and jointly process two types of features: (1) scalar features, which encode scalar information like atom types and distances; (2) vector features, which capture directional geometric information. Based on this, our key enhancements focus on improving its attention mechanism and incorporating a 3D update layer. More details about 3D-ReTrans are provided in Appendix B. + +3D Relational-Attention. Each atom is represented by concatenating its types and coordinates: $\mathbf{n}_i = [\mathbf{a}_i;\mathbf{x}_i]$ . The interaction between atoms $i$ and $j$ is captured by the pair feature $\mathbf{e}_{ij}\in \mathbb{R}^d$ and their Euclidean distance $r_{ij}$ . 3D Relational-Attention is defined as: + +$$ +\mathbf {q} _ {i j} = \left[ \mathbf {n} _ {i}, \mathbf {e} _ {i j} \right] \mathbf {W} ^ {q}, \tag {6} \quad \left[ \mathbf {k} _ {i j}; \mathbf {v} _ {i j} \right] = \left[ \mathbf {n} _ {j}; \mathbf {e} _ {i j} \right] \left[ \mathbf {W} ^ {k}; \mathbf {W} ^ {v} \right], \tag{7} +$$ + +$$ +\left[ \mathbf {d} _ {i j} ^ {k}; \mathbf {d} _ {i j} ^ {v} \right] = \operatorname {S i L U} \left(\left[ \mathbf {W} ^ {d k}; \mathbf {W} ^ {d v} \right] e ^ {\mathrm {R B F}} \left(r _ {i j}\right)\right), \tag {9} +$$ + +$$ +\boldsymbol {\alpha} _ {i j} = \operatorname {S i L U} _ {j} \left(\frac {\mathbf {q} _ {i j} \cdot \left(\mathbf {k} _ {i j} \odot \mathbf {d} _ {i j} ^ {k}\right)}{\sqrt {d}}\right), \quad (1 0) \quad \left[ \mathbf {o} _ {i} ^ {1}, \mathbf {o} _ {i} ^ {2}, \mathbf {o} _ {i} ^ {3} \right] = \mathbf {W} ^ {f} \left(\sum_ {j = 1} ^ {N} \boldsymbol {\alpha} _ {i j} \mathbf {s} _ {i j} ^ {3}\right), \tag {11} +$$ + +where $e^{\mathrm{RBF}}(\cdot):[0,\infty)\to \mathbb{R}^d$ is a distance expansion function to encode the distance variable into a $d$ -dimensional vector [24] (see Appendix B). The terms $\mathbf{W}^q, \mathbf{W}^k, \mathbf{W}^v, \mathbf{W}^{dk}, \mathbf{W}^{dv}$ , and $\mathbf{W}^f$ are learnable linear projectors, and $\odot$ denotes element-wise product. The output scalar features $\mathbf{o}_i^1, \mathbf{o}_i^2$ , and $\mathbf{o}_i^3$ encode pairwise interactions and interatomic distances. Moreover, the attention mechanism facilitates the integration of distance information into the vector features via scalar filters $\mathbf{s}_{ij}^1$ and $\mathbf{s}_{ij}^2$ within the subsequent 3D Update Layer. + +3D Update Layer. The 3D Update Layer facilitates information exchange between scalar and vector features. Vector features $\mathbf{vec}_i\in \mathbb{R}^{* \times 3}$ are initialized as zeros and jointly updated with scalar features $x_{i}$ . The update $\Delta x_{i}$ and $\Delta \mathbf{vec}_i$ are defined as: + +$$ +\left[ \mathbf {u} _ {i} ^ {1}, \mathbf {u} _ {i} ^ {2}, \mathbf {u} _ {i} ^ {3} \right] = \mathbf {W} ^ {v} \left(\mathbf {v e c} _ {i}\right), \quad (1 2) \quad \mathbf {w} _ {j} = \sum_ {j = 1} ^ {N} \left(\mathbf {v e c} _ {j} \odot \mathbf {s} _ {i j} ^ {1}\right) + \mathbf {s} _ {i j} ^ {2} \odot \frac {\mathbf {r} _ {i} - \mathbf {r} _ {j}}{\| \mathbf {r} _ {i} - \mathbf {r} _ {j} \|}, \tag {13} +$$ + +$$ +\Delta x _ {i} = \mathbf {o} _ {i} ^ {2} + \mathbf {o} _ {i} ^ {3} \odot \left(\mathbf {u} _ {i} ^ {1} \cdot \mathbf {u} _ {i} ^ {2}\right), \quad (1 4) \quad \Delta \mathbf {v e c} _ {i} = \mathbf {u} _ {i} ^ {3} \odot \mathbf {o} _ {i} ^ {1} + \mathbf {w} _ {j}, \tag {15} +$$ + +![](images/3d4f58f4648d66d98f19fc7ffdad5fda6379bcc2de42102b05b75ae8eb2cbf41.jpg) +Figure 6: Probe encoder for masked atom coordinates when pretrained with/without SRD. + +![](images/9db12e747bd3c8396f5a51d25578184a952c279fb17f0583e2f4c03f7c9d85ef.jpg) +Figure 5: Reconstruction loss across four pretraining settings. + +![](images/08b6482547f2c9bd239d87000d22283510e9ee2b7bd61d139aaf4ba2775ed028.jpg) +Figure 7: Reconstructing 2D-PE representation using 3D encoder representation. + +where $\mathbf{W}^v$ are learnable linear projectors. Scalar features incorporate vector information via elementwise multiplication with the scalar product of vector components, while vector features are updated using both directional features $\mathbf{w}_j$ and a scalar filter $\mathbf{o}_i^1$ . + +As shown in Figure 4, the 3D-ReTrans is constructed by stacking multiple 3D Relational-Attention and 3D Update Layers. Each layer performs residual updates on both scalar features $\Delta x_{i}$ and vector features $\Delta \mathbf{vec}_i$ , allowing the model to simultaneously capture scalar properties (e.g., atom types, distances) and expressive directional geometric information. For pretraining, the final outputs $x_{i}$ and $\mathbf{vec}_i$ are fed into the decoder, while for finetuning, they are passed to the prediction head. + +Learning 3D Equivariance and Invariance by Data Augmentations. Considering that our model lacks built-in 3D equivariance or invariance, we leverage data augmentations to instill these symmetries, following AlphaFold3 [53]. During MGM pretraining, atomic coordinates $\mathbf{x}$ are randomly rotated using transformations sampled from the SO(3) group and translated with offsets drawn from $\mathrm{t} \sim \mathcal{N}(\mathbf{0}, 0.01\mathbf{I}_3)$ . These augmentations encourage the model to adjust its predictions equivariantly with any rotations and small translations. During fine-tuning for property prediction, the same augmentations are applied, but the model is trained to predict consistent properties, thereby learning invariance to rotations and translations. + +We favor data-augmented equivariance with relational attention over fully E(3)-equivariant message passing. While built-in equivariant architectures offer formal guarantees, they often restrict how pair features are parameterized and incur non-trivial computational overhead (e.g., tensor bases, spherical harmonics), which can hinder scaling and complicate integration with diverse molecular cues. In contrast, our approach instills rotational and translation robustness through augmentations, allowing encoder to operate with lightweight vector and scalar updates and to flexibly ingest high-dimensional pair features without architectural surgery. This yields a plug-and-play backbone that is easier to optimize, accommodates sparsity and density variations, and remains representation-rich: relational attention can expand or swap pair features as downstream tasks evolve, while maintaining competitive robustness to pose changes at a substantially lower training and inference cost. + +# 5 Analyzing Selective Re-mask Decoding and Structure-Independent Decoder + +In this section, we conduct extensive experiments to evaluate the components of 3D-GSRD, focusing on the structure-independent decoder and SRD, including its key part 2D-from-3D distillation. We analyze their effects on the overall performance of the 3D-GSRD framework and how they contribute to 3D MGM pretraining. + +Analysis 1. The structure-dependent decoder can diminish the encoder's role in MRL. We pretrain the auto-encoder framework under four settings, all using 3D-ReTrans as the encoder: (1) a frozen encoder with a Transformer decoder; (2) a frozen encoder with a 3D-ReTrans decoder; (3) a trainable encoder with a Transformer decoder; and (4) a trainable encoder with a 3D-ReTrans decoder. Figure 5 reports the reconstruction loss of masked atom coordinates, averaged over every 50 batches. When the encoder is frozen, the Transformer decoder (i.e., structure-independent decoder) struggles to reconstruct masked atoms coordinates, yielding high reconstruction loss due to poor input representations and the absence of 2D molecular structural information. In contrast, with the 3D-ReTrans decoder (i.e., structure-dependent decoder), which leverages 2D molecular structures as input, the loss decreases rapidly during pretraining, even with a frozen encoder. This demonstrates that a powerful, structure-dependent decoder can compensate for weak encoder representations, diminishing the encoder's role in MRL. + +The structure-independent decoder heavily relies on high-quality encoder representations. When paired with a trainable encoder, the structure-independent decoder achieves a much lower reconstruction loss, indicating that it relies heavily on the encoder to provide informative representations. In contrast, with a structure-dependent decoder, the loss remains low regardless of the encoder representation's quality, highlighting that such decoders reduce the learning pressure on the encoder and can hinder its ability to learn meaningful molecular features. + +Analysis 2. Structure-independent decoder improves downstream performance compared to structure-dependent decoder. We pretrain the auto-encoder using either a structure-dependent or a structure-independent decoder while keeping all other settings constant, then finetune the pretrained encoder on molecular property prediction tasks. As Table 1 shows, the structure-independent decoder achieves better performance, outperforming the structure-dependent decoder by $5\%$ in Toluene. This demonstrates that using a structure-independent decoder encourages the encoder to learn more informative representations, leading to improved performance on downstream tasks. + +Analysis 3. 2D-PE and 2D-from-3D distillation boost downstream performance. We pretrain the auto-encoder with and without 2D-PE and 2D-from-3D distillation, followed by finetuning the encoder on MD17 datasets. As shown in Table 1, combining 2D-PE and distillation consistently improves performance. In contrast, using 2D-PE alone leads to degradation, likely due to unintended leakage of 2D structural information into the decoder. Moreover, 2D-from-3D distillation guides the 2D-PE to focus on encoding + +positional information for re-masked tokens, rather than learning molecular representation, which allows the 3D graph encoder to better specialize in MRL. + +Table 1: Analyzing the decoder and SRD. Performance (MAE ↓) on MD17. The variant without SRD and $\mathcal{L}_{\mathrm{distill}}$ corresponds to the model ablated without 2D-PE. + +
DecoderSRDLdistillSalicylicTolueneUracil
Structure-dependent✓✓0.04010.02910.0334
Structure-independent✗✗0.04040.02920.0328
Structure-independent✓✗0.04160.02930.0329
Structure-independent✓✓0.03870.02750.0315
+ +Analysis 4. SRD prevents the encoder representation from containing information about 3D coordinates. To examine whether the encoder captures detailed 3D coordinate information, we train an MLP probe to predict masked atom coordinates from the encoder's representations. We compare the reconstruction loss for encoders pretrained with and without SRD. As shown in Figure 6, the reconstruction loss for the encoder pretrained with SRD is much higher, suggesting that SRD suppresses direct encoding of 3D coordinate details. This forces the encoder to focus on learning higher-level molecular representations that are better aligned with downstream tasks. + +Analysis 5. 2D-PE produces 2D structural context without introducing information leakage. To assess whether 2D-PE introduces additional information beyond what the 3D graph encoder already captures, we train an MLP to reconstruct the 2D-PE representation from the 3D graph encoder representation and compute the reconstruction error, measured by the cosine similarity between the two representations. During this process, both the 3D graph encoder and 2D-PE are frozen, and only the MLP is trainable. Figure 7 shows that the cosine similarity is very close to 1.0, indicating that the context 2D-PE provided is mostly contained by the 3D graph encoder. + +Analysis 6. 2D-PE encodes structural information for decoding. We probe whether the pretrained 2D-PE captures structural information required by the decoder. Specifically, we freeze the 2D-PE and train two MLP classifiers to predict atom and bond types from its outputs. Both tasks achieve prediction accuracy above $99.99\%$ , demonstrating that the 2D-PE indeed encodes the structural information necessary for decoding. + +# 6 Experiments + +# 6.1 Experimental Setup + +Datasets. For pretraining, we use a large-scale molecular dataset PCQM4Mv2 [55], which contains approximately 3.37 million equilibrium 3D molecular graph structures. For downstream tasks, we evaluate our model on two widely used molecular property prediction datasets: QM9 [13] and MD17 [26]. Specifically, QM9 is a quantum chemistry dataset comprising 134k small molecules, each with its equilibrium conformation and 12 molecular properties (e.g., homo, lumo, dipole moment etc.) calculated using density functional theory (DFT). Following prior works [30, 31], we split + +Table 2: Performance (MAE $\downarrow$ ) on MD17 force prediction. The best results are bold. The second-best results are underline. Results marked with * are reproduced by us. + +
ModelsAspirinBenzeneEthanolMalonaldehydeNaphthaleneSalicylicTolueneUracil
TorchMD-NET0.12160.14790.04920.06950.03900.06550.03930.0484
3D-EMGP0.15600.16480.03890.07370.08290.11870.06190.0773
3D-EMGP(TorchMD-NET)0.11240.14170.04450.06180.03520.05860.03850.0477
Frad*0.08250.13550.04320.05350.04310.05690.04330.0482
3D-ReTrans0.07260.16190.05560.06590.04230.05230.04170.0427
3D-GSRD0.05830.14350.03550.04680.02660.03560.02740.0292
+ +Table 3: Performance (MAE ↓) on QM9. The best results are bold. The second-best results are underline. Results marked with * are reproduced by us. + +
Modelsμ(D)α(a03)homo (meV)lumo (meV)gap (meV)<R2>(a02)ZPVE (meV)U0 (meV)U (meV)H (meV)G (meV)Cv(cal molK)
Uni-Mol20.0890.305---5.26-----0.144
SchNet0.0330.23541.034.063.00.071.7014.0019.0014.0014.000.033
E(n)-GNN0.0290.07129.025.048.00.111.5511.0012.0012.0012.000.031
DimeNet++0.0300.04324.619.532.60.331.216.326.286.537.560.023
PaiNN0.0120.04527.620.445.70.071.285.855.835.987.350.024
SphereNet0.0250.04522.818.931.10.271.126.266.366.337.780.022
ComENet0.0250.04523.119.832.40.2591.206.596.826.867.980.024
TorchMD-NET0.0110.05920.318.636.10.0331.846.156.386.167.620.026
3D-ReTrans0.0160.05522.017.838.00.3411.856.186.366.517.890.029
Transformer-M0.0370.04117.516.227.40.0751.189.379.419.399.630.022
SE(3)-DDM0.0150.04623.519.540.20.1221.316.926.997.097.650.024
3D-EMGP0.0200.05721.318.237.10.0921.388.608.608.709.300.026
Coord0.0160.05217.714.731.80.4501.716.576.116.456.910.020
Frad*0.0120.04515.413.730.60.4281.5615.8814.6714.8713.520.023
SliDe*0.0150.05018.716.228.80.6061.7810.0510.7911.3411.800.025
Mol-AE*0.1520.434---6.962-----0.215
Uni-GEM0.0190.06020.916.734.5------0.023
3D-GSRD0.0090.03818.014.531.10.0471.385.485.675.846.900.020
+ +the dataset into 11,000/1,000/10,831 molecules for training, validation, and testing, respectively. MD17 provides simulated dynamical trajectories for 8 small molecules, including their energy, forces, and conformations. During finetuning, our model first predicts the molecular energy and subsequently derives the forces using the relationship $F = -\nabla_{r}E$ , where $r$ represents the 3D coordinates. For finetuning, we split the dataset into 9500/950 samples for training and validation, and use the remaining samples for testing. + +Baselines. To evaluate the effectiveness of our proposed framework, we adopt state-of-the-art 3D molecular pretraining methods and supervised models for molecular property prediction as baselines. For 3D molecular pretraining methods, we include Transformer-M [52], SE(3)-DDM [56], 3D-EMGP [57], Coordinate Denoising [32], Fractional Denoising [30], Sliced Denoising [31], UniGEM [58], Mol-AE [19]. For supervised models, we include SchNet [59], E(n)-GNN [60], DimeNet [61], DimeNet++ [62], PaiNN [23], SphereNet [63], TorchMD-NET [24], Uni-Mol2 [16], ComENet [64]. We also include the results of training our backbone (i.e., 3D-ReTrans) from scratch to evaluate the effectiveness of our pretraining methods. We reproduce the results for Frid, SliDe and Mol-AE, while the results for other baselines are directly taken from the referenced papers. More details about baselines and implementation are provided in Appendix D. + +# 6.2 Results on MD17 + +The MD17 dataset contains diverse non-equilibrium molecular structures that are highly sensitive to geometry, making it a challenging benchmark for 3D MRL. As shown in Table 2, 3D-ReTrans achieves performance comparable to TorchMD-NET and surpasses 3D-EMGP on 7 of 8 molecules, demonstrating its strength as a 3D graph encoder for 3D MGM. Moreover, 3D-GSRD attains state-of-the-art results on 7 of 8 molecules except Benzene, exceeding the strongest baseline (i.e., Frad) by a large margin. These results confirm the effectiveness of our pretraining method. + +Table 4: Ablation on 3D-ReTrans components. Performance (MAE $\downarrow$ ) on QM9. + +
Model Componentshomolumozpve
Relational-Transformer27.724.01.97
+3D Data Augmentation24.623.21.92
+3D Relational-Attention23.420.31.90
+3D Update Layer (3D-ReTrans)22.017.81.85
+ +Table 5: Analyzing SRD on the Relational-Transformer. Performance (MAE ↓) on MD17. + +
DecoderSRDL_distillTolueneUracil
Structure-dependent✓✓0.11440.0813
Structure-independent✗✗0.12500.0828
Structure-independent✓✗0.09980.0843
Structure-independent✓✓0.07450.0733
+ +# 6.3 Results on QM9 + +We also evaluate the effectiveness of 3D-ReTrans and our pretraining strategy on the QM9 dataset, as shown in Table 3. 3D-ReTrans achieves performance comparable to TorchMD-NET, validating the effectiveness of our proposed backbone architecture. Moreover, 3D-GSRD sets a new state-of-the-art on 7 out of 12 properties, surpassing most baselines, including methods with and without pretraining. These results demonstrate that 3D-GSRD is a highly effective pretraining strategy for MRL, offering advantages over coordinate denoising based approaches. + +# 6.4 Ablation Studies and Analysis + +Ablation on Backbone. To assess the effectiveness of our improvements to the Relational-Transformer [25], we perform ablation studies on each component of 3D-ReTrans, as summarized in Table 4. The results show that 3D Relational-Attention, the 3D Update Layer, and 3D data augmentation each enhance molecular property prediction, collectively boosting overall performance. + +Generalization of SRD Across 3D Graph Encoders. To evaluate the generalization of SRD, we replace 3D-ReTrans with the Relational-Transformer and conduct additional experiments. As shown in Table 5, incorporating SRD consistently improves downstream performance, demonstrating its effectiveness as a general pretraining strategy applicable to diverse 3D graph encoder architectures. + +Ablation on 2D Graph Position Encoder. We compare our 2D-PE against alternative structural embeddings, such as those in GraphGPS [37]. Following prior results, we adopt RWSE [65] as a representative baseline due to its strong performance on ZINC and PCQM4Mv2 with relatively low computational cost. As shown in Table 6, replacing 2D-PE with RWSE leads to consistently lower performance, confirming the advantage of 2D-PE in providing 2D structural context. + +Table 6: Ablation on 2D graph position encoder. Performance (MAE $\downarrow$ ) on MD17. + +
2D EncoderSalicylicUracil
RWSE0.03680.0310
2D-PE0.03560.0292
+ +# 7 Conclusion and Future Works + +In this work, we introduce 3D-GSRD, a 3D MGM framework with three key components: (1) the Selective Re-mask Decoding that selectively re-masks 3D-relevant information while preserving 2D graph structures; (2) a structure-independent decoder that eliminates all structural information by relying solely on encoder representation; and (3) 3D-ReTrans as the 3D graph encoder for MRL. Our detailed analysis reveals the internal mechanisms of SRD and the structure-independent decoder. Extensive experiments demonstrate that 3D-GSRD significantly outperforms baselines on downstream datasets such as QM9 and MD17. + +Despite promising results, several limitations remain. Our pretraining is conducted on PCQM4Mv2 [55] with 3.37M molecules, which is smaller than large-scale datasets such as PubChemQC [66] with 230M molecules, potentially constraining performance. Scaling to larger and more diverse datasets is an important direction. In addition, we focus on molecular property prediction, while other tasks like 3D molecule generation [67-70] and multi-modal molecule-text modeling [71-74] could also benefit from our pretrained autoencoder. Beyond molecular applications, our pretraining paradigm can be extended to broader biological modalities such as single-cell [75] and protein [76]. + +# Acknowledgement + +This research is supported by the National Natural Science Foundation of China (62572449, 624B1012) and National University of Singapore SoC (grant no: A-0010308-00-00). + +# References + +[1] Xiaomin Fang, Lihang Liu, Jieqiong Lei, Donglong He, Shanzhuo Zhang, Jingbo Zhou, Fan Wang, Hua Wu, and Haifeng Wang. Geometry-enhanced molecular representation learning for property prediction. Nature Machine Intelligence, 4(2):127-134, 2022. +[2] Jiying Zhang, Zijing Liu, Yu Wang, and Yu Li. Subgdiff: A subgraph diffusion model to improve molecular representation learning. arXiv preprint arXiv:2405.05665, 2024. +[3] Zewei Ji, Runhan Shi, Jiarui Lu, Fang Li, and Yang Yang. Relmole: Molecular representation learning based on two-level graph similarities. Journal of Chemical Information and Modeling, 62(22):5361-5372, 2022. +[4] Mingyang Wang, Zhe Wang, Huiyong Sun, Jike Wang, Chao Shen, Gaoqi Weng, Xin Chai, Honglin Li, Dongsheng Cao, and Tingjun Hou. Deep learning approaches for de novo drug design: An overview. Current opinion in structural biology, 72:135-144, 2022. +[5] Qifeng Bai, Shuo Liu, Yanan Tian, Tingyang Xu, Antonio Jesús Banegas-Luna, Horacio Pérez-Sánchez, Junzhou Huang, Huanxiang Liu, and Xiaojun Yao. Application advances of deep learning methods for de novo drug design and molecular dynamics simulation. Wiley Interdisciplinary Reviews: Computational Molecular Science, 12(3):e1581, 2022. +[6] Shuangli Li, Jingbo Zhou, Tong Xu, Dejing Dou, and Hui Xiong. Geomgl: Geometric graph contrastive learning for molecular property prediction. In Proceedings of the AAAI conference on artificial intelligence, volume 36, pages 4541-4549, 2022. +[7] Hannes Stärk, Dominique Beaini, Gabriele Corso, Prudencio Tossou, Christian Dallago, Stephan Gunnemann, and Pietro Liò. 3d infomax improves gnns for molecular property prediction. In International Conference on Machine Learning, pages 20479-20502. PMLR, 2022. +[8] Shikun Feng, Lixin Yang, Yanwen Huang, Yuyan Ni, Weiying Ma, and Yanyan Lan. Unimap: universal smiles-graph representation learning. arXiv preprint arXiv:2310.14216, 2023. +[9] Mario Krenn, Florian Häse, Akshit Kumar Nigam, Pascal Friederich, and Alan Aspuru-Guzik. Self-referencing embedded strings (selfies): A $100\%$ robust molecular string representation. Machine Learning: Science and Technology, 1(4):045024, 2020. +[10] Jun Xia, Chengshuai Zhao, Bozhen Hu, Zhangyang Gao, Cheng Tan, Yue Liu, Siyuan Li, and Stan Z Li. Mole-bert: Rethinking pre-training graph neural networks for molecules. 2023. +[11] Mengying Sun, Jing Xing, Huijun Wang, Bin Chen, and Jiayu Zhou. Mocl: data-driven molecular fingerprint via knowledge-aware contrastive learning from molecular graph. In Proceedings of the 27th ACM SIGKDD conference on knowledge discovery & data mining, pages 3585-3594, 2021. +[12] Yuyang Wang, Rishikesh Magar, Chen Liang, and Amir Barati Farimani. Improving molecular contrastive learning via faulty negative mitigation and decomposed fragment contrast. Journal of Chemical Information and Modeling, 62(11):2713-2725, 2022. +[13] Raghunathan Ramakrishnan, Pavlo O Dral, Matthias Rupp, and O Anatole Von Lilienfeld. Quantum chemistry structures and properties of 134 kilo molecules. Scientific data, 1(1):1-7, 2014. +[14] Gengmo Zhou, Zhifeng Gao, Qiankun Ding, Hang Zheng, Hongteng Xu, Zhewei Wei, Linfeng Zhang, and Guolin Ke. Uni-mol: A universal 3d molecular representation learning framework. 2023. +[15] Zhenyu Hou, Xiao Liu, Yukuo Cen, Yuxiao Dong, Hongxia Yang, Chunjie Wang, and Jie Tang. Graphmae: Self-supervised masked graph autoencoders. In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 594-604, 2022. +[16] Xiaohong Ji, Zhen Wang, Zhifeng Gao, Hang Zheng, Linfeng Zhang, Guolin Ke, et al. Uni-mol2: Exploring molecular pretraining model at scale. arXiv preprint arXiv:2406.14969, 2024. + +[17] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. +[18] Zhiyuan Liu, Yaorui Shi, An Zhang, Enzhi Zhang, Kenji Kawaguchi, Xiang Wang, and Tat-Seng Chua. Rethinking tokenizer and decoder in masked graph modeling for molecules. Advances in Neural Information Processing Systems, 36, 2024. +[19] Junwei Yang, Kangjie Zheng, Siyu Long, Zaiqing Nie, Ming Zhang, Xinyu Dai, Wei-Ying Ma, and Hao Zhou. Mol-AE: Auto-encoder based molecular representation learning with 3d cloze test objective. In *Forty-first International Conference on Machine Learning*, 2024. URL https://openreview.net/forum?id=inEuvSg0y1. +[20] Zhenyu Hou, Yufei He, Yukuo Cen, Xiao Liu, Yuxiao Dong, Evgeny Kharlamov, and Jie Tang. Graphmae2: A decoding-enhanced masked self-supervised graph learner. In Proceedings of the ACM web conference 2023, pages 737-746, 2023. +[21] Keyulu Xu, Weihua Hu, Jure Leskovec, and Stefanie Jegelka. How powerful are graph neural networks? In ICLR, 2019. +[22] Ashish Vavwani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NIPS, pages 5998-6008, 2017. +[23] Kristof Schütt, Oliver Unke, and Michael Gastegger. Equivariant message passing for the prediction of tensorial properties and molecular spectra. In International Conference on Machine Learning, pages 9377-9388. PMLR, 2021. +[24] Philipp Thölke and Gianni De Fabritiis. Torchmd-net: equivariant transformers for neural network based molecular potentials. arXiv preprint arXiv:2202.02541, 2022. +[25] Cameron Diao and Ricky Loynd. Relational attention: Generalizing transformers for graph-structured tasks. arXiv preprint arXiv:2210.05062, 2022. +[26] Stefan Chmiela, Alexandre Tkatchenko, Huziel E Sauceda, Igor Politavsky, Kristof T Schütt, and Klaus-Robert Müller. Machine learning of accurate energy-conserving molecular force fields. Science advances, 3(5):e1603015, 2017. +[27] Yu Rong, Yatao Bian, Tingyang Xu, Weiyang Xie, Ying Wei, Wenbing Huang, and Junzhou Huang. Self-supervised graph transformer on large-scale molecular data. Advances in neural information processing systems, 33:12559-12571, 2020. +[28] Pengyong Li, Jun Wang, Yixuan Qiao, Hao Chen, Yihuan Yu, Xiaojun Yao, Peng Gao, Guotong Xie, and Sen Song. Learn molecular representations from large-scale unlabeled molecules for drug discovery. arXiv preprint arXiv:2012.11175, 2020. +[29] Chaohao Yuan, Kangfei Zhao, Ercan Engin Kuruoglu, Liang Wang, Tingyang Xu, Wenbing Huang, Deli Zhao, Hong Cheng, and Yu Rong. A survey of graph transformers: Architectures, theories and applications. arXiv, abs/2502.16533, 2025. +[30] Shikun Feng, Yuyan Ni, Yanyan Lan, Zhi-Ming Ma, and Wei-Ying Ma. Fractional denoising for 3d molecular pre-training. In International Conference on Machine Learning, pages 9938-9961. PMLR, 2023. +[31] Yuyan Ni, Shikun Feng, Wei-Ying Ma, Zhi-Ming Ma, and Yanyan Lan. Sliced denoising: A physics-informed molecular pre-training method. arXiv preprint arXiv:2311.02124, 2023. +[32] Sheheryar Zaidi, Michael Schaarschmidt, James Martens, Hyunjik Kim, Yee Whye Teh, Alvaro Sanchez-Gonzalez, Peter Battaglia, Razvan Pascanu, and Jonathan Godwin. Pre-training via denoising for molecular property prediction. arXiv preprint arXiv:2206.00133, 2022. +[33] Liang Wang, Shaozhen Liu, Yu Rong, Deli Zhao, Qiang Liu, Shu Wu, and Liang Wang. Molspectra: Pre-training 3d molecular representation with multi-modal energy spectra. In ICLR, 2025. + +[34] Rui Jiao, Xiangzhe Kong, Ziyang Yu, Wenbing Huang, and Yang Liu. Equivariant pretrained transformer for unified geometric learning on multi-domain 3d molecules. arXiv preprint arXiv:2402.12714, 2024. +[35] Xu Wang, Huan Zhao, Wei-wei Tu, and Quanming Yao. Automated 3d pre-training for molecular property prediction. In Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 2419–2430, 2023. +[36] Shengchao Liu, Hanchen Wang, Weiyang Liu, Joan Lasenby, Hongyu Guo, and Jian Tang. Pre-training molecular graph representation with 3d geometry. arXiv preprint arXiv:2110.07728, 2021. +[37] Ladislav Rampášek, Michael Galkin, Vijay Prakash Dwivedi, Anh Tuan Luu, Guy Wolf, and Dominique Beaini. Recipe for a general, powerful, scalable graph transformer. Advances in Neural Information Processing Systems, 35:14501-14515, 2022. +[38] Devin Kreuzer, Dominique Beaini, Will Hamilton, Vincent Létourneau, and Prudencio Tossou. Rethinking graph transformers with spectral attention. Advances in Neural Information Processing Systems, 34:21618-21629, 2021. +[39] Vijay Prakash Dwivedi, Chaitanya K Joshi, Anh Tuan Luu, Thomas Laurent, Yoshua Bengio, and Xavier Bresson. Benchmarking graph neural networks. Journal of Machine Learning Research, 24(43):1-48, 2023. +[40] Pan Li, Yanbang Wang, Hongwei Wang, and Jure Leskovec. Distance encoding: Design provably more powerful neural networks for graph representation learning. Advances in Neural Information Processing Systems, 33:4465-4478, 2020. +[41] Grégoire Mialon, Dexiong Chen, Margot Selosse, and Julien Mairal. Graphit: Encoding graph structure in transformers. arXiv preprint arXiv:2106.05667, 2021. +[42] Dominique Beaini, Saro Passaro, Vincent Létourneau, Will Hamilton, Gabriele Corso, and Pietro Lio. Directional graph networks. In International Conference on Machine Learning, pages 748-758. PMLR, 2021. +[43] Chengxuan Ying, Tianle Cai, Shengjie Luo, Shuxin Zheng, Guolin Ke, Di He, Yanming Shen, and Tie-Yan Liu. Do transformers really perform badly for graph representation? Advances in neural information processing systems, 34:28877-28888, 2021. +[44] Cristian Bodnar, Fabrizio Frasca, Nina Otter, Yuguang Wang, Pietro Lio, Guido F Montufar, and Michael Bronstein. Weisfeiler and lehman go cellular: Cw networks. Advances in neural information processing systems, 34:2625-2640, 2021. +[45] Yuning You, Tianlong Chen, Zhangyang Wang, and Yang Shen. When does self-supervision help graph convolutional networks? In international conference on machine learning, pages 10871-10880. PMLR, 2020. +[46] Weihua Hu, Bowen Liu, Joseph Gomes, Marinka Zitnik, Percy Liang, Vijay Pande, and Jure Leskovec. Strategies for pre-training graph neural networks. arXiv preprint arXiv:1905.12265, 2019. +[47] Jintang Li, Ruofan Wu, Wangbin Sun, Liang Chen, Sheng Tian, Liang Zhu, Changhua Meng, Zibin Zheng, and Weiqiang Wang. Maskgae: Masked graph modeling meets graph autoencoders. arXiv preprint arXiv:2205.10053, 9:13, 2022. +[48] Chun Wang, Shirui Pan, Guodong Long, Xingquan Zhu, and Jing Jiang. Mgae: Marginalized graph autoencoder for graph clustering. In Proceedings of the 2017 ACM on Conference on Information and Knowledge Management, pages 889-898, 2017. +[49] Sixiao Zhang, Hongxu Chen, Haoran Yang, Xiangguo Sun, Philip S Yu, and Guandong Xu. Graph masked autoencoders with transformers. arXiv preprint arXiv:2202.08391, 2022. +[50] Jinhua Zhu, Yingce Xia, Lijun Wu, Shufang Xie, Wengang Zhou, Tao Qin, Houqiang Li, and Tie-Yan Liu. Dual-view molecular pre-training. In Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 3615-3627, 2023. + +[51] Yin Fang, Qiang Zhang, Haihong Yang, Xiang Zhuang, Shumin Deng, Wen Zhang, Ming Qin, Zhuo Chen, Xiaohui Fan, and Huajun Chen. Molecular contrastive learning with chemical element knowledge graph. In Proceedings of the AAAI conference on artificial intelligence, volume 36, pages 3968-3976, 2022. +[52] Shengjie Luo, Tianlang Chen, Yixian Xu, Shuxin Zheng, Tie-Yan Liu, Liwei Wang, and Di He. One transformer can understand both 2d & 3d molecular data. In The Eleventh International Conference on Learning Representations, 2022. +[53] Josh Abramson, Jonas Adler, Jack Dunger, Richard Evans, Tim Green, Alexander Pritzel, Olaf Ronneberger, Lindsay Willmore, Andrew J Ballard, Joshua Bambrick, et al. Accurate structure prediction of biomolecular interactions with alphafold 3. Nature, pages 1-3, 2024. +[54] Bowen Jing, Stephan Eismann, Patricia Suriana, Raphael JL Townshend, and Ron Dror. Learning from protein structure with geometric vector perceptrons. arXiv preprint arXiv:2009.01411, 2020. +[55] Weihua Hu, Matthias Fey, Marinka Zitnik, Yuxiao Dong, Hongyu Ren, Bowen Liu, Michele Catasta, and Jure Leskovec. Open graph benchmark: Datasets for machine learning on graphs. Advances in neural information processing systems, 33:22118-22133, 2020. +[56] Shengchao Liu, Hongyu Guo, and Jian Tang. Molecular geometry pretraining with se (3)-invariant denoising distance matching. arXiv preprint arXiv:2206.13602, 2022. +[57] Rui Jiao, Jiaqi Han, Wenbing Huang, Yu Rong, and Yang Liu. Energy-motivated equivariant pretraining for 3d molecular graphs. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 37, pages 8096-8104, 2023. +[58] Shikun Feng, Yuyan Ni, Yan Lu, Zhi-Ming Ma, Wei-Ying Ma, and Yanyan Lan. Unigem: A unified approach to generation and property prediction for molecules. arXiv preprint arXiv:2410.10516, 2024. +[59] Kristof T Schütt, Huziel E Sauceda, P-J Kindermans, Alexandre Tkatchenko, and K-R Müller. Schnet-a deep learning architecture for molecules and materials. The Journal of Chemical Physics, 148(24), 2018. +[60] Victor Garcia Satorras, Emiel Hoogeboom, and Max Welling. E (n) equivariant graph neural networks. In International conference on machine learning, pages 9323-9332. PMLR, 2021. +[61] Johannes Gasteiger, Janek Groß, and Stephan Gunnemann. Directional message passing for molecular graphs. arXiv preprint arXiv:2003.03123, 2020. +[62] Johannes Gasteiger, Shankari Giri, Johannes T Margraf, and Stephan Gunnemann. Fast and uncertainty-aware directional message passing for non-equilibrium molecules. arXiv preprint arXiv:2011.14115, 2020. +[63] Benjamin Coors, Alexandru Paul Condurache, and Andreas Geiger. Spherenet: Learning spherical representations for detection and classification in omnidirectional images. In Proceedings of the European conference on computer vision (ECCV), pages 518-533, 2018. +[64] Limei Wang, Yi Liu, Yuchao Lin, Haoran Liu, and Shuiwang Ji. Comenet: Towards complete and efficient message passing for 3d molecular graphs. Advances in Neural Information Processing Systems, 35:650-664, 2022. +[65] Vijay Prakash Dwivedi, Anh Tuan Luu, Thomas Laurent, Yoshua Bengio, and Xavier Bresson. Graph neural networks with learnable structural and positional representations. arXiv preprint arXiv:2110.07875, 2021. +[66] Maho Nakata and Tomomi Shimazaki. Pubchemqc project: a large-scale first-principles electronic structure database for data-driven chemistry. Journal of chemical information and modeling, 57(6):1300-1308, 2017. + +[67] Zhiyuan Liu, Yanchen Luo, Han Huang, Enzhi Zhang, Sihang Li, Junfeng Fang, Yaorui Shi, Xiang Wang, Kenji Kawaguchi, and Tat-Seng Chua. NEXT-MOL: 3d diffusion meets 1d language modeling for 3d molecule generation. In The Thirteenth International Conference on Learning Representations, 2025. URL https://openreview.net/forum?id=p66a00KLWN. +[68] Yanchen Luo, Zhiyuan Liu, Yi Zhao, Sihang Li, Hengxing Cai, Kenji Kawaguchi, Tat-Seng Chua, Yang Zhang, and Xiang Wang. Towards unified and lossless latent space for 3d molecular latent diffusion modeling. arXiv preprint arXiv:2503.15567, 2025. +[69] Liang Wang, Chao Song, Zhiyuan Liu, Yu Rong, Qiang Liu, Shu Wu, and Liang Wang. Diffusion models for molecules: A survey of methods and tasks. arXiv, abs/2502.09511, 2025. +[70] Guikun Xu, Yankai Yu, Yongquan Jiang, Yan Yang, and Yatao Bian. CoFM: Molecular conformation generation via flow matching in SE(3)-invariant latent space. In ICML 2025 Generative AI and Biology (GenBio) Workshop, 2025. URL https://openreview.net/forum?id=C0jrjy4F1D. +[71] Zhiyuan Liu, Sihang Li, Yanchen Luo, Hao Fei, Yixin Cao, Kenji Kawaguchi, Xiang Wang, and Tat-Seng Chua. Molca: Molecular graph-language modeling with cross-modal projector and unimodal adapter. In EMNLP, 2023. URL https://openreview.net/forum?id=14WRhMNq7H. +[72] Sihang Li, Zhiyuan Liu, Yanchen Luo, Xiang Wang, Xiangnan He, Kenji Kawaguchi, Tat-Seng Chua, and Qi Tian. 3d-molm: Towards 3d molecule-text interpretation in language models. In ICLR, 2024. URL https://openreview.net/forum?id=xI4yN1kaqh. +[73] Zhiyuan Liu, Yaorui Shi, An Zhang, Sihang Li, Enzhi Zhang, Xiang Wang, Kenji Kawaguchi, and Tat-Seng Chua. Reactxt: Understanding molecular "reaction-ship" via reaction-contextualized molecule-text pretraining. In Findings of the Association for Computational Linguistics: ACL 2024. Association for Computational Linguistics, 2024. URL https://openreview.net/forum?id=V-ejDfLwe. +[74] Yongqiang Chen, Quanming Yao, Juzheng Zhang, James Cheng, and Yatao Bian. Hierarchical graph tokenization for molecule-language alignment. In *Forty-second International Conference on Machine Learning*, 2025. URL https://openreview.net/forum?id=wpbNczwAwV. +[75] Yaorui Shi, Jiaqi Yang, Changhao Nai, Sihang Li, Junfeng Fang, Xiang Wang, Zhiyuan Liu, and Yang Zhang. Language-enhanced representation learning for single-cell transcriptomics. arXiv preprint arXiv:2503.09427, 2025. +[76] Zhiyuan Liu, An Zhang, Hao Fei, Enzhi Zhang, Xiang Wang, Kenji Kawaguchi, and Tat-Seng Chua. Prott3: Protein-to-text generation for text-based protein understanding. In ACL. Association for Computational Linguistics, 2024. URL https://openreview.net/forum?id=ZmIj0Pil2b. + +# NeurIPS Paper Checklist + +# 1. Claims + +Question: Do the main claims made in the abstract and introduction accurately reflect the paper's contributions and scope? + +Answer: [Yes] + +Justification: We have included the paper's contributions and scope in the abstract and introduction. + +Guidelines: + +- The answer NA means that the abstract and introduction do not include the claims made in the paper. +- The abstract and/or introduction should clearly state the claims made, including the contributions made in the paper and important assumptions and limitations. A No or NA answer to this question will not be perceived well by the reviewers. +- The claims made should match theoretical and experimental results, and reflect how much the results can be expected to generalize to other settings. +- It is fine to include aspirational goals as motivation as long as it is clear that these goals are not attained by the paper. + +# 2. Limitations + +Question: Does the paper discuss the limitations of the work performed by the authors? + +Answer: [Yes] + +Justification: The limitations are included in Section 7. + +Guidelines: + +- The answer NA means that the paper has no limitation while the answer No means that the paper has limitations, but those are not discussed in the paper. +- The authors are encouraged to create a separate "Limitations" section in their paper. +- The paper should point out any strong assumptions and how robust the results are to violations of these assumptions (e.g., independence assumptions, noiseless settings, model well-specification, asymptotic approximations only holding locally). The authors should reflect on how these assumptions might be violated in practice and what the implications would be. +- The authors should reflect on the scope of the claims made, e.g., if the approach was only tested on a few datasets or with a few runs. In general, empirical results often depend on implicit assumptions, which should be articulated. +- The authors should reflect on the factors that influence the performance of the approach. For example, a facial recognition algorithm may perform poorly when image resolution is low or images are taken in low lighting. Or a speech-to-text system might not be used reliably to provide closed captions for online lectures because it fails to handle technical jargon. +- The authors should discuss the computational efficiency of the proposed algorithms and how they scale with dataset size. +- If applicable, the authors should discuss possible limitations of their approach to address problems of privacy and fairness. +- While the authors might fear that complete honesty about limitations might be used by reviewers as grounds for rejection, a worse outcome might be that reviewers discover limitations that aren't acknowledged in the paper. The authors should use their best judgment and recognize that individual actions in favor of transparency play an important role in developing norms that preserve the integrity of the community. Reviewers will be specifically instructed to not penalize honesty concerning limitations. + +# 3. Theory assumptions and proofs + +Question: For each theoretical result, does the paper provide the full set of assumptions and a complete (and correct) proof? + +Answer: [NA] + +Justification: The paper does not include theoretical results. + +# Guidelines: + +- The answer NA means that the paper does not include theoretical results. +- All the theorems, formulas, and proofs in the paper should be numbered and cross-referenced. +- All assumptions should be clearly stated or referenced in the statement of any theorems. +- The proofs can either appear in the main paper or the supplemental material, but if they appear in the supplemental material, the authors are encouraged to provide a short proof sketch to provide intuition. +- Inversely, any informal proof provided in the core of the paper should be complemented by formal proofs provided in appendix or supplemental material. +- Theorems and Lemmas that the proof relies upon should be properly referenced. + +# 4. Experimental result reproducibility + +Question: Does the paper fully disclose all the information needed to reproduce the main experimental results of the paper to the extent that it affects the main claims and/or conclusions of the paper (regardless of whether the code and data are provided or not)? + +# Answer: [Yes] + +Justification: The information needed to reproduce the main experimental results is provided in Section 6.1 and Appendix D. + +# Guidelines: + +- The answer NA means that the paper does not include experiments. +- If the paper includes experiments, a No answer to this question will not be perceived well by the reviewers: Making the paper reproducible is important, regardless of whether the code and data are provided or not. +- If the contribution is a dataset and/or model, the authors should describe the steps taken to make their results reproducible or verifiable. +- Depending on the contribution, reproducibility can be accomplished in various ways. For example, if the contribution is a novel architecture, describing the architecture fully might suffice, or if the contribution is a specific model and empirical evaluation, it may be necessary to either make it possible for others to replicate the model with the same dataset, or provide access to the model. In general, releasing code and data is often one good way to accomplish this, but reproducibility can also be provided via detailed instructions for how to replicate the results, access to a hosted model (e.g., in the case of a large language model), releasing of a model checkpoint, or other means that are appropriate to the research performed. +- While NeurIPS does not require releasing code, the conference does require all submissions to provide some reasonable avenue for reproducibility, which may depend on the nature of the contribution. For example +(a) If the contribution is primarily a new algorithm, the paper should make it clear how to reproduce that algorithm. +(b) If the contribution is primarily a new model architecture, the paper should describe the architecture clearly and fully. +(c) If the contribution is a new model (e.g., a large language model), then there should either be a way to access this model for reproducing the results or a way to reproduce the model (e.g., with an open-source dataset or instructions for how to construct the dataset). +(d) We recognize that reproducibility may be tricky in some cases, in which case authors are welcome to describe the particular way they provide for reproducibility. In the case of closed-source models, it may be that access to the model is limited in some way (e.g., to registered users), but it should be possible for other researchers to have some path to reproducing or verifying the results. + +# 5. Open access to data and code + +Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material? + +# Answer: [Yes] + +Justification: The code of the paper is released at https://github.com/WuChang0124/3D-GSRD and the data used are publicly available. + +# Guidelines: + +- The answer NA means that paper does not include experiments requiring code. +- Please see the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details. +- While we encourage the release of code and data, we understand that this might not be possible, so "No" is an acceptable answer. Papers cannot be rejected simply for not including code, unless this is central to the contribution (e.g., for a new open-source benchmark). +- The instructions should contain the exact command and environment needed to run to reproduce the results. See the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details. +- The authors should provide instructions on data access and preparation, including how to access the raw data, preprocessed data, intermediate data, and generated data, etc. +- The authors should provide scripts to reproduce all experimental results for the new proposed method and baselines. If only a subset of experiments are reproducible, they should state which ones are omitted from the script and why. +- At submission time, to preserve anonymity, the authors should release anonymized versions (if applicable). +- Providing as much information as possible in supplemental material (appended to the paper) is recommended, but including URLs to data and code is permitted. + +# 6. Experimental setting/details + +Question: Does the paper specify all the training and test details (e.g., data splits, hyperparameters, how they were chosen, type of optimizer, etc.) necessary to understand the results? + +# Answer: [Yes] + +Justification: The experimental settings are provided in Section 6.1 and Appendix D. + +# Guidelines: + +- The answer NA means that the paper does not include experiments. +- The experimental setting should be presented in the core of the paper to a level of detail that is necessary to appreciate the results and make sense of them. +- The full details can be provided either with the code, in appendix, or as supplemental material. + +# 7. Experiment statistical significance + +Question: Does the paper report error bars suitably and correctly defined or other appropriate information about the statistical significance of the experiments? + +# Answer: [No] + +Justification: We follow the experimental design of previous works, which did not involve repeated runs and did not report error bars. And due to limited computational resources, we are unable to afford the overhead of running multiple trials for each experiment. + +# Guidelines: + +- The answer NA means that the paper does not include experiments. +- The authors should answer "Yes" if the results are accompanied by error bars, confidence intervals, or statistical significance tests, at least for the experiments that support the main claims of the paper. +- The factors of variability that the error bars are capturing should be clearly stated (for example, train/test split, initialization, random drawing of some parameter, or overall run with given experimental conditions). +- The method for calculating the error bars should be explained (closed form formula, call to a library function, bootstrap, etc.) + +- The assumptions made should be given (e.g., Normally distributed errors). +- It should be clear whether the error bar is the standard deviation or the standard error of the mean. +- It is OK to report 1-sigma error bars, but one should state it. The authors should preferably report a 2-sigma error bar than state that they have a $96\%$ CI, if the hypothesis of Normality of errors is not verified. +- For asymmetric distributions, the authors should be careful not to show in tables or figures symmetric error bars that would yield results that are out of range (e.g. negative error rates). +- If error bars are reported in tables or plots, The authors should explain in the text how they were calculated and reference the corresponding figures or tables in the text. + +# 8. Experiments compute resources + +Question: For each experiment, does the paper provide sufficient information on the computer resources (type of compute workers, memory, time of execution) needed to reproduce the experiments? + +Answer: [Yes] + +Justification: The information about compute resources are provided in Appendix D. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The paper should indicate the type of compute workers CPU or GPU, internal cluster, or cloud provider, including relevant memory and storage. +- The paper should provide the amount of compute required for each of the individual experimental runs as well as estimate the total compute. +- The paper should disclose whether the full research project required more compute than the experiments reported in the paper (e.g., preliminary or failed experiments that didn't make it into the paper). + +# 9. Code of ethics + +Question: Does the research conducted in the paper conform, in every respect, with the NeurIPS Code of Ethics https://neurips.cc/public/EthicsGuidelines? + +Answer: [Yes] + +Justification: The research in this paper conform with the NeurIPS Code of Ethics. + +Guidelines: + +- The answer NA means that the authors have not reviewed the NeurIPS Code of Ethics. +- If the authors answer No, they should explain the special circumstances that require a deviation from the Code of Ethics. +- The authors should make sure to preserve anonymity (e.g., if there is a special consideration due to laws or regulations in their jurisdiction). + +# 10. Broader impacts + +Question: Does the paper discuss both potential positive societal impacts and negative societal impacts of the work performed? + +Answer: [Yes] + +Justification: We discuss the societal impacts of the work performed in Appendix A. + +Guidelines: + +- The answer NA means that there is no societal impact of the work performed. +- If the authors answer NA or No, they should explain why their work has no societal impact or why the paper does not address societal impact. +- Examples of negative societal impacts include potential malicious or unintended uses (e.g., disinformation, generating fake profiles, surveillance), fairness considerations (e.g., deployment of technologies that could make decisions that unfairly impact specific groups), privacy considerations, and security considerations. + +- The conference expects that many papers will be foundational research and not tied to particular applications, let alone deployments. However, if there is a direct path to any negative applications, the authors should point it out. For example, it is legitimate to point out that an improvement in the quality of generative models could be used to generate deepfakes for disinformation. On the other hand, it is not needed to point out that a generic algorithm for optimizing neural networks could enable people to train models that generate Deepfakes faster. +- The authors should consider possible harms that could arise when the technology is being used as intended and functioning correctly, harms that could arise when the technology is being used as intended but gives incorrect results, and harms following from (intentional or unintentional) misuse of the technology. +- If there are negative societal impacts, the authors could also discuss possible mitigation strategies (e.g., gated release of models, providing defenses in addition to attacks, mechanisms for monitoring misuse, mechanisms to monitor how a system learns from feedback over time, improving the efficiency and accessibility of ML). + +# 11. Safeguards + +Question: Does the paper describe safeguards that have been put in place for responsible release of data or models that have a high risk for misuse (e.g., pretrained language models, image generators, or scraped datasets)? + +Answer: [No] + +Justification: The paper poses no such risks. + +Guidelines: + +- The answer NA means that the paper poses no such risks. +- Released models that have a high risk for misuse or dual-use should be released with necessary safeguards to allow for controlled use of the model, for example by requiring that users adhere to usage guidelines or restrictions to access the model or implementing safety filters. +- Datasets that have been scraped from the Internet could pose safety risks. The authors should describe how they avoided releasing unsafe images. +- We recognize that providing effective safeguards is challenging, and many papers do not require this, but we encourage authors to take this into account and make a best faith effort. + +# 12. Licenses for existing assets + +Question: Are the creators or original owners of assets (e.g., code, data, models), used in the paper, properly credited and are the license and terms of use explicitly mentioned and properly respected? + +Answer: [Yes] + +Justification: The PCQM4Mv2 dataset are used by https://ogb.stanford.edu/docs/1sc/pcqm4mv2/ under CC BY 4.0 License. The QM9 dataset are used via https://deepchemdata.s3-us-west-1.amazon.com/ under the MIT License. + +Guidelines: + +- The answer NA means that the paper does not use existing assets. +- The authors should cite the original paper that produced the code package or dataset. +- The authors should state which version of the asset is used and, if possible, include a URL. +- The name of the license (e.g., CC-BY 4.0) should be included for each asset. +- For scraped data from a particular source (e.g., website), the copyright and terms of service of that source should be provided. +- If assets are released, the license, copyright information, and terms of use in the package should be provided. For popular datasets, paperswithcode.com/datasets has curated licenses for some datasets. Their licensing guide can help determine the license of a dataset. + +- For existing datasets that are re-packaged, both the original license and the license of the derived asset (if it has changed) should be provided. +- If this information is not available online, the authors are encouraged to reach out to the asset's creators. + +# 13. New assets + +Question: Are new assets introduced in the paper well documented and is the documentation provided alongside the assets? + +Answer: [NA] + +Justification: The paper does not release new assets. + +Guidelines: + +- The answer NA means that the paper does not release new assets. +- Researchers should communicate the details of the dataset/code/model as part of their submissions via structured templates. This includes details about training, license, limitations, etc. +- The paper should discuss whether and how consent was obtained from people whose asset is used. +- At submission time, remember to anonymize your assets (if applicable). You can either create an anonymized URL or include an anonymized zip file. + +# 14. Crowdsourcing and research with human subjects + +Question: For crowdsourcing experiments and research with human subjects, does the paper include the full text of instructions given to participants and screenshots, if applicable, as well as details about compensation (if any)? + +Answer: [NA] + +Justification: The paper does not involve crowdsourcing nor research with human subjects. + +Guidelines: + +- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. +- Including this information in the supplemental material is fine, but if the main contribution of the paper involves human subjects, then as much detail as possible should be included in the main paper. +- According to the NeurIPS Code of Ethics, workers involved in data collection, curation, or other labor should be paid at least the minimum wage in the country of the data collector. + +# 15. Institutional review board (IRB) approvals or equivalent for research with human subjects + +Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals (or an equivalent approval/review based on the requirements of your country or institution) were obtained? + +Answer: [NA] + +Justification: The paper does not involve crowdsourcing nor research with human subjects. + +Guidelines: + +- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. +- Depending on the country in which research is conducted, IRB approval (or equivalent) may be required for any human subjects research. If you obtained IRB approval, you should clearly state this in the paper. +- We recognize that the procedures for this may vary significantly between institutions and locations, and we expect authors to adhere to the NeurIPS Code of Ethics and the guidelines for their institution. +- For initial submissions, do not include any information that would break anonymity (if applicable), such as the institution conducting the review. + +# 16. Declaration of LLM usage + +Question: Does the paper describe the usage of LLMs if it is an important, original, or non-standard component of the core methods in this research? Note that if the LLM is used only for writing, editing, or formatting purposes and does not impact the core methodology, scientific rigorousness, or originality of the research, declaration is not required. + +Answer: [NA] + +Justification: The core method development in this research does not involve LLMs as any important, original, or non-standard components. + +Guidelines: + +- The answer NA means that the core method development in this research does not involve LLMs as any important, original, or non-standard components. +- Please refer to our LLM policy (https://neurips.cc/Conferences/2025/LLM) for what should or should not be described. + +# A Broader Impacts + +This work advances molecular representation learning and has the potential to accelerate downstream applications such as molecular property prediction, drug discovery, molecular dynamics simulation, and material design, helping reduce the cost and time of wet-lab experiments. However, there is a risk of over-reliance on model predictions without sufficient interpretability or domain validation. Additionally, models trained on biased datasets may produce structural or chemical biases, limiting the model's generalization across molecular spaces. We encourage the community to test our models strictly before applying them in scientific scenarios. + +# B More Details on Methodology + +In this section, we describe the embedding layer preceding the attention mechanism in 3D-ReTrans. The input 3D molecule graph is defined as $\bar{G} = (\mathbf{x},\mathbf{a},\mathbf{e})$ , where $\mathbf{x}$ denotes atomic coordinates, $\mathbf{a}$ represents atom types, and $\mathbf{e}$ denotes pairwise edge features. Our goal is to obtain atomic and edge embeddings that encode both chemical and geometric context. + +The initial node embedding $e^{\mathrm{node}}$ jointly encodes atom coordinates and types: + +$$ +e ^ {\text {n o d e}} = \operatorname {E m b e d} ^ {\text {n o d e}} ([ \mathbf {x}, \mathbf {a} ]). \tag {16} +$$ + +To incorporate local geometric context, we compute the neighborhood embedding $e_i^{\mathrm{neigh}}$ for each atom $i$ based on the radial distances $d_{ij}$ to neighbors atom $j$ . The radial basis function is given by: + +$$ +e ^ {\mathrm {R B F}} \left(d _ {i j}\right) = \phi \left(d _ {i j}\right) \exp \left(- \beta \left(\exp \left(- d _ {i j}\right) - \mu\right) ^ {2}\right) \tag {17} +$$ + +$$ +\phi \left(d _ {i j}\right) = \left\{ \begin{array}{l l} \frac {1}{2} \left(\cos \left(\frac {\pi d _ {i j}}{d _ {\mathrm {c u t}}}\right) + 1\right), & \text {i f} d _ {i j} \leq d _ {\mathrm {c u t}} \\ 0, & \text {i f} d _ {i j} > d _ {\mathrm {c u t}} \end{array} \right. \tag {18} +$$ + +where $\phi(d_{ij})$ is a smooth cutoff function ensuring locality and $\beta$ and $\mu$ are fixed parameters. + +The neighborhood embedding $e_i^{\mathrm{near}}$ for atom $i$ is then defined as: + +$$ +e _ {i} ^ {\text {n e i g h}} = \sum_ {j = 1} ^ {N} \operatorname {E m b e d} ^ {\text {n e i g h}} ([ \mathbf {x}, \mathbf {a} ]) \odot \mathbf {W} ^ {r} e ^ {\mathrm {R B F}} \left(d _ {i j}\right), \tag {19} +$$ + +where $\odot$ denotes element-wise product and $\mathbf{W}^r$ is a learnable projector. We set the cutoff distance $d_{\mathrm{cut}} = 5\AA$ , ensuring that each atom only attends to neighbors within this spatial range. + +The final atomic embedding $e_i^{\mathrm{atomic}}$ combines node and neighborhood information: + +$$ +e _ {i} ^ {\text {a t o m i c}} = \mathbf {W} ^ {a} \left(\left[ e ^ {\text {n o d e}}, e _ {i} ^ {\text {n e i g h}} \right]\right), \tag {20} +$$ + +where $\mathbf{W}^a$ denotes learnable projector. + +We also obtain the edge embedding $e^{\mathrm{edge}}$ via: + +$$ +e ^ {\mathrm {e d g e}} = \operatorname {E m b e d} ^ {\mathrm {e d g e}} (\mathbf {e}). \tag {21} +$$ + +# C Pseudo Code + +We present the pseudocode for pretraining (see Algorithm 1) and finetuning (see Algorithm 2) algorithms in this section. + +# D Experimental Setup + +# D.1 Computational Resource + +All experiments are conducted on NVIDIA A6000-48G GPUs. Pretraining requires a total of 48 GPU hours. For downstream tasks, finetuning on QM9 and MD17 takes approximately 48 and 8 GPU hours per experiment, respectively. + +Algorithm 1 Pretraining of 3D-GSRD +Require: 3D graph encoder $\phi_{\theta}^{3D}$ , 2D graph encoder $\phi_{\theta}^{2D}$ , decoder $\phi_{\theta}^{De}$ , pretraining dataset $D$ , input 3D molecule graph $G = (\mathbf{x}, \mathbf{a}, \mathbf{e})$ , masked coordinates prediction head $PosHead_{\theta}$ , denoising prediction head $DenoiseHead_{\theta}$ , mask ratio $p$ , denoising loss weight $w$ . +1: while training is not finished do +2: $G_i = (\mathbf{x}, \mathbf{a}, \mathbf{e}) = \text{dataloger}(D)$ +3: randomly mask $p$ atoms and add Gaussian noise ( $\Delta x_i \sim 0.04 \cdot \mathcal{N}(0, \sigma^2 I_m)$ ) to the unmasked atomic coordinates +4: input molecule $\tilde{G}_i = (\tilde{\mathbf{a}}, \tilde{\mathbf{x}}, \tilde{\mathbf{e}})$ +5: $\mathbf{h}_{3D}$ , $\mathbf{vec} = \phi_{\theta}^{3D}(\tilde{G}_i)$ +6: $\mathbf{h}_{2D} = \phi_{\theta}^{2D}(\mathbf{a}, \mathbf{e})$ +7: SRD( $\mathbf{h}_{3D}$ , $\tilde{G}_i$ ) = re-mask( $\mathbf{h}_{3D}$ ) + stop-grad( $\mathbf{h}_{2D}$ ) +8: rep, vec = $\phi_{\theta}^{De}$ (SRD( $\mathbf{h}_{3D}$ , $\tilde{G}_i$ ), vec) +9: For masked atoms: $x_i^{pred} = PosHead_{\theta}(\text{rep}, \text{vec})$ +10: For unmasked atoms: $\Delta x_i^{pred} = DenoiseHead_{\theta}(\text{rep}, \text{vec})$ +11: Loss = $||x_i^{pred} - x_i||_2^2 + w \cdot ||\Delta x_i^{pred} - \Delta x_i||_2^2 - CosineSimilarity(\text{stop-grad}(\mathbf{h}_{3D}), \mathbf{h}_{2D})$ +12: Optimise(Loss) +13: end while + +Algorithm 2 Finetuning of 3D-GSRD +Require: 3D graph encoder $\phi_{\theta}^{3D}$ , finetuning dataset $D$ , input 3D molecule graph $G = (\mathbf{x}, \mathbf{a}, \mathbf{e})$ , label prediction head LabelHead $_{\theta}$ . +1: while training is not finished do +2: $G_i = (\mathbf{x}, \mathbf{a}, \mathbf{e}), y_i = \text{dataloder}(D)$ +3: $\mathbf{h}_{\mathbf{3D}}, \mathbf{vec} = \phi_{\theta}^{3D}(G_i)$ +4: $y_i^{pred} = \text{LabelHead}_\theta(\mathbf{h}_{\mathbf{3D}}, \mathbf{vec})$ +5: Loss = || $y_i^{pred} - y_i$ ||_2^2 +6: Optimise(Loss) +7: end while + +# D.2Baselines + +We describe the details of our reported baseline methods in this section. + +SchNet [59] proposes continuous-filter convolutional layers, which enables the model to capture local correlations in molecules without grid-based data. + +$\mathbf{E}(\mathbf{n})$ -GNN [60] introduces an architecture that is equivariant to rotations, translations, reflections, and permutations. Notably, its equivariance extends to higher dimensions with affordable computational overhead increase. + +DimeNet [61] applies directional message passing which enables graph neural networks to incorporate directional information for molecular predictions and using spherical Bessel functions and spherical harmonics to representation distances and angles. + +DimeNet++ [62] improves upon DimeNet by being $8 \times$ faster and achieving $10\%$ higher accuracy, while maintaining strong generalization across molecular configurations and compositions. + +PaiNN [23] addresses the limitations of invariant representations in message passing neural networks by extending the message passing framework to rotationally equivariant representations. + +SphereNet [63] analyzes 3D molecular graphs in the spherical coordinate system and propose the spherical message passing (SMP) scheme to efficiently distinguish molecular structures while reducing training complexity. + +TorchMD-NET [24] introduces an equivariant Transformer architecture with a modified attention mechanism that incorporates interatomic distances directly into the attention weights. + +Transformer-M [52] is a Transformer based architecture that that handles multiple molecular data modalities within a unified model by using two separate channels to encode 2D and 3D molecular structures. + +SE(3)-DDM [56] leverages an SE(3)-invariant score matching method to transform coordinate denoising into denoising the pairwise atomic distances within a molecule. + +3D-EMGP [57] introduces an equivariant energy-based model and develops a self-supervised pretraining framework including a physics-inspired node-level force prediction task and a graph-level noise scale prediction task. + +Coord [32] proposes a pretraining technique based on denoising for 3D molecular structures, showing it is equivalent to learning a force field. + +Frad [30] introduces a new hybrid noise strategy by first adding Gaussian noise to the dihedral angles of the rotatable bonds, followed by traditional noise to the atom coordinates, with pretraining focused solely on denoising the latter. + +SliDe [31] develops a novel sliced denoising method that adds Gaussian noise to bond lengths, angles, and torsion angles with their variances determined by parameters within the energy function. + +Uni-Mol2 [16] is a molecular pretraining model that uses a two-track transformer to jointly capture atomic-level, graph-level, and geometry-level features, while systematically investigating scaling laws in molecular pretraining. + +ComENet [64] introduces a graph neural network for 3D molecular graphs that adopts rotation angles and local completeness in the 1-hop neighborhood, while integrating quantum-inspired basis functions into its message passing mechanism. + +Mol-AE [19] addresses the gap between pretraining and downstream objectives in encoder-only 3D molecular models by introducing an auto-encoder with positional encodings as atomic identifiers and a 3D Cloze Test objective that drops atoms to better capture real substructures. + +Uni-Geom [58] unifies molecular generation and property prediction through a diffusion-based two-phase process of scaffold nucleation and molecule growth, using a multi-branch network with oversampling to balance tasks. + +# D.3 Implementation details + +We employ the 3D-ReTrans as the 3D graph encoder and implement 2D-PE as 2D-ReTrans, a simplified version of the 3D-ReTrans that excludes 3D coordinates and distance inputs, while using the Transformer as the structure-independent decoder. The 3D graph encoder is configured with a hidden dimension of 256, 8 attention heads, and 12 layers. The 2D-PE shares most of its configuration with the 3D graph encoder, except for a hidden dimension of 64 and 4 attention heads. The decoder consists of 2 layers, with the hidden dimension and number of attention heads same as the 3D graph encoder. The detailed hyper-parameters configuration for pretraining and finetuning are shown in Table 7, Table 8, and Table 9, respectively. + +# E More Experimental Results + +# E.1 Ablation on Alternative Approach to Eliminating 2D Leakage + +We investigate an alternative approach to fully eliminate 2D leakage by directly combining 3D embeddings with 2D positional encodings as molecular representations during downstream tasks. We test this setup with SRD both with and without distillation. As shown in Table 10, simply fusing 2D-PE with 3D-ReTrans under SRD without distillation consistently underperforms our original design, highlighting the necessity of SRD with distillation for effective 2D and 3D alignment. In contrast, when distillation is applied, downstream performance becomes insensitive to whether 2D-PE is explicitly included, suggesting that pretraining distillation sufficiently aligns the modalities and renders additional 2D input unnecessary. + +Table 7: Hyper-parameters for pretraining on PCQM4MV2. + +
ParameterValue
DatasetPCQM4MV2
Train/Val/Test SplitOthers/100/100
Batch size128
Inference Batch size128
Accumulate grad batches2
OptimizerAdamW
Weight decay1e-16
SchedulerCosineAnnealingLR
Init learning rate5e-5
Min learning rate1e-6
Warm up steps10000
Max epochs30
Masked ratio0.25
Masked coordinates reconstruction loss typeMSE loss
Coordinate noise scale(type: Gaussian)0.04
Denoising loss weight0.1
+ +Table 8: Hyper-parameters for finetuning on QM9. + +
ParameterValue
DatasetQM9
Train/Val/Test Split11000/1000/10831
Batch size128
Inference Batch size128
Accumulate grad batches1
OptimizerAdamW
Weight decay1e-16
SchedulerCosineAnnealingLR
Init learning rate5e-4
Min learning rate1e-6
Warm up steps1000
Learning rate cosine length2,000,000
Max steps2,000,000
Max epochs2000
Finetuning loss typeMSE loss
+ +# E.2 Evolution of 2D-PE's representations + +To examine how 2D-PE's representation evolves during pretraining, we track the cosine similarity between 2D and 3D representations. As shown in Figure 8, the similarity increases sharply in the initial training steps, approaching 1.0, and then grows gradually, indicating progressive alignment between the two modalities. + +Table 9: Hyper-parameters for finetuning on MD17. + +
ParameterValue
DatasetMD17
Train/Val Split9500/500/remaining data
Batch size80
Inference batch size64
Accumulate grad batches1
OptimizerAdamW
Weight decay0.0
SchedulerCosineAnnealingLR
Init learning rate5e-4
Min learning rate1e-6
Warm up steps1000
Max epochs1200
Force weight0.8
Energy weight0.2
Finetuning loss typeMAE loss
Ema alpha dy1.0
Ema alpha y0.05
+ +Table 10: Ablation on alternative approach to eliminate 2D leakage. Performance (MAE $\downarrow$ ) on MD17. + +
Downstream ModelSRDLdistillSalicylicTolueneUracil
2D-PE+3D-ReTrans✓✗0.04200.02930.0334
3D-ReTrans✓✗0.04160.02930.0329
2D-PE+3D-ReTrans✓✓0.03840.02750.0311
3D-ReTrans✓✓0.03870.02750.0315
+ +![](images/f678db1a82ec14119627841cadcf8bbf9aa29e59cf44cc1f1b824ad6e90a204e.jpg) +Figure 8: Evolution of the 2D-PE's representation. \ No newline at end of file diff --git a/NeurIPS/2025/3D-GSRD_ 3D Molecular Graph Auto-Encoder with Selective Re-mask Decoding/images.zip b/NeurIPS/2025/3D-GSRD_ 3D Molecular Graph Auto-Encoder with Selective Re-mask Decoding/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..b2e6838574c3f3becdd5274814eb807720f12248 --- /dev/null +++ b/NeurIPS/2025/3D-GSRD_ 3D Molecular Graph Auto-Encoder with Selective Re-mask Decoding/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:da2ee79d0dddea02d34c0908fd17001e127262bd8f7c1607613f9a6b159451eb +size 702028 diff --git a/NeurIPS/2025/3D-GSRD_ 3D Molecular Graph Auto-Encoder with Selective Re-mask Decoding/layout.json b/NeurIPS/2025/3D-GSRD_ 3D Molecular Graph Auto-Encoder with Selective Re-mask Decoding/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..637705da7129b515584e10742162424d3797cbe9 --- /dev/null +++ b/NeurIPS/2025/3D-GSRD_ 3D Molecular Graph Auto-Encoder with Selective Re-mask Decoding/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b69602f155ca1a64db00289b18421129548dd880a86b85929cfade7e5e02e9f6 +size 847899 diff --git a/NeurIPS/2025/3D-Prover_ Diversity Driven Theorem Proving With Determinantal Point Processes/c3c61440-6109-4e21-a29e-b79185447424_content_list.json b/NeurIPS/2025/3D-Prover_ Diversity Driven Theorem Proving With Determinantal Point Processes/c3c61440-6109-4e21-a29e-b79185447424_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..0578f14a3e52b400d1b6e55a147d80540b7bb909 --- /dev/null +++ b/NeurIPS/2025/3D-Prover_ Diversity Driven Theorem Proving With Determinantal Point Processes/c3c61440-6109-4e21-a29e-b79185447424_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c4c37328d64e4a1642c0a9e6289270264f1324e9792b78793eeb65362d429e2c +size 138612 diff --git a/NeurIPS/2025/3D-Prover_ Diversity Driven Theorem Proving With Determinantal Point Processes/c3c61440-6109-4e21-a29e-b79185447424_model.json b/NeurIPS/2025/3D-Prover_ Diversity Driven Theorem Proving With Determinantal Point Processes/c3c61440-6109-4e21-a29e-b79185447424_model.json new file mode 100644 index 0000000000000000000000000000000000000000..3b0ed321eaf5b5665708767bcee183024bb3aad5 --- /dev/null +++ b/NeurIPS/2025/3D-Prover_ Diversity Driven Theorem Proving With Determinantal Point Processes/c3c61440-6109-4e21-a29e-b79185447424_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9085950b0448a1c31695f232a1dd288182dcfbea904154483fb22b6b74493fbb +size 174878 diff --git a/NeurIPS/2025/3D-Prover_ Diversity Driven Theorem Proving With Determinantal Point Processes/c3c61440-6109-4e21-a29e-b79185447424_origin.pdf b/NeurIPS/2025/3D-Prover_ Diversity Driven Theorem Proving With Determinantal Point Processes/c3c61440-6109-4e21-a29e-b79185447424_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..8a0cced5a353c55ef569234fdf5ab62c3b3efa76 --- /dev/null +++ b/NeurIPS/2025/3D-Prover_ Diversity Driven Theorem Proving With Determinantal Point Processes/c3c61440-6109-4e21-a29e-b79185447424_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d8dce27d41cd429952a22f25303415de22112fa3c006fd7664e9012c2e9847a8 +size 996710 diff --git a/NeurIPS/2025/3D-Prover_ Diversity Driven Theorem Proving With Determinantal Point Processes/full.md b/NeurIPS/2025/3D-Prover_ Diversity Driven Theorem Proving With Determinantal Point Processes/full.md new file mode 100644 index 0000000000000000000000000000000000000000..831e3638257e9cb5603c84c3c9d2062e5c396649 --- /dev/null +++ b/NeurIPS/2025/3D-Prover_ Diversity Driven Theorem Proving With Determinantal Point Processes/full.md @@ -0,0 +1,564 @@ +# 3D-Prover: Diversity Driven Theorem Proving With Determinantal Point Processes + +Sean Lamont1,2, Christian Walder3, Amir Dezfouli4, Paul Montague2, Michael Norrish1 + +$^{1}$ Australian National University + +2Defence Science and Technology Group + +3Google DeepMind + +4BIMLOGIQ + +sean.lamont@anu.edu.au + +# Abstract + +A key challenge in automated formal reasoning is the intractable search space, which grows exponentially with the depth of the proof. This branching is caused by the large number of candidate proof tactics which can be applied to a given goal. Nonetheless, many of these tactics are semantically similar or lead to an execution error, wasting valuable resources in both cases. We address the problem of effectively pruning this search, using only synthetic data generated from previous proof attempts. We first demonstrate that it is possible to generate semantically aware tactic representations which capture the effect on the proving environment, likelihood of success, and execution time. We then propose a novel filtering mechanism which leverages these representations to select semantically diverse and high quality tactics, using Determinantal Point Processes. Our approach, 3D-Prover, is designed to be general, and to augment any underlying tactic generator. We demonstrate the effectiveness of 3D-Prover on the miniF2F and LeanDojo benchmarks by augmenting popular open source proving LLMs. We show that our approach leads to an increase in the overall proof rate, as well as a significant improvement in the tactic success rate, execution time and diversity. We make our code available at https://github.com/sean-lamont/3D-Prover. + +# 1 Introduction + +Interactive Theorem Proving (ITP) traditionally involves a human guiding an ITP system to verify a formal proposition. The applications range from secure software (Tan et al., 2019) to the verification of mathematical results (Hales et al., 2017). There has been significant interest in automating this process, with formalization efforts requiring a high level of human expertise (Klein et al., 2009). It is also considered a 'grand challenge' for AI, requiring a high level of reasoning and planning to be successful (Reddy, 1988). Even large general purpose models struggle with the complexity of the task, with for example GPT-4 only able to solve $13.5\%$ (Thakur et al., 2023) of the high school level miniF2F-test (Zheng et al., 2021) benchmark. This has motivated specialized models and search algorithms to address the unique challenges of the domain (see e.g. (Li et al., 2024) for a review). + +With most non-trivial proofs requiring long chains of correct reasoning, it is a challenge to generate them in one pass without mistakes. The addition of a search algorithm is common for addressing this (Xin et al., 2024; Wu et al., 2024). Under this paradigm, candidate tactics are generated and executed in the proving system, which (if successful) results in new subgoals to prove. This generates a tree of possible proof paths, where a search algorithm selects the most promising nodes to expand. The primary challenge faced by these approaches is the exponential growth in the number of proof paths, limiting the complexity of the problems that can be solved efficiently. Many of the generated + +![](images/1cf8a4f94e535cfa3a6fc0719311d36b3538ce15adc5c2cb12362dd49cbd649a.jpg) +Figure 1: An example node expansion for a failed ReProver attempt, which 3D-Prover was able to prove. Tactics on the left result in the same proof state, tactics on the right result in an error, and tactics in the centre result in a unique proof state. The high error rate and tactic similarity motivates our filtering approach, which prunes the search space to give a diverse set of subgoals. + +tactics are equivalent, modulo variable renaming and other semantics-preserving transformations. See Figure 1 for a sample search tree from ReProver (Yang et al., 2023), where several semantically similar paths are explored, wasting valuable resources. Simple lexical similarity scores fail to cover the semantics (meaning) of a tactic, as captured by the effect of the tactic on the environment. For example, an expression and its negation vary by only a single character, but have a large semantic difference. It is therefore desirable to filter tactics by their semantic rather than syntactic diversity. In addition, many tactics lead to an execution error from the prover. From our experiments with miniF2F, we find approximately $75 - 85\%$ of tactics result in an execution error (Section 2.2). As tactic execution can be expensive, this further restricts the space of proofs which can be explored efficiently. + +These challenges motivate our proposed approach, Diversity Driven Determinantal Point Process Prover (3D-Prover). 3D-Prover adds an extra 'dimension' to existing proving systems by including a filtering mechanism on top of the existing tactic generation and search components. 3D-Prover uses Determinantal Point Processes (Kulesza, 2012) to prune the search space by filtering tactics to diverse and high quality subsets. The rich synthetic data generated from proof attempts enables us to learn the effect tactics have on the environment, including the error likelihood and execution time. We leverage this to generate tactic representations which reflect their semantics, which 3D-Prover uses to filter tactics based on a combination of their diversity and quality. 3D-Prover allows for a direct tradeoff between search objectives, with hyperparameters controlling the weighting of error, time and diversity in the filtering process. 3D-Prover is a general approach which can be used to augment any underlying tactic generator. We demonstrate this by augmenting the popular ReProver and InternLM2.5-Step-Prover LLMs to obtain a significant improvement in the success rate, execution time and diversity of tactics, and the overall proof success rate. To summarise our contributions: + +- We study the feasibility of learning the environment dynamics of proving systems. We demonstrate tactic representations which capture the likely effect on the environment, using them to predict the likelihood of success and execution time of a tactic, as well as the resulting proof state or error message. +- We propose a novel edge filtering approach using Determinantal Point Processes (Kulesza & Taskar, 2011), which leverage these representations to select semantically diverse subsets of quality tactics. Our method is modular and can be used with any tactic generator. +- We evaluate our approach by augmenting ReProver (Yang et al., 2023) on the miniF2F (Zheng et al., 2021) benchmarks, where we demonstrate a significant improvement in the tactic success rate, diversity and overall proof rate. + +# 1.1 Related work + +There is little prior work on learning the effect of a tactic on the proving environment, with only Xin et al. (2024) using successful environment responses as an auxiliary objective. We investigate the task in detail, as well as learning the error likelihood, error messages and execution time, which + +we use to generate useful tactic representations. Several approaches use previous proof attempts to improve performance, using the sparse binary signal from the proof result (Li et al., 2024). This has been used to improve search (Lample et al., 2022; Wang et al., 2023), however these approaches do not consider node diversity, with nothing preventing the exploration of semantically similar paths. First & Brun (2022) examine a diverse ensemble of tactic models, whereas we focus on diversity with respect to the search, given an arbitrary underlying tactic model (or models). Recently, Yang et al. (2025) select subgoals based on their diversity, with a simple embedding model over the subgoal text. Our approach does not need to execute the tactics, as we learn embeddings reflecting the environment dynamics and use these to select tactics before execution, thereby saving resources. + +# 1.2 Background: Determinantal Point Processes + +Determinantal Point Processes (DPPs) are a class of probabilistic models for sampling subsets from a ground set $\mathcal{V}$ . They provide an inherent trade-off between the diversity and quality of the sampled subsets, successfully being applied to this end across a variety of domains (Kulesza, 2012; Hsiao & Grauman, 2018; Zhang et al., 2016). This motivates their use in our filtering approach (Section 3). + +In line with Kulesza (2012), for $|\mathcal{V}| = n$ we define the kernel $L \in \mathbb{R}^{n \times n}$ of a DPP as the Gram matrix $L = B^T B$ for $B \in \mathbb{R}^{n \times d}$ , where row $\pmb{b}_i \in \mathbb{R}^d$ of $B$ represents element $i \in \{1, \dots, n\}$ of $\mathcal{V}$ . The $\pmb{b}_i$ are commonly decomposed into a set of unit norm diversity features $\phi_i \in \mathbb{R}^d$ and quality scores $q_i \in \mathbb{R}^+$ , so that $\pmb{b}_i = q_i \phi_i$ , $||\phi_i|| = 1$ for all $i \in \{1, \dots, n\}$ . The similarity matrix $S$ is then defined as $S_{ij} = \phi_i^T \phi_j$ . The probability of sampling $A \subseteq \mathcal{V}$ is then proportional to the determinant of the submatrix of $L$ indexed by $A$ , $\mathbb{P}(A) \propto \det(L_A) = (\prod_{i \in A} q_i^2) \det(S_A)$ . Geometrically, this determinant is the volume of the parallelepiped spanned by the submatrix $L_A$ , which as we see in Figure 3, is maximised based on a combination of the similarity and length (quality) of the chosen elements. In this way, DPPs elegantly trade off between the quality and diversity of elements. Normally the size of the sampled subset $|A|$ is variable, however Kulesza & Taskar (2011) introduce $k$ -DPPs which restricts the size of the subset to a fixed $k \in \mathbb{N}$ , and where the probability of sampling $A$ is normalised over subsets of size $k$ . That is, for a $k$ -DPP, $\mathbb{P}(A) = \det(L_A) / \sum_{|A'| = k} \det(L_{A'})$ . + +# 2 Transition Aware Representation Learning + +One proof attempt can generate large amounts of data. We found a single pass of ReProver on the miniF2F-valid benchmark of 244 proofs results in approximately 500,000 transitions, capturing rich information about the error likelihood, execution time and resulting proof state or error message. We now explore the feasibility of using this data to learn how tactics affect the environment, operationalising this as a supervised learning task: given a goal and tactic, we predict the error status, execution time and environment output. We effectively learn these targets from only this synthetic data, and embed this information into a compact tactic representation. The upshot, as we show in Section 3, is that this can be used to improve the performance of subsequent proof attempts. + +# 2.1 Transition Models + +The result of a proof attempt (formalised in A) is the dataset $\mathcal{D}$ of transitions $\{(g,t,s,\tau ,o)\}$ , which captures the results of applying tactics $t\in \mathcal{T}$ to goals $g\in S$ . The status $s\in \{0,1\}$ , indicates a success (1) or failure (0), $\tau \in \mathbb{R}$ gives the execution time and the output $o\in \mathcal{O}$ is the environment response, which is an error message, new subgoals, or a proof success. We propose a method to learn tactic representations $\pmb {e}\in \mathbb{R}^d$ which capture the result $(s,\tau ,o)$ of applying $t$ to $g$ . By using these as features for DPP, we can filter tactics based on their expected outcome, before they are executed. + +We define our transition model $\xi : \mathcal{S} \times \mathcal{T} \to [0,1] \times \mathbb{R} \times \mathcal{O}$ as a mapping from a goal $g$ and tactic $t$ to an estimate of the status $s$ , time $\tau$ and output $o$ . To ensure $\xi$ admits effective representations in $e$ , we construct it as follows. The Encoder $E: \mathcal{S} \times \mathcal{T} \to \mathbb{R}^d$ takes the goal $g$ and tactic $t$ as input, and outputs our representation $E(g,t) = e$ . As $e$ will be used as the diversity feature for DPP, it is constrained to unit norm $||e|| = 1$ . The Predictor $P: \mathbb{R}^d \to [0,1] \times \mathbb{R}$ maps $e$ to an error probability for the status and a score for the time prediction, with $P(e) = (\hat{s},\hat{\tau})$ . The Decoder $D: \mathbb{R}^d \times \mathcal{S} \to \mathcal{O}$ maps $e$ and $g$ to the output prediction, such that $D(e,g) = \hat{o}$ . The transition model is then + +$$ +\xi (g, t) = \left(P (E (g, t)), D (E (g, t), g)\right) = (\hat {s}, \hat {\tau}, \hat {o}). \tag {1} +$$ + +![](images/90ddef23b163ba0d9e2193ca04937cc93476d40a01b2f9f528a72442b21ce651.jpg) +Figure 2: Our COMBINED architecture for learning transition aware tactic embeddings. The tactic t and goal g are concatenated and passed through the Encoder $E$ . A representation vector e is generated by mean-pooling over the tactic token embeddings $\mathbf{t}'$ . The Predictor $P$ takes this embedding and predicts whether the tactic results in an error (Status), and the execution time (Time). The Decoder $D$ takes the embedding and goal to predict the environment response (Output), which is either an error message or new goals to prove. This setup yields a compact representation of the tactic which captures its effect on the proving environment, enabling our proposed filtering model. + +We note that the Decoder and Predictor can only access information of $t$ through $e$ . Hence our architecture requires the Encoder to learn an effective representation for $e$ , so that the Decoder and Predictor can use this to determine the subsequent effect of the tactic on the environment. + +# 2.2 Experiments + +For our experiments, we use an Encoder-Decoder Transformer for the Decoder $D$ , and an Encoder-Only Transformer for the Encoder $E$ . We take the pretrained ReProver (Yang et al., 2023) LLM to initialise both components. We implement the Predictor $P$ as a single hidden layer MLP, with hidden dimension $d / 2$ (where $d = 1472$ ) and two real valued output nodes. The time prediction $\hat{\tau}$ is the output of the first node, and the status prediction $\hat{s}$ is taken as the sigmoid of the second. We use this simple Predictor architecture to speed up our filtering algorithm presented in Section 3. + +We investigate several variations of the transition model $\xi$ . For the COMBINED model (Figure 2), the tactic is concatenated with the goal, and the embeddings from the Encoder are computed for all tokens. We then generate a single tactic embedding by mean-pooling over the tactic tokens. We examine the COMBINED model both with the full goal text, and a variation COMBINED (SMALL GOAL) which embeds the goal first, and concatenates it as a single token vector to the tactic. This variation allows for more efficient batching when used for filtering in Section 1, as the goal need only be embedded once for multiple tactics, however gives less information to the model. The SEPARATE model encodes the tactic without attending to the goal. We hypothesise that allowing the tactic tokens to attend to the goal will allow the Encoder to better represent the semantics of the tactic. To form a naive baseline, we implement a No TACTIC model which does not use the tactic at all, and instead uses only the goal tokens. We do this to account for any inherent patterns in the goal which may be predictive of the outcome, for example a particular goal which has a high error rate. This allows us to ground our results in the performance of this baseline, so we can observe the direct effect of the tactic in predictive performance. We also compare with an ALL TOKENS model which uses all tactic tokens for the Decoder without reducing to a single embedding. We implement this comparison to see the degree of information loss induced by reducing tactics to a single vector. Given $\alpha_{s}, \alpha_{\tau}, \alpha_{o} \in \mathbb{R}^{+}$ , with estimates $\hat{s}, \hat{\tau}, \hat{o}$ and for minibatch $\mathcal{B} \subseteq \mathcal{D}$ , we optimise the following: + +$$ +\sum_ {(g, t, s, \tau , o) \in \mathcal {B}} \alpha_ {s} \mathcal {L} _ {s} (s, \hat {s}) + \alpha_ {\tau} \mathcal {L} _ {\tau} (\tau , \hat {\tau}) + \alpha_ {o} \mathcal {L} _ {o} (o, \hat {o}). \tag {2} +$$ + +Table 1: Results for predicting unseen environment responses given a goal and tactic, for transitions from miniF2F-valid. The No TACTIC result forms a baseline to assess the impact of the tactic representation. We observe that any tactic representation enables far better predictions, and constraining these to a single vector (COMBINED and SEPARATE) does not hurt the performance gain. This demonstrates tactic representations which capture their effect on the environment, enabling our filtering model in Section 3. Comparing the COMBINED and SEPARATE models, allowing the representation to attend to the goal leads to a large improvement. + +
EmbeddingOutputStatusTime
BLEUROUGE-L F1Top-4F1TPRTNRMSE
ALL TAXENS0.310.380.310.850.820.960.17
COMBINED0.330.390.320.880.850.970.16
COMBINED (SMALL GOAL)0.300.360.290.850.810.960.20
SEPARATE0.270.340.270.760.710.940.28
NO TACTIC0.170.220.130.220.140.960.37
+ +The hyperparameters $\alpha_{s},\alpha_{\tau},\alpha_{o}$ control the weighting of the status, time and output losses. For simplicity, we set these to 1, however they could be tuned to reflect the relative importance of each task. We use the binary cross-entropy loss $\mathcal{L}_s$ for the status prediction, the mean squared error (MSE) $\mathcal{L}_{\tau}$ for the time prediction, and the cross-entropy loss $\mathcal{L}_o$ for the output prediction. + +We obtain $\mathcal{D}$ from a single ReProver attempt on miniF2F-valid, yielding 498,236 transitions split randomly into $95\%$ training, $5\%$ testing. For the error prediction task, we reweight classes to account for imbalance, which is approximately $75\%$ error, $25\%$ success. We use the AdamW optimizer, with a learning rate of $10^{-5}$ and a batch size of 1, train for 2 epochs, and report the results on the test set. To assess the Output prediction, we generate 4 outputs with beam search for each transition. We use BLEU (Papineni et al., 2002) and ROUGE-L (Lin, 2004) to assess the quality of the highest scoring beam in comparison to the ground truth, which is either an error message or new subgoals. The Top-4 accuracy is the proportion of samples with one beam identical to the ground truth. For Status, we take the prediction as 1 if $\hat{s}_{ki} > 0.5$ and 0 otherwise, reporting the F1 score, true positive rate (TPR) and true negative rate (TNR). For time, we take the Mean Squared Error (MSE) of the prediction. + +Table 1 summarises the performance of our transition models. Our results suggest tactic representations which capture useful information about their effect on the environment ${}^{1}$ ,which we can see by the clear improvement across all approaches compared to the NO TACTIC baseline. The improvement of COMBINED over SEPARATE supports our hypothesis that we can better predict transitions when the tactic embedding attends to the goal. As expected, COMBINED outperforms COMBINED (SMALL GOAL), with COMBINED (SMALL GOAL) significantly outperforming the SEPARATE model. COMBINED (SMALL GOAL) therefore gives an effective compromise between the accurate but expensive COMBINED model, and the goal-unaware SEPARATE model. We note the ALL TOKENS model, with the Decoder attending to the full tactic, does not improve upon the full COMBINED model. This shows our architecture effectively represents the tactic as a single embedding without losing any relevant information. These results are the first to demonstrate the feasibility of learning the environment dynamics of proof systems. To illustrate the difficulty of this task, all predictions for the COMBINED model and their ground truth are provided with our code. + +# 3 3D-Prover + +Algorithm 1 defines 3D-Prover, which maps tactics $T$ from the underlying tactic policy $\pi_0$ to a subset $T'$ of size $K$ . We use the Encoder $E$ and Predictor $P$ from Section 2 to generate tactic embeddings $\phi_i$ and predict the time and error likelihood. As they are unit norm, the embeddings $\phi_i$ encode the predicted environment response through their direction. The quality score $q_i$ then scales $\phi_i$ based on the tactic model logits $m_i$ , as well as the predicted error likelihood $s_i$ and execution time $\tau_i$ . We have hyperparameters for normalisation temperature $\theta$ , as well as error and time weights + +![](images/47a74a3c688e49d7399cc7c5ea56fe4cdb3e808b4667679b6a11bdc0ef1b3dd2.jpg) +(a) Initial tactic embeddings $\phi_{i}$ , representing predicted outcome. + +![](images/fc4b8c5c9cde32dbf9817ce4734c1bde20c8967644adfb9d59b58453b41e1016.jpg) +(b) Quality scaled embeddings, $q_{i}\phi_{i}$ , to be filtered by $k$ -DPP +Figure 3: Visualisation of DPP for tactic filtering. The tactic embeddings from the transition model are scaled by quality scores, before a subset of tactics are selected using $k$ -DPP. Subsets are chosen proportionally to the area spanned by their elements, giving a combination of quality and diversity. For this simplified example, we take the 2D PCA projection of embeddings for tactics in Figure 1, setting the quality to the scaled generator logits. Comparing the shaded areas in (b) and assuming subst c and rw h1 have been selected, we see that symmetry is favoured over simp [h11] is scored higher by the generator, it is less diverse with respect to subst c and rw h1. + +Algorithm 1: 3D-Prover +```txt +Input: Goal $g$ candidate tactics $T = \{t_i\}_{i=1}^N$ , filter size $K$ , Encoder $E$ , Predictor $P$ , error weight $\lambda_s$ , time weight $\lambda_{\tau}$ , temperature $\theta$ , tactic policy $\pi_0$ +Output:Filtered tactics $T' \subset T$ +// Compute embeddings and scores for each tactic +for $i$ in $\{1,\dots,N\}$ do + $\begin{array}{l}\phi_i\gets E(g,t_i)\\ (s_i,\tau_i)\gets P(\phi_i)\\ \tau_i\gets 1 - \frac{\tau_i}{||\tau||},\tau = (\tau_1,..,\tau_N)\\ m_i\gets \frac{\exp(\pi_0(t_i|g) / \theta)}{\sum_{j=1}^N\exp(\pi_0(t_j|g) / \theta)}\\ q_i\gets m_i + \lambda_s s_i + \lambda_\tau \tau_i \end{array}$ +//Compute kernel matrix + $L\gets B^{T}B$ , where $B = [q_{1}\phi_{1},\ldots ,q_{N}\phi_{N}]$ +Compute eigenvalues $\lambda_{i}$ and eigenvectors $\mathbf{v}_i$ of $L$ +Sample $J\subset \{1,\dots,N\}$ using Algorithm 2 of Kulesza & Taskar (2011), with $\{(v_i,\lambda_i)\}$ $k = K$ +return $T^{\prime} = \{t_j\}_{j\in J}$ +``` + +$(\lambda_{s}, \lambda_{\tau})$ . $\theta$ controls the scaling temperature of the model logits, with a higher temperature flattening the distribution. It therefore adjusts the diversity bias of 3D-Prover by reducing the impact of the quality scores when sampling. We then compute the kernel $L$ from $q_{i}$ and $\phi_{i}$ , and sample a subset of tactics $T'$ using the $k$ -DPP algorithm (Kulesza & Taskar, 2011). Figure 3 visualises this process, where tactics subsets are sampled in proportion to their shaded area. + +# 3.1 Experiments + +We test the performance of 3D-Prover with two setups. We first use ReProver (Yang et al., 2023) as the underlying tactic policy $\pi_0$ , as it is a small ( $\sim$ 300M parameters) and popular open source proving model, allowing us to run extensive experiments and ablations in a reasonable time frame. To evaluate our approach over a large (7B) state-of-the-art model, we also present a smaller scale experiment using InternLM2.5-Step-Prover (Wu et al., 2024). We run our experiments in Lean (De Moura et al., 2015) using the BAIT (Lamont et al., 2024) platform with a modified LeanDojo (Yang et al., 2023) environment, where we set an environment timeout of 600 seconds per proof attempt. We train a COMBINED model for Reprover and a COMBINED (SMALL GOAL) model for InternLM + +Table 2: Pass@1 results on miniF2F, with $K$ tactics selected per node from ReProver. 3D-Prover uses a transition model trained from miniF2F-valid transitions. For miniF2F-test, we report the mean along with minumum and maximum over four runs, noting that Top- $K$ is deterministic given ReProver uses Beam Search. The Gain column reports the relative improvement over the Top- $K$ baseline. Results for No Filtering were ${27.8}\%$ for miniF2F-test and ${27.9}\%$ for miniF2F-valid. We observe a clear improvement using 3D-Prover, which increases as more filtering is applied (lower $K$ ). Our results on miniF2F-test show that 3D-Prover can improve search even for proofs out of distribution of the transition model. + +
KTop-KRandom3D-ProverGain
miniF2F-test (mean, minimum and maximum over four runs)
822.419.0 (18.4, 19.6)24.4 (23.7, 24.9)+8.9%
1626.525.4 (24.5, 25.7)27.3 (26.9, 27.8)+3.0%
3227.827.4 (26.9, 28.2)28.2 (27.3, 28.6)+1.4%
miniF2F-valid (single run)
821.719.325.0+15.2%
1626.624.229.1+9.4%
3227.927.528.7+2.9%
+ +from Section 2.2, using transitions from their respective base models. The Encoder and Predictor components then generate tactic embeddings and quality scores as per Algorithm 1. We first examine the performance of 3D-Prover without hyperparameter tuning, setting $\lambda_s = \lambda_\tau = 0$ , $\theta = 1$ . We then perform ablation studies (3.1.2) with ReProver over miniF2F-valid to examine the influence of the hyperparameters on the tactic success rate, execution time and diversity of the environment response. For miniF2F-test, we allow the model multiple attempts per proof to increase confidence in the results, while for miniF2F-valid we allow one attempt per configuration to allow a wider set of ablations. We also present an additional experiment over the larger LeanDojo benchmark in Appendix D. + +The default search policy is set to be Best First Search (BFS), with nodes expanded in order of their cumulative log probability. For each node selected, we generate $N = 64$ candidate tactics from the underlying model $^2$ . Following the original implementations, we use beam search for ReProver and sampling with $T = 0.7$ for InternLM. These form the ground set for the node, to be sub-sampled by the filtering algorithm. As beam search decoding is deterministic, the ground set for a given node is fixed across runs, allowing us to better isolate and compare approaches. We maintain sampling for InternLM to represent the original deployment scenario and test our approach under realistic usage. The filtering algorithm returns $K$ tactics, which are executed in the environment before updating the proof tree. For ReProver, we test three levels of filtering, with $K \in \{8,16,32\}$ . Lower $K$ corresponds to more filtering, for which the filtering algorithm will have a greater impact. We compare 3D-Prover, as outlined in Algorithm 1, with three baselines. The No Filtering baseline represents the original approach with no filtering. The Top-K baseline takes the top $K$ tactics from the ground set as judged by their log probabilities, corresponding to the top $K$ beams. We take $K$ tactics at random from the ground set to form the Random baseline, as an exploration-focused comparison. For InternLM, we test $K = 8$ with the Top- $K$ and No Filtering baselines, and perform an additional experiment with Critic Guided search from (Wu et al., 2024) in place of BFS. + +# 3.1.1 Proof Performance + +Table 2 and 3 shows the results of our ReProver and InternLM experiments on miniF2F. We observe 3D-Prover outperforming every baseline over both models. The influence of filtering becomes more apparent as $K$ is decreased as there are more tactics filtered out. Reflecting this, the magnitude of improvement given by 3D-Prover increases for lower $K$ . 3D-Prover is able to outperform both baselines by providing a tradeoff between the quality, as represented by Top- $K$ , and the diversity of the tactics. The choice of $K$ also controls the depth of the proof search, with larger $K$ giving broader search, and smaller $K$ deeper search. As most discovered proofs are short (favouring broad search), we see the Pass@1 performance for lower values of $K$ is generally lower, however over multiple attempts it can be beneficial to use deeper searches (see Appendix B). Our InternLM results + +Table 3: Results for miniF2F-test using tactics selected from InternLM2.5-Step-Prover (mean, maximum and minimum for Pass@1). We compare results using standard Best First Search model (BFS) and with the InternLM2.5 Critic Guided model for goal selection. 3D-Prover uses a transition model trained from transitions on miniF2F-test with the No Filtering model. The Pass@1 result for BFS for No Filtering was 44.8. We observe 3D-Prover outperforming both the No Filtering and Top- $K$ baselines, with a notable improvement over more attempts for the Critic Guided search model. + +
Top-K (K = 8)3D-Prover (K = 8)Gain
InternLM2.5-StepProver (BFS)
Pass@144.3 (44.1, 44.5)45.7 (45.3, 46.1)+3.2%
Pass@244.947.3+5.3%
No Filtering (K = 64)3D-Prover (K = 8)Gain
InternLM2.5-StepProver (Critic Guided)
Pass@143.7 (42.4, 44.5)45.7 (44.5, 47.0)+4.6%
Pass@649.053.1+8.4%
+ +(Table 3) demonstrate this, with the improvements from filtering growing over more attempts. This indicates a better variety of paths being explored, as each different attempt is more likely to take a more diverse approach. Finding deep proofs is a significant challenge Polu et al. (2022), with the search tree growing exponentially with proof depth. The large improvements from 3D-Prover for deeper search is a step towards addressing this. + +Tree search should be considered an augmentation of the base model, with improvements generally much smaller than those found by improving the generator itself. This is unsurprising, as the generator forms the base set of candidates for the search to explore. Improved search algorithms do, however, have the advantage of being applicable to different base models, which is important given the rapid advance of new and better generators. For example, DeepSeek-Prover-V1.5 obtains around $2 - 4\%$ relative improvements in proof success over miniF2F-test with its novel tree search algorithm, compared to no search. In comparison, improving their base model yields a $\sim 36\%$ relative improvement (Figure 5 and Table 1 in (Xin et al., 2024)). Similarly, Table 1 from Polu et al. (2022) shows their search approach yielding $0.04 - 5.7\%$ relative improvements for miniF2F-valid, with $\sim 40,000$ GPU hours required for their best results. We were able to find our improvements with significantly less resources, training our transition model on only a single attempt per proof. + +We emphasise that these results were obtained without any hyperparameter tuning, only using the representations as diversity features and model logits as quality scores. We present ablation studies looking closer at these hyperparameters, however a comprehensive sweep is prohibitively expensive (Appendix I). Despite this, we were able to obtain our improvements without tuning, demonstrating the effectiveness of our approach. For completeness, Appendix C details the Pass@1 performance over the hyperparameter configurations we tested for our ablations. We also highlight that the miniF2F-test ReProver results were obtained by training with transitions from miniF2F-valid, showing that 3D-Prover remains effective for proofs out of distribution. The results on miniF2F-valid represent the more common online scenario, with previous attempts on the same dataset being used to improve performance (see, for example, Lample et al. (2022); Polu et al. (2022)). We also note that our approach is lightweight with minimal overhead, as we detail in Appendix G. + +# 3.1.2 Ablation Study + +Effect of the Transition Model To demonstrate the utility of our tactic representations, we compare to an ablated 3D-Prover where the transition model Encoder is replaced by an Autoencoder of the same size. The Autoencoder is trained to reconstruct the original tactic, and therefore generates representations which reflect only the syntax of the tactic. This tests our hypothesis that semantically aware tactic representations are useful for proofs, justifying the inclusion of the transition model. From Table 4, the performance of 3D-Prover with the transition model embeddings is indeed superior to that of the Autoencoder across all values of $K$ . This shows that selecting for diversity with respect to the predicted semantics, rather than the syntax, leads to a direct improvement in proof performance. + +Table 4: Percentage of proofs found after one attempt (Pass@1) on miniF2F-valid, comparing 3D-Prover with a Transition Model Encoder to an Autoencoder trained to reconstruct the original tactics. We see that 3D-Prover with the Transition Model gives a clear improvement in proof success over the Autoencoder, demonstrating the utility of our representation architecture in Section 2. + +
K=8K=16K=32
Autoencoder23.027.927.0
3D-Prover25.029.128.7
+ +We have demonstrated that 3D-Prover improves proof success rate without any hyperparameter tuning, with a fixed $\lambda_{s} = \lambda_{\tau} = 0$ , $\theta = 1$ . We now examine whether we can use 3D-Prover to direct search to optimise secondary objectives, namely the execution time, tactic success rate and the diversity of environment response. + +Success Rate We observe from Table 5 3D-Prover significantly improves the success rate of chosen tactics. As $K$ decreases, this improvement increases in magnitude, reflecting the heightened influence of the filtering model. We see that this improvement increases with the error term $\lambda_{s}$ , which scales the quality scores of tactics by their predicted probability of success, as was intended. + +Table 5: Tactic success rate per node for miniF2F-valid (Mean ± Standard Error), where $\lambda_{s}$ controls the error weight of quality score in 3D-Prover. No filtering gives $27.7\% \pm 0.2\%$ . We see that 3D-Prover leads to fewer errors on average, which can be controlled by increasing $\lambda_{s}$ . + +
KTop-KRandom3D-Prover (λs=0.1)3D-Prover (λs=0.5)
839.0 ± 0.133.4 ± 0.143.3 ± 0.156.5 ± 0.1
1639.0 ± 0.130.9 ± 0.140.0 ± 0.151.7 ± 0.1
3235.0 ± 0.229.7 ± 0.135.7 ± 0.141.7 ± 0.1
+ +Diversity To measure diversity, we examine the percentage of successful tactics which result in a new proof path. We restrict to successful tactics to account for the discrepancy in success rate between approaches. We observe 3D-Prover gives more unique subgoals per successful tactic, which is noteworthy given the higher rate of successful tactics from 3D-Prover overall (Table 5). As intended, increasing $\theta$ gives further improvements under these metrics. This demonstrates that our approach is effective at avoiding redundant tactics, instead selecting tactics which yield more unique proof paths. Appendix F provides additional analysis, further supporting our claim of improved diversity. + +Table 6: Percentage of successful tactics per node resulting in unique subgoal(s) over miniF2F-valid (Mean ± Standard Error). No filtering gives $67.8\% \pm 0.3\%$ . We observe 3D-Prover results in more unique subgoals per tactic, leading to a more diverse set of proof paths, with larger $\theta$ controlling this. + +
KTop-KRandom3D-Prover (θ = 1)3D-Prover (θ = 4)
885.3 ± 0.189.9 ± 0.190.1 ± 0.191.1 ± 0.1
1677.5 ± 0.184.1 ± 0.184.9 ± 0.185.5 ± 0.1
3272.3 ± 0.276.3 ± 0.276.9 ± 0.277.5 ± 0.2
+ +Execution Time Table 7 shows the execution time for tactics over miniF2F-valid transitions. Again we see that 3D-Prover outperforms the baselines, with the improvement increasing with more filtering. Increasing the time weight $\lambda_{\tau}$ results in further reductions to the average execution time, demonstrating the accuracy of the predictions, and that they can directly result in faster tactics when filtering. It has been noted that preferring faster tactics can prevent the excessive application of powerful automation tactics such as simp (Lample et al., 2022, Appendix E). As these generally take longer to run, using faster tactics can help models learn underlying proof arguments which are often hidden by the automation. It can also greatly reduce the number of timeout errors. + +Table 7: Tactic execution time in milliseconds over miniF2F-valid proof attempts (Mean ± Standard Error). No filtering resulted in $232 \pm 0.9$ milliseconds. $\lambda_{\tau}$ controls the time weighting of the quality score in 3D-Prover. 3D-Prover selects faster tactics on average, with larger $\lambda_{\tau}$ magnifying this. + +
KTop-KRandom3D-Prover (λτ=0.1)3D-Prover (λτ=1.0)
8206 ± 0.8198 ± 0.9155 ± 0.5136 ± 0.5
16220 ± 0.8218 ± 0.9176 ± 0.6152 ± 0.5
32224 ± 0.8215 ± 0.8191 ± 0.7181 ± 0.6
+ +# 4 Conclusion + +**Limitations** Our main limitation was the scale of experiments we were able to run, with other results often requiring thousands of hours of train time using hundreds of provers and larger models (Lample et al., 2022; Polu et al., 2022; Xin et al., 2024). Given the large computational cost of evaluations (as we outline in Appendix I), we were only able to test InternLM2.5 up to Pass@2 for BFS and Pass@6 for Critic Guided, and ReProver up to Pass@4 on miniF2F-test, with a hyperparameter analysis of 15 runs on miniF2F-valid (Table C). We could only evaluate over the large LeanDojo benchmark (Appendix D) for single run with each approach. + +Summary We demonstrate the feasibility of learning proof system dynamics, where we generate tactic representations reflecting the response of the proof environment. We then leverage these with 3D-Prover, which filters candidate tactics to diverse and high quality subsets based on their likely outcome. We evaluate 3D-Prover by augmenting popular proving LLMs on the standard miniF2F and LeanDojo benchmarks, where we find an improvement in the overall proof success rate, particularly for deeper searches. Our ablation studies confirm the utility of our tactic representations, enabling the selection of tactics with improved success rates, diversity, and/or execution time. By effectively pruning the search space, 3D-Prover is a step towards enabling deeper automated proofs. + +# 5 Acknowledgements + +We acknowledge Defence Science and Technology Group (DSTG) for their support in this project. + +# References + +Chen, M., Tworek, J., Jun, H., Yuan, Q., Pinto, H. P. d. O., Kaplan, J., Edwards, H., Burda, Y., Joseph, N., Brockman, G., Ray, A., Puri, R., Krueger, G., Petrov, M., Khlaaf, H., Sastry, G., Mishkin, P., Chan, B., Gray, S., Ryder, N., Pavlov, M., Power, A., Kaiser, L., Bavarian, M., Winter, C., Tillet, P., Such, F. P., Cummings, D., Plappert, M., Chantzis, F., Barnes, E., Herbert-Voss, A., Guss, W. H., Nichol, A., Paino, A., Tezak, N., Tang, J., Babuschkin, I., Balaji, S., Jain, S., Saunders, W., Hesse, C., Carr, A. N., Leike, J., Achiam, J., Misra, V., Morikawa, E., Radford, A., Knight, M., Brundage, M., Murati, M., Mayer, K., Welinder, P., McGrew, B., Amodei, D., McCandlish, S., Sutskever, I., and Zaremba, W. Evaluating Large Language Models Trained on Code, 2021. URL https://arxiv.org/abs/2107.03374. Version Number: 2. +De Moura, L., Kong, S., Avigad, J., Van Doorn, F., and Von Raumer, J. The Lean Theorem Prover (System Description). volume 9195, pp. 378-388, Cham, 2015. Springer International Publishing. ISBN 978-3-319-21400-9 978-3-319-21401-6. doi: 10.1007/978-3-319-21401-6_26. URL http://link.springer.com/10.1007/978-3-319-21401-6_26. Book Title: Automated Deduction - CADE-25 Series Title: Lecture Notes in Computer Science. +First, E. and Brun, Y. Diversity-driven automated formal verification. In Proceedings of the 44th International Conference on Software Engineering, ICSE '22, pp. 749-761, New York, NY, USA, July 2022. Association for Computing Machinery. ISBN 978-1-4503-9221-1. doi: 10.1145/3510003.3510138. URL https://dl.acm.org/doi/10.1145/3510003.3510138. +Hales, T., Adams, M., Bauer, G., Dang, T. D., Harrison, J., Hoang, L. T., Kaliszyk, C., Magron, V., Mclaughlin, S., Nguyen, T. T., Nguyen, Q. T., Nipkow, T., Obua, S., Pleso, + +J., Rute, J., Solovyev, A., Ta, T. H. A., Tran, N. T., Trieu, T. D., Urban, J., Vu, K., and Zumkeller, R. A FORMAL PROOF OF THE KEPLER CONJECTURE. Forum of Mathematics, Pi, 5:e2, January 2017. ISSN 2050-5086. doi: 10.1017/fmp.2017.1. URL https://www.cambridge.org/core/journals/forum-of-mathematics/pi/article/formal-proof-of-the-kepler-conjecture/78FBD5E1A3D1BCCB8E0D5B0C463C9FBC. Publisher: Cambridge University Press. +Hsiao, W.-L. and Grauman, K. Creating capsule wardrobes from fashion images. In CVPR, pp. 7161-7170, 06 2018. doi: 10.1109/CVPR.2018.00748. +Klein, G., Elphinstone, K., Heiser, G., Andronick, J., Cock, D., Derrin, P., Elkaduwe, D., Engelhardt, K., Kolanski, R., Norrish, M., Sewell, T., Tuch, H., and Winwood, S. seL4: formal verification of an OS kernel. In Proceedings of the ACM SIGOPS 22nd symposium on Operating systems principles, SOSP '09, pp. 207-220, New York, NY, USA, October 2009. Association for Computing Machinery. ISBN 978-1-60558-752-3. doi: 10.1145/1629575.1629596. URL https://doi.org/10.1145/1629575.1629596. +Kulesza, A. Determinantal Point Processes for Machine Learning. Foundations and Trends® in Machine Learning, 5(2-3):123-286, 2012. ISSN 1935-8245. doi: 10.1561/220000044. URL http://dx.doi.org/10.1561/220000044. Publisher: Now Publishers. +Kulesza, A. and Taskar, B. k-DPPs: fixed-size determinantal point processes. In Proceedings of the 28th International Conference on International Conference on Machine Learning, ICML'11, pp. 1193-1200, Madison, WI, USA, June 2011. Omnipress. ISBN 978-1-4503-0619-5. +Lamont, S., Norrish, M., Dezfouli, A., Walder, C., and Montague, P. BAIT: Benchmarking (Embedding) Architectures for Interactive Theorem-Proving. Proceedings of the AAAI Conference on Artificial Intelligence, 38:10607-10615, March 2024. doi: 10.1609/aaai.v38i9.28931. +Lample, G., Lacroix, T., Lachaux, M.-a., Rodriguez, A., Hayat, A., Lavril, T., Ebner, G., and Martinet, X. HyperTree Proof Search for Neural Theorem Proving. In Advances in Neural Information Processing Systems, October 2022. URL https://openreview.net/forum?id=J4pX8Q8cxHH. +Li, Z., Sun, J., Murphy, L., Su, Q., Li, Z., Zhang, X., Yang, K., and Si, X. A survey on deep learning for theorem proving, 2024. URL https://arxiv.org/abs/2404.09939. +Lin, C.-Y. ROUGE: A Package for Automatic Evaluation of Summaries. In Text Summarization Branches Out, pp. 74-81, Barcelona, Spain, July 2004. Association for Computational Linguistics. URL https://aclanthology.org/W04-1013. +Papineni, K., Roukos, S., Ward, T., and Zhu, W.-J. Bleu: a Method for Automatic Evaluation of Machine Translation. In Isabelle, P., Charniak, E., and Lin, D. (eds.), Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics, pp. 311-318, Philadelphia, Pennsylvania, USA, July 2002. Association for Computational Linguistics. doi: 10.3115/1073083.1073135. URL https://aclanthology.org/P02-1040. +Polu, S., Han, J. M., Zheng, K., Baksys, M., Babuschkin, I., and Sutskever, I. Formal mathematics statement curriculum learning. In The Eleventh International Conference on Learning Representations, 2022. URL https://openreview.net/forum?id=-P7G-8dmSh4. +Reddy, R. Foundations and Grand Challenges of Artificial Intelligence: AAAI Presidential Address. AI Magazine, 9(4):9, December 1988. doi: 10.1609/ajmag.v9i4.950. URL https://ojs(aaai.org/ajmagazine/index.php/ajmagazine/article/view/950. Section: Articles. +Tan, Y. K., Myreen, M. O., Kumar, R., Fox, A., Owens, S., and Norrish, M. The verified CakeML compiler backend. Journal of Functional Programming, 29:e2, January 2019. ISSN 0956-7968, 1469-7653. doi: 10.1017/S0956796818000229. URL https://www.cambridge.org/core/journals/journal-of-functional-programming/article/verified-cakeml-compiler-backend/E43ED3EA740D2DF970067F4E2BB9EF7D. Publisher: Cambridge University Press. + +Thakur, A., Tsoukalas, G., Wen, Y., Xin, J., and Chaudhuri, S. An In-Context Learning Agent for Formal Theorem-Proving, 2023. URL https://arxiv.org/abs/2310.04353. Version Number: 5. +Wang, H., Yuan, Y., Liu, Z., Shen, J., Yin, Y., Xiong, J., Xie, E., Shi, H., Li, Y., Li, L., Yin, J., Li, Z., and Liang, X. DT-Solver: Automated Theorem Proving with Dynamic-Tree Sampling Guided by Proof-level Value Function. In Rogers, A., Boyd-Graber, J., and Okazaki, N. (eds.), Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pp. 12632–12646, Toronto, Canada, July 2023. Association for Computational Linguistics. doi: 10.18653/v1/2023.acl-long.706. URL https://aclanthology.org/2023.acl-long.706. +Wu, Z., Huang, S., Zhou, Z., Ying, H., Wang, J., Lin, D., and Chen, K. InternLM2.5-StepProver: Advancing Automated Theorem Proving via Expert Iteration on Large-Scale LEAN Problems, October 2024. URL http://arxiv.org/abs/2410.15700. arXiv:2410.15700 [cs]. +Xin, H., Ren, Z. Z., Song, J., Shao, Z., Zhao, W., Wang, H., Liu, B., Zhang, L., Lu, X., Du, Q., Gao, W., Zhu, Q., Yang, D., Gou, Z., Wu, Z. F., Luo, F., and Ruan, C. DeepSeek-Prover-V1.5: Harnessing Proof Assistant Feedback for Reinforcement Learning and Monte-Carlo Tree Search, 2024. URL https://arxiv.org/abs/2408.08152. Version Number: 1. +Yang, K., Swope, A., Gu, A., Chalamala, R., Song, P., Yu, S., Godil, S., Prenger, R., and Anandkumar, A. LeanDojo: Theorem proving with retrieval-augmented language models. In Neural Information Processing Systems (NeurIPS), 2023. +Yang, X.-W., Zhou, Z., Wang, H., Li, A., Wei, W.-D., Jin, H., Li, Z., and Li, Y.-F. Carts: Advancing neural theorem proving with diversified tactic calibration and bias-resistant tree search. *ICLR*, 2025. +Zhang, K., Chao, W.-L., Sha, F., and Grauman, K. Video summarization with long short-term memory. In Leibe, B., Matas, J., Sebe, N., and Welling, M. (eds.), Computer Vision - ECCV 2016, pp. 766-782, Cham, 2016. Springer International Publishing. ISBN 978-3-319-46478-7. +Zheng, K., Han, J. M., and Polu, S. miniF2F: a cross-system benchmark for formal olympiad-level mathematics. In International Conference on Learning Representations, 2021. URL https://openreview.net/forum?id=9ZPegFuFTFv. + +# A Proof Search Setup + +We first define the space of goals $S$ , tactics $\mathcal{T}$ and failures $\mathcal{F}$ . For our purposes, these all contain arbitrary strings, with the goal being a formal proposition, the tactic a command and the failure an error message. We then define the output space as $\mathcal{O} := \mathcal{P}(S) \cup \mathcal{F}$ . A proof tree is a DAG $G = (V, E)$ where $V \subset S$ is the set of goals and $E$ the edges between them. A proof attempt for a goal $g_0$ first internalises the proof tree with $V = \{g_0\}$ , $E = \emptyset$ . The search policy $\pi_S : G \times V \to \mathbb{R}^+$ is a distribution over goals given a proof tree, being used to select a goal $g$ to expand. The tactic policy $\pi_T : S \times T \to \mathbb{R}^+$ is a distribution over tactics given a goal, where $N \in \mathbb{N}$ tactics are sampled to give tactics $\{t_i\}_{i=1}^N \subset T$ . The goal, tactic pairs $(g, t_i)$ are then passed to the environment $\mathcal{E} : S \times T \to \mathcal{O}$ . For each pair, after $\tau_i \in \mathbb{R}$ seconds, it returns either a new set of goals $g_i' \subset S$ or an error, $e_i \in \mathcal{F}$ . We define this response as the output $o_i \in \mathcal{O}$ . We further define the status $s_i \in \{0, 1\}$ as 0 if $o_i \in \mathcal{F}$ , 1 if $o_i \in \mathcal{P}(S)$ and the transition as the tuple $(g, t_i, s_i, \tau_i, o_i)$ . The proof tree is then updated with $G = G \cup g_i'$ for all $g_i'$ , and the associated transitions are added as edges to $E$ . This is repeated until a proof is found, or a budget is exhausted. A proof of $g$ is found when $\mathcal{E}(g, t_i) = \emptyset$ for any $t_i$ , or if all $\{g_i'\}$ are proven for $\mathcal{E}(g, t_i) = \{g_i'\} \subset S$ . The result of a proof attempt is then the set of transitions $\{(g_k, t_{ki}, s_{ki}, \tau_{ki}, o_{ki})\}$ for all selected goals $g_k$ and their expanded tactics $t_i$ . For simplicity, we drop the indices to denote the set of transitions as $\{(g, t, s, \tau, o)\}$ . + +# B Pass@k + +Table 8 summarises the Pass@4 results for ReProver over miniF2F-test, which is the number of proofs found at least once over four attempts, with Table 9 showing the Pass@ $k$ up to $k = 4$ . We compare + +Table 8: Percentage of proofs found after four attempts (Pass@4) with ReProver on miniF2F-test, with $K$ tactics selected per node. + +
KRandom3D-ProverGain
825.728.6+11.3%
1630.231.0+2.6%
3229.829.8+0.0%
+ +Table 9: Pass@k rates for proof attempts on miniF2F-test + +
Pass@k3D-ProverRandom
K=8K=16K=32K=8K=16K=32
124.927.828.618.021.228.1
226.129.429.022.928.629.0
326.529.829.824.929.429.8
428.631.029.825.730.229.8
+ +3D-Prover to the Random baseline, taking the same four runs from Table 2, where $\lambda_s = \lambda_\tau = 0$ , $\theta = 1$ . With Top- $K$ being deterministic, the Pass@ $k$ rate is the same as the Pass@1 rate. Given several attempts, $K = 16$ appears to provide a good tradeoff between breadth and depth, performing the best overall. 3D-Prover maintains a large improvement for $K = 8$ , with a modest improvement for $K = 16$ . + +As discussed by Chen et al. (2021), the Pass@k metric favours exploratory approaches as $k$ increases, at the cost of lower performance for smaller $k$ . This is because, over many attempts, a highly exploratory approach is more likely to find at least one proof of a given goal, even though it may find fewer proofs in a single attempt than a more exploitative approach. Further discussion in Lample et al. (2022) finds that randomly sampling search parameters also improves Pass@k. With Pass@k being expensive to estimate, we fix our parameters over the four runs to give a more accurate estimate of Pass@1. Given this, a large scale experiment sampling these hyperparameters could lead to improved Pass@k results, as Lample et al. (2022) show for their HTTPS approach. + +# C Proof success rate over hyperparameters + +Table 10: Pass@1 results on miniF2F-valid, over different hyperparameter configurations for 3D-Prover with ReProver. + +
K(λs, λτ, θ)
(0.0, 0.0, 1.0)(0.1, 0.1, 1.0)(0.5, 0.1, 1.0)(0.1, 1.0, 1.0)(0.1, 0.1, 4.0)
825.025.025.822.523.8
1629.128.727.927.026.6
3228.728.328.727.927.0
+ +Table 10 shows the Pass@1 results on miniF2F-valid for 3D-Prover for our limited hyperparameter sweep. These results suggest that a lower time weight $\lambda_{\tau}$ leads to better proving results. The diversity parameter $\theta$ hinders performance for the larger value, consistent with what was observed by Chen et al. (2021), where they observe a tradeoff between exploration and Pass@1. Although these parameters may not improve Pass@1, different proofs may favour different configurations, with some requiring e.g. more depth or exploration than others. As discussed above, a higher Pass@k can usually be obtained by sampling a wide set of these parameters. For the set of hyperparameters we tested here, we found a cumulative proof rate (or Pass@15) of $32.8\%$ on miniF2F-valid. + +# D Evaluation on LeanDojo Benchmark + +We ran an additional experiment on the LeanDojo Novel Premises Yang et al. (2023) benchmark testing 3D-Prover on a larger dataset. This dataset has 2000 evaluation proofs in comparison to the 244 from miniF2F-valid and miniF2F-test, allowing us to evaluate the performance of 3D-Prover on a larger scale. + +We trained a transition model from a single ReProver attempt on LeanDojo Novel Premises, before evaluating 3D-Prover using ReProver with the methodology in Section 3. We set $K = 32$ for 3D-Prover, and compare to the model with No Filtering (i.e. $K = 64$ ), and Top- $K = 32$ . We further examine the distribution of proof lengths found from this experiment. To account for different proofs of the same goal, we adjust proof lengths to be the shortest found from any attempt (e.g. if 3D-Prover finds a proof of length 10, which was found in 3 steps by No Filtering, we count it as length 3). Hence, all proof lengths reported are the shortest found by any method. We report the number of proofs found by each approach, organised by the proof length in Table 11. + +Table 11: Number of Proofs found on LeanDojo Novel Premises, sorted by proof length. + +
Proof Length3D-Prover (K=32)Top-K (K=32)No Filtering (K=64)
1236233237
2167162174
3134126131
4606054
5403924
6762
7200
Total646626622
Pass@132.3%31.3%31.1%
+ +3D-Prover obtains a relative improvement of $3.2\%$ over Top- $K$ , and a $3.9\%$ relative improvement over No Filtering in terms of the number of proofs found, comparable to the same performance gain for $K = 32$ found in miniF2F-valid (Table 2). We see that 3D-Prover finds deeper proofs, while maintaining a high proof success rate for shallower proofs, unlike Top- $K$ . The No Filtering approach, as expected, finds the most shallow proofs, however quickly drops off in performance for deeper proofs. We also note that 3D-Prover found the 2 longest proofs of length 7, with neither baseline finding any. + +# E Embedding Discussion + +Embedding Comparison We now investigate whether the transition model (Figure 2) captures tactic semantics rather than syntax in its tactic embeddings. To test this, we examine the cosine similarity of tactic embeddings which lead to unique subgoals. Figure 4 takes an example node, examining all tactics which lead to a unique subgoal. The upper value displays the cosine similarity given by the transition model, while the lower value displays that given by the Autoencoder in Section 3.1.2. We observe that in most cases, the similarity given by the transition model is much lower than that given by the Autoencoder, which is only considering the syntax of the tactic. For example, the similarity between tactic 3 and 4 is very high for the Autoencoder, given the similar syntax between the two as they use the same lemma. Despite this similar syntax, the transition model embeddings show a high degree of dissimilarity, reflecting the different outcome they have on the environment. We present additional examples in the supplementary code. To generalise beyond these examples, we ran this comparison over the tactic embeddings which lead to unique subgoals for all 244 root nodes in minF2F-valid. Figure 5 shows the distribution of the average cosine similarity for each node, for both the transition model and the Autoencoder. The average cosine similarity for the transition model embeddings was 0.44 while the Autoencoder gave 0.57. While this comparison does not account for similarity between the unique subgoals, it is still clear that the transition model + +![](images/7adcb090877ee7acd81de784fc043fd7bc5b90333b47c30f5e4dea03a1b1b0d7.jpg) +Figure 4: Cosine similarity between tactic embeddings resulting in unique subgoals, for a sample root node in miniF2F-valid. The top value gives the similarity for embeddings from 3D-Prover, while the bottom gives the similarity for embeddings from an Autoencoder. We see that 3D-Prover better separates these semantically distinct tactics, in comparison to the Autoencoder, which only separates based on their syntax. + +![](images/35ca1ca546b0097ea7a5abf3db6efd41808ab181ec8d10569c030910c3a5dc22.jpg) +Figure 5: Distribution of cosine similarity for tactic embeddings resulting in unique subgoals, averaged over root nodes in miniF2F-valid. We see that 3D-Prover gives embeddings which better separate these semantically distinct tactics, in comparison the the syntax focused embeddings of the Autoencoder. + +embeddings better separate unique tactics than Autoencoder embeddings which are based on syntax alone. The result of this is a higher likelihood of 3D-Prover selecting tactics which give unique subgoals, which as we show in Section 3.1.2, results in the transition model outperforming the Autoencoder for proof discovery. + +**Embedding Objective** As outlined in Section 2, we train our embeddings to be reflective of the tactic semantics across all three components of Status, Time and Output. Hence 3D-Prover, which selects diverse embeddings, may lead to tactics predicted to have errors, where the errors are diverse in terms of their predicted message. The hyperparameter $\lambda_{s}$ can alleviate this by weighting the scores based on their likelihood of success. From our experiments (Table 10), there is not necessarily a benefit to Pass@1 by filtering out strongly based on the predicted error likelihood. To speculate, the error prediction, although quite good, is imperfect with false negatives (Table 1). This can lead to potentially useful tactics being ignored if the error prediction is overly trusted, even though there is a higher tactic success rate overall as in Table 5. Given these prediction errors, it may be the case that selecting goals which are predicted to lead to (diverse) errors may be preferable, given the possibility they result in successful new subgoals. These subgoals may be quite different from those previously selected, as they are mispredicted, so are clearly outside the space of tactics + +where the transition model is confident about the outcome. Further analysis could be worthwhile to investigate this. An embedding architecture trained only on successful tactics could be used, however given the high error rate of tactics, this would ignore a large proportion of the transition data. + +# F Additional Diversity Analysis + +To further examine diversity, we first look at the percentage of unique environment responses to tactics executed per node, including responses with unique errors (Table 12), using the same ReProver setup in 3. As it is difficult to select tactics guaranteed to be successful (see Table 5), we would expect a good exploratory policy to select tactics which result in more varied outputs (both errors and successes alike), so as to better explore the space. We also examine the degree of cosine similarity across unique subgoals (Table 13), using a simple text embedding model (a11-MiniLM-L6-v2). This more precisely quantifies the diversity in the contents of the subgoals. We do this to account for simple changes (such as variable renaming), which would not be differentiated under a simple check for uniqueness, and so allows us to examine how varied the contents of the resulting subgoals are. + +In both cases, we see that 3D-Prover results in more diverse responses. As intended, increasing $\theta$ results in further improvements to diversity under both metrics. The increased diversity in subgoal content (Table 13) is strengthened by the fact that 3D-Prover also gives more unique subgoals on average (Table 6), suggesting that our approach yields more unique subgoals, and that these subgoals are more varied in their contents. + +Table 12: Percentage of unique environment responses per node in miniF2F-valid (Mean ± Standard Error). Unique defines either syntactically distinct error messages or responses including at least one previously unseen subgoal. No filtering results in $63.3\% \pm 0.2\%$ . We see that 3D-Prover gives a higher diversity of environment responses, increasing with the diversity parameter $\theta$ . + +
KTop-KRandom3D-Prover
θ = 1θ = 4
883.9 ± 0.188.6 ± 0.190.8 ± 0.091.7 ± 0.0
1677.5 ± 0.181.4 ± 0.185.9 ± 0.186.6 ± 0.1
3271.1 ± 0.172.7 ± 0.177.6 ± 0.178.1 ± 0.1
+ +Table 13: Average cosine similarity between subgoal embeddings for ReProver over miniF2F-valid (Mean ± Standard Error). We observe consistently lower similarity among subgoals generated from 3D-Prover, increasing with the diversity parameter $\theta$ . + +
KTop-KRandom3D-Prover
θ = 1θ = 4
80.939 ± 0.0000.945 ± 0.0000.931 ± 0.0000.930 ± 0.000
160.916 ± 0.0000.926 ± 0.0000.910 ± 0.0000.905 ± 0.000
320.904 ± 0.0000.909 ± 0.0010.900 ± 0.0010.896 ± 0.001
+ +# G Computational Overhead + +3D-Prover adds a constant but minimal time and memory overhead, the majority of which is in generating embeddings for the candidate tactics. Taking our first run for 3D-Prover with InternLM, we found the filtering time over nodes to be 0.07s with a standard deviation of 0.02s. In comparison, the average tactic generation time was 7.5s with a standard deviation of 4s. This gives us a time overhead of approximately $0.9\%$ . The memory overhead was approximately 3GB of VRAM, while the tactic model took 44GB of VRAM, giving approximately $7\%$ memory overhead. + +We also note that the average time to execute a tactic from the No Filtering model was approximately 0.13s (with 0.12 standard deviation). With up to 64 candidate tactics per node, the filtering time + +of 0.07 seconds is around half the average time to execute a single tactic. Given the improvements in success rate (Table 5), our filtering model therefore gives an effective way to reduce the total computational resources by preventing the wasteful execution of erroneous tactics. To further support this, we observed the average total time for a proof search attempt with 3D-Prover with InternLM $(K = 8)$ to be 993.8 seconds (1761.5 Standard Deviation), while the total time for Top- $K = 8$ was 1116.8 seconds (2126.0 Standard Deviation). + +# H Number of tactic candidates for InternLM + +As noted in 3, our experiments with InternLM2.5-Step-Prover use sampling rather than beam search for tactic generation, as done in Wu et al. (2024). As a result, there is no guarantee there will be 64 unique tactics, with many samples being identical. As a result, we sample 128 initial tactics, and take the total unique candidates as our ground set (up to 64 total). We plot the distribution of unique tactic candidates per node in Figure 6. This informed our choice of $K = 8$ for our experiments, as higher $K$ would give minimal filtering in comparison to a beam search approach. + +![](images/7c63e8bfbb380927dc12a17de2f6be9226aa7f94bf6dde7d44942af06026e1f8.jpg) +Figure 6: Distribution of unique tactic candidates per node for InternLM2.5-Step-Prover over miniF2F-test $(\mu = 35.4, \sigma = 11.5)$ . + +# I Computational Resources and Usage + +For training our transition models, we used a single RTX4090 GPU with an Intel i9 13900k processor. For evaluating, we used two internal machines each with two RTXA6000 GPUs, and a Intel Xeon W-2223. For each evaluation experiment in 3, we assigned a single RTX A6000 GPU which contained both the tactic generator and transition model, which served 2 CPU provers which request tactics and evaluate in the Lean environment. + +For each transition model, training for 2 epochs took approximately 2 days for each run, giving a total of 10 days of 4090 training for our results in 2.2. Each evaluation run for ReProver over miniF2F took around 12h, while each evaluation run for InternLM2.5-Step-Prover around 2 days. Counting the number of runs for all baselines and 3D-Prover, we have 28 runs for ReProver with miniF2F-test, 22 for miniF2F-valid and 17 runs for InternLM. The additional result over the large LeanDojo dataset in D took around 5 days per run, with 7 runs total. We therefore estimate the total evaluation time (for a single machine with 2 RTX6000s and a Xeon W-223) to be 94 days, which was around 47 days with our 2 machines. The full research project included additional compute and experiments, as different architectures and approaches were prototyped, however we do not have an estimate for the amount. + +# NeurIPS Paper Checklist + +# 1. Claims + +Question: Do the main claims made in the abstract and introduction accurately reflect the paper's contributions and scope? + +Answer: [Yes] + +Justification: Section 2 demonstrates our first claim of proof outcome learning, with Section 3 demonstrating our second and third claims, showing improvements over two base models with our approach. + +Guidelines: + +- The answer NA means that the abstract and introduction do not include the claims made in the paper. +- The abstract and/or introduction should clearly state the claims made, including the contributions made in the paper and important assumptions and limitations. A No or NA answer to this question will not be perceived well by the reviewers. +- The claims made should match theoretical and experimental results, and reflect how much the results can be expected to generalize to other settings. +- It is fine to include aspirational goals as motivation as long as it is clear that these goals are not attained by the paper. + +# 2. Limitations + +Question: Does the paper discuss the limitations of the work performed by the authors? + +Answer: [Yes] + +Justification: Please see the limitations section 4, and the section on computational cost $G$ . + +Guidelines: + +- The answer NA means that the paper has no limitation while the answer No means that the paper has limitations, but those are not discussed in the paper. +- The authors are encouraged to create a separate "Limitations" section in their paper. +- The paper should point out any strong assumptions and how robust the results are to violations of these assumptions (e.g., independence assumptions, noiseless settings, model well-specification, asymptotic approximations only holding locally). The authors should reflect on how these assumptions might be violated in practice and what the implications would be. +- The authors should reflect on the scope of the claims made, e.g., if the approach was only tested on a few datasets or with a few runs. In general, empirical results often depend on implicit assumptions, which should be articulated. +- The authors should reflect on the factors that influence the performance of the approach. For example, a facial recognition algorithm may perform poorly when image resolution is low or images are taken in low lighting. Or a speech-to-text system might not be used reliably to provide closed captions for online lectures because it fails to handle technical jargon. +- The authors should discuss the computational efficiency of the proposed algorithms and how they scale with dataset size. +- If applicable, the authors should discuss possible limitations of their approach to address problems of privacy and fairness. +- While the authors might fear that complete honesty about limitations might be used by reviewers as grounds for rejection, a worse outcome might be that reviewers discover limitations that aren't acknowledged in the paper. The authors should use their best judgment and recognize that individual actions in favor of transparency play an important role in developing norms that preserve the integrity of the community. Reviewers will be specifically instructed to not penalize honesty concerning limitations. + +# 3. Theory assumptions and proofs + +Question: For each theoretical result, does the paper provide the full set of assumptions and a complete (and correct) proof? + +# Answer: [NA] + +Justification: No new theoretical results are presented. + +# Guidelines: + +- The answer NA means that the paper does not include theoretical results. +- All the theorems, formulas, and proofs in the paper should be numbered and cross-referenced. +- All assumptions should be clearly stated or referenced in the statement of any theorems. +- The proofs can either appear in the main paper or the supplemental material, but if they appear in the supplemental material, the authors are encouraged to provide a short proof sketch to provide intuition. +- Inversely, any informal proof provided in the core of the paper should be complemented by formal proofs provided in appendix or supplemental material. +- Theorems and Lemmas that the proof relies upon should be properly referenced. + +# 4. Experimental result reproducibility + +Question: Does the paper fully disclose all the information needed to reproduce the main experimental results of the paper to the extent that it affects the main claims and/or conclusions of the paper (regardless of whether the code and data are provided or not)? + +# Answer: [Yes] + +Justification: Our transition model and 3D-Prover architectures are fully described in section 2 and 3, with code provided in the supplementary material. + +# Guidelines: + +- The answer NA means that the paper does not include experiments. +- If the paper includes experiments, a No answer to this question will not be perceived well by the reviewers: Making the paper reproducible is important, regardless of whether the code and data are provided or not. +- If the contribution is a dataset and/or model, the authors should describe the steps taken to make their results reproducible or verifiable. +- Depending on the contribution, reproducibility can be accomplished in various ways. For example, if the contribution is a novel architecture, describing the architecture fully might suffice, or if the contribution is a specific model and empirical evaluation, it may be necessary to either make it possible for others to replicate the model with the same dataset, or provide access to the model. In general, releasing code and data is often one good way to accomplish this, but reproducibility can also be provided via detailed instructions for how to replicate the results, access to a hosted model (e.g., in the case of a large language model), releasing of a model checkpoint, or other means that are appropriate to the research performed. +- While NeurIPS does not require releasing code, the conference does require all submissions to provide some reasonable avenue for reproducibility, which may depend on the nature of the contribution. For example +(a) If the contribution is primarily a new algorithm, the paper should make it clear how to reproduce that algorithm. +(b) If the contribution is primarily a new model architecture, the paper should describe the architecture clearly and fully. +(c) If the contribution is a new model (e.g., a large language model), then there should either be a way to access this model for reproducing the results or a way to reproduce the model (e.g., with an open-source dataset or instructions for how to construct the dataset). +(d) We recognize that reproducibility may be tricky in some cases, in which case authors are welcome to describe the particular way they provide for reproducibility. In the case of closed-source models, it may be that access to the model is limited in some way (e.g., to registered users), but it should be possible for other researchers to have some path to reproducing or verifying the results. + +# 5. Open access to data and code + +Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material? + +Answer: [Yes] + +Justification: All code is provided in the supplementary material, with instructions for reproducing the main results. + +Guidelines: + +- The answer NA means that paper does not include experiments requiring code. +- Please see the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details. +- While we encourage the release of code and data, we understand that this might not be possible, so "No" is an acceptable answer. Papers cannot be rejected simply for not including code, unless this is central to the contribution (e.g., for a new open-source benchmark). +- The instructions should contain the exact command and environment needed to run to reproduce the results. See the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details. +- The authors should provide instructions on data access and preparation, including how to access the raw data, preprocessed data, intermediate data, and generated data, etc. +- The authors should provide scripts to reproduce all experimental results for the new proposed method and baselines. If only a subset of experiments are reproducible, they should state which ones are omitted from the script and why. +- At submission time, to preserve anonymity, the authors should release anonymized versions (if applicable). +- Providing as much information as possible in supplemental material (appended to the paper) is recommended, but including URLs to data and code is permitted. + +# 6. Experimental setting/details + +Question: Does the paper specify all the training and test details (e.g., data splits, hyperparameters, how they were chosen, type of optimizer, etc.) necessary to understand the results? + +Answer: [Yes] + +Justification: We specify all training and test details for our results, as outlined in each relevant section. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The experimental setting should be presented in the core of the paper to a level of detail that is necessary to appreciate the results and make sense of them. +- The full details can be provided either with the code, in appendix, or as supplemental material. + +# 7. Experiment statistical significance + +Question: Does the paper report error bars suitably and correctly defined or other appropriate information about the statistical significance of the experiments? + +Answer: [Yes] + +Justification: We clearly state the standard error for the relevant results. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The authors should answer "Yes" if the results are accompanied by error bars, confidence intervals, or statistical significance tests, at least for the experiments that support the main claims of the paper. +- The factors of variability that the error bars are capturing should be clearly stated (for example, train/test split, initialization, random drawing of some parameter, or overall run with given experimental conditions). + +- The method for calculating the error bars should be explained (closed form formula, call to a library function, bootstrap, etc.) +- The assumptions made should be given (e.g., Normally distributed errors). +- It should be clear whether the error bar is the standard deviation or the standard error of the mean. +- It is OK to report 1-sigma error bars, but one should state it. The authors should preferably report a 2-sigma error bar than state that they have a $96\%$ CI, if the hypothesis of Normality of errors is not verified. +- For asymmetric distributions, the authors should be careful not to show in tables or figures symmetric error bars that would yield results that are out of range (e.g. negative error rates). +- If error bars are reported in tables or plots, The authors should explain in the text how they were calculated and reference the corresponding figures or tables in the text. + +# 8. Experiments compute resources + +Question: For each experiment, does the paper provide sufficient information on the computer resources (type of compute workers, memory, time of execution) needed to reproduce the experiments? + +Answer: [Yes] + +Justification: Computational resources are detailed in I + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The paper should indicate the type of compute workers CPU or GPU, internal cluster, or cloud provider, including relevant memory and storage. +- The paper should provide the amount of compute required for each of the individual experimental runs as well as estimate the total compute. +- The paper should disclose whether the full research project required more compute than the experiments reported in the paper (e.g., preliminary or failed experiments that didn't make it into the paper). + +# 9. Code of ethics + +Question: Does the research conducted in the paper conform, in every respect, with the NeurIPS Code of Ethics https://neurips.cc/public/EthicsGuidelines? + +Answer: [Yes] + +Justification: Yes, all research conforms to the NeurIPS Code of Ethics. + +Guidelines: + +- The answer NA means that the authors have not reviewed the NeurIPS Code of Ethics. +- If the authors answer No, they should explain the special circumstances that require a deviation from the Code of Ethics. +- The authors should make sure to preserve anonymity (e.g., if there is a special consideration due to laws or regulations in their jurisdiction). + +# 10. Broader impacts + +Question: Does the paper discuss both potential positive societal impacts and negative societal impacts of the work performed? + +Answer: [Yes] + +Justification: Positive societal impacts are discussed in the introduction (i.e. furthering research in the area of Neural Theorem Proving). As Neural Theorem Proving models are used only for the benign task of proving theorems, we find no clear negative societal impacts. + +Guidelines: + +- The answer NA means that there is no societal impact of the work performed. +- If the authors answer NA or No, they should explain why their work has no societal impact or why the paper does not address societal impact. + +- Examples of negative societal impacts include potential malicious or unintended uses (e.g., disinformation, generating fake profiles, surveillance), fairness considerations (e.g., deployment of technologies that could make decisions that unfairly impact specific groups), privacy considerations, and security considerations. +- The conference expects that many papers will be foundational research and not tied to particular applications, let alone deployments. However, if there is a direct path to any negative applications, the authors should point it out. For example, it is legitimate to point out that an improvement in the quality of generative models could be used to generate deepfakes for disinformation. On the other hand, it is not needed to point out that a generic algorithm for optimizing neural networks could enable people to train models that generate Deepfakes faster. +- The authors should consider possible harms that could arise when the technology is being used as intended and functioning correctly, harms that could arise when the technology is being used as intended but gives incorrect results, and harms following from (intentional or unintentional) misuse of the technology. +- If there are negative societal impacts, the authors could also discuss possible mitigation strategies (e.g., gated release of models, providing defenses in addition to attacks, mechanisms for monitoring misuse, mechanisms to monitor how a system learns from feedback over time, improving the efficiency and accessibility of ML). + +# 11. Safeguards + +Question: Does the paper describe safeguards that have been put in place for responsible release of data or models that have a high risk for misuse (e.g., pretrained language models, image generators, or scraped datasets)? + +Answer: [NA] + +Justification: There is no clear risk of misuse of our data or models, as they are used only for theorem proving, which is a benign application. + +Guidelines: + +- The answer NA means that the paper poses no such risks. +- Released models that have a high risk for misuse or dual-use should be released with necessary safeguards to allow for controlled use of the model, for example by requiring that users adhere to usage guidelines or restrictions to access the model or implementing safety filters. +- Datasets that have been scraped from the Internet could pose safety risks. The authors should describe how they avoided releasing unsafe images. +- We recognize that providing effective safeguards is challenging, and many papers do not require this, but we encourage authors to take this into account and make a best faith effort. + +# 12. Licenses for existing assets + +Question: Are the creators or original owners of assets (e.g., code, data, models), used in the paper, properly credited and are the license and terms of use explicitly mentioned and properly respected? + +Answer: [Yes] + +Justification: All code, data and models used in the paper are properly cited, with their usage properly respected. + +Guidelines: + +- The answer NA means that the paper does not use existing assets. +- The authors should cite the original paper that produced the code package or dataset. +- The authors should state which version of the asset is used and, if possible, include a URI. +- The name of the license (e.g., CC-BY 4.0) should be included for each asset. +- For scraped data from a particular source (e.g., website), the copyright and terms of service of that source should be provided. + +- If assets are released, the license, copyright information, and terms of use in the package should be provided. For popular datasets, paperswithcode.com/datasets has curated licenses for some datasets. Their licensing guide can help determine the license of a dataset. +- For existing datasets that are re-packaged, both the original license and the license of the derived asset (if it has changed) should be provided. +- If this information is not available online, the authors are encouraged to reach out to the asset's creators. + +# 13. New assets + +Question: Are new assets introduced in the paper well documented and is the documentation provided alongside the assets? + +Answer: [Yes] + +Justification: The code for the paper is the primary asset introduced, which is documented and provided in the supplementary material. + +Guidelines: + +- The answer NA means that the paper does not release new assets. +- Researchers should communicate the details of the dataset/code/model as part of their submissions via structured templates. This includes details about training, license, limitations, etc. +- The paper should discuss whether and how consent was obtained from people whose asset is used. +- At submission time, remember to anonymize your assets (if applicable). You can either create an anonymized URL or include an anonymized zip file. + +# 14. Crowdsourcing and research with human subjects + +Question: For crowdsourcing experiments and research with human subjects, does the paper include the full text of instructions given to participants and screenshots, if applicable, as well as details about compensation (if any)? + +Answer: [NA] + +Justification: No crowdsourcing or human subjects were involved in this research. + +Guidelines: + +- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. +- Including this information in the supplemental material is fine, but if the main contribution of the paper involves human subjects, then as much detail as possible should be included in the main paper. +- According to the NeurIPS Code of Ethics, workers involved in data collection, curation, or other labor should be paid at least the minimum wage in the country of the data collector. + +# 15. Institutional review board (IRB) approvals or equivalent for research with human subjects + +Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals (or an equivalent approval/review based on the requirements of your country or institution) were obtained? + +Answer: [NA] + +Justification: No crowdsourcing or human subjects were involved in this research. + +Guidelines: + +- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. +- Depending on the country in which research is conducted, IRB approval (or equivalent) may be required for any human subjects research. If you obtained IRB approval, you should clearly state this in the paper. + +We recognize that the procedures for this may vary significantly between institutions and locations, and we expect authors to adhere to the NeurIPS Code of Ethics and the guidelines for their institution. +- For initial submissions, do not include any information that would break anonymity (if applicable), such as the institution conducting the review. + +# 16. Declaration of LLM usage + +Question: Does the paper describe the usage of LLMs if it is an important, original, or non-standard component of the core methods in this research? Note that if the LLM is used only for writing, editing, or formatting purposes and does not impact the core methodology, scientific rigorousness, or originality of the research, declaration is not required. + +Answer: [Yes] + +Justification: The usage of LLMs is fully described for our approaches, with LLMs being a component of our transition model (forming the base architecture to fine-tune in 2) and our 3D-Prover model (being used as tactic generators in 1). + +Guidelines: + +- The answer NA means that the core method development in this research does not involve LLMs as any important, original, or non-standard components. +- Please refer to our LLM policy (https://neurips.cc/Conferences/2025/LLM) for what should or should not be described. \ No newline at end of file diff --git a/NeurIPS/2025/3D-Prover_ Diversity Driven Theorem Proving With Determinantal Point Processes/images.zip b/NeurIPS/2025/3D-Prover_ Diversity Driven Theorem Proving With Determinantal Point Processes/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..da9f9b17ced75c444803699dae04b22a5a733d49 --- /dev/null +++ b/NeurIPS/2025/3D-Prover_ Diversity Driven Theorem Proving With Determinantal Point Processes/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3f9455d1a4d8890dea8c30d9c64194b7106110cf4c803eeacb2c547f418118aa +size 514614 diff --git a/NeurIPS/2025/3D-Prover_ Diversity Driven Theorem Proving With Determinantal Point Processes/layout.json b/NeurIPS/2025/3D-Prover_ Diversity Driven Theorem Proving With Determinantal Point Processes/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..fa6a2a3ecc6d23a365457fddec1f8e48aaaa650b --- /dev/null +++ b/NeurIPS/2025/3D-Prover_ Diversity Driven Theorem Proving With Determinantal Point Processes/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f51dae52fde417f7415516400a2849d12be6f0fee367178722493fe100493734 +size 785577 diff --git a/NeurIPS/2025/3DID_ Direct 3D Inverse Design for Aerodynamics with Physics-Aware Optimization/ad5d49a0-afaa-499e-850d-e17453f22cb7_content_list.json b/NeurIPS/2025/3DID_ Direct 3D Inverse Design for Aerodynamics with Physics-Aware Optimization/ad5d49a0-afaa-499e-850d-e17453f22cb7_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..f04fd8b9017087fd6ddb6335f7ecfd4b674a1e6e --- /dev/null +++ b/NeurIPS/2025/3DID_ Direct 3D Inverse Design for Aerodynamics with Physics-Aware Optimization/ad5d49a0-afaa-499e-850d-e17453f22cb7_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6c2c6cafdcec4b0f31d30bd9db1cac219335dfbff3bec6bce8a878de973982a6 +size 165450 diff --git a/NeurIPS/2025/3DID_ Direct 3D Inverse Design for Aerodynamics with Physics-Aware Optimization/ad5d49a0-afaa-499e-850d-e17453f22cb7_model.json b/NeurIPS/2025/3DID_ Direct 3D Inverse Design for Aerodynamics with Physics-Aware Optimization/ad5d49a0-afaa-499e-850d-e17453f22cb7_model.json new file mode 100644 index 0000000000000000000000000000000000000000..f4b0072a478cd5422f3635b44017bff03c297677 --- /dev/null +++ b/NeurIPS/2025/3DID_ Direct 3D Inverse Design for Aerodynamics with Physics-Aware Optimization/ad5d49a0-afaa-499e-850d-e17453f22cb7_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fa04e6ad7f34d57da2d3bae4066e6a4a716e4f3b94ddff7c117237ed1d64dbd9 +size 216202 diff --git a/NeurIPS/2025/3DID_ Direct 3D Inverse Design for Aerodynamics with Physics-Aware Optimization/ad5d49a0-afaa-499e-850d-e17453f22cb7_origin.pdf b/NeurIPS/2025/3DID_ Direct 3D Inverse Design for Aerodynamics with Physics-Aware Optimization/ad5d49a0-afaa-499e-850d-e17453f22cb7_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..10d3dcfaef76a8f9f0f82b7cc710486ed21cca95 --- /dev/null +++ b/NeurIPS/2025/3DID_ Direct 3D Inverse Design for Aerodynamics with Physics-Aware Optimization/ad5d49a0-afaa-499e-850d-e17453f22cb7_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:46af5832dceb0619a58d811eaa9a021f4ffd20dc1df00c5393dbbdae1e99a4a8 +size 21154066 diff --git a/NeurIPS/2025/3DID_ Direct 3D Inverse Design for Aerodynamics with Physics-Aware Optimization/full.md b/NeurIPS/2025/3DID_ Direct 3D Inverse Design for Aerodynamics with Physics-Aware Optimization/full.md new file mode 100644 index 0000000000000000000000000000000000000000..2b81a846795993c73ba85d14f217c51b01a25a6c --- /dev/null +++ b/NeurIPS/2025/3DID_ Direct 3D Inverse Design for Aerodynamics with Physics-Aware Optimization/full.md @@ -0,0 +1,748 @@ +# 3DID: Direct 3D Inverse Design for Aerodynamics with Physics-Aware Optimization + +Yuze Hao $^{1}$ , Linchao Zhu $^{1,2}$ , Yi Yang $^{1,2*}$ + +1 College of Computer Science and Technology, Zhejiang University +2 The State Key Lab of Brain-Machine Intelligence, Zhejiang University + +# Abstract + +Inverse design aims to design the input variables of a physical system to optimize a specified objective function, typically formulated as a search or optimization problem. However, in 3D domains, the design space grows exponentially, rendering exhaustive grid-based searches infeasible. Recent advances in deep learning have accelerated inverse design by providing powerful generative priors and differentiable surrogate models. Nevertheless, current methods tend to approximate the 3D design space using 2D projections or fine-tune existing 3D shapes. These approaches sacrifice volumetric detail and constrain design exploration, preventing true 3D design from scratch. In this paper, we propose a 3D Inverse Design (3DID) framework that directly navigates the 3D design space by coupling a continuous latent representation with a physics-aware optimization strategy. We first learn a unified physics-geometry embedding that compactly captures shape and physical field data in a continuous latent space. Then, we introduce a two-stage strategy to perform physics-aware optimization. In the first stage, a gradient-guided diffusion sampler explores the global latent manifold. In the second stage, an objective-driven, topology-preserving refinement further sculpts each candidate toward the target objective. This enables 3DID to generate high-fidelity 3D geometries, outperforming existing methods in both solution quality and design versatility. + +# 1 Introduction + +Inverse design seeks to identify the initial variables of a physical system that, under given constraints, optimizes a specified objective function. This fundamental challenge occurs across many scientific and engineering disciplines, such as materials science, mechanical engineering, aerospace design, and supports applications ranging from automotive body shaping [1] and nano-photonic device engineering [2] to mechanical materials design [3, 4] and physics detector development [5]. + +Despite its broad impact, efficiently exploring the design space toward a target objective presents significant challenges. First, inverse design must contend with the inherent complexity of simulating physical systems for evaluation. These simulations are often nonlinear and high-dimensional, requiring fine discretizations that dominate computational resources [6, 7]. Second, the design landscape is extremely large, inherently nonconvex, and riddled with local minima, making exhaustive global search infeasible [8, 9]. In 3D domains, where inverse design usually involves direct geometry optimization, the number of degrees of freedom grows exponentially [10]. This rapid growth in geometric complexity drives up simulation expense and intensifies the search challenge. + +To tackle inverse design in the 3D domain, various techniques have been proposed [1, 11, 12], yet they fall short of addressing the above challenges. Traditional approaches such as adjoint-based gradient methods [13, 14, 15] and Bayesian optimization [16, 17] provide broad applicability but depend on + +*Corresponding author. + +repeated high-fidelity simulations that incur prohibitive computational cost. With recent advances in deep learning, pretrained surrogate models [18, 19, 20, 21] can efficiently approximate the forward physical process and support end-to-end backpropagation to update design variables, speeding up convergence by orders of magnitude. However, many prior methods adopt two simplifications. One replaces the 3D design with 2D proxies [1, 11] (multi-view renderings or silhouettes), which removes geometry information. The other requires an initial geometry as the starting point for subsequent refinement [22, 23, 24]. In practice, both assumptions restrict design exploration and hinder a thorough search of complex 3D design spaces (see Fig. 1). As a result, they cannot support the full exploration of complex three-dimensional design spaces, limiting coverage to a narrow subset of feasible geometries. + +We identify two primary challenges in 3D inverse design. 1) The high dimensionality of 3D physics-geometry-coupled spaces impedes exploration. Inverse design must simultaneously optimize geometric structures while accurately evaluating their resulting physical properties. This coupling, combined with the continuous high-resolution nature of both shape and physical fields, makes the + +direct 3D exploration extraordinarily difficult. 2) The lack of optimization strategies that balance the exploration-validity trade-off. Refining a baseline geometry with a surrogate model ensures constraint compliance and design validity, but it confines the search to a local neighborhood and can introduce adversarial artifacts when driven too far [25, 26]. On the other hand, sampling candidates with a generative model offers broader exploration yet leaves results vulnerable to biases in the training data [27]. Consequently, samples stay tethered to the prior and tend to imitate prevalent patterns rather than pushing toward novel optima. + +To address these challenges, we introduce 3DID, a 3D inverse-design framework that explores the design space without relying on simplified parameterizations or predefined shapes. Rather than directly searching the prohibitively large, continuous physics-geometry-coupled space, we first learn a continuous physics-geometry unified latent representation. This compact embedding preserves fine-grained shape and physical field variations while dramatically reducing both dimensionality and computational cost, thereby overcoming the dual obstacles of large-scale shape optimization and physics-aware simulation. Building on this latent space, we then deploy a two-stage optimization pipeline to tackle the exploration-validity trade-off. It begins with a gradient-guided diffusion sampler that traverses the manifold from pure noise to generate diverse, physics-informed candidates by steering sampling toward high-performance regions using objective gradients. Each candidate then undergoes topology-preserving optimization, which further improves objective performance under strict mesh-quality and connectivity constraints, ensuring geometric integrity and preventing adversarial artifacts. Together, these components enable 3DID to discover novel, high-fidelity 3D designs that reliably meet target objectives. In summary, our contributions are threefold: + +![](images/216e9a03e743bda2b5aef95b422e262c6d9fd75e5095339462feabb17b227051.jpg) +Figure 1: Motivation of 3DID. Existing 3D inverse-design methods either rely on reduced-dimensional representations (2D projections or fixed parameterizations) that constrain design freedom, or require an initial geometry as a starting point for local refinement, which highly constrains the search space. In contrast, 3DID overcomes these limitations by directly exploring the full 3D design space from random initialization. + +(1) We propose a continuous latent embedding that jointly encodes detailed 3D geometry and high-fidelity physical fields, enabling an efficient, unified search within a compact latent manifold. +(2) We develop a two-stage optimization pipeline that begins with gradient-guided diffusion sampling for global exploration and followed by topology-preserving refinement, optimizing each candidate toward the desired objective while strictly maintaining structural integrity. + +(3) We validate our 3DID framework on aerodynamic shape optimization, demonstrating that it consistently generates novel geometries whose superior performance is confirmed through surrogate evaluations and high-fidelity CFD simulations, significantly surpassing baseline methods. + +# 2 Related work + +# 2.1 Inverse Design + +Compared with the forward PDE problem, which predicts the physical response of a given design using numerical solvers or learned surrogates [28, 29, 30, 31, 32, 33], inverse design seeks the design variables that achieve a target objective under engineering constraints [26, 18, 34]. Inverse design is a fundamental problem in many domains of science and engineering disciplines, including mechanical engineering [35, 36], material science [37, 38, 39], chemical engineering [40], medical engineering [41], and aerospace engineering [22, 6, 42]. Classical approaches typically combine high fidelity physics solvers with sampling-based optimization methods such as the Cross Entropy Method [43] or Gaussian-process model with Bayesian optimization [44] to explore the design space. With the advent of differentiable simulators, inverse design can be posed directly as a gradient-based optimization problem [41, 45]. More recently, deep learning driven methods have shown great promise by learning surrogate models that approximate forward physics and allow end-to-end backpropagation [18, 19]. Furthermore, generative models, including variational autoencoders [46], GAN [47, 48] and diffusion models [26, 49] have been applied to inverse design. While prior methods mainly excel in 2D or low-dimensional settings, we propose a framework that directly navigates the full 3D design space via physics-aware optimization. + +# 2.2 Aerodynamic Shape Optimization + +Aerodynamic shape optimization is a classical inverse design task that seeks geometries minimizing drag while satisfying constraints on lift, stability, and other performance criteria [50, 51, 52, 53]. Generally, effective optimization critically depends on two key components: shape representation and optimization strategy. Traditional approaches typically employ simplified, low-dimensional representations such as 2D projections [11, 54, 1] or spline-based parameterizations [55, 56, 57] to reduce dimensionality and computational costs. Optimization is then performed using gradient-based adjoint solvers for efficient local refinement [24, 58, 59]. Additionally, to accelerate convergence, many methods optimize from a pre-selected baseline geometry [22, 23, 24]. In contrast, we propose a guided diffusion model over a latent shape representation, enabling the design of unconstrained 3D geometries directly from noise, without relying on initial shapes or 2D profiles. + +# 3 Methods + +In this section, we first formalize the 3D inverse design problem (Section 3.1). We then introduce our physics-geometry unified representation (Section 3.2), describe the gradient-guided diffusion sampling process (Section 3.3), and detail the topology-preserving refinement stage (Section 3.4). Finally, we provide our implementation details (Section 3.5). + +# 3.1 Problem Formulation + +We consider the problem of 3D inverse design, where the goal is to identify a solid input geometry $M$ for a physical system that optimizes specified performance objectives while satisfying geometric constraints. Formally, let $M \subset \mathbb{R}^3$ denote a solid geometry, and let $\mathcal{F}(M)$ be the corresponding steady-state physical field (e.g., pressure or temperature distribution) governed by a partial differential equation (PDE) or an ordinary differential equation (ODE). We define the design objective as: + +$$ +\mathcal {J} (M) := \mathcal {J} (M, \mathcal {F} (M)), \tag {1} +$$ + +which may measure quantities such as drag, lift, and structural compliance. Specifically, $\mathcal{I}$ depends on $M$ in two ways: implicitly, via the resulting physical field $\mathcal{F}(M)$ on which $\mathcal{I}$ is evaluated, and explicitly, via direct geometric properties defined on $M$ . The classical inverse design aims to solve: + +$$ +M ^ {*} = \arg \min _ {M} \mathcal {J} (M, \mathcal {F} (M)) \tag {2} +$$ + +$$ +\begin{array}{l} \text {s . t .} C (M, \mathcal {F} (M)) \leq 0, \end{array} +$$ + +![](images/e7b259576f4d63204b8242e516d6116f39c5c98c3ebf8b82d4e32286c3df28a9.jpg) +Figure 2: The overview of PG-VAE. We use transformers to encode the design geometry and its associated physical field, along with learnable tokens, into a compact triplane latent representation $z$ . A decoder then upsamples the latent $z$ into high-resolution triplane feature maps, which can be reshaped into three orthogonal planes. Finally, a physics–geometry mapping network is applied to reconstruct both the occupancy field and the corresponding physical field from these feature maps. + +where the solution $M^{*}$ is the geometry that minimizes the performance objective, $C$ aggregates design constraints, such as volume, manufacturability, and boundary conditions. + +# 3.2 Physics-Geometry Unified Representation + +To compactly encode 3D geometry and its physical field, we adopt a triplane representation learned via our Physics-Geometry VAE (PG-VAE). As shown in Figure 2, it includes: (1) A physics-geometry encoder that maps input geometry and physical field into a latent code. (2) A latent-to-triplane decoder that reconstructs triplane feature maps from the latent code. (3) A physics-geometry mapping network that recovers the 3D geometry and physical field from the triplane. + +Physics-Geometry Encoder. The physics-geometry encoder comprises two parallel branches: one for processing raw 3D geometry, and the other for encoding the associated physical field. Inspired by [60], each branch uses learnable tokens to capture both local and global context. For the geometry branch, given uniformly sampled point clouds $P_{geo} \in \mathbb{R}^{N_g \times C_g}$ from the geometry, where $N_g$ is the number of points and $C_g$ represents features including normalized positions and surface normals, we utilize Fourier features [61] to encode positional structure. Then, a set of learnable tokens $e_g \in \mathbb{R}^{(3 \times r \times r) \times d_e}$ queries information from these points via cross-attention and self-attention layers, resulting in geometry latent tokens $z_g \in \mathbb{R}^{(3 \times r \times r) \times d_z}$ . The physical-field branch follows the same structure, processing uniformly sampled physical-field points $P_{phy} \in \mathbb{R}^{N_p \times C_p}$ , where $N_p$ is the number of points and $C_p$ is the dimension of physical-field features. Learnable tokens $e_p \in \mathbb{R}^{(3 \times r \times r) \times d_e}$ undergo similar attention layers to produce physical-field latents $z_p \in \mathbb{R}^{(3 \times r \times r) \times d_z}$ . Finally, geometry and physical tokens are concatenated and passed through MLP layers to form the unified latent representation $z = \mathrm{MLP}(\mathrm{Concat}(z_g, z_p))$ , where $z \in \mathbb{R}^{(3 \times r \times r) \times d_z}$ . + +Latent-to-Triplane Decoder. After obtaining the unified physics-geometry latent representation, we apply a decoder to formulate the triplane representation. Prior to decoding, we reshape the latent tokens by vertically concatenating three planes, yielding the reshaped latent tensor $z \in \mathbb{R}^{r \times (3 \times r) \times d_z}$ , following [62]. Subsequently, the latent tensor is passed through a series of convolutional layers for upsampling. The output is then reshaped into the final triplane features $T_{xy}, T_{xz}, T_{yz} \in \mathbb{R}^{R \times R \times d_t}$ , where $R$ denotes the resolution of each plane and $d_t$ is the feature dimension per pixel. + +Physics-Geometry Mapping Network. The mapping network serves to reconstruct 3D geometry and the associated physical field from the learned triplane representation. We utilize two parallel MLP branches: one for predicting geometric occupancy and one for estimating the physical field. Given a query point $q \in \mathbb{R}^3$ , we project it onto the three orthogonal planes and extract features. The aggregated feature is computed as $t_q = T_{xy}(q_{xy}) + T_{xz}(q_{xz}) + T_{yz}(q_{yz})$ . where $q_{xy}, q_{xz}, q_{yz}$ denote the 2D projections of $q$ onto the respective planes. The aggregated feature $t_q$ is then fed into the MLP branches to predict the occupancy field and physical field values at the corresponding point $q$ . + +![](images/61f0b13d47f4c3e3af384f2c040868d4c66fd73ea677c79cdf8292dad0a0b1ae.jpg) +Stage 1: Initial Design Generation + +![](images/53ac8b81ba0341212d919917b253b0c8b55df1ab97416af241ebafff8d9762eb.jpg) +Stage 2: Topology Preserving Refinement +Figure 3: The optimization framework of 3DID. Starting from noise, we guide diffusion using objective gradients to steer the latent toward high-performance regions. The decoded triplane then yields an initial mesh $M_0$ and its surface physical field $\varphi$ , which is then refined via backpropagation over a free-form deformation lattice to improve performance while preserving topology. + +End-to-End Training. Our VAE model is trained end-to-end to jointly reconstruct 3D geometry and the associated physical field. For occupancy field reconstruction, we employ the Binary Cross-Entropy (BCE) loss $\mathcal{L}_{\mathrm{BCE}}$ to supervise the predicted occupancy. To reconstruct the physical field, we utilize the Mean Squared Error (MSE) loss $\mathcal{L}_{\mathrm{MSE}}$ . Additionally, we incorporate a KL divergence loss $\mathcal{L}_{\mathrm{KL}}$ to regularize the latent space. Overall, our training loss can be formulated as: + +$$ +\mathcal {L} _ {\mathrm {P G - V A E}} = \lambda_ {\mathrm {B C E}} \mathcal {L} _ {\mathrm {B C E}} + \lambda_ {\mathrm {M S E}} \mathcal {L} _ {\mathrm {M S E}} + \lambda_ {\mathrm {K L}} \mathcal {L} _ {\mathrm {K L}}, \tag {3} +$$ + +where $\lambda_{\mathrm{BCE}},\lambda_{\mathrm{MSE}},\lambda_{\mathrm{KL}}$ are weighting coefficients. + +# 3.3 Objective Guided Diffusion + +Once the PG-VAE is trained, it provides a compact, expressive latent code $z$ that jointly captures 3D geometry and its physical field. We then train a diffusion model [63, 64] on these latents, enabling direct generation of samples on the learned manifold from pure noise. To drive inverse design, we inject gradients of the task objective $\mathcal{I}$ into the diffusion sampling, as shown in Figure 3. + +In standard diffusion sampling, each denoising step predicts noise via the learned score function: + +$$ +\epsilon_ {\phi} \left(z _ {t}, t\right) = - \sqrt {1 - \alpha_ {t}} \nabla_ {z _ {t}} \log p \left(z _ {t}\right), \tag {4} +$$ + +where $\nabla_{z_t}\log p(z_t)$ denotes the score function, i.e., the gradient of the log-probability density of the latent variable $z_{t}$ . By iteratively denoising, the model guides samples toward high-probability regions of the data manifold. In our case, we need to consider not only guiding the noise towards the feasible data manifold, but we also need to incorporate optimization of $\mathcal{I}$ during the sampling. Therefore, inspired by [65, 66], we replace the unconditional score with the conditional score $\nabla_{z_t}\log p(z_t\mid \mathcal{I})$ . By Bayes' rule, we can derive: + +$$ +\nabla_ {z _ {t}} \log p \left(z _ {t} \mid \mathcal {J}\right) \propto \nabla_ {z _ {t}} \log p \left(z _ {t}\right) + \nabla_ {z _ {t}} \log p (\mathcal {J} \mid z _ {t}). \tag {5} +$$ + +Here, $\nabla_{z_t}\log p(z_t)$ corresponds to the standard score function learned by the diffusion model, while $\nabla_{z_t}\log p(\mathcal{J} \mid z_t)$ acts as an additional guidance term that incorporates the influence of the design objective. Since $\nabla_{z_t}\log p(\mathcal{J} \mid z_t)$ is unknown, we approximate it by: + +$$ +\nabla_ {z _ {t}} \log p (\mathcal {J} \mid z _ {t}) \simeq \nabla_ {z _ {t}} \log p \left(\mathcal {J} \mid \hat {z} _ {0} (z _ {t})\right) \propto - \nabla_ {z _ {t}} \mathcal {J} \left(\hat {z} _ {0} (z _ {t})\right), \tag {6} +$$ + +where $\hat{z}_0(z_t)$ denotes the estimate of the clean latent code given the noisy latent $z_{t}$ , following [66]. Accordingly, we adjust the predicted noise to incorporate the influence of the design objective $\mathcal{I}$ , resulting in the guided noise prediction: + +$$ +\epsilon_ {\phi} ^ {\prime} \left(z _ {t}, t\right) = \epsilon_ {\phi} \left(z _ {t}, t\right) + \gamma \nabla_ {z _ {t}} \mathcal {J}, \tag {7} +$$ + +where $\epsilon_{\phi}^{\prime}$ is the modified noise prediction, and $\gamma$ is a scaling coefficient that controls the strength of the guidance. This objective-aware adjustment steers the sampling trajectory toward latent regions that both conform to the learned data distribution and advance the target design objective $\mathcal{I}$ . + +# 3.4 Topology-Preserving Refinement + +After objective-guided diffusion, we obtain an optimized latent code $z^{*}$ , which is reshaped into triplane feature maps and decoded by the physics-geometry mapping network to generate an initial 3D mesh $M_0$ with vertex set $V_{0} = \{v_{j}\}_{j = 1}^{N}$ and its associated physical field $\varphi = \{\varphi_{j}\}_{j = 1}^{N}$ , as shown in Figure 3. Although guided by the design objective, the generated designs remain highly biased by the prior distribution of designs from the training data [27]. We introduce a topology-preserving refinement stage based on free-form deformation (FFD) [67, 68], controlled by gradient descent. + +Specifically, we first wrap $M_0$ in a 3D lattice of control points $C = \{c_i\}_{i=1}^K$ . These control points form a flexible control grid that allows smooth and structured adjustment of the mesh shape while preserving its topology. The deformation of each vertex $v_j$ is computed as: + +$$ +v _ {j} ^ {\prime} (\boldsymbol {C}) = \sum_ {i = 1} ^ {K} \mathcal {B} _ {i} \left(v _ {j}\right) c _ {i}, \tag {8} +$$ + +where $\mathcal{B}_i(v_j)$ denotes the $i$ -th trivariate Bernstein basis function [68] evaluated at the normalized parametric coordinate of vertex $v_j$ . These basis functions provide smooth, localized influence from the control points, enabling flexible yet coherent deformation across the mesh. + +At the beginning of the refinement process, the control points are unmodified, so $v_{j}^{\prime}(C) = v_{j}$ . The initial vertex-field pairs $\{(v_{j}^{\prime}(C),\varphi_{j})\}$ are then fed into a pretrained GNN surrogate $f_{\mathrm{GNN}}$ which estimates the current design objective based on the mesh geometry and physical attributes: + +$$ +\hat {\mathcal {J}} (\boldsymbol {C}) = f _ {\text {G N N}} \left(\left(v _ {j} ^ {\prime} (\boldsymbol {C}), \varphi_ {j}\right) _ {j = 1} ^ {N}\right). \tag {9} +$$ + +With this differentiable surrogate model, we optimize the control points to improve the design objective. The overall refinement loss is defined as: + +$$ +\mathcal {L} (\boldsymbol {C}) = \hat {\mathcal {J}} (\boldsymbol {C}) + \lambda_ {\text {s m o o t h}} \sum_ {i = 1} ^ {K} \| \Delta c _ {i} \| ^ {2} + \lambda_ {\text {v o l}} \sum_ {\text {c e l l s}} \left(\frac {V _ {\text {d e f}}}{V _ {\text {o r i g}}} - 1\right) ^ {2}, \tag {10} +$$ + +where $\Delta c_{i}$ are control-point displacements, and the term weighted by $\lambda_{\mathrm{smooth}}$ penalizes large displacements for smooth deformations, while the term weighted by $\lambda_{\mathrm{vol}}$ penalizes cell-wise volume changes to ensure geometric consistency. Control points are updated via: + +$$ +\boldsymbol {C} ^ {(t + 1)} = \boldsymbol {C} ^ {(t)} - \eta \nabla_ {\boldsymbol {C}} \mathcal {L} (\boldsymbol {C} ^ {(t)}), \tag {11} +$$ + +We optimize using AdamW with a cosine-annealed learning rate $\eta$ . Iteration continues until an iteration count $T_{o}$ is reached. The resulting vertices $\mathbf{V}^{*} = \{v_{j}^{*}\}$ , obtained via the FFD mapping (Eq. 8), define the refined mesh $M^{*}$ , which preserves topology and improves target performance. + +# 3.5 Implementation Details + +To train our PG-VAE, we sample $N_{g} = N_{p} = 50,000$ points as input and use the physics–geometry encoder with one cross-attention layer and 8 self-attention layers with 12 heads and $d_{z} = 64$ , plus $r = 64$ learnable tokens of dimension $d_{e} = 768$ , yielding a latent code of $d_{z} = 32$ . The decoder upsamples via one self-attention layer and five ResNet blocks [69] to a triplane with $R = 256$ and channel $d_{t} = 64$ . Each branch's mapping network has five linear layers with a hidden dimension of 32. We train the VAE model with loss weights $\lambda_{\mathrm{BCE}} = 10^{-3}$ , $\lambda_{\mathrm{MSE}} = 10^{-5}$ , $\lambda_{\mathrm{KL}} = 10^{-6}$ . During training, we sample 50,000 points from the unit domain to supervise both occupancy and physical-field predictions. For occupancy, we adopt the semi-continuous formulation following [60]. We use a learning rate of $1e - 4$ , a batch size of 8 per GPU, and train for 100K steps. For the diffusion model, we employ 10 layers of DiT blocks [70], each with 16 attention heads of dimension 72. We train the diffusion model with 1000 denoising steps. For objective-guided sampling, an auxiliary U-Net surrogate predicts the task objective directly in latent code $z$ . To train the diffusion model, we use a learning rate of $5e - 5$ , a batch size of 4 per GPU, and train for 300K. In topology-preserving refinement, we deform candidates via a $20\times 6\times 6$ control-point grid along the x, y, and z axes. For the surrogate model $f_{\mathrm{GNN}}$ , we adopt MeshGraphNet [30] as our surrogate given its strong performance in mesh-based physical simulations. The surrogate is trained to predict aerodynamic drag from paired samples of geometry and ground-truth physical fields collected from the dataset. The model is trained with a learning rate of $1e - 5$ , a batch size of 8 per GPU, and trained for 100K. All models are trained with AdamW optimizer [71]. More training details of 3DID are included in the Appendix. + +Table 1: Quantitative comparison for aerodynamic vehicle design. The confidence interval information is detailed in the Appendix. Note that our method shows a slight drop in coverage, mainly because the topology-preserving refinement pushes designs beyond the original distribution to achieve better aerodynamic performance. + +
MethodPred-Drag↓Sim-Drag↓Novelty↑Coverage↑
GP, Voxel0.29970.42541.03990.5200
GP, Voxel+PCA0.30590.43630.97340.5850
CEM, Voxel0.29510.40970.97920.4350
CEM, Voxel+PCA0.30880.43930.98640.5100
CEM, TripNet0.31540.41611.03990.6050
Backprop, Voxel0.29790.41460.98600.4750
Backprop, Voxel+PCA0.30610.46140.97980.4950
Backprop, TripNet0.31530.41701.02940.5900
3DID-NoTopoRefine (ours)0.26230.37660.91950.6950
3DID (ours)0.26070.35361.17090.4300
+ +# 4 Experiments + +In the experiments, we aim to answer the following questions: (1) Does 3DID outperform traditional, sampling-based, and backpropagation methods in finding high-quality designs? (2) Does our unified physics-geometry triplane representation yield better objectives than alternative latent or purely geometric embeddings? (3) Does our two-stage pipeline outperform single-stage diffusion sampling and other standard optimization methods? To answer these questions, we evaluate our method on the vehicle aerodynamic shape optimization task, a representative example of 3D inverse design. + +In the following sections, we first introduce our dataset and evaluation metrics (Section 4.1). Next, we describe our experimental setup and compare against baseline methods (Section 4.2). Finally, we present two ablation studies: one on the unified physics-geometry representation (Section 4.3) and one on the two-stage optimization strategy (Section 4.4). + +# 4.1 Dataset and Evaluation Details + +Dataset. We conduct our experiments on the $\mathrm{DrivAerNet++}$ dataset [72, 73], the largest available collection for aerodynamic car design, comprising over 8,000 diverse geometries paired with high-fidelity CFD simulations. For training, we use the entire dataset. We first normalize each geometry to fit within a unit cube and apply the same scaling to the simulation fields. We then uniformly sample surface point clouds with corresponding normals from the geometry. Finally, for both the occupancy field and the physical field, we adopt the data extraction strategy of Park et al. [74], using a grid resolution of 256. Further details of our data preparation are given in the Appendix. + +Evaluation Metrics. We evaluate 3DID using four metrics. Pred-Drag is the drag coefficient estimated by a trained surrogate model, offering an approximation of the design objective. Sim-Drag is the drag coefficient obtained via high-fidelity CFD simulation, delivering an unbiased evaluation of aerodynamic performance. Novelty computes the average nearest neighbor distance from each generated design to its closest training example, indicating how distinct the designs are from existing ones. Coverage captures how well the generated designs cover the training distribution by measuring, for each training sample, the distance to its nearest generated design (using a k-nearest neighbor lookup) and reporting the fraction of training examples that fall within a predefined threshold. To extract features from the generated geometries, we use the pretrained PointNet model from [75]. Detailed evaluation procedures and simulation parameters are provided in the Appendix. + +# 4.2 3D Vehicle Aerodynamic Design + +In this experiment, we evaluate each inverse-design method on the aerodynamic shape optimization task, where the objective $\mathcal{I}$ is to reduce the drag force of the designed vehicles. All methods are trained on the same collection of car geometries paired with high-fidelity CFD simulations. As baselines, we compare against the Cross-Entropy Method(CEM) [43], the Gaussian-process surrogate with + +![](images/20d09dcabc13bbe3689b3a74ffd91341bebd8097137eee6ba7670dae908a249d.jpg) +Figure 4: Qualitative comparisons of different representations. Each row shows four candidates with geometry (left) and simulated velocity field (right) with Sim-Drag in the top-right. Despite equal resolution, voxel methods incur higher drag and often yield non-watertight shapes (red box) due to coarse discretization. Our continuous latent representation produces watertight, smooth designs with superior aerodynamic performance, outperforming both voxel-based and geometry-only approaches. + +Table 2: Ablation study on representation choices. + +
MethodPred-Drag↓Sim-Drag↓Novelty↑Coverage↑
Voxel0.27220.43181.06830.3450
Voxel+PCA0.27200.45650.98580.5750
TripNet0.26980.40661.05800.5500
3DID-NoTopoRefine (ours)0.26230.37660.91950.6950
3DID (ours)0.26070.35361.17090.4300
+ +Bayesian optimization(GP) [44], and the gradient-based backpropagation method(Backprop) [18]. To evaluate the impact of 3D encoding, the optimizer is instantiated with three representations: a dense voxel grid [76], a PCA-compressed voxel grid (Voxel+PCA) [22], and a pure geometry triplane network (TripNet) [21]. GP with the triplane representation is omitted due to its high computational cost. For fairness, we generate 64 candidate designs per method and report the average performance in Table 1. Architectures of baselines and training details are provided in the Appendix. + +From Table 1, it can be observed that 3DID delivers the best drag force result for both pred-drag and sim-drag compared to all baselines. Specifically, our full 3DID model reduces simulated drag by $13.6\%$ relative to the strongest baseline. These results demonstrate the effectiveness of our pipeline in discovering high-performance designs. Furthermore, our method achieves the highest novelty score (1.1709), indicating its ability to explore diverse design variations. Note that the drop in coverage occurs because topology-preserving refinement pushes designs beyond the training distribution to boost aerodynamic performance. A detailed ablation on the cascade optimization strategy is presented in Section 4.4. More visualization results and evaluation are presented in the Appendix. + +# 4.3 Ablation Study on Physics-Geometry Unified Representation + +In this experiment, we evaluate the performance with different representations, including Voxel [76], Voxel with PCA [22] and the pure geometry triplane (TripNet) [21]. For Voxel and Voxel with PCA, we first train a variational autoencoder (VAE) to embed the raw data into a compact latent space and then learn a diffusion model within that space. Finally, we employ our two-stage optimization pipeline consisting of gradient-guided diffusion sampling followed by topology-preserving refinement to generate diverse design candidates. Because the baseline representations do not include physical field information, we retrain a surrogate graph neural network that takes only geometry as input for the refinement stage. The results are reported in Table 2. + +As shown in Table 2, our 3DID method outperforms all baselines by a wide margin in both simulation drag and novelty. Compared to the best baseline TripNet, 3DID reduces Sim Drag from 0.4066 to 0.3536, a $13.0\%$ improvement, and lowers Pred Drag by $3.4\%$ (from 0.2698 to 0.2607). It also increases novelty from 1.0683 to 1.1709, a $9.6\%$ gain. In Figure 5, our continuous latent representation consistently yields watertight smooth geometries with superior aerodynamic performance, whereas voxel-based methods suffer higher drag and non-watertight artifacts. Although TripNet embeds continuous geometry, its optimized designs remain inferior. We attribute this to the absence of physical field guidance, which weakens the optimization gradients in the refinement stage. + +![](images/41cdc5e5ffeb994f4fb7277d1a8ad49a2a9f92113e418cc1f276655ad94aa0af.jpg) +Figure 5: Qualitative comparisons of topology-preserving refinement. Each row presents two design candidates comparisons with their geometry and simulated velocity field heatmaps. Sim-Drag values are shown in the top-right corner of each panel. Refined designs exhibit a more significant fastback profile (yellow box), reduced low-velocity recirculation zones (blue box), and stronger downward flow (green box), indicating improved aerodynamic performance. + +![](images/d05ad30b00e1019ad8047e696e2615453a5f808fb9a5e889541f020ee0bf5136.jpg) + +Table 3: Ablation study on design strategies. + +
MethodPred-Drag↓Sim-Drag↓Novelty↑Coverage↑
CEM0.31520.39871.07300.6800
GD0.30230.40951.08780.5800
W/O Guidance0.29710.39440.91770.7104
3DID-NoTopoRefine (ours)0.26230.37660.91950.6950
3DID (ours)0.26070.35361.17090.4300
+ +# 4.4 Ablation Study on Two-Stage Optimization Strategy + +In this experiment, we validate the effectiveness of our two-stage optimization pipeline by comparing it against two alternative design strategies, including Cross-Entropy Method (CEM) and gradient descent (GD), as well as a diffusion-only sampling approach without objective gradient guidance. All methods are based on our physics-geometry unified representation. + +From Table 3, we see that our full 3DID outperforms all baselines in Pred-Drag, Sim-Drag, and Novelty. Notably, when designing without guidance, our diffusion model attains the highest coverage value of 0.7104, as it captures the data manifold comprehensively. When generating without guidance, the diffusion model tends to mimic the distribution of the dataset, which leads to an increase of coverage but lacks the targeted optimization for aerodynamic performance and novelty that our 3DID method provides. Furthermore, to better validate the effectiveness of our topology-preserving refinement stage, we provide side-by-side qualitative comparisons in Figure 5. It can be seen that after the refinement stage, each candidate develops a more significant fastback profile, with diminished low-velocity recirculation regions, and stronger downward flow patterns, all indicators of improved aerodynamic performance as confirmed by the reduced Sim-Drag values. + +# 5 Limitations + +Limited to static physical fields. Despite the fact that 3DID achieves impressive results, a significant limitation is its focus on static fields. The current framework does not support inverse design involving time-dependent or dynamic physical fields. Time-dependent physical systems often involve solid geometries coupled with evolving physical properties over time. This would pose challenges for representation and optimization within our framework. Enhancing 3DID with time-aware representations and models may address these limitations, which we leave as an important direction for future work. + +Limited to single objective optimization. In 3DID, we address the inverse problem with a single objective, which may limit its applicability for broader scenarios. Although it is straightforward to aggregate multiple objectives into a single composite loss, this approach may overlook potential conflicts and trade-offs between objectives. Extending 3DID to support true multi-objective optimization is a promising direction for future research. + +Limited to surrogate-based physics awareness. We incorporate physical fields via data-driven surrogates and joint geometry-physics embeddings during generation and refinement, rather than enforcing governing laws explicitly. This guides designs toward plausible, high-performing regions but does not enforce hard physical constraints. Exploring hard-constraint mechanisms or tighter PDE-consistent couplings is an important direction for future work. + +# 6 Conclusion + +In this paper, we tackle the problem of 3D inverse design, which faces challenges from the high-dimensional physics-geometry coupling and the exploration-validity trade-off. To represent the coupled space, we propose a physics-geometry unified representation that preserves fine-grained shape details and physical-field information while significantly reducing dimensionality. Based on this representation, we introduce a two-stage physics-aware optimization strategy that first explores the latent manifold via gradient-guided diffusion sampling and then refines candidates through topology-preserving refinement. Extensive experiments demonstrate that our 3DID framework generates high-fidelity 3D models with greater versatility and superior performance on target objectives. + +# Acknowledgments + +This work is partially supported by National Science and Technology Major Project (2022ZD0117802). This work was also supported in part by "Pioneer" and "Leading Goose" R&D Program of Zhejiang (No. 2025C02032), the Fundamental Research Funds for the Central Universities (226-2025-00080), and the Earth System Big Data Platform of the School of Earth Sciences, Zhejiang University. + +# References + +[1] Binyang Song, Chenyang Yuan, Frank Permenter, Nikos Arechiga, and Faez Ahmed. Surrogate modeling of car drag coefficient with depth and normal renderings. In International Design Engineering Technical Conferences and Computers and Information in Engineering Conference, volume 87301, page V03AT03A029. American Society of Mechanical Engineers, 2023. +[2] Sean Molesky, Zin Lin, Alexander Y Piggott, Weiliang Jin, Jelena Vuckovic, and Alejandro W Rodriguez. Inverse design in nanophotonics. Nature Photonics, 12(11):659-670, 2018. +[3] Jan-Hendrik Bastek and Dennis M Kochmann. Inverse design of nonlinear mechanical metamaterials via video denoising diffusion models. Nature Machine Intelligence, 5(12):1466-1475, 2023. +[4] Claudio Zeni, Robert Pinsler, Daniel Zügner, Andrew Fowler, Matthew Horton, Xiang Fu, Zilong Wang, Aliaksandra Shysheya, Jonathan Crabbé, Shoko Ueda, et al. A generative model for inorganic materials design. Nature, pages 1-3, 2025. +[5] Ann-Kathrin Schuetz, AWP Poon, and Aobo Li. Resum: A rare event surrogate model for physics detector design. In International Conference on Learning Representations (ICLR). +[6] W Kyle Anderson and V Venkatakrishnan. Aerodynamic design optimization on unstructured grids with a continuous adjoint formulation. Computers & Fluids, 28(4-5):443-480, 1999. +[7] Chae M Rhie and Wei-Liang Chow. Numerical study of the turbulent flow past an airfoil with trailing edge separation. AIAA journal, 21(11):1525-1532, 1983. +[8] Martin Philip Bendsoe and Ole Sigmund. Topology optimization: theory, methods, and applications. Springer Science & Business Media, 2013. + +[9] Antony Jameson. Aerodynamic design via control theory. Journal of scientific computing, 3:233-260, 1988. +[10] Max D Gunzburger. Perspectives in flow control and optimization. SIAM, 2002. +[11] Nicolas Rosset, Guillaume Cordonnier, Regis Duvigneau, and Adrien Bousseau. Interactive design of 2d car profiles with aerodynamic feedback. In Computer Graphics Forum, volume 42, pages 427-437. Wiley Online Library, 2023. +[12] Liane Makatura. Pareto gamuts: Exploring optimal designs across varying contexts. PhD thesis, Massachusetts Institute of Technology, 2020. +[13] Antony Jameson. Aerodynamic shape optimization using the adjoint method. *Lectures at the Von Karman Institute, Brussels*, 6, 2003. +[14] Gaetan KW Kenway, Charles A Mader, Ping He, and Joaquim RRA Martins. Effective adjoint approaches for computational fluid dynamics. Progress in Aerospace Sciences, 110:100542, 2019. +[15] Lana Osusky, Howard Buckley, Thomas Reist, and David W Zingg. Drag minimization based on the navier-stokes equations using a newton-krylov approach. AIAA Journal, 53(6):1555-1577, 2015. +[16] Timothy MS Jim, Ghifari A Faza, Pramudita S Palar, and Koji Shimoyama. Bayesian optimization of a low-boom supersonic wing planform. AIAA journal, 59(11):4514-4529, 2021. +[17] Grégoire Mariethoz, Philippe Renard, and Jef Caers. Bayesian inverse problem and optimization with iterative spatial resampling. Water Resources Research, 46(11), 2010. +[18] Kelsey Allen, Tatiana Lopez-Guevara, Kimberly L Stachenfeld, Alvaro Sanchez Gonzalez, Peter Battaglia, Jessica B Hamrick, and Tobias Pfaff. Inverse design for fluid-structure interactions using graph network simulators. Advances in Neural Information Processing Systems, 35:13759-13774, 2022. +[19] Tailin Wu, Willie Neiswanger, Hongtao Zheng, Stefano Ermon, and Jure Leskovec. Uncertainty quantification for forward and inverse problems of pdes via latent global evolution. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 320-328, 2024. +[20] Tailin Wu, Takashi Maruyama, and Jure Leskovec. Learning to accelerate partial differential equations via latent global evolution. Advances in Neural Information Processing Systems, 35:2240-2253, 2022. +[21] Qian Chen, Mohamed Elrefaie, Angela Dai, and Faez Ahmed. Tripnet: Learning large-scale high-fidelity 3d car aerodynamics with triplane networks. arXiv preprint arXiv:2503.17400, 2025. +[22] Jonathan Tran, Kai Fukami, Kenta Inada, Daisuke Umehara, Yoshimichi Ono, Kenta Ogawa, and Kunihiko Taira. Aerodynamics-guided machine learning for design optimization of electric vehicles. Communications Engineering, 3(1):174, 2024. +[23] Nikita Durasov, Artem Lukoyanov, Jonathan Donier, and Pascal Fua. Debosh: Deep bayesian shape optimization. arXiv preprint arXiv:2109.13337, 2021. +[24] Pierre Baque, Edoardo Remelli, Francois Fleuret, and Pascal Fua. Geodesic convolutional shape optimization. In International Conference on Machine Learning, pages 472-481. PMLR, 2018. +[25] Qingqing Zhao, David B Lindell, and Gordon Wetzstein. Learning to solve pde-constrained inverse problems with graph networks. arXiv preprint arXiv:2206.00711, 2022. +[26] Tailin Wu, Takashi Maruyama, Long Wei, Tao Zhang, Yilun Du, Gianluca Iaccarino, and Jure Leskovec. Compositional generative inverse design. In International Conference on Learning Representations (ICLR), 2024. + +[27] Jan Pawel Stanczuk, Georgios Batzolis, Teo Deveney, and Carola-Bibiane Schonlieb. Diffusion models encode the intrinsic dimension of data manifolds. In International Conference on Machine Learning (ICML), 2024. +[28] John David Anderson, John Wendt, et al. Computational fluid dynamics, volume 206. Springer, 1995. +[29] Maziar Raissi, Paris Perdikaris, and George E Karniadakis. Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations. Journal of Computational physics, 378:686-707, 2019. +[30] Tobias Pfaff, Meire Fortunato, Alvaro Sanchez-Gonzalez, and Peter Battaglia. Learning mesh-based simulation with graph networks. In International Conference on Learning Representations (ICLR), 2020. +[31] Zongyi Li, Nikola Kovachki, Kamyar Azizzadenesheli, Burigede Liu, Kaushik Bhattacharya, Andrew Stuart, and Anima Anandkumar. Fourier neural operator for parametric partial differential equations. arXiv preprint arXiv:2010.08895, 2020. +[32] Xihang Yue, Linchao Zhu, and Yi Yang. Deltaphi: Learning physical trajectory residual for pde solving. arXiv preprint arXiv:2406.09795, 2024. +[33] Xihang Yue, Yi Yang, and Linchao Zhu. Holistic physics solver: Learning pdes in a unified spectral-physical space. In *Forty-second International Conference on Machine Learning*, 2025. +[34] Navid Ansari, Hans-Peter Seidel, Nima Vahidi Ferdowsi, and Vahid Babaei. Autoinverse: Uncertainty aware inversion of neural networks. Advances in Neural Information Processing Systems, 35:8675-8686, 2022. +[35] Changyu Deng, Yizhou Wang, Can Qin, Yun Fu, and Wei Lu. Self-directed online machine learning for topology optimization. Nature communications, 13(1):388, 2022. +[36] Stelian Coros, Bernhard Thomaszewski, Gioacchino Noris, Shinjiro Sueda, Moira Forberg, Robert W Sumner, Wojciech Matusik, and Bernd Bickel. Computational design of mechanical characters. ACM Transactions on Graphics (TOG), 32(4):1-12, 2013. +[37] Zekun Ren, Siyu Isaac Parker Tian, Juhwan Noh, Felipe Oviedo, Guangzong Xing, Jiali Li, Qiaohao Liang, Ruiming Zhu, Armin G Aberle, Shijing Sun, et al. An invertible crystallographic representation for general inverse design of inorganic crystals with targeted properties. Matter, 5(1):314-335, 2022. +[38] Claudio Zeni, Robert Pinsler, Daniel Zügner, Andrew Fowler, Matthew Horton, Xiang Fu, Sasha Shysheya, Jonathan Crabbé, Lixin Sun, Jake Smith, et al. Mattergen: a generative model for inorganic materials design. arXiv preprint arXiv:2312.03687, 2023. +[39] Alex Zunger. Inverse design in search of materials with target functionalities. Nature Reviews Chemistry, 2(4):0121, 2018. +[40] Arghya Bhowmik, Ivano E Castelli, Juan Maria Garcia-Lastra, Peter Bjørn Jørgensen, Ole Winther, and Tejs Vegge. A perspective on inverse design of battery interphases using multi-scale modelling, experiments and generative deep learning. Energy Storage Materials, 21:446-456, 2019. +[41] Yifei Li, Yuchen Sun, Pingchuan Ma, Eftychios Sifakis, Tao Du, Bo Zhu, and Wojciech Matusik. Neural fluid: Neural fluidic system design and control with differentiable simulation. In Advances in Neural Information Processing Systems (NeurIPS). +[42] Tiejun Li, Junjun Yan, Xinhai Chen, Zhichao Wang, Qingyang Zhang, Enqiang Zhou, Chunye Gong, and Jie Liu. Accelerating aerodynamic design optimization based on graph convolutional neural network. International Journal of Modern Physics C, 35(01):2450007, 2024. +[43] Reuven Y Rubinstein and Dirk P Kroese. The cross-entropy method: a unified approach to combinatorial optimization, Monte-Carlo simulation and machine learning. Springer Science & Business Media, 2004. + +[44] Peter ZG Qian and CF Jeff Wu. Bayesian hierarchical modeling for integrating low-accuracy and high-accuracy experiments. Technometrics, 50(2):192–204, 2008. +[45] Yuanming Hu, Luke Anderson, Tzu-Mao Li, Qi Sun, Nathan Carr, Jonathan Ragan-Kelley, and Frédo Durand. Difftaichi: Differentiable programming for physical simulation. arXiv preprint arXiv:1910.00935, 2019. +[46] Kazuo Yonekura, Kazunari Wada, and Katsuyuki Suzuki. Generating various airfoils with required lift coefficients by combining naca and joukowski airfoils using conditional variational autoencoders. Engineering Applications of Artificial Intelligence, 108:104560, 2022. +[47] Wei Chen, Kevin Chiu, and Mark D Fuge. Airfoil design parameterization and optimization using bezier generative adversarial networks. AIAA journal, 58(11):4723-4735, 2020. +[48] Gabriel Achour, Woong Je Sung, Olivia J Pinon-Fischer, and Dimitri N Mavris. Development of a conditional generative adversarial network for airfoil shape optimization. In AIAA Scitech 2020 Forum, page 2261, 2020. +[49] Jian Liu, Jianyu Wu, Hairun Xie, Jing Wang, Liu Wei, Wanli Ouyang, Junjun Jiang, Xianming Liu, SHIXIANG TANG, Miao Zhang, et al. Afbench: A large-scale benchmark for airfoil design. Advances in Neural Information Processing Systems (NeurIPS), 37:82757-82780, 2024. +[50] Jichao Li, Xiaosong Du, and Joaquim RRA Martins. Machine learning in aerodynamic shape optimization. Progress in Aerospace Sciences, 134:100849, 2022. +[51] Xinghui Yan, Jihong Zhu, Minchi Kuang, and Xiangyang Wang. Aerodynamic shape optimization using a novel optimizer based on machine learning techniques. Aerospace Science and Technology, 86:826-835, 2019. +[52] Stephan Schmidt, Caslav Ilic, Volker Schulz, and Nicolas R Gauger. Three-dimensional large-scale aerodynamic shape optimization based on shape calculus. AIAA journal, 51(11):2615-2627, 2013. +[53] Xiaolong He, Jichao Li, Charles A Mader, Anil Yildirim, and Joaquim RRA Martins. Robust aerodynamic shape optimization—from a circle to an airfoil. Aerospace Science and Technology, 87:48-61, 2019. +[54] Nikos Arechiga, Frank Permenter, Binyang Song, and Chenyang Yuan. Drag-guided diffusion models for vehicle image generation. arXiv preprint arXiv:2306.09935, 2023. +[55] Jichao Li and Jinsheng Cai. Massively multipoint aerodynamic shape design via surrogate-assisted gradient-based optimization. AIAA Journal, 58(5):1949-1963, 2020. +[56] Jichao Li and Mengqi Zhang. Adjoint-free aerodynamic shape optimization of the common research model wing. AIAA Journal, 59(6):1990-2000, 2021. +[57] Thomas P Dussauge, Woong Je Sung, Olivia J Pinon Fischer, and Dimitri N Mavris. A reinforcement learning approach to airfoil shape optimization. Scientific Reports, 13(1):9753, 2023. +[58] Mohamed Amine Bouhlel, Sicheng He, and Joaquim RRA Martins. Scalable gradient-enhanced artificial neural networks for airfoil shape design in the subsonic and transonic regimes. Structural and Multidisciplinary Optimization, 61(4):1363-1376, 2020. +[59] Xiaosong Du, Ping He, and Joaquim RRA Martins. A b-spline-based generative adversarial network model for fast interactive airfoil aerodynamic optimization. In AIAA scitech 2020 forum, page 2128, 2020. +[60] Shuang Wu, Youtian Lin, Feihu Zhang, Yifei Zeng, Jingxi Xu, Philip Torr, Xun Cao, and Yao Yao. Direct3d: Scalable image-to-3d generation via 3d latent diffusion transformer. Advances in Neural Information Processing Systems (NeurIPS), 2024. +[61] Andrew Jaegle, Felix Gimeno, Andy Brock, Oriol Vinyals, Andrew Zisserman, and Joao Carreira. Perceiver: General perception with iterative attention. In International Conference on Machine Learning (ICML), pages 4651-4664. PMLR, 2021. + +[62] Tengfei Wang, Bo Zhang, Ting Zhang, Shuyang Gu, Jianmin Bao, Tadas Baltrusaitis, Jingjing Shen, Dong Chen, Fang Wen, Qifeng Chen, et al. Rodin: A generative model for sculpting 3d digital avatars using diffusion. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), pages 4563-4573, 2023. +[63] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. +[64] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. +[65] Hyungjin Chung, Jeongsol Kim, Michael T McCann, Marc L Klasky, and Jong Chul Ye. Diffusion posterior sampling for general noisy inverse problems. arXiv preprint arXiv:2209.14687, 2022. +[66] Arpit Bansal, Hong-Min Chu, Avi Schwarzschild, Soumyadip Sengupta, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Universal guidance for diffusion models. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), pages 843–852, 2023. +[67] Gentaro Hirota, Renee Maheshwari, and Ming C Lin. Fast volume-preserving free form deformation using multi-level optimization. In Proceedings of the fifth ACM symposium on Solid modeling and applications, pages 234-245, 1999. +[68] William M Hsu, John F Hughes, and Henry Kaufman. Direct manipulation of free-form deformations. ACM Siggraph Computer Graphics, 26(2):177-184, 1992. +[69] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2016. +[70] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the International Conference on Computer Vision (ICCV), pages 4195-4205, 2023. +[71] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. +[72] Mohamed Elrefaie, Florin Morar, Angela Dai, and Faez Ahmed. Drivaernet++: A large-scale multimodal car dataset with computational fluid dynamics simulations and deep learning benchmarks. Advances in Neural Information Processing Systems (NeurIPS), 37:499-536, 2024. +[73] Mohamed Elrefaie, Angela Dai, and Faez Ahmed. Drivaernet: A parametric car dataset for data-driven aerodynamic design and graph-based drag prediction. In International Design Engineering Technical Conferences and Computers and Information in Engineering Conference, volume 88360, page V03AT03A019. American Society of Mechanical Engineers, 2024. +[74] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), pages 165-174, 2019. +[75] Alex Nichol, Heewoo Jun, Prafulla Dhariwal, Pamela Mishkin, and Mark Chen. Point-e: A system for generating 3d point clouds from complex prompts. arXiv preprint arXiv:2212.08751, 2022. +[76] Premith Kumar Chilukuri, Binyang Song, SungKu Kang, and Ran Jin. Generating optimized 3d designs for manufacturing using a guided voxel diffusion model. In International Manufacturing Science and Engineering Conference, volume 88117, page V002T07A006. American Society of Mechanical Engineers, 2024. +[77] Diederik P Kingma, Max Welling, et al. Auto-encoding variational bayes, 2013. +[78] Hervé Abdi and Lynne J Williams. Principal component analysis. Wiley interdisciplinary reviews: computational statistics, 2(4):433-459, 2010. + +[79] Christopher Greenshields. OpenFOAM v11 User Guide. The OpenFOAM Foundation, London, UK, 2023. +[80] Florian R Menter, Martin Kuntz, Robin Langtry, et al. Ten years of industrial experience with the sst turbulence model. Turbulence, heat and mass transfer, 4(1):625-632, 2003. + +# NeurIPS Paper Checklist + +# 1. Claims + +Question: Do the main claims made in the abstract and introduction accurately reflect the paper's contributions and scope? + +Answer: [Yes] + +Justification: In this paper, we propose a novel 3D inverse-design framework that can directly navigate the full 3D design space. Extensive experiments in aerodynamic shape optimization demonstrate the effectiveness of our 3DID framework. + +Guidelines: + +- The answer NA means that the abstract and introduction do not include the claims made in the paper. +- The abstract and/or introduction should clearly state the claims made, including the contributions made in the paper and important assumptions and limitations. A No or NA answer to this question will not be perceived well by the reviewers. +- The claims made should match theoretical and experimental results, and reflect how much the results can be expected to generalize to other settings. +- It is fine to include aspirational goals as motivation as long as it is clear that these goals are not attained by the paper. + +# 2. Limitations + +Question: Does the paper discuss the limitations of the work performed by the authors? + +Answer: [Yes] + +Justification: The limitations of 3DID are detailed in the main paper (Sec. 5). + +Guidelines: + +- The answer NA means that the paper has no limitation while the answer No means that the paper has limitations, but those are not discussed in the paper. +- The authors are encouraged to create a separate "Limitations" section in their paper. +- The paper should point out any strong assumptions and how robust the results are to violations of these assumptions (e.g., independence assumptions, noiseless settings, model well-specification, asymptotic approximations only holding locally). The authors should reflect on how these assumptions might be violated in practice and what the implications would be. +- The authors should reflect on the scope of the claims made, e.g., if the approach was only tested on a few datasets or with a few runs. In general, empirical results often depend on implicit assumptions, which should be articulated. +- The authors should reflect on the factors that influence the performance of the approach. For example, a facial recognition algorithm may perform poorly when image resolution is low or images are taken in low lighting. Or a speech-to-text system might not be used reliably to provide closed captions for online lectures because it fails to handle technical jargon. +- The authors should discuss the computational efficiency of the proposed algorithms and how they scale with dataset size. +- If applicable, the authors should discuss possible limitations of their approach to address problems of privacy and fairness. +- While the authors might fear that complete honesty about limitations might be used by reviewers as grounds for rejection, a worse outcome might be that reviewers discover limitations that aren't acknowledged in the paper. The authors should use their best judgment and recognize that individual actions in favor of transparency play an important role in developing norms that preserve the integrity of the community. Reviewers will be specifically instructed to not penalize honesty concerning limitations. + +# 3. Theory assumptions and proofs + +Question: For each theoretical result, does the paper provide the full set of assumptions and a complete (and correct) proof? + +# Answer: [NA] + +Justification: The paper does not include theoretical results. + +# Guidelines: + +- The answer NA means that the paper does not include theoretical results. +- All the theorems, formulas, and proofs in the paper should be numbered and cross-referenced. +- All assumptions should be clearly stated or referenced in the statement of any theorems. +- The proofs can either appear in the main paper or the supplemental material, but if they appear in the supplemental material, the authors are encouraged to provide a short proof sketch to provide intuition. +- Inversely, any informal proof provided in the core of the paper should be complemented by formal proofs provided in appendix or supplemental material. +- Theorems and Lemmas that the proof relies upon should be properly referenced. + +# 4. Experimental result reproducibility + +Question: Does the paper fully disclose all the information needed to reproduce the main experimental results of the paper to the extent that it affects the main claims and/or conclusions of the paper (regardless of whether the code and data are provided or not)? + +# Answer: [Yes] + +Justification: We provide a full description of 3DID's training setup and our dataset preparation pipeline in the Appendix. + +# Guidelines: + +- The answer NA means that the paper does not include experiments. +- If the paper includes experiments, a No answer to this question will not be perceived well by the reviewers: Making the paper reproducible is important, regardless of whether the code and data are provided or not. +- If the contribution is a dataset and/or model, the authors should describe the steps taken to make their results reproducible or verifiable. +- Depending on the contribution, reproducibility can be accomplished in various ways. For example, if the contribution is a novel architecture, describing the architecture fully might suffice, or if the contribution is a specific model and empirical evaluation, it may be necessary to either make it possible for others to replicate the model with the same dataset, or provide access to the model. In general, releasing code and data is often one good way to accomplish this, but reproducibility can also be provided via detailed instructions for how to replicate the results, access to a hosted model (e.g., in the case of a large language model), releasing of a model checkpoint, or other means that are appropriate to the research performed. +- While NeurIPS does not require releasing code, the conference does require all submissions to provide some reasonable avenue for reproducibility, which may depend on the nature of the contribution. For example +(a) If the contribution is primarily a new algorithm, the paper should make it clear how to reproduce that algorithm. +(b) If the contribution is primarily a new model architecture, the paper should describe the architecture clearly and fully. +(c) If the contribution is a new model (e.g., a large language model), then there should either be a way to access this model for reproducing the results or a way to reproduce the model (e.g., with an open-source dataset or instructions for how to construct the dataset). +(d) We recognize that reproducibility may be tricky in some cases, in which case authors are welcome to describe the particular way they provide for reproducibility. In the case of closed-source models, it may be that access to the model is limited in some way (e.g., to registered users), but it should be possible for other researchers to have some path to reproducing or verifying the results. + +# 5. Open access to data and code + +Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material? + +Answer: [Yes] + +Justification: Our code will be open-sourced and available on GitHub upon publication. + +Guidelines: + +- The answer NA means that paper does not include experiments requiring code. +- Please see the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details. +- While we encourage the release of code and data, we understand that this might not be possible, so "No" is an acceptable answer. Papers cannot be rejected simply for not including code, unless this is central to the contribution (e.g., for a new open-source benchmark). +- The instructions should contain the exact command and environment needed to run to reproduce the results. See the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details. +- The authors should provide instructions on data access and preparation, including how to access the raw data, preprocessed data, intermediate data, and generated data, etc. +- The authors should provide scripts to reproduce all experimental results for the new proposed method and baselines. If only a subset of experiments are reproducible, they should state which ones are omitted from the script and why. +- At submission time, to preserve anonymity, the authors should release anonymized versions (if applicable). +- Providing as much information as possible in supplemental material (appended to the paper) is recommended, but including URLs to data and code is permitted. + +# 6. Experimental setting/details + +Question: Does the paper specify all the training and test details (e.g., data splits, hyperparameters, how they were chosen, type of optimizer, etc.) necessary to understand the results? + +Answer: [Yes] + +Justification: The details of training and evaluation are included in Appendix. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The experimental setting should be presented in the core of the paper to a level of detail that is necessary to appreciate the results and make sense of them. +- The full details can be provided either with the code, in appendix, or as supplemental material. + +# 7. Experiment statistical significance + +Question: Does the paper report error bars suitably and correctly defined or other appropriate information about the statistical significance of the experiments? + +Answer: [Yes] + +Justification: The confidence interval information is detailed in the Appendix. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The authors should answer "Yes" if the results are accompanied by error bars, confidence intervals, or statistical significance tests, at least for the experiments that support the main claims of the paper. +- The factors of variability that the error bars are capturing should be clearly stated (for example, train/test split, initialization, random drawing of some parameter, or overall run with given experimental conditions). +- The method for calculating the error bars should be explained (closed form formula, call to a library function, bootstrap, etc.) + +- The assumptions made should be given (e.g., Normally distributed errors). +- It should be clear whether the error bar is the standard deviation or the standard error of the mean. +- It is OK to report 1-sigma error bars, but one should state it. The authors should preferably report a 2-sigma error bar than state that they have a $96\%$ CI, if the hypothesis of Normality of errors is not verified. +- For asymmetric distributions, the authors should be careful not to show in tables or figures symmetric error bars that would yield results that are out of range (e.g. negative error rates). +- If error bars are reported in tables or plots, The authors should explain in the text how they were calculated and reference the corresponding figures or tables in the text. + +# 8. Experiments compute resources + +Question: For each experiment, does the paper provide sufficient information on the computer resources (type of compute workers, memory, time of execution) needed to reproduce the experiments? + +Answer: [Yes] + +Justification: The details of compute resources are included in Appendix. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The paper should indicate the type of compute workers CPU or GPU, internal cluster, or cloud provider, including relevant memory and storage. +- The paper should provide the amount of compute required for each of the individual experimental runs as well as estimate the total compute. +- The paper should disclose whether the full research project required more compute than the experiments reported in the paper (e.g., preliminary or failed experiments that didn't make it into the paper). + +# 9. Code of ethics + +Question: Does the research conducted in the paper conform, in every respect, with the NeurIPS Code of Ethics https://neurips.cc/public/EthicsGuidelines? + +Answer: [Yes] + +Justification: We have reviewed the Code of Ethics and it conforms with the Code of Ethics. + +Guidelines: + +- The answer NA means that the authors have not reviewed the NeurIPS Code of Ethics. +- If the authors answer No, they should explain the special circumstances that require a deviation from the Code of Ethics. +- The authors should make sure to preserve anonymity (e.g., if there is a special consideration due to laws or regulations in their jurisdiction). + +# 10. Broader impacts + +Question: Does the paper discuss both potential positive societal impacts and negative societal impacts of the work performed? + +Answer: [Yes] + +Justification: The details of broader impacts are included in Appendix. + +Guidelines: + +- The answer NA means that there is no societal impact of the work performed. +- If the authors answer NA or No, they should explain why their work has no societal impact or why the paper does not address societal impact. +- Examples of negative societal impacts include potential malicious or unintended uses (e.g., disinformation, generating fake profiles, surveillance), fairness considerations (e.g., deployment of technologies that could make decisions that unfairly impact specific groups), privacy considerations, and security considerations. + +- The conference expects that many papers will be foundational research and not tied to particular applications, let alone deployments. However, if there is a direct path to any negative applications, the authors should point it out. For example, it is legitimate to point out that an improvement in the quality of generative models could be used to generate deepfakes for disinformation. On the other hand, it is not needed to point out that a generic algorithm for optimizing neural networks could enable people to train models that generate Deepfakes faster. +- The authors should consider possible harms that could arise when the technology is being used as intended and functioning correctly, harms that could arise when the technology is being used as intended but gives incorrect results, and harms following from (intentional or unintentional) misuse of the technology. +- If there are negative societal impacts, the authors could also discuss possible mitigation strategies (e.g., gated release of models, providing defenses in addition to attacks, mechanisms for monitoring misuse, mechanisms to monitor how a system learns from feedback over time, improving the efficiency and accessibility of ML). + +# 11. Safeguards + +Question: Does the paper describe safeguards that have been put in place for responsible release of data or models that have a high risk for misuse (e.g., pretrained language models, image generators, or scraped datasets)? + +Answer: [NA] + +Justification: Our model focuses on 3D inverse design, with minimal risk of misuse. + +Guidelines: + +- The answer NA means that the paper poses no such risks. +- Released models that have a high risk for misuse or dual-use should be released with necessary safeguards to allow for controlled use of the model, for example by requiring that users adhere to usage guidelines or restrictions to access the model or implementing safety filters. +- Datasets that have been scraped from the Internet could pose safety risks. The authors should describe how they avoided releasing unsafe images. +- We recognize that providing effective safeguards is challenging, and many papers do not require this, but we encourage authors to take this into account and make a best faith effort. + +# 12. Licenses for existing assets + +Question: Are the creators or original owners of assets (e.g., code, data, models), used in the paper, properly credited and are the license and terms of use explicitly mentioned and properly respected? + +Answer: [Yes] + +Justification: The licenses are mentioned in Appendix. + +Guidelines: + +- The answer NA means that the paper does not use existing assets. +- The authors should cite the original paper that produced the code package or dataset. +- The authors should state which version of the asset is used and, if possible, include a URL. +- The name of the license (e.g., CC-BY 4.0) should be included for each asset. +- For scraped data from a particular source (e.g., website), the copyright and terms of service of that source should be provided. +- If assets are released, the license, copyright information, and terms of use in the package should be provided. For popular datasets, paperswithcode.com/datasets has curated licenses for some datasets. Their licensing guide can help determine the license of a dataset. +- For existing datasets that are re-packaged, both the original license and the license of the derived asset (if it has changed) should be provided. + +- If this information is not available online, the authors are encouraged to reach out to the asset's creators. + +# 13. New assets + +Question: Are new assets introduced in the paper well documented and is the documentation provided alongside the assets? + +Answer: [Yes] + +Justification: The details of the model are provided in this paper. + +Guidelines: + +- The answer NA means that the paper does not release new assets. +- Researchers should communicate the details of the dataset/code/model as part of their submissions via structured templates. This includes details about training, license, limitations, etc. +- The paper should discuss whether and how consent was obtained from people whose asset is used. +- At submission time, remember to anonymize your assets (if applicable). You can either create an anonymized URL or include an anonymized zip file. + +# 14. Crowdsourcing and research with human subjects + +Question: For crowdsourcing experiments and research with human subjects, does the paper include the full text of instructions given to participants and screenshots, if applicable, as well as details about compensation (if any)? + +Answer: [NA] + +Justification: The paper does not involve crowdsourcing nor research with human subjects. + +Guidelines: + +- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. +- Including this information in the supplemental material is fine, but if the main contribution of the paper involves human subjects, then as much detail as possible should be included in the main paper. +- According to the NeurIPS Code of Ethics, workers involved in data collection, curation, or other labor should be paid at least the minimum wage in the country of the data collector. + +# 15. Institutional review board (IRB) approvals or equivalent for research with human subjects + +Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals (or an equivalent approval/review based on the requirements of your country or institution) were obtained? + +Answer: [NA] + +Justification: The paper does not involve crowdsourcing nor research with human subjects. + +Guidelines: + +- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. +- Depending on the country in which research is conducted, IRB approval (or equivalent) may be required for any human subjects research. If you obtained IRB approval, you should clearly state this in the paper. +- We recognize that the procedures for this may vary significantly between institutions and locations, and we expect authors to adhere to the NeurIPS Code of Ethics and the guidelines for their institution. +- For initial submissions, do not include any information that would break anonymity (if applicable), such as the institution conducting the review. + +# 16. Declaration of LLM usage + +Question: Does the paper describe the usage of LLMs if it is an important, original, or non-standard component of the core methods in this research? Note that if the LLM is used only for writing, editing, or formatting purposes and does not impact the core methodology, scientific rigorousness, or originality of the research, declaration is not required. + +# Answer: [NA] + +Justification: LLM is only used for editing. + +# Guidelines: + +- The answer NA means that the core method development in this research does not involve LLMs as any important, original, or non-standard components. +- Please refer to our LLM policy (https://neurips.cc/Conferences/2025/LLM) for what should or should not be described. + +# Appendix + +In Appendix A, we provide additional experiments results. + +- In Appendix A.1, we further provide the full results for 3D vehicle aerodynamic design. +- In Appendix A.2, we provide more visualization of design results. +- In Appendix A.3, we provide more qualitative comparisons of topology-preserving refinement. + +Appendix B: The implementation details of baseline methods. + +Appendix C: The dataset processing details. + +Appendix D: The implementation details of 3DID. + +Appendix E: The evaluation details of 3DID. + +Appendix F: The broader impact of 3DID. + +Appendix G: The licenses of datasets, codes, and models used in this paper. + +# A Additional Results + +# A.1 Full Results for 3D Vehicle Aerodynamic Design + +Here we present the full statistical results of our experiments, including $95\%$ confidence intervals for all compared methods, shown in Table 4. A box plot of the simulation-derived drag coefficient (Sim-Drag) is shown in Figure 6, illustrating the distribution, variability, and outlier behavior across different approaches. + +Table 4: Quantitative comparison for aerodynamic vehicle design. + +
MethodPred-Drag↓Sim-Drag↓Novelty↑Coverage↑
GP, Voxel0.2997±0.04360.4254 ± 0.03511.0399±0.05720.5200±0.0675
GP, Voxel+PCA0.3059±0.04900.4363 ± 0.04250.9734±0.01950.5850±0.0675
CEM, Voxel0.2951±0.04210.4097 ± 0.02790.9792±0.02130.4350±0.0676
CEM, Voxel+PCA0.3088±0.04780.4393 ± 0.04690.9864±0.02500.5100±0.0600
CEM, TripNet0.3154±0.04760.4161 ± 0.04151.0399±0.03230.6050± 0.0725
Backprop, Voxel0.2979±0.03140.4146 ± 0.02440.9860±0.02040.4750±0.0675
Backprop, Voxel+PCA0.3061±0.05760.4614 ± 0.03160.9798±0.02080.4950±0.0675
Backprop, TripNet0.3153±0.04720.4170 ± 0.04441.0294±0.02900.5900±0.0700
3DID-NoTopoRefine (ours)0.2623±0.03730.3766 ± 0.03930.9195±0.02130.6950±0.0627
3DID (ours)0.2607±0.03310.3536± 0.03131.1709±0.02820.4300±0.0650
+ +![](images/747a8479eb829fd1d2560531b795af8e9d544352422ee3cea29eecfb6e431797.jpg) +Figure 6: The box plot of the simulation-derived drag coefficient. + +# A.2 Visualization of 3DID Design + +Additional visualizations of our designs are provided in Figure 7, where each design is shown alongside its geometry and corresponding physical fields. + +# A.3 Comparisons of Topology-Preserving Refinement + +We provide additional qualitative comparisons in Figure 8 to demonstrate the effectiveness of our refinement stage. As shown, the design candidates consistently evolve toward a fastback profile after refinement, exhibiting reduced low-velocity recirculation regions and enhanced downward flow patterns, which indicate improved aerodynamic performance. + +![](images/61064ceaf6f8e184f94e791bc467c2acd070da9e68452d073371640f6638b101.jpg) +Figure 7: Qualitative results of our 3DID. Each row displays a design candidate along with its corresponding velocity and pressure field heatmaps. + +![](images/152f19da1f34e5116bcc8f06d5b9e69faa3eb04ceb8624e60f446f6efd3962ab.jpg) +Figure 8: Qualitative comparisons of topology-preserving refinement. Each row presents two design candidates comparisons with their geometry and simulated velocity field heatmaps. + +# B Baseline implementation details + +In our experiments, we compare our method against traditional sampling-based and backpropagation-based approaches using various design representations. As baselines, we include the Cross-Entropy Method(CEM) [43], the Gaussian-process surrogate with Bayesian optimization(GP) [44], and the gradient-based backpropagation method(Backprop) [18]. For representations, the optimizer is instantiated with three representations: a dense voxel grid [76], a PCA-compressed voxel grid (Voxel+PCA) [22], and a pure geometry triplane network (TripNet) [21]. For each representation, we train a VAE model [77] to compress the high-dimensional geometry into a compact latent code, which serves as the optimization space for inverse design. + +# B.1 Representation Baseline + +Voxel. We train a voxel VAE [77] model directly on dense voxelized geometry to learn a latent embedding, as demonstrated in Figure 9. To train the model, we utilize the entire DrivAerNet++ [72] dataset, and voxelize the provided geometry with $256^3$ resolution. For the Encoder, we leverage a sequence of 3D convolution layers followed by batch normalization and LeakyReLU to encode the voxel grid into a compact latent $z_{\mathrm{voxel}}$ . For voxel decoder, the latent vector $z_{\mathrm{voxel}}$ is first projected to a high-dimensional feature space and reshaped into a 3D tensor. A sequence of 3D transposed convolutional layers is then applied to reconstruct the voxel grid from this intermediate representation. Additionally, a separate drag prediction head, implemented as a multi-layer perceptron (MLP), is applied to estimate the target drag coefficient. We train the VAE model with reconstruction loss $\mathcal{L}_{\mathrm{recon}}$ , KL loss $\mathcal{L}_{\mathrm{KL}}$ , and the drag coefficient prediction loss $\mathcal{L}_{\mathrm{drag}}$ . The hyperparameters of the model and training are provided in Table 5 + +![](images/511b7c4355b88a592b7c2bd00f8b5a890363e6ff150f6951b5551bfa8cfee2b4.jpg) +Figure 9: The overview of Voxel-VAE. + +Voxel-PCA. Our Voxel-PCA representation is based on the representation proposed by [22] with modifications. In contrast to the Voxel-VAE, which directly uses voxel grids as input, the Voxel-PCA model first applies a dimensionality reduction step before downstream processing, as shown in Figure 10. Specifically, given the voxel data, we perform PCA [78] to obtain a compact representation of each geometry. Then, with this representation, we leverage a series of MLPs to encode the reduced features into a latent code $z_{\mathrm{voxel - pca}}$ . For reconstruction, an MLP decoder is first applied to reconstruct the PCA features from the latent code, which are then projected back to the voxel grid using the inverse PCA transformation. For drag prediction, similar to Voxel-VAE, a separate drag prediction head is applied to estimate the target drag coefficient. Our Voxel-PCA model is also trained with reconstruction loss $\mathcal{L}_{\mathrm{recon}}$ , KL loss $\mathcal{L}_{\mathrm{KL}}$ and drag prediction loss $\mathcal{L}_{\mathrm{drag}}$ . The hyperparameters of the model and training are provided in Table 6. + +TripNet. Our TripNet representation is a pure geometry-based triplane representation, similar to the one proposed in [21], where it was used for forward prediction. The training procedure mirrors that of our unified physics-geometry framework, but excludes the physical field prediction branch, as illustrated in Figure 11. To obtain the representation, we utilize transformers with learnable tokens to extract features from the input point cloud. These features are then decoded using a transformer-based decoder and a geometry mapping network to predict the occupancy field of the design geometry. We + +![](images/ae66b1f7932d436c577410f7cd4f953ce54f9d154be3ddde6df04a516ca91186.jpg) +Figure 10: The overview of Voxel-PCA-VAE. + +utilize the Binary Cross-Entropy loss $\mathcal{L}_{\mathrm{BCE}}$ and KL loss $\mathcal{L}_{\mathrm{KL}}$ to supervise the training of VAE. To further predict the drag coefficient, we adopt the same U-Net architecture used in our objective-guided diffusion model. The TripNet-VAE architecture adopts the same hyperparameter configuration as the geometry branch of our PG-VAE. More training hyperparameters are provided in Table 7. + +![](images/db32eb6badae931bac6ea9a794d09e334ed826ce5288a23915f4a31f32ebda4f.jpg) +Figure 11: The overview of TripNet-VAE. + +# B.2 Optimization Baseline + +CEM. Cross Entropy Method [43] is a traditional sampling-based optimization method widely used in classical inverse design problems. It starts with an initial distribution, and in each iteration, it samples multiple candidates from the current distribution. Then, these candidates are evaluated against the target objective function to select a subset of elite samples with the best performance. The distribution parameters are updated based on these elite samples. A smoothing coefficient controls the rate of distribution updates between iterations. This process continues until convergence or a maximum number of iterations is reached. In our experiment, we utilize a Gaussian distribution derived from encoded randomly selected samples as the initial distribution to provide a valid starting point. + +GP. Gaussian-process surrogate with Bayesian optimization is a classical optimization method for black-box optimization [16, 17]. Bayesian optimization (BO) operates by constructing a probabilistic surrogate model, commonly a Gaussian-process model, to approximate the objective function based on past observations. At each iteration, an acquisition function is used to balance exploration of uncertain regions and exploitation of promising areas, guiding the selection of the next evaluation point. This strategy enables efficient optimization in high-cost or sample-limited scenarios by focusing evaluations on the most informative regions of the design space. While this method is effective in low-dimensional settings, constructing an accurate GP model becomes computationally expensive and challenging as the dimensionality of the design space increases. Therefore, GP-based Bayesian optimization is typically limited to small-scale or low-dimensional problems, where the surrogate can be reliably trained. In our experiment, our Gaussian process employs a Matérn kernel with constant and white noise components to model the objective function. At each iteration, Expected Improvement (EI) is used as the acquisition function. + +Backprop. With the trained surrogate models, end-to-end backpropagation enables efficient gradient-based optimization of the design, leveraging the differentiability of the surrogate to guide updates [18, 19]. In our experiments, we use the trained drag predictor as the surrogate and update the latent code using the Adam optimizer. + +# C Dataset processing details. + +In this work, we conduct experiments on $\mathrm{DrivAerNet}++$ [72], which is the largest aerodynamic car design dataset, comprising diverse car designs with corresponding CFD simulations. To train our model, we use the dataset with 8085 car designs to extract the point cloud and physical field. We first normalize each geometry of cars to fit within a unit cube, then uniformly sample 50,000 points with corresponding normals from the geometry surface. For the physical field, we apply the same scaling factor to ensure alignment with the normalized geometry. Subsequently, we randomly sample 50,000 points within the unit cube and interpolate the physical field values at each location. These points serve as the input of our PG-VAE. For supervision, we additionally sample another 50,000 points, each annotated with both occupancy values and physical field data. In this work, we focus on the pressure and velocity fields for the physical field representation, as wall shear stress is defined only on the surface of the geometry and is thus not suitable for volumetric sampling. During physical field interpolation, since some $\mathrm{DrivAerNet}++$ samples are simulated using only half of the geometry, we map each sampled point to its symmetric counterpart when necessary. For the U-Net and GNN surrogate models used in guided diffusion sampling and topology-preserving optimization, we employ the drag coefficient values provided by the $\mathrm{DrivAerNet}++$ dataset as ground truth supervision during training. + +# D Implementation details. + +Our framework consists of three key components: the Physics-Geometry VAE (PG-VAE), Objective-Guided Diffusion, and Topology-Preserving Refinement. Below, we provide detailed implementation descriptions for each component. + +PG-VAE. The PG-VAE serves to compress both the design geometry and the corresponding physical field into a unified latent representation. We sample $N_{g} = N_{p} = 50,000$ points for the geometry and physical field branches, respectively. The encoder consists of one cross-attention layer and eight self-attention layers, each with 12 attention heads and an embedding dimension of $d_{z} = 64$ . We use $r = 64$ for learnable tokens, and each with a channel dimension of $d_{e} = 768$ , to enhance representation expressiveness. The latent code dimension is set to $d_{z} = 32$ . The decoder architecture consists of one self-attention layer followed by five ResNet blocks [69], which upsample the latent vector into a triplane representation with resolution $R = 256$ and channel dimension $d_{t} = 64$ . The output triplane is then queried using a mapping network composed of five fully connected layers with a hidden size of 32 per branch. We adopt a semi-continuous occupancy formulation [60] and supervise both occupancy and physical field predictions using 50,000 sampled points within the normalized unit cube. We optimize the VAE using a combination of three loss terms: binary cross-entropy loss ( $\lambda_{\mathrm{BCE}} = 10^{-3}$ ), mean squared error for field regression ( $\lambda_{\mathrm{MSE}} = 10^{-5}$ ), and KL divergence ( $\lambda_{\mathrm{KL}} = 10^{-6}$ ). Training is performed using the AdamW optimizer [71] with a learning rate of $1 \times 10^{-4}$ , batch size 8 per GPU, for 100,000 steps. We use four NVIDIA RTX A6000 GPUs to train the model. + +Objective-Guided Diffusion. To explore the latent design space efficiently, we employ a latent-space diffusion model composed of 10 DiT blocks [70], each containing 16 attention heads with a head dimension of 72. The diffusion process includes 1,000 denoising steps. During inference, an auxiliary U-Net surrogate network is used to predict the task objective directly from the latent code $z$ , thereby guiding the sampling process toward optimal designs. The diffusion model is trained using a learning rate of $5 \times 10^{-5}$ , batch size of 4 per GPU, for 300,000 steps with the AdamW optimizer. We use four NVIDIA RTX A6000 GPUs to train the diffusion model. + +Topology-Preserving Refinement. To refine the initial design candidates while maintaining mesh topology, we apply a Free-Form Deformation (FFD) grid with $20 \times 6 \times 6$ control points along the x, y, and z axes, respectively. The deformation is guided by a surrogate model based on MeshGraphNet [30], which comprises 8 message-passing blocks and operates on the surface mesh. The + +MeshGraphNet is trained to predict the drag force from a deformed mesh, serving as a differentiable objective function during refinement. This model is trained with a learning rate of $1 \times 10^{-5}$ , batch size of 8 per GPU, for 100,000 steps, using AdamW as the optimizer. We use two NVIDIA RTX A6000 GPUs to train the MeshGraphNet. + +# E Evaluation details. + +In our experiments, we evaluate the design candidates using four metrics: predicted drag force (Pred-Drag), simulated drag force (Sim-Drag), novelty, and coverage. + +Pred-Drag. We use the pretrained surrogate model to estimate the drag force of each candidate mesh. Given the mesh of designed candidates $M^{*}$ , our surrogate model directly predict the objective drag force $\hat{\mathcal{I}}$ which can be formalized as: + +$$ +\hat {\mathcal {J}} = \mathcal {F} _ {\text {s u r r o g a t e}} \left(M ^ {*}\right), \tag {12} +$$ + +where $\mathcal{F}_{\mathrm{surrogate}}$ denotes the learned mapping from 3D mesh geometry to the predicted drag coefficient. For our surrogate model, we adopt a MeshGraphNet [30] with 8 message passing blocks as the surrogate model. Unlike the model used in our topology-preserving refinement stage, this predictor operates solely on geometry, without requiring the associated physical field. To train the model, we use the entire $\mathrm{DrivAerNet}++$ [72] dataset. Given that different representations may produce varying topological structures, we apply remeshing and simplification to all candidates for fair comparison. + +Sim-Drag. To obtain an unbiased evaluation of the generated designs, we perform high-fidelity Computational Fluid Dynamics (CFD) simulations and compute the corresponding drag coefficients. Following DrivAerNet++ [72], we employ the OpenFOAM®V11 [79] to conduct steady-state incompressible simulation using the $k - \omega$ SST turbulence model, based on Menter's formulation [80]. We performed a series of quality checks to ensure the generated geometries were simulation-ready and properly aligned within the CFD domain. During simulation, considering the computation cost, we set the maximum local cells to 10 million and the maximum global cells to 50 million in snappyHexMesh. The simulation iterates for 1000s, and we use the final $30\%$ simulation data to calculate the average drag coefficient. The hyperparameters of our simulation are provided in Table 8. + +Novelty. To quantitatively assess how different the generated designs are from the training data, we measure the novelty of each candidate. Specifically, novelty is computed as the average distance from each generated design to its nearest neighbor in the training set, reflecting how distinct the generated designs are from existing ones. Let $\{g_i\}_{i=1}^{N_g}$ denote the set of generated designs and $\{t_j\}_{j=1}^{N_t}$ denote the set of training designs in the feature space. The novelty is defined as: + +$$ +\text {N o v e l t y} = \frac {1}{N _ {g}} \sum_ {i = 1} ^ {N _ {g}} \min _ {j} d \left(\boldsymbol {g} _ {i}, \boldsymbol {t} _ {j}\right), \tag {13} +$$ + +where $d(\cdot, \cdot)$ denotes the distance between feature embeddings, computed using the pretrained PointNet encoder [75]. + +**Coverage.** The coverage metric (also known as recall) evaluates how well the generated designs cover the training distribution by measuring, for each training sample, the distance to its nearest generated design (using a k-nearest neighbor lookup) and reporting the fraction of training examples that fall within a predefined threshold. Let $\{g_i\}_{i=1}^{N_g}$ denote the set of generated designs and $\{t_j\}_{j=1}^{N_t}$ denote the set of training designs in the feature space. The coverage is defined as: + +$$ +\text {C o v e r a g e} = \frac {1}{N _ {t}} \sum_ {j = 1} ^ {N _ {t}} \mathbf {1} \left[ \min _ {i} d \left(\boldsymbol {t} _ {j}, \boldsymbol {g} _ {i}\right) \leq \tau \right] \tag {14} +$$ + +where $d(\cdot, \cdot)$ is a distance metric, $\tau$ is a predefined threshold, and $\mathbf{1}[\cdot]$ is the indicator function that equals 1 if the condition is true and 0 otherwise. + +# F Broader Impacts + +Academic Impact. 3DID's methodology, which enables direct navigation through 3D physics-geometry space, simplifies the 3D inverse design process. With the unified physics-geometry representation, the computation gap between 3D and lower-dimensional inverse design is narrowed, allowing researchers to focus more on exploring cutting-edge inverse design strategies rather than being constrained by computational limitations. With the two-stage optimization strategy, our method balances between exploration and validity, offering researchers an effective approach for inverse design involving 3D geometry. + +Social Impact. The proposed 3D Inverse Design (3DID) framework extends the scope of geometry-driven design by enabling direct optimization of full 3D structures from scratch. By combining unified physics-geometry representations with physics-aware optimization, our method opens the door to more efficient, automated design workflows in fields such as aerospace engineering, biomedicine, additive manufacturing, and nanophotonics. In particular, 3DID can be applied to complex design tasks that traditionally rely on expert-crafted initial geometries and time-consuming simulation-based evaluations. In mechanical engineering, it can be used to optimize structural components for strength, weight, and thermal performance without manual trial-and-error. In the medical field, 3DID enables the fabrication of patient-specific implants by automatically generating geometries tailored to individual physiological and functional requirements. + +# G License + +The code will be publicly accessible. We use standard licenses from the community. We include the following licenses for the codes, datasets, and models we used in this paper. + +# 1. Dataset + +- DrivAerNet++ [72]: CC BY-NC 4.0 + +# 2. Codes + +NVIDIA PhysicsNeMo: Apache License 2.0 + +# 3. Evaluation + +- OpenFOAM [79]: GNU General Public License + +Table 5: Hyperparameters for Voxel-VAE + +
Hyperparameter nameValue
Hyperparameters for Voxel-VAE architecture:
Input shape[8, 256, 256, 256]
Output shape[8, 256, 256, 256]
Number of 3D convolution layer5
Dimension of latent z_voxel512
Number of 3D transposed convolutional layer5
Number of MLPs in drag predictor5
Batch size8
Dimension of encoder(1, 32, 64, 128, 256, 512)
Dimension of voxel decoder(512, 256, 128, 64, 32, 1)
Dimension of drag predictor(512, 256, 128, 64, 32, 1)
Hyperparameters for Voxel-VAE training:
OptimizerAdamW
Learning rate1e-4
Learning steps100K
Learning rate adjustment strategyCosine
Warm-up steps5K
L_recon weight10-3
L_KL weight10-4
L_drag weight10-3
+ +Table 6: Hyperparameters for Voxel-PCA-VAE + +
Hyperparameter nameValue
Hyperparameters for Voxel-PCA-VAE architecture:
PCA output dimension400
Number of MLP layers in encoder4
Dimension of latent z_voxel-pca64
Number of MLP layers in decoder4
Number of MLPs in drag predictor2
Batch size32
Dimension of encoder(400, 256, 128, 64, 64)
Dimension of PCA decoder(64, 64, 128, 256, 400)
Dimension of drag predictor(64, 32, 1)
Hyperparameters for Voxel-PCA-VAE training:
OptimizerAdamW
Learning rate5e-4
Learning steps100K
Learning rate adjustment strategyCosine
Warm-up steps5K
L_recon weight10-2
L_KL weight10-4
Ldrag weight10-3
+ +Table 7: Hyperparameters for TripNet VAE + +
Hyperparameter nameValue
Hyperparameters for TripNet-VAE training:
Batch size8
OptimizerAdamW
Learning rate1e-4
Learning steps100K
Learning rate adjustment strategyCosine
Warm-up steps5K
L BCE weight10-3
L KL weight10-6
+ +Table 8: CFD Simulation Parameters for OpenFOAM + +
Parameter nameValue
Solver Configuration:
OpenFOAM versionv11
SolverincompressibleFluid
AlgorithmSIMPLE
Turbulence modelk-ω-SST
Simulation typeSteady-state RANS
Flow Conditions:
Flow velocity (u∞)30 m/s
Kinematic viscosity (ν)1.56 × 10-5m2/s
Air density (ρ)1.184 kg/m3
Turbulent kinetic energy (k)0.375 m2/s2
Specific dissipation rate (ω)1.78 s-1
Computational Domain:
Domain dimensions44×8×6.4 m
Inlet distance12 m upstream
Outlet distance32 m downstream
Solver Tolerances:
Pressure absolute tolerance1 × 10-6
Pressure relative tolerance3 × 10-2
Velocity absolute tolerance1 × 10-8
Velocity relative tolerance5 × 10-3
Turbulence absolute tolerance1 × 10-8
Turbulence relative tolerance1 × 10-3
Potential solver absolute tolerance1 × 10-7
Potential solver relative tolerance1 × 10-2
Mesh Refinement:
Surface refinement level3-4
Feature refinement level4
Regional refinement level2
Wake refinement level2
Boundary layers5 layers
Layer expansion ratio1.2
Final layer thickness0.5
Force Calculation:
Reference length (lref)4.777 m
Reference area (Aref)2.0 m2
Reference center(0,0,0)
Drag direction(1,0,0)
Lift direction(0,0,1)
Simulation Control:
End time1000 s
Time step1 s
Write interval100 steps
Force coefs write interval10 steps
\ No newline at end of file diff --git a/NeurIPS/2025/3DID_ Direct 3D Inverse Design for Aerodynamics with Physics-Aware Optimization/images.zip b/NeurIPS/2025/3DID_ Direct 3D Inverse Design for Aerodynamics with Physics-Aware Optimization/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..30598fc1367eafee72865b38c17c54615edcbf82 --- /dev/null +++ b/NeurIPS/2025/3DID_ Direct 3D Inverse Design for Aerodynamics with Physics-Aware Optimization/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3efacefc4855837e54ef465deb8eb13aff9f0638c87703c39cc71002de121157 +size 1131452 diff --git a/NeurIPS/2025/3DID_ Direct 3D Inverse Design for Aerodynamics with Physics-Aware Optimization/layout.json b/NeurIPS/2025/3DID_ Direct 3D Inverse Design for Aerodynamics with Physics-Aware Optimization/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..6deaad97d524c3ce92fde293e6892dad10cec0a3 --- /dev/null +++ b/NeurIPS/2025/3DID_ Direct 3D Inverse Design for Aerodynamics with Physics-Aware Optimization/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f61707787ea7a642f8a47fa93a3e7c70f3629fa26e4f8e5721dac9fad76ea99b +size 847510 diff --git a/NeurIPS/2025/3DLLM-Mem_ Long-Term Spatial-Temporal Memory for Embodied 3D Large Language Model/2cfab1b2-0b1f-41fb-85e5-a311afcdecda_content_list.json b/NeurIPS/2025/3DLLM-Mem_ Long-Term Spatial-Temporal Memory for Embodied 3D Large Language Model/2cfab1b2-0b1f-41fb-85e5-a311afcdecda_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d7e30ff514219e0092b446483012cac41d9bdabf --- /dev/null +++ b/NeurIPS/2025/3DLLM-Mem_ Long-Term Spatial-Temporal Memory for Embodied 3D Large Language Model/2cfab1b2-0b1f-41fb-85e5-a311afcdecda_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d49c8d56c57093b90b545e27a5a0c344c15fec34f18fe2ba10bf8cb5310279c2 +size 181040 diff --git a/NeurIPS/2025/3DLLM-Mem_ Long-Term Spatial-Temporal Memory for Embodied 3D Large Language Model/2cfab1b2-0b1f-41fb-85e5-a311afcdecda_model.json b/NeurIPS/2025/3DLLM-Mem_ Long-Term Spatial-Temporal Memory for Embodied 3D Large Language Model/2cfab1b2-0b1f-41fb-85e5-a311afcdecda_model.json new file mode 100644 index 0000000000000000000000000000000000000000..a05434b8bac664faac96442ee61db8c4a94bb792 --- /dev/null +++ b/NeurIPS/2025/3DLLM-Mem_ Long-Term Spatial-Temporal Memory for Embodied 3D Large Language Model/2cfab1b2-0b1f-41fb-85e5-a311afcdecda_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d866bc9a1ba3ae21257ec7197b2973f5f99def433d4c5823c07278135c5290a0 +size 232009 diff --git a/NeurIPS/2025/3DLLM-Mem_ Long-Term Spatial-Temporal Memory for Embodied 3D Large Language Model/2cfab1b2-0b1f-41fb-85e5-a311afcdecda_origin.pdf b/NeurIPS/2025/3DLLM-Mem_ Long-Term Spatial-Temporal Memory for Embodied 3D Large Language Model/2cfab1b2-0b1f-41fb-85e5-a311afcdecda_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..d753c12c6225c6fef284f265a078efef659a86b9 --- /dev/null +++ b/NeurIPS/2025/3DLLM-Mem_ Long-Term Spatial-Temporal Memory for Embodied 3D Large Language Model/2cfab1b2-0b1f-41fb-85e5-a311afcdecda_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a5c31df08d5ef9c7dc5da668c7d52fa9b319b6463efba8d9f52a684399198054 +size 7976557 diff --git a/NeurIPS/2025/3DLLM-Mem_ Long-Term Spatial-Temporal Memory for Embodied 3D Large Language Model/full.md b/NeurIPS/2025/3DLLM-Mem_ Long-Term Spatial-Temporal Memory for Embodied 3D Large Language Model/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e1db7c0dfc404b476de607970b1c09b8a83bca3c --- /dev/null +++ b/NeurIPS/2025/3DLLM-Mem_ Long-Term Spatial-Temporal Memory for Embodied 3D Large Language Model/full.md @@ -0,0 +1,871 @@ +# 3DLLM-Mem: Long-Term Spatial-Temporal Memory for Embodied 3D Large Language Model + +Wenbo Hu1 Yining Hong1 Yanjun Wang1 Leison Gao1 Zibu Wei1 +Xingcheng Yao1 Nanyun Peng1 Yonatan Bitton2 Idan Szpektor2 Kai-Wei Chang1 + +1University of California, Los Angeles, 2Google Research + +https://3d11m-mem.github.io + +![](images/21be0b4dc178ccf4043837d07df9b61ba2332d484d08c70f38733d9b17b3b2e3.jpg) +Figure 1: We propose 3DLLM-MEM, a memory-enhanced 3D embodied agent that explores and incorporates feedback from the environment, interacts with objects, and incrementally builds and maintains a task-relevant long-term memory throughout its trajectory. For illustration purposes, agents from multiple time steps are shown simultaneously. + +# Abstract + +Humans excel at performing complex tasks by leveraging long-term memory across temporal and spatial experiences. In contrast, current Large Language Models (LLMs) struggle to effectively plan and act in dynamic, multi-room 3D environments. We posit that part of this limitation is due to the lack of proper 3D spatial-temporal memory modeling in LLMs. To address this, we first introduce 3DMEM-BENCH, a comprehensive benchmark comprising over 26,000 trajectories and 2,892 embodied tasks, question-answering and captioning, designed to evaluate an agent's ability to reason over long-term memory in 3D environments. Second, we propose 3DLLM-MEM, a novel dynamic memory management and fusion model for embodied spatial-temporal reasoning and actions in LLMs. Our model uses working memory tokens, which represents current observations, as queries to selectively attend to and fuse the most useful spatial and temporal features from episodic memory, which stores past observations and interactions. Our approach allows the agent to focus on task-relevant information while maintaining memory efficiency in complex, long-horizon environments. Experimental results + +demonstrate that 3DLLM-MEM achieves state-of-the-art performance across various tasks, outperforming the strongest baselines by $16.5\%$ in success rate on 3DMEM-BENCH's most challenging in-the-wild embodied tasks. + +# 1 Introduction + +Picture yourself traversing an unfamiliar home, as illustrated in Figure 1, on a mission to explore multiple rooms and evaluate various gift boxes to find the most suitable one for wrapping a teddy bear. As you navigate from room to room, your brain instinctively creates a 3D cognitive map of the environment, maintains a working memory of objects you've encountered, forms episodic memories that link observations across space and time, and plans efficient actions. This seamless integration of 3D spatial understanding, long-term memory encoding and retrieval, fluid switching between working and episodic memory, and purposeful action planning — cognitive processes that humans take for granted — remain formidable challenges for embodied AI systems today. + +Recent extensions of Large Language Models (LLMs) to 3D environments have birthed 3D-LLMs (Hong et al., 2023b; Guo et al., 2023; Gu et al., 2024; Huang et al., 2024b; Xu et al., 2025a) that can perceive and reason about 3D spaces, while 3D Vision-Language-Action models (Zhen et al., 2024; Zhao et al., 2025; Intelligence et al., 2025) further incorporate the ability to plan and act within these environments. Despite these advances, several critical limitations persist that prevent models from performing the kinds of tasks described above. First, current models struggle to maintain long-term memory chains when performing complex tasks that unfold across multiple visual scenarios, such as several rooms in a house, and extended time frames. Real-world 3D physical scenes are remarkably vast and information-dense, where every detail can matter for long-horizon embodied tasks — for instance, in Figure 1, finding the most suitable gift box requires remembering all the gift boxes encountered along the way and their characteristics and interaction with teddy bear. Dense 3D representations are particularly valuable as they capture comprehensive spatial information, preserving intricate geometric relationships and environmental details that sparse or object-centric approaches might miss. However, how to accurately and efficiently store dense 3D memory remains a fundamental challenge - retrieving the entire history would overwhelm the model's context limits, while selective retrieval (Xie et al., 2024; Wang et al., 2024; Yang et al., 2025b) risks omitting critical information needed for accurate reasoning and decision-making. The second challenge resides in the entanglement of spatial and temporal memory — agents must track not only where objects are, but how they change over time through exploration and interaction. As environments evolve, maintaining coherent representations of previously seen spaces while incorporating new information continues to exceed the capabilities of current embodied AI models. + +Our efforts at solving this challenge are two-fold. First, we introduce a novel benchmark for reasoning, planning and acting with long-term spatial-temporal memory in embodied environments. Our benchmark, 3DMEM-BENCH, encompasses multi-room 3D scenes from the Habitat environment, augmented with interactive objects to enable manipulation tasks across extended spatial-temporal horizons. Notably, we define fine-grained embodied tasks across varying levels of difficulty—from simple to hard—enabling deeper insight into model performance, which we believe is not addressed in prior benchmarks as shown in Table 1. Our task set spans a wide range of complexities, from straightforward object collection to challenging comparative reasoning tasks that require integrating observations across multiple rooms and time steps. Additionally, we include in-the-wild challenge tasks to evaluate the model's generalization capabilities beyond seen environments. The benchmark includes three evaluation categories: (1) embodied tasks requiring extended action sequences across multiple rooms, (2) spatial-temporal embodied question answering (EQA) that evaluates understanding of spatial relationships over time, and (3) long-term scene captioning that tests memorization of previously observed environments. Our dataset includes $26,000+$ trajectory examples spanning $182+$ unique scenes with an average of 18 rooms per scene. + +Second, we introduce 3DLLM-MEM, a 3D embodied LLM with dynamic memory management capabilities designed specifically for embodied spatial-temporal reasoning, planning and acting. To our knowledge, we are among the first to explore dense 3D representations as memory for embodied 3D LLMs — addressing a significant gap in current research as noted in recent 3D memory studies (Yang et al., 2025b). Unlike standard approaches that rely solely on context windows (Hong et al., 2023b; Huang et al., 2024b; Zhu et al., 2024), 3DLLM-MEM implements a dual-memory system: a limited-capacity working memory for current observations and an expandable episodic + +
Benchmark#Test Tasks#Train TrajectoriesLong-term MemoryFine-grained complexityEQACaptioning
ALFWorld (Shridhar et al., 2021)2743,553××NANA
Behavior-1K (Li et al., 2024a)1,000NA××NANA
VisualAgentBench (Liu et al., 2024)7464,482××NANA
EmbodiedBench (Yang et al., 2025a)1,128NA××NANA
3DMEM-BENCH (ours)1,86026,276✓✓865167
+ +Table 1: Comparison with related benchmarks. 3DMEM-BENCH focus on spatial-temporal memory through fine-grained embodied tasks and EQA that span multiple "pieces" of long-term memory, distinguishing it from prior benchmarks that typically target single-step or short-horizon reasoning. Fine-grained complexity indicates our embodied task spans from simple to medium to hard. + +memory that stores past spatial-temporal information as dense 3D representations. The key innovation is our memory fusion module that actively integrates information from both memory systems based on task relevance and spatial-temporal relationships. This allows the model to leverage the benefits of dense 3D representations while mitigating their computational demands, maintaining coherent spatial-temporal understanding across extended task horizons. The fusion process preserves critical spatial relationships while accounting for their evolvement through agent interactions over time. + +We evaluate popular 3D-LLMs and memory mechanisms on 3DMEM-BENCH. Experimental results demonstrate 3DLLM-MEM significantly outperforms all existing approaches in both in-domain and in-the-wild embodied tasks. Notably, while the performance of other methods drops sharply in the challenging in-the-wild setting, our method remains robust, achieving an average success rate of $32.1\%$ —demonstrating strong generalization capabilities. As task complexity increases from simple to hard, all existing approaches degrade significantly, achieving only $\sim 5\%$ success rate in hard in-the-wild tasks. In contrast, 3DLLM-MEM maintains a strong performance of $27.8\%$ , demonstrating its scalability and effectiveness in managing longer-term memory representations. + +Our contributions can be summarized as below: + +- We propose a novel task that requires agents to execute action chains while maintaining and utilizing long-term spatial-temporal memory. +- We construct 3DMEM-BENCH, a comprehensive benchmark comprising over 26,000 trajectories and 1,860 fine-grained long-term memory embodied tasks—ranging from simple to hard—along with question-answering tasks that target memory changes across time and space, and captioning tasks in complex 3D environments. +- We propose 3DLLM-MEM, an embodied 3D LLM with a novel memory fusion module for spatial-temporal reasoning, planning, and acting-which utilizes working memory tokens as queries to selectively fuse relevant features from episodic memory for efficient, task-aware decision-making. +- Experimental results on embodied tasks, question-answering, and captioning demonstrate that 3DLLM-MEM outperforms baselines by a large margin. + +# 2 The Embodied 3D Long-Term Spatial-Temporal Memory Benchmark + +# 2.1 Overview of 3DMEM-BENCH + +Design principles Long-term memory (Camina and Guell, 2017; Friedman et al., 2018; Zlotnik and Vansintjan, 2019) can be categorized into explicit memory and implicit memory. Explicit memory includes semantic memory, which stores general knowledge and facts about the world, and episodic memory, which consists of personal experiences that are time-stamped and context-specific. In contrast, implicit memory primarily involves procedural memory, such as learned skills and habits. + +To comprehensively evaluate 3D long-term memory for real-world applications, we design 3DMEM-BENCH following three core task categories: embodied tasks, long-term memory EQA, and captioning. As illustrated in Figure 2, embodied tasks require an embodied agent to solve realistic indoor environment challenges by leveraging both implicit and explicit long-term memory. Long-term memory EQA tests the agent's ability to answer complex embodied questions using spatial-temporal memory. This task includes five subcategories: spatial reasoning questions, long-term object navigation, comparative reasoning, multi-room layout understanding, and semantic object counting. Captioning tasks involve summarizing the agent's episodic memory to highlight shared and distinctive features across experiences, enabling more informed decision-making under the current task context. + +![](images/52c0fddb47ad7205fd914b62f5bf1ea761be275401340d672043f3469b476f34.jpg) +Figure 2: Overview of 3DMEM-BENCH. For long-term memory embodied tasks, we further incorporate in-the-wild challenges to test 3D agent's generalization abilities. Text inside $<>$ indicates high-level action tokens. For complete embodied task trajectories, please refer to Appendix C. + +# 2.2 Data Collection + +Base environment construction We build our scenes on top of the Habitat-Matterport 3D (HM3D) semantics dataset (Ramakrishnan et al., 2021), which has 1000 3D spaces and 10,600 rooms within those spaces. Pre-processing for the axis-aligned bounding box and using valid semantic label annotation, we filter to 182 3D spaces and 2,602 rooms. However, existing objects in HM3D scene are not interactive in Habitat-sim (Szot et al., 2021). To expand our task diversity and enable embodied tasks, we add interactive objects from Objaverse (Deitke et al., 2023) which consists of 800K 3D objects spanning rich categories. More environment construction details are illustrated in Appendix B. + +Generating task trajectories Following Hong et al. (2023b, 2024), we adopt box-demonstration-instruction-based prompting, which utilizes the axis-aligned bounding boxes (AABB) of both rooms and objects within the 3D scenes to prompt Gemini (Team et al., 2023) to generate diverse tasks. We further prompt Gemini to incorporate interactive objects based on task requirements and their appropriateness within indoor environments. Detailed prompt instructions and few-shot demonstration examples are provided in Appendix E. To ensure the validity of the generated trajectories, we develop a trajectory simulation pipeline that verifies each trajectory step-by-step. At every step, the simulator checks: (1) the correctness of the agent's location, (2) the existence and validity of referenced objects, and (3) the correctness of pick-up and put-down actions. Finally, we ensure that high-level actions can be executed in the simulator, following (Szot et al., 2024; Yang et al., 2025a). Details of this implementation are in Appendix F.1. On average, our filtering process yields a validation rate of approximately $24\%$ , ensuring the correctness and feasibility of the generated trajectories. + +Embodied data collection In our task settings, an embodied agent first performs random exploration within the environment to collect RGB-D observations and corresponding camera poses. + +![](images/7b3f9d9cce78494e425d6255f3d461b21ff139d6b96e8961b3e8156907efc546.jpg) +Figure 3: (a) We propose 3DLLM-MEM, a memory-enhanced 3D embodied agent that gradually form its long-term memory while executing tasks. Multiple timesteps are shown together but in different colors, with each timestep's memory including the prior one. The task is "prepare a simple breakfast" as shown in Figure 2. (b) Overview of our memory fusion mechanism. + +![](images/c160b022f13c985a7cd36a80ac58c63df94a4565daf87519773998b3a5ede23b.jpg) + +Then the agent follows the task trajectory, incrementally exploring new environments, executing interaction actions, and receiving feedback with new RGB-D observation data. All interaction results are recorded and the reconstructed point cloud data is precomputed and stored locally to enable faster loading during both training and inference. + +# 2.3 Data Curation + +As mentioned previously, we collect embodied data by prompting Gemini. To enable a fine-grained analysis of long-term memory capacity, we divide the tasks into three subcategories: simple, medium, and hard, comprising of 3, 5 and 10 multi-room scene settings respectively. In total, we collect 51K trajectories, with 31K in the simple setting, 10K in the medium, and 10K in the hard. + +To construct in-domain evaluation sets, we first remove training tasks and filter for instances that never shown in the agent's working memory. For the in-the-wild evaluation set, we apply additional filtering to assess the agent's generalization capabilities. Specifically, we select instances involving unseen objects and entirely unseen memory context, and we introduce novel in-the-wild challenges that differ from those encountered during training, as illustrated in Figure 2. + +For EQA data curation, we extract complete trajectories explored by agents and then prompt Gemini to generate question-answer pairs. The questions are categorized into spatial reasoning, long-term object navigation, comparative reasoning, multi-room layout understanding, and semantic object counting. As shown in Figure 2, these questions evaluate models on spatial-temporal changes in memory during embodied task execution. For long-term memory captioning, which primarily targets semantic episodic memory, we collect data across multiple rooms before and after the execution of each trajectory, enabling comparison and summarization of memory-relevant experiences. + +Quality control After constructing the entire benchmark, we implement two quality control procedures: automatic validation using trajectory simulation rules and a manual review of each benchmark instance. The automatic check involves re-running the trajectory simulation validation pipeline, as described in §2.2, particularly for the in-the-wild tasks. For human validation, four student experts in the field manually inspect each benchmark example. We render multi-view images of the entire scene using the simulator and verify whether the benchmark annotations accurately correspond to the simulated environment. More details are in Appendix F.2. + +# 3 3D Long-Term Spatial-Temporal Memory Model (3DLLM-MEM) + +# 3.1 Preliminary + +Recent work on 3D Large Language Models (3D-LLMs) has showcased robust capabilities. We choose LLaVA-3D (Zhu et al., 2024) as the base model to build our long-term memory 3D-LLM. + +LLaVA-3D directly builds on 2D-LLM with multi-view images as input and utilizing the 3D position embeddings to bring the 2D patches within a 3D spatial context to construct 3D patches. For each frame image, a CLIP encoder splits the image $X \in \mathbb{R}^{3 \times W \times H}$ into patches at the patch size $P$ . For each 3D scene, $V$ multi-view image patch features are encoded and then projected into LLM space as $X_{p} \in \mathbb{R}^{V \times d \times w \times h}$ , where $h = \left\lfloor \frac{H}{P} \right\rfloor$ , $w = \left\lfloor \frac{W}{P} \right\rfloor$ , and $d$ represents LLM's hidden dimension. The 3D positions in the 3D world are obtained with known depth image, camera intrinsic and extrinsic parameters and are further encoded into 3D position embeddings $P \in \mathbb{R}^{V \times d \times w \times h}$ . These are directly added to the 2D patch visual tokens $X_{p}$ , resulting in pixel-aligned 3D patches $X_{3D} \in \mathbb{R}^{V \times d \times w \times h}$ . To reduce redundancy in 3D patches, we adopt the Farthest Point Sampling (FPS) strategy to downsample the 3D features to a fixed number of tokens, resulting in $X_{3D\mathrm{Feat}} \in \mathbb{R}^{N \times d}$ . + +# 3.2 3DLLM-MEM Memory Module + +A 3D embodied agent gradually explores the environment by collecting observations and interacting with surrounding environments. For humans, current observations are held in working memory, while longer-term observations and experiences are stored in episodic memory. Inspired by human cognitive structure, 3DLLM-MEM is designed with a similar paradigm as illustrated in Figure 3. The current observation at time step $t = i$ , denoted as $X^{[t = i]} \in \mathbb{R}^{N \times d}$ , remains within the context window and serves as the agent's working memory. As the agent accumulates more experiences, past observations from time steps 1 to $T$ , represented as $X^{[t = 1:T]} \in \mathbb{R}^{T \times N \times d}$ , are stored as part of its episodic memory, where $T$ denotes the total number of timesteps. + +Episodic memory To manage episodic memory, we propose the use of a memory feature bank. For each observation at time step $j$ , where $1 \leq j \leq T$ , we first apply a multi-layer perceptron (MLP) layer to project the observation into a memory-specific feature space, which is then stored in the memory bank for future retrieval. To further enhance the temporal understanding of the agent's exploration, we incorporate sinusoidal positional embeddings to encode each time step $t = j$ , and then directly added to the corresponding memory feature representations. + +Memory fusion Our motivation is that an agent should leverage its current observations to recall the most relevant information from its episodic memory in order to complete the current task. To achieve this, we propose a mechanism called 3D memory fusion. Specifically, we encode the 3D features from the working memory into a shared memory space and use this representation as the query feature, denoted as $f_{t}^{Q} \in \mathbb{R}^{N \times M}$ , where $M$ is the dimensionality of the memory feature space. + +The episodic memory bank stores the corresponding key and value features from past observations: $f^{K} \in \mathbb{R}^{T \times N \times M}$ and $f^{V} \in \mathbb{R}^{T \times N \times M}$ , respectively. Here, $T$ is the number of past timesteps and $N$ is the number of memory tokens per timestep. This structure allows the agent to retrieve task-relevant information through memory-query attention. The fused memory feature is then concatenated with the working memory feature to produce the final memory-enhanced representation $f^{M}$ for the agent: + +$$ +f _ {\text {f u s e}} ^ {Q} = \operatorname {S o f t m a x} \left(\frac {f _ {t} ^ {Q} \left(f ^ {K}\right) ^ {\top}}{\sqrt {C}}\right) f ^ {V}, \quad f ^ {M} = \operatorname {C o n c a t} \left[ f _ {\text {f u s e}} ^ {Q}; f _ {t} ^ {Q} \right] \tag {1} +$$ + +Memory update The working memory is dynamic and updated online. As the agent interacts with the environment, changes in the environment are immediately reflected in the working memory through updated 3D representations. When the agent moves to a new environment, the previous working memory is transferred to the episodic memory bank. If the corresponding environment already exists in the memory bank and has been modified by the agent, the memory entry is updated accordingly. Thus, the memory bank remains dynamic and reflects the latest state of the explored environments. As described in §2.2, environment changes and corresponding observations are pre-collected and stored locally to facilitate efficient data loading during both training and inference. + +# 4 Experiments + +In this section, we first introduce the experimental setup and existing memory management baselines in §4.1. Then, we benchmark existing approaches on 3DMEM-BENCH, and present comprehensive results on embodied tasks, EQA, and captioning tasks to demonstrate the effectiveness of our + +
ModelSimpleMediumHardAverage
In-domainIn-the-wildIn-domainIn-the-wildIn-domainIn-the-wildIn-domainIn-the-wild
SRSub-SRSRSub-SRSRSub-SRSRSub-SRSRSub-SRSRSub-SRSRSub-SRSRSub-SR
3D-LLM (Finetuned)10.420.39.118.5------------
Everything in Context35.563.932.445.2------------
Most Recent Memory32.862.323.438.620.134.812.425.310.420.75.412.121.139.313.725.3
Retrieval-Augmented Memory34.263.028.346.221.840.213.728.010.821.64.810.622.341.615.628.3
3DLLM-MEM (Ours)45.573.437.065.436.867.831.657.430.546.227.842.137.662.532.155.0
+ +(a) Results on 3DMEM-BENCH embodied tasks. SR stands for success rate. Sub-SR stands for sub-success rate. Our model outperforms existing approaches by a large margin. + +
ModelEmbodied TaskEmbodied Question Answering (EQA)Captioning
In-domainIn-the-wildSpatialNav.ComparativeLayoutCountBLEU1BLEU4METEOR
3D-LLM (Finetuned)--2.95.80.07.70.042.312.030.6
3D-Mem (GPT4-o)--39.911.025.819.17.841.74.731.8
3D-Mem (Gemini-2.5-Flash)--41.618.237.630.212.742.84.829.6
3D-Mem (Gemini-2.5-Pro)--39.727.736.035.216.441.53.028.6
Most Recent Memory21.113.727.530.224.320.110.532.410.125.6
Retrieval-Augmented Memory22.315.638.033.431.829.715.640.811.529.3
3DLLM-MEM (Ours)37.632.162.840.641.439.926.358.218.837.3
+ +(b) Results on all tasks in 3DMEM-BENCH. Average success rate is reported for embodied tasks. Nav. stands for long-term object navigation. We report accuracy score for open-ended EQA evaluation and follow the standard LLM-as-judge evaluation protocol by prompting Gemini. Evaluation details are provided in Appendix E. + +Table 2: Comparison with 3D memory models and standard memory management approaches. Our model, 3DLLM-MEM, achieves the best performance across embodied, EQA and captioning tasks. + +3DLLM-MEM in §4.2, along with qualitative results. Finally, in §4.3, we conduct an ablation study of key design choices in 3DLLM-MEM, demonstrating the effectiveness of our proposed memory fusion mechanism. + +# 4.1 Experimental Setup + +Implementation details We implement our model based on LLaVA-3D (Zhu et al., 2024), modifying it to be compatible with Google TPUs with PyTorch/XLA frameworks (Paszke et al., 2019; team, 2017-2025). We first expand the model's context window to 8192 tokens to accommodate long-term memory inputs. We then fine-tune our proposed memory module along with the LLM decoder using our training split, initializing from LLaVA-3D's pretrained weights. Training is conducted on 8 Google Cloud TPU v5p cores with a batch size of 256. Our model is trained using supervised fine-tuning (SFT) with a standard language modeling loss. More details are provided in Appendix D. + +Baselines We compare 3DLLM-MEM against a broad range of memory management approaches: + +- Everything in Context. For a small subset of scenes, it is feasible to fit all observations directly into the model's context window. +- Most Recent Memory. Since retaining all observations in context is infeasible, we keep only the most recent observations, assuming they are most relevant to the current task. +- Retrieval-Augmented Memory. Inspired by retrieval-based techniques, we adopt a memory bank that stores past observations. During inference, the most relevant memory entries are retrieved and appended before the working memory to augment reasoning. +- 3D-LLM (Hong et al., 2023b). A popular 3D LLM recognized by the community. We finetune it on our training data and report its performance using the "everything in context" strategy with the longest context window supported. Further details are provided in Appendix G. +- 3D-Mem (Yang et al., 2025b). A framework designed for 3D scene memory in embodied exploration and reasoning. However, this method does not support embodied interaction or action execution. + +# 4.2 Experimental Results + +Results on embodied tasks As shown in Table 2a, 3DLLM-MEM significantly outperforms all existing approaches on both in-domain and in-the-wild embodied tasks. Notably, while the performance of other methods drops sharply in the in-the-wild setting, our method demonstrates strong generalization capabilities with a average success rate of $32.1\%$ . 3D-LLM showcases the lowest + +![](images/cb982ef7248cf7c2663290dc2a042dfa17b69d961b9cfbc004bb2dfddd85f7ce.jpg) +Figure 4: Qualitative example of 3DLLM-MEM, which maintains and utilizes a long-term memory to complete the task. Detailed task execution trajectory can be found in Figure 6. + +performance even under simple task settings, highlighting the necessity of incorporating an explicit memory module. Both the Most Recent Memory and Retrieval-Augmented Memory (RAG) baselines perform poorly in this setting, with RAG showing only a slight improvement, highlighting the challenges of retrieving relevant episodic memory. Interestingly, the Everything in Context baseline performs better than both recent memory and RAG approaches, suggesting that when all information can fit within the context window, the model can effectively utilize it. However, 3DLLM-MEM still outperforms Everything in Context, indicating the benefits of selectively fusing task-relevant memory features to better guide embodied reasoning and execution. As task complexity increases from simple to hard, all existing approaches degrade significantly, achieving only $\sim 5\%$ success rate in hard in-the-wild tasks. In contrast, 3DLLM-MEM maintains a strong performance of $27.8\%$ , demonstrating its scalability and effectiveness in managing longer-term memory representations. + +Results on long-term EQA and captioning As shown in Table 2b, 3DLLM-MEM consistently outperforms all existing approaches across all tasks in our benchmark. Notably, 3D-LLM achieves the second-best performance on the captioning task, highlighting its strong ability to summarize object-centric semantic memory. However, due to limited context length, it performs poorly on the EQA tasks, which require long-term spatial-temporal reasoning. 3D-Mem demonstrates improved performance in EQA over other baseline approaches. However, it falls short on spatial relation, navigation and object counting tasks, indicating the limitation of relying solely on aggregated image-centric memories. 3DLLM-MEM significantly outperforms both Most Recent Memory and RAG Memory, which further demonstrates the effectiveness of our memory fusion technique. + +Qualitative results We provide qualitative examples in Figure 4 and a more detailed version with explanations in Figure 6 (Appendix H), demonstrating that 3DLLM-MEM is capable of maintaining long-term memory and executing complex tasks in embodied environments. + +# 4.3 Ablation Study + +Our approach initializes the fused memory using working memory features, aiming to fuse the most relevant memories for the current task. We ablate several design choices for initializing the fusion query, as shown in Table 3. When using either the most recent episodic memory or learnable zero parameters, performance degrades compared to our proposed method. Interestingly, using the most recent memory outperforms zero initialization in the simple setting but underperforms in the hard setting. One possible explanation is that recent memory initialization encourages fusion with nearby observations, which may be sufficient for simple tasks and leads to faster convergence. In contrast, zero initialization is guided solely by training supervision to learn which memories are most useful. In summary, the ablation results demonstrate that initializing fusion queries with working memory tokens provides the most effective and robust design choice for long-term memory fusion. + +# 5 Related Works + +3D Large Language Models 3D Large Language Models (3D-LLMs) have demonstrated promising results across a wide variety of tasks, including 3D scene understanding, object detection, and segmentation (Hong et al., 2023b; Zhou et al., 2024; Huang et al., 2024a; Chen et al., 2024b; Xu et al., 2025a). In parallel, 3D embodied agents have expanded these capabilities to planning and action in interactive environments (Brohan et al., 2023; Huang et al., 2024b; Chen et al., 2024a; Black et al., + +
ModelSimpleMediumHardAverage
In-domainIn-the-wildIn-domainIn-the-wildIn-domainIn-the-wildIn-domainIn-the-wild
SRSub-SRSRSub-SRSRSub-SRSRSub-SRSRSub-SRSRSub-SRSRSub-SRSRSub-SR
3DLLM-MEM45.573.437.065.436.867.831.657.430.546.227.842.137.662.532.155.0
Init with Most Recent Episodic Memory42.369.428.650.732.458.623.745.122.637.815.331.432.455.322.542.4
Init with Learnable Zero Parameters41.467.227.950.033.059.223.445.824.240.418.635.632.955.623.343.8
+ +Table 3: Ablation study of query initialization designs in our memory fusion module. + +2024). Yet, existing models face significant challenges when performing long-horizon embodied tasks in densely populated 3D environments that require reasoning over long-term spatial-temporal memory. To address this, we propose an explicit memory module inspired by the structure of human implicit and explicit memory. Our model employs a memory fusion mechanism that efficiently retrieves and learns task-relevant information, resulting in enhanced performance on complex embodied tasks. + +Long-term Embodied Trajectories Embodied AI simulators (Chang et al., 2017; Kolve et al., 2017; Szot et al., 2021; Shen et al., 2021) have fostered the development of embodied AI agents. Grounded in these environments, some existing benchmarks focus on high-level planning tasks, typically involving short trajectories that can often be completed within single-room settings, thereby requiring minimal spatial-temporal memory (Shridhar et al., 2020, 2021; Li et al., 2024a; Szot et al., 2024; Li et al., 2024b; Yang et al., 2025a). Other benchmarks emphasize long-term scene exploration with extended trajectories, but are primarily centered around navigation tasks and often lack embodied interaction support (Deitke et al., 2020; Ramakrishnan et al., 2021; Krantz et al., 2022; Khanna et al., 2024). To bridge this gap, we introduce 3DMEM-BENCH, a benchmark specifically designed to evaluate long-horizon task execution that requires rich spatial-temporal memory and full embodied task support, as summarized in Table 1. + +Embodied Question Answering Benchmark Embodied Question Answering (EQA) benchmarks (Das et al., 2018; Wijmans et al., 2019; Yu et al., 2019) have been developed to advance goal-driven agents that can perceive their environment. Some EQA benchmarks also include embodied memory QA evaluation, such as OpenEQA (Majumdar et al., 2024), which includes an episodic memory QA split, and Yang et al. (2024), which focuses on spatial memory QA. In contrast, our benchmark, 3DMEM-BENCH jointly targets both spatial and episodic memory, especially their changes over time, while also supporting embodied action tasks, EQA and captioning. For specific comparison on EQA, our long-term memory EQA tasks are designed to require reasoning over multiple "pieces" of memory and their changes across time and space. Additionally, we consider the agent's location in the scene at the moment of answering each question during evaluation. + +Memory System Memory is a fundamental component of AI systems, with early work in the context of LLM agents that utilize memory for decision-making in web-based and sandbox environments (Shinn et al., 2023; Zhang et al., 2023; Packer et al., 2023; Zhang et al., 2024). Most existing approaches construct an experience pool or memory bank and focus on improving the retrieval of useful past information (Zhao et al., 2024; Gao et al., 2024; Xu et al., 2025b). In the computer vision domain, temporal memory has been studied extensively in video understanding and generation tasks (Wang et al., 2021; Diao et al., 2025), while spatial memory has been applied to scene-level visual understanding and 3D reconstruction (Wang and Agapito, 2024; Zou et al., 2025). Recent work such as 3D-Mem (Yang et al., 2025b) has investigated 3D scene memory for exploration and reasoning by prompting vision-language models. In contrast, our work focuses on dense 3D memory representations that are critical for real-world embodied scenarios, where task execution depends heavily on maintaining and reasoning over long-term spatial-temporal memory. + +# 6 Conclusion + +In this work, we introduce 3DMEM-BENCH, a comprehensive benchmark containing fine-grained long-term memory embodied tasks—ranging from simple to hard—along with question-answering tasks that target memory changes across time and space, and captioning task in complex 3D environments. We propose 3DLLM-MEM, an embodied 3D-LLM with novel memory fusion approach for spatial-temporal reasoning, planning, and acting. One limitation of our model is that currently 3DLLM-MEM does not involve low-level navigation and control policy, but utilizes high-level pre + +defined policies in simulator for carrying out the actions. We think that such aspects are orthogonal to our study, and could be explored and seamlessly integrated into our framework in the future. + +# Acknowledgments and Disclosure of Funding + +We thank anonymous reviewers and other members of UCLA-NLP+ group for their helpful comments. This work was partially supported by U.S. DARPA ECOLE Program No. HR00112390060, ONR grant N00014-23-1-2780, Amazon Research Award, and a Google gift fund. Peng and Chang have financial COI with Google and Amazon and were supported in part by a grant from DARPA to the Simons Institute for the Theory of Computing. + +# References + +Kevin Black, Noah Brown, Danny Driess, Adnan Esmail, Michael Equi, Chelsea Finn, Niccolo Fusai, Lachy Groom, Karol Hausman, Brian Ichter, Szymon Jakubczak, Tim Jones, Liyiming Ke, Sergey Levine, Adrian Li-Bell, Mohith Mothukuri, Suraj Nair, Karl Pertsch, Lucy Xiaoyang Shi, James Tanner, Quan Vuong, Anna Walling, Haohuan Wang, and Ury Zhilinsky. 2024. $\pi_0$ : A vision-language-action flow model for general robot control. 8 +Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Xi Chen, Krzysztof Choromanski, Tianli Ding, Danny Driess, Avinava Dubey, Chelsea Finn, Pete Florence, Chuyuan Fu, Montse Gonzalez Arenas, Keerthana Gopalakrishnan, Kehang Han, Karol Hausman, Alex Herzog, Jasmine Hsu, Brian Ichter, Alex Irpan, Nikhil Joshi, Ryan Julian, Dmitry Kalashnikov, Yuheng Kuang, Isabel Leal, Lisa Lee, Tsang-Wei Edward Lee, Sergey Levine, Yao Lu, Henryk Michalewski, Igor Mordatch, Karl Pertsch, Kanishka Rao, Krista Reymann, Michael Ryoo, Grecia Salazar, Pannag Sanketi, Pierre Sermanet, Jaspiar Singh, Anikait Singh, Radu Soricut, Huong Tran, Vincent Vanhoucke, Quan Vuong, Ayzaan Wahid, Stefan Welker, Paul Wohlhart, Jialin Wu, Fei Xia, Ted Xiao, Peng Xu, Sichun Xu, Tianhe Yu, and Brianna Zitkovich. 2023. Rt-2: Vision-language-action models transfer web knowledge to robotic control. volume abs/2307.15818. 8 +Eduardo Camina and Francisco Güell. 2017. The neuroanatomical, neurophysiological and psychological basis of memory: Current models and their origins. Frontiers in Pharmacology, 8:438. 3 +Angel Chang, Angela Dai, Thomas Funkhouser, Maciej Halber, Matthias Niessner, Manolis Savva, Shuran Song, Andy Zeng, and Yinda Zhang. 2017. Matterport3d: Learning from rgb-d data in indoor environments. International Conference on 3D Vision (3DV). 9 +Boyuan Chen, Zhuo Xu, Sean Kirmani, Brian Ichter, Dorsa Sadigh, Leonidas J. Guibas, and Fei Xia. 2024a. Spatialvlm: Endowing vision-language models with spatial reasoning capabilities. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 14455-14465. IEEE. 8 +Yilun Chen, Shuai Yang, Haifeng Huang, Tai Wang, Ruiyuan Lyu, Runsen Xu, Dahua Lin, and Jiangmiao Pang. 2024b. Grounded 3d-llm with referent tokens. ArXiv preprint, abs/2405.10370. 8 +Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Yunxuan Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, et al. 2022. Scaling instruction-finetuned language models. ArXiv preprint, abs/2210.11416. 24 +Abhishek Das, Samyak Datta, Georgia Gkioxari, Stefan Lee, Devi Parikh, and Dhruv Batra. 2018. Embodied question answering. In 2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018, pages 1-10. IEEE Computer Society. 9 +Matt Deitke, Winson Han, Alvaro Herrasti, Aniruddha Kembhavi, Eric Kolve, Roozbeh Mottaghi, Jordi Salvador, Dustin Schwenk, Eli VanderBilt, Matthew Wallingford, Luca Weihs, Mark Yatskar, and Ali Farhadi. 2020. Robothor: An open simulation-to-real embodied AI platform. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2020, Seattle, WA, USA, June 13-19, 2020, pages 3161-3171. IEEE. 9 + +Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. 2023. Objverse: A universe of annotated 3d objects. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, pages 13142-13153. IEEE. 4 +Xingjian Diao, Chunhui Zhang, Weiyi Wu, Zhongyu Ouyang, Peijun Qing, Ming Cheng, Soroush Vosoughi, and Jiang Gui. 2025. Temporal working memory: Query-guided segment refinement for enhanced multimodal understanding. 9 +Gary N Friedman, Luke Johnson, and Zachary M Williams. 2018. Long-term visual memory and its role in learning suppression. Frontiers in Psychology, 9:1896. 3 +Jinglong Gao, Xiao Ding, Yiming Cui, Jianbai Zhao, Hepeng Wang, Ting Liu, and Bing Qin. 2024. Self-evolving gpt: A lifelong autonomous experiential learner. 9 +Qiao Gu, Ali Kuwajerwala, Sacha Morin, Krishna Murthy Jatavallabhula, Bipasha Sen, Aditya Agarwal, Corban Rivera, William Paul, Kirsty Ellis, Rama Chellappa, et al. 2024. Conceptgraphs: Open-vocabulary 3d scene graphs for perception and planning. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 5021-5028. IEEE. 2 +Ziyu Guo, Renrui Zhang, Xiangyang Zhu, Yiwen Tang, Xianzheng Ma, Jiaming Han, Kexin Chen, Peng Gao, Xianzhi Li, Hongsheng Li, and Pheng-Ann Heng. 2023. Point-bind & point-llm: Aligning point cloud with multi-modality for 3d understanding, generation, and instruction following. 2 +Yining Hong, Chunru Lin, Yilun Du, Zhenfang Chen, Joshua B. Tenenbaum, and Chuang Gan. 2023a. 3d concept learning and reasoning from multi-view images. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, pages 9202-9212. IEEE. 21 +Yining Hong, Haoyu Zhen, Peihao Chen, Shuhong Zheng, Yilun Du, Zhenfang Chen, and Chuang Gan. 2023b. 3d-llm: Injecting the 3d world into large language models. In Advances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, NeurIPS 2023, New Orleans, LA, USA, December 10 - 16, 2023. 2, 4, 7, 8, 24 +Yining Hong, Zishuo Zheng, Peihao Chen, Yian Wang, Junyan Li, and Chuang Gan. 2024. Multiply: A multisensory object-centric embodied large language model in 3d world. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 26396-26406. IEEE. 4 +Haifeng Huang, Yilun Chen, Zehan Wang, Rongjie Huang, Runsen Xu, Tai Wang, Luping Liu, Xize Cheng, Yang Zhao, Jiangmiao Pang, and Zhou Zhao. 2024a. Chat-scene: Bridging 3d scene and large language models with object identifiers. In Advances in Neural Information Processing Systems 38: Annual Conference on Neural Information Processing Systems 2024, NeurIPS 2024, Vancouver, BC, Canada, December 10 - 15, 2024. 8 +Jiangyong Huang, Silong Yong, Xiaojian Ma, Xiongkun Linghu, Puhao Li, Yan Wang, Qing Li, Song-Chun Zhu, Baoxiong Jia, and Siyuan Huang. 2024b. An embodied generalist agent in 3d world. In *Forty-first International Conference on Machine Learning*, ICML 2024, Vienna, Austria, July 21-27, 2024. OpenReview.net. 2, 8 +Physical Intelligence, Kevin Black, Noah Brown, James Darpinian, Karan Dhabalia, Danny Driess, Adnan Esmail, Michael Equi, Chelsea Finn, Niccolo Fusai, Manuel Y. Galliker, Dibya Ghosh, Lachy Groom, Karol Hausman, Brian Ichter, Szymon Jakubczak, Tim Jones, Liyiming Ke, Devin LeBlanc, Sergey Levine, Adrian Li-Bell, Mohith Mothukuri, Suraj Nair, Karl Pertsch, Allen Z. Ren, Lucy Xiaoyang Shi, Laura Smith, Jost Tobias Springenberg, Kyle Stachowicz, James Tanner, Quan Vuong, Homer Walke, Anna Walling, Haohuan Wang, Lili Yu, and Ury Zhilinsky. 2025. $\pi_{0.5}$ : a vision-language-action model with open-world generalization. 2 +Mukul Khanna, Ram Ramrakhya, Gunjan Chhablani, Sriram Yenamandra, Théophile Gervet, Matthew Chang, Zsolt Kira, Devendra Singh Chaplot, Dhruv Batra, and Roozbeh Mottaghi. 2024. Goat-bench: A benchmark for multi-modal lifelong navigation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 16373-16383. IEEE. 9 + +Eric Kolve, Roozbeh Mottaghi, Winson Han, Eli VanderBilt, Luca Weihs, Alvaro Herrasti, Daniel Gordon, Yuke Zhu, Abhinav Gupta, and Ali Farhadi. 2017. AI2-THOR: An Interactive 3D Environment for Visual AI. arXiv. 9 +Jacob Krantz, Stefan Lee, Jitendra Malik, Dhruv Batra, and Devendra Singh Chaplot. 2022. Instance-specific image goal navigation: Training embodied agents to find object instances. 9 +Chengshu Li, Ruohan Zhang, Josiah Wong, Cem Gokmen, Sanjana Srivastava, Roberto Martín-Martín, Chen Wang, Gabrael Levine, Wensi Ai, Benjamin Martinez, Hang Yin, Michael Lingelbach, Minjune Hwang, Ayano Hiranaka, Sujay Garlanka, Arman Aydin, Sharon Lee, Jiankai Sun, Mona Anvari, Manasi Sharma, Dhruva Bansal, Samuel Hunter, Kyu-Young Kim, Alan Lou, Caleb R Matthews, Ivan Villa-Renteria, Jerry Huayang Tang, Claire Tang, Fei Xia, Yunzhu Li, Silvio Savarese, Hyowon Gweon, C. Karen Liu, Jiajun Wu, and Li Fei-Fei. 2024a. Behavior-1k: A human-centered, embodied ai benchmark with 1,000 everyday activities and realistic simulation. 3, 9 +Manling Li, Shiyu Zhao, Qineng Wang, Kangrui Wang, Yu Zhou, Sanjana Srivastava, Cem Gokmen, Tony Lee, Li Erran Li, Ruohan Zhang, Weiyu Liu, Percy Liang, Li Fei-Fei, Jiayuan Mao, and Jiajun Wu. 2024b. Embodied agent interface: Benchmarking llms for embodied decision making. In Advances in Neural Information Processing Systems 38: Annual Conference on Neural Information Processing Systems 2024, NeurIPS 2024, Vancouver, BC, Canada, December 10 - 15, 2024. 9 +Xiao Liu, Tianjie Zhang, Yu Gu, Iat Long Long, Yifan Xu, Xixuan Song, Shudan Zhang, Hanyu Lai, Xinyi Liu, Hanlin Zhao, et al. 2024. Visualagentbench: Towards large multimodal models as visual foundation agents. ArXiv preprint, abs/2408.06327. 3 +Arjun Majumdar, Anurag Ajay, Xiaohan Zhang, Pranav Putta, Sriram Yenamandra, Mikael Henaff, Sneha Silwal, Paul McVay, Oleksandr Maksymets, Sergio Arnaud, Karmesh Yadav, Qiyang Li, Ben Newman, Mohit Sharma, Vincent-Pierre Berges, Shiqi Zhang, Pulkit Agrawal, Yonatan Bisk, Dhruv Batra, Mrinal Kalakrishnan, Franziska Meier, Chris Paxton, Alexander Sax, and Aravind Rajeswaran. 2024. Openea: Embodied question answering in the era of foundation models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 16488-16498. IEEE. 9 +Charles Packer, Sarah Wooders, Kevin Lin, Vivian Fang, Shishir G. Patil, Ion Stoica, and Joseph E. Gonzalez. 2023. Memgpt: Towards llms as operating systems. 9 +Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Florian Gribonval, Rafal Jozefowicz, et al. 2019. Pytorch. https://pytorch.org/. 7, 23 +Santhosh Kumar Ramakrishnan, Aaron Gokaslan, Erik Wijmans, Oleksandr Maksymets, Alexander Clegg, John M Turner, Eric Undersander, Wojciech Galuba, Andrew Westbury, Angel X Chang, Manolis Savva, Yili Zhao, and Dhruv Batra. 2021. Habitat-matterport 3d dataset (HM3d): 1000 large-scale 3d environments for embodied AI. volume abs/2109.08238. 4, 9 +Bokui Shen, Fei Xia, Chengshu Li, Roberto Martin-Martin, Linxi Fan, Guanzhi Wang, Claudia Pérez-D'Arpino, Shyamal Buch, Sanjana Srivastava, Lyne P. Tchapmi, Micael E. Tchapmi, Kent Vainio, Josiah Wong, Li Fei-Fei, and Silvio Savarese. 2021. igibson 1.0: a simulation environment for interactive tasks in large realistic scenes. In 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), page accepted. IEEE. 9 +Noah Shinn, Federico Cassano, Ashwin Gopinath, Karthik Narasimhan, and Shunyu Yao. 2023. Reflexion: language agents with verbal reinforcement learning. In Advances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, NeurIPS 2023, New Orleans, LA, USA, December 10 - 16, 2023. 9 +Mohit Shridhar, Jesse Thomason, Daniel Gordon, Yonatan Bisk, Winson Han, Roozbeh Mottaghi, Luke Zettlemoyer, and Dieter Fox. 2020. ALFRED: A benchmark for interpreting grounded instructions for everyday tasks. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2020, Seattle, WA, USA, June 13-19, 2020, pages 10737-10746. IEEE. 9 + +Mohit Shridhar, Xingdi Yuan, Marc-Alexandre Côté, Yonatan Bisk, Adam Trischler, and Matthew J. Hausknecht. 2021. Alfworld: Aligning text and embodied environments for interactive learning. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021. OpenReview.net. 3, 9 +Andrew Szot, Alexander Clegg, Eric Undersander, Erik Wijmans, Yili Zhao, John Turner, Noah Maestre, Mustafa Mukadam, Devendra Singh Chaplot, Oleksandr Maksymets, Aaron Gokaslan, Vladimir Vondrus, Sameer Dharur, Franziska Meier, Wojciech Galuba, Angel X. Chang, Zsolt Kira, Vladlen Koltun, Jitendra Malik, Manolis Savva, and Dhruv Batra. 2021. Habitat 2.0: Training home assistants to rearrange their habitat. In Advances in Neural Information Processing Systems 34: Annual Conference on Neural Information Processing Systems 2021, NeurIPS 2021, December 6-14, 2021, virtual, pages 251-266. 4, 9 +Andrew Szot, Max Schwarzer, Harsh Agrawal, Bogdan Mazoure, Rin Metcalf, Walter Talbott, Natalie Mackraz, R. Devon Hjelm, and Alexander T. Toshev. 2024. Large language models as generalizable policies for embodied tasks. In The Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024. OpenReview.net. 4, 9 +Gemini Team, Rohan Anil, Sebastian Borgeaud, Yonghui Wu, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, et al. 2023. Gemini: a family of highly capable multimodal models. ArXiv preprint, abs/2312.11805. 4 +XLA team. 2017-2025. Xla: Optimizing compiler for machine learning. https://www.tensorflow.org/xla.7,23 +Hao Wang, Weining Wang, and Jing Liu. 2021. Temporal memory attention for video semantic segmentation. In 2021 IEEE International Conference on Image Processing (ICIP), pages 2254-2258. IEEE. 9 +Hengyi Wang and Lourdes Agapito. 2024. 3d reconstruction with spatial memory. ArXiv preprint, abs/2408.16061. 9 +Zixuan Wang, Bo Yu, Junzhe Zhao, Wenhao Sun, Sai Hou, Shuai Liang, Xing Hu, Yinhe Han, and Yiming Gan. 2024. Karma: Augmenting embodied ai agents with long-and-short term memory systems. 2 +Erik Wijmans, Samyak Datta, Oleksandr Maksymets, Abhishek Das, Georgia Gkioxari, Stefan Lee, Irfan Essa, Devi Parikh, and Dhruv Batra. 2019. Embodied question answering in photorealistic environments with point cloud perception. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019, pages 6659-6668. Computer Vision Foundation / IEEE. 9 +Quanting Xie, So Yeon Min, Pengliang Ji, Yue Yang, Tianyi Zhang, Aarav Bajaj, Ruslan Salakhutdinov, Matthew Johnson-Roberson, and Yonatan Bisk. 2024. Embodied-rag: General non-parametric embodied memory for retrieval and generation. 2 +Runsen Xu, Xiaolong Wang, Tai Wang, Yilun Chen, Jiangmiao Pang, and Dahua Lin. 2025a. Pointllm: Empowering large language models to understand point clouds. In Computer Vision - ECCV 2024, pages 131-147. Springer Nature Switzerland. 2, 8 +Wujiang Xu, Zujie Liang, Kai Mei, Hang Gao, Juntao Tan, and Yongfeng Zhang. 2025b. A-mem: Agentic memory for llm agents. ArXiv preprint, abs/2502.12110.9 +Jihan Yang, Shusheng Yang, Anjali W. Gupta, Rilyn Han, Li Fei-Fei, and Saining Xie. 2024. Thinking in Space: How Multimodal Large Language Models See, Remember and Recall Spaces. ArXiv preprint, abs/2412.14171. 9 +Rui Yang, Hanyang Chen, Junyu Zhang, Mark Zhao, Cheng Qian, Kangrui Wang, Qineng Wang, Teja Venkat Koripella, Marziyeh Movahedi, Manling Li, Heng Ji, Huan Zhang, and Tong Zhang. 2025a. Embodiedbench: Comprehensive benchmarking multi-modal large language models for vision-driven embodied agents. 3, 4, 9 + +Yuncong Yang, Han Yang, Jiachen Zhou, Peihao Chen, Hongxin Zhang, Yilun Du, and Chuang Gan. 2025b. 3d-mem: 3d scene memory for embodied exploration and reasoning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 2, 7, 9, 26 +Licheng Yu, Xinlei Chen, Georgia Gkioxari, Mohit Bansal, Tamara L. Berg, and Dhruv Batra. 2019. Multi-target embodied question answering. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019, pages 6309-6318. Computer Vision Foundation / IEEE. 9 +Danyang Zhang, Lu Chen, Situo Zhang, Hongshen Xu, Zihan Zhao, and Kai Yu. 2023. Large language model is semi-parametric reinforcement learning agent. ArXiv preprint, abs/2306.07929. 9 +Zeyu Zhang, Xiaohe Bo, Chen Ma, Rui Li, Xu Chen, Quanyu Dai, Jieming Zhu, Zhenhua Dong, and Ji-Rong Wen. 2024. A survey on the memory mechanism of large language model based agents. 9 +Andrew Zhao, Daniel Huang, Quentin Xu, Matthieu Lin, Yong-Jin Liu, and Gao Huang. 2024. Expel: LLM agents are experiential learners. In Thirty-Eighth AAAI Conference on Artificial Intelligence, AAAI 2024, Thirty-Sixth Conference on Innovative Applications of Artificial Intelligence, IAAI 2024, Fourteenth Symposium on Educational Advances in Artificial Intelligence, EAAI 2014, February 20-27, 2024, Vancouver, Canada, pages 19632-19642. AAAI Press. 9 +Qingqing Zhao, Yao Lu, Moo Jin Kim, Zipeng Fu, Zhuoyang Zhang, Yecheng Wu, Zhaoshuo Li, Qianli Ma, Song Han, Chelsea Finn, Ankur Handa, Ming-Yu Liu, Donglai Xiang, Gordon Wetzstein, and Tsung-Yi Lin. 2025. Cot-vla: Visual chain-of-thought reasoning for vision-language-action models. 2 +Haoyu Zhen, Xiaowen Qiu, Peihao Chen, Jincheng Yang, Xin Yan, Yilun Du, Yining Hong, and Chuang Gan. 2024. 3d-vla: 3d vision-language-action generative world model. ArXiv preprint, abs/2403.09631. 2 +Junsheng Zhou, Jinsheng Wang, Baorui Ma, Yu-Shen Liu, Tiejun Huang, and Xinlong Wang. 2024. Uni3d: Exploring unified 3d representation at scale. In The Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024. OpenReview.net. 8 +Chenming Zhu, Tai Wang, Wenwei Zhang, Jiangmiao Pang, and Xihui Liu. 2024. Llava-3d: A simple yet effective pathway to empowering Imms with 3d-awareness. ArXiv preprint, abs/2409.18125. 2, 5, 7, 23 +Guillermo Zlotnik and Aaron Vansintjan. 2019. Memory: An extended definition. Frontiers in Psychology, 10:2523. 3 +Xueyan Zou, Yuchen Song, Ri-Zhao Qiu, Xuanbin Peng, Jianglong Ye, Sifei Liu, and Xiaolong Wang. 2025. M3: 3d-spatial multimodal memory. In ICLR. 9 + +# NeurIPS Paper Checklist + +# 1. Claims + +Question: Do the main claims made in the abstract and introduction accurately reflect the paper's contributions and scope? + +Answer:[Yes] + +Justification: We clearly state our main claims in abstract and introduction. + +Guidelines: + +- The answer NA means that the abstract and introduction do not include the claims made in the paper. +- The abstract and/or introduction should clearly state the claims made, including the contributions made in the paper and important assumptions and limitations. A No or NA answer to this question will not be perceived well by the reviewers. +- The claims made should match theoretical and experimental results, and reflect how much the results can be expected to generalize to other settings. +- It is fine to include aspirational goals as motivation as long as it is clear that these goals are not attained by the paper. + +# 2. Limitations + +Question: Does the paper discuss the limitations of the work performed by the authors? + +Answer: [Yes] + +Justification: We discuss the limitations in Section 6. + +Guidelines: + +- The answer NA means that the paper has no limitation while the answer No means that the paper has limitations, but those are not discussed in the paper. +- The authors are encouraged to create a separate "Limitations" section in their paper. +- The paper should point out any strong assumptions and how robust the results are to violations of these assumptions (e.g., independence assumptions, noiseless settings, model well-specification, asymptotic approximations only holding locally). The authors should reflect on how these assumptions might be violated in practice and what the implications would be. +- The authors should reflect on the scope of the claims made, e.g., if the approach was only tested on a few datasets or with a few runs. In general, empirical results often depend on implicit assumptions, which should be articulated. +- The authors should reflect on the factors that influence the performance of the approach. For example, a facial recognition algorithm may perform poorly when image resolution is low or images are taken in low lighting. Or a speech-to-text system might not be used reliably to provide closed captions for online lectures because it fails to handle technical jargon. +- The authors should discuss the computational efficiency of the proposed algorithms and how they scale with dataset size. +- If applicable, the authors should discuss possible limitations of their approach to address problems of privacy and fairness. +- While the authors might fear that complete honesty about limitations might be used by reviewers as grounds for rejection, a worse outcome might be that reviewers discover limitations that aren't acknowledged in the paper. The authors should use their best judgment and recognize that individual actions in favor of transparency play an important role in developing norms that preserve the integrity of the community. Reviewers will be specifically instructed to not penalize honesty concerning limitations. + +# 3. Theory assumptions and proofs + +Question: For each theoretical result, does the paper provide the full set of assumptions and a complete (and correct) proof? + +Answer: [NA]. + +Justification: This paper doesn't introduce new theorems. + +# Guidelines: + +- The answer NA means that the paper does not include theoretical results. +- All the theorems, formulas, and proofs in the paper should be numbered and cross-referenced. +- All assumptions should be clearly stated or referenced in the statement of any theorems. +- The proofs can either appear in the main paper or the supplemental material, but if they appear in the supplemental material, the authors are encouraged to provide a short proof sketch to provide intuition. +- Inversely, any informal proof provided in the core of the paper should be complemented by formal proofs provided in appendix or supplemental material. +- Theorems and Lemmas that the proof relies upon should be properly referenced. + +# 4. Experimental result reproducibility + +Question: Does the paper fully disclose all the information needed to reproduce the main experimental results of the paper to the extent that it affects the main claims and/or conclusions of the paper (regardless of whether the code and data are provided or not)? + +# Answer: [Yes] + +Justification: Yes, we fully disclose all the information, please refer to Section 3 and our experimental setup in Section 4.1 + +# Guidelines: + +- The answer NA means that the paper does not include experiments. +- If the paper includes experiments, a No answer to this question will not be perceived well by the reviewers: Making the paper reproducible is important, regardless of whether the code and data are provided or not. +- If the contribution is a dataset and/or model, the authors should describe the steps taken to make their results reproducible or verifiable. +- Depending on the contribution, reproducibility can be accomplished in various ways. For example, if the contribution is a novel architecture, describing the architecture fully might suffice, or if the contribution is a specific model and empirical evaluation, it may be necessary to either make it possible for others to replicate the model with the ssame dataset, or provide access to the model. In general, releasing code and data is often one good way to accomplish this, but reproducibility can also be provided via detailed instructions for how to replicate the results, access to a hosted model (e.g., in the case of a large language model), releasing of a model checkpoint, or other means that are appropriate to the research performed. +- While NeurIPS does not require releasing code, the conference does require all submissions to provide some reasonable avenue for reproducibility, which may depend on the nature of the contribution. For example +(a) If the contribution is primarily a new algorithm, the paper should make it clear how to reproduce that algorithm. +(b) If the contribution is primarily a new model architecture, the paper should describe the architecture clearly and fully. +(c) If the contribution is a new model (e.g., a large language model), then there should either be a way to access this model for reproducing the results or a way to reproduce the model (e.g., with an open-source dataset or instructions for how to construct the dataset). +(d) We recognize that reproducibility may be tricky in some cases, in which case authors are welcome to describe the particular way they provide for reproducibility. In the case of closed-source models, it may be that access to the model is limited in some way (e.g., to registered users), but it should be possible for other researchers to have some path to reproducing or verifying the results. + +# 5. Open access to data and code + +Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material? + +# Answer: [Yes] + +Justification: We will release our code and data publicly after the review process. We also provide data sample and example code in the submitted supplemental material. + +# Guidelines: + +- The answer NA means that paper does not include experiments requiring code. +- Please see the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details. +- While we encourage the release of code and data, we understand that this might not be possible, so "No" is an acceptable answer. Papers cannot be rejected simply for not including code, unless this is central to the contribution (e.g., for a new open-source benchmark). +- The instructions should contain the exact command and environment needed to run to reproduce the results. See the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details. +- The authors should provide instructions on data access and preparation, including how to access the raw data, preprocessed data, intermediate data, and generated data, etc. +- The authors should provide scripts to reproduce all experimental results for the new proposed method and baselines. If only a subset of experiments are reproducible, they should state which ones are omitted from the script and why. +- At submission time, to preserve anonymity, the authors should release anonymized versions (if applicable). +- Providing as much information as possible in supplemental material (appended to the paper) is recommended, but including URLs to data and code is permitted. + +# 6. Experimental setting/details + +Question: Does the paper specify all the training and test details (e.g., data splits, hyperparameters, how they were chosen, type of optimizer, etc.) necessary to understand the results? + +# Answer: [Yes] + +Justification: Please refer to our implementation details in Section 4.1 and Appendix D. + +# Guidelines: + +- The answer NA means that the paper does not include experiments. +- The experimental setting should be presented in the core of the paper to a level of detail that is necessary to appreciate the results and make sense of them. +- The full details can be provided either with the code, in appendix, or as supplemental material. + +# 7. Experiment statistical significance + +Question: Does the paper report error bars suitably and correctly defined or other appropriate information about the statistical significance of the experiments? + +# Answer: [No] + +Justification: We conduct experiments on Google TPU and the training of long-term memory 3D-LLM is expensive, we don't have the resources to run the experiments multiple times and calculate the error bar. + +# Guidelines: + +- The answer NA means that the paper does not include experiments. +- The authors should answer "Yes" if the results are accompanied by error bars, confidence intervals, or statistical significance tests, at least for the experiments that support the main claims of the paper. +- The factors of variability that the error bars are capturing should be clearly stated (for example, train/test split, initialization, random drawing of some parameter, or overall run with given experimental conditions). +- The method for calculating the error bars should be explained (closed form formula, call to a library function, bootstrap, etc.) + +- The assumptions made should be given (e.g., Normally distributed errors). +- It should be clear whether the error bar is the standard deviation or the standard error of the mean. +- It is OK to report 1-sigma error bars, but one should state it. The authors should preferably report a 2-sigma error bar than state that they have a $96\%$ CI, if the hypothesis of Normality of errors is not verified. +- For asymmetric distributions, the authors should be careful not to show in tables or figures symmetric error bars that would yield results that are out of range (e.g. negative error rates). +- If error bars are reported in tables or plots, The authors should explain in the text how they were calculated and reference the corresponding figures or tables in the text. + +# 8. Experiments compute resources + +Question: For each experiment, does the paper provide sufficient information on the computer resources (type of compute workers, memory, time of execution) needed to reproduce the experiments? + +Answer: [Yes] + +Justification: Please refer to our implementation details in Section 4.1 and Appendix D. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The paper should indicate the type of compute workers CPU or GPU, internal cluster, or cloud provider, including relevant memory and storage. +- The paper should provide the amount of compute required for each of the individual experimental runs as well as estimate the total compute. +- The paper should disclose whether the full research project required more compute than the experiments reported in the paper (e.g., preliminary or failed experiments that didn't make it into the paper). + +# 9. Code of ethics + +Question: Does the research conducted in the paper conform, in every respect, with the NeurIPS Code of Ethics https://neurips.cc/public/EthicsGuidelines? + +Answer: [Yes] + +Justification: We discuss ethics concern and broader impact in Appendix A. + +Guidelines: + +- The answer NA means that the authors have not reviewed the NeurIPS Code of Ethics. +- If the authors answer No, they should explain the special circumstances that require a deviation from the Code of Ethics. +- The authors should make sure to preserve anonymity (e.g., if there is a special consideration due to laws or regulations in their jurisdiction). + +# 10. Broader impacts + +Question: Does the paper discuss both potential positive societal impacts and negative societal impacts of the work performed? + +Answer: [Yes] + +Justification: We discuss broader impact in Appendix A. + +Guidelines: + +- The answer NA means that there is no societal impact of the work performed. +- If the authors answer NA or No, they should explain why their work has no societal impact or why the paper does not address societal impact. +- Examples of negative societal impacts include potential malicious or unintended uses (e.g., disinformation, generating fake profiles, surveillance), fairness considerations (e.g., deployment of technologies that could make decisions that unfairly impact specific groups), privacy considerations, and security considerations. + +- The conference expects that many papers will be foundational research and not tied to particular applications, let alone deployments. However, if there is a direct path to any negative applications, the authors should point it out. For example, it is legitimate to point out that an improvement in the quality of generative models could be used to generate deepfakes for disinformation. On the other hand, it is not needed to point out that a generic algorithm for optimizing neural networks could enable people to train models that generate Deepfakes faster. +- The authors should consider possible harms that could arise when the technology is being used as intended and functioning correctly, harms that could arise when the technology is being used as intended but gives incorrect results, and harms following from (intentional or unintentional) misuse of the technology. +- If there are negative societal impacts, the authors could also discuss possible mitigation strategies (e.g., gated release of models, providing defenses in addition to attacks, mechanisms for monitoring misuse, mechanisms to monitor how a system learns from feedback over time, improving the efficiency and accessibility of ML). + +# 11. Safeguards + +Question: Does the paper describe safeguards that have been put in place for responsible release of data or models that have a high risk for misuse (e.g., pretrained language models, image generators, or scraped datasets)? + +Answer: [Yes] + +Justification: We require users to follow the guidelines such as Llama's guidelines when release the model. + +# Guidelines: + +- The answer NA means that the paper poses no such risks. +- Released models that have a high risk for misuse or dual-use should be released with necessary safeguards to allow for controlled use of the model, for example by requiring that users adhere to usage guidelines or restrictions to access the model or implementing safety filters. +- Datasets that have been scraped from the Internet could pose safety risks. The authors should describe how they avoided releasing unsafe images. +- We recognize that providing effective safeguards is challenging, and many papers do not require this, but we encourage authors to take this into account and make a best faith effort. + +# 12. Licenses for existing assets + +Question: Are the creators or original owners of assets (e.g., code, data, models), used in the paper, properly credited and are the license and terms of use explicitly mentioned and properly respected? + +Answer: [Yes] + +Justification: We properly credited the original owners and followed their license. + +# Guidelines: + +- The answer NA means that the paper does not use existing assets. +- The authors should cite the original paper that produced the code package or dataset. +- The authors should state which version of the asset is used and, if possible, include a URL. +- The name of the license (e.g., CC-BY 4.0) should be included for each asset. +- For scraped data from a particular source (e.g., website), the copyright and terms of service of that source should be provided. +- If assets are released, the license, copyright information, and terms of use in the package should be provided. For popular datasets, paperswithcode.com/datasets has curated licenses for some datasets. Their licensing guide can help determine the license of a dataset. +- For existing datasets that are re-packaged, both the original license and the license of the derived asset (if it has changed) should be provided. + +- If this information is not available online, the authors are encouraged to reach out to the asset's creators. + +# 13. New assets + +Question: Are new assets introduced in the paper well documented and is the documentation provided alongside the assets? + +Answer: [Yes]. + +Justification: We document our new introduced benchmark in Section 2.1 and Section 2.2. We also provide the documentation in our dataset along with the data sample in the supplementary material. + +Guidelines: + +- The answer NA means that the paper does not release new assets. +- Researchers should communicate the details of the dataset/code/model as part of their submissions via structured templates. This includes details about training, license, limitations, etc. +- The paper should discuss whether and how consent was obtained from people whose asset is used. +- At submission time, remember to anonymize your assets (if applicable). You can either create an anonymized URL or include an anonymized zip file. + +# 14. Crowdsourcing and research with human subjects + +Question: For crowdsourcing experiments and research with human subjects, does the paper include the full text of instructions given to participants and screenshots, if applicable, as well as details about compensation (if any)? + +Answer: [NA] + +Justification: We authors conducted the human validation with clear instructions, no other research with external human subjects is performed. + +Guidelines: + +- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. +- Including this information in the supplemental material is fine, but if the main contribution of the paper involves human subjects, then as much detail as possible should be included in the main paper. +- According to the NeurIPS Code of Ethics, workers involved in data collection, curation, or other labor should be paid at least the minimum wage in the country of the data collector. + +# 15. Institutional review board (IRB) approvals or equivalent for research with human subjects + +Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals (or an equivalent approval/review based on the requirements of your country or institution) were obtained? + +Answer: [NA] + +Justification: This paper does not involve crowdsourcing nor research with human subjects. + +Guidelines: + +- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. +- Depending on the country in which research is conducted, IRB approval (or equivalent) may be required for any human subjects research. If you obtained IRB approval, you should clearly state this in the paper. +- We recognize that the procedures for this may vary significantly between institutions and locations, and we expect authors to adhere to the NeurIPS Code of Ethics and the guidelines for their institution. + +- For initial submissions, do not include any information that would break anonymity (if applicable), such as the institution conducting the review. + +# 16. Declaration of LLM usage + +Question: Does the paper describe the usage of LLMs if it is an important, original, or non-standard component of the core methods in this research? Note that if the LLM is used only for writing, editing, or formatting purposes and does not impact the core methodology, scientific rigorousness, or originality of the research, declaration is not required. + +Answer: [NA]. + +Justification: We only used LLMs for improving grammar in paper writing. + +Guidelines: + +- The answer NA means that the core method development in this research does not involve LLMs as any important, original, or non-standard components. +- Please refer to our LLM policy (https://neurips.cc/Conferences/2025/LLM) for what should or should not be described. + +# A Broader Impact + +The deployment and release of 3DLLM-MEM carry both potential benefits and risks. These considerations include visual aspects as well as common issues found in existing LLMs like Alpaca and Vicuna. Since 3DLLM-MEM is built on LLaMA, Vicuna, and CLIP, it inherits certain challenges associated with LLMs and vision encoders. Below, we outline the risks and the mitigation strategies implemented for the release of this model. + +Hallucination Similar to other LLMs, 3DLLM-MEM might produce outputs that are not based on factual information or input data. This raises concerns about the accuracy of inferences, particularly in critical applications such as medical fields. + +Biases Biases present in the base models can be brought to 3DLLM-MEM, stemming from both the vision encoder (CLIP) and the language decoder (LLaMA / Vicuna). This may result in biased outcomes or unfair representations of diverse content. + +Energy Consumption We train our model on our training data split which contains about 26K trajectories. The training time only takes less than one day, which makes energy consumption not a primary concern. + +# B Environment Construction + +To support navigation-centric interaction, the agent requires precise knowledge of two things: the traversable layout of each scene and the exact locations of all movable objects. Following 3D-CLR (Hong et al., 2023a), we build this spatial substrate from HM3D's richly annotated indoor scans. We rely on the semantic surface mesh that accompanies each scene to calculate the room and objects' locations. The mesh scan has a unique (24-bit) hexadecimal color for every surface triangle that provides a semantic table that links each color to a surface label (e.g., floor or ceiling) and a room identifier. + +We first derive an axis-aligned bounding box for every room in each HM3D scene. We query the semantic table to retrieve, for every room identifier, the hex colors assigned to its floor and ceiling surfaces. Before processing individual rooms we aggregate the minimum heights of all floor surfaces in the entire scan, producing a global set of candidate floor elevations. For each room we then load the point clouds of its floor and ceiling. When both are available, the vertical bounds are given by the floor's lowest point and the ceiling's highest point. If the floor is missing, the lower bound is set to the highest global floor elevation that still lies below the room's ceiling; if the ceiling is missing, the upper bound defaults to the highest point in the available cloud (i.e., the surface immediately above the floor). Rooms for which both surfaces are absent are discarded because no reliable vertical evidence is present. Horizontal limits are simply the minimum and maximum coordinates of the + +room's floor and ceiling points, and together these six coordinates constitute the room's axis-aligned bounding box. + +Based on the bounding boxes for each room, we calculate the bounding boxes of objects in that room by gathering the corresponding vertices and computing the bounding boxes of the object by taking the minimum and maximum coordinates along the three axes. We merge room-level and object-level bounding boxes by their shared room index. After the aforementioned processing, the dataset comprises 182 distinct 3-D scenes covering a total of 2,602 rooms. + +# C Embodied Task Data Examples + +In this section, we provide complete embodied task trajectories. + +Here is an example of in-the-wild task: + +Task: Rearrange desk items by size. The items to be rearranged are: vase, box, and printer. + +# Trajectory: + +"I'm in room 10. I will start exploring the room for the items I need.", "I see a printer(0) on the desk(0). I also see a box(0) on the cabinet(0)." "I need to remember their locations and sizes for comparison later." + +" ", + +"Now I'm in room 11. I'm exploring this room for the items.", "I don't see a vase in the room, but I see a printer(0) on a rack(1)." "I will remember these objects and their locations." + +" ", + +"Now I'm in room 6. I'm exploring this room.", "I do not see any of the items I am looking for.", "", + +"I'm in room 8. I am exploring this room.", "I see flower vase(0) on the floor. I will remember the object and its location.", + +" ", + +"I'm in room 7. I am exploring this room.", "I do not see any of the items I am looking for.", "", + +"I'm in room 9. I am exploring this room.", "I do not see any of the items I am looking for.", "", + +"Now I am in room 5. I am exploring this room.", "I do not see any of the items I am looking for.", "", + +"Now I am in room 2. I am exploring this room.", "I see a box(0) on the armchair(1). I will remember these objects and their locations.", + +" ", + +"I'm in room 1. I do not see any items that are part of the rearrangement.", "", + +"Now I'm in room 12. I will start exploring this room.", "I think I've seen the vase, printer, and box in other rooms, so I'll head back to room 8 to pick up the vase first.", "", + +"I need to pick up flower vase(0) in this room first. I pick up flower vase(0)." +"PICK UP flower vase(0) from room(8) in room(8)>", +"I'm carrying flower vase(0) to room 10 where I saw printer and box.", +"GO TO ROOM(10)>", + +"Now I am in room 10. I will put flower vase(0) on the desk(0) for now so I can pick up box(0)." +"PUT DOWN flower vase(0) from room(8) on desk(0) in room(10)>", + +"Now I will pick up box(0) from cabinet(0)." + +"PICK UP box(0) from room(10) in room(10)>", + +"Now I will put box(0) on the desk(0) so I can pick up printer(0)." + +"PUT DOWN box(0) from room(10) on desk(0) in room(10)>", + +"Now I will pick up printer(0) from beside the desk(0)." + +"PICK UP printer(0) from room(10) in room(10)>", + +"I need to compare the sizes of the vase, box and printer now to determine the rearranging order.", "I recall that printer(0) in this room had the largest bounding box dimensions based on my earlier observation. I need to put down printer(0).", + +"PUT DOWN printer(0) from room(10) on desk(0) in room(10)>", + +"Now I will pick up box(0)." + +"Pick UP box(0) from room(10) in room(10)>", + +"I remember that box(0) from cabinet(0) had medium bounding box dimensions, so I will put it on the desk.", + +"PUT DOWN box(0) from room(10) on desk(0) in room(10)>", + +"Now I will pick up flower vase(0)." + +"PICK UP flower vase(0) from room(8) in room(10)>", + +"I remember that flower vase(0) had the smallest bounding box dimensions, so I will put it on the desk.", + +"PUT DOWN flower vase(0) from room(8) on desk(0) in room(10)>", + +"I will now arrange the objects from small to large: flower vase(0), box(0) and printer(0) on the desk(0)." "Now I pick up the medium object, box(0)." + +"Pick UP box(0) from room(10) in room(10)>", + +"I will put the medium object on the left of the largest object, printer(0)." + +"PUT DOWN box(0) from room(10) on desk(0) in room(10)>", + +"Now I pick up the smallest object, flower vase(0)." + +"PICK UP flower vase(0) from room(8) in room(10)>", + +"I will put the smallest object on the left of the medium object, box(0)." + +"PUT DOWN flower vase(0) from room(8) on desk(0) in room(10)>", + +"The objects are now arranged from small to large: flower vase(0), box(0) and printer(0) on the desk(0). Task Complete." + +Here is an example of in-domain task: + +Task: Preparing for Cooking Stir-fried Tomato and Scrambled Eggs. You need to prepare the ingredients, seasonings and cooking tools. + +Trajectory: "I'm at room 4. It seems to be living room, I need to prepare ingredients and seasonings and cooking tools and setup them in the kitchen", "There is seasonings on the table.", + +"Pick UP seasonings(0) from room(4) in room(4)>", + +"No other things found for my task", + +" " + +"I'm at room 5. It seems to be dining room. I saw an apron on dining table.", "I need the apron but I have seasonings(0) on my hand. Lets find kitchen to put the seasonings first and go back for apron", "", + +"I'm at room 8. It seems to be kitchen with two cabinets. I found tomatoes and eggs on the cabinet(1) that is far away from me and cooking pan in the sink.", "There are two countertops in this room. countertop(1) is closer to stove(0). I should setup on countertop(1)." + +"PUT DOWN seasonings(0) from room(4) on countertop(1) in room(8)>", + +"Pick up tomatoes(0) from room(8) in room(8)>", + +"PUT DOWN tomatoes(0) from room(8) on countertop(1) in room(8)>", + +"Pick UP eggs(0) from room(8) in room(8)>", + +"PUT DOWN eggs(0) from room(8) on countertop(1) in room(8)>", + +"PICK UP cooking pan(0) from room(8) in room(8)>", + +"PUT DOWN cooking pan(0) from room(8) on stove(0) in room(8)>", + +"I remember I saw apron in room 5. I need to find it back.", + +" $< \mathrm{GO}$ TO ROOM(5)>", + +"PICK UP apron(0) from room(5) in room(5)>", + +" $<$ GO TO ROOM(8)>", + +"PUT DOWN apron(0) from room(5) on countertop(1) in room(8)>", + +"The setup in kitchen has been prepared. Task Complete." + +# D Implementation Details + +We implement our model based on LLaVA-3D (Zhu et al., 2024), modifying it to be compatible with Google TPUs with PyTorch/XLA frameworks (Paszke et al., 2019; team, 2017-2025). We first expand the model's context window to 8192 tokens to accommodate long-term memory inputs. We then fine-tune our proposed memory module along with the LLM decoder using our training split, initializing from LLaVA-3D's pretrained weights. Training is conducted on 8 Google Cloud TPU v5p cores with a batch size of 256 for 1000 steps, which takes about 1 day to complete. We use Adam optimizer with learning rate of 2e-5 with no weight decay. Additionally, we apply a linear warmup + +of the learning rate during the initial $3\%$ steps, increasing from $10^{-8}$ to $10^{-5}$ , followed by a cosine decay scheduler. + +# E Prompts for Gemini + +As mentioned in § 2.2, we prompt Gemini to generate the long-term trajectories as illustrated in Table 4, generate the question-answering tasks as shown in Table 5, and generate caption tasks as shown in Table 6. For open-ended QA evaluation, we followed standard LLM-as-judge protocol by prompting Gemini as illustrated in Table 7. + +# F Data Validation + +# F.1 Trajectory Validation + +We implement a trajectory simulation pipeline driven by the commands listed in Table 4. For each command, the simulator records the agent's current room and the full set of objects it is holding, then updates the set of objects in each room to reflect pick-up and put-down actions. A pick-up removes the specified object (along with any nested items) from the room the agent occupies and adds it to the agent's hand; a put-down removes the object from the agent's hand and places it into the designated room. The pipeline validates each command based on these criteria: (1) the agent's location; (2) the referenced object and (3) the correctness of pick-up and put-down actions. For location validation, a command is marked as invalid if the agent attempts to pick up an object from a room that does not match its current room, or tries to drop an object into a room other than the one it currently occupies. Additionally, if the agent tries to visit a room that does not exist in the scene, or attempts to enter a new room when all rooms have already been explored, the trajectory is also considered invalid. For object validation, a pick-up command is invalid if the target object does not exist in the current room, and a put-down command is invalid if the agent is not currently holding the specified object. For pick-up and put-down validation, the agent is allowed to hold only one object at a time. A command is considered invalid if the agent attempts to pick up an object while already holding one, or tries to put down an object when its hand is empty. Finally, after all commands have been executed, if the trajectory ends with the agent still holding an object that was never put down, the entire trajectory is marked as invalid. + +# F.2 Human Validation + +As mentioned in §2.3 After automatic trajectory validation, we further conduct human validation, in which four student experts in the field manually inspect each benchmark example. We render multi-view images of the entire scene using the simulator and verify whether the benchmark annotations accurately correspond to the simulated environment as illustrated in Figure 5. + +# G Evaluation Setup Details + +3D-LLM Similar to the 3D-LLM work (Hong et al., 2023b), we use their direct reconstruction method to extract the 3D features from each scene in our training data. To process our long-term memory data, which requires multi-scene input across each task, we feed each room in the task through the 3D-LLm Q-Former head independently to get separate 32-token dense representation of each room with per-room 3d positional embeddings injected into the features. Then we concatenate the representations before feeding the input into the frozen t5-flanxl (Chung et al., 2022) backbone like the original work. + +The 3D-LLM model also included learned location tokens used to describe certain locations within each room in the scene. To fit 3D-LLM to our task data, we substitute the location tokens with our specific interaction tokens (eg. used by all models in our experiments) and train the model to learn the new tokens to stay consistent with our higher level interaction used across our training data. Analysis of the 3D-LLM model evaluation output, indicated the primary struggle for the model was retaining long term memory of semantic observations in the scene, so we prioritized aligning 3D-LLM with the high level long-term memory representation in our data over low level spatial understanding of the scene. + +# System message + +You are an AI assistant and task generator for a 3D embodied agent operating in a multi-room environment. The environment provides detailed object instance information, including bounding boxes and IDs. Your goal is to generate a complex task that requires the agent to explore multiple rooms, navigate, and crucially use long-term memory to recall details observed earlier. + +# Prompt + +1. Environment and Object Information + +Object Representation: Each object is given with a bounding box in the format: "(num): [x_min, y_min, z_min], [x_max, y_max, z_max] Here, (num) indicates the ID, with (0) being the closest to the origin [0,0,0]. IDs reset for each room (e.g., sofa(0) in room 2 and sofa(0) in room 4 if each room has one sofa). + +Actions Available: : Navigate to a new, unexplored room (and unlock its objects). Do not use this for rooms that have been visited before. $ (num): [x min, y min, z min], [x max, y max, z max] with units of meters, and each represents left-bottom corner and the right-top corner coordinate. + +You will also receive a trajectory composed of the following tokens and reasoning chains. : Pick up an object that originally belongs to a specific room while in that same room. : Place an object (that originally belongs to a room) onto another object (such as a table or floor) in a room. : which navigates to a new room you haven't explored and unlocks objects there. + +This trajectory is what the agent have executed over the past. You need to propose several questions and answers that focused on the reasoning abilities of the long-term memory of the agent. These reasoning questions should focus on what have changed temporally or spatially in this agent's memory. It's important that this change challenged the agent's memory. For example the questions should contain object counting, spatial relation, comparison between objects across rooms, long-term multi-room room layout, long-term multi-room object navigation. Remember spatial memory is important, you should design questions that asked about the 3D object spatial relation and layout in the room that need the agent to perform a hard reasoning for the final answer. + +For clarity, consider these examples: {In-context examples} + +Here is the scene information: {Input scene information} + +Here is the agent's trajectory: {Input agent's trajectory} + +Table 5: Prompt template for generate QA data. {In-context examples} are in-context examples. +{Input scene information} are scene, room and object semantics along with their bounding boxes. +{Input agent's trajectory} is the 3D agent's explored trajectories and action chains. + +For finetuning on our data, we use the hyperparameters provided by 3D-LLM and finetune until model loss stops decreasing. Due to compute limitations, we trained on captioning task for 15 epochs, question-answering task for 20 epochs, and allocated most of the compute time on the embodied task, which we trained on for 75 epochs. + +3D-Mem We benchmark 3D-Mem (Yang et al., 2025b) on the question-answering and captioning splits of 3DMEM-BENCH. 3D-Mem is a snapshot-based 3D memory architecture originally developed for embodied exploration and reasoning; it keeps two complementary stores—memory snapshots, a compact set of multi-view RGB-D frames with per-object bounding boxes summarizing the areas the agent has inspected, and frontier snapshots, boundary views that suggest where useful new information may be found next. In its native setting the agent navigates an unfamiliar scene by selecting the frontier view most likely to advance its task and then answers visual questions using the most relevant memory snapshots. Because our evaluation focuses on post-exploration reasoning rather than active exploration, we disable the frontier component and retain only the memory snapshots. For these two tasks, the system will capture memory snapshots in each room from the room center, and finish the QA and captioning base on the memory snapshots of all the explored rooms. + +# H Qualitative Examples + +We provide qualitative examples as shown in Figure 6. It demonstrates that 3DLLM-MEM can maintain a long-term memory and perform complex tasks in the embodied environments. More examples can be found in the supplementary materials. + +# Prompt + +You are provided with a scene description containing multiple rooms. Each room includes a list of objects along with their positions in the room, represented by bounding boxes. Each object's bounding box is defined by a 3D coordinate in the format: (num): [x min, y min, z min],[x max, y max, z max] with units in meters (defining the left-bottom and right-top corners). Your task is to generate an object caption for each room in the form of a coherent, descriptive paragraph that conveys the 3D spatial arrangement and relative positions of all objects within that room. + +Then, you will receive the object descriptions and caption for the current 3D room you are in. You will also be provided with the previous rooms' captions as well. Your task is to generate new captions covering the summarization of the common features across all rooms based on your current room and important difference based on your current room. The reasons of generating the new caption is to help the agent to remind of what are in previous rooms memories can help the agent in this current room. The past objects and observations should be related to current room by examining the summarization of common things and differences. For clarity, consider these examples: {In-context examples} + +Here is the scene information: {Input scene information} + +Here is current room you are in and previous rooms you went: {Input agent's location} + +Table 6: Prompt template for generate QA data. {In-context examples} are in-context examples. {Input scene information} are scene, room and object semantics along with their bounding boxes. {Input agent's location} is the location for current room in the scene and the past explored rooms. + +# System message + +Please act as an impartial judge and evaluate the quality of the response provided by an AI assistant to the user question. Your evaluation should consider correctness and helpfulness. You will be given a reference answer and the assistant's answer. You evaluation should focus on the assistant's answer to the second question. Begin your evaluation by comparing the assistant's answer with the reference answer. Identify and correct any mistakes. Be as objective as possible. After providing your explanation, you must rate the response on a scale of 1 to 10 by strictly following this format: "[rating]", for example: "Rating: [[5]]". + +# Prompt + +```txt +<|The Start of Reference Answer|> + ### User: + question_1 + ### Reference answer: + ref_answer_1 + ### User: + question_2 + ### Reference answer: + ref_answer_2 + + +``` + +Table 7: Prompt template for open-ended QA evaluation following standard LLM-as-judge protocol. + +![](images/9ee560098b84ed80ec0f0e1735f16571e6d1e27fc3d78b05083e6131bef20531.jpg) +Room 8 + +![](images/c4cb279be287459bf9cbff6c552380d7ac2f40d4119445592ad3594188f71e95.jpg) + +![](images/6b8e46774391fe96cd9b3c0a7c1874c664047ba23e66655090a0bee17ef9dad3.jpg) + +![](images/0c0f3d2f3daaedea8a9468d0c2d14137e6ed2b8f90c357231b6b944d8f7178bf.jpg) + +![](images/2459167ec3b9a5c4f146c16a83865e234205297e0d5d6dddc498b651d6d4b306.jpg) +Room 9 + +![](images/b3229a59ba663d59ee4ab31c2900ecf50e9555884ce6435065b7b8d286240c10.jpg) +Figure 5: Example of human annotators manually check the data quality on QA and captioning tasks through multiple rendered multi-view images from each room. + +![](images/73c18d306694874ef57abfc7b0aff4e0dc8ed843e7d52d4d7750e9505e1805ea.jpg) + +![](images/4fd45325b9f8ad4d1c75942721b5f7de37a58058bc725926d6de20e3e5c1ffc5.jpg) + +# Object Captioning + +Room 8 contains distinct sleeping, working, and lounging zones. Bed(0) extends along one side, with pillow(0) near its head and cabinet(1) beside it, which supports decoration(0) and is flanked by book(0) and book(1). Near the foot of bed(0), yoga mat(0), bag(0), and box(0) form a small activity area. At the far end, wardrobe(0) stands near wall(4), next to beanbag chair(0) and exercise ball(0), with clothes(0) draped nearby, creating a relaxed corner. The office area, positioned across from bed(0), includes desk(0), office chair(0), and cabinet(0), with trashcan(0) nearby and ceiling lamp(0) overhead. Window(0) on the parapet(0) brings in natural light. The arrangement ensures functional use of space while maintaining clear separation between different activities. + +# Question Answering + +# Object Counting + +Q: How many chairs are there in Room 8? +A: There are 2 chairs in total: one office chair(0) and one beanbag chair(0). + +# Spatial Relation + +- Q: If you are sitting on the office chair(0) working at the desk(0) in Room 8, where is the window(0) relative to you? +A: The window(0) is behind you to the left. + +# Comparative QA + +Q: Which room has more books, Room 8 or Room 9? +- A: Room 8 has more books. It contains two books, book(0) and book(1), both placed on top of the cabinet(0) near the desk(0). + +# Object Navigation + +Q: If you are sitting on the beanbag chair(0) in Room 8, how do you get to the books? +- A: You need to stand up, turn left, and walk across the room toward the cabinet(0) near the desk(0), where the books are placed. + +# Room Layout + +- Q: In Room 8, what is the relative layout of the bed(0), desk(0), and office chair(0)? Which one is closest to the window(0)? +- A: The bed(0) is placed along the side wall near the window(0), the desk(0) is located in the far corner of the room along wall(5), and the office chair(0) is positioned directly in front of the desk(0). Among these, the bed(0) is closest to the window(0), which is embedded in the wall beside it. + +![](images/0f0a569da600b0fff6467c46fb72760ee6a80610b2656ac734a00a63225b75b8.jpg) +(1) + +![](images/8f4b5f9b78dcf12ff8ba04c4bcf85f1e76eab4d47ea3b8e8fe1db0813d89ad8a.jpg) +(2) + +![](images/154b9c07b192017b0d2b0fe2cc686ffc724f5963b553593642fe422e02479a68.jpg) +(3) + +![](images/9f70166c25f1c5c85c0631e9572c060536f08f841108d7e7b9b30790981ad80e.jpg) +(4) + +![](images/3210665f4d1625f5ea18f0a05ab84b2ce2862ad64e1b2efcbe75e6c74c12f689.jpg) +(5) + +![](images/10fdcada2f77900be37e9a2f3519161627186b8068b5217b9ff1d62dc433eaab.jpg) +(6) + +![](images/9007f06bc26531d60a68fbefa1fe5548762ba0cfc77c6b6c2c6653bcf636eab9.jpg) +(7) + +![](images/efb6f45207a44876cfc62fe4acc4d1f06f6ab4acddf99b024406a11f994790a2.jpg) +(8) + +![](images/588b4da065022d695ee43248057ac1594fbc11af88c02515b3f4c1cc358bc517.jpg) +(9) +Figure 6: Qualitative example of 3DLLM-MEM. The task instruction is: Prepare a cozy reading nook in the living room with two books and a teacup. In images (1) and (2), the agent explores the environment randomly, forming an initial memory of the scene. After receiving the task instruction, it recalls its memory and navigates to the bedroom to pick up a book from the cabinet, as shown in images (3) and (4). The agent then returns to the living room and places the book on the table in front of the sofa (image 5). Unable to recall any additional books, the agent resumes exploration and finds a second book on the bed, which it picks up (image 6) and stacks on top of the first book (image 7). Finally, the agent recalls seeing a teacup in the kitchen, navigates to retrieve it (image 8), and places it on the table in the living room (image 9). The task is successfully completed. \ No newline at end of file diff --git a/NeurIPS/2025/3DLLM-Mem_ Long-Term Spatial-Temporal Memory for Embodied 3D Large Language Model/images.zip b/NeurIPS/2025/3DLLM-Mem_ Long-Term Spatial-Temporal Memory for Embodied 3D Large Language Model/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..a74992d3fd09c396f6714b91088c7ddf80ad6fbb --- /dev/null +++ b/NeurIPS/2025/3DLLM-Mem_ Long-Term Spatial-Temporal Memory for Embodied 3D Large Language Model/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a478bf992f0d61ab693e1e1959508925f743404468c3f114673d742e517da446 +size 897264 diff --git a/NeurIPS/2025/3DLLM-Mem_ Long-Term Spatial-Temporal Memory for Embodied 3D Large Language Model/layout.json b/NeurIPS/2025/3DLLM-Mem_ Long-Term Spatial-Temporal Memory for Embodied 3D Large Language Model/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..bb4b534d7b355c9519a14a5eeef6405a5cd657d2 --- /dev/null +++ b/NeurIPS/2025/3DLLM-Mem_ Long-Term Spatial-Temporal Memory for Embodied 3D Large Language Model/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:71dda8f9030b0aacf5e23206e9f266c205f99d633edcf0f0ead5d98ed8c5db69 +size 823946 diff --git a/NeurIPS/2025/3DOT_ Texture Transfer for 3DGS Objects from a Single Reference Image/c493dbd3-b439-4fca-ba8f-663b83be5ce4_content_list.json b/NeurIPS/2025/3DOT_ Texture Transfer for 3DGS Objects from a Single Reference Image/c493dbd3-b439-4fca-ba8f-663b83be5ce4_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..68aa0e32642e73cebe2c63864521b91c178a59a7 --- /dev/null +++ b/NeurIPS/2025/3DOT_ Texture Transfer for 3DGS Objects from a Single Reference Image/c493dbd3-b439-4fca-ba8f-663b83be5ce4_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b0d1189c637b0e23c0b0f3329ec634a323eb8da03884d7781c07a8f2e6163224 +size 115998 diff --git a/NeurIPS/2025/3DOT_ Texture Transfer for 3DGS Objects from a Single Reference Image/c493dbd3-b439-4fca-ba8f-663b83be5ce4_model.json b/NeurIPS/2025/3DOT_ Texture Transfer for 3DGS Objects from a Single Reference Image/c493dbd3-b439-4fca-ba8f-663b83be5ce4_model.json new file mode 100644 index 0000000000000000000000000000000000000000..2912f8fc0e1bf2aac15220f9c137ea3e3454f490 --- /dev/null +++ b/NeurIPS/2025/3DOT_ Texture Transfer for 3DGS Objects from a Single Reference Image/c493dbd3-b439-4fca-ba8f-663b83be5ce4_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f6f8576746857429a7165ae3dc40ff0cbac174a0b625d80fe6967148a74ee2be +size 157290 diff --git a/NeurIPS/2025/3DOT_ Texture Transfer for 3DGS Objects from a Single Reference Image/c493dbd3-b439-4fca-ba8f-663b83be5ce4_origin.pdf b/NeurIPS/2025/3DOT_ Texture Transfer for 3DGS Objects from a Single Reference Image/c493dbd3-b439-4fca-ba8f-663b83be5ce4_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..42f271331c1f872b1555d2e00dc6a0b4a26a3dec --- /dev/null +++ b/NeurIPS/2025/3DOT_ Texture Transfer for 3DGS Objects from a Single Reference Image/c493dbd3-b439-4fca-ba8f-663b83be5ce4_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:17d2bac327896dc71a0cd0cee0e9a20d0ebf0cc82cb780bb97a4fe27fae7b388 +size 24403384 diff --git a/NeurIPS/2025/3DOT_ Texture Transfer for 3DGS Objects from a Single Reference Image/full.md b/NeurIPS/2025/3DOT_ Texture Transfer for 3DGS Objects from a Single Reference Image/full.md new file mode 100644 index 0000000000000000000000000000000000000000..096b33cbf0d7cb4bdb6d49763e3bb31efeaf60bd --- /dev/null +++ b/NeurIPS/2025/3DOT_ Texture Transfer for 3DGS Objects from a Single Reference Image/full.md @@ -0,0 +1,589 @@ +# 3DOT: Texture Transfer for 3DGS Objects from a Single Reference Image + +Xiao Cao1 Beibei Lin1 Bo Wang2 Zhiyong Huang1 Robby T. Tan1,3 + +$^{1}$ National University of Singapore $^{2}$ University of Mississippi $^{3}$ ASUS Intelligent Cloud Services + +{xiaocao, beibei.lin}@u.nus.edu + +hawk.rsrch@gmail.com {dcshuang,robby.tan}@nus.edu.sg + +![](images/6d32693f7ed237a9611a49a05b0dde07eea5a277cfad54a16b1c33be87f59c5f.jpg) +Reference Image + +![](images/684db3c5ce70aae52e616ee71d5524dc177507f7bb3c1df0ed52f65dacf0684c.jpg) +Target Scene + +![](images/7cd076101fc600df8c60f7d92437b5f9370bc4b3b5ef3e9526eede0c7faa5d18.jpg) +Plug-n-Play + +![](images/59fb794643b027dd9baf4e501f897f6df7b498d63d9c63335810efa8b5c8ee6e.jpg) +IGS2GS + +![](images/d7dc70b79bbcc3198ae25aa2e3aa2d9a79045b353e13c56d4ef33296f22621a7.jpg) +GaussCtrl + +![](images/2e23fcb89be40a6d05aa2bfa97288ec5805e32afa9b214dd4f8829c12b92cd6b.jpg) +Ours + +![](images/10dd311aba2b88ed7780c06c6198e72af173aaf1b78f714e4366bdf82130063a.jpg) +Figure 1: Comparison of 2D and 3D image-based texture editing methods. Prompts are "moss-covered table" and "pink plastic bear". 2D methods Plug-n-Play [36] suffers from view inconsistency problem; 3D text-driven editing methods IGS2GS [37] and GaussCtrl [41] struggle to preserve texture characteristics. Ours faithfully edit the texture, material appearance, and color. + +![](images/d0f23b5cdd1b6d7b3df025ede5ce9c2fb03ebfeb19b900e17edf55edc5ce92d1.jpg) + +![](images/6489e77cf7259445a92e277e28c21bb740a486b729ff43f0cda0c7fce6d8d7f1.jpg) + +![](images/087cc3799edaa5045a0e7cabb5f827ba5edd9710b0b8d5befa90c67b4dccee54.jpg) + +![](images/36c81416fa7ed36b2b18cc86f1b57db41d25616aa57d1361803c24ce661bbf24.jpg) + +![](images/98a9f97fdc1e085499acf49d1fd5e97331c0f8a0763663d0d628b82674eedce3.jpg) + +# Abstract + +Image-based 3D texture transfer from a single 2D reference image enables practical customization of 3D object appearances with minimal manual effort. Adapted 2D editing and text-driven 3D editing approaches can serve this purpose. However, 2D editing typically involves frame-by-frame manipulation, often resulting in inconsistencies across views, while text-driven 3D editing struggles to preserve texture characteristics from reference images. To tackle these challenges, we introduce 3DOT, a 3D Gaussian Splating Object Texture Transfer method based on a single reference image, integrating: 1) progressive generation, 2) view-consistency gradient guidance, and 3) prompt-tuned gradient guidance. To ensure view consistency, progressive generation starts by transferring texture from the reference image and gradually propagates it to adjacent views. View-consistency gradient guidance further reinforces coherence by conditioning the generation model on feature differences between consistent and inconsistent outputs. To preserve texture characteristics, prompt-tuning-based gradient guidance learns a token that describes differences between original and reference textures, guiding the transfer for faithful texture preservation across views. Overall, 3DOT combines these strategies to achieve effective texture transfer while maintaining structural coherence across viewpoints. Extensive qualitative and quantitative evaluations confirm that our three components enable convincing and effective 2D-to-3D texture transfer. Our project page is available here: https://massyzs.github.io/3DOT_web/. + +# 1 Introduction + +Transferring texture from a 2D image to a 3D object is a valuable yet underexplored capability in 3D editing. It enables efficient texture manipulation and benefits applications such as virtual reality, CG films, and 3D games [2, 31, 40, 23, 28, 39, 34, 5]. Despite advances in 2D texture and 3D editing techniques, transferring texture from a single 2D image to a 3D object remains challenging due to difficulties in ensuring view consistency and preserving texture characteristics, particularly for unseen views beyond the reference image. + +2D image-based editing methods [36, 46, 25, 32, 14, 51, 24, 42] perform texture transfer by finetuning a diffusion model (e.g., DreamBooth [30], Textual Inversion [13]) and editing images rendered from a 3D object to create a finetuning dataset. The resulting 3D object often suffers from view inconsistency and identity loss due to the absence of constraints enforcing multi-view coherence and identity preservation, as shown in Figure 1. 3D editing methods [15, 37, 9, 41, 8, 27, 6], especially text-driven ones, guide editing using prompts derived from reference images via visual language models or manual descriptions. However, these prompts are typically coarse and miss fine-grained features, resulting in identity mismatch and inconsistent appearance across views. + +Motivated by these challenges, we propose 3DOT, a novel framework for transferring texture from a single 2D reference image to a 3D object represented by 3D Gaussian Splatting [21]. 3DOT comprises three key components: 1) a progressive generation process, 2) view-consistency gradient guidance, and 3) prompt-tuning-based gradient guidance. The first two components enforce view consistency, while the third preserves texture characteristics. + +In the progressive generation process, we first obtain reference images either by directly pasting the reference image onto the 3D object or by generating candidate views using a depth-conditioned model [47] based on the unedited view's depth. The image that best matches the target attributes is then selected. To facilitate prompt tuning and sparse cross-attention, we remove backgrounds from both the unedited training images and the reference images, and project them into the latent space for k-step partial diffusion. The generation begins from the reference view and progressively propagates to neighboring views, guided by sparse cross-attention on previously edited views. This strategy maximizes overlap between adjacent reference images to enforce view consistency. + +To enhance view consistency in 3D editing, we introduce view-consistency gradient guidance. The core idea is to guide the diffusion model toward view-consistent generation by minimizing texture inconsistency features in intermediate outputs. Specifically, we initialize two diffusion modules: one conditioned on reference views via cross-attention, and the other guided only by a text prompt. Since cross-attention is the only differing component, the discrepancy between their intermediate results captures view-consistency features. During each denoising step, these features are scaled and injected as gradient guidance, steering the generation toward consistent outputs across views. + +Since the reference image reveals no texture for unseen views, coarse text prompts often lead to inconsistency. To overcome this, we propose prompt-tuning-based gradient guidance that captures texture differences as additional prompt tokens. Specifically, we compute the difference between reference and unedited images in the CLIP feature space [11], encoding the texture transformation direction. This signal is injected into the diffusion denoising process as gradient guidance, enabling consistent texture transfer across views. The fine-tuned prompt improves style coherence in unseen views while preserving details in the reference view. + +We evaluate our method on the face-forwarding [38] and 360-degree [3] datasets. Results show effective texture transfer with fine detail preservation and strong view consistency. Our key contributions: + +- 3DOT, an image-based 3D Gaussian Splitting (3DGS) texture transfer framework that enables efficient and flexible texture editing. +- A progressive generation process with view-consistency gradient guidance to address view inconsistency across novel views. +- Prompt-tuning-based gradient guidance preserves texture characteristics in seen views and enforces style consistency in unseen views. +- Extensive experiments demonstrate that 3DOT achieves state-of-the-art visual quality and quantitative performance. + +![](images/507354dd2ce1aa5dcbe43b797425b275f18ccaee749c5862ca82218e61a45fae.jpg) +Figure 2: 3DOT. Our framework enables texture transfer from a single image to a 3D object. The left panels illustrate the selection of the reference image using a generative approach. Then, our method employs a progressive generation process guided by view-consistency and prompt-tuning-based gradient guidance to preserve both cross-view consistency and texture identity. $\mathbb{R},\mathbb{T}$ , and $\mathbb{T}'$ denote the reference set, text prompt, and learned texture difference token, respectively. + +![](images/f9ac2ddc5600e457cd699d56f05203f33d7435071af968482f9ef65fab3814de.jpg) + +# 2 Related Work + +2D Diffusion-based Editing DragDiffusion [32] defines target edits using keypoints and replaces them with reference images. A-Tale-of-Two-Features [46] combines dense DINO [7] features and sparse diffusion features by merging reference-view semantics with target-view structures. Plug-and-Play [36] refines fine-grained details by injecting diffusion features into DINO features, while DiffEditor [25] improves 2D editing precision via differential equation-based sampling with regional gradient guidance. The most relevant work, SwapAnything [14], leverages DreamBooth [30] and AdaIN [19] to encode source images and maintain style consistency during 2D edits. Although effective for image editing, these methods operate on individual views without enforcing view consistency, highlighting the need for 3D-aware texture editing techniques. + +3D Editing Most 3D editing methods leverage 2D diffusion models for guidance and adopt dataset-updating strategies to finetune pretrained 3D scenes. Instruct-NeRF2NeRF [15] and Instruct-GS2GS [37] use instruct-pix2pix [4] to guide updates for NeRF or Gaussian Splitting. GaussianEditor [9] introduces hierarchical representations for more stable edits under stochastic guidance. Direct Gaussian Editor (DGE) [8] addresses view consistency via epipolar cross-attention, but its initial independent generation introduces artifacts. GaussCtrl [41] injects features from unedited views to preserve consistency, but this can cause the diffusion model to retain original textures, limiting editability. StyleSplat [20] achieves texture edits without a generative model but requires altering the 3D representation, which falls outside our setting of editing a fixed 3D object using a single reference image. Methods that ignore view consistency [15, 37, 9] can be extended with image captioning, while consistency-aware approaches [8, 41] can inject latent reference features during denoising. However, such modifications offer only coarse control. High-quality, identity-preserving edits require more precise and targeted designs. + +# 3 Proposed Method + +Fig. 2 illustrates our 3DOT pipeline, consisting of three key modules: 1) a progressive generation process, 2) view-consistency gradient guidance for enforcing texture coherence across different views, and 3) prompt-tuning-based gradient guidance for preserving object identity. + +To obtain the reference image, we either generate depth-conditioned candidates or extract textures by directly cropping texture into object shape in a certain rendered view. In the generative approach, users select the candidate that best matches the desired attributes. In the texture-based approach, extracted textures are directly mapped onto the object surface. Following [41], both reference and unedited images are encoded into latent space to initialize the denoising process. We then apply + +prompt-tuning to capture texture differences between the reference and the 3D object, guiding diffusion to preserve identity. Edited views are progressively generated, starting from the reference view. The resulting dataset is utilized to finetune 3D Gaussian model and the above procedure is iteratively conducted [15] for smooth results. + +# 3.1 Progressive Generation + +Existing methods [15, 37, 9, 41, 8] struggle to balance view consistency and editing flexibility. For example, GaussCtrl [41] conditions diffusion on unedited images to enforce consistency but often retains original textures, limiting editability. DGE [8] avoids reliance on unedited inputs but edits non-adjacent views, introducing inconsistencies. + +To overcome these limitations, we propose a progressive generation process that removes dependency on unedited images and avoids isolated generation steps, achieving both consistency and flexibility. + +We first generate reference images by conditioning a generative model on depth maps with background masking to ensure geometric alignment. To improve quality, we refine depth maps using dilated and blurred masks to address black-edge artifacts and apply the original mask to remove redundant content from the outputs. + +For a selected reference view $\tau$ and target view $\mathbb{I}_i$ , we construct a sparse reference set $\mathbb{R}_i = \{\mathbb{I}_{\tau}, \mathbb{I}_{i-1}, \mathbb{F}(\mathbb{I})_{\tau}\}$ , excluding backgrounds. Including $\mathbb{I}_{i-1}$ maintains local consistency via minimal angular changes. As edits propagate to distant views, errors accumulate. For symmetric case, we can include $\mathbb{F}(\mathbb{I})_{\tau}$ , a horizontally flipped variant of the reference, to preserve alignment with fewer conditioning views. + +The generative model is conditioned on $\mathbb{R}$ using weighted fused cross-attention: + +$$ +\operatorname {W e i g h t e d A t t n} _ {e} = \lambda \operatorname {A t t n} _ {e, e} + (1 - \lambda) \sum_ {i \in \mathbb {R}} w _ {i} \operatorname {A t t n} _ {e, i}, \tag {1} +$$ + +where $e$ denotes the image that is currently editing, $\mathrm{Attn}_{i,j}$ denotes the attention score between images $i$ and $j$ , and $\lambda$ balances self- and cross-attention. + +Partial denoising (Eq. 3) begins with the reference view and progressively extends to adjacent views. These edited, view-consistent images are then used to finetune the 3D Gaussian model, and the process is repeated iteratively. + +# 3.2 View-Consistency Gradient Guidance + +Existing generative methods [47, 45, 30, 13] rely on many reference views to maintain consistency. In contrast, our progressive generation begins with a single reference and uses only a few views. To enhance cross-attention effectiveness under this constraint, we propose a consistency-aware gradient guidance mechanism inspired by classifier-free guidance [16], modifying the noise estimate [16] to amplify cross-view signals without additional training. + +Given a target view $\mathbb{I}_i$ and reference set $\mathbb{R}_i = \{\mathbb{I}_{\tau},\mathbb{I}_{i - 1},\mathbb{F}(\mathbb{I})_{\tau}\}$ (as in Sec. 3.1), we define the denoising prediction as: + +$$ +\begin{array}{l} \epsilon_ {\theta} ^ {t} (z _ {\lambda}, \mathbb {T}, \mathbb {R}) = \epsilon_ {\hat {\theta}} ^ {t} (z _ {\lambda}) \\ + w _ {\mathbb {T}} \left(\epsilon_ {\theta} ^ {t} \left(z _ {\lambda}, \mathbb {T}, \mathbb {R}\right) - \epsilon_ {\theta} ^ {t} \left(z _ {\lambda}, \mathbb {R}\right)\right) \\ + w _ {\mathbb {R}} \big (\epsilon_ {\theta} ^ {t} (z _ {\lambda}, \mathbb {T}, \mathbb {R}) - \epsilon_ {\hat {\theta}} ^ {t} (z _ {\lambda}, \mathbb {T}) \big), \tag {2} \\ \end{array} +$$ + +where $\mathbb{T}$ is the text prompt, $\theta$ and $\hat{\theta}$ refer to diffusion with and without fused cross-attention, and $w_{\mathbb{T}}, w_{\mathbb{R}}$ are scaling factors. + +We perform partial denoising as: + +$$ +\begin{array}{l} z ^ {(t - 1 | \kappa)} = \sqrt {\alpha_ {t - 1 | \kappa}} \frac {z ^ {t | \kappa} - \sqrt {1 - \alpha_ {t | \kappa}} \cdot \epsilon_ {\theta} ^ {t | \kappa} (z , \mathbb {T} , \mathbb {R})}{\sqrt {\alpha_ {t | \kappa}}} \\ + \sqrt {1 - \alpha_ {t - 1 | \kappa}} \epsilon_ {\theta} ^ {t | \kappa} (z, \mathbb {T}, \mathbb {R}), \tag {3} \\ \end{array} +$$ + +![](images/a640fefaf84f9561e6e2f906f86df7735d28aa61d757ae0806a0e3e09752918a.jpg) +Figure 3: Qualitative comparison on 360-degree scenes (material and color edits): Our 3DOT method faithfully edits 3D objects' texture based on reference images. + +where $t \in [0, \kappa]$ , $\kappa_{\tau} < \kappa_{i \neq \tau}$ , and $\alpha$ is the DDIM scheduler coefficient. The latent input $z^t$ is initialized via: + +$$ +z ^ {t + 1} = \sqrt {\alpha_ {t + 1}} \frac {z ^ {t} - \sqrt {1 - \alpha_ {t}} \cdot \epsilon^ {t}}{\sqrt {\alpha_ {t}}} + \sqrt {1 - \alpha_ {t + 1}} \epsilon^ {t}. \tag {4} +$$ + +In Eq. 2, the second term reflects differences between guidance with and without the unconditional prompt [16], improving adherence to textual instructions. The third term captures variations induced by reference conditioning, and amplifying it strengthens view consistency. This gradient-based mechanism guides generation toward coherent multi-view results, enhancing consistency without additional training overhead. + +# 3.3 Prompt-Tuning-Based Gradient Guidance + +Text prompts provide only coarse control during diffusion, often leading to identity loss and inconsistent texture fidelity. For example, the phrase "stone bear" may yield highly variable textures across different generations. These coarse text descriptions result in view-inconsistent generations and further cause 3D inconsistency. + +Among fine-tuning methods [30, 45, 17, 13], textual inversion [13] learns a custom token to represent object-specific textures but requires multiple images to achieve reasonable quality (see supplementary for single-image results). + +To address this, we introduce prompt-tuning-based gradient guidance, which reduces the need for multiple images while encoding texture differences more effectively. The key idea is to learn a new token that captures the texture discrepancy between the unedited 3D object and the reference image, and to use this token to guide denoising toward the desired style. + +Given the reference image $\mathbb{I}_{\tau}$ and its corresponding unedited rendering $\hat{\mathbb{I}}_{\tau}$ , we compute the texture difference in CLIP feature space: + +$$ +\Delta_ {\hat {\mathbb {I}} _ {\tau} \rightarrow \mathbb {I} _ {\tau}} = \operatorname {C L I P} \left(\hat {\mathbb {I}} _ {\tau}\right) - \operatorname {C L I P} \left(\mathbb {I} _ {\tau}\right). \tag {5} +$$ + +We initialize the text token $\hat{\mathbb{T}}$ using a base prompt (e.g., from a VLM), and optimize it by aligning with the texture difference via: + +$$ +L _ {\mathrm {c l i p}} = \operatorname {c o s i n e} \left(\Delta_ {\hat {\mathbb {I}} _ {\tau} \rightarrow \mathbb {I} _ {\tau}}, \hat {\mathbb {T}}\right). \tag {6} +$$ + +To reduce misalignment between image and text representations in CLIP space, we apply further prompt tuning in the diffusion feature space, following [26]: + +$$ +L _ {\text {d i f f}} = \epsilon_ {\theta} \left(z _ {\lambda}, \mathbb {T} ^ {\prime}, \mathbb {R}\right) - \epsilon_ {\theta} ^ {\prime} \left(z _ {\lambda}, \mathbb {T} ^ {\prime}, \mathbb {R}\right). \tag {7} +$$ + +
MetricsIN2NIGS2GSGaussCtrlDGEOurs
CLIP Score ↑0.89170.89080.86380.85720.9333
Lpips(Alex) ↓0.17080.16830.16920.17130.1166
Lpips(VGG) ↓0.16760.15940.15910.16030.1247
Vision-GPT ↑45.552485476
User study ↑2.03752.43752.37502.00004.5750
+ +Table 1: Quantitative results evaluated by CLIP score, VGG-based and Alex-based LPIPS scores, Vision-GPT and user studies given reference image with rendered edited objects. Bold text refers to the best performance and underlined text refers to the second best performance. Detailed results can be found in Supplementary Material Section 2. + +The fine-tuned token $\mathbb{T}'$ acts as a style-aware prompt enriched by texture differences. While not meaningful in textual form, it encodes critical style information for guiding generation. During inference, we extract and amplify this information at $t$ -step via the difference: + +$$ +\epsilon_ {\theta} ^ {t} (z _ {\lambda}, \mathbb {T} ^ {\prime}, \mathbb {R}) - \epsilon_ {\hat {\theta}} ^ {t} (z _ {\lambda}, \mathbb {T}, \mathbb {R}), +$$ + +and integrate it into the denoising process. The final prediction becomes: + +$$ +\begin{array}{l} \epsilon_ {\theta} ^ {t} (z _ {\lambda}, \mathbb {T}, \mathbb {R}, \mathbb {T} ^ {\prime}) = \epsilon_ {\hat {\theta}} ^ {t} (z _ {\lambda}) \\ + w _ {\mathbb {T}} \left(\epsilon_ {\theta} ^ {t} \left(z _ {\lambda}, \mathbb {T}, \mathbb {R}\right) - \epsilon_ {\theta} ^ {t} \left(z _ {\lambda}, \mathbb {R}\right)\right) \\ + w _ {\mathbb {R}} \left(\epsilon_ {\theta} ^ {t} \left(z _ {\lambda}, \mathbb {T}, \mathbb {R}\right) - \epsilon_ {\theta} ^ {t} \left(z _ {\lambda}, \mathbb {T}\right)\right) \\ + w _ {\mathbb {T} ^ {\prime}} \left(\epsilon_ {\theta} ^ {t} \left(z _ {\lambda}, \mathbb {T} ^ {\prime}, \mathbb {R}\right) - \epsilon_ {\hat {\theta}} ^ {t} \left(z _ {\lambda}, \mathbb {T}, \mathbb {R}\right)\right). \tag {8} \\ \end{array} +$$ + +This term strengthens style consistency across views while preserving fine texture details aligned with the reference. + +# 4 Experiments + +We compare our method with state-of-the-art text-driven editing approaches, including GaussCtrl [41], DGE [8], IGS2GS [37], and IN2N [15]. Since these methods rely on text inputs, we use captioned descriptions as editing prompts to enable image-based 3D texture editing functionality. For quantitative evaluation, we employ AlexNet-based [22] and VGG-based [33] LPIPS scores [48], CLIP score [29], and Vision-GPT score [1], supplemented by user studies. Comparisons are conducted across multiple scenes from different datasets to ensure a comprehensive assessment following [41]. + +# 4.1 Evaluation + +Quantitative For each edit, we compute AlexNet-based and VGG-based LPIPS scores, CLIP score, Vision-GPT score, and conduct user studies, as summarized in Table 1. Detailed per-scene scores are provided in the supplementary material. LPIPS and CLIP scores serve as perceptual evaluation metrics, measuring feature similarity between rendered edited objects and reference images. LPIPS ranges from 0 to 1, with lower values indicating better perceptual quality, while higher CLIP scores are preferred. Vision-GPT assesses the faithfulness of edited textures from the reasoning perspective, scoring from 0 to 100, where higher values indicate better alignment. For user studies, participants are informed of the edited object and required to rate the 3D result on a scale of 1 to 5, with higher scores reflecting better quality. Quantitative results show that our method achieves the highest performance across all metrics. + +Qualitative We present qualitative results of 360-degree dataset in Figs.1, 3 and 4. Figs.3 and 3 includes reference images with texture color or material variations, while Fig.4 features those with complex textures and significant semantic changes. Fig.5 shows the results of "face-forward" case. Our method enables more precise 3D object editing without unintended texture leakage between objects. In the 360-degree color and material editing scenario (e.g., bear and table), IN2N [15] and IGS2GS [37] suffer from incorrect color saturation and inaccurate material representation. In the bear scenarios (Fig.1, 3a), their results are undersaturated, whereas in the table scenarios (Fig.3b, 1), + +![](images/b4ec242d841256490c8c091690ab61d0fe21e135d5923c60189c89bf2278bb8a.jpg) +Figure 4: Qualitative comparison on 360-degree scenes (complicated texture edits): Our 3DOT method successfully edits 3D objects' texture to complicated reference textures. + +![](images/1f76871204b546fc9ce5c8723d8aedf837df84281f5ac102c75b6fb9f71c4de4.jpg) +Figure 5: Qualitative comparison on face-forwarding scenes: Our 3DOT method faithfully edits 3D objects' texture to reference textures and generates the most plausible texture edits for unseen views. + +they are over-saturated. None of the baseline methods accurately reproduces the intended material attributes (i.e., plastic, moss in Fig. 1 and metallic in Fig. 3a). GaussCtrl [41] excessively preserves the original 3D object's appearance, resulting in minimal modifications due to its unedited reference set. Our method effectively edits textures while achieving realistic material appearances, such as + +![](images/6539756c9f5235bf7afdf35ab3889d33019b242e752d7299778067520d989862.jpg) +$w_{\mathbb{T}'} = 0$ + +![](images/ff6dbccc8665cbabdab9ea0f796f6c8a0717fa31ad0e40dc3a1b4cfefbaeabd2.jpg) +$w_{\mathbb{R}} = 0$ + +![](images/fae9a06f2ab0656f759062d369c33ea825774b27412a55fb6b9e11cb7b6dc9e4.jpg) +Ours + +![](images/e3442b411d513654e947ef0e3f785b8e26935d3682d8f84060cee32de278529d.jpg) +Ref. Image + +![](images/3993ed52fb79166e2df4890e1e50bae900cf03e475354bd274f462190814e458.jpg) +View1 +Figure 6: Ablation studies on proposed two gradient guidances. + +![](images/228ee1cb825ba3db750bf8879f24bf6ec3e72dcfe1e44fdda8ec60ada6572c9c.jpg) +(a) Performance when set $w_{\mathbb{T}'}$ and $w_{\mathbb{R}}$ to 0. +View 2 +(b) Using Ref.1 for prompt-tuning guidance and Ref.2 as reference image + +![](images/873e99034be05d43c80107840c372b0a6eb43ce4b59f24b1beba144b56aaa6f8.jpg) +Ref. Image 1 + +![](images/ec571dec8037f262f7d72d19f4d0e51ebaf07238b9e796d667a18568b0d9cd69.jpg) +Ref. Image 2 + +specular highlights on the bear and the lush, velvety moss on the table. In moss-covered table scenario (Fig.1), all 3D baseline methods only attempt to edit texture while ours can also modify geometry to better match the "moss material". + +In the large semantic change editing scenario (Fig.4), baseline methods struggle with significant transformations. DGE [8] often fails, as its edits remain nearly unchanged. Its initial independent editing stage leads to inconsistent results and further causes the epipolar attention mechanism to break down in highly dissimilar views, resulting in minimal overall changes. Our method achieves precise 3D object editing with a texture style that closely matches the reference image, enabled by our proposed prompt-tuning, consistency guidance, and progressive process. Prompt-tuning preserves intricate texture details, while consistency guidance and progressive generation mitigate blurriness from view inconsistency. + +In the face-forward case (Fig.5), our method preserves fine details, such as the black lower eyelid and feather-like cloth in the hawk scenario (Fig.5a). IN2N [15] and IGS2GS [37] generate erroneous results due to their independent diffusion process and full diffusion steps. The independent generation process leads to inconsistent images, while full diffusion steps cause excessive texture changes and identity loss. Finetuning NeRF with these inconsistent and identity-lost images can result in network collapse. For GaussCtrl [41] and DGE [8], particularly in the hawk case, large texture differences break their view-consistency mechanisms, resulting in outputs that retain the original object's appearance instead of the intended modifications. + +# 4.2 Editing Speed + +We compare the editing time of our method with two baselines: the Gaussian Splatting model (GaussCtrl [41]) and a representative NeRF-based model (IN2N [15]). GaussCtrl requires 15min 47s, while IN2N takes 20h 51min 20s. Our approach preserves the efficiency advantage of Gaussian Splatting, with an editing time of 23min 33s, introducing only a modest additional overhead from the incorporation of view-consistency and prompt-tuning-based guidance. In particular, the additional time required by our guidance components is approximately 8min in total. Since our full pipeline typically involves two iterations of image editing, the added cost per iteration is about 4min. We consider this overhead efficient given the improvements in texture fidelity and view consistency achieved by our method. + +
AblationLPIPS (Alex)↓LPIPS (VGGT)↓CLIP Score↑
W/O prompt-tuning (Sec.3.3)0.03550.06790.9203
W/O consistency (Sec.3.2)0.05580.08440.9340
W/O Prog. Gen.0.13300.12900.9160
Ours0.03510.06200.9445
+ +Table 2: Ablation studies: Performance evaluation when removing (1) prompt-tuning guidance, (2) view-consistency guidance, and (3) progressive generation mechanism. + +# 4.3 Ablation Studies + +We evaluate the effectiveness of prompt-tuning-based gradient guidance and view-consistency gradient guidance from both qualitative and quantitative perspectives. We further illustrate the effectiveness of prompt-tuning-based guidance by using two distinct images: one serving as the reference image during image generation, and the other as the reference for prompt tuning. + +Prompt-tuning based Gradient Guidance We first evaluate the effect of removing prompt-tuning-based guidance by setting the guidance scale $w_{\mathbb{T}'} = 0$ , as shown in Table 2 and Figure 6a. Without this guidance, the rendered images exhibit blurry surface highlights, and the distribution direction does not align with the reference image. + +Additionally, we demonstrate the effectiveness of prompt-tuning guidance by using a prompt token trained on reference image Ref-2 to edit the 3D object and setting Ref-1 as the target reference, as illustrated in Figure 6b. Ref-2 depicts a bear characterized by sharp metallic edges, whereas Ref-1 shows a bear with a rusted metallic texture. In Figure 6b, View-1 aligns closely with Ref-1, producing a rendering that closely matches the rusted metal appearance, which demonstrates the effectiveness of the utilized fused cross-attention. Conversely, View-2, which represents an unseen viewpoint without explicit texture guidance, utilizes the learned token to guide the rendering towards the "colorful metallic bear" appearance consistent with Ref-2. This demonstrates the effectiveness of the proposed prompt-tuning-based guidance when editing parts of the 3D object not visible in the reference image. + +View-consistency Gradient Guidance We evaluate the effectiveness of view-consistency guidance by setting $w_{\mathbb{R}} = 0$ , as reported in Table 2 and illustrated in Figure 6a. When this guidance is disabled, the performance significantly degrades, resulting in edited images exhibiting notable undersaturation. This undersaturation primarily arises due to inconsistencies within overlapping regions in the intermediate outputs. These findings underscore the crucial role of view-consistency gradient guidance in maintaining editing quality and color fidelity. + +Progressive Generation To evaluate the effectiveness of the progressive generation, we disable progressive view propagation and perform editing using only the initial reference image. The degraded performance (as shown in Table 2) highlights the importance of propagating texture across neighboring views to enforce view consistency and mitigate artifacts from single-view editing. + +# 4.4 Discussion + +Differences from Multi-View Diffusion While existing multi-view diffusion methods are designed to generate multi-view consistent 3D objects, the task setting, constraints, and diffusion components in our work are fundamentally different. Specifically, 3DOT takes as input a single 2D reference image and a fixed 3D Gaussian Splitting (3DGS) representation, and transfers high-fidelity texture onto this existing geometry. In contrast, multi-view diffusion methods are typically designed to synthesize novel views or reconstruct 3D scenes from a few texture-consistent input images, without anchoring to an explicit 3D representation. These methods rely on implicit geometry learned from priors, which makes them unsuitable for editing tasks that require consistency with a given 3D structure. Moreover, multi-view diffusion approaches assume that texture appearance is consistent across views, an assumption that does not hold in texture transfer scenarios where the reference image and the 3D representation exhibit different textures. Directly applying them in this setting leads to collapsed reconstructions, highlighting the need for a specialized framework such as ours. + +Different Geometric Reference Image Geometry mismatches between the 2D reference image and the 3D object can occur when the reference is generated via a depth-conditioned diffusion model with a small conditioning scale factor. We consider two common scenarios: + +- Slightly Different Geometry: Minor deviations (e.g., slight pose or scale differences, such as a standing bear with legs in a different position) can be effectively handled by 3DOT. (1) During progressive generation, cross-attention extracts texture features while being constrained by the depth map and partially reversed latent features, preserving appearance cues from the original image. (2) During 3D fine-tuning, overlapping regions from adjacent edited views iteratively correct geometric discrepancies and reinforce consistent texture transfer. +- Significantly Different Geometry: When the reference depicts a substantially different shape (e.g., a running bear instead of a standing one), transfer quality may degrade. Such cases usually result from failures in depth-conditioned option generation or from unsuitable user-provided references. In practice, these poor references can be easily identified during the interactive selection stage, and regenerated using our partial denoising strategy with a larger depth control factor. + +In summary, 3DOT is robust to minor geometric mismatches, and mitigates larger discrepancies through reference regeneration and iterative correction during 3D fine-tuning. + +**Limitation** Dark border around edited regions occurs in some cases. We attribute this artifact to language-based object segmentation methods (e.g., LangSAM) to generate masks for isolating objects in the reference views. These segmentation methods often include a narrow band of background pixels near object boundaries due to imperfect boundary localization. As a result, during the diffusion-based editing stage, this narrow band is misinterpreted as valid texture content, leading to the appearance of dark borders in the final renderings even after applying the soften mask. This may be addressed by utilizing a more advanced language-based segmentation method [49, 43, 35], depth as additional information [44, 50, 10] with a semantic 3D representation [18, 12]. The quality of unedited 3D Gaussians also impacts editing performance. Undertrained Gaussian spheres (e.g., floating Gaussians in empty space) degrade rendered images, disrupting the mask generation process. Incorrect segmentation can result in edits with significantly altered geometry, ultimately causing 3D Gaussian collapse. + +# 5 Conclusion + +We introduced 3DOT, a framework for image-based 3DGS texture transfer from a single reference image, an underexplored capability in the 3D editing domain. To enable high-quality and view-consistent texture transfer, we proposed three key components: (1) progressive generation, (2) view-consistency gradient guidance, and (3) prompt-tuning-based gradient guidance. These components effectively address challenges of view-consistency and texture characteristic preservation during transferring process. We evaluated 3DOT across various scenes involving color, material, and large semantic changes. 3DOT consistently outperforms existing baselines, both visually and quantitatively. + +# References + +[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. +[2] Shivangi Aneja, Justus Thies, Angela Dai, and Matthias Nießner. Clipface: Text-guided editing of textured 3d morphable models. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-11, 2023. +[3] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mipnerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5470–5479, 2022. +[4] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18392-18402, 2023. +[5] Xiao Cao, Beibei Lin, Bo Wang, Zhiyong Huang, and Robby T Tan. Ssnerf: Sparse view semi-supervised neural radiance fields with augmentation. arXiv preprint arXiv:2408.09144, 2024. +[6] Xiao Cao, Yuyang Zhao, Robby T Tan, and Zhiyong Huang. Bridging 3d editing and geometry-consistent paired dataset creation for 2d nighttime-to-daytime translation. In [CVPR 2025 Workshop] SyntaGen: 2nd Workshop on Harnessing Generative Models for Synthetic Visual Datasets. +[7] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. +[8] Minghao Chen, Iro Laina, and Andrea Vedaldi. Dge: Direct gaussian 3d editing by consistent multi-view editing. arXiv preprint arXiv:2404.18929, 2024. +[9] Yiwen Chen, Zilong Chen, Chi Zhang, Feng Wang, Xiaofeng Yang, Yikai Wang, Zhongang Cai, Lei Yang, Huaping Liu, and Guosheng Lin. Gaussianeditor: Swift and controllable 3d editing with gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21476-21485, 2024. +[10] Zitan Chen, Zhuang Qi, Xiao Cao, Xiangxian Li, Xiangxu Meng, and Lei Meng. Class-level structural relation modeling and smoothing for visual representation learning. In Proceedings of the 31st ACM International Conference on Multimedia, pages 2964-2972, 2023. +[11] Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2818-2829, 2023. +[12] Shaohui Dai, Yansong Qu, Zheyan Li, Xinyang Li, Shengchuan Zhang, and Liujuan Cao. Training-free hierarchical scene understanding for gaussian splatting with superpoint graphs. arXiv preprint arXiv:2504.13153, 2025. +[13] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. +[14] Jing Gu, Yilin Wang, Nanxuan Zhao, Wei Xiong, Qing Liu, Zhifei Zhang, He Zhang, Jianming Zhang, HyunJoon Jung, and Xin Eric Wang. Swapanything: Enabling arbitrary object swapping in personalized visual editing. arXiv preprint arXiv:2404.05717, 2024. +[15] Ayaan Haque, Matthew Tancik, Alexei A Efros, Aleksander Holynski, and Angjoo Kanazawa. Instruct-nerf2nerf: Editing 3d scenes with instructions. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19740–19750, 2023. + +[16] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. +[17] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. +[18] Chi Huang, Xinyang Li, Shengchuan Zhang, Liujuan Cao, and Rongrong Ji. Nerf-dets: Enhancing multi-view 3d object detection with sampling-adaptive network of continuous nerf-based representation. arXiv e-prints, pages arXiv-2404, 2024. +[19] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In Proceedings of the IEEE international conference on computer vision, pages 1501-1510, 2017. +[20] Sahil Jain, Avik Kuthiala, Prabhdeep Singh Sethi, and Prakanshul Saxena. Stylesplat: 3d object style transfer with gaussian splatting. arXiv preprint arXiv:2407.09473, 2024. +[21] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. +[22] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. Advances in neural information processing systems, 25, 2012. +[23] Xinyang Li, Zhangyu Lai, Linning Xu, Yansong Qu, Liujuan Cao, Shengchuan Zhang, Bo Dai, and Rongrong Ji. Director3d: Real-world camera trajectory and 3d scene generation from text. Advances in Neural Information Processing Systems, 37:75125-75151, 2024. +[24] Beibei Lin, Tingting Chen, and Tan Robby T. Geocomplete: Geometry-aware diffusion for reference-driven image completion. The Thirty-Ninth Annual Conference on Neural Information Processing Systems, 2025. +[25] Chong Mou, Xintao Wang, Jiechong Song, Ying Shan, and Jian Zhang. Diffeditor: Boosting accuracy and flexibility on diffusion-based image editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8488-8497, 2024. +[26] Thao Nguyen, Yuheng Li, Utkarsh Ojha, and Yong Jae Lee. Visual instruction inversion: Image editing via image prompting. Advances in Neural Information Processing Systems, 36, 2024. +[27] Yansong Qu, Dian Chen, Xinyang Li, Xiaofan Li, Shengchuan Zhang, Liujuan Cao, and Rongrong Ji. Drag your gaussian: Effective drag-based editing with score distillation for 3d gaussian splatting. arXiv preprint arXiv:2501.18672, 2025. +[28] Yansong Qu, Shaohui Dai, Xinyang Li, Jianghang Lin, Liujuan Cao, Shengchuan Zhang, and Rongrong Ji. Goi: Find 3d gaussians of interest with an estimizable open-vocabulary semantic-space hyperplane. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 5328-5337, 2024. +[29] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. +[30] Nataniel Ruiz, Yuzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22500-22510, 2023. +[31] Taha Samavati and Mohsen Soryani. Deep learning-based 3d reconstruction: a survey. Artificial Intelligence Review, 56(9):9175-9219, 2023. +[32] Yujun Shi, Chuhui Xue, Jun Hao Liew, Jiachun Pan, Hanshu Yan, Wenqing Zhang, Vincent YF Tan, and Song Bai. Dragdiffusion: Harnessing diffusion models for interactive point-based image editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8839-8849, 2024. + +[33] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. +[34] Weilin Sun, Manyi Li, Peng Li, Xiao Cao, Xiangxu Meng, and Lei Meng. Sequential selection and calibration of video frames for 3d outdoor scene reconstruction. CAAI Transactions on Intelligence Technology, 9(6):1500-1514, 2024. +[35] Lei Tan, Pingyang Dai, Jie Chen, Liujuan Cao, Yongjian Wu, and Rongrong Ji. Partformer: Awakening latent diverse representation from vision transformer for object re-identification. arXiv preprint arXiv:2408.16684, 2024. +[36] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1921-1930, 2023. +[37] Cyrus Vachha and Ayaan Haque. Instruct-gs2gs: Editing 3d gaussian splats with instructions, 2024. +[38] Can Wang, Ruixiang Jiang, Mengei Chai, Mingming He, Dongdong Chen, and Jing Liao. Nerf-art: Text-driven neural radiance fields stylization. IEEE Transactions on Visualization and Computer Graphics, 2023. +[39] Yuze Wang, Junyi Wang, Ruicheng Gao, Yansong Qu, Wantong Duan, Shuo Yang, and Yue Qi. Look at the sky: Sky-aware efficient 3d gaussian splatting in the wild. IEEE Transactions on Visualization and Computer Graphics, 2025. +[40] Yuze Wang, Junyi Wang, Yansong Qu, and Yue Qi. Rip-nerf: Learning rotation-invariant point-based neural radiance field for fine-grained editing and compositing. In Proceedings of the 2023 ACM international conference on multimedia retrieval, pages 125–134, 2023. +[41] Jing Wu, Jia-Wang Bian, Xinghui Li, Guangrun Wang, Ian Reid, Philip Torr, and Victor Adrian Prisacariu. Gaussctrl: multi-view consistent text-driven 3d gaussian splatting editing. arXiv preprint arXiv:2403.08733, 2024. +[42] Yihang Wu, Xiao Cao, Kaixin Li, Zitan Chen, Haonan Wang, Lei Meng, and Zhiyong Huang. Towards better text-to-image generation alignment via attention modulation. In International Conference on Neural Information Processing, pages 332-347. Springer, 2024. +[43] Jiaer Xia, Lei Tan, Pingyang Dai, Mingbo Zhao, Yongjian Wu, and Liujuan Cao. Attention disturbance and dual-path constraint network for occluded person re-identification. In Proceedings of the AAAI conference on artificial intelligence, volume 38, pages 6198–6206, 2024. +[44] Weilong Yan, Ming Li, Haipeng Li, Shuwei Shao, and Robby T Tan. Synthetic-to-real self-supervised robust depth estimation via learning with motion and structure priors. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 21880-21890, 2025. +[45] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. +[46] Junyi Zhang, Charles Herrmann, Junhwa Hur, Luisa Polania Cabrera, Varun Jampani, Deqing Sun, and Ming-Hsuan Yang. A tale of two features: Stable diffusion complements dino for zero-shot semantic correspondence. Advances in Neural Information Processing Systems, 36, 2024. +[47] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. +[48] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. +[49] Xin Zhang and Robby T Tan. Mamba as a bridge: Where vision foundation models meet vision language models for domain-generalized semantic segmentation. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 14527-14537, 2025. + +[50] Xin Zhang, Jinheng Xie, Yuan Yuan, Michael Bi Mi, and Robby T Tan. Heap: unsupervised object discovery and localization with contrastive grouping. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 7323-7331, 2024. +[51] Chenyang Zhu, Kai Li, Yue Ma, Longxiang Tang, Chengyu Fang, Chubin Chen, Qifeng Chen, and Xiu Li. Instantswap: Fast customized concept swapping across sharp shape differences. arXiv preprint arXiv:2412.01197, 2024. + +# NeurIPS Paper Checklist + +# 1. Claims + +Question: Do the main claims made in the abstract and introduction accurately reflect the paper's contributions and scope? + +Answer:[Yes] + +Justification: Please refer to abstract and introduction. + +Guidelines: + +- The answer NA means that the abstract and introduction do not include the claims made in the paper. +- The abstract and/or introduction should clearly state the claims made, including the contributions made in the paper and important assumptions and limitations. A No or NA answer to this question will not be perceived well by the reviewers. +- The claims made should match theoretical and experimental results, and reflect how much the results can be expected to generalize to other settings. +- It is fine to include aspirational goals as motivation as long as it is clear that these goals are not attained by the paper. + +# 2. Limitations + +Question: Does the paper discuss the limitations of the work performed by the authors? + +Answer: [Yes] + +Justification: Please refer to Discussion section + +Guidelines: + +- The answer NA means that the paper has no limitation while the answer No means that the paper has limitations, but those are not discussed in the paper. +- The authors are encouraged to create a separate "Limitations" section in their paper. +- The paper should point out any strong assumptions and how robust the results are to violations of these assumptions (e.g., independence assumptions, noiseless settings, model well-specification, asymptotic approximations only holding locally). The authors should reflect on how these assumptions might be violated in practice and what the implications would be. +- The authors should reflect on the scope of the claims made, e.g., if the approach was only tested on a few datasets or with a few runs. In general, empirical results often depend on implicit assumptions, which should be articulated. +- The authors should reflect on the factors that influence the performance of the approach. For example, a facial recognition algorithm may perform poorly when image resolution is low or images are taken in low lighting. Or a speech-to-text system might not be used reliably to provide closed captions for online lectures because it fails to handle technical jargon. +- The authors should discuss the computational efficiency of the proposed algorithms and how they scale with dataset size. +- If applicable, the authors should discuss possible limitations of their approach to address problems of privacy and fairness. +- While the authors might fear that complete honesty about limitations might be used by reviewers as grounds for rejection, a worse outcome might be that reviewers discover limitations that aren't acknowledged in the paper. The authors should use their best judgment and recognize that individual actions in favor of transparency play an important role in developing norms that preserve the integrity of the community. Reviewers will be specifically instructed to not penalize honesty concerning limitations. + +# 3. Theory assumptions and proofs + +Question: For each theoretical result, does the paper provide the full set of assumptions and a complete (and correct) proof? + +Answer: [NA] + +Justification: No theory assumption. + +# Guidelines: + +- The answer NA means that the paper does not include theoretical results. +- All the theorems, formulas, and proofs in the paper should be numbered and cross-referenced. +- All assumptions should be clearly stated or referenced in the statement of any theorems. +- The proofs can either appear in the main paper or the supplemental material, but if they appear in the supplemental material, the authors are encouraged to provide a short proof sketch to provide intuition. +- Inversely, any informal proof provided in the core of the paper should be complemented by formal proofs provided in appendix or supplemental material. +- Theorems and Lemmas that the proof relies upon should be properly referenced. + +# 4. Experimental result reproducibility + +Question: Does the paper fully disclose all the information needed to reproduce the main experimental results of the paper to the extent that it affects the main claims and/or conclusions of the paper (regardless of whether the code and data are provided or not)? + +Answer: [Yes] + +Justification: Please refer to implementation detail and code upon acceptance. + +# Guidelines: + +- The answer NA means that the paper does not include experiments. +- If the paper includes experiments, a No answer to this question will not be perceived well by the reviewers: Making the paper reproducible is important, regardless of whether the code and data are provided or not. +- If the contribution is a dataset and/or model, the authors should describe the steps taken to make their results reproducible or verifiable. +- Depending on the contribution, reproducibility can be accomplished in various ways. For example, if the contribution is a novel architecture, describing the architecture fully might suffice, or if the contribution is a specific model and empirical evaluation, it may be necessary to either make it possible for others to replicate the model with the same dataset, or provide access to the model. In general, releasing code and data is often one good way to accomplish this, but reproducibility can also be provided via detailed instructions for how to replicate the results, access to a hosted model (e.g., in the case of a large language model), releasing of a model checkpoint, or other means that are appropriate to the research performed. +- While NeurIPS does not require releasing code, the conference does require all submissions to provide some reasonable avenue for reproducibility, which may depend on the nature of the contribution. For example +(a) If the contribution is primarily a new algorithm, the paper should make it clear how to reproduce that algorithm. +(b) If the contribution is primarily a new model architecture, the paper should describe the architecture clearly and fully. +(c) If the contribution is a new model (e.g., a large language model), then there should either be a way to access this model for reproducing the results or a way to reproduce the model (e.g., with an open-source dataset or instructions for how to construct the dataset). +(d) We recognize that reproducibility may be tricky in some cases, in which case authors are welcome to describe the particular way they provide for reproducibility. In the case of closed-source models, it may be that access to the model is limited in some way (e.g., to registered users), but it should be possible for other researchers to have some path to reproducing or verifying the results. + +# 5. Open access to data and code + +Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material? + +# Answer: [No] + +Justification: We will release our code upon acceptance. + +# Guidelines: + +- The answer NA means that paper does not include experiments requiring code. +- Please see the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details. +- While we encourage the release of code and data, we understand that this might not be possible, so "No" is an acceptable answer. Papers cannot be rejected simply for not including code, unless this is central to the contribution (e.g., for a new open-source benchmark). +- The instructions should contain the exact command and environment needed to run to reproduce the results. See the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details. +- The authors should provide instructions on data access and preparation, including how to access the raw data, preprocessed data, intermediate data, and generated data, etc. +- The authors should provide scripts to reproduce all experimental results for the new proposed method and baselines. If only a subset of experiments are reproducible, they should state which ones are omitted from the script and why. +- At submission time, to preserve anonymity, the authors should release anonymized versions (if applicable). +- Providing as much information as possible in supplemental material (appended to the paper) is recommended, but including URLs to data and code is permitted. + +# 6. Experimental setting/details + +Question: Does the paper specify all the training and test details (e.g., data splits, hyperparameters, how they were chosen, type of optimizer, etc.) necessary to understand the results? + +# Answer: [Yes] + +Justification: Please refer to experimental section. + +# Guidelines: + +- The answer NA means that the paper does not include experiments. +- The experimental setting should be presented in the core of the paper to a level of detail that is necessary to appreciate the results and make sense of them. +- The full details can be provided either with the code, in appendix, or as supplemental material. + +# 7. Experiment statistical significance + +Question: Does the paper report error bars suitably and correctly defined or other appropriate information about the statistical significance of the experiments? + +# Answer: [No] + +Justification: Our experiment section does not include error bar related experiment. + +# Guidelines: + +- The answer NA means that the paper does not include experiments. +- The authors should answer "Yes" if the results are accompanied by error bars, confidence intervals, or statistical significance tests, at least for the experiments that support the main claims of the paper. +- The factors of variability that the error bars are capturing should be clearly stated (for example, train/test split, initialization, random drawing of some parameter, or overall run with given experimental conditions). +- The method for calculating the error bars should be explained (closed form formula, call to a library function, bootstrap, etc.) +- The assumptions made should be given (e.g., Normally distributed errors). +- It should be clear whether the error bar is the standard deviation or the standard error of the mean. + +- It is OK to report 1-sigma error bars, but one should state it. The authors should preferably report a 2-sigma error bar than state that they have a $96\%$ CI, if the hypothesis of Normality of errors is not verified. +- For asymmetric distributions, the authors should be careful not to show in tables or figures symmetric error bars that would yield results that are out of range (e.g. negative error rates). +- If error bars are reported in tables or plots, The authors should explain in the text how they were calculated and reference the corresponding figures or tables in the text. + +# 8. Experiments compute resources + +Question: For each experiment, does the paper provide sufficient information on the computer resources (type of compute workers, memory, time of execution) needed to reproduce the experiments? + +Answer: [Yes] + +Justification: We include in the supplementary material. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The paper should indicate the type of compute workers CPU or GPU, internal cluster, or cloud provider, including relevant memory and storage. +- The paper should provide the amount of compute required for each of the individual experimental runs as well as estimate the total compute. +- The paper should disclose whether the full research project required more compute than the experiments reported in the paper (e.g., preliminary or failed experiments that didn't make it into the paper). + +# 9. Code of ethics + +Question: Does the research conducted in the paper conform, in every respect, with the NeurIPS Code of Ethics https://neurips.cc/public/EthicsGuidelines? + +Answer: [Yes] + +Justification: We use public datasets. + +Guidelines: + +The answer NA means that the authors have not reviewed the NeurIPS Code of Ethics. +- If the authors answer No, they should explain the special circumstances that require a deviation from the Code of Ethics. +- The authors should make sure to preserve anonymity (e.g., if there is a special consideration due to laws or regulations in their jurisdiction). + +# 10. Broader impacts + +Question: Does the paper discuss both potential positive societal impacts and negative societal impacts of the work performed? + +Answer: [NA] + +Justification: Our work is unrelated to social matters. + +Guidelines: + +- The answer NA means that there is no societal impact of the work performed. +- If the authors answer NA or No, they should explain why their work has no societal impact or why the paper does not address societal impact. +- Examples of negative societal impacts include potential malicious or unintended uses (e.g., disinformation, generating fake profiles, surveillance), fairness considerations (e.g., deployment of technologies that could make decisions that unfairly impact specific groups), privacy considerations, and security considerations. +- The conference expects that many papers will be foundational research and not tied to particular applications, let alone deployments. However, if there is a direct path to any negative applications, the authors should point it out. For example, it is legitimate to point out that an improvement in the quality of generative models could be used to + +generate deepfakes for disinformation. On the other hand, it is not needed to point out that a generic algorithm for optimizing neural networks could enable people to train models that generate Deepfakes faster. + +- The authors should consider possible harms that could arise when the technology is being used as intended and functioning correctly, harms that could arise when the technology is being used as intended but gives incorrect results, and harms following from (intentional or unintentional) misuse of the technology. +- If there are negative societal impacts, the authors could also discuss possible mitigation strategies (e.g., gated release of models, providing defenses in addition to attacks, mechanisms for monitoring misuse, mechanisms to monitor how a system learns from feedback over time, improving the efficiency and accessibility of ML). + +# 11. Safeguards + +Question: Does the paper describe safeguards that have been put in place for responsible release of data or models that have a high risk for misuse (e.g., pretrained language models, image generators, or scraped datasets)? + +Answer: [NA] + +Justification: We does not pose such risk. + +Guidelines: + +- The answer NA means that the paper poses no such risks. +- Released models that have a high risk for misuse or dual-use should be released with necessary safeguards to allow for controlled use of the model, for example by requiring that users adhere to usage guidelines or restrictions to access the model or implementing safety filters. +- Datasets that have been scraped from the Internet could pose safety risks. The authors should describe how they avoided releasing unsafe images. +- We recognize that providing effective safeguards is challenging, and many papers do not require this, but we encourage authors to take this into account and make a best faith effort. + +# 12. Licenses for existing assets + +Question: Are the creators or original owners of assets (e.g., code, data, models), used in the paper, properly credited and are the license and terms of use explicitly mentioned and properly respected? + +Answer: [Yes] + +Justification: All datasets are public datasets. + +Guidelines: + +- The answer NA means that the paper does not use existing assets. +- The authors should cite the original paper that produced the code package or dataset. +- The authors should state which version of the asset is used and, if possible, include a URL. +- The name of the license (e.g., CC-BY 4.0) should be included for each asset. +- For scraped data from a particular source (e.g., website), the copyright and terms of service of that source should be provided. +- If assets are released, the license, copyright information, and terms of use in the package should be provided. For popular datasets, paperswithcode.com/datasets has curated licenses for some datasets. Their licensing guide can help determine the license of a dataset. +- For existing datasets that are re-packaged, both the original license and the license of the derived asset (if it has changed) should be provided. +- If this information is not available online, the authors are encouraged to reach out to the asset's creators. + +# 13. New assets + +Question: Are new assets introduced in the paper well documented and is the documentation provided alongside the assets? + +Answer: [No] + +Justification: No new asset. + +Guidelines: + +- The answer NA means that the paper does not release new assets. +- Researchers should communicate the details of the dataset/code/model as part of their submissions via structured templates. This includes details about training, license, limitations, etc. +- The paper should discuss whether and how consent was obtained from people whose asset is used. +- At submission time, remember to anonymize your assets (if applicable). You can either create an anonymized URL or include an anonymized zip file. + +# 14. Crowdsourcing and research with human subjects + +Question: For crowdsourcing experiments and research with human subjects, does the paper include the full text of instructions given to participants and screenshots, if applicable, as well as details about compensation (if any)? + +Answer: [No] + +Justification: Not applicable. + +Guidelines: + +- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. +- Including this information in the supplemental material is fine, but if the main contribution of the paper involves human subjects, then as much detail as possible should be included in the main paper. +- According to the NeurIPS Code of Ethics, workers involved in data collection, curation, or other labor should be paid at least the minimum wage in the country of the data collector. + +# 15. Institutional review board (IRB) approvals or equivalent for research with human subjects + +Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals (or an equivalent approval/review based on the requirements of your country or institution) were obtained? + +Answer: [NA] + +Justification: Not applicable. + +Guidelines: + +- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. +- Depending on the country in which research is conducted, IRB approval (or equivalent) may be required for any human subjects research. If you obtained IRB approval, you should clearly state this in the paper. +- We recognize that the procedures for this may vary significantly between institutions and locations, and we expect authors to adhere to the NeurIPS Code of Ethics and the guidelines for their institution. +- For initial submissions, do not include any information that would break anonymity (if applicable), such as the institution conducting the review. + +# 16. Declaration of LLM usage + +Question: Does the paper describe the usage of LLMs if it is an important, original, or non-standard component of the core methods in this research? Note that if the LLM is used only for writing, editing, or formatting purposes and does not impact the core methodology, scientific rigorousness, or originality of the research, declaration is not required. + +Answer: [NA] + +Justification: We only use it to revise our paper. + +# Guidelines: + +- The answer NA means that the core method development in this research does not involve LLMs as any important, original, or non-standard components. +- Please refer to our LLM policy (https://neurips.cc/Conferences/2025/LLM) for what should or should not be described. \ No newline at end of file diff --git a/NeurIPS/2025/3DOT_ Texture Transfer for 3DGS Objects from a Single Reference Image/images.zip b/NeurIPS/2025/3DOT_ Texture Transfer for 3DGS Objects from a Single Reference Image/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..abbfe6a41f7f902df62d12a96523dcebe706c2b1 --- /dev/null +++ b/NeurIPS/2025/3DOT_ Texture Transfer for 3DGS Objects from a Single Reference Image/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:346326ea41e492ba61d03a2a0951b2b222e1b32d7aeb88c19bdea846a61ca2b1 +size 745976 diff --git a/NeurIPS/2025/3DOT_ Texture Transfer for 3DGS Objects from a Single Reference Image/layout.json b/NeurIPS/2025/3DOT_ Texture Transfer for 3DGS Objects from a Single Reference Image/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..20533f0f6d1f52c0b674060f50aad5c5b9a829f2 --- /dev/null +++ b/NeurIPS/2025/3DOT_ Texture Transfer for 3DGS Objects from a Single Reference Image/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:605edc0cf168efd63d83567d2268acf96b60ae777807691385248e296c7cda4f +size 600586 diff --git a/NeurIPS/2025/3DPE-Gaze_Unlocking the Potential of 3D Facial Priors for Generalized Gaze Estimation/2ac394c3-60fc-4f41-8020-0b1c99fd8d5c_content_list.json b/NeurIPS/2025/3DPE-Gaze_Unlocking the Potential of 3D Facial Priors for Generalized Gaze Estimation/2ac394c3-60fc-4f41-8020-0b1c99fd8d5c_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..484321932dc176e2ed8edabf3de2423531095c16 --- /dev/null +++ b/NeurIPS/2025/3DPE-Gaze_Unlocking the Potential of 3D Facial Priors for Generalized Gaze Estimation/2ac394c3-60fc-4f41-8020-0b1c99fd8d5c_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d231d3420dc8579d95252e125a5033424d810958b4109f069bf353413047255e +size 117083 diff --git a/NeurIPS/2025/3DPE-Gaze_Unlocking the Potential of 3D Facial Priors for Generalized Gaze Estimation/2ac394c3-60fc-4f41-8020-0b1c99fd8d5c_model.json b/NeurIPS/2025/3DPE-Gaze_Unlocking the Potential of 3D Facial Priors for Generalized Gaze Estimation/2ac394c3-60fc-4f41-8020-0b1c99fd8d5c_model.json new file mode 100644 index 0000000000000000000000000000000000000000..c97655eff8f9834a5525873fdb80e6323e7cb13a --- /dev/null +++ b/NeurIPS/2025/3DPE-Gaze_Unlocking the Potential of 3D Facial Priors for Generalized Gaze Estimation/2ac394c3-60fc-4f41-8020-0b1c99fd8d5c_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e4b6a529b33ca5049784a67a9f1d0c58af4b4e17f9c76c427376a87dfff21d0c +size 154673 diff --git a/NeurIPS/2025/3DPE-Gaze_Unlocking the Potential of 3D Facial Priors for Generalized Gaze Estimation/2ac394c3-60fc-4f41-8020-0b1c99fd8d5c_origin.pdf b/NeurIPS/2025/3DPE-Gaze_Unlocking the Potential of 3D Facial Priors for Generalized Gaze Estimation/2ac394c3-60fc-4f41-8020-0b1c99fd8d5c_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..532c748f4eaa86d7977d406784ffa2b6fbdfec80 --- /dev/null +++ b/NeurIPS/2025/3DPE-Gaze_Unlocking the Potential of 3D Facial Priors for Generalized Gaze Estimation/2ac394c3-60fc-4f41-8020-0b1c99fd8d5c_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:794538e5eca7cbac5e9dd5cb21088336235213145cf9479f43c6676da5e081ea +size 2821701 diff --git a/NeurIPS/2025/3DPE-Gaze_Unlocking the Potential of 3D Facial Priors for Generalized Gaze Estimation/full.md b/NeurIPS/2025/3DPE-Gaze_Unlocking the Potential of 3D Facial Priors for Generalized Gaze Estimation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..4bd4d444f6c7317fd09f349d4e3bf4ce3167d642 --- /dev/null +++ b/NeurIPS/2025/3DPE-Gaze_Unlocking the Potential of 3D Facial Priors for Generalized Gaze Estimation/full.md @@ -0,0 +1,573 @@ +# 3DPE-Gaze:Unlocking the Potential of 3D Facial Priors for Generalized Gaze Estimation + +Yangshi Ge Yiwei Bao Feng Lu* + +State Key Laboratory of VR Technology and Systems, School of CSE, Beihang University {geyangshi, baoyiwei, lufeng}@buaa.edu.cn + +# Abstract + +In recent years, face-based deep-learning gaze estimation methods have achieved significant advancements. However, while face images provide supplementary information beneficial for gaze inference, the substantial extraneous information they contain also increases the risk of overfitting during model training and compromises generalization capability. To alleviate this problem, we propose the 3DPE-Gaze framework, explicitly modeling 3D facial priors for feature decoupling and generalized gaze estimation. The 3DPE-Gaze framework consists of two core modules: the 3D Geometric Prior Module (3DGP) incorporating the FLAME model to parameterize facial structures and gaze-irrelevant facial appearances while extracting gaze features; the Semantic Concept Alignment Module (SCAM) separates gaze-related and unrelated concepts through CLIP-guided contrastive learning. Finally, the 3DPE-Gaze framework combines 3D facial landmark as prior for generalized gaze estimation. Experimental results show that 3DPE-Gaze outperforms existing state-of-the-art methods on four major cross-domain tasks, with particularly outstanding performance in challenging scenarios such as lighting variations, extreme head poses, and glasses occlusion. + +# 1 Introduction + +Gaze estimation has wide applications in computer vision, being an essential technology in scenarios such as human-computer interaction[13, 28, 29], virtual/augmented reality[25, 31], and driving monitoring[1, 19, 27]. Gaze estimation methods can be broadly categorized into two types: model-based approaches and appearance-based approaches. Model-based approaches calculate gaze direction by simulating the anatomical structure of the eyeball, offering higher accuracy but typically requiring specialized hardware and controlled environments. Appearance-based approaches, on the other hand, learn gaze mapping relationships directly from image features, offering broader application potential. + +Appearance-based gaze estimation research has undergone a critical transition from focusing solely on eye regions to utilizing full-face information. Early research primarily extracted local features from eye regions[26, 20], but this approach overlooked the important contextual information provided by other facial areas. Zhang et al.[38] pioneered the use of full-face input and introduced a spatial weighting mechanism, significantly improving prediction accuracy. Since then, full-face input has become the mainstream approach in gaze estimation research. + +However, the full-face input strategy is a double-edged sword: while it contains rich gaze-related structural information, the eye region occupies only a small proportion of the image, simultaneously introducing numerous domain-specific interferences (such as lighting variations, skin color differences, expression changes, etc.)[33]. These irrelevant information easily lead to models overfitting to + +![](images/9ffacad5f7844925fd033a474025ab3a8c4f7c2d2f53d4cd5b5fccff6eb80ab1.jpg) +Figure 1: Left shows traditional CNN exhibiting clear performance collapse under standard conditions versus domain shift conditions. Right briefly demonstrates how we introduce 3D facial priors. + +appearance features in the training data, severely affecting cross-domain generalization capability. To address this challenge, researchers have proposed various solutions, such as adversarial learning[7], data perturbation[33], feature separation[18], and multimodal fusion[36, 35]. However, most of these methods remain at the image feature level and fail to fully utilize 3D facial priors, making it difficult for them to capture the inherent geometric structure of the face and more susceptible to surface appearance changes. Therefore, how to fully utilize structured facial information while effectively filtering domain-specific noise remains a key challenge in the field of gaze estimation. + +To address this challenge, we propose the Three-Dimensional Prior Enhanced Gaze Estimation framework (3DPE-Gaze), consisting of two complementary core modules: 3D Geometric Prior Module (3DGP) and Semantic Concept Alignment Module (SCAM). As shown in Figure 1, the 3DGP module decouples head pose, expression, shape, and gaze parameters from 2D facial images through the FLAME parametric face model, transforming gaze estimation from a prediction process based on domain-susceptible 2D appearance features to one based on stable geometric structures. However, FLAME-based encoders [9, 8] have inherent limitations in precisely modeling eyeball rotation and capturing subtle eye details, making it difficult to fully express complex gaze behaviors. To compensate for this deficiency, we designed the SCAM module as a complement. This module, based on CLIP semantic representations, distinguishes between gaze-relevant and irrelevant features through contrastive learning. Our FacePrior-Gaze Predictor effectively fuses these two complementary types of information, enabling the model to simultaneously leverage structured geometric constraints and high-level semantic concepts, thereby enhancing cross-domain generalization capability. + +In summary, 3DPE-Gaze systematically injects 3D facial priors and semantic concept priors into the gaze estimation process, effectively solving the cross-domain generalization problem in gaze prediction. Our experiments on multiple benchmark datasets have validated the excellent cross-domain performance of this method, achieving significant performance improvements in various challenging scenarios compared to current state-of-the-art methods. The main contributions of this research are as follows: + +- We propose the 3D Geometric Prior Module (3DGP), a novel parametric encoder that transforms the gaze estimation problem from pixel space to parameters such as head pose, expression, and shape, establishing the foundation for associating gaze prediction with 3D facial structure and achieving effective extraction of domain-invariant features. +- We propose the Semantic Concept Alignment Module (SCAM), specifically addressing the limitations of pure geometric modeling in capturing complex gaze semantics. This module innovatively leverages CLIP pre-trained knowledge and contrastive learning strategies to explicitly distinguish gaze-relevant concepts from domain interference factors in the feature space, allowing the model to adapt to new environments without any target domain data. +- Experimental results show that our 3DPE-Gaze framework achieves significant improvements in various cross-domain settings. In cross-dataset evaluations, it reduces average error by more than $27\%$ compared to the baseline and improves by $6\%$ over SOTA methods, with particularly outstanding performance in highly challenging scenarios. + +# 2 Related Work + +Gaze Estimation. Gaze estimation has shifted from using local eye features [26, 20] to full-face inputs [38], which, despite including richer structural information, introduced domain-specific noise that impairs generalization. Existing solutions have largely focused on either purifying 2D image features [7, 33, 18] or introducing prior knowledge. For instance, some methods incorporate geometric cues like head pose as auxiliary information [40, 38, 22, 32, 17]. However, these approaches often treat geometric priors as supplementary inputs rather than core representations. In contrast, our 3DGP module fundamentally reframes the problem by using a parametric 3D model to map faces from an unstructured pixel space to a physically meaningful and decoupled parameter space. This transforms the task from simple "pixel-to-gaze" regression into a more robust "structure-to-gaze" prediction. + +Other works have leveraged large-scale pre-trained models like CLIP to enhance robustness [35, 36]. These methods typically apply visual-linguistic knowledge to general 2D image features. Our work differs significantly in its design. The SCAM module performs a targeted semantic purification on a specific gaze code that has already been geometrically isolated by the 3DGP module. This novel "secondary purification" on a pre-decoupled feature is a key contribution of our framework. + +Face Parametric Model. 3D face models have evolved from the 3DMM[2] first proposed by Blanz and Vetter, to BFM[23], FaceWarehouse[3], and then to FLAME[16], gradually enhancing the modeling capabilities for facial shape, expression, and movement. Although recent models like ICT[15] and FaceScape[34] offer higher geometric detail in appearance, they lack FLAME's capabilities in feature disentanglement and efficiency. FLAME's key advantage lies in its explicit parametric decoupling of shape, pose, and expression, a characteristic that highly aligns with the need in gaze estimation to separate gaze-related features from domain-specific interference. Additionally, it achieves a good balance between expressiveness and computational efficiency. + +# 3 Method + +# 3.1 3DPE-Gaze Framework Overview + +To unlock the potential of 3D facial priors for generalized gaze estimation, we propose the 3DPE-Gaze framework. As shown in Figure 2, the 3DPE-Gaze framework consists of two core modules and a final gaze predictor: + +3D Geometric Prior Module (3DGP) The 3DGP module, based on the parametric face model FLAME, decomposes facial images into shape, expression, pose, lighting, and gaze parameters, achieving explicit decoupling of gaze features, facial structure features, and domain interference factors in physical space. This module also provides physically constrained anatomical foundations for gaze estimation through the reconstruction of 3D facial structures. + +Semantic Concept Alignment Module (SCAM) The SCAM module utilizes CLIP pre-trained knowledge and contrastive learning strategies. This module explicitly distinguishes between "gaze-related concepts" (such as looking up-left) and "domain interference concepts" (such as lighting variations, facial expressions, etc.) in the feature space, providing cross-domain stable representations at the semantic level. + +Facial Prior-Gaze Predictor The Facial Prior-Gaze Predictor integrates facial structural features extracted by 3DGP with semantic concepts identified by SCAM. This decoder focuses on highly gaze-relevant geometric regions through attention mechanisms while suppressing domain-specific interference such as background and lighting, generating the final accurate gaze prediction results. + +This design successfully unlocks the potential of 3D facial priors in gaze estimation: the 3DGP module provides stable facial prior constraints, while the SCAM module compensates for FLAME's limitations in eyeball modeling, offering semantic-level supplements. The complementary fusion of both significantly enhances the model's cross-domain generalization capability. Technical details of each module will be elaborated in the following sections. + +# 3.2 3D Geometric Prior Module (3DGP) + +The 3DGP module decomposes input images into structured geometric parameter representations through a parametric face model, thereby transferring gaze estimation from the pixel space, which + +![](images/543491eed254ea1853ea5ed6d7b2b7521855b2b6852e22b046f6c2a7ff1d6049.jpg) +Figure 2: Overview of the proposed 3DPE-Gaze framework. The framework contains two core modules: (1) The 3D Geometric Prior Module (3DGP) utilizes the FLAME model to decompose facial images into geometric parameters, achieving physical decoupling of gaze features and domain interference; (2) The Semantic Concept Alignment Module (SCAM) separates gaze-related and unrelated concepts through CLIP-guided contrastive learning. + +is susceptible to domain interference, to a more stable, decoupled parameter space. Through this decomposition, we can explicitly separate gaze code, facial structure parameters that affect gaze appearance, and domain-specific lighting parameters. + +Parametric Representation and Encoding: Based on the FLAME model[16], we designed a gaze-specific multi-task encoder that receives preprocessed facial images II, and outputs a set of geometric parameters: + +$$ +[ \beta , \theta , \psi , \gamma , l ] \tag {1} +$$ + +where, $\beta \in \mathbb{R}^{100}$ controls static facial shape, $\theta \in \mathbb{R}^6$ represents head pose, $\psi \in \mathbb{R}^{50}$ captures dynamic expressions, $\gamma \in \mathbb{R}^{256}$ is a gaze-specific encoding containing eyeball movement information, and $l\in \mathbb{R}^{27}$ represents lighting parameters modeling environmental light conditions. + +This parametric design decomposes facial images into domain-invariant geometric factors and susceptible appearance factors. For example, head pose $\theta$ and expression parameters $\psi$ remain relatively consistent under different lighting or skin color conditions, while lighting parameters $l$ explicitly capture environmental variation information. This explicit decoupling significantly reduces the model's dependence on domain-specific appearance features. + +To provide stable and structurally constrained facial parametric representations, we reused the encoder architecture of the DECA[9] pre-trained model (based on a ResNet backbone) and initialized all layers corresponding to the $\beta, \theta, \psi, l$ parameters in the multi-task encoder with its pre-trained weights (including the shared ResNet backbone and corresponding output branches). During training, these DECA-initialized parts are completely frozen (non-trainable) to ensure the accuracy and stability of facial structure parameters. In contrast, only the encoding branch responsible for outputting the gaze code $\gamma$ is trainable. Considering that the $\gamma$ parameter is a specific mapping of generic features extracted by the ResNet backbone to gaze direction, we designed this branch as a multi-layer perceptron (MLP) connected after the shared ResNet backbone output features, able to directly learn the mapping from robust features to gaze parameters, focusing on gaze feature learning. + +3D Reconstruction and Structural Constraints: After obtaining FLAME parameters, we reconstruct the 3D facial mesh through the FLAME decoder: + +$$ +M _ {\text {p r e d}} = \operatorname {F L A M E} (\beta , \theta , \psi) = \bar {T} + B _ {S} (\beta) + B _ {P} (\theta) + B _ {E} (\psi , \theta) \tag {2} +$$ + +where $\bar{T}$ is the average face template, $B_{S}(\beta)$ is the shape blend shape controlled by identity parameters, $B_{P}(\theta)$ is the pose-related deformation, and $B_{E}(\psi ,\theta)$ is the expression blend shape. From the reconstructed facial mesh, we extract key 3D facial landmarks $K\in \mathbb{R}^{N\times 3}$ , which provide anatomy-based constraints for subsequent gaze decoding, ensuring that gaze predictions conform to physical laws. + +Compared to traditional methods, the 3DGP module transforms gaze estimation from a black-box "pixel-to-gaze" mapping to a more structured "structure-to-gaze" prediction. By decoupling known geometric factors, it enhances model interpretability and significantly improves cross-domain stability. + +# 3.3 Semantic Concept Alignment Module (SCAM) + +The SCAM module compensates for 3DGP's limitations in semantic understanding by explicitly distinguishing between "gaze-related" and "gaze-irrelevant" features at the concept level through the powerful vision-language alignment ability of the CLIP[24]. + +Systematic Generation of Semantic Concepts: To provide robust semantic supervision, we employ a systematic methodology for generating language descriptions for both gaze-related ( $S_{\text{gaze}}$ ) and domain interference ( $S_{\text{domain}}$ ) concepts. + +For gaze-related concepts, a template-based approach is utilized, conditioned on the ground-truth gaze direction. We discretize gaze directions into canonical zones (e.g., "upper-right," "lower-left"), from which textual prompts such as, "A person looking towards the upper-right," are generated. This ensures the semantic representation is dynamically and accurately aligned with the precise directional gaze information for each sample. + +For domain interference concepts, we leverage the physical parameters already extracted by our 3DGP module to programmatically generate a diverse set of gaze-irrelevant descriptions. This strategy creates a strong, coherent link between our geometric decoupling and semantic purification stages. Specifically, these prompts describe: (1) Lighting conditions, derived from the light parameters $l$ (e.g., "under strong light," "in dim light"); (2) Facial expressions, based on the expression parameters $\psi$ (e.g., "a neutral expression," "a person with a significant facial expression"); (3) Facial shape, informed by the shape parameters $\beta$ ; and (4) Head pose, based on the pose parameters $\theta$ (e.g., "head tilted to the left"). + +These systematically generated sets of descriptions are then passed through the CLIP text encoder to obtain their semantic representations: + +$$ +t _ {\text {g a z e}} = \Psi_ {\text {C L I P}} \left(s _ {\text {g a z e}}\right), \quad s _ {\text {g a z e}} \in S _ {\text {g a z e}} \tag {3} +$$ + +$$ +t _ {\text {d o m a i n}} = \Psi_ {\text {C L I P}} \left(s _ {\text {d o m a i n}}\right), \quad s _ {\text {d o m a i n}} \in S _ {\text {d o m a i n}} \tag {4} +$$ + +where $\Psi_{\mathrm{CLIP}}(\cdot)$ represents the CLIP encoder. + +Contrastive Learning and Concept Separation: To align gaze code $\gamma$ with gaze concepts in semantic space while keeping them away from domain interference concepts, we designed a bidirectional contrastive loss: + +Gaze Concept Alignment Loss: + +$$ +\mathcal {L} _ {\text {g a z e - a l i g n}} = - \log \frac {\exp \left(\sin \left(z _ {\gamma} , t _ {\text {g a z e}}\right) / \tau\right)}{\sum_ {j} \exp \left(\sin \left(z _ {\gamma} , t _ {j}\right) / \tau\right)} \tag {5} +$$ + +where $z_{\gamma}$ is the projected gaze parameter, sim calculates cosine similarity, and $\tau$ is a temperature parameter. + +Domain Interference Repulsion Loss: + +$$ +\mathcal {L} _ {\text {d o m a i n - r e p e l}} = \frac {1}{| \mathcal {S} _ {\text {d o m a i n}} |} \sum_ {s \in \mathcal {S} _ {\text {d o m a i n}}} \max (0, \operatorname {s i m} (z \gamma , \Psi_ {\mathrm {C L I P}} (s)) - m) \tag {6} +$$ + +where $m$ is a boundary parameter, ensuring that gaze features maintain sufficient distance from domain interference concepts. + +Through this contrastive mechanism, the SCAM module establishes an explicit boundary between gaze features and domain interference features in semantic space, enabling the model to focus on + +high-level concepts related to gaze while filtering out domain-specific interferences. Complementary to geometric priors, semantic priors demonstrate unique advantages in handling abstract visual concepts and complex visual interferences. + +# 3.4 Facial Prior-Gaze Predictor + +To fully utilize both geometric and semantic priors, we designed a collaborative decoder that makes the two representations mutually enhancing through an attention mechanism, jointly guiding gaze prediction. + +Cross-modal Feature Fusion: We adopt a cross-attention mechanism, using semantically guided gaze code $\gamma$ as the Query, and three-dimensional geometric landmarks $K$ as the Key and Value: + +$$ +F _ {\text {f u s e d}} = \text {C r o s s A t t e n t i o n} (Q = W _ {q} \gamma ;; K = W _ {k} K,; V = W _ {v} K) \tag {7} +$$ + +where $W_{q}, W_{k}, W_{v}$ are learnable parameter matrices. This design enables the model to selectively focus on key structural regions based on semantic understanding, increasing sensitivity to gaze-related features. + +Gaze Vector Prediction: The fused features generate gaze predictions through the final MLP decoder: + +$$ +g = \frac {f _ {\text {r e g}} \left(F _ {\text {f u s e d}}\right)}{\left| f _ {\text {r e g}} \left(F _ {\text {f u s e d}}\right) \right| _ {2}} \tag {8} +$$ + +This collaborative mechanism achieves complementary enhancement of geometric and semantic priors: the structural features provided by 3DGP ensure that gaze predictions conform to physical constraints, while the semantic concepts contributed by SCAM guide the model to focus on the most relevant regions and suppress interference. Working together, they effectively alleviate the problem of traditional methods over-relying on domain-specific appearance features while fully utilizing structured information within the full-face range. + +# 3.5 Training Objectives and Implementation + +We adopt a multi-objective joint optimization strategy to train the 3DPE-Gaze framework, with the overall loss function: + +$$ +\mathcal {L} = \lambda_ {1} \mathcal {L} _ {\text {a n g l e}} + \lambda_ {2} \mathcal {L} _ {\text {g a z e - a l i g n}} + \lambda_ {3} \mathcal {L} _ {\text {d o m a i n - r e p e l}} \tag {9} +$$ + +Gaze Angular Loss $\mathcal{L}$ angle measures the angular difference between predicted gaze and true gaze: + +$$ +\mathcal {L} _ {\text {a n g l e}} = \arccos \left(\frac {\hat {g} ^ {T} g}{\| \hat {g} \| _ {2} \| g \| _ {2}}\right) \tag {10} +$$ + +Gaze Concept Alignment Loss $\mathcal{L}_{\mathrm{gaze - align}}$ encourages gaze parameters $\gamma$ to align with gaze-related concepts. + +Domain Interference Repulsion Loss $\mathcal{L}_{\mathrm{domain - repel}}$ ensures that gaze parameters $\gamma$ stay away from interference concepts. + +Hyperparameters $\lambda_1, \lambda_2$ , and $\lambda_3$ are used to balance the contributions of each loss term. Empirically, we set $\lambda_1 = \lambda_2 = \lambda_3 = 1.0$ . + +Training Implementation Details We conducted experiments on a single NVIDIA A100 GPU. Specifically, we adopted a staged training strategy: first freezing the FLAME parameter branches and pre-training the gaze parameter $\gamma$ ; then introducing contrastive learning of the SCAM module and jointly optimizing the entire model. We used the Adam optimizer with an initial learning rate of $10^{-4}$ and a batch size of 256. + +# 4 Experiments + +Datasets and Evaluation Protocol. We adopt experimental settings consistent with cutting-edge research in cross-domain gaze estimation [7, 33, 36, 35, 18], evaluating our method on four + +cross-domain tasks. Specifically, we use ETH-XGaze[39] and Gaze360[14] as training datasets, and MPIIFaceGaze[37] and EyeDiap[12] as testing datasets. For concise representation, we denote these four cross-domain tasks as $\mathcal{D}_E(\mathrm{ETH - XGaze})\to \mathcal{D}_M(\mathrm{MPIIfaceGaze}),\mathcal{D}_E\to \mathcal{D}_D(\mathrm{EyeDiap}),$ $\mathcal{D}_G(\mathrm{Gaze}360)\rightarrow \mathcal{D}_M,$ and $\mathcal{D}_G\to \mathcal{D}_D$ . This standardized cross-domain setup ensures that our experimental results can be directly compared with related research. + +Data Preprocessing. For $\mathcal{D}_E$ , $\mathcal{D}_M$ , and $\mathcal{D}_D$ , we normalize facial images following the standard method in [37]; for $\mathcal{D}_G$ , we only select frontal face images to match the distribution characteristics of other datasets, which is consistent with former researches [7, 18]. All images are resized to a uniform resolution of $224\times 224$ and normalized to the $[0,1]$ range, thereby eliminating the influence of differences in acquisition devices and resolutions across different datasets. + +# 4.1 Performance Comparison with State-of-the-Art Methods + +Table 1: Performance comparison on cross-domain gaze estimation tasks (unit: degrees) + +
MethodDE→DMDE→DDDG→DMDG→DDAvg
CNN Baseline8.568.909.518.488.86
PureGaze [7]7.087.449.289.328.28
CDG [30]6.737.957.037.277.25
Xu et al. [33]6.507.447.559.037.63
Liang et al. [18]5.796.967.067.996.95
CLIP-Gaze [36]6.417.516.897.066.96
LG-Gaze [35]6.457.226.836.866.84
Our 3DPE-Gaze6.666.136.716.236.43
+ +Table1 shows the performance comparison between 3DPE-Gaze and existing state-of-the-art methods on four cross-domain gaze estimation tasks. The results demonstrate that our method surpasses SOTA methods in 3 out of 4 cross-domain settings. From the overall performance, our approach exhibits stronger general generalization capability, achieving the lowest average error across all four cross-domain tasks, fully validating the excellent performance of our proposed framework in addressing cross-domain challenges. This result also confirms that effective utilization of facial priors can significantly enhance the cross-domain generalization capability of gaze estimation models. + +Table 2: In-domain gaze estimation performance comparison (unit: degrees). + +
Methodwithin DMwithin Ddwithin DGwithin DE
Dilated-Net [4]4.426.1913.73N/A
Gaze360 [14]4.065.3611.044.46
RT-Gene [11]4.666.0212.26N/A
FullFace [38]4.936.5314.997.38
RCNN [21]4.105.3111.23N/A
CA-Net [6]4.275.2711.20N/A
GazeTR-Pure [5]4.745.7213.58N/A
GazeTR-Hybrid [5]4.005.1710.62N/A
CNN Baseline4.747.4913.235.69
Our 3DPE-Gaze4.035.0611.834.39
+ +While our 3DPE-Gaze framework is primarily designed to enhance cross-domain generalization, it is also crucial to validate that this improvement does not compromise its performance within a single domain. Therefore, we conducted in-domain experiments, with the results presented in Table 2. The results show that our method achieves highly competitive, and in some cases state-of-the-art, performance. This demonstrates that our approach of leveraging 3D facial priors not only significantly boosts cross-domain robustness but also maintains excellent accuracy for in-domain gaze estimation. + +# 4.2 Diagnostic Analysis of Geometric Prior Quality + +To understand the performance discrepancy across different tasks, particularly the weaker result on the $\mathcal{D}_E\rightarrow \mathcal{D}_M$ task, we conducted a diagnostic analysis. We hypothesize that the model's final accuracy is strongly correlated with the quality of the geometric priors extracted by the 3DGP module. + +To test this, we use the stability of the reconstructed 3D Inter-ocular Distance (IOD) as a proxy for prior quality. For any given subject, their physical IOD is fixed; therefore, a lower standard deviation of the IOD across multiple images indicates a more stable and higher-quality 3D prior. As shown in Table 3, we grouped subjects from each target dataset into "High-Quality" and "Low-Quality" prior groups based on this metric. + +The results clearly show that for both target domains, the High-Quality Prior group achieves significantly lower gaze error than the Low-Quality Prior group. This analysis confirms that our model's performance is indeed dependent on the quality of the extracted 3D geometry, and the instability of priors from the $\mathcal{D}_M$ dataset is the primary reason for the higher error in that specific cross-domain task. + +Table 3: Diagnostic analysis of geometric prior quality and its impact on gaze error. Avg. IOD Std Dev refers to the average standard deviation of the reconstructed 3D Inter-ocular Distance. + +
DomainPrior Quality GroupAvg. IOD Std DevAverage Gaze ErrorNumber of Subjects
DMHigh-Quality Prior4.15.50°5
Low-Quality Prior10.57.24°10
Overall8.46.66°15
D DHigh-Quality Prior3.85.38°9
Low-Quality Prior9.27.10°7
Overall6.26.13°16
+ +# 4.3 Ablation Studies + +# Effectiveness of Core Modules + +Our proposed framework relies on three complementary loss functions to achieve high-performance cross-domain gaze estimation: Gaze Angular Loss $(\mathcal{L}_{\mathrm{angle}})$ , Gaze Concept Alignment Loss $(\mathcal{L}_{\mathrm{gaze - align}})$ , and Domain Interference Repulsion Loss $(\mathcal{L}_{\mathrm{domain - repel}})$ . To verify the contribution of each loss function and their synergistic effect, we designed ablation experiments as shown in Table 4. + +Table 4: Ablation experiments for different loss function combinations (unit: degrees) + +
Model ConfigurationDE→DMDE→DdDG→DMDG→DD
CNN Baseline8.568.909.518.48
+ Ldomain-repel7.327.197.637.44
+ Lgaze-align8.038.358.407.87
+ SCAM (Lgaze-align + Ldomain-repel)7.186.867.327.17
3DGP Only (Langle)7.457.197.606.80
+ Ldomain-repel6.906.507.056.35
+ Lgaze-align7.416.387.606.72
+ SCAM (Lgaze-align + Ldomain-repel)6.666.136.716.23
+ +The results show that using the FLAME prior-based $\mathcal{L}_{\mathrm{angle}}$ alone can significantly reduce errors, validating the fundamental value of geometric constraints. Further introducing $\mathcal{L}_{\mathrm{gaze - align}}$ enhances model performance by leveraging semantic information. Finally, integrating $\mathcal{L}_{\mathrm{domain - repel}}$ effectively separates gaze-related and unrelated features, bringing comprehensive performance improvements, with the complete model achieving the best average error. This indicates that the synergistic effect of geometric constraints, semantic guidance, and feature decoupling is crucial for improving cross-domain robustness. + +# Impact of Backbone Model Choices + +To analyze the impact of our backbone model choices, we conducted ablation studies as suggested by reviewers, with results shown in Table 5. For the semantic encoder, we found that while more powerful CLIP architectures can improve performance, our chosen ViT-B-16 provides a strong balance between accuracy and efficiency. Similarly, for the FLAME parameter regressor, other high-quality models like SPECTRE also achieve competitive results, demonstrating our framework's flexibility. We selected DECA as our primary regressor due to its wide availability and recognized strong performance. + +Table 5: Ablation study on the impact of different backbone models (CLIP text encoders and FLAME parameter regressors). Unit: degrees. + +
Backbone ComponentDE→DMDE→DdDG→DMDG→DD
Semantic Encoder (CLIP)
RN506.906.556.956.44
ViT-B-16 (Ours)6.666.136.716.23
ViT-L-146.556.086.606.17
ViT-H-146.506.056.616.12
FLAME Parameter Regressor
EMOCA [8]6.856.406.906.45
DECA (Ours) [9]6.666.136.716.23
SPECTRE [10]6.596.106.656.18
+ +# Optimization of Geometric Feature Transfer Paths + +Table 6: Ablation experiments for model architecture configurations (unit: degrees). + +
Model ConfigurationDE→DMDE→DDDG→DMDG→DD
Landmarks Only6.666.136.716.23
Landmarks + Pose7.206.706.856.47
Landmarks + Full Parameters7.326.536.856.94
+ +To identify the optimal geometric representation for generalization, we compared three feature configurations from the FLAME model: using only 3D landmarks, landmarks with head pose parameters, and landmarks with the full parameter set. As shown in Table 6, using only facial landmarks performed best across all tasks. This result suggests that landmarks provide a sufficiently structured abstraction of facial geometry, capturing key gaze-related relationships while filtering the domain-specific noise present in the more detailed parameters. Including the full parameter set introduced additional domain-specific biases that hindered performance, confirming that focusing on abstract, domain-invariant features is more effective for cross-domain learning. + +# 4.4 Robustness Verification of Facial Priors in Extreme Scenarios + +The core idea of our 3DPE-Gaze framework is to decouple gaze from other irrelevant features and incorporating 3D facial priors for generalized gaze estimation. Thus, in this section, we conduct experiments to verify the robustness of 3DPE-Gaze framework regarding varies factors, including head pose, lighting conditions, expression changes and glasses on the $\mathcal{D}_E\to \mathcal{D}_M$ task. + +Robustness under Extreme Lighting Conditions. As shown in Figure 3 (left), our method outperforms the baseline under all lighting conditions, especially in extremes. It reduces errors by $8.3\%$ in low-light and $10.6\%$ in high-light areas, as our 3D geometric representation effectively separates facial structure from environmental lighting effects. + +Adaptability to Expression Changes. Figure 3 (right) shows that our 3DPE-Gaze method consistently outperforms the baseline across the entire range of facial expressions. For extreme expressions (L1 intensity $>17$ ), our method maintains a stable error level while the baseline's error significantly increases. This is due to our model's ability to decouple expression muscle activity from gaze direction. + +![](images/73d0ed24dd549fb75aa6aa81c81bf2302d94a7b90266baafb79e63f6c12a04c7.jpg) +Figure 3: Robustness analysis under extreme scenarios. Left: Analysis of gaze estimation accuracy under different lighting intensities; Right: Impact of facial expression variations on gaze estimation precision. + +![](images/98308bdfa37467894a7864e4503a3b853637e1424761466f35fe6bcb5e4b432f.jpg) + +![](images/44159b489b20daaed99dc9f6fa11783720facf81672235515abd5e1c329b7566.jpg) +Figure 4: Robustness analysis under extreme scenarios. Left: Gaze estimation accuracy under large head rotations; Right: Impact of glasses occlusion on gaze estimation precision. + +![](images/8ccf59c3760a950538df3b58cc975035d0181286aad0096b09c72a4b1c04bb2c.jpg) + +Large Head Rotations. Head pose variation is a major challenge in gaze estimation. As shown in Figure 4 (left), our method outperforms the baseline across all head poses, with the advantage growing at larger angles. By effectively separating the compound effects of head pose and eyeball movement, our method maintains stable performance even in extreme poses where the baseline fails. + +Generalization Capability in Glasses Occlusion Scenarios. As shown in Figure 4 (right), our method significantly outperforms the baseline both without glasses (20.9% error reduction, $10.60^{\circ} \rightarrow 8.39^{\circ}$ ) and with glasses (12.3% error reduction, $9.30^{\circ} \rightarrow 8.16^{\circ}$ ). Most importantly, our model demonstrates excellent cross-condition stability, with an error difference of only $0.23^{\circ}$ between the two conditions, compared to $1.30^{\circ}$ for the baseline, ensuring a more reliable user experience. + +# 5 Conclusion + +This paper introduces 3DPE-Gaze, a novel framework for cross-domain gaze estimation that integrates 3D geometric and semantic priors. By leveraging a 3DGP module for geometric decoupling and a SCAM module for semantic purification, our method shifts the task from an unstable appearance-based space to a more robust geometric and semantic one. This design achieves state-of-the-art cross-domain performance without requiring any target domain data. + +Limitations and Future Work. Limitations of our work include the FLAME model's difficulty in precisely modeling fine eye details and our reliance on a predefined set of semantic concepts for contrastive learning. Future work will focus on three areas: developing specialized parametric eye models for finer detail; exploring adaptive semantic learning strategies, such as concept generation; extending the framework to dynamic, real-time applications on mobile and AR/VR devices. + +Acknowledgement. This research has been supported by Beijing Natural Science Foundation (L242019). + +# References + +[1] Stefano Alletto, Andrea Palazzi, Francesco Solera, Simone Calderara, and Rita Cucchiara. Dr(eye)ve: A dataset for attention-based tasks with applications to autonomous and assisted driving. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, June 2016. +[2] Volker Blanz and Thomas Vetter. A Morphable Model For The Synthesis Of 3D Faces. Association for Computing Machinery, New York, NY, USA, 1 edition, 2023. ISBN 9798400708978. URL https://doi.org/10.1145/3596711.3596730. +[3] Chen Cao, Yanlin Weng, Shun Zhou, Yiying Tong, and Kun Zhou. Facewarehouse: A 3d facial expression database for visual computing. IEEE Transactions on Visualization and Computer Graphics, 20(3): 413-425, 2014. doi: 10.1109/TVCG.2013.249. +[4] Zhaokang Chen and Bertram E Shi. Appearance-based gaze estimation using dilated-convolutions. In Asian Conference on Computer Vision, pages 309–324. Springer, 2018. +[5] Yihua Cheng and Feng Lu. Gaze estimation using transformer. In 2022 26th International Conference on Pattern Recognition (ICPR), pages 3341-3347. IEEE, 2022. +[6] Yihua Cheng, Shiyao Huang, Fei Wang, Chen Qian, and Feng Lu. A coarse-to-fine adaptive network for appearance-based gaze estimation. In Proceedings of the AAAI conference on artificial intelligence, volume 34, pages 10623-10630, 2020. +[7] Yihua Cheng, Yiwei Bao, and Feng Lu. Puregaze: Purifying gaze feature for generalizable gaze estimation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 436-443, 2022. +[8] Radek Daneček, Michael J Black, and Timo Bolkart. Emoca: Emotion driven monocular face capture and animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20311-20322, 2022. +[9] Yao Feng, Haiwen Feng, Michael J Black, and Timo Bolkart. Learning an animatable detailed 3d face model from in-the-wild images. ACM Transactions on Graphics (ToG), 40(4):1-13, 2021. +[10] Panagiotis P Filntisis, George Retsinas, Foivos Paraperas-Papantoniou, Athanasios Katsamanis, Anastasios Roussos, and Petros Maragos. Visual speech-aware perceptual 3d facial expression reconstruction from videos. arXiv preprint arXiv:2207.11094, 2022. +[11] Tobias Fischer, Hyung Jin Chang, and Yiannis Demiris. Rt-gene: Real-time eye gaze estimation in natural environments. In Proceedings of the European conference on computer vision (ECCV), pages 334-352, 2018. +[12] Kenneth Alberto Funes Mora, Florent Monay, and Jean-Marc Odobez. Eyediap: a database for the development and evaluation of gaze estimation algorithms from rgb and rgb-d cameras. In Proceedings of the Symposium on Eye Tracking Research and Applications, ETRA '14, page 255-258, New York, NY, USA, 2014. Association for Computing Machinery. ISBN 9781450327510. doi: 10.1145/2578153.2578190. URL https://doi.org/10.1145/2578153.2578190. +[13] Christina Katsini, Yasmeen Abdrabou, George E. Raptis, Mohamed Khamis, and Florian Alt. The role of eye gaze in security and privacy applications: Survey and future hci research directions. In Proceedings of the 2020 CHI Conference on Human Factors in Computing Systems, CHI '20, page 1-21, New York, NY, USA, 2020. Association for Computing Machinery. ISBN 9781450367080. doi: 10.1145/3313831.3376840. URL https://doi.org/10.1145/3313831.3376840. +[14] Petr Kellnhofer, Adria Recasens, Simon Stent, Wojciech Matusik, and Antonio Torralba. Gaze360: Physically unconstrained gaze estimation in the wild. In IEEE International Conference on Computer Vision (ICCV), October 2019. +[15] Ruilong Li, Karl Bladin, Yajie Zhao, Chinmay Chinara, Owen Ingraham, Pengda Xiang, Xinglei Ren, Pratusha Prasad, Bipin Kishore, Jun Xing, et al. Learning formation of physically-based face attributes. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3410-3419, 2020. +[16] Tianye Li, Timo Bolkart, Michael J. Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4d scans. ACM Trans. Graph., 36(6), November 2017. ISSN 0730-0301. doi: 10.1145/3130800.3130813. URL https://doi.org/10.1145/3130800.3130813. + +[17] Yingxi Li, Xiaowei Bai, Liang Xie, Xiaodong Wang, Feng Lu, Feitian Zhang, Ye Yan, and Erwei Yin. Real-time gaze tracking via head-eye cues on head mounted devices. IEEE Transactions on Mobile Computing, 23(12):13292-13309, 2024. doi: 10.1109/TMC.2024.3425928. +[18] Ziyang Liang, Yiwei Bao, and Feng Lu. De-confounded gaze estimation. In Ales Leonardis, Elisa Ricci, Stefan Roth, Olga Russakovsky, Torsten Sattler, and Gül Varol, editors, Computer Vision – ECCV 2024, pages 219–235, Cham, 2025. Springer Nature Switzerland. ISBN 978-3-031-73337-6. +[19] Congcong Liu, Yuying Chen, Lei Tai, Haoyang Ye, Ming Liu, and Bert Shi. A gaze model improves autonomous driving. 06 2019. doi: 10.1145/3314111.3319846. +[20] Feng Lu, Yusuke Sugano, Takahiro Okabe, and Yoichi Sato. Adaptive linear regression for appearance-based gaze estimation. IEEE transactions on pattern analysis and machine intelligence, 36(10):2033-2046, 2014. +[21] C Palmero, J Selva, MA Bagheri, and S Escalera. Recurrent cnn for 3d gaze estimation using appearance and shape cues. arxiv 2018. arXiv preprint arXiv:1805.03064. +[22] Seonwook Park, Shalini De Mello, Pavlo Molchanov, Umar Iqbal, Otmar Hilliges, and Jan Kautz. Few-shot adaptive gaze estimation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9368-9377, 2019. +[23] Pascal Paysan, Reinhard Knothe, Brian Amberg, Sami Romdhani, and Thomas Vetter. A 3d face model for pose and illumination invariant face recognition. In 2009 sixth IEEE international conference on advanced video and signal based surveillance, pages 296-301. IEEE, 2009. +[24] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PmLR, 2021. +[25] Vincent Sitzmann, Ana Serrano, Amy Pavel, Maneesh Agrawala, Diego Gutierrez, and Gordon Wetzstein. Saliency in vr: How do people explore virtual environments? IEEE Transactions on Visualization and Computer Graphics, PP, 12 2016. doi: 10.1109/TVCG.2018.2793599. +[26] Kar-Han Tan, David J Kriegman, and Narendra Ahuja. Appearance-based eye gaze estimation. In Sixth IEEE Workshop on Applications of Computer Vision, 2002.(WACV 2002). Proceedings., pages 191-195. IEEE, 2002. +[27] Ashish Tawari, Kuo Chen, and Mohan Manubhai Trivedi. Where is the driver looking: Analysis of head, eye and iris for robust gaze zone estimation. 17th International IEEE Conference on Intelligent Transportation Systems (ITSC), pages 988-994, 2014. URL https://api-semanticscholar.org/CorpusID:19012458. +[28] Yunus Terzioglu, Bilge Mutlu, and Erol Şahin. Designing social cues for collaborative robots: The role of gaze and breathing in human-robot collaboration. In Proceedings of the 2020 ACM/IEEE International Conference on Human-Robot Interaction, HRI '20, page 343–357, New York, NY, USA, 2020. Association for Computing Machinery. ISBN 9781450367462. doi: 10.1145/3319502.3374829. URL https://doi.org/10.1145/3319502.3374829. +[29] Haofei Wang, Xu jiong Dong, Zhaokang Chen, and Bert Shi. Hybrid gaze/eeg brain computer interface for robot arm control on a pick and place task. volume 2015, pages 1476-1479, 08 2015. doi: 10.1109/EMBC.2015.7318649. +[30] Yaoming Wang, Yangzhou Jiang, Jin Li, Bingbing Ni, Wenrui Dai, Chenglin Li, Hongkai Xiong, and Teng Li. Contrastive regression for domain adaptation on gaze estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19376-19385, 2022. +[31] Zhimin Wang and Feng Lu. Tasks reflected in the eyes: Egocentric gaze-aware visual task type recognition in virtual reality. IEEE Transactions on Visualization and Computer Graphics, 30(11): 7277-7287, November 2024. ISSN 1077-2626. doi: 10.1109/TVCG.2024.3456164. URL https://doi.org/10.1109/TVCG.2024.3456164. +[32] Yunfeng Xiao, Xiaowei Bai, Baojun Chen, Hao Su, Hao He, Liang Xie, and Erwei Yin. De2gaze: Deformable and decoupled representation learning for 3d gaze estimation. In 2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3091-3100, 2025. doi: 10.1109/CVPR52734.2025.00294. + +[33] Mingjie Xu, Haofei Wang, and Feng Lu. Learning a generalized gaze estimator from gaze-consistent feature. In Proceedings of the Thirty-Seventh AAAI Conference on Artificial Intelligence and Thirty-Fifth Conference on Innovative Applications of Artificial Intelligence and Thirteenth Symposium on Educational Advances in Artificial Intelligence, AAAI'23/AAAI'23/EAAI'23. AAAI Press, 2023. ISBN 978-1-57735-880-0. doi: 10.1609/aaai.v37i3.25406. URL https://doi.org/10.1609/aaai.v37i3.25406. +[34] Haotian Yang, Hao Zhu, Yanru Wang, Mingkai Huang, Qiu Shen, Ruigang Yang, and Xun Cao. Facescape: a large-scale high quality 3d face dataset and detailed riggable 3d face prediction. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 601-610, 2020. +[35] Pengwei Yin, Jingjing Wang, Guanzhong Zeng, Di Xie, and Jiang Zhu. Lg-gaze: Learning geometry-aware continuous prompts for language-guided gaze estimation. In European Conference on Computer Vision, pages 1-17. Springer, 2024. +[36] Pengwei Yin, Guanzhong Zeng, Jingjing Wang, and Di Xie. Clip-gaze: towards general gaze estimation via visual-linguistic model. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 6729–6737, 2024. +[37] Xuong Zhang, Yusuke Sugano, Mario Fritz, and Andreas Bulling. It's written all over your face: Full-face appearance-based gaze estimation. In Proc. IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 2299-2308, 2017. doi: 10.1109/CVPRW.2017.284. +[38] Xuong Zhang, Yusuke Sugano, Mario Fritz, and Andreas Bulling. It's written all over your face: Full-face appearance-based gaze estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 51-60, 2017. +[39] Xucong Zhang, Seonwook Park, Thabo Beeler, Derek Bradley, Siyu Tang, and Otmar Hilliges. Eth-xgaze: A large scale dataset for gaze estimation under extreme head pose and gaze variation. In Computer Vision - ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part V, page 365-381, Berlin, Heidelberg, 2020. Springer-Verlag. ISBN 978-3-030-58557-0. doi: 10.1007/978-3-030-58558-7_22. URL https://doi.org/10.1007/978-3-030-58558-7_22. +[40] Wangjiang Zhu and Haoping Deng. Monocular free-head 3d gaze tracking with deep learning and geometry constraints. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), Oct 2017. + +# NeurIPS Paper Checklist + +# 1. Claims + +Question: Do the main claims made in the abstract and introduction accurately reflect the paper's contributions and scope? + +Answer: [Yes] + +Justification: In the abstract and introduction, we present previous work on cross-domain gaze estimation and claim that our work improves gaze estimation accuracy by introducing facial priors. We provide evidence for our algorithm in the experimental chapter, which reflects the claims made in the abstract and introduction. + +Guidelines: + +- The answer NA means that the abstract and introduction do not include the claims made in the paper. +- The abstract and/or introduction should clearly state the claims made, including the contributions made in the paper and important assumptions and limitations. A No or NA answer to this question will not be perceived well by the reviewers. +- The claims made should match theoretical and experimental results, and reflect how much the results can be expected to generalize to other settings. +- It is fine to include aspirational goals as motivation as long as it is clear that these goals are not attained by the paper. + +# 2. Limitations + +Question: Does the paper discuss the limitations of the work performed by the authors? + +Answer: [Yes] + +Justification: We have a subsection in the conclusion to discuss the limitations of our method and future research directions. + +Guidelines: + +- The answer NA means that the paper has no limitation while the answer No means that the paper has limitations, but those are not discussed in the paper. +- The authors are encouraged to create a separate "Limitations" section in their paper. +- The paper should point out any strong assumptions and how robust the results are to violations of these assumptions (e.g., independence assumptions, noiseless settings, model well-specification, asymptotic approximations only holding locally). The authors should reflect on how these assumptions might be violated in practice and what the implications would be. +- The authors should reflect on the scope of the claims made, e.g., if the approach was only tested on a few datasets or with a few runs. In general, empirical results often depend on implicit assumptions, which should be articulated. +- The authors should reflect on the factors that influence the performance of the approach. For example, a facial recognition algorithm may perform poorly when image resolution is low or images are taken in low lighting. Or a speech-to-text system might not be used reliably to provide closed captions for online lectures because it fails to handle technical jargon. +- The authors should discuss the computational efficiency of the proposed algorithms and how they scale with dataset size. +- If applicable, the authors should discuss possible limitations of their approach to address problems of privacy and fairness. +- While the authors might fear that complete honesty about limitations might be used by reviewers as grounds for rejection, a worse outcome might be that reviewers discover limitations that aren't acknowledged in the paper. The authors should use their best judgment and recognize that individual actions in favor of transparency play an important role in developing norms that preserve the integrity of the community. Reviewers will be specifically instructed to not penalize honesty concerning limitations. + +# 3. Theory assumptions and proofs + +Question: For each theoretical result, does the paper provide the full set of assumptions and a complete (and correct) proof? + +Answer: [NA] + +Justification: This paper does not contain theoretical results. + +Guidelines: + +- The answer NA means that the paper does not include theoretical results. +- All the theorems, formulas, and proofs in the paper should be numbered and cross-referenced. +- All assumptions should be clearly stated or referenced in the statement of any theorems. +- The proofs can either appear in the main paper or the supplemental material, but if they appear in the supplemental material, the authors are encouraged to provide a short proof sketch to provide intuition. +- Inversely, any informal proof provided in the core of the paper should be complemented by formal proofs provided in appendix or supplemental material. +- Theorems and Lemmas that the proof relies upon should be properly referenced. + +# 4. Experimental result reproducibility + +Question: Does the paper fully disclose all the information needed to reproduce the main experimental results of the paper to the extent that it affects the main claims and/or conclusions of the paper (regardless of whether the code and data are provided or not)? + +Answer: [Yes] + +Justification: This paper fully discloses all the information needed to reproduce the main experimental results. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- If the paper includes experiments, a No answer to this question will not be perceived well by the reviewers: Making the paper reproducible is important, regardless of whether the code and data are provided or not. +- If the contribution is a dataset and/or model, the authors should describe the steps taken to make their results reproducible or verifiable. +- Depending on the contribution, reproducibility can be accomplished in various ways. For example, if the contribution is a novel architecture, describing the architecture fully might suffice, or if the contribution is a specific model and empirical evaluation, it may be necessary to either make it possible for others to replicate the model with the same dataset, or provide access to the model. In general, releasing code and data is often one good way to accomplish this, but reproducibility can also be provided via detailed instructions for how to replicate the results, access to a hosted model (e.g., in the case of a large language model), releasing of a model checkpoint, or other means that are appropriate to the research performed. +- While NeurIPS does not require releasing code, the conference does require all submissions to provide some reasonable avenue for reproducibility, which may depend on the nature of the contribution. For example +(a) If the contribution is primarily a new algorithm, the paper should make it clear how to reproduce that algorithm. +(b) If the contribution is primarily a new model architecture, the paper should describe the architecture clearly and fully. +(c) If the contribution is a new model (e.g., a large language model), then there should either be a way to access this model for reproducing the results or a way to reproduce the model (e.g., with an open-source dataset or instructions for how to construct the dataset). +(d) We recognize that reproducibility may be tricky in some cases, in which case authors are welcome to describe the particular way they provide for reproducibility. In the case of closed-source models, it may be that access to the model is limited in some way (e.g., to registered users), but it should be possible for other researchers to have some path to reproducing or verifying the results. + +# 5. Open access to data and code + +Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material? + +Answer: [No] + +Justification: We have provided sufficient information to reproduce the results, and we will release the data and code upon cleanup. + +Guidelines: + +- The answer NA means that paper does not include experiments requiring code. +- Please see the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details. +- While we encourage the release of code and data, we understand that this might not be possible, so "No" is an acceptable answer. Papers cannot be rejected simply for not including code, unless this is central to the contribution (e.g., for a new open-source benchmark). +- The instructions should contain the exact command and environment needed to run to reproduce the results. See the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details. +- The authors should provide instructions on data access and preparation, including how to access the raw data, preprocessed data, intermediate data, and generated data, etc. +- The authors should provide scripts to reproduce all experimental results for the new proposed method and baselines. If only a subset of experiments are reproducible, they should state which ones are omitted from the script and why. +- At submission time, to preserve anonymity, the authors should release anonymized versions (if applicable). +- Providing as much information as possible in supplemental material (appended to the paper) is recommended, but including URLs to data and code is permitted. + +# 6. Experimental setting/details + +Question: Does the paper specify all the training and test details (e.g., data splits, hyperparameters, how they were chosen, type of optimizer, etc.) necessary to understand the results? + +Answer: [Yes] + +Justification: We have described all training details in the paper. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The experimental setting should be presented in the core of the paper to a level of detail that is necessary to appreciate the results and make sense of them. +- The full details can be provided either with the code, in appendix, or as supplemental material. + +# 7. Experiment statistical significance + +Question: Does the paper report error bars suitably and correctly defined or other appropriate information about the statistical significance of the experiments? + +Answer: [No] + +Justification: Due to limitations in computational resource access and training time, we were unable to calculate error bars. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The authors should answer "Yes" if the results are accompanied by error bars, confidence intervals, or statistical significance tests, at least for the experiments that support the main claims of the paper. + +- The factors of variability that the error bars are capturing should be clearly stated (for example, train/test split, initialization, random drawing of some parameter, or overall run with given experimental conditions). +- The method for calculating the error bars should be explained (closed form formula call to a library function, bootstrap, etc.) +- The assumptions made should be given (e.g., Normally distributed errors). +- It should be clear whether the error bar is the standard deviation or the standard error of the mean. +- It is OK to report 1-sigma error bars, but one should state it. The authors should preferably report a 2-sigma error bar than state that they have a $96\%$ CI, if the hypothesis of Normality of errors is not verified. +- For asymmetric distributions, the authors should be careful not to show in tables or figures symmetric error bars that would yield results that are out of range (e.g. negative error rates). +- If error bars are reported in tables or plots, The authors should explain in the text how they were calculated and reference the corresponding figures or tables in the text. + +# 8. Experiments compute resources + +Question: For each experiment, does the paper provide sufficient information on the computer resources (type of compute workers, memory, time of execution) needed to reproduce the experiments? + +Answer: [Yes] + +Justification: The computational resources required for the experiments are detailed in our paper. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The paper should indicate the type of compute workers CPU or GPU, internal cluster or cloud provider, including relevant memory and storage. +- The paper should provide the amount of compute required for each of the individual experimental runs as well as estimate the total compute. +- The paper should disclose whether the full research project required more compute than the experiments reported in the paper (e.g., preliminary or failed experiments that didn't make it into the paper). + +# 9. Code of ethics + +Question: Does the research conducted in the paper conform, in every respect, with the NeurIPS Code of Ethics https://neurips.cc/public/EthicsGuidelines? + +Answer: [Yes] + +Justification: We fully comply with the NeurIPS Code of Ethics. + +Guidelines: + +The answer NA means that the authors have not reviewed the NeurIPS Code of Ethics +- If the authors answer No, they should explain the special circumstances that require a deviation from the Code of Ethics. +- The authors should make sure to preserve anonymity (e.g., if there is a special consideration due to laws or regulations in their jurisdiction). + +# 10. Broader impacts + +Question: Does the paper discuss both potential positive societal impacts and negative societal impacts of the work performed? + +Answer: [NA] + +There is no societal impact of the work performed. + +Guidelines: + +- The answer NA means that there is no societal impact of the work performed. + +- If the authors answer NA or No, they should explain why their work has no societal impact or why the paper does not address societal impact. +- Examples of negative societal impacts include potential malicious or unintended uses (e.g., disinformation, generating fake profiles, surveillance), fairness considerations (e.g., deployment of technologies that could make decisions that unfairly impact specific groups), privacy considerations, and security considerations. +- The conference expects that many papers will be foundational research and not tied to particular applications, let alone deployments. However, if there is a direct path to any negative applications, the authors should point it out. For example, it is legitimate to point out that an improvement in the quality of generative models could be used to generate deepfakes for disinformation. On the other hand, it is not needed to point out that a generic algorithm for optimizing neural networks could enable people to train models that generate Deepfakes faster. +- The authors should consider possible harms that could arise when the technology is being used as intended and functioning correctly, harms that could arise when the technology is being used as intended but gives incorrect results, and harms following from (intentional or unintentional) misuse of the technology. +- If there are negative societal impacts, the authors could also discuss possible mitigation strategies (e.g., gated release of models, providing defenses in addition to attacks, mechanisms for monitoring misuse, mechanisms to monitor how a system learns from feedback over time, improving the efficiency and accessibility of ML). + +# 11. Safeguards + +Question: Does the paper describe safeguards that have been put in place for responsible release of data or models that have a high risk for misuse (e.g., pretrained language models, image generators, or scraped datasets)? + +Answer: [NA] + +Justification: Our models are not high risk and do not require safeguards. + +Guidelines: + +- The answer NA means that the paper poses no such risks. +- Released models that have a high risk for misuse or dual-use should be released with necessary safeguards to allow for controlled use of the model, for example by requiring that users adhere to usage guidelines or restrictions to access the model or implementing safety filters. +- Datasets that have been scraped from the Internet could pose safety risks. The authors should describe how they avoided releasing unsafe images. +- We recognize that providing effective safeguards is challenging, and many papers do not require this, but we encourage authors to take this into account and make a best faith effort. + +# 12. Licenses for existing assets + +Question: Are the creators or original owners of assets (e.g., code, data, models), used in the paper, properly credited and are the license and terms of use explicitly mentioned and properly respected? + +Answer: [Yes] + +Justification: We cited the creators and complied with the licenses and terms of use. + +Guidelines: + +- The answer NA means that the paper does not use existing assets. +- The authors should cite the original paper that produced the code package or dataset. +- The authors should state which version of the asset is used and, if possible, include a URI. +- The name of the license (e.g., CC-BY 4.0) should be included for each asset. +- For scraped data from a particular source (e.g., website), the copyright and terms of service of that source should be provided. + +- If assets are released, the license, copyright information, and terms of use in the package should be provided. For popular datasets, paperswithcode.com/datasets has curated licenses for some datasets. Their licensing guide can help determine the license of a dataset. +- For existing datasets that are re-packaged, both the original license and the license of the derived asset (if it has changed) should be provided. +- If this information is not available online, the authors are encouraged to reach out to the asset's creators. + +# 13. New assets + +Question: Are new assets introduced in the paper well documented and is the documentation provided alongside the assets? + +Answer: [NA] + +Justification: Our paper does not release new assets. + +Guidelines: + +- The answer NA means that the paper does not release new assets. +- Researchers should communicate the details of the dataset/code/model as part of their submissions via structured templates. This includes details about training, license, limitations, etc. +- The paper should discuss whether and how consent was obtained from people whose asset is used. +- At submission time, remember to anonymize your assets (if applicable). You can either create an anonymized URL or include an anonymized zip file. + +# 14. Crowdsourcing and research with human subjects + +Question: For crowdsourcing experiments and research with human subjects, does the paper include the full text of instructions given to participants and screenshots, if applicable, as well as details about compensation (if any)? + +Answer: [NA] + +Justification: The paper does not involve crowdsourcing or research with human subjects. + +Guidelines: + +- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. +- Including this information in the supplemental material is fine, but if the main contribution of the paper involves human subjects, then as much detail as possible should be included in the main paper. +- According to the NeurIPS Code of Ethics, workers involved in data collection, curation, or other labor should be paid at least the minimum wage in the country of the data collector. + +# 15. Institutional review board (IRB) approvals or equivalent for research with human subjects + +Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals (or an equivalent approval/review based on the requirements of your country or institution) were obtained? + +Answer: [NA] + +Justification: The paper does not involve crowdsourcing or research with human subjects. + +Guidelines: + +- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. +- Depending on the country in which research is conducted, IRB approval (or equivalent) may be required for any human subjects research. If you obtained IRB approval, you should clearly state this in the paper. + +We recognize that the procedures for this may vary significantly between institutions and locations, and we expect authors to adhere to the NeurIPS Code of Ethics and the guidelines for their institution. +- For initial submissions, do not include any information that would break anonymity (if applicable), such as the institution conducting the review. + +# 16. Declaration of LLM usage + +Question: Does the paper describe the usage of LLMs if it is an important, original, or non-standard component of the core methods in this research? Note that if the LLM is used only for writing, editing, or formatting purposes and does not impact the core methodology, scientific rigorousness, or originality of the research, declaration is not required. + +Answer: [NA] + +Justification: The core method development in this research does not involve LLMs as any important, original, or non-standard components. + +Guidelines: + +- The answer NA means that the core method development in this research does not involve LLMs as any important, original, or non-standard components. +- Please refer to our LLM policy (https://neurips.cc/Conferences/2025/LLM) for what should or should not be described. \ No newline at end of file diff --git a/NeurIPS/2025/3DPE-Gaze_Unlocking the Potential of 3D Facial Priors for Generalized Gaze Estimation/images.zip b/NeurIPS/2025/3DPE-Gaze_Unlocking the Potential of 3D Facial Priors for Generalized Gaze Estimation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..c91dd222d9afeea1cbee708eeb54ac2919516f77 --- /dev/null +++ b/NeurIPS/2025/3DPE-Gaze_Unlocking the Potential of 3D Facial Priors for Generalized Gaze Estimation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:942692148e4821a1824d9bfb3af2734974f53693d49eac914068e6f65660117d +size 535535 diff --git a/NeurIPS/2025/3DPE-Gaze_Unlocking the Potential of 3D Facial Priors for Generalized Gaze Estimation/layout.json b/NeurIPS/2025/3DPE-Gaze_Unlocking the Potential of 3D Facial Priors for Generalized Gaze Estimation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..d3043ca382e18b1b1b224bfd50ffaff5b271c30b --- /dev/null +++ b/NeurIPS/2025/3DPE-Gaze_Unlocking the Potential of 3D Facial Priors for Generalized Gaze Estimation/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dbe4fd25ca1917b9529ade30686483963fbba06af025c031bdcfecc16cb4df73 +size 585978 diff --git a/NeurIPS/2025/3DRS_ MLLMs Need 3D-Aware Representation Supervision for Scene Understanding/69be3fec-bdc4-4278-88b7-31546d832ba5_content_list.json b/NeurIPS/2025/3DRS_ MLLMs Need 3D-Aware Representation Supervision for Scene Understanding/69be3fec-bdc4-4278-88b7-31546d832ba5_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..fe7b7173abe6a26d106b1737628b0010a85b8770 --- /dev/null +++ b/NeurIPS/2025/3DRS_ MLLMs Need 3D-Aware Representation Supervision for Scene Understanding/69be3fec-bdc4-4278-88b7-31546d832ba5_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:74aad362bfd1789f2b7e1dabd9af897e1dfe983ceedd8322ad85bcc20b767930 +size 180889 diff --git a/NeurIPS/2025/3DRS_ MLLMs Need 3D-Aware Representation Supervision for Scene Understanding/69be3fec-bdc4-4278-88b7-31546d832ba5_model.json b/NeurIPS/2025/3DRS_ MLLMs Need 3D-Aware Representation Supervision for Scene Understanding/69be3fec-bdc4-4278-88b7-31546d832ba5_model.json new file mode 100644 index 0000000000000000000000000000000000000000..79ed9fdd6ae8293b7b811c363dff75c8c0926808 --- /dev/null +++ b/NeurIPS/2025/3DRS_ MLLMs Need 3D-Aware Representation Supervision for Scene Understanding/69be3fec-bdc4-4278-88b7-31546d832ba5_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6794498e050e0a31c5bce9d8a8f378459790494eb40fe9f62e49e8fd0e5eb104 +size 229900 diff --git a/NeurIPS/2025/3DRS_ MLLMs Need 3D-Aware Representation Supervision for Scene Understanding/69be3fec-bdc4-4278-88b7-31546d832ba5_origin.pdf b/NeurIPS/2025/3DRS_ MLLMs Need 3D-Aware Representation Supervision for Scene Understanding/69be3fec-bdc4-4278-88b7-31546d832ba5_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..082085ecedf87d75aaea016fd245f335cd291b6b --- /dev/null +++ b/NeurIPS/2025/3DRS_ MLLMs Need 3D-Aware Representation Supervision for Scene Understanding/69be3fec-bdc4-4278-88b7-31546d832ba5_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fb63f021b65549fb6539a4a5a9978277390227f87f86c525e074c4404a2d0210 +size 6958174 diff --git a/NeurIPS/2025/3DRS_ MLLMs Need 3D-Aware Representation Supervision for Scene Understanding/full.md b/NeurIPS/2025/3DRS_ MLLMs Need 3D-Aware Representation Supervision for Scene Understanding/full.md new file mode 100644 index 0000000000000000000000000000000000000000..4df0daa9811264b1b5b82e5ee7cb1ae93f0fef5d --- /dev/null +++ b/NeurIPS/2025/3DRS_ MLLMs Need 3D-Aware Representation Supervision for Scene Understanding/full.md @@ -0,0 +1,899 @@ +# 3DRS: MLLMs Need 3D-Aware Representation Supervision for Scene Understanding + +Xiaohu Huang + +Jingjing Wu2 + +Qunyi Xie + +Kai Han* + +Visual AI Lab, The University of Hong Kong + +$^{2}$ Department of Computer Vision Technology (VIS), Baidu Inc. huangxiaohu@connect.hku.hk, kaihanx@hku.hk + +# Abstract + +Recent advances in scene understanding have leveraged multimodal large language models (MLLMs) for 3D reasoning by capitalizing on their strong 2D pretraining. However, the lack of explicit 3D data during MLLM pretraining limits 3D representation capability. In this paper, we investigate the 3D-awareness of MLLMs by evaluating multi-view correspondence and reveal a strong positive correlation between the quality of 3D-aware representation and downstream task performance. Motivated by this, we propose 3DRS, a framework that enhances MLLM 3D Representation learning by introducing Supervision from pretrained 3D foundation models. Our approach aligns MLLM visual features with rich 3D knowledge distilled from 3D models, effectively improving scene understanding. Extensive experiments across multiple benchmarks and MLLMs—including visual grounding, captioning, and question answering—demonstrate consistent performance gains. + +Project page: https://visual-ai.github.io/3drs + +![](images/0e161aa6341c8f1bb9eda1aad711a5d82024fc73f041693abd5fdcec92be2666.jpg) +(a) Overview of 3DRS. + +![](images/0da1c3c58757c8b06468fe168c96a231e430e1bff533ba087dd70490e02b8033.jpg) +(b) Performance Improvement. +Figure 1: Enhancing 3D awareness of MLLMs to improve downstream performance. (a) Besides the common text supervision for MLLMs, 3DRS adopts 3D foundation models to supervise 3D-aware visual representation learning in MLLMs. (b) Combined with 3DRS, we achieve consistent performance improvement across multiple MLLMs and benchmarks. + +# 1 Introduction + +Scene understanding serves as a cornerstone for interpreting 3D environments, enabling a wide range of critical applications ranging from robotic navigation to augmented reality. The recent emergence of large language models (LLMs) [42, 17, 18] has sparked innovative research aimed at endowing these models with scene comprehension capabilities. One major line of research [50, 24, 36, 19, 24, 40, 41, 9, 14, 8, 54] utilizes point cloud encoders—either independently or in combination with multi-view images—to extract 3D representations, which are subsequently projected into a language-aligned space for LLMs. However, these approaches are constrained by the scarcity of paired 3D-text datasets, which impedes effective cross-modal feature alignment. + +In response to this challenge, recent state-of-the-art methods [64, 62, 16, 22, 37] have shifted towards leveraging multi-view images exclusively, drawing inspiration from the success of large-scale visual-language pretraining in multimodal LLMs (MLLMs) [29, 27, 59, 1, 46?]. These approaches aim to transfer 2D visual understanding to 3D scene comprehension by injecting 3D priors, such as 3D positional embeddings, into the models, thereby allowing MLLMs to capitalize on their extensive pretrained 2D knowledge for 3D interpretation. + +Despite these advancements, genuine 3D scene understanding fundamentally requires models to capture intrinsic 3D attributes and spatial structures to comprehend scenes. The absence of explicit 3D data during MLLM pretraining reveals a significant gap, which motivates our core investigation centered around the following questions: (1) How can we evaluate the ability of MLLMs to learn 3D-aware representations? (2) How does the quality of 3D feature learning influence downstream scene understanding performance? (3) What methods can enhance 3D-aware representation learning within MLLM frameworks? While several prior works [53, 28, 15, 33] have attempted to probe the 3D awareness of 2D vision foundation models, systematic investigation into 3D-aware representation learning in MLLMs remains largely unexplored. This gap is particularly crucial given the growing adoption of MLLMs in multimodal 3D understanding tasks. Our study aims to address this overlooked area and provide new insights into 3D representation learning within the MLLM paradigm. + +For the first question, we conduct comprehensive experiments to evaluate the 3D awareness on three representative MLLMs, including LLaVA-Next-Video [59], LLaVA-One-Vision [27], and Qwen2-VL [46], following the finetuning settings of Video-3D LLM [62]. Specifically, we assess 3D awareness via view equivariance, quantifying it by computing the feature similarity between corresponding pairs from the same 3D voxel across different views. This evaluation requires MLLMs to associate the same object across multiple views, thereby reflecting their capacity for 3D representation. Our analysis encompasses six datasets spanning tasks such as 3D grounding [5], captioning [12], and question answering [2]. + +To address the second question, we systematically analyze model performance across these datasets and observe that samples with higher correspondence scores—i.e., those exhibiting stronger 3D awareness—consistently lead to improved performance. This finding demonstrates a strong positive correlation between the quality of 3D-aware representations and downstream scene understanding performance, highlighting the necessity of enhancing 3D feature learning in MLLMs. + +In response to the third question and building upon our earlier findings, we first introduce a view equivalence supervision strategy for MLLMs, encouraging alignment between feature pairs corresponding to the same 3D location across different views (positive pairs) while discouraging similarity among unrelated pairs (negative pairs). While this approach results in some performance gains, the supervision provided by such handcrafted, single-task objectives is inherently limited for 3D learning. + +In contrast, recent 3D foundation models such as VGGT [45] and FLARE [57] are pretrained end-to-end on multi-view image sequences spanning a diverse set of 3D geometric tasks—including not only correspondence learning, but also depth estimation and camera parameter prediction. This comprehensive pretraining enables them to encode rich 3D properties within their features. Building on this, we propose a framework, 3DRS, that leverages these pretrained models by using their features as alignment targets for the visual outputs of MLLMs, thereby facilitating more effective 3D-aware representation learning. Unlike previous 3D MLLM approaches, in addition to traditional text token supervision, our framework employs explicit 3D-specific supervision directly on scene visual tokens. As demonstrated in our experiments (see Fig. 1), incorporating this form of 3D supervision consistently improves performance across a range of MLLMs and benchmarks. Notably, our approach incurs no additional training overhead, since the supervisory features can be pre-extracted offline. + +We believe this design offers valuable new insights for applying 3D foundation models in scene understanding. The key contribution of this paper can be summarized as follows: + +- We conduct a systematic evaluation of the 3D-awareness of MLLMs using multi-view correspondence metrics, and observe a strong positive correlation between 3D-aware representation quality and downstream scene understanding performance across diverse tasks, datasets, and models. +- We propose a 3D-aware representation supervision framework that aligns MLLM visual features with those of a 3D geometry-pretrained model, enabling effective 3D feature learning. +- Extensive experiments demonstrate consistent performance improvements across multiple MLLMs and 3D scene understanding benchmarks, validating the effectiveness and generality of our approach. + +# 2 Method + +# 2.1 Investigating 3D-Aware Representation Learning in MLLMs + +# 2.1.1 Preliminaries + +A MLLM typically consists of two main components: an image encoder $\mathcal{E}_{\mathrm{img}}$ and a text decoder $\mathcal{T}$ . In this work, the input to our MLLM comprises a set of $N$ multi-view images $\mathcal{I} = \{I_1, I_2, \dots, I_N\}$ , each associated with per-pixel 3D coordinates $\mathcal{C} = \{\mathbf{C}_1, \mathbf{C}_2, \dots, \mathbf{C}_N\}$ , where $\mathbf{C}_i \in \mathbb{R}^{H \times W \times 3}$ for image $I_i$ of size $H \times W$ . The 3D coordinates for each pixel are computed from the depth map and the corresponding camera intrinsic and extrinsic parameters; detailed formulas and procedures can be found in the App. A.1. + +The MLLM receives both multi-view images and language instructions as input. Internally, for each image $I_{i}$ , the image encoder $\mathcal{E}_{\mathrm{img}}$ extracts visual features $\mathbf{F}_i\in \mathbb{R}^{H\times W\times d}$ , where $d$ is the feature dimension. Following Video3DLLM [62], we encode the per-pixel 3D coordinates via a positional encoding function $\phi (\cdot)$ and inject this information into the image features by addition: + +$$ +\mathbf {F} _ {i} ^ {3 D} = \mathbf {F} _ {i} + \phi (\mathbf {C} _ {i}). +$$ + +This design allows the MLLM to inherit 2D perceptual knowledge from pretraining while equipping it with explicit 3D priors. + +During finetuning, the MLLM—which we denote as $f_{\theta}$ —passes visual features $\{\mathbf{F}_i^{3D}\}_{i=1}^N$ with the instruction tokens to the text decoder for autoregressive text generation. After the processing of the text decoder, we refer to the final per-pixel visual embedding of pixel $p$ in image $I_i$ from LLM as $\mathbf{f}_i(p)$ . The model is optimized by minimizing the standard cross-entropy loss: + +$$ +\mathcal {L} _ {\mathrm {C E}} = - \sum_ {t = 1} ^ {T} \log p _ {\theta} \left(y _ {t} \mid y _ {< t}, \left\{I _ {i}, \mathbf {C} _ {i} \right\} _ {i = 1} ^ {N}, \text {i n s t r u c t i o n}\right), +$$ + +where $y_{t}$ is the $t$ -th output token, and $p_{\theta}$ is the probability predicted by the model given all previous tokens and the multimodal context (i.e., images and instructions). + +# 2.1.2 Assessing 3D Feature Learning via Multi-View Correspondence + +Inspired by the crucial role of cross-view correspondences in 3D modeling [20], we propose a correspondence-based evaluation framework. Multi-view correspondences are fundamental in 3D vision, serving as essential cues for core tasks such as ray retriangulation [20], bundle adjustment [43], and pose estimation [39]. They are also critical for downstream applications like instance recognition and retrieval [39, 51, 49]. Therefore, we adopt multi-view correspondence analysis as a proxy to evaluate the 3D representations of MLLMs. This approach requires the model to accurately associate and align objects or regions that occupy the same position in 3D space across different viewpoints. + +Voxelization and correspondence pair construction. We first voxelize the 3D scene into a regular grid of voxels $\mathcal{V} = \{v_{1},\dots,v_{M}\}$ . For each view $I_{i}$ , given its per-pixel 3D coordinates $\mathbf{C}_i$ , we assign + +every pixel's feature $\mathbf{f}_i(p)$ to a voxel according to its 3D position. Features from different views that fall into the same voxel $v_{k}$ are regarded as correspondence pairs. + +Feature similarity and correspondence scores. Let $\mathcal{P}_k$ denote all correspondence feature pairs in voxel $v_{k}$ , i.e., all pairs $(\mathbf{f}_i(p),\mathbf{f}_j(q))$ where both pixels $p$ and $q$ from images $I_{i}$ and $I_{j}$ are assigned to $v_{k}$ with $i\neq j$ . For any pair of visual features $(\mathbf{f}_a,\mathbf{f}_b)$ from the last layer of MLLM, feature similarity is measured by the cosine similarity: + +$$ +S (\mathbf {f} _ {a}, \mathbf {f} _ {b}) = \frac {\mathbf {f} _ {a} ^ {\top} \mathbf {f} _ {b}}{\| \mathbf {f} _ {a} \| \cdot \| \mathbf {f} _ {b} \|}. +$$ + +For each sequence, we compute: + +$$ +\bar {S} = \frac {1}{| \mathcal {P} |} \sum_ {(\mathbf {f} _ {a}, \mathbf {f} _ {b}) \in \mathcal {P}} S (\mathbf {f} _ {a}, \mathbf {f} _ {b}), +$$ + +where $\bar{S}$ and $\mathcal{P}$ denote the correspondence score for each sequence and all the correspondence pairs in this sequence. A higher correspondence score indicates that the model produces more consistent features across views for the same 3D spatial location, reflecting stronger 3D-aware representation learning. + +# 2.1.3 Quality of 3D Feature vs. Downstream Task Performance. + +We evaluate three representative MLLMs, LLaVA-Next-Video, LLaVA-OneVision, and Qwen2-VL, on five diverse 3D scene understanding benchmarks, including visual grounding (Multi3DRefer, ScanRefer), captioning (Scan2Cap), and question answering (ScanQA, SQA3D). All benchmarks are based on multi-view RGBD sequences. The three MLLMs respectively emphasize video understanding, joint image-video reasoning, and advanced arbitrary-resolution visual encoding. + +To analyze the relationship between 3D feature learning and downstream task performance, we sort samples within each dataset by their correspondence scores and divide them into four quartiles (Q1-Q4, lowest to highest). From Fig. 2, we observe a clear trend: as the correspondence score increases, the model's performance on the downstream task consistently improves. This strong positive correlation demonstrates the critical importance of 3D-aware representation quality for effective scene understanding in MLLMs. + +![](images/d739331ec7874af687553c9f480fa9d734f125b4b2fedb4b6c1623b94158f8aa.jpg) + +![](images/3d2ffa28c0c5496dbb57e48a28a1604f10451325b58cb20bdad595347e0c9d3c.jpg) + +![](images/f3cb86b2291b58d15c4392ac98e69bcafa35757d213c431bb4989b6bef0f8a08.jpg) +Figure 2: Performance across correspondence score quartiles. Model performance across correspondence score quartiles (Q1–Q4, lowest to highest) for each dataset. Samples were divided into quartiles by their correspondence scores. A clear trend is observed: model accuracy improves as the correspondence score increases. + +These findings highlight the need for strategies to further enhance 3D-aware representation learning in MLLMs, which we address in the next section. + +# 2.2 Enhancing 3D-Aware Representation Learning in MLLMs + +# 2.2.1 Correspondence-based 3D Supervision Loss + +Inspired by our correspondence-based evaluation, a straightforward approach is to directly supervise the MLLM's visual features to be consistent for matched 3D locations across views and dissimilar for mismatched locations. We let $\mathcal{P}_k^+$ denote all positive feature pairs in voxel $v_{k}$ , i.e., all pairs $(\mathbf{f}_i(p),\mathbf{f}_j(q))$ where pixels $p$ and $q$ from images $I_{i}$ and $I_{j}$ are assigned to $v_{k}$ with $i\neq j$ . Similarly, $\mathcal{P}_k^-$ denotes negative pairs between $v_{k}$ and any other voxel $v_{l}$ ( $l\neq k$ ). We supervise these objectives directly using a simple loss function by maximizing the feature similarity in $\mathcal{P}^+$ and minimizing that in $\mathcal{P}^{-}$ : + +$$ +\mathcal {L} _ {\mathrm {c o r r}} ^ {+} = \frac {1}{| \mathcal {P} ^ {+} |} \sum_ {(\mathbf {f} _ {a}, \mathbf {f} _ {b}) \in \mathcal {P} ^ {+}} \left[ 1 - S (\mathbf {f} _ {a}, \mathbf {f} _ {b}) \right], +$$ + +$$ +\mathcal {L} _ {\mathrm {c o r r}} ^ {-} = \frac {1}{| \mathcal {P} ^ {-} |} \sum_ {(\mathbf {f} _ {a}, \mathbf {f} _ {b}) \in \mathcal {P} ^ {-}} S (\mathbf {f} _ {a}, \mathbf {f} _ {b}). +$$ + +The overall correspondence loss is a weighted sum: + +$$ +\mathcal {L} _ {\mathrm {c o r r}} = \mathcal {L} _ {\mathrm {c o r r}} ^ {+} + \mathcal {L} _ {\mathrm {c o r r}} ^ {-}. +$$ + +By directly supervising positive pairs to be similar and negative pairs to be dissimilar, this correspondence loss encourages the model to learn multi-view 3D correspondences, thus enhancing the 3D-awareness of the learned representations. As will be shown in the experiments Sec. 3, supplementing the standard cross-entropy objective with $\mathcal{L}_{\mathrm{corr}}$ leads to improvements in downstream task performance. However, as this loss primarily targets view equivariance, the range of 3D properties captured remains limited, motivating the need for richer supervision. + +# 2.2.2 3D Foundation Model-Guided Feature Distillation + +To overcome the inherent limitations of single-task supervision, we further introduce a knowledge distillation framework, 3DRS, that leverages the rich 3D priors embedded in 3D foundation models, e.g., FLARE and VGGT. These models are pretrained on a wide array of 3D geometric tasks—including correspondence learning, camera parameter estimation, multi-view depth prediction, and dense point cloud reconstruction—which enables them to extract robust and highly 3D-aware visual features from multi-view image sequences. + +Distillation target preparation. As shown in Fig. 3a, given a set of multi-view images $\mathcal{I}$ for a scene, we first input them into a pretrained 3D foundation model $g$ , which outputs a collection of per-pixel visual features $\{\mathbf{f}_i^{3\mathrm{D}}(p)\}$ for each image $I_{i}$ and pixel $p$ . Since the spatial resolution of these features may differ from those of the MLLM outputs $\{\mathbf{f}_i(p)\}$ , we apply 2D average pooling to the 3D foundation model's output to match the MLLM feature map size. + +Feature alignment and loss. To align the MLLM's per-pixel visual features with the 3D foundation model, we first process each $\mathbf{f}_i(p)$ with a two-layer MLP (denoted as $\mathrm{MLP}_{\mathrm{align}}$ ) to ensure compatibility in feature dimension: + +$$ +\hat {\mathbf {f}} _ {i} (p) = \mathrm {M L P} _ {\mathrm {a l i g n}} (\mathbf {f} _ {i} (p)). +$$ + +We then employ a distillation loss based on euclidean similarity to maximize the alignment between the MLLM features $\hat{\mathbf{f}}_i(p)$ and the corresponding 3D foundation model features $\mathbf{f}_i^{3\mathrm{D}}(p)$ : + +$$ +\mathcal {L} _ {\mathrm {a l i g n}} = - \frac {1}{N H W} \sum_ {i = 1} ^ {N} \sum_ {p \in I _ {i}} S \left(\hat {\mathbf {f}} _ {i} (p), \mathbf {f} _ {i} ^ {\mathrm {3 D}} (p)\right), +$$ + +where $S(\cdot, \cdot)$ denotes cosine similarity, and the sum is calculated over all pixels and views in the batch. + +![](images/aeb1492fb948b4276e810c3d746986d489cb702b9e94d40e536d4a405e44aa8c.jpg) +(a) Details of 3DRS. + +![](images/815a9fe25da285a8ba135a3989d150d12084607a86c4bfcf806e97ea1964b48f.jpg) +(b) Comparison of correspondence learning. +Figure 3: (a) 3DRS uses a 3D foundation model to supervise the visual representation of the MLLM. (b) 3DRS effectively improves the correspondence learning for MLLMs. + +Overall training objective. The final training objective for the MLLM combines the standard cross-entropy loss for text generation and the 3D foundation model distillation loss: + +$$ +\mathcal {L} _ {\text {t o t a l}} = \mathcal {L} _ {\text {C E}} + \mathcal {L} _ {\text {a l i g n}}. +$$ + +This approach enables the MLLM to inherit comprehensive 3D knowledge from powerful geometry-pretrained models, facilitating the learning of richer and more robust 3D-aware representations. Importantly, the distillation targets from the 3D foundation model can be precomputed offline, introducing no additional overhead during MLLM fine-tuning. + +As illustrated in Fig. 3b, we compare the correspondence scores before and after applying our 3DRS, where VGGT serves as the foundation model. The results consistently demonstrate that introducing 3DRS leads to substantial improvements in correspondence learning ability across all evaluated MLLMs and benchmarks. This proves the effectiveness of leveraging a pretrained 3D foundation model as a teacher model for enhancing 3D-aware representation learning in MLLMs. More comprehensive experimental results and analyses are detailed in Sec. 3. + +# 3 Experiments + +# 3.1 Datasets and Evaluation Metrics + +Datasets. We evaluate our approach on six benchmarks that collectively span key challenges in 3D scene understanding. ScanRefer [5] focuses on localizing objects using free-form language, while Multi3DRefer [58] generalizes this to queries referencing zero, one, or multiple objects, better reflecting real-world ambiguity. Scan2Cap [12] addresses dense captioning by pairing detected objects in 3D scans with natural language descriptions. For question answering, ScanQA [2] tasks models with answering open-ended questions grounded in 3D geometry and semantics, and SQA3D [32] goes further by requiring situated reasoning: agents must interpret their position and context to answer complex queries. All these datasets are sourced from the richly annotated ScanNet [13] corpus, and we follow standard validation and test splits as established in prior work [24, 64, 9, 62]. Besides, VSI-Bench [52] is used to evaluate the performance on visual-based spatial understanding tasks, which are composed of numerical and multiple-choice questions. The statistics of training sets are detailed in the App. A.2. + +Evaluation metrics. For ScanRefer, we report accuracy at IoU thresholds of 0.25 and 0.5 (Acc@0.25, Acc@0.5). Multi3DRefer uses F1 scores at matching IoU thresholds. Scan2Cap is evaluated by CIDEr and BLEU-4 scores at 0.5 IoU (C@0.5, B-4@0.5). ScanQA is assessed by CIDEr and exact match accuracy (C, EM), while SQA3D uses exact match accuracy as the metric. + +# 3.2 Implementation Details + +Our experiments leverage several MLLMs, including LLaVA-Next-Video 7B [59], LLaVA-OneVision 7B [27], and Qwen2-VL 7B [46]. In addition to these baselines, we systematically compare the + +Table 1: Performance comparison on 3D scene understanding benchmarks. Specialists are single-task methods, while generalists target multiple tasks. Bold denotes best performance. + +
MethodScanReferMulti3DReferScan2CapScanQASQA3D
Acc@0.25Acc@0.5F1@0.25F1@0.5C@0.5B-4@0.5CEMEM
Specialists
ScanRefer [5]37.324.3-------
MVT [26]40.833.3-------
3DVG-Trans [60]45.934.5-------
ViL3DRel [7]47.937.7-------
M3DRef-CLIP [58]51.944.742.8-38.4----
Scan2Cap [12]----35.222.4---
ScanQA [2]------64.921.147.2
3D-VisTA [65]50.645.8--66.934.069.622.448.5
Generalists
3D-LLM(Flamingo) [22]21.2-----59.220.4-
3D-LLM(BLIP2-flant5) [22]30.3-----69.420.5-
Chat-3D [47]------53.2--
Chat-3D v2 [24]42.538.445.141.663.931.887.6-54.7
LL3DA [8]----62.936.076.8--
SceneLLM [16]------80.027.253.6
LEO [25]----72.438.2101.421.550.0
Grounded 3D-LLM [9]47.944.145.240.670.635.572.7--
PQ3D [66]57.051.2-50.180.336.0--47.1
ChatScene [24]55.550.257.152.477.136.387.721.654.6
LLaVA-3D [64]54.142.4--79.241.191.727.055.6
Inst3D-LLM [54]57.851.658.353.579.738.388.624.6-
3D-LLaVA [14]51.240.6--78.836.992.6-54.5
Video-3D LLM [62]58.151.758.052.783.841.3102.130.158.6
3DRS62.956.160.454.986.141.6104.830.360.6
+ +effect of using 2D versus 3D foundation models as teachers for MLLM finetuning. The 2D teacher models include DINOv2 [34], MAE [21], and SigLIP [44], while the 3D teacher models comprise FLARE [57] and VGGT [45]. Unless stated otherwise, we use LLaVA-Next-Video as the MLLM and VGGT as the representation teacher for our experiments. + +For both training and inference, we uniformly sample 32 frames per scan to construct multi-view image sets. For evaluating the correspondence score, we use the voxel size of 0.1 for voxelization. All models are optimized using Adam, with a batch size of 16 and a warm-up ratio of 0.03. The learning rates are set to a maximum of $1 \times 10^{-5}$ for the language model and $2 \times 10^{-6}$ for the visual backbone during the warm-up period. During training for visual grounding and dense captioning, ground truth object regions are used as candidates, whereas during inference, we follow the procedure of [24, 25, 62] and employ Mask3D [38] to generate object proposals. For LLaVA-Next-Video and LLaVA-OneVision, we finetune all model parameters. For Qwen2-VL, due to GPU memory constraints, we finetune only the projector and the LLM components. We use 8 H100 NVIDIA GPUs for all experiments. + +# 3.3 Comparison with State-of-the-Art Models + +Table 1 presents a comprehensive comparison between our approach, task-specific specialist models—which require fine-tuning on individual datasets—and 3D generalist models that are capable of handling multiple tasks. Compared to specialist models, our approach achieves substantial performance improvements. This demonstrates the significant benefits brought by joint training and the LLM-based architecture, which contribute to superior generalization and feature integration compared to methods tailored for specific tasks. Furthermore, our method consistently outperforms 3D generalist approaches that utilize point clouds as input, such as LL3DA, Chat-3D, Grounded 3D-LLM, and 3D-LLaVA. Compared to Inst3D-LLM—which fuses multi-view images and point clouds—our approach also shows clear advantages, highlighting the strength of leveraging MLLMs as the backbone. Additionally, our method achieves considerable improvements over other MLLM-based methods, including LLaVA-3D and Video-3D LLM. These results collectively indicate that enhancing the 3D-awareness of MLLMs is highly effective for 3D scene understanding tasks, further validating the effectiveness of our proposed strategy. + +Table 2: Performance comparison on VSI-Bench. + +
MethodAvg.Obj. CountAbs. DistanceObj. SizeRoom SizeRel. Dist.Rel. Dir.Route PlanAppr. Order
GPT-4o [1]34.046.25.343.838.237.041.331.528.5
Gemini-1.5 Pro [? ]45.456.230.964.143.651.346.336.034.6
LongVA-7B [56]29.238.022.233.143.325.415.733.117.7
InternVL2-40B [11]37.041.326.248.227.547.632.727.844.7
LLaVA-Video-7B [59]35.648.514.047.824.243.542.434.030.6
LLaVA-Video-72B [59]40.948.922.857.435.342.436.735.048.6
LLaVA-OneVision-7B [27]32.447.720.747.442.535.229.424.4
LLaVA-OneVision-72B [27]40.243.523.957.637.542.539.932.544.6
3DRS45.968.734.853.656.640.943.230.439.2
+ +Table 2 compares our method with leading proprietary APIs and open-source models on VSI-Bench [52], covering tasks such as object counting, spatial distance estimation, size reasoning, and sequence understanding. 3DRS achieves the best open-source results on most metrics—including object count, absolute distance, room size, and appearance oder—and remains competitive with proprietary models. These results demonstrate the strong spatial reasoning, generalization, and comprehensive scene understanding capabilities of our approach across diverse 3D vision tasks. + +Table 3: Performance comparison of 3DRS when using with different MLLMs. + +
MethodScanReferMulti3DRefScan2CapScanQASQA3D
Acc@0.25Acc@0.5F1@0.25F1@0.5B-4@0.5C@0.5CEMEM
LLaVA-Next-Video [59]58.151.758.052.741.383.8102.130.158.6
LLaVA-Next-Video w/ 3DRS62.956.160.454.941.686.1104.830.360.6
LLaVA-OneVision [27]57.351.057.151.940.481.7101.729.057.4
LLaVA-OneVision w/ 3DRS61.855.060.555.041.283.1104.029.559.4
Qwen2-VL [46]57.050.856.251.439.579.497.528.758.6
Qwen2-VL w/ 3DRS60.153.558.554.540.981.699.228.960.0
+ +# 3.4 Diagnostic Study + +Effectiveness with different MLLMs. Table 3 demonstrates that integrating 3DRS with different MLLMs—LLaVA-Next-Video, LLaVA-OneVision, and Qwen2-VL—consistently boosts performance across all evaluated benchmarks. For example, LLaVA-Next-Video w/ 3DRS improves ScanRefer Acc@0.25 from 58.1 to 62.9, and Multi3DRef F1@0.25 from 58.0 to 60.4. Similar gains are observed for LLaVA-OneVision and Qwen2-VL, where 3DRS brings improvements on every dataset and metric. These results highlight the general applicability of our approach and its effectiveness in enhancing 3D scene understanding for various MLLMs. + +Comparison between 2D and 3D foundation models. Table 4 compares the performance of using 2D and 3D foundation models as representation supervisors. It is clear that 3D foundation models (FLARE and VGGT) outperform all 2D foundation models (MAE, Siglip2, Dinov2) across almost every metric. This performance gap can be attributed to the inherent difference in the prior knowledge captured by 2D and 3D foundation models. 3D models are pre-trained on large-scale 3D data and thus better capture geometric structure, spatial relationships, and depth information, which are critical for 3D scene understanding tasks. In contrast, 2D foundation models, trained on images, lack explicit 3D spatial priors and struggle to provide effective supervision for learning 3D-aware representations. This highlights the importance of 3D-specific foundation models for achieving strong results in downstream 3D tasks. + +Comparison of supervision signal. Table 5 shows that using correspondence loss for supervision brings improvements over the baseline, demonstrating the effectiveness of encouraging the model to learn multi-view correspondences. However, when 3D foundation model supervision is applied, the performance increases even further across all metrics. This indicates that 3D foundation models, with their rich 3D prior knowledge learned during pre-training, can more effectively enhance the 3D representation ability of MLLMs and yield greater gains for 3D understanding tasks. + +Comparison of supervision at different layers. Table 6 examines the effect of applying 3D foundation model supervision at different layers of the network. The results reveal that supervision + +at deeper layers, especially the last layer, leads to the highest performance. This is likely because deeper layers are closer to the output and thus have a more direct impact on the final predictions. Additionally, these layers possess more parameters and a greater capacity to fit 3D features, which results in larger improvements on downstream tasks. + +Table 4: Ablation study on using different 2D/3D foundation models as the representation supervisor. Bold denotes the best in each group. + +
Representation SupervisorScanReferMulti3DRefScan2CapScanQASQA3D
Acc@0.25Acc@0.5F1@0.25F1@0.5C@0.5B-4@0.5CEMEM
Baseline58.151.758.052.783.841.3102.130.158.6
2D Foundation Models
Siglip2 [44]58.252.959.753.181.740.2100.229.159.4
MAE [21]59.153.760.053.782.840.4102.529.559.2
Dinov2 [34]59.853.358.553.580.339.3103.529.660.1
3D Foundation Models
FLARE [57]62.155.759.854.886.642.5104.430.160.1
VGGT [45]62.956.160.454.986.141.6104.830.360.6
+ +# 4 Related Work + +# 4.1 Scene Understanding with Large Language Models + +LLMs, owing to their strong reasoning capabilities and remarkable success in 2D image understanding, have been widely applied to scene understanding tasks. Early works such as PointLLM [50], Point-Bind [19], GPT4Point [36], MiniGPT-3D [40], and Chat-3D [47] leverage the alignment between point cloud and text features to facilitate 3D scene comprehension. Building on this foundation, methods like Grounded 3D-LLM [9], LL3DA [8], 3D-LLaVA [14], and Inst3D-LLM [54] design more advanced cross-modal modules to better fuse multi-modal features, thereby enhancing scene representations. Furthermore, Chat-Scene [24] and Inst3D-LLM [54] exploit the complementary nature of 2D and 3D features to further boost scene understanding. + +Some recent approaches, such as 3D-LLM [22] and Scene LLM [16], employ multi-view inputs and introduce 3D priors to transform 2D representations into a 3D-aware format. Thanks to pre-training on large-scale image-text datasets, methods based on MLLMs are gaining increasing popularity in the field of scene understanding. For instance, LLaVA-3D [64] takes multi-view images as input and utilizes voxelization to reduce the dimensionality of representations, thus lowering computational costs while leveraging the strengths of MLLMs. However, many MLLMs require specially structured inputs, making them incompatible with certain approaches. Video 3D-LLM [62] and GPT4Scene [37] more naturally inherit the MLLM pipeline by introducing 3D priors—such as positional embeddings or spatial markers—enabling the model to better comprehend 3D scene content. + +Our work follows this line of MLLM-based scene understanding, aiming to probe the 3D-awareness of MLLMs and analyze their relationships with downstream tasks. In particular, we demonstrate that introducing guidance from 3D foundation models can effectively enhance the representational capability of MLLMs for 3D scene understanding. + +# 4.2 3D-Awareness in Vision Models + +Several studies have investigated 3D-awareness; however, most prior work has focused on pure vision models rather than MLLMs, and primarily leveraged 2D foundation models instead of 3D ones. For example, FiT3D [55], Probe3D [15], and Lexicon3D [33] empirically analyze the 3D-awareness of visual foundation models. CUA-O3D [28] proposes integrating multiple 2D foundation models for 3D scene understanding, while Yang et al.[53] evaluates and enhances the 3D-awareness of ViT-based models in various downstream tasks. Some previous 3D detection works [10, 48, 3] have focused on improving 3D representations for pure vision models or CLIP-style vision-language models (VLMs), primarily aiming to enhance geometric understanding and spatial localization within unimodal or early multimodal frameworks. In addition, several studies on scene understanding [30, 23, 35] have + +Table 5: Comparison of different supervision strategies. + +
MethodMulti3DRefScanRefer
Acc@0.25Acc@0.5F1@0.25F1@0.5
Baseline58.052.758.151.7
w/ Correspondence Loss60.153.359.153.7
w/ 3D Supervision62.956.160.454.9
+ +Table 6: 3D foundation model supervision at different layers. + +
LayerMulti3DRefScanRefer
Acc@0.25Acc@0.5F1@0.25F1@0.5
Last Layer62.956.160.454.9
3rd Last Layer61.754.959.754.3
5th Last Layer61.454.859.354.0
10th Last Layer59.153.653.353.8
+ +investigated various strategies for distilling 3D representations, such as transferring knowledge from 3D models to 2D networks or promoting cross-modal alignment. However, these efforts have not addressed the unique challenges presented by MLLMs, which require a more holistic integration of visual, linguistic, and spatial information. As a result, the potential of distilling 3D awareness into MLLMs for richer and more generalizable scene understanding remains largely unexplored. + +In contrast, our work specifically targets the 3D-awareness of MLLMs. Rather than enhancing 3D feature learning via 2D foundation models, we introduce 3D foundation models as supervisors to directly guide and improve the 3D representation capabilities of MLLMs. + +# 5 Conclusion + +In this paper, we present a comprehensive study of the 3D representation capabilities of multi-modal large language models (MLLMs) in the context of scene understanding. While most existing research has centered on leveraging 2D foundation models to improve visual reasoning in MLLMs, the role and utility of 3D foundation models in this setting remain largely unexplored. To bridge this gap, we propose 3DRS, a novel framework that introduces direct 3D-aware supervision to MLLMs by leveraging pretrained 3D foundation models as teachers. Our approach enables MLLMs to acquire richer geometric and spatial representations, facilitating more accurate and robust understanding of complex 3D scenes. Through extensive experiments on diverse 3D scene understanding benchmarks, we demonstrate that 3DRS consistently improves performance across a variety of tasks, such as object localization, spatial reasoning, and 3D question answering. These results highlight the unique advantages and significant potential of integrating 3D foundation models for advancing multimodal scene understanding. + +# 6 Limitation + +While our paper aims to enhance the 3D-awareness of MLLMs, the relatively limited size of the dataset used for finetuning—especially when compared to that used during the MLLM pretraining stage—may restrict the full realization of our approach's potential. Consequently, the improvements demonstrated in this work may only represent an initial step toward more robust 3D understanding. A promising direction for future research is to incorporate 3D-awareness learning into the pretraining stage of MLLMs, which could lead to fundamentally stronger models with deeper 3D comprehension. Besides, due to the distillation-based nature of our approach, the performance of our method is upper-bounded by the quality of the teacher 3D foundation model. Any limitations or failure modes of the teacher—such as inaccurate correspondence, erroneous depth estimation, or incomplete geometric representations—can be propagated to the student MLLM and may potentially mislead it during training. While our experiments demonstrate consistent improvements over strong baselines, it is possible that errors or biases in the teacher's predictions can negatively impact the downstream 3D reasoning abilities of the student model. We believe that, as 3D foundation models continue to rapidly advance, this limitation becomes less pronounced over time. + +# Acknowledgments + +This work is supported by Hong Kong Research Grant Council - General Research Fund (Grant No. 17213825). We would like to thank Weining Ren for the valuable and insightful discussions. + +# References + +[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. +[2] Daichi Azuma, Taiki Miyanishi, Shuhei Kurita, and Motoaki Kawanabe. Scanqa: 3d question answering for spatial scene understanding. In CVPR, 2022. License: Creative Commons Attribution-NonCommercial-ShareAlike 3.0 Unported License. +[3] Geonho Bang, Kwangjin Choi, Jisong Kim, Dongsuk Kum, and Jun Won Choi. Radardistill: Boosting radar-based object detection performance via knowledge distillation from lidar features. In CVPR, 2024. +[4] Daigang Cai, Lichen Zhao, Jing Zhang, Lu Sheng, and Dong Xu. 3djcg: A unified framework for joint dense captioning and visual grounding on 3d point clouds. In CVPR, 2022. +[5] Dave Zhenyu Chen, Angel X Chang, and Matthias Nießner. Scanrefer: 3d object localization in rgb-d scans using natural language. In ECCV, 2020. License: Creative Commons Attribution-NonCommercial-ShareAlike 3.0 Unported License. +[6] Dave Zhenyu Chen, Qirui Wu, Matthias Nießner, and Angel X. Chang. D3net: A unified speaker-listener architecture for 3d dense captioning and visual grounding. In ECCV, 2022. +[7] Shizhe Chen, Pierre-Louis Guhur, Makarand Tapaswi, Cordelia Schmid, and Ivan Laptev. Language conditioned spatial relation reasoning for 3d object grounding. In NeurIPS, 2022. +[8] Sijin Chen, Xin Chen, Chi Zhang, Mingsheng Li, Gang Yu, Hao Fei, Hongyuan Zhu, Jiayuan Fan, and Tao Chen. LL3DA: visual interactive instruction tuning for omni-3d understanding, reasoning, and planning. In CVPR, 2024. +[9] Yilun Chen, Shuai Yang, Haifeng Huang, Tai Wang, Runsen Xu, Ruiyuan Lyu, Dahua Lin, and Jiangmiao Pang. Grounded 3d-llm with referent tokens. arXiv preprint arXiv:2405.10370, 2024. +[10] Zehui Chen, Zhenyu Li, Shiquan Zhang, Liangji Fang, Qinhong Jiang, and Feng Zhao. Bevdistill: Cross-modal bev distillation for multi-view 3d object detection. arXiv preprint arXiv:2211.09386, 2022. +[11] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In CVPR, 2024. +[12] Zhenyu Chen, Ali Gholami, Matthias Nießner, and Angel X Chang. Scan2cap: Context-aware dense captioning in rgb-d scans. In CVPR, 2021. License: Creative Commons Attribution-NonCommercial-ShareAlike 3.0 Unported License. +[13] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In CVPR, 2017. License: ScanNet Terms of Use. +[14] Jiajun Deng, Tianyu He, Li Jiang, Tianyu Wang, Feras Dayoub, and Ian Reid. 3d-llava: Towards generalist 3d Imms with omni superpoint transformer. In CVPR, 2025. +[15] Mohamed El Banani, Amit Raj, Kevis-Kokitsi Maninis, Abhishek Kar, Yanzhen Li, Michael Rubinstein, Deqing Sun, Leonidas Guibas, Justin Johnson, and Varun Jampani. Probing the 3d awareness of visual foundation models. In CVPR, 2024. +[16] Rao Fu, Jingyu Liu, Xilun Chen, Yixin Nie, and Wenhan Xiong. Scene-llm: Extending language model for 3d visual understanding and reasoning. arXiv preprint arXiv:2403.11401, 2024. +[17] Aaron Grattafori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. + +[18] Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948, 2025. +[19] Ziyu Guo, Renrui Zhang, Xiangyang Zhu, Yiwen Tang, Xianzheng Ma, Jiaming Han, Kexin Chen, Peng Gao, Xianzhi Li, Hongsheng Li, et al. Point-bind & point-llm: Aligning point cloud with multi-modality for 3d understanding, generation, and instruction following. arXiv preprint arXiv:2309.00615, 2023. +[20] Richard Hartley. Multiple view geometry in computer vision, volume 665. Cambridge university press, 2003. +[21] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, 2022. +[22] Yining Hong, Haoyu Zhen, Peihao Chen, Shuhong Zheng, Yilun Du, Zhenfang Chen, and Chuang Gan. 3d-llm: Injecting the 3d world into large language models. In NeurIPS, 2023. +[23] Ji Hou, Saining Xie, Benjamin Graham, Angela Dai, and Matthias Nießner. Pri3d: Can 3d priors help 2d representation learning? In ICCV, 2021. +[24] Haifeng Huang, Zehan Wang, Rongjie Huang, Luping Liu, Xize Cheng, Yang Zhao, Tao Jin, and Zhou Zhao. Chat-3d v2: Bridging 3d scene and large language models with object identifiers. CoRR, 2023. +[25] Jiangyong Huang, Silong Yong, Xiaojian Ma, Xiongkun Linghu, Puhao Li, Yan Wang, Qing Li, Song-Chun Zhu, Baoxiong Jia, and Siyuan Huang. An embodied generalist agent in 3d world. In ICML, 2024. +[26] Shijia Huang, Yilun Chen, Jiaya Jia, and Liwei Wang. Multi-view transformer for 3d visual grounding. In CVPR, 2022. +[27] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, et al. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. +[28] Jinlong Li, Cristiano Saltori, Fabio Poiesi, and Nicu Sebe. Cross-modal and uncertainty-aware agglomeration for open-vocabulary 3d scene understanding. In CVPR, 2025. +[29] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023. +[30] Zhengzhe Liu, Xiaojuan Qi, and Chi-Wing Fu. 3d-to-2d distillation for indoor scene parsing. In CVPR, 2021. +[31] Zuyan Liu, Yuhao Dong, Ziwei Liu, Winston Hu, Jiwen Lu, and Yongming Rao. Oryx mllm: On-demand spatial-temporal understanding at arbitrary resolution. arXiv preprint arXiv:2409.12961, 2024. +[32] Xiaojian Ma, Silong Yong, Zilong Zheng, Qing Li, Yitao Liang, Song-Chun Zhu, and Siyuan Huang. SQA3D: situated question answering in 3d scenes. In ICLR, 2023. License: CC-BY-4.0. +[33] Yunze Man, Shuhong Zheng, Zhipeng Bao, Martial Hebert, Liangyan Gui, and Yu-Xiong Wang. Lexicon3d: Probing visual foundation models for complex 3d scene understanding. In NeurIPS, 2024. +[34] Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. +[35] Songyou Peng, Kyle Genova, Chiyu Jiang, Andrea Tagliasacchi, Marc Pollefeys, Thomas Funkhouser, et al. Openscene: 3d scene understanding with open vocabularies. In CVPR, 2023. +[36] Zhangyang Qi, Ye Fang, Zeyi Sun, Xiaoyang Wu, Tong Wu, Jiaqi Wang, Dahua Lin, and Hengshuang Zhao. Gpt4point: A unified framework for point-language understanding and generation. In CVPR, 2024. +[37] Zhangyang Qi, Zhixiong Zhang, Ye Fang, Jiaqi Wang, and Hengshuang Zhao. Gpt4scene: Understand 3d scenes from videos with vision-language models. arXiv preprint arXiv:2501.01428, 2025. +[38] Jonas Schult, Francis Engelmann, Alexander Hermans, Or Litany, Siyu Tang, and Bastian Leibe. Mask3d: Mask transformer for 3d semantic instance segmentation. In ICRA, 2023. +[39] Stefan Stojanov, Anh Thai, Zixuan Huang, and James M Rehg. Learning dense object descriptors from multiple views for low-shot category generalization. In NeurIPS, 2022. + +[40] Yuan Tang, Xu Han, Xianzhi Li, Qiao Yu, Yixue Hao, Long Hu, and Min Chen. Minigpt-3d: Efficiently aligning 3d point clouds with large language models using 2d priors. In ACM Multi, 2024. +[41] Yuan Tang, Xu Han, Xianzhi Li, Qiao Yu, Jinfeng Xu, Yixue Hao, Long Hu, and Min Chen. More text, less point: Towards 3d data-efficient point-language understanding. In AAAI, 2025. +[42] Qwen Team et al. Qwen2 technical report. arXiv preprint arXiv:2407.10671, 2(8), 2024. +[43] Bill Triggs, Philip F McLauchlan, Richard I Hartley, and Andrew W Fitzgibbon. Bundle adjustment—a modern synthesis. In International workshop on vision algorithms. Springer, 1999. +[44] Michael Tschannen, Alexey Gritsenko, Xiao Wang, Muhammad Ferjad Naeem, Ibrahim Alabdulmohsin, Nikhil Parthasarathy, Talfan Evans, Lucas Beyer, Ye Xia, Basil Mustafa, et al. Siglip 2: Multilingual vision-language encoders with improved semantic understanding, localization, and dense features. arXiv preprint arXiv:2502.14786, 2025. +[45] Jianyuan Wang, Minghao Chen, Nikita Karaev, Andrea Vedaldi, Christian Rupprecht, and David Novotny. Vggt: Visual geometry grounded transformer. In CVPR, 2025. +[46] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. +[47] Zehan Wang, Haifeng Huang, Yang Zhao, Ziang Zhang, and Zhou Zhao. Chat-3d: Data-efficiently tuning large language model for universal dialogue of 3d scenes. arXiv preprint arXiv:2308.08769, 2023. +[48] Zeyu Wang, Dingwen Li, Chenxu Luo, Cihang Xie, and Xiaodong Yang. Distillbev: Boosting multi-camera 3d object detection with cross-modal knowledge distillation. In ICCV, 2023. +[49] Hao Wu, Ruochong Li, Hao Wang, and Hui Xiong. Com3d: Leveraging cross-view correspondence and cross-modal mining for 3d retrieval. In ICME, 2024. +[50] Runsen Xu, Xiaolong Wang, Tai Wang, Yilun Chen, Jiangmiao Pang, and Dahua Lin. Pointllm: Empowering large language models to understand point clouds. In ECCV, 2024. +[51] Yong Xu, Chaoda Zheng, Ruotao Xu, Yuhui Quan, and Haibin Ling. Multi-view 3d shape recognition via correspondence-aware deep learning. IEEE Transactions on Image Processing, 2021. +[52] Jihan Yang, Shusheng Yang, Anjali W Gupta, Rilyn Han, Li Fei-Fei, and Saining Xie. Thinking in space: How multimodal large language models see, remember, and recall spaces. In CVPR, 2025. +[53] Yang You, Yixin Li, Congyue Deng, Yue Wang, and Leonidas Guibas. Multiview equivariance improves 3d correspondence understanding with minimal feature finetuning. arXiv preprint arXiv:2411.19458, 2024. +[54] Hanxun Yu, Wentong Li, Song Wang, Junbo Chen, and Jianke Zhu. Inst3d-lmm: Instance-aware 3d scene understanding with multi-modal instruction tuning. In CVPR, 2025. +[55] Yuanwen Yue, Anurag Das, Francis Engelmann, Siyu Tang, and Jan Eric Lenssen. Improving 2d feature representations by 3d-aware fine-tuning. In ECCV, 2024. +[56] Peiyuan Zhang, Kaichen Zhang, Bo Li, Guangtao Zeng, Jingkang Yang, Yuanhan Zhang, Ziyue Wang, Haoran Tan, Chunyuan Li, and Ziwei Liu. Long context transfer from language to vision. arXiv preprint arXiv:2406.16852, 2024. +[57] Shangzhan Zhang, Jianyuan Wang, Yinghao Xu, Nan Xue, Christian Rupprecht, Xiaowei Zhou, Yujun Shen, and Gordon Wetzstein. Flare: Feed-forward geometry, appearance and camera estimation from uncalibrated sparse views. In CVPR, 2025. +[58] Yiming Zhang, ZeMing Gong, and Angel X. Chang. Multi3drefer: Grounding text description to multiple 3d objects. In ICCV, 2023. License: MIT. +[59] Yuanhan Zhang, Jinming Wu, Wei Li, Bo Li, Zejun Ma, Ziwei Liu, and Chunyuan Li. Video instruction tuning with synthetic data. arXiv preprint arXiv:2410.02713, 2024. +[60] Lichen Zhao, Daigang Cai, Lu Sheng, and Dong Xu. 3dvg-transformer: Relation modeling for visual grounding on point clouds. In ICCV, 2021. +[61] Duo Zheng, Shijia Huang, Yanyang Li, and Liwei Wang. Learning from videos for 3d world: Enhancing mllms with 3d vision geometry priors. In NeurIPS, 2025. + +[62] Duo Zheng, Shijia Huang, and Liwei Wang. Video-3d llm: Learning position-aware video representation for 3d scene understanding. In CVPR, 2025. +[63] Duo Zheng, Shijia Huang, Lin Zhao, Yiwu Zhong, and Liwei Wang. Towards learning a generalist model for embodied navigation. In CVPR, 2024. +[64] Chenming Zhu, Tai Wang, Wenwei Zhang, Jiangmiao Pang, and Xihui Liu. Llava-3d: A simple yet effective pathway to empowering Imms with 3d-awareness. arXiv preprint arXiv:2409.18125, 2024. +[65] Ziyu Zhu, Xiaojian Ma, Yixin Chen, Zhidong Deng, Siyuan Huang, and Qing Li. 3d-vista: Pre-trained transformer for 3d vision and text alignment. In ICCV, 2023. +[66] Ziyu Zhu, Zhuofan Zhang, Xiaojian Ma, Xuesong Niu, Yixin Chen, Baoxiong Jia, Zhidong Deng, Siyuan Huang, and Qing Li. Unifying 3d vision-language understanding via promptable queries. In ECCV, 2024. + +# A Technical Appendices and Supplementary Material + +# A.1 World Coordinate Computation + +Given a set of $N$ images $\mathcal{I} = \{I_1, I_2, \dots, I_N\}$ , each image $I_i$ is paired with its depth map $D_i \in \mathbb{R}^{H \times W}$ , camera intrinsic matrix $K_i \in \mathbb{R}^{3 \times 3}$ , and camera-to-world extrinsic matrix $T_i \in \mathbb{R}^{4 \times 4}$ . For a pixel at $(u, v)$ in image $I_i$ , the corresponding 3D coordinate in the global coordinate system, denoted as $\mathbf{C}_i(u, v) \in \mathbb{R}^3$ , is computed as: + +$$ +\left[ \begin{array}{c} \mathbf {C} _ {i} (u, v) \\ 1 \end{array} \right] = T _ {i} \left[ \begin{array}{c} D _ {i} (u, v) \cdot K _ {i} ^ {- 1} \left[ \begin{array}{l} u \\ v \\ 1 \end{array} \right] \\ 1 \end{array} \right] \tag {1} +$$ + +Repeating this process for all pixels yields the per-pixel 3D coordinate map $\mathbf{C}_i\in \mathbb{R}^{H\times W\times 3}$ for each image $I_{i}$ . The complete set of coordinate maps is denoted as $\mathcal{C} = \{\mathbf{C}_1,\mathbf{C}_2,\dots ,\mathbf{C}_N\}$ . + +# A.2 Datsests for Training + +For model fine-tuning, we utilize a collection of well-established 3D vision-language datasets. Specifically, we follow the model finetuning settings of Video-3D LLM [62] by using the validation splits of ScanRefer, Multi3DRefer, Scan2Cap, and ScanQA, as well as the test split of SQA3D. Across these datasets, the number of data samples varies significantly: ScanRefer and Scan2Cap each provide 36,665 samples, while Multi3DRefer offers 43,838 entries. ScanQA contains 26,515 instances, and SQA3D is the largest with 79,445 samples. Most datasets are derived from 562 unique scans, except SQA3D, which includes 518 scans. We further report the average lengths of questions and answers for each dataset. For example, question lengths range from approximately 13 to 38 words, with Scan2Cap and ScanQA also providing answer texts, averaging 17.9 and 2.4 words in length, respectively. In SQA3D, the average question and answer lengths are 37.8 and 1.1 words. For the evaluation on VSI-Bench, we use the pre-training data from VG-LLM [61]. + +# A.3 Detailed Comparison + +In this section, we provide a detailed comparison with other methods using all metrics across 5 benchmarks. + +Scanrefer. Tab. 7 shows that our method 3DRS achieves the best overall performance on the ScanRefer validation set, especially in the challenging "Multiple" scenario where precise target + +Table 7: Performance comparison on the validation set of ScanRefer [5]. "Unique" and "Multiple" depends on whether there are other objects of the same class as the target object. + +
MethodUniqueMultipleOverall
Acc@0.25Acc@0.5Acc@0.25Acc@0.5Acc@0.25Acc@0.5
ScanRefer [5]76.353.532.721.141.227.4
MVT [26]77.766.431.925.340.833.3
3DVG-Transformer [60]81.960.639.328.447.634.7
ViL3DRel [7]81.668.640.330.747.937.7
3DJCG [4]83.564.341.430.849.637.3
D3Net [6]-72.0-30.1-37.9
M3DRef-CLIP [58]85.377.243.836.851.944.7
3D-VisTA [65]81.675.143.739.150.645.8
3D-LLM (Flamingo) [22]----21.2-
3D-LLM (BLIP2-flant5) [22]----30.3-
Grounded 3D-LLM [9]----47.944.1
PQ3D [66]86.778.351.546.257.051.2
ChatScene [24]89.682.547.842.955.550.2
LLaVA-3D [64]----54.142.2
Video 3D-LLM [62]88.078.350.945.358.151.7
3DRS (Ours)87.477.957.050.862.956.1
+ +Table 8: Performance comparison on the validation set of Multi3DRefer [58]. ZT: zero-target, ST: single-target, MT: multi-target, D: distractor. + +
MethodZT w/o D F1ZT w/D F1ST w/o D F1@0.25ST w/D F1@0.5ST w/D F1@0.25MT F1@0.5MT F1@0.25ALL F1@0.25F1@0.5
M3DRef-CLIP [58]81.839.453.547.834.630.643.637.942.8
D3Net [6]81.632.5-38.6-23.3-35.0-
3DJCG [4]94.166.9-26.0-16.7-26.2-
Grounded 3D-LLM [9]--------45.2
PQ3D [66]85.457.7-68.5-43.6-40.9-
ChatScene [24]90.362.682.975.949.144.545.741.157.1
Video 3D-LLM [62]94.778.582.673.452.147.240.835.758.0
3DRS (Ours)95.679.479.671.457.051.343.037.860.4
+ +Table 9: Performance comparison on the validation set of ScanQA [2]. EM indicates exact match accuracy, and B-1, B-2, B-3, B-4 denote BLEU-1, -2, -3, -4, respectively. + +
MethodEMB-1B-2B-3B-4ROUGE-LMETEORCIDEr
ScanQA [2]21.0530.2420.4015.1110.0833.3313.1464.86
3D-VisTA [65]22.40---10.4035.7013.9069.60
Oryx-34B [31]-38.0024.60--37.3015.0072.30
LLaVA-Video-7B [59]-39.7126.579.333.0944.6217.7288.70
3D-LLM (Flamingo) [22]20.4030.3017.8012.007.2032.3012.2059.20
3D-LLM (BLIP2-flant5) [22]20.5039.3025.2018.4012.0035.7014.5069.40
Chat-3D [47]-29.10--6.4028.5011.9053.20
NaviLLM [63]23.00---12.5038.4015.4075.90
LL3DA [8]----13.5337.3115.8876.79
Scene-LLM [16]27.2043.6026.8019.1012.0040.0016.6080.00
LEO [25]----11.5039.3016.2080.00
Grounded 3D-LLM [9]----13.40--72.70
ChatScene [24]21.6243.2029.0620.5714.3141.5618.0087.70
LLaVA-3D [64]27.00---14.5050.1020.7091.70
Video 3D-LLM [62]30.1047.0531.7022.8316.1749.0219.84102.06
3DRS (Ours)30.3048.3732.6723.7917.2249.8220.47104.78
+ +discrimination is required. These results demonstrate that 3DRS effectively leverages multi-view images for robust spatial understanding and accurate object localization. + +Multi3DRefer. In Tab. 8, 3DRS achieves the best overall results on the Multi3DRefer validation set, with top F1 scores in both standard and challenging scenarios. Our method consistently outperforms previous approaches, especially in the difficult zero-target and distractor settings, demonstrating superior robustness and spatial understanding. + +ScanQA. In Tab. 9, 3DRS achieves the best performance on the ScanQA validation set across almost all metrics, including EM, BLEU scores, METEOR, and CIDEr, demonstrating its strong effectiveness for 3D question answering. + +SQA3D. In Tab. 10, 3DRS achieves the highest scores on the SQA3D test set, outperforming all previous approaches on almost every question type as well as in the overall average, which demonstrates its superior capability for 3D question answering across diverse scenarios. + +Scan2cap. In Tab. 11, 3DRS achieves the best performance on the Scan2Cap validation set in terms of CIDEr (C), and remains highly competitive on other metrics such as BLEU-4, METEOR, and ROUGE-L, demonstrating strong overall effectiveness for 3D captioning. + +# A.4 Ablation Study + +Table 12 shows the effect of applying supervision to different numbers of network layers across multiple 3D scene understanding tasks, including object localization, captioning, and question answering. Supervising only the last layer consistently achieves the best performance on all benchmarks. As more intermediate layers are added for supervision, the results degrade. This suggests that multi-layer supervision may over-constrain geometric features and weaken semantic representations, ultimately hindering downstream performance. Future work may explore more advanced strategies to balance geometric and semantic cues. + +Table 10: Performance comparison on the test set of SQA3D [32]. + +
MethodTest setAvg.
WhatIsHowCanWhichOthers
SQA3D [32]31.6063.8046.0069.5043.9045.3046.60
3D-VisTA [4]34.8063.3045.4069.8047.2048.1048.50
LLaVA-Video[59]42.7056.3047.5055.3050.1047.2048.50
Scene-LLM [16]40.9069.1045.0070.8047.2052.3054.20
LEO [25]------50.00
ChatScene [24]45.4067.0052.0069.5049.9055.0054.60
LLaVA-3D [64]------55.60
Video 3D-LLM [62]51.1072.4055.5069.8051.3056.0058.60
3DRS (Ours)54.4075.2057.0072.2049.9059.0060.60
+ +Table 11: Performance comparison on the validation set of Scan2Cap [12]. + +
Method@0.5
CB-4MR
Scan2Cap [12]39.0823.3221.9744.48
3DJCG [4]49.4831.0324.2250.80
D3Net [6]62.6435.6825.7253.90
3D-VisTA [65]66.9034.0027.1054.30
LL3DA [8]65.1936.7925.9755.06
LEO [25]68.4036.9027.7057.80
ChatScene [24]77.1936.3428.0158.12
LLaVA-3D [64]79.2141.1230.2163.41
Video 3D-LLM [62]83.7742.4328.8762.34
3DRS (Ours)86.1141.6328.9762.29
+ +Table 12: Distillation on multiple layers. + +
SupervisionScanReferMulti3DReferScan2CapScanQASQA3D
Acc@0.25Acc@0.5F1@0.25F1@0.5C@0.5B-4@0.5CEMEM
Last layer62.956.160.454.941.686.1104.830.360.6
Last layer + last 3rd layer61.554.860.154.941.484.4101.429.260.5
Last layer + last 3rd + last 5th layer60.553.959.053.840.081.1102.930.059.6
+ +Table 13 reports the impact of different distillation loss functions, including euclidean loss, cosine loss, and their combination, across various 3D scene understanding benchmarks. The results show that all loss types yield very similar performance, indicating that the choice of feature distance metric has limited influence in our setting. + +# A.5 Qualitative Results + +Visualizations Fig. 4 illustrates qualitative results of our method across three tasks: visual grounding, object captioning, and question answering. + +For the visual grounding task (top two rows), the model is required to localize objects within a 3D scene based on natural language descriptions. Each example shows the ground truth bounding box (blue), the result from a baseline method (red), and our prediction (green). In both cases, our method's predictions match the ground truth more closely than the baseline, demonstrating improved grounding accuracy. + +In the object captioning task (middle two rows), the model generates descriptive captions for specific objects in the scene. The captions from the ground truth, the baseline, and our method are shown alongside their corresponding regions. We also report CIDEr scores to measure caption quality. Our approach produces more accurate and detailed descriptions with significantly higher CIDEr scores compared to the baseline. + +For the question answering task (bottom two rows), the model answers questions about the scene. Ground truth answers, baseline outputs, and our results are provided for each question. Red rectangles highlight the visual evidence used by our model to generate the answers. Our method provides correct answers that align with the ground truth, whereas the baseline often fails to do so. + +Overall, the visualizations demonstrate that our approach consistently outperforms the baseline across all tasks, delivering more accurate grounding, richer object descriptions, and more reliable answers to visual questions. + +Table 13: Performance with different distillation losses. + +
SupervisionScanReferMulti3DReferScan2CapScanQASQA3D
Acc@0.25Acc@0.5F1@0.25F1@0.5B-4@0.5C@0.5CEMEM
Euclidean loss62.956.160.454.941.686.1104.830.360.6
Cosine loss62.255.560.455.241.885.9104.530.160.7
Cosine + Euclidean62.355.760.355.042.185.8102.729.760.2
+ +Figs. 5 and 6 provide a visual summary of how our method performs on three challenging 3D scene understanding tasks. These tasks include identifying objects based on language, generating descriptions for specific regions, and answering spatial questions about the scene. + +In the visual grounding examples at the top, the model is challenged to find the correct object in a complex 3D environment based on a textual description. The comparison highlights three bounding boxes for each case: blue for the ground truth, red for the baseline, and green for our result. Our predictions consistently align with the intended targets, showing our model's ability to accurately interpret spatial and semantic cues from language. + +The object captioning section in the middle presents how each model describes a highlighted object or area. For each instance, the ground truth, baseline output, and our generated caption are shown, along with their respective CIDEr scores. Our model's captions are both more precise and more faithful to the scene's content, as reflected in the higher evaluation scores. + +At the bottom, the question answering task demonstrates the model's reasoning abilities within a 3D environment. The figures show the posed question, the correct answer, the baseline's response, and our model's answer. Even for questions that require counting or locating objects, our approach tends to provide accurate answers, often supported by clear visual evidence in the scene. + +Altogether, these qualitative results illustrate that our approach delivers more reliable scene understanding across a variety of tasks, outperforming the baseline in both accuracy and descriptive quality. + +# A.6 Broader Impacts + +Positive impacts. The advancement of 3D perception in AI systems holds significant positive societal potential. Enhanced 3D understanding can benefit applications such as assistive robotics for the elderly and disabled, safer autonomous navigation, improved medical imaging, and immersive educational tools. These technologies have the capacity to improve quality of life, boost accessibility, and enable new forms of human-computer interaction. + +Negative impacts. However, the adoption of enhanced 3D perception also raises important privacy concerns, especially in surveillance and monitoring contexts where individuals' activities or environments could be reconstructed and analyzed without their consent. To address these risks, it is crucial to apply robust data anonymization methods—such as blurring faces or removing identifiable features—ensure informed consent from data subjects, enforce strict access controls and data security protocols, and adhere to relevant privacy regulations to protect individual rights. + +# Visual Grounding + +![](images/4789f7795c04bd8d6bb1586c2ab2618936ac562c2168e6c72865b0da3552a261.jpg) + +Question: Identify the object according to the following description. A small blue cabinet. there is a chair and bed near it. + +GT: Blue box. + +Baseline: Red box. + +Ours:Green box. + +![](images/09a36e7983b0af6c111af310ce4cb793b17db530f495f13fa579ac20bfbb2869.jpg) + +Question: Identify the object according to the following description. There is a rectangular white kitchen cabinet. it is to the right of a stove. + +GT: Blue box. + +Baseline: Red box. + +Ours: Green box. + +# Object Captioning + +![](images/09cb72113579d13fe7de12ff4171e7a3d99aa9bea709e386f5158ed9474b460d.jpg) + +GT: This is a rectangular ottoman. The ottoman is in front of the sofa. + +Baseline: This is a round end table. It is to the left of the couch. (CIDER: 0.01) + +Ours: This is a brown ottoman. It is in front of a sofa. (CIDER: 3.52) + +![](images/d6db9a4cc2c8610a8ba11215f9a387c568f41010507f52064203c0e6047ce853.jpg) + +GT: The cabinet is in the middle of the room. The cabinet is a white box with a blue rectangular top. + +Baseline: This is a long counter. it is behind a backpack. (CIDER: 0.00) + +Ours: The cabinet is in the center of the room. the cabinet is a white rectangular prism. (CIDER: 3.07) + +# Question Answering + +![](images/98d0de74478b6169d48055c9b4e77ab39cac6bb10bd495a485084daadff1ea55.jpg) + +Question: How many windows? Answer the question simply. + +GT: 4. + +Baseline: 2. + +Ours: 4. + +![](images/36958199d28cd214e8666f7063276f26c3f94e8e58bdc0f1b149360288a2fcc9.jpg) +Figure 4: Visualization of Results Across Different Tasks. (a) Visual Grounding: The predicted bounding box closely aligns with the ground truth. (b) Object Captioning: Our method generates accurate captions for each referred object. (c) Question Answering: The model provides precise answers, where we use the red rectangles to indicate the visual cues utilized for each response. Best viewed when zoomed in. + +Question: How many levels does the shelf have? Answer the question simply. + +GT: 3. + +Baseline: 6. + +Ours: 3. + +# Visual Grounding + +![](images/e2ecba2c1bb7cc0320d8a93e1fc9d077bc5a92a4f5947cf1a45bc8748aa54e5d.jpg) + +Question: The couch is north of the round coffee table. the couch has three seats and two armrests. + +GT: Blue box. + +Baseline: Red box. + +Ours:Green box. + +![](images/c6cf574c764b419da30351478980fe880cb6224a7d00923656a8991336da0318.jpg) + +Question: Identify the object according to the following description. This is a small end table with a book or magazine laying on it. If you were sitting in the black loveseat facing the end table, the short, long bookshelf would be on your right. + +GT: Blue box. + +Baseline: Red box. + +Ours: Green box. + +# Object Captioning + +![](images/9006ff3854ee10fcc504736e1146097b88dc4aed6268abb2a269e63cec8274e4.jpg) + +GT: This is a brown desk. it is behind a chair. + +Baseline: This is a blue chair. it is at a desk. (CIDER: 0.8247) + +Ours: This is a wooden desk. it is to the right of another desk. (CIDER: 2.968) + +![](images/a5d8951af92bbdf4e1a0c51155db2bb21419e1ec9b59c3edc5d15020f3062dc4.jpg) + +GT: This is a wooden table. It is against a wall. + +Baseline: The chair is the second closest one to the whiteboard. the chair has a curved backside and four legs. (CIDER: 0.0009) + +Ours: This is a wooden table. it is against the wall. (CIDER: 2.4132) + +# Question Answering + +![](images/3892c9392129c48ac95e54c518a54b082de9173587c9d09211f7f118c6c4b261.jpg) + +Question: How many chairs are next to the curtain? Answer the question simply. + +GT: 1. + +Baseline: 2. + +Ours: 1. + +![](images/370c52c2ff386603eeb715c4918ea49d570af750d11544a7cc2d2435b9621b56.jpg) +Figure 5: Visualization of Results Across Different Tasks. (a) Visual Grounding: The predicted bounding box closely aligns with the ground truth. (b) Object Captioning: Our method generates accurate captions for each referred object. (c) Question Answering: The model provides precise answers, where we use the red rectangles to indicate the visual cues utilized for each response. Best viewed when zoomed in. + +Question: How many chairs are on the right side of table? Answer the question simply. + +GT: 3. + +Baseline: 2. + +Ours: 3. + +# Visual Grounding + +![](images/c1ae47c1963112cd1c0843ba4d76e7e930213a2fdbec2eb94edc37f6d5e2250a.jpg) + +Question: The chair is grey, the chair is on the left side of the long table. + +GT: Blue box. + +Baseline: Red box. + +Ours: Green box. + +![](images/93448f7ea4041f6d39541ce2fa626dbf8bbae7a1f9ebb99cfd587ed83298719c.jpg) + +Question: This round table is in the middle of two surfaces. it is brown. + +GT: Blue box. + +Baseline: Red box. + +Ours: Green box. + +# Object Captioning + +![](images/5f2105b2113206fe6ec4210c005e881e8bde91897d839b2cddc3ed76b7d84cd3.jpg) + +GT: This pillow is on the bed. It is bright. + +Baseline: This is a bed with white sheets. Tt is to the right of a nightstand. (CIDER: 0.2095) + +Ours: This is a white pillow. it is on the bed. (CIDER: 2.7067) + +![](images/d8268fc354089a13bb0fc3b9f59b019ca521c04d48a38e7d389c902f0ebdd3df.jpg) + +GT: This is a red ottoman. it is next to a round table. + +Baseline: The table is round. it is in between the two couches. (CIDER: 0.1327) + +Ours: This is a red ottoman. it is to the left of a round table. (CIDER: 2.4524) + +# Question Answering + +![](images/e5c1063a60faa280b4a2b6857d4df16789e85d118213be33ffd8d300b3fb254d.jpg) + +Question: What chairs are closest to plant? Answer the question simply. + +GT: 2. + +Baseline: 4. + +Ours: 2. + +![](images/c46b458f520ac7e1199441c28b2905f8d71e70fbcbcc68889558a5614eb28956.jpg) +Figure 6: Visualization of Results Across Different Tasks. (a) Visual Grounding: The predicted bounding box closely aligns with the ground truth. (b) Object Captioning: Our method generates accurate captions for each referred object. (c) Question Answering: The model provides precise answers, where we use the red rectangles to indicate the visual cues utilized for each response. Best viewed when zoomed in. + +Question: Where is the octagagon shape table located? Answer the question simply. + +GT: In between 2 chairs. + +Baseline: In front of black chair. + +Ours: Between 2 chairs. + +# NeurIPS Paper Checklist + +The checklist is designed to encourage best practices for responsible machine learning research, addressing issues of reproducibility, transparency, research ethics, and societal impact. Do not remove the checklist: The papers not including the checklist will be desk rejected. The checklist should follow the references and follow the (optional) supplemental material. The checklist does NOT count towards the page limit. + +Please read the checklist guidelines carefully for information on how to answer these questions. For each question in the checklist: + +- You should answer [Yes], [No], or [NA]. +- [NA] means either that the question is Not Applicable for that particular paper or the relevant information is Not Available. +- Please provide a short (1–2 sentence) justification right after your answer (even for NA). + +The checklist answers are an integral part of your paper submission. They are visible to the reviewers, area chairs, senior area chairs, and ethics reviewers. You will be asked to also include it (after eventual revisions) with the final version of your paper, and its final version will be published with the paper. + +The reviewers of your paper will be asked to use the checklist as one of the factors in their evaluation. While "[Yes]" is generally preferable to "[No]", it is perfectly acceptable to answer "[No]" provided a proper justification is given (e.g., "error bars are not reported because it would be too computationally expensive" or "we were unable to find the license for the dataset we used"). In general, answering "[No]" or "[NA]" is not grounds for rejection. While the questions are phrased in a binary way, we acknowledge that the true answer is often more nuanced, so please just use your best judgment and write a justification to elaborate. All supporting evidence can appear either in the main paper or the supplemental material, provided in appendix. If you answer [Yes] to a question, in the justification please point to the section(s) where related material for the question can be found. + +IMPORTANT, please: + +- Delete this instruction block, but keep the section heading "NeurIPS Paper Checklist", +- Keep the checklist subsection headings, questions/answers and guidelines below. +- Do not modify the questions and only use the provided macros for your answers. + +# 1. Claims + +Question: Do the main claims made in the abstract and introduction accurately reflect the paper's contributions and scope? + +Answer: [Yes] + +Justification: The main claims in the abstract and introduction accurately reflect the paper's contributions and scope about 3D-aware representation learning in MLLMs for scene understanding. + +Guidelines: + +- The answer NA means that the abstract and introduction do not include the claims made in the paper. +- The abstract and/or introduction should clearly state the claims made, including the contributions made in the paper and important assumptions and limitations. A No or NA answer to this question will not be perceived well by the reviewers. +- The claims made should match theoretical and experimental results, and reflect how much the results can be expected to generalize to other settings. +- It is fine to include aspirational goals as motivation as long as it is clear that these goals are not attained by the paper. + +# 2. Limitations + +Question: Does the paper discuss the limitations of the work performed by the authors? + +Answer: [Yes] + +Justification: The limitation section is provided in Sec. 6. + +# Guidelines: + +- The answer NA means that the paper has no limitation while the answer No means that the paper has limitations, but those are not discussed in the paper. +- The authors are encouraged to create a separate "Limitations" section in their paper. +- The paper should point out any strong assumptions and how robust the results are to violations of these assumptions (e.g., independence assumptions, noiseless settings, model well-specification, asymptotic approximations only holding locally). The authors should reflect on how these assumptions might be violated in practice and what the implications would be. +- The authors should reflect on the scope of the claims made, e.g., if the approach was only tested on a few datasets or with a few runs. In general, empirical results often depend on implicit assumptions, which should be articulated. +- The authors should reflect on the factors that influence the performance of the approach. For example, a facial recognition algorithm may perform poorly when image resolution is low or images are taken in low lighting. Or a speech-to-text system might not be used reliably to provide closed captions for online lectures because it fails to handle technical jargon. +- The authors should discuss the computational efficiency of the proposed algorithms and how they scale with dataset size. +- If applicable, the authors should discuss possible limitations of their approach to address problems of privacy and fairness. +- While the authors might fear that complete honesty about limitations might be used by reviewers as grounds for rejection, a worse outcome might be that reviewers discover limitations that aren't acknowledged in the paper. The authors should use their best judgment and recognize that individual actions in favor of transparency play an important role in developing norms that preserve the integrity of the community. Reviewers will be specifically instructed to not penalize honesty concerning limitations. + +# 3. Theory assumptions and proofs + +Question: For each theoretical result, does the paper provide the full set of assumptions and a complete (and correct) proof? + +Answer: [NA] + +Justification: This paper does not contain theoretical assumptions and proofs. + +Guidelines: + +- The answer NA means that the paper does not include theoretical results. +- All the theorems, formulas, and proofs in the paper should be numbered and cross-referenced. +- All assumptions should be clearly stated or referenced in the statement of any theorems. +- The proofs can either appear in the main paper or the supplemental material, but if they appear in the supplemental material, the authors are encouraged to provide a short proof sketch to provide intuition. +- Inversely, any informal proof provided in the core of the paper should be complemented by formal proofs provided in appendix or supplemental material. +- Theorems and Lemmas that the proof relies upon should be properly referenced. + +# 4. Experimental result reproducibility + +Question: Does the paper fully disclose all the information needed to reproduce the main experimental results of the paper to the extent that it affects the main claims and/or conclusions of the paper (regardless of whether the code and data are provided or not)? + +Answer: [Yes] + +Justification: The detailed method design is given in Sec. 2, and experimental details are all provided in Sec. 3 and App. A.2. + +Guidelines: + +- The answer NA means that the paper does not include experiments. + +- If the paper includes experiments, a No answer to this question will not be perceived well by the reviewers: Making the paper reproducible is important, regardless of whether the code and data are provided or not. + +- If the contribution is a dataset and/or model, the authors should describe the steps taken to make their results reproducible or verifiable. + +- Depending on the contribution, reproducibility can be accomplished in various ways. For example, if the contribution is a novel architecture, describing the architecture fully might suffice, or if the contribution is a specific model and empirical evaluation, it may be necessary to either make it possible for others to replicate the model with the same dataset, or provide access to the model. In general, releasing code and data is often one good way to accomplish this, but reproducibility can also be provided via detailed instructions for how to replicate the results, access to a hosted model (e.g., in the case of a large language model), releasing of a model checkpoint, or other means that are appropriate to the research performed. + +- While NeurIPS does not require releasing code, the conference does require all submissions to provide some reasonable avenue for reproducibility, which may depend on the nature of the contribution. For example + +(a) If the contribution is primarily a new algorithm, the paper should make it clear how to reproduce that algorithm. +(b) If the contribution is primarily a new model architecture, the paper should describe the architecture clearly and fully. +(c) If the contribution is a new model (e.g., a large language model), then there should either be a way to access this model for reproducing the results or a way to reproduce the model (e.g., with an open-source dataset or instructions for how to construct the dataset). +(d) We recognize that reproducibility may be tricky in some cases, in which case authors are welcome to describe the particular way they provide for reproducibility. In the case of closed-source models, it may be that access to the model is limited in some way (e.g., to registered users), but it should be possible for other researchers to have some path to reproducing or verifying the results. + +# 5. Open access to data and code + +Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in the supplemental material? + +Answer: [No] + +Justification: We will publicly release the code and related instructions in the near future. + +Guidelines: + +- The answer NA means that paper does not include experiments requiring code. +- Please see the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details. +- While we encourage the release of code and data, we understand that this might not be possible, so "No" is an acceptable answer. Papers cannot be rejected simply for not including code, unless this is central to the contribution (e.g., for a new open-source benchmark). +- The instructions should contain the exact command and environment needed to run to reproduce the results. See the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details. +- The authors should provide instructions on data access and preparation, including how to access the raw data, preprocessed data, intermediate data, and generated data, etc. +- The authors should provide scripts to reproduce all experimental results for the new proposed method and baselines. If only a subset of experiments are reproducible, they should state which ones are omitted from the script and why. +- At submission time, to preserve anonymity, the authors should release anonymized versions (if applicable). + +- Providing as much information as possible in supplemental material (appended to the paper) is recommended, but including URLs to data and code is permitted. + +# 6. Experimental setting/details + +Question: Does the paper specify all the training and test details (e.g., data splits, hyperparameters, how they were chosen, type of optimizer, etc.) necessary to understand the results? + +Answer: [Yes] + +Justification: All the necessary details are provided in Sec. 3 and App. A.2. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The experimental setting should be presented in the core of the paper to a level of detail that is necessary to appreciate the results and make sense of them. +- The full details can be provided either with the code, in appendix, or as supplemental material. + +# 7. Experiment statistical significance + +Question: Does the paper report error bars suitably and correctly defined or other appropriate information about the statistical significance of the experiments? + +Answer: [No] + +Justification: Our experimental settings are consistent with those used in prior works published at top conferences, which do not report error bars as well. Besides, due to resource constraints, we were unable to perform the multiple runs required to report error bars. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The authors should answer "Yes" if the results are accompanied by error bars, confidence intervals, or statistical significance tests, at least for the experiments that support the main claims of the paper. +- The factors of variability that the error bars are capturing should be clearly stated (for example, train/test split, initialization, random drawing of some parameter, or overall run with given experimental conditions). +- The method for calculating the error bars should be explained (closed form formula, call to a library function, bootstrap, etc.) +- The assumptions made should be given (e.g., Normally distributed errors). +- It should be clear whether the error bar is the standard deviation or the standard error of the mean. +- It is OK to report 1-sigma error bars, but one should state it. The authors should preferably report a 2-sigma error bar than state that they have a $96\%$ CI, if the hypothesis of Normality of errors is not verified. +- For asymmetric distributions, the authors should be careful not to show in tables or figures symmetric error bars that would yield results that are out of range (e.g. negative error rates). +- If error bars are reported in tables or plots, The authors should explain in the text how they were calculated and reference the corresponding figures or tables in the text. + +# 8. Experiments compute resources + +Question: For each experiment, does the paper provide sufficient information on the computer resources (type of compute workers, memory, time of execution) needed to reproduce the experiments? + +Answer: [Yes] + +Justification: The information of GPUs we use for experiments is provided in Sec. 2. + +Guidelines: + +- The answer NA means that the paper does not include experiments. + +- The paper should indicate the type of compute workers CPU or GPU, internal cluster, or cloud provider, including relevant memory and storage. +- The paper should provide the amount of compute required for each of the individual experimental runs as well as estimate the total compute. +- The paper should disclose whether the full research project required more compute than the experiments reported in the paper (e.g., preliminary or failed experiments that didn't make it into the paper). + +# 9. Code of ethics + +Question: Does the research conducted in the paper conform, in every respect, with the NeurIPS Code of Ethics https://neurips.cc/public/EthicsGuidelines? + +Answer: [Yes] + +Justification: We strictly follow the NeurIPS Code of Ethics. + +Guidelines: + +- The answer NA means that the authors have not reviewed the NeurIPS Code of Ethics. +- If the authors answer No, they should explain the special circumstances that require a deviation from the Code of Ethics. +- The authors should make sure to preserve anonymity (e.g., if there is a special consideration due to laws or regulations in their jurisdiction). + +# 10. Broader impacts + +Question: Does the paper discuss both potential positive societal impacts and negative societal impacts of the work performed? + +Answer: [Yes] + +Justification: We discuss the broader impacts in App. A.6. + +Guidelines: + +- The answer NA means that there is no societal impact of the work performed. +- If the authors answer NA or No, they should explain why their work has no societal impact or why the paper does not address societal impact. +- Examples of negative societal impacts include potential malicious or unintended uses (e.g., disinformation, generating fake profiles, surveillance), fairness considerations (e.g., deployment of technologies that could make decisions that unfairly impact specific groups), privacy considerations, and security considerations. +- The conference expects that many papers will be foundational research and not tied to particular applications, let alone deployments. However, if there is a direct path to any negative applications, the authors should point it out. For example, it is legitimate to point out that an improvement in the quality of generative models could be used to generate deepfakes for disinformation. On the other hand, it is not needed to point out that a generic algorithm for optimizing neural networks could enable people to train models that generate Deepfakes faster. +- The authors should consider possible harms that could arise when the technology is being used as intended and functioning correctly, harms that could arise when the technology is being used as intended but gives incorrect results, and harms following from (intentional or unintentional) misuse of the technology. +- If there are negative societal impacts, the authors could also discuss possible mitigation strategies (e.g., gated release of models, providing defenses in addition to attacks, mechanisms for monitoring misuse, mechanisms to monitor how a system learns from feedback over time, improving the efficiency and accessibility of ML). + +# 11. Safeguards + +Question: Does the paper describe safeguards that have been put in place for responsible release of data or models that have a high risk for misuse (e.g., pretrained language models, image generators, or scraped datasets)? + +Answer: [NA] + +Justification: This paper uses all public datasets for training models and has no such risk. + +# Guidelines: + +- The answer NA means that the paper poses no such risks. +- Released models that have a high risk for misuse or dual-use should be released with necessary safeguards to allow for controlled use of the model, for example by requiring that users adhere to usage guidelines or restrictions to access the model or implementing safety filters. +- Datasets that have been scraped from the Internet could pose safety risks. The authors should describe how they avoided releasing unsafe images. +- We recognize that providing effective safeguards is challenging, and many papers do not require this, but we encourage authors to take this into account and make a best faith effort. + +# 12. Licenses for existing assets + +Question: Are the creators or original owners of assets (e.g., code, data, models), used in the paper, properly credited and are the license and terms of use explicitly mentioned and properly respected? + +Answer: [Yes] + +Justification: The original papers of the datasets are cited, and their licenses are provided in the reference. + +# Guidelines: + +- The answer NA means that the paper does not use existing assets. +- The authors should cite the original paper that produced the code package or dataset. +- The authors should state which version of the asset is used and, if possible, include a URL. +- The name of the license (e.g., CC-BY 4.0) should be included for each asset. +- For scraped data from a particular source (e.g., website), the copyright and terms of service of that source should be provided. +- If assets are released, the license, copyright information, and terms of use in the package should be provided. For popular datasets, paperswithcode.com/datasets has curated licenses for some datasets. Their licensing guide can help determine the license of a dataset. +- For existing datasets that are re-packaged, both the original license and the license of the derived asset (if it has changed) should be provided. +- If this information is not available online, the authors are encouraged to reach out to the asset's creators. + +# 13. New assets + +Question: Are new assets introduced in the paper well documented and is the documentation provided alongside the assets? + +Answer: [Yes] + +Justification: We detail the training and inference settings in Sec. 3 and App. A.2. As for the code, we will release it in the near future. + +# Guidelines: + +- The answer NA means that the paper does not release new assets. +- Researchers should communicate the details of the dataset/code/model as part of their submissions via structured templates. This includes details about training, license, limitations, etc. +- The paper should discuss whether and how consent was obtained from people whose asset is used. +- At submission time, remember to anonymize your assets (if applicable). You can either create an anonymized URL or include an anonymized zip file. + +# 14. Crowdsourcing and research with human subjects + +Question: For crowdsourcing experiments and research with human subjects, does the paper include the full text of instructions given to participants and screenshots, if applicable, as well as details about compensation (if any)? + +# Answer: [NA] + +Justification: We do not involve human subjects in our experiments. + +# Guidelines: + +- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. +- Including this information in the supplemental material is fine, but if the main contribution of the paper involves human subjects, then as much detail as possible should be included in the main paper. +- According to the NeurIPS Code of Ethics, workers involved in data collection, curation, or other labor should be paid at least the minimum wage in the country of the data collector. + +# 15. Institutional review board (IRB) approvals or equivalent for research with human subjects + +Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals (or an equivalent approval/review based on the requirements of your country or institution) were obtained? + +# Answer: [NA] + +Justification: We do not involve human subjects in our experiments. + +# Guidelines: + +- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. +- Depending on the country in which research is conducted, IRB approval (or equivalent) may be required for any human subjects research. If you obtained IRB approval, you should clearly state this in the paper. +- We recognize that the procedures for this may vary significantly between institutions and locations, and we expect authors to adhere to the NeurIPS Code of Ethics and the guidelines for their institution. +- For initial submissions, do not include any information that would break anonymity (if applicable), such as the institution conducting the review. + +# 16. Declaration of LLM usage + +Question: Does the paper describe the usage of LLMs if it is an important, original, or non-standard component of the core methods in this research? Note that if the LLM is used only for writing, editing, or formatting purposes and does not impact the core methodology, scientific rigorousness, or originality of the research, declaration is not required. + +# Answer: [NA] + +Justification: We only use LLM to improve the wording of the paper and implement code for visualizations. + +# Guidelines: + +- The answer NA means that the core method development in this research does not involve LLMs as any important, original, or non-standard components. +- Please refer to our LLM policy (https://neurips.cc/Conferences/2025/LLM) for what should or should not be described. \ No newline at end of file diff --git a/NeurIPS/2025/3DRS_ MLLMs Need 3D-Aware Representation Supervision for Scene Understanding/images.zip b/NeurIPS/2025/3DRS_ MLLMs Need 3D-Aware Representation Supervision for Scene Understanding/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..7b71a5f58ef2c788db53fabdbf5bf7e3163fae33 --- /dev/null +++ b/NeurIPS/2025/3DRS_ MLLMs Need 3D-Aware Representation Supervision for Scene Understanding/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:156f5c620fed5f56b616ed542708b9a910f2aec558e719938cf24da7718c5de1 +size 1020789 diff --git a/NeurIPS/2025/3DRS_ MLLMs Need 3D-Aware Representation Supervision for Scene Understanding/layout.json b/NeurIPS/2025/3DRS_ MLLMs Need 3D-Aware Representation Supervision for Scene Understanding/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..56e2ae69e143ced5fe2cc0b8776188ad5c077e26 --- /dev/null +++ b/NeurIPS/2025/3DRS_ MLLMs Need 3D-Aware Representation Supervision for Scene Understanding/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:08ad2fb896f79fcd04c6dc9d371a915a372c5d922922d0e88b049c68b79c6212 +size 869981 diff --git a/NeurIPS/2025/4D-LRM_ Large Space-Time Reconstruction Model From and To Any View at Any Time/0bfe7188-533a-4c0b-b157-3907f500cfe9_content_list.json b/NeurIPS/2025/4D-LRM_ Large Space-Time Reconstruction Model From and To Any View at Any Time/0bfe7188-533a-4c0b-b157-3907f500cfe9_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d181008e0a5b52570224199cd5165aa8c1fc86ed --- /dev/null +++ b/NeurIPS/2025/4D-LRM_ Large Space-Time Reconstruction Model From and To Any View at Any Time/0bfe7188-533a-4c0b-b157-3907f500cfe9_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:26a380b30446ee8fa0d63b7c21f8f951f71d5b861346d87bc45d655859c5fada +size 176694 diff --git a/NeurIPS/2025/4D-LRM_ Large Space-Time Reconstruction Model From and To Any View at Any Time/0bfe7188-533a-4c0b-b157-3907f500cfe9_model.json b/NeurIPS/2025/4D-LRM_ Large Space-Time Reconstruction Model From and To Any View at Any Time/0bfe7188-533a-4c0b-b157-3907f500cfe9_model.json new file mode 100644 index 0000000000000000000000000000000000000000..235c41251d12f4520c0b698844ce7bab983f06ba --- /dev/null +++ b/NeurIPS/2025/4D-LRM_ Large Space-Time Reconstruction Model From and To Any View at Any Time/0bfe7188-533a-4c0b-b157-3907f500cfe9_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:249904b65ffaa50494c88201d4e9c3062cfd9e0be9964fa3d9ca1ca414ef50d8 +size 230387 diff --git a/NeurIPS/2025/4D-LRM_ Large Space-Time Reconstruction Model From and To Any View at Any Time/0bfe7188-533a-4c0b-b157-3907f500cfe9_origin.pdf b/NeurIPS/2025/4D-LRM_ Large Space-Time Reconstruction Model From and To Any View at Any Time/0bfe7188-533a-4c0b-b157-3907f500cfe9_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..2e88c2b3643ea5093d9bf86b9d8dc7e397f6e00c --- /dev/null +++ b/NeurIPS/2025/4D-LRM_ Large Space-Time Reconstruction Model From and To Any View at Any Time/0bfe7188-533a-4c0b-b157-3907f500cfe9_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a89d81894d778a631ecc756e3bc51788f5f4bb7984ea9eb3d8cdd40c51e83259 +size 3585405 diff --git a/NeurIPS/2025/4D-LRM_ Large Space-Time Reconstruction Model From and To Any View at Any Time/full.md b/NeurIPS/2025/4D-LRM_ Large Space-Time Reconstruction Model From and To Any View at Any Time/full.md new file mode 100644 index 0000000000000000000000000000000000000000..cae998f90ea3e7248e8d931318f3483eccbd5193 --- /dev/null +++ b/NeurIPS/2025/4D-LRM_ Large Space-Time Reconstruction Model From and To Any View at Any Time/full.md @@ -0,0 +1,823 @@ +# 4D-LRM: Large Space-Time Reconstruction Model From and To Any View at Any Time + +Ziqiao Ma $^{1,2}$ Xuweiyi Chen $^{4}$ Shoubin Yu $^{1,3}$ + +Sai Bi $^{1}$ Kai Zhang $^{1}$ Ziwen Chen $^{1,5}$ Sihan Xu $^{2}$ Jianing Yang $^{1,2}$ + +Zexiang Xu $^{1}$ Kalyan Sunkavalli $^{1}$ Mohit Bansal $^{3}$ Joyce Chai $^{2}$ Hao Tan $^{1}$ + +$^{1}$ Adobe Research $^{2}$ University of Michigan + +$^{3}$ UNC Chapel Hill $^{4}$ University of Virginia $^{5}$ Oregon State University + +https://4dlrm.github.io/ + +![](images/f57d3292a2ecef2c6532601ec7b2ed51a814fede5a0b9b89d354b644c2e1c0e6.jpg) + +![](images/75b98e74889eb7d25f803dc72b60c06a98ad08d14b1d6ebd93c97421657c9413.jpg) + +![](images/91331939dc98e9276257d67ce7c99fcc32ed5bca3c3356bab8384ed77c25e058.jpg) + +![](images/5ca8bde311e820a9b8d9ff916dbde58ae9da96ca978a01bb2c8840d95f65c648.jpg) +Figure 1: Large Space-Time Reconstruction Model (4D-LRM) is a data-driven 4D reconstruction model that takes sparse input views at any time and renders arbitrary novel view-time combinations. + +![](images/d274759d95e00c04e79c8162d4e83324fb28aa74bf966dc5d9b19154ceabf3e8.jpg) + +![](images/639809845143c99723431dd057276974b24a5e249bb993f4429889d40ec5b14c.jpg) + +# Abstract + +Can we scale 4D pretraining to learn general space-time representations that reconstruct an object from a few views at some times to any view at any time? We provide an affirmative answer with 4D-LRM, the first large-scale 4D reconstruction model that takes input from unconstrained views and timestamps and renders arbitrary novel view-time combinations. Unlike prior 4D approaches, e.g., optimization-based, geometry-based, or generative, that struggle with efficiency, generalization, or faithfulness, 4D-LRM learns a unified space-time representation and directly predicts per-pixel 4D Gaussian primitives from posed image tokens across time, enabling fast, high-quality rendering at, in principle, infinite frame rate. Our results demonstrate that scaling spatiotemporal pretraining enables accurate and efficient 4D reconstruction. We show that 4D-LRM generalizes to novel objects, interpolates across time, and handles diverse camera setups. It reconstructs 24-frame sequences in one forward pass with less than 1.5 seconds on a single A100 GPU. + +# 1 Introduction + +Reconstructing dynamic objects and scenes from video is a fundamental problem in computer vision full of challenges. The ability to accurately capture both spatial structure and temporal dynamics to build a complete 4D representation from limited visual inputs, across varying views and time, would significantly advance applications, such as 4D asset generation [78, 53] for video games, film production, and AR/VR, as well as world modeling [32, 96] for embodied AI and robotics. + +Prior work on 4D modeling generally follows three directions, each shaped by different assumptions, target applications, or inherent limitations. The first direction is optimization-based. These methods reconstruct space and time by optimizing per scene or object from multi-view videos [3, 28, 82]. While these methods can produce high-quality results, they require dense spatial and temporal sampling, limiting their practicality with sparse inputs. The second direction is geometry-based. Motivated by Geometry Transformers [71, 66], they aim to estimate dynamic geometry and extract camera poses or depth maps directly from videos [90, 19, 70]. This line is orthogonal to the previous line of work, as these methods are not intended for novel view or time synthesis and instead focus on per-frame geometry estimation [90]. The third direction is generation-based, aiming to produce perceptually plausible 4D assets using visual generative models, particularly video diffusion models [73, 76, 78, 83]. These methods require fewer inputs, but are still computationally expensive, sensitive to prompts [36], or tailored for single-view monocular videos (Figure 2a) [53, 78]. As Yao et al. [83] noted, generating dynamic 3D content from a single-view video is inherently ill-posed for reconstruction due to motion ambiguity. Thus, these methods prioritize perceptual quality over faithful 4D reconstruction. + +Recent advances in Large Reconstruction Models (LRMs) [24, 92, 98] offer a promising rendering-based alternative toward efficient and high-quality 3D reconstruction. Based on Transformer architectures, LRMs have shown strong performance on 3D reconstruction tasks by learning powerful priors over shape and appearance from large-scale 3D datasets. This also enables them to reconstruct detailed objects and scenes from only a few posed images. However, existing LRMs are designed for static 3D objects or scenes. Although recent work has explored adapting them for 4D asset generation [53] and scene-level reconstruction with limited camera dynamics [81, 39], extending LRMs to general 4D reconstruction remains challenging, particularly when dealing with sparse multi-views and missing timestamps. We envision that an ideal 4D reconstruction model should be able to learn accurate spatiotemporal representations from a limited set of input views at a few timestamps, enabling reconstruction at novel view-times by effectively sharing information across both viewpoint and time (Figure 2b). This motivates a fundamental question for 4D modeling: + +Can we scale 4D pretraining to learn a generic space-time representation that reconstructs an object from a few views at some time points, to any view at any time? + +We introduce 4D-LRM, a Transformer-based large space-time reconstruction model for dynamic object reconstruction, trained in a data-driven manner. Inspired by 4D Gaussian Splatting (4DGS; [82]), 4D-LRM adopts a unified treatment of space and time, representing a dynamic object as a cloud of anisotropic 4D Gaussians. As illustrated in Figure 3, we patchify temporally posed input images into image tokens, which are processed by Transformer blocks. The model then directly regresses per-view, per-pixel 4D Gaussian primitives from contextualized multi-view tokens across time. These predicted 4DGS primitives enable fast, high-quality reconstruction and rendering from any viewpoint at any timestamp, with, in principle, an infinite frame rate. We train 4D-LRM on a curated subset of Objaverse4D [13], consisting of dynamic, articulated objects captured over time. The model scales effectively with more data and larger model size, and generalizes well to novel objects. + +To the best of our knowledge, 4D-LRM is the first large-scale 4D reconstruction model that supports input from unconstrained posed views and timestamps, and renders arbitrary novel view-time combinations. With around 300M parameters, 4D-LRM achieves a high-quality reconstruction on Consistent4D [28] and the hold-out test set of Objaverse4D using only one input view per frame. It reconstructs a 24-frame dynamic object in one forward pass with less than 1.5 seconds on a single A100 GPU. Compared to per frame 3D reconstruction, 4D-LRM exhibits strong and robust performance under diverse input camera configurations. We attribute this to 4D-LRM's ability to jointly model spatial and temporal contexts, effectively resolving motion ambiguities by sharing information across views and time. We further unlock its application to 4D asset generation, which surpasses baselines in both faithfulness and inference speed. Finally, we present detailed scaling behavior analyses for both training and inference, examining the scalability of different design choices and how inference + +![](images/9a89888bd9036bbc6f09b436013d9ea4e83c0d9d5b61e6eb03df191032a994b8.jpg) +Figure 2: Comparison between previous generative 4D modeling methods (e.g., L4GM [53] and SV4D [78, 83]) and our goal of generic 4D reconstruction. Prior approaches take a single monocular video as input and use generative priors to synthesize multi-view images for the first frame. In contrast, our objective is to reconstruct dynamic objects from any viewpoint at any timestamp. + +![](images/b6ac86c9d9d1c4534c3a2730ec5524f323a67a8eb12a6f079f3c5d3d6846d919.jpg) + +performance varies with the number of input views. This highlights future directions in developing 4D-LRM variants that can handle longer contexts [98] and support test-time training [11]. + +# 2 Large Space-Time Reconstruction Model (4D-LRM) + +# 2.1 Preliminary: 4D Gaussian Splatting (4DGS) + +3D Gaussian Splatting. With 3D Gaussian Splatting (GS; [31]), a static 3D scene can be represented as a cloud of anisotropic 3D Gaussians. Each Gaussian is in principle unbounded and, unless filtered, contributes to a point $x \in \mathbb{R}^3$ via an unnormalized density: + +$$ +p (x | \mu , \Sigma) = \exp \left[ - \frac {1}{2} (x - \mu) ^ {T} \Sigma^ {- 1} (x - \mu) \right], \tag {1} +$$ + +where $\mu = (\mu_x, \mu_y, \mu_z) \in \mathbb{R}^3$ is the mean and $\Sigma \in \mathbb{R}^{3 \times 3}$ is the covariance. The covariance is factorized as $\Sigma = R S S^T R^T$ , with $S = \mathrm{diag}(s_x, s_y, s_z)$ and $R$ derived from a unit quaternion $q$ . + +Pixel-Aligned Gaussians. GS-based reconstruction models [6, 60, 92] adopt pixel-aligned Gaussian rendering, where the center of each 3D Gaussian is computed from the ray distance and known camera parameters. Given ray origin $\mathrm{ray}_o$ , direction $\mathrm{ray}_d$ , and distance $\delta$ , the center is $\mu = \mathrm{ray}_o + \delta \cdot \mathrm{ray}_d$ . Therefore, each Gaussian can be parameterized with $\dim_{3\mathrm{DGS}} = 12$ , including 3-channel RGB color, 3-channel scale, 4-channel rotation quaternion, 1-channel opacity, and 1-channel ray distance. + +4D Gaussian Splatting. When considering a dynamic scene, Yang et al. [82] observed that treating space and time as independent, i.e., assuming $p_i(x,y,z|t) = p_i(x,y,z)$ for the $i$ -th visible Gaussian, is undesirable. Instead, they extend the formulation of Kerbl et al. [31] to model dynamic scenes with a unified treatment of spatial and temporal dimensions using a coherent 4D Gaussian representation (4DGS; [82]). The mean of 4DGS is given by $\mu = (\mu_x,\mu_y,\mu_z,\mu_t)\in \mathbb{R}^4$ , which captures both the spatial and temporal centers. 4DGS parameterizes its covariance matrix $\Sigma$ as a 4D ellipsoid $\Sigma = RSS^{T}RT$ , where $S = \mathrm{diag}(s_x,s_y,s_z,s_t)$ is a diagonal scaling matrix and $R\in \mathbb{R}^{4\times 4}$ is a 4D rotation matrix. In 4D Euclidean space, $R$ can be decomposed into a pair of left and right isoclinic rotations, each represented by a quaternion. Together, a general 4D Gaussian can be parameterized with $\dim_{4\mathrm{DGS}} = 20$ , including 3-channel RGB color, 4-channel scale, two 4-channel quaternions, 1-channel opacity, and the 4-channel space-time centers of order xyzt. + +Sampling Conditional 3DGS from 4DGS. As is shown in Figure 3, the marginal probability $p_i(t)$ at time $t$ is a one-dimension Gaussian $\mathcal{N}(t; \mu_4, \Sigma_{4,4})$ . The conditional 3DGS can be derived from the properties of the multivariate Gaussian with: + +$$ +\mu_ {x y z | t} = \mu_ {1: 3} + \Sigma_ {1: 3, 4} \Sigma_ {4, 4} ^ {- 1} (t - \mu_ {4}), \tag {2} +$$ + +$$ +\Sigma_ {x y z \mid t} = \Sigma_ {1: 3, 1: 3} - \Sigma_ {1: 3, 4} \Sigma_ {4, 4} ^ {- 1} \Sigma_ {4, 1: 3}. +$$ + +This decomposition enables direct adaptation of the 3DGS tile-based rasterizer by first evaluating the marginal distribution $p_i(t)$ , allowing for the accumulation of color and opacity over time. More details are available in Appendix B.2. + +![](images/89ebdbf31b65d59e7f92f7791a7d18d6e32692e9041d4644c84e63958d4bf68f.jpg) +Figure 3: Overview of 4D-LRM. 4D-LRM adopts a unified treatment of space and time, representing a dynamic object as a cloud of anisotropic 4D Gaussians [82]. We train a simple Transformer to regress 4D Gaussian primitives from a set of images with camera poses and timestamps. Each input image is tokenized by patchifying the temporally posed frames. The resulting multi-view image tokens are concatenated in temporal order and passed through a series of transformer blocks. An optional set of $N$ learnable free Gaussian tokens append the image tokens for greater generative flexibility. + +# 2.2 Transformer-Based Image-to-4DGS Decoder + +Tokenizing Temporally Posed Images. As shown in Figure 3, the inputs to our model are $V$ images from arbitrary view-time combinations, denoted as $\{\mathbf{I}_j\in \mathbb{R}^{H\times W\times 3}\}$ for $j = 1,2,\dots ,V$ , along with their corresponding camera intrinsic and extrinsic parameters. Here, $H$ and $W$ represent the image height and width, respectively. For pose conditioning, we compute Plücker ray coordinates [48] for each image, resulting in $\{\mathbf{P}_j\in \mathbb{R}^{H\times W\times 6}\}$ . Instead of the standard canonical Plücker coordinates $[\mathrm{ray}_d,\mathrm{ray}_o\times \mathrm{ray}_d]$ , we follow GS-LRM [92] and represent each ray as a direction plus its closest point to the origin $[\mathrm{ray}_d,\mathrm{ray}_o - \langle \mathrm{ray}_o,\mathrm{ray}_d\rangle \cdot \mathrm{ray}_d]$ , which is suitable for pose-sensitive learning and pixel alignment. Temporal conditioning is provided by a timestamp map $\{\mathbf{T}_j\in \mathbb{R}^{H\times W\times 1}\}$ . We channel-wise concatenate the RGB values, Plücker coordinates, and time to obtain a per-view feature map $\widetilde{\mathbf{I}}_j = \mathrm{Concat}(\mathbf{I}_j,\mathbf{P}_j,\mathbf{T}_j)$ of 10 channels, enabling per-pixel pose and time conditioning, naturally serving as spatial and temporal embeddings to distinguish different patches. Therefore, we do not use additional positional embeddings. Following Vision Transformer (ViT) architectures [16], we divide each per-view feature map into non-overlapping patches of size $PS^2$ . Each patch is flattened into a vector embedding of size $10\cdot PS^2$ , and then projected to a token of dimension $D$ (the Transformer width) using a linear layer. + +Decoding Per-Pixel Gaussians with Transformer. We concatenate the image tokens and feed them through $L$ layers of Transformer blocks. Each block follows a standard architecture consisting of PreLayerNorm [2], multi-head self-attention [64], and MLP, all equipped with residual connections [23]. The output tokens are then unpatchified and decoded into pixel-aligned 4D Gaussian primitives using a single linear layer. Given the same patch size, this results in $V \times H \times W$ Gaussians, each with $\dim_{4\mathrm{DGS}} = 20$ . As in prior GS-based LRMs, we adopt pixel-aligned Gaussian rendering. From each decoded 4D Gaussian parameter $\mathbf{g} \in \mathbb{R}^{20}$ , we split the 4-channel space-time vector $(\mathbf{g}_{\mathrm{x}}, \mathbf{g}_{\mathrm{y}}, \mathbf{g}_{\mathrm{z}}, \mathbf{g}_{\mathrm{t}})$ , retain the time $\mu_t = \mathbf{g}_t$ , and normalize the xyz features to a scalar distance $\delta$ . Further details on 4DGS parameter initialization and differentiable rasterization are provided in Appendix B. + +Optional Free Gaussians. Pixel-aligned Gaussians scale naturally with input resolution, making them well-suited for generalization to higher resolutions [92]. However, this design becomes suboptimal for very sparse views or setups with limited view coverage over motion, such as those used in standard generative 4D modeling [53, 78]. To address this, we optionally introduce an additional set of $N$ + +![](images/506a1b60c9622a7bc7e7173476808a75491c6e9b1671019e7fb5d2172aa0605f.jpg) + +![](images/04cdc12a54523a45dd99d6754bc0f5684f5d0617f51b1443a63bb9fa3b47ee4d.jpg) + +![](images/e19613965d3fe072e9dc6a0d00ddadfc1fe34797a6b34a266c15a912eac6195b.jpg) + +![](images/e3e86f363a16ecebca66954d0c0e500c42df6308f7e48f60f265b879c6faf795.jpg) +Figure 4: Different types of camera setups in evaluations: Alternating Canonical Views (4 camera poses, 24/24 timestamps seen, 24 input views in total); Frame Interpolation (4 camera poses, 12/24 timestamps seen, 24 input views in total); Two Rotating Cameras (24 camera poses, 24/24 timestamps seen, 24 input views in total); Single View Video [53, 78] (4 camera poses on the first frame plus a single view video for subsequent frames, 24/24 timestamps seen, 27 input views in total); Random Input Views (random poses, 24/24 timestamps seen, 24 input views in total). + +![](images/48f6604bc475a9b292fbce3b40590b5c9641e7c954fa5c29a2edc5920a7b648e.jpg) + +learnable Gaussian tokens, concatenated with the image tokens. These tokens allow the model to generate freeform 4D Gaussian primitives. Unlike pixel-aligned Gaussians, these do not rely on pose or time conditioning. Instead, we use a separate linear layer to decode the 4-channel space-time vector $(\mathbf{g}_{\mathrm{x}},\mathbf{g}_{\mathrm{y}},\mathbf{g}_{\mathrm{z}},\mathbf{g}_{\mathrm{t}})$ that directly defines the Gaussian center $\mu = (\mu_{x},\mu_{y},\mu_{z},\mu_{t})$ after activation. We describe in Section 3.1 how 4D-LRM can be fine-tuned for 4D asset generation with free Gaussians. + +# 2.3 Training Objectives + +During training, we render images at $U$ supervision views using the predicted 4D Gaussians and minimize the image reconstruction loss. Let $\{\mathbf{I}_{i'}^* \mid i' = 1,2,\dots,U\}$ denote the ground truth views and $\{\widehat{\mathbf{I}}_{i'}^*\}$ the corresponding rendered images. The training loss combines Mean Squared Error (MSE) and Perceptual loss [9]: + +$$ +\mathcal {L} = \frac {1}{U} \sum_ {i ^ {\prime} = 1} ^ {U} \left(\operatorname {M S E} \left(\widehat {\mathbf {I}} _ {i ^ {\prime}} ^ {*}, \mathbf {I} _ {i ^ {\prime}} ^ {*}\right) + \lambda \cdot \text {P e r c e p t u a l} \left(\widehat {\mathbf {I}} _ {i ^ {\prime}} ^ {*}, \mathbf {I} _ {i ^ {\prime}} ^ {*}\right)\right), \tag {3} +$$ + +where $\lambda$ controls the weight of the perceptual loss and is set to 0.5 empirically. + +# 3 Experiments + +# 3.1 Implementation Details + +Training Data. To enable large-scale training, we construct a 4D dataset derived from Objaverse [13], which provides a subset of animated 3D assets. However, the raw dataset is not directly suitable for 4D modeling: object motions are often inconsistent, and the dataset contains duplicates and artifacts. To address this, we build upon the filtered subset curated by Diffusion4D [38], which removes static objects and unstable motion sequences, leading to 32,000 animated objects. For each object, we render a 24-frame video from diverse camera trajectories. We augment the dataset with 783,000 static 3D objects from Objaverse by treating each as a 24-frame video, applying minor frame-by-frame displacements along a single random direction. More details are available in Appendix B.3. + +Benchmark Data. We use the Consistent4D [28] objects for evaluation. Additionally, we hold out 56 challenging objects exhibiting more complex motion as an extended test set to support future benchmarking. For evaluation, we re-render the first $48(2 \times 24)$ frames of the Consistent4D dataset and the first 24 frames of the Objaverse4D (Test) split. During training, we exclude 6 (out of the 7) Consistent4D test objects that also appear in Objaverse from our training set to avoid data leakage. + +Curriculum Learning. We adopt a curriculum learning strategy to reduce computational cost. Specifically, we pretrain the model at a resolution of $128 \times 128$ for 100,000 steps, and then continue at $256 \times 256$ for an additional 20,000 steps. The continual pretraining stage uses the same model architecture, initialized with the same pre-trained weights, but processes more tokens due to more + +Table 1: Breakdown evaluation of each camera setup on Consistent4D (Re-rendered) dataset. For each setup, we evaluate and average the score on 4 canonical views and 1 randomly sampled view. We consider different Resolutions and model Initialization strategies and compare to GS-LRM [92]. + +
Res.ModelsInit.Alter.FrameInterpolationTwo Rotating CamerasRandom Input Views
PSNRLPIPSSSIMPSNRLPIPSSSIMPSNRLPIPSSSIMPSNRLPIPSSSIM
1284D-LRM-BaseNo29.2330.0470.96129.1940.0470.96125.0140.0710.92625.7880.0810.926
4D-LRM-LargeNo30.2740.0380.96930.2600.0380.96925.9040.0610.93526.5250.0700.934
4D-LRM-LargeYes31.0230.0310.97430.9170.0310.97328.7030.0420.95928.7890.0490.957
256SoM [69]-25.5860.0550.906---22.7560.0890.94117.6370.2080.875
GS-LRM (Per Fr.)-19.3270.0970.902---20.0370.0940.93516.8260.2120.801
GS-LRM (All in)-21.6060.0860.92521.5900.0860.92520.6410.1000.90919.6650.1320.897
4D-LRM-BaseNo27.4430.0620.95227.3940.0620.95323.4290.0880.91823.8820.0960.916
4D-LRM-LargeNo27.8600.0490.95927.8220.0480.95925.0950.0690.93425.7760.0730.934
4D-LRM-FreeYes30.3960.0360.97330.3760.0360.97326.1840.0610.94326.3370.0670.939
4D-LRM-LargeYes32.1770.0280.98032.1450.0280.98027.6640.0500.95727.9900.0570.954
+ +![](images/271f2df867cdae059fc313f31e24f5816f379953ad309a058da7ff4d722425cd.jpg) +Figure 5: Visual comparison with GS-LRM [92] using (a) all input views across time and (b) three random views from the same timestamp. 4D-LRM reconstructs novel view-time combinations by learning spatiotemporal representations from sparse inputs, outperforming per-frame reconstruction by effectively sharing information across both space and time. + +pixel-aligned Gaussians in higher resolution. At each training step for object-level data, we randomly sample 36 images (from 144 renderings over 24 frames) as a training example. From this set, we independently select 12 input views and 24 supervision views, allowing overlap to improve convergence. Both pretraining stages are performed on 160 A100 GPUs. The first stage adopted a per-GPU batch size of 16 and took approximately 5 days, while the second stage adopted a per-GPU batch size of 8 with an additional 5 days. Unless otherwise specified, most training configurations follow GS-LRM [92]. Additional implementation and training details are provided in Appendix B.4. + +Fine-tuning 4D-LRM for 4D Generation. In the pre-training stage, we do not include any free Gaussians. We fine-tune 4D-LRM with $N = 4096$ free Gaussians for the 4D generation task similar to the setting of [53, 78]. For each training example over 24 frames, we select 4 canonical views at the initial frame and all views from a single-view monocular video as input. We randomly sample 8 supervision views. This is trained on 64 A100 GPUs with a per-GPU batch size of 8 for 16,000 steps. + +# 3.2 Experiment Setups + +Input Camera Setup. Since 4D-LRM supports arbitrary input views at any time, we define several camera setup configurations fore broad and systematic evaluation (see Figure 4 for illustrations): + +- Alternating Canonical Views. The input view alternates cyclically among four canonical directions (front, left, back, and right) across frames; + +Table 2: Breakdown evaluation of each camera setup on Objaverse4D (Test) dataset. For each setup, we evaluate and average the score on 4 canonical views and 1 randomly sampled view. We consider different Resolutions and model Initialization strategies and compare to GS-LRM [92]. + +
Res.ModelInit.Alter.Frame InterpolationTwo Rotating CamerasRandom Input Views
PSNRLPIPSSSIMPSNRLPIPSSSIMPSNRLPIPSSSIMPSNRLPIPSSSIM
1284D-LRM-BaseNo27.3740.0680.93727.2870.0680.93725.4960.0750.92424.1740.1120.893
4D-LRM-LargeNo28.4890.0550.94928.4400.0550.94926.3120.0640.93425.0230.0960.905
4D-LRM-LargeYes29.2510.0420.95929.1690.0430.95828.6760.0430.95727.5860.0640.939
256GS-LRM (Per Frame)-18.7960.1640.854---19.7290.1430.91118.3000.1760.845
GS-LRM (All in)-19.3880.1380.88819.4120.1390.88819.4280.1380.88819.3790.1380.888
4D-LRM-BaseNo25.8060.0850.92825.7110.0850.92923.9740.0850.92422.4090.1250.889
4D-LRM-LargeNo26.6580.0660.94226.5800.0660.94225.5240.0670.93724.4610.0950.913
4D-LRM-FreeYes28.8380.0500.95828.7900.0510.95826.9980.0560.94925.2670.0820.923
4D-LRM-LargeYes30.0940.0410.96730.0280.0420.96727.8100.0490.95726.6940.0720.939
+ +![](images/b7fed90335a16fe7e4881160465df3599bdcab7e8bd36e1a678c45c2ab758ec0.jpg) +Figure 6: Qualitative examples of 4D-LRM taking views with missing frames, including Frame Interpolation (only half timestamps seen, 24 input views) and Random Views at Random Frames (24 input views from random views and times). + +- Frame Interpolation. Two canonical views are provided at even-numbered frames, while all odd-numbered frames are omitted from the input. This setup is designed to evaluate the model's interpolation ability across time; +- Two Rotating Cameras. Two virtual cameras rotate from front to back, one sweeping from the left and the other from the right. The left-rotating camera provides views at odd-numbered frames and the other for even-numbered frames, creating a complementary dual-view sequence; +- Random Input Views. At each frame, a single view is randomly selected from the available camera positions. This setting introduces high variability that requires robustness to unstructured input. + +Baselines. To the best of our knowledge, 4D-LRM is the first large-scale 4D reconstruction model that supports inputs from unconstrained views and timestamps, and enables rendering at arbitrary novel view-time combinations. The closest baselines are recent multi-view diffusion models [73, 76] and feedforward models [39] that support camera pose and time control, though these methods are not publicly available. For setups from moving cameras, we compare 4D-LRM to GS-LRMs run in a per-frame fashion or directly with all views jointly. We also include Shape of Motion (SoM) [69], an optimization-based method that models videos as causal, one-way motion trajectories. It represents scene motion using compact SE(3) bases, enforcing consistent forward movement throughout dynamic scenes. Following its guidelines, we manually segment the dynamic region and run 3,000 optimization iterations. In addition, we perform systematic ablation studies on 4D-LRM, evaluating two model scales: Large (300M parameters; 1024 hidden size, 24 layers, 16 attention heads) and Base (85M parameters; 768 hidden size, 12 layers, 12 attention heads). We also explore the effects of different Resolution settings and Initialization strategies, i.e., whether initializing the Transformer with weights from GS-LRM improves performance, using the same training setup and number of steps. Unless otherwise specified, we use 4D-LRM-Large with initialization by default. More design choices are discussed with training-time scaling in Section 4 and Appendix B.4. + +Table 3: Comparison between 4D-LRM and GS-LRM under varying numbers of input views per frame (VPF). Rand.: randomly selected input views; Canon.: 4 canonical input views at each frame. + +
InputVPFModelObjaverse4D (Test)Consistent4D (Re.)
PSNRLPIPSSSIMPSNRLPIPSSSIM
Rand.1GS-LRM18.7380.1650.85217.2010.1660.846
4D-LRM28.3430.0400.96430.5130.0360.972
2GS-LRM22.2520.1050.89822.4250.0970.904
4D-LRM28.6220.0510.95730.6010.0350.973
3GS-LRM24.3810.0830.91724.6610.0790.924
4D-LRM28.2120.0520.95430.5540.0350.972
4GS-LRM26.1180.0700.93326.1260.0700.935
4D-LRM27.9400.0530.95330.4450.0340.972
Canon.4GS-LRM28.7100.0470.96230.0670.0380.967
4D-LRM27.8390.0550.95230.8500.0390.968
+ +Table 4: Application to 4D generation on the original Consistent4D benchmark. ${}^{ \ddagger }$ Using ground truth multi-view reference in the first frame as the skyline. + +
ModelPSNRLPIPSSSIMFVDCLIP
Consistent4D [28]-0.160-1,133.440.87
DG4D [52]-0.160--0.87
4Diffusion [89]-0.165--0.88
Efficient4D [45]-0.130--0.92
GaussianFlow [22]-0.140--0.91
4DGen [85]-0.130--0.89
STAG4D [88]-0.130-992.210.91
SV4D [78]-0.129-677.680.93
L4GM [53]-0.120-691.870.94
4D-LRM-Large20.0940.1380.8851,063.880.87
4D-LRM-Free23.7770.1170.916677.580.94
4D-LRM-Free†26.1180.0550.947674.590.96
+ +![](images/07a41b655ae64f9ba39a78455d89c8fc7cac98db5a862021b300a23b623cf01a.jpg) +Figure 7: Visual comparisons of 4D-LRM(-Free) to generation-based 4D models. For fair comparisons, we initialize each model with the first frame with the ground truth multi-view images. + +![](images/d897133d4ad89e1f578a5bc0944a65395cdf31e98f793a0f66121d56fb4b6a05.jpg) + +![](images/bc0b43a6e5df3b3a267a8e0428b7756e4d9dd93cd7f7928ad6ff6b3acabaeee5.jpg) + +Metrics. We follow previous work to adopt PSNR [5], SSIM [72], LPIPS [93] metrics. For a broad coverage of evaluation views, we evaluated and averaged the metrics on 4 canonical views and 1 randomly sampled view at each frame. + +# 3.3 Main Results: 4D Reconstruction + +On both the re-rendered Consistent4D dataset (Table 1) and the Objaverse4D dataset (Table 2), 4D-LRM demonstrates strong performance across a variety of camera configurations. Among the four tested setups, alternating canonical views provide the greatest view coverage over time and represent the least challenging configuration, under which 4D-LRM achieves PSNR scores exceeding 30. Notably, 4D-LRM remains robust even under more difficult settings, such as when half of the frames are omitted, or in configurations with limited spatiotemporal coverage, including two rotating cameras and random input views. We also find that increasing model size leads to improved performance under identical data and training steps. Additionally, initializing the Transformer with weights from GS-LRM further enhances performance and accelerates convergence. + +The above experiments use 24 input views over a 24-frame motion sequence, corresponding to a only one view per frame. Since GS-LRM is designed for sparse-view reconstruction, we further compare it with 4D-LRM under denser input conditions. Specifically, we evaluate both models using: (1) multiple randomly selected input views per frame, and (2) four canonical views per frame, with the task of rendering a single randomly chosen novel view per frame. The results of these comparisons are presented in Table 3. We observe that GS-LRM performs comparably only when sufficient view coverage is available, as illustrated in Figure 5. In contrast, 4D-LRM consistently outperforms per-frame 3D reconstruction methods by leveraging spatial and temporal information jointly. As shown in Table 1, while SoM performs fair under structured settings such as canonical or rotating views, it struggles in the random input view scenario due to its limited capacity to handle unconstrained spatiotemporal inputs. This highlights the advantage of 4D-LRM's learned, generalizable representation for arbitrary view-time combinations. Finally, qualitative results in Figure 6 highlight 4D-LRM's ability to generalize to novel objects and interpolate effectively over time, even when input frames are missing. + +# 3.4 Application: 4D Generation. + +We demonstrate that 4D-LRM can be extended to a 4D generation setup. Specifically, we chain 4D-LRM (fine-tuned with free Gaussians) with SV3D [65], and compare it against existing generation- + +![](images/c501039903868a5f9dad4a2efb43cf91db9b71895c45f889431945679aa78181.jpg) +(a) $\mu_t$ w/o interpolation. + +![](images/7f988e2436aa779a2c8f45f42b9766bb680c3dbdbb86ccfd3d0ee6ec900a7168.jpg) +(b) $\mu_t$ w/ interpolation. + +![](images/f1842f3862710ce11ae19617e655fcf011f1197e2bfbdab73da8846b6978899c.jpg) +(c) $\Sigma_t$ w/o interpolation. + +![](images/3803c7787512c58d62de81f807521be6ccc0a0ff1b8e90ca29eb9d170821022c.jpg) +(d) $\Sigma_{t}\mathrm{w}/$ interpolation. + +![](images/b2acfb34c1ac905958b84b9a954c4914168551aaf438a0d0623bb24120d90c63.jpg) +Figure 8: We visualize the distributions of $\mu_t$ and $\Sigma_t$ under Alternating Canonical Views and Frame Interpolation setups on 24 frames with the same dynamic object. +(a) PSNR vs. #Training Steps. +Figure 9: Training-time scaling curves. Tested on Consistent4D (re-rendered). We compute the PSNR, SSIM and LPIPS for different number of different setups, with a 4D-LRM-base as the model. = + +![](images/bf8560a62e7dd1c34fd44de70175647141f5399f90c46c6ae3673e130ea015b8.jpg) +(b) SSIM vs. #Training Steps. + +![](images/488041423e27e9b51f92794a19a186d743c8bff3cbccc215dcb8d0356706a52e.jpg) +(c) LPIPS vs. #Training Steps. + +based methods. When paired with a diffusion model as a generative prior, 4D-LRM outperforms baseline 4D generation approaches on the original Consistent4D benchmark. This improvement is due to 4D-LRM's ability to produce much more faithful reconstructions of dynamic objects, even in the presence of motion ambiguity. Since diffusion models introduce high variance, we provide a comparison in Figure 7, where we evaluate 4D-LRM alongside other generative 4D models [53, 78] using ground-truth multi-view inputs from the first frame to ensure a fair comparison. Moreover, the core 4D-LRM model is efficient, requiring less than 1.5 seconds per forward pass; the primary computational bottleneck lies in the diffusion model. + +# 4 Analyses and Discussions + +Why Can 4D-LRM Interpolate Frames? 4D-LRM demonstrates strong frame interpolation capabilities, which aligns with its design: time is modeled as a continuous distribution rather than as discrete steps. To better understand this behavior, we analyze the 4DGS primitives predicted for the first 24 frames of the Guppie object in Consistent4D under two settings: Alternating Canonical Views (24 known timestamps) and Frame Interpolation (12 known timestamps). We visualize the distributions of the temporal mean $\mu_t = \mu_4$ and variance $\Sigma_t = \Sigma_{4,4}$ in Figure 10. Interestingly, when some timestamps are missing, 4D-LRM learns to reallocate certain Gaussians toward these missing regions, effectively filling the temporal gaps. Moreover, in the interpolation setting, the predicted 4DGS primitives tend to have larger $\Sigma_t$ , increasing their temporal support. This allows each Gaussian to influence a broader range of neighboring timestamps after sampling, improving interpolation quality and temporal coverage. + +Training-Time Scaling. To understand how different design considerations affect the training efficiency, we provide the scaling behavior with the following configurations to 4D-LRM-Base. + +- 4D-LRM-Base: Transformer with 768 hidden dimensions, 12 layers, and 12 attention heads, trained with 12 random input views and 12 random target views. No free Gaussians. +- # Target $\times$ 2: Trained with 12 random input views and 24 random target views. +- w/ Hexplane: Instead of unified space-time representation, Wu et al. [75] proposed an alternative 4DGS representation with decomposed neural voxel encoding inspired by HexPlane [3]. +- w/ Temp Align: Similar to the idea of pixel-aligned Gaussians, we force $\mu_T$ to the input frame time, reducing the parameterization to $\dim_{4\mathrm{DGS}} = 19$ . +- w/ Free GS: Trained with $N = 1024$ free Gaussian tokens from scratch. + +We observe that increasing the number of target views slightly improves convergence speed, though at the cost of increased iteration time. Introducing free Gaussians from scratch does not significantly impact reconstruction quality but substantially slows down training. Additionally, we find that the 4DGS representation from [75] is less expressive than the unified space-time formulation proposed + +![](images/6b7db1107a205b83781e599be17c1b45b01e9a0c80350f02174e7b5774213da1.jpg) +(a) PSNR vs. #Input View. + +![](images/2b1a237ed7db77ff394457c625381586da04b21751f6061c452e2a8ea3ef744e.jpg) +(b) SSIM vs. #Input View. +Figure 10: Inference-time scaling curves. Tested on Consistent4D (re-rendered). We compute the PSNR, SSIM, and LPIPS for different numbers of randomly selected input views. + +![](images/cf808a2f8fd5309719a14d2ab965f552b1f6a0207704e67549f3d9b81d44ae52.jpg) +(c) LPIPS vs. #Input View. + +by [82], which informed our final design choice. We also note that enforcing strict temporal alignment degrades performance, whereas pixel alignment improves reconstruction quality. This supports our earlier observation that 4D-LRM effectively redistributes Gaussians to unseen time intervals to handle sparse temporal supervision. + +Inference-Time Scaling. Finally, we analyze inference-time scaling as the number of input views varies. In terms of PSNR and SSIM, performance improves with more input views and peaks at 48, after which it begins to decline slightly. We attribute this to two factors: (1) excessive Gaussians may overcrowd the 4D representation, reducing its quality, and (2) the Transformer struggles with very long input sequences. This observation suggests a promising future direction: designing 4D-LRM variants that can handle longer contexts with hybrid models [98] and incorporate test-time training [11, 94]. + +# 5 Related Work + +Prior work on 4D modeling generally falls into three broad directions: optimization-based, geometry-based, and generation-based, each shaped by different assumptions, data requirements, and target applications. The first direction is optimization-based, mostly covered in previous discussions of 4D representations, where methods reconstruct dynamic scenes by optimizing per-object or per-scene representations using multi-view video [3, 28, 82, 69]. While capable of producing high-quality reconstructions, they are typically constrained by the need for dense spatial and temporal supervision. To improve generalization and reduce test-time cost, recent methods incorporate depth supervision or lightweight tuning [95, 62]. DyST [56] adopts a fully feed-forward approach, learning a latent decomposition of content, dynamics, and pose from monocular videos via a Transformer. However, it models space-time implicitly and remains limited in novel view synthesis quality. Although recent work has explored adapting LRMs for 4D asset generation [53] and scene-level reconstruction with limited input and target camera dynamics [81, 39], extending LRMs to general 4D reconstruction remains challenging, particularly when considering any target view at any time from sparse multi-views and missing timestamps. The second direction is geometry-based, which aims to estimate dynamic scene geometry, such as depth, flow, or camera motion, directly from input videos. Inspired by static sparse-view geometry methods like DUSt3R [71], recent work has extended this paradigm to dynamic scenes [90, 19, 70]. These methods often incorporate correspondence-based supervision or monocular depth priors to recover frame-wise geometry or trajectories. Unlike optimization-based approaches, they do not model the full spatiotemporal volume and are not intended for novel view or time synthesis. The third direction is generation-based, which leverages video generative models to synthesize perceptually plausible 4D assets. This includes both explicit 4D asset synthesis [53, 78, 83] and controllable video generation [36, 73, 76]. These methods reduce dependence on multi-view inputs by relying on learned priors, typically from large-scale video diffusion models. However, they are often computationally intensive at inference time, prompt-sensitive [36], and mostly limited to monocular inputs. Reconstructing faithful 4D geometry from a single-view video remains fundamentally ill-posed due to motion ambiguity [83]. Our goal is to learn a generic space-time representation that reconstructs an object from a few views at some time points, to any view at any time. We include an expanded related work section in Appendix due to the page limitation. + +# 6 Conclusion + +This work introduces 4D-LRM, the first large-scale 4D reconstruction model capable of processing unconstrained views and timestamps to render arbitrary novel view-time combinations. By learning a unified spatiotemporal representation and directly predicting per-pixel 4D Gaussian primitives from posed image tokens over time, 4D-LRM enables fast, high-quality rendering at, in principle, infinite frame rates. + +# References + +[1] Sameer Agarwal, Yasutaka Furukawa, Noah Snavely, Ian Simon, Brian Curless, Steven M Seitz, and Richard Szeliski. Building rome in a day. Communications of the ACM, 54(10):105-112, 2011. +[2] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016. +[3] Ang Cao and Justin Johnson. Hexplane: A fast representation for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 130-141, 2023. +[4] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16123-16133, 2022. +[5] Luen C Chan and Peter Whiteman. Hardware-constrained hybrid coding of video imagery. IEEE Transactions on Aerospace and Electronic Systems, (1):71-84, 1983. +[6] David Charatan, Sizhe Lester Li, Andrea Tagliasacchi, and Vincent Sitzmann. pixelsplat: 3d gaussian splats from image pairs for scalable generalizable 3d reconstruction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19457-19467, 2024. +[7] Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14124-14133, 2021. +[8] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European conference on computer vision, pages 333-350, 2022. +[9] Qifeng Chen and Vladlen Koltun. Photographic image synthesis with cascaded refinement networks. In Proceedings of the IEEE international conference on computer vision, pages 1511-1520, 2017. +[10] Tianqi Chen, Bing Xu, Chiyuan Zhang, and Carlos Guestrin. Training deep nets with sublinear memory cost. arXiv preprint arXiv:1604.06174, 2016. +[11] Karan Dalal, Daniel Koceja, Gashon Hussein, Jiarui Xu, Yue Zhao, Youjin Song, Shihao Han, Ka Chun Cheung, Jan Kautz, Carlos Guestrin, et al. One-minute video generation with test-time training. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2025. +[12] Tri Dao. Flashattention-2: Faster attention with better parallelism and work partitioning. In The Twelfth International Conference on Learning Representations, 2023. +[13] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13142-13153, 2023. +[14] Matt Deitke, Ruoshi Liu, Matthew Wallingford, Huong Ngo, Oscar Michel, Aditya Kusupati, Alan Fan, Christian Laforte, Vikram Voleti, Samir Yitzhak Gadre, et al. Objverse-xl: A universe of $10\mathrm{m} + 3\mathrm{d}$ objects. In Conference on Neural Information Processing Systems, 2024. +[15] Kangle Deng, Andrew Liu, Jun-Yan Zhu, and Deva Ramanan. Depth-supervised nerf: Fewer views and faster training for free. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12882-12891, 2022. +[16] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2020. + +[17] Laura Downs, Anthony Francis, Nate Koenig, Brandon Kinman, Ryan Hickman, Krista Reymann, Thomas B McHugh, and Vincent Vanhoucke. Google scanned objects: A high-quality dataset of 3d scanned household items. In 2022 International Conference on Robotics and Automation (ICRA), pages 2553-2560. IEEE, 2022. +[18] Bardienus P Duisterhof, Zhao Mandi, Yunchao Yao, Jia-Wei Liu, Jenny Seidenschwarz, Mike Zheng Shou, Deva Ramanan, Shuran Song, Stan Birchfield, Bowen Wen, et al. Deformgs: Scene flow in highly deformable scenes for deformable object manipulation. In The 16th International Workshop on the Algorithmic Foundations of Robotics, 2024. +[19] Haiwen Feng, Junyi Zhang, Qianqian Wang, Yufei Ye, Pengcheng Yu, Michael J Black, Trevor Darrell, and Angjoo Kanazawa. St4rtrack: Simultaneous 4d reconstruction and tracking in the world. arXiv preprint arXiv:2504.13152, 2025. +[20] Sara Fridovich-Keil, Giacomo Meanti, Frederik Rahbæk Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12479–12488, 2023. +[21] Yasutaka Furukawa and Jean Ponce. Accurate, dense, and robust multiview stereopsis. IEEE transactions on pattern analysis and machine intelligence, 32(8):1362-1376, 2009. +[22] Quankai Gao, Qiangeng Xu, Zhe Cao, Ben Mildenhall, Wenchao Ma, Le Chen, Danhang Tang, and Ulrich Neumann. Gaussianflow: Splitting gaussian dynamics for 4d content creation. arXiv preprint arXiv:2403.12365, 2024. +[23] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. +[24] Yicong Hong, Kai Zhang, Jiuxiang Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. Lrm: Large reconstruction model for single image to 3d. In The Twelfth International Conference on Learning Representations, 2024. +[25] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically accurate radiance fields. In *Special Interest Group on Computer Graphics and Interactive Techniques Conference*, pages 1-11, 2024. +[26] Hanwen Jiang, Qixing Huang, and Georgios Pavlakos. Real3d: Scaling up large reconstruction models with real-world images. arXiv preprint arXiv:2406.08479, 2024. +[27] Hanwen Jiang, Hao Tan, Peng Wang, Haian Jin, Yue Zhao, Sai Bi, Kai Zhang, Fujun Luan, Kalyan Sunkavalli, Qixing Huang, et al. Rayzer: A self-supervised large view synthesis model. arXiv preprint arXiv:2505.00702, 2025. +[28] Yanqin Jiang, Li Zhang, Jin Gao, Weiming Hu, and Yao Yao. Consistent4d: Consistent $360^{\circ}$ dynamic object generation from monocular video. In The Twelfth International Conference on Learning Representations, 2024. +[29] Haian Jin, Hanwen Jiang, Hao Tan, Kai Zhang, Sai Bi, Tianyuan Zhang, Fujun Luan, Noah Snavely, and Zexiang Xu. Lvsm: A large view synthesis model with minimal 3d inductive bias. arXiv preprint arXiv:2410.17242, 2024. +[30] Mohammad Mahdi Johari, Yann Lepoittevin, and François Fleuret. Geonerf: Generalizing nerf with geometry priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18365-18375, 2022. +[31] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42(4):1-14, 2023. +[32] Justin Kerr, Chung Min Kim, Mingxuan Wu, Brent Yi, Qianqian Wang, Ken Goldberg, and Angjoo Kanazawa. Robot see robot do: Imitating articulated object manipulation with monocular 4d reconstruction. In 8th Annual Conference on Robot Learning, 2024. + +[33] Agelos Kratimenos, Jiahui Lei, and Kostas Daniilidis. Dynmf: Neural motion factorization for real-time dynamic view synthesis with 3d gaussian splattering. In European Conference on Computer Vision, pages 252-269, 2024. +[34] Benjamin Lefaudeau, Francisco Massa, Diana Liskovich, Wenhan Xiong, Vittorio Caggiano, Sean Naren, Min Xu, Jieru Hu, Marta Tintore, Susan Zhang, Patrick Labatut, Daniel Haziza, Luca Wehrstedt, Jeremy Reizenstein, and Grigory Sizov. xformers: A modular and hackable transformer modelling library. https://github.com/facebookresearch/xformers, 2022. +[35] Vincent Leroy, Yohann Cabon, and Jérôme Revaud. Grounding image matching in 3d with mast3r. In European Conference on Computer Vision, pages 71–91, 2024. +[36] Bing Li, Cheng Zheng, Wenxuan Zhu, Jinjie Mai, Biao Zhang, Peter Wonka, and Bernard Ghanem. Vivid-zoo: Multi-view video generation with diffusion model. In Conference on Neural Information Processing Systems, pages 62189-62222, 2024. +[37] Jiahao Li, Hao Tan, Kai Zhang, Zexiang Xu, Fujun Luan, Yinghao Xu, Yicong Hong, Kalyan Sunkavalli, Greg Shakhnarovich, and Sai Bi. Instant3d: Fast text-to-3d with sparse-view generation and large reconstruction model. In The Twelfth International Conference on Learning Representations, 2024. +[38] Hanwen Liang, Yuyang Yin, Dejia Xu, Hanxue Liang, Zhangyang Wang, Konstantinos N Plataniotis, Yao Zhao, and Yunchao Wei. Diffusion4d: Fast spatial-temporal consistent 4d generation via video diffusion models. In Conference on Neural Information Processing Systems, 2024. +[39] Hanxue Liang, Jiawei Ren, Ashkan Mirzaei, Antonio Torralba, Ziwei Liu, Igor Gilitschenski, Sanja Fidler, Cengiz Oztireli, Huan Ling, Zan Gojcic, et al. Feed-forward bullet-time reconstruction of dynamic scenes from monocular videos. arXiv preprint arXiv:2412.03526, 2024. +[40] Yiqing Liang, Numair Khan, Zhengqin Li, Thu Nguyen-Phuoc, Douglas Lanman, James Tompkin, and Lei Xiao. Gaufre: Gaussian deformation fields for real-time dynamic novel view synthesis. In 2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 2642–2652. IEEE, 2025. +[41] Xiaoxiao Long, Cheng Lin, Peng Wang, Taku Komura, and Wenping Wang. Sparseneus: Fast generalizable neural surface reconstruction from sparse views. In European Conference on Computer Vision, pages 210-227, 2022. +[42] Baorui Ma, Junsheng Zhou, Yu-Shen Liu, and Zhizhong Han. Towards better gradient consistency for neural signed distance functions via level set alignment. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 17724-17734, 2023. +[43] Paulius Micikevicius, Sharan Narang, Jonah Alben, Gregory Diamos, Erich Elsen, David Garcia, Boris Ginsburg, Michael Houston, Oleksii Kuchaiev, Ganesh Venkatesh, et al. Mixed precision training. In International Conference on Learning Representations, 2018. +[44] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European Conference on Computer Vision, 2020. +[45] Zijie Pan, Zeyu Yang, Xiatian Zhu, and Li Zhang. Efficient4d: Fast dynamic 3d object generation from a single-view video. arXiv preprint arXiv 2401.08742, 2024. +[46] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 165-174, 2019. +[47] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5865-5874, 2021. + +[48] Julius Plücker. XVii. on a new geometry of space. Philosophical Transactions of the Royal Society of London, (155):725-791, 1865. +[49] Marc Pollefeys, Luc Van Gool, Maarten Vergauwen, Frank Verbiest, Kurt Cornelis, Jan Tops, and Reinhard Koch. Visual modeling with a hand-held camera. International Journal of Computer Vision, 59:207-232, 2004. +[50] Marc Pollefeys, David Nistér, J-M Frahm, Amir Akbarzadeh, Philippines Mordohai, Brian Clipp, Chris Engels, David Gallup, S-J Kim, Paul Merrell, et al. Detailed real-time urban 3d reconstruction from video. International Journal of Computer Vision, 78:143-167, 2008. +[51] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10318–10327, 2021. +[52] Jiawei Ren, Liang Pan, Jiaxiang Tang, Chi Zhang, Ang Cao, Gang Zeng, and Ziwei Liu. Dreamgaussian4d: Generative 4d gaussian splatting. arXiv preprint arXiv:2312.17142, 2023. +[53] Jiawei Ren, Cheng Xie, Ashkan Mirzaei, Karsten Kreis, Ziwei Liu, Antonio Torralba, Sanja Fidler, Seung Wook Kim, Huan Ling, et al. L4gm: Large 4d gaussian reconstruction model. In Conference on Neural Information Processing Systems, pages 56828-56858, 2024. +[54] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4104-4113, 2016. +[55] Johannes L Schonberger, Enliang Zheng, Jan-Michael Frahm, and Marc Pollefeys. Pixelwise view selection for unstructured multi-view stereo. In European Conference on Computer Vision, pages 501-518, 2016. +[56] Maximilian Seitzer, Sjoerd van Steenkiste, Thomas Kipf, Klaus Greff, and Mehdi S. M. Sajjadi. DyST: Towards dynamic neural scene representations on real-world videos. In The Twelfth International Conference on Learning Representations, 2024. +[57] Noah Snavely, Steven M Seitz, and Richard Szeliski. Photo tourism: exploring photo collections in 3d. ACM Transactions on Graphics (TOG), 25(3):835-846, 2006. +[58] Mohammed Suhail, Carlos Esteves, Leonid Sigal, and Ameesh Makadia. Generalizable patch-based neural rendering. In European Conference on Computer Vision, pages 156-174, 2022. +[59] Mohammed Suhail, Carlos Esteves, Leonid Sigal, and Ameesh Makadia. Light field neural rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8269-8279, 2022. +[60] Jiaxiang Tang, Zhaoxi Chen, Xiaokang Chen, Tengfei Wang, Gang Zeng, and Ziwei Liu. Lgm: Large multi-view gaussian model for high-resolution 3d content creation. In European Conference on Computer Vision, pages 1-18, 2024. +[61] Zhenggang Tang, Yuchen Fan, Dilin Wang, Hongyu Xu, Rakesh Ranjan, Alexander Schwing, and Zhicheng Yan. Mv-dust3r+: Single-stage scene reconstruction from sparse views in 2 seconds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. +[62] Fengrui Tian, Shaoyi Du, and Yueqi Duan. Mononerf: Learning a generalizable dynamic radiance field from monocular videos. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17903-17913, 2023. +[63] Edgar Tretschk, Ayush Tewari, Vladislav Golyanik, Michael Zollhöfer, Christoph Lassner, and Christian Theobalt. Non-rigid neural radiance fields: Reconstruction and novel view synthesis of a dynamic scene from monocular video. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12959-12970, 2021. + +[64] Ashish Vavwani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Conference on Neural Information Processing Systems, 2017. +[65] Vikram Voleti, Chun-Han Yao, Mark Boss, Adam Letts, David Pankratz, Dmitry Tochilkin, Christian Laforte, Robin Rombach, and Varun Jampani. Sv3d: Novel multi-view synthesis and 3d generation from a single image using latent video diffusion. In European Conference on Computer Vision, pages 439–457, 2024. +[66] Jianyuan Wang, Minghao Chen, Nikita Karaev, Andrea Vedaldi, Christian Rupprecht, and David Novotny. Vggt: Visual geometry grounded transformer. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 5294-5306, 2025. +[67] Peng Wang, Hao Tan, Sai Bi, Yinghao Xu, Fujun Luan, Kalyan Sunkavalli, Wenping Wang, Zexiang Xu, and Kai Zhang. Pf-lrm: Pose-free large reconstruction model for joint pose and shape prediction. In The Twelfth International Conference on Learning Representations, 2024. +[68] Qianqian Wang, Zhicheng Wang, Kyle Genova, Pratul P Srinivasan, Howard Zhou, Jonathan T Barron, Ricardo Martin-Brualla, Noah Snavely, and Thomas Funkhouser. Ibrnet: Learning multi-view image-based rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4690-4699, 2021. +[69] Qianqian Wang, Vickie Ye, Hang Gao, Jake Austin, Zhengqi Li, and Angjoo Kanazawa. Shape of motion: 4d reconstruction from a single video. arXiv preprint arXiv:2407.13764, 2024. +[70] Qianqian Wang, Yifei Zhang, Aleksander Holynski, Alexei A Efros, and Angjoo Kanazawa. Continuous 3d perception model with persistent state. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2025. +[71] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidrovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20697-20709, 2024. +[72] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4): 600-612, 2004. +[73] Daniel Watson, Saurabh Saxena, Lala Li, Andrea Tagliasacchi, and David J Fleet. Controlling space and time with diffusion models. In The Thirteenth International Conference on Learning Representations, 2024. +[74] Xinyue Wei, Kai Zhang, Sai Bi, Hao Tan, Fujun Luan, Valentin Deschaintre, Kalyan Sunkavalli, Hao Su, and Zexiang Xu. Meshlrm: Large reconstruction model for high-quality mesh. arXiv preprint arXiv:2404.12385, 2024. +[75] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 20310-20320, 2024. +[76] Rundi Wu, Ruiqi Gao, Ben Poole, Alex Trevithick, Changxi Zheng, Jonathan T Barron, and Aleksander Holynski. Cat4d: Create anything in 4d with multi-view video diffusion models. arXiv preprint arXiv:2411.18613, 2024. +[77] Desai Xie, Sai Bi, Zhixin Shu, Kai Zhang, Zexiang Xu, Yi Zhou, Soren Pirk, Arie Kaufman, Xin Sun, and Hao Tan. Lrm-zero: Training large reconstruction models with synthesized data. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. +[78] Yiming Xie, Chun-Han Yao, Vikram Voleti, Huaizu Jiang, and Varun Jampani. Sv4d: Dynamic 3d content generation with multi-frame and multi-view consistency. arXiv preprint arXiv:2407.17470, 2024. + +[79] Yinghao Xu, Hao Tan, Fujun Luan, Sai Bi, Peng Wang, Jiahao Li, Zifan Shi, Kalyan Sunkavalli, Gordon Wetzstein, Zexiang Xu, et al. Dmv3d: Denoising multi-view diffusion using 3d large reconstruction model. In The Twelfth International Conference on Learning Representations, 2024. +[80] Jianing Yang, Alexander Sax, Kevin J Liang, Mikael Henaff, Hao Tang, Ang Cao, Joyce Chai, Franziska Meier, and Matt Feiszli. Fast3r: Towards 3d reconstruction of $1000+$ images in one forward pass. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2025. +[81] Jiawei Yang, Jiahui Huang, Yuxiao Chen, Yan Wang, Boyi Li, Yurong You, Apoorva Sharma, Maximilian Igl, Peter Karkus, Danfei Xu, et al. Storm: Spatio-temporal reconstruction model for large-scale outdoor scenes. In The Thirteenth International Conference on Learning Representations, 2025. +[82] Zeyu Yang, Hongye Yang, Zijie Pan, and Li Zhang. Real-time photorealistic dynamic scene representation and rendering with 4d gaussian splatting. In The Twelfth International Conference on Learning Representations, 2024. +[83] Chun-Han Yao, Yiming Xie, Vikram Voleti, Huaizu Jiang, and Varun Jampani. Sv4d 2.0: Enhancing spatio-temporal consistency in multi-view video diffusion for high-quality 4d generation. arXiv preprint arXiv:2503.16396, 2025. +[84] Wang Yifan, Noam Aigerman, Vladimir G Kim, Siddhartha Chaudhuri, and Olga Sorkine-Hornung. Neural cages for detail-preserving 3d deformations. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 75-83, 2020. +[85] Yuyang Yin, Dejia Xu, Zhangyang Wang, Yao Zhao, and Yunchao Wei. 4dgen: Grounded 4d content generation with spatial-temporal consistency. arXiv preprint arXiv:2312.17225, 2023. +[86] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4578-4587, 2021. +[87] Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-splatting: Alias-free 3d gaussian splatting. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19447-19456, 2024. +[88] Yifei Zeng, Yanqin Jiang, Siyu Zhu, Yuanxun Lu, Youtian Lin, Hao Zhu, Weiming Hu, Xun Cao, and Yao Yao. Stag4d: Spatial-temporal anchored generative 4d gaussians. In European Conference on Computer Vision, pages 163–179, 2024. +[89] Haiyu Zhang, Xinyuan Chen, Yaohui Wang, Xihui Liu, Yunhong Wang, and Yu Qiao. 4diffusion: Multi-view video diffusion model for 4d generation. In Neural Information Processing Systems, volume 37, pages 15272-15295, 2024. +[90] Junyi Zhang, Charles Herrmann, Junhwa Hur, Varun Jampani, Trevor Darrell, Forrester Cole, Deqing Sun, and Ming-Hsuan Yang. Monst3r: A simple approach for estimating geometry in the presence of motion. In The Thirteenth International Conference on Learning Representations, 2025. +[91] Kai Zhang, Nick Kolkin, Sai Bi, Fujun Luan, Zexiang Xu, Eli Shechtman, and Noah Snavely. Arf: Artistic radiance fields. In European Conference on Computer Vision, pages 717-733, 2022. +[92] Kai Zhang, Sai Bi, Hao Tan, Yuanbo Xiangli, Nanxuan Zhao, Kalyan Sunkavalli, and Zexiang Xu. Gs-lrm: Large reconstruction model for 3d gaussian splattering. In European Conference on Computer Vision, pages 1-19, 2024. +[93] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. + +[94] Tianyuan Zhang, Sai Bi, Yicong Hong, Kai Zhang, Fujun Luan, Songlin Yang, Kalyan Sunkavalli, William T Freeman, and Hao Tan. Test-time training done right. arXiv preprint arXiv:2505.23884, 2025. +[95] Xiaoming Zhao, R Alex Colburn, Fangchang Ma, Miguel Ángel Bautista, Joshua M. Susskind, and Alex Schwing. Pseudo-generalized dynamic view synthesis from a video. In The Twelfth International Conference on Learning Representations, 2024. +[96] Haoyu Zhen, Qiao Sun, Hongxin Zhang, Junyan Li, Siyuan Zhou, Yilun Du, and Chuang Gan. Tesseract: Learning 4d embodied world models. arXiv preprint arXiv:2504.20995, 2025. +[97] Wojciech Zielonka, Timur Bagautdinov, Shunsuke Saito, Michael Zollhöfer, Justus Thies, and Javier Romero. Drivable 3d gaussian avatars. In International Conference on 3D Vision, 2025. +[98] Chen Ziwen, Hao Tan, Kai Zhang, Sai Bi, Fujun Luan, Yicong Hong, Li Fuxin, and Zexiang Xu. Long-lrm: Long-sequence large reconstruction model for wide-coverage gaussian splats. arXiv preprint arXiv:2410.12781, 2024. + +# NeurIPS Paper Checklist + +# 1. Claims + +Question: Do the main claims made in the abstract and introduction accurately reflect the paper's contributions and scope? + +Answer: [Yes] + +Justification: The abstract and introduction clearly state the central goal of scaling 4D pretraining to enable generalizable space-time reconstruction. The proposed 4D-LRM model directly addresses this goal, demonstrating high-quality view-time rendering, generalization across scenes and time, and fast reconstruction speed. The claims made are well-supported by the model design, experiments, and results throughout the paper. + +Guidelines: + +- The answer NA means that the abstract and introduction do not include the claims made in the paper. +- The abstract and/or introduction should clearly state the claims made, including the contributions made in the paper and important assumptions and limitations. A No or NA answer to this question will not be perceived well by the reviewers. +- The claims made should match theoretical and experimental results, and reflect how much the results can be expected to generalize to other settings. +- It is fine to include aspirational goals as motivation as long as it is clear that these goals are not attained by the paper. + +# 2. Limitations + +Question: Does the paper discuss the limitations of the work performed by the authors? + +Answer: [Yes] + +Justification: Please refer to Appendix D. + +Guidelines: + +- The answer NA means that the paper has no limitation while the answer No means that the paper has limitations, but those are not discussed in the paper. +- The authors are encouraged to create a separate "Limitations" section in their paper. +- The paper should point out any strong assumptions and how robust the results are to violations of these assumptions (e.g., independence assumptions, noiseless settings, model well-specification, asymptotic approximations only holding locally). The authors should reflect on how these assumptions might be violated in practice and what the implications would be. +- The authors should reflect on the scope of the claims made, e.g., if the approach was only tested on a few datasets or with a few runs. In general, empirical results often depend on implicit assumptions, which should be articulated. +- The authors should reflect on the factors that influence the performance of the approach. For example, a facial recognition algorithm may perform poorly when image resolution is low or images are taken in low lighting. Or a speech-to-text system might not be used reliably to provide closed captions for online lectures because it fails to handle technical jargon. +- The authors should discuss the computational efficiency of the proposed algorithms and how they scale with dataset size. +- If applicable, the authors should discuss possible limitations of their approach to address problems of privacy and fairness. +- While the authors might fear that complete honesty about limitations might be used by reviewers as grounds for rejection, a worse outcome might be that reviewers discover limitations that aren't acknowledged in the paper. The authors should use their best judgment and recognize that individual actions in favor of transparency play an important role in developing norms that preserve the integrity of the community. Reviewers will be specifically instructed to not penalize honesty concerning limitations. + +# 3. Theory assumptions and proofs + +Question: For each theoretical result, does the paper provide the full set of assumptions and a complete (and correct) proof? + +Answer: [NA] + +Justification: This paper provides empirical study of 4D representation. There is no proof for theoretical results. + +Guidelines: + +- The answer NA means that the paper does not include theoretical results. +- All the theorems, formulas, and proofs in the paper should be numbered and cross-referenced. +- All assumptions should be clearly stated or referenced in the statement of any theorems. +- The proofs can either appear in the main paper or the supplemental material, but if they appear in the supplemental material, the authors are encouraged to provide a short proof sketch to provide intuition. +- Inversely, any informal proof provided in the core of the paper should be complemented by formal proofs provided in appendix or supplemental material. +- Theorems and Lemmas that the proof relies upon should be properly referenced. + +# 4. Experimental result reproducibility + +Question: Does the paper fully disclose all the information needed to reproduce the main experimental results of the paper to the extent that it affects the main claims and/or conclusions of the paper (regardless of whether the code and data are provided or not)? + +Answer: [Yes] + +Justification: See Appendix B foe details. + +Guidelines: + +- The answer NA means that the paper does not include experiments. + +- If the paper includes experiments, a No answer to this question will not be perceived well by the reviewers: Making the paper reproducible is important, regardless of whether the code and data are provided or not. + +- If the contribution is a dataset and/or model, the authors should describe the steps taken to make their results reproducible or verifiable. + +- Depending on the contribution, reproducibility can be accomplished in various ways. For example, if the contribution is a novel architecture, describing the architecture fully might suffice, or if the contribution is a specific model and empirical evaluation, it may be necessary to either make it possible for others to replicate the model with the same dataset, or provide access to the model. In general, releasing code and data is often one good way to accomplish this, but reproducibility can also be provided via detailed instructions for how to replicate the results, access to a hosted model (e.g., in the case of a large language model), releasing of a model checkpoint, or other means that are appropriate to the research performed. + +- While NeurIPS does not require releasing code, the conference does require all submissions to provide some reasonable avenue for reproducibility, which may depend on the nature of the contribution. For example + +(a) If the contribution is primarily a new algorithm, the paper should make it clear how to reproduce that algorithm. +(b) If the contribution is primarily a new model architecture, the paper should describe the architecture clearly and fully. +(c) If the contribution is a new model (e.g., a large language model), then there should either be a way to access this model for reproducing the results or a way to reproduce the model (e.g., with an open-source dataset or instructions for how to construct the dataset). +(d) We recognize that reproducibility may be tricky in some cases, in which case authors are welcome to describe the particular way they provide for reproducibility. In the case of closed-source models, it may be that access to the model is limited in some way (e.g., to registered users), but it should be possible for other researchers to have some path to reproducing or verifying the results. + +# 5. Open access to data and code + +Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material? + +Answer: [Yes] + +Justification: We will release code and data in order to support reproducibility. + +Guidelines: + +- The answer NA means that paper does not include experiments requiring code. +- Please see the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details. +- While we encourage the release of code and data, we understand that this might not be possible, so "No" is an acceptable answer. Papers cannot be rejected simply for not including code, unless this is central to the contribution (e.g., for a new open-source benchmark). +- The instructions should contain the exact command and environment needed to run to reproduce the results. See the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details. +- The authors should provide instructions on data access and preparation, including how to access the raw data, preprocessed data, intermediate data, and generated data, etc. +- The authors should provide scripts to reproduce all experimental results for the new proposed method and baselines. If only a subset of experiments are reproducible, they should state which ones are omitted from the script and why. +- At submission time, to preserve anonymity, the authors should release anonymized versions (if applicable). +- Providing as much information as possible in supplemental material (appended to the paper) is recommended, but including URLs to data and code is permitted. + +# 6. Experimental setting/details + +Question: Does the paper specify all the training and test details (e.g., data splits, hyperparameters, how they were chosen, type of optimizer, etc.) necessary to understand the results? + +Answer: [Yes] + +Justification: We discuss the training and test details in Section 3.1. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The experimental setting should be presented in the core of the paper to a level of detail that is necessary to appreciate the results and make sense of them. +- The full details can be provided either with the code, in appendix, or as supplemental material. + +# 7. Experiment statistical significance + +Question: Does the paper report error bars suitably and correctly defined or other appropriate information about the statistical significance of the experiments? + +Answer: [No] + +Justification: Majority of experiments are deterministic or based on fixed evaluation protocols (e.g., reconstruction from fixed views and times), making standard deviation or error bars less informative. Due to computational cost and the nature of the 4D task, we focus on representative results rather than repeated trials. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The authors should answer "Yes" if the results are accompanied by error bars, confidence intervals, or statistical significance tests, at least for the experiments that support the main claims of the paper. + +- The factors of variability that the error bars are capturing should be clearly stated (for example, train/test split, initialization, random drawing of some parameter, or overall run with given experimental conditions). +- The method for calculating the error bars should be explained (closed form formula call to a library function, bootstrap, etc.) +- The assumptions made should be given (e.g., Normally distributed errors). +- It should be clear whether the error bar is the standard deviation or the standard error of the mean. +- It is OK to report 1-sigma error bars, but one should state it. The authors should preferably report a 2-sigma error bar than state that they have a $96\%$ CI, if the hypothesis of Normality of errors is not verified. +- For asymmetric distributions, the authors should be careful not to show in tables or figures symmetric error bars that would yield results that are out of range (e.g. negative error rates). +- If error bars are reported in tables or plots, The authors should explain in the text how they were calculated and reference the corresponding figures or tables in the text. + +# 8. Experiments compute resources + +Question: For each experiment, does the paper provide sufficient information on the computer resources (type of compute workers, memory, time of execution) needed to reproduce the experiments? + +Answer: [Yes] + +Justification: We discuss all computing requirements in the Section 3.1. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The paper should indicate the type of compute workers CPU or GPU, internal cluster, or cloud provider, including relevant memory and storage. +- The paper should provide the amount of compute required for each of the individual experimental runs as well as estimate the total compute. +- The paper should disclose whether the full research project required more compute than the experiments reported in the paper (e.g., preliminary or failed experiments that didn't make it into the paper). + +# 9. Code of ethics + +Question: Does the research conducted in the paper conform, in every respect, with the NeurIPS Code of Ethics https://neurips.cc/public/EthicsGuidelines? + +Answer: [Yes] + +Justification: Our research involves 4D representation learning and does not involve human subjects, sensitive data, or deployment-related risks. We adhere all ethical standards regarding transparency, reproducibility in accordance with the NeurIPS Code of ethics. + +Guidelines: + +The answer NA means that the authors have not reviewed the NeurIPS Code of Ethics +- If the authors answer No, they should explain the special circumstances that require a deviation from the Code of Ethics. +- The authors should make sure to preserve anonymity (e.g., if there is a special consideration due to laws or regulations in their jurisdiction). + +# 10. Broader impacts + +Question: Does the paper discuss both potential positive societal impacts and negative societal impacts of the work performed? + +Answer: [Yes] + +Justification: See Appendix E. + +Guidelines: + +- The answer NA means that there is no societal impact of the work performed. + +- If the authors answer NA or No, they should explain why their work has no societal impact or why the paper does not address societal impact. +- Examples of negative societal impacts include potential malicious or unintended uses (e.g., disinformation, generating fake profiles, surveillance), fairness considerations (e.g., deployment of technologies that could make decisions that unfairly impact specific groups), privacy considerations, and security considerations. +- The conference expects that many papers will be foundational research and not tied to particular applications, let alone deployments. However, if there is a direct path to any negative applications, the authors should point it out. For example, it is legitimate to point out that an improvement in the quality of generative models could be used to generate deepfakes for disinformation. On the other hand, it is not needed to point out that a generic algorithm for optimizing neural networks could enable people to train models that generate Deepfakes faster. +- The authors should consider possible harms that could arise when the technology is being used as intended and functioning correctly, harms that could arise when the technology is being used as intended but gives incorrect results, and harms following from (intentional or unintentional) misuse of the technology. +- If there are negative societal impacts, the authors could also discuss possible mitigation strategies (e.g., gated release of models, providing defenses in addition to attacks, mechanisms for monitoring misuse, mechanisms to monitor how a system learns from feedback over time, improving the efficiency and accessibility of ML). + +# 11. Safeguards + +Question: Does the paper describe safeguards that have been put in place for responsible release of data or models that have a high risk for misuse (e.g., pretrained language models, image generators, or scraped datasets)? + +Answer: [NA] + +Justification: Our work does not involve the release of models or datasets with known high risk for misuse. Our research focuses on 4D representation learning which does not include any language models and generative models. + +# Guidelines: + +- The answer NA means that the paper poses no such risks. +- Released models that have a high risk for misuse or dual-use should be released with necessary safeguards to allow for controlled use of the model, for example by requiring that users adhere to usage guidelines or restrictions to access the model or implementing safety filters. +- Datasets that have been scraped from the Internet could pose safety risks. The authors should describe how they avoided releasing unsafe images. +- We recognize that providing effective safeguards is challenging, and many papers do not require this, but we encourage authors to take this into account and make a best faith effort. + +# 12. Licenses for existing assets + +Question: Are the creators or original owners of assets (e.g., code, data, models), used in the paper, properly credited and are the license and terms of use explicitly mentioned and properly respected? + +Answer: [Yes] + +Justification: We acknowledge all external code and models used in this paper by properly citing their original sources, and we provide a detailed discussion of how each asset is used in the appendix. + +# Guidelines: + +- The answer NA means that the paper does not use existing assets. +- The authors should cite the original paper that produced the code package or dataset. +- The authors should state which version of the asset is used and, if possible, include a URI. + +- The name of the license (e.g., CC-BY 4.0) should be included for each asset. +- For scraped data from a particular source (e.g., website), the copyright and terms of service of that source should be provided. +- If assets are released, the license, copyright information, and terms of use in the package should be provided. For popular datasets, paperswithcode.com/datasets has curated licenses for some datasets. Their licensing guide can help determine the license of a dataset. +- For existing datasets that are re-packaged, both the original license and the license of the derived asset (if it has changed) should be provided. +- If this information is not available online, the authors are encouraged to reach out to the asset's creators. + +# 13. New assets + +Question: Are new assets introduced in the paper well documented and is the documentation provided alongside the assets? + +Answer: [Yes] + +Justification: We clearly discuss and document all new assets. We only used open-sourced data in order to train our model and we will release our code. + +Guidelines: + +- The answer NA means that the paper does not release new assets. +- Researchers should communicate the details of the dataset/code/model as part of their submissions via structured templates. This includes details about training, license, limitations, etc. +- The paper should discuss whether and how consent was obtained from people whose asset is used. +- At submission time, remember to anonymize your assets (if applicable). You can either create an anonymized URL or include an anonymized zip file. + +# 14. Crowdsourcing and research with human subjects + +Question: For crowdsourcing experiments and research with human subjects, does the paper include the full text of instructions given to participants and screenshots, if applicable, as well as details about compensation (if any)? + +Answer: [NA] + +Justification: Our work does not involve human subjects or any form of crowdsourcing. All experiments are conducted using publicly available datasets without human participation. + +Guidelines: + +- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. +- Including this information in the supplemental material is fine, but if the main contribution of the paper involves human subjects, then as much detail as possible should be included in the main paper. +- According to the NeurIPS Code of Ethics, workers involved in data collection, curation, or other labor should be paid at least the minimum wage in the country of the data collector. + +# 15. Institutional review board (IRB) approvals or equivalent for research with human subjects + +Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals (or an equivalent approval/review based on the requirements of your country or institution) were obtained? + +Answer: [NA] + +Justification: This work does not involve human subjects or study participants. Therefore, no ethical risks or IRB approvals are applicable. + +Guidelines: + +- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. +- Depending on the country in which research is conducted, IRB approval (or equivalent) may be required for any human subjects research. If you obtained IRB approval, you should clearly state this in the paper. +- We recognize that the procedures for this may vary significantly between institutions and locations, and we expect authors to adhere to the NeurIPS Code of Ethics and the guidelines for their institution. +- For initial submissions, do not include any information that would break anonymity (if applicable), such as the institution conducting the review. + +# 16. Declaration of LLM usage + +Question: Does the paper describe the usage of LLMs if it is an important, original, or non-standard component of the core methods in this research? Note that if the LLM is used only for writing, editing, or formatting purposes and does not impact the core methodology, scientific rigorousness, or originality of the research, declaration is not required. + +Answer: [NA] + +Justification: LLMs were not used as part of the core methodology, experimental design, or analysis in this research. Any LLM usage was limited to minor writing or formatting support and did not affect the scientific contributions of the paper. + +Guidelines: + +- The answer NA means that the core method development in this research does not involve LLMs as any important, original, or non-standard components. +- Please refer to our LLM policy (https://neurips.cc/Conferences/2025/LLM) for what should or should not be described. + +# A Related Work (Expanded) + +3D Representations. At the heart of 3D reconstruction lies the choice of scene representation. Classical multi-view geometry methods [49, 57, 21, 1, 54] rely on calibrated images and epipolar constraints to recover structure via structure-from-motion (SfM; [49]) and multi-view stereo (MVS; [50, 21, 55]). While robust under densely sampled views, these pipelines degrade in the presence of occlusion, textureless surfaces, or wide baselines. To improve generalization, recent geometry-based models [71, 35] extend reconstruction to unposed and sparse view settings by leveraging large-scale pretraining. In parallel, neural optimization approaches fit implicit representations per scene, such as neural radiance fields (NeRF; [44, 4, 8, 15]) or signed distance functions [46, 42]. In this work, we adopt Gaussian Splitting (GS; [31, 25, 87]) as our 3D representation, which prioritizes efficiency and real-time performance through explicit point-based representations and offers faster rendering and competitive quality with minimal optimization per scene. + +4D Representations. Modeling dynamic scenes in 4D has evolved from implicit, per-scene neural fields to more efficient, explicit spatio-temporal representations. Early methods [51, 47, 63] extend NeRF to dynamic scenes by conditioning on time or deformation fields, but they remain computationally heavy and lack real-time inference. A more practical shift arrives with HexPlane [3] and K-Planes [20], which decompose the 4D volume into planar factorizations. Concurrent with these NeRF-based approaches, dynamic Gaussian Splating emerges as an explicit, high-performance framework for dynamic scenes [18, 33, 40]. D3GA [97] adapts this paradigm to human avatars by embedding Gaussian splats into deformable tetrahedral cages [84], allowing anatomical consistency and real-time control via joint-driven pose signals. Notably, 4D-GS [75] represents dynamic scenes using a canonical set of 3D Gaussians combined with 4D voxel-based features. These voxel encodings, inspired by the spatio-temporal factorization of HexPlane [3], are decoded by a lightweight MLP to predict per-Gaussian deformations across time. In contrast, we adapted 4DGS [82], which introduces a more elegant and unified formulation by directly parameterizing 4D Gaussians over space and time, allowing anisotropic deformation and smooth, time-varying appearance via 4D spherical harmonics. + +Feed-Forward Reconstruction Models. Recent advances in generalizable radiance field-based methods have achieved state-of-the-art quality in novel view synthesis by leveraging NeRF-style volume rendering [86]. These approaches typically employ 3D-to-2D geometric projections to sample per-view image features, using architectural priors such as epipolar feature sampling [68, 59, 58] or plane-swept cost volumes [7, 30, 41] inspired by MVS. In contrast, we explore a simpler and more flexible design: a large Transformer-based model without explicit 3D inductive biases, which directly regresses Gaussian primitives. Parallel to radiance field approaches, a separate line of research investigates feed-forward, geometry-centric reconstruction models [61, 80], building upon DUSt3R [71] and leveraging large-scale training. This work aligns more closely with large reconstruction models (LRMs), which have recently emerged as a unified framework for producing view-consistent 3D reconstructions. These models are trained on massive 3D datasets and use triplane-based NeRFs [37, 24, 79, 67, 26] or 3D Gaussian Splatting [92, 60, 77, 98] to encode strong priors over shape and appearance, enabling high-quality reconstruction from just a few posed views. While early efforts have begun extending LRMs to generate 4D assets [53], we present the first LRM for general 4D reconstruction that can handle sparse multi-views and missing timestamps. + +# B 4D-LRM Implementation and Training Details + +# B.1 4DGS Parameterizing and Initialization + +Given the decoded 4D Gaussian parameter $\mathbf{g} \in \mathbb{R}^{20}$ , we split it into $(\mathbf{g}_{\mathrm{xyz}} \in \mathbb{R}^3, \mathbf{g}_{\mathrm{t}} \in \mathbb{R}, \mathbf{g}_{\mathrm{rgb}} \in \mathbb{R}^3, \mathbf{g}_{\mathrm{scale,xyz}} \in \mathbb{R}^3, \mathbf{g}_{\mathrm{scale,t}} \in \mathbb{R}, \mathbf{g}_{\mathrm{rotation,left}} \in \mathbb{R}^4, \mathbf{g}_{\mathrm{rotation,right}} \in \mathbb{R}^4, \mathbf{g}_{\mathrm{opacity}} \in \mathbb{R})$ . + +Space and Distance. Given the ray origin $\mathrm{ray}_o$ , direction $\mathrm{ray}_d$ , and a distance scalar $\delta$ , the pixel-aligned Gaussian center in space along the ray is computed as $\mu_{\mathrm{xyz}} = \mathrm{ray}_o + \delta \cdot \mathrm{ray}_d$ . To derive $\delta$ from the decoded 4D Gaussian primitive $(\mathbf{g}_{\mathrm{x}},\mathbf{g}_{\mathrm{y}},\mathbf{g}_{\mathrm{z}})$ , we adopt an interpolation scheme bounded by two empirically determined depth limits, $\delta_{\mathrm{near}}$ and $\delta_{\mathrm{far}}$ , which define the permissible depth interval along each ray. This range constrains the model's predictions to lie within a spatially valid and semantically meaningful region of 3D space, consistent with the geometry captured during training. + +$$ +\omega = \operatorname {s i g m o i d} \left[ \left(\mathbf {g} _ {\mathrm {x}} + \mathbf {g} _ {\mathrm {y}} + \mathbf {g} _ {\mathrm {z}}\right) / 3 \right], \tag {4} +$$ + +$$ +\delta = (1 - \omega) \delta_ {\text {n e a r}} + \omega \delta_ {\text {f a r}}. \tag {5} +$$ + +Following the setup in [92], we set $\delta_{\mathrm{near}} = 0.1$ and $\delta_{\mathrm{far}} = 4.5$ . The predicted $\mu_{\mathrm{xyz}}$ values are further clipped to the range $[-1,1]^3$ . + +Scale and Opacity. We follow the default activation functions used in 3DGS to ensure the predicted scale and opacity values fall within valid ranges: all scales are mapped to $\mathbb{R}^+$ and opacity to $(0,1)$ . Specifically, we apply the exponential activation to scale, mapping real-valued inputs to positive values, and the sigmoid activation to opacity. We observe similar training dynamics reported in [92], that the learned scale of 3D Gaussians can become excessively large. In such cases, the Gaussian degenerates into a highly anisotropic distribution, resembling a thin stick or line stretched across space and time. This can lead to unstable training dynamics, slow convergence, and temporal ghosting artifacts. To mitigate this, we apply constant biases to the Transformer's output to shift the initialization, and we clip the predicted scales to remain within a reasonable range. + +$$ +\operatorname {s c a l e} _ {x y z} = \min \left\{\exp \left(\mathbf {g} _ {\text {s c a l e , x y z}} - 2. 3\right), 0. 3 \right\}, \tag {6} +$$ + +$$ +\operatorname {s c a l e} _ {t} = \min \left\{\exp \left(\mathbf {g} _ {\text {s c a l e}, t} - 2. 3\right), 1. 0 \right\}, \tag {7} +$$ + +$$ +\text {o p a c i t y} = \sigma \left(\mathbf {g} _ {\text {o p a c i t y}} - 2. 0\right), \tag {8} +$$ + +All hyperparameters are empirically chosen and primarily serve to stabilize training. We observe that model performance is relatively insensitive to their specific values. + +Rotation. We predict unnormalized quaternions and apply L2 normalization to ensure they lie on the unit hypersphere, producing valid unit quaternions for rotation. This approach simplifies optimization, as the model can freely output real-valued 4D vectors while normalization guarantees valid rotations. Following [82], we use a pair of $q_{l} = (a,b,c,d)$ and $q_{r} = (p,q,r,s)$ for the left and right unit quaternions, respectively. We use them to represent isotropic rotations in a symmetric form. $R$ can be constructed by: + +$$ +R = L \left(q _ {l}\right) R \left(q _ {r}\right) = \left( \begin{array}{c c c c} a & - b & - c & - d \\ b & a & - d & c \\ c & d & a & - b \\ d & - c & b & a \end{array} \right) \left( \begin{array}{c c c c} p & - q & - r & - s \\ q & p & s & - r \\ r & - s & p & q \\ s & r & - q & p \end{array} \right). \tag {9} +$$ + +Spherical Harmonics / RGB. A 3D Gaussian stores a set of spherical harmonics (SH) coefficients to represent view-dependent color, along with a scalar opacity value $\alpha$ . 4DGS extends this representation to enable both view-dependent appearance and its temporal evolution, by incorporating a time-variant extension of the SH basis. This allows the appearance of each Gaussian to change smoothly over both viewpoint and time. In our implementation, we directly interpret the model's output as the zero-order SH coefficients, following the convention used in 4DGS [82]. For simplicity, we do not include higher-order SH terms in this work. + +# B.2 Differentiable Rasterization and Deferred Rendering + +In the rendering process, given a pixel $(u,v)$ in an image $\mathbf{I}$ at time $t$ , along with the camera's extrinsic matrix $E$ and intrinsic matrix $K$ , the pixel color $\mathbf{I}(u,v,t)$ is computed by blending the contributions of all visible conditional 3D Gaussians. We build on the tile-based rasterization pipeline introduced in 3DGS and 4DGS [31, 82], and adopt deferred backpropagation [91] during rendering to reduce GPU memory consumption. We describe more details for completeness. + +Filtering. At inference time, the final opacity of each conditional 3D Gaussian is weighted by its temporal marginal $p(t)$ , which reflects its relevance at the rendered time step. To improve rendering efficiency and visual clarity, we apply two filtering strategies: (1) Gaussians with marginal probability $p(t) < 0.05$ are removed, and (2) Gaussians whose weighted opacity $\alpha < 0.05$ are removed. These filters are applied only during inference. Applying them during training would prematurely eliminate potentially useful Gaussians, leading to degraded convergence or dead Gaussians that fail to get optimized in training. + +Rasterization. These filtered Gaussians are projected onto the image plane and sorted in front-to-back order based on their depth. For each pixel, the final color is computed using alpha blending, where the contribution of each Gaussian is weighted by its projected 2D density $p_i(u,v,t)$ , its opacity $\alpha_{i}$ and its view-dependent color $c_{i}(d_{i},t)$ . Additionally, a transmittance term $\prod_{j = 1}^{i - 1}[1 - p_j(u,v,t)\alpha_j]$ + +Algorithm 1 Image-to-4DGS Pseudo Code. +```ini +# Input dimensions: +# b = batch size; v = number of views; h, w = image height and width +# Input tensors: +# images : [b, v, h, w, 3] # RGB image frames +# frame_time : [b, v, 1] # Frame timestamp per view +# extrinsics : [b, v, 4, 4] # Camera-to-world (c2w) transformation matrices +# intrinsics : [b, v, 4] # Camera intrinsics (fx, fy, cx, cy) +# Output tensors (4DGS parameters): +# xyzt : [b, *, 4] # 4D Gaussian centers (x, y, z, t) +# rgb : [b, *, 3] # RGB color +# scale : [b, *, 4] # Anisotropic Gaussian scale (3D space + time) +# rotation : [b, *, 4, 4] # Rotation matrix +# opacity : [b, *, 1] # Opacity value +# Output tensors (4DGS parameters): +# xyzt : [b, *, 4] # 4D Gaussian centers (x, y, z, t) +# rgb : [b, *, 3] # RGB color +# scale : [b, *, 4] # Anisotropic Gaussian scale (3D space + time) +# rotation : [b, *, 4, 4] # Rotation matrix +# opacity : [b, * , 1] # Opacity value +# Augment and patchify input for 4D representation +x_grid, y_grid = meshgrid(h, w) # [h, w] +ray_dirCAM = compute_camera Rays(x_grid, y_grid, intrinsics) # [b, v, 3, h, w] +ray_dir_world = transform Directions (ray_dirCAM, extrinsics) # [b, v, 3, h, w] +ray_origin = extract CAMERA_origin(extrinsics) # [b, v, 3, h, w] +o jot_d = dot_product(-ray.origin, ray_dir_world, dim=2) # [b, v, 1, h, w] +nearest_pts = ray.origin + o jot_d * ray_dir_world # [b, v, 3, h, w] +# Concatenate and patchify augmented images into transformer input +x = concatenate( +normalize_rgb/images), # [b, v, 3, h, w], RGB scaled to [-1, 1] +normalize_t(frame_time), # [b, v, 1, h, w], time scaled to [-1, 1] +ray_dir_world, # [b, v, 3, h, w] +nearest_ptss # [b, v, 3, h, w] +) # Final: [b, v, 10, h, w] +x = patchify(x, patch_size=8) # [b * v, numpatches, patch_dim] +# Transformer +x = linear(x) # [b, v * numpatches, hidden_dim] +x = transformer(LN(x)) # LayerNorm + Transformer +x = depatchify(LN(x), out_dim=20) # [b, v * h * w, 20] +# 4DGS Parameterization +# Step 1: Split the transformer output into individual 4DGS fields +xyzt, t, rgb, scale_xyz, scale_t, rotation_left, rotation_right, opacity = \ split(x, sizes=[3, 1, 3, 3, 1, 4, 4, 1], dim=-1) +# Step 2: Compute center position (xyz + t) +w = sigmoid(norm(xyz)) # Soft depth interpolation weight +delta = near * (1 - w) + far * w # Range interpolation [near, far] +xyz = ray_origin + ray_dir_world * delta # 3D center point +xyzt = concatenate(xyz, t) # [b, v * h * w, 4] +# Step 3: Compute scale (clipped exp) +scale_xyz = clip(exp(scale_xyz - 2.3), max=0.3) # Spatial scale +scale_t = clip(exp(scale_t - 2.3), max=1.0) # Temporal scale +scale = concatenate(scale_xyz, scale_t) # [b, v * h * w, 4] +# Step 4:Normalize quaternions +q_1 = normalize(rotation_left) # [b, v * h * w, 4] +q_r = normalize(rotation_right) # [b, v * h * w, 4] +# Step 5:Construct rotation matrix R = L(q_1) * R(q_r) +L = build_left_quaternion_matrix(q_1) # [b, v * h * w, 4, 4] +R = build_right_quaternion_matrix(q_r) # [b, v * h * w, 4, 4] +rotation = matmul(L,R) # [b,v*h*w，4，4] +# Step 6:Compute opacity +opacity = sigmoid(opacity - 2.0) # [b,v*h*w，1] +# Final 4DGS outputs +return xyzt, rgb, scale, rotation, opacity +``` + +models the amount of light that reaches the $i$ -th Gaussian after being attenuated by all previous ones. This formulation enables differentiable, order-dependent compositing. Yang et al. [82] noted that $p_i(u,v,t)$ can be factorized as the product of a conditional and a marginal probability at time $t$ : + +$$ +\mathcal {I} (u, v, t) = \sum_ {i = 1} ^ {N} p _ {i} (t) p _ {i} (u, v | t) \alpha_ {i} c _ {i} (d, t) \prod_ {j = 1} ^ {i - 1} [ 1 - p _ {j} (t) p _ {j} (u, v | t) \alpha_ {j} ]. \tag {10} +$$ + +To compute the image-space density $p_i(u,v|t)$ , we start with the 4-channel space-time features in the order of xyzt, and compute the conditional 3DGS: + +$$ +\mu_ {i, x y z \mid t} = \mu_ {i, 1: 3} + \Sigma_ {i, 1: 3, 4} \Sigma_ {i, 4, 4} ^ {- 1} (t - \mu_ {i, 4}), \tag {11} +$$ + +$$ +\Sigma_ {i, x y z | t} = \Sigma_ {i, 1: 3, 1: 3} - \Sigma_ {i, 1: 3, 4} \Sigma_ {i, 4, 4} ^ {- 1} \Sigma_ {i, 4, 1: 3} +$$ + +We approximate the projection of a 3D Gaussian $\mathcal{N}(\mu_i,\Sigma_i)$ using a linearized perspective transformation as is in [31]. The resulting 2D Gaussian is + +$$ +p _ {i} (u, v) \sim \mathcal {N} \left(\mu_ {\mathrm {i}, \mathrm {u v}}, \Sigma_ {\mathrm {i}, \mathrm {u v}}\right), \tag {12} +$$ + +where the mean and covariance are computed as $\mu_{\mathrm{i,uv}} = \mathrm{Proj}(\mu_{i,xyz|t},E,K)_{1:2}$ and $\Sigma_{i,uv} = (JE\Sigma_{i,xyz|t}E^{\top}J^{\top})_{1:2,1:2}$ , with $\mathrm{Proj}(\cdot ,\cdot ,\cdot)$ denoting projection from world to image coordinates using extrinsic $E$ and intrinsic $K$ , and $J$ the Jacobian of the perspective projection at $\mu_{i,xyz|t}$ . + +# B.3 Dataset Curation + +To enable large-scale training, we construct a 4D dataset derived from Objaverse [13], which provides a subset of animated 3D assets. However, the raw dataset is not directly suitable for 4D modeling: object motions are often inconsistent, and the dataset contains duplicates and artifacts. We build upon the filtered subset curated by Diffusion4D [38], which removes static objects and unstable motion sequences. For each object, we render a 24-frame video from diverse camera trajectories. If the original animation has fewer frames, we pad by repeating the last frame. The views include four canonical static views (front, back, left, right), elevated moving trajectories, and randomized orbits at varying distances, similar to [92]. This design encourages robustness to both viewpoint and motion variation. To further ensure quality, we compute the maximum L1 distance across frames for each sequence to select a high-quality subset with sufficient but not overly aggressive motion. Our final dataset contains 3,000 high-quality animated objects (HQ4D) selected from 32,000 animated objects (4D). We augment the dataset with 783,000 static 3D objects from Objaverse by treating each as a 24-frame video, applying minor frame-by-frame displacements along a single random direction. During pretraining, we sample from HQ4D, 4D, and 3D with a mixing ratio of 200:50:1. + +# B.4 Training Details + +We keep most settings identical to GS-LRM [92] so we can initialize 4D-LRM training upon it. Below, we describe the details for completeness. We use a patch size of $8 \times 8$ for the image tokenizer. 4D-LRM-Large employs a 24-layer Transformer with a hidden dimension of 1024, 16 attention heads, and a two-layer MLP with GeLU activation. 4D-LRM-Base uses 12 layers with a hidden dimension of 768 and 12 attention heads, sharing the same MLP design. All Transformer blocks are equipped with Pre-Layer Normalization and residual connections. Additionally, Layer Normalization is applied after the patchifying linear layer and before the unpatchifying linear layer to stabilize training. To enable efficient training and inference, we adopt Flash-Attention v2 [12] via the xFormers library [34], along with gradient checkpointing [10] and mixed-precision training using the BF16 data type [43]. + +# C Additional Results + +# C.1 Evaluation on 3D Reconstruction + +Table 5 reports performance on the GSO dataset [17], comparing various models under two resolutions. Notably, when adapting 4D-LRM for static 3D reconstruction by setting all timestamps to zero, we observe a modest drop in performance relative to GS-LRM at $256 \times 256$ resolution. Despite this, 4D-LRM still outperforms the LGM and many models evaluated at higher resolution. This suggests that the spatiotemporal representations learned by 4D-LRM remain effective for conventional 3D tasks, highlighting its versatility and robustness. + +# C.2 Failure Cases + +![](images/65d97042d30da78520846963b4ecea7bdcf785f0db4ea811e57c70e579f4c700.jpg) +Figure 11: A typical failure case as 4D-LRM sometimes struggles with non-linear motion trajectories. When an object follows a non-linear path, the structures cannot be efficiently captured by a single ellipsoidal Gaussian. As a result, the model requires multiple Gaussians placed along the trajectory to approximate the motion, increasing complexity and often leading to artifacts if not properly aligned. + +We provide a typical failure case in Figure 11. 4D-LRM still struggles with challenging 4D reconstruction scenarios involving self-occlusion and fast motion, which often result in temporal ghosting artifacts. These failures are primarily due to the limitations of both (1) under-training (the current model has not yet reached the saturation for scaling) as well as (2) the current Gaussian representation, which models appearance and motion as smooth, continuous functions. In the presence of rapid or discontinuous changes, such as objects moving in and out of occlusion or undergoing abrupt non-rigid deformation, the model + +cannot accurately localize or update the corresponding Gaussians in time. As a result, outdated Gaussians persist across frames, leading to visually noticeable residuals and motion trails. 4D-LRM sometimes falls short in non-linear trajectories. The kernel density of a Gaussian is ellipsoidal, it defines mass aligned with principal directions. When an object follows a non-linear path, the optimal support is curved or branched, not ellipsoidal. To approximate a non-linear trajectory, the model needs more Gaussians along the curve. + +Table 5: Performance on the GSO dataset [17] for 3D reconstruction. 4D-LRM is evaluated by collapsing frame times to 0. + +
Res.ModelPSNRLPIPSSSIM
512SparseNeus [41]20.620.1990.836
Triplane-LRM [37]26.540.0640.893
Mesh-LRM [74]27.930.0810.925
GS-LRM [92]30.520.0500.952
256LGM [60]21.440.1220.832
GS-LRM [92]29.590.0510.944
4D-LRM27.350.0610.929
+ +# C.3 Additional Qualitative Examples + +We provide additional qualitative examples in Figure 12 and 13. + +# D Limitations + +We highlight the following future directions: + +Long Context. Issues such as limited resolution, short video duration, and occlusion are fundamentally challenges of memory and long-range dependencies in sequence modeling. Although 4D-LRM achieves high-quality reconstruction from sparse posed images, several limitations remain. First, it cannot yet efficiently process hundreds of input images in a single forward pass. Second, its maximum training resolution is $256 \times 256$ , though it generalizes up to $512 \times 512$ . Unlike GS-LRM, which was fine-tuned at 512 resolution, fine-tuning 4D-LRM at this scale is significantly more expensive, requiring approximately 75 seconds per training step. A promising direction for future work is to develop high-resolution 4D reconstruction models capable of handling hundreds of 1K or 2K resolution inputs. This will require fundamental architectural advances, such as hybrid models for long-context handling [98] and test-time training strategies [11, 94]. + +Removing 3D Inductive Bias. Currently, 4D-LRM relies on posed images and explicitly learns 4D Gaussian primitives for rendering. To scale up 4D representations from in-the-wild videos, future work should aim to remove strong 3D inductive biases. This includes learning to reconstruct from unposed images [67, 27], and designing architectures that forgo explicit 3D representations such as NeRF or 3DGS [29, 27, 94]. + +From Objects to Scenes. Currently, 4D-LRM is not trained at the scene level, as the concept of "any view" is less well-defined, e.g., we cannot observe what lies behind walls. Although GS-LRM has shown that this architecture can scale to scene-level reconstruction, we currently lack + +![](images/3365bfb7ceaa57a5ab805a553cd0bef007452ad0d05fee3fb826b6a9f8f27e2c.jpg) +Figure 12: Qualitative examples of 4D-LRM under varying camera setups. We show the performance of 4D-LRM when taking input views captured with different camera configurations, demonstrating its robustness to diverse spatial arrangements and viewpoints. + +a license-compliant, high-quality 4D scene dataset for training. Moreover, the data augmentation strategies used for object-level data do not directly transfer to scene-level setups. While we start to see attempts on this line with limited camera movement or domain-specific applications and limited input/target camera dynamics [81, 39], more future work should investigate both scalable 4D datasets and training methods for extending 4D-LRM to scene-level reconstruction. + +# E Broader Impact + +While this paper does not explicitly address societal impact, the proposed 4D representation learning method has the potential to benefit a range of downstream applications, including robotics, AR/VR, and digital content creation, by enabling more accurate and efficient modeling of dynamic scenes. However, the work primarily focuses on technical contributions, and we do not identify immediate ethical concerns or direct social implications. As with any foundation model capable of detailed spatial-temporal understanding, future applications should consider issues of privacy, surveillance, and potential misuse, especially if deployed in real-world environments involving human data. + +# F Acknowledgment. + +The authors would like to thank Ang Cao, Junyi Zhang, and Wenhao Chai for their helpful discussions and feedback. + +![](images/c088922e32b90c17f6a014fb108b05c1d0d8de5d6bfe25e7834d988884235a42.jpg) +Figure 13: Additional frame interpolation examples. We insert $4 \times$ denser frames between Alternating Canonical Views as input. \ No newline at end of file diff --git a/NeurIPS/2025/4D-LRM_ Large Space-Time Reconstruction Model From and To Any View at Any Time/images.zip b/NeurIPS/2025/4D-LRM_ Large Space-Time Reconstruction Model From and To Any View at Any Time/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..d5c37288c9402212f89241810afb9ba9720afd13 --- /dev/null +++ b/NeurIPS/2025/4D-LRM_ Large Space-Time Reconstruction Model From and To Any View at Any Time/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3fa982f967ec300b78bd959cba4b253c88197fed28506fa33bd4198fae3966fd +size 1428916 diff --git a/NeurIPS/2025/4D-LRM_ Large Space-Time Reconstruction Model From and To Any View at Any Time/layout.json b/NeurIPS/2025/4D-LRM_ Large Space-Time Reconstruction Model From and To Any View at Any Time/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..0ea952f3a2b824fa639fef677e1d446468c2c62b --- /dev/null +++ b/NeurIPS/2025/4D-LRM_ Large Space-Time Reconstruction Model From and To Any View at Any Time/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6171cd2001a9b812098ecd4d988614d11eeafb913803ab70a7104f9334f44a33 +size 892513 diff --git a/NeurIPS/2025/4D-VLA_ Spatiotemporal Vision-Language-Action Pretraining with Cross-Scene Calibration/1f403ad0-c85b-4b10-8b9e-091c9e64bd05_content_list.json b/NeurIPS/2025/4D-VLA_ Spatiotemporal Vision-Language-Action Pretraining with Cross-Scene Calibration/1f403ad0-c85b-4b10-8b9e-091c9e64bd05_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..eebcc60641bcfe8a3868e1709eed1c9fe22d88ec --- /dev/null +++ b/NeurIPS/2025/4D-VLA_ Spatiotemporal Vision-Language-Action Pretraining with Cross-Scene Calibration/1f403ad0-c85b-4b10-8b9e-091c9e64bd05_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a77b1a7133851bb669607968ccdef05ddf6c22f9a2a7e0b94734e0d26252c37d +size 149148 diff --git a/NeurIPS/2025/4D-VLA_ Spatiotemporal Vision-Language-Action Pretraining with Cross-Scene Calibration/1f403ad0-c85b-4b10-8b9e-091c9e64bd05_model.json b/NeurIPS/2025/4D-VLA_ Spatiotemporal Vision-Language-Action Pretraining with Cross-Scene Calibration/1f403ad0-c85b-4b10-8b9e-091c9e64bd05_model.json new file mode 100644 index 0000000000000000000000000000000000000000..8fc8392d90848b7220efcfbb810b456790e35b80 --- /dev/null +++ b/NeurIPS/2025/4D-VLA_ Spatiotemporal Vision-Language-Action Pretraining with Cross-Scene Calibration/1f403ad0-c85b-4b10-8b9e-091c9e64bd05_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bdb4af9f79ff2b054ccb7c9587f017f3b5f9d78b0136bf6bd555c7cdbedccfa8 +size 189204 diff --git a/NeurIPS/2025/4D-VLA_ Spatiotemporal Vision-Language-Action Pretraining with Cross-Scene Calibration/1f403ad0-c85b-4b10-8b9e-091c9e64bd05_origin.pdf b/NeurIPS/2025/4D-VLA_ Spatiotemporal Vision-Language-Action Pretraining with Cross-Scene Calibration/1f403ad0-c85b-4b10-8b9e-091c9e64bd05_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..c701cff598c2fd44d7c60c3d109aa77ce7bd6eb4 --- /dev/null +++ b/NeurIPS/2025/4D-VLA_ Spatiotemporal Vision-Language-Action Pretraining with Cross-Scene Calibration/1f403ad0-c85b-4b10-8b9e-091c9e64bd05_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cb1156e97aba7955f5873efddfd7c79a2d787c6277de2482d35efeec18b0a0b5 +size 2328032 diff --git a/NeurIPS/2025/4D-VLA_ Spatiotemporal Vision-Language-Action Pretraining with Cross-Scene Calibration/full.md b/NeurIPS/2025/4D-VLA_ Spatiotemporal Vision-Language-Action Pretraining with Cross-Scene Calibration/full.md new file mode 100644 index 0000000000000000000000000000000000000000..9ad564dff381fa2c74e550fcf483e88ef0fac531 --- /dev/null +++ b/NeurIPS/2025/4D-VLA_ Spatiotemporal Vision-Language-Action Pretraining with Cross-Scene Calibration/full.md @@ -0,0 +1,695 @@ +# 4D-VLA: Spatiotemporal Vision-Language-Action Pretraining with Cross-Scene Calibration + +Jiahui Zhang $^{1*}$ Yurui Chen $^{1*}$ Yueming Xu $^{1}$ Ze Huang $^{1}$ Yanpeng Zhou $^{2}$ Yu-Jie Yuan $^{2}$ Xinyue Cai $^{2}$ Guowei Huang $^{2}$ Xingyue Quan $^{2}$ Hang Xu $^{2}$ Li Zhang $^{1\dagger}$ + +$^{1}$ School of Data Science, Fudan University $^{2}$ Huawei Noah's Ark Lab + +https://github.com/LogosRoboticsGroup/4D-VLA + +![](images/efcf2a60a49089ee47f82be52bab50102b2a8011b66e1c6399007a68596de683.jpg) + +![](images/d106fc4c3cfba0363dd78ed0fc46ffefe0481ce1e8d64211807ed5c692349e62.jpg) +Figure 1: Top: Our pretraining design philosophy highlights that prior methods often lack key cues in their input for accurate action inference. This leads to target action distributions $A_{t}(\cdot)$ exhibiting high variance or non-smoothness, which negatively impacts pretraining performance. A rough analysis shows that in the DROID dataset, $67\%$ of the samples have the robot's base occluded, causing coordinate system chaos. Bottom: We verify our method in both simulated and real-world robotic settings and report the performance for the OpenVLA baseline and our 4D-VLA approach. + +# Abstract + +Leveraging diverse robotic data for pretraining remains a critical challenge. Existing methods typically model the dataset's action distribution using simple observations as inputs. However, these inputs are often incomplete, resulting in a dispersed conditional action distribution—an issue we refer to as coordinate system chaos and state chaos. This inconsistency significantly hampers pretraining efficiency. To address this, we propose 4D-VLA, a novel approach that effectively integrates 4D information into the input to mitigate these sources of chaos. Our model introduces + +depth and temporal information into visual features with sequential RGB-D inputs, aligning the coordinate systems of the robot and the scene. This alignment endows the model with strong spatiotemporal reasoning capabilities while minimizing training overhead. Additionally, we introduce memory bank sampling, a frame sampling strategy designed to extract informative frames from historical images, further improving effectiveness and efficiency. Experimental results demonstrate that our pretraining method and architectural components substantially enhance model performance. In both simulated and real-world experiments, our model achieves a significant increase in success rate over OpenVLA [1]. To further assess spatial perception and generalization to novel views, we introduce MV-Bench, a multi-view simulation benchmark. Our model consistently outperforms existing methods, demonstrating stronger spatial understanding and adaptability. + +# 1 Introduction + +The emergence of pretrained vision-language models has established an effective framework for aligning human language with visual data, advancing embodied intelligence. Meanwhile, the opensourcing of diverse robotic datasets [2, 3, 4, 5] has enabled data-driven robot control. However, efficiently extracting useful information from these datasets remains a challenge for improving generalization across diverse scenarios. + +A real-world action distribution can be interpreted as a response function conditioned on observations or input, denoted as $A_{t}$ (input). The objective of pretraining is to learn a model $\mathrm{F}_{\theta}$ (input) that approximates this function using large-scale data $A_{t}^{data}$ (input). However, existing pretraining paradigms often suffer from incomplete or under-informative input, lacking critical contextual cues required for reliable action reasoning. Consequently, the resulting target distribution $A_{t}(\cdot)$ may exhibit undesirable characteristics—such as lack of smoothness, high variance, or multimodality—which hinder the model's ability to learn robust and generalizable behaviors. + +Previous approaches, such as OpenVLA [1], use only a single RGB image and a textual instruction as input. This limited input setting leads to two prominent types of chaos. The first is coordinate system chaos, which arises when actions are defined in the robot's coordinate frame, yet the visual input lacks sufficient spatial context. For instance, if the image does not fully capture the robot's body, it becomes challenging to infer the robot's exact position and orientation. The second is state chaos, which arises in scenarios where a single frame lacks the necessary temporal or contextual cues to resolve action ambiguity. This includes symmetric trajectories—where it is difficult to infer the direction of motion—as well as cases where visually similar observations correspond to entirely different actions. These ambiguities hinder the effectiveness of pretraining, as illustrated in Fig. 1. In contrast, HPT [6] extends the input by incorporating some dataset-specific parameters, which helps address the issue of inconsistent coordinate systems across different datasets. However, this approach lacks scalability and increases the complexity of training. + +To address this, we propose 4D-VLA, a framework that integrates 4D spatiotemporal information to resolve such ambiguities. By combining spatial coordinate embeddings with a 3D-aware module, it generates spatial vision tokens that align the robot's coordinate system with the scene, enhancing 3D perception. This also enables efficient encoding of multiple historical frames, improving temporal reasoning. Additionally, we introduce memory bank sampling, a frame sampling strategy that selects key frames based on historical similarity, boosting model efficiency. + +To further investigate the spatial understanding and generalization of VLA models, we introduce MV-Bench, a multi-view dataset that evaluates performance across diverse viewpoints. Our approach enables robust pretraining, improving generalization to novel scenarios while outperforming baselines. + +Our contributions are: (i) We propose 4D-VLA, an efficient VLA model that integrates a spatial module with vision features to generate 3D-aware spatial vision tokens, effectively mitigating coordinate system and state chaos, thereby significantly enhancing pretraining efficiency. Additionally, we introduce memory bank sampling, a simple but effective method for historical information sampling. (ii) Our model has been validated both in the simulated and real-world environment, demonstrating its superiority. (iii) We develop a multi-view simulation dataset MV-Bench to evaluate spatial understanding and generalization, on which our model achieves outstanding performance. + +# 2 Related works + +Vision-language models Recent advancements in vision-language models (VLMs) have significantly enhanced the integration of vision and language understanding across diverse domains. Models like Flamingo [7], LLAVA [8], and BLIP-2 [9] focus on aligning text and image feature spaces to facilitate effective image understanding. More recently, there has been growing interest in expanding language models to support multi-image inputs, enabling them to handle more complex tasks and real-world scenarios. For instance, [10, 11, 12, 13] utilize multi-image sequences to capture temporal and action-related dynamics across frames, providing robust video comprehension. In addition, some recent models have begun incorporating 3D information to bolster spatial reasoning, as seen in 3D-LLM [14] and Scene-LLM [15]. These models leverage 3D inputs to enhance understanding of spatial relationships within a scene. However, none of these models explicitly integrate 4D spatiotemporal information to fully capture both spatial and temporal dynamics within the architecture. + +Vision-language-action models The vision-language-action (VLA) model represents a significant advancement in vision-language research, enabling more complex and interactive tasks aimed at facilitating real-world environment interactions. [16, 17, 18] complete tasks by directly predicting trajectories. [19, 20, 21, 22, 23, 24] predict the robot's current actions to enable closed-loop control, while [25, 26] enhance action prediction by training a world model to forecast future states. Some works [27, 28] also drives VLA policies using simple concatenation of historical observations. + +Recent works leverage diverse robotic datasets from various scenes and robot types to pretrain models for better generalization in novel environments. [29, 30, 3, 1] use the single image as input and are pretrained on large-scale datasets, while Octo [31] incorporates historical context and uses a diffusion head to predict the next $n$ actions. HPT [6] addresses the heterogeneity among different datasets by introducing dataset-specific parameters during pretraining to improve training efficiency. However, these methods overlook that the inefficiency in prior pretraining arises from insufficient input context, resulting in a high variance of the conditioned action distribution $A_{t}(\cdot)$ and ultimately hindering pretraining effectiveness. Our approach tackles this issue by introducing 4D information to mitigate coordinate system chaos and state chaos. This enables the model to learn meaningful action distributions from diverse datasets, thereby enhancing performance. + +# 3 Method + +This section provides a comprehensive overview of our proposed 4D-VLA. As shown in Fig. 2, our model processes sequential RGB-D images as input, converting them into corresponding spatial vision tokens. These tokens, together with task-specific text tokens, serve as feature inputs for the subsequent Transformer. After decoding through the VLM Transformer and action head, it ultimately generates the action output. + +# 3.1 Preliminary + +Problem definition The vision-language action (VLA) model takes a language instruction as input and aims to control a robot to accomplish the specified task. Specifically, VLA with a low-level control policy refers to a class of models that use the current observations as input to predict an action for the robot in its present state, enabling end-to-end, closed-loop control. The action is defined by three components: $\Delta \pmb{x} \in \mathbb{R}^3$ , $\Delta \pmb{\theta} \in \mathbb{R}^3$ , and $g \in [0,1]$ , representing the control translation, rotation offset, and the open-close state of the robot's end-effector, respectively. + +Vision-language model backbone We leverage a pretrained large vision-language model (VLM) as the backbone, specifically InternVL-4B [12], which consists of a text tokenizer $\mathcal{T}$ , a vision encoder $\mathcal{E}$ , and a Transformer decoder $\mathcal{D}$ . The vision encoder processes visual observations, which are subsequently compressed by an MLP projector $\mathcal{P}$ to generate vision embeddings, while text inputs are tokenized and embedded to form structured textual tokens. These multimodal tokens are then fed into the decoder $\mathcal{D}$ for next-token prediction. This backbone provides a robust foundation for aligning visual information with the shared semantic space, enabling effective robot action generation. + +![](images/858b9d43ceca57890a55e6c59fa5d61028f304591d53252788388fb842bd167f.jpg) +Figure 2: Our 4D-VLA pipeline. Our memory bank sampling method selects informative frames from sequential RGB-D inputs. A vision encoder with 3D coordinate embeddings generates spatial-aware tokens, which are fused into a 4D spatiotemporal representation. Combined with text tokens, these are processed by the LLM to decode actions via an action head. + +# 3.2 Spatial-aware visual tokens + +A reasonable action prediction requires awareness of both semantic perception and spatial perception of the scene. A fundamental aspect of spatial perception is that the robot must accurately determine its relative position within the scene, effectively resolving coordinate system chaos. To address this issue while enhancing spatial perception without compromising semantic awareness, we fuse the coordinate information derived from depth with vision tokens, resulting in spatial vision tokens. + +In our method, the input image $\mathbf{I} \in \mathbb{R}^{3 \times h \times w}$ is first encoded by $\mathcal{E}$ into a feature map with a downsampling rate of $c$ , yielding $\mathbf{f}_v = \mathcal{E}(\mathbf{I}) \in \mathbb{R}^{k \times \frac{h}{c} \times \frac{w}{c}}$ , where $k$ denotes the feature dimension. Next, we obtain a downsampled depth map $\mathbf{D} \in \mathbb{R}^{\frac{h}{c} \times \frac{w}{c}}$ , which assigns depth values to each corresponding feature volume. Using the camera's extrinsic $[\mathbf{R}|\mathbf{T}]$ and intrinsics $\mathbf{K}$ , we back-project the depth value into 3D coordinates $\mathbf{P}_w \in \mathbb{R}^{3 \times \frac{h}{c} \times \frac{w}{c}}$ within the world (or robot) coordinate system: + +$$ +\mathbf {P} _ {w} (\cdot , u, v) = \mathbf {R} \left(\mathbf {D} (u, v) \cdot \mathbf {K} ^ {- 1} \left[ \begin{array}{l} u \\ v \\ 1 \end{array} \right]\right) + \mathbf {T}. \tag {1} +$$ + +We apply a learnable positional embedding $\mathcal{E}_S$ to encode the 3D coordinates and integrate it with the original visual feature map via element-wise addition, forming spatial vision features that enhance spatial representation. These features are then processed by the MLP projector $\mathcal{P}$ within InternVL [12], generating spatial vision tokens $e^{ST} = \mathcal{P}(\mathcal{E}(\mathbf{I}) + \mathcal{E}_S(\mathbf{P}_w))$ . + +# 3.3 4D representation with multi-frame encoding + +As spatial vision tokens are designed to represent information from a single frame, and the LLM is fine-tuned with video data, we can naturally extend our encoding process to incorporate multi-frame information using the same approach. This allows us to integrate coherent 4D spatiotemporal representations seamlessly. A naive approach is to use the spatial vision tokens from the current frame along with those from uniformly sampled historical frames as input $\{e_{t - j}^{ST}\mid j = 0,1,2,\dots ,n - 1\}$ , where $n$ denotes the total number of temporal frames used, and $t$ represents the index of the current frame in the sequence. These tokens, together with the corresponding text tokens from task instructions, are then fed into the VLM's Transformer for decoding. + +Memory bank sampling. However, our experiments reveal that the performance of the model is highly sensitive to both the sampling interval and the total temporal window $n$ . While denser frame sampling improves performance, it significantly increases memory consumption and reduces inference speed. Moreover, since a robot's movement speed varies over time, a naive uniform sampling strategy often leads to redundant information, resulting in inefficiencies. + +To effectively exploit temporal information, we propose an adaptive historical frame sampling method based on a memory bank, aiming to capture rich historical information with a minimal number of frames. Specifically, given current timestamp $t$ , all image observations $\{\mathbf{I}_{t-j} \mid j = 0,1,2,\dots,n-1\}$ with a temporal window $n$ , memory bank sampling $\mathcal{M}$ returns a set of $k$ sampled timestamps $\mathcal{H} = \mathcal{M}(t,\{\mathbf{I}\},k,\phi)$ , where $k$ is the number of sample frames based on certain feature extractor $\phi$ . + +The algorithm is detailed in Alg. 1, which sequentially traverses image groups while maintaining a similarity queue, ensuring each newly added frame has lower similarity than the current maximum. + +Temporal positional encoding. Since memory bank sampling follows a non-uniform strategy, it is essential to encode the temporal position of each sampled spatial vision token relative to the current frame. To enhance the model's flexibility and generalization, we introduce a time encoding token $e^T$ , which captures the relative temporal offset between the historical and current frames, $e_j^T = \mathcal{E}_T(t - j)$ , where $j$ denotes the timestamp of a historical frame, $t$ represents the current observation time, and $\mathcal{E}_t$ is the learnable temporal encoding function. Accordingly, the final input token set is structured as: + +$$ +\mathcal {X} = \bigcup_ {i \in \mathcal {H}} \left[ \boldsymbol {e} _ {i} ^ {T} \mid \boldsymbol {e} _ {i} ^ {S T} \right] \cup \left\{\boldsymbol {e} ^ {\text {t e x t}} \right\}, \tag {2} +$$ + +where $e^{text}$ represents the text instruction tokens. Each sampled frame $i$ contributes a token pair $[e_i^T \mid e_i^{ST}]$ , with the temporal encoding token preceding the spatial vision token. This design ensures that the model effectively captures temporal relationships while maintaining spatial-awareness. + +# 3.4 Loss functions + +To accelerate control policy generation, we use an action head with two MLP layers that predict the action $[\Delta \hat{x},\Delta \hat{\theta},\hat{y} ]$ using the hidden features of the VLM Transformer's last token. + +Our total training loss can be written as follows: + +$$ +\mathcal {L} = \mathcal {L} _ {t} + \mathcal {L} _ {r} + \mathcal {L} _ {g} + \lambda_ {d} \mathcal {L} _ {d}, \tag {3} +$$ + +where the translation loss $\mathcal{L}_t = \|\Delta \hat{\pmb{x}} - \Delta \pmb{x}\|_2$ , the rotation loss $\mathcal{L}_r = \|\Delta \hat{\pmb{\theta}} - \Delta \pmb{\theta}\|_2$ , and the grip loss $\mathcal{L}_g = \mathrm{BCE}(\hat{g}, g)$ . Since the translation $\|\Delta \pmb{x}\|$ in action is often small, we place greater emphasis on directional awareness within the action by introducing a directional loss: + +$$ +\mathcal {L} _ {d} = \left\| \boldsymbol {d} (\Delta \hat {\boldsymbol {x}}) - \boldsymbol {d} (\Delta \boldsymbol {x}) \right\| _ {2}, \tag {4} +$$ + +where $\pmb {d}(\pmb {x}) = \frac{\pmb{x}}{\|\pmb{x}\|_2 + \epsilon}$ $\epsilon$ is a small value ensuring smoothness at zero. + +# Algorithm 1: memory bank sampling + +Input: $t, \{\mathbf{I}_{t-j} \mid j = 0,1,\dots,n-1\}$ , sample size $k$ , feature extractor $\phi$ + +Output: A set of sampled timestamps $\mathcal{H}$ + +Initialize $\mathcal{H}\gets [t]$ Start with the current frame Initialize $\mathbf{S}\gets [-\inf ]$ Similarity list + +for $j = 1$ to $n - 1$ do + +$$ +s = \operatorname {S i m i l a r i t y} \left(\phi \left(\mathbf {I} _ {\mathcal {H} [ - 1 ]}\right), \phi \left(\mathbf {I} _ {t - j}\right)\right) +$$ + +if $\operatorname{len}(\mathcal{H}) < k$ then + +Append $t - j$ to $\mathcal{H}$ ; Append $s$ to S + +else + +$$ +m = \arg \max (\mathbf {S}) +$$ + +if $s < \mathbf{S}[m]$ then $\triangleright$ Insert and reorganize + +Append $t - j$ to $\mathcal{H}$ ; Append $s$ to $\mathbf{S}$ + +$$ +s ^ {\prime} = \text {S i m i l a r i t y} (\phi (\mathbf {I} _ {\mathcal {H} [ m - 1 ]}), \phi (\mathbf {I} _ {\mathcal {H} [ m + 1 ]})) +$$ + +Remove $\mathcal{H}[m]$ + +Replace $\mathbf{S}[m + 1]$ with $s^\prime$ ; Remove $\mathbf{S}[m]$ + +else $\triangleright$ Replace the last frame + +$$ +s ^ {\prime} = \operatorname {S i m i l a r i t y} \left(\phi \left(\mathbf {I} _ {\mathcal {H} [ - 2 ]}\right), \phi \left(\mathbf {I} _ {\mathcal {H} [ t - j ]}\right)\right) +$$ + +Remove $\mathcal{H}[-1]$ + +Append $t - j$ to $\mathcal{H}$ ; Append $s'$ to $\mathbf{S}$ + +return $\mathcal{H}$ + +# 3.5 MV-Bench + +We propose the MV-Bench to provide a comprehensive evaluation of model capabilities in learning control policies across diverse viewpoints and generalizing to novel views, while also assessing the model's spatial understanding and adaptability. + +Benchmark settings. We build a multi-view dataset based on LIBERO-SPATIAL [4]. For each trajectory, we sample 6 training and 6 testing viewpoints uniformly within a $270^{\circ}$ front-facing range. Evaluation includes two tasks: In-View, where training and testing use the same views; and Cross-View, where testing is done on unseen viewpoints. Our camera settings are shown in Fig. 3. + +# 4 Experiments + +We first introduce the datasets and simulation environment, then describe pretraining and fine-tuning. Our model is pretrained on real-world data and fine-tuned with both simulation and real-world trajectories. We run closed-loop evaluations in diverse environments and report task performance. + +In addition, we evaluate on the ARM4R [32] benchmark for a direct comparison with ARM4R. Note that ARM4R pretrains with 3D inputs, while our approach differs in supervision and architecture; results and a detailed discussion are provided in Appx. 7.1. + +![](images/916532d6449e777b17bd9b161f0c8940b24613cd7b980faa9c3bd6153f2f2a73.jpg) +Figure 3: Our MV-Bench camera setting. We select 6 diverse viewpoints as training views and render images for all LIBERO-SPATIAL tasks. Novel inference views are placed near the training views. To avoid occlusion from the black box, test views in blocked areas are excluded. + +![](images/2eed5909dc098a379d6ea2b3312c9b9f447295a89bee7f2ce2f78cec23c9336b.jpg) + +
MethodSpatialObjectGoalLongAvg.
UniAct-0.5B [33]†64.577.568.046.564.1
SparseVLM [34]†79.867.072.639.464.7
FastV [35]†83.484.074.251.673.3
VLA-Cache [36]†83.885.876.452.874.7
DiffusionPolicy [16]78.3 ± 1.192.5 ± 0.768.3 ± 1.250.5 ± 1.372.4 ± 0.7
TraceVLA [37]84.6 ± 0.285.2 ± 0.475.1 ± 0.354.1 ± 1.074.8 ± 0.5
SpatialVLA [38]88.2 ± 0.589.9 ± 0.778.6 ± 0.655.5 ± 1.078.1 ± 0.7
Octo [31]78.9 ± 1.085.7 ± 0.984.6 ± 0.951.1 ± 1.375.1 ± 0.6
OpenVLA [1]84.7 ± 0.988.4 ± 0.879.2 ± 1.053.7 ± 1.376.5 ± 0.6
4D-VLA(Ours)88.9 ± 0.595.2 ± 0.390.9 ± 0.479.1 ± 1.288.6 ± 0.3
+ +Table 1: Evaluation of success rate on LIBERO. Bold indicates the best-performing model. Our model significantly outperforms other competitors, with an average success rate 12.1 higher than OpenVLA. $\dagger$ Denotes no available standard deviation data. + +# 4.1 Datasets and simulation environments + +DROID [2] A diverse real-world robot manipulation dataset with 76,000 demonstration trajectories, or 350 hours of interaction data, spanning a total of 564 scenes and 86 tasks, each featuring RGB-D data from two third-person and one wrist-mounted camera. + +LIBERO [4] The LIBERO benchmark is a simulation suite with 4 task sets designed to advance life-long learning in robotic manipulation. LIBERO-SPATIAL, LIBERO-OBJECT, and LIBERO-GOAL explore knowledge transfer in spatial reasoning, object understanding, and task goals. LIBERO-100 includes 90 short-horizon (LIBERO-90) and 10 long-horizon (LIBERO-LONG) tasks, covering 130 subtasks, each with 50 trajectories captured from both a main and wrist-mounted camera. + +# 4.2 Pretraining setups + +Data process. We pretrain our model on the DROID [2] dataset. RGB-D frames are resized to $448 \times 252$ , and each trajectory is uniformly downsampled to 100 actions. We remove frames with unchanged proprioception, specifically the stationary frames, and exclude trajectories with a total action count exceeding 600. Actions are defined as the difference between the end-effector's current and target states, with translation scaled by 15 and rotation (Euler angles) by 5 for normalization. + +Pretraining details. Our pretrained model is based on InternVL-4B. We use a temporal window of 20 and apply memory bank sampling to select 5 past frames along with the current frame. The RGB-D inputs are processed by the original vision encoder in InternVL, with 3D positional encodings incorporated before being passed to a projector—an MLP with a downsampling rate of 4, initialized from the pretrained InternVL weights. To handle sparse or incomplete depth from DROID, we compute the mean depth within each vision patch. If over $90\%$ of a patch is masked, we skip adding 3D features, effectively acting as dropout and data augmentation. + +In the training process, we freeze the vision encoder but training all other parameters. $\lambda_{d}$ is set to 1. We utilize a cosine learning rate scheduler with an initial learning rate of 2e-5. Our model was trained for 1 epoch with a batch size of 512, requiring around 20k steps to complete. Training was conducted on 8 NVIDIA A6000 GPUs over 96 hours. Inference with FlashAttention in bf16 requires approximately 8 GB of GPU memory. + +![](images/e42f7c82a86c1a50a297617a89dc617e8e1f6eaf9ff9cc53b11c49b84aa4eea2.jpg) +Task1: Spatial generalization + +![](images/046e2893d18a91d8fd77f0d1f7b2472564a9e3dcb48a6ad089d96a2720f61a5a.jpg) +Placing the yellow cube into a plate from unseen spatial positions. + +![](images/e6e431fca8083c5820378c088a9cfcb93a8b05249366b32706f4a8846b171e92.jpg) +Task2: Robustness to distractors + +![](images/f98f1ceb61cf7d519d3a1ef34e361f496e5cc30f6ce122ca845fda8d53425ec7.jpg) +Placing two green cubes into the plate under unseen background. + +![](images/46d934f415ec21aa461ad16341f7238a14a634a76fc50484a98ee5e0af89f37e.jpg) + +![](images/45fe61e1e4790f77603b679fbeb7e848eae431271414366885844580b6d3ce2e.jpg) + +![](images/7a24e91111bc27e8cf8d0b43f384f95039fd8132604546a4f89d22c0f4bcfa6c.jpg) +Task3: Precise placement +Stacking the yellow cube precisely onto the red cube. + +![](images/00cfc6dd08d280e1f4063b8f50326a7254987c622e8875648747046495947548.jpg) + +![](images/a2d762f07f4429db270001a375e40e7591e4972dea35adc1cf060b1ebe947d80.jpg) +Task4: Instruction following + +![](images/d1efffd01ff162867b072d78cf9958937ea7603b398258ad6a34e850d9e8097d.jpg) + +![](images/db88bbae4b1e51ee978c06e3f16cc9acf7ec7f9bc39c3ae8d67cebc85b34b5c4.jpg) + +![](images/5e4a4d71685e87c700c90c26c6168e2b6febc7f1455abb41ef50ea8748e87a6b.jpg) + +![](images/2d64b3699bca69e8cf34d85ec1fd116e2634e063d4eb3dff863cf2e15a101562.jpg) +Following color-ordered pick-up instructions +Figure 4: Our real-world experiment settings. These settings aim to evaluate the model's spatial generalization, robustness to distractors, precision in placement, and ability to follow instructions. Each row presents a 3-frame execution snapshot. + +# 4.3 LIBERO evaluation + +After pretraining, we fine-tune and conduct close-loop testing in the LIBERO simulation environment, using task success rate as the metric to evaluate the performance. + +Evaluation protocol. For each task in LIBERO, there are 10 test subtasks, and each subtask contains 50 different object layouts. This requires 1500 simulation tracks to evaluate a single task. The random seed is used to alter the initial state of the objects. To evaluate each task, we randomly sample 3 different seeds to calculate the mean and standard deviation of the success rate. + +Fine-tuning details. Unlike the pretraining phase, we used the simplest input settings to enable our model to learn the interaction effects between 3D information and historical data on actions. In the subsequent fine-tuning phase, we set the number of sequential frames $k = 5$ with a window size $n = 20$ . The current state RGB-D image is resized to $448 \times 448$ , while historical images are resized to $224 \times 224$ to optimize memory usage and model efficiency. Following [1], we normalize the ground truth action label using the 99th and 1st percentiles. We employed a cosine learning rate scheduler with a learning rate of 4e-5, using a batch size of 128 and training for 20 epochs. $\lambda_{d}$ is set to 0. During training, we activated all network parameters except the vision encoder. + +Experimental results. As shown in Tab. 1, our $4D$ -VLA shows clear performance gains over prior methods, with $5.2\%$ over OpenVLA[1] on LIBERO-SPATIAL, $2.7\%$ over DiffusionPolicy [16] on LIBERO-OBJECT, and $6.3\%$ over Octo [31] on LIBERO-GOAL. On the most challenging LIBERO-LONG task, it surpasses OpenVLA by $25.4\%$ . On average, $4D$ -VLA improves success rate by $12.1\%$ than OpenVLA, demonstrating stronger stability and spatiotemporal reasoning in complex settings. + +# 4.4 MV-Bench evaluation + +As shown in Tab. 2, our model achieves a $81.0\%$ success rate in the In-View setting, demonstrating its capability to handle diverse training views effectively. This result shows that the integration of spatiotemporal information enables our model to manage challenging and conflicting perspectives, outperforming OpenVLA [1] by a significant margin of $28.8\%$ . In the Cross-View evaluation, our model also achieves the best per + +formance, indicating a notable generalization capability across novel viewpoints. It highlights the robustness of our model in handling diverse viewpoints. + +
MethodIn-View, Δ0°Avg.
Angle0°60°120°270°300°330°
OpenVLA [1]57.450.050.643.553.857.852.2
4D-VLA (Ours)83.287.079.570.275.890.281.0
Cross-View, Δ15°Δ30°
Angle15°45°75°105°30°90°
OpenVLA [1]64.048.254.234.263.039.250.5
4D-VLA (Ours)83.483.274.065.675.860.873.8
+ +Table 2: Evaluation of success rate on MV-Bench. $\Delta$ symbol representing the angular deviation from the nearest training viewpoint along the z-axis. + +# 4.5 Real-world evaluation + +To evaluate models in real-world scenarios, we conducted physical experiments using a Franka robotic arm. Specifically, we designed 4 representative tasks to assess the model's capabilities, using success rate as the primary evaluation metric. For each task, we manually collected 50 diverse trajectories for + +
MethodTask 1Task 2Task 3Task 4Avg.
OpenVLA [1]45.0022.5030.0013.3327.70
Base VLA35.0020.005.002.6715.67
+ Pretraining60.0060.0040.0028.0047.00
+ Pretraining + Coord.75.0060.0085.0034.6763.67
+ Pretraining + Hist.80.0077.5070.0036.0065.88
+ Pretraining + Coord. + Hist. (full model)90.0082.5090.0080.0085.63
+ +Table 3: Real-world evaluation results. We incrementally improve the Base VLA by adding pretraining, coordinate encoding, and historical frames selected via memory bank sampling. + +training and trained the model for 20 epochs. For Task 4, involving color-ordered pick-up instructions, we used 5 color sequences with 10 demonstrations each, totaling 50 trajectories. + +Task descriptions. As shown in Fig. 4, we design 4 real-world manipulation tasks to evaluate different aspects of the model's spatial reasoning and generalization capabilities. Task 1: Spatial generalization. The robot is tasked with placing a yellow cube into a plate from positions not seen during training, evaluating its ability to generalize across novel spatial positions. Task 2: Robustness to distractors. The robot is asked to place two green cubes into the plate in scenes with cluttered backgrounds. This task evaluates the model's robustness to environments distractors. Task 3: Precise placement. The robot is asked to precisely stack the yellow cube onto the red one. This task emphasizes the need for fine-grained action prediction. Task 4: Instruction following. The robot is asked to execute color-ordered pick-and-place commands (e.g., red $\rightarrow$ green $\rightarrow$ blue). The task assesses the model's ability to follow structured instructions correctly. + +Evaluation metrics. Tasks 1 and 3 are evaluated by success rate (i.e., successful trials / total trials). For Task 2, each correctly placed green block earns 1 point (maximum 2 per trial), with performance measured as a total score out of 40. Task 4 involves following color-ordered instructions across 5 combinations (5 trials each). Each correctly placed block earns 1 point (maximum 3 per trial), for a total score out of 75. The final performance is reported as score divided by 75. + +Experimental results. We set InternVL-4B with single RGB image inputs followed by an action head as our Base VLA model. Based on this, we incrementally add different modules to investigate their individual contributions to performance. OpenVLA [1] is set as our main competitor. + +As shown in Tab. 3, the base VLA model without pretraining underperforms OpenVLA on all tasks. Our spatially grounded pretraining significantly boosts performance—even with a single RGB-D frame—confirming its effectiveness. In short-horizon tasks (Task 1 and 3), adding coordinate information notably improves results, especially in Task 3, which demands precise spatial alignment. This shows that coordinate encoding enhances spatial grounding and action accuracy. In long-horizon tasks (Task 2 and 4), the model often succeeds in the first step but fails the second without access to history, due to latent state ambiguity. memory bank sampling alleviates this by providing temporally relevant frames, improving multi-step reasoning. Overall, we find that matching the downstream input to the pretraining setting (e.g., coordinate-aware, temporally structured) leads to better transfer of spatial representations. Even with mismatched inputs, our model still outperforms baselines, showing strong generalizability. Coordinate encoding offers explicit spatial cues, while multi-frame inputs provide temporal context to disambiguate action intent. + +# 4.6 Multi-view real-world evaluation + +In this section, we conduct additional real-world experiments under a multi-view camera setup. We design two more challenging tasks to evaluate the model's generalization ability with respect to: (i) variations in object locations together with the changed background environments; (ii) inputs from novel camera viewpoints. We use 4 fixed cameras to capture each demonstration from different angles, collecting 50 trajectories per task per camera—resulting in a total of 200 trajectories per task for training. All models are trained for 20 epochs, and performance is measured by success rate. + +Task descriptions. These two more challenging tasks are shown in Fig. 5. Task 1: Out-of-distribution generalization. The robot is tasked with placing a yellow cube into a plate under conditions where both the spatial configuration and the surrounding environment differ from those seen during training. These variations include changes in the plate's position, the presence and location + +![](images/2a0b668fd51da82d497b9014719cd4959a2021f8056be918fa2d538f0b8c6645.jpg) +Training + +![](images/843174fa21f42faafd781f8e39720c3e07bf619b1fdd97c410941470dc9432a5.jpg) + +![](images/eafbe75d0bd88097aa2c6101862121d6ea512438acf15bef7fca74c78a953136.jpg) + +![](images/2799a858ca19f4bbfdc6b6c9fb8ed6daa81c296ccf43623bade71b8a795c964e.jpg) + +![](images/8ffeaa5d39ca8086082a24bfb96793ae7c7309427baea81c030c74290477ae42.jpg) +Placing the yellow cube into the plate. + +![](images/7dd0d1fd47f83e97526ef32553f3fb9a20fe5591f8c59aa7aec646e0f19a16c7.jpg) +Inference-Task 1 + +![](images/8eb3be4920698ef675a514a47b353f82cd457a10316c57608cea0d3143b8632f.jpg) +Placing the yellow cube into the plate. (Training view, different cubes spatial positions) + +![](images/02c5594460e51bd0d1281bf89fea48eb716387a6770ec649d3f8786b872d57b8.jpg) +Inference-Task 2 +Placing the yellow cube into the plate. (Novel-view, different cubes spatial positions) +Figure 5: Our multi-view real-world experiment settings. These settings aim to evaluate the model's out-of-distribution and novel-view generalization ability. + +![](images/6b96f05ad25d3f23b405645fcc917a9a8d37532717f1daa3a332e8b13837a99c.jpg) + +
MethodIn-ViewCross-ViewAvg.
Angle0°90°180°225°Δ15°(−15°)Δ25°(−25°)Δ45°(135°)
OpenVLA [1]251530103010518
4D-VLA (Ours)6050656550554055
+ +Table 4: Real-world multi-view evaluation. We test our model's spatial generalization across varying viewpoints and object layouts. 4D-VLA shows strong in-view and cross-view performance, highlighting its robustness under real-world distribution shifts. + +of distractor objects (e.g., other cubes). This task evaluates the model's ability to generalize to unseen object arrangements and background contexts, testing its robustness in real-world deployments beyond the training distribution. Task 2: Novel-view generalization. Similar to Task 1, the robot is asked to place a yellow cube into a plate, with spatial setup and surroundings differing from training. However, during inference, data input is captured exclusively from an additional novel, unseen camera viewpoint that was not used during training. This task evaluates the model's viewpoint robustness—its ability to generalize across camera perspectives and accurately interpret the scene from unfamiliar angles. Success in this task reflects strong spatial understanding and invariance to viewpoint changes, both critical for real-world multi-camera deployment. To simplify the setup, the target block is only moved within a small spatial range, while the background is fully randomized. + +Evaluation metrics. Each multi-view task is evaluated over 20 trials. In every trial, both the background and object positions are randomly shuffled to assess the model's robustness and generalization. The evaluation metric is the task success rate, computed as the ratio of successful trials to the total number of trials. + +Experiments results. As shown in Tab. 4, $4D$ -VLA significantly outperforms OpenVLA in both in-view and cross-view settings, demonstrating strong generalization to viewpoint shifts and layout variations. In the in-view setting, where the camera is fixed but object layouts change, our model maintains consistently high success rates, indicating robustness to spatial perturbations. In the more challenging cross-view setting involving unseen viewpoints, $4D$ -VLA continues to perform stably across different angles. Although performance slightly drops at larger viewpoint shifts (e.g., $\Delta 45^{\circ}$ ), it remains stable compared to OpenVLA, whose success rate fluctuates more severely under such conditions. These results suggest that our model effectively captures spatial consistency across views, leading to more reliable visuomotor control in real-world environments. + +# 5 Discussion + +Building upon previous experiments, we further analyze three key questions: (i) The Role of historical information -How does historical context influence model's effectiveness? (ii) Ablation study on model components -How do different architectural components contribute to the overall + +performance of the model? (iii) The impact of coordinate system chaos -How does coordinate system inconsistency affect the model's performance, and can the introduction of a 3D representation mitigate this issue? The first question is addressed in Sec. 5.1; the second is covered in Appx. 7.2, and the third in Appx. 7.3. Our discussion begins with the simple model, i.e., InternVL-4B with an MLP action head, using single RGB image as the vision input. + +# 5.1 Exploring historical information utilization + +As highlighted in Octo [31], incorporating historical context can significantly boost model performance. However, this has been underexplored. We conduct a comprehensive analysis to assess its true impact on model effectiveness. We define window size $n$ as the number of historical frames available to the model and $k$ as the number of frames sampled from them. To systematically investigate their impact, we design two experimental settings: (i) fixing $n$ while progressively increasing $k$ , and (ii) fixing $k$ while expanding $n$ . + +![](images/63fe7004ab49e3d13771de527f0015bf49a365d4464c058dacc48470effdf7c6.jpg) +Figure 6: Historical image analysis. Larger points indicate lower efficiency. + +
EncodingPositionFusionSuccess rate
learnablerelativeconcat75.6
learnablerelativeadditive74.8
sinusoidalrelativeadditive71.6
learnableabsoluteadditive0.0
-No temporal encoding-63.0
+ +Table 5: Ablation on temporal encoding method. + +As illustrated in Fig. 6, our findings reveal that the historical window size $n$ plays a pivotal role in determining performance, whereas the effect of $k$ is comparatively minor. This suggests that uniform sampling may introduce excessive redundancy, adversely diminishing the model's efficiency. memory bank sampling strategy, which effectively reduces redundancy and improves performance, even with a smaller $k$ , showcasing its effectiveness in maximizing the utility of historical context. + +Ablation on temporal encoding method. Since MBS employs non-uniform sampling, the absence of explicit temporal position encodings can lead to ambiguity in the historical sequence, making it difficult for the model to fully leverage past information. To address this, we investigate the effect of different temporal encoding strategies on model performance, as summarized in Tab. 5. Specifically, additive means the temporal encoding is directly added to the image tokens, while concat appends it before the image token, and absolute refers to encoding based on the sampled timestamps rather than relative to the current frame. Clearly, the concat method achieves the best performance, with the key advantage of seamlessly integrating with spatial vision tokens to enhance the model's representation. + +# 6 Conclusion + +In this paper, we present 4D-VLA, which incorporates 4D information to address challenges in leveraging diverse robotic datasets for pretraining, such as coordinate system chaos and state chaos. Our model encodes sequential RGB-D images into visual features with corresponding 3D coordinates, aligning the robot's coordinate system with the scene. This alignment enables strong spatiotemporal reasoning with minimal training. We also introduce memory bank sampling, a frame sampling strategy to extract informative and diverse key frames from sequences, improving efficiency. In the simulated LIBERO environment, 4D-VLA outperforms existing methods. Additionally, our multiview simulation dataset, MV-Bench, demonstrates superior spatial perception and generalization to novel viewpoints, surpassing current approaches. A limitation of our approach is its reliance on RGB-D input, which introduces hardware restriction. + +# Acknowledgments + +This work was supported in part by National Natural Science Foundation of China (Grant No. 62376060). + +# References + +[1] Moo Jin Kim, Karl Pertsch, Siddharth Karamcheti, Ted Xiao, Ashwin Balakrishna, Suraj Nair, Rafael Rafailov, Ethan Foster, Grace Lam, Pannag Sanketi, et al. Openvla: An open-source vision-language-action model. arXiv preprint, 2024. 2, 3, 6, 7, 8, 9 +[2] Alexander Khazatsky, Karl Pertsch, Suraj Nair, Ashwin Balakrishna, Sudeep Dasari, Siddharth Karamcheti, Soroush Nasiriany, Mohan Kumar Srirama, Lawrence Yunliang Chen, Kirsty Ellis, Peter David Fagan, Joey Hejna, Masha Itkina, Marion Lepert, Yecheng Jason Ma, Patrick Tree Miller, Jimmy Wu, Suneel Belkhale, Shivin Dass, Huy Ha, Arhan Jain, Abraham Lee, Youngwoon Lee, Marius Memmel, Sungjae Park, Ilija Radosavovic, Kaiyuan Wang, Albert Zhan, Kevin Black, Cheng Chi, Kyle Beltran Hatch, Shan Lin, Jingpei Lu, Jean Mercat, Abdul Rehman, Pannag R Sanketi, Archit Sharma, Cody Simpson, Quan Vuong, Homer Rich Walke, Blake Wulfe, Ted Xiao, Jonathan Heewon Yang, Arefeh Yavary, Tony Z. Zhao, Christopher Agia, Rohan Baijal, Mateo Guaman Castro, Daphne Chen, Qiuyu Chen, Trinity Chung, Jaimyn Drake, Ethan Paul Foster, Jensen Gao, David Antonio Herrera, Minho Heo, Kyle Hsu, Jiaheng Hu, Donovon Jackson, Charlotte Le, Yunshuang Li, Kevin Lin, Roy Lin, Zehan Ma, Abhiram Maddukuri, Suvir Mirchandani, Daniel Morton, Tony Nguyen, Abigail O'Neill, Rosario Scalise, Derick Seale, Victor Son, Stephen Tian, Emi Tran, Andrew E. Wang, Yilin Wu, Annie Xie, Jingyun Yang, Patrick Yin, Yunchu Zhang, Osbert Bastani, Glen Berseth, Jeannette Bohg, Ken Goldberg, Abhinav Gupta, Abhishek Gupta, Dinesh Jayaraman, Joseph J Lim, Jitendra Malik, Roberto Martin-Martin, Subramanian Ramamoorthy, Dorsa Sadigh, Shuran Song, Jiajun Wu, Michael C. Yip, Yuke Zhu, Thomas Kollar, Sergey Levine, and Chelsea Finn. Droid: A large-scale in-the-wild robot manipulation dataset. 2024. 2. 6 +[3] Open X-Embodiment Collaboration, Abhishek Padalkar, Acorn Pooley, Ajinkya Jain, Alex Bewley, Alex Herzog, Alex Irpan, Alexander Khazatsky, Anant Rai, Anikait Singh, Anthony Brohan, Antonin Raffin, Ayzaan Wahid, Ben Burgess-Limerick, Beomjoon Kim, Bernhard Scholkopf, Brian Ichter, Cewu Lu, Charles Xu, Chelsea Finn, Chenfeng Xu, Cheng Chi, Chenguang Huang, Christine Chan, Chuer Pan, Chuyuan Fu, Coline Devin, Danny Driess, Deepak Pathak, Dhruv Shah, Dieter Buchler, Dmitry Kalashnikov, Dorsa Sadigh, Edward Johns, Federico Ceola, Fei Xia, Freek Stulp, Gaoyue Zhou, Gaurav S. Sukhatme, Gautam Salhotra, Ge Yan, Giulio Schiavi, Hao Su, Hao-Shu Fang, Haochen Shi, Heni Ben Amor, Henrik I Christensen, Hiroki Furuta, Homer Walke, Hongjie Fang, Igor Mordatch, Iija Radosavovic, Isabel Leal, Jacky Liang, Jaehyung Kim, Jan Schneider, Jasmine Hsu, Jeannette Bohg, Jeffrey Bingham, Jiajun Wu, Jialin Wu, Jianlan Luo, Jiayuan Gu, Jie Tan, Jihoon Oh, Jitendra Malik, Jonathan Tompson, Jonathan Yang, Joseph J. Lim, João Silverio, Junhyek Han, Kanishka Rao, Karl Pertsch, Karol Hausman, Keegan Go, Keerthana Gopalakrishnan, Ken Goldberg, Kendra Byrne, Kenneth Oslund, Kento Kawaharazuka, Kevin Zhang, Keyvan Majd, Krishan Rana, Krishnan Srinivasan, Lawrence Yunliang Chen, Lerrel Pinto, Liam Tan, Lionel Ott, Lisa Lee, Masayoshi Tomizuka, Maximilian Du, Michael Ahn, Mingtong Zhang, Mingyu Ding, Mohan Kumar Srirama, Mohit Sharma, Moo Jin Kim, Naoaki Kanazawa, Nicklas Hansen, Nicolas Heess, Nikhil J Joshi, Niko Suenderhauf, Norman Di Palo, Nur Muhammad Mahi Shafiullah, Oier Mees, Oliver Kroemer, Pannag R Sanketi, Paul Wohlhart, Peng Xu, Pierre Sermanet, Priya Sundaresan, Quan Vuong, Rafael Rafailov, Ran Tian, Ria Doshi, Roberto Martín-Martín, Russell Mendonca, Rutav Shah, Ryan Hoque, Ryan Julian, Samuel Bustamante, Sean Kirmani, Sergey Levine, Sherry Moore, Shikhar Bahl, Shivin Dass, Shuran Song, Sichun Xu, Siddhant Haldar, Simeon Adebola, Simon Guist, Soroush Nasiriany, Stefan Schaal, Stefan Welker, Stephen Tian, Sudeep Dasari, Suneel Belkhale, Takayuki Osa, Tatsuya Harada, Tatsuya Matsushima, Ted Xiao, Tianhe Yu, Tianli Ding, Todor Davchev, Tony Z. Zhao, Travis Armstrong, Trevor Darrell, Vidhi Jain Vincent Vanhoucke Wei Zhan Wenxuan Zhou Wolfram Burgard Xi Chen Xiaolong Wang Xinghao Zhu Xuanlin Li Yao Lu Yevgen Chebotar Yifan Zhou Yifeng Zhu Ying Xu Yixuan Wang Yonatan Bisk Yoonyoung Cho Youngwoo Lee Yuchen Cui Yueh hua Wu Yujin Tang Yuke Zhu Yunzhu Li Yusuke Iwasawa Yutaka Matsuo Zhuo Xu and Zichen Jeff Cui. Open X-Embodiment: Robotic learning datasets and RT-X models. https://arxiv.org/abs/2310.08864 2023. 2.3 +[4] Bo Liu, Yifeng Zhu, Chongkai Gao, Yihao Feng, Qiang Liu, Yuke Zhu, and Peter Stone. Libero: Benchmarking knowledge transfer for lifelong robot learning. In NeurIPS, 2024. 2, 5, 6 + +[5] Oier Mees, Lukas Hermann, Erick Rosete-Beas, and Wolfram Burgard. Calvin: A benchmark for language-conditioned policy learning for long-horizon robot manipulation tasks. In RA-L, 2022. 2 +[6] Lirui Wang, Xinlei Chen, Jialiang Zhao, and Kaiming He. Scaling proprioceptive-visual learning with heterogeneous pre-trained transformers. arXiv preprint, 2024. 2, 3 +[7] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. In NeurIPS, 2022. 3 +[8] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023. 3 +[9] Junnan Li, Dongxu Li, Silvio Savarese, and Steven C. H. Hoi. BLIP-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, 2023. 3 +[10] Tim Dettmers, Mike Lewis, Younes Belkada, and Luke Zettlemoyer. Gpt3. int8(): 8-bit matrix multiplication for transformers at scale. In NeurIPS, 2022. 3 +[11] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint, 2023. 3 +[12] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In CVPR, 2024. 3, 4 +[13] Guo Chen, Yin-Dong Zheng, Jiahao Wang, Jilan Xu, Yifei Huang, Junting Pan, Yi Wang, Yali Wang, Yu Qiao, Tong Lu, et al. Videollm: Modeling video sequence with large language models. arXiv preprint, 2023. 3 +[14] Yining Hong, Haoyu Zhen, Peihao Chen, Shuhong Zheng, Yilun Du, Zhenfang Chen, and Chuang Gan. 3d-llm: Injecting the 3d world into large language models. In NeurIPS, 2023. 3 +[15] Rao Fu, Jingyu Liu, Xilun Chen, Yixin Nie, and Wenhan Xiong. Scene-llm: Extending language model for 3d visual understanding and reasoning. arXiv preprint, 2024. 3 +[16] Cheng Chi, Siyuan Feng, Yilun Du, Zhenjia Xu, Eric Cousineau, Benjamin Burchfiel, and Shuran Song. Diffusion policy: Visuomotor policy learning via action diffusion. In RSS, 2023. 3, 6, 7 +[17] Wenlong Huang, Chen Wang, Ruohan Zhang, Yunzhu Li, Jiajun Wu, and Li Fei-Fei. Voxposer: Composable 3d value maps for robotic manipulation with language models. arXiv preprint, 2023. 3 +[18] Wenlong Huang, Chen Wang, Yunzhu Li, Ruohan Zhang, and Li Fei-Fei. Rekep: Spatiotemporal reasoning of relational keypoint constraints for robotic manipulation. arXiv preprint, 2024. 3 +[19] Chuan Wen, Xingyu Lin, John So, Kai Chen, Qi Dou, Yang Gao, and Pieter Abbeel. Any-point trajectory modeling for policy learning. arXiv preprint, 2023. 3 +[20] Homanga Bharadhwaj, Roozbeh Mottaghi, Abhinav Gupta, and Shubham Tulsiani. Track2act: Predicting point tracks from internet videos enables generalizable robot manipulation. arXiv preprint, 2024. 3 +[21] Yunfan Jiang, Agrim Gupta, Zichen Zhang, Guanzhi Wang, Yongqiang Dou, Yanjun Chen, Li Fei-Fei, Anima Anandkumar, Yuke Zhu, and Linxi Fan. Vima: General robot manipulation with multimodal prompts. arXiv preprint, 2022. 3 +[22] Xinghang Li, Minghuan Liu, Hanbo Zhang, Cunjun Yu, Jie Xu, Hongtao Wu, Chilam Cheang, Ya Jing, Weinan Zhang, Huaping Liu, et al. Vision-language foundation models as effective robot imitators. arXiv preprint, 2023. 3 + +[23] Mengda Xu, Zhenjia Xu, Yinghao Xu, Cheng Chi, Gordon Wetzstein, Manuela Veloso, and Shuran Song. Flow as the cross-domain manipulation interface. arXiv preprint, 2024. 3 +[24] Jiangyong Huang, Silong Yong, Xiaojian Ma, Xiongkun Linghu, Puhao Li, Yan Wang, Qing Li, Song-Chun Zhu, Baoxiong Jia, and Siyuan Huang. An embodied generalist agent in 3d world. In ICML, 2024. 3 +[25] Haoyu Zhen, Xiaowen Qiu, Peihao Chen, Jincheng Yang, Xin Yan, Yilun Du, Yining Hong, and Chuang Gan. 3d-vla: A 3d vision-language-action generative world model. In ICML, 2024. 3 +[26] Hongtao Wu, Ya Jing, Chilam Cheang, Guangzeng Chen, Jiafeng Xu, Xinghang Li, Minghuan Liu, Hang Li, and Tao Kong. Unleashing large-scale video generative pre-training for visual robot manipulation. In ICLR, 2024. 3 +[27] Ilija Radosavovic, Bike Zhang, Baifeng Shi, Jathushan Rajasegaran, Sarthak Kamat, Trevor Darrell, Koushil Sreenath, and Jitendra Malik. Humanoid locomotion as next token prediction. Advances in neural information processing systems, 37:79307-79324, 2024. 3 +[28] Ilija Radosavovic, Baifeng Shi, Letian Fu, Ken Goldberg, Trevor Darrell, and Jitendra Malik. Robot learning with sensorimotor pre-training. In CoRL, 2023. 3 +[29] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Joseph Dabis, Chelsea Finn, Keerthana Gopalakrishnan, Karol Hausman, Alex Herzog, Jasmine Hsu, Julian Ibarz, Brian Ichter, Alex Irpan, Tomas Jackson, Sally Jesmonth, Nikhil Joshi, Ryan Julian, Dmitry Kalashnikov, Yuheng Kuang, Isabel Leal, Kuang-Huei Lee, Sergey Levine, Yao Lu, Utsav Malla, Deeksha Manjunath, Igor Mordatch, Ofir Nachum, Carolina Parada, Jodilyn Peralta, Emily Perez, Karl Pertsch, Jornell Quiambao, Kanishka Rao, Michael Ryoo, Grecia Salazar, Pannag Sanketi, Kevin Sayed, Jaspiar Singh, Sumedh Sontakke, Austin Stone, Clayton Tan, Huong Tran, Vincent Vanhoucke, Steve Vega, Quan Vuong, Fei Xia, Ted Xiao, Peng Xu, Sichun Xu, Tianhe Yu, and Brianna Zitkovich. Rt-1: Robotics transformer for real-world control at scale. arXiv preprint, 2022. 3 +[30] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Xi Chen, Krzysztof Choromanski, Tianli Ding, Danny Driess, Avinava Dubey, Chelsea Finn, Pete Florence, Chuyuan Fu, Montse Gonzalez Arenas, Keerthana Gopalakrishnan, Kehang Han, Karol Hausman, Alex Herzog, Jasmine Hsu, Brian Ichter, Alex Irpan, Nikhil Joshi, Ryan Julian, Dmitry Kalashnikov, Yuheng Kuang, Isabel Leal, Lisa Lee, Tsang-Wei Edward Lee, Sergey Levine, Yao Lu, Henryk Michalewski, Igor Mordatch, Karl Pertsch, Kanishka Rao, Krista Reymann, Michael Ryoo, Grecia Salazar, Pannag Sanketi, Pierre Sermanet, Jaspiar Singh, Anikait Singh, Radu Soricut, Huong Tran, Vincent Vanhoucke, Quan Vuong, Ayzaan Wahid, Stefan Welker, Paul Wohlhart, Jialin Wu, Fei Xia, Ted Xiao, Peng Xu, Sichun Xu, Tianhe Yu, and Brianna Zitkovich. Rt-2: Vision-language-action models transfer web knowledge to robotic control. arXiv preprint, 2023. 3 +[31] Octo Model Team, Dibya Ghosh, Homer Walke, Karl Pertsch, Kevin Black, Oier Mees, Sudeep Dasari, Joey Hejna, Charles Xu, Jianlan Luo, Tobias Kreiman, You Liang Tan, Dorsa Sadigh, Chelsea Finn, and Sergey Levine. Octo: An open-source generalist robot policy. https://octo-models.github.io, 2023. 3, 6, 7, 10 +[32] Dantong Niu, Yuvan Sharma, Haoru Xue, Giscard Biamby, Junyi Zhang, Ziteng Ji, Trevor Darrell, and Roei Herzig. Pre-training auto-regressive robotic models with 4d representations. In Forty-second International Conference on Machine Learning. 5, 22 +[33] Jinliang Zheng, Jianxiong Li, Dongxiu Liu, Yinan Zheng, Zhihao Wang, Zhonghong Ou, Yu Liu, Jingjing Liu, Ya-Qin Zhang, and Xianyuan Zhan. Universal actions for enhanced embodied foundation models. arXiv preprint, 2025. 6 +[34] Yuan Zhang, Chun-Kai Fan, Junpeng Ma, Wenzhao Zheng, Tao Huang, Kuan Cheng, Denis Gudovskiy, Tomoyuki Okuno, Yohei Nakata, Kurt Keutzer, et al. Sparsevm: Visual token sparsification for efficient vision-language model inference. arXiv preprint, 2024. 6 +[35] Liang Chen, Haozhe Zhao, Tianyu Liu, Shuai Bai, Junyang Lin, Chang Zhou, and Baobao Chang. An image is worth 1/2 tokens after layer 2: Plug-and-play inference acceleration for large vision-language models. In ECCV, 2024.6 + +[36] Siyu Xu, Yunke Wang, Chenghao Xia, Dihao Zhu, Tao Huang, and Chang Xu. Vla-cache: Towards efficient vision-language-action model via adaptive token caching in robotic manipulation. arXiv preprint, 2025. 6 +[37] Ruijie Zheng, Yongyuan Liang, Shuaiyi Huang, Jianfeng Gao, Hal Daumé III, Andrey Kolobov, Furong Huang, and Jianwei Yang. Tracevla: Visual trace prompting enhances spatial-temporal awareness for generalist robotic policies. arXiv preprint, 2024. 6 +[38] Delin Qu, Haoming Song, Qizhi Chen, Yuanqi Yao, Xinyi Ye, Yan Ding, Zhigang Wang, JiaYuan Gu, Bin Zhao, Dong Wang, et al. Spatialvla: Exploring spatial representations for visual-language-action model. arXiv preprint, 2025. 6 +[39] Stephen James, Kentaro Wada, Tristan Laidlow, and Andrew J Davison. Coarse-to-fine q-attention: Efficient learning for visual robotic manipulation via discretisation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13739–13748, 2022. 22 +[40] Dantong Niu, Yuvan Sharma, Giscard Biamby, Jerome Quenum, Yutong Bai, Baifeng Shi, Trevor Darrell, and Roei Herzig. Llarva: Vision-action instruction tuning enhances robot learning. In Conference on Robot Learning, pages 3333-3355. PMLR, 2025. 22 +[41] Mohit Shridhar, Lucas Manuelli, and Dieter Fox. Perceiver-actor: A multi-task transformer for robotic manipulation. In Conference on Robot Learning, pages 785–799. PMLR, 2023. 22 +[42] Xiaogian Shen, Yunyang Xiong, Changsheng Zhao, Lemeng Wu, Jun Chen, Chenchen Zhu, Zechun Liu, Fanyi Xiao, Balakrishnan Varadarajan, Florian Bordes, et al. Longvu: Spatiotemporal adaptive compression for long video-language understanding. In *Forty-second International Conference on Machine Learning*. 22, 23 +[43] Yanwei Li, Chengyao Wang, and Jiaya Jia. Llama-vid: An image is worth 2 tokens in large language models. In European Conference on Computer Vision, pages 323–340. Springer, 2024. 22, 23 + +# NeurIPS Paper Checklist + +# 1. Claims + +Question: Do the main claims made in the abstract and introduction accurately reflect the paper's contributions and scope? + +Answer: [Yes] + +Justification: We have clearly stated the contribution and scope of the paper in the abstract. + +Guidelines: + +- The answer NA means that the abstract and introduction do not include the claims made in the paper. +- The abstract and/or introduction should clearly state the claims made, including the contributions made in the paper and important assumptions and limitations. A No or NA answer to this question will not be perceived well by the reviewers. +- The claims made should match theoretical and experimental results, and reflect how much the results can be expected to generalize to other settings. +- It is fine to include aspirational goals as motivation as long as it is clear that these goals are not attained by the paper. + +# 2. Limitations + +Question: Does the paper discuss the limitations of the work performed by the authors? + +Answer: [Yes] + +Justification: We have discussed the limitations of the work in the conclusion part. + +Guidelines: + +- The answer NA means that the paper has no limitation while the answer No means that the paper has limitations, but those are not discussed in the paper. +- The authors are encouraged to create a separate "Limitations" section in their paper. +- The paper should point out any strong assumptions and how robust the results are to violations of these assumptions (e.g., independence assumptions, noiseless settings, model well-specification, asymptotic approximations only holding locally). The authors should reflect on how these assumptions might be violated in practice and what the implications would be. +- The authors should reflect on the scope of the claims made, e.g., if the approach was only tested on a few datasets or with a few runs. In general, empirical results often depend on implicit assumptions, which should be articulated. +- The authors should reflect on the factors that influence the performance of the approach. For example, a facial recognition algorithm may perform poorly when image resolution is low or images are taken in low lighting. Or a speech-to-text system might not be used reliably to provide closed captions for online lectures because it fails to handle technical jargon. +- The authors should discuss the computational efficiency of the proposed algorithms and how they scale with dataset size. +- If applicable, the authors should discuss possible limitations of their approach to address problems of privacy and fairness. +- While the authors might fear that complete honesty about limitations might be used by reviewers as grounds for rejection, a worse outcome might be that reviewers discover limitations that aren't acknowledged in the paper. The authors should use their best judgment and recognize that individual actions in favor of transparency play an important role in developing norms that preserve the integrity of the community. Reviewers will be specifically instructed to not penalize honesty concerning limitations. + +# 3. Theory assumptions and proofs + +Question: For each theoretical result, does the paper provide the full set of assumptions and a complete (and correct) proof? + +Answer: [Yes] + +Justification: We have provided the theoretical assumption and proof of this paper in the method part. + +# Guidelines: + +- The answer NA means that the paper does not include theoretical results. +- All the theorems, formulas, and proofs in the paper should be numbered and cross-referenced. +- All assumptions should be clearly stated or referenced in the statement of any theorems. +- The proofs can either appear in the main paper or the supplemental material, but if they appear in the supplemental material, the authors are encouraged to provide a short proof sketch to provide intuition. +- Inversely, any informal proof provided in the core of the paper should be complemented by formal proofs provided in appendix or supplemental material. +- Theorems and Lemmas that the proof relies upon should be properly referenced. + +# 4. Experimental result reproducibility + +Question: Does the paper fully disclose all the information needed to reproduce the main experimental results of the paper to the extent that it affects the main claims and/or conclusions of the paper (regardless of whether the code and data are provided or not)? + +# Answer: [Yes] + +Justification: We have disclosed all the information needed to reproduce the main experimental results of the paper in the experiment part. + +# Guidelines: + +- The answer NA means that the paper does not include experiments. +- If the paper includes experiments, a No answer to this question will not be perceived well by the reviewers: Making the paper reproducible is important, regardless of whether the code and data are provided or not. +- If the contribution is a dataset and/or model, the authors should describe the steps taken to make their results reproducible or verifiable. +- Depending on the contribution, reproducibility can be accomplished in various ways. For example, if the contribution is a novel architecture, describing the architecture fully might suffice, or if the contribution is a specific model and empirical evaluation, it may be necessary to either make it possible for others to replicate the model with the same dataset, or provide access to the model. In general, releasing code and data is often one good way to accomplish this, but reproducibility can also be provided via detailed instructions for how to replicate the results, access to a hosted model (e.g., in the case of a large language model), releasing of a model checkpoint, or other means that are appropriate to the research performed. +- While NeurIPS does not require releasing code, the conference does require all submissions to provide some reasonable avenue for reproducibility, which may depend on the nature of the contribution. For example +(a) If the contribution is primarily a new algorithm, the paper should make it clear how to reproduce that algorithm. +(b) If the contribution is primarily a new model architecture, the paper should describe the architecture clearly and fully. +(c) If the contribution is a new model (e.g., a large language model), then there should either be a way to access this model for reproducing the results or a way to reproduce the model (e.g., with an open-source dataset or instructions for how to construct the dataset). +(d) We recognize that reproducibility may be tricky in some cases, in which case authors are welcome to describe the particular way they provide for reproducibility. In the case of closed-source models, it may be that access to the model is limited in some way (e.g., to registered users), but it should be possible for other researchers to have some path to reproducing or verifying the results. + +# 5. Open access to data and code + +Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material? + +Answer: [No] + +Justification: We have not released the code of this paper in the submission. + +Guidelines: + +- The answer NA means that paper does not include experiments requiring code. +- Please see the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details. +- While we encourage the release of code and data, we understand that this might not be possible, so "No" is an acceptable answer. Papers cannot be rejected simply for not including code, unless this is central to the contribution (e.g., for a new open-source benchmark). +- The instructions should contain the exact command and environment needed to run to reproduce the results. See the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details. +- The authors should provide instructions on data access and preparation, including how to access the raw data, preprocessed data, intermediate data, and generated data, etc. +- The authors should provide scripts to reproduce all experimental results for the new proposed method and baselines. If only a subset of experiments are reproducible, they should state which ones are omitted from the script and why. +- At submission time, to preserve anonymity, the authors should release anonymized versions (if applicable). +- Providing as much information as possible in supplemental material (appended to the paper) is recommended, but including URLs to data and code is permitted. + +# 6. Experimental setting/details + +Question: Does the paper specify all the training and test details (e.g., data splits, hyperparameters, how they were chosen, type of optimizer, etc.) necessary to understand the results? + +Answer: [Yes] + +Justification: We have specified all the training and test details of this paper in the experiment part. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The experimental setting should be presented in the core of the paper to a level of detail that is necessary to appreciate the results and make sense of them. +- The full details can be provided either with the code, in appendix, or as supplemental material. + +# 7. Experiment statistical significance + +Question: Does the paper report error bars suitably and correctly defined or other appropriate information about the statistical significance of the experiments? + +Answer: [Yes] + +Justification: We have reported appropriate information about the statistical significance of the experiments in the experiment part. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The authors should answer "Yes" if the results are accompanied by error bars, confidence intervals, or statistical significance tests, at least for the experiments that support the main claims of the paper. +- The factors of variability that the error bars are capturing should be clearly stated (for example, train/test split, initialization, random drawing of some parameter, or overall run with given experimental conditions). + +- The method for calculating the error bars should be explained (closed form formula, call to a library function, bootstrap, etc.) +- The assumptions made should be given (e.g., Normally distributed errors). +- It should be clear whether the error bar is the standard deviation or the standard error of the mean. +- It is OK to report 1-sigma error bars, but one should state it. The authors should preferably report a 2-sigma error bar than state that they have a $96\%$ CI, if the hypothesis of Normality of errors is not verified. +- For asymmetric distributions, the authors should be careful not to show in tables or figures symmetric error bars that would yield results that are out of range (e.g. negative error rates). +- If error bars are reported in tables or plots, The authors should explain in the text how they were calculated and reference the corresponding figures or tables in the text. + +# 8. Experiments compute resources + +Question: For each experiment, does the paper provide sufficient information on the computer resources (type of compute workers, memory, time of execution) needed to reproduce the experiments? + +Answer: [Yes] + +Justification: We have provided sufficient information on the computer resources of this method in the experiment part. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The paper should indicate the type of compute workers CPU or GPU, internal cluster, or cloud provider, including relevant memory and storage. +- The paper should provide the amount of compute required for each of the individual experimental runs as well as estimate the total compute. +- The paper should disclose whether the full research project required more compute than the experiments reported in the paper (e.g., preliminary or failed experiments that didn't make it into the paper). + +# 9. Code of ethics + +Question: Does the research conducted in the paper conform, in every respect, with the NeurIPS Code of Ethics https://neurips.cc/public/EthicsGuidelines? + +Answer: [Yes] + +Justification: We have conducted research under the NeurIPS Code Of Ethics. + +Guidelines: + +- The answer NA means that the authors have not reviewed the NeurIPS Code of Ethics. +- If the authors answer No, they should explain the special circumstances that require a deviation from the Code of Ethics. +- The authors should make sure to preserve anonymity (e.g., if there is a special consideration due to laws or regulations in their jurisdiction). + +# 10. Broader impacts + +Question: Does the paper discuss both potential positive societal impacts and negative societal impacts of the work performed? + +Answer: [Yes] + +Justification: We have discussed the societal impacts of this paper in the introduction and method parts. + +Guidelines: + +- The answer NA means that there is no societal impact of the work performed. +- If the authors answer NA or No, they should explain why their work has no societal impact or why the paper does not address societal impact. + +- Examples of negative societal impacts include potential malicious or unintended uses (e.g., disinformation, generating fake profiles, surveillance), fairness considerations (e.g., deployment of technologies that could make decisions that unfairly impact specific groups), privacy considerations, and security considerations. +- The conference expects that many papers will be foundational research and not tied to particular applications, let alone deployments. However, if there is a direct path to any negative applications, the authors should point it out. For example, it is legitimate to point out that an improvement in the quality of generative models could be used to generate deepfakes for disinformation. On the other hand, it is not needed to point out that a generic algorithm for optimizing neural networks could enable people to train models that generate Deepfakes faster. +- The authors should consider possible harms that could arise when the technology is being used as intended and functioning correctly, harms that could arise when the technology is being used as intended but gives incorrect results, and harms following from (intentional or unintentional) misuse of the technology. +- If there are negative societal impacts, the authors could also discuss possible mitigation strategies (e.g., gated release of models, providing defenses in addition to attacks, mechanisms for monitoring misuse, mechanisms to monitor how a system learns from feedback over time, improving the efficiency and accessibility of ML). + +# 11. Safeguards + +Question: Does the paper describe safeguards that have been put in place for responsible release of data or models that have a high risk for misuse (e.g., pretrained language models, image generators, or scraped datasets)? + +Answer: [NA] + +Justification: This paper has not released data or models that have a high risk for misuse. + +# Guidelines: + +- The answer NA means that the paper poses no such risks. +- Released models that have a high risk for misuse or dual-use should be released with necessary safeguards to allow for controlled use of the model, for example by requiring that users adhere to usage guidelines or restrictions to access the model or implementing safety filters. +- Datasets that have been scraped from the Internet could pose safety risks. The authors should describe how they avoided releasing unsafe images. +- We recognize that providing effective safeguards is challenging, and many papers do not require this, but we encourage authors to take this into account and make a best faith effort. + +# 12. Licenses for existing assets + +Question: Are the creators or original owners of assets (e.g., code, data, models), used in the paper, properly credited and are the license and terms of use explicitly mentioned and properly respected? + +Answer: [Yes] + +Justification: All the assets used in the paper are properly licensed. Our paper cites the original paper that produced the code package or dataset. + +# Guidelines: + +- The answer NA means that the paper does not use existing assets. +- The authors should cite the original paper that produced the code package or dataset. +- The authors should state which version of the asset is used and, if possible, include a URL. +- The name of the license (e.g., CC-BY 4.0) should be included for each asset. +- For scraped data from a particular source (e.g., website), the copyright and terms of service of that source should be provided. + +- If assets are released, the license, copyright information, and terms of use in the package should be provided. For popular datasets, paperswithcode.com/datasets has curated licenses for some datasets. Their licensing guide can help determine the license of a dataset. +- For existing datasets that are re-packaged, both the original license and the license of the derived asset (if it has changed) should be provided. +- If this information is not available online, the authors are encouraged to reach out to the asset's creators. + +# 13. New assets + +Question: Are new assets introduced in the paper well documented and is the documentation provided alongside the assets? + +Answer: [NA] + +Justification: Our paper does not release new assets, including data and models. We also did not publish or submit the code. + +Guidelines: + +- The answer NA means that the paper does not release new assets. +- Researchers should communicate the details of the dataset/code/model as part of their submissions via structured templates. This includes details about training, license, limitations, etc. +- The paper should discuss whether and how consent was obtained from people whose asset is used. +- At submission time, remember to anonymize your assets (if applicable). You can either create an anonymized URL or include an anonymized zip file. + +# 14. Crowdsourcing and research with human subjects + +Question: For crowdsourcing experiments and research with human subjects, does the paper include the full text of instructions given to participants and screenshots, if applicable, as well as details about compensation (if any)? + +Answer: [NA] + +Justification: The research topic of this paper is vision language action model, not involving crowdsourcing nor research with human subjects. + +Guidelines: + +- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. +- Including this information in the supplemental material is fine, but if the main contribution of the paper involves human subjects, then as much detail as possible should be included in the main paper. +- According to the NeurIPS Code of Ethics, workers involved in data collection, curation, or other labor should be paid at least the minimum wage in the country of the data collector. + +# 15. Institutional review board (IRB) approvals or equivalent for research with human subjects + +Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals (or an equivalent approval/review based on the requirements of your country or institution) were obtained? + +Answer: [NA] + +Justification: The research topic of this paper is vision language action model, not involving crowdsourcing nor research with human subjects. + +Guidelines: + +- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. + +- We recognize that the procedures for this may vary significantly between institutions and locations, and we expect authors to adhere to the NeurIPS Code of Ethics and the guidelines for their institution. +- For initial submissions, do not include any information that would break anonymity (if applicable), such as the institution conducting the review. + +# 16. Declaration of LLM usage + +Question: Does the paper describe the usage of LLMs if it is an important, original, or non-standard component of the core methods in this research? Note that if the LLM is used only for writing, editing, or formatting purposes and does not impact the core methodology, scientific rigorousness, or originality of the research, declaration is not required. + +Answer: [NA] + +Justification: LLM is used only for writing, editing, or formatting purposes and does not impact the core methodology, scientific rigorousness, or originality of the research + +Guidelines: + +- The answer NA means that the core method development in this research does not involve LLMs as any important, original, or non-standard components. +- Please refer to our LLM policy (https://neurips.cc/Conferences/2025/LLM) for what should or should not be described. + +# 7 Appendix section + +# 7.1 Comparison with ARM4R + +Our approach differs from ARM4R in both design and practice. Instead of representation learning with full 3D point clouds and proprioception, we adopt an end-to-end, action-centric pretraining pipeline that maps vision-language inputs directly to low-level actions. Inputs are RGB with depth-aligned spatial coordinates (2D patches + $(x,y,z)$ embeddings), keeping the model lightweight and VLM-compatible while avoiding point-cloud encoders. + +
MethodMeet-off-grillPush ButtonsPlace-wineOpen-drawerAvg.
C2FARM-BC [39]20.072.018.020.032.5
LLARVA [40]80.056.012.060.052.0
PerAct [41]84.048.012.080.056.0
ARM4R [32]94.467.236.088.871.9
4D-VLA(Ours)95.279.245.689.677.4
+ +Analysis. To balance efficiency and spatial coverage, we use our streaming memory bank sampling strategy that keeps 5 history frames within a 20-frame context. The current step uses dual views (front + wrist), while history stores only the front view. We also adopt OpenVLA's random-crop augmentation. This configuration provides strong spatial grounding and good runtime, offering a strategy different from ARM4R yet comparably effective on its benchmark. + +Note on supervision. Our method leverages calibrated intrinsics/extrinsics during training and inference. While such calibration is standard in many robotic datasets, it is an additional source of supervision not explicitly used by ARM4R. + +# 7.2 More ablation study + +Table 6: RLBench results under the ARM4R protocol. + +
Action headFPSSuccess rate
MLP12.629.5
auto regression0.628.1
diffusion8.727.0
PretrainCoord.Prop.Success rate
XXX29.5
XX✓16.4
X✓X36.6
✓✓X45.0
+ +Table 7: Ablations on heads and inputs (Libero-Long). Left: action head vs. FPS and success (MLP, autoregressive, diffusion). Right: effect of pretraining, 3D coordinate embedding, and proprioceptive tokens on success. + +
Sampling MethodSuccess ↑Latency (ms) ↓Peak Mem (MB) ↓
Single frame0.73876.58342.5
Adaptive Pooling [42]0.604150.78949.1
Grid Pooling [43]0.620208.88852.7
Q-Former (per-frame)0.556223.38812.8
MBS (Ours)0.866160.08682.9
+ +Table 8: Frame sampling ablations on Libero-Spatial. MBS attains the highest success (0.866) with competitive efficiency, while single-frame is fastest and most memory-light but less accurate. + +Action heads. We conduct additional ablations on LIBERO-LONG. For the action heads (Tab. 7, left), the MLP head yields the highest inference speed with a relatively high success rate. In contrast, the autoregressive head-predicting text tokens for actions—runs slower due to multi-token reasoning. + +Pretrain and spatial components. Results for other components/operations are shown in Tab. 7 (right). Proprioceptive tokens hurt performance, likely because proprioceptive states (e.g., joint positions, velocities) are highly correlated with action labels, which encourages overfitting to dataset-specific motion patterns rather than learning generalizable visuomotor features. + +Memory bank sampling. We compare our memory bank sampling with three video-style samplers: Adaptive Pooling [42], Grid Pooling [43], and a per frame $Q$ -Former, using InternVL2.5-4B on Libero-Spatial (20 epochs). We report success, latency, and peak memory in Tab. 8. Adaptive Pooling caches the full trajectory and uses two thresholds (0.99 for frame selection and 0.96 for spatial compression). We omit its second-stage cross-modal filter due to low language diversity (10 tasks) and observe hyperparameter sensitivity and variable-length tokens that increase memory and reduce accuracy. Grid Pooling replaces language tokens with 8 learnable queries and compresses each frame to a $2 \times 2$ grid (4 tokens), yielding $20 \times (4 + 8)$ historical tokens; it is simple but nonadaptive and can drop key spatial cues. $Q$ -Former (per frame) uses 10 learnable queries per frame to extract features but converges slowly and discards too much spatial context, which hurts success. MBS (ours) operates online with a window of 20 and a memory of 5. It selects informative key frames from the stream, stores only 5 history frames at $224 \times 224$ , and keeps the current frame at $448 \times 448$ . As a result, MBS achieves the highest success (0.866) with competitive cost (160.0 ms and 8,682.9 MB). Single frame is faster (76.5 ms) but weaker (0.738). The other baselines are slower and less accurate. MBS fits VLA's streaming and causal nature because it avoids full-clip caching, uses relative temporal encoding, and remains compatible with closed-loop control. It yields stronger long horizon performance without extra retraining or heavy compute. + +# 7.3 Coordinate system chaos impact analysis + +To assess the impact of coordinate system chaos on VLA's performance, we conduct a controlled experiment. The experiment consists of two main steps: First, we deliberately introduce controlled chaos into the LIBERO environment. Then, we compare the performance of models with and without 3D information. Our findings demonstrate that chaotic coordinate transformations significantly degrade model performance while incorporating 3D information effectively alleviates it. + +Chaos generation. To simulate the diverse viewpoints in the pretraining dataset—where the robot's body is absent from the image, and its coordinate system varies unpredictably—we select trajectories from a specific task in LIBERO-SPATIAL and render each trajectory from 30 distinct viewpoints. For each trajectory and its corresponding viewpoint, we introduce coordinate system chaos by applying a random translation $\mathbf{t} \in \mathbb{R}^3$ and rotation $\mathbf{q} \in \mathbb{S}\mathbb{O}(3)$ to the robot's coordinate frame. + +After applying the coordinate transformation, the gripper's grasping state in the ground-truth action remains unchanged. However, the rotation offset $\Delta \theta$ and translation $\Delta x$ , along with the proprioceptive and camera pose information, undergo the following transformation: + +![](images/d04310f5c26ca1c1237ae7080a76b410d57830888973d42d67544f50637542a3.jpg) +Figure 7: Success rates under varying coordinate chaos levels. + +$$ +\begin{array}{l} \Delta \boldsymbol {\theta} ^ {\prime} = \psi^ {- 1} (\mathbf {q} \psi (\Delta \boldsymbol {\theta}) \mathbf {q} ^ {\top}), \\ \boldsymbol {\theta} ^ {\prime} = \psi^ {- 1} (\mathbf {q} \psi (\boldsymbol {\theta})), \\ \mathbf {R} ^ {\prime} = \mathbf {q R}, \\ \Delta \boldsymbol {x} ^ {\prime} = \mathbf {q} \Delta \boldsymbol {x}, \\ \boldsymbol {x} ^ {\prime} = \mathbf {q} \boldsymbol {x} + \mathbf {t}, \\ \mathbf {T} ^ {\prime} = \mathbf {q} \mathbf {T} + \mathbf {t}. \tag {5} \\ \end{array} +$$ + +Here, the function $\psi : \mathbb{R}^3 \to \mathbb{SO}(3)$ maps an Euler angle to its corresponding rotation matrix. The terms $\Delta \theta'$ and $\Delta x'$ denote the transformed action values, while $\theta', \theta$ and $x'$ , $x$ represent the transformed and original rotation and position, respectively. Additionally, $[\mathbf{R}'|\mathbf{T}']$ and $[\mathbf{R}|\mathbf{T}]$ correspond to the transformed and original camera poses. + +In the subsequent training process, we utilize the transformed action values $\Delta \theta^{\prime}$ and $\Delta x^{\prime}$ , along with the transformed camera parameters, for model training. + +Implementation details. We employ a simple baseline, controlling for the involvement of 3D information. The baseline model extracts tokens from a single RGB view, while the alternative model converts an RGB-D frame into spatial vision tokens as input for the LLM Transformer. During testing, we do not apply random rotations or translations to the world coordinate system. + +Experimental results. We control chaos levels by adjusting the magnitude of random rotations. Level 0 applies no rotation, while levels 1-3 introduce random $z$ -axis rotations of $15^{\circ}$ , $30^{\circ}$ , and $90^{\circ}$ , respectively. Translation $t$ is randomly set within a range of 0.5. As shown in Fig. 7, without chaos, both models perform well, with 3D information further boosting success rates. Notably, the 3D model shows lower variance across viewpoints. As chaos increases, the non-3D model's performance drops sharply, while the 3D model maintains relatively high success—highlighting the value of 3D cues in handling coordinate system chaos. \ No newline at end of file diff --git a/NeurIPS/2025/4D-VLA_ Spatiotemporal Vision-Language-Action Pretraining with Cross-Scene Calibration/images.zip b/NeurIPS/2025/4D-VLA_ Spatiotemporal Vision-Language-Action Pretraining with Cross-Scene Calibration/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..d47290fd4956c26b30df23dfa6aa53d18698e087 --- /dev/null +++ b/NeurIPS/2025/4D-VLA_ Spatiotemporal Vision-Language-Action Pretraining with Cross-Scene Calibration/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:89a2b9787bee2c0744250a67267f031e973ee0173cbe288e1b7a425d354b7d84 +size 627035 diff --git a/NeurIPS/2025/4D-VLA_ Spatiotemporal Vision-Language-Action Pretraining with Cross-Scene Calibration/layout.json b/NeurIPS/2025/4D-VLA_ Spatiotemporal Vision-Language-Action Pretraining with Cross-Scene Calibration/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..79e8a274c156d3678b2106995ba126b07b763214 --- /dev/null +++ b/NeurIPS/2025/4D-VLA_ Spatiotemporal Vision-Language-Action Pretraining with Cross-Scene Calibration/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ad82de33b68899033111a0c6c3a39ef8064c8bf18b6887dacd6e2686a255ab8b +size 782142 diff --git a/NeurIPS/2025/4D3R_ Motion-Aware Neural Reconstruction and Rendering of Dynamic Scenes from Monocular Videos/5eff43d6-ba58-4f66-b9f4-0ea4774effc5_content_list.json b/NeurIPS/2025/4D3R_ Motion-Aware Neural Reconstruction and Rendering of Dynamic Scenes from Monocular Videos/5eff43d6-ba58-4f66-b9f4-0ea4774effc5_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..c84d63061b349a76b30fb461199021fdb7f853fa --- /dev/null +++ b/NeurIPS/2025/4D3R_ Motion-Aware Neural Reconstruction and Rendering of Dynamic Scenes from Monocular Videos/5eff43d6-ba58-4f66-b9f4-0ea4774effc5_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7da8483305abd88101a84a0e0211f47b5b93b787ccc0d349bc750eec8b53ad1f +size 141321 diff --git a/NeurIPS/2025/4D3R_ Motion-Aware Neural Reconstruction and Rendering of Dynamic Scenes from Monocular Videos/5eff43d6-ba58-4f66-b9f4-0ea4774effc5_model.json b/NeurIPS/2025/4D3R_ Motion-Aware Neural Reconstruction and Rendering of Dynamic Scenes from Monocular Videos/5eff43d6-ba58-4f66-b9f4-0ea4774effc5_model.json new file mode 100644 index 0000000000000000000000000000000000000000..2b671f3e724f06a463f2dd21f490730638b8cdf0 --- /dev/null +++ b/NeurIPS/2025/4D3R_ Motion-Aware Neural Reconstruction and Rendering of Dynamic Scenes from Monocular Videos/5eff43d6-ba58-4f66-b9f4-0ea4774effc5_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:90108a2a9b84b4fba199b2f709e9e8e78902a711265c68e54b33ed16bebfef7a +size 189623 diff --git a/NeurIPS/2025/4D3R_ Motion-Aware Neural Reconstruction and Rendering of Dynamic Scenes from Monocular Videos/5eff43d6-ba58-4f66-b9f4-0ea4774effc5_origin.pdf b/NeurIPS/2025/4D3R_ Motion-Aware Neural Reconstruction and Rendering of Dynamic Scenes from Monocular Videos/5eff43d6-ba58-4f66-b9f4-0ea4774effc5_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..91d4a504776c9f1ff00947dc68effdddc83df07b --- /dev/null +++ b/NeurIPS/2025/4D3R_ Motion-Aware Neural Reconstruction and Rendering of Dynamic Scenes from Monocular Videos/5eff43d6-ba58-4f66-b9f4-0ea4774effc5_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b288abe3484f854db63606c5944ed64ffee8f02d3fbb7881799060aa27c82ade +size 5691610 diff --git a/NeurIPS/2025/4D3R_ Motion-Aware Neural Reconstruction and Rendering of Dynamic Scenes from Monocular Videos/full.md b/NeurIPS/2025/4D3R_ Motion-Aware Neural Reconstruction and Rendering of Dynamic Scenes from Monocular Videos/full.md new file mode 100644 index 0000000000000000000000000000000000000000..3680baed641d5924d9c2cff8d536790dbb5e662d --- /dev/null +++ b/NeurIPS/2025/4D3R_ Motion-Aware Neural Reconstruction and Rendering of Dynamic Scenes from Monocular Videos/full.md @@ -0,0 +1,714 @@ +# 4D3R: Motion-Aware Neural Reconstruction and Rendering of Dynamic Scenes from Monocular Videos + +Mengqi Guo $^{1}$ , Bo Xu $^{2}$ , Yanyan Li $^{1}$ , Gim Hee Lee $^{1}$ + +$^{1}$ National University of Singapore + +$^{2}$ Wuhan University + +{mengqi.guo, gimhee.lee}@comp.nus.edu.sg + +![](images/76438a1f766521ba2796fa5a8dc459d84bcebc0dd5275da3cc907efa7d0660c7.jpg) +Figure 1: Overview of our pose-free 4D Gaussian Splitting. Given a monocular video sequence of a dynamic scene (left), our method directly reconstructs the 4D scene without pre-computed camera poses (right). Dynamic control points guide the deformation of Gaussian points to model motion, producing high-quality novel views across different time steps. + +# Abstract + +Novel view synthesis from monocular videos of dynamic scenes with unknown camera poses remains a fundamental challenge in computer vision and graphics. While recent advances in 3D representations such as Neural Radiance Fields (NeRF) and 3D Gaussian Splatting (3DGS) have shown promising results for static scenes, they struggle with dynamic content and typically rely on pre-computed camera poses. We present 4D3R, a pose-free dynamic neural rendering framework that decouples static and dynamic components through a two-stage approach. Our method first leverages 3D foundational models for initial pose and geometry estimation, followed by motion-aware refinement. 4D3R introduces two key technical innovations: (1) a motion-aware bundle adjustment (MA-BA) module that combines transformer-based learned priors with SAM2 for robust dynamic object segmentation, enabling more accurate camera pose refinement; and (2) an efficient Motion-Aware Gaussian Splatting (MA-GS) representation that uses control points with a deformation field MLP and linear blend skinning to model dynamic motion, significantly reducing computational cost while maintaining high-quality reconstruction. Extensive experiments on real-world dynamic datasets demonstrate that our approach achieves up to 1.8dB PSNR improvement over state-of-the-art methods, particularly in challenging scenarios with large dynamic objects, while reducing computational requirements by $5\times$ compared to previous dynamic scene representations. + +# 1 Introduction + +Novel view rendering from monocular videos of dynamic scenes remains a fundamental challenge in both computer graphics and computer vision communities. While static scene reconstruction has seen significant advances with methods like 3D Gaussian Splatting (3DGS) [25, 58, 29] and Neural Radiance Fields (NeRF) [38, 20, 3, 1, 2], dynamic scenes present substantially greater challenges. Unlike static environments, dynamic scenes require modeling intricate 3D environments with temporal coherence while handling complex camera viewpoints. These approaches leverage Gaussian primitives or neural layers but face severe challenges in capturing scene evolution with high fidelity, especially when addressing substantial changes in handling dynamic content such as managing motion, ensuring temporal consistency, and maintaining efficient scene representations. + +Adapting 3DGS to dynamic scenes has led to the development of various 4D-GS approaches [60, 22] that incorporate deformation fields modeled by multi-layer perceptrons (MLPs), motion bases, or 4D representations. Despite achieving quantitative improvements, these methods typically rely on precomputed camera poses from multi-view systems or Structure-from-Motion pipelines, which often fail in scenes with significant dynamic content. This dependency on ground truth or pre-estimated camera poses severely limits their applicability in real-world environments. + +Estimation of 6-DoF camera poses typically involves establishing 2D-3D correspondences followed by solving the Perspective-n-Point (PnP) [19] problem with RANSAC [14]. Approaches for predicting 2D-3D correspondences can be broadly categorized into two main directions: Structure-from-Motion (SfM) methods such as COLMAP [47, 48], and scene coordinate regression (SCR) [49]. SfM methods detect and describe key points in 2D images [53, 11], linking them to corresponding 3D coordinates [34, 46]. Although these methods are effective, they still face challenges including high computational overhead, significant storage requirements, and potential privacy concerns when processing sensitive data [51]. In contrast, SCR methods [49, 5, 56, 71] utilize deep neural networks (DNNs) to directly predict the 3D coordinates of the image pixels, followed by running PnP with RANSAC for camera pose estimation. These approaches offer notable advantages such as higher accuracy in smaller scenes, reduced training times, and minimizing storage requirements. Taking advantage of these benefits, this paper adopts SCR over traditional SfM methods for camera pose estimation. Furthermore, DUSt3R [56] employs a Vision Transformer (ViT)-based architecture to predict 3D coordinates using a data-driven approach and in the following work, MonST3R [71] extends DUSt3R to dynamic scenes by fine-tuning the model on suitable dynamic datasets. However, treating pose estimation and scene reconstruction as separate tasks in dynamic environments typically leads to suboptimal performance. + +A straightforward approach to addressing the pose-free novel view synthesis problem is to directly combine MonST3R [71] with 4D-GS methods [22]. However, the accuracy of camera poses predicted by MonST3R is not sufficiently stable, causing 4D-GS methods to struggle with reconstructing accurate scenes and resulting in poor rendering quality. Particularly, these methods often fail in scenarios involving moving objects that occupy a significant portion of the image. Since correspondences between 2D keypoints and 3D points are established based on static scene elements, dynamic objects are commonly deemed to be outliers during the RANSAC process. This assumption breaks down in the presence of large or dominant moving objects, further degrading performance in dynamic environments. + +To overcome these issues, this paper proposes a novel architecture for pose-free dynamic Gaussian Splitting that integrates transformer-based motion priors for initial pose estimation and then refines it using a Motion-Aware Bundle Adjustment (MA-BA) module. Our key insight is that the motion mask and scene reconstruction should be jointly optimized rather than treated as isolated processes, allowing for more accurate camera pose estimation and higher-quality novel view synthesis. Specifically, the ViT-based transformer gives the initial dynamic mask. We sample the top-K values and turn their location into K-point prompts for pretrained SAM2 [43]. Finally, the final dynamic mask is the combination of the output dynamic object segments from the SAM2 and the confidence map from the transformer. The dynamic mask serves as a static point selection when performing RANSAC, which can reduce the noise introduced by the dynamic points. + +For 4D-GS, the expensive computational cost is a huge burden since millions of GS points need to learn a set of motion parameters. Some works design compact representations to solve this problem, such as the sparse motion basis [23], sparse-control points [22], and k-plane [60]. We design our 4D + +![](images/0584d73b07133f6a1fc579736c184e291aa9b16124f9babf831e4659e7f46f5a.jpg) +Figure 2: Overview of our motion-aware 4D gaussian splatting pipeline. Our framework consists of three main modules: (1) A 4D-aware information extractor that processes input frames through parallel ViT encoders and decoders to extract geometric and motion information; (2) A motion-aware bundle adjustment module that leverages motion predictions for robust camera estimation; and (3) A motion-aware gaussian splatting module that enables dynamic scene modeling through adaptive control points. + +representation of Motion-Aware Gaussian Splatting (MA-GS) based on motion masks generated from the MA-BA module. Specifically, we model the dynamic motion by hundreds of control points with a deformation field MLP, and Gaussian points are using Linear Blend Skinning (LBS). The model first trains the control points for the dynamic part and then trains the GS points. + +We show the effectiveness of our approach through extensive experiments on both synthetic and real-world dynamic scenes. Our findings demonstrate notable gains in pose estimation accuracy and reconstruction quality over current methods. In particular, we outperform state-of-the-art techniques on difficult dynamic scenes by achieving improvements of 1.8 in PSNR in novel view rendering quality and more accurate pose estimation. + +Our key contributions include: 1) We propose a novel motion-aware pipeline that fundamentally integrates pose estimation with scene reconstruction, eliminating the traditional separation that causes failures in highly dynamic scenes. 2) We introduce a theoretical framework for motion-aware bundle adjustment that jointly optimizes camera poses and scene representation, enabling robust performance even when moving objects dominate the scene. 3) We design a compact and efficient 4D representation using motion-aware gaussian splatting that significantly reduces memory requirements while maintaining rendering quality. 4) Our approach demonstrates state-of-the-art performance on challenging dynamic scenes without requiring pre-computed poses, enabling truly monocular novel view synthesis. + +# 2 Related Work + +# 2.1 Static Scene Novel View Rendering. + +Novel View Synthesis aims to generate novel viewpoints from a set of observations. Recently, neural implicit representations have shown impressive capabilities. NeRF [38] achieved breakthrough results by representing scenes through MLPs. Subsequent work focused on acceleration through methods like baking [21] and explicit representations [39]. 3DGS [25] introduced efficient rasterization of anisotropic 3D Gaussians, enabling real-time rendering without quality degradation. Recent extensions have ireal-time rendering [68, 44], camera modeling [61], faster training [35, 16, 39, 7], and sparse view [69, 45]. However, these methods assume static scenes and known camera parameters, limiting their practical applications. + +# 2.2 Dynamic Scene Novel View Rendering. + +Research has expanded to capture both motion and geometry in dynamic scenes [64, 37, 66, 31, 15, 32, 30]. Initial approaches [42, 40] learned additional time-varying deformation fields. Alternative methods [28, 16, 62] encode scene dynamics through multi-dimensional feature fields without explicit motion modeling. Following 3DGS, recent work [60] proposes learning individual Gaussian trajectories over time. More efficient representations have emerged, including factorized motion bases [27] and sparse control points [22]. Another approach [65] extends spherical harmonics to 4D. As noted in Dycheck [18], many methods focus on unrealistic scenarios, while real-world capture involves substantial motion. To resolve motion ambiguity, recent work leverages pretrained depth estimation [63] or trajectory tracking [24]. Our approach utilizes DUSt3R [56] for initialization and incorporates SAM2 [43] for dynamic motion segmentation. + +# 2.3 Pose-Free Neural Field. + +Traditional NVS relies heavily on SfM [47] for camera parameters. Recent research explores optimizing neural fields without pre-calibrated poses. iNeRF[67] estimates camera poses from pre-trained NeRF through photometric optimization. NeRF- [59] jointly optimize camera and scene parameters with geometric regularization. BARF [33] and GARF [10] address gradient inconsistency in positional embeddings. Nope-NeRF [4] leverages geometric priors for accurate camera estimation. In the 3DGS domain, CF3DGS [17] introduces progressive optimization, while InstantSplat [12] uses DUSt3R [56] for initialization but remains limited to static scenes. ZeroGS [9] utilizes the DUSt3R and progressive image registration for pose-free 3D GS. Our approach differs by introducing a pose-free pipeline for dynamic scenes that decouples static backgrounds from dynamic objects. We utilize DUSt3R's geometric foundation model and leverage 3DGS's explicit nature for enhanced geometric regularization. + +# 3 Method + +# 3.1 Preliminaries + +3D Gaussian splatting represents scenes using a collection of colored 3D Gaussian primitives. Each Gaussian $G_{j}$ is characterized by its center position $\pmb{\mu}_{j}$ , covariance matrix $\Sigma_{j}$ (parameterized by rotation quaternion $\mathbf{q}_j$ and scaling vector $\mathbf{s}_j$ ), opacity value $\sigma_{j}$ , and spherical harmonic coefficients $\mathbf{sh}_j$ for view-dependent appearance. The scene representation is thus $\mathcal{G} = \{G_j:\pmb {\mu}_j,\mathbf{q}_j,\mathbf{s}_j,\sigma_j,\mathbf{sh}_j\}$ . + +During rendering, these 3D Gaussians are projected onto the image plane with transformed 2D covariance matrices $\pmb{\Sigma}'$ . The final color at each pixel is computed through $\alpha$ -blending: + +$$ +C (\mathbf {u}) = \sum_ {i \in \mathcal {N}} T _ {i} \alpha_ {i} \operatorname {S H} \left(\mathbf {s h} _ {i}, \mathbf {v} _ {i}\right), \quad \text {w h e r e} T _ {i} = \prod_ {j = 1} ^ {i - 1} (1 - \alpha_ {j}) \tag {1} +$$ + +Our framework extends Gaussian Splatting to dynamic scenes with sparse control while maintaining computational efficiency and rendering quality. For more details, please refer to the supplementary. + +# 3.2 Problem Definition and Overview + +Given a monocular video sequence $\mathcal{V} = \{I_t\}_{t=1}^{\top}$ capturing a dynamic scene with moving objects and camera motion, our goal is to reconstruct a complete 4D representation of the scene. Our pipeline estimates the following parameters: 1) Camera parameters $\mathcal{C}_t = \{K_t, R_t, T_t\}$ , where $K_t \in \mathbb{R}^{3 \times 3}$ represents the intrinsic matrix, and $[R_t \mid T_t] \in SE(3)$ denotes the extrinsic parameters. 2) Dense depth map $D_t \in \mathbb{R}^{H \times W}$ and motion map $M_t \in \{0,1\}^{H \times W}$ capturing per-pixel information. 3) Dynamic scene representation through motion-aware Gaussian Splatting parameters $\mathcal{G}$ , motion-aware control points $\mathbb{P}$ , and a deformation field MLP $\Theta$ . + +As illustrated in Fig 2, our framework combines implicit geometric estimation and explicit motion understanding to address the unique challenges of dynamic scene reconstruction from monocular video through three primary components: 4D-aware information extractor, Motion-Aware Bundle Adjustment (MA-BA) and Motion-Aware Gaussian Splatting (MA-GS) representation. + +# 3.3 4D-Aware Information Extraction + +Our 4D-aware information extractor serves as the foundation for both camera pose estimation and scene reconstruction by leveraging pre-trained vision models to extract geometric and motion information from monocular inputs. To have a good initialization of the Gaussian splatting, we first employ a pre-trained ViT-based transformer model from MonST3R [71] that sequentially processes each input frame $I_{t}$ through an encoder-decoder block that extracts deep features to give a scene coordinate map $X_{t} \in \mathbb{R}^{H \times W \times 3}$ representing the 3D structure and a confidence map $W_{t} \in \mathbb{R}^{H \times W}$ indicating the reliability of the predictions. We also include the optical flow from SEA-RAFT [57]. + +Using the scene coordinate map $X_{t}$ and confidence map $W_{t}$ , we obtain high-confidence points $S$ through a principled two-step filtering process: + +$$ +\mathcal {S} = \left\{p _ {i} \mid W _ {t} \left(p _ {i}\right) > \tau_ {c} \wedge D _ {t} \left(p _ {i}\right) < \tau_ {d} \right\}, \tag {2} +$$ + +where $\tau_{c}$ and $\tau_{d}$ are the confidence and depth thresholds, respectively. This filtering strategy is based on two key insights: 1) The selection of points with high confidence scores are more likely to yield reliable geometric estimates. 2) Filtering out points at infinity with zero disparity that give unreliable depth estimates. + +Unlike previous approaches that rely solely on motion estimators, we leverage SAM2 [43] semantic understanding ability to generate pixel-precise dynamic object segmentation $M_{t} \in \{0,1\}^{H \times W}$ , with high-confidence points $S$ as prompts. This method critically enables our motion-aware processing pipeline to handle scenes dominated by dynamic content. + +# 3.4 Motion-Aware Bundle Adjustment + +Our MA-BA module introduces an approach to camera pose estimation that explicitly models the separation between static and dynamic scene components, addressing a fundamental limitation in traditional bundle adjustment methods. We leverage the dynamic region mask $M_{t}$ to enhance the accuracy of camera pose estimation. For frame pairs $(I_{t}, I_{t'})$ , we introduce a masked PnP-RANSAC approach that focuses solely on static regions: + +$$ +\mathcal {P} _ {\text {s t a t i c}} = \left\{p _ {i} \in \mathcal {S} \mid M _ {t} \left(p _ {i}\right) = 0 \right\}. \tag {3} +$$ + +By restricting correspondence matching to static regions, we significantly reduce the likelihood of incorrect matches due to dynamic objects. The optimization objective becomes: + +$$ +E \left(R _ {t}, T _ {t}\right) = \sum_ {p _ {i} \in \mathcal {P} _ {\text {s t a t i c}}} \| \Pi \left(R _ {t} p _ {i} + T _ {t}\right) - p _ {i} ^ {\prime} \| ^ {2}. \tag {4} +$$ + +where $\Pi (\cdot)$ is the camera model mapping a set of 3D points onto the image. + +We further refine the camera poses through Differentiable Dense Bundle Adjustment (DBA) layer [54]. For more details, please refer to the supplementary. + +# 3.5 Motion-Aware Gaussian Splitting + +Our Motion-Aware Gaussian Splitting (MA-GS) module introduces a principled approach to dynamic scene representation that significantly reduces the parameter space by focusing computational resources on regions requiring deformation modeling. We adopt a set of control points $\mathcal{P} = (p_i\in \mathbb{R}^3,\sigma_i\in \mathbb{R}^+)_{i = 1}^{N_P}$ , where $p_i$ represents the 3D coordinate in the canonical space and $\sigma_{i}$ defines the radius of the Radial Basis Function (RBF) kernel. These control points are initialized from our motion map $M_{t}$ , where the static and dynamic regions are handled distinctly through our MA-GS module. + +In the first stage, we optimize the control points on the dynamic regions with the following loss: + +$$ +L _ {\text {c o n t r o l}} = \sum_ {p _ {i} \in \mathcal {P}} M _ {t} \left(p _ {i}\right) L _ {\text {r e n d e r}} \left(p _ {i}\right), \tag {5} +$$ + +where $M_{t}(p_{i})$ acts as a binary mask to selectively optimize only the control points in dynamic regions and $L_{render}$ refers to the standard photometric loss between rendered and ground truth pixels, using L1 and DSSIM metrics. The equation selectively applies this loss only to dynamic regions via the + +Mt(pi) binary mask. This targeted optimization significantly reduces the complexity by focusing only on regions that require deformation modeling. For dynamic control points, we learn time-varying transformations through a specialized mapping function: + +$$ +\Theta : \left(p _ {t}, t\right)\rightarrow \left(R _ {t} ^ {k}, T _ {t} ^ {k}\right), \tag {6} +$$ + +where $[R_t^k\mid T_t^k ]\in SE(3)$ represents the six-degrees-of-freedom transformation at time $t$ $\Theta$ is the deformation field MLP. For numerical stability and continuous interpolation, we represent rotations using unit quaternions $r_t^k\in \mathbb{H}$ , where $\mathbb{H}$ denotes the space of unit quaternions. The dynamic scene rendering process utilizes these control points through weighted transformation blending. + +In the second stage, we optimize the Gaussian parameters $G_{j} = \{\pmb{\mu}_{j},\mathbf{q}_{j},\mathbf{s}_{j},\sigma_{j},\mathbf{sh}_{j}\}$ by applying motion-aware constraints through a detached gradient path: + +$$ +\mu_ {j} ^ {\prime} = \left\{ \begin{array}{l l} \mu_ {j} & \text {i f} M _ {t} (\mu_ {j}) = 0 \\ \sum_ {k \in \mathcal {N} _ {j}} w _ {j k} \left(R _ {t} ^ {k} \left(\mu_ {j} - p _ {k}\right) + p _ {k} + T _ {t} ^ {k}\right) & \text {o t h e r w i s e ,} \end{array} \right. \tag {7} +$$ + +where $w_{ij}$ is Linear Blend Skinning (LBS) weight [52]. Crucially, this constraint is implemented with gradient detachment to ensure the motion-aware transformation does not affect the parameter updates of the MLP and vice versa. This prevents competing optimization objectives between Gaussian parameters and deformation field parameters, leading to more stable convergence and better results. The LBS weights are computed with a normalized exponential kernel: + +$$ +w _ {j k} = \tilde {w} _ {j k} / \sum_ {k \in \mathcal {N} j} \tilde {w} _ {j k}, \quad \text {w h e r e} \tilde {w} _ {j k} = \exp \left(- d _ {j k} ^ {2} / 2 \sigma_ {k} ^ {2}\right). \tag {8} +$$ + +$d_{jk}$ represents the Euclidean distance between Gaussian center $\mu_j$ and control point $p_k$ , and $\mathcal{N}_j$ denotes the set of $K$ -nearest neighboring control points for Gaussian $j$ . + +For orientation updates, we employ quaternion blending to ensure smooth rotational deformation: + +$$ +q _ {j} ^ {\prime} = \left(\sum_ {k \in \mathcal {N} j} w _ {j k} r _ {t} ^ {k}\right) \otimes q _ {j}, \tag {9} +$$ + +where $\otimes$ denotes quaternion multiplication. This formulation ensures the entire dynamic scene experiences smooth and physically realistic deformations while maintaining computational efficiency through our motion-aware design. + +# 3.6 Training Strategy + +As mentioned in the previous section, our training process optimizes the entire scene representation through the two-stage procedure. Specifically, we optimize for the control points in the first stage with $L_{control}$ . The second stage of optimization for the motion-aware Gaussian parameters is achieved in the rendering loss $L_{render}$ using L1 distance and DSSIM metrics as follows: + +$$ +L = L _ {\text {r e n d e r}} + \lambda_ {\text {a r a p}} L _ {\text {a r a p}} + \lambda_ {\text {r i g i d}} L _ {\text {r i g i d}}, \tag {10} +$$ + +where $L_{\text{arap}}$ enforces local rigidity with as-rigid-as-possible regularization [50]: + +$$ +L _ {a r a p} \left(p _ {i}, t _ {1}, t _ {2}\right) = \sum w _ {i k} \| \left|\left(p _ {i} ^ {t _ {1}} - p _ {k} ^ {t _ {1}}\right) - R _ {i} \left(p _ {i} ^ {t _ {2}} - p _ {k} ^ {t _ {2}}\right)\right|\left. \right| ^ {2}, \tag {11} +$$ + +and $L_{\text{rigid}}$ enforces rigidity in static regions: + +$$ +L _ {r i g i d} = \sum_ {j} \left(1 - M _ {t} \left(\mu_ {j}\right)\right) \left\| \mu_ {j} ^ {\prime} - \mu_ {j} \right\| ^ {2}. \tag {12} +$$ + +We employ an adaptive control point strategy that optimizes point distribution based on reconstruction impact. We compute the gradient magnitude of the rendering loss with respect to Gaussian positions, weighted by their influence radius: + +$$ +g _ {k} = \sum \tilde {w} _ {j} \left\| \frac {\partial L}{\partial \mu_ {j}} \right\| ^ {2}, \tag {13} +$$ + +where $\tilde{w}_j$ represents the LBS weights connecting Gaussian points to control points, while $\frac{\partial L}{\partial \mu_j}$ is the gradient of the loss with respect to Gaussian positions. This gradient magnitude sum $(g_k)$ measures each control point's influence on reconstruction quality. During optimization, we add control points in areas with high $g_k$ , adaptively refining the representation where needed most. This adaptive approach ensures effective representation while maintaining high reconstruction quality across diverse dynamic scenes with varying motion complexity. + +Table 1: Quantitative results on HyperNeRF's [41] dataset. The best and the second best results are denoted by pink and yellow. The rendering resolution is $960 \times 540$ . "Time" in the table stands for training times plus camera pose estimation time. + +
ModelCOLMAPPSNR(dB)↑MS-SSIM↑Times↓FPS↑Storage(MB)↓
Nerfies [40]✓22.20.803~ hours< 1-
HyperNeRF [41]✓22.40.81432 hours< 1-
TiNeuVox-B [13]✓24.30.8363.5 hours148
3D-GS [25]✓19.70.6804 hours5552
FDNeRF [70]✓24.20.842-0.05440
4DGS [60]✓25.20.8455 hours3461
SC-GS [22]✓25.30.8414 hours4585
MonST3R+SC-GS×20.40.6972 hours45153
RoDynRF [36]×23.80.82028 hours<1200
Ours×25.60.84450 mins4580
+ +Table 2: Quantitative results on DyNeRF's [28] dataset. The best and the second best results are denoted by pink and yellow. "Time" in the table stands for training times plus pose estimation time. + +
ModelCOLMAPPSNR(dB)↑MS-SSIM↑Times↓FPS↑Storage(MB)↓
HyperNeRF [41]✓16.90.63832 hours< 1-
TiNeuVox-B [13]✓18.20.7123.5 hours148
3D-GS [25]✓15.30.5414 hours5552
4DGS [60]✓18.90.7405 hours3461
SC-GS [22]✓19.00.7464 hours4585
MonST3R+SC-GS×16.40.6242 hours45153
RoDynRF [36]×17.80.62028 hours<1200
Ours×19.60.75550 mins4580
+ +# 4 Experiments + +# 4.1 Experimental Settings + +Datasets. We evaluate our approach on three representative datasets: HyperNeRF's dataset [41], DyNeRF dataset [28], and the MPI Sintel dataset [6]. The DyNeRF dataset features controlled dynamic scenes captured with synchronized cameras, offering a solid baseline for evaluation. We use only one camera view for training, treating it as a monocular sequence. The HyperNeRF dataset presents more challenging scenarios with complex object deformations and camera movements. The MPI Sintel dataset provides ground truth camera poses, enabling quantitative evaluation of our pose estimation accuracy. + +Evaluation Metrics We evaluate using standard metrics for Novel View Synthesis: Peak Signal-to-Noise Ratio (PSNR), Structural Similarity (SSIM), and Multi-Scale SSIM (MS-SSIM). For camera pose estimation, we report the same metrics as [8]: Absolute Translation Error (ATE), Relative Translation Error (RPE trans), and Relative Rotation Error (RPE rot), after applying a Sim(3) Umayama alignment on prediction to the ground truth. + +Baselines. For a comprehensive evaluation, we compare our approach against state-of-the-art methods in both novel view rendering and pose estimation. For novel view rendering, we consider methods designed for static scenes like 3DGS [25], and dynamic scene methods including Nerfies [40], HyperNeRF [41], TiNeuVox [13], 4DGS [60], FDNeRF [70], and SC-GS [22], which represent the current state-of-the-art in dynamic scene modeling. We further compare with pose-free 4D novel view rendering baselines RoDynRF [36] and a strong baseline combining MonST3R [71] (one of the best dynamic pose estimation methods) with SC-GS [22] (one of the best dynamic scene modeling methods). For pose estimation evaluation, we compare against established methods such as DROID-SLAM [54], DPVO [55], and ParticleSFM [73], noting that these methods require camera intrinsics as input. + +Implementation Details. Our implementation uses a ViT-based transformer for 4D information extraction, pre-trained on DUSt3R [56] and fine-tuned on MonST3R [71] datasets. For dynamic segmentation refinement, we employ SAM2 [43] with multi-point prompting. The Motion-Aware + +![](images/760255ace5b7c873a05425089b7fde53833614fc1c3727c2a63f723bf1445605.jpg) +(a) MonST3R Mask + +![](images/9e8e150cfe40c5bcf290ae80640bb3e765f426040f5ba1ce2de1a9f392f8ad9d.jpg) +(b) Input Image + +![](images/39c3106abb10bb9e6c757721df3a2d15eb56394c0153b1310240834401e61334.jpg) +(c) Depth Map +Figure 3: Our motion mask refinement pipeline: (a) Initial dynamic mask from MonST3R showing coarse segmentation, (b) Input image of tomato cutting scene, (c) Estimated depth map highlighting object boundaries, (d) Confidence map indicating regions of dynamic motion, (e) Strategically sampled points for mask refinement, and (f) Final refined mask after SAM2 processing showing improved object boundary delineation. The pipeline effectively captures the dynamic nature of the cutting motion while maintaining precise object boundaries. + +![](images/c65b538000e25edb77e4ac84b69c35d698b9b4aa00c4489e2c6a25768bb58445.jpg) +(d) Confidence + +![](images/5da93db6d190266527006461b68efef4ef106c7ec6eb102ed506e81d27eb8e36.jpg) +(e) Sampled Points + +![](images/232936b17335abc86f105ff26a559226e2a64c800a98a8ae9e83cf9cf5ecec5f.jpg) +(f)Refined Mask + +Table 3: Ablation study of our proposed modules on HyperNeRF dataset. + +
MethodPSNR(dB)↑MS-SSIM↑Times↓
w/o motion-map20.40.6972 hours
w/o SAM-refine23.80.7651 hours
w/o MA-GS23.50.7341.5 hours
Ours25.60.84450mins
+ +Gaussian Splatting uses 512 control points, optimized using Adam with learning rates from 1e-4 to 1e-7 (exponential decay). All experiments run on a single NVIDIA RTX3090 GPU. + +# 4.2 Results on Novel View Rendering + +A key advantage of our approach is its superior performance on challenging scenes with dynamic objects. On the DyNeRF dataset (Tab 2), we achieve state-of-the-art results with a PSNR of 19.6dB and MS-SSIM of 0.775, outperforming existing methods regardless of their reliance on known camera poses. This superior performance comes from our motion-aware components, which effectively handle dynamic objects while maintaining accurate camera pose estimation. + +Our method demonstrates exceptional computational efficiency, achieving $5 \times$ faster training time compared to COLMAP-dependent methods while maintaining comparable quality. On the HyperNeRF dataset (Tab 1 and Fig 4), we achieve results (PSNR of 25.6dB and MS-SSIM of 0.844) competitive with SC-GS (25.3dB/0.841) and 4DGS (25.2dB/0.845), while significantly outpacing other COLMAP-free approaches like RoDynRF (23.8dB/0.820) and MonST3R+SC-GS (20.4dB/0.697). + +Furthermore, our approach excels in resource utilization, maintaining a competitive 45 FPS during inference while requiring only 80MB of memory. This is substantially more efficient than competing methods such as MonST3R+SC-GS (153MB) and RoDynRF (200MB). The efficiency stems from our compact motion-aware representation and efficient motion mask generation pipeline, as illustrated in Fig 2. + +# 5 Ablation Studies + +To validate the effectiveness of our key components, we conduct comprehensive ablation studies shown in Tab 3: + +Motion-aware Map Removing the motion-aware map leads to a significant performance drop of 5.2dB in terms of PSNR. This substantial drop confirms our theoretical insight that accurate dynamic-static decomposition is fundamental for handling scenes with large moving objects, addressing the limitation of previous methods like MonST3R that assume dynamic objects occupy only a small portion of the scene. + +![](images/840b6dc19f26f6177033da34cb3455390910ee3a3b7359af01975dc9aa331369.jpg) +Figure 4: Qualitative comparison with baselines. + +SAM-based Refinement Without our SAM2-based refinement module, performance decreases to 23.8dB/0.765. The 1.8dB performance gap validates our hypothesis that combining transformer-based motion priors with foundation model segmentation creates a synergistic effect, as shown in Fig 3. The refined masks enable more reliable static point selection during RANSAC, reducing noise from dynamic regions. + +Motion-aware Gaussian Splatting (MA-GS) Excluding MA-GS results in degraded performance (23.5dB/0.734) while increasing training time by 2x. These results confirm our theoretical prediction that focusing computational resources on motion-significant regions through our control point mechanism substantially improves both efficiency and quality. The improved efficiency (50 mins vs 1.5 hours) demonstrates that our compact representation successfully reduces the search space compared to methods requiring optimization of motion parameters for all Gaussian points. + +# 6 Limitation and Broader Impact + +Despite our method's improvements, limitations exist: it requires textured frames with sufficient static regions, assumes distinguishable dynamic objects, and struggles with complex non-rigid deformations. While enabling positive applications in AR/VR and remote collaboration, potential misuse exists in surveillance or unauthorized 3D reconstruction. We recommend consent mechanisms for human-centric applications and privacy-preserving rendering techniques. Future work should explore multi-modal sensing, self-supervised segmentation, and privacy-aware reconstruction protocols. + +# 7 Conclusion + +We presented a novel motion-aware framework for pose-free dynamic novel view synthesis from monocular videos. Our method integrates three key innovations: a 4D-aware information extractor leveraging pre-trained foundation models, a motion-aware bundle adjustment module for robust camera pose estimation, and a compact motion-aware Gaussian splatting representation. Extensive experiments show that our method significantly reduces computational overhead while achieving state-of-the-art performance in pose estimation accuracy and novel view synthesis quality. Compared to COLMAP-dependent methods, our approach achieves $5\mathrm{x}$ faster training times and outperforms existing methods by 1.8dB in PSNR. For more intricate dynamic scenes, future research might investigate adding temporal consistency constraints. Furthermore, examining self-supervised learning strategies for motion mask creation may help lessen dependency on pre-trained models while preserving strong performance. + +# Acknowledgment + +This research / project is supported by the National Research Foundation (NRF) Singapore, under its NRF-Investigatorship Programme (Award ID. NRF-NRFI09-0008), and the Tier 2 grant MOET2EP20124-0015 from the Singapore Ministry of Education. + +# References + +[1] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In ICCV, 2021. +[2] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In CVPR, 2022. +[3] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Zip-nerf: Antialiased grid-based neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19697-19705, 2023. +[4] Wenjing Bian, Zirui Wang, Kejie Li, Jia-Wang Bian, and Victor Adrian Prisacariu. Nope-nerf: Optimising neural radiance field with no pose prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4160–4169, 2023. +[5] Eric Brachmann, Tommaso Cavallari, and Victor Adrian Prisacariu. Accelerated coordinate encoding: Learning to relocalize in minutes using rgb and poses. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5044-5053, 2023. +[6] Daniel J Butler, Jonas Wulff, Garrett B Stanley, and Michael J Black. A naturalistic open source movie for optical flow evaluation. In ECCV. Springer, 2012. +[7] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. ECCV, 2022. +[8] Weirong Chen, Le Chen, Rui Wang, and Marc Pollefeys. Leap-vo: Long-term effective any point tracking for visual odometry. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19844-19853, 2024. +[9] Yu Chen, Rolandos Alexandros Potamias, Evangelos Ververas, Jifei Song, Jiankang Deng, and Gim Hee Lee. Zerogs: Training 3d gaussian splatting from unposed images. arXiv preprint arXiv:2411.15779, 2024. +[10] Shin-Fang Chng, Sameera Ramasinghe, Jamie Sherrah, and Simon Lucey. Gaussian activated neural radiance fields for high fidelity reconstruction and pose estimation. In European Conference on Computer Vision, pages 264-280. Springer, 2022. +[11] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superpoint: Self-supervised interest point detection and description. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 224-236, 2018. +[12] Zhiwen Fan, Wenyan Cong, Kairun Wen, Kevin Wang, Jian Zhang, Xinghao Ding, Danfei Xu, Boris Ivanovic, Marco Pavone, Georgios Pavlakos, et al. Instantsplat: Unbounded sparse-view pose-free gaussian splatting in 40 seconds. arXiv preprint arXiv:2403.20309, 2(3):4, 2024. +[13] Jiemin Fang, Taoran Yi, Xinggang Wang, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Matthias Nießner, and Qi Tian. Fast dynamic radiance fields with time-aware neural voxels. In SIGGRAPH Asia 2022 Conference Papers, pages 1-9, 2022. +[14] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 24(6):381-395, 1981. +[15] Sara Fridovich-Keil, Giacomo Meanti, Frederik Rahbæk Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. +[16] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In CVPR, 2022. + +[17] Yang Fu, Sifei Liu, Amey Kulkarni, Jan Kautz, Alexei A. Efros, and Xiaolong Wang. Colmap-free 3d gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20796-20805, June 2024. +[18] Hang Gao, Ruilong Li, Shubham Tulsiani, Bryan Russell, and Angjoo Kanazawa. Monocular dynamic view synthesis: A reality check. Advances in Neural Information Processing Systems, 35:33768-33780, 2022. +[19] Xiao-Shan Gao, Xiao-Rong Hou, Jianliang Tang, and Hang-Fei Cheng. Complete solution classification for the perspective-three-point problem. IEEE transactions on pattern analysis and machine intelligence, 25(8):930–943, 2003. +[20] Mengqi Guo, Chen Li, Hanlin Chen, and Gim Hee Lee. Unikd: Uncertainty-filtered incremental knowledge distillation for neural implicit representation. In European Conference on Computer Vision, pages 237-254. Springer, 2025. +[21] Peter Hedman, Pratul P. Srinivasan, Ben Mildenhall, Jonathan T. Barron, and Paul Debevec. Baking neural radiance fields for real-time view synthesis. ICCV, 2021. +[22] Yi-Hua Huang, Yang-Tian Sun, Ziyi Yang, Xiaoyang Lyu, Yan-Pei Cao, and Xiaojuan Qi. Sc-gs: Sparse-controlled gaussian splatting for editable dynamic scenes. In CVPR, pages 4220-4230, 2024. +[23] Yoonwoo Jeong, Junmyeong Lee, Hoseung Choi, and Minsu Cho. Rodygs: Robust dynamic gaussian splatting for casual videos. arXiv preprint arXiv:2412.03077, 2024. +[24] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Cotracker: It is better to track together. In European Conference on Computer Vision, pages 18-35. Springer, 2025. +[25] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42(4), July 2023. +[26] Johannes Kopf, Xuejian Rong, and Jia-Bin Huang. Robust consistent video depth estimation. In Computer Vision and Pattern Recognition, 2021. +[27] Agelos Kratimenos, Jiahui Lei, and Kostas Daniilidis. Dynmf: Neural motion factorization for real-time dynamic view synthesis with 3d gaussian splattering. In ECCV. Springer, 2025. +[28] Tianye Li, Mira Slavcheva, Michael Zollhoefer, Simon Green, Christoph Lassner, Changil Kim, Tanner Schmidt, Steven Lovegrove, Michael Goesele, Richard Newcombe, and Zhaoyang Lv. Neural 3d video synthesis from multi-view video. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2022. +[29] Yanyan Li, Chenyu Lyu, Yan Di, Guangyao Zhai, Gim Hee Lee, and Federico Tombari. Geogaussian: Geometry-aware gaussian splatting for scene rendering. In European Conference on Computer Vision, pages 441-457. Springer, 2024. +[30] Zhan Li, Zhang Chen, Zhong Li, and Yi Xu. Spacetime gaussian feature splatting for real-time dynamic view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. +[31] Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021. +[32] Zhengqi Li, Qianqian Wang, Forrester Cole, Richard Tucker, and Noah Snavely. Dynibar: Neural dynamic image-based rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. +[33] Chen-Hsuan Lin, Wei-Chiu Ma, Antonio Torralba, and Simon Lucey. Barf: Bundle-adjusting neural radiance fields. In ICCV, 2021. +[34] Philipp Lindenberger, Paul-Edouard Sarlin, and Marc Pollefeys. Lightglue: Local feature matching at light speed. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17627-17638, 2023. +[35] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. NeurIPS, 2020. + +[36] Yu-Lun Liu, Chen Gao, Andreas Meuleman, Hung-Yu Tseng, Ayush Saraf, Changil Kim, Yung-Yu Chuang, Johannes Kopf, and Jia-Bin Huang. Robust dynamic radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13-23, 2023. +[37] Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. In 3DV, 2024. +[38] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. +[39] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. SIGGRAPH, 2022. +[40] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In ICCV, pages 5865-5874, 2021. +[41] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Ricardo Martin-Brualla, and Steven M Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. arXiv preprint arXiv:2106.13228, 2021. +[42] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. In CVPR, 2021. +[43] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Radle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junting Pan, Kalyan Vasudev Alwala, Nicolas Carion, Chao-Yuan Wu, Ross Girshick, Piotr Dólar, and Christoph Feichtenhofer. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. +[44] Christian Reiser, Songyou Peng, Yiyi Liao, and Andreas Geiger. Kilonerf: Speeding up neural radiance fields with thousands of tiny mlp's. In ICCV, 2021. +[45] Barbara Roessle, Jonathan T Barron, Ben Mildenhall, Pratul P Srinivasan, and Matthias Nießner. Dense depth priors for neural radiance fields from sparse input views. In CVPR, 2022. +[46] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4938–4947, 2020. +[47] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. +[48] Johannes Lutz Schonberger, Enliang Zheng, Marc Pollefeys, and Jan-Michael Frahm. Pixelwise view selection for unstructured multi-view stereo. In European Conference on Computer Vision (ECCV), 2016. +[49] Jamie Shotton, Ben Glocker, Christopher Zach, Shahram Izadi, Antonio Criminisi, and Andrew Fitzgibbon. Scene coordinate regression forests for camera relocalization in rgb-d images. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2930-2937, 2013. +[50] Olga Sorkine and Marc Alexa. As-rigid-as-possible surface modeling. In Symposium on Geometry processing, volume 4, pages 109–116. CiteSeer, 2007. +[51] Pablo Speciale, Johannes L Schonberger, Sing Bing Kang, Sudipta N Sinha, and Marc Pollefeys. Privacy preserving image-based localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5493-5503, 2019. +[52] Robert W Sumner, Johannes Schmid, and Mark Pauly. Embedded deformation for shape manipulation. In ACM siggraph 2007 papers, pages 80–es. ACM, 2007. +[53] Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. Loftr: Detector-free local feature matching with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8922-8931, 2021. +[54] Zachary Teed and Jia Deng. Droid-slam: Deep visual slam for monocular, stereo, and rgb-d cameras. Advances in neural information processing systems, 34:16558-16569, 2021. +[55] Zachary Teed, Lahav Lipson, and Jia Deng. Deep patch visual odometry. Advances in Neural Information Processing Systems, 36, 2024. + +[56] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20697-20709, 2024. +[57] Yihan Wang, Lahav Lipson, and Jia Deng. Sea-raft: Simple, efficient, accurate raft for optical flow. In ECCV, pages 36–54, 2024. +[58] Yunsong Wang, Tianxin Huang, Hanlin Chen, and Gim Hee Lee. Freesplat: Generalizable 3d gaussian splatting towards free view synthesis of indoor scenes. Advances in Neural Information Processing Systems, 37:107326-107349, 2024. +[59] Zirui Wang, Shangzhe Wu, Weidi Xie, Min Chen, and Victor Adrian Prisacariu. Nerf-: Neural radiance fields without known camera parameters. arXiv preprint arXiv:2102.07064, 2021. +[60] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. In CVPR, 2024. +[61] Bo Xu, Ziao Liu, Mengqi Guo, Jiancheng Li, and Gim Hee Lee. Urs-nerf: Unordered rolling shutter bundle adjustment for neural radiance fields. In ECCV. Springer, 2025. +[62] Zhiwen Yan, Chen Li, and Gim Hee Lee. Nerf-ds: Neural radiance fields for dynamic specular objects. In CVPR, pages 8285-8295, 2023. +[63] Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything v2. arXiv preprint arXiv:2406.09414, 2024. +[64] Zeyu Yang, Hongye Yang, Zijie Pan, and Li Zhang. Real-time photorealistic dynamic scene representation and rendering with 4d gaussian splatting. arXiv preprint arXiv:2310.10642, 2023. +[65] Zeyu Yang, Hongye Yang, Zijie Pan, and Li Zhang. Real-time photorealistic dynamic scene representation and rendering with 4d gaussian splatting. In ICLR, 2024. +[66] Ziyi Yang, Xinyu Gao, Wen Zhou, Shaohui Jiao, Yuqing Zhang, and Xiaogang Jin. Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2024. +[67] Lin Yen-Chen, Pete Florence, Jonathan T Barron, Alberto Rodriguez, Phillip Isola, and Tsung-Yi Lin. Inerf: Inverting neural radiance fields for pose estimation. In IROS, 2021. +[68] Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. Plenoctrees for real-time rendering of neural radiance fields. In ICCV, 2021. +[69] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In CVPR, 2021. +[70] Jingbo Zhang, Xiaoyu Li, Ziyu Wan, Can Wang, and Jing Liao. Fdnerf: Few-shot dynamic neural radiance fields for face reconstruction and expression editing. In SIGGRAPH Asia 2022 Conference Papers, 2022. +[71] Junyi Zhang, Charles Herrmann, Junhwa Hur, Varun Jampani, Trevor Darrell, Forrester Cole, Deqing Sun, and Ming-Hsuan Yang. Monst3r: A simple approach for estimating geometry in the presence of motion. arXiv preprint arXiv:2410.03825, 2024. +[72] Zhoutong Zhang, Forrester Cole, Zhengqi Li, Michael Rubinstein, Noah Snavely, and William T Freeman. Structure and motion from casual videos. In European Conference on Computer Vision, pages 20-37. Springer, 2022. +[73] Wang Zhao, Shaohui Liu, Hengkai Guo, Wenping Wang, and Yong-Jin Liu. *Particlesfm: Exploiting dense point trajectories for localizing moving cameras in the wild*. In European Conference on Computer Vision, pages 523–542. Springer, 2022. + +# A Preliminaries + +The representation of 3D scenes with Gaussian splatting employs a collection of colored 3D Gaussian primitives. Each Gaussian primitive $G$ is characterized by its center $\mu$ in 3D space and a corresponding 3D covariance matrix $\Sigma$ , conforming to: + +$$ +G (\mathbf {x}) = \exp \left(- \frac {1}{2} (\mathbf {x} - \boldsymbol {\mu}) ^ {\top} \boldsymbol {\Sigma} ^ {- 1} (\mathbf {x} - \boldsymbol {\mu})\right). \tag {14} +$$ + +To facilitate optimization, we decompose the covariance matrix $\pmb{\Sigma}$ into $\mathbf{RSS}^{\top}\mathbf{R}^{\top}$ , where $\mathbf{R}$ represents a rotation matrix encoded by a quaternion $\mathbf{q} \in SO(3)$ , and $\mathbf{S}$ denotes a scaling matrix parameterized by a 3D vector $\mathbf{s}$ . The complete parameterization of each Gaussian includes an opacity value $\sigma$ governing its rendering influence, alongside spherical harmonic (SH) coefficients $\mathbf{sh}$ that capture view-dependent appearance variations. + +The scene representation therefore consists of a set $\mathcal{G} = \{G_j : \pmb{\mu}_j, \mathbf{q}_j, \mathbf{s}_j, \sigma_j, \mathbf{sh}_j\}$ , where $\pmb{\mu}_j$ is the position, $\mathbf{q}_j$ is the orientation, $\mathbf{s}_j$ is the scale, $\sigma_j$ is the standard deviation, and $\mathbf{sh}_j$ is the spherical harmonics coefficients. The rendering pipeline projects these 3D Gaussians onto the image plane, where they undergo efficient $\alpha$ -blending. During projection, the 2D covariance matrix $\pmb{\Sigma}'$ and center $\pmb{\mu}'$ are computed as: + +$$ +\boldsymbol {\Sigma} ^ {\prime} = \mathbf {J} \mathbf {W} \boldsymbol {\Sigma} \mathbf {W} ^ {\top} \mathbf {J} ^ {\top}, \quad \boldsymbol {\mu} ^ {\prime} = J W \boldsymbol {\mu}, \tag {15} +$$ + +where $\mathbf{J}$ is the Jacobian matrix of the linear approximation of the projective transformation and $\mathbf{W}$ is the rotation matrix of the viewpoint. The final color $C(\mathbf{u})$ at pixel $\mathbf{u}$ emerges from neural point-based $\alpha$ -blending: + +$$ +C (\mathbf {u}) = \sum_ {i \in \mathcal {N}} T _ {i} \alpha_ {i} \operatorname {S H} \left(\mathbf {s h} _ {i}, \mathbf {v} _ {i}\right), \quad \text {w h e r e} T _ {i} = \prod_ {j = 1} ^ {i - 1} (1 - \alpha_ {j}). \tag {16} +$$ + +Here, $\mathcal{N}$ shows the number of Gaussians that overlap with the pixel $\mathbf{u}$ . In this formulation, $\mathrm{SH}(\cdot ,\cdot)$ represents the spherical harmonic function evaluated with respect to the view-direction $\mathbf{v}_i$ . The $\alpha$ -value for each Gaussian is determined by: + +$$ +\alpha_ {i} = \sigma_ {i} \exp \left(- \frac {1}{2} \left(\mathbf {p} - \boldsymbol {\mu} _ {i} ^ {\prime}\right) ^ {\top} \boldsymbol {\Sigma} _ {i} ^ {\prime - 1} \left(\mathbf {p} - \boldsymbol {\mu} _ {i} ^ {\prime}\right)\right), \tag {17} +$$ + +where $\pmb{\mu}_i^{\prime}$ and $\Sigma_i^\prime$ correspond to the projected center and covariance matrix of Gaussian $G_{i}$ . Real-time and high-fidelity image synthesis is achieved through the optimization of Gaussian parameters $\{G_j:\pmb {\mu}_j,\mathbf{q}_j,\mathbf{s}_j,\sigma_j,\mathbf{sh}_j\}$ coupled with adaptive density adjustment. We propose a framework that builds on the foundation of Gaussian Splatting for dynamic scenes by adding sparse control, without compromising computational efficiency and rendering quality. + +# B Differentiable Dense Bundle Adjustment (DBA) layer + +We further refine the camera poses through the Differentiable Dense Bundle Adjustment (DBA) layer [54], which incorporates optical flow information to improve geometry estimation. This approach is particularly effective in dynamic scenes as it allows us to focus optimization on static regions while accounting for motion consistency in dynamic areas: + +$$ +E _ {D B A} \left(C _ {t} ^ {\prime}, d _ {t} ^ {\prime}\right) = \sum_ {(i, j) \in \mathcal {E}} \left(1 - M _ {t} (i)\right) \| p _ {i j} ^ {*} - \Pi_ {e} \left(C _ {i j} ^ {\prime} \circ \Pi_ {e} ^ {- 1} \left(p _ {i}, d _ {i} ^ {\prime}\right)\right) \| _ {\Sigma_ {i j}} ^ {2}, \tag {18} +$$ + +where $C_t'$ is the camera pose and $d_t'$ is the depth values. $\| \cdot \|_{\Sigma_{ij}}$ is the Mahalanobis distance weighted by the confidence scores, $(i,j) \in \mathcal{E}$ denotes an overlapping field-of-view with shared points between image $I_i$ and $I_j$ , $p_{ij}^*$ is the sum of optical flow $r_{ij}$ and $p_{ij}$ , and the term $(1 - M_t(i))$ ensures that only static regions contribute to the optimization. + +The system optimizes for updated camera pose $C_t'$ and depth values $d_t'$ through a sparse matrix formulation $\Delta \xi_t$ and $\Delta d_t$ , which is the normal equation derived from the cost function in Eqn.(18): + +$$ +\left[ \begin{array}{l l} B & E \\ E ^ {\top} & C \end{array} \right] \left[ \begin{array}{l} \Delta \xi_ {t} \\ \Delta d _ {t} \end{array} \right] = \left[ \begin{array}{l} v \\ w \end{array} \right] \tag {19} +$$ + +# Two-Stage Optimization Process + +![](images/5f08a9c37ccbbba16d2e74c09401b41cab2afd7cd48ffe14360e1c1423aefb14.jpg) +Stage 1: Control Point Optimization + +Stage 2: Gaussian Point Optimization with LBS +![](images/c33c6574cffcb0094627d2d0803a923ec43d3059bae8b15051d9146f1b3726c3.jpg) +Dynamic Control Points +$\bigcirc$ Static Control Points + +Figure 5: Two-Stage Optimization Process for Motion-Aware Gaussian Splatting. Stage 1 optimizes control points in dynamic regions (red) using control point loss, while static regions (gray) remain fixed. Stage 2 performs Gaussian optimization (blue ellipses) through Linear Blend Skinning, with connection lines showing influence weights between control points and Gaussians. +![](images/65e3d51a9f7c4baf7a07b947fe647ef2ee5945611f2e181aa3f955a574e889c2.jpg) +Gaussian Points +LBS Weights + +where $E$ models the coupling between pose and depth parameters, $C$ captures the depth-depth relationships, $B$ represents the pose-pose interactions, $v$ and $w$ are the gradient terms corresponding to pose updates and depth updates, respectively. + +# C Two-stage Optimization Strategy in MAGS + +Our two-stage optimization strategy reduces computational cost by focusing on dynamic regions only, as described in Sec. 3.5 of the main paper. We illustrate the strategy in Fig 5, which demonstrates our approach to efficient motion-aware scene reconstruction. In the first stage, we selectively optimize control points only within regions identified as dynamic by our motion segmentation module, significantly reducing the computational burden compared to methods that optimize all scene parameters simultaneously. The dynamic region boundary (indicated by the dashed red outline) separates areas requiring deformation modeling from static background elements. During the second stage, Gaussian primitives are optimized using Linear Blend Skinning weights computed through normalized exponential kernels based on spatial proximity to control points. The varying opacity of connection lines visualizes the influence magnitude of each control point on nearby Gaussian primitives, with stronger connections (darker lines) indicating higher LBS weights. This strategic separation of optimization stages enables our method to achieve both computational efficiency and high-quality reconstruction by concentrating computational resources where motion occurs while maintaining stable anchoring in static regions. + +Table 4: Quantitative evaluation of camera poses estimation on the MPI Sintel dataset. The best and the second best results are denoted by pink and yellow. The methods of the top block discard the dynamic components and do not reconstruct the dynamic scenes; thus they cannot render novel views. We exclude the COLMAP results since it fails to produce poses in 5 out of 14 sequences. + +
MethodATE↓RPE trans↓RPE rot↓
DROID-SLAM* [54]0.1750.0841.912
DPVO* [55]0.1150.0721.975
ParticleSFM [73]0.1290.0310.535
LEAP-VO* [8]0.0890.0661.250
Robust-CVD [26]0.3600.1543.443
CasualSAM [72]0.1410.0350.615
DUSt3R [56] w/ mask0.4170.2505.796
MonST3R [71]0.1080.0420.732
NeRF- [67]0.4330.2203.088
BARF [33]0.4470.2036.353
RoDynRF [36]0.0890.0731.313
Ours0.0860.0350.639
+ +* requires ground truth camera intrinsics as input + +# D LBS parameter learning + +As shown in Eqn. (8) of the main paper, the LBS weights are computed with a normalized exponential kernel. Our motion-aware framework significantly improves the efficiency and stability of LBS parameter learning through three key mechanisms: + +1) Focusing control point optimization exclusively on dynamic regions identified by our motion mask, which concentrates computational resources where they're most needed. +2) Preserving static points' positions during the GS optimization process, which provides stable anchors for the scene representation. +3) Substantially reducing the number of points requiring LBS parameter learning, which improves both computational efficiency and optimization stability. + +# E Results on Pose Estimation + +Tab 4 presents our pose estimation results on the MPI Sintel dataset, where our method demonstrates exceptional performance across all metrics. The methods in the upper portion of the table discard dynamic components and cannot render novel views. COLMAP results are excluded because it fails to produce poses in 5 out of 14 sequences, which further illustrates the challenges faced by COLMAP-dependent methods in handling general scenes with complex camera motions. We achieved an ATE of 0.086, showing comparable accuracy to the SOTA method LEAP-VO (0.089) and significantly outperforming methods like BARF (0.447) and NeRF- (0.433). For relative pose errors, our method achieves 0.035 for translation and 0.639 for rotation, matching or exceeding the performance of specialized pose estimation methods. + +The strong pose estimation performance can be attributed to several factors. First, our MA-BA effectively leverages the dynamic masks refined by our mask refinement pipeline, reducing the noise introduced by dynamic objects during pose optimization. Second, integrating SAM2 for mask refinement significantly improves the accuracy of dynamic object segmentation, leading to more reliable static point selection for RANSAC-based pose estimation. + +# F Technical Discussion and Analysis + +# F.1 Technical Contributions + +Our motion-aware framework delivers three key innovations: (1) an integrated MA-BA module that combines transformer-based motion priors with SAM2's segmentation capabilities in a unified + +pipeline; (2) a gradient-detached dynamic control point mechanism that strategically allocates computational resources to motion-significant regions; and (3) an end-to-end pose-free approach that eliminates dependency on pre-computed camera poses. Empirical validation confirms these innovations deliver substantial improvements, with a 1.3dB PSNR gain over combined baseline components. + +# F.2 Scene Coordinate Regression Advantages + +Our approach leverages Scene Coordinate Regression (SCR) over Structure from Motion (SfM) for its dual benefits: SCR not only delivers faster and more accurate results but also generates high-quality dynamic masks through our motion-aware bundle adjustment module. This creates a synergistic pipeline where dynamic region identification directly informs reconstruction, enabling more efficient processing while maintaining or exceeding the accuracy of traditional SfM approaches. + +# F.3 Two-Stage Optimization Strategy + +Our framework employs a two-stage strategy that separates motion estimation from reconstruction. The first stage establishes coherent motion estimates (16-19 PSNR) without optimizing Gaussian points directly. Control points define the deformation field but don't serve as Gaussian centers. Gaussians are initialized independently during the second stage, guided by the established motion field, enabling more stable convergence and higher-quality results. + +# F.4 Control Point Efficiency + +Our control point approach delivers two significant advantages: (1) a $5 \times$ improvement in training time by focusing optimization efforts only on dynamic regions; and (2) a remarkably compact representation requiring only 80MB of storage compared to 153-200MB for competing methods. These efficiency gains enable handling of challenging scenes with large motion magnitudes and longer sequences, as demonstrated by superior performance on the DyNeRF dataset. + +# F.5 Deformation Modeling + +Our approach employs As-Rigid-As-Possible (ARAP) regularization as a flexible prior for both rigid and non-rigid deformations. For soft-body objects, this acts as a smoothness constraint rather than enforcing strict rigidity. The system adaptively allocates more control points to highly deformable regions, creating a finer deformation grid where needed. This approach successfully handles soft objects, as demonstrated in the "peel-banana" sequence. + +# F.6 Implementation Parameters + +Our implementation uses 512 control points by default, initialized uniformly with higher density in dynamic regions. Performance remains robust across a range of control point quantities (100-1000), reflecting the effectiveness of our adaptive allocation strategy that focuses computational resources based on motion significance rather than a fixed distribution. + +# NeurIPS Paper Checklist + +# 1. Claims + +Question: Do the main claims made in the abstract and introduction accurately reflect the paper's contributions and scope? + +Answer:[Yes] + +Justification: The claims do reflect the paper's contributions and scope. + +Guidelines: + +- The answer NA means that the abstract and introduction do not include the claims made in the paper. +- The abstract and/or introduction should clearly state the claims made, including the contributions made in the paper and important assumptions and limitations. A No or NA answer to this question will not be perceived well by the reviewers. +- The claims made should match theoretical and experimental results, and reflect how much the results can be expected to generalize to other settings. +- It is fine to include aspirational goals as motivation as long as it is clear that these goals are not attained by the paper. + +# 2. Limitations + +Question: Does the paper discuss the limitations of the work performed by the authors? + +Answer: [Yes] + +Justification: Discuss in the main paper. + +Guidelines: + +- The answer NA means that the paper has no limitation while the answer No means that the paper has limitations, but those are not discussed in the paper. +- The authors are encouraged to create a separate "Limitations" section in their paper. +- The paper should point out any strong assumptions and how robust the results are to violations of these assumptions (e.g., independence assumptions, noiseless settings, model well-specification, asymptotic approximations only holding locally). The authors should reflect on how these assumptions might be violated in practice and what the implications would be. +- The authors should reflect on the scope of the claims made, e.g., if the approach was only tested on a few datasets or with a few runs. In general, empirical results often depend on implicit assumptions, which should be articulated. +- The authors should reflect on the factors that influence the performance of the approach. For example, a facial recognition algorithm may perform poorly when image resolution is low or images are taken in low lighting. Or a speech-to-text system might not be used reliably to provide closed captions for online lectures because it fails to handle technical jargon. +- The authors should discuss the computational efficiency of the proposed algorithms and how they scale with dataset size. +- If applicable, the authors should discuss possible limitations of their approach to address problems of privacy and fairness. +- While the authors might fear that complete honesty about limitations might be used by reviewers as grounds for rejection, a worse outcome might be that reviewers discover limitations that aren't acknowledged in the paper. The authors should use their best judgment and recognize that individual actions in favor of transparency play an important role in developing norms that preserve the integrity of the community. Reviewers will be specifically instructed to not penalize honesty concerning limitations. + +# 3. Theory assumptions and proofs + +Question: For each theoretical result, does the paper provide the full set of assumptions and a complete (and correct) proof? + +Answer: [Yes] + +Justification: Yes. paper provide the assumptions and a complete proof. + +# Guidelines: + +- The answer NA means that the paper does not include theoretical results. +- All the theorems, formulas, and proofs in the paper should be numbered and cross-referenced. +- All assumptions should be clearly stated or referenced in the statement of any theorems. +- The proofs can either appear in the main paper or the supplemental material, but if they appear in the supplemental material, the authors are encouraged to provide a short proof sketch to provide intuition. +- Inversely, any informal proof provided in the core of the paper should be complemented by formal proofs provided in appendix or supplemental material. +- Theorems and Lemmas that the proof relies upon should be properly referenced. + +# 4. Experimental result reproducibility + +Question: Does the paper fully disclose all the information needed to reproduce the main experimental results of the paper to the extent that it affects the main claims and/or conclusions of the paper (regardless of whether the code and data are provided or not)? + +# Answer: [Yes] + +Justification: the paper discloses all the information needed to reproduce the main experimental results. + +# Guidelines: + +- The answer NA means that the paper does not include experiments. +- If the paper includes experiments, a No answer to this question will not be perceived well by the reviewers: Making the paper reproducible is important, regardless of whether the code and data are provided or not. +- If the contribution is a dataset and/or model, the authors should describe the steps taken to make their results reproducible or verifiable. +- Depending on the contribution, reproducibility can be accomplished in various ways. For example, if the contribution is a novel architecture, describing the architecture fully might suffice, or if the contribution is a specific model and empirical evaluation, it may be necessary to either make it possible for others to replicate the model with the same dataset, or provide access to the model. In general, releasing code and data is often one good way to accomplish this, but reproducibility can also be provided via detailed instructions for how to replicate the results, access to a hosted model (e.g., in the case of a large language model), releasing of a model checkpoint, or other means that are appropriate to the research performed. +- While NeurIPS does not require releasing code, the conference does require all submissions to provide some reasonable avenue for reproducibility, which may depend on the nature of the contribution. For example +(a) If the contribution is primarily a new algorithm, the paper should make it clear how to reproduce that algorithm. +(b) If the contribution is primarily a new model architecture, the paper should describe the architecture clearly and fully. +(c) If the contribution is a new model (e.g., a large language model), then there should either be a way to access this model for reproducing the results or a way to reproduce the model (e.g., with an open-source dataset or instructions for how to construct the dataset). +(d) We recognize that reproducibility may be tricky in some cases, in which case authors are welcome to describe the particular way they provide for reproducibility. In the case of closed-source models, it may be that access to the model is limited in some way (e.g., to registered users), but it should be possible for other researchers to have some path to reproducing or verifying the results. + +# 5. Open access to data and code + +Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material? + +# Answer: [Yes] + +Justification: The paper will provide open access to the data and code after accepted. + +# Guidelines: + +- The answer NA means that paper does not include experiments requiring code. +- Please see the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details. +- While we encourage the release of code and data, we understand that this might not be possible, so "No" is an acceptable answer. Papers cannot be rejected simply for not including code, unless this is central to the contribution (e.g., for a new open-source benchmark). +- The instructions should contain the exact command and environment needed to run to reproduce the results. See the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details. +- The authors should provide instructions on data access and preparation, including how to access the raw data, preprocessed data, intermediate data, and generated data, etc. +- The authors should provide scripts to reproduce all experimental results for the new proposed method and baselines. If only a subset of experiments are reproducible, they should state which ones are omitted from the script and why. +- At submission time, to preserve anonymity, the authors should release anonymized versions (if applicable). +- Providing as much information as possible in supplemental material (appended to the paper) is recommended, but including URLs to data and code is permitted. + +# 6. Experimental setting/details + +Question: Does the paper specify all the training and test details (e.g., data splits, hyperparameters, how they were chosen, type of optimizer, etc.) necessary to understand the results? + +# Answer: [Yes] + +Justification: the paper specify all the training and test details. + +# Guidelines: + +- The answer NA means that the paper does not include experiments. +- The experimental setting should be presented in the core of the paper to a level of detail that is necessary to appreciate the results and make sense of them. +- The full details can be provided either with the code, in appendix, or as supplemental material. + +# 7. Experiment statistical significance + +Question: Does the paper report error bars suitably and correctly defined or other appropriate information about the statistical significance of the experiments? + +# Answer: [No] + +Justification: I mainly report PSNR and SSIM. + +# Guidelines: + +- The answer NA means that the paper does not include experiments. +- The authors should answer "Yes" if the results are accompanied by error bars, confidence intervals, or statistical significance tests, at least for the experiments that support the main claims of the paper. +- The factors of variability that the error bars are capturing should be clearly stated (for example, train/test split, initialization, random drawing of some parameter, or overall run with given experimental conditions). +- The method for calculating the error bars should be explained (closed form formula, call to a library function, bootstrap, etc.) +- The assumptions made should be given (e.g., Normally distributed errors). +- It should be clear whether the error bar is the standard deviation or the standard error of the mean. + +- It is OK to report 1-sigma error bars, but one should state it. The authors should preferably report a 2-sigma error bar than state that they have a $96\%$ CI, if the hypothesis of Normality of errors is not verified. +- For asymmetric distributions, the authors should be careful not to show in tables or figures symmetric error bars that would yield results that are out of range (e.g. negative error rates). +- If error bars are reported in tables or plots, The authors should explain in the text how they were calculated and reference the corresponding figures or tables in the text. + +# 8. Experiments compute resources + +Question: For each experiment, does the paper provide sufficient information on the computer resources (type of compute workers, memory, time of execution) needed to reproduce the experiments? + +Answer: [Yes] + +Justification: I include the resource details like GPU and training time. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The paper should indicate the type of compute workers CPU or GPU, internal cluster, or cloud provider, including relevant memory and storage. +- The paper should provide the amount of compute required for each of the individual experimental runs as well as estimate the total compute. +- The paper should disclose whether the full research project required more computer than the experiments reported in the paper (e.g., preliminary or failed experiments that didn't make it into the paper). + +# 9. Code of ethics + +Question: Does the research conducted in the paper conform, in every respect, with the NeurIPS Code of Ethics https://neurips.cc/public/EthicsGuidelines? + +Answer: [Yes] + +Justification: The paper follows the NeurIPS Code of Ethics. + +Guidelines: + +- The answer NA means that the authors have not reviewed the NeurIPS Code of Ethics. +- If the authors answer No, they should explain the special circumstances that require a deviation from the Code of Ethics. +- The authors should make sure to preserve anonymity (e.g., if there is a special consideration due to laws or regulations in their jurisdiction). + +# 10. Broader impacts + +Question: Does the paper discuss both potential positive societal impacts and negative societal impacts of the work performed? + +Answer: [Yes] + +Justification: Discuss in the supp. + +Guidelines: + +- The answer NA means that there is no societal impact of the work performed. +- If the authors answer NA or No, they should explain why their work has no societal impact or why the paper does not address societal impact. +- Examples of negative societal impacts include potential malicious or unintended uses (e.g., disinformation, generating fake profiles, surveillance), fairness considerations (e.g., deployment of technologies that could make decisions that unfairly impact specific groups), privacy considerations, and security considerations. +- The conference expects that many papers will be foundational research and not tied to particular applications, let alone deployments. However, if there is a direct path to any negative applications, the authors should point it out. For example, it is legitimate to point out that an improvement in the quality of generative models could be used to + +generate deepfakes for disinformation. On the other hand, it is not needed to point out that a generic algorithm for optimizing neural networks could enable people to train models that generate Deepfakes faster. + +- The authors should consider possible harms that could arise when the technology is being used as intended and functioning correctly, harms that could arise when the technology is being used as intended but gives incorrect results, and harms following from (intentional or unintentional) misuse of the technology. +- If there are negative societal impacts, the authors could also discuss possible mitigation strategies (e.g., gated release of models, providing defenses in addition to attacks, mechanisms for monitoring misuse, mechanisms to monitor how a system learns from feedback over time, improving the efficiency and accessibility of ML). + +# 11. Safeguards + +Question: Does the paper describe safeguards that have been put in place for responsible release of data or models that have a high risk for misuse (e.g., pretrained language models, image generators, or scraped datasets)? + +Answer: [NA] + +Justification: the paper poses no such risks. + +Guidelines: + +- The answer NA means that the paper poses no such risks. +- Released models that have a high risk for misuse or dual-use should be released with necessary safeguards to allow for controlled use of the model, for example by requiring that users adhere to usage guidelines or restrictions to access the model or implementing safety filters. +- Datasets that have been scraped from the Internet could pose safety risks. The authors should describe how they avoided releasing unsafe images. +- We recognize that providing effective safeguards is challenging, and many papers do not require this, but we encourage authors to take this into account and make a best faith effort. + +# 12. Licenses for existing assets + +Question: Are the creators or original owners of assets (e.g., code, data, models), used in the paper, properly credited and are the license and terms of use explicitly mentioned and properly respected? + +Answer: [Yes] + +Justification: The paper cites all original papers. + +Guidelines: + +- The answer NA means that the paper does not use existing assets. +- The authors should cite the original paper that produced the code package or dataset. +- The authors should state which version of the asset is used and, if possible, include a URL. +- The name of the license (e.g., CC-BY 4.0) should be included for each asset. +- For scraped data from a particular source (e.g., website), the copyright and terms of service of that source should be provided. +- If assets are released, the license, copyright information, and terms of use in the package should be provided. For popular datasets, paperswithcode.com/datasets has curated licenses for some datasets. Their licensing guide can help determine the license of a dataset. +- For existing datasets that are re-packaged, both the original license and the license of the derived asset (if it has changed) should be provided. +- If this information is not available online, the authors are encouraged to reach out to the asset's creators. + +# 13. New assets + +Question: Are new assets introduced in the paper well documented and is the documentation provided alongside the assets? + +# Answer: [NA] + +Justification: the paper does not release new assets. + +# Guidelines: + +- The answer NA means that the paper does not release new assets. +- Researchers should communicate the details of the dataset/code/model as part of their submissions via structured templates. This includes details about training, license, limitations, etc. +- The paper should discuss whether and how consent was obtained from people whose asset is used. +- At submission time, remember to anonymize your assets (if applicable). You can either create an anonymized URL or include an anonymized zip file. + +# 14. Crowdsourcing and research with human subjects + +Question: For crowdsourcing experiments and research with human subjects, does the paper include the full text of instructions given to participants and screenshots, if applicable, as well as details about compensation (if any)? + +# Answer: [NA] + +Justification: paper does not involve crowdsourcing nor research with human subjects. + +# Guidelines: + +- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. +- Including this information in the supplemental material is fine, but if the main contribution of the paper involves human subjects, then as much detail as possible should be included in the main paper. +- According to the NeurIPS Code of Ethics, workers involved in data collection, curation, or other labor should be paid at least the minimum wage in the country of the data collector. + +# 15. Institutional review board (IRB) approvals or equivalent for research with human subjects + +Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals (or an equivalent approval/review based on the requirements of your country or institution) were obtained? + +# Answer: [NA] + +Justification: the paper does not involve crowdsourcing nor research with human subjects. + +# Guidelines: + +- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. +- Depending on the country in which research is conducted, IRB approval (or equivalent) may be required for any human subjects research. If you obtained IRB approval, you should clearly state this in the paper. +- We recognize that the procedures for this may vary significantly between institutions and locations, and we expect authors to adhere to the NeurIPS Code of Ethics and the guidelines for their institution. +- For initial submissions, do not include any information that would break anonymity (if applicable), such as the institution conducting the review. + +# 16. Declaration of LLM usage + +Question: Does the paper describe the usage of LLMs if it is an important, original, or non-standard component of the core methods in this research? Note that if the LLM is used only for writing, editing, or formatting purposes and does not impact the core methodology, scientific rigorousness, or originality of the research, declaration is not required. + +# Answer: [NA] + +Justification: the core method development in this research does not involve LLMs as any important, original, or non-standard components. + +# Guidelines: + +- The answer NA means that the core method development in this research does not involve LLMs as any important, original, or non-standard components. +- Please refer to our LLM policy (https://neurips.cc/Conferences/2025/LLM) for what should or should not be described. \ No newline at end of file diff --git a/NeurIPS/2025/4D3R_ Motion-Aware Neural Reconstruction and Rendering of Dynamic Scenes from Monocular Videos/images.zip b/NeurIPS/2025/4D3R_ Motion-Aware Neural Reconstruction and Rendering of Dynamic Scenes from Monocular Videos/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..ca79bfeafe9d51181325f8c57f25ce61cef48ee0 --- /dev/null +++ b/NeurIPS/2025/4D3R_ Motion-Aware Neural Reconstruction and Rendering of Dynamic Scenes from Monocular Videos/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:da6697ff235376eb62486e845341382533393c5e28a8a8d69262689e8d861c3d +size 589346 diff --git a/NeurIPS/2025/4D3R_ Motion-Aware Neural Reconstruction and Rendering of Dynamic Scenes from Monocular Videos/layout.json b/NeurIPS/2025/4D3R_ Motion-Aware Neural Reconstruction and Rendering of Dynamic Scenes from Monocular Videos/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c1c51e27410892653c07e6d388a2ab5c8429a517 --- /dev/null +++ b/NeurIPS/2025/4D3R_ Motion-Aware Neural Reconstruction and Rendering of Dynamic Scenes from Monocular Videos/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:977a4c5d227e8d13ed3b68b1aea2d4f0883262d0fa4ed28e54addffb648f2748 +size 765483 diff --git a/NeurIPS/2025/4DGCPro_ Efficient Hierarchical 4D Gaussian Compression for Progressive Volumetric Video Streaming/1472e6bb-e10f-4ba7-bd48-854364138628_content_list.json b/NeurIPS/2025/4DGCPro_ Efficient Hierarchical 4D Gaussian Compression for Progressive Volumetric Video Streaming/1472e6bb-e10f-4ba7-bd48-854364138628_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..6623f1374a0017b921da9c767a69c34c9a4dfede --- /dev/null +++ b/NeurIPS/2025/4DGCPro_ Efficient Hierarchical 4D Gaussian Compression for Progressive Volumetric Video Streaming/1472e6bb-e10f-4ba7-bd48-854364138628_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7c8f2a2c86d463a7cd1e1da0ea5834e3daf052dfa772314b0e756171fe0a7b20 +size 158705 diff --git a/NeurIPS/2025/4DGCPro_ Efficient Hierarchical 4D Gaussian Compression for Progressive Volumetric Video Streaming/1472e6bb-e10f-4ba7-bd48-854364138628_model.json b/NeurIPS/2025/4DGCPro_ Efficient Hierarchical 4D Gaussian Compression for Progressive Volumetric Video Streaming/1472e6bb-e10f-4ba7-bd48-854364138628_model.json new file mode 100644 index 0000000000000000000000000000000000000000..0c1ea3f5c994e7c9cd724652d96034f61a92ba69 --- /dev/null +++ b/NeurIPS/2025/4DGCPro_ Efficient Hierarchical 4D Gaussian Compression for Progressive Volumetric Video Streaming/1472e6bb-e10f-4ba7-bd48-854364138628_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1c1248d3e354224bc73f0bb5828896111344ea31c105721d65b0b9236131c18a +size 207115 diff --git a/NeurIPS/2025/4DGCPro_ Efficient Hierarchical 4D Gaussian Compression for Progressive Volumetric Video Streaming/1472e6bb-e10f-4ba7-bd48-854364138628_origin.pdf b/NeurIPS/2025/4DGCPro_ Efficient Hierarchical 4D Gaussian Compression for Progressive Volumetric Video Streaming/1472e6bb-e10f-4ba7-bd48-854364138628_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..f78308754cb91ea8be915ae2df54273cb798cb4d --- /dev/null +++ b/NeurIPS/2025/4DGCPro_ Efficient Hierarchical 4D Gaussian Compression for Progressive Volumetric Video Streaming/1472e6bb-e10f-4ba7-bd48-854364138628_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c348f929eca5c836aa40fead58cb33a8916f7e64f70334d48adda812d6af1228 +size 42047514 diff --git a/NeurIPS/2025/4DGCPro_ Efficient Hierarchical 4D Gaussian Compression for Progressive Volumetric Video Streaming/full.md b/NeurIPS/2025/4DGCPro_ Efficient Hierarchical 4D Gaussian Compression for Progressive Volumetric Video Streaming/full.md new file mode 100644 index 0000000000000000000000000000000000000000..c1bcf703c272546e98cc4f5d9f812bd4050df2fa --- /dev/null +++ b/NeurIPS/2025/4DGCPro_ Efficient Hierarchical 4D Gaussian Compression for Progressive Volumetric Video Streaming/full.md @@ -0,0 +1,697 @@ +# 4DGCPro: Efficient Hierarchical 4D Gaussian Compression for Progressive Volumetric Video Streaming + +Zihan Zheng $^{1}$ , Zhenlong Wu $^{1}$ , Houqiang Zhong $^{2}$ , Yuan Tian $^{2,3}$ , Ning Cao $^{4}$ , Lan Xu $^{5}$ , Jiangchao Yao $^{1}$ , Xiaoyun Zhang $^{1}$ , Qiang Hu $^{1*}$ , Wenjun Zhang $^{1,2}$ + +Cooperative Medianet Innovation Center, Shanghai Jiaotong University1 +Department of Electronics, Shanghai Jiaotong University2 +Shanghai AI Lab3 + +Cloud platform department, E-surfing Vision Technology Co., Ltd. School of Information Science and Technology, ShanghaiTech University + +![](images/8c1975c12eb7c1e636beb5ed5a3c656407e3f1e848e369084c2c56240aac1b65.jpg) +Figure 1: Left: Our method enables progressive streaming of hierarchical 4D Gaussians within a single bitstream, where incremental enhancement layers (e.g., +0.10MB) gradually improve visual quality (e.g., from 30.04dB to 31.18dB) with minimal bitrate overhead. Right: The streamed content is adaptively decoded and rendered in real-time on various devices (e.g., tablets 44FPS, desktops 58FPS) by dynamically selecting layers (L1-L6) based on available bandwidth and compute. + +# Abstract + +Achieving seamless viewing of high-fidelity volumetric video, comparable to 2D video experiences, remains an open challenge. Existing volumetric video compression methods either lack the flexibility to adjust quality and bitrate within a single model for efficient streaming across diverse networks and devices, or struggle with real-time decoding and rendering on lightweight mobile platforms. To address these challenges, we introduce 4DGCPro, a novel hierarchical 4D Gaussian compression framework that facilitates real-time mobile decoding and high-quality rendering via progressive volumetric video streaming in a single bitstream. Specifically, we propose a perceptually-weighted and compression-friendly hierarchical 4D Gaussian representation with motion-aware adaptive grouping to reduce temporal redundancy, preserve coherence, and enable scalable multi-level detail streaming. Furthermore, we present an end-to-end entropy-optimized training scheme, which incorporates layer-wise rate-distortion (RD) supervision and attribute-specific entropy modeling for efficient bitstream generation. Extensive experiments show that 4DGCPro enables flexible quality and multiple bitrate within a single model, achieving real-time decoding and rendering on mobile devices while outperforming existing methods in RD performance across multiple datasets. + +# 1 Introduction + +Volumetric video enables immersive 3D experiences with free-viewpoint navigation, but streaming and rendering high-quality long sequences with large motions remains challenging, especially on lightweight devices like mobile phones. Compared to 2D video, volumetric content demands higher bandwidth, storage, and real-time decoding capabilities, making fixed-bitrate solutions inadequate to handle the variability across heterogeneous devices and network conditions. Therefore, the core challenge lies in achieving real-time, high-fidelity playback with low computational cost, while enabling scalable and progressive streaming under constrained resources. + +Traditional volumetric reconstruction methods, including surface estimation [8], point clouds [57], meshes [68], light field [28, 44] and depth-based techniques [60, 79], struggle to faithfully capture the geometric complexity and temporal dynamics of real-world scenes. Neural radiance field (NeRF) [40] address these limitations by modeling view-dependent appearance without relying on explicit geometry, enabling photorealistic rendering. While extensions [49, 30, 12, 24, 14, 5, 78] adapt NeRF to dynamic scenes, they remain constrained by the difficulty of handling long sequences and efficient streaming. Some works [66, 67, 73, 81, 80] compress dynamic NeRFs to enable streaming, but the high computational cost of decoding and rendering limits their practicality in real-time applications. + +Recent work on 3D Gaussian Splatting (3DGS) [27] introduces an explicit scene representation using anisotropic Gaussian primitives with real-time, differentiable rasterization, achieving unprecedented rendering speed and visual quality. Subsequent studies [34, 72, 19, 75] extend 3DGS to dynamic scenes by incorporating temporal attributes into Gaussian parameters, but require full-sequence pre-loading during training and rendering, limiting streaming practicality. Alternative approaches [38, 61, 18, 16] model temporal Gaussian variations via deformable fields or residual tracking to enable streamable representations, yet incur substantial bandwidth overhead. A few studies [20, 26, 25, 69] have explored compression for dynamic 3DGS. For example, 4DGC [20] jointly optimizes representation and entropy models via RD loss, improving efficiency but struggling with high decoding latency and poor robustness to large motions due to rigid modeling. More fundamentally, existing dynamic Gaussian compression methods lack the flexibility to adjust video quality and bitrate within a single model, and typically require separate models for each bitrate, leading to high storage costs and limited adaptability under varying network and device conditions. + +To tackle the above challenges, we propose 4DGCPro, a novel hierarchical 4D Gaussian compression approach for progressive volumetric video streaming. As illustrated in Fig. 1, our method achieves multiple bitrate using a single model and enables real-time decoding and high-fidelity rendering on lightweight devices for large-motion sequences. We realize this through three key innovations. First, we introduce a perceptually-weighted hierarchical Gaussian representation for keyframes, guided by a significance metric that combines geometric volume and opacity. This enables scalable representation across detail levels and establishes the foundation for dynamic modeling. Second, we propose a hierarchical motion modeling strategy, where motion in subsequent frames is decomposed into rigid transformations and residual deformations to capture large displacements and preserve temporal coherence. We further adopt motion-aware adaptive Gaussian grouping to handle topological changes and long-term dynamics, ensuring compact and consistent temporal representation. + +Third, we propose a joint entropy-optimized training and progressive coding framework for efficient and scalable bitstream generation. Specifically, we introduce layer-wise rate-distortion (RD) supervision into the training pipeline using differentiable quantization and attribute-specific entropy modeling. For keyframes, we utilize FFT-accelerated Gaussian kernel density estimation (KDE) for precise bitrate prediction of Gaussian attributes, with hierarchical Gaussian optimization guided by per-layer RD trade-offs. For inter-frame, we apply Gaussian-distribution-based entropy estimation and temporal consistency constraints to maintain compactness and coherence. After training, we quantize attributes and convert multi-layer representations into stacked 2D single-channel maps, which are encoded with 2D CODEs into a progressive bitstream, enabling scalable real-time decoding and rendering via hardware video CODEs and shaders. Experimental results show that our 4DGCPro supports multiple bitrates using a single model and achieves state-of-the-art RD performance across various datasets. Compared to the SOTA method HPC [80], our approach achieves a $3\mathbf{x}$ compression rate without quality degradation, while enabling real-time decoding and rendering on mobile devices. + +In summary, our contributions are as follows: + +- We propose 4DGCPro, a novel framework for progressive volumetric video streaming that supports multiple bitrates with a single compact model, enabling real-time decoding and rendering on mobile devices with superior RD performance. +- We introduce a compact hierarchical 4D Gaussian representation with motion-aware adaptive grouping for scalable and high-fidelity modeling of dynamic scenes. +- We present an end-to-end entropy-optimized training scheme with layer-wise RD supervision and attribute-specific entropy modeling, enabling fine-grained RD optimization across layers and better overall compression. + +# 2 Related Work + +# 2.1 NeRF-based Volumetric Video Modeling + +NeRF [40] have revolutionized 3D scene representation using differentiable volume rendering with implicit neural representations. While recent advances in static scene representation [2-4, 7, 39, 41, 46, 52, 53] have improved compactness and reconstruction speed, several works have extended these methods to dynamic scenes. Flow-based approaches [32, 33] construct 3D features from monocular video, reducing data collection complexity but requiring additional priors for complex scenes. Deformation field methods [10, 45, 49, 59] warp dynamic frames into a canonical space to capture temporal features, yet suffer from slow training and rendering. To accelerate performance, recent methods [12, 24, 14, 5, 58, 31, 47, 63, 65] adopt explicit 4D radiance field representations based on structured volumetric decompositions (e.g., voxel grids, multi-plane projections, or tensor factorizations), yet these unified frameworks remain incompatible with streaming scenarios. + +# 2.2 3DGS-based Volumetric Video Modeling + +3DGS [27] and its variants [22, 13, 23, 6, 17] enable photorealistic static scene reconstruction through their efficiency and physical interpretability, with recent extensions to dynamic scenes. Current dynamic 3DGS approaches mainly follow two paradigms. Some studies [34, 72, 76, 19, 75, 74, 77] model Gaussian attributes as temporal functions to create unified dynamic representations, which achieve exceptional RD performance but neglect streaming feasibility. Alternative approaches [38, 61, 18, 16] employ frame-wise modeling with explicit rigid motion estimation, enabling streamable Gaussian volumetric video at the cost of increased data volume and compromised reconstruction quality. While $\mathrm{V}^3$ [69] optimizes the full Gaussian coefficients to model complex motions, its fixed group length leads to error accumulation or data redundancy. Meanwhile, it lacks the capability to support multiple-bitrate selection within a single bitstream. Our approach introduces a compact hierarchical motion-aware Gaussian representation coupled with adaptive Gaussian grouping that dynamically responds to topological changes and achieves multiple bitrate using a single model. + +# 2.3 Volumetric Video Compression + +Volumetric video compression is crucial for reducing massive data requirements, where traditional approaches employ octree [55, 62] and wavelet [42] techniques (later standardized as MPEG-PCC [56]), while subsequent learning-based methods [35, 50, 51, 29, 1, 15, 64] focus on improved efficiency. While recent advances [66, 67, 73, 59, 48, 9, 54, 81, 80, 21] have made progress in compressing dynamic NeRF features for storage optimization, they commonly suffer from poor quality and slow decoding/rendering. For instance, HPC [80] employs learned compression for progressive coding of residual feature grids representing dynamic scenes. However, its high decoding latency limits real-time applications. For 3DGS-based methods, static scene techniques [43, 11, 37, 71] dominate, whereas dynamic scene approaches [26, 25, 20] face inefficiency and single-rate constraints. Our method delivers superior RD performance and computational efficiency using standard video codecs, supporting both hardware-accelerated real-time decoding and progressive streaming for quality adaptation across dynamic bandwidth. + +# 3 Method + +In this section, we present the technical details of our 4DGCPro architecture (Fig. 2). The framework begins with a perceptually-weighted hierarchical Gaussian representation for keyframes (Sec. 3.1), + +![](images/85dde56a8eead98685ec509261db7ea59cc18be9bf80e45a8f7e9e9b1dbc3a3f.jpg) +Figure 2: Our 4DGCPro framework. (a) Perceptually-weighted hierarchical 4D Gaussian representation models keyframes at multi-level detail for progressive reconstruction. (b) Hierarchical motion modeling decomposes dynamic scenes into rigid transformations and residual deformations based on the previous frame, while (c) motion-aware adaptive grouping dynamically adjusts to topological changes to enhance temporal consistency and reduce error accumulation. The entire pipeline is end-to-end optimized with layer-wise RD supervision and attribute-specific entropy modeling. + +which establishes the foundation for dynamic scene characterization and progressive streaming. We then introduce a hierarchical motion modeling approach with adaptive grouping (Sec. 3.2), decomposing motions into rigid transformations and residual deformations. This motion-aware adaptive Gaussian grouping mechanism effectively handles diverse motion patterns in complex scenes. To generate efficient and scalable bitstreams, we incorporate layer-wise RD optimization into the training pipeline through differentiable quantization and attribute-specific entropy modeling, followed by compression using standard 2D video CODECs (Sec. 3.3). + +# 3.1 Perceptually-Weighted Hierarchical Gaussian Keyframe Representation + +Recall that 3DGS represents scenes explicitly through 3D Gaussians $\mathbf{G}$ , defined by a set of learnable parameters, including center position $\mu$ , rotation matrix $\mathbf{R}$ representing orientation, spherical harmonic coefficients $\mathbf{f}$ for view-dependent appearance modeling, scaling factors $\mathbf{s}$ controlling spatial extent, and opacity value $\alpha$ . The spatial influence at point $\mathbf{x}$ follows $\mathbf{G}(\mathbf{x})$ , expressed as: + +$$ +\mathbf {G} (\mathbf {x}) = \exp \left(- \frac {1}{2} (\mathbf {x} - \boldsymbol {\mu}) ^ {T} \boldsymbol {\Sigma} ^ {- 1} (\mathbf {x} - \boldsymbol {\mu})\right). \tag {1} +$$ + +The covariance matrix $\pmb{\Sigma}$ is constructed through $\pmb{\Sigma} = \mathbf{R}\mathbf{s}\mathbf{s}^T\mathbf{R}^T$ . With $\alpha_{i}^{\prime}$ being the projection of the opacity of the $i$ -th Gaussian onto the image plane and $\mathbf{c}_i$ denoting the color of the $i$ -th Gaussian in the viewing direction, the pixel color $\mathbf{c}$ is computed by differentiable splatting of $N$ ordered Gaussians as follows: + +$$ +\mathbf {c} = \sum_ {i \in N} \mathbf {c} _ {i} \alpha^ {\prime} _ {i} \prod_ {j = 1} ^ {i - 1} \left(1 - \alpha^ {\prime} _ {j}\right), \tag {2} +$$ + +When reconstructing long-duration dynamic scenes, we first reconstruct high-quality keyframes to serve as references for subsequent inter-frames. Inspired by $\mathbf{V}^3$ [69], we initialize keyframe 3D Gaussians through NeuS2-based [70] surface mesh extraction. After pre-training, low-opacity Gaussians are pruned to achieve compact representations. While this optimized 3DGS delivers high-fidelity reconstruction, its substantial data footprint becomes problematic for smooth viewing under fluctuating bandwidth conditions. We therefore propose a perceptually-weighted hierarchical Gaussian representation guided by significance metric $\Psi$ , which serves as the basis for progressive transmission and rendering. The proposed metric $\Psi$ analytically evaluates each Gaussian's visual importance through two geometrically-grounded attributes: (1) spatial volume $S$ , representing the 3D volume occupied by the Gaussian and computed as $\frac{4}{3}\pi abc$ (where $a, b, c$ are its scale parameters along the three principal axes), which reflects its structural contribution to the scene geometry; and (2) opacity $\alpha$ , which determines its perceptual weight in final rendering. These orthogonal factors are integrated with the weight $\lambda_{\Psi}$ : + +$$ +\Psi = \alpha + \lambda_ {\Psi} S. \tag {3} +$$ + +![](images/e01ca85a13c67935004c1ac669d696d7f79af71554e9648f619b072d445ee758.jpg) +Keyframe Inter-frame + +![](images/2d996230be542130947b488818f06a1131808d12441e78e358d4f42162140afe.jpg) + +![](images/b26d4657db7dc5c217c78a986ac6c514a3726dcaf8e1f26eeee5ae84d25052ac.jpg) +# +Figure 3: Analysis of group size and attributes distributions. (a) Large groups suffer from error accumulation while small groups exhibit data redundancy. (b) Keyframe Gaussian attributes display irregular spatial distributions, whereas (c) residual attributes follow Gaussian distributions. + +![](images/45b43af99e64fe51558dd448c75197570f12d7220885565a9c4518f74dffff4b.jpg) + +After sorting all Gaussians in descending order using our significance metric, we partition them into $L$ hierarchical layers $\mathbf{G} = \{\mathbf{G}^l\}_{l=1}^L$ . The base layer $\mathbf{G}^1$ preserves essential scene structures, while subsequent layers progressively enhance details. This hierarchical representation facilitates adaptive streaming, where the client dynamically selects the optimal number of layers $l$ to decode based on network conditions and computational resources. This approach ensures smooth playback while efficiently balancing transmission overhead and reconstruction fidelity across diverse network environments. + +Progressive Rendering. Our method supports progressive rendering from a single compressed representation, enabling scalable visual output with adjustable levels of detail. Starting from the base layer $(l = 1)$ , which contains essential structural and appearance information, each subsequent Gaussian layer incrementally refines the reconstruction. Specifically, when decoding up to level $l$ , only the Gaussians up to that layer (denoted by the index set $N^l$ ) are used for rendering. The color $\mathbf{c}^l$ at this stage is computed as: + +$$ +\mathbf {c} ^ {l} = \sum_ {i \in N ^ {l}} \mathbf {c} ^ {i} \alpha^ {\prime} _ {i} \prod_ {j = 1} ^ {i - 1} \left(1 - \alpha^ {\prime} _ {j}\right). \tag {4} +$$ + +Each additional layer introduces a compact set of Gaussians that enhance detail without redundant transmission. This hierarchical refinement strategy allows the model to adapt dynamically to available computational and bandwidth resources, balancing reconstruction quality and efficiency in real time. The result is a scalable, high-fidelity rendering system capable of maintaining seamless visual enhancement under varying resource constraints, all within a unified bitstream. + +# 3.2 Hierarchical Motion Modeling with Adaptive Grouping + +Building upon the hierarchical keyframe Gaussian representation, we employ it as a reference basis for training subsequent inter-frames. Our proposed hierarchical motion modeling strategy effectively captures large-scale complex motions while maintaining temporal coherence through decomposition of frame differences into rigid transformations and residual deformations. We further introduces motion-aware adaptive Gaussian grouping, which dynamically responds to varying scene changes to achieve dual benefits: enhanced representation fidelity and reduced model size for efficient streaming. + +Rigid Transformation. To estimate the rigid transformations of Gaussians between frames, we utilize the Gaussian positions $\pmb{\mu}_{t-1} = \{\pmb{\mu}_{t-1}^l\}_{l=1}^L$ from the previous frame as input and predicts both translation $\Delta \pmb{\mu}_t = \{\Delta \pmb{\mu}_t^l\}_{l=1}^L$ and rotation $\Delta \mathbf{R}_t = \{\Delta \mathbf{R}_t^l\}_{l=1}^L$ . The module first employs a multi-resolution hash grid $\mathbf{H}_t = \{\mathbf{H}_t^l\}_{l=1}^{L_h}$ with $L_h$ levels to capture motion features $\mathbf{h}_t$ at different scales through hash coding as: + +$$ +\mathbf {h} _ {t} ^ {l} = \left\{\mathbf {h} _ {t} ^ {l _ {h}} \right\} _ {l _ {h} = 1} ^ {L _ {h}} = \left\{\operatorname {i n t e r p} \left(\boldsymbol {\mu} _ {t - 1} ^ {l}, \mathbf {H} _ {t} ^ {l _ {h}}\right) \right\} _ {l _ {h} = 1} ^ {L _ {h}}, \tag {5} +$$ + +where $\mathrm{interp}(\cdot)$ refers to the hash grid interpolation operation. Subsequently, $\mathbf{h}_t$ is input into two lightweight MLPs, namely $\Phi_t^\mu$ and $\Phi_t^{\mathbf{R}}$ , to calculate the translation $\Delta \pmb{\mu}_t^l$ and rotation $\Delta \mathbf{R}_t^l$ for each Gaussian: + +$$ +\Delta \boldsymbol {\mu} _ {t} ^ {l} = \Phi_ {t} ^ {\boldsymbol {\mu}} \left(\mathbf {h} _ {t} ^ {l}\right), \quad \Delta \mathbf {R} _ {t} ^ {l} = \Phi_ {t} ^ {\mathbf {R}} \left(\mathbf {h} _ {t} ^ {l}\right). \tag {6} +$$ + +In this manner, the position and rotation of frame $t$ can be determined using the equations $\pmb{\mu}_t^l = \pmb{\mu}_{t - 1}^l +\Delta \pmb{\mu}_t^l$ and $\mathbf{R}_t^l = \Delta \mathbf{R}_t^l\mathbf{R}_{t - 1}^l$ + +Residual Deformation. Existing motion-aware 3DGS streaming methods [20, 61] primarily focus on rigid motion simulation and Gaussian compensation, which often fail to accommodate object deformation and frequently introduces visual artifacts and temporal instability. To address these limitations, our approach incorporates a residual deformation framework following rigid transformation. The framework learns Gaussian deformations via adaptive scaling, opacity and color adjustments while predicting attribute residuals $(\Delta \mathbf{s}_t^l,\Delta \alpha_t^l,\Delta \mathbf{f}_t^l)$ relative to parameters of t-1 frame, ensuring both local detail preservation and temporal stability. + +$$ +\mathbf {s} _ {t} ^ {l} = \mathbf {s} _ {t - 1} ^ {l} + \Delta \mathbf {s} _ {t} ^ {l}, \quad \alpha_ {t} ^ {l} = \alpha_ {t - 1} ^ {l} + \Delta \alpha_ {t} ^ {l}, \quad \mathbf {f} _ {t} ^ {l} = \mathbf {f} _ {t - 1} ^ {l} + \Delta \mathbf {f} _ {t} ^ {l}. \tag {7} +$$ + +By combining both rigid transformations and residual deformations, our method effectively captures both large displacements and subtle scene variations, significantly reducing visual artifacts while maintaining temporal coherence. + +Motion-aware Adaptive Gaussian Grouping. For long-sequence dynamic scenes with substantial motion, using only the initial frame as reference becomes inadequate due to accumulating scene variations. Meanwhile, as shown in Fig. 3(a), fixed-length group structures inevitably introduce two competing artifacts: error accumulation across frames in large groups, and data redundancy in small groups due to repeated parameter transmission. We address this through motion-aware adaptive Gaussian grouping, where the group size is dynamically determined by rigid transformation results. When the average Gaussian translation $\overline{\Delta\mu_t}$ exceeds a predefined threshold $\tau_{\mu}$ , indicating substantial scene changes, we initiate a new group with an updated reference frame. This adaptive grouping strategy automatically adjusts to motion intensity, employing shorter groups during rapid changes for better reference quality, while maintaining longer groups for stable segments to optimize compression efficiency. The resulting representation achieves both accuracy and compactness by balancing temporal coherence with adaptive topology updates. + +In summary, our 4DGCPro dynamically structures the scene into variable-length groups for efficient temporal modeling. For a group starting at frame T with length N, we sequentially represent it as $\mathbf{G}_T$ , $\{\Delta \pmb{\mu}_t, \Delta \mathbf{R}_t, \Delta \mathbf{f}_t, \Delta \mathbf{s}_t, \Delta \alpha_t\}_{t = T + 1}^T + N - 1$ , where $\mathbf{G}_T$ is the keyframe Gaussian and $\{\Delta \pmb{\mu}_t, \Delta \mathbf{R}_t, \Delta \mathbf{f}_t, \Delta \mathbf{s}_t, \Delta \alpha_t\}$ are the hierarchical residual attributes. This design optimally exploits inter-frame similarities while preserving reconstruction quality under complex motions. + +# 3.3 End-to-end Entropy-optimized Training + +We propose an end-to-end entropy-optimized training scheme, which attains the optimal RD performance by incorporating layer-wise RD supervision and attribute-specific entropy modeling. To facilitate gradient back-propagation, we utilize differentiable quantization along with attribute-specific entropy modeling method to accurately estimate the bitrates of diverse attributes. Furthermore, we carry out progressive compression with 2D codec on the hierarchical representation of Gaussians, enabling scalable real-time decoding and rendering. Next, we will introduce keyframe optimization, inter-keyframe optimization, and progressive bitstream generation in details. + +Keyframe Optimization. During the optimization process of keyframes, we first use $\mathcal{L}_{color}$ as a supervision term to pretrain the Gaussians: + +$$ +\mathcal {L} _ {\text {c o l o r}} = \left(1 - \lambda_ {\text {s s i m}}\right) \| \mathbf {c} _ {g} - \hat {\mathbf {c}} \| _ {1} + \lambda_ {\text {s s i m}} \mathcal {L} _ {\mathrm {D} - \mathrm {S S I M}}, \tag {8} +$$ + +where $\mathbf{c}_g$ and $\hat{\mathbf{c}}$ denote the ground truth and reconstructed colors respectively, and $\lambda_{\mathrm{ssim}}$ weights the D-SSIM[36] metric. After pre-training and pruning, we hierarchically organize Gaussians and perform joint entropy-optimized hierarchical training to maximize the RD performance per level. To ensure differentiable gradient flow and enhance quantization robustness, we implement uniform noise injection $u \sim U\left(-\frac{1}{2q}, \frac{1}{2q}\right)$ to simulate quantization effects with step size $q$ . Additionally, we introduce entropy estimation of Gaussian attributes into our loss function to improve compression efficiency. As shown in Fig. 3(b), keyframe Gaussian attributes exhibit irregular spatial distributions, necessitating KDE-based density estimation. Our implementation first computes the cumulative distribution function (CDF) through Silverman-rule bandwidth selection and FFT convolution, then obtains the probability density function (PDF) via numerical differentiation of the CDF: + +$$ +P _ {\mathrm {P M F}} (\hat {y}) = P _ {\mathrm {C D F}} (\hat {y} + \frac {1}{2}) - P _ {\mathrm {C D F}} (\hat {y} - \frac {1}{2}). \tag {9} +$$ + +To ensure optimal RD performance at each level, the keyframe optimization loss function $\mathcal{L}_{\mathrm{key}}$ is formulated as a weighted sum of per-level losses $\mathcal{L}_{\mathrm{key}}^l$ , where each level's loss combines a photometric + +![](images/4f7b8d49694fd5567dbe15bf2a9e2cb43c95e6ff9b01fbb014f05e2a66ee1728.jpg) + +![](images/a2dceae20ab2b16f7a690f7049c82bc30a19c64b894c78b36f02e16c40ab5ce4.jpg) + +![](images/25f6879e400fe0a45853059673757d135a7f08290f62cdc3e29a9d750b2e7197.jpg) +\( \left\{ {1,2,3}\right\} \Rightarrow \;\left\{ {1,3,4}\right\} \Rightarrow \;\left\{ {2,3,4}\right\} \Rightarrow \;\left\{ {3,4}\right\} \Rightarrow \left\{ {4,5}\right\} \Rightarrow \left\{ {5,6}\right\} \Rightarrow \left\{ {6,7}\right\} \Rightarrow \left\{ {7,8}\right\} \Rightarrow \left\{ {8,9}\right\} \Rightarrow \left\{ {9,10}\right\} \Rightarrow \left\{ {10,11}\right\} \Rightarrow \left\{ {11,12}\right\} \Rightarrow \left\{ {12,13}\right\} \Rightarrow \left\{ {13,14}\right\} \Rightarrow \left\{ {14,15}\right\} \Rightarrow \left\{ {15,16}\right\} \Rightarrow \left\{ {16,17}\right\} \Rightarrow \left\{ {17,18}\right\} \Rightarrow \left\{ {18,19}\right\} \Rightarrow \left\{ {19,20}\right\} \Rightarrow \left\{ {20,21}\right\} \Rightarrow \left\{ {21,22}\right\} \Rightarrow \left\{ {22,23}\right\} \Rightarrow \left\{ {23,24}\right\} \Rightarrow \left\{ {24,25}\right\} \Rightarrow \left\{ {25,26}\right\} \Rightarrow \left\{ {26,27}\right\} \Rightarrow \left\{ {27,28}\right\} \Rightarrow \left\{ {28,29}\right\} \Rightarrow \left\{ {29,30}\right\} \Rightarrow \left\{ {30,31}\right\} \Rightarrow \left\{ {31,32}\right\} \Rightarrow \left\{ {32,33}\right\} \Rightarrow \left\{ {33,34}\right\} \Rightarrow \left\{ {34,35}\right\} \Rightarrow \left\{ {35,36}\right\} \Rightarrow \left\{ {36,37}\right\} \Rightarrow \left\{ {37,38}\right\} \Rightarrow \left\{ {38,39}\right\} \Rightarrow \left\{ {39,40}\right\} \Rightarrow \left\{ {40,41}\right\} \Rightarrow \left\{ {41,42}\right\} \Rightarrow } + +![](images/8c78e164f0f354d0747cd0b4b29bcc70f5f66310d791074967fcc0b53daea804.jpg) +8.10MB + +![](images/c085cd7df2546d7b3c3c788e03b924f8589ec700ed8149af96a22fd452e05678.jpg) +$\therefore m = \frac{3}{11}$ + +![](images/f1fa95d27290e6668d23cf87818fb8172baa4c371525470773728f2195f29e0a.jpg) + +![](images/409de40092a82f748e792c09c34cd9722b8a6d1a46d62bd175770073d72f72d9.jpg) +Figure 4: Qualitative comparison on our 4DGCPro and HiFi4G [26] datasets against ReRF [66], HPC [80], 3DGStream [61] and $\mathrm{V}^3$ [69]. + +![](images/2c52face964b318b723a9dcab01bfb4c95540086b59a10dfd681e1077e734ebb.jpg) + +![](images/4ddcf234f0ff0bff2612f4b879541516b4a2073b161bb3a6e67ee40c89bd312a.jpg) + +![](images/e275718cccdbb0fec82f1c0bb81ec26f5c92a16a2e459dc2591015a82cf632cd.jpg) + +![](images/61a20629fddb22d26b8f37a792b5401d35103aae10fc839b486f51b4dc4c03a7.jpg) +0.95MB +V3 + +![](images/fabeedafa56afc6d1556a9a3f75b677bfd5b478e380d091a7448bcc8a2653d15.jpg) +0.60MB +Ours + +term $\mathcal{L}_{\mathrm{color}}^l$ and a rate term $\mathcal{L}_{\mathrm{rate\_key}}^l$ : + +$$ +\mathcal {L} _ {\text {k e y}} = \sum_ {l = 1} ^ {L} \lambda^ {l} \mathcal {L} _ {\text {k e y}} ^ {l} = \sum_ {l = 1} ^ {L} \lambda^ {l} \left(\mathcal {L} _ {\text {c o l o r}} ^ {l} + \lambda_ {\text {r a t e} \_ \text {k e y}} \mathcal {L} _ {\text {r a t e} \_ \text {k e y}} ^ {l}\right), \tag {10} +$$ + +$$ +\mathcal {L} _ {\text {c o l o r}} ^ {l} = \left(1 - \lambda_ {\text {s s i m}}\right) \| \mathbf {c} _ {g} - \hat {\mathbf {c}} ^ {l} \| _ {1} + \lambda_ {\text {s s i m}} \mathcal {L} _ {\mathrm {D} - \text {S S I M}} ^ {l}, \tag {11} +$$ + +$$ +\mathcal {L} _ {\text {r a t e} \_ \text {k e y}} ^ {l} = - \frac {1}{N} \sum_ {\hat {\mathbf {y}} _ {t} ^ {l} \in \left\{\hat {\mathbf {R}} _ {t} ^ {l}, \hat {\mathbf {s}} _ {t} ^ {l}, \hat {\mathbf {f}} _ {t} ^ {l}, \hat {\alpha} _ {t} ^ {l} \right\}} \log_ {2} \left(P _ {\mathrm {P M F}} (\hat {y} _ {t} ^ {l})\right). \tag {12} +$$ + +Here, $\mathcal{L}_{\mathrm{color}}^l$ measures the photometric difference between the ground truth and the Gaussian rendering results at level $l$ , $\mathcal{L}_{\mathrm{rate\_key}}^l$ denotes the entropy estimated via KDE from Gaussian attributes at the same level. $\lambda_{\mathrm{rate\_key}}$ is the weight of the entropy loss, and $\lambda^l$ is the weight parameter for the loss at different levels. With this training objective, we obtain the keyframe Gaussians that achieve the optimal RD performance at each level. + +Inter-frame Optimization. Building upon the hierarchically trained Gaussians of keyframes, we optimize subsequent Gaussians within each group. Since Gaussian positions and rotations are particularly crucial for rendering quality, we only employ simulated quantization deliberately exclude entropy constraints and rely solely on $\mathcal{L}_{\mathrm{color}}$ for supervision. + +In the residual deformation stage, to maximize both accuracy and compactness at each level, we maintain hierarchical supervision by augmenting color constraints with both entropy loss $\mathcal{L}_{\mathrm{rate\_inter}}^l$ and temporal loss $\mathcal{L}_{\mathrm{reg}}^l$ . As illustrated in Fig. 3(c), we validate the Gaussian distribution of residual attributes, which allows us to simplify the entropy estimation to merely calculating the mean and variance of residuals, significantly streamlining training. To further enhance temporal coherence, we impose the temporal loss on residual attributes, explicitly enforcing inter-frame consistency. This deliberate smoothness constraint not only improves reconstruction quality but also reduces residual magnitudes during subsequent coding, ultimately optimizing storage efficiency. Thus, the training objective $\mathcal{L}_{\mathrm{inter}}$ for this stage can be summarized as: + +$$ +\mathcal {L} _ {\text {i n t e r}} = \sum_ {l = 1} ^ {L} \lambda^ {l} \mathcal {L} _ {\text {i n t e r}} ^ {l} = \sum_ {l = 1} ^ {L} \lambda^ {l} \left(\mathcal {L} _ {\text {c o l o r}} ^ {l} + \lambda_ {\text {r a t e - i n t e r}} \mathcal {L} _ {\text {r a t e - i n t e r}} ^ {l} + \lambda_ {\text {r e g}} \mathcal {L} _ {\text {r e g}} ^ {l}\right), \tag {13} +$$ + +$$ +\mathcal {L} _ {\text {r a t e} \_ \text {i n t e r}} ^ {l} = - \frac {1}{N} \sum_ {\hat {y} _ {t} ^ {l} \in \left\{\Delta \hat {\mathrm {s}} _ {t} ^ {l}, \Delta \hat {\mathrm {f}} _ {t} ^ {l}, \Delta \hat {\alpha} _ {t} ^ {l} \right\}} \log_ {2} \left(P _ {\mathrm {P M F}} \left(\hat {y} _ {t} ^ {l}\right)\right), \tag {14} +$$ + +$$ +\mathcal {L} _ {\text {r e g}} ^ {l} = \sum_ {\hat {y} _ {t} ^ {l} \in \left\{\Delta \hat {\mathbf {s}} _ {t} ^ {l}, \Delta \hat {\mathbf {f}} _ {t} ^ {l}, \Delta \hat {\alpha} _ {t} ^ {l} \right\}} | | \hat {y} _ {t} ^ {l} | | _ {1}, \tag {15} +$$ + +where $\mathcal{L}_{\mathrm{inter}}^l$ denotes the inter-frame loss for the $l$ -th layer of Gaussians, while $\mathcal{L}_{\mathrm{rate\_inter}}^l$ and $\mathcal{L}_{\mathrm{reg}}^l$ are weighted by $\lambda_{\mathrm{rate\_inter}}$ and $\lambda_{\mathrm{reg}}$ , respectively. Through this joint entropy-optimized training framework, we obtain a compact yet high-fidelity hierarchical 4D Gaussian representation, enabling efficient volumetric video compression for storage and transmission. + +Efficient Progressive Bitstream Generation. Once the training is completed, we explicitly separate Gaussians at different levels and implement differential quantization for Gaussian attributes, where we + +Table 1: Quantitative comparison on our 4DGCPro, HiFi4G [26] and N3DV [31] datasets. Our method achieves the best rendering quality against other methods, achieving a progressive rendering results within one single model. + +
Method4DGCProHiFi4G[26]N3DV[31]
PSNR(dB)↑SSIM↑Size(MB)↓PSNR(dB)↑SSIM↑Size(MB)↓PSNR(dB)↑SSIM↑Size(MB)↓
ReRF[66]27.570.9471.7030.300.9770.9729.710.9180.77
HPC[80]27.680.9481.0834.140.9870.72---
3DGStream[61]21.080.8378.121.020.9468.131.540.9428.10
4DGC[20]21.480.8500.9721.050.9460.9431.580.9430.50
HiCoM[16]24.650.9262.6129.370.9681.9431.170.9390.70
V3[69]28.110.9551.6036.260.9940.92---
Ours(High)29.470.9631.3136.380.9950.7531.640.9440.64
Ours(Mid)28.680.9580.6635.480.9910.3731.140.9380.43
Ours(Low)27.690.9520.3334.620.9880.1930.680.9260.21
+ +![](images/0e7b609443bd4b850df78231eb817056fe871c777934afb3c6c355fc0dbfce26.jpg) +Figure 5: Rate-distortion curves across various datasets. Rate-distortion curves not only illustrate the superiority of our method compared to the multiple-bitrate approaches ReRF [66], HPC [80], 4DGC [20], and $\mathrm{V}^3$ [69], but also demonstrate the efficiency of various components within our method. + +employ uint16 or uint32 precision for position information due to its higher sensitivity to errors while using uint8 for all other attributes. Since each Gaussian parameter contains multiple channels, we carefully flatten each feature channel into a separate 2D single-channel image while strictly preserving the 2D spatial continuity. The flattened feature images are systematically arranged into temporal sequences by aligning same-group, same-level, and same-channel features. These sequences are then compressed using an H.264 video encoder, enabling scalable real-time decoding and rendering via hardware video codecs and shaders. These bitstreams are transmitted collectively, enabling clients to selectively receive and decode different quality levels for adaptive rendering, thereby supporting smooth quality transitions, flexible viewing experiences, and real-time presentation. + +# 4 Experiments + +To comprehensively evaluate our method, we conducted experiments on two kinds of distinct datasets: (1) the N3DV dataset [31] featuring subtle motions with background, and (2) the HiFi4G dataset [26] containing complex motions without background. We additionally captured a new dataset using 81 synchronized Z-CAM cinema cameras $3840 \times 2160$ , recording diverse performances including dance, sports, and instrument playing. Our dataset contains not only solo performances but also multi-person interactions, which poses higher demands on the modeling method. We further introduce the detailed experimental settings in Sec. A. + +# 4.1 Comparison + +To validate the effectiveness of the proposed method, we compare it against several SOTA approaches including NeRF-based methods ReRF [66], HPC [80] and 3DGS-based methods 3DGStream [61], 4DGC [20], HiCoM [16], $\mathrm{V}^3$ [69], presenting results in Fig. 4. It can be observed that due to the limitations of the finite neural representation of NeRF when dealing with complex motions, both ReRF [66] and HPC [80] produce blurry results and over-smoothing. Meanwhile, 3DGStream [61] is limited to modeling only rigid motion of Gaussians and relies exclusively on the first frame as a universal reference across all frames. This leads to severe error accumulation over time, particularly in regions with motion, causing pronounced visual artifacts. Due to its inability to capture non-rigid deformations and significant displacements, the approach produces inconsistencies, including trajectory fragmentation and residual errors propagated from previous frames. Compared with + +Table 2: The BD-PSNR results of our 4DGCPro, HPC [80], 4DGC [20] and $\mathbf{V}^3$ [69] when compared with ReRF [66] on different datasets. + +
Dataset4DGCProHiFi4G[26]N3DV[31]
MethodHPC4DGCV3OursHPC4DGCV3OursHPC4DGCV3Ours
BD-PSNR(dB)↑3.42-6.151.904.205.84-9.107.197.87-1.99-2.07
+ +Table 3: Complexity comparison of our method with dynamic scene compression methods, ReRF [66], HPC [80], 4DGC [20] and $\mathrm{V}^3$ [69] on 4DGCPro dataset. + +
TimeReRF[66]4DGC[20]V3[69]HPC[80]Ours
HighMidLowHighMidLow
Encode(ms)8202700390330028702420408205102
Decode(ms)61942012110390291912
Train(min)42.730.830.979393934.34.34.3
Render(ms)525.62.82311671103.12.52.2
+ +$\mathrm{V}^3$ [69], our method achieves better reconstruction quality while reducing the model bitrate, and it can render multi-quality reconstruction results from a single model. We further provide more demonstrations of our method in Sec. B.1. + +For quantitative comparison, as demonstrated in Tab. 1, our method achieves superior performance compared to other approaches on diverse datasets. On the 4DGCPro dataset, our method achieves superior performance across all quality levels: the high-quality model attains the best PSNR (29.47dB) and SSIM (0.963) with compact size (1.31MB); the medium-quality version maintains high PSNR performance (28.68dB) with improved compression (0.66MB); while the low-quality configuration further reduces model size to 0.33MB while retaining competitive quality (27.69dB). Notably, our framework supports multiple bitrates within a single model while outperforming baselines on all metrics. The rate-distortion superiority of our approach is further demonstrated in Fig. 5 and quantitatively validated through BD-PSNR measurements in Tab. 2. Our method achieves consistent RD improvements across all datasets and bitrates, with BD-PSNR gains of 4.20dB, 7.87dB, and 2.07dB over ReRF on the 4DGCPro, HiFi4G, and N3DV datasets, respectively. These results significantly exceed those of other compared methods, including HPC (3.42dB on 4DGCPro) and $\mathrm{V}^3$ (1.90dB on 4DGCPro). The superior RD performance demonstrates the effectiveness of our significance metric, adaptive grouping, and entropy modeling strategies. + +As validated in Tab. 3, our method demonstrates exceptional computational efficiency across all quality levels. The medium-quality configuration achieves 19ms decoding and 2.5ms rendering per frame, enabling real-time performance at over 52 FPS, while even the high-quality setting maintains practical efficiency with 29ms decoding and 3.1ms rendering. As shown in Tab. 7, our approach also delivers remarkable performance on lightweight devices: on mobile platforms, the complete pipeline requires only 43ms for high-quality rendering, reduced to 39ms and 34ms for medium and low quality, demonstrating real-time decoding and rendering capability even under strict resource constraints. + +These results collectively demonstrate that our method achieves the best trade-off between reconstruction quality, compression ratio, and computational efficiency among all compared approaches. The progressive coding capability further enhances practical applicability, enabling adaptive quality adjustment based on available computational resources and bandwidth constraints. + +# 4.2 Ablation Studies + +We conducted four ablation studies to evaluate the effectiveness of each component of our method. These experiments focus on the significance metric, motion-aware adaptive Gaussian grouping, the number of Gaussian layers, and joint entropy-optimized training. Using the full model as the baseline, we first ablated the components of the significance metric, including the weight parameter $\lambda_{\Psi}$ . We then compared our adaptive grouping strategy against different fixed group lengths. The third experiment examines the impact of different number of Gaussian layers. Finally, we assessed various entropy modeling methods and underscored the importance of simulated quantization. + +Tab. 4 presents ablation results for the significance metric and adaptive grouping strategy. The left section evaluates the significance metric for low-level Gaussians. Compared to our full model (27.69dB), using only opacity or volume leads to clear performance degradation (26.71dB and 25.83dB, respectively). Simply multiplying these two factors also causes a significant PSNR drop of -1.33dB. Furthermore, improper weighting of $\lambda_{\Psi}$ results in measurable PSNR reductions ranging + +Table 4: Ablation studies of our perceptually-weighted hierarchical Gaussian representation and adaptive Gaussian grouping. + +
Significance MetricPSNR(dB)↑Size(MB)↓
w/o Opacity26.710.39
w/o Volume25.830.38
Multiplication26.360.33
λΨ=2×10527.510.34
λΨ=5×10427.570.35
Ours(Full)27.690.33
+ +
Group SizeBDBR(%)↓BD-PSNR(dB)↑
148.37-0.96
511.81-0.25
1016.34-0.32
1514.95-0.33
2011.42-0.34
258.11-0.31
+ +Table 5: Ablation studies of the number of layers and end-to-end entropy-optimized training scheme. + +
LBD-PSNR(dB)↑Training Time(min)↓
4-0.873.1
5-0.383.5
6-4.3
70.064.9
80.095.5
+ +
TrainingBDBR(%)↓BD-PSNR(dB)↑
w/o R-E32.73-0.60
w/o H-S61.21-2.89
w/o S-Q4.36-0.1
Only KDE0.58-0.02
Only Gaussian--
+ +from -0.18dB to -0.12dB. The right section validates our motion-aware adaptive grouping approach. Even the best-performing fixed group size strategy shows consistent degradation, with a minimum BDBR of $8.11\%$ and a maximum BD-PSNR reduction of -0.25dB, confirming the clear advantage of our adaptive grouping method in rate-distortion performance. + +Table 5 examines the impact of the number of Gaussian layers and the entropy modeling strategy. The left panel shows that deeper hierarchies (e.g., $L = 8$ ) only slightly improve BD-PSNR (up to 0.09dB) at the cost of a noticeable increase in training time (+1.2 min). In contrast, shallower configurations (e.g., $L = 4$ ) lead to a more substantial degradation in BD-PSNR (-0.87dB). The right panel highlights the importance of entropy modeling: omitting it ("w/o R-E") introduces redundancy, while removing hierarchical supervision ("w/o H-S") severely degrades the quality of low-level Gaussians, resulting in a BD-PSNR drop of -2.89dB. Moreover, the absence of simulated quantization (S-Q) leads to a 4.36% increase in BDBR, confirming its essential role in enhancing resilience to quantization errors. For Gaussian parameter modeling, using only KDE estimation ("Only KDE") achieves similar performance but prolongs training by 1.2 minutes per frame, whereas assuming a universal Gaussian distribution ("Only Gaussian") causes training failures. Additional ablation studies are provided in Section B.3. + +# 5 Discussion + +Limitation. Although 4DGCPro presents an innovative and efficient approach to progressive streaming of volumetric video, it has several limitations. First, the Gaussian optimization process suffers from prolonged training times (several minutes) due to hierarchical supervision which leads to repeated rendering passes. Accelerating this procedure remains an essential research objective. Second, our method relies on multi-view video input and faces challenges in sparse-view reconstruction, limiting its applicability in scenarios with insufficient camera coverage. Finally, the current framework underperforms in spatially extensive scenes, necessitating further exploration to enhance its scalability. + +Conclusion. We present 4DGCPro, a novel hierarchical 4D Gaussian compression approach for progressive volumetric video streaming. Our framework accomplishes multiple bitrate control within a single model while supporting both real-time decoding and high-fidelity rendering on mobile platforms, even for sequences containing large motion displacements. Our approach begins by constructing a perceptually-weighted hierarchical Gaussian representation using the importance metric. We then model inter-frame Gaussians by rigid transformations and residual deformations, enhanced by a motion-aware adaptive Gaussian grouping strategy for efficient sequence-wide modeling. Furthermore, we introduce a joint entropy-optimized training and progressive coding framework, employing attribute-specific entropy modeling to ensure precise and efficient optimization. Thanks to its multiple bitrate capability, 4DGCPro enables progressive streaming and high-efficiency decoding/rendering across multiple quality levels, making it ideal for bandwidth-fluctuating scenarios. This work establishes a critical foundation for broader volumetric video adoption. + +# 6 Acknowledgements + +This work is supported by National Natural Science Foundation of China (62571322, 62431015, 62271308), STCSM (24ZR1432000, 24511106902, 24511106900, 22DZ2229005), 111 plan (BP0719010), and State Key Laboratory of UHD Video and Audio Production and Presentation. + +# References + +[1] Akhtar, A., Gao, W., Li, L., Li, Z., Jia, W., & Liu, S. (2022) Video-based point cloud compression artifact removal. IEEE Transactions on Multimedia 24:2866-2876. +[2] Barron, J. T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., & Srinivasan, P. P. (2021) Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. ICCV +[3] Barron, J. T., Mildenhall, B., Verbin, D., Srinivasan, P. P., & Hedman, P. (2022) Mip-nerf 360: Unbounded anti-aliased neural radiance fields. CVPR +[4] Barron, J. T., Mildenhall, B., Verbin, D., Srinivasan, P. P., & Hedman, P. (2023) Zip-nerf: Anti-aliased grid-based neural radiance fields. ICCV +[5] Cao, A. & Johnson, J. (2023) Hexplane: A fast representation for dynamic scenes. In CVPR pages 130-141. +[6] Charatan, D., Li, S., Tagliasacchi, A., & Sitzmann, V. (2023) pixelsplat: 3d gaussian splats from image pairs for scalable generalizable 3d reconstruction. In arXiv +[7] Chen, Y. & Lee, G. H. (2023) Dbarf: Deep bundle-adjusting generalizable neural radiance fields. In CVPR pages 24-34. +[8] Dai, A., Nießner, M., Zollhöfer, M., Izadi, S., & Theobalt, C. (2017) Bundlefusion: Real-time globally consistent 3d reconstruction using on-the-fly surface reintegration. ACM Transactions on Graphics (ToG) 36(4): 1. +[9] Deng, C. L. & Tartaglione, E. (2023) Compressing explicit voxel grid representations: fast nerfs become also small. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision pages 1236-1245. +[10] Du, Y., Zhang, Y., Yu, H.-X., Tenenbaum, J. B., & Wu, J. (2021) Neural radiance flow for 4d view synthesis and video processing. In Proceedings of the IEEE/CVF International Conference on Computer Vision +[11] Fan, Z., Wang, K., Wen, K., Zhu, Z., Xu, D., & Wang, Z. (2024) Lightgaussian: Unbounded 3d gaussian compression with 15x reduction and 200+ fps. In Proc. Advances in Neural Information Processing Systems (NeurIPS) +[12] Fang, J., Yi, T., Wang, X., Xie, L., Zhang, X., Liu, W., Nießner, M., & Tian, Q. (2022) Fast dynamic radiance fields with time-aware neural voxels. In SIGGRAPH Asia 2022 Conference Papers ACM. +[13] Feng, G., Chen, S., Fu, R., Liao, Z., Wang, Y., Liu, T., Pei, Z., Li, H., Zhang, X., & Dai, B. 2024. +[14] Fridovich-Keil, S., Meanti, G., Warburg, F. R., Recht, B., & Kanazawa, A. (2023) K-planes: Explicit radiance fields in space, time, and appearance. In CVPR pages 12479-12488. +[15] Fu, C., Li, G., Song, R., Gao, W., & Liu, S. (2022) Octattention: Octree-based large-scale contexts model for point cloud compression. In the AAAI Conference on Artificial Intelligence 36, pp. 625-633. +[16] Gao, Q., Meng, J., Wen, C., Chen, J., & Zhang, J. (2024.) Hicom: Hierarchical coherent motion for dynamic streamable scenes with 3d gaussian splatting. In Advances in Neural Information Processing Systems (NeurIPS) +[17] Gao, Z., Planche, B., Zheng, M., Choudhuri, A., Chen, T., & Wu, Z. 2024. +[18] Girish, S., Li, T., Mazumdar, A., Shrivastava, A., Luebke, D., & De Mello, S. (2024) Queen: Quantized efficient encoding of dynamic gaussians for streaming free-viewpoint videos. In A. Globerson, L. Mackey, D. Belgrave, A. Fan, U. Paquet, J. Tomczak, and C. Zhang, (eds.), Advances in Neural Information Processing Systems 37, pp. 43435-43467. Curran Associates, Inc. +[19] Guo, Z., Zhou, W., Li, L., Wang, M., & Li, H. (2024) Motion-aware 3d gaussian splatting for efficient dynamic scene reconstruction. ArXiv abs/2403.11447. + +[20] Hu, Q., Zheng, Z., Zhong, H., Fu, S., Song, L., XiaoyunZhang, Zhai, G., & Wang, Y. (2025.) 4dgc: Rate-aware 4d gaussian compression for efficient streamable free-viewpoint video. In CVPR +[21] Hu, Q., Zhong, H., Zheng, Z., Zhang, X., Cheng, Z., Song, L., Zhai, G., & Wang, Y. (2025.) Vrvvc: Variable-rate nef-based volumetric video compression. Proceedings of the AAAI Conference on Artificial Intelligence 39(4):3563-3571. +[22] Huang, B., Yu, Z., Chen, A., Geiger, A., & Gao, S. (2024) 2d gaussian splatting for geometrically accurate radiance fields. In SIGGRAPH 2024 Conference Papers Association for Computing Machinery. +[23] Hollein, L., Božić, A., Zollhöfer, M., & Nießner, M. 2024. +[24] Isik, M., Rünz, M., Georgopoulos, M., Khakhulin, T., Starck, J., Agapito, L., & Nießner, M. (2023) Humanrf: High-fidelity neural radiance fields for humans in motion. ACM Transactions on Graphics (TOG) 42 (4). +[25] Jiang, Y., Shen, Z., Hong, Y., Guo, C., Wu, Y., Zhang, Y., Yu, J., & Xu, L. (2024.) Robust dual gaussian splatting for immersive human-centric volumetric videos. arXiv preprint arXiv:2409.08353 +[26] Jiang, Y., Shen, Z., Wang, P., Su, Z., Hong, Y., Zhang, Y., Yu, J., & Xu, L. (2024.) Hifi4g: High-fidelity human performance rendering via compact gaussian splatting. In CVPR pages 19734-19745. +[27] Kerbl, B., Kopanas, G., Leimkuhler, T., & Drettakis, G. (2023) 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics 42(4). +[28] Levoy, M. & Hanrahan, P. Light Field Rendering. Association for Computing Machinery, New York, NY, USA, 2023. +[29] Li, L., Li, Z., Zakharchenko, V., Chen, J., & Li, H. (2020) Advanced 3d motion prediction for video-based dynamic point cloud compression. IEEE Transactions on Image Processing 29:289-302. +[30] Li, L., Shen, Z., Wang, Z., Shen, L., & Tan, P. (2022.) Streaming radiance fields for 3d video synthesis. Advances in Neural Information Processing Systems 35:13485-13498. +[31] Li, T., Slavcheva, M., Zollhoefer, M., Green, S., Lassner, C., Kim, C., Schmidt, T., Lovegrove, S., Goesele, M., Newcombe, R., & others (2022.) Neural 3d video synthesis from multi-view video. In CVPR pages 5521-5531. +[32] Li, Z., Niklaus, S., Snavely, N., & Wang, O. (2021) Neural scene flow fields for space-time view synthesis of dynamic scenes. In CVPR pages 6494-6504. +[33] Li, Z., Wang, Q., Cole, F., Tucker, R., & Snavely, N. (2023) Dynibar: Neural dynamic image-based rendering. In CVPR pages 4273-4284. +[34] Li, Z., Chen, Z., Li, Z., & Xu, Y. (2024) Spacetime gaussian feature splatting for real-time dynamic view synthesis. In CVPR pages 8508-8520. +[35] Liang, Z. & Liang, F. (2022) Transpcc: Towards deep point cloud compression via transformers. In Proceedings of the 2022 International Conference on Multimedia Retrieval page 1-5, New York, NY, USA: Association for Computing Machinery. +[36] Loza, A., Mihaylova, L., Canagarajah, N., & Bull, D. (2006) Structural similarity-based object tracking in video sequences. In 2006 9th International Conference on Information Fusion pages 1-6. +[37] Lu, T., Yu, M., Xu, L., Xiangli, Y., Wang, L., Lin, D., & Dai, B. (2024) Scaffold-gs: Structured 3d gaussians for view-adaptive rendering. In CVPR pages 20654-20664. +[38] Luiten, J., Kopanas, G., Leibe, B., & Ramanan, D. (2024) Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. In 3DV +[39] Martin-Brualla, R., Radwan, N., Sajjadi, M. S. M., Barron, J. T., Dosovitskiy, A., & Duckworth, D. (2021) NeRF in the Wild: Neural Radiance Fields for Unconstrained Photo Collections. In CVPR +[40] Mildenhall, B., Srinivasan, P. P., Tancik, M., Barron, J. T., Ramamoorthi, R., & Ng, R. (2021) Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65(1):99-106. +[41] Müller, T., Evans, A., Schied, C., & Keller, A. (2022) Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph. 41(4):102:1-102:15. + +[42] Nadenau, M., Reichel, J., & Kunt, M. (2003) Wavelet-based color image compression: Exploiting the contrast sensitivity function. IEEE transactions on image processing : a publication of the IEEE Signal Processing Society 12:58-70. +[43] Navaneet, K., Meibodi, K. P., Koohpayegani, S. A., & Pirsiavash, H. (2024) Compgs: Smaller and faster gaussian splatting with vector quantization. ECCV +[44] Overbeck, R. S., Erickson, D., Evangelakos, D., Pharr, M., & Debevec, P. (2018) A system for acquiring, processing, and rendering panoramic light field stills for virtual reality. ACM Trans. Graph. 37(6). +[45] Park, K., Sinha, U., Barron, J. T., Bouaziz, S., Goldman, D. B., Seitz, S. M., & Martin-Brualla, R. (2021) Nerfies: Deformable neural radiance fields. In ICCV (ICCV) pages 5865-5874. +[46] Park, K., Henzler, P., Mildenhall, B., & Barron, R. (2023.) Camp: Camera preconditioning for neural radiance fields. ACM Trans. Graph. +[47] Park, S., Son, M., Jang, S., Ahn, Y. C., Kim, J.-Y., & Kang, N. (2023.) Temporal interpolation is all you need for dynamic neural radiance fields. CVPR pages 4212-4221. +[48] Peng, S., Yan, Y., Shuai, Q., Bao, H., & Zhou, X. (2023) Representing volumetric videos as dynamic mlp maps. In CVPR pages 4252-4262. +[49] Pumarola, A., Corona, E., Pons-Moll, G., & Moreno-Noguer, F. (2020) D-NeRF: Neural Radiance Fields for Dynamic Scenes. In CVPR +[50] Quach, M., Valenzise, G., & Dufaux, F. (2019) Learning convolutional transforms for lossy point cloud geometry compression. In 2019 IEEE International Conference on Image Processing (ICIP) pages 4320-4324. +[51] Quach, M., Valenzise, G., & Dufaux, F. (2020) Improved deep point cloud geometry compression. In 2020 IEEE 22nd International Workshop on Multimedia Signal Processing (MMSP) pages 1-6. +[52] Rabich, S., Stotko, P., & Klein, R. (2023) Fpo++: Efficient encoding and rendering of dynamic neural radiance fields by analyzing and enhancing fourier plenoctrees. arXiv preprint arXiv:2310.20710 +[53] Reiser, C., Szeliski, R., Verbin, D., Srinivasan, P., Mildenhall, B., Geiger, A., Barron, J., & Hedman, P. (2023) Merf: Memory-efficient radiance fields for real-time view synthesis in unbounded scenes. ACM Transactions on Graphics (TOG) 42(4):1-12. +[54] Rho, D., Lee, B., Nam, S., Lee, J. C., Ko, J. H., & Park, E. (2023) Masked wavelet representation for compact neural radiance fields. In CVPR pages 20680-20690. +[55] Schnabel, R. & Klein, R. (2006) Octree-based point-cloud compression. In Proceedings of the 3rd Eurographics / IEEE VGTC Conference on Point-Based Graphics page 111-121, Goslar, DEU: Eurographics Association. +[56] Schwarz, S., Preda, M., Baroncini, V., Budagavi, M., Cesar, P., Chou, P. A., Cohen, R. A., Krivokuca, M., Lasserre, S., Li, Z., Llach, J., Mamou, K., Mekuria, R., Nakagami, O., Siahaan, E., Tabatabai, A., Tourapis, A. M., & Zakharchenko, V. (2019) Emerging mpeg standards for point cloud compression. IEEE Journal on Emerging and Selected Topics in Circuits and Systems 9(1):133-148. +[57] Schonberger, J. L. & Frahm, J.-M. (2016) Structure-from-motion revisited. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) pages 4104-4113. +[58] Shao, R., Zheng, Z., Tu, H., Liu, B., Zhang, H., & Liu, Y. (2023) Tensor4d: Efficient neural 4d decomposition for high-fidelity dynamic reconstruction and rendering. In CVPR pages 16632-16642. +[59] Song, L., Chen, A., Li, Z., Chen, Z., Chen, L., Yuan, J., Xu, Y., & Geiger, A. (2023) Nerfplayer: A streamable dynamic scene representation with decomposed neural radiance fields. IEEE Transactions on Visualization and Computer Graphics 29(5):2732-2742. +[60] Su, Z., Xu, L., Zheng, Z., Yu, T., Liu, Y., & Fang, L. (2020) Robustfusion: Human volumetric capture with data-driven visual cues using a rgbd camera. In Computer Vision - ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part IV page 246-264, Berlin, Heidelberg: Springer-Verlag. +[61] Sun, J., Jiao, H., Li, G., Zhang, Z., Zhao, L., & Xing, W. (2024) 3dgstream: On-the-fly training of 3d gaussians for efficient streaming of photo-realistic free-viewpoint videos. In CVPR pages 20675-20685. +[62] Thanou, D., Chou, P. A., & Frossard, P. (2016) Graph-based compression of dynamic 3d point cloud sequences. IEEE Transactions on Image Processing 25(4):1765-1778. + +[63] Wang, F., Tan, S., Li, X., Tian, Z., Song, Y., & Liu, H. (2023.) Mixed neural voxels for fast multi-view video synthesis. In ICCV pages 19649-19659. +[64] Wang, J., Zhu, H., Liu, H., & Ma, Z. (2021) Lossy point cloud geometry compression via end-to-end learning. IEEE Transactions on Circuits and Systems for Video Technology 31(12):4909-4923. +[65] Wang, L., Zhang, J., Liu, X., Zhao, F., Zhang, Y., Zhang, Y., Wu, M., Yu, J., & Xu, L. (2022) Fourier plenoctrees for dynamic radiance field rendering in real-time. In CVPR pages 13514-13524. +[66] Wang, L., Hu, Q., He, Q., Wang, Z., Yu, J., Tuytelaars, T., Xu, L., & Wu, M. (2023.) Neural residual radiance fields for streamably free-viewpoint videos. In CVPR pages 76-87. +[67] Wang, L., Yao, K., Guo, C., Zhang, Z., Hu, Q., Yu, J., Xu, L., & Wu, M. 2023. +[68] Wang, N., Zhang, Y., Li, Z., Fu, Y., Liu, W., & Jiang, Y.-G. (2018) Pixel2mesh: Generating 3d mesh models from single rgb images. In Proceedings of the European Conference on Computer Vision (ECCV) +[69] Wang, P., Zhang, Z., Wang, L., Yao, K., Xie, S., Yu, J., Wu, M., & Xu, L. (2024.) V^3: Viewing volumetric videos on mobiles via streamable 2d dynamic gaussians. ACM Transactions on Graphics (TOG) 43 (6):1-13. +[70] Wang, Y., Han, Q., Habermann, M., Daniilidis, K., Theobalt, C., & Liu, L. (2023.) Neus2: Fast learning of neural implicit surfaces for multi-view reconstruction. In ICCV +[71] Wang, Y., Li, Z., Guo, L., Yang, W., Kot, A. C., & Wen, B. (2024.) Contextgs: Compact 3d gaussian splatting with anchor level context model. arXiv preprint arXiv:2405.20721 +[72] Wu, G., Yi, T., Fang, J., Xie, L., Zhang, X., Wei, W., Liu, W., Tian, Q., & Wang, X. (2024.) 4d gaussian splatting for real-time dynamic scene rendering. In CVPR pages 20310-20320. +[73] Wu, M., Wang, Z., Kouros, G., & Tuytelaars, T. (2024.) Tetrif: Temporal tri-plane radiance fields for efficient free-viewpoint video. In CVPR pages 6487-6496. +[74] Xu, Z., Peng, S., Lin, H., He, G., Sun, J., Shen, Y., Bao, H., & Zhou, X. (2024) 4k4d: Real-time 4d view synthesis at 4k resolution. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pages 20029-20040. +[75] Yan, J., Peng, R., Tang, L., & Wang, R. (2024) 4d gaussian splatting with scale-aware residual field and adaptive optimization for real-time rendering of temporally complex dynamic scenes. In ACM MM pages 7871-7880. +[76] Yang, Z., Gao, X., Zhou, W., Jiao, S., Zhang, Y., & Jin, X. (2023) Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction. arXiv preprint arXiv:2309.13101 +[77] Yang, Z., Yang, H., Pan, Z., & Zhang, L. (2024) Real-time photorealistic dynamic scene representation and rendering with 4d gaussian splatting. In +[78] Zhang, J., Liu, X., Ye, X., Zhao, F., Zhang, Y., Wu, M., Zhang, Y., Xu, L., & Yu, J. (2021) Editable free-viewpoint video using a layered neural representation. ACM Transactions on Graphics (TOG) 40(4):1-18. +[79] Zhang, Z. (2012) Microsoft Kinect sensor and its effect. IEEE MultiMedia 19(2):4-10. +[80] Zheng, Z., Zhong, H., Hu, Q., Zhang, X., Song, L., Zhang, Y., & Wang, Y. (2024). Hpc: Hierarchical progressive coding framework for volumetric video. In ACM MM page 7937-7946, New York, NY, USA: Association for Computing Machinery. +[81] Zheng, Z., Zhong, H., Hu, Q., Zhang, X., Song, L., Zhang, Y., & Wang, Y. (2024.) Jointrf: End-to-end joint optimization for dynamic neural radiance field representation and compression. In 2024 IEEE International Conference on Image Processing (ICIP) pages 3292-3298. + +# NeurIPS Paper Checklist + +# 1. Claims + +Question: Do the main claims made in the abstract and introduction accurately reflect the paper's contributions and scope? + +Answer: [Yes] + +Justification: The main claims made in the abstract and introduction accurately reflect the paper's contributions and scope. + +Guidelines: + +- The answer NA means that the abstract and introduction do not include the claims made in the paper. +- The abstract and/or introduction should clearly state the claims made, including the contributions made in the paper and important assumptions and limitations. A No or NA answer to this question will not be perceived well by the reviewers. +- The claims made should match theoretical and experimental results, and reflect how much the results can be expected to generalize to other settings. +- It is fine to include aspirational goals as motivation as long as it is clear that these goals are not attained by the paper. + +# 2. Limitations + +Question: Does the paper discuss the limitations of the work performed by the authors? + +Answer: [Yes] + +Justification: The limitations has been discussed in the Discussion section. + +Guidelines: + +- The answer NA means that the paper has no limitation while the answer No means that the paper has limitations, but those are not discussed in the paper. +- The authors are encouraged to create a separate "Limitations" section in their paper. +- The paper should point out any strong assumptions and how robust the results are to violations of these assumptions (e.g., independence assumptions, noiseless settings, model well-specification, asymptotic approximations only holding locally). The authors should reflect on how these assumptions might be violated in practice and what the implications would be. +- The authors should reflect on the scope of the claims made, e.g., if the approach was only tested on a few datasets or with a few runs. In general, empirical results often depend on implicit assumptions, which should be articulated. +- The authors should reflect on the factors that influence the performance of the approach. For example, a facial recognition algorithm may perform poorly when image resolution is low or images are taken in low lighting. Or a speech-to-text system might not be used reliably to provide closed captions for online lectures because it fails to handle technical jargon. +- The authors should discuss the computational efficiency of the proposed algorithms and how they scale with dataset size. +- If applicable, the authors should discuss possible limitations of their approach to address problems of privacy and fairness. +- While the authors might fear that complete honesty about limitations might be used by reviewers as grounds for rejection, a worse outcome might be that reviewers discover limitations that aren't acknowledged in the paper. The authors should use their best judgment and recognize that individual actions in favor of transparency play an important role in developing norms that preserve the integrity of the community. Reviewers will be specifically instructed to not penalize honesty concerning limitations. + +# 3. Theory assumptions and proofs + +Question: For each theoretical result, does the paper provide the full set of assumptions and a complete (and correct) proof? + +Answer: [Yes] + +Justification: All the theorems, formulas in the paper are numbered and cross-referenced. + +# Guidelines: + +- The answer NA means that the paper does not include theoretical results. +- All the theorems, formulas, and proofs in the paper should be numbered and cross-referenced. +- All assumptions should be clearly stated or referenced in the statement of any theorems. +- The proofs can either appear in the main paper or the supplemental material, but if they appear in the supplemental material, the authors are encouraged to provide a short proof sketch to provide intuition. +- Inversely, any informal proof provided in the core of the paper should be complemented by formal proofs provided in appendix or supplemental material. +- Theorems and Lemmas that the proof relies upon should be properly referenced. + +# 4. Experimental result reproducibility + +Question: Does the paper fully disclose all the information needed to reproduce the main experimental results of the paper to the extent that it affects the main claims and/or conclusions of the paper (regardless of whether the code and data are provided or not)? + +Answer: [Yes] + +Justification: : We will open source our code to replicate our results. + +# Guidelines: + +- The answer NA means that the paper does not include experiments. +- If the paper includes experiments, a No answer to this question will not be perceived well by the reviewers: Making the paper reproducible is important, regardless of whether the code and data are provided or not. +- If the contribution is a dataset and/or model, the authors should describe the steps taken to make their results reproducible or verifiable. +- Depending on the contribution, reproducibility can be accomplished in various ways. For example, if the contribution is a novel architecture, describing the architecture fully might suffice, or if the contribution is a specific model and empirical evaluation, it may be necessary to either make it possible for others to replicate the model with the same dataset, or provide access to the model. In general, releasing code and data is often one good way to accomplish this, but reproducibility can also be provided via detailed instructions for how to replicate the results, access to a hosted model (e.g., in the case of a large language model), releasing of a model checkpoint, or other means that are appropriate to the research performed. +- While NeurIPS does not require releasing code, the conference does require all submissions to provide some reasonable avenue for reproducibility, which may depend on the nature of the contribution. For example +(a) If the contribution is primarily a new algorithm, the paper should make it clear how to reproduce that algorithm. +(b) If the contribution is primarily a new model architecture, the paper should describe the architecture clearly and fully. +(c) If the contribution is a new model (e.g., a large language model), then there should either be a way to access this model for reproducing the results or a way to reproduce the model (e.g., with an open-source dataset or instructions for how to construct the dataset). +(d) We recognize that reproducibility may be tricky in some cases, in which case authors are welcome to describe the particular way they provide for reproducibility. In the case of closed-source models, it may be that access to the model is limited in some way (e.g., to registered users), but it should be possible for other researchers to have some path to reproducing or verifying the results. + +# 5. Open access to data and code + +Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material? + +# Answer: [Yes] + +Justification: The paper provides open access to both the data and the code. + +# Guidelines: + +- The answer NA means that paper does not include experiments requiring code. +- Please see the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details. +- While we encourage the release of code and data, we understand that this might not be possible, so "No" is an acceptable answer. Papers cannot be rejected simply for not including code, unless this is central to the contribution (e.g., for a new open-source benchmark). +- The instructions should contain the exact command and environment needed to run to reproduce the results. See the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details. +- The authors should provide instructions on data access and preparation, including how to access the raw data, preprocessed data, intermediate data, and generated data, etc. +- The authors should provide scripts to reproduce all experimental results for the new proposed method and baselines. If only a subset of experiments are reproducible, they should state which ones are omitted from the script and why. +- At submission time, to preserve anonymity, the authors should release anonymized versions (if applicable). +- Providing as much information as possible in supplemental material (appended to the paper) is recommended, but including URLs to data and code is permitted. + +# 6. Experimental setting/details + +Question: Does the paper specify all the training and test details (e.g., data splits, hyperparameters, how they were chosen, type of optimizer, etc.) necessary to understand the results? + +# Answer: [Yes] + +Justification: The details of training and testing are provided in the Sec. A. + +# Guidelines: + +- The answer NA means that the paper does not include experiments. +- The experimental setting should be presented in the core of the paper to a level of detail that is necessary to appreciate the results and make sense of them. +- The full details can be provided either with the code, in appendix, or as supplemental material. + +# 7. Experiment statistical significance + +Question: Does the paper report error bars suitably and correctly defined or other appropriate information about the statistical significance of the experiments? + +# Answer: [No] + +Justification: We ran all our experiments 3 times and reported the mean metrics. However, since some of the results are directly cited from other papers, we did not include error bars in the main text to maintain consistency. Nevertheless, we reported the standard deviation of the available results in the appendix, please see Tab. 6. + +# Guidelines: + +- The answer NA means that the paper does not include experiments. +- The authors should answer "Yes" if the results are accompanied by error bars, confidence intervals, or statistical significance tests, at least for the experiments that support the main claims of the paper. +- The factors of variability that the error bars are capturing should be clearly stated (for example, train/test split, initialization, random drawing of some parameter, or overall run with given experimental conditions). +- The method for calculating the error bars should be explained (closed form formula, call to a library function, bootstrap, etc.) + +- The assumptions made should be given (e.g., Normally distributed errors). +- It should be clear whether the error bar is the standard deviation or the standard error of the mean. +- It is OK to report 1-sigma error bars, but one should state it. The authors should preferably report a 2-sigma error bar than state that they have a $96\%$ CI, if the hypothesis of Normality of errors is not verified. +- For asymmetric distributions, the authors should be careful not to show in tables or figures symmetric error bars that would yield results that are out of range (e.g. negative error rates). +- If error bars are reported in tables or plots, The authors should explain in the text how they were calculated and reference the corresponding figures or tables in the text. + +# 8. Experiments compute resources + +Question: For each experiment, does the paper provide sufficient information on the computer resources (type of compute workers, memory, time of execution) needed to reproduce the experiments? + +Answer: [Yes] + +Justification: The details of compute resources are provided in Sec. A. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The paper should indicate the type of compute workers CPU or GPU, internal cluster, or cloud provider, including relevant memory and storage. +- The paper should provide the amount of compute required for each of the individual experimental runs as well as estimate the total compute. +- The paper should disclose whether the full research project required more compute than the experiments reported in the paper (e.g., preliminary or failed experiments that didn't make it into the paper). + +# 9. Code of ethics + +Question: Does the research conducted in the paper conform, in every respect, with the NeurIPS Code of Ethics https://neurips.cc/public/EthicsGuidelines? + +Answer: [Yes] + +Justification: The research conducted in the paper conform, in every respect, with the NeurIPS Code of Ethics. + +Guidelines: + +- The answer NA means that the authors have not reviewed the NeurIPS Code of Ethics. +- If the authors answer No, they should explain the special circumstances that require a deviation from the Code of Ethics. +- The authors should make sure to preserve anonymity (e.g., if there is a special consideration due to laws or regulations in their jurisdiction). + +# 10. Broader impacts + +Question: Does the paper discuss both potential positive societal impacts and negative societal impacts of the work performed? + +Answer: [No] + +Justification: This study mainly focuses on the innovation of volumetric video reconstruction and compression algorithms, aiming to improve the reconstruction accuracy and efficiency and expand the generality of the algorithms. The datasets used in the research are all publicly available standard datasets and self-made datasets that will be made public, which do not contain any data involving personal privacy or sensitive information. + +Guidelines: + +- The answer NA means that there is no societal impact of the work performed. +- If the authors answer NA or No, they should explain why their work has no societal impact or why the paper does not address societal impact. + +- Examples of negative societal impacts include potential malicious or unintended uses (e.g., disinformation, generating fake profiles, surveillance), fairness considerations (e.g., deployment of technologies that could make decisions that unfairly impact specific groups), privacy considerations, and security considerations. +- The conference expects that many papers will be foundational research and not tied to particular applications, let alone deployments. However, if there is a direct path to any negative applications, the authors should point it out. For example, it is legitimate to point out that an improvement in the quality of generative models could be used to generate deepfakes for disinformation. On the other hand, it is not needed to point out that a generic algorithm for optimizing neural networks could enable people to train models that generate Deepfakes faster. +- The authors should consider possible harms that could arise when the technology is being used as intended and functioning correctly, harms that could arise when the technology is being used as intended but gives incorrect results, and harms following from (intentional or unintentional) misuse of the technology. +- If there are negative societal impacts, the authors could also discuss possible mitigation strategies (e.g., gated release of models, providing defenses in addition to attacks, mechanisms for monitoring misuse, mechanisms to monitor how a system learns from feedback over time, improving the efficiency and accessibility of ML). + +# 11. Safeguards + +Question: Does the paper describe safeguards that have been put in place for responsible release of data or models that have a high risk for misuse (e.g., pretrained language models, image generators, or scraped datasets)? + +Answer: [No] + +Justification: The volumetric video reconstruction and compression algorithms proposed in this study, along with the related data, carry a low risk of being maliciously misused. Both the publicly available standard datasets used in the study and the self-made datasets planned for public release do not involve privacy-sensitive information or data content that could be directly used for harmful purposes. Additionally, the algorithms of this study mainly focus on improving the reconstruction accuracy and compression efficiency, and do not have the direct ability to generate misleading content or infringe on personal rights. At the current stage, no high-risk scenarios requiring special security safeguards have been found, so no specific security protection mechanisms have been designed for the release of data or models. However, if potential risks are discovered in the future, we will actively explore and implement appropriate protective measures. + +# Guidelines: + +- The answer NA means that the paper poses no such risks. +- Released models that have a high risk for misuse or dual-use should be released with necessary safeguards to allow for controlled use of the model, for example by requiring that users adhere to usage guidelines or restrictions to access the model or implementing safety filters. +- Datasets that have been scraped from the Internet could pose safety risks. The authors should describe how they avoided releasing unsafe images. +- We recognize that providing effective safeguards is challenging, and many papers do not require this, but we encourage authors to take this into account and make a best faith effort. + +# 12. Licenses for existing assets + +Question: Are the creators or original owners of assets (e.g., code, data, models), used in the paper, properly credited and are the license and terms of use explicitly mentioned and properly respected? + +Answer: [Yes] + +Justification: The creators or original owners of assets (e.g., code, data, models), used in the paper, are properly credited, and the license and terms of use explicitly are mentioned and properly respected. + +# Guidelines: + +- The answer NA means that the paper does not use existing assets. +- The authors should cite the original paper that produced the code package or dataset. +- The authors should state which version of the asset is used and, if possible, include a URL. +- The name of the license (e.g., CC-BY 4.0) should be included for each asset. +- For scraped data from a particular source (e.g., website), the copyright and terms of service of that source should be provided. +- If assets are released, the license, copyright information, and terms of use in the package should be provided. For popular datasets, paperswithcode.com/datasets has curated licenses for some datasets. Their licensing guide can help determine the license of a dataset. +- For existing datasets that are re-packaged, both the original license and the license of the derived asset (if it has changed) should be provided. +- If this information is not available online, the authors are encouraged to reach out to the asset's creators. + +# 13. New assets + +Question: Are new assets introduced in the paper well documented and is the documentation provided alongside the assets? + +Answer: [NA] + +Justification: N/A. + +Guidelines: + +- The answer NA means that the paper does not release new assets. +- Researchers should communicate the details of the dataset/code/model as part of their submissions via structured templates. This includes details about training, license, limitations, etc. +- The paper should discuss whether and how consent was obtained from people whose asset is used. +- At submission time, remember to anonymize your assets (if applicable). You can either create an anonymized URL or include an anonymized zip file. + +# 14. Crowdsourcing and research with human subjects + +Question: For crowdsourcing experiments and research with human subjects, does the paper include the full text of instructions given to participants and screenshots, if applicable, as well as details about compensation (if any)? + +Answer: [NA] + +Justification: N/A. + +Guidelines: + +- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. +- Including this information in the supplemental material is fine, but if the main contribution of the paper involves human subjects, then as much detail as possible should be included in the main paper. +- According to the NeurIPS Code of Ethics, workers involved in data collection, curation, or other labor should be paid at least the minimum wage in the country of the data collector. + +# 15. Institutional review board (IRB) approvals or equivalent for research with human subjects + +Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals (or an equivalent approval/review based on the requirements of your country or institution) were obtained? + +Answer: [NA] + +Justification: N/A. + +Guidelines: + +- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. +- Depending on the country in which research is conducted, IRB approval (or equivalent) may be required for any human subjects research. If you obtained IRB approval, you should clearly state this in the paper. +- We recognize that the procedures for this may vary significantly between institutions and locations, and we expect authors to adhere to the NeurIPS Code of Ethics and the guidelines for their institution. +- For initial submissions, do not include any information that would break anonymity (if applicable), such as the institution conducting the review. + +# 16. Declaration of LLM usage + +Question: Does the paper describe the usage of LLMs if it is an important, original, or non-standard component of the core methods in this research? Note that if the LLM is used only for writing, editing, or formatting purposes and does not impact the core methodology, scientific rigorousness, or originality of the research, declaration is not required. + +Answer: [NA] + +Justification: N/A. + +Guidelines: + +- The answer NA means that the core method development in this research does not involve LLMs as any important, original, or non-standard components. +- Please refer to our LLM policy (https://neurips.cc/Conferences/2025/LLM) for what should or should not be described. + +![](images/67944f87689483be0bc017f12d91500d20d6e0590f6d29c18db4bbed84fb4704.jpg) +Figure 6: Gallery of our results. 4DGCPro delivers real-time high-fidelity rendering of scenes across challenging motions, such as "playing instruments", "dancing" and "playing sports". + +![](images/ca45adfaac0c8d6ff46ecf32a96b19e4f6988dc8f6ec1cccc1144600bab477cd.jpg) +Figure 7: Multi-bitrate results of our method under a single bitstream. + +This appendix provides additional material to supplement the main text. We will first introduce implementation details in Sec. A. Then we provide additional experimental results in Sec. B. + +# A Implementation Details + +Our code is primarily based on the open-source codes of 3DGS [27] and $\mathrm{V}^3$ [69] and is also inspired by 3DGStream [61] and 4DGC [20]. In the initialization phase, due to the limitations of the NeuS2 [70] method, it is challenging to obtain high-quality surface meshes for scenes with backgrounds, making it difficult to initialize Gaussians effectively. Therefore, on the N3DV dataset, we still initialize Gaussians based on the results of COLMAP. Additionally, we observed that in scenes with multiple interacting people, NeuS2 may occasionally fail to capture all individuals. To enhance the stability of our method, we train the NeuS2 network parameters of each keyframe to learn residuals from a pre-trained, known-correct NeuS2 network. During the pre-training phase of key frames, we first train for 12,000 steps under the setting of $\lambda_{\mathrm{ssim}} = 0.2$ . In the Gaussian pruning phase, we remove $40\%$ of the Gaussians with lower opacity on HiFi4G dataset [26] and 4DGCPro dataset but not remove any Gaussians on N3DV dataset [31]. During the hierarchical process, we set $\lambda_{\Psi} = 1\times 10^{5}$ to ensure a balance between volume and opacity, and divide the Gaussians into $L = 6$ layers. Regarding the motion-aware adaptive Gaussian grouping, we have selected different $\tau_{\mu}$ values for different + +![](images/111a09c16cee8040cd61c3c9f706d5ebfbaa8ea146ba4f2aaefce246a05eeb07.jpg) +Figure 8: The results of our method on long sequences. We show that the performance of our method does not decrease as the number of frames increases. + +Table 6: PSNR performance stability across three runs. Results are reported as the mean and standard deviation from three runs on the 4DGCPro dataset. + +
MethodDance1Dance2Coser1Coser2Boxing1Band1Mean
Ours(High)31.79 ±0.3729.78 ±0.1129.84 ±0.1825.65 ±0.2432.54 ±0.0627.19 ±0.1529.47 ±0.28
Ours(Mid)30.66 ±0.2829.05 ±0.1429.39 ±0.1325.40 ±0.2130.65 ±0.0826.93 ±0.1028.68 ±0.22
Ours(Low)29.26 ±0.1727.97 ±0.0928.60 ±0.0825.06 ±0.1628.99 ±0.0526.23 ±0.0827.69 ±0.14
+ +datasets: $\tau_{\mu} = 0.0025$ for 4DGCpro, $\tau_{\mu} = 0.001$ for HiFi4G, and $\tau_{\mu} = 0.01$ for N3DV. Then, we conduct end-to-end entropy-optimized training on keyframes for 1,500 steps with the supervision of $\lambda^l = \left\{ \begin{array}{ll}0.5 / l, & l < L\\ 1, & l = L \end{array} \right.$ , and $\lambda_{\mathrm{rate\_key}} = 1\times 10^{-7}$ . In the subsequent inter-frame Optimization., we set $\lambda_{\mathrm{rate\_inter}} = 1\times 10^{-4}$ , $\lambda_{\mathrm{reg}} = 1\times 10^{-3}$ , and train for 800 and 2,000 steps in the rigid transformation and residual deformation phases respectively. During the compression process, we have adopted different precisions for the compression of positions depending on the dataset. Since the Gaussian positions in the N3DV dataset have a larger range, we have quantized the Gaussian positions of the N3DV dataset using uint32 precision and compressed all attributes with qp = 10. For the other two datasets, we have applied uint16 precision and qp = 20. The H.264 encoder was configured using the x264 library with the following settings: I/P-frames only (no B-frames), 3 reference frames, color space in YUV4:4:4, and preset set to "medium." + +Our experimental setup includes an Intel(R) Xeon(R) W-2245 CPU running at 3.90 GHz and an RTX 3090 graphics card. In the experiment, we conducted evaluations on a total of 12 sequences from the 4DGCPro and N3DV datasets. Also, we selected the Greeting and Umbrella sequences from the HiFi4G dataset. We selected the 48th view as the test view in the HiFi4G and 4DGCPro datasets, while the 0th view was chosen in the N3DV dataset. Due to the relatively high resolution of the HiFi4G and 4DGCPro datasets, we downsample them by a factor of 2 for training. In the experimental results, we reproduce and run the comparison methods on the HiFi4G and 4DGCPro datasets, and the results on the N3DV dataset are reported in the papers of 4DGC and HiCoM [16]. Additionally, since the HPC [80] and $\mathrm{V}^3$ methods cannot operate properly on datasets with backgrounds, we do not report their metrics on the N3DV dataset. + +# B Additional Experimental Results + +# B.1 Addition Demonstrations of 4DGCPro + +In this section, we additionally present some results of our method to comprehensively demonstrate the advantages of our method as much as possible. First of all, in Fig. 6, the render gallery of our method is shown. It can be seen that our method can achieve high-quality reconstruction for scenarios + +Table 7: Runtime analysis on multiple platforms of rendering. + +
PlatformDesktopTabletPhone
HighMidLowHighMidLowHighMidLow
Decoding(ms)241711272210352719
Rendering(ms)2.62.32.1161411433934
+ +Table 8: Quantitative comparison of average PSNR values(dB) and model size(MB) across all sequences in the 4DGCPro dataset. + +
MethodDance1Dance2Coser1Coser2Boxing1Band1Mean
ReRF30.01/0.9826.30/2.1727.30/1.3224.35/1.9931.12/1.4626.35/2.1027.57/1.70
HPC30.78/0.8326.95/0.9226.90/0.8624.01/1.2931.00/1.1526.45/1.4227.68/1.08
3DGStream22.65/8.1018.45/8.1023.62/8.1019.64/8.1023.49/8.1018.62/8.1021.08/8.10
4DGC22.53/0.8519.77/1.0024.28/0.9619.06/0.9424.06/0.8419.18/1.2021.48/0.97
HiCoM25.06/1.6223.81/1.7224.55/2.0721.51/3.8728.87/2.9224.07/3.4724.65/2.61
V331.37/1.1029.19/1.1129.52/1.4518.97/1.8532.58/1.6127.11/2.4628.11/1.60
Ours(High)31.79/0.8929.78/0.7829.84/1.1925.65/1.8032.54/1.3327.19/1.8829.47/1.31
Ours(Mid)30.66/0.4529.05/0.3929.39/0.6025.40/0.9130.65/0.6626.93/0.9428.68/0.66
Ours(Low)29.26/0.2227.97/0.2128.60/0.3025.06/0.4528.99/0.3326.23/0.4727.69/0.33
+ +Table 9: Quantitative comparison across two sequences in the HiFi4G dataset. + +
MethodGreetingUmbrellaMean
PSNR(dB)↑Size(MB)↓PSNR(dB)↑Size(MB)↓PSNR(dB)↑Size(MB)↓
ReRF29.900.9630.690.9830.300.97
HPC35.470.7532.810.6934.140.72
3DGStream21.388.1020.658.1021.028.10
4DGC20.940.9921.150.8921.050.94
HiCoM29.121.7029.612.1829.371.94
V337.060.8935.450.9536.260.92
Ours(High)37.210.6835.520.8136.380.75
Ours(Middle)36.830.3434.130.3935.480.37
Ours(Low)35.960.1733.280.2034.620.19
+ +with many complex motions. Fig. 7 shows the multi-bitrate results of our method under a single bitstream. From the figure, it can be observed that due to the high effectiveness of our hierarchical representation and layerwise supervision, 4DGCPro maintains excellent rendering quality at each level of the results. Fig. 8 presents the multi-bitrate results of our method in a long sequence. Our method can maintain a very high reconstruction quality in long sequences. As illustrated in Fig. 9, we present supplementary qualitative results for scenes from the N3DV [31] dataset. These visual outcomes vividly showcase the resilience of our approach in accurately capturing and effectively representing dynamic scenes. In addition, to verify the stability of the method, we conducted multiple experiments under the same setting and reported the fluctuations of the results, as shown in Tab. 6. + +We also conducted an efficiency analysis of our 4DGCPro across multiple platforms. The test platforms consist of an Ubuntu desktop PC featuring an Intel i9-10920X processor and an NVIDIA GeForce RTX 3090 GPU, an Apple iPad powered by an Apple M2 processor, and an Apple iPhone equipped with an A15 Bionic processor. As presented in Tab. 7, we detail the time consumption of each thread within the rendering pipeline. During the decoding process, on the desktop, multi-threaded decoding combined with CUDA memory copying takes between 11ms and 24ms. For Apple's mobile devices, leveraging parallel decoding via compute shaders, these devices achieve a decoding time consumption comparable to that of the desktop (22ms on the tablet and 27ms on the phone). Regarding the rendering thread, the desktop with a CUDA-enabled device demonstrates an extremely rapid rendering speed (2.3ms), while mobile devices are also capable of rendering at a satisfactory frame rate (14ms and 39ms). + +# B.2 Additional Comparison Results + +We present the quantitative results for each scene from three datasets respectively in Tab. 8, 9 and 10 to provide a more detailed comparison. We categorize the datasets into two groups: large-motion, background-free datasets such as 4DGCPro and HiFi4G [26], and small-motion, background-containing datasets like N3DV [31]. Our method demonstrates a distinct advantage in the former group, which can be attributed to its precise motion modeling capabilities. In contrast, methods like 3DGStream [61] and 4DGC [20] perform poorly. Specifically, they completely loses its modeling ability after processing only a very short sequence due to the error accumulation. This shortcoming stems from its inherent limitations of using only the first frame as a reference and being capable of modeling only rigid motion. Notably, $\mathrm{V}^3$ [69] shows subpar performance on the coser2 sequence. + +Table 10: Quantitative comparison of average PSNR values(dB) and model size(MB) across all sequences in the N3DV dataset. + +
MethodCoffee MartiniCook SpinachCut BeefFlame SalmonFlame SteakSear SteakMean
ReRF26.24/0.7931.23/0.8431.82/0.8126.80/0.7832.08/0.9130.03/0.5129.71/0.77
3DGStream27.96/8.0032.88/8.0532.99/8.1928.52/8.0733.41/8.1933.58/8.1631.54/8.11
HiCoM28.04/0.8032.45/0.6032.72/0.6028.37/0.9032.87/0.6032.57/0.6031.17/0.70
4DGC27.98/0.5832.81/0.4433.03/0.4728.49/0.5133.58/0.4433.60/0.5031.58/0.49
Ours(High)27.91/0.6432.93/0.6733.10/0.6128.73/0.6133.77/0.6533.72/0.6631.64/0.64
Ours(Mid)27.63/0.4332.30/0.4532.51/0.4128.28/0.4333.09/0.4333.03/0.4431.14/0.43
Ours(Low)26.88/0.2131.95/0.2232.17/0.2027.95/0.2132.49/0.2032.64/0.2330.68/0.21
+ +Table 11: Results of more ablation studies. + +
Ablation StudiesHighMidLow
PSNR(dB)Size(MB)PSNR(dB)Size(MB)PSNR(dB)Size(MB)
w/o Motion Decomposition28.841.3128.170.6627.040.33
w/o Layer-wise Supervision29.531.3326.490.6724.980.34
Ours Full29.471.3128.680.6627.690.33
+ +The reason lies in the fact that NeuS2 [70] fails to initialize the two interacting objects correctly during the training process, resulting in only a portion of the image being successfully modeled. Our approach addresses this issue by introducing residual NeuS2. Within the N3DV dataset, given the small motion amplitudes, 3DGStream and 4DGC achieve good results. Nevertheless, our method still manages to deliver comparable performance. + +# B.3 Additional Ablation Studies + +In this section, we conducted additional experiments to validate the efficacy of motion decomposition and layer-wise supervision, as shown in Tab. 11. The results reveal that motion decomposition enables precise modeling of dynamic scenes by distinctly separating the processes of rigid transformation and residual deformation. However, simultaneously training these two components makes the positions of Gaussians less accurate and increases the training difficulty, and accurate results cannot be obtained with the same number of training steps. Regarding layer-wise supervision, while it has a marginal impact on the reconstruction quality of complete Gaussians, it yields a substantial boost in the quality of lower-level Gaussians. + +![](images/2720693c6c61c14a2c746b2368725ee933a6d200ebe74874406e645a827b95cc.jpg) +Figure 9: Qualitative results of scenes from N3DV dataset. Frames shown are the $50^{\mathrm{th}}$ , $100^{\mathrm{th}}$ , $150^{\mathrm{th}}$ , $200^{\mathrm{th}}$ , and $250^{\mathrm{th}}$ from the test video. \ No newline at end of file diff --git a/NeurIPS/2025/4DGCPro_ Efficient Hierarchical 4D Gaussian Compression for Progressive Volumetric Video Streaming/images.zip b/NeurIPS/2025/4DGCPro_ Efficient Hierarchical 4D Gaussian Compression for Progressive Volumetric Video Streaming/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..3368d5bf5572d2378cb1b689a3a31f3023e60d23 --- /dev/null +++ b/NeurIPS/2025/4DGCPro_ Efficient Hierarchical 4D Gaussian Compression for Progressive Volumetric Video Streaming/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e488e46616dd04ba56b70c154c29abeddb15621e07fff2fd85f550b5b0ca76f0 +size 1368799 diff --git a/NeurIPS/2025/4DGCPro_ Efficient Hierarchical 4D Gaussian Compression for Progressive Volumetric Video Streaming/layout.json b/NeurIPS/2025/4DGCPro_ Efficient Hierarchical 4D Gaussian Compression for Progressive Volumetric Video Streaming/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b02a08601899bf40803ea7838a914f3d565f0c10 --- /dev/null +++ b/NeurIPS/2025/4DGCPro_ Efficient Hierarchical 4D Gaussian Compression for Progressive Volumetric Video Streaming/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7e162805a67f0b674d34cf55b1a02846d11d97fc8cf2b7da838fbcd54927e894 +size 804131 diff --git a/NeurIPS/2025/4DGT_ Learning a 4D Gaussian Transformer Using Real-World Monocular Videos/95c5a7fc-10b4-46f9-9e2a-6ec1f1d22c84_content_list.json b/NeurIPS/2025/4DGT_ Learning a 4D Gaussian Transformer Using Real-World Monocular Videos/95c5a7fc-10b4-46f9-9e2a-6ec1f1d22c84_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..5e25dad8a855d626fc451ae92285071169c433d6 --- /dev/null +++ b/NeurIPS/2025/4DGT_ Learning a 4D Gaussian Transformer Using Real-World Monocular Videos/95c5a7fc-10b4-46f9-9e2a-6ec1f1d22c84_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0ffe929cde884a8ff881f0c0b8f3c769f89c8f4929fd06af3618b4923e842fb5 +size 161587 diff --git a/NeurIPS/2025/4DGT_ Learning a 4D Gaussian Transformer Using Real-World Monocular Videos/95c5a7fc-10b4-46f9-9e2a-6ec1f1d22c84_model.json b/NeurIPS/2025/4DGT_ Learning a 4D Gaussian Transformer Using Real-World Monocular Videos/95c5a7fc-10b4-46f9-9e2a-6ec1f1d22c84_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7b9f6c3049579c25f4c6318ad0cd4946cfa2c99a --- /dev/null +++ b/NeurIPS/2025/4DGT_ Learning a 4D Gaussian Transformer Using Real-World Monocular Videos/95c5a7fc-10b4-46f9-9e2a-6ec1f1d22c84_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:481f0080594238597b5a14b7b1e6d4617a1b55b4b67ed817585b65efc62dd55d +size 219110 diff --git a/NeurIPS/2025/4DGT_ Learning a 4D Gaussian Transformer Using Real-World Monocular Videos/95c5a7fc-10b4-46f9-9e2a-6ec1f1d22c84_origin.pdf b/NeurIPS/2025/4DGT_ Learning a 4D Gaussian Transformer Using Real-World Monocular Videos/95c5a7fc-10b4-46f9-9e2a-6ec1f1d22c84_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..d1abd06f3d3c019a8a8f8d6a5b52aa52c0f5d3a4 --- /dev/null +++ b/NeurIPS/2025/4DGT_ Learning a 4D Gaussian Transformer Using Real-World Monocular Videos/95c5a7fc-10b4-46f9-9e2a-6ec1f1d22c84_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e712f5e30e79253903c2d3fe47b07670da70ce9ca64573add712e3f18045bbf4 +size 47311347 diff --git a/NeurIPS/2025/4DGT_ Learning a 4D Gaussian Transformer Using Real-World Monocular Videos/full.md b/NeurIPS/2025/4DGT_ Learning a 4D Gaussian Transformer Using Real-World Monocular Videos/full.md new file mode 100644 index 0000000000000000000000000000000000000000..ddb3186e4ebd02429221099d23e3559eb67857ae --- /dev/null +++ b/NeurIPS/2025/4DGT_ Learning a 4D Gaussian Transformer Using Real-World Monocular Videos/full.md @@ -0,0 +1,778 @@ +# 4DGT: Learning a 4D Gaussian Transformer Using Real-World Monocular Videos + +Zhen Xu $^{1,2,*}$ Zhengqin Li $^{1}$ Zhao Dong $^{1}$ Xiaowei Zhou $^{2}$ Richard Newcombe $^{1}$ Zhaoyang Lv $^{1}$ + +$^{1}$ Reality Labs Research, Meta $^{2}$ Zhejiang University + +Project page: https://4dgt.github.io + +Mono Video + +Supervision + +Network Input + +![](images/40b7d89b8d140a1a559c25a770a930d396e1db6ecd5787cb70508e2d130aea23.jpg) + +![](images/3a08143d2c64fff3ee6428e35ac085b17120cc825b11ff73c0d956faacaff4d7.jpg) + +![](images/595232cff464db6d4349fe90a015caf442595b3ebae87bacb4e913b86712aa36.jpg) +Figure 1: We propose a scalable 4D dynamic reconstruction model trained only on real-world monocular RGB videos. The feed-forward 4DGS (section 3.1) representation enables us to render the geometry and appearance of the dynamic scene from novel views in real-time. Even without explicit supervision, the model can learn to distinguish dynamic contents from the background and produce realistic optical flows. The figure shows an enlarged set of Gaussians for the purpose of visualization. The embedded rendered videos only play in Adobe Reader or KDE Okular. + +![](images/909755fcb4db206de3bdaac46bfd2241411f17e50b7e4d4a2d041a0488ec2491.jpg) + +# Abstract + +We propose 4DGT, a 4D Gaussian-based Transformer model for dynamic scene reconstruction, trained entirely on real-world monocular posed videos. Using 4D Gaussian as an inductive bias, 4DGT unifies static and dynamic components, enabling the modeling of complex, time-varying environments with varying object lifespans. We proposed a novel density control strategy in training, which enables our 4DGT to handle longer space-time input and remain efficient rendering at runtime. Our model processes 64 consecutive posed frames in a rolling-window fashion, predicting consistent 4D Gaussians in the scene. Unlike optimization-based methods, 4DGT performs purely feed-forward inference, reducing reconstruction time from hours to seconds and scaling effectively to long video sequences. Trained only on large-scale monocular posed video datasets, 4DGT can outperform prior Gaussian-based networks significantly in real-world videos and achieve on-par accuracy with optimization-based methods on cross-domain videos. + +# 1 Introduction + +Humans record videos to digitize interactions with their surroundings. The ability to recover persistent geometry and 4D motion from videos has a profound impact on AR/VR, robotics, and content creation. Modeling 4D dynamic interactions from general-purpose videos remains a long-standing challenge. Prior work relying on multi-view synchronized capture or depth sensing is constrained to specific application domains. Recent progress in monocular dynamic video reconstruction via per-video optimization shows promise, but lacks scalability due to its time-consuming inference. + +In this paper, we propose 4D Gaussian Transformer (4DGT), a novel transformer-based model that reconstructs dynamic scenes from posed monocular videos in a feedforward manner. We assume camera calibration and 6-degree-of-freedom (6DoF) poses are available from on-device SLAM [9] or offline pipelines [23, 31]. Inspired by recent feedforward reconstruction methods for static 3D scenes [69, 72], 4DGT learns reconstruction from data and adopts 4D Gaussian Splatting (4DGS) [55] as a unified representation for both static and dynamic content, differing only in lifespan. This design enables fast 4D reconstruction from short videos in seconds. For longer videos with global pose consistency, 4DGT predicts consistent world-aligned 4DGS using 64-frame rolling windows. + +Training a 4D representation is challenging in defining appropriate supervision. While dynamic monocular videos are abundant, they lack space-time constraints. Multi-view video datasets [4, 22] are limited in both quantity and diversity, making them insufficient for training models that generalize in the wild. Prior methods [43] trained on synthetic object-level data suffer from a generalization gap when applied to complex real-world dynamics. + +To address this, we train 4DGT exclusively on posed monocular videos from public datasets. We use two key strategies to mitigate space-time ambiguity. First, we leverage depth and normal predictions from expert models [40, 64, 66] as auxiliary supervision for guiding geometry learning. Second, we regularize predicted Gaussian properties to favor longer lifespans and reduce overfitting to specific views. These enable 4DGT to effectively disentangle space-time structure, yielding high-quality geometry, novel view synthesis, and emergent motion properties such as segmentation and flow. Our reconstructions also show better metric consistency than the expert models used for supervision. + +Scaling a transformer to predict dense, pixel-aligned 4DGS presents two main challenges. First, dense pixel-aligned 4DGS predictions are computationally expensive for training and rendering. Inspired by density control in 3DGS [19], we introduce a pruning strategy that removes the redundant pixel-aligned 4DGS and further increases tokens in a second training with denser space-time samples. This effectively reduces $80\%$ of Gaussians and enables a $16\times$ higher sampling rate with the same compute. Second, as space-time samples increase, the number of tokens grows, and vanilla self-attention scales quadratically. To address this, we propose a level-of-detail structure via multi-level spatiotemporal attention, achieving an additional $2\times$ reduction in computational cost. + +4DGT is the first transformer-based method for predicting 4DGS in a feedforward manner using only real-world posed monocular videos in training. Extensive evaluations across datasets and domains show that 4DGT achieves comparable reconstruction quality to optimization-based methods while being three orders of magnitude faster, making it practical for long video reconstruction. Compared to prior methods that only train on synthetic object-level data [43], 4DGT generalizes better to complex real-world dynamics. Compared to the per-frame prediction pipeline [25], it also exhibits emergent motion properties. + +In summary, we make the following technical contributions: + +- We introduce 4DGT, a novel 4DGS transformer trained on posed monocular videos at scale, which produces consistent 4D video reconstructions in seconds at inference. +- We propose a training strategy to densify and prune space-time pixel-aligned Gaussians, reducing $80\%$ of predictions, achieving $16\times$ higher sampling rate during training and a $5\times$ speed-up in rendering. +- We design a multi-level attention module to efficiently fuse space-time tokens, further reducing training time by half. +- Our experiments demonstrate strong scalability of 4DGT across real-world domains using mixed training datasets and can outperform the previous Gaussian network significantly. The performance of 4DGT is on par with optimization-based methods in accuracy in cross-domain videos recorded by similar devices used in training, while being 3 orders of magnitude faster. + +# 2 Related Work + +Nonrigid reconstruction. Recovering dynamic content from video has long been a holy grail challenge in 3D vision. Early approaches demonstrated promising non-rigid shape reconstruction from RGB-D videos [32, 47, 3], but relied heavily on depth input and struggled with complex dynamic scenes. Since the seminal work on Neural Radiance Fields (NeRF) [30], several methods have extended NeRF to 4D using multi-view videos [22, 10] or posed monocular videos [36, 37]. However, 4D NeRFs are slow to train and render, limiting their scalability for complex dynamic scenes. Recently, generative priors have shown promise in aiding 4D reconstruction [58], offering strong regularization for shape and motion across space and time. Still, optimizing 4D representations remains time-consuming, and reconstruction quality depends heavily on the generalizability of priors across different scene domains. + +Dynamic Gaussian representations. Since the introduction of 3D Gaussian Splitting [19], several methods have extended it to 4DGS variants [65, 62, 55, 8], showing promising dynamic scene reconstruction from multi-view videos with faster training and real-time rendering support. However, optimizing dynamic Gaussians for monocular videos remains challenging. Recently, a few works have shown that complex 4D scenes composed of moving Gaussians can be recovered by leveraging depth, segmentation, and tracking priors from 2D expert models [57, 53, 21]. Despite strong performance, these methods involve complex processing, including manual annotation on dynamic regions. It further requires lengthy optimization, limiting its scalability in practical applications. + +Large reconstruction models. Transformer-based 3D large reconstruction models (LRMs) have shown strong potential for learning high-quality 3D reconstruction from data, at the object level [14, 16, 35, 60, 24] and static scenes [59, 69, 72, 61]. LRMs can generate reconstructions in seconds from a few input views, achieving quality comparable to optimization-based neural methods. However, training LRMs requires large-scale multi-view supervision of the same instance, which is readily available in synthetic datasets or static scene captures, but remains scarce for real-world videos. + +Some recent efforts explored training transformers to predict time-dependent 3D-GS [43, 25, 63, 41] using animated synthetic data [43], self-curated real-world internet videos [25] and street-level data [63], making them the closest related works in motivation. In contrast to these methods, which predict time-dependent 3D-GS representations, our 4DGT offers a holistic 4D scene representation that captures geometry better and enables motion understanding capabilities lacking in prior approaches. [63] requires multi-camera input and only focuses on street-level scenes. Compared to training on synthetic data in [43], our real-world training approach generalizes better to real-world scenes. Compared to B-Timer [25], which adopts a per-frame Gaussian prediction pipeline, our method produces explicit dynamic Gaussians thus can model explicit motion, showing emergent capabilities like motion segmentation. Compared to Pred. 3D Repr. [41], which adopts a tri-plane based implicit representation, our Gaussian model enables fast rendering after the feed-forward reconstruction. + +# 3 Method + +Given a posed monocular video, 4DGT uses a transformer to predict a set of 4DGS, which can be rendered in real time. We first describe the architecture and dynamic scene representation in section 3.1. To enable efficient training and rendering at scale, we introduce a pixel density control strategy and a level-of-detail structure based on spatial attention. Both techniques improve space-time sampling rates under fixed compute budgets. In section 3.3, we detail our training process and regularization strategies designed to resolve space-time ambiguities in monocular videos. + +# 3.1 Feed-Forward Dynamic Gaussian Prediction + +Input encoding. Given a posed monocular video, we extract a set of image frames $\mathbf{I}_i$ with camera calibration $\mathbf{P}_i$ and timestamp $\mathbf{T}_i$ , denoted as $\{\mathbf{I}_i \in R^{H \times W \times 3}, \mathbf{P}_i \in R^{H \times W \times 6}, \mathbf{T}_i \in R^{H \times W \times 1} | i = 1 \dots N\}$ , where $\mathbf{P}_i$ represents the Plücker coordinates [18] and $N$ is the total number of frames. We convert them into patches. For frame $i$ , the patches are denoted as $\{\mathbf{I}_{i,j} \in R^{p \times p \times 3} | j = 1 \dots HW / p^2\}$ , $\{\mathbf{T}_{i,j}\}$ and $\{\mathbf{P}_{i,j}\}$ , where $p$ is the patch size. + +Feature fusion. We use the pretrained DINOv2 image encoder [33] to extract high-level $C$ -dimensional features $\mathbf{F}_{i,j} \in R^C$ . These are concatenated with the temporal and spatial encoding $\mathbf{T}_{i,j}$ + +![](images/f5fda35d75df3545d22066bdb2ebffa055e415e6fb15d6aea52e457dea47c0ef.jpg) +Figure 2: An overview of our method in training and rendering. 4DGT takes a series of monocular frames with poses as input. During training, we subsample the temporal frames at different granularity and use all images in training. We first train 4DGT to predict pixel-aligned Gaussians at coarse resolution in stage one. In stage two training, we pruned a majority of non-activated Gaussians according to the histograms of per-patch activation channels, and densify the Gaussian prediction by increasing the input token samples in both space and time. At inference time, we run the 4DGT network trained after stage two. It can support dense video frames input at high resolution. + +and $\mathbf{P}_{i,j}$ as well as the input RGB image $\mathbf{I}_{i,j}$ to form the fused transformer input: + +$$ +\left\{\mathbf {X} _ {i, j} \right\} = \mathcal {F} \left(\left\{\mathbf {I} _ {i, j} \oplus \mathbf {T} _ {i, j} \oplus \mathbf {P} _ {i, j} \oplus \mathbf {F} _ {i, j} \mid i = 1 \dots N, j = 1 \dots H W / p ^ {2} \right\}\right), \tag {1} +$$ + +where $\oplus$ denotes the concatenation operation and $\mathcal{F}$ denotes the all-to-all self-attention transformer module. In contrast to ViT, input used static LRMs in that it uses only Plücker rays [69] or DINO feature [14], our transformer takes timestamp-aware Plücker rays with DINO feature together as input, which we found to be beneficial to provide the best prediction in view synthesis as well as geometry prediction. + +Dynamic Gaussians. We use a variant of 4DGS [65, 55] to unify the various components in dynamic scene predictions. To better represent geometry, we adopt the 2DGS [15] defined by the center $\mathbf{x} \in R^3$ , scale $\mathbf{s} \in R^2$ , opacity $\mathbf{o} \in R^1$ and orientation $\mathbf{q} \in R^4$ (quaternion) of the Gaussians. Compared to 3DGS[19] used in previous work [65, 55], 2DGS yields better geometry predictions. To represent motion, we use four temporal attributes, namely the temporal center $\mathbf{c} \in R^1$ , life-span $\mathbf{l} \in R^1$ , velocity $\mathbf{v} \in R^3$ , and angular velocity $\boldsymbol{\omega} \in R^3$ (as angle-axis) for each Gaussian. Given a specific timestamp $t_s$ for rendering a particular dynamic Gaussian point $\mathbf{g} = \{\mathbf{x}, \mathbf{s}, \mathbf{q}, \mathbf{o}, \mathbf{c}, \mathbf{l}, \mathbf{v}, \boldsymbol{\omega}\}$ , we first calculate the offset to the opacity, location and orientation of the Gaussian point from the temporal attributes [65]. Specifically, the life-span $\mathbf{l}$ is used to influence the Gaussian opacity $\mathbf{o}$ over time: + +$$ +\boldsymbol {\sigma} = \sqrt {- \frac {1}{2} \cdot \frac {\left(\mathbf {1} / 2\right) ^ {2}}{\log \left(o _ {t h}\right)}}, \quad \mathbf {o} _ {t _ {s}} = \mathbf {o} \cdot e ^ {- \frac {1}{2} \cdot \frac {\left(t _ {s} - \mathbf {c}\right) ^ {2}}{\sigma^ {2}}}, \tag {2} +$$ + +where $o_{th}$ is the opacity multiplier at the life-span boundary and $\sigma$ is the standard deviation of the Gaussian distribution in the temporal domain. Intuitively, the Gaussian retains its full opacity at its temporal center and fades in a Gaussian distribution along the temporal axis. At 1/2 time relative to the temporal center $\mathbf{c}$ , the opacity of the point is reduced by multiplying a small factor $o_{th}$ , which is set to 0.05 for all experiments. The location and orientation of each Gaussian is adjusted by the velocity $\mathbf{v}$ and angular velocity $\omega$ to account for the motion: + +$$ +\mathbf {x} _ {t _ {s}} = \mathbf {x} + \mathbf {v} \cdot (t _ {s} - \mathbf {c}), \quad \mathbf {q} _ {t _ {s}} = \mathbf {q} \cdot \phi (\boldsymbol {\omega} \cdot (t _ {s} - \mathbf {c})), \tag {3} +$$ + +where $\phi$ denotes converting the angle-axis representation to a quaternion. For each 2DGS with $\mathbf{x}_{t_s},\mathbf{s}_{t_s},\mathbf{q}_{t_s},\mathbf{o}_{t_s}$ at timestamp $t_s$ , we apply 2DGS rasterizer implemented in [67] to render image. + +Dynamic Gaussian decoding. Given the pixel-aligned features $\{\mathbf{X}_{i,j}\}$ , we use a transformer to decode the pixel-aligned 4DGS for each frame as + +$$ +\left\{\mathbf {G} _ {i, j} \right\} = \mathcal {D} _ {\mathbf {x}, \mathbf {s}, \mathbf {q}, \mathbf {o}, \mathbf {c}, \mathbf {l}, \mathbf {v}, \omega} \left(\left\{\mathbf {X} _ {i, j} \mid i = 1 \dots N, j = 1 \dots H W / p ^ {2} \right\}\right), \tag {4} +$$ + +where $\mathcal{D}_{\mathbf{x},\mathbf{s},\mathbf{q},\mathbf{o},\mathbf{c},\mathbf{l},\mathbf{v},\omega}$ denotes the MLP-based decoder head for producing the full suite of dynamic Gaussian parameters $\{\mathbf{G}_{i,j}\}$ . + +Our proposed 4DGS can unify the prediction of appearance and geometry properties of both static and dynamic elements. For static scenes, the network can learn to predict Gaussians with a long-living lifespan $1 \to \infty$ , $\mathbf{v} \to 0$ , $\omega \to 0$ . For complex dynamic motions with occlusions, it predicts transient dynamic objects with short-living Gaussians $1 \to 0$ . + +# 3.2 Multi-level Pixel & Token Density Control + +While pixel-aligned Gaussian has been a standard choice in prior work [69, 72], it has a severe limitation in representing video frames that require dense, long-term sampling to capture motion effectively. Naively sampling frames spatial-temporally would result in two key issues. First, the increasing number of aligned Gaussians degrades optimization and rendering performance, leading to suboptimal training and blurry details in dynamic regions. Second, the growing number of input tokens significantly increases computational cost, resulting in under-trained models. + +Two-stage training. We introduce a two-stage approach to address these challenges. First, we train on coarsely sampled low-resolution images from scratch until convergence. In the second stage, inspired by 3DGS [19], we propose to filter pixel-aligned Gaussians by pruning low-opacity predictions per patch based on the histograms and increasing the token count to predict more Gaussians across space-time. Additionally, we introduce a multi-level spatiotemporal attention mechanism to further reduce the computational cost of self-attention layers. + +Pruning. After the initial training stage at coarse resolution, we compute a histogram of activated Gaussians per patch and observe that only a few channels are activated. A similar pattern emerges in other pixel-aligned Gaussian methods [69], motivating us to decode only a small set of Gaussians using the activated channels. Formally, for each patch of Gaussian parameters $\mathbf{G}_{i,j}$ , we consider the standard deviation of their opacity values $\mathbf{o}_{i,j} = \{\mathbf{o}_{i,j,k}|k = 1\dots p^2\}$ : + +$$ +\mu \left(\mathbf {o} _ {i, j}\right) = \frac {1}{p ^ {2}} \sum_ {k = 1} ^ {p ^ {2}} \mathbf {o} _ {i, j, k}, \quad \sigma \left(\mathbf {o} _ {i, j}\right) = \sqrt {\mu \left(\mathbf {o} _ {i , j} ^ {2}\right) - \mu \left(\mathbf {o} _ {i , j}\right) ^ {2}}, \tag {5} +$$ + +where $\mu (\mathbf{o}_{i,j})$ is the mean of the opacity values and $\sigma (\mathbf{o}_{i,j})$ is the standard deviation. A particular pixel $k$ is considered activated if it has a value larger than 1 unit of the standard deviation: + +$$ +\mathbf {m} _ {i, j, k} = \left\{ \begin{array}{l l} 1, & \mathbf {o} _ {i, j, k} > \mu \left(\mathbf {o} _ {i, j}\right) + \sigma \left(\mathbf {o} _ {i, j}\right), \\ 0, & \text {o t h e r w i s e .} \end{array} \right. \tag {6} +$$ + +where $k$ is the index of the pixel in the patch and $\mathbf{m}_{i,j,k}$ indicates whether the pixel is activated. A histogram $\mathbf{h}_{i,j}$ of all activation mask $\mathbf{m}_{i,j} = \{\mathbf{m}_{i,j,k}|k = 1\dots p^2\}$ for the patch output is: + +$$ +\mathbf {H} = \sum_ {i = 1, j = 1} ^ {N, H W / p ^ {2}} \mathbf {M} _ {i, j}, \tag {7} +$$ + +where $\mathbf{M}_{i,j} \in \mathbb{N}^{p^2}$ is the activation mask for patch $(i,j)$ , $\mathbf{H} \in \mathbb{N}^{p^2}$ is the aggregated histogram of all activation masks, and $p$ is the patch size. We select $S$ channel from the histogram $\mathbf{h}_{i,j}$ for the patch output for all patches onward in training. This effectively implements an $S / p^2$ times reduction in the number of Gaussians for each patch, mimicking the pruning strategy of 3DGS [19]. We provide more in-depth analysis for this histogram-based pruning strategy compared to alternatives with visualizations in the appendix. + +Densification. The predicted Gaussian number can naturally increase with more space-time token inputs, either in resolution per frame or temporal frame numbers. The initially trained model provides a good scaffolding for pixel-aligned Gaussians when we increase the input token number in space and time. In the second stage of training, we increase the spatial and temporal resolution by a factor + +of $R_{s}$ and $R_{t}$ , respectively. Combining the densification process and pruning strategy, this would result in $R_{s}^{2} \cdot R_{t} \cdot S / p^{2}$ times of the Gaussians compared to the first stage. We select $R_{s} = 2$ , $R_{t} = 4$ , $S = 10$ , and $p = 14$ for all experiments, leading to only $80\%$ of the original number of Gaussians while greatly increasing the sampling rate of space-time by 16 times. + +Multi-level spatial-temporal attention. The number of patches participating in the self-attention module $\mathcal{F}$ increases by a factor of $R_s^2 \cdot R_t$ , which will slow down optimization and inference significantly. To mitigate this, we propose a temporal level-of-detail attention mechanism to reduce the computational cost. We propose to divide the $N$ input frames into $M$ equal trunks in the highest level. This division limits the attention mechanism in the temporal dimension, but reduces the computation of calculating $n$ total tokens to $O\left(\frac{n^2}{M}\right)$ . To balance spatial-temporal samples, we construct a temporal level-of-detail structure by alternating the temporal range and spatial resolution, achieving a much smaller overhead while maintaining the ability to handle long temporal windows. For each level $l$ , we reduce the spatial resolution by a factor of $2^l$ and increase temporal samples by 2. Empirically, we use level $L = 3$ and $M = 4$ , which leads to an approximately 2 times reduction in the computational cost. + +# 3.3 Training + +Loss and regularization. We train 4DGT using segments of $W = 128$ consecutive frames from the monocular video and subsample every 8 frames as input, resulting in $N = 16$ input frames. Notably, for the second stage training where we apply techniques mentioned in section 3.2 and section 3.2, we increase the number of input frames to $N = 64$ . After obtaining all Gaussian parameters $\{\mathbf{G}_{i,j}\}$ from each of the $N$ input frames, we render them to all $W = 128$ images for self-supervision and compute the MSE loss. Additionally, we add the perceptual LPIPS loss [17] $\mathcal{L}_{lpips}$ and SSIM loss [56] $\mathcal{L}_{ssim}$ for better perceptual quality. + +$$ +\mathcal {L} _ {\mathrm {m s e}} = \sum_ {i = 1} ^ {W} \frac {\left\| \mathbf {I} _ {i} - \mathbf {I} _ {i} ^ {\prime} \right\| _ {2}}{W}, \quad \mathcal {L} _ {\mathrm {l p i p s}} = \sum_ {i = 1} ^ {W} \frac {\left\| \psi (\mathbf {I} _ {i}) - \psi \left(\mathbf {I} _ {i} ^ {\prime}\right) \right\| _ {1}}{W}, \quad \mathcal {L} _ {\mathrm {s s i m}} = \sum_ {i = 1} ^ {W} \frac {\operatorname {S S I M} \left(\mathbf {I} _ {i} , \mathbf {I} _ {i} ^ {\prime}\right)}{W}, \tag {8} +$$ + +where $\mathbf{I}_i$ denotes the input image and $\mathbf{I}_i^{\prime}$ is the rendered image, $\psi$ is the pre-trained layers AlexNet [20] and SSIM is the SSIM function [56]. To better regularize the training, we encourage the points to be static and have a long lifespan using: + +$$ +\mathcal {L} _ {\mathbf {v}} = \sum_ {i = 1, j = 1, k = 1} ^ {N, H W / p ^ {2}, p ^ {2}} \frac {\left\| \mathbf {v} _ {i , j , k} \right\| _ {1}}{N H W}, \quad \mathcal {L} _ {\boldsymbol {\omega}} = \sum_ {i = 1, j = 1, k = 1} ^ {N, H W / p ^ {2}, p ^ {2}} \frac {\left\| \boldsymbol {\omega} _ {i , j , k} \right\| _ {1}}{N H W}, \quad \mathcal {L} _ {1} = \sum_ {i = 1, j = 1, k = 1} ^ {N, H W / p ^ {2}, p ^ {2}} \frac {\left\| \frac {1}{I _ {i , j , k}} \right\| _ {1}}{N H W}, \tag {9} +$$ + +where $\mathbf{v}_{i,j,k}$ is the velocity of the Gaussian point, $\omega_{i,j,k}$ is the angular velocity of the Gaussian point and $\mathbf{l}_{i,j,k}$ is the life-span of the Gaussian point. + +Expert guidance. We observe that training can benefit from leveraging monocular export models in geometry prediction. We extract the depth map $\mathbf{D}_i$ and normal map $\mathbf{N}_i$ from all $W$ frames using DepthAnythingV2 [64] and StableNormal [66] and use them as a pseudo-supervision signal: + +$$ +\mathcal {L} _ {\mathbf {D}} = \sum_ {i = 1} ^ {W} \frac {\left\| \mathbf {D} _ {i} - \mathbf {D} _ {i} ^ {\prime} \right\| _ {2}}{W}, \quad \mathcal {L} _ {\mathbf {N}} = \sum_ {i = 1} ^ {W} \frac {\left\| \mathbf {N} _ {i} - \mathbf {N} _ {i} ^ {\prime} \right\| _ {2}}{W}, \tag {10} +$$ + +where $\mathbf{D}_i^{\prime}$ and $\mathbf{N}_i^{\prime}$ are the predicted depth and normal map rendered using the 2DGS rasterizer [15]. The final loss function for training the feed-forward prediction pipeline is: + +$$ +\mathcal {L} = \mathcal {L} _ {\mathrm {m s e}} + \lambda_ {\mathrm {l p i p s}} \mathcal {L} _ {\mathrm {l p i p s}} + \lambda_ {\mathrm {s s i m}} \mathcal {L} _ {\mathrm {s s i m}} + \lambda_ {\mathbf {v}} \mathcal {L} _ {\mathbf {v}} + \lambda_ {\boldsymbol {\omega}} \mathcal {L} _ {\boldsymbol {\omega}} + \lambda_ {\mathrm {l}} \mathcal {L} _ {\mathrm {l}} + \lambda_ {\mathbf {D}} \mathcal {L} _ {\mathbf {D}} + \lambda_ {\mathbf {N}} \mathcal {L} _ {\mathbf {N}}, \tag {11} +$$ + +where $\lambda_{\mathrm{lpips}}$ , $\lambda_{\mathrm{ssim}}$ , $\lambda_{\mathbf{v}}$ , $\lambda_{\omega}$ , $\lambda_{\mathrm{l}}$ , $\lambda_{\mathbf{D}}$ and $\lambda_{\mathbf{N}}$ are the weights for the corresponding loss functions. We set $\lambda_{\mathrm{lpips}} = 2.0$ , $\lambda_{\mathrm{ssim}} = 0.2$ , $\lambda_{\mathbf{v}} = 1.0$ , $\lambda_{\omega} = 1.0$ , $\lambda_{\mathrm{l}} = 1.0$ , $\lambda_{\mathbf{D}} = 0.1$ and $\lambda_{\mathbf{N}} = 0.01$ for all experiments. All weights for the regularization losses are warmed up linearly from 0 to their final values during the first 2500 iterations of training. + +# 4 Implementation Detail + +Architecture. We use a modified ViT architecture [7] for our fusion network $\mathcal{F}$ . Specifically, we use 12 layers of all-to-all self-attention with 16 heads, each head having a hidden dimension of 96, + +and the fully connected layers have a $4\times$ wider hidden channel size. Since the Plücker coordinates [18] $\mathbf{P}$ and timestamps $\mathbf{T}$ already provide the 4D position of each pixel, we do not use additional embedding for the positional information. For the second stage training, where we enable the multilevel spatial-temporal attention module, we copy the weights of the first-stage transformer $L$ times and train them independently. The $l$ -th transformer is responsible for the $l$ -th level of spatial-temporal attention, with 1 classification token for passing information between different levels. For the MLP decoders $\mathcal{D}$ , we use 2 fully connected layers with a hidden dimension of 256 for each channel. Both the transformer modules $\mathcal{F}$ and $\mathcal{G}$ use GELU [13] as the activation function and layer normalization [1] as the normalization function. We also disable the bias parameters for all the layers. + +Training & Inference. We implement 4DGT in PyTorch framework [38]. We employ FlashAttentionV3 [45] and the GSplat Rasterizer [67] for efficient attention and Gaussian optimization respectively. For optimization, we use the AdamW optimizer [27] with a learning rate of $5e^{-4}$ and a weight decay of 0.05. For the second stage training, the learning rate is set to $1e^{-5}$ . Additionally, we linearly warm-up the learning rate of each stage in the first 2500 steps and then apply the cosine decaying schedule [26] for the remaining steps. During the second strange training, we additionally augment the input and output to the network by varying the aspect ratio and field of view of the images. Specifically, we randomly sample an aspect ratio from the uniform distribution on $\left[\frac{1}{3},\frac{3}{1}\right]$ and a field of view ratio on the original image on $[30\%, 100\%]$ . We train our reconstruction model $100k$ iterations for the first stage and $30k$ iterations for the second stage, using a total batch size of 64. With 64 Nvidia H100 GPUs, the first stage training takes roughly 9 days and the second stage training takes roughly 6 days. For all other experiments on inference speed, we use a single 80 GB A100 GPU. + +# 5 Experiments + +Training Datasets. We use the following real-world monocular videos with high-quality calibrations: + +- Project Aria datasets with closed-loop trajectories: the EgoExo4D [12], Nymeria [29], Hot3D [2] and Aria Everyday Activities (AEA) [28]. +- Video data with COLMAP [44] camera parameters: Epic-Fields [50, 5] and Cop3D [46]. +- Phone videos with ARKit camera poses: ARKitTrack [71]. + +Evaluation datasets. We use the synthetic rendering provided in ADT [34] datasets, which provides metric ground truth depth. To evaluate cross-domain generalization, we use DyCheck [11] (DyC) datasets and the dynamic scene in TUM-SLAM [48] (TUM) to evaluate novel view synthesis. We further hold out a test split from EgoExo4D, AEA, and Hot3D, which we refer to as the Aria test set. + +Metrics. For appearance evaluation, we compare the PSNR and LPIPS [70] metrics on novel view and time rendering results. For geometry evaluation, we compare against the depth RMSE [64] and normal angle error [66]. We additionally provide qualitative comparisons of motion rendered in 2D as optical flow and motion segmentation. All comparison experiments are conducted on 128-frame subsequences of the monocular videos, with 64 frames used as input and the remaining 64 frames used for testing, with images resized to $504 \times 504$ resolution or a similar pixel number for controlled comparison unless specified otherwise. For the DyCheck [11] dataset, we additionally compare the rendering results on the provided test view cameras, which show signals on extreme view synthesis. We provide more details about evaluation implementations in the appendix. + +Baselines. We consider the following baselines as the most relevant work for evaluation. + +1. L4GM [43]: It is the closest prior 4D Gaussian model that generalizes to real-world videos. Different from ours trained using real-world data only, they trained on a synthetic dataset and leveraged additional multi-view diffusion priors from ImageDream [52]. +2. Static-LRM: We trained a static scene LRM following [69] on the same real world data as our 4DGT. We use 2DGS instead of 3DGS as the representation that shows more similarity to our approach, except that we further model the dynamic content. +3. Expert monocular models: We compared each individual expert model we used during training in the same setting, including DepthAnythingV2 [64] aligned with the metric scale of UniDepth [39] and normals provided by StableNormal [66]. For novel view evaluation, we unproject the image using the nearest depth and normal frame. + +![](images/7bf68a929423d1ebb09fc240638a5ac582f9aca4e8eed026cbe41cb86a114a4a.jpg) +Figure 3: From left-to-right, we show the novel space-time view comparisons on ADT [34], EgoExo4D [12], DyCheck [11] and the DyCheck test-view (rightmost). We render the depth (upper right) and normal (below right) next to each synthesized novel view. For ground truth depth and normal on EgoExo4D and DyCheck, we use predictions from the expert models from the ground truth image for reference. Please refer to the appendix for more visual comparisons. + +4. MonST3R [68]: We compare to the dynamic point based representation [68] which highlights the representation difference in using 4DGS. We use ground truth camera poses as input to their model using the official implementation and using PyTorch3D [42] for normal estimation. +5. Shape of Motion (SoM) [53]: We use SoM to represent the top-tier per-scene optimization method as a reference for best dynamic reconstruction quality. We follow SoM's instructions to manually segment the dynamic part. It requires running expert models as input, including mask, depth, and tracking, which we do not use. We include the preprocessing time for time comparisons. + +We do not make comparisons with CAT4D [58], BulletTimer [25] and Pred. 3D Repr. [41] since they provide neither the source code nor the pre-trained models. + +Comparisons to baselines. Table 1 and Figure 3 present our comparisons to baselines. Compared to L4GM, which also predicts dynamic Gaussians from a trained transformer, 4DGT shows much better generalization across real scenes. We found that a static-LRM can provide a strong baseline for static scenes but will fail when dynamic motion is present, while 4DGT can do well in both. It is further validated in Table 2c on dynamic regions. The geometry predicted from 4DGT is more consistent with the world coordinate compared to expert models when evaluated in metric scale. Compared to the optimization-based method SoM, 4DGT can offer on-par quality in view synthesis as well + +Table 1: Comparisons to baselines. We mark SoM in grey as a reference for optimization-based methods, and rank the other baselines with ours as comparisons in learning based approaches. + +
MethodPSNR (Render) ↑LPIPS (Render) ↓RMSE (Depth) ↓Deg. ↓Recon. Time ↓
ADTTUMDyCAriaAvgADTTUMDyCAriaAvgADTTUMAvgADTTUM
SoM [53]±1.640±0.266±2.597±3.701±0.024±0.069±0.060±0.067±3.000±1.354±1.633±3.607
30.3021.0316.4926.6923.630.2420.3370.3920.3070.3204.1582.7563.43435.0560000 ms / f
L4GM [43]±2.713±0.112±0.629±1.666±0.051±0.013±0.062±0.052±0.660±0.791±0.386±4.181200 ms / f
7.3489.2268.7708.6178.4900.6880.6700.5870.6980.6612.6061.6982.09463.51
±1.651±1.673±1.673±3.355±0.011±0.023±0.045±0.135±0.842±0.243±0.543±2.699
MonST3R [68, 66]25.1320.6111.3219.9019.240.2460.2730.4290.3230.3182.1110.6531.38225.004500 ms / f
±3.195±3.117±1.258±2.419±0.082±0.073±0.058±0.046±0.621±0.074±0.348±0.748
Experts [40, 66]23.3218.6411.5322.3218.960.2990.3180.4230.2360.3192.9310.9191.92526.26350 ms / f
±1.508±0.048±2.034±1.591±0.021±0.009±0.067±0.019±0.463±0.048±0.255±1.831
Ours28.3121.0216.1227.3623.200.2430.3490.4080.2300.3080.9340.3940.66425.9225 ms / f
+ +Table 2: Ablation study on our method components using ADT and DyCheck (DyC). + +(a) Ablation on dynamic Gaussian in stage one training. + +
MethodPSNR↑LPIPS↓RMSEDeg.
ADTDyCAvgADTDyCAvgADT↓ADT↓
Naive15.4913.9514.720.6120.5170.5641.15642.25
+ EgoExo4D [12]22.7215.2719.000.2290.3850.3071.27841.11
Static LRM [69]19.2914.2116.750.3990.4630.4310.83031.96
Per-frame [43]10.0712.1011.080.7480.7140.7313.11761.77
+ LN,D,l,ω,v26.4515.8621.150.1700.3990.2840.77321.59
+ +(c) Evaluation on the dynamic foreground. + +
MethodPSNR†LPIPS↓RMSEDeg.
ADTDyCAvgADTDyCAvgADT↓ADT↓
Static LRM [69] (masked)17.3013.5616.760.0590.2200.1020.61349.35
Ours (masked)27.2914.9322.860.0300.1950.0750.38833.32
+ +(b) Ablation on stage two training. + +
MethodPSNR↑LPIPS↓RMSEDeg.
ADTDyCAvgADTDyCAvgADT↓ADT↓
NaiveOut of Memory
Random.25.7914.9720.380.3330.4800.4060.75025.84
+ D&P28.7915.5222.160.2420.4340.3380.72225.95
+ Multi-level28.7115.3422.030.2420.4390.3410.78327.30
+ Mix. (Ours)28.3116.1222.220.2430.4080.3260.93425.92
+ +(d) Motion segmentation results. + +
Methodw/o Lv,ω,1MegaSaM [23]Ours
mIoU × 100↑9.4±4.177.4±4.081.2±1.8
+ +as geometry prediction while being 3 orders of magnitude faster in runtime, which makes it more favorable to process long-time videos in practice. + +Ablation study in stage one training. In Table 2a for stage one training, we start from a Naive training baseline at coarse resolution using the image rendering losses in Eq. 8 trained only using the AEA dataset with only 7 hours of data. After further scaling to using the EgoExo4D dataset (300 hours), we find that increasing the scale of the dataset can significantly improve the performance. We also compare to a static LRM [69] and per-frame LRM [43] counterpart in the same setting. We can already see the benefits over the baselines at this stage by a large margin. We further include the full training loss in Eq. 9 and Eq. 11, and we can see significant improvements in the quality across all metrics. The regularization terms $\mathcal{L}_{\mathrm{N,D,l},\omega,\mathrm{v}}$ can be categorized into two groups: (1) $\mathcal{L}_{\mathrm{v},\omega,1}$ regularizes the motion of the dynamic Gaussian predictions. Without this term, although the quantitative results are not greatly affected, the model would fall into the trivial local minima of making every Gaussian transient and dynamic, not correctly modeling the scene's static or slow-moving parts. This would result in a purely-white motion mask. In table 2d, we evaluate the quality of the motion mask on the ADT dataset [34] of our method without this term. Thanks to the explicit modeling of the motion parameters in our representation, our model can produce comparable motion segmentation against MegaSaM [23], which has explicit flow supervision, while being $200\times$ faster (1.5s v.s. 300s). (2) $\mathcal{L}_{\mathrm{N,D}}$ provides expert guidance for the geometry of the dynamic Gaussian prediction. These terms can greatly improve the quality of the reconstructed geometry. Please refer to the appendix for visual comparisons. + +Ablation study in stage two training. Table. 2b shows the ablation study of key design in stage two training. Starting from a Naive model without density control to prune and densify $(D\& P)$ Gaussians and the multi-level attention proposed in 3.2, naively scaling up resolution and temporal samples will run out of memory in training. Compared with variants using a Random sampled Gaussian from predicted patches and further densifying, using the proposed D&P strategy to decode sparse activated Gaussian will lead to better results while being efficient in training. Adding the proposed multi-level attention can further speed up training with only a minor sacrifice in quality, but can speed up training two times faster. Finally, we mixed all the proposed datasets in training at stage two. Compared to model training only using EgoExo4D, mixing datasets improves generalization across domains. + +Table 3: Comparison with a more comprehensive upper-bound on the ADT dataset [34]. + +
MethodPSNR ↑LPIPS ↓RMSE ↓Degree ↓Recon. Time ↓
SoM [53]30.30±1.640.242±0.0244.16±3.0034.07±3.6160,000 ms/f
SoM* [53, 19]28.40±2.600.281±0.0274.18±2.9818.46±1.7660,000 ms/f
Ours28.31±1.510.243±0.0210.93±0.4625.94±1.8425 ms/f
Ours_tune10s31.98±1.010.220±0.0120.85±0.4519.25±2.6525 + 150 ms/f
+ +Initializing optimization-based methods with 4DGT. Table 3 presents a quantitative comparison on the ADT dataset [34] with stronger baselines and 4DGT-initialized optimization-based method. SoM-2DGS-Geometry augments SoM [53] with 2DGS [19] and employs the same normal regularization as ours to provide a stronger geometry upper-bound. This results in improved normal quality, but our feed-forward 4DGT prediction method still achieves comparable—or superior—results while being orders of magnitude faster. After finetuning the feed-forward prediction for only 10 seconds (100 iterations, $150\mathrm{ms}$ per frame, denoted as $Ours_{tune10s}$ ), performance further improves beyond all optimization-based baselines. While SoM and its 2DGS-augmented variant require 30,000 optimization steps per scene (and a sophisticated tracking expert like TAPIR [6]), our finetuned feed-forward prediction achieves a $350\times$ speedup, and pure feed-forward performance ( $Ours$ ) is $2,400\times$ faster. More finetuning further improves reconstruction quality. + +# 6 Conclusion + +We introduced 4DGT, a novel dynamic scene reconstruction method that predicts 4DGS from an input posed video frame in a feed-forward manner. The representation power of 4DGT enables it to handle general dynamic scenes using 4DGS with varying lifespans, and support it to handle complex dynamics in long videos. Different from prior work that heavily depends on multi-view supervision from synthetic datasets, 4DGT is trained only using real-world monocular videos. We demonstrate that 4DGT can generalize well to videos recorded from similar devices, and the ability of generalization can improve when mixing datasets for training in scale. + +Limitations and future work. We do not claim 4DGT can generalize to all videos in the wild. We assume the availability of a reliable calibration to train 4DGT and deploy it for inference. For this requirement, the training datasets have been limited to data sources from a few egocentric devices and phone captures. We observe that the quality may degrade in videos recorded by an unseen type of device due to the inaccurate metric scale calibrations. We believe this can be significantly improved by further scaling up the method using more diverse datasets recorded by different devices with curated calibrations. Similar to most monocular reconstruction methods, we still observe significant artifacts when viewing Gaussians from extreme view angles, departing far from the input trajectory. Future directions can propose better representations to address it or learn to distill more priors from multi-view expert models, such as generative video models. + +# Acknowledgments and Disclosure of Funding + +We thank Aljaz Bozic for the insightful discussions. We thank the Project Aria team for their open-source and dataset contributions. This work was also partially supported by NSFC (No. U24B20154). + +# References + +[1] J. L. Ba, J. R. Kiros, and G. E. Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016. +[2] P. Banerjee, S. Shkodrani, P. Moulon, S. Hampali, S. Han, F. Zhang, L. Zhang, J. Fountain, E. Miller, S. Basol, et al. Hot3d: Hand and object tracking in 3d from egocentric multi-view videos. arXiv preprint arXiv:2411.19167, 2024. +[3] A. Bozic, M. Zollhofer, C. Theobalt, and M. Niessner. Deepdeform: Learning non-rigid rgb-d reconstruction with semi-supervised data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. +[4] M. Broxton, J. Flynn, R. Overbeck, D. Erickson, P. Hedman, M. Duvall, J. Dourgarian, J. Busch, M. Whalen, and P. Debevec. Immersive light field video with a layered mesh representation. ACM Transactions on Graphics (TOG), 39(4):86-1, 2020. +[5] D. Damen, H. Doughty, G. M. Farinella, S. Fidler, A. Furnari, E. Kazakos, D. Moltisanti, J. Munro, T. Perrett, W. Price, et al. Scaling egocentric vision: The epic-kitchens dataset. In Proceedings of the European conference on computer vision (ECCV), pages 720-736, 2018. +[6] C. Doersch, Y. Yang, M. Vecerik, D. Gokay, A. Gupta, Y. Aytar, J. Carreira, and A. Zisserman. Tapir: Tracking any point with per-frame initialization and temporal refinement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10061-10072, 2023. +[7] A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. +[8] Y. Duan, F. Wei, Q. Dai, Y. He, W. Chen, and B. Chen. 4d-rotor gaussian splatting: towards efficient novel view synthesis for dynamic scenes. In ACM SIGGRAPH 2024 Conference Papers, pages 1–11, 2024. +[9] J. Engel, K. Somasundaram, M. Goesele, A. Sun, A. Gamino, A. Turner, A. Talattof, A. Yuan, B. Souti, B. Meredith, et al. Project aria: A new tool for egocentric multi-modal ai research. arXiv preprint arXiv:2308.13561, 2023. +[10] S. Fridovich-Keil, G. Meanti, F. R. Warburg, B. Recht, and A. Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12479–12488, 2023. +[11] H. Gao, R. Li, S. Tulsiani, B. Russell, and A. Kanazawa. Monocular dynamic view synthesis: A reality check. Advances in Neural Information Processing Systems, 35:33768-33780, 2022. +[12] K. Grauman, A. Westbury, L. Torresani, K. Kitani, J. Malik, T. Afouras, K. Ashutosh, V. Baiyya, S. Bansal, B. Boote, et al. Ego-exo4d: Understanding skilled human activity from first-and third-person perspectives. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19383-19400, 2024. +[13] D. Hendrycks and K. Gimpel. Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415, 2016. +[14] Y. Hong, K. Zhang, J. Gu, S. Bi, Y. Zhou, D. Liu, F. Liu, K. Sunkavalli, T. Bui, and H. Tan. Lrm: Large reconstruction model for single image to 3d. arXiv preprint arXiv:2311.04400, 2023. +[15] B. Huang, Z. Yu, A. Chen, A. Geiger, and S. Gao. 2d gaussian splatting for geometrically accurate radiance fields. In ACM SIGGRAPH 2024 conference papers, pages 1-11, 2024. +[16] Y. Jiang, L. Zhang, J. Gao, W. Hu, and Y. Yao. Consistent4d: Consistent 360 $\{\backslash \mathrm{deg}\}$ dynamic object generation from monocular video. arXiv preprint arXiv:2311.02848, 2023. +[17] J. Johnson, A. Alahi, and L. Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In ECCV, 2016. +[18] P. Julius. On a new geometry of space. Julius Plückers Gesammelte wissenschaftliche Abhandlungen, op. cit.(24), 1:462-545, 1865. +[19] B. Kerbl, G. Kopanas, T. Leimkuhler, and G. Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (TOG), 42(4):1-14, 2023. +[20] A. Krizhevsky, I. Sutskever, and G. E. Hinton. Imagenet classification with deep convolutional neural networks. Advances in neural information processing systems, 25, 2012. +[21] J. Lei, Y. Weng, A. Harley, L. Guibas, and K. Daniilidis. MoSca: Dynamic gaussian fusion from casual videos via 4D motion scaffolds. arXiv preprint arXiv:2405.17421, 2024. + +[22] T. Li, M. Slavcheva, M. Zollhoefer, S. Green, C. Lassner, C. Kim, T. Schmidt, S. Lovegrove, M. Goesele, R. Newcombe, et al. Neural 3d video synthesis from multi-view video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5521-5531, 2022. +[23] Z. Li, R. Tucker, F. Cole, Q. Wang, L. Jin, V. Ye, A. Kanazawa, A. Holynski, and N. Snavely. Megasam: Accurate, fast, and robust structure and motion from casual dynamic videos. arXiv preprint arXiv:2412.04463, 2024. +[24] Z. Li, D. Wang, K. Chen, Z. Lv, T. Nguyen-Phuoc, M. Lee, J.-B. Huang, L. Xiao, C. Zhang, Y. Zhu, et al. Lirm: Large inverse rendering model for progressive reconstruction of shape, materials and view-dependent radiance fields. arXiv preprint arXiv:2504.20026, 2025. +[25] H. Liang, J. Ren, A. Mirzaei, A. Torralba, Z. Liu, I. Gilitschenski, S. Fidler, C. Oztireli, H. Ling, Z. Gojcic, et al. Feed-forward bullet-time reconstruction of dynamic scenes from monocular videos. arXiv preprint arXiv:2412.03526, 2024. +[26] I. Loshchilov and F. Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016. +[27] I. Loshchilov and F. Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. +[28] Z. Lv, N. Charron, P. Moulon, A. Gamino, C. Peng, C. Sweeney, E. Miller, H. Tang, J. Meissner, J. Dong, et al. Aria everyday activities dataset. arXiv preprint arXiv:2402.13349, 2024. +[29] L. Ma, Y. Ye, F. Hong, V. Guzov, Y. Jiang, R. Postyeni, L. Pesqueira, A. Gamino, V. Baiyya, H. J. Kim, et al. Nymeria: A massive collection of multimodal egocentric daily motion in the wild. In European Conference on Computer Vision, pages 445-465. Springer, 2024. +[30] B. Mildenhall, P. P. Srinivasan, M. Tancik, J. T. Barron, R. Ramamoorthi, and R. Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. +[31] R. Murai, E. Dexheimer, and A. J. Davison. MASt3R-SLAM: Real-time dense SLAM with 3D reconstruction priors. In CVPR, June 2025. +[32] R. A. Newcombe, D. Fox, and S. M. Seitz. Dynamicfusion: Reconstruction and tracking of non-rigid scenes in real-time. In CVPR, 2015. +[33] M. Oquab, T. Darcet, T. Moutakanni, H. Vo, M. Szafraniec, V. Khalidov, P. Fernandez, D. Haziza, F. Massa, A. El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. +[34] X. Pan, N. Charron, Y. Yang, S. Peters, T. Whelan, C. Kong, O. Parkhi, R. Newcombe, and Y. C. Ren. Aria digital twin: A new benchmark dataset for egocentric 3d machine perception. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20133-20143, 2023. +[35] Z. Pan, Z. Yang, X. Zhu, and L. Zhang. Efficient4d: Fast dynamic 3d object generation from a single-view video. arXiv preprint arXiv:2401.08742, 2024. +[36] K. Park, U. Sinha, J. T. Barron, S. Bouaziz, D. B. Goldman, S. M. Seitz, and R. Martin-Brualla. Nerfies: Deformable neural radiance fields. In ICCV, 2021. +[37] K. Park, U. Sinha, P. Hedman, J. T. Barron, S. Bouaziz, D. B. Goldman, R. Martin-Brualla, and S. M. Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. arXiv preprint arXiv:2106.13228, 2021. +[38] A. Paszke, S. Gross, F. Massa, A. Lerer, J. Bradbury, G. Chanan, T. Killeen, Z. Lin, N. Gimelshein, L. Antiga, A. Desmaison, A. Kopf, E. Yang, Z. DeVito, M. Raison, A. Tejani, S. Chilamkurthy, B. Steiner, L. Fang, J. Bai, and S. Chintala. Pytorch: An imperative style, high-performance deep learning library. In NeurIPS, 2019. +[39] L. Piccinelli, Y.-H. Yang, C. Sakaridis, M. Segu, S. Li, L. Van Gool, and F. Yu. Undepth: Universal monocular metric depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10106-10116, 2024. +[40] L. Piccinelli, C. Sakaridis, Y.-H. Yang, M. Segu, S. Li, W. Abbeloos, and L. Van Gool. Unidepthv2: Universal monocular metric depth estimation made simpler. arXiv preprint arXiv:2502.20110, 2025. +[41] D. Qi, T. Yang, B. Wang, X. Zhang, and W. Zhang. Predicting 3d representations for dynamic scenes. arXiv preprint arXiv:2501.16617, 2025. +[42] N. Ravi, J. Reizenstein, D. Novotny, T. Gordon, W.-Y. Lo, J. Johnson, and G. Gkioxari. Accelerating 3d deep learning with pytorch3d. arXiv:2007.08501, 2020. +[43] J. Ren, C. Xie, A. Mirzaei, K. Kreis, Z. Liu, A. Torralba, S. Fidler, S. W. Kim, H. Ling, et al. L4gm: Large 4d gaussian reconstruction model. Advances in Neural Information Processing Systems, 37:56828-56858, 2024. + +[44] J. L. Schonberger and J.-M. Frahm. Structure-from-motion revisited. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4104-4113, 2016. +[45] J. Shah, G. Bikshandi, Y. Zhang, V. Thakkar, P. Ramani, and T. Dao. Flashattention-3: Fast and accurate attention with asynchrony and low-precision. Advances in Neural Information Processing Systems, 37: 68658–68685, 2024. +[46] S. Sinha, R. Shapovalov, J. Reizenstein, I. Rocco, N. Neverova, A. Vedaldi, and D. Novotny. Common pets in 3d: Dynamic new-view synthesis of real-life deformable categories. CVPR, 2023. +[47] M. Slavcheva, M. Baust, D. Cremers, and S. Ilic. Killingfusion: Non-rigid 3d reconstruction without correspondences. In CVPR, July 2017. +[48] J. Sturm, N. Engelhard, F. Endres, W. Burgard, and D. Cremers. A benchmark for the evaluation of rgb-d slam systems. In 2012 IEEE/RSJ international conference on intelligent robots and systems, pages 573-580. IEEE, 2012. +[49] Z. Teed and J. Deng. Raft: Recurrent all-pairs field transforms for optical flow. In European conference on computer vision, pages 402-419. Springer, 2020. +[50] V. Tschernezki, A. Darkhalil, Z. Zhu, D. Fouhey, I. Laina, D. Larlus, D. Damen, and A. Vedaldi. Epic fields: Marrying 3d geometry and video understanding. Advances in Neural Information Processing Systems, 36: 26485–26500, 2023. +[51] A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. +[52] P. Wang and Y. Shi. Imagedream: Image-prompt multi-view diffusion for 3d generation. arXiv preprint arXiv:2312.02201, 2023. +[53] Q. Wang, V. Ye, H. Gao, J. Austin, Z. Li, and A. Kanazawa. Shape of motion: 4d reconstruction from a single video. arXiv preprint arXiv:2407.13764, 2024. +[54] Q. Wang, Y. Zhang, A. Holynski, A. A. Efros, and A. Kanazawa. Continuous 3d perception model with persistent state, 2025. +[55] Y. Wang, P. Yang, Z. Xu, J. Sun, Z. Zhang, C. Yuan, H. Bao, S. Peng, and X. Zhou. Freetimegs: Free gaussians at anytime anywhere for dynamic scene reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2025. +[56] Z. Wang, A. C. Bovik, H. R. Sheikh, and E. P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE TIP, 2004. +[57] G. Wu, T. Yi, J. Fang, L. Xie, X. Zhang, W. Wei, W. Liu, Q. Tian, and W. Xinggang. 4d gaussian splatting for real-time dynamic scene rendering. arXiv preprint arXiv:2310.08528, 2023. +[58] R. Wu, R. Gao, B. Poole, A. Trevithick, C. Zheng, J. T. Barron, and A. Holynski. Cat4d: Create anything in 4d with multi-view video diffusion models. arXiv preprint arXiv:2411.18613, 2024. +[59] D. Xie, S. Bi, Z. Shu, K. Zhang, Z. Xu, Y. Zhou, A. Kaufman, X. Sun, and H. Tan. Lrm-zero: Training large reconstruction models with synthesized data. arXiv preprint arXiv:2406.09371, 2024. +[60] Y. Xie, C.-H. Yao, V. Voleti, H. Jiang, and V. Jampani. Sv4d: Dynamic 3d content generation with multi-frame and multi-view consistency. arXiv preprint arXiv:2407.17470, 2024. +[61] Y. Xu, Z. Shi, W. Yifan, H. Chen, C. Yang, S. Peng, Y. Shen, and G. Wetzstein. Grm: Large gaussian reconstruction model for efficient 3d reconstruction and generation. In European Conference on Computer Vision, pages 1–20. Springer, 2024. +[62] Z. Xu, Y. Xu, Z. Yu, S. Peng, J. Sun, H. Bao, and X. Zhou. Representing long volumetric video with temporal gaussian hierarchy. ACM Transactions on Graphics, 43(6), November 2024. URL https://zju3dv.github.io/longvolcap. +[63] J. Yang, J. Huang, Y. Chen, Y. Wang, B. Li, Y. You, M. Igl, A. Sharma, P. Karkus, D. Xu, B. Ivanovic, Y. Wang, and M. Pavone. Storm: Spatio-temporal reconstruction model for large-scale outdoor scenes. arXiv preprint arXiv:2501.00602, 2025. +[64] L. Yang, B. Kang, Z. Huang, Z. Zhao, X. Xu, J. Feng, and H. Zhao. Depth anything v2. Advances in Neural Information Processing Systems, 37:21875-21911, 2024. +[65] Z. Yang, H. Yang, Z. Pan, and L. Zhang. Real-time photorealistic dynamic scene representation and rendering with 4d gaussian splatting. In International Conference on Learning Representations (ICLR), 2024. +[66] C. Ye, L. Qiu, X. Gu, Q. Zuo, Y. Wu, Z. Dong, L. Bo, Y. Xiu, and X. Han. Stabilnormal: Reducing diffusion variance for stable and sharp normal. ACM Transactions on Graphics (TOG), 43(6):1-18, 2024. +[67] V. Ye, R. Li, J. Kerr, M. Turkulainen, B. Yi, Z. Pan, O. Seiskari, J. Ye, J. Hu, M. Tancik, et al. gsplat: An open-source library for gaussian splatting. Journal of Machine Learning Research, 26(34):1-17, 2025. + +[68] J. Zhang, C. Herrmann, J. Hur, V. Jampani, T. Darrell, F. Cole, D. Sun, and M.-H. Yang. Monst3r: A simple approach for estimating geometry in the presence of motion. arXiv preprint arXiv:2410.03825, 2024. +[69] K. Zhang, S. Bi, H. Tan, Y. Xiangli, N. Zhao, K. Sunkavalli, and Z. Xu. Gs-lrm: Large reconstruction model for 3d gaussian splatting. European Conference on Computer Vision, 2024. +[70] R. Zhang, P. Isola, A. A. Efros, E. Shechtman, and O. Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. +[71] H. Zhao, J. Chen, L. Wang, and H. Lu. Arkitrack: A new diverse dataset for tracking using mobile rgb-data. CVPR, 2023. +[72] C. Ziwen, H. Tan, K. Zhang, S. Bi, F. Luan, Y. Hong, L. Fuxin, and Z. Xu. Long-lrm: Long-sequence large reconstruction model for wide-coverage gaussian splats. arXiv preprint arXiv:2410.12781, 2024. + +# NeurIPS Paper Checklist + +# 1. Claims + +Question: Do the main claims made in the abstract and introduction accurately reflect the paper's contributions and scope? + +Answer: [Yes] + +Justification: Our abstract and introduction has clearly stated the contributions and claims we made. We have articulated the technical challenges we have addressed compared to prior work, and stated the novelty of our methods both in results end-to-end, and analyzed via careful ablations. + +Guidelines: + +- The answer NA means that the abstract and introduction do not include the claims made in the paper. +- The abstract and/or introduction should clearly state the claims made, including the contributions made in the paper and important assumptions and limitations. A No or NA answer to this question will not be perceived well by the reviewers. +- The claims made should match theoretical and experimental results, and reflect how much the results can be expected to generalize to other settings. +- It is fine to include aspirational goals as motivation as long as it is clear that these goals are not attained by the paper. + +# 2. Limitations + +Question: Does the paper discuss the limitations of the work performed by the authors? + +Answer: [Yes] + +Justification: We have included the limitations of our current methods in the conclusion section. Overall we observe two main limitations for the current methods. First, we observe the generalization capabilities of our method is related to the training data. The real world videos are captured from different devices and have various characteristics. We observed that the generalization capability of the model can improve drastic to videos captured by the devices when there are training data collected by the same device used in training. In our experiments, we have extensively used data from Project Aria, Iphone and GoPro, which however is not representative enough for general in the wild videos. We see domain gap due to this reason and we believe it can be resolved in the future to further scale our training to videos with larger varieties. Second, we have further highlighted the limitation of our predicted 4DGS when viewed from extreme view angles. This has been a persistent challenge in monocular 4D reconstruction and remain as a challenge. + +Guidelines: + +- The answer NA means that the paper has no limitation while the answer No means that the paper has limitations, but those are not discussed in the paper. +- The authors are encouraged to create a separate "Limitations" section in their paper. +- The paper should point out any strong assumptions and how robust the results are to violations of these assumptions (e.g., independence assumptions, noiseless settings, model well-specification, asymptotic approximations only holding locally). The authors should reflect on how these assumptions might be violated in practice and what the implications would be. +- The authors should reflect on the scope of the claims made, e.g., if the approach was only tested on a few datasets or with a few runs. In general, empirical results often depend on implicit assumptions, which should be articulated. +- The authors should reflect on the factors that influence the performance of the approach. For example, a facial recognition algorithm may perform poorly when image resolution is low or images are taken in low lighting. Or a speech-to-text system might not be used reliably to provide closed captions for online lectures because it fails to handle technical jargon. +- The authors should discuss the computational efficiency of the proposed algorithms and how they scale with dataset size. + +- If applicable, the authors should discuss possible limitations of their approach to address problems of privacy and fairness. +- While the authors might fear that complete honesty about limitations might be used by reviewers as grounds for rejection, a worse outcome might be that reviewers discover limitations that aren't acknowledged in the paper. The authors should use their best judgment and recognize that individual actions in favor of transparency play an important role in developing norms that preserve the integrity of the community. Reviewers will be specifically instructed to not penalize honesty concerning limitations. + +# 3. Theory assumptions and proofs + +Question: For each theoretical result, does the paper provide the full set of assumptions and a complete (and correct) proof? + +Answer: [NA] + +Justification: N/A + +Guidelines: + +- The answer NA means that the paper does not include theoretical results. +- All the theorems, formulas, and proofs in the paper should be numbered and cross-referenced. +- All assumptions should be clearly stated or referenced in the statement of any theorems. +- The proofs can either appear in the main paper or the supplemental material, but if they appear in the supplemental material, the authors are encouraged to provide a short proof sketch to provide intuition. +- Inversely, any informal proof provided in the core of the paper should be complemented by formal proofs provided in appendix or supplemental material. +- Theorems and Lemmas that the proof relies upon should be properly referenced. + +# 4. Experimental result reproducibility + +Question: Does the paper fully disclose all the information needed to reproduce the main experimental results of the paper to the extent that it affects the main claims and/or conclusions of the paper (regardless of whether the code and data are provided or not)? + +Answer: [Yes] + +Justification: We have provided details of our methods in implementation with the hyperparameters. We have full disclosed our training datasets and training settings. We also provided more details in training and evaluation that cannot be sufficiently described in the main paper due to page limit. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- If the paper includes experiments, a No answer to this question will not be perceived well by the reviewers: Making the paper reproducible is important, regardless of whether the code and data are provided or not. +- If the contribution is a dataset and/or model, the authors should describe the steps taken to make their results reproducible or verifiable. +- Depending on the contribution, reproducibility can be accomplished in various ways. For example, if the contribution is a novel architecture, describing the architecture fully might suffice, or if the contribution is a specific model and empirical evaluation, it may be necessary to either make it possible for others to replicate the model with the same dataset, or provide access to the model. In general, releasing code and data is often one good way to accomplish this, but reproducibility can also be provided via detailed instructions for how to replicate the results, access to a hosted model (e.g., in the case of a large language model), releasing of a model checkpoint, or other means that are appropriate to the research performed. +- While NeurIPS does not require releasing code, the conference does require all submissions to provide some reasonable avenue for reproducibility, which may depend on the nature of the contribution. For example + +(a) If the contribution is primarily a new algorithm, the paper should make it clear how to reproduce that algorithm. +(b) If the contribution is primarily a new model architecture, the paper should describe the architecture clearly and fully. +(c) If the contribution is a new model (e.g., a large language model), then there should either be a way to access this model for reproducing the results or a way to reproduce the model (e.g., with an open-source dataset or instructions for how to construct the dataset). +(d) We recognize that reproducibility may be tricky in some cases, in which case authors are welcome to describe the particular way they provide for reproducibility. In the case of closed-source models, it may be that access to the model is limited in some way (e.g., to registered users), but it should be possible for other researchers to have some path to reproducing or verifying the results. + +# 5. Open access to data and code + +Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material? + +Answer: No + +Justification: We cannot provide the code implementation at the submission time due to legal process. We will intend to release our model upon the acceptance of the paper and legal review. + +Guidelines: + +- The answer NA means that paper does not include experiments requiring code. +- Please see the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details. +- While we encourage the release of code and data, we understand that this might not be possible, so "No" is an acceptable answer. Papers cannot be rejected simply for not including code, unless this is central to the contribution (e.g., for a new open-source benchmark). +- The instructions should contain the exact command and environment needed to run to reproduce the results. See the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details. +- The authors should provide instructions on data access and preparation, including how to access the raw data, preprocessed data, intermediate data, and generated data, etc. +- The authors should provide scripts to reproduce all experimental results for the new proposed method and baselines. If only a subset of experiments are reproducible, they should state which ones are omitted from the script and why. +- At submission time, to preserve anonymity, the authors should release anonymized versions (if applicable). +- Providing as much information as possible in supplemental material (appended to the paper) is recommended, but including URLs to data and code is permitted. + +# 6. Experimental setting/details + +Question: Does the paper specify all the training and test details (e.g., data splits, hyperparameters, how they were chosen, type of optimizer, etc.) necessary to understand the results? + +Answer: [Yes] + +Justification: We have provided details of training datasets, evaluation datasets, our implementation details and hyperparameters. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The experimental setting should be presented in the core of the paper to a level of detail that is necessary to appreciate the results and make sense of them. +- The full details can be provided either with the code, in appendix, or as supplemental material. + +# 7. Experiment statistical significance + +Question: Does the paper report error bars suitably and correctly defined or other appropriate information about the statistical significance of the experiments? + +Answer: [Yes] + +Justification: We provided the full evaluations with error bars in standard deviation for our evaluation against baselines. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The authors should answer "Yes" if the results are accompanied by error bars, confidence intervals, or statistical significance tests, at least for the experiments that support the main claims of the paper. +- The factors of variability that the error bars are capturing should be clearly stated (for example, train/test split, initialization, random drawing of some parameter, or overall run with given experimental conditions). +- The method for calculating the error bars should be explained (closed form formula, call to a library function, bootstrap, etc.) +- The assumptions made should be given (e.g., Normally distributed errors). +- It should be clear whether the error bar is the standard deviation or the standard error of the mean. +- It is OK to report 1-sigma error bars, but one should state it. The authors should preferably report a 2-sigma error bar than state that they have a $96\%$ CI, if the hypothesis of Normality of errors is not verified. +- For asymmetric distributions, the authors should be careful not to show in tables or figures symmetric error bars that would yield results that are out of range (e.g. negative error rates). +- If error bars are reported in tables or plots, The authors should explain in the text how they were calculated and reference the corresponding figures or tables in the text. + +# 8. Experiments compute resources + +Question: For each experiment, does the paper provide sufficient information on the computer resources (type of compute workers, memory, time of execution) needed to reproduce the experiments? + +Answer: [Yes] + +Justification: We have provided details of our compute resources in reproducing the model in implementation details. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The paper should indicate the type of compute workers CPU or GPU, internal cluster, or cloud provider, including relevant memory and storage. +- The paper should provide the amount of compute required for each of the individual experimental runs as well as estimate the total compute. +- The paper should disclose whether the full research project required more compute than the experiments reported in the paper (e.g., preliminary or failed experiments that didn't make it into the paper). + +# 9. Code of ethics + +Question: Does the research conducted in the paper conform, in every respect, with the NeurIPS Code of Ethics https://neurips.cc/public/EthicsGuidelines? + +Answer: [Yes] + +Justification: We have fully reviewed the code of ethics. + +Guidelines: + +- The answer NA means that the authors have not reviewed the NeurIPS Code of Ethics. + +- If the authors answer No, they should explain the special circumstances that require a deviation from the Code of Ethics. +- The authors should make sure to preserve anonymity (e.g., if there is a special consideration due to laws or regulations in their jurisdiction). + +# 10. Broader impacts + +Question: Does the paper discuss both potential positive societal impacts and negative societal impacts of the work performed? + +Answer: [NA] + +Justification: We include a short discussion of social impacts in the supplementary materials. As an early research efforts, we have not foresee our method will significantly impact the social topics at this stage. + +Guidelines: + +- The answer NA means that there is no societal impact of the work performed. +- If the authors answer NA or No, they should explain why their work has no societal impact or why the paper does not address societal impact. +- Examples of negative societal impacts include potential malicious or unintended uses (e.g., disinformation, generating fake profiles, surveillance), fairness considerations (e.g., deployment of technologies that could make decisions that unfairly impact specific groups), privacy considerations, and security considerations. +- The conference expects that many papers will be foundational research and not tied to particular applications, let alone deployments. However, if there is a direct path to any negative applications, the authors should point it out. For example, it is legitimate to point out that an improvement in the quality of generative models could be used to generate deepfakes for disinformation. On the other hand, it is not needed to point out that a generic algorithm for optimizing neural networks could enable people to train models that generate Deepfakes faster. +- The authors should consider possible harms that could arise when the technology is being used as intended and functioning correctly, harms that could arise when the technology is being used as intended but gives incorrect results, and harms following from (intentional or unintentional) misuse of the technology. +- If there are negative societal impacts, the authors could also discuss possible mitigation strategies (e.g., gated release of models, providing defenses in addition to attacks, mechanisms for monitoring misuse, mechanisms to monitor how a system learns from feedback over time, improving the efficiency and accessibility of ML). + +# 11. Safeguards + +Question: Does the paper describe safeguards that have been put in place for responsible release of data or models that have a high risk for misuse (e.g., pretrained language models, image generators, or scraped datasets)? + +Answer: [NA] + +Justification: Although we intend to release model if upon approval, we do not foresee misuse of our model in sensitive topics. Our models can only reconstruct 4D from input an user provider and will not generate or hallucinate sensitive content. Our training data also has strong compliance to privacy. + +Guidelines: + +- The answer NA means that the paper poses no such risks. +- Released models that have a high risk for misuse or dual-use should be released with necessary safeguards to allow for controlled use of the model, for example by requiring that users adhere to usage guidelines or restrictions to access the model or implementing safety filters. +- Datasets that have been scraped from the Internet could pose safety risks. The authors should describe how they avoided releasing unsafe images. +- We recognize that providing effective safeguards is challenging, and many papers do not require this, but we encourage authors to take this into account and make a best faith effort. + +# 12. Licenses for existing assets + +Question: Are the creators or original owners of assets (e.g., code, data, models), used in the paper, properly credited and are the license and terms of use explicitly mentioned and properly respected? + +Answer: [Yes] + +Justification: We have properly cited all the work, including data, model and code implementation that influence this work. All datasets, models and code have public research license to be used safely. + +Guidelines: + +- The answer NA means that the paper does not use existing assets. +- The authors should cite the original paper that produced the code package or dataset. +- The authors should state which version of the asset is used and, if possible, include a URL. +- The name of the license (e.g., CC-BY 4.0) should be included for each asset. +- For scraped data from a particular source (e.g., website), the copyright and terms of service of that source should be provided. +- If assets are released, the license, copyright information, and terms of use in the package should be provided. For popular datasets, paperswithcode.com/datasets has curated licenses for some datasets. Their licensing guide can help determine the license of a dataset. +- For existing datasets that are re-packaged, both the original license and the license of the derived asset (if it has changed) should be provided. +- If this information is not available online, the authors are encouraged to reach out to the asset's creators. + +# 13. New assets + +Question: Are new assets introduced in the paper well documented and is the documentation provided alongside the assets? + +Answer: [NA] + +Justification: N/A + +Guidelines: + +- The answer NA means that the paper does not release new assets. +- Researchers should communicate the details of the dataset/code/model as part of their submissions via structured templates. This includes details about training, license, limitations, etc. +- The paper should discuss whether and how consent was obtained from people whose asset is used. +- At submission time, remember to anonymize your assets (if applicable). You can either create an anonymized URL or include an anonymized zip file. + +# 14. Crowdsourcing and research with human subjects + +Question: For crowdsourcing experiments and research with human subjects, does the paper include the full text of instructions given to participants and screenshots, if applicable, as well as details about compensation (if any)? + +Answer: [NA] + +Justification: N/A + +Guidelines: + +- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. +- Including this information in the supplemental material is fine, but if the main contribution of the paper involves human subjects, then as much detail as possible should be included in the main paper. + +- According to the NeurIPS Code of Ethics, workers involved in data collection, curation, or other labor should be paid at least the minimum wage in the country of the data collector. + +# 15. Institutional review board (IRB) approvals or equivalent for research with human subjects + +Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals (or an equivalent approval/review based on the requirements of your country or institution) were obtained? + +Answer: [NA] + +Justification: N/A + +Guidelines: + +- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. +- Depending on the country in which research is conducted, IRB approval (or equivalent) may be required for any human subjects research. If you obtained IRB approval, you should clearly state this in the paper. +- We recognize that the procedures for this may vary significantly between institutions and locations, and we expect authors to adhere to the NeurIPS Code of Ethics and the guidelines for their institution. +- For initial submissions, do not include any information that would break anonymity (if applicable), such as the institution conducting the review. + +# 16. Declaration of LLM usage + +Question: Does the paper describe the usage of LLMs if it is an important, original, or non-standard component of the core methods in this research? Note that if the LLM is used only for writing, editing, or formatting purposes and does not impact the core methodology, scientific rigorousness, or originality of the research, declaration is not required. + +Answer: [NA] + +Justification: N/A + +Guidelines: + +- The answer NA means that the core method development in this research does not involve LLMs as any important, original, or non-standard components. +- Please refer to our LLM policy (https://neurips.cc/Conferences/2025/LLM) for what should or should not be described. + +# Contents + +1 Introduction 2 +2 Related Work 3 +3 Method 3 + +3.1 Feed-Forward Dynamic Gaussian Prediction 3 +3.2 Multi-level Pixel & Token Density Control 5 +3.3 Training 6 + +4 Implementation Detail 6 + +5 Experiments 7 +6 Conclusion 10 + +A Additional Details 23 + +A.1 Additional Details on the Multi-Level Attention Module 23 +A.2 Additional Details on the Densification & Pruning of the Dynamic Gaussian 23 +A.3 Additional Details on Datasets and Baselines 24 +A.4 Additional Details on the Number of Gaussians 24 + +B Additional Results 25 + +B.1 Qualitative Results of the Ablation Study 25 +B.2 Results and Discussion on Pruning Pattern Selection 25 +B.3 Supplementary Videos 26 + +C Additional DisCUssions 26 + +C.1 Additional Discussions on Expert Models 26 +C.2 Additional Discussions on the Choice of the Dynamic Gaussian Representation 26 +C.3 Additional Discussions on Metric Scale Cameras 27 +C.4 Additional Discussions on Limitations 27 + +D Social Impact 27 + +![](images/85d6484ff7c9524d59978719493695ec9d79c3ec55deff5661e9d436e43aecee.jpg) +Figure 4: The predicted opacity map $(\in R^{N\times H\times W})$ of the pixel-aligned dynamic Gaussians from 4DGT and the computed histogram $(\in R^{p\times p})$ of the activation distribution. The right section shows the difference between histogram thresholding (Ours) and other filtering methods (randomly or uniformly selecting the Gaussians to keep) for reducing the number of Gaussians. + +# A Additional Details + +# A.1 Additional Details on the Multi-Level Attention Module + +As mentioned in the Method section of the main paper, aside from the number of Gaussians, another efficiency-limiting factor is the number of tokens in the large number of high-resolution images. In the second stage of training, we increase the spatial and temporal resolution by a factor of $R_{s}$ and $R_{t}$ , respectively. The number of patches participating in the self-attention module $\mathcal{F}$ increases by a factor of $R_{s}^{2} \cdot R_{t}$ , which will slow down optimization and inference significantly. To mitigate this, we propose a temporal level-of-detail attention mechanism to reduce the computational cost. Specifically, noticing that the computational complexity of the self-attention module is $O(n^{2})$ (simplified from $O(n^{2} + n)$ ) where $n$ is the number of tokens [51], We propose to divide the $N$ input frames into $M$ equal trunks in the highest level. This division limits the attention mechanism in the temporal dimension, but reduces the computation of calculating $n$ total tokens to $O\left(\frac{n^2}{M}\right)$ . To balance spatial-temporal samples, we construct a temporal level-of-detail structure by alternating the temporal range and spatial resolution, achieving a much smaller overhead while maintaining the ability to handle long temporal windows. For each level $l$ , we reduce the spatial resolution by a factor of $2^{l}$ and increase temporal samples by 2. This results in a computational complexity of: + +$$ +O \left(\frac {n ^ {2}}{M} + \dots + \frac {n ^ {2}}{M \cdot 2 ^ {L - 1}}\right) = O \left(\frac {n ^ {2}}{M \cdot 2 ^ {L - 1}} \cdot \sum_ {l = 0} ^ {L - 1} 2 ^ {l}\right) = O \left(n ^ {2} \cdot \frac {2 ^ {L} - 1}{M \cdot 2 ^ {L - 1}}\right) \approx O \left(\frac {2 n ^ {2}}{M}\right). \tag {12} +$$ + +Empirically, we use level $L = 3$ and $M = 4$ , which leads to an approximately 2 times reduction in the computational cost. + +# A.2 Additional Details on the Densification & Pruning of the Dynamic Gaussian + +In appendix A.2, we visualize the predicted opacity map of the pixel-aligned dynamic Gaussians. It shows clear patterns of the activation of the pixel inside each patch, especially for the dynamic regions. Notably, randomly or uniformly selecting the Gaussians to keep will lead to a significant number of active Gaussians being pruned, effectively removing the ability of the model to model the dynamic parts, while our histogram thresholding scheme can effectively keep the Gaussians that are contributing. These strategies blend the densification and pruning strategies of Gaussian representations [19] and the multi-stage training strategy of ViT models [54], effectively introducing a density control scheme for the feed-forward prediction pipeline. + +# A.3 Additional Details on Datasets and Baselines + +For each dataset used in training [28, 12, 2, 71, 29, 46, 50], we select $99.15\%$ of the sequences as the training set and hold out the rest. For the datasets used in evaluation: + +- ADT [34]: We select 4 subsequences for validating the reconstruction performance: + +-Apartment_release-multiuser_cook_seq141_M1292 +-Apartment_releaseMULTiskeleton_party_seq114_M1292 +-Apartment_releasemealSkeleton_seq135_M1292 +-Apartment_release_work}skeleton_seq137_M1292 + +- DyCheck [11]: We use all 6 sequences with 3 views, and follow [53, 11] to apply the covisibility mask before computing metrics on novel views: + +- apple, block, space-out, spin, paper-windmill, teddy + +- TUM [48]: We sclet 3 subsequences for evaluation: + +- rgbd_dataset_freiburg2Desk_with_person +- rgbd_dataset_freiburg3_walking_halfsphere +- rgbd_dataset_freiburg3_sitting_halfsphere + +- EgoExo4D [12]: We select 3 subsequences from the hold-out sequences: + +- cmu_bike01_2, sfu_cooking015_2, uniandes_bouldering_003_10 + +- Nymeria [29]: We select 2 sequences from the hold-out set: + +- 20230607_s0_james_johnson_ACT1_7xwm28 +- 20230612_s1_christina_jones_ACT0_u2r0z8 + +- AEA [28]: We select the loc5.script5_seq7_rec1 sequence from the hold-out set. +- Hot3D [2]: We select the P0020_ff537251 sequence from the hold-out set. + +The testing sequences from EgoExo4D, AEA, and Hot3D are denoted as $Aria$ in all comparisons. + +Note that we do not make comparisons with CAT4D [58], BulletTimer [25] since they provide neither the source code nor the pre-trained models as of writing. + +# A.4 Additional Details on the Number of Gaussians + +The number of dynamic Gaussians predicted by the first stage is pixel-aligned, and can be computed as (derived from eq. (4)): + +$$ +N _ {\mathbf {g}} = N \times H \times W. \tag {13} +$$ + +For a resolution of $252 \times 252$ and 16 images (half spatial resolution and $1/4 \times$ temporal resolution of the second stage), this results in 508,032 (0.5M) Gaussians. + +In the second stage, the resolution is increased to $504 \times 504$ and 64 images. With the proposed patch-based pruning strategy, the number of Gaussians can be computed as (derived from eq. (4)): + +$$ +N _ {\mathbf {g}} = N \times H \times W \times \frac {S}{p ^ {2}}, \tag {14} +$$ + +which results in a total of 829,440 Gaussians. + +Finally, the proposed multi-level spatial attention mechanism introduces two additional downsampled outputs with $1/4 \times$ and $1/16 \times$ the spatial resolution, respectively. This leads to a final Gaussian count of: + +$$ +N _ {\mathbf {g}} = 8 2 9, 4 4 0 \times \left(1 + \frac {1}{4} + \frac {1}{1 6}\right) = 1, 0 8 8, 6 4 0 (1 \mathrm {M}). \tag {15} +$$ + +for the second stage. + +Thanks to our proposed selective activation pruning strategy, the number of Gaussians only increases by $1 \times$ while the space-time resolution increases $15 \times$ . + +![](images/620cc68a8b6486b6c442b1a625d000d3963ca7af96ba05f271a063d90cd9412e.jpg) +Figure 5: Ablation study on proposed components. + +# B Additional Results + +# B.1 Qualitative Results of the Ablation Study + +In fig. 5, we show the ablation study results for the first and second stage training, respectively. As shown in the table and figure, our proposed loss and representation effectively model the dynamic regions and improve the reconstruction quality. Moreover, the proposed density control scheme effectively regularized the number of Gaussians with increased input count and input resolution, greatly improving details and avoiding using too much memory. Adding larger-scale datasets [12, 71] helps generalization for both in-domain and out-of-domain datasets. + +# B.2 Results and Discussion on Pruning Pattern Selection + +In eq. (7), after computing $\mathbf{H}$ once following the first-stage training, the top $S$ entries are selected to define a shared pruning pattern for the second-stage training. This approach effectively shares the same pattern across all patches. + +The motivation for this design is twofold: + +- Empirical activation consistency: Across the $p^2$ pixels within each patch, the model consistently favors similar pixels for activation and subsequent use in Gaussian rendering across all patches. This aggregation leads to a clear shared activation pattern, as visualized + +in appendix A.2 (right), where the predicted opacity maps exhibit this pattern consistently. Quantitative results (table 4) show that the shared pruning pattern achieves on-par or better performance compared to recalculating the pattern for each patch on-the-fly. + +- Implementation efficiency: Using a shared pruning pattern enables efficient implementation by discarding unused rows in the weight matrix of the final fully-connected layer of the decoder heads. This is considerably less resource-intensive regarding both memory usage and computation time, compared to dynamically sorting the opacity values for every patch during runtime. + +Table 4: Comparison of pruning strategies on the ADT [34] dataset. "On-the-fly" computes a unique pruning pattern for each patch. "Shared" (ours) uses a single pattern for all patches. + +
MethodPSNR↑LPIPS↓RMSE↓Degree↓Speed Overhead
On-the-fly28.360.2410.7825.95Yes
Shared (Ours)28.790.2420.7225.84No
+ +# B.3 Supplementary Videos + +We attach additional video results in the supplementary video material. The supplementary video is structured as follows: + +- 00:00:00-00:00:15: Brief introduction to the input & output setting and goal of the paper. +- 00:00:15-00:00:45: Reconstruction and novel view rendering results for the depth, normal, optical flow, dynamic mask, and appearance inferred in a rolling window fashion over a long video. +- 00:00:45-00:01:05: More qualitative video results from other datasets. +- 00:01:05-00:01:35: Comparison with baseline methods StaticLRM [69], L4GM [43] and Shape-of-Motion [53]. +- 00:01:35-00:02:00: Ablation study of the proposed components. + +# C Additional Disussions + +# C.1 Additional Discussions on Expert Models + +Aside from the expert normal and depth guidance, one natural way to improve the outputs' temporal consistency is to incorporate a flow expert's guidance [49]. It seems an optical flow expert like RAFT [49] can easily be plugged into our pipelin, but it's non-trivial in practice. There are two reasons we did not adopt such a flow model for guidance: + +- In our preliminary experiments, we discovered that the estimated optical flow exhibits strong inconsistency, often reaching more than 10 pixels in cycle consistency errors. This in turn makes the training of the 4DGT model unstable and leads to NaN values in the prediction. +- A tracking expert model like TAPIR [6] would produce much more consistent results for guiding the prediction of dynamic Gaussians, as shown by Shape-of-Motion [53]. However, the computation of such dense all-to-all tracking is extremely time-consuming (a few hours for a 128-frame clip), making it impractical for our large-scale training setup (1000 hours of video data). + +Due to these reasons, we leave the addition of the flow expert model's guidance to improve the temporal consistency to future work. + +# C.2 Additional Discussions on the Choice of the Dynamic Gaussian Representation + +The main purpose for our choice of dynamic Gaussian representation is to enable seamless integration to a feed-forward prediction pipeline, which can be used for self-supervised training on general + +dynamic videos. Compared to per-frame 3DGS, static 3DGS, flow vector field or decomposed motion bases, we found the explicit modeling of the motion terms of dynamic Gaussians (adapted from FTGS [65], originally proposed in 4DGS [19] (L124, L129)) to be a better fit for this purpose. + +- Compared to a per-frame 3DGS [65] representation (denoted as the *per-frame* variant in the ablation studies of the paper), our representation enables the integration of space-time information as a 4D Gaussian with a non-zero life-span, automatically encodes information across multiple frames, making it possible to train the 4DGT model in a self-supervised manner on monocular videos. In the most extreme case with infinite life-span, the representation is reduced to a purely static 3DGS (denoted as the *Static-LRM* baseline in the paper) and would only work on static scenes. Compared to per-frame 3DGS and static 3DGS, our representation can freely encode the different levels of motion speed (from 0 to $\infty$ ) of the dynamic scene. Comparison against the *StaticLRM* baseline and *per-frame 3DGS* variant can be found in Table 2(a) of the main paper. +- Compared to a 3D flow vector field representation, like DynamicGaussians [65], our representation can be easily integrated into the pixel-aligned feed-forward prediction pipeline for patch-based vision transformers. However, their flow vector field, which is typically encoded by an MLP, would be much harder to predict in a feed-forward manner. Similar problems exist for the rigid motion representation used in Shape-of-Motion [53], since it's extremely ill-posed to accurately predict their motion bases and coefficients without complicated initialization. In comparison, our representation does not require such careful initialization. In practice, we simply set all $\mathbf{v},\omega$ to zero, $\mathbf{t}$ to the timestamp of the corresponding frame, and $\mathbf{l}$ to a large value (50s). +- Compared to other implicit network-based methods like NeRF [30] (as used in Pred. 3D Repr. [41]), a Gaussian-based representation would enable much more efficient rendering and training. + +# C.3 Additional Discussions on Metric Scale Cameras + +We empirically find the model works best when trained and inferred with metric-scale cameras due to the ambiguity in depth-scale estimation. Notably, the model doesn't rely on fully accurate scaling to perform well, as shown by experiments on the COP3D and EPIC-FIELDS datasets [46, 50], showing the ability to handle slight deviation from metric-scale calibrations. By introducing such non-metric datasets in training, we force the model to reason from the relative relation of the input cameras and the input images. However, the model would fail to predict coherent results when there exists an order-of-magnitude scale error. + +# C.4 Additional Discussions on Limitations + +Due to the ill-posed nature of monocular reconstruction (e.g., limited viewpoint coverage and low frame rates), our method, like other monocular approaches, can still exhibit some blurriness and artifacts, especially during sudden or very fast movements and when visualizing reconstructions from challenging viewpoints. These issues are largely inherent to current monocular reconstruction paradigms. Notably, however, as also pointed out by reviewers, our approach already surpasses prior state-of-the-art baselines in terms of sharpness and artifact reduction, and demonstrates results comparable to optimization-based methods. Further improvements, such as scaling up the training datasets and introducing additional expert or supervisory signals to the 4DGT model, are promising directions for alleviating these remaining limitations. + +# D Social Impact + +As an early-stage research on 4D reconstruction, we do not foresee any immediate social impact from this work. However, it's worth noting that such a feed-forward pipeline could be used to synthesize more convincing fake videos by introducing novel views. \ No newline at end of file diff --git a/NeurIPS/2025/4DGT_ Learning a 4D Gaussian Transformer Using Real-World Monocular Videos/images.zip b/NeurIPS/2025/4DGT_ Learning a 4D Gaussian Transformer Using Real-World Monocular Videos/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..73507ace45fb4bafa6b37b3a8b16ce6a83d16e88 --- /dev/null +++ b/NeurIPS/2025/4DGT_ Learning a 4D Gaussian Transformer Using Real-World Monocular Videos/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:67ec0450fa17356645128c49402e2c7c6f9308f3dca7cabc23523b0ed39e8db9 +size 967019 diff --git a/NeurIPS/2025/4DGT_ Learning a 4D Gaussian Transformer Using Real-World Monocular Videos/layout.json b/NeurIPS/2025/4DGT_ Learning a 4D Gaussian Transformer Using Real-World Monocular Videos/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..08aca9ccb451fbfb7d130c5b42eb12ade9016c33 --- /dev/null +++ b/NeurIPS/2025/4DGT_ Learning a 4D Gaussian Transformer Using Real-World Monocular Videos/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f7ffa5df83780a9a91b112a0b2f8f08209ebfca6a8199dafad461aa6f060d533 +size 908822 diff --git a/NeurIPS/2025/4KAgent_ Agentic Any Image to 4K Super-Resolution/9a1cd916-3b96-4f2d-b210-f802fa3e6e47_content_list.json b/NeurIPS/2025/4KAgent_ Agentic Any Image to 4K Super-Resolution/9a1cd916-3b96-4f2d-b210-f802fa3e6e47_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..9d115299331d120270a0c22c75978f0755c87b2e --- /dev/null +++ b/NeurIPS/2025/4KAgent_ Agentic Any Image to 4K Super-Resolution/9a1cd916-3b96-4f2d-b210-f802fa3e6e47_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:72bc96075335fbd48767611aa7d49fbf9f9d5fb81a3869a525ad69a7b4f77c6f +size 544155 diff --git a/NeurIPS/2025/4KAgent_ Agentic Any Image to 4K Super-Resolution/9a1cd916-3b96-4f2d-b210-f802fa3e6e47_model.json b/NeurIPS/2025/4KAgent_ Agentic Any Image to 4K Super-Resolution/9a1cd916-3b96-4f2d-b210-f802fa3e6e47_model.json new file mode 100644 index 0000000000000000000000000000000000000000..598324a406ea95db6fb526bac4fd5e94d14bcd88 --- /dev/null +++ b/NeurIPS/2025/4KAgent_ Agentic Any Image to 4K Super-Resolution/9a1cd916-3b96-4f2d-b210-f802fa3e6e47_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:09684a32464abfa6397134dd6b3f5028c2d5347492bfcd52e2dd651f71cf81bc +size 711426 diff --git a/NeurIPS/2025/4KAgent_ Agentic Any Image to 4K Super-Resolution/9a1cd916-3b96-4f2d-b210-f802fa3e6e47_origin.pdf b/NeurIPS/2025/4KAgent_ Agentic Any Image to 4K Super-Resolution/9a1cd916-3b96-4f2d-b210-f802fa3e6e47_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..b3900ae1fd2882de1c7137227e0f95a5b404c570 --- /dev/null +++ b/NeurIPS/2025/4KAgent_ Agentic Any Image to 4K Super-Resolution/9a1cd916-3b96-4f2d-b210-f802fa3e6e47_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d21400548964c091e4c7541d1e66ea4d77b4fe56fdb12ad1978ceb3d049b196f +size 38561794 diff --git a/NeurIPS/2025/4KAgent_ Agentic Any Image to 4K Super-Resolution/full.md b/NeurIPS/2025/4KAgent_ Agentic Any Image to 4K Super-Resolution/full.md new file mode 100644 index 0000000000000000000000000000000000000000..c9586232ec21c26ae4e92db27a5ba903d23a183e --- /dev/null +++ b/NeurIPS/2025/4KAgent_ Agentic Any Image to 4K Super-Resolution/full.md @@ -0,0 +1,2364 @@ +Yushen Zuo $^{1}$ , Qi Zheng $^{1\dagger}$ , Mingyang Wu $^{1\dagger}$ , Xinrui Jiang $^{2\dagger}$ , Renjie Li $^{1}$ , Jian Wang $^{3}$ , Yide Zhang $^{4}$ , Gengchen Mai $^{5}$ , Lihong V. Wang $^{6}$ , James Zou $^{2}$ , Xiaoyu Wang $^{7}$ , Ming-Hsuan Yang $^{8}$ , Zhengzhong Tu $^{1\star}$ + +$^{1}$ Texas A&M University $^{2}$ Stanford University $^{3}$ Snap Inc. $^{4}$ CU Boulder $^{5}$ UT Austin $^{6}$ California Institute of Technology $^{7}$ Topaz Labs $^{8}$ UC Merced $^{*}$ Corresponding Author: tzz@amu.edu. $\dagger$ Equal contributions. + +Project Website: 4kagent.github.io + +![](images/dee99e0388fe89a3e625dac3ede1fcd9b4e1124a3bd9b859410ed5311074d581.jpg) + +![](images/5ec95a88c52c61c1b860832f606ad8a898279c425266a2d067ef9b709352840f.jpg) + +![](images/0550005f31dc4403ce5f59e1de6792afa5076e82e5d432c93e3d7a69db1211f3.jpg) +Figure 1: We present 4KAgent, an agentic image super-resolution generalist designed to universally upscale any image to 4K, regardless of input type, degradation level, or domain. That is, 4KAgent effectively restores diverse imagery, spanning from natural scenes, severely degraded captures (e.g., old photos), human/pet portraits, AI-generated content (AIGC), as well as specialized scientific imaging domains, such as remote sensing, fluorescence microscopy, pathology, and various medical modalities like X-ray, ultrasound, and funduscopy—all without the need for any re-training or domain-specific adaptation. + +# Abstract + +We present 4KAgent, a unified agentic super-resolution generalist system designed to universally upscale any image to 4K resolution (and even higher, if applied iteratively). Our system can transform images from extremely low resolutions with severe degradations, for example, highly distorted inputs at $256 \times 256$ , into crystal-clear, photorealistic 4K outputs. 4KAgent comprises three core components: (1) Profiling, a module that customizes the 4KAgent pipeline based on bespoke use cases; (2) A Perception Agent, which leverages vision-language models alongside image quality assessment experts to analyze the input image and make a tailored restoration plan; and (3) A Restoration Agent, which executes the plan, following a recursive execution-reflection paradigm, guided by a quality-driven mixture-of-expert policy to select the optimal output for each step. Additionally, 4KAgent embeds a specialized face restoration pipeline, significantly enhancing facial details in portrait and selfie photos. We rigorously evaluate our 4KAgent across 11 distinct task categories encompassing a total of 26 diverse benchmarks, setting new state-of-the-art on a broad spectrum of imaging domains. Our evaluations cover natural images, portrait photos, AI-generated content, satellite imagery, fluorescence microscopy, and medical imaging like fundoscopy, ultrasound, and X-ray, demonstrating superior performance in terms of both perceptual (e.g., NIQE, MUSIQ) and fidelity (e.g., PSNR) metrics. By establishing a novel agentic paradigm for low-level vision tasks, we aim to catalyze broader interest and innovation within vision-centric autonomous agents across diverse research communities. We release all the code, models, and results at: https://4kagent.github.io. + +# 1 Introduction + +Image super-resolution (SR) is a fundamental task in computer vision that aims to reconstruct high-resolution (HR) images from their low-resolution (LR) counterparts [120, 20, 21, 54, 106, 140, 102, 101, 82, 98]. It serves as a bedrock for various low-level vision tasks [135, 63, 128, 93], including deblurring [90, 15], dehazing [31, 58], deraining [81, 41], and low-light enhancement [109, 29]. Beyond its classical role in computational photography and imaging, SR techniques significantly influence numerous domains, such as biomedical imaging [25, 83], remote sensing [85, 32, 51], surveillance [136], and embodied artificial intelligence applications [30, 84, 39]. + +Traditional SR methods [20, 102] typically assume known synthetic degradation during training, which limits their generalization to real-world captures that suffer from complex, heterogeneous, and unpredictable degradations [101]. Recent research has increasingly shifted to a more practical real-world super-resolution (RealSR) task [8, 113], which attempts to explicitly address diverse and unknown degradations in naturally captured photo- and video-graphs. RealSR requires models not only to handle multiple combined degradations effectively but also to exhibit strong adaptability and generalization across varied scenarios [125, 94]. Many effective solutions have been proposed to solve the RealSR problem, via simulating complex real-world degradations [134, 101], leveraging the powerful generative prior of pre-trained diffusion models [43, 113, 114, 89, 72], enabling robust restoration under unknown conditions. Inspired by the advanced planning and reasoning capabilities of large language models (LLMs) [111, 36, 124, 22], agentic restoration frameworks [9, 144] have emerged as an advanced tool that can adaptively handle multiple degradations through sequential planning and dynamic restoration strategies. + +Despite their successes in certain scenarios, existing performant generative approaches [113, 89] can only handle limited degradation ranges, e.g., up to $4 \times$ upscaling, failing to recover extremely low-quality images with highly complex and diverse degradations in the wild. Moreover, SR specialist models are known to generalize poorly to out-of-distribution domains [10], let alone when applied on a different scaling factor. This is mainly due to heavy reliance on supervised learning on synthetic image pairs that cannot fully capture the complex real-world image degradations, not to mention other domains, ranging from AI-generated imagery, scientific computing, to biomedical images. Last but not least, practically, users often demand highly specific workflows, e.g., either denoising, upscaling, or prioritizing high fidelity over perceptual quality, and hence a one-size-fits-all system that can flexibly adapt to satisfy diverse user needs and application scenarios is in pressing need. + +To fill this gap, we present 4KAgent, the first-of-its-kind agentic framework for generic, flexible, and interpretable super-resolution of any image to 4K. As illustrated in Fig. 1, 4KAgent is capable of upscaling any low resolution image (e.g., 0.065 megapixels) to $4\mathrm{K}\times 4\mathrm{K}$ , (i.e., 16.7 megapixels) by $16\times$ upscaling factors1 (\$3.4). It also sets new state-of-the-art (SoTA) on classical image superresolution (\$3.1), real-world image super-resolution (\$3.2), face restoration (Appendix), and multiple-degradation image restoration (\$3.3) benchmarks, in terms of perceptual quality. We also show that 4KAgent generalizes across widespread applications in low-level tasks, such as joint restoration & 4K upscaling (\$3.5), and AI-generated content 4K upscaling (\$3.6). Lastly, thanks to the mixture-of-experts and profile design, 4KAgent demonstrates broader impacts on interdisciplinary areas such as scientific super-resolution (\$3.7), including Ⓞ Satellite image super-resolution, Ⓞ fluorescence microscopy super-resolution, and Ⓞ medical image super-resolution. + +# Our contributions are as follows: + +- [Framework] We present 4KAgent, the first AI agent framework for universal any-image-to-4K upscaling, capable of handling all image categories, ranging from classical and realistic degradations, extreme low-quality inputs, AI-generated imagery, to scientific imaging tasks such as remote sensing, microscopy, and biomedical inputs. +- [System Design] We design a multi-agent system in 4KAgent, the Perception Agent employs large vision-language models (VLMs) to analyze the content and distortion in the image and provide the restoration plan for the restoration agent to execute. The Restoration Agent, which sets up an execution—reflection—rollback procedure for recursive restoration and upscaling. +- [Q-MoE & Face Restoration pipeline] In each restoration step of the restoration plan, we propose a Quality-Driven Mixture-of-Expert (Q-MoE) policy in execution and reflection to select the optimal image. We further develop a face restoration pipeline to enhance faces in images. +- [Profile Module] To expand the applicability of 4KAgent, we propose a Profile Module to bring the availability to customize the system for different restoration tasks. 4KAgent can adapt to different restoration tasks without extra training. +- [DIV4K-50 Dataset] To evaluate 4K super-resolution performances, we build the DIV4K-50 dataset as a challenging testset to upscale a low-quality (LQ) image in $256 \times 256$ resolution with multiple degradations to a high-quality (HQ) 4K image in $4096 \times 4096$ resolution. +- [Experiments] Extensive experimental results demonstrate the superiority of 4KAgent as a generalist 4K upscaling agent: 4KAgent sets new state-of-the-art on a variety of real-world image super-resolution benchmarks, multiple-degradation restoration benchmarks, face restoration, 4K upscaling task, and various scientific imaging tasks, including satellite image super-resolution, fluorescence microscopic imaging, X-ray radiography, and bio-medical imaging super-resolution. + +# 2 Method + +# 2.1 System Overview + +We introduce 4KAgent, a multi-agent framework designed to upscale any real-world image to 4K resolution. Fig. 2 illustrates the overall workflow of our proposed 4KAgent, which decomposes the restoration pipeline into a collection of specialized agents. The Perception Agent analyzes degradations (noise, blur, etc.), extracts semantic/structural cues, and schedules a restoration plan containing a sequence of operators (denoising, deblurring, super-resolution, etc.). The Restoration Agent follows the restoration plan using our proposed Quality-driven Mixture-of-Experts (Q-MoE) to pick the best output from multiple restoration tools. The Rollback mechanism will be activated if the quality of the restored image falls below a threshold. Additionally, a dedicated Face Restoration Pipeline further enhances facial regions by triggering expert face restoration models. A user-configurable Profile Module allows users to customize the system (e.g., prioritize fidelity or perceptual quality), enabling robust, high-quality 4K SR across diverse content and degradation types. + +# 2.2 Perception Agent + +The Perception Agent is designed as a four-stage analytical module that bridges low-level image quality assessment with high-level reasoning. Its core function is to extract a robust and holistic + +![](images/71c95be7f36d3edf259e07240027022f35403fabbfa87ed1aaaf0f0a7eeee8ed.jpg) +Figure 2: 4KAgent system overview. + +understanding of the input image in terms of both semantic content and low-level degradations, and to create a restoration plan that will guide the subsequent restoration process. + +Image Analyzer. Perception agent invokes a suite of expert Image Quality Assessment (IQA) tools that evaluate the input image $I$ across multiple quality dimensions $Q_{I} = (Q_{1}, Q_{2}, \ldots)$ . Specifically, we adopt CLIPIQA [97], TOPIQ [7], MUSIQ [45], and NIQE [137] as the IQA metrics. These metrics represent perceptual quality from diverse aspects (due to their different model designs and training data), which will be employed as context for the next step of degradation reasoning. + +Degradation Reasoning. Perception agent leverages a VLM $M_R$ to reason over the obtained IQA metrics. Specifically, by incorporating the input image $I$ , IQA metrics $Q_I$ , the VLM $M_R$ will predict the degradation list $D_I$ from the input image, which will correspond to an initial restoration agenda $A_I'$ . Meanwhile, $M_D$ also analyzes the content in the image and outputs the corresponding image descriptions $C_I$ (i.e., captioning). The whole process can be expressed as $C_I, D_I, A_I' = M_R(I, Q_I)$ . + +Upscaling Factor Configuration. 4KAgent is able to automatically determines and applies an appropriate super-resolution scale to reach 4K. Given an input image $I$ with height $H_{I}$ and width $W_{I}$ , the scale factor $s$ is calculated by $s = \min \left(\{s\in \{2,4,8,16\} \mid \max (H_I,W_I)*s\geq 4000\} \cup \{16\}\right)$ . After obtaining the initialize agenda $A_{I}'$ from $M_D$ , 4KAgent will calculate the scale factor $s$ and append super-resolution task(s) into $A_{I}'$ to obtain the final agenda $A_{I}$ . Under this setting, 4KAgent is able to upscale any image (resolution larger than $250\times 250$ ) to 4K resolution in a single process. + +Task Planning. After obtaining the degradation list $D_{I}$ present in the input image and the restoration agenda $A_{I}$ , the perception agent employs an LLM / VLM $M_P$ to provide the restoration plan. Specifically, by coprating image descriptions $C_I$ , degradation list $D_{I}$ , restoration experience $E$ , and input image itself $I$ (available when using VLM as $M_P$ ), $M_P$ outputs an initial restoration plan $P_{I} = M_{P}(C_{I},D_{I},A_{I},E,I)$ , which contains a sequence of restoration tasks. + +# 2.3 Restoration Agent + +Building upon the task plan $P_{I}$ provided by the Perception Agent, the Restoration Agent executes an iterative process, each stage of which tightly couples restoration and evaluation using an execution-reflection-rollback triplet. Within this agent, we propose a quality-driven mixture-of-experts (Q-MoE) policy, both in execution and reflection, to select the optimal image for each restoration step. We also employ a rollback mechanism to adjust the restoration plan if necessary. + +Execution. Guided by the task plan $P_{I}$ , this agent executes the restoration step by step. In each restoration step, the input image will go through all tools in the toolbox, which contains a number of + +advanced restoration models (detailed in Appendix) corresponding to each individual restoration task. In 4KAgent, we have curated 9 different restoration tasks that are useful to enhance picture quality: Brightening, Defocus Deblurring, Motion Deblurring, Dehazing, Denoising, Deraining, JPEG Compression Artifact Removal, Super Resolution, and Face Restoration. Specifically, for step $k$ in the restoration plan, it produces multiple restoration results $\{T_{i}(I_{k - 1}), i = 1\sim N\}$ ( $T_{i}$ is $i$ -th tool in the toolbox, $N$ is the number of tools in the toolbox) based on the input image $I_{k - 1}$ . + +Reflection. After obtaining restoration results $\{T_{i}(I_{k - 1}), i = 1 \sim N\}$ , the Restoration Agent will select the optimal image based on their quality. To evaluate the quality of image $T_{i}(I_{k - 1})$ , we compute the image quality score by combining the preference model HPSv2 [115] and no-reference IQA metrics. Specifically, we use HPSv2 to assess the human preference of the resulting image $T_{i}(I_{k - 1})$ based on the image content description $C_I$ . For no-reference IQA metrics, we employ NIQE [74], MANIQA [121], MUSIQ [45], and CLIPIQA [97] to calculate a weighted sum as its no-reference quality score: $Q_{s}(T_{i}(I_{k - 1})) = \mathrm{H}(T_{i}(I_{k - 1}), C_I) + Q_{nr}(T_{i}(I_{k - 1})) / 4$ , where $Q_{nr}(T_{i}(I_{k - 1})) = w_{\mathrm{NIQE}} * (1 - Q_{\mathrm{NIQE}} / 10) + \sum_{j \in \Omega} w_{j} * Q_{j}$ , $\Omega = \{\mathrm{MUSIQ}, \mathrm{MANIQE}, \mathrm{CLIPIQA}\}$ , $\mathrm{H}$ indicates the HPSv2 evaluation. After obtaining the quality score of each result image, the final result of this restoration step is obtained by the highest quality score: $I_{k} = \arg \max_{Q_{s}}(T_{1}(I_{k - 1}), T_{2}(I_{k - 1}), \dots, T_{N}(I_{k - 1}))$ . The combination of execution and reflection can be viewed as a Mixture-of-Expert (MoE) system, which we refer to as a quality-driven mixture-of-expert (Q-MoE): the input image is processed through each expert (execution), and the Reflection function selects the optimal image among all. + +Rollback. Following previous AI Agent systems [79, 144, 141, 61, 34], we also design a rollback mechanism in the 4KAgent system. Specifically, if the quality score of $I_{k}$ is lower than a threshold $\eta$ , i.e., $Q_{s}(I_{k}) \leq \eta$ , the restoration step will be seen as a failure step and 4KAgent will generate a failure message $S_{I}$ . Then the system will rollback to the input image $I_{k-1}$ and assign a different restoration task in this step. (Detailed process is shown in the Appendix) + +# 2.4 Face Restoration Pipeline + +Human face regions are often the most visually sensitive and semantically important components in an image. However, conventional super-resolution methods struggle to maintain identity consistency, natural skin textures, and perceptual quality when restoring faces, especially in heavily degraded portraits. To address this, 4KAgent incorporates a dedicated Face Restoration pipeline, which is selectively triggered within the restoration workflow. The Face Restoration pipeline is embedded as a submodule in 4KAgent and will only be invoked after the super-resolution restoration step, ensuring that face quality refinement is seamlessly integrated into the iterative restoration loop. + +The overall framework of the face restoration pipeline in the 4KAgent system is shown in Fig. 3. First, 4KAgent will detect and crop faces in the input image $\{F_I^l,l = 1\sim L\}$ ( $L$ is the number of faces in the image $I$ ). Then, if super-resolution is in the restoration plan and the resulting image $I_{k}$ of super-resolution step does not trigger the rollback mechanism, 4KAgent will detect and crop faces in the resulting image $\{F_{I_k}^l,l = 1\sim L'\}$ ( $L'$ is the number of faces in the image $I_{k}$ ). If $L = L'$ , then for each face in $I_{k}$ , different advanced face restoration + +![](images/9520755e53aff74a942950f5b0638423b4ae7e4615e4a69c10866588fda44512.jpg) +Figure 3: Face restoration pipeline overview. + +methods are applied, yielding restored faces $\{T_i^f (F_{I_k}^l),i = 0\sim N^f\}$ . Here, $T_{i}^{f}$ is a face restoration tool in the toolbox, $T_0^f$ is an identical function, and $N^{f}$ is the number of face restoration tools in the toolbox. Likewise, we also conduct Q-MoE policy here: 4KAgent selects the best face based on the quality score $Q_{s}^{f}$ . The quality score $Q_{s}^{f}$ not only considers face quality, but also identity preservation: + +$$ +Q _ {s} ^ {f} \left(T _ {i} ^ {f} \left(F _ {I _ {k}} ^ {l}\right)\right) = w _ {\mathrm {I P}} * \mathrm {I P} \left(T _ {i} ^ {f} \left(F _ {I _ {k}} ^ {l}\right), F _ {I} ^ {l}\right) + w _ {\mathrm {I Q A}} * \left(Q _ {n r} \left(T _ {i} ^ {f} \left(F _ {I _ {k}} ^ {l}\right)\right) / 4 + Q _ {\mathrm {C F}} \left(T _ {i} ^ {f} \left(F _ {I _ {k}} ^ {l}\right)\right)\right), \tag {1} +$$ + +where $l = 1 \sim L$ . IP calculates the cosine similarity of face features, extracted using ArcFace [19]. CF indicates CLIB-FIQA [76], which is an advanced face IQA metric. 4KAgent combines the no-reference quality score used in the reflection stage and the CLIB-FIQA score to assess the face quality. After obtaining quality score $Q_{s}^{f}$ , 4KAgent selects the best face $F_{out}^{l}$ : $F_{out}^{l} =$ + +$\arg \max_{Q_s^f}(T_0^f (F_{I_k}^l),T_1^f (F_{I_k}^l),\dots,T_{N_f}^f (F_{I_k}^l))$ 4KAgent will paste $F_{out}^{l}$ back to the original image $I_{k}$ , then proceeding to the next step. + +# 2.5 Profile Module + +To enhance the flexibility and applicability of our 4KAgent system, we develop the Profile Module, enabling dynamic customization for diverse image restoration scenarios, per user's needs. Specifically, the Profile Module acts like a system prompt for LLM applications, allowing fine-grained control through the following seven configuration parameters: + +1. Perception Agent: Specifies the choice of LLM / VLM employed by the Perception Agent. [Default: Llama-vision] +2. Upscale to 4K: Determines whether to upscale to 4K resolution. [Default: True] +3. Scale Factor: Explicitly defines the upscale factor for the entire pipeline. (Default: 4, Options: [2, 4, 8, 16]). This parameter overrides "Upscale to 4K" when specified. +4. Restore Option: Explicitly sets the restoration task(s) to be applied. If set to None, restoration task(s) are determined automatically by the Perception Agent. (Default: None) +5. Face Restore: Toggles activation of the dedicated face restore pipeline. (Default: True) +6. Brightening: Controls the activation of image brightening, which may cause color shifts in restored images. Provided as [Optional] to maintain image color fidelity. (Default: False) +7. Restore Preference: Defines whether to prioritize higher perceptual quality or higher fidelity in image restoration. (Options: [Perception, Fidelity], Default: Perception). Here we respect the perception-distortion tradeoff [4, 142], deeming models that optimize for distortion metrics (e.g., PSNR, SSIM [107]) as Fidelity models while methods trained for perceptual quality (e.g., NIQE [74], MUSIQ [45]) as Perception models. + +The Profile Module offers exceptional configurability, enabling seamless adaptation to a wide range of restoration tasks without requiring model retraining or domain-specific fine-tuning. To the best of our knowledge, 4KAgent is a first-of-its-kind framework that enjoys unprecedented robustness and generalizability: each distinct restoration scenario can be addressed by simply selecting an appropriate configuration profile, thanks to which 4KAgent consistently achieves excellent performance across a variety of challenging restoration domains. Comprehensive details on predefined profiles in 4KAgent and their naming conventions are further elaborated in the Appendix. + +# 3 Experiments + +We evaluate 4KAgent on a wide range of 11 image super-resolution tasks, including classical image SR $(4\times)$ (§3.1), real-world image SR $(4\times)$ (§3.2), multiple-degradation image restoration (§3.3), face restoration $(4\times)$ (Appendix), large-scale factor SR $(16\times)$ (§3.4), joint restoration with 4K upscaling (§3.5), and AI-Generated Content 4K SR (§3.6), remote sensing image SR (Appendix), fluorescence microscopy image SR (Appendix), and medical image SR (Appendix). Totally, we test 4KAgent on 26 benchmarks. The summary of datasets used in experiments is shown in the Appendix. We also present the parameter setting, the details of the toolbox, and prompts used in LLM / VLM component of 4KAgent in the Appendix. + +# 3.1 Classical Image Super-Resolution + +In this section, we follow the classical SR experiment setting [139, 63], evaluating on the classic SR datasets, including Set5 [3], Set14 [130], B100 [70], Urban100 [35], and Manga109 [71]. Besides PSNR and SSIM [107], we also evaluate images on LPIPS [138], FID [33], NIQE [74], and MUSIQ [45] for more comprehensive evaluation. + +Thanks to the high flexibility of our 4KAgent, which is governed by configurable profiles, we use 3 different pro + +Table 1: Quantitative comparisons in classical image SR benchmark (B100). The best and second-best performances are marked in bold and underline. + +
MethodPSNR↑SSIM↑LPIPS↓FID↓NIQE↓MUSIQ↑
SwinIR [63]27.920.74890.354894.576.2757.71
X-Restormer [13]27.990.75080.352190.526.2157.91
HAT-L [14]28.080.75470.344089.526.2058.71
DiffBIR [65]24.990.61560.271984.993.9268.23
OSEDiff [113]24.350.64950.240873.234.0868.54
AgenticIR [144]22.510.58530.3078102.924.0868.36
4KAgent (ExpSR-s4-F)28.090.75400.345388.896.0259.12
4KAgent (ExpSR-s4-P)24.640.62940.238773.643.8669.42
4KAgent (GenSR-s4-P)23.640.62460.257278.803.9369.44
+ +files to customize the 4KAgent in this experiment: ExpSR-s4-F, ExpSR-s4-P, and GenSR-s4-P. For comparison, we use state-of-the-art fidelity-based methods (e.g., SwinIR [63], X-Restormer [13], HAT [14]) and perception-based methods (e.g., DiffBIR [65], OSEDiff [113]). We also include AgenticIR [144] for agentic system-level comparison. We present experimental results on the B100 dataset in Tab. 1. Detailed experimental results on these datasets and visual comparisons are shown in the Appendix. Based on quantitative comparisons, 4KAgent outperforms AgenticIR in almost every metric on classic SR benchmarks. Moreover, 4KAgent can easily output images with high perception quality or high fidelity by setting different profiles. + +# 3.2 Real-World Image Super-Resolution + +In this section, we follow previous RealISR methods and use real-world image super-resolution datasets (RealSR [5], DrealSR [112]) for evaluation. We use 2 different profiles to customize 4KAgent in this experiment: ExpSR-s4-P and GenSR-s4-P. We compare 4KAgent with state-of-the-art methods, including ResShift [127], StableSR [98], DiffBIR [65], PASD [122], SeeSR [114], SinSR [104], and OSEDiff [113]. We also employ AgenticIR in this experiment for agentic system comparison. We use the + +Table 2: Quantitative comparison in Realistic image superresolution benchmark (RealSR). The best and second best performances are marked in bold and underline. + +
MethodPSNR↑SSIM↑LPIPS↓FID↓NIQE↓MUSIQ↑
ResShift [127]26.310.74110.3489142.817.2758.10
StableSR [98]24.690.70520.3091127.205.7665.42
DiffBIR [65]24.880.66730.3567124.565.6364.66
PASD [122]25.220.68090.3392123.085.1868.74
SeeSR [114]25.330.72730.2985125.665.3869.37
SinSR [104]26.300.73540.3212137.056.3160.41
OSEDiff [113]25.150.73410.2921123.505.6569.09
AgenticIR [144]22.450.64470.3745140.385.8165.87
4KAgent (ExpSR-s4-P)24.600.68390.3253127.645.0970.97
4KAgent (GenSR-s4-P)22.550.65570.3509134.634.7871.77
+ +same metrics to evaluate result images. Experimental results on RealSR dataset are shown in Tab. 2. 4KAgent outperforms AgenticIR in every metric regardless of profile setting. Moreover, 4KAgent sets a new state-of-the-art performance on no-reference metrics (e.g., NIQE, MUSIQ). Detailed experimental results on both datasets and visual comparisons are shown in the Appendix. + +# 3.3 Multiple-Degradataion Image Restoration + +In this section, we follow the setting of AgenticIR, using the Group A, B, and C test set which contains 1,440 LQ images processed with 16 combinations of degradations applied to images from MiO100 dataset [50]. In this experiment, we configure 4KAgent with GenMIR-P profile. We compare 4KAgent with several all-in-one models: AirNet [59], PromptIR [80], MiOIR [49], DA-CLIP [68], InstructIR [16], AutoDIR [43], and agentic systems: + +Table 3: Quantitative comparison of multiple image restoration tasks on Group C subset. The best and second-best performances are marked in bold and underline. + +
MethodPSNR↑SSIM↑LPIPS↓MANIQA↑CLIPIQA↑MUSIQ↑
AirNet [59]17.950.51450.57820.18540.311330.12
PromptIR [80]18.510.51660.57560.19060.310429.71
MiOIR [49]15.630.48960.53760.17170.289137.95
DA-CLIP [68]18.530.53200.53350.19160.347633.87
InstructIR [16]17.090.51350.55820.17320.253733.69
AutoDIR [43]18.610.54430.50190.20450.293937.86
AgenticIR [144]18.820.54740.44930.26980.394848.68
MAIR [42]19.420.55440.41420.27980.423951.36
4KAgent (GenMIR-P)19.770.56290.42710.35450.523355.56
+ +AgenticIR [144] and MAIR [42]. Experiment results are shown in Tab. 3 and Fig. 4. + +![](images/86b14af936d8304d70cbc50143a921b3cca678e2e710989fe16f3f683bf207f3.jpg) +Figure 4: Visual comparisons on MiO100 dataset. (Please zoom in to see details.) + +![](images/1f8ab4746fcc58b426bf491ebe1efed67bcefbc75c255b01cd69c250146d8631.jpg) + +![](images/3378d34760660d84e6e380439dc875ed7680a9a6f3671e179f6b46207af3b58a.jpg) + +![](images/4e1db5734f485636e7cda6299467acdaab2f5d23cfdb482d0413f7fdb961a6d4.jpg) + +![](images/0fd24f29bf07c9aa40ac4dcd9d8854e93813a649f0a93cda6514a0fbdd1cac00.jpg) + +![](images/bde66e7ed8a9ea9c66d2e9961a07e4d56d6508cf18764b444e9087ff84d69420.jpg) + +We present experimental results on Group C here. Experimental results on all groups are shown in the Appendix. As summarized in Tab. 3, 4KAgent achieves state-of-the-art performance on all metrics excluding LPIPS. As shown in Fig. 4, 4KAgent generates images with high-grained details and more + +consistent with the HQ image (e.g., the hand and bottle on the left and the skin of the lizard on the right). These results demonstrate the superiority of 4KAgent under complex distortion. + +# 3.4 Large Scale-Factor $(16\times)$ Image Super-Resolution + +In this section, we target a challenging setting, $16\times$ real-world image upscaling. For the dataset used in this experiment, we select RealSRSet [134], a 20-image real-world low-quality image dataset for a large-scale super-resolution experiment. Specifically, we configure 4KAgent with the Gen4K-P profile for this experiment. Based on the resolution of input images in the dataset, 4KAgent will upscale each image with a scale factor of 16. + +We compare 4KAgent against HAT-L [14], DiffBIR [65] under different settings: (1) $4 \times \rightarrow 4 \times$ : first upscale the low-quality image by a scale factor of 4, then upscale the upscaled images to obtain the $16 \times$ upscaled image. (2) $16 \times$ : directly upscale the low-quality image by a scale factor of 16. We also extend AgenticIR for large scale-factor $(16 \times)$ image super-resolution for agentic system comparison. Visual comparisons of different methods are shown in Fig. 5. 4KAgent generates high-grained and realistic details which is more visually pleasant than comparing methods (e.g., the rock and grass textures). Quantitative results and more visual comparisons are shown in the Appendix. + +![](images/94548e6d3ee25bcc0f9c9f8bd62a7b2ad26c13f75f491012fe1a506b78b3bcfe.jpg) +Figure 5: Visual comparisons on RealSRSet dataset. (Please zoom in to see details) + +![](images/0f969d74f64b10a6b232910ddab8333b9616fc7fd2c0be15e96912ee50a27fad.jpg) +LQ + +![](images/1ea42d7963802028a3b464cdfe50479cf0b381aa0855fafc30705bac2bc81470.jpg) +HAT-L $(4\times \rightarrow 4\times)$ + +![](images/8042eb2ec9ebac3d5274ea5cfd9c59c26f8592da04d88672cd3dbbfede493ab5.jpg) +DiffBIR $(16\times)$ + +![](images/a8c46864d67370ed4b6f59f406101e1f25b2b6317ecb537f2433a45524453b47.jpg) +AgenticIR + +![](images/0e444bae28e578afc59a15be17336b28f85984d040b7a0aa2eee3a0a56d039ce.jpg) +4KAgent (Gen4K-P) + +# 3.5 Joint Restoration & 4K Upscaling + +In this section, we bring 4KAgent to the most challenging setting: Joint multiple image restoration and 4K upscaling. As there are no previous methods and datasets targeted at this setting, we construct a new evaluation dataset, DIV4K-50, constructed from the Aesthetic-4K dataset [132] to rigorously test end-to-end restoration and ultra-high-scale SR. Specifically, we select 50 images from the Aesthetic-4K dataset based on its content, then center-crop each image to $4096 \times 4096$ as the high-quality (HQ) ground truth. Then we downsample HQ images to $256 \times 256$ and randomly apply combinations of defocus blur, motion blur, additive Gaussian noise, and JPEG compression to generate the corresponding low-quality (LQ) images. In this experiment, we also configure 4KAgent with the Gen4K-P profile. We compare 4KAgent with more methods in this experiment (OSEDiff [113], PiSA-SR [89]) with more upscaling settings. + +As shown in Fig. 6, 4KAgent reconstructs finer and more natural details than comparing methods (e.g., the facial features). Quantitative results and more visual comparisons are shown in the Appendix. + +![](images/084e460393fd5a01d3555706ab159786473c191fb1ba53f7193a09b74380ad1a.jpg) + +![](images/9655111de4c90a03754abd3484aa3164feed079d600d38d1ab75fdaadc685aca.jpg) +LQ + +![](images/08871122c40aabe4de574d899504abf734bec5e77d315dff87e627408d740345.jpg) +HAT-L $(4\times \rightarrow 4\times)$ + +![](images/b8602a8c9a2e5680f1fa903c0fa60199d61284d12c46b69466108acd5ff31f35.jpg) +DiffBIR $(4\times \rightarrow 4\times)$ + +![](images/402619784e320305348e5dbc541e3d3fb31145db246d8a2dedc40355774f8925.jpg) +OSEDiff $(4\times \rightarrow 4\times)$ + +![](images/1b7d55182cfd2760b89efec61f1c4b836992152a75acc512810f075b1cb36d07.jpg) +PiSA-SR $(4\times \rightarrow 4\times)$ + +![](images/9bcd62020d939517061a185c3323d433b4a44d54913f82af2b7259cba3889db3.jpg) +AgenticIR +Figure 6: Visual comparisons on DIV4K-50 dataset (Please zoom in to see details). + +![](images/ff342ff4e05050e14ebd082ffcfa04caf22e8c479b1b7d320b247204d0405200.jpg) +4KAgent (Gen4K-P) + +![](images/6eab9ed64498f6750fcd1100b527c2008111a5b5938fe015acc24a51086fbc88.jpg) +HQ + +# 3.6 AI-Generated Content (AIGC) 4K Super-Resolution + +In this section, we investigate super-resolution in AIGC scenarios by comparing direct 4K generation with 1K upscaling using our 4KAgent. As no prior method and dataset targets this setting, we sample 200 prompts from two standard AIGC benchmarks [108, 55] and generate both 1K and native 4K images using several generative models [53, 23, 11, 119, 37, 132], which we refer to as GenAIBench-4K and DiffusionDB-4K. We use the ExpSR-s4-P profile in 4KAgent in this experiment. To better capture local degradations in 4K images, we introduce MUSIQ-P, a patch-based metric that averages MUSIQ scores over non-overlapping $512 \times 512$ regions. + +Table 4: Quantitative comparison of AIGC $4 \times$ SR in GenAIBench-4K [55]. Bold denotes top performers; underlines indicate second and third. MUSIQ- $\mathbf{P}^*$ is a patch-applied MUSIQ metric for evaluating 4K images. + +
ModelNIQE↓MANIQA↑MUSIQ-P*↑CLIPIQA↑
SANA-1K [119]4.180.481466.300.7147
+ 4KAgent3.030.473557.970.7050
GPT4o [37]5.690.499764.430.6607
+ 4KAgent3.560.497658.280.7016
FLUX.1-dev [53]6.180.501861.020.6768
+ 4KAgent2.980.503458.190.7078
PixArt-Σ [11]4.120.441563.740.6960
+ 4KAgent2.760.469956.710.7077
SD3-Medium [23]5.030.476764.680.6922
+ 4KAgent2.990.515560.220.7169
+ +Fig. 7 presents models enhanced by 4KAgent produce finer-grained details compared to native 4K generation methods [132, 119]. As shown in Tab. 4, 4KAgent consistently boosts both perceptual quality and semantic alignment on GenAIBench-4K, surpassing their original 1K methods. In addition, we observe that 4KAgent demonstrates stronger alignment with human preferences compared to native 4K generation methods, as evidenced by higher PickScore [48]. Additional comparisons with native 4K methods and results on DiffusionDB-4K are presented in the Appendix. + +![](images/ff591549af72072429dd70af06e718839c0513a73af3cf27f2211be1cc2db0b7.jpg) +Figure 7: Qualitative comparison between native 4K image generation and 1K image generation methods with 4KAgent, using identical prompts. + +# 3.7 Scientific Images + +In this section, we extend the evaluation of 4KAgent across a broad spectrum of cross-domain scientific image super-resolution applications, where high spatial fidelity is crucial but often limited by sensor cost and physical constraints [91, 92, 95]. The explored imaging domains and corresponding benchmark datasets are as follows: For remote sensing image super-resolution, we evaluate on four benchmark datasets covering varied land-use patterns and sensing characteristics, including AID [118], DIOR [60], DOTA [117], and WorldStrat [17]. For fluorescence microscopy super-resolution, we compare the performance of 4KAgent with a set of representative deep learning-based single-image SR (SISR) methods on SR-CACO-2 dataset [2]. For biomedical super-resolution, datasets from 4 distinct modalities are considered: bcSR [40] in pathology microscopy, Chest X-ray 2017 [46] and Chest X-ray 14 [100] in X-ray, US-Case [86] and MMUS1K [75] in ultrasound, and DRIVE retinal vessel dataset [87] in fundoscopy. + +Visual comparisons are shown in Fig. 8, showcasing six imaging modalities side by side: remote sensing, fundoscopy, confocal fluorescence microscopy, pathology, X-ray, and ultrasound from left to + +![](images/72c04c7f53e30121ab7b9c465ccfeae1fd46616794da693764e0fb57fdea6d54.jpg) +Figure 8: 4KAgent in processing scientific images. (Low-quality input vs. 4KAgent result) + +right. The top and bottom rows display the LQ inputs and 4KAgent outputs, respectively. Detailed quantitative and qualitative results are presented in the Appendix. + +# 4 Related Works + +We briefly review related works in three key areas. In image super-resolution, early CNN-based methods like SRCNN [20] laid the foundation, followed by architectural innovations such as residual and dense connections [47, 64, 140], attention mechanisms [6, 18, 139], and generative models including GAN-based methods [54, 101, 134, 8, 62, 63, 57] and diffusion-based approaches [98, 122, 65, 114, 82, 104, 113]. In image restoration, progress has been driven by residual learning [133, 67, 135], attention modules [126, 12, 93, 27, 129], transformers [96, 145, 128, 105], and both GAN [28, 1, 26, 38, 73, 77] and diffusion models [65, 116, 43, 103, 44, 24], with recent efforts addressing multi-degradation scenarios via unified models [59, 69, 78, 123, 131]. In LLM agents, foundational reasoning frameworks like CoT [110], ReAct [124], and CoALA [88] have inspired domain-specific systems such as MMCTAgent [52], VideoAgent [99], and MedCoT [66]. Notably, RestoreAgent [9], AgenticIR [144], MAIR [42], HybridAgent [56], and Q-Agent [143] demonstrate how LLMs can orchestrate multi-step visual restoration workflows via perception, planning, and quality-aware decision-making. For a full discussion, please refer to the Appendix. + +# 5 Concluding Remarks + +In this paper, we introduced 4KAgent, a versatile agentic image super-resolution generalist model designed to universally upscale images of diverse types and degradation levels to 4K resolution. By leveraging advanced multi-expert integration, adaptive decision-making, and specialized tools for perception and fidelity optimization, 4KAgent significantly enhances restoration quality across various challenging domains, including severely degraded images, natural scenes, portraits, AI-generated content, and specialized scientific modalities such as remote sensing, microscopy, and medical imaging. Our extensive evaluations on standard benchmarks and specialized datasets confirm that 4KAgent consistently outperforms existing state-of-the-art approaches, especially in complex scenarios where conventional super-resolution methods fall short. This robust performance, achieved without domain-specific retraining, demonstrates the models unique generalizability and practical utility for generic deployment in both consumer, commercial, and scientific applications. + +Future Work Looking ahead, we have identified several promising directions that can further enhance the capabilities and applicability of 4KAgent, enabling broader use cases. First, we will optimize the efficiency of 4KAgent by designing more accurate distortion-perception models and refining the execution-reflection-rollback mechanism to achieve faster and higher-quality image restoration. Second, we will prioritize enhancing the safety and robustness of 4KAgent to mitigate risks such as privacy breaches and the generation of harmful imagery. Lastly, we will continuously expand the toolbox of 4KAgent by integrating additional domain-specific restoration methods and developing targeted restoration profiles to improve performance and user experience. + +Acknowledgements Gengchen Mai is supported by the NSF under Grant No. 2521631. + +# References + +[1] David Bau, Hendrik Strobelt, William Peebles, Jonas Wulff, Bolei Zhou, Jun-Yan Zhu, and Antonio Torralba. Semantic photo manipulation with a generative image prior. arXiv preprint arXiv:2005.07727, 2020. 10 +[2] Soufiane Belharbi, Mara Whitford, Phuong Hoang, Shakeeb Murtaza, Luke McCaffrey, and Eric Granger. Sr-caco-2: A dataset for confocal fluorescence microscopy image super-resolution. Advances in Neural Information Processing Systems, 37:59948-59983, 2024. 9 +[3] Marco Bevilacqua, Aline Roumy, Christine Guillemot, and Marie Line Alberi-Morel. Low-complexity single-image super-resolution based on nonnegative neighbor embedding. 2012. 6 +[4] Yochai Blau and Tomer Michaeli. The perception-distortion tradeoff. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6228-6237, 2018. 6 +[5] Jianrui Cai, Hui Zeng, Hongwei Yong, Zisheng Cao, and Lei Zhang. Toward real-world single image super-resolution: A new benchmark and a new model. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3086-3095, 2019. 7 +[6] Chaofeng Chen, Dihong Gong, Hao Wang, Zhifeng Li, and Kwan-Yee K Wong. Learning spatial attention for face super-resolution. IEEE Transactions on Image Processing, 30:1219-1231, 2020. 10 +[7] Chaofeng Chen, Jiadi Mo, Jingwen Hou, Haoning Wu, Liang Liao, Wenxiu Sun, Qiong Yan, and Weisi Lin. Topiq: A top-down approach from semantics to distortions for image quality assessment. IEEE Transactions on Image Processing, 2024. 4 +[8] Chaofeng Chen, Xinyu Shi, Yipeng Qin, Xiaoming Li, Xiaoguang Han, Tao Yang, and Shihui Guo. Realworld blind super-resolution via feature matching with implicit high-resolution priors. In Proceedings of the 30th ACM International Conference on Multimedia, pages 1329-1338, 2022. 2, 10 +[9] Haoyu Chen, Wenbo Li, Jinjin Gu, Jingjing Ren, Sixiang Chen, Tian Ye, Renjing Pei, Kaiwen Zhou, Fenglong Song, and Lei Zhu. Restoreagent: Autonomous image restoration agent via multimodal large language models. arXiv preprint arXiv:2407.18035, 2024. 2, 10 +[10] Haoyu Chen, Wenbo Li, Jinjin Gu, Jingjing Ren, Haoze Sun, Xueyi Zou, Zhensong Zhang, Youliang Yan, and Lei Zhu. Low-res leads the way: Improving generalization for super-resolution by self-supervised learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25857-25867, 2024. 2 +[11] Junsong Chen, Chongjian Ge, Enze Xie, Yue Wu, Lewei Yao, Xiaozhe Ren, Zhongdao Wang, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart-σ: Weak-to-strong training of diffusion transformer for 4k text-to-image generation. In European Conference on Computer Vision, pages 74–91. Springer, 2024. 9 +[12] Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun. Simple baselines for image restoration. In European conference on computer vision, pages 17-33. Springer, 2022. 10 +[13] Xiangyu Chen, Zheyuan Li, Yuandong Pu, Yihao Liu, Jiantao Zhou, Yu Qiao, and Chao Dong. A comparative study of image restoration networks for general backbone network design. In European Conference on Computer Vision, pages 74-91. Springer, 2024. 6, 7 +[14] Xiangyu Chen, Xintao Wang, Wenlong Zhang, Xiangtao Kong, Yu Qiao, Jiantao Zhou, and Chao Dong. Hat: Hybrid attention transformer for image restoration. arXiv preprint arXiv:2309.05239, 2023. 6, 7, 8 +[15] Sung-Jin Cho, Seo-Won Ji, Jun-Pyo Hong, Seung-Won Jung, and Sung-Jea Ko. Rethinking coarse-to-fine approach in single image deblurring. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4641-4650, 2021. 2 +[16] Marcos V Conde, Gregor Geigle, and Radu Timofte. Instructir: High-quality image restoration following human instructions. In European Conference on Computer Vision, pages 1-21. Springer, 2024. 7 +[17] Julien Cornebise, Ivan Oršolíć, and Freddie Kalaitzis. Open high-resolution satellite imagery: The worldstrat dataset—with application to super-resolution. Advances in Neural Information Processing Systems, 35:25979-25991, 2022. 9 +[18] Tao Dai, Jianrui Cai, Yongbing Zhang, Shu-Tao Xia, and Lei Zhang. Second-order attention network for single image super-resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11065-11074, 2019. 10 + +[19] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690-4699, 2019. 5 +[20] Chao Dong, Chen Change Loy, Kaiming He, and Xiaou Tang. Learning a deep convolutional network for image super-resolution. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part IV 13, pages 184-199. Springer, 2014. 2, 10 +[21] Chao Dong, Chen Change Loy, Kaiming He, and Xiaou Tang. Image super-resolution using deep convolutional networks. IEEE transactions on pattern analysis and machine intelligence, 38(2):295-307, 2015. 2 +[22] Zane Durante, Qiuyuan Huang, Naoki Wake, Ran Gong, Jae Sung Park, Bidipta Sarkar, Rohan Taori, Yusuke Noda, Demetri Terzopoulos, Yejin Choi, et al. Agent ai: Surveying the horizons of multimodal interaction. arXiv preprint arXiv:2401.03568, 2024. 2 +[23] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In *Forty-first international conference on machine learning*, 2024. 9 +[24] Ben Fei, Zhaoyang Lyu, Liang Pan, Junzhe Zhang, Weidong Yang, Tianyue Luo, Bo Zhang, and Bo Dai. Generative diffusion prior for unified image restoration and enhancement. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9935-9946, 2023. 10 +[25] Hayit Greenspan. Super-resolution in medical imaging. The computer journal, 52(1):43-63, 2009. 2 +[26] Jinjin Gu, Yujun Shen, and Bolei Zhou. Image processing using multi-code gan prior. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3012-3021, 2020. 10 +[27] Shuhang Gu, Yawei Li, Luc Van Gool, and Radu Timofte. Self-guided network for fast image denoising. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2511-2520, 2019. 10 +[28] Ishaan Gulrajani, Faruk Ahmed, Martin Arjovsky, Vincent Dumoulin, and Aaron C Courville. Improved training of wasserstein gans. Advances in neural information processing systems, 30, 2017. 10 +[29] Xiaojie Guo, Yu Li, and Haibin Ling. Lime: Low-light image enhancement via illumination map estimation. IEEE Transactions on image processing, 26(2):982-993, 2016. 2 +[30] Muhammad Haris, Greg Shakhnarovich, and Norimichi Ukita. Task-driven super resolution: Object detection in low-resolution images. In Neural Information Processing: 28th International Conference, ICONIP 2021, Sanur, Bali, Indonesia, December 8–12, 2021, Proceedings, Part V 28, pages 387–395. Springer, 2021. 2 +[31] Kaiming He, Jian Sun, and Xiaou Tang. Single image haze removal using dark channel prior. IEEE transactions on pattern analysis and machine intelligence, 33(12):2341-2353, 2010. 2 +[32] Yutong He, Dingjie Wang, Nicholas Lai, William Zhang, Chenlin Meng, Marshall Burke, David Lobell, and Stefano Ermon. Spatial-temporal super-resolution of satellite imagery via conditional pixel synthesis. Advances in Neural Information Processing Systems, 34:27903-27915, 2021. 2 +[33] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6 +[34] Minda Hu, Tianqing Fang, Jianshu Zhang, Junyu Ma, Zhisong Zhang, Jingyan Zhou, Hongming Zhang, Haitao Mi, Dong Yu, and Irwin King. Webcot: Enhancing web agent reasoning by reconstructing chain-of-thought in reflection, branching, and rollback. arXiv preprint arXiv:2505.20013, 2025. 5 +[35] Jia-Bin Huang, Abhishek Singh, and Narendra Ahuja. Single image super-resolution from transformed self-exemplars. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5197-5206, 2015. 6 +[36] Xu Huang, Weiwen Liu, Xiaolong Chen, Xingmei Wang, Hao Wang, Defu Lian, Yasheng Wang, Ruiming Tang, and Enhong Chen. Understanding the planning of llm agents: A survey. arXiv preprint arXiv:2402.02716, 2024. 2 +[37] Aaron Hurst, Adam Lerner, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, et al. Gpt-4o system card. arXiv preprint arXiv:2410.21276, 2024. 9 + +[38] Shady Abu Hussein, Tom Tirer, and Raja Giryes. Image-adaptive gan based reconstruction. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 3121-3129, 2020. 10 +[39] Md Jahidul Islam, Sadman Sakib Enan, Peigen Luo, and Junaed Sattar. Underwater image super-resolution using deep residual multipliers. In 2020 IEEE international conference on robotics and automation (ICRA), pages 900-906. IEEE, 2020. 2 +[40] Feng Jia, Lei Tan, Guang Wang, Cheng Jia, and Zhi Chen. A super-resolution network using channel attention retention for pathology images. PeerJ Computer Science, 9:e1196, 2023. Published 2023 Jan 17. 9 +[41] Kui Jiang, Zhongyuan Wang, Peng Yi, Chen Chen, Baojin Huang, Yimin Luo, Jiayi Ma, and Junjun Jiang. Multi-scale progressive fusion network for single image deraining. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8346-8355, 2020. 2 +[42] Xu Jiang, Gehui Li, Bin Chen, and Jian Zhang. Multi-agent image restoration. arXiv preprint arXiv:2503.09403, 2025. 7, 10 +[43] Yitong Jiang, Zhaoyang Zhang, Tianfan Xue, and Jinwei Gu. Autodir: Automatic all-in-one image restoration with latent diffusion. In European Conference on Computer Vision, pages 340-359. Springer, 2024. 2, 7, 10 +[44] Bahjat Kawar, Michael Elad, Stefano Ermon, and Jiaming Song. Denoising diffusion restoration models. Advances in Neural Information Processing Systems, 35:23593-23606, 2022. 10 +[45] Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. Musiq: Multi-scale image quality transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5148-5157, 2021. 4, 5, 6 +[46] Daniel S Kermany, Michael Goldbaum, Wenjia Cai, Carolina CS Valentim, Huiying Liang, Sally L Baxter, Alex McKeown, Ge Yang, Xiaokang Wu, Fangbing Yan, et al. Identifying medical diagnoses and treatable diseases by image-based deep learning. cell, 172(5):1122-1131, 2018. 9 +[47] Jiwon Kim, Jung Kwon Lee, and Kyoung Mu Lee. Accurate image super-resolution using very deep convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1646-1654, 2016. 10 +[48] Yuval Kirstain, Adam Polyak, Uriel Singer, Shahbuland Matiana, Joe Penna, and Omer Levy. Pick-a-pic: An open dataset of user preferences for text-to-image generation. Advances in Neural Information Processing Systems, 36:36652-36663, 2023. 9 +[49] Xiangtao Kong, Chao Dong, and Lei Zhang. Towards effective multiple-in-one image restoration: A sequential and prompt learning strategy. arXiv preprint arXiv:2401.03379, 2024. 7 +[50] Xiangtao Kong, Jinjin Gu, Yihao Liu, Wenlong Zhang, Xiangyu Chen, Yu Qiao, and Chao Dong. A preliminary exploration towards general image restoration. arXiv preprint arXiv:2408.15143, 2024. 7 +[51] Pawel Kowaleczko, Tomasz Tarasiewicz, Maciej Ziaja, Daniel Kostrzewa, Jakub Nalepa, Przemyslaw Rokita, and Michal Kawulok. A real-world benchmark for sentinel-2 multi-image super-resolution. Scientific Data, 10(1):644, 2023. 2 +[52] Somnath Kumar, Yash Gadhia, Tanuja Ganu, and Akshay Nambi. Mmctagent: Multi-modal critical thinking agent framework for complex visual reasoning, 2024. 10 +[53] Black Forest Labs. Flux. https://github.com/black-forest-labs/flux, 2024.9 +[54] Christian Ledig, Lucas Theis, Ferenc Huszár, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, et al. Photo-realistic single image superresolution using a generative adversarial network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4681-4690, 2017. 2, 10 +[55] Baiqi Li, Zhiqiu Lin, Deepak Pathak, Jiayao Li, Yixin Fei, Kewen Wu, Tiffany Ling, Xide Xia, Pengchuan Zhang, Graham Neubig, et al. Genai-bench: Evaluating and improving compositional text-to-visual generation. arXiv preprint arXiv:2406.13743, 2024. 9 +[56] Bingchen Li, Xin Li, Yiting Lu, and Zhibo Chen. Hybrid agents for image restoration. arXiv preprint arXiv:2503.10120, 2025. 10 + +[57] Bingchen Li, Xin Li, Hanxin Zhu, Yeying Jin, Ruoyu Feng, Zhizheng Zhang, and Zhibo Chen. Sed: Semantic-aware discriminator for image super-resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 25784-25795, 2024. 10 +[58] Boyi Li, Wenqi Ren, Dengpan Fu, Dacheng Tao, Dan Feng, Wenjun Zeng, and Zhangyang Wang. Benchmarking single-image dehazing and beyond. IEEE Transactions on Image Processing, 28(1):492-505, 2018. 2 +[59] Boyun Li, Xiao Liu, Peng Hu, Zhongqin Wu, Jiancheng Lv, and Xi Peng. All-in-one image restoration for unknown corruption. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 17452-17462, 2022. 7, 10 +[60] Ke Li, Gang Wan, Gong Cheng, Liqui Meng, and Junwei Han. Object detection in optical remote sensing images: A survey and a new benchmark. ISPRS journal of photogrammetry and remote sensing, 159:296-307, 2020. 9 +[61] Xingzuo Li, Kehai Chen, Yunfei Long, Xuefeng Bai, Yong Xu, and Min Zhang. Generator-assistant stepwise rollback framework for large language model agent. arXiv preprint arXiv:2503.02519, 2025. 5 +[62] Jie Liang, Hui Zeng, and Lei Zhang. Efficient and degradation-adaptive network for real-world image super-resolution. In European Conference on Computer Vision, pages 574-591. Springer, 2022. 10 +[63] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swin transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1833-1844, 2021. 2, 6, 7, 10 +[64] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Kyoung Mu Lee. Enhanced deep residual networks for single image super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 136-144, 2017. 10 +[65] Xinqi Lin, Jingwen He, Ziyan Chen, Zhaoyang Lyu, Bo Dai, Fanghua Yu, Yu Qiao, Wanli Ouyang, and Chao Dong. Diffbir: Toward blind image restoration with generative diffusion prior. In European Conference on Computer Vision, pages 430-448. Springer, 2024. 6, 7, 8, 10 +[66] Jiaxiang Liu, Yuan Wang, Jiawei Du, Joey Tianyi Zhou, and Zuozhu Liu. Medcot: Medical chain of thought via hierarchical expert, 2024. 10 +[67] Xing Liu, Masanori Suganuma, Zhun Sun, and Takayuki Okatani. Dual residual networks leveraging the potential of paired operations for image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7007-7016, 2019. 10 +[68] Ziwei Luo, Fredrik K Gustafsson, Zheng Zhao, Jens Sjolund, and Thomas B Schon. Controlling vision-language models for multi-task image restoration. arXiv preprint arXiv:2310.01018, 2023. 7 +[69] Jiaqi Ma, Tianheng Cheng, Guoli Wang, Qian Zhang, Xinggang Wang, and Lefei Zhang. Prores: Exploring degradation-aware visual prompt for universal image restoration. arXiv preprint arXiv:2306.13653, 2023. 10 +[70] David Martin, Charless Fowlkes, Doron Tal, and Jitendra Malik. A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. In Proceedings eighth IEEE international conference on computer vision. ICCV 2001, volume 2, pages 416-423. IEEE, 2001. 6 +[71] Yusuke Matsui, Kota Ito, Yuji Aramaki, Azuma Fujimoto, Toru Ogawa, Toshihiko Yamasaki, and Kiyoharu Aizawa. Sketch-based manga retrieval using manga109 dataset. Multimedia tools and applications, 76:21811-21838, 2017. 6 +[72] Kangfu Mei, Hossein Talebi, Mojtaba Ardakani, Vishal M Patel, Peyman Milanfar, and Mauricio Delbracio. The power of context: How multimodality improves image super-resolution. arXiv preprint arXiv:2503.14503, 2025. 2 +[73] Sachit Menon, Alexandru Damian, Shijia Hu, Nikhil Ravi, and Cynthia Rudin. Pulse: Self-supervised photo upsampling via latent space exploration of generative models. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 2437-2445, 2020. 10 +[74] Anish Mittal, Rajiv Soundararajan, and Alan C Bovik. Making a completely blind image quality analyzer. IEEE Signal processing letters, 20(3):209-212, 2012. 5, 6 + +[75] Zhangkai Ni, Runyu Xiao, Wenhan Yang, Hanli Wang, Zhihua Wang, Lihua Xiang, and Liping Sun. M2trans: Multi-modal regularized coarse-to-fine transformer for ultrasound image super-resolution. IEEE Journal of Biomedical and Health Informatics, pages 1-12, 2024. 9 +[76] Fu-Zhao Ou, Chongyi Li, Shiqi Wang, and Sam Kwong. Clib-fiaq: face image quality assessment with confidence calibration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1694-1704, 2024. 5 +[77] Xingang Pan, Xiaohang Zhan, Bo Dai, Dahua Lin, Chen Change Loy, and Ping Luo. Exploiting deep generative prior for versatile image restoration and manipulation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(11):7474-7489, 2021. 10 +[78] Dongwon Park, Byung Hyun Lee, and Se Young Chun. All-in-one image restoration for unknown degradations using adaptive discriminative filters for specific degradations. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5815-5824. IEEE, 2023. 10 +[79] Shishir G Patil, Tianjun Zhang, Vivian Fang, Roy Huang, Aaron Hao, Martin Casado, Joseph E Gonzalez, Raluca Ada Popa, Ion Stoica, et al. Goex: Perspectives and designs towards a runtime for autonomous llm applications. arXiv preprint arXiv:2404.06921, 2024. 5 +[80] Vaishnav Potlapalli, Syed Waqas Zamir, Salman H Khan, and Fahad Shahbaz Khan. Prompt: Prompting for all-in-one image restoration. Advances in Neural Information Processing Systems, 36:71275-71293, 2023. 7 +[81] Dongwei Ren, Wangmeng Zuo, Qinghua Hu, Pengfei Zhu, and Deyu Meng. Progressive image deraining networks: A better and simpler baseline. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3937-3946, 2019. 2 +[82] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image super-resolution via iterative refinement. IEEE transactions on pattern analysis and machine intelligence, 45(4):4713-4726, 2022. 2, 3, 10 +[83] Lothar Schermelleh, Alexia Ferrand, Thomas Huser, Christian Eggeling, Markus Sauer, Oliver Biehlmaier, and Gregor PC Drummen. Super-resolution microscopy demystified. Nature cell biology, 21(1):72-84, 2019. 2 +[84] Tixiao Shan, Jinkun Wang, Fanfei Chen, Paul Szenher, and Brendan Englot. Simulation-based lidar super-resolution for ground vehicles. Robotics and Autonomous Systems, 134:103647, 2020. 2 +[85] Jacob Shermeyer and Adam Van Etten. The effects of super-resolution on object detection performance in satellite imagery. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, 2019. 2 +[86] FUJIFILM Healthcare Europe & SonoSkills. US-CASE: Ultrasound Cases Dataset. http://www.ultrasoundcases.info/Cases-Home.aspx, 2025.9 +[87] J. J. Staal, M. D. Abramoff, M. Niemeijer, M. A. Viergever, and B. van Ginneken. Ridge-based vessel segmentation in color images of the retina. IEEE Transactions on Medical Imaging, 23(4):501-509, 2004. 9 +[88] Theodore R. Sumers, Shunyu Yao, Karthik Narasimhan, and Thomas L. Griffiths. Cognitive architectures for language agents, 2024. 10 +[89] Lingchen Sun, Rongyuan Wu, Zhiyuan Ma, Shuaizheng Liu, Qiaosi Yi, and Lei Zhang. Pixel-level and semantic-level adjustable super-resolution: A dual-lora approach. arXiv preprint arXiv:2412.03017, 2024. 2, 8 +[90] Xin Tao, Hongyun Gao, Xiaoyong Shen, Jue Wang, and Jiaya Jia. Scale-recurrent network for deep image deblurring. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8174-8182, 2018. 2 +[91] Jennifer A Thorley, Jeremy Pike, and Joshua Z Rappoport. Super-resolution microscopy: a comparison of commercially available options. In *Fluorescence microscopy*, pages 199-212. Elsevier, 2014. 9 +[92] Kalina L Tosheva, Yue Yuan, Pedro Matos Pereira, Sián Culley, and Ricardo Henriques. Between life and death: strategies to reduce phototoxicity in super-resolution microscopy. Journal of Physics D: Applied Physics, 53(16):163001, 2020. 9 + +[93] Zhengzhong Tu, Hossein Talebi, Han Zhang, Feng Yang, Peyman Milanfar, Alan Bovik, and Yinxiao Li. Maxim: Multi-axis mlp for image processing. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5769-5780, 2022. 2, 10 +[94] Zhengzhong Tu, Yilin Wang, Neil Birkbeck, Balu Adsumilli, and Alan C Bovik. Ugc-vqa: Benchmarking blind video quality assessment for user generated content. IEEE Transactions on Image Processing, 30:4449-4464, 2021. 2 +[95] Sabina Umirzakova, Shabir Ahmad, Latif U Khan, and Taegkeun Whangbo. Medical image superresolution for smart healthcare applications: A comprehensive survey. Information Fusion, 103:102075, 2024. 9 +[96] Jeya Maria Jose Valanarasu, Rajeev Yasarla, and Vishal M Patel. Transweather: Transformer-based restoration of images degraded by adverse weather conditions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2353-2363, 2022. 10 +[97] Jianyi Wang, Kelvin CK Chan, and Chen Change Loy. Exploring clip for assessing the look and feel of images. In Proceedings of the AAAI conference on artificial intelligence, volume 37, pages 2555–2563, 2023. 4, 5 +[98] Jianyi Wang, Zongsheng Yue, Shangchen Zhou, Kelvin CK Chan, and Chen Change Loy. Exploiting diffusion prior for real-world image super-resolution. International Journal of Computer Vision, 132(12):5929-5949, 2024. 2, 7, 10 +[99] Xiaohan Wang, Yuhui Zhang, Orr Zohar, and Serena Yeung-Levy. Videoagent: Long-form video understanding with large language model as agent. In Ales Leonardis, Elisa Ricci, Stefan Roth, Olga Russakovsky, Torsten Sattler, and Gul Varol, editors, Computer Vision – ECCV 2024, pages 58–76, Cham, 2025. Springer Nature Switzerland. 10 +[100] Xiaosong Wang, Yifan Peng, Le Lu, Zhiyong Lu, Mohammadhadi Bagheri, and Ronald M Summers. Chestx-ray8: Hospital-scale chest x-ray database and benchmarks on weakly-supervised classification and localization of common thorax diseases. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2097-2106, 2017. 9 +[101] Xintao Wang, Liangbin Xie, Chao Dong, and Ying Shan. Real-esrgan: Training real-world blind super-resolution with pure synthetic data. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1905–1914, 2021. 2, 10 +[102] Xintao Wang, Ke Yu, Shixiang Wu, Jinjin Gu, Yihao Liu, Chao Dong, Yu Qiao, and Chen Change Loy. Esrgan: Enhanced super-resolution generative adversarial networks. In Proceedings of the European conference on computer vision (ECCV) workshops, 2018. 2 +[103] Yinhuai Wang, Jiwen Yu, and Jian Zhang. Zero-shot image restoration using denoising diffusion null-space model. arXiv preprint arXiv:2212.00490, 2022. 10 +[104] Yufei Wang, Wenhan Yang, Xinyuan Chen, Yaohui Wang, Lanqing Guo, Lap-Pui Chau, Ziwei Liu, Yu Qiao, Alex C Kot, and Bihan Wen. Sinsr: diffusion-based image super-resolution in a single step. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 25796-25805, 2024. 7, 10 +[105] Zhendong Wang, Xiaodong Cun, Jianmin Bao, Wengang Zhou, Jianzhuang Liu, and Houqiang Li. Uformer: A general u-shaped transformer for image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 17683-17693, 2022. 10 +[106] Zhihao Wang, Jian Chen, and Steven CH Hoi. Deep learning for image super-resolution: A survey. IEEE transactions on pattern analysis and machine intelligence, 43(10):3365-3387, 2020. 2 +[107] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 6 +[108] Zijie J Wang, Evan Montoya, David Munechika, Haoyang Yang, Benjamin Hoover, and Duen Horng Chau. Diffusiondb: A large-scale prompt gallery dataset for text-to-image generative models. arXiv preprint arXiv:2210.14896, 2022. 9 +[109] Chen Wei, Wenjing Wang, Wenhan Yang, and Jiaying Liu. Deep retinex decomposition for low-light enhancement. arXiv preprint arXiv:1808.04560, 2018. 2 + +[110] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, brian richter, Fei Xia, Ed Chi, Quoc V Le, and Denny Zhou. Chain-of-thought prompting elicits reasoning in large language models. In S. Koyejo, S. Mohamed, A. Agarwal, D. Belgrave, K. Cho, and A. Oh, editors, Advances in Neural Information Processing Systems, volume 35, pages 24824-24837. Curran Associates, Inc., 2022. 10 +[111] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824-24837, 2022. 2 +[112] Pengxu Wei, Ziwei Xie, Hannan Lu, Zongyuan Zhan, Qixiang Ye, Wangmeng Zuo, and Liang Lin. Component divide-and-conquer for real-world image super-resolution. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VIII 16, pages 101-117. Springer, 2020. 7 +[113] Rongyuan Wu, Lingchen Sun, Zhiyuan Ma, and Lei Zhang. One-step effective diffusion network for real-world image super-resolution. Advances in Neural Information Processing Systems, 37:92529-92553, 2024. 2, 6, 7, 8, 10 +[114] Rongyuan Wu, Tao Yang, Lingchen Sun, Zhengqiang Zhang, Shuai Li, and Lei Zhang. Seesr: Towards semantics-aware real-world image super-resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 25456-25467, 2024. 2, 7, 10 +[115] Xiaoshi Wu, Yiming Hao, Keqiang Sun, Yixiong Chen, Feng Zhu, Rui Zhao, and Hongsheng Li. Human preference score v2: A solid benchmark for evaluating human preferences of text-to-image synthesis. arXiv preprint arXiv:2306.09341, 2023. 5 +[116] Bin Xia, Yulun Zhang, Shiyin Wang, Yitong Wang, Xinglong Wu, Yapeng Tian, Wenming Yang, and Luc Van Gool. Diffir: Efficient diffusion model for image restoration. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13095-13105, 2023. 10 +[117] Gui-Song Xia, Xiang Bai, Jian Ding, Zhen Zhu, Serge Belongie, Jiebo Luo, Mihai Datcu, Marcello Pelillo, and Liangpei Zhang. Dota: A large-scale dataset for object detection in aerial images. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3974-3983, 2018. 9 +[118] Gui-Song Xia, Jingwen Hu, Fan Hu, Baoguang Shi, Xiang Bai, Yanfei Zhong, Liangpei Zhang, and Xiaoqiang Lu. Aid: A benchmark data set for performance evaluation of aerial scene classification. IEEE Transactions on Geoscience and Remote Sensing, 55(7):3965-3981, 2017. 9 +[119] Enze Xie, Junsong Chen, Junyu Chen, Han Cai, Haotian Tang, Yujun Lin, Zhekai Zhang, Muyang Li, Ligeng Zhu, Yao Lu, et al. Sana: Efficient high-resolution image synthesis with linear diffusion transformers. arXiv preprint arXiv:2410.10629, 2024. 9 +[120] Jianchao Yang, John Wright, Thomas S Huang, and Yi Ma. Image super-resolution via sparse representation. IEEE transactions on image processing, 19(11):2861-2873, 2010. 2 +[121] Sidi Yang, Tianhe Wu, Shuwei Shi, Shanshan Lao, Yuan Gong, Mingdeng Cao, Jiahao Wang, and Yujiu Yang. Maniaq: Multi-dimension attention network for no-reference image quality assessment. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1191-1200, 2022. 5 +[122] Tao Yang, Rongyuan Wu, Peiran Ren, Xuansong Xie, and Lei Zhang. Pixel-aware stable diffusion for realistic image super-resolution and personalized stylization. In European Conference on Computer Vision, pages 74-91. Springer, 2024. 7, 10 +[123] Mingde Yao, Ruikang Xu, Yuanshen Guan, Jie Huang, and Zhiwei Xiong. Neural degradation representation learning for all-in-one image restoration. IEEE Transactions on Image Processing, 2024. 10 +[124] Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. React: Synergizing reasoning and acting in language models. In International Conference on Learning Representations (ICLR), 2023. 2, 10 +[125] Zhenqiang Ying, Haoran Niu, Praful Gupta, Dhruv Mahajan, Deepti Ghadiyaram, and Alan Bovik. From patches to pictures (paq-2-piq): Mapping the perceptual space of picture quality. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3575–3585, 2020. 2 +[126] Jiahui Yu, Zhe Lin, Jimei Yang, Xiaohui Shen, Xin Lu, and Thomas S Huang. Free-form image inpainting with gated convolution. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4471-4480, 2019. 10 + +[127] Zongsheng Yue, Jianyi Wang, and Chen Change Loy. Reshift: Efficient diffusion model for image superresolution by residual shifting. Advances in Neural Information Processing Systems, 36:13294-13307, 2023. 7 +[128] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5728-5739, 2022. 2, 10 +[129] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling Shao. Learning enriched features for real image restoration and enhancement. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXV 16, pages 492-511. Springer, 2020. 10 +[130] Roman Zeyde, Michael Elad, and Matan Protter. On single image scale-up using sparse-representations. In International conference on curves and surfaces, pages 711-730. Springer, 2010. 6 +[131] Cheng Zhang, Yu Zhu, Qingsen Yan, Jinqiu Sun, and Yanning Zhang. All-in-one multi-degradation image restoration network via hierarchical degradation representation. In Proceedings of the 31st ACM international conference on multimedia, pages 2285–2293, 2023. 10 +[132] Jinjin Zhang, Qiuyu Huang, Junjie Liu, Xiefan Guo, and Di Huang. Diffusion-4k: Ultra-high-resolution image synthesis with latent diffusion models. arXiv preprint arXiv:2503.18352, 2025. 8, 9 +[133] Kai Zhang, Yawei Li, Wangmeng Zuo, Lei Zhang, Luc Van Gool, and Radu Timofte. Plug-and-play image restoration with deep denoiser prior. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(10):6360-6376, 2021. 10 +[134] Kai Zhang, Jingyun Liang, Luc Van Gool, and Radu Timofte. Designing a practical degradation model for deep blind image super-resolution. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4791-4800, 2021. 2, 8, 10 +[135] Kai Zhang, Wangmeng Zuo, Yunjin Chen, Deyu Meng, and Lei Zhang. Beyond a gaussian denoiser: Residual learning of deep cnn for image denoising. IEEE transactions on image processing, 26(7):3142-3155, 2017. 2, 10 +[136] Liangpei Zhang, Hongyan Zhang, Huanfeng Shen, and Pingxiang Li. A super-resolution reconstruction algorithm for surveillance images. Signal Processing, 90(3):848-859, 2010. 2 +[137] Lin Zhang, Lei Zhang, and Alan C Bovik. A feature-enriched completely blind image quality evaluator. IEEE Transactions on Image Processing, 24(8):2579-2591, 2015. 4 +[138] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6 +[139] Yulun Zhang, Kunpeng Li, Kai Li, Lichen Wang, Bineng Zhong, and Yun Fu. Image super-resolution using very deep residual channel attention networks. In Proceedings of the European conference on computer vision (ECCV), pages 286-301, 2018. 6, 10 +[140] Yulun Zhang, Yapeng Tian, Yu Kong, Bineng Zhong, and Yun Fu. Residual dense network for image super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2472-2481, 2018. 2, 10 +[141] Zhisong Zhang, Tianqing Fang, Kaixin Ma, Wenhao Yu, Hongming Zhang, Haitao Mi, and Dong Yu. Enhancing web agents with explicit rollback mechanisms. arXiv preprint arXiv:2504.11788, 2025. 5 +[142] Qi Zheng, Yibo Fan, Leilei Huang, Tianyu Zhu, Jiaming Liu, Zhijian Hao, Shuo Xing, Chia-Ju Chen, Xiongkuo Min, Alan C Bovik, et al. Video quality assessment: A comprehensive survey. arXiv preprint arXiv:2412.04508, 2024. 6 +[143] Yingjie Zhou, Jiezhang Cao, Zicheng Zhang, Farong Wen, Yanwei Jiang, Jun Jia, Xiaohong Liu, Xiongkuo Min, and Guangtao Zhai. Q-agent: Quality-driven chain-of-thought image restoration agent through robust multimodal large language model. arXiv preprint arXiv:2504.07148, 2025. 10 +[144] Kaiwen Zhu, Jinjin Gu, Zhiyuan You, Yu Qiao, and Chao Dong. An intelligent agentic system for complex image restoration problems. arXiv preprint arXiv:2410.17809, 2024. 2, 5, 6, 7, 10 +[145] Ruoxi Zhu, Zhengzhong Tu, Jiaming Liu, Alan C Bovik, and Yibo Fan. Mwformer: Multi-weather image restoration using degradation-aware transformers. IEEE Transactions on Image Processing, 2024. 10 + +# NeurIPS Paper Checklist + +# 1. Claims + +Question: Do the main claims made in the abstract and introduction accurately reflect the paper's contributions and scope? + +Answer: [Yes] + +Justification: The abstract and introduction covers the contribution and scope of this paper. Guidelines: + +- The answer NA means that the abstract and introduction do not include the claims made in the paper. +- The abstract and/or introduction should clearly state the claims made, including the contributions made in the paper and important assumptions and limitations. A No or NA answer to this question will not be perceived well by the reviewers. +- The claims made should match theoretical and experimental results, and reflect how much the results can be expected to generalize to other settings. +- It is fine to include aspirational goals as motivation as long as it is clear that these goals are not attained by the paper. + +# 2. Limitations + +Question: Does the paper discuss the limitations of the work performed by the authors? + +Answer: [Yes] + +Justification: We discuss the limitation of our proposed method. + +Guidelines: + +- The answer NA means that the paper has no limitation while the answer No means that the paper has limitations, but those are not discussed in the paper. +- The authors are encouraged to create a separate "Limitations" section in their paper. +- The paper should point out any strong assumptions and how robust the results are to violations of these assumptions (e.g., independence assumptions, noiseless settings, model well-specification, asymptotic approximations only holding locally). The authors should reflect on how these assumptions might be violated in practice and what the implications would be. +- The authors should reflect on the scope of the claims made, e.g., if the approach was only tested on a few datasets or with a few runs. In general, empirical results often depend on implicit assumptions, which should be articulated. +- The authors should reflect on the factors that influence the performance of the approach. For example, a facial recognition algorithm may perform poorly when image resolution is low or images are taken in low lighting. Or a speech-to-text system might not be used reliably to provide closed captions for online lectures because it fails to handle technical jargon. +- The authors should discuss the computational efficiency of the proposed algorithms and how they scale with dataset size. +- If applicable, the authors should discuss possible limitations of their approach to address problems of privacy and fairness. +- While the authors might fear that complete honesty about limitations might be used by reviewers as grounds for rejection, a worse outcome might be that reviewers discover limitations that aren't acknowledged in the paper. The authors should use their best judgment and recognize that individual actions in favor of transparency play an important role in developing norms that preserve the integrity of the community. Reviewers will be specifically instructed to not penalize honesty concerning limitations. + +# 3. Theory assumptions and proofs + +Question: For each theoretical result, does the paper provide the full set of assumptions and a complete (and correct) proof? + +Answer: [NA] + +Justification: Paper does not include theoretical results. + +# Guidelines: + +- The answer NA means that the paper does not include theoretical results. +- All the theorems, formulas, and proofs in the paper should be numbered and cross-referenced. +- All assumptions should be clearly stated or referenced in the statement of any theorems. +- The proofs can either appear in the main paper or the supplemental material, but if they appear in the supplemental material, the authors are encouraged to provide a short proof sketch to provide intuition. +- Inversely, any informal proof provided in the core of the paper should be complemented by formal proofs provided in appendix or supplemental material. +- Theorems and Lemmas that the proof relies upon should be properly referenced. + +# 4. Experimental result reproducibility + +Question: Does the paper fully disclose all the information needed to reproduce the main experimental results of the paper to the extent that it affects the main claims and/or conclusions of the paper (regardless of whether the code and data are provided or not)? + +Answer: [Yes] + +Justification: Our proposed method in this paper is reproducible. + +# Guidelines: + +- The answer NA means that the paper does not include experiments. +- If the paper includes experiments, a No answer to this question will not be perceived well by the reviewers: Making the paper reproducible is important, regardless of whether the code and data are provided or not. +- If the contribution is a dataset and/or model, the authors should describe the steps taken to make their results reproducible or verifiable. +- Depending on the contribution, reproducibility can be accomplished in various ways. For example, if the contribution is a novel architecture, describing the architecture fully might suffice, or if the contribution is a specific model and empirical evaluation, it may be necessary to either make it possible for others to replicate the model with the same dataset, or provide access to the model. In general, releasing code and data is often one good way to accomplish this, but reproducibility can also be provided via detailed instructions for how to replicate the results, access to a hosted model (e.g., in the case of a large language model), releasing of a model checkpoint, or other means that are appropriate to the research performed. +- While NeurIPS does not require releasing code, the conference does require all submissions to provide some reasonable avenue for reproducibility, which may depend on the nature of the contribution. For example +(a) If the contribution is primarily a new algorithm, the paper should make it clear how to reproduce that algorithm. +(b) If the contribution is primarily a new model architecture, the paper should describe the architecture clearly and fully. +(c) If the contribution is a new model (e.g., a large language model), then there should either be a way to access this model for reproducing the results or a way to reproduce the model (e.g., with an open-source dataset or instructions for how to construct the dataset). +(d) We recognize that reproducibility may be tricky in some cases, in which case authors are welcome to describe the particular way they provide for reproducibility. In the case of closed-source models, it may be that access to the model is limited in some way (e.g., to registered users), but it should be possible for other researchers to have some path to reproducing or verifying the results. + +# 5. Open access to data and code + +Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material? + +# Answer: [No] + +Justification: We will release our code and data after our paper gets accepted. + +# Guidelines: + +- The answer NA means that paper does not include experiments requiring code. +- Please see the NeurIPS code and data submission guidelines (https://nips.cc/public/subjects/CodeSubmissionPolicy) for more details. +- While we encourage the release of code and data, we understand that this might not be possible, so No is an acceptable answer. Papers cannot be rejected simply for not including code, unless this is central to the contribution (e.g., for a new open-source benchmark). +- The instructions should contain the exact command and environment needed to run to reproduce the results. See the NeurIPS code and data submission guidelines (https: //nips.cc/public/subjects/CodeSubmissionPolicy) for more details. +- The authors should provide instructions on data access and preparation, including how to access the raw data, preprocessed data, intermediate data, and generated data, etc. +- The authors should provide scripts to reproduce all experimental results for the new proposed method and baselines. If only a subset of experiments are reproducible, they should state which ones are omitted from the script and why. +- At submission time, to preserve anonymity, the authors should release anonymized versions (if applicable). +- Providing as much information as possible in supplemental material (appended to the paper) is recommended, but including URLs to data and code is permitted. + +# 6. Experimental setting/details + +Question: Does the paper specify all the training and test details (e.g., data splits, hyperparameters, how they were chosen, type of optimizer, etc.) necessary to understand the results? + +# Answer: [Yes] + +Justification: We describe the details of benchmark and test metrics in our paper. + +# Guidelines: + +- The answer NA means that the paper does not include experiments. +- The experimental setting should be presented in the core of the paper to a level of detail that is necessary to appreciate the results and make sense of them. +- The full details can be provided either with the code, in appendix, or as supplemental material. + +# 7. Experiment statistical significance + +Question: Does the paper report error bars suitably and correctly defined or other appropriate information about the statistical significance of the experiments? + +# Answer: [No] + +Justification: Experiment results in the paper are not accompanied by error bars. + +# Guidelines: + +- The answer NA means that the paper does not include experiments. +- The authors should answer "Yes" if the results are accompanied by error bars, confidence intervals, or statistical significance tests, at least for the experiments that support the main claims of the paper. +- The factors of variability that the error bars are capturing should be clearly stated (for example, train/test split, initialization, random drawing of some parameter, or overall run with given experimental conditions). +- The method for calculating the error bars should be explained (closed form formula, call to a library function, bootstrap, etc.) +- The assumptions made should be given (e.g., Normally distributed errors). +- It should be clear whether the error bar is the standard deviation or the standard error of the mean. + +- It is OK to report 1-sigma error bars, but one should state it. The authors should preferably report a 2-sigma error bar than state that they have a $96\%$ CI, if the hypothesis of Normality of errors is not verified. +- For asymmetric distributions, the authors should be careful not to show in tables or figures symmetric error bars that would yield results that are out of range (e.g. negative error rates). +- If error bars are reported in tables or plots, The authors should explain in the text how they were calculated and reference the corresponding figures or tables in the text. + +# 8. Experiments compute resources + +Question: For each experiment, does the paper provide sufficient information on the computer resources (type of compute workers, memory, time of execution) needed to reproduce the experiments? + +Answer: [Yes] + +Justification: We describe the computing resources to conduct experiments using our proposed method in the paper. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The paper should indicate the type of compute workers CPU or GPU, internal cluster, or cloud provider, including relevant memory and storage. +- The paper should provide the amount of compute required for each of the individual experimental runs as well as estimate the total compute. +- The paper should disclose whether the full research project required more computer than the experiments reported in the paper (e.g., preliminary or failed experiments that didn't make it into the paper). + +# 9. Code of ethics + +Question: Does the research conducted in the paper conform, in every respect, with the NeurIPS Code of Ethics https://neurips.cc/public/EthicsGuidelines? + +Answer: [Yes] + +Justification: The research conducted in the paper does not violate the NeurIPS Code of Ethics. + +Guidelines: + +- The answer NA means that the authors have not reviewed the NeurIPS Code of Ethics. +- If the authors answer No, they should explain the special circumstances that require a deviation from the Code of Ethics. +- The authors should make sure to preserve anonymity (e.g., if there is a special consideration due to laws or regulations in their jurisdiction). + +# 10. Broader impacts + +Question: Does the paper discuss both potential positive societal impacts and negative societal impacts of the work performed? + +Answer: [Yes] + +Justification: We discuss the social impacts of our proposed method in the paper. + +Guidelines: + +- The answer NA means that there is no societal impact of the work performed. +- If the authors answer NA or No, they should explain why their work has no societal impact or why the paper does not address societal impact. +- Examples of negative societal impacts include potential malicious or unintended uses (e.g., disinformation, generating fake profiles, surveillance), fairness considerations (e.g., deployment of technologies that could make decisions that unfairly impact specific groups), privacy considerations, and security considerations. + +- The conference expects that many papers will be foundational research and not tied to particular applications, let alone deployments. However, if there is a direct path to any negative applications, the authors should point it out. For example, it is legitimate to point out that an improvement in the quality of generative models could be used to generate deepfakes for disinformation. On the other hand, it is not needed to point out that a generic algorithm for optimizing neural networks could enable people to train models that generate Deepfakes faster. +- The authors should consider possible harms that could arise when the technology is being used as intended and functioning correctly, harms that could arise when the technology is being used as intended but gives incorrect results, and harms following from (intentional or unintentional) misuse of the technology. +- If there are negative societal impacts, the authors could also discuss possible mitigation strategies (e.g., gated release of models, providing defenses in addition to attacks, mechanisms for monitoring misuse, mechanisms to monitor how a system learns from feedback over time, improving the efficiency and accessibility of ML). + +# 11. Safeguards + +Question: Does the paper describe safeguards that have been put in place for responsible release of data or models that have a high risk for misuse (e.g., pretrained language models, image generators, or scraped datasets)? + +Answer: [NA] + +Justification: The paper poses no such risks on the release of data and models. + +Guidelines: + +- The answer NA means that the paper poses no such risks. +- Released models that have a high risk for misuse or dual-use should be released with necessary safeguards to allow for controlled use of the model, for example by requiring that users adhere to usage guidelines or restrictions to access the model or implementing safety filters. +- Datasets that have been scraped from the Internet could pose safety risks. The authors should describe how they avoided releasing unsafe images. +- We recognize that providing effective safeguards is challenging, and many papers do not require this, but we encourage authors to take this into account and make a best faith effort. + +# 12. Licenses for existing assets + +Question: Are the creators or original owners of assets (e.g., code, data, models), used in the paper, properly credited and are the license and terms of use explicitly mentioned and properly respected? + +Answer: [Yes] + +Justification: We properly cite assets used in the paper. + +Guidelines: + +- The answer NA means that the paper does not use existing assets. +- The authors should cite the original paper that produced the code package or dataset. +- The authors should state which version of the asset is used and, if possible, include a URL. +- The name of the license (e.g., CC-BY 4.0) should be included for each asset. +- For scraped data from a particular source (e.g., website), the copyright and terms of service of that source should be provided. +- If assets are released, the license, copyright information, and terms of use in the package should be provided. For popular datasets, paperswithcode.com/datasets has curated licenses for some datasets. Their licensing guide can help determine the license of a dataset. +- For existing datasets that are re-packaged, both the original license and the license of the derived asset (if it has changed) should be provided. + +- If this information is not available online, the authors are encouraged to reach out to the asset's creators. + +# 13. New assets + +Question: Are new assets introduced in the paper well documented and is the documentation provided alongside the assets? + +Answer: [Yes] + +Justification: We release a new dataset for image restoration and 4K upscaling. + +Guidelines: + +- The answer NA means that the paper does not release new assets. +- Researchers should communicate the details of the dataset/code/model as part of their submissions via structured templates. This includes details about training, license, limitations, etc. +- The paper should discuss whether and how consent was obtained from people whose asset is used. +- At submission time, remember to anonymize your assets (if applicable). You can either create an anonymized URL or include an anonymized zip file. + +# 14. Crowdsourcing and research with human subjects + +Question: For crowdsourcing experiments and research with human subjects, does the paper include the full text of instructions given to participants and screenshots, if applicable, as well as details about compensation (if any)? + +Answer: [NA] + +Justification: Our paper does not include crowdsourcing experiments. + +Guidelines: + +- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. +- Including this information in the supplemental material is fine, but if the main contribution of the paper involves human subjects, then as much detail as possible should be included in the main paper. +- According to the NeurIPS Code of Ethics, workers involved in data collection, curation, or other labor should be paid at least the minimum wage in the country of the data collector. + +# 15. Institutional review board (IRB) approvals or equivalent for research with human subjects + +Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals (or an equivalent approval/review based on the requirements of your country or institution) were obtained? + +Answer: [NA] + +Justification: Our paper does not involve crowdsourcing nor research with human subjects. + +Guidelines: + +- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. +- Depending on the country in which research is conducted, IRB approval (or equivalent) may be required for any human subjects research. If you obtained IRB approval, you should clearly state this in the paper. +- We recognize that the procedures for this may vary significantly between institutions and locations, and we expect authors to adhere to the NeurIPS Code of Ethics and the guidelines for their institution. +- For initial submissions, do not include any information that would break anonymity (if applicable), such as the institution conducting the review. + +# 16. Declaration of LLM usage + +Question: Does the paper describe the usage of LLMs if it is an important, original, or non-standard component of the core methods in this research? Note that if the LLM is used only for writing, editing, or formatting purposes and does not impact the core methodology, scientific rigorousness, or originality of the research, declaration is not required. + +# Answer: [Yes] + +Justification: We described the usage of LLM component in the proposed method in the paper. + +# Guidelines: + +- The answer NA means that the core method development in this research does not involve LLMs as any important, original, or non-standard components. +- Please refer to our LLM policy (https://neurips.cc/Conferences/2025/LLM) for what should or should not be described. + +Yushen Zuo $^{1}$ , Qi Zheng $^{1\dagger}$ , Mingyang Wu $^{1\dagger}$ , Xinrui Jiang $^{2\dagger}$ , Renjie Li $^{1}$ , Jian Wang $^{3}$ , Yide Zhang $^{4}$ , Gengchen Mai $^{5}$ , Lihong V. Wang $^{6}$ , James Zou $^{2}$ , Xiaoyu Wang $^{7}$ , Ming-Hsuan Yang $^{8}$ , Zhengzhong Tu $^{1\star}$ + +$^{1}$ Texas A&M University $^{2}$ Stanford University $^{3}$ Snap Inc. $^{4}$ CU Boulder $^{5}$ UT Austin $^{6}$ California Institute of Technology $^{7}$ Topaz Labs $^{8}$ UC Merced $^{*}$ Corresponding Author: tzz@amu.edu. $\dagger$ Equal contributions. + +Project Website: 4kagent.github.io + +# Contents + +A Model Card 3 +A.1 Profile 3 +A.2 Model Zoo 3 +A.3 Inference Details 4 +A.3.1 Rollback 4 +A.3.2 Implementation Details 4 +A.3.3 Prompts 4 +BExperiment Overview 5 +C Experiment Part I: $4\times$ Natural Image Super-Resolution 6 +C.1 Classical Image Super-Resolution 6 +C.2 Real-World Image Super-Resolution 9 +C.3 Multiple-Degradation Image Restoration 10 +C.4 Face Restoration 12 +D Experiment Part II: $16\times$ Natural Image Super-Resolution 13 +D.1 Large Scale Factor $(16\times)$ Image Super-Resolution 13 +D.2 Joint restoration & 4K Upscaling 14 +E Experiment Part III: AI-Generated Content (AIGC) 4K Super-Resolution 16 +E.1 AI-Generated Content 4K Super-Resolution Experiment 16 +F Experiment Part IV: Scientific Imagery Super-Resolution 18 + +F.1 Remote Sensing Image Super-Resolution 18 +F.2 Fluorescence Microscopic Image Super-Resolution 25 +F.3 Pathology Image Super-Resolution 27 +F.4 Medical Image Super-Resolution: X-Ray, Ultrasound, and Fundoscopy 29 + +# G Ablation Studies 33 + +# H Applications and Broader Impacts 34 + +H.1 Applications 34 +H.2 Broader Impacts 37 +H.3 Limitations and Potential Negative Societal Impacts 37 + +# I Related Works 38 + +I.1 Image Super-Resolution 38 +I.2 Image Restoration 38 +I.3 LLM Agents 39 + +# A Model Card + +# A.1 Profile + +4KAgent is highly flexible based on the profile setting. Users can easily customize 4KAgent by pre-selecting a profile defined in 4KAgent. We pre-define profile examples in Tab. 1, which cover most use cases and include all the modes used in our experiments. This feature offers easy and intuitive customization for new, unseen scenarios identified by the customers. + +Table 1: Pre-defined Profiles in 4KAgent. + +
Profile NicknamePerception AgentUpscale to 4KScale FactorRestore OptionFace RestoreBrighteningRestore Preference
Gen4K-PDepictQA [205]TrueNoneNoneTrueFalsePerception
Gen4K-FDepictQA [205]TrueNoneNoneTrueFalseFidelity
Aer4K-PLlama-3.2-Vision [6]TrueNoneNoneFalseFalsePerception
Aer4K-FLlama-3.2-Vision [6]TrueNoneNoneFalseFalseFidelity
AerSR-s4-PLlama-3.2-Vision [6]False4NoneFalseFalsePerception
AerSR-s4-FLlama-3.2-Vision [6]False4NoneFalseFalseFidelity
ExpSR-s4-PLlama-3.2-Vision [6]False4super-resolutionFalseFalsePerception
ExpSR-s4-FLlama-3.2-Vision [6]False4super-resolutionFalseFalseFidelity
ExpSR-s2-FLlama-3.2-Vision [6]False2super-resolutionFalseFalseFidelity
ExpSR-s8-FLlama-3.2-Vision [6]False8super-resolutionFalseFalseFidelity
GenSR-s4-PDepictQA [205]False4NoneFalseFalsePerception
GenMIR-PDepictQA [205]False4NoneFalseTruePerception
ExpSRFR-s4-PLlama-3.2-Vision [6]False4super-resolutionTrueFalsePerception
GenSRFR-s4-PDepictQA [205]False4NoneTrueFalsePerception
+ +Profile naming convention: We combine restoration type, restoration task, and restoration preference to construct the profile name. For example, Gen indicates a General image, 4K indicates "Upscale to 4K" on, and P indicates to restore the image with high Perceptual quality. Aer indicates Aerial image, Exp corresponds to Explicit setting, indicating that the profile has explicitly set the restoration task (e.g., SR, which indicates Super-Resolution). MIR indicates Multiple Image Restoration. FR indicates Face Restoration. s4 indicates to upscale the image by a scale factor of 4. + +4KAgent supports various VLMs or LLMs in the Perception Agent, enabling effective analysis of image content and degradation. Specifically, users can select either DepictQA [205] or Llama-3.2-Vision (11B) [6] as available options, but can also be extended to other more recent VLM models, e.g., Qwen2.5-VL [10]. For the VLMs or LLMs to schedule the restoration plan, users can choose from GPT-4 [4], or Llama-3.2-Vision. This is configured by the Perception Agent in the profile module. For example, when it is set to Llama-3.2-Vision, the Llama-3.2-Vision model serves as the core engine to perceive image content and degradation, and then schedules the restoration plan $P_{I}$ . As DepictQA is fine-tuned for image quality assessment (IQA), when it is set as the VLM in the perception agent, 4KAgent will use Llama-3.2-Vision to obtain the image description $C_{I}$ and use GPT-4 [4] to schedule the restoration plan. + +# A.2 Model Zoo + +The 4KAgent system supports nine distinct image restoration models in the toolbox: ① Brightening, ② Defocus Deblurring, ③ Motion Deblurring, ④ Dehazing, ⑤ Denoising, ⑥ Deraining, ⑦ JPEG Compression Artifact Removal (JPEG CAR), ⑧ Super Resolution, and ⑨ Face Restoration. For each of these tasks, we integrate advanced state-of-the-art methods into our comprehensive restoration toolbox. Detailed correspondences between restoration tasks and their respective methods are presented below, where 'QF' denotes the Quality Factor and 'BQF' indicates methods that are blind to the Quality Factor in the JPEG CAR task. + +![](images/ffbfcedacdee7bd75cf38042ab213cd866d2563856c4f08bb015cd07b245e63f.jpg) + +As previously mentioned in Appendix A.1, users can tailor the model by adjusting the Restore Preference setting, which prioritizes either perceptual quality or fidelity. We achieve this by partitioning our toolbox methods into perception-oriented and fidelity-oriented categories. For example, the Super-Resolution tools in the toolbox are split into: + +1. Fidelity-based: HAT-L [26], X-Restormer [25], SwinFIR [217], HMA [30], DRCT [57] +2. Perception-based: DiffBIR [105], HAT-GAN [26], OSEDiff [183], PiSA-SR [153], SwinIR (Real-ISR) [101] + +Therefore, when Restore Preference is set to Perception, 4KAgent will only use the Perception-based methods to restore the image, efficiently meeting the users restoration requirements. + +We further develop a Fast4K mode for 4KAgent. Specifically, when the size of the input image at the current step of the restoration plan exceeds a predefined threshold $s_t$ , 4KAgent automatically excludes methods with long inference times from the toolbox, such as DiffBIR (a 50-step diffusion-based method) in the super-resolution toolbox. Users can adjust $s_t$ to control the running time of 4KAgent. To comprehensively evaluate the performance of 4KAgent, we disable it in all our experiments. + +# A.3 Inference Details + +# A.3.1 Rollback + +In this section, we present the detailed workflow of the rollback mechanism in 4KAgent. If the quality score of $I_{k}$ after step $k$ in the initial restoration plan $P_{I}$ is lower than a threshold $\eta$ , i.e., $Q_{s}(I_{k}) \leq \eta$ , this step will be seen as a failure step and 4KAgent will generate a failure message $S_{I}$ . Then 4KAgent will employ the perception agent to adjust the subsequent plan based on the degradation list $D_{I}$ , the remaining restoration tasks $A_{I}^{R}$ of the restoration agenda $A_{I}$ , restoration experience $E$ , and failure message $F_{I}$ : $P_{I}^{adj} = M_{P}(D_{I}, A_{I}^{R}, E, S_{I})$ . After that, 4KAgent will assign a different restoration task in this step. If all subsequent restoration tasks assigned to this step lead to rollback, 4KAgent will take a compromise policy and go back to the original plan to execute subsequent tasks. + +# A.3.2 Implementation Details + +Computing Resource. As a multi-agent system, 4KAgent supports multi-GPU deployment. Specifically, 4KAgent assigns different agents (Perception Agent, Restoration Agent) on different GPUs to conserve memory. Most of our experiments were conducted using two NVIDIA RTX 4090 GPUs. + +Hyper-parameters. Hyperparameters in 4KAgent reside in the Restoration Agent, namely the weights used to compute the execution quality scores $Q_{s}$ and $Q_{s}^{f}$ in execution, as well as the quality threshold $\eta$ in rollback. Specifically, in 4KAgent, we set $w_{\mathrm{NIQE}} = 1.0$ , $w_{\mathrm{MUSIQ}} = 0.01$ , $w_{\mathrm{MANIQA}} = 1.0$ , $w_{\mathrm{CLIPQA}} = 1.0$ for $Q_{s}$ , $w_{\mathrm{IP}} = 0.001$ , $w_{\mathrm{IQA}} = 1.0$ for $Q_{s}^{f}$ , and $\eta = 0.5$ for rollback. + +# A.3.3 Prompts + +In 4KAgent, we enable the VLM / LLM to perceive image degradations and formulate a restoration plan via customized system prompts. In this section, we present the details of these prompts in + +4KAgent. When the Perception Agent in the profile module selects DepictQA, we use the same prompt as in AgenticIR [235] for DepictQA to assess the image degradations, and GPT-4 to construct the restoration plan. When setting Llama-3.2-Vision in Perception Agent, we design tailored prompts for degradation reasoning and planning, as shown below, where $\{\cdot\}$ represents slots to fill according to the context, and the content inside comes from external input. For the restoration experience $E$ in 4KAgent, we employ the restoration experience from AgenticIR. + +# Prompt for Llama-Vision in Degradation Reasoning + +You are an expert tasked with image quality assessment (IQA) and well-versed in popular IQA metrics, including CLIPIQA+, TOPIQ_NR, MUSIQ, and NIQE. Note that for NIQE, a lower score indicates better image quality, whereas for the other metrics, higher scores generally reflect better quality. Here's an image to restore, along with its corresponding quality scores evaluated using the aforementioned IQA metrics. First, please describe the content and style of the input image, the description must not contain its image quality. Second, please assess the image based on both the metric scores and your prior visual knowledge, with respect to the following two degradations: noise, motion blur, defocus blur, haze, rain, jpeg compression artifact. Images may suffer from one or more of these degradations. **Do not output any explanations or comments.** **Strictly return only a JSON object** containing degradation types and image content/style description. The keys in the JSON object should be: 'degradations' and 'image_description'. Information about the input image: IQA metrics: {iqa_result}. ({iqa_result} corresponds to $Q_I$ .) + +# Prompt for Llama-Vision in Planning (Rollback) + +You are an expert in image restoration. Given an image of low quality, your task is to guide the user to utilize various tools to enhance its quality. The input image requires a list of restoration tasks. Your goal is to make a plan (the order of the tasks) based on the task list. The final output should be formatted as a JSON object containing the restoration plan (the correct order of the tasks). The key in the JSON object should be: 'plan'. Information about the input image: Its description is: {image_description} $(C_I)$ , It suffers from degradations {degradations} $(D_I)$ , The list of restoration tasks: {tasks} $(A_I / A_I^R)$ , For your information, based on past trials, we have the following experience in making a restoration plan: {experience} $(E)$ . Based on this experience, please give the correct order of the tasks in the restoration plan. The restoration plan must be a permutation of {tasks} in the order you determine. (Besides, in attempts just now, we found the result is unsatisfactory if {failed_tries} $(S_I)$ is conducted first. Remember not to arrange {failed_tries} in the first place.) **Do not output any explanations or comments.** **Strictly return only a JSON object** containing plan. The keys in the JSON object should be: 'plan'. + +# B Experiment Overview + +We evaluate 4KAgent on a variety of complex degradation and super-resolution (SR) tasks, demonstrate its flexible profile-driven modes for different restoration requirements, validate its generalization to multiple image domains, and quantify the contributions of each core component via ablation studies. Specifically, we test 4KAgent on a wide range of 11 image SR tasks on 26 benchmarks. The summary of datasets used in experiments is shown in Tab. 2, which can be classified as natural degraded images (\$C,D), AI-generated images (\$E), and scientific images (\$F). Then, we perform an ablation study on Q-MoE policy and Face restoration pipeline in 4KAgent with a runtime analysis (\$G). + +First, we evaluate 4KAgent on natural image restoration / super-resolution tasks under general settings, including classical image super-resolution $(4\times)$ (§C.1), real-world image super-resolution $(4\times)$ (§C.2), multiple-degradation image restoration (§C.3), and face restoration $(4\times)$ (§C.4). Next, we assess its performance in more challenging scenarios, such as large scale factor super-resolution $(16\times)$ (§D.1) and joint restoration with 4K upscaling (§D.2). Finally, we extend 4KAgent to diverse domains by testing its capabilities on AIGC images (§E.1) and scientific imagery (§F), including remote sensing (§F.1), microscopy (§F.2), pathology (§F.3), and medical images (§F.4). To comprehensively evaluate the performance of 4KAgent, we disable the Fast4K mode. + +# C Experiment Part I: $4 \times$ Natural Image Super-Resolution + +# C.1 Classical Image Super-Resolution + +Settings In this section, we provide a more detailed experiment result on Set5 [13], Set14 [213], B100 [121], Urban100 [58], and Manga109 [123] datasets. For a more comprehensive comparison, we employ more recent methods, including HMA [30], DRCT [57], SwinFIR [217], as they are contained in the toolbox. Except for metrics used in the main paper (PSNR, SSIM [177], LPIPS [225], FID [55], NIQE [223], MUSIQ [68]), we employ DISTS [38], CLIPIQA [166], and MANIQA-pipal [198] for evaluation. + +Specifically, PSNR and SSIM are computed on the Y channel in the YCbCr space. They are used to measure the fidelity of images. LPIPS and DISTS are computed in the RGB space, are used to measure the perceptual quality of images. FID is used to evaluate the distance of distributions between the ground truth and the restored images. NIQE, CLIP-IQA, MUSIQ, and MANIQA-pipal are used to evaluate the perceptual quality of images without reference images. + +Quantitative Comparison It should be noted that, once the user sets the Restore Option to super-resolution in the profile, the 4KAgent system can be seen as a quality-driven Mixture-of-Expert system for image super-resolution. In this mode, the system sequentially invokes every super-resolution tool in its toolbox based on the Restore Option setting in the profile, then selects the best result based on the quality score $Q_{s}$ . Accordingly, we group 4KAgent with ExpSR-s4-F and ExpSR-s4-P profile to Fidelity based m + +Table 2: Dataset summary in 4KAgent experiments. + +
TaskDataset#Test Images
Classical SR (§C.1)Set5 [13]5
Set14 [213]14
B100 [121]100
Urban100 [58]100
Manga109 [123]109
Real-World SR (§C.2)RealSR [16]100
DrealSR [181]93
Multiple-Degradation IR (§C.3)MiO-Group A [235]640
MiO-Group B [235]400
MiO-Group C [235]400
Face Restoration (§C.4)WebPhoto-Test [171]407
Large Scale Factor SR (§D.1)RealSRSet [220]20
Joint Image Restoration + 4K Upscaling (§D.2)DIV4K-50 (Ours)50
AI-Generated Content 4K SR (§E.1)GenAIBench-4K [87]100
DiffusionDB-4K [179]100
Remote Sensing SR (§F.1)AID [189]51
DIOR [94]154
DOTA [188]183
WorldStrat [33]85
Fluorescence Microscopy Image SR (§F.2)SR-CACO-2 [12]300
Pathology Image SR (§F.3)bcSR [63]200
Medical Image SR (§F.4)Chest X-ray 2017 [70]624
Chest X-ray 14 [170]880
US-Case [150]111
MMUS1K [133]100
DRIVE [151]20
+ +Experimental results are shown in Tabs. 3 and 4. For the commonly used fidelity metrics PSNR and SSIM in the classical image SR task, 4KAgent with ExpSR-s4-F profile shows competitive performance compared to state-of-the-art fidelity-based methods, ranking among the top three on Set5, B100, Urban100, and Manga109 datasets. For perception-based methods, we focus more on perceptual metrics such as NIQE, CLIPIQA, MUSIQ, and MANIQA. By simply switching the profile from ExpSR-s4-F to ExpSR-s4-P, 4KAgent achieves strong performance among state-of-the-art perception-based methods, ranking among the top two across most metrics on all classical SR benchmarks. For comparison across agentic systems, 4KAgent outperforms AgenticIR in most metrics on classical SR benchmarks, especially on the Set14 and B100 datasets. + +Table 3: Quantitative comparison on classical image super-resolution benchmarks (Set5, Set14, B100). The top three performances of each metric are marked in **bold**, **underline**, **italic** respectively. For Agentic systems, we only **bold** the best performance. + +
DatasetMethodPSNR↑SSIM↑LPIPS↓DISTS↓FID↓NIQE↓CLIPIAQ↑MUSIQ↑MANIQA↑
Set5Fidelity based method
SwinIR [101]32.920.90440.16690.156757.377.240.617959.980.6095
X-Restormer [25]33.150.90570.16360.156460.247.070.636860.090.6169
DRCT [57]33.260.90670.16160.152652.256.940.640660.210.6100
HAT-L [26]33.290.90820.15820.154256.957.110.638960.440.6212
HMA [30]33.390.90890.15870.153554.617.110.633860.390.6241
4KAgent (ExpSR-s4-F)33.340.90810.15890.154956.626.900.629460.020.6177
Perception based method
SwinIR (Real-ISR) [101]28.480.84460.16320.159063.587.460.707262.430.6153
DiffBIR [105]26.410.75100.20590.188872.796.060.840570.230.6767
OSEDiff [183]26.210.80630.15830.164767.505.780.797368.760.6698
PiSA-SR [153]27.560.81890.13180.151662.945.870.808669.870.6904
4KAgent (ExpSR-s4-P)26.880.78990.15910.165770.635.790.824569.930.6808
Agentic System
AgenticIR [235]23.680.67110.27370.2190124.966.590.775071.880.7079
4KAgent (GenSR-s4-P)26.250.76720.17850.183689.026.720.739670.390.6811
Set14Fidelity based method
SwinIR [101]29.090.79500.26710.157470.496.190.525263.100.5891
X-Restormer [25]29.160.79630.26590.155769.866.220.533262.910.5925
DRCT [57]29.570.80090.26170.152467.846.090.536263.120.5932
HAT-L [26]29.460.80140.25650.151666.616.110.526763.230.5986
HMA [30]29.510.80190.25670.151069.416.250.527863.000.6012
4KAgent (ExpSR-s4-F)29.430.79890.25930.152867.835.950.531563.450.5970
Perception based method
SwinIR (Real-ISR) [101]25.910.71870.22440.150896.194.450.650666.820.6054
DiffBIR [105]24.730.63490.23380.1545100.514.340.755372.970.6869
OSEDiff [183]24.300.66630.23890.1524101.034.610.726470.020.6674
PiSA-SR [153]24.760.67160.19930.134389.914.160.764371.810.7015
4KAgent (ExpSR-s4-P)24.760.64710.21580.1467101.994.010.774072.540.6956
Agentic System
AgenticIR [235]21.980.60640.28070.1812129.294.580.744972.480.6804
4KAgent (GenSR-s4-P)23.400.63400.24840.1749125.294.290.760473.640.7061
B100Fidelity based method
SwinIR [101]27.920.74890.35480.200594.576.270.537357.710.5860
X-Restormer [25]27.990.75080.35210.197290.526.210.542757.910.5935
DRCT [57]28.100.75350.34800.194787.766.060.549958.780.5895
HAT-L [26]28.080.75470.34400.195289.526.200.547758.710.5991
HMA [30]28.120.75590.34420.195388.466.170.553459.110.6043
4KAgent (ExpSR-s4-F)28.090.75400.34530.195088.896.020.551659.120.5994
Perception based method
SwinIR (Real-ISR) [101]25.420.67110.25000.169992.654.000.632262.780.6085
DiffBIR [105]24.990.61560.27190.166684.993.920.748368.230.6750
OSEDiff [183]24.350.64950.24080.163473.234.080.742268.540.6725
PiSA-SR [153]25.000.65200.21110.147161.824.040.738468.470.6829
4KAgent (ExpSR-s4-P)24.640.62940.23870.160673.643.860.754669.420.6851
Agentic System
AgenticIR [235]22.510.58530.30780.1907102.924.080.747468.360.6752
4KAgent (GenSR-s4-P)23.640.62460.25720.170278.803.930.735469.440.6844
+ +Qualitative Comparison For visual comparison, we select two leading fidelity-based methods (X-Restormer, HAT-L), two perception-based methods (SwinIR (Real-ISR), DiffBIR), as well as one agentic system (AgenticIR) as baselines. For 4KAgent, we present images under GenSR-s4-P profile for a comprehensive comparison. Visual comparisons in Fig. 1 reveal that fidelity-based methods tend to produce overly smooth or blurred details (e.g., HAT-L), even trained with real-world image SR setting (e.g., SwinIR (Real-ISR)), which is visually unpleasant. Diffusion-based method (e.g., DiffBIR) generates rich but unrealistic details. AgenticIR performs well in detail generation but still lacks realism and exhibits noticeable color shifts. 4KAgent delivers both richer and more accurate + +Table 4: Quantitative comparison on classical image super-resolution benchmarks (Urban100 and Manga109). The top three performances of each metric are marked in **bold**, **underline**, **italic** respectively. For Agentic systems, we only **bold** the best performance. + +
DatasetMethodPSNR↑SSIM↑LPIPS↓DISTS↓FID↓NIQE↓CLIPIAQ↑MUSIQ↑MANIQA↑
Urban100Fidelity based method
SwinIR [101]27.450.82540.18400.15333.585.500.500370.000.6693
X-Restormer [25]27.640.82880.18050.15043.655.610.495370.000.6746
DRCT [57]28.780.84920.16230.13882.925.450.527170.480.6778
HAT-L [26]28.580.84950.15980.14112.875.550.505470.620.6866
HMA [30]28.690.85110.15830.14052.935.610.508470.750.6893
4KAgent (ExpSR-s4-F)28.590.84790.15990.13992.975.310.523570.700.6833
Perception based method
SwinIR (Real-ISR) [101]23.240.71840.19080.136525.364.290.616971.990.6578
DiffBIR [105]22.510.63970.20110.139526.104.790.718573.100.6956
OSEDiff [183]21.880.65720.21850.147938.134.670.659372.350.6822
PiSA-SR [153]22.360.67040.18230.129728.514.430.681472.930.7020
4KAgent (ExpSR-s4-P)22.560.65820.19550.137825.554.530.709273.650.6981
Agentic System
AgenticIR [235]22.030.66150.21470.150731.094.650.679073.100.6873
4KAgent (GenSR-s4-P)22.270.65450.20730.144432.294.430.700173.570.6961
Manga109Fidelity based method
SwinIR [101]32.050.92600.09260.07611.885.320.638570.320.6117
X-Restormer [25]32.400.92790.09090.07481.885.480.632570.050.6123
DRCT [57]32.840.93070.08890.06851.495.080.636269.770.6087
HAT-L [26]33.080.93340.08450.06841.485.260.616069.760.6145
HMA [30]33.200.93440.08350.06821.475.240.620869.920.6196
4KAgent (ExpSR-s4-F)32.870.93160.08600.06831.484.950.632969.990.6125
Perception based method
SwinIR (Real-ISR) [101]26.290.85530.13670.094824.594.300.631670.280.5868
DiffBIR [105]23.570.72970.19230.127530.114.550.780474.510.6787
OSEDiff [183]23.740.79800.17030.118141.544.780.687472.510.6538
PiSA-SR [153]24.020.81190.14500.116134.114.350.727774.760.6779
4KAgent (ExpSR-s4-P)23.760.76150.17760.123133.364.320.767875.080.6801
Agentic System
AgenticIR [235]23.700.75500.18620.124634.014.380.745073.980.6597
4KAgent (GenSR-s4-P)23.120.75560.18340.126434.584.230.765275.020.6797
+ +![](images/2c3028cfcb743fff74f7a6356ea28e65323f5c5729ef200639eb3a893f505b25.jpg) + +![](images/788c989da4dd93507cf26c7b6566ec10f07956a0860a3515768c4e3ffa40c9a8.jpg) + +![](images/abcbebd594ea690d4515dee5b25a00e0058b1e1970186f6252ead8ce3e89e594.jpg) + +![](images/ff215ac84d240e2c231d7bfdf020c38af8c4b15dbd235ee04814c250c5ec8aaa.jpg) + +![](images/583405567c07974abd98ce35e746463b01ea5384d288318e0fadab760aed3eec.jpg) + +![](images/8c037124103a9d397384dc6b31f58c4c7654b8ccf2fceb85ddfd976efd6438e1.jpg) +Figure 1: Visual comparisons on the classical image SR task (Please zoom in to see details). + +![](images/abce892cf94003999ab37301baaba1f195f55a4904bb2de4955fe213a8474227.jpg) + +![](images/3cf263da2e6ef1722c56516d1fdc661c2aba8c493ed60d1872c599342af986ac.jpg) + +![](images/4f777ae31523a832a6e236df3b4df7339703796427f7b0a3cd4df2ce1b4ffc6c.jpg) + +![](images/77408508bb67a6618c3ad1f0e373a8034a1590254faae6cb0cbe28832b39c14b.jpg) + +![](images/6bbf21b08cfd642231534af4f21ddd84ca2d124a3a6f31f361feecbe8990e180.jpg) +DiffBIR + +![](images/a79b766c8c83dfdc12707f0aab7c7e42831ef0dfdd19875ce69110432ed42f6d.jpg) +AgenticIR + +![](images/552449de139670e6e929ba42adccd76844aab8d63c166778485f20302f714c3c.jpg) +4KAgent (GenSR-s4-P) + +![](images/950e6a964728e522267147519569bf7a7522625ad5b9d9b67551fe718d764dbf.jpg) +HQ + +details than these methods. For instance, it faithfully reproduces the fine stripes on tree bark in the top row and the intricate structure of antlers in the bottom row. + +**Discussions** In the context of classical image super-resolution (SR), fidelity-based methods prioritize reconstruction accuracy, measured by PSNR and SSIM, resulting in outputs that often appear overly smooth or blurred. In contrast, perception-based methods optimize for high perceptual quality, reflected in metrics like NIQE, CLIPIQA, MUSIQ, and MANIQA, though often at the expense of fidelity. For example, diffusion-based approaches (e.g., DiffBIR) may hallucinate rich but unrealistic textures. AgenticIR, while capable of generating sharper details, sometimes introduces color shifts or artifacts that undermine visual plausibility. 4KAgent offers configurable flexibility through its profile system, allowing it to operate either as a fidelity-based system (ExpSR-s4-F) or as a perception-based system (ExpSR-s4-P). Quantitatively, 4KAgent delivers competitive PSNR and SSIM scores under the fidelity-based profile, and achieves leading performance in perceptual metrics (NIQE, CLIPIQA, MUSIQ, MANIQA) under the perception-based profile. Qualitatively, 4KAgent consistently produces images with rich, realistic details. The flexibility of 4KAgent allows it to strike a superior balance: it can be easily tuned for maximum visual fidelity or for maximum perceptual appeal without extra training or adaptation, which avoids the common drawbacks of existing SR systems. + +# C.2 Real-World Image Super-Resolution + +Settings In this section, we present the detailed analysis of 4KAgent in the real-world image super-resolution task by presenting detailed experiment results on both the RealSR and DRealSR datasets, as well as visual comparisons. + +Quantitative Comparison Experiment results on real-world image super-resolution datasets are shown in Tab. 5. For the real-world image super-resolution task, we concern more about perceptual metrics, such as NIQE, CLIPIQA, MUSIQ, and MANIQA. Real-world image SR methods have achieved promising results on these metrics. AgenticIR, which contains the DiffBIR in its toolbox, outperforms DiffBIR in most perceptual metrics, proving that agentic systems have better potential in solving the real-world SR problem. 4KAgent goes a step further and outperforms AgenticIR in most metrics, achieving better perceptual quality with better fidelity (PSNR and SSIM), regardless of profile setting. In addition, 4KAgent sets a new state-of-the-art performance on perceptual metrics. + +Table 5: Quantitative comparison on real-world image super-resolution benchmarks (RealSR and DrealSR). The top three performances of each metric are marked in bold, underline, italic respectively. + +
DatasetMethodPSNR↑SSIM↑LPIPS↓DISTS↓FID↓NIQE↓CLIPIAQ↑MUSIQ↑MANIQA↑
RealSRResShift [209]26.310.74110.34890.2498142.817.270.545058.100.5305
StableSR [167]24.690.70520.30910.2167127.205.760.619565.420.6211
DiffBIR [105]24.880.66730.35670.2290124.565.630.641264.660.6231
PASD [199]25.220.68090.33920.2259123.085.180.650268.740.6461
SeeSR [184]25.330.72730.29850.2213125.665.380.659469.370.6439
SinSR [175]26.300.73540.32120.2346137.056.310.620460.410.5389
OSEDiff [183]25.150.73410.29210.2128123.505.650.669369.090.6339
PiSA-SR [153]25.500.74170.26720.2044124.095.500.670270.150.6560
AgenticIR [235]22.450.64470.37450.2503140.385.810.650665.870.6210
4KAgent (ExpSR-s4-P)24.600.68390.32530.2292127.645.090.707870.970.6602
4KAgent (GenSR-s4-P)22.550.65570.35090.2468134.634.780.666671.770.6564
DrealSRResShift [209]28.450.76320.40730.2700175.928.280.525949.860.4573
StableSR [167]28.040.74600.33540.2287147.036.510.617158.500.5602
DiffBIR [105]26.840.66600.44460.2706167.386.020.629260.680.5902
PASD [199]27.480.70510.38540.2535157.365.570.671464.550.6130
SeeSR [184]28.260.76980.31970.2306149.866.520.667264.840.6026
SinSR [175]28.410.74950.37410.2488177.057.020.636755.340.4898
OSEDiff [183]27.920.78350.29680.2165135.296.490.696364.650.5899
PiSA-SR [153]28.310.78040.29600.2169130.616.200.697066.110.6156
AgenticIR [235]23.060.61450.47750.2973182.026.110.654263.590.5927
4KAgent (ExpSR-s4-P)26.000.65350.42570.2717170.195.510.716767.720.6397
4KAgent (GenSR-s4-P)23.110.61260.45790.2866178.364.650.709269.300.6219
+ +Qualitative Comparison For visual comparison, we select four leading real-world image superresolution methods (StableSR, DiffBIR, SinSR, OSEDiff) as well as one agentic system (AgenticIR) as baselines. The visual results are presented in Fig. 2. While previous methods are able to recover rich details from the LQ image, their results often lack realism and fidelity. For example, in the top + +row, OSEDiff reconstructs clothing that appears more like jackets, whereas the HQ reference image shows down jackets. 4KAgent produces sharper and more realistic details, such as the texture of the down jacket in the top row and the clarity of the number '27' in the bottom row. + +![](images/fc472c2d016207137a87a82de25fe21f4a6ff4b16ff200922a9fd943506b4e4f.jpg) + +![](images/8e524e950b6c529aad167492f799012b601bc70457b2da6f67fae9dd1b13f7b2.jpg) +LO + +![](images/2fc679a7b8b10d62384f99a15938ba75afa645730bd402ab15b6c034539bd965.jpg) +StableSR + +![](images/be18b6fe9c75aa361a697669f54c57837b40e25749d5fe6f51a57237dea2fa72.jpg) +DiffBIR + +![](images/2bc9f94da78d1e83c72202f4c5a86ac6b6d47bf066b5f84dfd5ec924e26aefc9.jpg) +SinSR + +![](images/b1bdb7b2663a921cf4f9e75c7bc482ea39cd1de85577569abfe6e94ae0e011cb.jpg) +OSEDiff + +![](images/26cb9fd6d6e469dd0efe570460b98b70eff516fb0aed0d89c3a7c017e21cc43e.jpg) +AgenticIR + +![](images/41b29b9585b2705ea57d48cc55bffc953f829b8f4e906d7808ee9ad1b6697007.jpg) +4KAgent (GenSR-s4-P) + +![](images/4a4001c5bf3cc282e13552e1b4f78f6c1c88ee89c7afc29133e19bed9864cecd.jpg) +HQ + +![](images/77a1854b7a99af18e5ba69b7c47bcae5a8e8b608fc2aa8f993c36b30124a426f.jpg) +Figure 2: Visual comparisons on the real-world image SR task (Please zoom in to see details). + +![](images/16ae9587c3c8e07fc59f7834ee6a462cb02c2fc87a3b14b150dc927ed0dd0df0.jpg) +LO + +OSEDiff + +![](images/947a20d06d688d3b28a76955121f59048b6fecccbb0215ef34792755d018cdc6.jpg) +StableSR + +![](images/f313861e1fa3bed6b7a3cc75715d89a227696329d0b31641a8bf53e3b03f493e.jpg) +DiffBIR + +![](images/0841fe1ea30b10e16ae3b5c952983cc8ce67dc352e02b16b47b26480cb2e499f.jpg) +SinSR + +OSEDiff + +![](images/4156e6ff8c630efcd605fbe3fcca03170830a65107719c7b56f7a7185141944e.jpg) +AgenticIR + +![](images/5f2fd2e407314d04d7711712488fef47c921ad5392d4cc26531f419724f466f6.jpg) +4KAgent (GenSR-s4-P) + +![](images/8f77f787fd26493579bbcd42da2454e877317bc9708bb799e3ecf43843a3b8a3.jpg) +HQ + +**Discussions** The real-world image super-resolution task is more challenging than the classical image super-resolution task as it contains more complex distortions than the synthetic downsample, which can also be seen from the comparison of the LQ image and HQ image in the dataset. Under this challenging setting, agentic systems prove their advantage by analyzing the distortion and restoring the image properly. 4KAgent further proves its superiority by consistently outperforming AgenticIR in most quantitative metrics. In particular, 4KAgent sets a new state-of-the-art performance for no-reference perceptual metrics, demonstrating that its design effectively elevates perceived realism. Qualitatively, these gains translate into visibly sharper and more believable details. By dynamically leveraging multiple SR experts and selecting the optimal result, 4KAgent shows its superiority in the challenging real-world image super-resolution task. + +# C.3 Multiple-Degradation Image Restoration + +Settings In this section, we present detailed experimental results on Group A, B, and C test sets from the MiO100 dataset [76] with more visual comparisons. + +Quantitative Comparison Experimental results are shown in Tab. 6. In the multiple-degradation image restoration (IR) task, agentic systems once again prove their superiority, outperforming all-in-one methods in all metrics. Among the agentic systems, 4KAgent performs the best, achieving a new state-of-the-art performance on PSNR, MANIQA, CLIPIQA, and MUSIQ. Specifically, for no-reference perceptual metrics (MANIQA, CLIPIQA, MUSIQ), 4KAgent outperforms all compared methods by a noticeable margin (e.g., 4.2 lead of MUSIQ on Group C). For SSIM and LPIPS metrics, 4KAgent remains competitive, ranking among the top two on Group A and Group C subsets. + +Qualitative Comparison For visual comparison, we select two leading all-in-one methods (DA-CLIP, AutoDIR) as well as an agentic system (AgenticIR) as baselines. Visual comparisons are shown in Fig. 3, all-in-one methods perform limited under this setting, especially when restoring complex distortions, such as raindrops. AgenticIR achieves promising results, proving the potential of agentic systems in dealing with complex distortion tasks. 4KAgent goes a step further, generating + +Table 6: Quantitative comparison of multiple-degradation image restoration tasks on three subsets (Group A, B, and C) from the MiO100 dataset. The top three performances of each metric are marked in bold, underline, italic respectively. + +
DegradationsMethodPSNR↑SSIM↑LPIPS↓MANIQA↑CLLIQA↑MUSIQ↑
Group AAirNet [90]19.130.60190.42830.25810.393042.46
PromptIR [140]20.060.60880.41270.26330.401342.62
MiOIR [75]20.840.65580.37150.24510.393347.82
DA-CLIP [117]19.580.60320.42660.24180.413942.51
InstructIR [32]18.030.57510.44290.26600.352845.77
AutoDIR [66]19.640.62860.39670.25000.376747.01
AgenicIR [235]21.040.68180.31480.30710.447456.88
MAIR [65]21.020.67150.29630.33300.475159.19
4KAgent (GenMIR-P)21.480.67200.30190.37480.554463.19
Group BAirNet [90]19.310.65670.36700.28820.427447.88
PromptIR [140]20.470.67040.33700.28930.428948.10
MiOIR [75]20.560.69050.32430.26380.433051.87
DA-CLIP [117]18.560.59460.44050.24350.415443.70
InstructIR [32]18.340.62350.40720.30220.379050.94
AutoDIR [66]19.900.66430.35420.25340.398649.64
AgenicIR [235]20.550.70090.30720.32040.464857.57
MAIR [65]20.920.70040.27880.35440.508460.98
4KAgent (GenMIR-P)20.950.67270.30170.37340.550562.69
Group CAirNet [90]17.950.51450.57820.18540.311330.12
PromptIR [140]18.510.51660.57560.19060.310429.71
MiOIR [75]15.630.48960.53760.17170.289137.95
DA-CLIP [117]18.530.53200.53350.19160.347633.87
InstructIR [32]17.090.51350.55820.17320.253733.69
AutoDIR [66]18.610.54430.50190.20450.293937.86
AgenicIR [235]18.820.54740.44930.26980.394848.68
MAIR [65]19.420.55440.41420.27980.423951.36
4KAgent (GenMIR-P)19.770.56290.42710.35450.523355.56
+ +![](images/47e23a9a88a2144679c1dba907e72bb1ddf0a406a08c789053bed25b44935517.jpg) + +![](images/963c2e93a4afe7781a5da23d9e18f28b81b1f897faa1c0c830db639f1e06741d.jpg) + +![](images/84b1d731df8b15b95c14bd5b89837fbdab11537511232443f034dbf37f8e50aa.jpg) + +![](images/4a20e89f3ccaad0418287cd0144ff2925e831273f786420bf86b7cb59168f4b4.jpg) + +![](images/086485be69533ee73fbc8670de78d129e83cbe571669bd5e1e86ee5d8cf54b37.jpg) +Figure 3: More visual comparisons on MiO100 dataset (Please zoom in to see details). + +![](images/bd0a275d92a0b6bf4a7f38e1fbdc7c9cf3fe49f552be85c3092ef5d098b7c963.jpg) +LQ + +![](images/91fab0375b54a88df7dfbb44b578a274630f5f8d02c1227ac755627af26413bc.jpg) +DA-CLIP + +![](images/5e12b6ffa0dbff1650ebb28bc8123f3e35eb4d03c9849850bd4b12480f5ca103.jpg) +AutoDIR + +images with high-grained details and more consistent with the high-quality (HQ) reference image. For instance, the natural stripes on the tree trunks and the fine leaf textures in the top row, as well as the intricate waterfall ripples and mountain contours in the bottom row. + +**Discussions** Concluding from experiment results, agentic systems have shown their superiority in the multi-degradation image restoration tasks, where each low-quality (LQ) image is affected by $2 \sim 3$ types of distortions. Under this challenging setting, 4KAgent exhibits a clear advantage in handling complex, multi-degraded inputs, outperforming both conventional all-in-one methods and previous agentic systems. Quantitatively, 4KAgent achieves state-of-the-art results across multiple metrics, including PSNR, MANIQA, CLIPIQA, and MUSIQ, highlighting its strong capability in enhancing both fidelity and perceptual quality. Qualitatively, this metric superiority translates into more faithful restoration of fine-grained patterns and textures, even under severe and heterogeneous distortions. + +# C.4 Face Restoration + +Settings In this section, we evaluate 4KAgent on the real-world face restoration benchmark, the WebPhoto-test [171] dataset, which contains 407 low-quality face images collected from the Internet. As face restoration pipeline is a module subsequent to the super-resolution task in 4KAgent, we first downsample the images by a factor of 4 to generate low-quality (LQ) images. In this experiment, we configure 4KAgent with the GenSRFR-s4-P profile. + +We compare 4KAgent with state-of-the-art face restoration methods, including CodeFormer [232], GFPGAN [171], and Difface [208], as well as an agentic system AgenticIR [235]. For face restoration methods, we set the scaling factor to 4. As there are no high-quality (HQ) references, we evaluate performance with four no-reference perceptual metrics (NIQE, CLIPIQA, MUSIQ, and MANIQA-pipal) and two advanced face-specific IQA metrics (CLIB-FIQA [135] and DSL-FIQA [24]). + +Quantitative Comparison Experimental results are shown in Tab. 7. AgenticIR performs worse than previous face restoration methods in terms of general perceptual metrics and face IQA metrics. 4KAgent outperforms AgenticIR on every metric by a clear margin (e.g., 19.67 lead on MUSIQ). Moreover, 4KAgent achieves the best scores on general no-reference perceptual metrics and delivers competitive results on face IQA metrics, ranking second in both CLIB-FIQA and DSL-FIQA. + +Table 7: Quantitative comparison on face restoration benchmark (WebPhoto-Test). The top three performances of each metric are marked in **bold**, **underline**, **italic** respectively. + +
DatasetMethodNIQE↓CLIPIQA↑MUSIQ↑MANIQA↑CLIB-FIQA↑DSL-FIQA↑
WebPhoto-TestGFPGAN [171]5.120.679274.210.63790.65900.7732
CodeFormer [232]4.580.688473.870.64150.68400.7435
DifFace [208]4.200.583165.310.58910.65110.6189
AgenticIR [235]6.850.573156.250.54650.59780.5289
4KAgent (GenSRFR-s4-P)4.150.707775.920.65760.66710.7683
+ +![](images/95bf92658901884ea0ed5b530d42e3647ea5d1ce155d073ef47f07d0a89bc28c.jpg) + +![](images/abeaf693882c5c8dde7dd142b019272eb2517d142357e0ea0dda0ffbf1275941.jpg) + +![](images/6c2c53c9ff3b7718a9da8a3c12a40a5f1c2e608a1f850d71938997e7f37197d0.jpg) + +![](images/5c09763e0c500268940660a1f886f65c98f9256f75e4c85d03d9d09786bc13f1.jpg) + +![](images/425e5eab2c672195d38edd8dca13dd32cc5deb3187014727d796725eeb8c3b9a.jpg) + +![](images/7406fb19c83b444733cf8969a8e600f4198962135f424251f09d5d146794fc7d.jpg) + +![](images/f23a4d39b81799a407235c5d7a9f748deba45bcc32c86398a71cc1b50cabc6cd.jpg) +LQ +Figure 4: Visual comparisons on the face restoration task (Please zoom in to see details). + +![](images/a06772de4312ae1ba13602fd0775d16e3f8bbe2271c1173d60118bc66293bbef.jpg) +GFPGAN + +![](images/2efdbd939b270e287983269127ecbf518f64b80de7a9a1025ed84d7371d2e33b.jpg) +CodeFormer + +![](images/3c4790bce1af269ba709518c3c81a0da349075d8f7ca33c4057237ae41e0cd48.jpg) +Difface + +![](images/bd7e0072054b1abf45820dceffac773a9232c4b4f03bc716c80e9744a61e39e7.jpg) +AgenticIR + +![](images/e79f0ba0f9fe557b20c4b2ca031f2fa997131936e5b445bcad885c9a675f42b8.jpg) +4KAgent (GenSRFR-s4-P) + +Qualitative Comparison Visual comparisons are shown in Fig. 4. Compared with other methods, 4KAgent demonstrates a clear advantage in restoring realistic facial details, such as fine hair strands + +and natural skin textures. Moreover, it achieves superior restoration performance in non-facial regions, such as the wall and leaves in the first row and the logo on the hat in the second row. By consistently delivering high-quality restoration in both facial and non-facial areas, 4KAgent produces more visually pleasing and perceptually balanced results overall. + +**Discussions** In the face restoration scenario, 4KAgent effectively addresses both facial and contextual degradations. Quantitatively, 4KAgent achieves the best performance on general no-reference perceptual metrics and delivers competitive scores on face IQA metrics, demonstrating the superiority of its system design and face Q-MoE policy. Qualitatively, this translates into more natural and richly detailed facial features, such as individual hair strands and realistic skin texture, while also enhancing background elements, producing overall more visually pleasing outputs. Among agentic systems, AgenticIR applies a uniform processing pipeline without a dedicated face restoration module, which limits its performance on face restoration tasks. Benefit from the face restoration pipeline and profile design, 4KAgent can be tailored as a face restoration expert, achieving superior results. We further present how these two designs enhance the face restoration ability of 4KAgent in the ablation study. + +# D Experiment Part II: $16 \times$ Natural Image Super-Resolution + +Experiment Part I (§C) demonstrates the flexibility and superiority of 4KAgent on general image restoration and super-resolution tasks. In this section, we evaluate 4KAgent on more challenging restoration tasks. First, we assess its performance on the $16\times$ real-world image super-resolution task. Next, we evaluate 4KAgent on our proposed DIV4K-50 dataset. + +# D.1 Large Scale Factor $(16\times)$ Image Super-Resolution + +Settings In the main paper, we present visual comparisons between 4KAgent and state-of-the-art image super-resolution methods (e.g., HAT-L, DiffBIR) and agentic systems (e.g., AgenticIR) on the RealSRSet dataset under a large-scale factor $(16\times)$ upscaling setting. Here, we extend our experiment with detailed quantitative evaluations with more methods and additional qualitative examples. As there are no corresponding high-quality (HQ) reference images in the dataset, we evaluate the result images on four no-reference perceptual metrics: NIQE, CLIPIQA, MUSIQ, and MANIQA-pipal. + +Quantitative Comparison Experiment results are shown in Tab. 8. As the largest scale factor of the pre-trained model in HAT-L is 4, we apply the $4 \times 4 \times$ setting for HAT-L for $16 \times$ upscaling. Fidelity-based method struggles to deliver satisfactory performance in perceptual metrics under this setting. Recent perceptual-based real-world image super-resolution methods perform well in these metrics, even with the $16 \times$ setting. For example, DiffBIR with $16 \times$ setting achieves the best NIQE and MANIQA scores. Among agentic systems, 4KAgent outperforms AgenticIR on every metric by a clear margin (e.g., 6.13 lead on MUSIQ). In addition, 4KAgent achieves the best performance on MUSIQ and the second-best performance on NIQE. For CLIPIQA and MANIQA metrics, 4KAgent also delivers competitive performance, ranking among the top three across all methods. + +Table 8: Quantitative comparison on RealSRSet dataset under $16\times$ upscaling. The top three performances of each metric are marked in bold,underline,italic respectively. + +
DatasetMethodNIQE↓CLIPIQA↑MUSIQ↑MANIQA↑
RealSRSetHAT-L [26] (4×→4×)10.590.388525.060.3060
DiffBIR [105] (4×→4×)3.630.786744.860.6076
DiffBIR [105] (16×)2.800.758347.540.6099
OSEDiff [183] (4×→4×)5.400.766548.420.5362
OSEDiff [183] (16×)4.660.648335.330.4581
PiSA-SR [153] (4×→4×)5.700.788348.200.5464
PiSA-SR [153] (16×)4.880.638435.900.4128
AgenticIR [235]4.860.677544.710.5236
4KAgent (Gen4K-P)3.530.779450.840.5913
+ +Qualitative Comparison For visual comparison, we select three representative methods to benchmark against 4KAgent: (1) HAT-L $(4\times \rightarrow 4\times)$ : As a representative fidelity-based method, we + +![](images/a16513b6ad71c39d814cd342f66bf2fc8c63e796b9bf1daa9a0517f3c3c54bc2.jpg) + +![](images/ebe34b33caeba8fd8b311ea6a87c08c88c9f9d8f99fade01783e6bbb71b20b46.jpg) + +![](images/3b01b11af5901535dedae5384c3ee704bb41a5151da9d16218ee86d02d781d03.jpg) + +![](images/ed7a00f092e3711a9829e404e776036b5b8a75514ad62262b392eb0d7f492dd3.jpg) + +![](images/328ef1cdf51853f3b1b3ca25c3a266bb134727eb0f9ae19bbd170abbcd3b99d7.jpg) + +![](images/89ad06b909560c18dc2e57b0cf4708f239add2e628cc1518eadb962c648496bf.jpg) + +![](images/2d9277a492931b3dd39f359a0ca20039652c9f8b0051bdbb2029221b366e686e.jpg) +Figure 5: Visual comparisons on RealSRSet dataset (16× upscaling) (Please zoom in to see details). + +![](images/ad2aa79ffdcc1d2eac92105a7e5a2397b18bf57c0a1fb36e62f66fd51c83dff2.jpg) +LQ + +![](images/d2108b29afaf88285aa85c45bcff766c70595424d6b9fe2d3dd58cbeb2fdfe67.jpg) +HAT-L $(4\times \rightarrow 4\times)$ + +![](images/b83cda4f3290e41b1f680b32b25cd5c8326416c54bd9e86a3e4a178e0687ceff.jpg) +DiffBIR $(16\times)$ + +![](images/084a58418472c17adb14d7a506e2d1c3a8117f7d3f87f5554cb139ec2b25a862.jpg) +AgenticIR + +![](images/ea331b4f105a57a5458084ae538b730c3bc8d2fbab9fe074af2346fa14498be0.jpg) +4KAgent (Gen4K-P) + +investigate its performance under a large-scale upscaling setting. (2) DiffBIR $(16\times)$ : As shown in Tab. 8, DiffBIR with $16\times$ setting achieves the best performance on NIQE and MANIQA. Therefore, we include it to assess its visual quality. (3) AgenticIR: Selected for agentic system comparison. Visual comparisons are shown in Fig. 5. + +HAT-L $(4\times \rightarrow 4\times)$ shows limited enhancement over the low-quality input, leading to notably blurred textures. DiffBIR $(16\times)$ produces visually rich but often unrealistic hallucinations, and in some cases even alters the semantic content of the scene (e.g., the first row), which is visually unappealing. AgenticIR generates visually plausible results but lacks sufficient sharpness and fine-grained details. 4KAgent generates high-grained and realistic details: the rock and grass textures in the first row, and the hair strands, eyebrow patterns, and naturally expressive eyes in the second and third rows are all much more faithfully restored with high-grained details. + +**Discussions** In the challenging $16 \times$ upscaling scenario, 4KAgent delivers competitive quantitative results alongside high-grained and realistic qualitative results, compared to other methods. In addition, traditional fidelity-oriented methods such as HAT-L struggle to recover fine details and instead produce overly smoothed and blurred results. This highlights the limitation of fidelity-driven pipelines under extreme magnification levels. Therefore, for such high-scale upscaling tasks, it is essential to configure 4KAgent with a perception-oriented profile (e.g., setting Restore Option to Perception in the profile module) to better prioritize realistic texture synthesis. As image resolution approaches 4K and beyond, existing no-reference perceptual metrics may become misaligned with human judgment of visual quality. This discrepancy underscores the need for developing new no-reference perceptual metrics specifically designed for ultra-high-resolution images. + +# D.2 Joint restoration & 4K Upscaling + +Settings In this section, we bring 4KAgent to the most challenging setting: Joint multiple image restoration and 4K upscaling. As there are no previous methods and datasets targeted at this setting, we propose a new evaluation dataset, DIV4K-50, constructed from the Aesthetic-4K dataset [218] to rigorously test end-to-end restoration and ultra-high-scale SR. In this experiment, we evaluate 4KAgent on DIV4K-50 and we configure 4KAgent with the Gen4K-P profile. Comparing methods and experimental settings are the same as in Appendix D.1. + +Quantitative Comparison Quantitative comparisons are shown in Tab. 9. Similar to the experiment result in Appendix D.1, real-world image super-resolution methods perform competitively on perceptual metrics under this challenging setting. For example, DiffBIR achieves the best score on NIQE and CLIPIQA metrics. For agentic systems, 4KAgent outperforms AgenticIR on every metric. Additionally, 4KAgent achieves the best performance on MUSIQ and MANIQA metrics, and the second-best performance on NIQE and CLIPIQA metrics. + +Qualitative Comparison As shown in Appendix D.1, directly upscaling images with the $16\times$ setting often produces visually rich but unrealistic artifacts. To ensure a fair and meaningful qualitative comparison in this experiment, we select previous methods with the $4\times \rightarrow 4\times$ setting, along with + +Table 9: Quantitative comparison on DIV4K-50 dataset. The top three performances of each metric are marked in bold, underline, italic respectively. + +
DatasetMethodNIQE↓CLIPIQA↑MUSIQ↑MANIQA↑
DIV4K-50HAT-L [26] (4× → 4×)11.860.469922.820.3270
DiffBIR [105] (4× → 4×)3.360.758837.170.5916
DiffBIR [105] (16×)2.650.707838.590.5858
OSEDiff [183] (4× → 4×)4.880.720139.880.5482
OSEDiff [183] (16×)8.370.568025.070.4210
PiSA-SR [153] (4× → 4×)5.010.714138.220.5364
PiSA-SR [153] (16×)9.300.554924.510.3861
AgenticIR [235]5.130.561439.550.4814
4KAgent (Gen4K-P)3.150.758544.160.5928
+ +![](images/64521ad7a2a4d13440dda374d8c067326f4d70262430c6a239e26de07131d183.jpg) + +![](images/40f5afa87abb82fc0361fac64c13656d3a7350ffb6aff50ef05a46b0a1b1c925.jpg) +LQ + +![](images/7cb07c7e6aca7453726d528d81f889c38660f83d7512ca94e32bcd2069ff3751.jpg) +HAT-L $(4\times \rightarrow 4\times)$ + +![](images/c04fecb5048fc7b2870ad3df9df79056c1220ee04e7079130d416b027d5d0808.jpg) +DiffBIR $(4\times \rightarrow 4\times)$ + +![](images/c2c53ced41f60d6f49e4a91abbf2eaddd21665ee78556405cc957bfab0319325.jpg) +OSEDiff $(4\times \rightarrow 4\times)$ + +![](images/86096999d6291ed29b24c3bbe3f37132798786f45cee6238591ce7d13f321d46.jpg) + +![](images/c28d5fea78c7f55a696ed00ce756faa2cefafd1e510b39c66525b3b4050ab22e.jpg) +LQ + +![](images/b6ea30d00e54af494c5868fb4d0a5498e3da883511b255333240afb2a1e7204c.jpg) +HAT-L $(4\times \rightarrow 4\times)$ + +![](images/6d4d5014e182571a7924cdb81fa03a29a955c710cc0e34784d47fa2a21fe101d.jpg) +DiffBIR $(4\times \rightarrow 4\times)$ + +![](images/237da557c1092a477ed2a1c3201a2e91573c76a1f1ac2fec31f7b17b4dab5da1.jpg) +OSEDiff $(4\times \rightarrow 4\times)$ + +![](images/8352ec95e28ededb6c41ae510885d32b0f95aad9cad2239d8d77602d655f1537.jpg) +PiSA-SR $(4\times \rightarrow 4\times)$ + +![](images/fc2a685071b363873097f770659b57df268e7fb37180a768078b99f454bb2448.jpg) +AgenticIR + +![](images/8248a41a904e7e66693f8fb7b82a180704bf5939ae0735ded94caacecdefcdd0.jpg) +4KAgent (Gen4K-P) +Figure 6: More visual comparisons on DIV4K-50 dataset (Please zoom in to see details). + +![](images/c4e549b096cd7fe932257d77ec9731b72cec338d8078134735e142c3045747d0.jpg) +HQ + +AgenticIR, as baselines. Qualitative comparisons are shown in Fig. 6. Real-world image super-resolution methods generally recover more details than fidelity-based method. However, their outputs still exhibit noticeable distortions. For example, the generated patches from OSEDiff and PiSA-SR in the middle row retain visible JPEG compression artifacts, which degrade the overall visual quality. While DiffBIR achieves the most favorable visual results among these methods, its outputs still suffer from either blurring or unrealistic artifacts. AgenticIR performs competitively but tends to produce insufficiently sharp details. 4KAgent consistently reconstructs finer and more natural details. Notable examples include the facial features in the top row, the bears fur in the middle row, and the intricate coral textures in the bottom row, highlighting the superiority of our method. Figure 6 presents additional visual comparisons: even in this challenging setting, 4KAgent faithfully reconstructs finer, more natural details, such as the bears fur in the top row and the intricate mountain textures in the bottom row, highlighting the superiority of our method. + +Discussions The DIV4K-50 benchmark presents an extremely challenging setting, where fully recovering low-quality $256 \times 256$ inputs to match the $4096 \times 4096$ ground-truth images is virtually unattainable. Therefore, our focus shifts towards generating richly detailed and visually authentic + +textures. Existing real-world image super-resolution methods struggle to fully correct the compounded degradations present in challenging scenarios. While these methods achieve competitive scores on perceptual metrics, their outputs tend to suffer from unrealistic hallucinated textures. Prior agentic systems, despite their effectiveness in handling multiple degradations, show limitations in maintaining sufficient sharpness when upscaling to 4K resolutions. 4KAgent demonstrates its capability to simultaneously address multiple degradations and extreme scaling factors, effectively reconstructing natural, fine-grained details with high visual fidelity. + +# E Experiment Part III: AI-Generated Content (AIGC) 4K Super-Resolution + +Text-to-visual models have ushered in a new era of high-quality image synthesis in AI-generated content. While existing models exhibit impressive capabilities in interpreting and following complex user instructions, their limitation to relatively low-resolution outputs (e.g., $1024 \times 1024$ ) poses significant challenges for applications requiring ultra-high visual fidelity, such as digital content creation and cinematic production. Scaling diffusion models for high-resolution generation entails computational overhead and access to large-scale high-resolution training data. As a practical alternative, pre-trained diffusion models can be repurposed for ultra-high-resolution image generation. + +# E.1 AI-Generated Content 4K Super-Resolution Experiment + +In this section, we present comprehensive experimental results comparing the effectiveness of end-to-end generation of ultra-high-resolution images with upscaling 1K images to 4K using our 4KAgent. The comparisons are conducted on two curated datasets: GenAIBench-4K and DiffusionDB-4K. + +Settings We curated a set of 200 prompts sampled from two widely-used AIGC benchmarks [179, 87]. For each prompt, 1K-resolution images were generated using several representative text-to-image models, including Flux.1-dev [79], Stable Diffusion 3 (SD3) [42], PixArt- $\Sigma$ [21], SANA [195], and GPT-4o [59]. In parallel, we employed native 4K-capable models such as SANA [195] and Diffusion4K [218] to directly synthesize 4K-resolution outputs. Due to more stringent safety protocols, GPT-4o yielded only 39 valid 1K-resolution images from DiffusionDB prompts. + +We use the ExpSR-s4-P profile in 4KAgent here. To assess perceptual quality, we employ no-reference perceptual metrics. However, we observe that these metrics — particularly MUSIQ — are not tailored for evaluating ultra-high-resolution images, likely due to their inability to capture fine-grained details through a multi-scale architecture. To mitigate this limitation, we introduce MUSIQ-P, a patch-applied variant that computes MUSIQ scores over non-overlapping $512 \times 512$ patches and averages them, thereby improving sensitivity to localized artifacts in ultra-high-resolution content. + +Table 10: Comprehensive quantitative comparison of AIGC $4 \times$ Super-Resolution. The top three performances of each metric are marked in **bold**, **underline**, **italic** respectively. MUSIQ- $\mathbf{P}^*$ indicates a patch-applied variant of the MUSIQ metric for evaluating ultra-high-resolution (4K) images. + +
DatasetGenAIBench-4K [87]DiffusionDB-4K [179]
ModelNIQE↓CLIPIQA↑MUSIQ-P*↑MANIA↑NIQE↓CLIPIQA↑MUSIQ-P*↑MANIA↑
SANA-4K [195]4.020.617247.930.36733.740.600548.660.3425
Diffusion-4K [218]6.380.504935.070.35356.550.505635.870.3404
SANA-1K [195]4.180.714766.300.48143.800.691067.990.5104
+ 4KAgent3.030.705057.970.47353.040.708260.480.4715
GPT4o [59]5.690.660764.430.49975.130.627562.530.4398
+ 4KAgent3.560.701658.280.49763.400.686756.670.4711
FLUX.1-dev [79]6.180.676861.020.50185.330.750969.690.5835
+ 4KAgent2.980.707858.190.50343.040.744060.880.5056
PixArt-Σ [21]4.120.696063.740.44153.660.689266.540.4386
+ 4KAgent2.760.707756.710.46992.880.709258.850.4659
SD3-Medium [42]5.030.692264.680.47674.380.666765.990.4413
+ 4KAgent2.990.716960.220.51552.990.706659.350.4747
+ +Quantitative Comparison. Tab. 10 presents the quantitative results across three strategies: (1) native 4K generation, (2) 1K-resolution generation, and (3) 1K-resolution images upscaled by 4KAgent. On GenAIBench-4K, the SANA-1K + 4KAgent pipeline achieves a NIQE score of 3.03 and a CLIPIQA of 0.7050, significantly outperforming SANA-4K (NIQE 4.02, CLIPIQA 0.6172). Similarly, PixArt- $\Sigma$ + 4KAgent obtains the best NIQE score (2.76) among all methods, while SD3-Medium + 4KAgent achieves the best CLIPIQA (0.7169) and MANIQA (0.5155) scores. On DiffusionDB-4K, several models such as SANA-1K, GPT-4o, PixArt- $\Sigma$ , and SD3-Medium, when upscaled with 4KAgent, achieve significantly lower NIQE and higher CLIPIQA scores. Although MUSIQ-P scores for the upscaled images show a slight decrease relative to their 1K counterparts, the difference remains marginal, suggesting limited perceptual degradation during upscaling. + +To further assess semantic and aesthetic fidelity, we report PickScore [73] in Tab. 11, which quantitatively captures diversity and human-aligned visual quality. On GenAIBench-4K, models enhanced with 4KAgent outperform their native 4K counterparts. On DiffusionDB-4K, the performance gap is smaller, which may be attributed to the dataset's richer and more descriptive prompt content. + +Table 11: Comparison of PickScore-Based [73] Quantitative Evaluation Between Native 4K and 4KAgent-Upscaled 1K Models. + +
DatasetAvg. Prompt LengthSANA-4KSANA-1K + 4KAgentDiffusion-4KFlux.1-dev + 4KAgent
GenAIBench-4K [87]12.130.44820.55180.23890.7611
DiffusionDB-4K [179]25.290.48930.51070.24060.7594
+ +![](images/50065d103640892e79eb0eea9b0994a13bec91a0a7ab212e2985dab514630a13.jpg) +A magician pulling a rabbit out of a hat. + +![](images/cf7a2390da5ffb16a9e4fc313e8ef0a53c318c0d9a78d352e067fbfa14b09f73.jpg) +An action photograph of a cat casting fur out if it's mouth. + +![](images/ebef88553eb60c0ed4c68710e6c713f9ca82c6251b69ba2d0013b68ff08d85ce.jpg) + +![](images/0474bbf5529151b42c5cf8d45742899096dd66b9d0bdbbd04642c17aa659d3a8.jpg) + +![](images/e40d4c25ea1f34e9a2622fbda44284be1a55980e0a0ee3b63fcc7abcd7364629.jpg) +Hyperrealistic portrait of a philippine character in a scenic environment, flowers, rain, war, by beksinski zdzislaw, yoro sean, buchholz, quint, christensen james c., yellow palette. +Figure 7: Visual comparison between native 4K image generation and 1K image generation methods with 4KAgent, using identical prompts. 4K images from Diffusion-4K and SANA-4K are displayed on the left, while the corresponding outputs enhanced by 4KAgent are shown on the right. + +Qualitative Comparison Fig. 7 shows qualitative results of applying 4KAgent to various base models under identical prompts. Across different models, 4KAgent consistently enhances visual fidelity and preserves fine-grained details. As shown in Fig. 8, images upscaled from SANA-1K using 4KAgent exhibit richer textures and stronger aesthetic alignment than those generated natively at 4K resolution by SANA-4K. + +![](images/0749317dd9b9044222762269760522171dc69e4f782482f14cb53e42ac4f60bd.jpg) +SANA-4K + +![](images/f76dc764aebdf1b40cd98f455acd7f7ddd680c23f4268b0dce9f14dbbdc27451.jpg) +SANA-1K + 4KAgent +Figure 8: Visual comparison of aesthetic preference alignment between SANA-4K and SANA-1K+4KAgent using identical prompts sampled from GenAIBench-4K. SANA-1K+4KAgent yields superior aesthetic alignment and richer high-resolution details, highlighted in the zoomed-in patches. + +**Discussions** The application of our 4KAgent in AIGC scenarios leads to substantial improvements in image quality when upscaling 1K-resolution images to 4K. First, when applied to 1K-resolution inputs, 4KAgent consistently achieves notable gains across multiple quantitative benchmarks, enabling more detailed and accurate reconstructions in the resulting 4K outputs. As traditional metrics are not specifically tuned for ultra-high-resolution images, we adopted the adaptive MUSIQ-P metric to enable a perceptually-focused evaluation. The results indicate that 4K-upscaled images achieve perceptual quality scores comparable to their original 1K counterparts. Second, 4KAgent demonstrates strong capability in synthesizing high-fidelity visual details and intricate textures. However, as with many traditional super-resolution methods, this training-free framework occasionally introduces unintended bokeh-like artifacts, particularly in blurred background regions. Given 4KAgent's modular and scalable design, we believe integrating task-specific profile configurations and perceptual alignment strategies could further reduce artifacts and improve robustness in diverse AIGC applications. + +# F Experiment Part IV: Scientific Imagery Super-Resolution + +# F.1 Remote Sensing Image Super-Resolution + +High-resolution satellite imagery is foundational for a wide spectrum of remote sensing tasks, including urban planning, environmental monitoring, and disaster response [53, 146]. However, due to cost, bandwidth, and sensing constraints, acquiring such high-resolution imagery globally at very high frequency remains very expensive or even impractical. Recent advances in deep learning-based super-resolution have provided a promising alternative by reconstructing high-fidelity imagery from lower-resolution observations [77]. In this section, we evaluate 4KAgent against state-of-the-art baselines on a diverse set of real-world satellite image super-resolution datasets. + +Settings We evaluate our models on four benchmark datasets covering varied land-use patterns and sensing characteristics: + +- AID [189] is a large-scale dataset constructed to benchmark aerial scene classification methods. It includes over 10,000 high-resolution aerial images across 30 scene categories, such as airports, industrial areas, and farmlands. Each image has a resolution of $600 \times 600$ with a spatial resolution + +of $0.5 - 8\mathrm{m}$ /pixel. Images exhibit high intra-class diversity and low inter-class variation, making them well-suited for evaluating generalization in SR tasks. + +- DIOR [94] is a comprehensive object detection benchmark in the remote sensing domain, containing 23,463 images and 192,472 annotated object instances across 20 categories. The resolution of images is $800 \times 800$ , and the spatial resolutions range from $0.5\mathrm{m}$ to $30\mathrm{m}$ . These images exhibit high diversity in resolution, imaging conditions, and object scale. +- DOTA [188] consists of 2,806 ultra-high-resolution aerial images collected from various sensors. It features over 188,000 labeled object instances with arbitrary orientations. Each image is in resolution about $4000 \times 4000$ . The combination of large-scale, fine-grained annotations and high inter-scene variability makes DOTA particularly valuable for evaluating perceptual fidelity. +- WorldStrat [33] is a unique dataset designed for real-world satellite image super-resolution tasks, with globally stratified land use coverage across $10,000\mathrm{km}^2$ . It pairs $1054\times 1054$ pixel high-resolution (SPOT 6/7, $1.5\mathrm{m/pixel}$ ) and temporally matched low-resolution (Sentinel-2, $10\mathrm{m/pixel}$ ) imagery for thousands of regions worldwide. Importantly, unlike synthetic degradation benchmarks, WorldStrat contains real cross-sensor-captured low-resolution (LR) and high-resolution (HR) image pairs, introducing natural misalignment and color mismatches due to different sensor characteristics. + +From each dataset, we select 100-200 representative scenes. Following [193, 194], for AID, DIOR, and DOTA, HR images are downsampled using bicubic interpolation to generate corresponding LR inputs. For WorldStrat, we adopt the datasets official pre-processing pipeline, selecting LR images that are temporally closest to each HR acquisition. Notably, due to the different sensors used for LR and HR captures, RGB content may exhibit significant variation, posing a realistic challenge for super-resolution. To test the generalization ability of our models, we evaluate on a spectrum of resolution scales: 1) $4 \times$ SR (128→512) on DIOR and DOTA datasets; 2) $4 \times$ SR (160→640) on AID and WorldStrat datasets; 3) $4 \times$ SR (512→2048) for high-res DOTA scenes; 4) $16 \times$ SR (e.g., 256→4096) for DOTA scenes. + +We evaluate 4KAgent with the AerSR-s4-F profile and AerSR-s4-P profile in $4\times$ super resolution for Fidelity and Perception preference, respectively. Then we evaluate 4KAgent with the Aer4K-F profile and Aer4K-P profile in $16\times$ super resolution for Fidelity and Perception preference, respectively. We benchmark 4KAgent against the following categories of SR models: 1) Expert aerial SR models: HAUNet [168], TransENet [86]; 2) Fidelity-based SR models: HAT-L [26], PiSA-SR-PSNR [153], and SwinIR [101]; 3) Perception-based SR Models: DiffBIR [105], OSEDiff [183], HAT-GAN [26], PiSA-SR [153], and SwinIR (Real-ISR) [101]. For 'PiSA-SR-PSNR', we set the pixel guidance factor $\lambda_{pix} = 1.0$ and semantic guidance factor $\lambda_{sem} = 0$ for PiSA-SR [153] in inference. Additionally, we include AgenticIR for agentic system comparison. Together, these diverse datasets and models enable a comprehensive evaluation of 4KAgent in terms of both pixel fidelity and perceptual quality, across synthetic and real-world degradation settings. + +Quantitative Comparison We report the comparison results on AID (4x SR, 160→640), DIOR (4× SR, 128→512), DOTA (4× SR, 128→512), WorldStrat (4× SR, 160→640), DOTA (4× SR, 512→2048), and DOTA (16× 4K SR) in Tabs. 12 to 17, respectively. Across all six benchmark settings, 4KAgent with Fidelity preference consistently demonstrates superior performance in terms of pixel-level reconstruction. It ranks within the top three PSNR in four out of six tasks, and ranks within the top three in SSIM across all synthetic scenarios. This confirms its ability to preserve structural details across scales and domains. Importantly, 4KAgent also consistently outperforms AgenticIR by a large margin across all fidelity metrics and tasks, highlighting its effectiveness. + +In terms of perceptual quality, 4KAgent with Perception preference achieves top performance in perceptual quality assessment metrics across multiple datasets. Notably, on the WorldStrat, 4KAgent (AerSR-s4-P) ranks first in MUSIQ, MANIQA, and CLIPIQA. These results indicate that 4KAgent not only produces photorealistic outputs but also maintains robustness in real-world, sensor-misaligned scenarios. Compared to AgenticIR, 4KAgent with Perception preference shows clear gains across all perceptual IQA metrics, reaffirming the value of 4KAgent in balancing realism and structure across diverse and challenging remote sensing settings. + +Qualitative Comparison Figs. 9 to 13 present a comprehensive visual comparison of all evaluated models across the tested datasets of $4 \times$ AID, $4 \times$ DIOR, $4 \times$ DOTA, $4 \times$ WorldStrat, and $16 \times$ DOTA. Firstly, 4KAgent with the perception preference consistently delivers superior perceptual quality on + +Table 12: $4 \times$ performance comparison of evaluated models on the AID dataset (160→640). The top three performances of each metric are marked in **bold**, **underline**, **italic** respectively. + +
DataSetModelPSNR↑SSIM↑LPIPS↓DISTS↓FID↓NIQE↓MUSIQ↑MANIA↑CLIPQA↑
Fidelity-based SRSwinIR [101]28.48870.74220.43550.2473186.18437.429550.12220.31080.2563
HAT-L [26]27.16300.68350.46350.2330126.99117.358636.42990.41490.4111
PiSA-SR-PSNR [153]27.90790.72510.42730.4273144.31097.237141.62390.40240.1831
Perception-based SRSwinIR (Real-ISR) [101]26.50900.67000.33440.1928129.38793.869060.65440.56410.5205
HAT-GAN [26]25.78600.66430.35220.2137140.23644.825855.48620.56180.3556
DiffBIR [105]24.83430.55540.44660.2374130.63864.887165.96360.63420.7302
OSEDiff [183]25.22200.61640.34970.349791.59573.666163.98550.62190.6074
PiSA-SR [153]24.59710.59030.35410.3541115.82873.485966.04330.65550.6346
Expert Aerial SRHAUNet [168]28.51360.71460.43270.2083122.16567.249735.50210.41620.1706
TransENet [86]28.03170.69830.41790.2109125.94956.770035.11400.37760.1162
Agentic SystemAgenticIR [235]21.34310.51470.46000.2539149.71914.532567.02570.62830.6693
4KAgent (AerSR-s4-F)28.54810.71570.44360.2263127.44117.591637.57130.37740.4322
4KAgent (AerSR-s4-P)24.42120.53540.46960.2566139.77754.891568.18580.64390.6897
+ +Table 13: $4 \times$ performance comparison of evaluated models on the DIOR dataset (128→512). The top three performances of each metric are marked in **bold**, **underline**, **italic** respectively. + +
TypeModelPSNR↑SSIM↑LPIPS↓DISTS↓FID↓NIQE↓MUSIQ↑MANIQA↑CLIPQA↑
Fidelity-based SRSwinIR [101]27.87510.72570.44740.2488223.46537.124751.64810.31220.2441
HAT-L [26]27.73550.69620.45860.2195134.46497.114237.17840.41510.4194
PiSA-SR-PSNR [153]27.41760.71180.43780.2202167.76046.945542.50870.40950.1881
Perception-based SRSwinIR (Real-ISR) [101]26.47080.66980.33910.1983144.39003.892160.63190.55520.5091
HAT-GAN [26]26.80150.68480.33980.2073149.81214.845955.80460.55920.3392
DiffBIR [105]24.92540.57420.42010.2317146.26424.919866.45720.63150.7078
OSEDiff [183]25.04700.62070.35060.1875127.78883.664165.39340.62450.5976
PiSA-SR [153]24.40780.59320.35340.3534129.27243.511167.63650.65710.6229
Expert Aerial SRHAUNet [168]27.82210.69920.45270.2100128.87706.958635.58850.40890.1572
TransENet [86]27.30020.68240.43910.2113129.48936.498634.77130.37500.0984
Agentic SystemAgenticIR [235]22.48110.56540.46680.2388169.23414.744663.13990.59380.6252
4KAgent (AerSR-s4-F)27.67610.70620.43090.2250146.56187.255537.55430.38110.4368
4KAgent (AerSR-s4-P)24.48930.57950.43740.2471160.60064.652268.01170.63580.6456
+ +Table 14: $4 \times$ performance comparison of evaluated models on the DOTA dataset (128→512). The top three performances of each metric are marked in **bold**, **underline**, **italic** respectively. + +
TypeModelPSNR↑SSIM↑LPIPS↓DISTS↓FID↓NIQE↓MUSIQ↑MANIA↑CLIPQA↑
Fidelity-based SRHAT-L [26]33.07200.86560.24480.147158.01056.652751.75470.58580.3725
PiSA-SR-PSNR [153]28.96230.79990.34150.2093133.76647.735047.48470.48780.3094
SwinIR [101]30.59690.82540.32750.2215143.28667.539154.31110.45260.2700
Perception-based SRDiffBIR [105]25.83260.64890.37240.2340115.14406.090664.95390.65350.6772
OSEDiff [183]26.36160.71560.33240.2133126.46705.425764.12200.62780.6736
HAT-GAN [26]28.65570.78690.27510.1818115.17435.724557.01590.59290.3691
PiSA-SR [153]25.84470.69210.32200.2081112.10424.906266.39010.66760.6855
SwinIR (Real-ISR) [101]28.90000.78830.26570.1769110.45524.771259.74890.58860.4519
Expert Aerial SRHAUNet [168]32.82860.86270.24800.142857.30086.591750.74920.57110.3824
TransENet [86]30.72140.81760.28830.155368.51206.287842.89570.48560.3441
Agentic SystemAgenticIR [235]19.96550.59730.42270.2620137.27776.312665.55960.63750.6198
4KAgent (AerSR-s4-F)31.35890.84780.28530.177688.03667.080850.68150.55150.3799
4KAgent (AerSR-s4-P)24.92240.64270.38840.2555131.03466.160967.03550.67010.6800
+ +low-resolution SR datasets, as demonstrated in Figs. 9 to 12. In contrast, 4KAgent with the fidelity preference excels on high-resolution 4K SR datasets, producing the most faithful reconstructions in Fig. 13. Secondly, 4KAgent exhibits a clear advantage in reconstructing fine structures such as lines and patterns, as evident in Figs. 10 and 11. Finally, in the challenging cross-sensor super-resolution scenario of WorldStrat, where LR and HR images originate from different sensors. 4KAgent with + +Table 15: $4 \times$ performance comparison of evaluated models on the WorldStrat dataset (160→640). The top three performances of each metric are marked in **bold**, **underline**, **italic** respectively. + +
TypeModelPSNR↑SSIM↑LPIPS↓DISTS↓FID↓NIQE↓MUSIQ↑MANIA↑CLIPQA↑
Fidelity-based SRHAT-L [26]21.22380.64800.34680.2108145.37989.002630.66550.28440.2339
PiSA-SR-PSNR [153]24.43120.72710.33120.2283142.887010.532729.64620.29940.2509
SwinIR [101]18.09370.61360.34510.2279180.49548.260727.19540.21040.2355
Perception-based SRDiffBIR [105]20.64850.51500.67810.3712227.67647.551453.56660.54750.6075
OSEDiff [183]25.97160.63160.44600.2562176.25898.634246.50920.49880.5096
HAT-GAN [26]27.07960.72410.31990.1978137.44419.847430.37410.35870.2623
PiSA-SR [153]23.93040.61790.45810.2748170.04267.221448.64140.51520.5010
SwinIR (Real-ISR) [101]27.90620.71200.34730.2074149.442810.123732.94840.34970.3060
Expert Aerial SRHAUNet [168]26.18950.71430.31410.1976128.274710.631828.44010.31290.2701
TransENet [86]24.48790.69430.32700.2106133.69597.776527.89650.31520.2246
Agentic SystemAgenticIR [235]19.58830.51880.67160.3686224.80428.707954.06490.51660.5402
4KAgent (AerSR-s4-F)22.35290.64700.37020.2324166.27319.536434.36980.30110.2875
4KAgent (AerSR-s4-P)20.15100.53790.63630.3664223.48668.552856.84210.55470.6236
+ +Table 16: $4 \times$ performance comparison of evaluated models on the DOTA dataset (512→2048). The top three performances of each metric are marked in **bold**, **underline**, **italic** respectively. + +
TypeModelPSNR↑SSIM↑LPIPS↓DISTS↓FID↓NIQE↓MUSIQ↑MANIQA↑CLIPQA↑
Fidelity-based SRHAT-L [26]38.48560.91010.19560.10070.27226.807639.55620.50060.3408
PiSA-SR-PSNR [153]31.53360.85190.31830.179740.78797.098138.62770.42260.2444
SwinIR [101]33.74630.87590.29900.180915.69646.759143.37240.42000.2847
Perception-based SRDiffBIR [105]26.30320.64900.40350.195376.35224.036057.21330.62650.7306
OSEDiff [183]28.57680.75670.36320.212770.42224.128651.44680.58710.6907
HAT-GAN [26]31.27350.84080.27200.152747.06064.679147.60110.55080.3551
PiSA-SR [153]27.67650.73030.34990.217554.58443.803651.76040.61350.6353
SwinIR (Real-ISR) [101]31.69330.84230.25640.145337.37144.105548.33760.55610.4038
Expert Aerial SRHAUNet [168]38.22370.90750.20020.09740.29846.690738.87760.49260.3471
TransENet [86]35.97760.88240.24310.12670.21376.388634.87750.43450.2942
Agentic SystemAgenticIR [235]21.47190.72840.41570.216777.72864.790249.59130.54960.4853
4KAgent (AerSR-s4-F)36.76550.90170.23430.128311.00177.036138.72860.47380.3493
4KAgent (AerSR-s4-P)28.42810.75130.34400.218141.14253.926752.17350.62640.6608
+ +Table 17: ${16} \times$ performance comparison of evaluated models on the DOTA dataset (4K resolution). The top three performances of each metric are marked in bold,underline,italic respectively. + +
TypeModelPSNR↑SSIM↑LPIPS↓DISTS↓FID↓NIQE↓MUSIQ↑MANIQA↑CLIPQA↑
Fidelity-based SRHAT-L [26]23.95860.63620.64710.321982.76449.080731.53940.25650.3062
PiSA-SR-PSNR [153]22.62650.59940.72790.3368110.51129.258324.01540.20660.1766
SwinIR [101]22.94250.60950.68600.3815162.11289.761337.02680.27920.2814
Perception-based SRDiffBIR [105]21.40930.46120.55950.2214114.85953.404657.57710.50300.7588
OSEDiff [183]22.06020.55440.54500.2647107.36224.166752.52780.44300.7287
HAT-GAN [26]21.75250.59010.55900.2668139.20475.446545.64110.27910.3448
PiSA-SR [153]22.10220.57610.55520.2517100.33364.272348.01510.31940.5972
SwinIR (Real-ISR) [101]21.67700.57310.54310.2431129.77453.737750.54130.30330.4885
Expert Aerial SRHAUNet [168]23.66490.62680.69220.330486.24879.001826.24890.22070.2567
TransENet [86]22.96900.59920.74490.353197.58957.693121.30920.19030.1765
Agentic SystemAgenticIR [235]17.87360.46750.59280.2451135.64373.895054.36850.43010.6551
4KAgent (Aer4K-F)23.43480.62550.65200.3312105.67109.006433.66450.27250.3314
4KAgent (Aer4K-P)21.98260.55150.55250.2415112.25183.723055.77300.51750.7159
+ +the perception preference still maintains promising visual performance, as shown in Fig. 12. This demonstrates the robustness of 4KAgent across both resolution scales and sensor domains. + +![](images/b604630d0a0b522d29dfb3a1e11c7df98c2854dcb5b4404670a8acecabd45977.jpg) +Parking Parking 41 + +![](images/3d8827f890daad7923889c399f8898e992433064c950fa1941340c7d4b14a861.jpg) +HAT-L + +![](images/fc25e0a1ae3d8f6446eaecaa66037b398e039ca7af03459b214ca51c51f1079b.jpg) +PiSA-SR-PSNR + +![](images/715338798b843f79024dc0ed93b93aaaecd553fd74634a1d392d9cce76896b01.jpg) +SwinIR + +![](images/2e0dc9a4f26995d644d82b8a5677ec150a07da8e6c29da107fb66f1be3e2580d.jpg) +DiffBIR + +![](images/46682219df8370ab5234a2fd20352ab10d043be8444499bf9a142d87c4410cf1.jpg) +OSEDiff + +![](images/d4c6f5500104d08fd2aa51147f706f59550ae9f9b87fa84fbf5aed8cb9a9dc7c.jpg) + +![](images/9d95b7881e1a3e9bb7487c59992f5bf4307edd66bb45a6523549768efe929380.jpg) + +![](images/a33cc626161ec82a23d604dbd4cd96286e70837bbb6c9c67a509e689f76097bb.jpg) +Stadium_stadium_220 + +![](images/1d6254c0afa8c95c0adb956b5e0bf23bd6caadf93a52f9c2617e9f61e5e7773d.jpg) +SwinIR (Real-ISR) + +![](images/ed64308f622af560e2bbd90d949659201adbb6297c1a8cf054acd0dd2edff14f.jpg) +HAUNet + +![](images/cd0d6fc6ba5e5212e857399e63601e107b1e3eb9da880bf6d1c3717982deb842.jpg) +TransENet + +![](images/ee08188918745ed9c825fbc47f781f22e926892e1df76d5e6193d3b75f592a27.jpg) +AgenticIR + +![](images/771a050c443304531c3599209e600084530bcff89a3f20d76a8307e16c2d86ae.jpg) +4KAgent (AerS) + +![](images/d5d26c7014089189d3e98952c1634fb0d2fbce300bc5b48159c5c96debfab3bb.jpg) +4KAgent (AerSR-s4-P) + +![](images/c4c92bc744ee57d966b588cbe0f7a603422fa3264d61aea17ebb7268cc3f7468.jpg) +PiSA-SR +Ground Truth + +![](images/857baff88b4996d0301b9dd0d158d005d74734a6259a8aa69500c1c7b6ea738c.jpg) +School_school_148 +Figure 9: Visual comparison on AID dataset (160→640). + +![](images/0d8f46d299817613f497a7b7d904409ea76cf8f2c694c1d29a31b7e2cade6584.jpg) +1111 + +![](images/2d168ff97769277242fa15b09530eddf23ce8ffb4fb4dc29f05a3a439f027415.jpg) +1 + +![](images/c9e435353d43cd99598f531ec2784e5cb15b61efccdb8775841b2793cca76f30.jpg) +# + +![](images/4523f00abe30a9f7ea16252d33843f017de4eaa14945158d489e46e6dcdce1b2.jpg) +# +# + +![](images/3feeb53061ad5d0fc2accbcf469670d4d9615e6cb3530dceb5b99115e0d1e228.jpg) +# + +![](images/c0b3aac75f91b7fa9c518b8fd938b2b2b8e177bfbc514f7e984eafc2525c81f3.jpg) +1041-01 + +![](images/1a7aa1dcd20f5b665acd9d51653591793b748b57f4b37b596a4ac885a4f2dcb1.jpg) +15181 + +![](images/e570769305227c91091729d2307a9c672e9d6c8979fe12f27dc2090d2bc1a408.jpg) + +![](images/d9cdff0f5e49ad9cdd269e361e7787e94aa09af932fd7d46727212fc9f5adf97.jpg) + +![](images/3893127e8516af1ed429a4188be0bb3e108987b8e49d11a9732fe898106c07aa.jpg) + +![](images/7ae4b31bea373ec7e2f24fc83d64cba1df80b9afae9e603217cfdb02ce5642e7.jpg) + +![](images/4e1f67a97d5d7e2390ea2a7cb083d6450a8745bf633a42235b174f802a73f78e.jpg) + +![](images/ef3fac52571cca32cf9560449983e95d99fc8cc98553812b78f68b064dab7de9.jpg) + +![](images/06a219690f6c0fbeaa21fc390fe2f017e310d40621a8d9bf2a3b8489e5437a90.jpg) +11951 + +![](images/0dad58cdd35ae9383d587ea865eff5c5927eb77c93ac8771a6a6b343a471b0d8.jpg) + +![](images/a71d71948ed9feb089496b466cef64dd6eae440089a9790ce72ee7f0ab0d3a7d.jpg) + +![](images/67430c32c8c7ca0da1d21220e620c638c75c6a08d325140685129897bd4745e4.jpg) + +![](images/91d7adf4b28776e1bc1f15a548f3fcf125f6b7452eeb12730e47c71ab24b583b.jpg) + +![](images/936366d0b2dfa06f285688d17deca7eb87fb05e129d91d7e2ed527473a7382d0.jpg) + +![](images/1b7f3f9ec39d508c905e8089b29a39e1fc77bd43e9f96e4ba82a48634a175feb.jpg) + +![](images/02caf87f44e16b9ef288589b3622d0b181ef16007a9b14dd647ad5358b244241.jpg) + +![](images/d09bd6653fff6eb8b588f9bea8617f7a6932a0c9c76ca74b8a7ee9c92bec3af5.jpg) +16401 +Figure 10: Visual comparison on DIOR dataset $(128\rightarrow 512)$ + +![](images/0b2140e17addcaf896e1bb821adad89f57d7599d54be3e4ae61002ceff07d6d8.jpg) +SwinIR (Real-ISR) +HAUNet + +![](images/219396dacfd1fa1ea25e3ce5d29f8549859aaadf52b963d5c70323a41aeb3d54.jpg) + +![](images/736bfd452f92dc136be3a2cea3217a3cc25faff17dc9ca0011d1b7a7d293cc91.jpg) +TransENet + +![](images/aab1dbd8ea8f24726baa175103f7019ecbaef04cca668055f12ecce4a936a946.jpg) +AgenticIR +4KAgent (AerS + +![](images/3776668ed057d3de90843c077127510b181f63a7b5397859e40b74a110bba991.jpg) +R-s4-F) 4KAgent (AerSR-s4-P). + +![](images/671d1f6651afd94ada060dc281c64cf04bfaea1f1d1fe72902c2e27373f83ad8.jpg) + +![](images/706daa309df0eb0abd7c5136ea78afab576687e656ad1033219ae543de249ef0.jpg) +Ground Truth + +![](images/fd8d08868d4aca0d332387db4350b60ddd0205f4afe8405bee587ad4de1e886f.jpg) +HAT-L + +![](images/55c4896fac093e1abc40607815e13a47bfb036b0ba8995ea8422d57e7b9694e2.jpg) + +![](images/0cc054f5900765650a56de81c27e3607ce721395b3c57d64d1caea8f384372d9.jpg) +PiSA-SR-PSNR + +![](images/964dabc931c4de292823af9814ab4eaa54ae84b2f1d54c56ecdfeaecb05ac755.jpg) +SwinIR + +![](images/4fd006a4a7119a49fdad14a0a84377dad0a11b276202c03714f6faf44fce9ada.jpg) +DiffBIR + +![](images/b2df9fdf4f6288e3af18ff81c11f937904a3d47e4e283b1f40c0dba7fccbc95c.jpg) +OSEDiff + +![](images/2cdaa693f715bb8d32c48abae23247523d960c896765ed641ad38da1bde0e4fe.jpg) +HAT-GAN + +![](images/70d99a364452ff37a98efc69aaf43fbe31adda5cc20432be3fb2b2fe40c1b3a5.jpg) +PiSA-SR + +![](images/4c2f70c452fa2c9f2162616391f8710280899e0cb6502f63a5a9b7ae6f320a02.jpg) +P0256 +P0176 + +![](images/6f3269235b09b53e751669a0abb29711ac8fcd39689cfaae1978a1ddbdb7bde1.jpg) +HAT-L + +![](images/802b7b8b053f44714a95ac11ffc4af803f62b0ef74a6d86dcd5453a516ff4a2d.jpg) +PiSA-SR-PSNR +SwinIR + +![](images/880857283197fb56f10884ae568469ee83dcc9d312731f68b973a61c33e7fc2f.jpg) +DiffBIR + +![](images/f92a8654578fddf8913ce27fed0b013dd9ab21198ba989531edf517bb283b9cb.jpg) +OSEDiff + +![](images/ce47a50ff26671e51e335cfdd74fe28a3aa101962e223d3ee30f7ecdeeff81f0.jpg) +PiSA-SR + +![](images/a80ba1a59cb7fdfae7ac1d1e6b1d82ac88637e1aad2d18cd3618a24d5a2f0e85.jpg) +P0222 + +![](images/7cde38469b26788c76c15bf6899d14cb11d4c17b7f2fb9e1483ca2f4865c6439.jpg) +HAT-L + +![](images/bb2fd2e2cd01e1af3792447e73e70b75514b1acf953cd60d0b68e42aedabe1ce.jpg) +PiSA-SR-PSNR +SwinIR + +![](images/29a3612e92d2f81a03118db8c2dcde64a9cb680b30ffe544e3fea88f4523626c.jpg) +DiffBIR + +![](images/3e8ff252017518f4ac7ffc072cb1c240f148082cb12e3b87ab47b033e168b961.jpg) +OSEDiff + +![](images/11e20b3580b769709be55888cb3495ee90eafafe14a3042466ef68be682892a3.jpg) +HAT-GAN + +![](images/ed48bbab9e0e49ec64a714e236c592dfb7ec2ffac1f3587ce5f67191195eca46.jpg) +PiSA-SR + +![](images/ce6a97af89f22c4d1f67f11a01590aa1c60503fb40928100cc685a173eb2cc59.jpg) +UNHCR-PAKs003493 + +![](images/2addcc6a6dd0e3126e85f2f8389fb91afba4c4f87de0386a70a0266a499deba4.jpg) +HAT-L + +![](images/b021f07b6d7d3ceda2609375beace2606e56455b41820bdf1f9512f483c915ab.jpg) +PiSA-SR-PSNR + +![](images/90dedd2f4961c55961399388670ae175768860756e53fbe1117f8b769288ff85.jpg) +SwinIR + +![](images/aeab71afd4e62d280e989984adfab8fd140930f75505515fe7dff4e028d090f6.jpg) +DiffBIR + +![](images/2b34c678a7bcb7a225283602c701acbe14dddade188b4141b535b428f29fb787.jpg) +OSEDiff + +![](images/c2ca2e5294f262bf8340d37087d2cfcaf0d5ae79a60e21870b44a36a9c609c7c.jpg) +HAT-GAN + +![](images/0910c76e58cfaadc00193e3b9954775c8bb14ff5a826604af8b7bd63a10761b6.jpg) +PiSA-SR + +![](images/d866992b36e814ae2ace9b3b1f3c823c6748e40041f641fe4ea75465c7e92a93.jpg) + +![](images/f8865e1b76e86587efa679cc6d88fb9a027b85ee5d0f8a411aed7faee27d512d.jpg) +HAT-L + +![](images/8243c7088b599fa6645cb4e2a29417acd85c09e73a7cd1e2858e43b0d9915fb9.jpg) +PiSA-SR-PSNR +SwinIR + +![](images/94179c50d62b865aeaf1b3bc915d196ce5142ee883d2860375bcd3ad568dc7e3.jpg) +DiffBIR + +![](images/6196714c44cc93a922bd3f05f32ffef915723794c0e958466cf1564286759151.jpg) +OSEDiff + +![](images/7d46ae0ed97ae3524d0b1cfd880f572ba0483e527a6a26a0844febe89e3ed3c6.jpg) +HAT-GAN + +![](images/b52c3bd2f24322e2c83d0d620bbf3de2ff92f5ee27cbe6b78ca928a7836a8346.jpg) +PiSA-SR + +![](images/ff7716b3bf3319278b0975e948696535e22e069d9281edeed9bcadba7b1757a1.jpg) +Landcover-1423442 +UNHCR-TURs004682 + +![](images/8c3f01d174e1db6ec1c704a9e5e43c37cbb0cf587a841d728da005d3359340e1.jpg) +HAT-J + +![](images/9095a1907f7bdaffce024c819895fd1f392c1e25ae4824c02300c50c6a29a10b.jpg) +SA-SR-PSNR +SwinJB + +![](images/27ae3ac391ad998f43e6c4e96e82cd5e27a61ce3d3b86f242e56a1ed231a3da6.jpg) +DiffBIR + +![](images/0d964c4668fe38132820f32b4b9969704ed330af79a709a829a1ad29b033c441.jpg) +OSEDiff + +![](images/5c0902f536cd788b4d81a04abd696fce897773dd4dca230a2b975b69e16ab081.jpg) +HAT-GAN + +![](images/ad61990fda3f690d916056b6c02b608c77d72ab54404550eb23eb06c77229ef2.jpg) +PiSA-SR + +![](images/c939dfc56dc5b17314488f7256aebc30f132b01e9775206a057440b0dd41ed47.jpg) +P0063 + +![](images/705f240abda6178c55d445d44053a677cc79c5daa4946db5699085a984f78b31.jpg) + +![](images/ae27b5585ce94bffdb1c269f150a40e658846476bc5e5b9aa0db5a939f8a5543.jpg) +SwinIR + +![](images/f5fc6f3610cc90493d34253c2d52648e68be51e91e8085649d988c5085b23a33.jpg) +$\therefore m = \frac{3}{11}$ ; + +![](images/65b105c5be9f3d28f727b3ede46f25cfd41cf6edcfb3eebad4320590817f1e4f.jpg) + +![](images/cd1c4685acca93b97f5c6b692fa85c6b3a700f8375a7cbb7ce1b485d4e89708d.jpg) +OSEDiff + +![](images/8feed24130122aad5568023bbcdd152ba7010cabd01186b3b34eb2ea1d950549.jpg) +HAT-GAN + +![](images/bd73364a5b6544bebd40c4d32805d854aef8aeef2a8e4242debe84c65a6cdade.jpg) +PiSA-SR + +![](images/a2d0f84d4501e9ba46e959edbdff64626e40c4c1a12325931f39e4d91cc71e75.jpg) +SwinIR (Real-ISR) + +![](images/4327464cd4e8384efcde3bec8f9e5b9a4ea63129c18139a7588b766feaf754c1.jpg) +HAUNet +TransENet + +![](images/d95a73ea5cd66673cd431ceca655d34dfb56a351f16ceed66e957df3b4bad585.jpg) + +![](images/979f44fcbc45d74d8f79e0267379ebac26307f49ac99bdd7806d5066530b5e53.jpg) + +![](images/1be7b3b2368386ad38a0f9b5bbeecaac3981eb6ca42ceebb84d65be0ba0f53fe.jpg) + +![](images/21f14f14368b1cd04a4f4c7b04ae2fc54e8f2bacf74ec9499d6f6247e9df8a75.jpg) + +![](images/614304d348708c8f7638273eddcbe96c4e225ce8be5f2e98761ec2860b40f884.jpg) + +![](images/17732b9a3bf2a9e11e966f79da7b52487a7587ef11450c91270f1b2cbe4245af.jpg) + +![](images/f7339c869a5b6108ccef5b22585c79608175c172b16d1da86763c94624638015.jpg) +PSA-SR-PSNR + +![](images/20b09770c680a8fb24ec9fc7419a92da748abc2794715623ce67e29101ce6f6a.jpg) + +![](images/810431e2207006b0580de9f10d8cb5060bc55319602c6447d24cc3c8731c895a.jpg) + +![](images/e72e72ed6f510f2eaa9e02bee78b815b450a0bbf2c36f289ff37cfad1c49dfe2.jpg) + +![](images/e3e793d22ee7b84c2b166de3510b6faf84ab2fca7e63a2f908bc51760ed831e9.jpg) + +![](images/bfc8d76dcb1fe7082aec0418038171c3fcf0caa9d5d7974b3d99ba6102e15df5.jpg) + +![](images/a4d539f0dd02a67793beea2b8c58b7c866cc97d1096c4212f0a490f42cca4b0d.jpg) + +![](images/9a897af63447ec056c85b67909f0724fc2c04bf9873e404b6c3f2298f6a04800.jpg) +P00867 +P0021 + +![](images/2e702b36c22a953f4dc3d5b5126ff168385f496ccb2ef4945a347451735e0b02.jpg) +HAUNet + +![](images/83ce9363a94a8b2f195d5b878caa8adf55c502758b9fc180491ddf1590689659.jpg) + +![](images/fd4de3f6ecb0801c0018117836b97131951cacf6757198a8404bb8a292134308.jpg) + +![](images/5f3a45158abe59416a8d0770e7256f611a7b99843891bf301ff43859a3c9a87f.jpg) + +![](images/e971ed5bbd43dc86469cea6121151fac865ce7e46d601803cfe3aa533767672a.jpg) + +![](images/0b5321a2fca034c2d70edf3486fb76074be4407ed202ba3398da47210aba1159.jpg) + +![](images/ccb6f3e841eb633466e9452af9cf7ddd55796c2f200e38bde9373184a42f4a50.jpg) +P0064 + +![](images/58cdd787c469de612aebf2c60d0674ed9d96e31d0ce92383adbca242da783fc4.jpg) +$\therefore m = \frac{3}{11}$ +PiSA-SR-PSNR + +![](images/e79a5cfa4aef0c943a84e3b30e229048482a9d214fd9e6d97922e4d95cd5f14a.jpg) +SwinIR +DiffBIR + +![](images/a06e44282b50eb3cfc94610b704870805ac009164798e871ba236d6e5fdcb929.jpg) + +![](images/02810fbbf8888f27294c2e0c61cd449b16f187fd8fe4789a436c7e290c050241.jpg) +OSEDiff + +![](images/a6bf353db8a880eb84e38319dae57f216136c3a4b392174f62f80ff011c50b8b.jpg) + +![](images/f560c744e28a1bb4e84f9cf98f7b28002583d3de5e72cf51bbf3c79a1c937a0d.jpg) +round Truth +PiSA-SR + +![](images/fcad25d3ed1b6aba20fcb27fd6b54bc0bfc9a5e6cd409c65e69a34162b5e20dc.jpg) +SwinIR (Real-ISR) + +![](images/047e532e4bf0de3991abbbf115c52dd95b22fbaa0d8f659289766dff714868ae.jpg) +HAUNet + +![](images/98d28d6ab45d0c1744a08b50ddd52dab2c0b8448070dc09c163980efcb960efe.jpg) +TransENet + +![](images/a0663ab0c11367d8700e48641539756fe6b55ea68f9e4520967315f07602a1a7.jpg) +4KAgent (Aer4K-F) + +![](images/d33ce5b109e2f6cad1a28e56f8aad28b9ed7ffe92771a72e51599d04be1c42ed.jpg) + +![](images/1fa0754a0815652ee64469640470bacfabdddebb38ee95f79ab269d87d8b7e89.jpg) + +![](images/9d1cca0f42d5561e6fc5f31fd64a605d2792604e9b04cb9c0d92a5c1f6277c00.jpg) + +![](images/bd328b9855758575ffefb9d7fcfe8e703e11c5a0c41c3abfa21a564b2fe2201b.jpg) + +![](images/440f2be1a22d88977479ff29b14cd137cbc9743e069a4625a0539fa4f395dbda.jpg) + +![](images/3c37efe44124f5ad7a02ed2c543c1a5eb9197a59ea11e0e7bf998c8aea042a35.jpg) + +![](images/2e7ff6a3c69ace03632f8deeb9abcb6acf1e967957da865bbd0169f90bcaf593.jpg) + +![](images/68315d7c7d4a3b25218e1256e34480e5a2709b41a37f78c9929f358489473c0f.jpg) + +![](images/57f7f7edb6aae3b880591c4f43b46d54f080bfb1aa73108c25224da1aabd5689.jpg) + +![](images/8b83630ed710051b54c003d88c946e56f5d080a42d97e9b064f8f77ba0c66f4d.jpg) + +![](images/c6ec4be9584db78603d67f4158b34dd292ca0e231a636218e366b6f54ab5a4cf.jpg) +HAT-L + +![](images/e2acbadc816b36d7a434137855268e4ea06f2663bb1339c0e6a8575d16c8cdea.jpg) +PiSA-SR-PSNR +HAUNet + +![](images/da83c3eb83893a5aa8475a6ca794f6a18df14e32052b74c3e26fbfa35f0cf120.jpg) + +![](images/e186c4bb805afc1b161253334982ac97a87c7bc1408f0e11d2a831ab73bb5aef.jpg) + +![](images/a0b67a5f9510b85fc3ebd71ea5784d283cef91045ceb7e9be9a9453ce2583f7f.jpg) + +![](images/cd6e92658ad4aec8f0e38728c476c86b37ce19a36b8fca470fb4c531f05056d4.jpg) + +![](images/47c61434ddce87f1b4e46d25b3900db2588aef418c1666f960c8090908099bc4.jpg) + +![](images/175c3a3a233a80b680daf7abb15e734c42ce298096683d70efee73babfa7da11.jpg) +SwinIR (Real-ISR) + +![](images/92360f838cbbcad32b8bbaf10ee6efafe2217a21f14acf13a1a898df2b0d1cd7.jpg) +PiSA-SR-PSNR + +![](images/ecc193d49517b07c0b8f1d40f7b2baa7a02a2a528bec5c72772527bce9e6f6fd.jpg) +SwinIR +${12}/{14}$ + +![](images/41df5d23ed5b3afe19a5e27786dcf73396716a9112bda8ecb11c76279a009073.jpg) + +![](images/d6e9f3c6d3fab7e183b689178ff265c9040076a8bf18be24b86f32932b1b1986.jpg) +OSEDiff + +![](images/05b2e05fc85647be50b0d08badeb6694a6b7922cf74deb60cad0bca02bcdb2cb.jpg) +HAT-GAN +PiSA-SR + +![](images/1f33b57b64cd6bb838d030914fd4ad372bf9d321f09f26088acdd93302cd94d9.jpg) + +![](images/7634d23c842867106c0dae7e0ac48866bb80e7fc131d7079adda17f3443dc2ef.jpg) +P00158 +Figure 13: Visual comparison on DOTA dataset (4K upscaling). + +![](images/d06332fe5826a37ebb6e597c1713d029485e8aa13dcd0459368a1e8e30c2936c.jpg) +SwinIR (Real-ISR) +HAT-L + +![](images/abd1ba2baae9e91215acbf32252910693e5feb4a67cc8e99c445b0bccb6ac5ca.jpg) +SwinIR (Real-ISR) + +![](images/41d1a62e5b522be177d158d3f43f8ba1d7e5537ac90235ea56b2a622a258f6d2.jpg) +HAUNet +PiSA-SR-PSNR + +![](images/6ed177211824fe69f8217b4520b9faa906d0290beb733c47503fba1152bc78b8.jpg) +HAUNet + +![](images/7034497cc77fc985850a22eda82114082f0ee6d6847b90726c914c0d33d459f5.jpg) +TransENet +SwinIR + +![](images/d63b0d823b53e5c96a5b2f5adfd673734a6a33d400d829bb7cbb321aded31766.jpg) +TransEN + +![](images/1fbe1a10997aef07e14ab340c9dc033b82c868f7cb3d07689848e09b5f60b991.jpg) +AgenticIR +DiffBIR + +![](images/7e19b9130488a9f0e5c6e7b586a236b56030fce40998a71c919f94ae81ee00f6.jpg) +AgenticIR + +![](images/3cdb7ab3cc7ad1ccb2923eda917f67207389ce810f52c08ca8be03f2330c386f.jpg) +OSEDiff + +![](images/b13ed23d9e1725916a2c144509a55b8f0c9e582594ad15f71c463f87c5ed25c2.jpg) +ent (Aer4K-F) + +![](images/b129cff1b7b096e4916205c3d50f3b32be27c40a428b96ef1d7fca0c1d83c44a.jpg) +HAT-GAN + +![](images/0f18608ac9f35e9fb0bc5f1a12dde1d66baef378add2ebe872ba5312198bcb07.jpg) +4KAgent (Aer4K-P) + +![](images/f659dfdb1e482c7aba62e245ea1bca92a9e95a208f9900743ca7d1af7595ee91.jpg) +Ground Truth +PiSA-SR + +![](images/e1898b8c381a12a4683c3eb370526580ab0e63dd36e0661ab045f9574d9fa78a.jpg) +Ground Truth + +**Discussions** These results across fidelity and perception metrics, combined with qualitative visual comparisons, provide several key insights into the advantages of 4KAgent. First, the consistent top-tier performance of 4KAgent with fidelity-based profile across a wide range of datasets and scaling factors suggests that the agent's analytical pipeline and adaptive control provide more precise reconstruction than traditional feedforward models. Unlike conventional SR networks that are typically optimized for either low-level fidelity or high-level realism, 4KAagents architecture decouples these objectives through specialized profiles, allowing it to excel in both domains without compromise. Second, the perceptual strength of 4KAgent is reflected not only in numerical scores but also in its sharper textures, reduced artifacts, and more semantically coherent outputs in qualitative results, demonstrating the value of integrating agentic reasoning with perceptual priors, especially in real-world datasets like WorldStrat. Finally, the margin by which both 4KAgent variants outperform AgenticIR, demonstrating the superiority of 4KAgent in terms of agentic system. Our contributions in the design of specialized profiles, adaptive modulation, and perceptual alignment mechanisms are crucial to bridging the gap between task generality and SR specialization. Together, these findings indicate that agentic architectures, when properly aligned with SR objectives, enable scalable, generalizable, and controllable super-resolution across diverse remote sensing domains. + +# F.2 Fluorescence Microscopic Image Super-Resolution + +Confocal fluorescence microscopy is one of the most accessible and widely used techniques for studying cellular and subcellular structures [56, 200]. It builds a sharp image by using either a single pinhole to scan point-by-point or an array of pinholes on a spinning disk to scan multiple points simultaneously to reject out-of-focus light, offering molecular specificity and 3D sectioning capabilities. However, it is constrained by diffraction-limited resolution down to $200\mathrm{nm}$ under visible light [157]. Meanwhile, high-intensity illumination required for improved resolution leads to photobleaching and phototoxicity, limiting live-cell imaging duration and data throughput [158]. Deep learning-based single-image super-resolution (SISR) methods have shown great promise in recovering high-frequency details from lower-resolution inputs in biological microscopy, overcoming some limitations of hardware-based SR techniques [120], despite the scarcity of large, publicly available fluorescence microscopy datasets. To extend the evaluation of our 4KAgent on this scientific application of different modalities, we conducted experiments on a representative dataset against baselines from major SISR families. + +Settings We evaluate 4KAgent on SR-CACO-2 benchmark dataset [12], which contains 2,200 unique images of the Caco-2 human epithelial cell line, labeled with three distinct fluorescent markers: Survivin (CELL0), E-cadherin / Tubulin (CELL1), and Histone H2B (CELL2), at $\times 2$ $(256\rightarrow 512)$ , $\times 4$ $(128\rightarrow 512)$ , and $\times 8$ $(64\rightarrow 512)$ scales. To generate high-resolution images, each tile was scanned with a $1024\times 1024$ pixel resolution, and 8 scans were captured and then averaged together to reduce noise. Meanwhile, low-resolution images were captured directly by the microscope at three different scales without averaging. The full dataset contains 9,937 patches for each cell extracted from scanning confocal volumes, with tiles 9, 10, 14, 20 used as the test set. In our experiments, we randomly sampled 100 patches from each marker category in the test set at three super-resolution scales. + +We evaluate 4KAgent with the ExpSR-s2-F profile, ExpSR-s4-F profile, and ExpSR-s8-F profile, considering the demands and requirements of the microscopy image super-resolution task. We benchmark 4KAgent against 15 representative SISR models, broadly spanning pre-upsampling, post-upsampling, iterative up-and-down sampling and progressive upsampling SR methods. Each model has been trained on the SR-CACO-2 training set before deployment. Encompassing a wide spectrum of upsampling strategies, this rigorous benchmark ensures a comprehensive comparison between our 4KAgent and other specialized and general-purpose SR methods, assessing 4KAgent's blind inference performance on the novel microscopy data domain. + +Quantitative Comparison All quantitative results are in Tab. 18. PSNR, SSIM, and NRMSE were selected as criteria. We noticed that the background of a cell constituted a significant proportion of an image, as was also reported in the original SR-CACO-2 benchmark. Because including the noninformative dark background in evaluation can lead to inflated and biased performance metrics, we adopted the masking strategy described in [12] to define our Regions-of-Interest (ROIs) and calculated performance metrics based only on these areas. Across different scales and cell types, 4KAgent with Fidelity preference consistently achieves top performance in pixel-level reconstruction in ROI. The superior result on PSNR, SSIM, and NRMSE confirms 4KAgent's effectiveness in reconstructing fine fluorescence-labeled structures and low-level pixel fidelity, under various downsampling conditions. Furthermore, when compared with ENLCN, one of the most competitive methods, our 4KAgent with fidelity mode consistently exhibits a clear advantage in all numerical metrics, underscoring its ability to handle the blind super-resolution task for real-world microscopy data. + +Qualitative Comparison Representative qualitative results for the highly challenging $8 \times$ superresolution task are shown in Fig. 14. At such a high magnification, where information loss is severe, the ability to reconstruct distinct biological structures for each of the three cellular markers becomes a critical test for any SISR method. + +Our visual analysis reveals clear performance differences. For the inherently dim and sparse CELL0 Survivin marker, 4KAagents reconstruction is markedly clearer and closer to the ground truth. It successfully restores the faint midbody structure with higher fidelity than top-performing baselines like ENLCN and ACT, which struggle to resolve this signal from the background. This superior performance is also evident for CELL1, where 4KAgent delineates the membrane and cytoskeletal framework with sharp, continuous lines. In contrast, the outputs from most other methods appear + +Table 18: Performance comparison of evaluated models on the selected sr-caco-2 test set on ROI only, i.e., cells. The top three performances of each metric are marked in bold, underline, and italic. + +
SISR MethodsScalePSNR↑NRMSE↓SSIM↑
CELLOCELLO1CELLO2MeanCELLOCELLO1CELLO2MeanCELLOCELLO1CELLO2Mean
BicubicX234.9332.6130.2432.590.08870.06810.06610.07430.78990.78260.70380.7588
X435.0831.9930.4332.500.07930.06900.06410.07080.84110.79980.77180.8042
X832.0128.7726.2729.020.13110.10710.12400.12070.72800.66770.68080.6922
Pre-upsampling SR
SRCNN [40]X237.0434.4733.0334.850.06100.05160.04710.05320.87330.85660.82830.8527
X435.3932.7331.4833.200.07040.06100.05630.06260.87070.81930.81040.8335
X832.5229.1626.5329.410.10740.09240.11430.10470.81170.72190.72070.7514
VDSR [72]X237.5034.2933.0034.930.06020.05540.04790.05450.89210.86080.83850.8638
X436.1832.5231.4433.380.06630.06380.05710.06240.87770.82180.81800.8392
X832.0328.8026.4229.080.13070.10620.12370.12020.72910.67120.68770.6960
DRRN [154]X237.3534.2333.0834.890.06090.05550.04750.05460.89170.85970.83860.8634
X436.0032.5031.4333.310.06780.06370.05700.06280.87720.82160.81670.8385
X831.9228.3126.3928.880.13100.10960.12300.12120.72900.65830.68600.6911
MemNet [155]X235.6933.4030.8133.300.07760.05570.06070.06470.82950.82800.77590.8111
X434.6132.4830.2632.450.08080.06100.06540.06900.84650.80670.76510.8061
X832.0028.7626.5229.090.12720.09930.11830.11490.75280.69720.71020.7201
Post-upsampling SR
NLSN [125]X237.5734.3133.1435.010.05880.05270.04650.05270.89110.85630.83750.8616
X436.3932.7531.6833.610.06300.06000.05480.05930.87540.81790.81310.8355
X832.5629.1326.3029.330.11470.09390.12050.10970.79090.70920.69890.7330
DFCAN [142]X237.2134.2032.7434.720.06140.05610.04930.05560.88990.86030.83750.8626
X435.9232.4931.2933.230.06840.06530.05820.06400.87700.82230.81740.8389
X831.2528.1525.4528.280.13440.10790.12760.12330.74470.67490.68410.7012
SwinIR [101]X224.5534.4833.0830.710.23490.05270.04730.11160.37850.86260.83850.6932
X435.9332.6631.5733.390.06730.06180.05590.06170.87720.81980.81610.8377
X831.3428.4325.8628.540.13140.10350.12300.11930.75160.68380.69230.7092
ENLCN [186]X237.5934.4133.1535.050.05740.05180.04620.05180.88760.85690.83400.8595
X436.3032.7431.6333.560.06380.06060.05530.05990.87660.81960.81480.8370
X832.6929.2826.3129.430.11080.09210.12050.10780.79980.71090.69840.7364
GRL [96]X231.2834.5432.8132.880.10880.05220.04920.07010.80430.86250.83370.8335
X435.7632.8131.4833.350.06780.05810.05650.06080.87740.81330.81440.8350
X828.1428.9326.2227.760.15550.09530.11040.12040.72960.71100.71970.7201
ACT [203]X237.2434.6633.1435.010.06190.04960.04590.05250.88900.86040.82880.8594
X436.1732.7631.5633.500.06520.05900.05540.05990.87610.81340.80650.8320
X832.7429.1326.3929.420.10640.09150.11520.10440.80830.71280.70630.7425
Omni-SR [165]X237.3534.1933.0234.850.05970.05480.04750.05400.88960.85620.83700.8609
X435.8632.5331.4933.290.06800.06350.05630.06260.87370.81650.81170.8340
X830.4428.2125.3227.990.14180.10750.12650.12530.72310.66730.68080.6904
Iterative up-and-down sampling SR
DBPN [51]X237.4434.5433.0235.000.05880.05120.04700.05230.88720.86010.83390.8604
X436.2232.8531.6433.570.06380.05910.05480.05930.87450.82160.80910.8351
X832.3928.8926.3629.210.11030.09460.11570.10690.80600.70840.71490.7431
SRFBN [99]X236.0233.4931.3833.630.07670.06110.05760.06510.83190.82050.75920.8038
X435.6632.4931.0533.070.07290.06360.05920.06530.85890.81310.79210.8214
X832.3329.0526.6229.330.12430.10190.11810.11480.75530.68690.70810.7168
Progressive upsampling SR
ProSR [173]X236.9234.6632.8034.790.06210.04940.04850.05330.88790.85770.83850.8614
X436.1132.7131.6133.480.06560.06140.05560.06090.87710.82170.81640.8384
X832.1329.4326.3629.310.12680.09090.12240.11340.75040.72000.69190.7208
MS-LapSRN [81]X232.7332.4928.3431.190.10140.05930.08050.08040.79570.81770.77350.7956
X430.9131.3630.6930.990.11180.06720.06110.08010.81240.78200.78580.7934
X830.6727.6424.6827.660.12060.10560.13050.11890.78290.69020.66490.7127
Agentic System
4KAgent (ExpSR-sN-F) (N ∈ [2, 4, 8])x239.9236.9533.9336.940.05080.03370.04260.04240.93210.91050.87450.9057
x441.2536.8635.0737.730.03890.03180.03660.03580.95550.93140.90890.9319
x838.9333.6631.9934.860.05320.04830.06020.05390.93780.90330.89290.9113
+ +noticeably blurry, failing to preserve the cells essential structural integrity. In the case of the bright nuclear marker CELL2, the diffuse nature of the chromatin structure means even the ground truth image itself lacks hard, well-defined edges. In this difficult context, 4KAgent reconstructs a complex, high-frequency textural pattern that is visually competitive with the other methods. While the intricate nature of the target makes absolute fidelity hard to judge, our method effectively generates a detailed result on par with other models, even under the zero-shot blind inference setting. + +**Discussions** Our experiments show that 4KAgent delivers leading performance on the challenging SR-CACO-2 dataset, with quantitative metrics and qualitative results surpassing those of the evaluated specialist models. The superior performance of 4KAgent underscores its strong applicability to fluorescence microscopy SISR. First, it showcases strong zero-shot generalization, achieving highly competitive super-resolution performance on microscopy data, and could be further strengthened by adapting more domain-specific tools. Second, 4KAgent exhibits impressive cross-domain transferability, successfully adapting methods originally optimized for natural scenes to the distinct characteristics of fluorescence microscopy images. Third, the agent-based architecture enables the flexible and + +![](images/226bd2692a2e2e17fe2e8563bbf6e9c922741c628ad7d2dbf24c47a0f5af8c6c.jpg) + +![](images/516f5c8caf11744cd020d95e205f4d95b9cfb5e1b7005382464310820c3b8872.jpg) +Omni-SR + +![](images/4f5d875911eee2464c3e668550ce52b71883dd180e88c1c5b9641d46bb6f874f.jpg) +DBPN + +![](images/fe10390beb62e376771057623dfc4e53c3f1585df5b2f30f4e1ec20c64df4424.jpg) +SRFBN + +![](images/940c3978ff7fea255752043e885c6b65538f480cbd2c2ababc829b23241332b1.jpg) +ProSR + +![](images/05ce23ea54307e3cb2d3d7bde77f0026244dda672ab68e812cafb014af7466e5.jpg) +MSLapSRN +4KAgent (ExpSR-s4-F) C + +![](images/58a5330db8f21d12263eb7690192797095e9688dba897fbe0012eda9a1aa9cf5.jpg) + +![](images/924f294ccab4709cf3ca357383a5a00156a6a3bc20a4e244f72cca49bfef2aaa.jpg) +GT + +![](images/0968cde639c5b96c84e91f817b2bba713c48ece0af02326a422540b6ff9c27d5.jpg) +10 127 2048 2688 7936 8576 CELLO +9381536217623042944CELL1 + +![](images/7a7c213d4fbf85c8837d31762eb1265fe9f4ce4bab316bf5d0452de7e85517b7.jpg) +Omni-SR + +![](images/1fc0340b2ceb44c54579bd5da9ecb21455895a3307eb7b306dc0fe7fed7ec0bd.jpg) +DBPN + +![](images/78c6a47b023b194293fdd7ec92fc5b3dd1b6f3d859beaa011eecf348cf11b678.jpg) +SRFBN + +![](images/4bfef9445e1d21e7bfc8fda0eb622215df331625e07e1d8f6193b6f32c480e51.jpg) +ProSR + +![](images/43aefece4b0beb51a6aff5108ac027efefaecd5e535bb6824812296e3894e7fd.jpg) +MSLapSRN +4KAgent (ExpSR-s4-F) C + +![](images/68ce96db39db23a481e173a89d80cb290d4e2d9ccb40fc9aaa3da5bb3e7e9664.jpg) + +![](images/2bbe4575545e9c89167a424f5d5cb07665b607d42e3ee9edb70bf4d79517152f.jpg) +GT + +![](images/627b2a036ecd0a2f6ae1b2b64c4044a3b454219c3e7f9bdd3272ffae4659a45b.jpg) +9 208 7936 8576 2304 2944 CELL2 +Figure 14: Visualization of fluorescence microscopy image SR on SR-CACO-2 dataset (64→512). + +![](images/745a9e8dd58291d858ce88887c2771b7ccbaf5efe4862504aec59c98019eda43.jpg) +Omni-SR + +![](images/5a6e53c227b9df1ba49c6024423773edc3dd61ed1b6766765a886dd638d73aab.jpg) +DBPN + +![](images/e33f780080d816f42b37337c695bfd041f11db994206b28ad93ff9d8446a83fc.jpg) +SRFBN + +![](images/ee1708ad846ffaad0c92c270c3fdc9e5a201960ec786ca8de3acb1653e24a54b.jpg) +ProSR + +![](images/925153e080f8d63477a2a5897efe70fbc796226950ad335cbd9ccf1a2fefe2b5.jpg) +MSLapSRN +4KAgent (ExpSR-s4-F) + +![](images/f9508412e0abb76fa993fe25a0d250df12a2b05e93103bc3b460f55974be3c69.jpg) + +![](images/2938edd1d7107759a7d8c813e12255e5d529d5dc988a7d1a80e3683d0b1ecddf.jpg) +GT + +modular integration of existing models without requiring expensive retraining or model modification. Beyond immediate applications to super-resolution, the modularity and domain-agnostic nature of 4KAgent also suggest its broad potential for other real-world biomedical imaging domains where data scarcity or retraining costs are limiting factors. + +# F.3 Pathology Image Super-Resolution + +Pathology images—particularly whole-slide images (WSIs) and their extracted patches—play a critical role in digital diagnostics and disease detection. Typically, glass slides containing tissue sections stained with hematoxylin and eosin are digitized using high-speed scanners at resolutions approaching $\sim 0.25\mu m$ per pixel, resulting in gigapixel-scale images characterized by distinct color profiles and high-frequency textures unique to cellular structures. However, the substantial costs and data storage requirements associated with ultra-high-resolution scanning have led many workflows to rely on computational upscaling from lower-resolution acquisitions. This task is challenging, as pathology images possess specialized characteristics that pose significant difficulties for conventional single-image super-resolution (SISR) methods originally optimized for natural scenes. To address this challenge, we evaluate our 4KAgent on the bcSR dataset [63], comparing it against several established techniques to assess its effectiveness in this specialized domain. + +Settings Our evaluation is conducted on the bcSR benchmark dataset curated for pathology image super-resolution. The bcSR dataset was derived from the larger CAMELYON [108] dataset, which contains WSIs of H&E-stained breast cancer sentinel lymph node sections. To create bcSR, the authors first sampled representative $1024 \times 1024$ patches from the original WSIs. Subsequently, a filtering process was applied to remove patches with large blank areas and to select for images with high color channel variance, ensuring the dataset was rich in informative and challenging tissue structures. The final bcSR dataset consists of 1,200 unique images, which were split into a 1,000-image training set and a 200-image test set. + +Following the standard protocol established by the bcSR benchmark, the high-resolution ground truth images were downsampled using bicubic interpolation to generate the low-resolution inputs. We evaluated 4KAgent using the ExpSR-s4-F and ExpSR-s8-F profiles for the $4\times$ and $8\times$ tasks, respectively, prioritizing pixel fidelity which is critical for preserving fine diagnostic details. Performance was measured using the PSNR and SSIM metrics. + +Quantitative Comparison Quantitative results for the pathology image super-resolution task are summarized in Tab. 19. Across both $4 \times$ and $8 \times$ upsampling tasks, our 4KAgent achieves the highest SSIM score among all evaluated methods. While CARN, a model specifically designed for this pathology dataset, attains a marginally higher Peak PSNR, 4KAgent's superior SSIM is more indicative of its ability to accurately preserve the complex tissue morphology and textures that are essential for pathological diagnosis, which is more critical for the reliability of features used for clinical assessment. This demonstrates the effectiveness of 4KAgent in recovering diagnostically relevant details from downsampled pathology patches, with performance comparable or superior to other specialized, fully-trained models. + +Table 19: Quantitative comparison on Pathology dataset. The top three performances of each metric are marked in bold, underline, italic respectively. + +
Method4×8×
PSNR↑SSIM↑PSNR↑SSIM↑
Bicubic27.0190.665922.4750.2776
SRCNN [40]27.4750.732922.4890.3624
SRGAN [82]28.6060.771923.7290.5580
EDSR [103]29.8300.805824.3660.5715
RDN [230]29.9130.807424.3920.5711
RCAN [229]29.9160.808524.4040.5749
SWD-Net [29]29.8530.800024.4650.5755
CARN [63]29.9640.840824.4790.5763
4KAgent (ExpSR-sN-F) (N ∈ [4, 8])29.7460.860224.3000.5826
+ +Qualitative Comparison Fig. 15 presents a qualitative comparison for $4 \times$ super-resolution on representative patches from the bcSR test set. While both methods significantly improve upon the heavily blurred low-resolution inputs, a closer inspection of the ROIs reveals that our training-free 4KAgent consistently produces results that are on par with, and often superior to, the fully-trained, domain-specific CARN model. This superior performance is particularly evident in challenging cases. In image 1010, 4KAgent successfully delineates individual cell boundaries and restores the heterogeneous texture of the tissue architecture. In contrast, CARN's output suffers from a loss of sharpness and definition, resulting in noisier and blurrier cell regions and poorly defined tissue architecture, while also introducing a slight color deviation. Similarly, in image 1175, 4KAgent accurately reconstructs the intricate internal structures, preserving the sharp outlines of the nuclei and cytoplasm. CARN's output, conversely, suffers from a loss of sharpness and inaccurate detail, while also exhibiting subtle grid-like artifacts. Across all examples, 4KAgent consistently generates nuclei with sharper boundaries and more distinct internal textures, along with clearer cell membranes, demonstrating a higher fidelity to the ground truth. + +These visual improvements also directly correlate with 4KAgent's higher SSIM scores, confirming its enhanced ability to preserve the structural integrity of the tissue. The accurate recovery of such fine-grained morphological details is critical for potential downstream clinical applications. High-fidelity reconstructions like those from 4KAgent can enable more reliable automated analysis, such as precise nuclei segmentation for cell counting, classification of cellular atypia, and grading of cancerous tissue, thereby highlighting the potential value of our approach in digital pathology workflows. + +**Discussions** The combined quantitative and qualitative results underscore the significant potential of 4KAgent for pathology image super-resolution. Although not leading in PSNR, 4KAgent's superior SSIM scores demonstrate a more accurate reconstruction of high-frequency textures and tissue morphology, which are paramount for pathological interpretation. Furthermore, because 4KAgent is not trained on a specific pathology dataset, it is less susceptible to overfitting to the characteristics of a single data source. This provides a significant advantage when performing SR + +![](images/0962e65788c09dde0dd5428f491783a74c10ad310049248e3e16c98bef84646b.jpg) +1010 + +![](images/226e86b599644f80cdad8c8c84c523972020055a676c01e5406f9e017ddcca76.jpg) +LR + +![](images/57a53f2978afe96aac5a0342b24f971e00c6208dbe8bee27813bd2087c17773c.jpg) +CARN + +![](images/c12d0d00c89b4d388ed61f6ebc13c3f8dbe42147893f23b79b4f5585bbe743f8.jpg) + +![](images/7fba0e47530b9678ef3cc0dc1c3c17a603d20f6eea6f6f5733c6471c259d1bd2.jpg) +GT + +![](images/aec0d8ca21509061d70f72a3156fe68c8d76159499ce50b7f4580de60a189cf5.jpg) +1057 + +![](images/b99c5b83c2c2b695d760302d243c6218db0e034744091efbefddcd8dd63364cc.jpg) +LR + +![](images/8bea0c21398362ae5c1de47b47ce0ed881edb5b34938a428d232a3bc6716553e.jpg) + +![](images/9f2dfa189cebb9630bb790c3702a3392a068bfdeac1975744e5e7dc5d18fdc3b.jpg) +4KAgent (ExpSR-s4-F) + +![](images/523a19d53647595b4da64571c80ea7678ea3775c31aa235b258c0aa79fbef2a6.jpg) + +![](images/f413d69204e0cafc7724914b6ef9a172d85ab6b84a58f1eaaf19bec31e7b14f0.jpg) +1123 + +![](images/77d4585bc7a8ff59a71991e3c10908e89f863fa5db5c861e31db92ce08f465e2.jpg) +LR + +![](images/e2e1d5048d1c76893d7b988fd3a0fc5b0152dced65740143d0464bf730e6006f.jpg) +CARN +CARN + +![](images/d68d53eadca656e419a999e620793048b5b1a118133a2b9ffed2eccfdb26e7a5.jpg) +4KAgent (ExpSR-s4-F) +4KAgent (ExpSR-s4-F) + +![](images/c3813c3fa4dfb2e55634ad32bb223fbd14b2a13b8978c33b58f065e2b4a6cfbe.jpg) +GT +GT + +![](images/31f94d22cfdb99a170b41a04d946e46a211ef8e09dcf3f85a53a7da08e737472.jpg) +1175 +Figure 15: Visual comparison of pathology image super-resolution on bcSR dataset (256→1024). + +![](images/bb9e3472e61fe3df83bc1771b63ce599b119a9ba1d64d0dc8eba090779bd3002.jpg) +LR + +![](images/c6a0e8811b7dac29b04fa0b3548173d9ba9187aa5f32664c435a292268e83771.jpg) +CARN + +![](images/b539c318fb4c05e4a2d5b9e467bedb3389f3a9e5e0f81cfc78e7f80613b84d60.jpg) +4KAgent (ExpSR-s4-F) + +![](images/60adcf89e62d41032f0017a4a61872aabcf3fb009c406d15d193193872964a7a.jpg) +GT + +on real-world pathology images acquired by different scanners and staining protocols. Additionally, 4KAgent's agentic framework allows for flexible expansion of its tool profile to better adapt to pathology imaging modalities. Its leading performance on the bcSR dataset validates the potential of this agentic approach as a robust and generalizable solution for biomedical imaging. + +The ability to generate reconstructions with high structural fidelity has direct implications for critical downstream applications in computational pathology. By restoring sharper nuclei, clearer cell membranes, and more intelligible tissue architecture, 4KAgent provides a more reliable input for automated analysis pipelines. This can enhance the accuracy of tasks such as nuclei segmentation, cell counting, and the classification of cancerous tissue, ultimately making it a more robust and practically useful tool for AI-assisted biomedical diagnostics. + +# F.4 Medical Image Super-Resolution: X-Ray, Ultrasound, and Fundoscopy + +In this chapter, we shift our focus to the super-resolution of clinical diagnostic imaging modalities, where the primary goal is to enhance anatomical and pathological details for improved diagnostic accuracy while minimizing patient burden, such as reducing exposure to ionizing radiation in X-ray imaging [160]. Although often grouped together, these modalities can be fundamentally diverse, operating on different physical principles: from X-rays utilizing ionizing radiation to ultrasound relying on acoustic waves. This diversity gives rise to unique image characteristics and modality-specific challenges, such as maintaining pathological invariance in chest X-rays or avoiding the generation of pseudo-structures in ultrasound images. + +The prevailing approach in medical image SR has been the development of highly specialized models, each trained on specific datasets for a single modality [207]. The rise of foundation models + +has led to more powerful specialist systems, such as those tailored for a single modality like CT or dermatology [136, 71], though a few pioneering works have begun to explore more universal solutions [104]. A significant drawback of such a specialized paradigm is poor generalization across different datasets and modalities, which creates a major bottleneck for practical clinical deployment and motivates our evaluation of 4KAgent's performance on these challenges. To this end, this chapter evaluates 4KAgent's performance across several distinct and clinically important modalities. + +Settings We evaluate 4KAgent with the ExpSR-s4-F profile across three medical imaging modalities with distinct imaging principles: X-ray, ultrasound, and fundoscopy, benchmarking against their respective baselines. Benchmark datasets of each imaging modality are summarized as follows: + +- X-ray. Chest X-ray 2017 [70] and Chest X-ray 14 [170]. Specifically, Chest X-ray 2017 is a dataset of 5,856 pediatric images from Guangzhou Women and Childrens Medical Centre, split into 5,232 images for training and 624 images for testing. Chest X-ray 14 contains 112,120 frontal-view X-rays of 30,805 patients with 14 disease labels mined from radiology reports. Among them, 880 images additionally contain expert-annotated bounding boxes. Following the settings in [196], we evaluate on the Chest X-ray 2017 test set and the 880 annotated data in Chest X-ray 14. +- Ultrasound Image. US-Case [150] and MMUS1K [133]. The US-Case collection comprises over 7,000 sonographic images spanning organs such as the liver, heart, and mediastinum. Adopting the selection protocol from [133], we reused the subset of 111 images in the test set for benchmarking, excluding 11 scans whose small field of view limited their diagnostic value. MMUS1K features 1,023 anonymized multi-organ ultrasound scans, including bladder, gallbladder, thyroid, kidney, etc., sourced from Shanghai Tenth Peoples Hospital. All images meet a minimum resolution of $448 \times 600$ px and were cleansed of watermarks and blurring artifacts via LabelImg. The test set with label numbers from 0801 to 0900 was used for evaluation. +- Fundoscopy Image. DRIVE [151] consists of 40 color fundus images from a diabetic retinopathy screening program in the Netherlands collected by a Canon CR5 non-mydriatic 3CCD camera. The dataset is equally divided into 20 for training and 20 for testing. As all images are in $584 \times 565$ resolution, following the setup in [5], the original high-resolution (HR) images were resized to $512 \times 512$ and before being further utilized to generate LR pairs. + +To consist with baseline methods for each dataset, the LR input images were generated by down-sampling HR images via bicubic interpolation. For X-ray images, we use SSIM, FSIM [224], and MSIM [178] as metrics, while PSNR and SSIM are used for Ultrasound and Fundoscopy images. + +Table 20: Quantitative comparison on X-ray datasets. The top three performances of each metric are marked in **bold**, **underline**, **italic** respectively. + +
MethodChest X-ray 2017Chest X-ray 14
SSIM↑FSIM↑MSIM↑SSIM↑FSIM↑MSIM↑
Nearest-Neighbor0.6370.6720.6680.7010.7240.713
Interpolation [197]0.6150.6630.6440.6870.6980.681
CTF [222]0.8890.9330.9540.9170.9550.943
ESPCN [147]0.7560.8250.8040.7950.8220.815
FSRCNN [41]0.8970.9430.9530.9170.9590.953
LapSRN [80]0.8930.9420.9540.9150.9560.949
SRGAN [82]0.8210.8960.8680.8440.9030.897
GAN-CIRCLE [204]0.8970.9470.9230.9190.9690.945
SNSRGAN [196]0.9110.9810.9830.9250.9950.986
4KAgent (ExpSR-s4-F)0.9330.9960.9870.9600.9990.993
+ +Quantitative Comparison X-ray Quantitative results are summarized in Tab. 20. Ultrasound Quantitative results are summarized in Tab. 21. Fundoscopy Quantitative results are summarized in Tab. 22, which collectively demonstrate 4KAgent's consistently superior performance across all + +three distinct medical imaging modalities. On the X-ray datasets, 4KAgent with Fidelity profile surpasses the specialized SNSRGAN [196] model across all structure-focused metrics. For Ultrasound imaging, it also achieves a significant performance leap, boosting the PSNR on the MMUS1K dataset by nearly 3 dB over the previous state-of-the-art, M2Trans [133]. Similarly, on the DRIVE Fundoscopy dataset, 4KAgent again sets a new performance benchmark, improving the PSNR from 37.72 to 41.52 and the SSIM from 0.91 to 0.95. This consistent outperformance across modalities, from the need for pathological invariance in X-rays to the clarity of fine vessels in fundoscopy, highlights the effectiveness and robustness of 4KAgent for diverse medical SR tasks. + +Qualitative Comparison Representative qualitative results for X-Ray SR, ultrasound SR, and fundoscopy SR are shown in Figs. 16 to 18. The visual outcomes generally align with our quantitative findings and suggest the potential benefits of 4KAgent for clinical imaging applications. + +For X-ray super-resolution, where maintaining pathological invariance is important, 4KAgent produces reconstructions with improved delineation of lung parenchyma and clearer visibility of rib cage contours. It appears to achieve this clarity while reducing some of the oversmoothing artifacts occasionally seen in other SR methods, thus helping to preserve diagnostic details that are crucial for identifying pulmonary abnormalities with greater confidence. + +In the ultrasound comparisons, 4KAgent shows a notable advantage. On the US-CASE example, it restores clearer tissue boundaries and more internal detail compared to the blurrier reconstruction from the M2Trans baseline. Similarly, for the MMUS1K image, 4KAgent appears to reduce speckle noise while enhancing anatomical definition, whereas the baseline result is affected by some noise and artifacts. In both cases, 4KAgent generates echogenic patterns that more closely resemble the ground truth, improving overall image fidelity. + +Table 21: Quantitative comparison on Ultrasound dataset. The top three performances of each metric are marked in bold, underline, italic respectively. + +
MethodUS-CASEMMUS1K
PSNR↑SSIM↑PSNR↑SSIM↑
Bicubic28.900.789228.240.7817
EDSR [103]30.820.849730.040.8326
SwinIR [101]28.500.783427.660.7758
ELAN [226]31.020.846430.400.8309
ESRT [115]30.840.837430.250.8235
HAT [27]28.720.781228.080.7582
M2Trans [133]31.320.851630.680.8392
4KAgent (ExpSR-s4-F)33.270.889533.580.8678
+ +Table 22: Quantitative comparison on Fundoscopy dataset. The top three performances of each metric are marked in bold, underline, italic respectively. + +
DatasetMethodPSNR↑SSIM↑
DRIVEBicubic25.200.86
SRGAN [82]34.220.88
Ahmad et al. [5]37.720.91
4KAgent (ExpSR-s4-F)41.520.95
+ +The fundoscopy results demonstrate 4KAgent's effectiveness in restoring details from degraded inputs. Compared to the LR image, 4KAgent's reconstruction of the retinal vascular network shows a clear improvement. The method produces sharper and more continuous vessels, resolving many of the fine micro-vessels and bifurcation points that are obscured in the LR version. The resulting image more closely resembles the HR ground truth, suggesting its potential to aid in retinopathy screening from lower-resolution captures without sacrificing critical diagnostic details. + +**Discussions** From both quantitative and qualitative perspectives, our evaluation suggests that 4KAgent is a capable system for cross-domain super-resolution across diverse clinical imaging modalities, showing competitive performance on X-ray, ultrasound, and fundoscopy datasets. This result is notable, as the prevailing approach often involves developing specialized models for each modality, a paradigm that can be limited by poor generalization across different scanners and protocols. By not relying on domain-specific training, 4KAgents agentic framework offers a flexible alternative, adaptively deploying its tools to address the unique challenges of each image, from enhancing the clarity of lung markings in chest radiographs to defining subtle echogenic interfaces in ultrasound and resolving fine vascular networks in fundoscopy. + +The ability to generate reconstructions with improved structural details may have implications for downstream applications. For example, improved sharpness in retinal vessels could aid in retinopathy screening; clearer ultrasound images help with tissue boundary delineation for segmentation; and more detailed X-rays could enhance the visibility of subtle pulmonary abnormalities. Ultimately, the + +![](images/d6fbd32d20c8471cd6800b970c30adf196d8371644e2b3bb41bab007535757e1.jpg) +Chest X-ray 2017, VIRUS-4275213-0001 + +![](images/95378f450086a7845f2651bc55733283a58b688559062d9d5dac570274d1282a.jpg) +LR + +![](images/b69bd6525f9e880907c161c3c4d9f98865888e29bc6a33a5a4eccdbfc39575d2.jpg) +SNSRGAN + +![](images/a425fd80129fb4a20076e342b2da0291440a116bbd00ca5af4679abfc1c013ee.jpg) +4KAgent (ExpSR-s4-F) + +![](images/109faf0a8165550f95b2705767eca047f653a7351efb608633afae92b60101d8.jpg) +GT + +![](images/bd3c99714afe424b3c16fdc732b8373c59c8335a2e264c19d553dfdc9e329bcb.jpg) +Chest X-ray 14,00028454 016 +Figure 16: Visual comparison of X-Ray image SR on Chest X-ray 2017 and Chest X-ray 14 dataset. + +![](images/907ba93d7400e94d98e9e931ccd64d92b1afa12d85eae346363a5e91034e7716.jpg) +LR + +![](images/1b88210a9f3b0968ae32e066a3a6459ed94ac83ee813d269534aad510c37e443.jpg) +SNSRGAN + +![](images/9317c37e1c4f75102268b5f5f5fe686e85f42ce39f00d450efaa427b3b468494.jpg) +4KAgent (ExpSR-s4-F) + +![](images/5e19d6a81fcdae94bdf89e6baf53c2c0d85f50414cae084b38e65004fb9b94f1.jpg) +GT + +![](images/2d80c27e0963050fb5bc1c0cf8a88f395f574b877f9146baf59a2c87a776dc68.jpg) +US-CASE,113 + +![](images/63bd9a2d2a6592c361f8fa6ac57de0d100387ea5074c39c0225d4f5d4a677c98.jpg) +LR + +![](images/a634c9462f3fc9b3920255647f4d343ef5ae0f928197624dd727af1d88883fd1.jpg) +M2Trans + +![](images/85d75481da7a0698b8a52b6db165c9402fa0b8b35bbcbd01b571fa1875a34744.jpg) + +![](images/8f7dbff82bcec49ddde70edfa7fdb74aa64e16dbf4b2f33944bc8bad6797422f.jpg) +GT + +![](images/866d622f6dbd11e0e38a3976cb8d792fe6f97076cfdc5115abca3520812c1e6a.jpg) +MMUS1K,0857 +Figure 17: Visual comparison of Ultrasound image SR on US-CASE and MMUS1K dataset. + +![](images/68b75e8ebcccd514ed2a23fa133704442d68b3439cc1476b39e0db397835b69d.jpg) +LR + +![](images/5306e4e361c32fadc190fe53ffa3e7fa57013377df7164222231bb9ed653fc88.jpg) +M2Trans + +![](images/a33fd815323e2e9350ffb18c5068725b4f1d0cadfd754dee20b067a5ab126322.jpg) +4KAgent (ExpSR-s4-F) +4KAgent (ExpSR-s4-F) + +![](images/b514fdd8de5d2e5e1de6039597873196a8b917641eff5d52a0505966e97d7e02.jpg) +GT + +![](images/414c5f9d3698e53fa377816cd671ddf0847a7d17dbae6c7961117794b86e8cf2.jpg) +Figure 18: Visual comparison of fundoscopy image SR on DRIVE dataset (128→512). + +performance of our agentic approach indicates its significant potential for robust deployment across more real-world clinical workflows and imaging modalities, driven by its adaptability and inherent extensibility for incorporating more specialized medical profiles. + +# G Ablation Studies + +In this section, we conduct ablation studies on core components in 4KAgent system: (1) Q-MoE policy and (2) Face restoration pipeline. Then, we perform running time analysis of 4KAgent. + +Q-MoE policy. To assess the contribution of our Q-MoE mechanism during execution and reflection, we perform an ablation study in which Q-MoE is replaced by the DFS strategy from AgenticIR [235], denoting this variant as 4KAgent (DFS). Experiments are conducted on the MiO-100 Group C dataset under the multiple-degradation image restoration setting. + +As shown in Tab. 23, integrating Q-MoE leads to substantial improvements in perceptual quality. Specifically, metrics such as LPIPS, MANIQA, CLIPIQA, and MUSIQ exhibit significant gains, with minimal impact on fidelity metrics like PSNR and SSIM. Furthermore, the visual comparisons presented in Fig. 19 provide additional evidence, showing that 4KAgent equipped with Q-MoE generates noticeably sharper and more realistic details compared to the DFS-based variant. + +Table 23: Ablation study on Q-MoE policy. The better performance is marked in bold. + +
DatasetMethodPSNR↑SSIM↑LPIPS↓MANIQA↑CLIPiQA↑MUSIQ↑
MiO100 - Group C4KAgent (DFS)19.810.57850.43810.32860.485454.03
4KAgent (Q-MoE)19.770.56290.42710.35450.523355.56
+ +![](images/f05dee3efd5ba272d564dab7167aa20d5f55598b97f5ae07ffe95fbf2b2d1aa1.jpg) +Figure 19: Visual comparisons for ablation study on Q-MoE. + +![](images/e6c719860f86613bf2c157f0cf5461d4ba201931ccd613f56b697d4d6a1c6e97.jpg) +LQ + +![](images/5ec634f284d5458fc404bc847e52ab36e7193234e515e02424322d53f52c6c4d.jpg) +4KAgent (DFS) + +![](images/0362f734b794fb5085ea768c758d8012fe330399b5ba1594e8f568bb2fedfd02.jpg) +4KAgent (Q-MoE) + +![](images/5ff1f5b72ec704077a590c93ae6b9fa8905c29c9b4fc962081fa691f6b5c229a.jpg) +HQ + +Face restoration pipeline. To evaluate the impact of our Face Restoration Pipeline, we conduct an ablation study on the WebPhoto-test dataset using three profiles: ExpSR-s4-P, ExpSRFR-s4-P, and GenSRFR-s4-P. Experiment result and the difference among these profiles are shown in Tab. 24. + +Enabling the face restoration module (i.e., switching profile from ExpSR-s4-P to ExpSRFR-s4-P and GenSRFR-s4-P) yields higher face IQA scores (CLIB-FIQA and DSL-FIQA). Moreover, when setting the 'Restore Option' to 'None' rather than 'super-resolution', we observe further improvements across both generic image IQA metrics (NIQE and MUSIQ) and face IQA metrics. + +Table 24: Ablation study on face restoration pipeline on WebPhoto-Test dataset. The best and second-best performances are marked in **bold** and **underline** respectively. + +
MethodRestore OptionFace RestoreNIQE↓CLIPIQA↑MUSIQ↑MANIQA↑CLIB-FIQA↑DSL-FIQA↑
4KAgent (ExpSR-s4-P)super-resolutionFalse5.110.711973.620.66010.64150.7194
4KAgent (ExpSRFR-s4-P)super-resolutionTrue4.530.660072.890.64050.66020.7237
4KAgent (GenSRFR-s4-P)NoneTrue4.150.707775.920.65760.66710.7683
+ +Visual comparisons are shown in Fig. 20. 4KAgent with GenSRFR-s4-P profile produces the finest facial details (hair details, harmony of facial area and background area). This trend indicates that WebPhoto-test images suffer from complex, mixed degradations and 4KAgent benefits from integrating multiple restoration tasks with Q-MoE driven selection to achieve superior visual quality. + +Running Time Analysis. The inference time of 4KAgent varies depending on the selected profile, the quality of the input image, and the length of the restoration plan. In this section, we analyze the inference time of 4KAgent using NVIDIA RTX 4090 GPUs. Specifically, we report the fastest and slowest cases observed in our experiments. The fastest case involves super-resolving images $(\times 4)$ from the B100 dataset using the ExpSR-s4-F profile. The slowest case corresponds to jointly + +![](images/3062bf3b5b484aaec09b40312a55b7ffe1b6c318249426a5274ea0518a342cbf.jpg) +LQ + +![](images/42427b92369c82122f6c2902d7c62bac854a521b85e3d09ef715cafc2f4c75ee.jpg) +4KAgent (ExpSR-s4-P) + +![](images/aafbc79af63b11cbb03b1f1e618ed73bbb4cba389e6cae00b46a6027f5165eea.jpg) +4KAgent (ExpSRFR-s4-P) +Figure 20: Visual comparisons for ablation study on face restoration pipeline. + +![](images/8e0eba52e936dc8479c25ac478ff96c11a6e03f1512d8a0fa0f5c113f2ee9228.jpg) +4KAgent (GenSRFR-s4-P) + +restoring and upscaling low-quality images from the DIV4K-50 dataset to 4K resolution under the Gen4K-P profile. The inference times for these two cases are summarized in Tab. 25. + +Table 25: Inference time of 4KAgent (fastest and slowest cases on our experiments). + +
Profile NameTaskResolutionBenchmarkLength of PlanInference Time (s)
ExpSR-s4-FSuper-resolution (4×)120 × 80 → 480 × 320B1001.0 ± 0.050.96 ± 2.01
Gen4K-PJoint restoration + 4K Upscaling256 × 256 → 4096 × 4096DIV4K-503.4 ± 0.61551.76 ± 230.73
+ +As 4KAgent currently executes its tools sequentially, there is substantial potential for acceleration, for example, by running independent restoration tools in parallel at each step. + +# H Applications and Broader Impacts + +# H.1 Applications + +High-resolution On-Demand Media Streaming 4KAgent offers significant potential for enabling network operators, such as YouTube, Netflix, Instagram, Amazon Prime Video, TikTok, Kwai, Snap, Twitch, to name a few, to deliver 4K-quality video services from much lower-bitrate streams. For example, edge-based SR can upscale a 1K stream to 4K at the user's device [214], allowing providers to store and transmit mostly lower-resolution content (e.g., 1K) which can then be upscaled to 4K quality on the end-user's device using edge-based processing. This approach dramatically cuts storage and bandwidth costs (and even energy use) compared to naively streaming native 4K [91]. Technologies like NVIDIA's Deep Learning Super Sampling (DLSS) [1] demonstrate the feasibility and usability of real-time super-resolution on GPU chips. Integrating such real-time upscaling into adaptive streaming protocols could also improve user experience by minimizing disruptive quality shifts often associated with variable network conditions, ensuring viewers consistently receive high-resolution playback on capable displays. + +Video Conferencing and Telepresence Network bandwidth constraints and limitations inherent in typical webcams or smartphone cameras often necessitate transmitting video streams at resolutions lower than 4K. Implementing SR algorithms, such as 4KAgent, on the receiver's end can effectively upscale these lower-resolution feeds. This process restores fine-grained details in facial features or gestures that might otherwise be lost, thereby enhancing the perceived visual quality and potentially aiding communication cues like lip-reading or the interpretation of subtle expressions [130, 191, 95]. Consequently, even devices with modest camera capabilities can deliver an experience approximating 4K quality to the viewer, without requiring increased upload bandwidth from the sender. This democratization of high-resolution video conferencing can improve remote collaboration, making it more accessible and effective for users constrained by network limitations or hardware capabilities. + +Surveillance and Security Image SR technologies like 4KAgent (with fidelity-based profile) offer significant value in enhancing footage from law enforcement operations, particularly from body-worn cameras and dashcams. These devices often capture video at resolutions like $720\mathrm{p}$ or $1080\mathrm{p}$ with wide fields of view, resulting in low-detail imagery, especially in challenging conditions such as low + +light [92]. Faces or license plates captured at a distance may span only a few dozen pixels, far below recommended thresholds for identification (e.g., $\sim 90\times 60$ pixels per face for courtroom evidence [2]). The quality is often further compromised by heavy compression and sensor limitations, introducing noise and motion blur. Modern SR approaches, particularly "blind" methods that model complex real-world degradations, can effectively mitigate these issues and restore detail in practical bodycam footage. By enhancing critical regions (faces, license plates) in police videos, SR can improve both human and automated identification, while preserving the veracity required for judicial use. + +Similarly, public surveillance systems, including city-wide CCTV networks, border security cameras, and transit hub monitoring, face comparable challenges related to resolution and image quality. Fixed cameras covering wide areas often render persons or objects of interest with very low pixel counts, with quality impacted by distance, illumination, camera motion, and aggressive compression techniques employed to manage bandwidth and storage [131]. SR provides a means to enhance detail retroactively without costly hardware upgrades. Field studies have also reported the effectiveness of SR. For example, a National Institute of Justice study [3] showed that multi-resolution SR could reconstruct identifiable features from extremely low-resolution facial images comparable to those from real-world security cameras. Overall, SR can act as a force multiplier for legacy surveillance infrastructure, enhancing situational awareness and forensic capabilities. However, the enhanced capability for identification also raises potential privacy concerns, which will be discussed in Appendix H.3. + +Gaming and Entertainment SR techniques are extensively utilized in the entertainment sector to enhance visual quality while sustaining high frame rates, particularly in demanding gaming, VR, and AR applications. A prominent example is NVIDIA's DLSS, a suite of AI-powered neural rendering techniques that upscale lower-resolution frames to higher target resolutions, as high as 4K. DLSS can significantly improve performance, often more than doubling GPU throughput and leading to substantial frame rate increases for instance, one report indicated a boost of up to approximately $360\%$ (e.g., from 8 to 36.8 fps on an RTX 2060 at 4K). Successive iterations like DLSS 3 with Frame Generation and DLSS 3.5 with Ray Reconstruction have introduced further advancements by using AI to generate additional frames or improve ray-traced effects. + +VR, AR, and XR This need for efficient, high-quality rendering extends critically to the domain of spatial intelligence and computing, as seen in advanced devices like the Apple Vision Pro, which aims to deliver experiences with more pixels than a 4K TV per eye. While such platforms boast high native display resolutions, SR techniques could play a crucial role in rendering complex mixed-reality scenes or high-fidelity passthrough video efficiently, maintaining visual clarity without overwhelming the processing capabilities. Similarly, as smart glasses like the Ray-Ban Meta Wayfarer evolve and potentially incorporate more advanced display capabilities for augmented reality overlays, SR will be key to delivering crisp digital information without excessive battery drain. Broader XR initiatives, such as Google's development of Android XR, also stand to benefit from robust SR solutions to enable a diverse ecosystem of devices to achieve compelling visual experiences. For all these platforms, from gaming consoles to sophisticated XR headsets and smart glasses, the ability of SR systems like 4KAgent to adaptively enhance visual quality from various inputs will be paramount in balancing immersive, high-resolution experiences with practical performance and power constraints. + +AI-Generated Content (AIGC) Production Industry Photographers, digital artists, and filmmakers increasingly leverage SR tools to enlarge, restore, and enhance the quality of both conventional (e.g., old photos, archival digital footage) and even AI-generated images and video footage. We have demonstrated in Appendix E.1 that 4KAgent is capable of synthesizing high-fidelity details in generated media, coinciding with a recent trend that generates a high-resolution advertisement for $\mathrm{KFC}^1$ , leveraging outputs from generative video models such as Google's Veo [163], Luma AI's Dream Machine [116], and OpenAI's Sora [134], further enhanced using Topaz Labs Video Upscaler to achieve higher resolutions (e.g., 4K) and professional quality suitable for broader use. SR techniques are crucial for bridging this gap towards generating ultra-high-resolution content, enabling creators to enhance these AI-generated visuals. For instance, images generated for concept art, marketing materials, or virtual environments can be significantly improved in detail and clarity through SR, making them suitable for 4K displays or large-format printing. Similarly, generative video content, which might be created at lower resolutions to manage computational costs, can be + +upscaled using specialized tools like Topaz Video AI to achieve crisp, higher-resolution results (e.g., 4K) ready for distribution or integration into larger productions. State-of-the-art SR methods, including GAN-based approaches, can synthesize photorealistic details, effectively transforming AIGC outputs into polished, professional-grade assets. The ability of robust and adaptive SR solutions like 4KAgent to handle the diverse and sometimes unpredictable nature of AIGC makes them particularly valuable for ensuring that AI-driven creative endeavors can meet high-quality benchmarks. + +Scientific Imaging High-resolution imagery is crucial across numerous scientific disciplines, particularly when native sensor capabilities are constrained. In remote sensing, deep super-resolution (SR) methods significantly enhance spatial details of satellite imagery, facilitating accurate land-use classification and environmental monitoring [132, 146]. For instance, self-supervised SR techniques trained on sequences of satellite images yield sharper and less noisy results compared to raw captures, substantially improving downstream geospatial analysis [132]. Microscopy and biomedical imaging similarly benefit from SR, particularly through novel quantum imaging techniques. Recent advancements by Zhang et al. and He et al. leverage quantum entanglement to achieve unprecedented imaging resolution, demonstrating quantum microscopy at the Heisenberg limit and significantly enhancing cellular and sub-cellular visualization [228, 54]. Additionally, computational SR methods applied to microscopy, like content-aware restoration for fluorescence images [182], complement these quantum techniques by computationally reconstructing detailed 3D biological structures from limited optical inputs. Thus, versatile and advanced SR frameworks such as 4KAgent, coupled with emerging quantum imaging methods, can revolutionize scientific research by providing richer, more precise imagery across multiple imaging modalities. + +Medical Image Applications In medical imaging, SR facilitates detailed diagnostics by transforming low-dose or rapidly acquired imaging scans into high-fidelity medical images. Techniques employing deep learning-based SR on modalities such as Magnetic Resonance Imaging (MRI) and Computed Tomography (CT) have shown promise in generating accurate, high-resolution images from suboptimal inputs, thus reducing patient radiation exposure and acquisition time without sacrificing diagnostic quality [28, 231]. For instance, generative adversarial networks (GANs) and transformer-based SR approaches like TransMRI demonstrate substantial improvements in enhancing anatomical details critical for diagnostic accuracy [44]. Consequently, methods like 4KAgent, which provide universal super-resolution capabilities, can significantly impact clinical diagnostics by offering highly detailed and diagnostically reliable imagery from resource-efficient imaging procedures. + +Embodied AI and Robotics Embodied AI systems, including robotics platforms, leverage SR to enhance visual perception, critical for tasks such as navigation, object manipulation, and human-robot interaction. Robotic visual systems frequently face limitations in sensor resolution and onboard processing capacity, challenges that SR methods can effectively address. Recent studies indicate that integrating SR into robotic vision pipelines notably improves object detection and localization, particularly for distant or small-scale objects critical in dynamic environments [122]. Furthermore, real-time lightweight SR models tailored for robotic platforms have been developed, improving perception accuracy and enabling robots to perform complex tasks efficiently, such as precise grasping and navigation through cluttered or visually challenging scenarios [7, 62]. Consequently, robust SR algorithms substantially advance robotic autonomy and operational effectiveness. + +Autonomous Vehicles, Drones, and Intelligent Transportation Systems (ITS) Autonomous vehicles, aerial drones, and intelligent transportation systems increasingly depend on high-quality visual inputs for safe and efficient operation. SR technologies significantly boost the capability of onboard cameras to discern critical details from lower-resolution captures under real-world constraints. For instance, SR-enhanced imagery improves the detection accuracy of pedestrian, vehicle, and traffic sign recognition systems, crucial for autonomous navigation and safety-critical decision-making [156, 128, 109]. Research demonstrates that training object detection models with super-resolved images significantly enhances their effectiveness in challenging scenarios, including poor visibility or long-range object detection from drones or vehicle-mounted cameras [128]. Furthermore, ITS can employ SR-enhanced camera networks for robust and precise traffic monitoring and anomaly detection, improving urban mobility management and public safety. + +GIScience and GeoAI Geographic Information Science (GIScience) heavily utilizes spatially explicit, high-resolution data for mapping, analysis, and policy-making. Deep SR has recently + +emerged as a powerful method for enhancing geospatial imagery resolution, substantially improving the interpretability and accuracy of satellite-derived GIS datasets. For example, SR methods significantly improve the extraction accuracy of buildings, roads, vegetation, and other geographic features from medium- to low-resolution imagery, providing a cost-effective alternative to deploying expensive high-resolution satellite sensors [146, 141]. Additionally, high-quality SR outputs are vital for applications such as precision agriculture and disaster response planning, demonstrating the broad utility of universal SR frameworks like 4KAgent within GIScience research and applications. + +# H.2 Broader Impacts + +Economic Impacts. SR technologies, exemplified by 4KAgent, drive significant economic advantages by enhancing operational efficiency, creating new markets, and supporting environmental sustainability. By reducing bandwidth, storage, and infrastructure costs associated with high-resolution content delivery, SR solutions enable economical, high-quality media dissemination over constrained networks, benefiting digital platforms and smaller businesses [98, 215]. Companies such as BytePlus and Maxar Intelligence have successfully leveraged SR technologies to open new markets in healthcare diagnostics, geospatial intelligence, and media restoration [15, 61]. Additionally, by minimizing data storage and transmission demands, SR contributes meaningfully to the environmental goals of reduced energy consumption and lower carbon emissions [85]. + +Accessibility. SR substantially promotes digital equity by enabling high-quality visual content access for users in regions with limited bandwidth or resource constraints without necessitating advanced infrastructure or expensive devices [139]. Particularly in education and healthcare, SR supports remote learning and telemedicine by delivering clearer instructional and diagnostic imagery, significantly benefiting underserved communities [14]. Moreover, SR-integrated assistive technologies offer improved accessibility for individuals with visual impairments by enhancing image clarity and text readability, facilitating greater inclusion and interaction within the digital sphere [102, 164]. + +Vertical Impacts to Industry. Demand for real-time, high-quality SR capabilities stimulates technological progress across various industries, including media, robotics, autonomous systems, and scientific visualization. SR advancement drives innovation in specialized neural hardware, edge computing solutions, and embedded AI, spurring the development of powerful and efficient Neural Processing Units (NPUs) in consumer devices [17, 83]. Additionally, SR techniques significantly enhance autonomous vehicle, drone, and robotic platform capabilities by improving object detection accuracy, scene understanding, and decision-making reliability, particularly in challenging operational environments [128, 146]. These impacts extend to scientific instrumentation, such as microscopy and geospatial imaging, where SR enables unprecedented detail and precision [132, 182]. + +# H.3 Limitations and Potential Negative Societal Impacts + +Efficiency and Computational Cost. SR methods, particularly deep learning-based and agentic frameworks like 4KAgent, often require substantial computational resources for training and inference. High-resolution SR models usually rely on resource-intensive GPU or TPU clusters, imposing significant energy consumption [148]. Even optimized inference can become computationally burdensome at the edge, potentially limiting deployment on low-powered or mobile devices unless significant model compression and acceleration techniques are applied [83, 227]. Balancing performance and efficiency remains a critical open challenge, particularly for real-time applications in resource-constrained environments. + +Bias, Fairness, and Model Drift. Data-driven SR models inherit biases from their training datasets, potentially leading to uneven quality across different image categories, demographics, or scenarios [47, 124]. Such biases might systematically disadvantage specific groups, for instance, by inadequately resolving images related to underrepresented populations or environments. Model drift over timewhere the model gradually becomes less accurate due to changing real-world distributions also poses a serious issue for practical deployment, requiring continuous monitoring and recalibration to ensure fairness and reliability [45, 114]. + +Ethical Issues and Privacy. Enhanced imaging capabilities enabled by SR, particularly in surveillance contexts, can amplify privacy risks. The ability to recover detailed features such as faces or + +license plates from previously anonymized or low-resolution imagery might lead to unauthorized or unethical identification of individuals [52]. This capability necessitates clear regulatory guidelines and ethical oversight to avoid misuse [97]. Concerns are especially pronounced in contexts of law enforcement, border surveillance, and public monitoring, where SR technologies must be carefully governed to prevent potential violations of civil liberties and personal privacy [34]. + +Failure Modes in High-stake Settings. The adoption of SR techniques in high-stakes environments, including medical diagnostics, autonomous vehicles, and security monitoring, introduces risks associated with model hallucinations or misleading image reconstructions [31]. SR models, particularly generative approaches, may produce plausible yet incorrect details absent from the original low-resolution inputs, potentially leading to erroneous interpretations or decisions [8]. In clinical settings, for example, SR-generated artifacts could result in incorrect diagnoses or overlooked medical conditions, underscoring the importance of rigorous validation and transparency regarding model uncertainty and reliability [69]. + +# I Related Works + +# I.1 Image Super-Resolution + +Deep learning has significantly advanced the field of single-image Super-Resolution (SR). The seminal work, SRCNN [40], introduced a convolutional net for SR, with a primary focus on optimizing the Mean Squared Error between the super-resolved and high-resolution images. Following this, numerous studies have enhanced reconstruction accuracy by improving network architectures, including residual and dense connections [72, 103, 230], attention mechanisms [18, 37, 229], and multi-scale networks [46, 93]. While these methods perform well in modeling the posterior distribution of the training data, they inevitably suffer from the issue of overly smooth visual results [82, 127, 192]. In recent years, significant efforts have been made to develop generative model-based SISR techniques that produce more visually appealing results. These include autoregressive models [36, 126, 162], GAN-based models [82, 172, 220, 19, 100, 89], and diffusion-based models [167, 199, 105, 184, 145, 175, 183]. SRGAN [82], as a pioneering GAN-based SR model, assumes image degradation through bicubic downsampling and generates photo-realistic images. BSRGAN [220] and Real-ESRGAN [172] achieve promising real-world SR results by using randomly shuffled degradation and higher-order degradation. SwinIR [101] replaces the CNN-based generator network with visual transformers, leading to more stable training and more realistic textures. Additionally, SeD [89] introduces a semantic-aware discriminator to capture fine-grained distributions by incorporating image semantics as a condition. Recent diffusion-based models have focused on fine-tuning the Stable Diffusion model for reconstructing high-quality images, using low-quality images as control signals. Notably, StableSR [167] fine-tunes a time-aware encoder and employs feature warping to balance fidelity and perceptual quality. SeeSR [184] introduces degradation-robust, tag-style text prompts to enhance the semantic awareness of the Real-ISR model. Furthermore, recent studies on diffusion-based models, such as SinSR [175], OSEDiff [183], PiSA-SR [153], and GuideSR [9] achieve one-step image super-resolution. + +# I.2 Image Restoration + +Recent advances in deep learning have led to remarkable progress in blind image restoration tasks, including denoising, deblurring, deraining, dehazing, and removal of JPEG compression artifacts. Early works such as ARCNN [39] demonstrated the potential of compact convolutional neural networks, particularly in the context of image denoising. Since then, a broad range of sophisticated network architectures and training strategies have been developed to further enhance restoration performance. These include the use of residual blocks [219, 112, 221], attention mechanisms [206, 22, 159, 49, 211], and Transformer-based designs [161, 236, 210, 176]; as well as generative paradigms such as GANs [50, 23, 11, 48, 60, 127, 137, 110] and diffusion models [105, 187, 66, 174, 67, 43]. Notably, general-purpose restoration models like Uformer [176], MAXIM [159], Restormer [210], and NAFNet [22] have demonstrated strong performance across diverse restoration tasks, often trained independently for each specific degradation type. However, such single-degradation methods often struggle in real-world scenarios where multiple types of degradations co-exist. This limitation has sparked growing interest in the emerging field of All-in-One image restoration, which aims to build unified models capable of handling a wide range of degradations with a single network [90, 119, 138, + +201, 216, 161]. For instance, AirNet [90] introduces a degradation classifier trained via contrastive learning to guide restoration, while ADMS [138] employs a multi-type degradation classifier to dynamically select Adaptive Discriminant filters, enabling degradation-specific parameter modulation within the restoration network. + +# I.3 LLM Agents + +Advancements in LLM-based frameworks have enabled more structured reasoning and agent designs, particularly for complex multimodal tasks. Initial efforts emphasized improving reasoning capabilities through refined prompting strategies and modular architecture. Chain-of-Thought (CoT) prompting [180] introduced stepwise reasoning, facilitating decomposition and interpretability across diverse tasks. ReAct [202] combined reasoning with tool interaction by interleaving thought traces and external actions, supporting more adaptive behavior. Extending this direction, CoALA [152] formalized components such as memory, reasoning, and control within a cognitive architecture, offering a modular design space for building general-purpose language agents. These developments established a basis for domain-specific agent systems with integrated reasoning pipelines. Building on these foundations, application-driven LLM agents have been developed to incorporate tool use and dynamic decision-making within specialized domains. In vision tasks, MMCTAgent [78], VideoAgent [169], and ReAgent-V [234] implement planning and evaluation pipelines for image and video analysis, incorporating external modules for retrieval and verification. In the medical domain, agents such as MedCoT [111], CLINICR [129], and MMedAgent-RL [190] employ hierarchical reasoning frameworks to address clinical questions, integrating structured logic and domain-specific knowledge to enhance interpretability and decision quality. + +Similarly, LLM-based agents have also emerged as a promising paradigm for tackling complex image restoration tasks involving multiple degradations. RestoreAgent [20] pioneered the use of MLLMs for autonomous task identification, model selection, and execution planning. AgenticIR [235] introduced a five-stage human-inspired workflowPerception, Scheduling, Execution, Reflection, and Reschedulingaugmented with self-exploration to build IR-specific experience. MAIR [65] advanced this by employing a multi-agent system guided by real-world degradation priors, improving both efficiency and scalability. HybridAgent [88] proposed a hybrid interaction scheme with fast and slow agents, along with a mixed-distortion removal strategy to mitigate error propagation. Q-Agent [233] further introduces a quality-driven chain-of-thought framework, leveraging no-reference IQA metrics to guide greedy restoration without costly rollbacks. These works demonstrate the growing potential of combining general-purpose language intelligence with visual tools for robust, adaptive image restoration. More recently, JarvisIR [106] and JarvisArt [107] leveraged intelligent agent workflows to perform task-oriented image restoration and creative photo retouching. + +# References + +[1] NVIDIA DLSS 4 Supreme Speed. Superior Visuals. Powered by AI. https://www.nvidia.com/en-us/geforce/technologies/dlss/.34 +[2] Video quality in public safety (VQiPS) workshop report. https://www.nist.gov/system/files/documents/2016/10/06/final_vqipsworkshopreport_092912.pdf.35 +[3] Ramzi Abiantun, Felix Juefei-Xu, Utsav Prabhu, and Marios Savvides. Ssr2: Sparse signal recovery for single-image super-resolution on faces with extreme low resolutions. Pattern Recognition, 90:308-324, 2019. 35 +[4] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 3 +[5] Waqar Ahmad, Hazrat Ali, Zubair Shah, and Shoaib Azmat. A new generative adversarial network for medical images super resolution. Scientific Reports, 12(1):9533, 2022. 30, 31 +[6] Meta AI. Llama 3.2-vision. https://huggingface.co/meta-llama/Llama-3.2-11B-Vision-Instruct, 2024.3 +[7] Simone Angarano, Francesco Salvetti, Mauro Martini, and Marcello Chiaberge. Generative adversarial super-resolution at the edge with knowledge distillation. Engineering Applications of Artificial Intelligence, 123:106407, 2023. 36 +[8] Vegard Antun, Francesco Renna, Clarice Poon, Ben Adcock, and Anders C Hansen. On instabilities of deep learning in image reconstruction and the potential costs of ai. Proceedings of the National Academy of Sciences, 117(48):30088-30095, 2020. 38 +[9] Aditya Arora, Zhengzhong Tu, Yufei Wang, Ruizheng Bai, Jian Wang, and Sizhuo Ma. Guidesr: Rethinking guidance for one-step high-fidelity diffusion-based super-resolution. arXiv preprint arXiv:2505.00687, 2025.38 +[10] Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, et al. Qwen2.5-vl technical report. arXiv preprint arXiv:2502.13923, 2025. 3 +[11] David Bau, Hendrik Strobelt, William Peebles, Jonas Wulff, Bolei Zhou, Jun-Yan Zhu, and Antonio Torralba. Semantic photo manipulation with a generative image prior. arXiv preprint arXiv:2005.07727, 2020.38 +[12] Soufiane Belharbi, Mara Whitford, Phuong Hoang, Shakeeb Murtaza, Luke McCaffrey, and Eric Granger. Sr-caco-2: A dataset for confocal fluorescence microscopy image super-resolution. Advances in Neural Information Processing Systems, 37:59948-59983, 2024. 6, 25 +[13] Marco Bevilacqua, Aline Roumy, Christine Guillemot, and Marie Line Alberi-Morel. Low-complexity single-image super-resolution based on nonnegative neighbor embedding. 2012. 6 +[14] BytePlus. Business growth through superior technology. https://www BYTEplus.com, 2025.37 +[15] BytePlus. Unleashing the power of super resolution: A game-changer for visual content. https://www.byteplus.com/en/topic/96403, 2025.37 +[16] Jianrui Cai, Hui Zeng, Hongwei Yong, Zisheng Cao, and Lei Zhang. Toward real-world single image super-resolution: A new benchmark and a new model. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3086-3095, 2019. 6 +[17] Jung-Woo Chang, Keon-Woo Kang, and Suk-Ju Kang. An energy-efficient fpga-based deconvolutional neural networks accelerator for single image super-resolution. IEEE Transactions on Circuits and Systems for Video Technology, 30(1):281-295, 2018. 37 +[18] Chaofeng Chen, Dihong Gong, Hao Wang, Zhifeng Li, and Kwan-Yee K Wong. Learning spatial attention for face super-resolution. IEEE Transactions on Image Processing, 30:1219-1231, 2020. 38 +[19] Chaofeng Chen, Xinyu Shi, Yipeng Qin, Xiaoming Li, Xiaoguang Han, Tao Yang, and Shihui Guo. Realworld blind super-resolution via feature matching with implicit high-resolution priors. In Proceedings of the 30th ACM International Conference on Multimedia, pages 1329-1338, 2022. 38 + +[20] Haoyu Chen, Wenbo Li, Jinjin Gu, Jingjing Ren, Sixiang Chen, Tian Ye, Renjing Pei, Kaiwen Zhou, Fenglong Song, and Lei Zhu. Restoreagent: Autonomous image restoration agent via multimodal large language models. arXiv preprint arXiv:2407.18035, 2024. 39 +[21] Junsong Chen, Chongjian Ge, Enze Xie, Yue Wu, Lewei Yao, Xiaozhe Ren, Zhongdao Wang, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart- $\sigma$ : Weak-to-strong training of diffusion transformer for 4k text-to-image generation. In European Conference on Computer Vision, pages 74–91. Springer, 2024. 16 +[22] Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun. Simple baselines for image restoration. In European conference on computer vision, pages 17-33. Springer, 2022. 4, 38 +[23] Shengjie Chen, Shuo Chen, Zhenhua Guo, and Yushen Zuo. Low-resolution palmprint image denoising by generative adversarial networks. Neurocomputing, 358:275-284, 2019. 38 +[24] Wei-Ting Chen, Gurunandan Krishnan, Qiang Gao, Sy-Yen Kuo, Sizhou Ma, and Jian Wang. Dsl-fiqa: Assessing facial image quality via dual-set degradation learning and landmark-guided transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2931-2941, 2024. 12 +[25] Xiangyu Chen, Zheyuan Li, Yuandong Pu, Yihao Liu, Jiantao Zhou, Yu Qiao, and Chao Dong. A comparative study of image restoration networks for general backbone network design. In European Conference on Computer Vision, pages 74-91. Springer, 2024. 4, 7, 8 +[26] Xiangyu Chen, Xintao Wang, Wenlong Zhang, Xiangtao Kong, Yu Qiao, Jiantao Zhou, and Chao Dong. Hat: Hybrid attention transformer for image restoration. arXiv preprint arXiv:2309.05239, 2023. 4, 7, 8, 13, 15, 19, 20, 21 +[27] Xiangyu Chen, Xintao Wang, Jiantao Zhou, Yu Qiao, and Chao Dong. Activating more pixels in image super-resolution transformer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22367-22377, 2023. 31 +[28] Yuhua Chen, Yibin Xie, Zhengwei Zhou, Feng Shi, Anthony G Christodoulou, and Debiao Li. Brain mri super resolution using 3d deep densely connected neural networks. In 2018 IEEE 15th international symposium on biomedical imaging (ISBI 2018), pages 739-742. IEEE, 2018. 36 +[29] Zhen Chen, Xiaoqing Guo, Chen Yang, Bulat Ibragimov, and Yixuan Yuan. Joint spatial-wavelet dual-stream network for super-resolution. In Medical Image Computing and Computer Assisted Intervention-MICCAI 2020: 23rd International Conference, Lima, Peru, October 4-8, 2020, Proceedings, Part V 23, pages 184-193. Springer, 2020. 28 +[30] Shu-Chuan Chu, Zhi-Chao Dou, Jeng-Shyang Pan, Shaowei Weng, and Junbao Li. Hmanet: Hybrid multi-axis aggregation network for image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6257-6266, 2024. 4, 6, 7, 8 +[31] Joseph Paul Cohen, Margaux Luck, and Sina Honari. Distribution matching losses can hallucinate features in medical image translation. In Medical Image Computing and Computer Assisted Intervention-MICCAI 2018: 21st International Conference, Granada, Spain, September 16-20, 2018, Proceedings, Part I, pages 529-536. Springer, 2018. 38 +[32] Marcos V Conde, Gregor Geigle, and Radu Timofte. Instructir: High-quality image restoration following human instructions. In European Conference on Computer Vision, pages 1-21. Springer, 2024. 11 +[33] Julien Cornebise, Ivan Oršić, and Freddie Kalaitzis. Open high-resolution satellite imagery: The worldstrat dataset—with application to super-resolution. Advances in Neural Information Processing Systems, 35:25979-25991, 2022. 6, 19 +[34] Kate Crawford. Atlas of AI: Power, Politics, and the Planetary Costs of Artificial Intelligence. Yale University Press, 2021. 38 +[35] Yuning Cui, Wenqi Ren, Xiaochun Cao, and Alois Knoll. Revitalizing convolutional network for image restoration. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 4 +[36] Ryan Dahl, Mohammad Norouzi, and Jonathon Shlens. Pixel recursive super resolution. In Proceedings of the IEEE international conference on computer vision, pages 5439-5448, 2017. 38 +[37] Tao Dai, Jianrui Cai, Yongbing Zhang, Shu-Tao Xia, and Lei Zhang. Second-order attention network for single image super-resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11065-11074, 2019. 38 + +[38] Keyan Ding, Kede Ma, Shiqi Wang, and Eero P Simoncelli. Image quality assessment: Unifying structure and texture similarity. IEEE transactions on pattern analysis and machine intelligence, 44(5):2567-2581, 2020. 6 +[39] Chao Dong, Yubin Deng, Chen Change Loy, and Xiaou Tang. Compression artifacts reduction by a deep convolutional network. In Proceedings of the IEEE international conference on computer vision, pages 576-584, 2015. 38 +[40] Chao Dong, Chen Change Loy, Kaiming He, and Xiaou Tang. Learning a deep convolutional network for image super-resolution. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part IV 13, pages 184-199. Springer, 2014. 26, 28, 38 +[41] Chao Dong, Chen Change Loy, and Xiaou Tang. Accelerating the super-resolution convolutional neural network. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 391-407. Springer, 2016. 30 +[42] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In *Forty-first international conference on machine learning*, 2024. 16 +[43] Ben Fei, Zhaoyang Lyu, Liang Pan, Junzhe Zhang, Weidong Yang, Tianyue Luo, Bo Zhang, and Bo Dai. Generative diffusion prior for unified image restoration and enhancement. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9935-9946, 2023. 38 +[44] Chun-Mei Feng, Yunlu Yan, Huazhu Fu, Li Chen, and Yong Xu. Task transformer network for joint mri reconstruction and super-resolution. In Medical Image Computing and Computer Assisted Intervention—MICCAI 2021: 24th International Conference, Strasbourg, France, September 27–October 1, 2021, Proceedings, Part VI 24, pages 307–317. Springer, 2021. 36 +[45] João Gama, Indre Žliobaite, Albert Bifet, Mykola Pechenizkiy, and Abdelhamid Bouchachia. A survey on concept drift adaptation. ACM computing surveys (CSUR), 46(4):1-37, 2014. 37 +[46] Shangqi Gao and Xiahai Zhuang. Multi-scale deep neural networks for real image super-resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, 2019. 38 +[47] Timnit Gebru, Jamie Morgenstern, Briana Vecchione, Jennifer Wortman Vaughan, Hanna Wallach, Hal Daumé Iii, and Kate Crawford. Datasheets for datasets. Communications of the ACM, 64(12):86-92, 2021. 37 +[48] Jinjin Gu, Yujun Shen, and Bolei Zhou. Image processing using multi-code gan prior. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3012-3021, 2020. 38 +[49] Shuhang Gu, Yawei Li, Luc Van Gool, and Radu Timofte. Self-guided network for fast image denoising. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2511-2520, 2019. 38 +[50] Ishaan Gulrajani, Faruk Ahmed, Martin Arjovsky, Vincent Dumoulin, and Aaron C Courville. Improved training of wasserstein gans. Advances in neural information processing systems, 30, 2017. 38 +[51] Muhammad Haris, Gregory Shakhnarovich, and Norimichi Ukita. Deep back-projection networks for super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1664-1673, 2018. 26 +[52] Woodrow Hartzog. Privacy's blueprint: The battle to control the design of new technologies. Harvard University Press, 2018. 38 +[53] Yutong He, Dingjie Wang, Nicholas Lai, William Zhang, Chenlin Meng, Marshall Burke, David Lobell, and Stefano Ermon. Spatial-temporal super-resolution of satellite imagery via conditional pixel synthesis. Advances in Neural Information Processing Systems, 34:27903-27915, 2021. 18 +[54] Zhe He, Yide Zhang, Xin Tong, Lei Li, and Lihong V Wang. Quantum microscopy of cells at the heisenberg limit. Nature Communications, 14(1):2441, 2023. 36 +[55] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6 + +[56] Shane M Hickey, Ben Ung, Christie Bader, Robert Brooks, Joanna Lazniewska, Ian RD Johnson, Alexandra Sorvina, Jessica Logan, Carmela Martini, Courtney R Moore, et al. Fluorescence microscopy: an outline of hardware, biological handling, and fluorophore considerations. *Cells*, 11(1):35, 2021. 25 +[57] Chih-Chung Hsu, Chia-Ming Lee, and Yi-Shiuan Chou. Drct: Saving image super-resolution away from information bottleneck. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6133-6142, 2024. 4, 6, 7, 8 +[58] Jia-Bin Huang, Abhishek Singh, and Narendra Ahuja. Single image super-resolution from transformed self-exemplars. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5197-5206, 2015. 6 +[59] Aaron Hurst, Adam Lerner, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, et al. Gpt-4o system card. arXiv preprint arXiv:2410.21276, 2024. 16 +[60] Shady Abu Hussein, Tom Tirer, and Raja Giryes. Image-adaptive gan based reconstruction. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 3121-3129, 2020. 38 +[61] Maxar Intelligence. Supercharging maxar intelligence's imagery basemaps with worldview legion satellite imagery. https://blog.maxar.com/earth-intelligence/2025/supercharging-maxar-intelligences-imagery-basemaps-with-worldview-legion-satellite-imagery, 2025.37 +[62] Md Jahidul Islam, Peigen Luo, and Junaed Sattar. Simultaneous enhancement and super-resolution of underwater imagery for improved visual perception. In 16th Robotics: Science and Systems, RSS 2020. MIT Press Journals, 2020. 36 +[63] Feng Jia, Lei Tan, Guang Wang, Cheng Jia, and Zhi Chen. A super-resolution network using channel attention retention for pathology images. PeerJ Computer Science, 9:e1196, 2023. Published 2023 Jan 17. 6, 27, 28 +[64] Jiaxi Jiang, Kai Zhang, and Radu Timofte. Towards flexible blindJPEG artifacts removal. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4997-5006, 2021. 4 +[65] Xu Jiang, Gehui Li, Bin Chen, and Jian Zhang. Multi-agent image restoration. arXiv preprint arXiv:2503.09403, 2025. 11, 39 +[66] Yitong Jiang, Zhaoyang Zhang, Tianfan Xue, and Jinwei Gu. Autodir: Automatic all-in-one image restoration with latent diffusion. In European Conference on Computer Vision, pages 340-359. Springer, 2024. 11, 38 +[67] Bahjat Kawar, Michael Elad, Stefano Ermon, and Jiaming Song. Denoising diffusion restoration models. Advances in Neural Information Processing Systems, 35:23593-23606, 2022. 38 +[68] Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. Musiq: Multi-scale image quality transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5148-5157, 2021. 6 +[69] Christopher J Kelly, Alan Karthikesalingam, Mustafa Suleyman, Greg Corrado, and Dominic King. Key challenges for delivering clinical impact with artificial intelligence. BMC medicine, 17:1-9, 2019. 38 +[70] Daniel S Kermany, Michael Goldbaum, Wenjia Cai, Carolina CS Valentim, Huiying Liang, Sally L Baxter, Alex McKeown, Ge Yang, Xiaokang Wu, Fangbing Yan, et al. Identifying medical diagnoses and treatable diseases by image-based deep learning. cell, 172(5):1122-1131, 2018. 6, 30 +[71] Chanwoo Kim, Soham U. Gadgil, Alex J. DeGrave, Jesutofunmi A. Omiye, Zhuo Ran Cai, Roxana Daneshjou, and Su-In Lee. Transparent medical image AI via an image-text foundation model grounded in medical literature. Nature Medicine, 30:1154-1165, 2024. 30 +[72] Jiwon Kim, Jung Kwon Lee, and Kyoung Mu Lee. Accurate image super-resolution using very deep convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1646-1654, 2016. 26, 38 +[73] Yuval Kirstain, Adam Polyak, Uriel Singer, Shahbuland Matiana, Joe Penna, and Omer Levy. Pick-a-pic: An open dataset of user preferences for text-to-image generation. Advances in Neural Information Processing Systems, 36:36652-36663, 2023. 17 + +[74] Lingshun Kong, Jiangxin Dong, Ming-Hsuan Yang, and Jinshan Pan. Efficient visual state space model for image deblurring. arXiv preprint arXiv:2405.14343, 2024. 4 +[75] Xiangtao Kong, Chao Dong, and Lei Zhang. Towards effective multiple-in-one image restoration: A sequential and prompt learning strategy. arXiv preprint arXiv:2401.03379, 2024. 11 +[76] Xiangtao Kong, Jinjin Gu, Yihao Liu, Wenlong Zhang, Xiangyu Chen, Yu Qiao, and Chao Dong. A preliminary exploration towards general image restoration. arXiv preprint arXiv:2408.15143, 2024. 10 +[77] Pawel Kowaleczko, Tomasz Tarasiewicz, Maciej Ziaja, Daniel Kostrzewa, Jakub Nalepa, Przemyslaw Rokita, and Michal Kawulok. A real-world benchmark for sentinel-2 multi-image super-resolution. Scientific Data, 10(1):644, 2023. 18 +[78] Somnath Kumar, Yash Gadhia, Tanuja Ganu, and Akshay Nambi. Mmctagent: Multi-modal critical thinking agent framework for complex visual reasoning. arXiv preprint arXiv:2405.18358, 2024. 39 +[79] Black Forest Labs. Flux. https://github.com/black-forest-labs/flux, 2024.16 +[80] Wei-Sheng Lai, Jia-Bin Huang, Narendra Ahuja, and Ming-Hsuan Yang. Deep laplacian pyramid networks for fast and accurate super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 624-632, 2017. 30 +[81] Wei-Sheng Lai, Jia-Bin Huang, Narendra Ahuja, and Ming-Hsuan Yang. Fast and accurate image superresolution with deep laplacian pyramid networks. IEEE transactions on pattern analysis and machine intelligence, 41(11):2599-2613, 2018. 26 +[82] Christian Ledig, Lucas Theis, Ferenc Huszár, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, et al. Photo-realistic single image superresolution using a generative adversarial network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4681-4690, 2017. 28, 30, 31, 38 +[83] Juhyoung Lee, Jinsu Lee, and Hoi-Jun Yoo. Srnpu: An energy-efficient cnn-based super-resolution processor with tile-based selective super-resolution in mobile devices. IEEE Journal on Emerging and Selected Topics in Circuits and Systems, 10(3):320-334, 2020. 37 +[84] Junyong Lee, Hyeongseok Son, Jaesung Rim, Sunghyun Cho, and Seungyong Lee. Iterative filter adaptive network for single image defocus deblurring. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2034-2042, 2021. 4 +[85] Royson Lee, Stylianos I Venieris, Lukasz Dudziak, Sourav Bhattacharya, and Nicholas D Lane. Mobisr: Efficient on-device super-resolution through heterogeneous mobile processors. In The 25th annual international conference on mobile computing and networking, pages 1-16, 2019. 37 +[86] Sen Lei, Zhenwei Shi, and Wenjing Mo. Transformer-based multistage enhancement for remote sensing image super-resolution. IEEE Transactions on Geoscience and Remote Sensing, 60:1-11, 2022. 19, 20, 21 +[87] Baiqi Li, Zhiqiu Lin, Deepak Pathak, Jiayao Li, Yixin Fei, Kewen Wu, Tiffany Ling, Xide Xia, Pengchuan Zhang, Graham Neubig, et al. Genai-bench: Evaluating and improving compositional text-to-visual generation. arXiv preprint arXiv:2406.13743, 2024. 6, 16, 17 +[88] Bingchen Li, Xin Li, Yiting Lu, and Zhibo Chen. Hybrid agents for image restoration. arXiv preprint arXiv:2503.10120, 2025. 39 +[89] Bingchen Li, Xin Li, Hanxin Zhu, Yeying Jin, Ruoyu Feng, Zhizheng Zhang, and Zhibo Chen. Sed: Semantic-aware discriminator for image super-resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 25784-25795, 2024. 38 +[90] Boyun Li, Xiao Liu, Peng Hu, Zhongqin Wu, Jiancheng Lv, and Xi Peng. All-in-one image restoration for unknown corruption. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 17452-17462, 2022. 11, 38, 39 +[91] Gen Li, Jie Ji, Minghai Qin, Wei Niu, Bin Ren, Fatemeh Afghah, Linke Guo, and Xiaolong Ma. Towards high-quality and efficient video super-resolution via spatial-temporal data overfitting. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10259-10269. IEEE, 2023. 34 +[92] Jinlong Li, Baolu Li, Zhengzhong Tu, Xinyu Liu, Qing Guo, Felix Juefei-Xu, Runsheng Xu, and Hongkai Yu. Light the night: A multi-condition diffusion framework for unpaired low-light enhancement in autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15205-15215, 2024. 35 + +[93] Juncheng Li, Faming Fang, Kangfu Mei, and Guixu Zhang. Multi-scale residual network for image super-resolution. In Proceedings of the European conference on computer vision (ECCV), pages 517-532, 2018. 38 +[94] Ke Li, Gang Wan, Gong Cheng, Liqui Meng, and Junwei Han. Object detection in optical remote sensing images: A survey and a new benchmark. ISPRS journal of photogrammetry and remote sensing, 159:296-307, 2020. 6, 19 +[95] Xin Li, Kun Yuan, Bingchen Li, Fengbin Guan, Yizhen Shao, Zihao Yu, Xijun Wang, Yiting Lu, Wei Luo, Suhang Yao, et al. Ntire 2025 challenge on short-formUGC video quality assessment and enhancement: Methods and results. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 1092-1103, 2025. 34 +[96] Yawei Li, Yuchen Fan, Xiaoyu Xiang, Denis Demandolx, Rakesh Ranjan, Radu Timofte, and Luc Van Gool. Efficient and explicit modelling of image hierarchies for image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18278-18289, 2023. 26 +[97] Yiming Li, Shuo Shao, Yu He, Junfeng Guo, Tianwei Zhang, Zhan Qin, Pin-Yu Chen, Michael Backes, Philip Torr, Dacheng Tao, and Kui Ren. Rethinking data protection in the (generative) artificial intelligence era. arXiv preprint arXiv:2507.03034, 2025. 38 +[98] Yinxiao Li, Pengchong Jin, Feng Yang, Ce Liu, Ming-Hsuan Yang, and Peyman Milanfar. Comiser: Compression-informed video super-resolution. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2543-2552, 2021. 37 +[99] Zhen Li, Jinglei Yang, Zheng Liu, Xiaomin Yang, Gwanggil Jeon, and Wei Wu. Feedback network for image super-resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3867-3876, 2019. 26 +[100] Jie Liang, Hui Zeng, and Lei Zhang. Efficient and degradation-adaptive network for real-world image super-resolution. In European Conference on Computer Vision, pages 574-591. Springer, 2022. 38 +[101] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swin transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1833-1844, 2021. 4, 7, 8, 19, 20, 21, 26, 31, 38 +[102] Lighthouse Guild. High tech for low vision: How technology is changing the world for people with vision loss. https://lighthouseguild.org/news/high-tech-for-low-vision-how-technology-is-changing-the-world-for-people-with-vision-loss/, 2025.37 +[103] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Kyoung Mu Lee. Enhanced deep residual networks for single image super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 136-144, 2017. 28, 31, 38 +[104] Tianwei Lin, Wenqiao Zhang, Sijing Li, Yuqian Yuan, Binhe Yu, Haoyuan Li, Wanggui He, Hao Jiang, Mengze Li, Xiaohui Song, et al. Healthgpt: A medical large vision-language model for unifying comprehension and generation via heterogeneous knowledge adaptation. arXiv preprint arXiv:2502.09838, 2025. 30 +[105] Xinqi Lin, Jingwen He, Ziyan Chen, Zhaoyang Lyu, Bo Dai, Fanghua Yu, Yu Qiao, Wanli Ouyang, and Chao Dong. Diffbir: Toward blind image restoration with generative diffusion prior. In European Conference on Computer Vision, pages 430-448. Springer, 2024. 4, 7, 8, 9, 13, 15, 19, 20, 21, 38 +[106] Yunlong Lin, Zixu Lin, Haoyu Chen, Panwang Pan, Chenxin Li, Sixiang Chen, Kairun Wen, Yeying Jin, Wenbo Li, and Xinghao Ding. Jarvisir: Elevating autonomous driving perception with intelligent image restoration. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 22369-22380, 2025. 39 +[107] Yunlong Lin, Zixu Lin, Kunjie Lin, Jinbin Bai, Panwang Pan, Chenxin Li, Haoyu Chen, Zhongdao Wang, Xinghao Ding, Wenbo Li, et al. Jarvisart: Liberating human artistic creativity via an intelligent photo retouching agent. arXiv preprint arXiv:2506.17612, 2025. 39 +[108] G. Litjens, P. Bandi, B. Ehteshami Bejnordi, O. Geessink, M. Balkenhol, P. Bult, A. Halilovic, M. Hermsen, R. van de Loo, R. Vogels, Q.F. Manson, N. Stathonikos, A. Baidoshvili, P. van Diest, C. Wauters, M. van Dijk, and J. van der Laak. 1399 H&E-stained sentinel lymph node sections of breast cancer patients: the CAMELYON dataset. *GigaScience*, 7(6):giy065, June 2018. 27 + +[109] Jiaming Liu, Zihao Liu, Xuan Huang, Ruoxi Zhu, Qi Zheng, Zhijian Hao, Tao Liu, Jun Tao, and Yibo Fan. Auto-isp: An efficient real-time automatic hyperparameter optimization framework for isp hardware system. In Proceedings of the 61st ACM/IEEE Design Automation Conference, pages 1–6, 2024. 36 +[110] Jiaming Liu, Qi Zheng, Zihao Liu, Yilian Zhong, Peiye Liu, Tao Liu, Shusong Xu, Yanheng Lu, Sicheng Li, Dimin Niu, et al. Frequency-biased synergistic design for image compression and compensation. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 12820-12829, 2025. 38 +[111] Jiaxiang Liu, Yuan Wang, Jiawei Du, Joey Tianyi Zhou, and Zuozhu Liu. Medcot: Medical chain of thought via hierarchical expert. arXiv preprint arXiv:2412.13736, 2024. 39 +[112] Xing Liu, Masanori Suganuma, Zhun Sun, and Takayuki Okatani. Dual residual networks leveraging the potential of paired operations for image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7007-7016, 2019. 38 +[113] Yuhao Liu, Zhanghan Ke, Fang Liu, Nanxuan Zhao, and Rynson WH Lau. Diff-plugin: Revitalizing details for diffusion-based low-level tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4197-4208, 2024. 4 +[114] Jie Lu, Anjin Liu, Fan Dong, Feng Gu, Joao Gama, and Guangquan Zhang. Learning under concept drift: A review. IEEE transactions on knowledge and data engineering, 31(12):2346-2363, 2018. 37 +[115] Zhisheng Lu, Juncheng Li, Hong Liu, Chaoyan Huang, Linlin Zhang, and Tieyong Zeng. Transformer for single image super-resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 457-466, 2022. 31 +[116] Luma. Dream-machine. https://lumalabs.ai/dream-machine, 2024.35 +[117] Ziwei Luo, Fredrik K Gustafsson, Zheng Zhao, Jens Sjolund, and Thomas B Schön. Controlling vision-language models for multi-task image restoration. arXiv preprint arXiv:2310.01018, 2023. 11 +[118] Xiaogian Lv, Shengping Zhang, Chenyang Wang, Yichen Zheng, Bineng Zhong, Chongyi Li, and Liqiang Nie. Fourier priors-guided diffusion for zero-shot joint low-light enhancement and deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25378-25388, 2024. 4 +[119] Jiaqi Ma, Tianheng Cheng, Guoli Wang, Qian Zhang, Xinggang Wang, and Lefei Zhang. Prores: Exploring degradation-aware visual prompt for universal image restoration. arXiv preprint arXiv:2306.13653, 2023. 38 +[120] Varun Mannam, Yide Zhang, Xiaotong Yuan, and Scott Howard. Deep learning-based super-resolution fluorescence microscopy on small datasets. In *Single Molecule Spectroscopy and Superresolution Imaging XIV*, volume 11650, pages 60–68. SPIE, 2021. 25 +[121] David Martin, Charless Fowlkes, Doron Tal, and Jitendra Malik. A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. In Proceedings eighth IEEE international conference on computer vision. ICCV 2001, volume 2, pages 416-423. IEEE, 2001. 6 +[122] Daniel Enrique Martinez, Waiman Meinhold, John Oshinski, Ai-Ping Hu, and Jun Ueda. Super resolution for improved positioning of an mri-guided spinal cellular injection robot. Journal of Medical Robotics Research, 6(01n02):2140002, 2021. 36 +[123] Yusuke Matsui, Kota Ito, Yuji Aramaki, Azuma Fujimoto, Toru Ogawa, Toshihiko Yamasaki, and Kiyoharu Aizawa. Sketch-based manga retrieval using manga109 dataset. Multimedia tools and applications, 76:21811-21838, 2017. 6 +[124] Ninareh Mehrabi, Fred Morstatter, Nripsuta Saxena, Kristina Lerman, and Aram Galstyan. A survey on bias and fairness in machine learning. ACM computing surveys (CSUR), 54(6):1-35, 2021. 37 +[125] Yiqun Mei, Yuchen Fan, and Yuqian Zhou. Image super-resolution with non-local sparse attention. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3517-3526, 2021. 26 +[126] Jacob Menick and Nal Kalchbrenner. Generating high fidelity images with subscale pixel networks and multidimensional upscaling. arXiv preprint arXiv:1812.01608, 2018. 38 +[127] Sachit Menon, Alexandru Damian, Shijia Hu, Nikhil Ravi, and Cynthia Rudin. Pulse: Self-supervised photo upsampling via latent space exploration of generative models. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 2437-2445, 2020. 38 + +[128] Yogendra Rao Musunuri, Oh-Seol Kwon, and Sun-Yuan Kung. Srodnet: Object detection network based on super resolution for autonomous vehicles. Remote Sensing, 14(24):6270, 2022. 36, 37 +[129] Saeel Sandeep Nachane, Ojas Gramopadhye, Prateek Chanda, Ganesh Ramakrishnan, Kshitij Sharad Jadhav, Yatin Nandwani, Dinesh Raghu, and Sachindra Joshi. Few shot chain-of-thought driven reasoning to prompt llms for open ended medical question answering. arXiv preprint arXiv:2403.04890, 2024. 39 +[130] Babak Naderi, Ross Cutler, Juhee Cho, Nabakumar Khongbantabam, and Dejan Ivkovic. Icme 2025 grand challenge on video super-resolution for video conferencing. arXiv preprint arXiv:2506.12269, 2025. 34 +[131] Valfride Nascimento, Rayson Laroca, Jorge de A Lambert, William Robson Schwartz, and David Menotti. Super-resolution of license plate images using attention modules and sub-pixel convolution layers. Computers & Graphics, 113:69-76, 2023. 35 +[132] Ngoc Long Nguyen, Jérémy Anger, Axel Davy, Pablo Arias, and Gabriele Facciolo. Self-supervised multi-image super-resolution for push-frame satellite images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1121-1131, 2021. 36, 37 +[133] Zhangkai Ni, Runyu Xiao, Wenhan Yang, Hanli Wang, Zhihua Wang, Lihua Xiang, and Liping Sun. M2trans: Multi-modal regularized coarse-to-fine transformer for ultrasound image super-resolution. IEEE Journal of Biomedical and Health Informatics, pages 1-12, 2024. 6, 30, 31 +[134] OpenAI, Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, Clarence Ng, Ricky Wang, and Aditya Ramesh. Video generation models as world simulators. https://openai.com/research/video-generation-models-as-world-simulators, 2024. 35 +[135] Fu-Zhao Ou, Chongyi Li, Shiqi Wang, and Sam Kwong. Clib-fiqa: face image quality assessment with confidence calibration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1694–1704, 2024. 12 +[136] Suraj Pai, Ibrahim Hadzic, Dennis Bontempi, Keno Bressem, Benjamin H Kann, Andriy Fedorov, Raymond H Mak, and Hugo JWL Aerts. Vision foundation models for computed tomography. arXiv preprint arXiv:2501.09001, 2025. 30 +[137] Xingang Pan, Xiaohang Zhan, Bo Dai, Dahua Lin, Chen Change Loy, and Ping Luo. Exploiting deep generative prior for versatile image restoration and manipulation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(11):7474-7489, 2021. 38 +[138] Dongwon Park, Byung Hyun Lee, and Se Young Chun. All-in-one image restoration for unknown degradations using adaptive discriminative filters for specific degradations. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5815-5824. IEEE, 2023. 38, 39 +[139] Leonardo Peroni and Sergey Gorinsky. An end-to-end pipeline perspective on video streaming in best-effort networks: a survey and tutorial. ACM Computing Surveys, 2024. 37 +[140] Vaishnav Potlapalli, Syed Waqas Zamir, Salman H Khan, and Fahad Shahbaz Khan. Prompt for all-in-one image restoration. Advances in Neural Information Processing Systems, 36:71275-71293, 2023. 11 +[141] Darren Pouliot, Rasim Latifovic, Jon Pasher, and Jason Duffe. Landsat super-resolution enhancement using convolution neural networks and sentinel-2 for training. Remote Sensing, 10(3):394, 2018. 37 +[142] Chang Qiao, Di Li, Yuting Guo, Chong Liu, Tao Jiang, Qionghai Dai, and Dong Li. Evaluation and development of deep neural networks for image super-resolution in optical microscopy. Nature methods, 18(2):194–202, 2021. 26 +[143] Lingyan Ruan, Mojtaba Bemana, Hans-peter Seidel, Karol Myszkowski, and Bin Chen. Revisiting image deblurring with an efficient convnet. arXiv preprint arXiv:2302.02234, 2023. 4 +[144] Lingyan Ruan, Bin Chen, Jizhou Li, and Miuling Lam. Learning to deblur using light field generated and real defocus images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16304–16313, 2022. 4 +[145] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image super-resolution via iterative refinement. IEEE transactions on pattern analysis and machine intelligence, 45(4):4713-4726, 2022. 38 + +[146] Jacob Shermeyer and Adam Van Etten. The effects of super-resolution on object detection performance in satellite imagery. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, 2019. 18, 36, 37 +[147] Wenzhe Shi, Jose Caballero, Ferenc Huszár, Johannes Totz, Andrew P Aitken, Rob Bishop, Daniel Rueckert, and Zehan Wang. Real-time single image and video super-resolution using an efficient subpixel convolutional neural network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1874-1883, 2016. 30 +[148] Dehua Song, Yunhe Wang, Hanting Chen, Chang Xu, Chunjing Xu, and DaCheng Tao. Addersr: Towards energy efficient image super-resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15648-15657, 2021. 37 +[149] Yuda Song, Zhuqing He, Hui Qian, and Xin Du. Vision transformers for single image dehazing. IEEE Transactions on Image Processing, 32:1927-1941, 2023. 4 +[150] FUJIFILM Healthcare Europe & SonoSkills. US-CASE: Ultrasound Cases Dataset. http://www.ultrasoundcases.info/Cases-Home.aspx, 2025. 6, 30 +[151] J. J. Staal, M. D. Abramoff, M. Niemeijer, M. A. Viergever, and B. van Ginneken. Ridge-based vessel segmentation in color images of the retina. IEEE Transactions on Medical Imaging, 23(4):501-509, 2004. 6, 30 +[152] Theodore Sumers, Shunyu Yao, Karthik Narasimhan, and Thomas Griffiths. Cognitive architectures for language agents. Transactions on Machine Learning Research, 2023. 39 +[153] Lingchen Sun, Rongyuan Wu, Zhiyuan Ma, Shuaizheng Liu, Qiaosi Yi, and Lei Zhang. Pixel-level and semantic-level adjustable super-resolution: A dual-lora approach. arXiv preprint arXiv:2412.03017, 2024. 4, 7, 8, 9, 13, 15, 19, 20, 21, 38 +[154] Ying Tai, Jian Yang, and Xiaoming Liu. Image super-resolution via deep recursive residual network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3147-3155, 2017. 26 +[155] Ying Tai, Jian Yang, Xiaoming Liu, and Chunyan Xu. Memnet: A persistent memory network for image restoration. In Proceedings of the IEEE international conference on computer vision, pages 4539-4547, 2017. 26 +[156] Muhammed Telçeken, Devrim Akgun, Sezgin Kacar, and Bunyamin Bingol. A new approach for super resolution object detection using an image slicing algorithm and the segment anything model. Sensors, 24(14):4526, 2024. 36 +[157] Jennifer A Thorley, Jeremy Pike, and Joshua Z Rappoport. Super-resolution microscopy: a comparison of commercially available options. In *Fluorescence microscopy*, pages 199–212. Elsevier, 2014. 25 +[158] Kalina L Tosheva, Yue Yuan, Pedro Matos Pereira, Sian Culley, and Ricardo Henriques. Between life and death: strategies to reduce phototoxicity in super-resolution microscopy. Journal of Physics D: Applied Physics, 53(16):163001, 2020. 25 +[159] Zhengzhong Tu, Hossein Talebi, Han Zhang, Feng Yang, Peyman Milanfar, Alan Bovik, and Yinxiao Li. Maxim: Multi-axis mlp for image processing. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5769-5780, 2022. 4, 38 +[160] Sabina Umirzakova, Shabir Ahmad, Latif U Khan, and Taegkeun Whangbo. Medical image superresolution for smart healthcare applications: A comprehensive survey. Information Fusion, 103:102075, 2024. 29 +[161] Jeya Maria Jose Valanarasu, Rajeev Yasarla, and Vishal M Patel. Transweather: Transformer-based restoration of images degraded by adverse weather conditions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2353-2363, 2022. 38, 39 +[162] Aaron Van den Oord, Nal Kalchbrenner, Lasse Espeholt, Oriol Vinyals, Alex Graves, et al. Conditional image generation with pixelCNN decoders. Advances in neural information processing systems, 29, 2016. 38 + +[163] Veo-Team, Agrim Gupta, Ali Razavi, Andeep Toor, Ankush Gupta, Dumitru Erhan, Eleni Shaw, Eric Lau, Frank Belletti, Gabe Barth-Maron, Gregory Shaw, Hakan Erdogan, Hakim Sidahmed, Henna Nandwani, Hernan Moraldo, Hyunjik Kim, Irina Blok, Jeff Donahue, José Lezama, Kory Mathewson, Kurtis David, Matthieu Kim Lorrain, Marc van Zee, Medhini Narasimhan, Miaosen Wang, Mohammad Babaeizadeh, Nelly Papalampidi, Nick Pezzotti, Nilpa Jha, Parker Barnes, Pieter-Jan Kindermans, Rachel Hornung, Ruben Villegas, Ryan Poplin, Salah Zaiem, Sander Dieleman, Sayna Ebrahimi, Scott Wisdom, Serena Zhang, Shlomi Fruchter, Signe Nørly, Weizhe Hua, Xinchen Yan, Yuqing Du, and Yutian Chen. Veo 2. https://deepmind.google/technologies/veo/veo-2, 2024. 35 +[164] W3C. Research - low vision accessibility task force. https://www.w3.org/WAI/GL/low-vision-allyt/wiki/Research, 2025.37 +[165] Hang Wang, Xuanhong Chen, Bingbing Ni, Yutian Liu, and Jinfan Liu. Omni aggregation networks for lightweight image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22378-22387, 2023. 26 +[166] Jianyi Wang, Kelvin CK Chan, and Chen Change Loy. Exploring clip for assessing the look and feel of images. In Proceedings of the AAAI conference on artificial intelligence, volume 37, pages 2555–2563, 2023. 6 +[167] Jianyi Wang, Zongsheng Yue, Shangchen Zhou, Kelvin CK Chan, and Chen Change Loy. Exploiting diffusion prior for real-world image super-resolution. International Journal of Computer Vision, 132(12):5929-5949, 2024. 9, 38 +[168] Jiarui Wang, Binglu Wang, Xiaoxu Wang, Yongqiang Zhao, and Teng Long. Hybrid attention-based u-shaped network for remote sensing image super-resolution. IEEE Transactions on Geoscience and Remote Sensing, 61:1-15, 2023. 19, 20, 21 +[169] Xiaohan Wang, Yuhui Zhang, Orr Zohar, and Serena Yeung-Levy. Videoagent: Long-form video understanding with large language model as agent. In European Conference on Computer Vision, pages 58-76. Springer, 2024. 39 +[170] Xiaosong Wang, Yifan Peng, Le Lu, Zhiyong Lu, Mohammadhadi Bagheri, and Ronald M Summers. Chestx-ray8: Hospital-scale chest x-ray database and benchmarks on weakly-supervised classification and localization of common thorax diseases. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2097-2106, 2017. 6, 30 +[171] Xintao Wang, Yu Li, Honglun Zhang, and Ying Shan. Towards real-world blind face restoration with generative facial prior. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9168-9178, 2021. 4, 6, 12 +[172] Xintao Wang, Liangbin Xie, Chao Dong, and Ying Shan. Real-esrgan: Training real-world blind super-resolution with pure synthetic data. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1905–1914, 2021. 38 +[173] Yifan Wang, Federico Perazzi, Brian McWilliams, Alexander Sorkine-Hornung, Olga Sorkine-Hornung, and Christopher Schroers. A fully progressive approach to single-image super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 864-873, 2018. 26 +[174] Yinhuai Wang, Jiwen Yu, and Jian Zhang. Zero-shot image restoration using denoising diffusion null-space model. arXiv preprint arXiv:2212.00490, 2022. 38 +[175] Yufei Wang, Wenhan Yang, Xinyuan Chen, Yaohui Wang, Lanqing Guo, Lap-Pui Chau, Ziwei Liu, Yu Qiao, Alex C Kot, and Bihan Wen. Sinsr: diffusion-based image super-resolution in a single step. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 25796-25805, 2024. 9, 38 +[176] Zhendong Wang, Xiaodong Cun, Jianmin Bao, Wengang Zhou, Jianzhuang Liu, and Houqiang Li. Uformer: A general u-shaped transformer for image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 17683-17693, 2022. 38 +[177] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 6 +[178] Zhou Wang, Eero P Simoncelli, and Alan C Bovik. Multiscale structural similarity for image quality assessment. In The Thirsty-Seventh Asilomar Conference on Signals, Systems & Computers, 2003, volume 2, pages 1398-1402. IEEE, 2003. 30 + +[179] Zijie J Wang, Evan Montoya, David Munechika, Haoyang Yang, Benjamin Hoover, and Duen Horng Chau. Diffusiondb: A large-scale prompt gallery dataset for text-to-image generative models. arXiv preprint arXiv:2210.14896, 2022. 6, 16, 17 +[180] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824-24837, 2022. 39 +[181] Pengxu Wei, Ziwei Xie, Hannan Lu, Zongyuan Zhan, Qixiang Ye, Wangmeng Zuo, and Liang Lin. Component divide-and-conquer for real-world image super-resolution. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VIII 16, pages 101-117. Springer, 2020. 6 +[182] Martin Weigert, Uwe Schmidt, Tobias Boothe, Andreas Müller, Alexandr Dibrov, Akanksha Jain, Benjamin Wilhelm, Deborah Schmidt, Coleman Broaddus, Sián Culley, et al. Content-aware image restoration: pushing the limits of fluorescence microscopy. Nature methods, 15(12):1090-1097, 2018. 36, 37 +[183] Rongyuan Wu, Lingchen Sun, Zhiyuan Ma, and Lei Zhang. One-step effective diffusion network for real-world image super-resolution. Advances in Neural Information Processing Systems, 37:92529-92553, 2024. 4, 7, 8, 9, 13, 15, 19, 20, 21, 38 +[184] Rongyuan Wu, Tao Yang, Lingchen Sun, Zhengqiang Zhang, Shuai Li, and Lei Zhang. Seesr: Towards semantics-aware real-world image super-resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 25456-25467, 2024. 9, 38 +[185] Rui-Qi Wu, Zheng-Peng Duan, Chun-Le Guo, Zhi Chai, and Chongyi Li. Ridcp: Revitalizing real image dehazing via high-quality codebook priors. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22282-22291, 2023. 4 +[186] Bin Xia, Yucheng Hang, Yapeng Tian, Wenming Yang, Qingmin Liao, and Jie Zhou. Efficient non-local contrastive attention for image super-resolution. In Proceedings of the AAAI conference on artificial intelligence, volume 36, pages 2759-2767, 2022. 26 +[187] Bin Xia, Yulun Zhang, Shiyin Wang, Yitong Wang, Xinglong Wu, Yapeng Tian, Wenming Yang, and Luc Van Gool. Diffir: Efficient diffusion model for image restoration. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13095-13105, 2023. 38 +[188] Gui-Song Xia, Xiang Bai, Jian Ding, Zhen Zhu, Serge Belongie, Jiebo Luo, Mihai Datcu, Marcello Pelillo, and Liangpei Zhang. Dota: A large-scale dataset for object detection in aerial images. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3974-3983, 2018. 6, 19 +[189] Gui-Song Xia, Jingwen Hu, Fan Hu, Baoguang Shi, Xiang Bai, Yanfei Zhong, Liangpei Zhang, and Xiaoqiang Lu. Aid: A benchmark data set for performance evaluation of aerial scene classification. IEEE Transactions on Geoscience and Remote Sensing, 55(7):3965-3981, 2017. 6, 18 +[190] Peng Xia, Jinglu Wang, Yibo Peng, Kaide Zeng, Xian Wu, Xiangru Tang, Hongtu Zhu, Yun Li, Shujie Liu, Yan Lu, et al. Mmedagent-rl: Optimizing multi-agent collaboration for multimodal medical reasoning. arXiv preprint arXiv:2506.00555, 2025. 39 +[191] Jun Xiao, Xinyang Jiang, Ningxin Zheng, Huan Yang, Yifan Yang, Yuqing Yang, Dongsheng Li, and Kin-Man Lam. Online video super-resolution with convolutional kernel bypass grafts. IEEE Transactions on Multimedia, 25:8972-8987, 2023. 34 +[192] Jun Xiao, Tianshan Liu, Rui Zhao, and Kin-Man Lam. Balanced distortion and perception in single-image super-resolution based on optimal transport in wavelet domain. Neurocomputing, 464:408-420, 2021. 38 +[193] Yi Xiao, Qiangqiang Yuan, Kui Jiang, Yuzeng Chen, Qiang Zhang, and Chia-Wen Lin. Frequency-assisted mamba for remote sensing image super-resolution. IEEE Transactions on Multimedia, 2024. 19 +[194] Yi Xiao, Qiangqiang Yuan, Kui Jiang, Jiang He, Xianyu Jin, and Liangpei Zhang. Ediffsr: An efficient diffusion probabilistic model for remote sensing image super-resolution. IEEE Transactions on Geoscience and Remote Sensing, 62:1-14, 2023. 19 +[195] Enze Xie, Junsong Chen, Junyu Chen, Han Cai, Haotian Tang, Yujun Lin, Zhekai Zhang, Muyang Li, Ligeng Zhu, Yao Lu, et al. Sana: Efficient high-resolution image synthesis with linear diffusion transformers. arXiv preprint arXiv:2410.10629, 2024. 16 + +[196] Liming Xu, Xianhua Zeng, Zhiwei Huang, Weisheng Li, and He Zhang. Low-dose chest x-ray image superresolution using generative adversarial nets with spectral normalization. Biomedical Signal Processing and Control, 55:101600, 2020. 30, 31 +[197] Feng Yang, Yue-Min Zhu, Jian-Hua Luo, Marc Robini, Jie Liu, and Pierre Croisille. A comparative study of different level interpolations for improving spatial resolution in diffusion tensor imaging. IEEE Journal of Biomedical and Health Informatics, 18(4):1317-1327, 2014. 30 +[198] Sidi Yang, Tianhe Wu, Shuwei Shi, Shanshan Lao, Yuan Gong, Mingdeng Cao, Jiahao Wang, and Yujiu Yang. Maniaq: Multi-dimension attention network for no-reference image quality assessment. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1191-1200, 2022. 6 +[199] Tao Yang, Rongyuan Wu, Peiran Ren, Xuansong Xie, and Lei Zhang. Pixel-aware stable diffusion for realistic image super-resolution and personalized stylization. In European Conference on Computer Vision, pages 74–91. Springer, 2024. 9, 38 +[200] Tianjie Yang, Yaoru Luo, Wei Ji, and Ge Yang. Advancing biological super-resolution microscopy through deep learning: a brief review. Biophysics Reports, 7(4):253, 2021. 25 +[201] Mingde Yao, Ruikang Xu, Yuanshen Guan, Jie Huang, and Zhiwei Xiong. Neural degradation representation learning for all-in-one image restoration. IEEE Transactions on Image Processing, 2024. 39 +[202] Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. React: Synergizing reasoning and acting in language models. In International Conference on Learning Representations (ICLR), 2023. 39 +[203] Jinsu Yoo, Taehoon Kim, Sihaeng Lee, Seung Hwan Kim, Honglak Lee, and Tae Hyun Kim. Enriched cnn-transformer feature aggregation networks for super-resolution. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 4956-4965, 2023. 26 +[204] Chenyu You, Guang Li, Yi Zhang, Xiaoliu Zhang, Hongming Shan, Mengzhou Li, Shenghong Ju, Zhen Zhao, Zhuiyang Zhang, Wenxiang Cong, et al. Ct super-resolution gan constrained by the identical, residual, and cycle learning ensemble (gan-circle). IEEE transactions on medical imaging, 39(1):188-203, 2019. 30 +[205] Zhiyuan You, Zheyuan Li, Jinjin Gu, Zhenfei Yin, Tianfan Xue, and Chao Dong. Depicting beyond scores: Advancing image quality assessment through multi-modal language models. In European Conference on Computer Vision, pages 259-276. Springer, 2024. 3 +[206] Jiahui Yu, Zhe Lin, Jamei Yang, Xiaohui Shen, Xin Lu, and Thomas S Huang. Free-form image inpainting with gated convolution. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4471-4480, 2019. 38 +[207] Miao Yu, Zhenghua Xu, and Thomas Lukasiewicz. A general survey on medical image super-resolution via deep learning. Computers in Biology and Medicine, 193:110345, Jul 2025. 29 +[208] Zongsheng Yue and Chen Change Loy. Difface: Blind face restoration with diffused error contraction. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 4, 12 +[209] Zongsheng Yue, Jianyi Wang, and Chen Change Loy. Reshift: Efficient diffusion model for image superresolution by residual shifting. Advances in Neural Information Processing Systems, 36:13294-13307, 2023. 9 +[210] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5728-5739, 2022. 4, 38 +[211] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling Shao. Learning enriched features for real image restoration and enhancement. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXV 16, pages 492-511. Springer, 2020. 38 +[212] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling Shao. Multi-stage progressive image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14821-14831, 2021. 4 + +[213] Roman Zeyde, Michael Elad, and Matan Protter. On single image scale-up using sparse-representations. In International conference on curves and surfaces, pages 711-730. Springer, 2010. 6 +[214] Zheng Zhan, Yifan Gong, Pu Zhao, Geng Yuan, Wei Niu, Yushu Wu, Tianyun Zhang, Malith Jayaweera, David Kaeli, Bin Ren, et al. Achieving on-mobile real-time super-resolution with neural architecture and pruning search. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4821-4831, 2021. 34 +[215] Aoyang Zhang, Qing Li, Ying Chen, Xiaoteng Ma, Longhao Zou, Yong Jiang, Zhimin Xu, and Gabriel-Miro Muntean. Video super-resolution and cachingan edge-assisted adaptive video streaming solution. IEEE Transactions on Broadcasting, 67(4):799-812, 2021. 37 +[216] Cheng Zhang, Yu Zhu, Qingsen Yan, Jinqiu Sun, and Yanning Zhang. All-in-one multi-degradation image restoration network via hierarchical degradation representation. In Proceedings of the 31st ACM international conference on multimedia, pages 2285-2293, 2023. 39 +[217] Dafeng Zhang, Feiyu Huang, Shizhuo Liu, Xiaobing Wang, and Zhezhu Jin. Swinfir: Revisiting the swinir with fast fourier convolution and improved training for image super-resolution. arXiv preprint arXiv:2208.11247, 2022. 4, 6 +[218] Jinjin Zhang, Qiuyu Huang, Junjie Liu, Xiefan Guo, and Di Huang. Diffusion-4k: Ultra-high-resolution image synthesis with latent diffusion models. arXiv preprint arXiv:2503.18352, 2025. 14, 16 +[219] Kai Zhang, Yawei Li, Wangmeng Zuo, Lei Zhang, Luc Van Gool, and Radu Timofte. Plug-and-play image restoration with deep denoiser prior. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(10):6360-6376, 2021. 38 +[220] Kai Zhang, Jingyun Liang, Luc Van Gool, and Radu Timofte. Designing a practical degradation model for deep blind image super-resolution. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4791-4800, 2021. 6, 38 +[221] Kai Zhang, Wangmeng Zuo, Yunjin Chen, Deyu Meng, and Lei Zhang. Beyond a gaussian denoiser: Residual learning of deep cnn for image denoising. IEEE transactions on image processing, 26(7):3142-3155, 2017. 38 +[222] Kaibing Zhang, Dacheng Tao, Xinbo Gao, Xuelong Li, and Jie Li. Coarse-to-fine learning for single-image super-resolution. IEEE transactions on neural networks and learning systems, 28(5):1109-1122, 2016. 30 +[223] Lin Zhang, Lei Zhang, and Alan C Bovik. A feature-enriched completely blind image quality evaluator. IEEE Transactions on Image Processing, 24(8):2579-2591, 2015. 6 +[224] Lin Zhang, Lei Zhang, Xuanqin Mou, and David Zhang. Fsim: A feature similarity index for image quality assessment. IEEE transactions on Image Processing, 20(8):2378-2386, 2011. 30 +[225] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6 +[226] Xindong Zhang, Hui Zeng, Shi Guo, and Lei Zhang. Efficient long-range attention network for image super-resolution. In European conference on computer vision, pages 649-667. Springer, 2022. 31 +[227] Xindong Zhang, Hui Zeng, and Lei Zhang. Edge-oriented convolution block for real-time super resolution on mobile devices. In Proceedings of the 29th ACM international conference on multimedia, pages 4034-4043, 2021. 37 +[228] Yide Zhang, Zhe He, Xin Tong, David C Garrett, Rui Cao, and Lihong V Wang. Quantum imaging of biological organisms through spatial and polarization entanglement. Science Advances, 10(10):eadk1495, 2024. 36 +[229] Yulun Zhang, Kunpeng Li, Kai Li, Lichen Wang, Bineng Zhong, and Yun Fu. Image super-resolution using very deep residual channel attention networks. In Proceedings of the European conference on computer vision (ECCV), pages 286-301, 2018. 28, 38 +[230] Yulun Zhang, Yapeng Tian, Yu Kong, Bineng Zhong, and Yun Fu. Residual dense network for image super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2472-2481, 2018. 28, 38 + +[231] Xiaole Zhao, Yulun Zhang, Tao Zhang, and Xueming Zou. Channel splitting network for single mr image super-resolution. IEEE transactions on image processing, 28(11):5649-5662, 2019. 36 +[232] Shangchen Zhou, Kelvin Chan, Chongyi Li, and Chen Change Loy. Towards robust blind face restoration with codebook lookup transformer. Advances in Neural Information Processing Systems, 35:30599-30611, 2022. 4, 12 +[233] Yingjie Zhou, Jiezhang Cao, Zicheng Zhang, Farong Wen, Yanwei Jiang, Jun Jia, Xiaohong Liu, Xiongkuo Min, and Guangtao Zhai. Q-agent: Quality-driven chain-of-thought image restoration agent through robust multimodal large language model. arXiv preprint arXiv:2504.07148, 2025. 39 +[234] Yiyang Zhou, Yangfan He, Yaofeng Su, Siwei Han, Joel Jang, Gedas Bertasius, Mohit Bansal, and Huaxiu Yao. Reagent-v: A reward-driven multi-agent framework for video understanding. arXiv preprint arXiv:2506.01300, 2025. 39 +[235] Kaiwen Zhu, Jinjin Gu, Zhiyuan You, Yu Qiao, and Chao Dong. An intelligent agentic system for complex image restoration problems. arXiv preprint arXiv:2410.17809, 2024. 5, 6, 7, 8, 9, 11, 12, 13, 15, 20, 21, 33, 39 +[236] Ruoxi Zhu, Zhengzhong Tu, Jiaming Liu, Alan C Bovik, and Yibo Fan. Mwformer: Multi-weather image restoration using degradation-aware transformers. IEEE Transactions on Image Processing, 2024. 38 +[237] Karel J Zuiderveld et al. Contrast limited adaptive histogram equalization. Graphics gems, 4(1):474-485, 1994. 4 \ No newline at end of file diff --git a/NeurIPS/2025/4KAgent_ Agentic Any Image to 4K Super-Resolution/images.zip b/NeurIPS/2025/4KAgent_ Agentic Any Image to 4K Super-Resolution/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..786db99b35c30516f74158227efada903e2220bf --- /dev/null +++ b/NeurIPS/2025/4KAgent_ Agentic Any Image to 4K Super-Resolution/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dff683ac68e75876d61c7385265f34bc4e71801c29915730c03be5449f15f695 +size 5090834 diff --git a/NeurIPS/2025/4KAgent_ Agentic Any Image to 4K Super-Resolution/layout.json b/NeurIPS/2025/4KAgent_ Agentic Any Image to 4K Super-Resolution/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..bdaca5b92fb23ac449e482af5e3e2361fafbb9b7 --- /dev/null +++ b/NeurIPS/2025/4KAgent_ Agentic Any Image to 4K Super-Resolution/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d0d54356964b101f3fae35cda51fe4546b0ae862654a34e826bdbafb1efc63bf +size 2957711 diff --git a/NeurIPS/2025/70% Size, 100% Accuracy_ Lossless LLM Compression for Efficient GPU Inference via Dynamic-Length Float (DFloat11)/98af0408-1e69-494e-8ae9-b093fff40a51_content_list.json b/NeurIPS/2025/70% Size, 100% Accuracy_ Lossless LLM Compression for Efficient GPU Inference via Dynamic-Length Float (DFloat11)/98af0408-1e69-494e-8ae9-b093fff40a51_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..dcde4679b4ccf7b4b9b04a39f8933203d7715dec --- /dev/null +++ b/NeurIPS/2025/70% Size, 100% Accuracy_ Lossless LLM Compression for Efficient GPU Inference via Dynamic-Length Float (DFloat11)/98af0408-1e69-494e-8ae9-b093fff40a51_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c61e4b30ac2edf47fbb0995355c26b0e183525ab5d6e373dfb45dbcb0bdbc0f2 +size 157772 diff --git a/NeurIPS/2025/70% Size, 100% Accuracy_ Lossless LLM Compression for Efficient GPU Inference via Dynamic-Length Float (DFloat11)/98af0408-1e69-494e-8ae9-b093fff40a51_model.json b/NeurIPS/2025/70% Size, 100% Accuracy_ Lossless LLM Compression for Efficient GPU Inference via Dynamic-Length Float (DFloat11)/98af0408-1e69-494e-8ae9-b093fff40a51_model.json new file mode 100644 index 0000000000000000000000000000000000000000..24ac4572811593a2798bbdb1c8339c3ce337ef34 --- /dev/null +++ b/NeurIPS/2025/70% Size, 100% Accuracy_ Lossless LLM Compression for Efficient GPU Inference via Dynamic-Length Float (DFloat11)/98af0408-1e69-494e-8ae9-b093fff40a51_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:49598124a440799643122ef96d5c1b7fea91992601f81391e7fbcff821cd4f2e +size 204716 diff --git a/NeurIPS/2025/70% Size, 100% Accuracy_ Lossless LLM Compression for Efficient GPU Inference via Dynamic-Length Float (DFloat11)/98af0408-1e69-494e-8ae9-b093fff40a51_origin.pdf b/NeurIPS/2025/70% Size, 100% Accuracy_ Lossless LLM Compression for Efficient GPU Inference via Dynamic-Length Float (DFloat11)/98af0408-1e69-494e-8ae9-b093fff40a51_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..0051eba1ffd6b9c0dc2c00f70284cc45dee793d0 --- /dev/null +++ b/NeurIPS/2025/70% Size, 100% Accuracy_ Lossless LLM Compression for Efficient GPU Inference via Dynamic-Length Float (DFloat11)/98af0408-1e69-494e-8ae9-b093fff40a51_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ce7f5d74e491d7d4de75971a222966307cbcbb37d535654f50305950d484993c +size 5647600 diff --git a/NeurIPS/2025/70% Size, 100% Accuracy_ Lossless LLM Compression for Efficient GPU Inference via Dynamic-Length Float (DFloat11)/full.md b/NeurIPS/2025/70% Size, 100% Accuracy_ Lossless LLM Compression for Efficient GPU Inference via Dynamic-Length Float (DFloat11)/full.md new file mode 100644 index 0000000000000000000000000000000000000000..5d08818c89a5918700276210de1a711abeb1b6a7 --- /dev/null +++ b/NeurIPS/2025/70% Size, 100% Accuracy_ Lossless LLM Compression for Efficient GPU Inference via Dynamic-Length Float (DFloat11)/full.md @@ -0,0 +1,756 @@ +# 70% Size, 100% Accuracy: Lossless LLM Compression for Efficient GPU Inference via Dynamic-Length Float (DFloat11) + +Tianyi Zhang $^{1}$ , Mohsen Hariri $^{2}$ , Shaochen (Henry) Zhong $^{1}$ , Vipin Chaudhary $^{2}$ , Yang Sui $^{1}$ , Xia Hu $^{1}$ , and Anshumali Shrivastava $^{1,3}$ + +$^{1}$ Department of Computer Science, Rice University + $^{2}$ Department of Computer and Data Sciences, Case Western Reserve University + $^{3}$ Ken Kennedy Institute + +{tz21, henry.zhong, yang.sui, xia.hu, anshumali}@rice.edu, {mohsen.hariri, vipin}@case.edu + +Code: https://github.com/LeanModels/DFloat11 +Models: https://huggingface.co/DFloat11 + +# Abstract + +Large-scale AI models, such as Large Language Models (LLMs) and Diffusion Models (DMs), have grown rapidly in size, creating significant challenges for efficient deployment on resource-constrained hardware. In this paper, we introduce Dynamic-Length Float (DFloat11), a lossless compression framework that reduces LLM and DM size by $30\%$ while preserving outputs that are bit-for-bit identical to the original model. DFloat11 is motivated by the low entropy in the BFloat16 weight representation of LLMs, which reveals significant inefficiency in the existing storage format. By applying entropy coding, DFloat11 assigns dynamic-length encodings to weights based on frequency, achieving near information-optimal compression without any loss of precision. To facilitate efficient inference with dynamic-length encodings, we develop a custom GPU kernel for fast online decompression. Our design incorporates the following: (i) compact, hierarchical lookup tables (LUTs) that fit within GPU SRAM for efficient decoding, (ii) a two-phase GPU kernel for coordinating thread read/write positions using lightweight auxiliary variables, and (iii) transformer-block-level decompression to minimize latency. Experiments on Llama 3.3, Qwen 3, Mistral 3, FLUX.1, and others validate our hypothesis that DFloat11 achieves around $30\%$ model size reduction while preserving bit-for-bit identical outputs. Compared to a potential alternative of offloading parts of an uncompressed model to the CPU to meet memory constraints, DFloat11 achieves $2.3 - 46.2\times$ higher throughput in token generation. With a fixed GPU memory budget, DFloat11 enables $5.7 - 14.9\times$ longer generation lengths than uncompressed models. Notably, our method enables lossless inference of Llama 3.1 405B, an 810GB model, on a single node equipped with $8\times 80\mathrm{GB}$ GPUs. + +# 1 Introduction + +Foundation models, such as Large Language Models (LLMs) and Diffusion Models (DMs), have demonstrated remarkable capabilities across a wide range of Natural Language Processing (NLP) [56] and Computer Vision (CV) tasks [57]. However, their huge model sizes create substantial obstacles + +for efficient deployment, especially in memory-constrained environments. For example, a competitive recent LLM, Llama 3.1 405B [20], has 405 billion parameters in 16-bit Brain Float (BFloat16) format and requires about 810 GB of memory for full inference, exceeding the capacity of a typical high-end GPU server (e.g., DGX A100/H100 with $8 \times 80$ GB GPUs). As a result, deploying this model requires multiple nodes, making it expensive and inaccessible. In this work, we present a solution that compresses any BFloat16 model to approximately $70\%$ of its original size while preserving $100\%$ of its accuracy on any task. + +Model compression via quantization has limitations. Quantization is a type of lossy compression method that lowers the precision of model weights by converting them into lower bit-width representations [15, 37, 36, 43]. Although it can significantly reduce memory usage and often improve inference speed, quantization is not a one-size-fits-all solution and presents several key limitations: Accuracy degradation. By design, quantization introduces approximation errors. The degree of accuracy loss depends on multiple factors, including the base model, quantization method, evaluation benchmark, and target bit-width [35]. These interactions make it difficult to predict or quantify the impact comprehensively. Even mild quantization can noticeably degrade performance. For example, applying 8-bit SmoothQuant [51] to DeepSeek-R1-Distill-Qwen-1.5B [21] results in a $9.09\%$ drop in average accuracy across reasoning tasks [39]. Behavioral shifts. Even when overall accuracy metrics appear roughly unchanged, quantized models may behave differently from their full-precision counterparts. For instance, Dutta et al. [13] observe a phenomenon called flips, where quantized models produce answers that change from correct to incorrect and vice versa. This indicates that quantization can significantly alter model behavior, even when standard accuracy metrics show minimal change. For example, the W8A16 GPTQ-quantized Qwen2-1.5B[15, 54] exhibits only a $0.3\%$ drop in GSM8K (8-shot) accuracy [5], yet $6.37\%$ of its answers flip in correctness [13]. Compliance and reliability concerns. In domains like finance or healthcare, quantized models may not satisfy regulatory or reliability standards, as their outputs may differ from those of the original models [31]. We refer readers to Appendix A for a more detailed discussion on quantization. + +Existing lossless model compression does not support efficient GPU inference. Unlike lossy compression, lossless compression reduces model size while preserving the full precision of the original weights. This ensures the model's output distribution remains identical to that of the uncompressed counterpart. However, most existing lossless methods focus on storage efficiency, such as compressing model checkpoints [22, 25], or target specialized hardware like FPGAs [59], rather than accelerating inference on general-purpose GPUs. While useful for tasks like checkpoint rollback during large-scale training [47] or reducing download time from model hubs [25], these methods offer little to no benefit for GPU-based inference. + +Our proposal, Dynamic-Length Float (DFloat11), is a lossless compression framework optimized for efficient GPU inference. We identify a key inefficiency in the commonly used BFloat16 format: its 8-bit exponent field carries only about 2.6 bits of actual information. This redundancy is consistent across a wide range of LLMs, as shown in Section 2.2. To exploit it, we apply Huffman coding [28] to the exponent bits of BFloat16 weights, while leaving the sign and mantissa bits uncompressed. The resulting exponents have dynamic-length encodings: frequent values are assigned shorter codes, while rarer ones use longer codes. However, standard Huffman decoding relies on sequential bit-by-bit tree traversal, which is inefficient on GPUs due to limited parallelism. Assigning one GPU thread per decompression task leads to severe hardware underutilization and high latency. To overcome this, we design a hardware-aware algorithm that enables efficient online decompression of dynamic-length floats on GPUs. Our solution includes three key components: 1. compact, hierarchical lookup tables (LUTs) that fit in GPU SRAM to support fast, table-based Huffman decoding, 2. a two-phase GPU kernel that uses lightweight auxiliary variables to coordinate thread-level read and write operations, and 3. batched decompression at the transformer-block level to maximize throughput. We summarize our contributions as follows: + +1. We propose Dynamic-Length Float (DFloat11), a losslessly compressed floating-point format that reduces BFloat16 weights to approximately 11 bits. This yields around $30\%$ model size reduction with bit-for-bit identical outputs. +2. We develop optimized, hardware-aware algorithms for efficient GPU inference with DFloat11-compressed models by leveraging GPU memory and compute hierarchies. + +![](images/5d92c2eb35a67a1288368801f972f7edcd6d99d253351fdeb1ac33d3c00267b3.jpg) +Figure 1: (Left) The allocation of bits for the components of BFloat16. (Right 3) The Shannon entropy of the components (sign, exponent, mantissa) of BFloat16 weights in various LLMs. + +![](images/826e55136c466707b317d88831978f8e8a8d256338945163c91dd97cad7ad4de.jpg) + +3. We evaluate DFloat11 across popular LLMs and diffusion transformers, including Llama 3, Qwen 3, Mistral 3, DeepSeek R1 Distilled, FLUX.1, and Stable Diffusion 3.5 [20, 46, 45, 21, 32, 2]. Our method consistently achieves $30\%$ compression without altering original outputs at all. Notably, it enables running Llama-3.1-405B on a single node ( $8 \times 80$ GB A100 GPUs), reducing hardware requirements by half without accuracy loss. + +# 2 Method + +In this section, we introduce our proposed floating-point format, Dynamic-Length Float (DFloat11), along with its custom decompression kernel designed for efficient GPU inference. + +# 2.1 Preliminary + +Brain Float (BFloat16) Recent state-of-the-art LLMs predominantly employ the 16-bit Brain Float format (BFloat16 or BF16) for storing weights, due to its balance of numerical precision with memory efficiency. BF16 allocates its 16 bits as follows: 1 sign bit, 8 exponent bits, and 7 mantissa bits. The numerical value represented by a BF16 number is computed as: + +$$ +(- 1) ^ {\text {s i g n}} \times 2 ^ {\text {e x p o n e n t} - 1 2 7} \times (1. \text {m a n t i s s a}), \tag {1} +$$ + +where mantissa is interpreted as a binary fractional value. + +Entropy Coding Entropy coding is a core technique in lossless data compression that leverages statistical redundancy to reduce data size. Several widely used methods fall under this category, including Huffman coding [28], arithmetic coding [33], and Asymmetric Numeral Systems (ANS) [12]. Among these, Huffman coding is one of the most widely adopted, which uses variable-length encoding to minimize the size of encoded data. It assigns shorter binary codes to more frequent symbols and longer codes to less frequent ones. The codes are decoded using a prefix-free binary tree, known as a Huffman tree. Due to the prefix-free property of Huffman codes, no code is a prefix of any other, which ensures unique decodability of the encoded bitstream without the need for delimiters. The tree is constructed based on symbol frequencies and is provably optimal for any given frequency distribution. However, decoding Huffman codes in a massively parallel manner is challenging due to its inherently sequential nature. + +GPU Computation and Memory Paradigm GPUs are designed to perform computations in a massively parallel manner. A modern GPU consists of thousands of threads, which are organized into blocks and executed on streaming multiprocessors (SMs). Each block has access to a small, fast, on-chip memory called shared memory (often referred to as SRAM), which provides much lower latency and higher bandwidth than the off-chip global memory, commonly known as high-bandwidth memory (HBM). The capacity of shared memory is limited, typically having up to 100 KB per block. In this work, we leverage the fast access characteristics of SRAM to enable efficient on-the-fly decompression of compressed weights during inference. + +![](images/64a987150e5ff65489c3e4a3dd5fea533895e7f812a8ebda129ea4126474ad19.jpg) +Figure 2: Our proposed format Dynamic-Length Float for compressing BFloat16 weights of LLMs losslessly down to 11 bits. The exponents are compressed via Huffman coding, while the sign and mantissa bits remain uncompressed. + +# 2.2 Motivation: BFloat16 Representation is Information Inefficient + +To motivate the lossless compression of LLM weights, we analyze the compressibility of the BFloat16 weights of recent LLMs. Specifically, we use Shannon entropy to quantify the information content of BFloat16 components (sign, exponent, and mantissa) for all linear projection matrices of an LLM. The Shannon entropy $H(\cdot)$ is defined as: + +$$ +H (X) = - \sum_ {x \in \mathcal {X}} p (x) \log_ {2} p (x) \tag {2} +$$ + +where $X$ is a discrete random variable with support $\mathcal{X}$ , and $p: \mathcal{X} \to [0,1]$ denotes its probability mass function. We present the computed entropy values in Figure 1. As shown, the entropy of the sign and mantissa bits is close to their respective bit widths, indicating limited potential for compression. In contrast, the exponent exhibits significantly lower entropy, approximately 2.6 bits versus its allocated 8 bits, suggesting substantial opportunities for lossless compression. + +To understand this discrepancy, we visualize the frequency distribution of all BFloat16 components in Figure 8 and the ranked frequency of exponent values in Figure 9, both in the Appendix. The sign and mantissa values are relatively uniform across their ranges, but the exponent distribution is highly imbalanced: only about 40 of the 256 possible 8-bit values are used, with the rest never appearing. Ranked frequencies also decay rapidly. These observations reveal the low entropy of the exponent and its potential for compression. + +# 2.3 Dynamic-Length Float: Lossless LLM Compression for Efficient GPU Inference + +To address the substantial information inefficiency in the BFloat16 representation of LLM weights, we propose a lossless compression framework that encodes floating-point parameters using entropy coding. Specifically, we build a Huffman tree based on the distribution of exponents in model weights. We then compress the exponents using Huffman coding, while preserving the original signs and mantissas. Exponents are encoded and tightly bit-packed into a byte array, EncodedExponent, while the sign and mantissa are left uncompressed and stored in a separate byte array PackedSignMantissa. Figure 2 illustrates Dynamic-Length Float (DFloat11 or DF11), our proposed format for compactly representing BFloat16 model parameters. + +The Core Challenge: Efficient GPU Inference with Compressed Weights While DFloat11 enables lossless compression of LLM weights, efficient GPU inference remains a key challenge. Entropy-coded weights use variable-length encoding and cannot be directly used in matrix multiplications. As a result, each weight matrix must be decompressed on-the-fly to its original BFloat16 format + +![](images/431d07b6849fd9935cc31c081470cd990bcc5e1c0d7ee44d1ec926bc5b361c1c.jpg) +Figure 3: (Left) The Huffman tree is decomposed into a set of non-overlapping subtrees, each corresponding to a compact lookup table (LUT). These hierarchical LUTs reside in GPU SRAM to enable efficient Huffman decoding via array lookups. (Right) Each thread decodes $n$ bytes of encoded exponents. The array Gaps stores the bit offset of the first element assigned to each thread, while the array Block Output Positions stores the index of the first element for each thread block. + +![](images/93b016cc3692241619afcca0ff3234344c5da3147783b55b4d074f2a7412ee80.jpg) + +before matrix multiplication, then discarded immediately after use to conserve memory. However, traditional Huffman decoding is inherently sequential, requiring bit-by-bit tree traversal for each element, which is ill-suited for GPUs' parallel architecture. Naively assigning a single thread for decompression leads to poor utilization and high latency. Addressing this bottleneck is essential for practical compressed inference. + +In the following paragraphs, we present our solution in detail: a set of hardware-aware algorithmic designs tailored for low-latency decoding of entropy-coded weights in a massively parallel manner. Our approach consists of three key components: ① leveraging compact lookup tables that fit within GPU SRAM for efficient, lookup-based decoding, ② introducing a two-phase kernel design to coordinate read/write operations for all threads using lightweight auxiliary variables, and ③ performing decompression at the transformer block level to minimize latency. + +# 2.3.1 Efficient Decoding with Hierarchical Lookup Tables + +The traditional approach to decoding Huffman codes involves reading the encoded bitstream bit by bit and traversing the Huffman tree accordingly. However, this method is inefficient on GPUs due to frequent branching and limited parallelism. To enable efficient decoding on GPUs, we adopt a lookup-table-based approach [53]. + +Assume the maximum Huffman code length is $L$ , and we construct a lookup table LUT of size $2^{L}$ . At each index $i$ , LUT stores the decoded exponent whose Huffman code matches the prefix of the binary representation of $i$ . To decode the next exponent, we read the next $L$ bits from the encoded bitstream, interpret them as an index into LUT, and retrieve the corresponding value. To determine how many bits to advance in the stream, we use a secondary lookup table CodeLengths, which maps each exponent to the length of its Huffman code. A detailed example of this decoding process is provided in Section I of the Appendix. + +In practice, the value of $L$ can be large. For LLMs, $L$ typically ranges from 24 to 32, resulting in a LUT with up to $2^{32}$ entries, which cannot fit within GPU SRAM for fast lookups. To address this, we decompose the monolithic LUT into a hierarchy of compact lookup tables [53]. We first partition the Huffman tree into non-overlapping subtrees of height 8. Each subtree corresponds to a compact LUT that decodes 8 bits, requiring only $2^{8} = 256$ entries. + +Figure 3 shows an example of how a Huffman tree of height 4 can be decomposed into a hierarchy of compact LUTs, each with 4 entries. Because the LUTs are organized hierarchically, some entries must serve as references to other LUTs lower in the hierarchy. We take advantage of the sparsity in 8-bit exponent usage: although 256 values are available, typically only around 40 are used in LLMs (see Figure 9 in the Appendix). We repurpose unused values (specifically, the range 240 to 255) as pointers to other LUTs. These values correspond to extremely large magnitudes ( $\pm 2^{113}$ to $\pm 2^{128}$ ) that do not occur in LLM weights, making them safe for use as internal markers. + +We use $k$ to denote the number of compact LUTs. In our experiments, we observe that $k$ ranges from 4 to 8 for the Huffman trees built from BFloat16 exponent values. Combined with CodeLengths, these LUTs occupy at most $(8 + 1) \times 256$ bytes of memory, which easily fits within SRAM and allows for fast repeated lookups. + +# 2.3.2 Two-Phase Kernel and Lightweight Auxiliary Variables + +To leverage the parallel processing capabilities of GPUs, we assign each thread to a contiguous, non-overlapping block of encoded exponents consisting of $n$ bytes ( $n = 8$ in our experiments). Each thread decodes elements whose Huffman codes begin within its assigned block. Since Huffman codes are variable-length, a thread may need to skip some bits at the start before decoding the first element. Similarly, the last element may span beyond the assigned byte range. + +This approach introduces two key challenges: 1. The starting bit position for each thread is unclear due to the variable-length nature of Huffman codes. 2. Except for the first thread, the index of decoded elements is unknown, making it difficult to determine their correct output locations. + +To address the first issue, we use a gap array [53] to specify the starting bit offset for each thread. The array Gaps has one entry per thread, where each entry indicates the offset of the first valid Huffman code relative to the thread's assigned starting byte. With a maximum code length of 32 bits, each offset lies in $[0, 31]$ and is stored using only 5 bits. + +For the second issue, maintaining an output position for each thread is straightforward but memory-intensive. Each position requires a 32-bit integer, and with tens of thousands of threads per weight matrix, this leads to significant overhead, undermining DFloat11's compression benefits. To reduce this overhead, we store the output position only for the first element of each thread block rather than for every thread. Since each block typically contains hundreds to thousands of threads, this optimization reduces the overhead from one 32-bit integer per thread to one per block, making the memory cost negligible. Figure 3 illustrates how the gap and block-level output position arrays encode the metadata associated with the encoded exponents. + +To support this design, we implement a two-phase kernel. In the first phase, each thread decodes its assigned block and counts the number of elements, without writing to the HBM. Afterward, threads within a block synchronize to compute per-thread output positions via a prefix sum over the element counts. We use the Blelloch algorithm [4] for this step. In the second phase, each thread re-decodes the same block, this time writing decoded values to a write buffer in SRAM at the calculated positions. To avoid redundant global memory access, the encoded exponents are loaded into SRAM before the first pass. Once all decoded exponents are written to SRAM, a single batch of coalesced writes is issued to HBM. Pseudocode for the two-phase kernel is provided in Algorithm 1 of the Appendix. + +# 2.3.3 Transformer-Block-Level Decompression + +We now have a complete recipe for decompressing entropy-coded exponents in a massively parallel manner. During inference, the LLM weights stored in DFloat11 format, along with auxiliary variables (the thread-level gap array and block-level output position array), reside entirely in GPU memory. When a weight matrix is needed for matrix multiplication, it is decompressed on-the-fly into the original BFloat16 format. Once the matrix multiplication is complete, the BFloat16 matrix is immediately discarded to conserve GPU memory. + +In practice, decompressing a single weight matrix often underutilizes GPU resources due to its relatively small size. As the matrix size increases, decompression throughput improves. Figure 7 illustrates this trend, showing how DFloat11 decompression scales with matrix size. To capitalize on this, we propose batching the decompression of multiple matrices together to improve throughput and hide latency. Specifically, we decompress all DFloat11 weight matrices within a transformer block as a single batch. This batched decompression occurs right before the forward pass of the transformer block. We also compress the token embedding and language modeling head of LLMs. Since these matrices are large enough to saturate GPU resources, batching their decompression is unnecessary. + +Table 1: DF11 statistics for various models. Model sizes are shown before and after compression. + +
ModelOriginal → DF11 CompressedCompression RatioAvg. Bit Width
Large Language Models
Llama 3.1 8B Instruct16.06 GB → 10.90 GB67.84%10.85
Llama 3.3 70B Instruct141.11 GB → 95.40 GB67.61%10.82
Llama 3.1 405B Instruct811.71 GB → 551.22 GB67.91%10.87
Qwen 3 14B29.54 GB → 20.14 GB68.17%10.91
QwQ 32B65.53 GB → 44.65 GB68.14%10.90
Mistral Nemo Instruct24.50 GB → 16.59 GB67.74%10.84
Mistral Small 347.14 GB → 31.86 GB67.58%10.81
Phi 4 Reasoning Plus29.32 GB → 19.83 GB67.64%10.82
DeepSeek R1 Distill Llama 8B16.06 GB → 10.89 GB67.81%10.85
Diffusion Transformers
FLUX.1 dev23.80 GB → 16.33 GB68.61%10.98
FLUX.1 schnell23.78 GB → 16.31 GB68.58%10.97
Stable Diffusion 3.5 Large16.29 GB → 11.33 GB69.52%11.12
+ +Table 2: Comparison of accuracy and perplexity for the BF16 and DF11 models on different benchmarks. DF11 compression results in absolutely no loss in accuracy or perplexity. + +
ModelData TypeAccuracyPerplexity
MMLUTruthfulQAWikiTextC4
Llama 3.1 8B InstructBF1668.010 ± 0.37536.965 ± 1.6908.64921.677
DF11 (Ours)68.010 ± 0.37536.965 ± 1.6908.64921.677
+ +# 3 Experiments + +We empirically evaluate the effectiveness of DF11 compression and its GPU inference efficiency. A range of recent LLMs and DMs are compressed from their original BFloat16 format into DF11, and we report the resulting compression ratios. We then compare the inference performance of DF11-compressed models against their uncompressed counterparts across different GPUs, followed by an ablation study to analyze the impact of compression. + +Software and Hardware We implement the DF11 decompression kernel in CUDA and $\mathrm{C + + }$ , and integrate it into the HuggingFace Transformers [48] inference framework. We evaluate the inference efficiency of our DF11 models against the original BF16 counterparts. We use the HuggingFace Accelerate framework to support CPU offloading and multi-GPU inference. To assess the performance of the DF11 kernel across different hardware configurations, we run experiments on multiple machines with varying GPU and CPU setups. The hardware specifications for all experimental machines are provided in Table 4 in the Appendix. + +# 3.1 Results + +DF11 compresses models to $70\%$ size. Table 1 presents the compression factors of DF11 for a wide selection of recent LLMs and DMs. Specifically, we apply compression to all weight matrices and token embeddings in LLMs and all weight matrices in the transformer blocks of DMs. The models we compress include Llama 3.1/3.3 [20], Qwen 3 [54], Mistral Nemo/Small [44, 45], Phi 4 [1], DeepSeek R1 Distilled [21], Stable Diffusion 3.5 [2], FLUX.1 [32]. DF11 achieves approximately $70\%$ compression across all models, corresponding to an effective bit width of around 11 bits. + +Accuracy and perplexity evaluations confirm DF11 compression is lossless. We verify the lossless property of DF11 compression through a series of accuracy and perplexity evaluations on standard benchmarks. Evaluations are conducted using lm.evaluation_harness [18], reporting accuracy on MMLU [24] and TruthfulQA [38], and word-level perplexity on WikiText [41] and C4 [42]. The results are shown in Table 2. As demonstrated, the compressed model achieves identical accuracy and perplexity to the original BF16 counterpart. We also present the text-to-image + +Table 3: Comparison of peak GPU memory usage and text-to-image generation time for diffusion transformers in BF16 and DF11, using a single A5000 GPU. + +
ModelPeak GPU Memory (GB)Generation Time (s)
BF16DF11 (Ours)BF16DF11 (Ours)
Stable Diffusion 3.5 Large16.4411.7866.36 ± 0.1369.08 ± 0.11
FLUX.1 dev23.1516.7274.41 ± 0.1578.53 ± 0.18
+ +![](images/783078f5bdd5c286f830963b4077284f4b8f4334a4a03ed2a14056b2bc446e38.jpg) + +![](images/f097bd7128e57136a5f3ee553ff483d0c2623555e024cc554f6f2a8b4dc277a5.jpg) +Qwen 3A4B (28GB) GPU: A5000 (24GB) + +![](images/7c3d46df1881a46425dbc36391ca83268eb67b4e0639f997f392c949b2df31de.jpg) + +![](images/4a614b789dd526604acd9a67e1ec1df6a164f04d2ff8b39dbc18f1104059987c.jpg) + +![](images/88745fd4e3328a57457087325a886c32438f4ce2f2e8994b5c7e65a19cc59eea.jpg) +Figure 4: Comparison of throughput (top row) and latency (bottom row) for token decoding using the original BF16 models and their DF11-compressed counterparts. Portions of the BF16 models are offloaded to the CPU due to GPU memory constraints. + +![](images/6891428b56b7561a134884064c9d8f00c696151d0868f8160004c6109491ccf3.jpg) +Mistral Small 3 (48GB) GPU: A100 (40GB) + +![](images/6bf360ebd7248f8b5ea117cdec3291dbf9d2668c3563d799cf8b6c9519270674.jpg) +QwQ32B (64GB) GPU:RTX8000 (48GB) + +generation results of BF16 and DF11 Stable Diffusion 3.5 Large model in Appendix J. Given the same random seed and text prompt, the image generated are pixel-wise identical with the original model. + +DF11 outperforms CPU offloading in inference efficiency. We compare the inference performance of DF11 and BF16 models across various hardware platforms. Due to memory constraints, BF16 models exceed the capacity of a single GPU and require partial CPU offloading, while DF11 models fit entirely within GPU memory. For fair comparison, we retain most computation on the GPU for BF16 models and offload only necessary components. Latency and throughput are measured after a 100-token warm-up run, followed by decoding 100 tokens from an empty prompt across varying batch sizes. Each configuration is run five times, and we report the average results. As shown in Figure 4, DF11 consistently outperforms BF16 with CPU offloading, achieving $2.31 - 46.24 \times$ lower latency or higher throughput. Multi-GPU comparisons are shown in Figure 10 in the Appendix. + +DF11 reduces memory usage for diffusion transformers with minimal latency impact. We assess the impact of DF11 compression on diffusion transformer models by measuring peak GPU memory usage and text-to-image generation latency for an $1024 \times 1024$ image across five runs. Neither the BF16 nor DF11 models employ CPU offloading. As shown in Table 3, DF11 reduces memory consumption by $28.3\%$ for Stable Diffusion 3.5 and $27.8\%$ for FLUX.1. The relative increase in latency is small: $4.1\%$ for Stable Diffusion and $5.5\%$ for FLUX.1. + +DF11 memory savings enable longer generation lengths. DF11 compression not only can reduce the number of GPUs needed for inference but can also support longer generation under the same VRAM budget. During decoding, the KV cache grows linearly with the number of tokens and quickly becomes a memory bottleneck. Figure 5 shows GPU memory usage for DF11 and BF16 models with batch size 1 as token count increases. DF11 allows 5.70 to $14.86 \times$ more tokens to be decoded before reaching memory limits. + +![](images/45836ff77df6a4ae58c952e1fd6e6f501c6bfde943c3d200ca6d9d4d34a0e33b.jpg) +Figure 5: Comparison of GPU memory consumption between BF16 models and DF11 counterparts. The DF11 models support $5.70 - 14.86 \times$ longer context lengths by allowing more GPU memory to be used for storing the KV cache. "O.O.M." means out of memory. + +![](images/e0e121565d16b0eab2f8478554556929b83db81b285f3ee831b28fe9a4d47c65.jpg) +Figure 6: Comparison of latency breakdown for DF11 and BF16 Llama 3.1 8B Instruct during GPU inference for different token batch sizes, using one A100-40GB GPU. + +# 3.2 Ablation Study + +Latency breakdown shows decompression overhead is amortized at larger batch sizes. We analyze the latency of Llama 3.1 8B Instruct in BF16 and DF11 formats across varying token batch sizes on an A100-40GB GPU. For each setting, we measure the average latency of each component over 10 runs, as shown in Figure 6. DF11 introduces additional latency from decompressing the token embedding, transformer blocks, and language modeling head. This overhead is constant and independent of batch size, so increasing the token batch size effectively amortizes the cost. + +DF11 decompression is significantly faster than CPU-to-GPU transfer and nvCOMP ANS. We compare DF11 decompression latency and throughput with two baselines: CPU-to-GPU weight transfer and ANS decompression [12] from NVIDIA's nvCOMP [6], using sliced weight matrices from the Llama 3.1 8B Instruct language modeling head. As shown in Figure 7, DF11 achieves up to $34.95 \times$ higher throughput than CPU transfer and up to $20.97 \times$ faster decompression than nvCOMP. DF11 also offers a better compression ratio (68%) compared to nvCOMP (79%). Moreover, DF11 decompression throughput improves with larger matrix sizes due to better GPU utilization. + +![](images/1f5e9774b0ea594a350a7d4303783e09f38060b912a9e3f6355f99e7bd5f3116.jpg) +Figure 7: Throughput (left two) and latency (right two) comparisons between transferring BF16 matrices from CPU to GPU and decompressing the same matrices on GPU using the NVIDIA nvCOMP ANS library and our proposed DF11 kernel, across matrix sizes and GPU types. + +# 4 Related Works + +Data Formats for Model Weights Full-precision model weights are typically stored in formats such as BF16, FP16, or FP32. Several works have proposed 4-bit compressed formats, including FP4, INT4, NF4 (NormalFloat) [9], AF4 (AbnormalFloat) [58], and SF4 (Student Float) [11], which represent each parameter with 4 bits. Unlike these lossy formats, the proposed DF11 format compresses weights losslessly. + +Lossless Model Compression While lossy compression methods such as pruning [14] and quantization [37, 15] are well-studied, lossless compression remains less explored. Four prior works have addressed this area. Deep Compression [22] applied Huffman coding [28] to quantized CNNs, achieving $22\%$ additional compression. ZipNN [25] extended this approach to language models with improved compression over classical methods. However, both techniques target storage efficiency and do not support inference-time gains. NeuZip [23] is the only prior work supporting GPU inference. It uses Asymmetric Numeral Systems (ANS) with layer-wise decompression and relies on NVIDIA's nvCOMP for GPU-based operations. nvCOMP is no longer open source, and its binary-only distribution limits adoption. Moreover, as shown in Figure 7, nvCOMP ANS incurs higher latency and lower throughput compared to our DFloat11 kernel. Huff-LLM [59] is designed for FPGA-like hardware and is not applicable to GPUs. Additional discussion of related formats is presented in Appendix B. + +# 5 Conclusion + +We introduce Dynamic-Length Float (DFloat11), a lossless compression framework designed for efficient GPU inference of BFloat16 models, including both large language models (LLMs) and diffusion models (DMs). DFloat11 exploits the information redundancy inherent in foundation model weights through entropy-coded, dynamic-length encoding, achieving compression rates close to the information-theoretic limit. To enable efficient deployment, we develop hardware-aware algorithms that support high-speed inference directly on compressed weights. Extensive experiments demonstrate that DFloat11 significantly reduces GPU memory requirements for LLMs and DMs, allowing for longer generation lengths, while maintaining bit-exact accuracy and incurring only negligible decompression overhead. + +# Acknowledgements + +This work was supported by National Science Foundation SHF-2211815 and Ken Kennedy Institute Cluster Grants. Additionally, Henry and Xia are supported by ITE-2429680, IIS-2310260, and US Department of Transportation (USDOT) Tier-1 University Transportation Center (UTC) Transportation Cybersecurity Center for Advanced Research and Education (CYBER-CARE) grant #69A3552348332. Mohsen and Vipin are supported by OAC-2320952, OAC-2112606, and OAC-2117439. The views and conclusions in this paper are those of the authors and do not represent the views of any funding or supporting agencies. + +# References + +[1] Marah Abdin, Jyoti Aneja, Harkirat Behl, Sébastien Bubeck, Ronen Eldan, Suriya Gunasekar, Michael Harrison, Russell J Hewett, Mojan Javaheripi, Piero Kauffmann, et al. Phi-4 technical report. arXiv preprint arXiv:2412.08905, 2024. +[2] Stability AI. Introducing stable diffusion 3.5. https://stability.ai/news/introducing-stable-diffusion-3-5, October 2024. Accessed: May 15, 2025. +[3] Anonymous. FAFO: Lossy KV cache compression for lossless inference acceleration via draftless fumble decoding. In Submitted to The Fourteenth International Conference on Learning Representations, 2025. under review. +[4] Guy E Blelloch. Scans as primitive parallel operations. IEEE Transactions on computers, 38(11):1526-1538, 1989. +[5] Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, Christopher Hesse, and John Schulman. Training verifiers to solve math word problems. arXiv preprint arXiv:2110.14168, 2021. +[6] NVIDIA Corporation. nvCOMP: GPU-accelerated compression and decompression library. https://developer.nvidia.com/nvcomp, 2025. Accessed: April 11, 2025. +[7] Tri Dao, Dan Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. Flashattention: Fast and memory-efficient exact attention with io-awareness. Advances in neural information processing systems, 35:16344-16359, 2022. +[8] Tim Dettmers, Mike Lewis, Younes Belkada, and Luke Zettlemoyer. Gpt3. int8(): 8-bit matrix multiplication for transformers at scale. Advances in neural information processing systems, 35:30318-30332, 2022. +[9] Tim Dettmers, Artidoro Pagnoni, Ari Holtzman, and Luke Zettlemoyer. Qlora: Efficient finetuning of quantized llms. Advances in neural information processing systems, 36:10088-10115, 2023. +[10] P. Deutsch and J.-L. Gailly. Rfc1950: Zlib compressed data format specification version 3.3, 1996. +[11] Jordan Dotzel, Yuzong Chen, Bahaa Kotb, Sushma Prasad, Gang Wu, Sheng Li, Mohamed S Abdelfattah, and Zhiru Zhang. Learning from students: Applying t-distributions to explore accurate and efficient formats for llms. In International Conference on Machine Learning, pages 11573-11591. PMLR, 2024. +[12] Jarek Duda. Asymmetric numeral systems: entropy coding combining speed of huffman coding with compression rate of arithmetic coding. arXiv preprint arXiv:1311.2540, 2013. +[13] Abhinav Dutta, Sanjeev Krishnan, Nipun Kwatra, and Ramachandran Ramjee. Accuracy is not all you need. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. +[14] Elias Frantar and Dan Alistarh. Sparsegpt: Massive language models can be accurately pruned in one-shot. In International Conference on Machine Learning, pages 10323-10337. PMLR, 2023. +[15] Elias Frantar, Saleh Ashkboos, Torsten Hoefler, and Dan Alisthar. Gptq: Accurate post-training quantization for generative pre-trained transformers. arXiv preprint arXiv:2210.17323, 2022. +[16] Yichao Fu, Peter Bailis, Ion Stoica, and Hao Zhang. Break the sequential dependency of LLM inference using lookahead decoding. In *Forty-first International Conference on Machine Learning*, 2024. +[17] Kazuki Fujii, Taishi Nakamura, and Rio Yokota. Balancing speed and stability: The trade-offs of fp8 vs. bf16 training in llms. arXiv preprint arXiv:2411.08719, 2024. + +[18] Leo Gao, Jonathan Tow, Baber Abbasi, Stella Biderman, Sid Black, Anthony DiPofi, Charles Foster, Laurence Golding, Jeffrey Hsu, Alain Le Noac'h, Haonan Li, Kyle McDonell, Niklas Muennighoff, Chris Ociepa, Jason Phang, Laria Reynolds, Hailey Schoelkopf, Aviya Skowron, Lintang Sutawika, Eric Tang, Anish Thite, Ben Wang, Kevin Wang, and Andy Zou. A framework for few-shot language model evaluation, 07 2024. +[19] Ruihao Gong, Yang Yong, Shiqiao Gu, Yushi Huang, Chengtao Lv, Yunchen Zhang, Xianglong Liu, and Dacheng Tao. Llmc: Benchmarking large language model quantization with a versatile compression toolkit. arXiv preprint arXiv:2405.06001, 2024. +[20] Aaron Grattafori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. +[21] Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948, 2025. +[22] Song Han, Huizi Mao, and William J Dally. Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding. arXiv preprint arXiv:1510.00149, 2015. +[23] Yongchang Hao, Yanshuai Cao, and Lili Mou. Neuzip: Memory-efficient training and inference with dynamic compression of neural networks. arXiv preprint arXiv:2410.20650, 2024. +[24] Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. Measuring massive multitask language understanding. In International Conference on Learning Representations. +[25] Moshik Hershcovitch, Andrew Wood, Leshem Choshen, Guy Girmonsky, Roy Leibovitz, Ilias Ennmouri, Michal Malka, Peter Chin, Swaminathan Sundararaman, and Danny Harnik. Zipnn: Lossless compression for ai models. arXiv preprint arXiv:2411.05239, 2024. +[26] Coleman Richard Charles Hooper, Sehoon Kim, Hiva Mohammadzadeh, Michael W. Mahoney, Sophia Shao, Kurt Keutzer, and Amir Gholami. KVQuant: Towards 10 million context length LLM inference with KV cache quantization. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. +[27] Edward J Hu, yelong shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuzhhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2022. +[28] David A Huffman. A method for the construction of minimum-redundancy codes. Proceedings of the IRE, 40(9):1098-1101, 1952. +[29] Renren Jin, Jiangcun Du, Wuwei Huang, Wei Liu, Jian Luan, Bin Wang, and Deyi Xiong. A comprehensive evaluation of quantization strategies for large language models. In *Findings of the Association for Computational Linguistics ACL* 2024, pages 12186–12215, 2024. +[30] Dhiraj Kalamkar, Dheevatsa Mudigere, Naveen Mellempudi, Dipankar Das, Kunal Banerjee, Sasikanth Avancha, Dharma Teja Vooturi, Nataraj Jammalamadaka, Jianyu Huang, Hector Yuen, et al. A study of bfloat16 for deep learning training. arXiv preprint arXiv:1905.12322, 2019. +[31] Artyom Kharinaev, Viktor Moskvoretskii, Egor Shvetsov, Kseniia Studenikina, Bykov Mikhail, and Evgeny Burnaev. Investigating the impact of quantization methods on the safety and reliability of large language models. arXiv preprint arXiv:2502.15799, 2025. +[32] Black Forest Labs. Flux. https://github.com/black-forest-labs/flux, 2024. +[33] G. G. Langdon. An introduction to arithmetic coding. IBM Journal of Research and Development, 28(2):135-149, 1984. +[34] Yaniv Leviathan, Matan Kalman, and Yossi Matias. Fast inference from transformers via speculative decoding. In Fortieth International Conference on Machine Learning, 2023. + +[35] Shiyao Li, Xuefei Ning, Luning Wang, Tengxuan Liu, Xiangsheng Shi, Shengen Yan, Guohao Dai, Huazhong Yang, and Yu Wang. Evaluating quantized large language models. In International Conference on Machine Learning, pages 28480-28524. PMLR, 2024. +[36] Xiuyu Li, Yijiang Liu, Long Lian, Huanrui Yang, Zhen Dong, Daniel Kang, Shanghang Zhang, and Kurt Keutzer. Q-diffusion: Quantizing diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17535-17545, 2023. +[37] Ji Lin, Jiaming Tang, Haotian Tang, Shang Yang, Wei-Ming Chen, Wei-Chen Wang, Guangxuan Xiao, Xingyu Dang, Chuang Gan, and Song Han. Awq: Activation-aware weight quantization for on-device llm compression and acceleration. Proceedings of Machine Learning and Systems, 6:87-100, 2024. +[38] Stephanie Lin, Jacob Hilton, and Owain Evans. Truthfulqa: Measuring how models mimic human falsehoods. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 3214-3252, 2022. +[39] Ruikang Liu, Yuxuan Sun, Manyi Zhang, Haoli Bai, Xianzhi Yu, Tiezheng Yu, Chun Yuan, and Lu Hou. Quantization hurts reasoning? an empirical study on quantized reasoning models. arXiv preprint arXiv:2504.04823, 2025. +[40] Zirui Liu, Jiayi Yuan, Hongye Jin, Shaochen Zhong, Zhaozhuo Xu, Vladimir Braverman, Beidi Chen, and Xia Hu. KIVI: A tuning-free asymmetric 2bit quantization for KV cache. In Forty-first International Conference on Machine Learning, 2024. +[41] Stephen Merity, Caiming Xiong, James Bradbury, and Richard Socher. Pointer sentinel mixture models. In International Conference on Learning Representations, 2017. +[42] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of Machine Learning Research, 21(140):1-67, 2020. +[43] Yang Sui, Yanyu Li, Anil Kag, Yerlan Idelbayev, Junli Cao, Ju Hu, Dhritiman Sagar, Bo Yuan, Sergey Tulyakov, and Jian Ren. Bitsfusion: 1.99 bits weight quantization of diffusion model. arXiv preprint arXiv:2406.04333, 2024. +[44] Mistral AI Team. Mistral NeMo. https://mistral.ai/news/mistral-nemo, July 2024. +[45] Mistral AI Team. Mistral Small 3. https://mistral.ai/news/mistral-small-3, January 2025. +[46] Qwen Team. Qwen3: Think deeper, act faster, April 2025. +[47] Zhuang Wang, Zhen Jia, Shuai Zhang, Zhen Zhang, Mason Fu, T. S. Eugene Ng, and Yida Wang. Gemini: Fast failure recovery in distributed training with in-memory checkpoints. 2023. +[48] Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumont, Clement Delangue, Anthony Moi, Pierrick Cistac, Tim Rault, Rémi Louf, Morgan Funtowicz, et al. Transformers: State-of-the-art natural language processing. In Proceedings of the 2020 conference on empirical methods in natural language processing: system demonstrations, pages 38-45, 2020. +[49] Heming Xia, Tao Ge, Peiyi Wang, Si-Qing Chen, Furu Wei, and Zhifang Sui. Speculative decoding: Exploiting speculative execution for accelerating seq2seq generation. arXiv preprint arXiv:2203.16487, 2022. +[50] Heming Xia, Zhe Yang, Qingxiu Dong, Peiyi Wang, Yongqi Li, Tao Ge, Tianyu Liu, Wenjie Li, and Zhifang Sui. Unlocking efficiency in large language model inference: A comprehensive survey of speculative decoding. arXiv preprint arXiv:2401.07851, 2024. +[51] Guangxuan Xiao, Ji Lin, Mickael Seznec, Hao Wu, Julien Demouth, and Song Han. Smoothquant: Accurate and efficient post-training quantization for large language models. In International Conference on Machine Learning, pages 38087-38099. PMLR, 2023. + +[52] Mengwei Xu, Wangsong Yin, Dongqi Cai, Rongjie Yi, Daliang Xu, Qipeng Wang, Bingyang Wu, Yihao Zhao, Chen Yang, Shihe Wang, et al. A survey of resource-efficient llm and multimodal foundation models. arXiv preprint arXiv:2401.08092, 2024. +[53] Naoya Yamamoto, Koji Nakano, Yasuaki Ito, Daisuke Takafuji, Akihiko Kasagi, and Tsuguchika Tabaru. Huffman coding with gap arrays forgpu acceleration. In Proceedings of the 49th International Conference on Parallel Processing, pages 1-11, 2020. +[54] An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, et al. Qwen2. 5 technical report. arXiv preprint arXiv:2412.15115, 2024. +[55] Ge Yang, Changyi He, Jinyang Guo, Jianyu Wu, Yifu Ding, Aishan Liu, Haotong Qin, Pengliang Ji, and Xianglong Liu. LLMCBench: Benchmarking large language model compression for efficient deployment. In The Thirty-eight Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2024. +[56] Jingfeng Yang, Haongye Jin, Ruixiang Tang, Xiaotian Han, Qizhang Feng, Haoming Jiang, Shaochen Zhong, Bing Yin, and Xia Hu. Harnessing the power of llms in practice: A survey on chatgpt and beyond. 2024. +[57] Ling Yang, Zhilong Zhang, Yang Song, Shenda Hong, Runsheng Xu, Yue Zhao, Wentao Zhang, Bin Cui, and Ming-Hsuan Yang. Diffusion models: A comprehensive survey of methods and applications. ACM Computing Surveys, 56(4):1-39, 2023. +[58] Davis Yoshida. Nf4 isn't information theoretically optimal (and that's good). arXiv preprint arXiv:2306.06965, 2023. +[59] Patrick Yubeaton, Tareq Mahmoud, Shehab Naga, Pooria Taheri, Tianhua Xia, Arun George, Yasmein Khalil, Sai Qian Zhang, Siddharth Joshi, Chinmay Hegde, et al. Huff-llm: End-to-end lossless compression for efficient llm inference. arXiv preprint arXiv:2502.00922, 2025. +[60] Haochen Zhang, Junze Yin, Guanchu Wang, Zirui Liu, Lin Yang, Tianyi Zhang, Anshumali Shrivastava, and Vladimir Braverman. Breaking the frozen subspace: Importance sampling for low-rank optimization in LLM pretraining. In The Thirty-ninth Annual Conference on Neural Information Processing Systems, 2025. +[61] Tianyi Zhang and Anshumali Shrivastava. Leanquant: Accurate and scalable large language model quantization with loss-error-aware grid. In The Thirteenth International Conference on Learning Representations, 2025. +[62] Tianyi Zhang, Junda Su, Aditya Desai, Oscar Wu, Zhaozhuo Xu, and Anshumali Shrivastava. Sketch to adapt: Fine-tunable sketches for efficient LLM adaptation. In *Forty-second International Conference on Machine Learning*, 2025. +[63] Tianyi Zhang, Jonah Wonkyu Yi, Zhaozhuo Xu, and Anshumali Shrivastava. KV cache is 1 bit per channel: Efficient large language model inference with coupled quantization. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. +[64] Tianyi Zhang, Jonah Wonkyu Yi, Bowen Yao, Zhaozhuo Xu, and Anshumali Shrivastava. NoMAD-attention: Efficient LLM inference on CPUs through multiply-add-free attention. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. +[65] Jiawei Zhao, Zhenyu Zhang, Beidi Chen, Zhangyang Wang, Anima Anandkumar, and Yuandong Tian. Galore: Memory-efficient LLM training by gradient low-rank projection. In *Forty-first International Conference on Machine Learning*, 2024. + +# Appendix + +# A Discussion: Is Quantization a Universal Solution? + +Much of the motivation behind our work lies in understanding whether lossless compression of large-scale models such as LLMs, which preserves $100\%$ identical output behavior compared to the original uncompressed model, is a practical direction worthy of further study. Specifically, how does DFloat11, which compresses LLMs to approximately 11 bits, compare to widely used lossy quantization techniques [15, 37], where models are typically reduced to even lower bit-widths (e.g., 8-bit or 4-bit)? + +The answer is far more nuanced than a simple "Yes/No" or a one-size-fits-all judgment about which approach is better. For instance, existing benchmark studies like [19, 55, 29] often suggest that 8-bit (weight-only or not) quantization is a relatively "safe" compression scheme. Although technically lossy, 8-bit models can often maintain strong task performance across a range of standard benchmarks. However, we must note these benchmarks typically focus on a narrow set of tasks (e.g., WikiText2 perplexity, MMLU, Commonsense Reasoning), and thus fail to offer a comprehensive view of real-world LLM usage, especially from the perspective of end-users. + +That being said, the argument that "current benchmarks fail to capture the performance gap between 8-bit compressed and 16-bit uncompressed models" is itself constrained by the limitations of the current benchmarking landscape, making it difficult to produce abundant supporting evidence. Nonetheless, some reports have begun to highlight such gaps. For example, human evaluations on LLM Arena show a notable performance drop between Llama-3.1-405B-Instruct [20] and its W8A8 counterpart (Llama-3.1-405B-Instruct-FP8), particularly under coding (1293 vs. 1277) and long-query (1282 vs. 1275) tasks. Similarly, quantizing DeepSeek-R1-Distill-Llama-70B [21] from 16 bits to 8 bits results in a $23.7\%$ drop on GPQA (from $9.51\%$ to $7.25\%$ ).2 Furthermore, reasoning, a core capability of modern LLMs, appears especially sensitive to compression loss. Recent benchmark [39] reveals that quantizing DeepSeek-R1-Distill-Qwen-1.5B with 8-bit SmoothQuant [51] (for weight, attention, and KV cache) leads to an average $9.09\%$ drop in reasoning tasks ( $48.82\%$ to $44.29\%$ ) across datasets like AIME, MATH-500, GPQA-Diamond, and LiveCodeBench. We leave more evidence exploring the performance gap between 8-bit quantized and uncompressed model in Appendix H. + +Although the broader question: "Which specific task, on which model, using which quantization technique, under what conditions, will lead to a noticeable drop compared to FP16/BF16?" is likely to remain open-ended simply due to the sheer amount of potential combinations. It is fair to say that lossy quantization introduces complexities that some end-users would prefer to avoid, since it creates uncontrolled variables that must be empirically stress-tested for each deployment scenario. + +To eliminate this burden, DFloat11 offers a compelling alternative: delivering $100\%$ identical performance to the original model, while consuming only $\sim 70\%$ of the memory footprint with throughput benefits, which is a unique and practical offering for resource-constrained deployment settings. + +# B Extended Related Works + +Data Formats for Model Weights LLM weights are typically stored in compact floating-point formats such as FP16 or BFloat16 (officially stylized as $bfloat16^3$ ). FP16 allocates 1 sign bit, 5 exponent bits, and 10 mantissa bits, whereas BFloat16 uses 1 sign bit, 8 exponent bits, and 7 mantissa bits. Compared to FP16, BFloat16 offers a wider dynamic range at the cost of precision, which improves numerical stability and mitigates overflow issues during training [17, 30]. + +Compressed data formats typically aim for lower bit-widths. For example, FP8—which comes in both E4M3 (4 exponent bits, 3 mantissa bits, plus 1 sign bit) and E5M2 configurations—has seen reasonable adoption in LLM training and development. Integer formats like INT8 have also been well explored, as in LLM. int8() [8] and its following works. Formats with a stronger emphasis on efficiency, such + +as FP4, INT4, NF4 [9], and AF4 [58], use only 4 bits. In this work, we primarily focus on formats with $\geq 8$ bits, as benchmark literature [55, 19, 39] often suggests that 8-bit quantization results in negligible performance drop—though we show in Section A that this claim is likely skewed due to evaluation selectiveness and benchmark limitations. + +Lossless Model Compression While lossy model compression techniques such as pruning and quantization [14, 37, 15] have received widespread attention, lossless model compression remains a relatively underexplored area. Upon careful investigation, we identified roughly four prior works that have made meaningful efforts in this space. Deep Compression [22] is a foundational work, applying Huffman coding [28] to quantized CNN models and achieving an additional $\sim 22\%$ compression gain for model checkpoints. ZipNN [25] extended this idea to language models, comparing its results to classic lossless compression tools such as zlib [10] and zstd4 and demonstrated superior compression gains. However, this line of work—including their industry counterparts, such as ezm7—is limited in that its efficiency gains only apply to storage (reducing the size of model checkpoints) but offer no benefits during inference. While such storage savings are meaningful in large-scale training settings—where frequent snapshotting and checkpoint rollbacks are needed [47)—they have limited impact for everyday LLM end-users. Model downloading is typically a one-time cost, so even if a model checkpoint is compressed by $50\%$ , it only cuts the download time at most by half, presumably over the model's entire lifecycle of deployment. Furthermore, checkpoints are usually stored on disk, where terabytes of capacity are easily available, making up a much looser constraint compared to GPU HBM (High Bandwidth Memory); one of the main resource constraints during inference. + +We argue that a lossless compression technique would be substantially more impactful if it could deliver efficiency gains during inference—particularly on GPU-based systems, which is the default setup for LLM serving. In this context, NeuZip [23] is the only prior work we identify that supports GPU inference. NeuZip applies entropy encoding with layer-wise decompression to maintain a reduced memory footprint throughout serving. However, it is built on NVIDIA's nvCOMP: "a high-speed data compression and decompression library optimized for NVIDIA GPUs".6 Unfortunately, nvCOMP is no longer open-source (only binary executables are available), which hinders future research. Moreover, we empirically find that nvCOMP's inference throughput and latency are significantly worse than our proposed DFloat11 kernel, resulting in a pipeline that trades memory efficiency for substantial inference overhead (see Figure 7). + +Another work referencing NeuZip is *Huff-LLM* [59], which also aims to reduce memory costs while maintaining efficient inference. However, its contributions are specific to FPGA-like architectures and do not apply to GPUs. To the best of our knowledge, the DFloat data format we presented (and its respective kernel support in DFloat11) shall serve as the only GPU-inference-friendly data format with lossless compression benefits. + +Efficient LLM Inference LLMs are computationally intensive and resource-demanding, making the efficiency of LLM inference a key research focus [52]. FlashAttention [7] accelerates exact attention computation on GPUs through kernel fusion, while NoMAD Attention [64] speeds up attention on CPUs using in-register lookups. Model compression is another effective strategy to reduce resource requirements for serving LLMs and diffusion models. Quantization methods such as GPTQ [15], AWQ [37], SmoothQuant [51], LeanQuant [61], CQ [63], KVQuant [26], and KIVI [40] lower memory usage and enhance efficiency by compressing model weights, activations, or KV cache. Compression is also applied in fine-tuning: methods like LoRA [27], QLoRA [9], and SketchTune [62] compress model weight deltas, whereas GaLore [65] and SARA [60] compress optimizer states during training. One additional line of work relevant to efficient LLM inference would be lossless efficient decoding, where paradigms such as speculative decoding [49, 34, 50] and $n$ -gram candidate decoding [16, 3] offer lossless generation quality with improved latency. DFloat11 mainly differs from these works in that it provides substantial savings in memory footprint while maintaining lossless generation quality, whereas most—if not all—lossless efficient decoding methods require memory consumption equal to or greater than that of the original model. + +https://encode.su/threads/ + +![](images/fab4295a226f8488d8cd85f3acd19b1d34cc6d62f2f4ce8a7635733f74f1a162.jpg) + +![](images/6fc88f26a4c3caecddf151d03badcb9654cbf8d5cee26428d29a2f644b8dcb9c.jpg) + +![](images/b3f97d5ecc376c8373b5aa86cb6441a5aabb9d36d04c89b6140942fd65d67f7a.jpg) + +![](images/c62a1ef3b49bb27ed272a5bc75ea8ebbea12ffec6d58f689dede3da1af94ab70.jpg) + +![](images/3084496a4c7373f08fb5b908223d31ceb6a0398f33bdefc05f4bf3b6cd763c0c.jpg) + +![](images/2222b99d099a4d81021773b49e31953c2cb3755a88cea97706d77a7a4104c164.jpg) + +![](images/4b5448c83c6b0736f1919800bf7a1ac957129271669a75470c5cbf36215cf2ba.jpg) + +![](images/b6fa69ef5ae243a9ea31e117929b94f3b1e7e11db5dcab82d349958558e771e8.jpg) + +![](images/ea2c32843c0684bb8a62314c078110d4b00504329f36225eff8f6cb56dc61583.jpg) + +![](images/685806e489498f3f9a5fb1dc73c55dae479e5a2ea9e59a335699bffbe6b2ec1e.jpg) + +![](images/7c260dab8537380cebe8bb399aceb5811e62129d64d40f69d53ac415652524b6.jpg) + +![](images/2bec79029dfd6da538fa351a3d2960585a0b94bd83e945d31e2aa338663470f8.jpg) + +![](images/efce7e95f664acf745534c02e468bd367500f703c87a6b52aeae9f89492959ff.jpg) +Figure 8: Relative frequency distribution of sign, exponent, and mantissa values in the BFloat16 weights of all linear projection layers across various LLMs. + +![](images/36ffb61ab4bd96e56d44b970a46949d61482a10ed9e4b991381cba6aa699a539.jpg) + +![](images/0aab2147ce6a7facfc1c8e784eb8c12c0db7dd80e8edede863b3b5268fdf8a23.jpg) + +# C Frequency Distribution of BFloat16 Values + +Figure 8 presents the frequency distribution for distinct values of sign, exponent, and mantissa bits in the BFloat16 weights of LLMs. Figure 9 shows the sorted frequency of exponent values of LLM weights. + +![](images/688877942cdcb25631b1a1a931fab7d4478a3bd22ef809bca9d72eb98433b3b1.jpg) +Figure 9: Distribution of BFloat16 exponent values across various models. The frequency of exponent values (shown in log scale) decays rapidly with exponent rank. + +![](images/5f54d3b6451b832d1b10153a2e989971f4dc4f5be9e571edd907fe017185fe84.jpg) + +![](images/8e7493757f2d8b3eb277e52aa57619923b3634c8ea00302d5db6299cb43d7aba.jpg) + +![](images/363ec831ba21bea4c3ab21b3cc45f92a30415fae59bf62714f983f518ec9a12b.jpg) + +![](images/1ff4ed6ba4adf872de6054812b45ca95857215b3b0c2c8ad994c764350ac92f7.jpg) + +# D Pseudo-code of the GPU kernel for DFloat11 Decompression + +Algorithm 1 presents the pseudo-code of the two-phase GPU kernel for decompressing DFloat11 to BFloat16. + +Table 4: System specifications of servers used for experiments. + +
GPUGPU MemoryCPUCPU Memory
Server 1NVIDIA RTX A500024564MiBAMD EPYC 7513 32-Core504GB
Server 2NVIDIA A10040960MiBAMD EPYC 7742 64-Core1.48TB
Server 3NVIDIA Quadro RTX 800049152MiBAMD EPYC 7742 64-Core1.48TB
+ +Algorithm 1 GPU kernel for decompressing DFloat11 to BFloat16 +1: procedure DF11ToBF16 require: -EncodedExponent, PackedSignMantissa: byte arrays - $\mathrm{LUT}_1,\dots ,\mathrm{LUT}_k$ , CodeLengths: 8-bit unsigned integer arrays of size 256 - Gaps: 5-bit unsigned integer array (one entry per thread in each block) - BlockOutputPos: 32-bit unsigned integer array (one entry per block) - Outputs: BFloat16 array, for storing results - $B,T,n,k$ : the number of thread blocks, number of threads, number of bytes processed by each thread, number of compact LUTs, respectively +2: Divide EncodedExponent into chunks: EncodedExponent,..EncodedExponent $^b$ of size nT bytes each +3: for all $b\gets 1,\ldots ,B$ (in parallel across blocks) do +4: Load EncodedExponentbinto SRAM +5: Divide EncodedExponentbinto chunks: EncodedExponentb,1.,.., EncodedExponentb,Tof size n bytes each +6: Load $\mathrm{LUT}_{1},\dots ,\mathrm{LUT}_{k}$ ,CodeLengths into SRAM +7: Initialize integer arrays NumElements[1...T], ThreadOutputPos[1...T] with all 0s +8: Initialize BFloat16 write buffer WriteBuffer in SRAM +9: for all $t\gets 1,\dots ,T$ (in parallel across threads) do $\triangleright$ Phase 1: Each thread determines its initial output position +10: BitOffset $\leftarrow$ Gaps[bT+t] +11: while BitOffset $< 8n$ do +12: Read the next 4 bytes of EncodedExponentb,t, starting from the BitOffset-th bit, into + $\mathrm{Byte}_{1\dots 4}$ +13: $i\gets 1$ +14: Exponent $\leftarrow$ LUT[Byte] +15: while Exponent $\geq 240$ do $\triangleright$ Exponent $\geq 240$ means that it is a pointer to the next LUT +16: $i\gets i + 1$ +17: Exponent $\leftarrow$ LUT(257-Exponent)[Byte] +18: end while +19: BitOffset $\leftarrow$ BitOffset + CodeLengths[Exponent] +20: NumElements[t] $\leftarrow$ NumElements[t] + 1 +21: end while +22: Thread Synchronization Barrier $\triangleright$ Compute prefix-sum using Blelloch's Algorithm: +23: ThreadOutputPos[t] $\leftarrow$ BlockOutputPos[b] $+ \sum_{i = 1}^{t - 1}$ NumElements[i] $\triangleright$ Phase 2: Writing decoded BFloat16s to the appropriate positions +24: BitOffset $\leftarrow$ Gaps[bT+t] +25: while BitOffset $< 8n$ do +26: Read the next 4 bytes of EncodedExponentb,t, starting from the BitOffset-th bit, into + $\mathrm{Byte}_{1\dots 4}$ +27: $i\gets 1$ +28: Exponent $\leftarrow$ LUT[Byte] +29: while Exponent $\geq 240$ do $\triangleright$ Exponent $\geq 240$ means that it is a pointer to the next LUT +30: $i\gets i + 1$ +31: Exponent $\leftarrow$ LUT(257-Exponent)[Byte] +32: end while +33: Byte $\leftarrow$ PackedSignMantissa [ThreadOutputPos[t]] +34: Sign $\leftarrow$ Byte bitwise_and 0b10000000 +35: Mantissa $\leftarrow$ Byte bitwise_and 0b01111111 +36: WriteBuffer[ThreadOutputPos[t] - BlockOutputPos[b]] $\leftarrow$ (Sign bitwise_left SHIFT 8) bitwise_or (Exponent bitwise_leftSHIFT 7) bitwise_or Mantissa +37: BitOffset $\leftarrow$ BitOffset + CodeLengths[Exponent] +38: ThreadOutputPos[t] $\leftarrow$ ThreadOutputPos[t] + 1 +39: end while +40: end for $\triangleright$ Perform coalesced writes to HBM: +41: Outputs[BlockOutputPos[b] . . . (BlockOutputPos[b + 1] - 1)] $\leftarrow$ WriteBuffer[0... (BlockOutputPos[b + 1] - BlockOutputPos[b] - 1)] +42: end for +43: end procedure + +# E Hardware for Experiments + +Table 4 presents the hardware configuration of servers used for experiments. + +# F DFloat11 Compression Time + +Table 5: Compression time per transformer block for different models. + +
ModelCompression Time per Transformer Block (s)
Llama 3.1 8B Instruct191
Llama 3.3 70B Instruct547
Llama 3.1 405B Instruct2133
+ +Table 5 reports the time required to compress a single transformer block for models of different sizes. Compression is a one-time preprocessing step for each model and is performed using a single CPU thread. Since transformer blocks are independent in terms of weight storage, their compression can be parallelized across multiple CPU threads, making the overall process highly scalable and efficient. + +![](images/02460277e336179931ff9668c18dd40a4c47b75dd4cf375eaf28580ba5abe255.jpg) +Figure 10: Comparison of average latency and throughput for token decoding between the original (BF16) models and their losslessly compressed (DF11) counterparts. The BF16 and DF11 models are run on the same GPU configurations, with Flash Attention [7] turned on for both methods. + +# G GPU Inference Efficiency Comparison: BF16 vs. DF11 + +We present the GPU inference efficiency of BF16 and DF11 models in Figure 10, for various models and batch sizes on A100 GPUs. + +# H Impact of Lossy Quantization + +An accuracy comparison of the original and INT8-quantized Llama model is presented in table 6. + +Table 6: INT8 quantization error on different tasks. "Math" denotes MATH Hard with 2 shots. "GPQA CoT" is with 2 shots. "Δ" denotes the error gap via INT8 quantization. + +
ModelData TypeMathGPQA CoT
Llama-3.1-8B-InstructBF1623.9215.18
INT819.9214.06
Δ4.01.12
+ +![](images/cad7ec4cc5eeec4ef21599e9b8326267fef34c76f2df9ab6c7a2d88e5011d2c3.jpg) +Figure 11: Decoding Huffman codes can be performed either by traversing the Huffman tree or by using two lookup tables: one that maps each $L$ -bit binary code to its corresponding symbol, and another that stores the code length for each symbol. + +# I Efficient Decoding of Huffman Codes Using Compact Lookup Tables + +# I.1 The Dual Lookup Table Approach + +Huffman decoding can be performed by traversing the Huffman tree: starting from the root, each bit of the encoded bitstream determines the branch to follow, and the symbol is fully decoded upon reaching a leaf node. While this bit-by-bit traversal is conceptually simple, it is inefficient in practice. Each branching decision depends on the previous one, leading to frequent memory accesses and conditional jumps. This pattern is especially problematic on GPUs, where it causes branch divergence and limits instruction-level parallelism. A widely adopted alternative is lookup-table-based decoding [53], which flattens the Huffman tree into two compact lookup tables. This enables decoding of each symbol using just two array lookups and a bit shift, significantly improving throughput. + +We employ two lookup tables, LUT and CodeLengths, to achieve efficient, branch-free Huffman decoding. Let $L$ denote the length of the longest codeword in the Huffman codebook. We construct the primary lookup table LUT as an array of size $2^{L}$ , where each entry maps an $L$ -bit binary sequence to the first symbol it encodes. + +Figure 11 shows an example with $L = 4$ and a set of symbols A, B, C, D, E, F. For clarity, we use letters to represent symbols, though in practice these correspond to exponent values in BFloat16 weights. The lookup table LUT contains $2^4 = 16$ entries, indexed by all possible 4-bit binary sequences. Each entry in LUT stores the symbol whose Huffman code matches the prefix of that index. If a symbol's Huffman code is shorter than $L$ bits, it will fill multiple consecutive entries. For example, if symbol A is encoded as the single bit 0, then all binary sequences from 0000 to 0111 begin with 0, so entries 0 through 7 in LUT are assigned to A. In contrast, symbols with Huffman codes of length $L$ occupy exactly one entry each. For instance, $\mathrm{E} = 1110$ and $\mathrm{F} = 1111$ map to entries 14 and 15, respectively. This construction yields a dense prefix table that allows decoding a symbol with a single array lookup using an $L$ -bit segment from the encoded bitstream. + +To advance the encoded bitstream for decoding the next symbol, we also store the code lengths of all symbols. The second lookup table, CodeLengths, maps each symbol to its Huffman code length. In + +![](images/1833f862a0c1672d2beaec6af18323ff2c4aa08048210d5b8e709adf102e84c1.jpg) +Figure 12: A Huffman tree can be decomposed into a hierarchy of subtrees, each represented by a compact lookup table (LUT). Each LUT may reference another lower-level LUT in the hierarchy. This hierarchical decoding approach is functionally equivalent to using a single monolithic LUT, but significantly more memory efficient. + +the example, the lengths are: A:1, B:3, C:3, D:3, E:4, F:4. Together, these two tables allow fast, deterministic decoding by repeating the following steps: + +1. Use the next $L$ bits from the encoded bitstream to index LUT and retrieve the decoded symbol. +2. Look up the code length of the decoded symbol from CodeLengths to determine how many bits to consume. +3. Advance the encoded bitstream and repeat. + +This approach eliminates conditional branches and pointer chasing during decoding, making it highly suitable for parallel computation on GPUs. + +# I.2 Decomposing LUT into Hierarchical, Compact Lookup Tables + +The primary lookup table LUT contains $2^{L}$ entries, where $L$ is the maximum code length in the Huffman codebook. While this enables constant-time decoding, the table size grows exponentially with $L$ . In practice, $L$ ranges from 24 to 32 for Huffman trees built with BFloat16 exponents. This results in table sizes of $2^{24}$ to $2^{32}$ entries, which far exceeds the capacity of GPU SRAM. To address this, we decompose LUT into multiple smaller lookup tables that fit within on-chip memory, while still enabling fast decoding. + +Hierarchical Table Structure Instead of storing a single flat table of size $2^{L}$ , we decompose LUT into a hierarchy of compact lookup tables. Each table corresponds to a subtree of the Huffman tree and is responsible for decoding $b$ bits. Each table processes the next $b$ bits and either (i) directly returns a decoded symbol, or (ii) delegates to a table next in the hierarchy for decoding the next $b$ bits. This hierarchical organization mirrors the structure of the original Huffman tree and significantly reduces total memory usage. + +Figure 12 illustrates an example where the Huffman tree is partitioned into three subtrees, each mapped to a separate lookup table responsible for 2 bits. The decoding process using these three LUTs proceeds as follows: + +- $\mathrm{LUT}_0$ : Uses the first and second bits of the encoded bitstream to determine how to proceed, leading to 3 possible cases: + +- 00, $01 \rightarrow$ decode the next symbol as A. +- $10 \rightarrow$ delegate to $\mathsf{LUT}_1$ . + +- $11 \rightarrow$ delegate to $\mathsf{LUT}_2$ . + +- LUT1: Uses the third and fourth bits of the encoded bitstream to continue decoding: + +- 00, 01 → decode the next symbol as B +- 10, $11\to$ decode the next symbol as C + +- LUT $_2$ : Uses the third and fourth bits of the encoded bitstream to continue decoding: + +- 00, 01 → decode the next symbol as D +- $10 \rightarrow$ decode the next symbol as E +- $11 \rightarrow$ decode the next symbol as F + +For decoding Huffman-coded BFloat16 exponents, we decompose the LUT into multiple compact lookup tables, each responsible for decoding 8 bits (i.e. $b = 8$ ). This allows us to read the next byte from the encoded bitstream and perform an array lookup from a 256-entry array in each step. In practice, the decomposition of LUT leads to 4 to 8 compact LUTs, each with 256 entries, which comfortably fits within fast SRAM. + +# J Text-to-image Results of BF16 and DF11 Diffusion Models + +![](images/a128440ea9d181425e0c5fac19fa0c3da4342374a442de2ec28453822282f590.jpg) +Figure 13: Images generated by Stable Diffusion 3.5 Large in the original BFloat16 precision (top 5) are pixel-wise identical to those produced by the DFloat11-compressed model (bottom 5), using the same prompt and random seed. +Figure 13 presents the comparison of images generated using Stable Diffusion 3.5 Large in BFloat16 and DFloat11 weight format. The images are pixel-wise identical, when using the same prompt and random seed. + +# K Limitations + +This work focuses exclusively on losslessly compressing BFloat16 weights. We do not consider other formats such as FP32, FP16, or FP8, which may require different compression strategies. While DF11 improves memory efficiency, it introduces a small but non-zero latency overhead due to decompression. This overhead is amortized at larger batch sizes but may impact latency-sensitive applications with small batches. Our evaluation is limited to GPUs. We do not assess performance on other hardware such as CPUs, TPUs, or custom accelerators, which may require platform-specific optimizations. + +# NeurIPS Paper Checklist + +# 1. Claims + +Question: Do the main claims made in the abstract and introduction accurately reflect the paper's contributions and scope? + +Answer: [Yes] + +Justification: The main claims accurately reflect the paper's contributions and scope, as supported by evidences in the Method and Experiments section. + +Guidelines: + +- The answer NA means that the abstract and introduction do not include the claims made in the paper. +- The abstract and/or introduction should clearly state the claims made, including the contributions made in the paper and important assumptions and limitations. A No or NA answer to this question will not be perceived well by the reviewers. +- The claims made should match theoretical and experimental results, and reflect how much the results can be expected to generalize to other settings. +- It is fine to include aspirational goals as motivation as long as it is clear that these goals are not attained by the paper. + +# 2. Limitations + +Question: Does the paper discuss the limitations of the work performed by the authors? + +Answer: [Yes] + +Justification: We have included a Limitations section. + +Guidelines: + +- The answer NA means that the paper has no limitation while the answer No means that the paper has limitations, but those are not discussed in the paper. +- The authors are encouraged to create a separate "Limitations" section in their paper. +- The paper should point out any strong assumptions and how robust the results are to violations of these assumptions (e.g., independence assumptions, noiseless settings, model well-specification, asymptotic approximations only holding locally). The authors should reflect on how these assumptions might be violated in practice and what the implications would be. +- The authors should reflect on the scope of the claims made, e.g., if the approach was only tested on a few datasets or with a few runs. In general, empirical results often depend on implicit assumptions, which should be articulated. +- The authors should reflect on the factors that influence the performance of the approach. For example, a facial recognition algorithm may perform poorly when image resolution is low or images are taken in low lighting. Or a speech-to-text system might not be used reliably to provide closed captions for online lectures because it fails to handle technical jargon. +- The authors should discuss the computational efficiency of the proposed algorithms and how they scale with dataset size. +- If applicable, the authors should discuss possible limitations of their approach to address problems of privacy and fairness. +- While the authors might fear that complete honesty about limitations might be used by reviewers as grounds for rejection, a worse outcome might be that reviewers discover limitations that aren't acknowledged in the paper. The authors should use their best judgment and recognize that individual actions in favor of transparency play an important role in developing norms that preserve the integrity of the community. Reviewers will be specifically instructed to not penalize honesty concerning limitations. + +# 3. Theory assumptions and proofs + +Question: For each theoretical result, does the paper provide the full set of assumptions and a complete (and correct) proof? + +Answer: [NA] + +Justification: This paper does not include theoretical results. + +# Guidelines: + +- The answer NA means that the paper does not include theoretical results. +- All the theorems, formulas, and proofs in the paper should be numbered and cross-referenced. +- All assumptions should be clearly stated or referenced in the statement of any theorems. +- The proofs can either appear in the main paper or the supplemental material, but if they appear in the supplemental material, the authors are encouraged to provide a short proof sketch to provide intuition. +- Inversely, any informal proof provided in the core of the paper should be complemented by formal proofs provided in appendix or supplemental material. +- Theorems and Lemmas that the proof relies upon should be properly referenced. + +# 4. Experimental result reproducibility + +Question: Does the paper fully disclose all the information needed to reproduce the main experimental results of the paper to the extent that it affects the main claims and/or conclusions of the paper (regardless of whether the code and data are provided or not)? + +Answer: [Yes] + +Justification: We have fully disclosed all information needed for result reproduction. + +# Guidelines: + +- The answer NA means that the paper does not include experiments. +- If the paper includes experiments, a No answer to this question will not be perceived well by the reviewers: Making the paper reproducible is important, regardless of whether the code and data are provided or not. +- If the contribution is a dataset and/or model, the authors should describe the steps taken to make their results reproducible or verifiable. +- Depending on the contribution, reproducibility can be accomplished in various ways. For example, if the contribution is a novel architecture, describing the architecture fully might suffice, or if the contribution is a specific model and empirical evaluation, it may be necessary to either make it possible for others to replicate the model with the same dataset, or provide access to the model. In general, releasing code and data is often one good way to accomplish this, but reproducibility can also be provided via detailed instructions for how to replicate the results, access to a hosted model (e.g., in the case of a large language model), releasing of a model checkpoint, or other means that are appropriate to the research performed. +- While NeurIPS does not require releasing code, the conference does require all submissions to provide some reasonable avenue for reproducibility, which may depend on the nature of the contribution. For example +(a) If the contribution is primarily a new algorithm, the paper should make it clear how to reproduce that algorithm. +(b) If the contribution is primarily a new model architecture, the paper should describe the architecture clearly and fully. +(c) If the contribution is a new model (e.g., a large language model), then there should either be a way to access this model for reproducing the results or a way to reproduce the model (e.g., with an open-source dataset or instructions for how to construct the dataset). +(d) We recognize that reproducibility may be tricky in some cases, in which case authors are welcome to describe the particular way they provide for reproducibility. In the case of closed-source models, it may be that access to the model is limited in some way (e.g., to registered users), but it should be possible for other researchers to have some path to reproducing or verifying the results. + +# 5. Open access to data and code + +Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material? + +# Answer: [Yes] + +Justification: Our code and models are publicly available. + +# Guidelines: + +- The answer NA means that paper does not include experiments requiring code. +- Please see the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details. +- While we encourage the release of code and data, we understand that this might not be possible, so "No" is an acceptable answer. Papers cannot be rejected simply for not including code, unless this is central to the contribution (e.g., for a new open-source benchmark). +- The instructions should contain the exact command and environment needed to run to reproduce the results. See the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details. +- The authors should provide instructions on data access and preparation, including how to access the raw data, preprocessed data, intermediate data, and generated data, etc. +- The authors should provide scripts to reproduce all experimental results for the new proposed method and baselines. If only a subset of experiments are reproducible, they should state which ones are omitted from the script and why. +- At submission time, to preserve anonymity, the authors should release anonymized versions (if applicable). +- Providing as much information as possible in supplemental material (appended to the paper) is recommended, but including URLs to data and code is permitted. + +# 6. Experimental setting/details + +Question: Does the paper specify all the training and test details (e.g., data splits, hyperparameters, how they were chosen, type of optimizer, etc.) necessary to understand the results? + +# Answer: [Yes] + +Justification: This paper does not include training experiments. All evaluation details necessary to reproduce and understand the results are provided in the full paper. + +# Guidelines: + +- The answer NA means that the paper does not include experiments. +- The experimental setting should be presented in the core of the paper to a level of detail that is necessary to appreciate the results and make sense of them. +- The full details can be provided either with the code, in appendix, or as supplemental material. + +# 7. Experiment statistical significance + +Question: Does the paper report error bars suitably and correctly defined or other appropriate information about the statistical significance of the experiments? + +# Answer: [Yes] + +Justification: We include error bars for experiments where appropriate. + +# Guidelines: + +- The answer NA means that the paper does not include experiments. +- The authors should answer "Yes" if the results are accompanied by error bars, confidence intervals, or statistical significance tests, at least for the experiments that support the main claims of the paper. +- The factors of variability that the error bars are capturing should be clearly stated (for example, train/test split, initialization, random drawing of some parameter, or overall run with given experimental conditions). +- The method for calculating the error bars should be explained (closed form formula, call to a library function, bootstrap, etc.) +- The assumptions made should be given (e.g., Normally distributed errors). + +- It should be clear whether the error bar is the standard deviation or the standard error of the mean. +- It is OK to report 1-sigma error bars, but one should state it. The authors should preferably report a 2-sigma error bar than state that they have a $96\%$ CI, if the hypothesis of Normality of errors is not verified. +- For asymmetric distributions, the authors should be careful not to show in tables or figures symmetric error bars that would yield results that are out of range (e.g. negative error rates). +- If error bars are reported in tables or plots, The authors should explain in the text how they were calculated and reference the corresponding figures or tables in the text. + +# 8. Experiments compute resources + +Question: For each experiment, does the paper provide sufficient information on the computer resources (type of compute workers, memory, time of execution) needed to reproduce the experiments? + +Answer: [Yes] + +Justification: We have provided detailed information on the computational resources required to reproduce our results. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The paper should indicate the type of compute workers CPU or GPU, internal cluster, or cloud provider, including relevant memory and storage. +- The paper should provide the amount of compute required for each of the individual experimental runs as well as estimate the total compute. +- The paper should disclose whether the full research project required more compute than the experiments reported in the paper (e.g., preliminary or failed experiments that didn't make it into the paper). + +# 9. Code of ethics + +Question: Does the research conducted in the paper conform, in every respect, with the NeurIPS Code of Ethics https://neurips.cc/public/EthicsGuidelines? + +Answer: [Yes] + +Justification: After reviewing the NeurIPS Code of Ethics, we have verified that our research is in compliance. + +Guidelines: + +- The answer NA means that the authors have not reviewed the NeurIPS Code of Ethics. +- If the authors answer No, they should explain the special circumstances that require a deviation from the Code of Ethics. +- The authors should make sure to preserve anonymity (e.g., if there is a special consideration due to laws or regulations in their jurisdiction). + +# 10. Broader impacts + +Question: Does the paper discuss both potential positive societal impacts and negative societal impacts of the work performed? + +Answer: [NA] + +Justification: This work focuses on improving the inference memory efficiency of existing pre-trained models and does not involve the development of new model, training data, or training method, which limits its direct societal impact. + +Guidelines: + +- The answer NA means that there is no societal impact of the work performed. +- If the authors answer NA or No, they should explain why their work has no societal impact or why the paper does not address societal impact. + +- Examples of negative societal impacts include potential malicious or unintended uses (e.g., disinformation, generating fake profiles, surveillance), fairness considerations (e.g., deployment of technologies that could make decisions that unfairly impact specific groups), privacy considerations, and security considerations. +- The conference expects that many papers will be foundational research and not tied to particular applications, let alone deployments. However, if there is a direct path to any negative applications, the authors should point it out. For example, it is legitimate to point out that an improvement in the quality of generative models could be used to generate deepfakes for disinformation. On the other hand, it is not needed to point out that a generic algorithm for optimizing neural networks could enable people to train models that generate Deepfakes faster. +- The authors should consider possible harms that could arise when the technology is being used as intended and functioning correctly, harms that could arise when the technology is being used as intended but gives incorrect results, and harms following from (intentional or unintentional) misuse of the technology. +- If there are negative societal impacts, the authors could also discuss possible mitigation strategies (e.g., gated release of models, providing defenses in addition to attacks, mechanisms for monitoring misuse, mechanisms to monitor how a system learns from feedback over time, improving the efficiency and accessibility of ML). + +# 11. Safeguards + +Question: Does the paper describe safeguards that have been put in place for responsible release of data or models that have a high risk for misuse (e.g., pretrained language models, image generators, or scraped datasets)? + +Answer: [NA] + +Justification: We do not introduce new data or models. + +# Guidelines: + +- The answer NA means that the paper poses no such risks. +- Released models that have a high risk for misuse or dual-use should be released with necessary safeguards to allow for controlled use of the model, for example by requiring that users adhere to usage guidelines or restrictions to access the model or implementing safety filters. +- Datasets that have been scraped from the Internet could pose safety risks. The authors should describe how they avoided releasing unsafe images. +- We recognize that providing effective safeguards is challenging, and many papers do not require this, but we encourage authors to take this into account and make a best faith effort. + +# 12. Licenses for existing assets + +Question: Are the creators or original owners of assets (e.g., code, data, models), used in the paper, properly credited and are the license and terms of use explicitly mentioned and properly respected? + +Answer: [Yes] + +Justification: The assets used are properly credited. + +# Guidelines: + +- The answer NA means that the paper does not use existing assets. +- The authors should cite the original paper that produced the code package or dataset. +- The authors should state which version of the asset is used and, if possible, include a URI. +- The name of the license (e.g., CC-BY 4.0) should be included for each asset. +- For scraped data from a particular source (e.g., website), the copyright and terms of service of that source should be provided. +- If assets are released, the license, copyright information, and terms of use in the package should be provided. For popular datasets, paperswithcode.com/datasets has curated licenses for some datasets. Their licensing guide can help determine the license of a dataset. + +- For existing datasets that are re-packaged, both the original license and the license of the derived asset (if it has changed) should be provided. +- If this information is not available online, the authors are encouraged to reach out to the asset's creators. + +# 13. New assets + +Question: Are new assets introduced in the paper well documented and is the documentation provided alongside the assets? + +Answer: [NA] + +Justification: No new assets are introduced by this paper. + +Guidelines: + +- The answer NA means that the paper does not release new assets. +- Researchers should communicate the details of the dataset/code/model as part of their submissions via structured templates. This includes details about training, license, limitations, etc. +- The paper should discuss whether and how consent was obtained from people whose asset is used. +- At submission time, remember to anonymize your assets (if applicable). You can either create an anonymized URL or include an anonymized zip file. + +# 14. Crowdsourcing and research with human subjects + +Question: For crowdsourcing experiments and research with human subjects, does the paper include the full text of instructions given to participants and screenshots, if applicable, as well as details about compensation (if any)? + +Answer: [NA] + +Justification: This paper does not involve crowdsourcing or research with human subjects. + +Guidelines: + +- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. +- Including this information in the supplemental material is fine, but if the main contribution of the paper involves human subjects, then as much detail as possible should be included in the main paper. +- According to the NeurIPS Code of Ethics, workers involved in data collection, curation, or other labor should be paid at least the minimum wage in the country of the data collector. + +# 15. Institutional review board (IRB) approvals or equivalent for research with human subjects + +Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals (or an equivalent approval/review based on the requirements of your country or institution) were obtained? + +Answer: [NA] + +Justification: This paper does not involve crowdsourcing or research with human subjects. + +Guidelines: + +- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. +- Depending on the country in which research is conducted, IRB approval (or equivalent) may be required for any human subjects research. If you obtained IRB approval, you should clearly state this in the paper. +- We recognize that the procedures for this may vary significantly between institutions and locations, and we expect authors to adhere to the NeurIPS Code of Ethics and the guidelines for their institution. +- For initial submissions, do not include any information that would break anonymity (if applicable), such as the institution conducting the review. + +# 16. Declaration of LLM usage + +Question: Does the paper describe the usage of LLMs if it is an important, original, or non-standard component of the core methods in this research? Note that if the LLM is used only for writing, editing, or formatting purposes and does not impact the core methodology, scientific rigourousness, or originality of the research, declaration is not required. + +Answer: [NA] + +Justification: LLMs are not involved in any origin or important parts of this research. + +Guidelines: + +- The answer NA means that the core method development in this research does not involve LLMs as any important, original, or non-standard components. +- Please refer to our LLM policy (https://neurips.cc/Conferences/2025/LLM) for what should or should not be described. \ No newline at end of file diff --git a/NeurIPS/2025/70% Size, 100% Accuracy_ Lossless LLM Compression for Efficient GPU Inference via Dynamic-Length Float (DFloat11)/images.zip b/NeurIPS/2025/70% Size, 100% Accuracy_ Lossless LLM Compression for Efficient GPU Inference via Dynamic-Length Float (DFloat11)/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..1db531ef44d6c7eb30f44e36a4fd91b686d85c9b --- /dev/null +++ b/NeurIPS/2025/70% Size, 100% Accuracy_ Lossless LLM Compression for Efficient GPU Inference via Dynamic-Length Float (DFloat11)/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:436b623c00b595af5ccdc340109b61a41440187bc207a1d3e12aec1e1ba6b93d +size 1063971 diff --git a/NeurIPS/2025/70% Size, 100% Accuracy_ Lossless LLM Compression for Efficient GPU Inference via Dynamic-Length Float (DFloat11)/layout.json b/NeurIPS/2025/70% Size, 100% Accuracy_ Lossless LLM Compression for Efficient GPU Inference via Dynamic-Length Float (DFloat11)/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..5b8b391273bd44f632c7eec41a843fb109f3bbab --- /dev/null +++ b/NeurIPS/2025/70% Size, 100% Accuracy_ Lossless LLM Compression for Efficient GPU Inference via Dynamic-Length Float (DFloat11)/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e4e4922aa6d3d12593e70ea3371a8802c1de7d6d3eb98ce5c2edba6db6f9f6e3 +size 851380 diff --git a/NeurIPS/2025/A Bayesian Approach to Contextual Dynamic Pricing using the Proportional Hazards Model with Discrete Price Data/06160620-923e-4420-9464-79a186a15f51_content_list.json b/NeurIPS/2025/A Bayesian Approach to Contextual Dynamic Pricing using the Proportional Hazards Model with Discrete Price Data/06160620-923e-4420-9464-79a186a15f51_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..561d598083bb47ac9809359398b5961af1ff9881 --- /dev/null +++ b/NeurIPS/2025/A Bayesian Approach to Contextual Dynamic Pricing using the Proportional Hazards Model with Discrete Price Data/06160620-923e-4420-9464-79a186a15f51_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2ef709f4ef2ff5b808fb1a96edf27f5e31068096c167e1286ccf8596fab98286 +size 478850 diff --git a/NeurIPS/2025/A Bayesian Approach to Contextual Dynamic Pricing using the Proportional Hazards Model with Discrete Price Data/06160620-923e-4420-9464-79a186a15f51_model.json b/NeurIPS/2025/A Bayesian Approach to Contextual Dynamic Pricing using the Proportional Hazards Model with Discrete Price Data/06160620-923e-4420-9464-79a186a15f51_model.json new file mode 100644 index 0000000000000000000000000000000000000000..5fe7ff208bcca062181d6dd90c32ad0e81f7ab8e --- /dev/null +++ b/NeurIPS/2025/A Bayesian Approach to Contextual Dynamic Pricing using the Proportional Hazards Model with Discrete Price Data/06160620-923e-4420-9464-79a186a15f51_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5c309bfeb39254bfe87d42539b93ab2662e41a0e2a3b922ecdcb9f5b7297ac71 +size 550324 diff --git a/NeurIPS/2025/A Bayesian Approach to Contextual Dynamic Pricing using the Proportional Hazards Model with Discrete Price Data/06160620-923e-4420-9464-79a186a15f51_origin.pdf b/NeurIPS/2025/A Bayesian Approach to Contextual Dynamic Pricing using the Proportional Hazards Model with Discrete Price Data/06160620-923e-4420-9464-79a186a15f51_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..36afe67e0f082df35e489daef7318d92f357a018 --- /dev/null +++ b/NeurIPS/2025/A Bayesian Approach to Contextual Dynamic Pricing using the Proportional Hazards Model with Discrete Price Data/06160620-923e-4420-9464-79a186a15f51_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9c072307b2d607849a3b731b4a2dd47bd0324e658f1bd2080993cecc0f00e34e +size 1584953 diff --git a/NeurIPS/2025/A Bayesian Approach to Contextual Dynamic Pricing using the Proportional Hazards Model with Discrete Price Data/full.md b/NeurIPS/2025/A Bayesian Approach to Contextual Dynamic Pricing using the Proportional Hazards Model with Discrete Price Data/full.md new file mode 100644 index 0000000000000000000000000000000000000000..c16b561a8c0dba86d50cb6060743764a0c3e4868 --- /dev/null +++ b/NeurIPS/2025/A Bayesian Approach to Contextual Dynamic Pricing using the Proportional Hazards Model with Discrete Price Data/full.md @@ -0,0 +1,2553 @@ +# A Bayesian Approach to Contextual Dynamic Pricing using the Proportional Hazards Model with Discrete Price Data + +Dongguen Kim1, Young-Geun Choi2, Minwoo Chae1* + +$^{1}$ Department of Industrial and Management Engineering, + +Pohang University of Science and Technology + +$^{2}$ Department of Mathematics Education, Sungkyunkwan University + +{dgkim, mchae}@postech.ac.kr; ygchoi@skku.edu + +# Abstract + +Dynamic pricing algorithms typically assume continuous price variables, which may not reflect real-world scenarios where prices are often discrete. This paper demonstrates that leveraging discrete price information within a semi-parametric model can substantially improve performance, depending on the size of the support set of the price variable relative to the time horizon. Specifically, we propose a novel semi-parametric contextual dynamic pricing algorithm, namely BayesCoxCP, based on a Bayesian approach to the Cox proportional hazards model. Our theoretical analysis establishes high-probability regret bounds that adapt to the sparsity level $\gamma$ , proving that our algorithm achieves a regret upper bound of $\widetilde{O}(T^{(1+\gamma)/2} + \sqrt{dT})$ for $\gamma < 1/3$ and $\widetilde{O}(T^{2/3} + \sqrt{dT})$ for $\gamma \geq 1/3$ , where $\gamma$ represents the sparsity of the price grid relative to the time horizon $T$ . Through numerical experiments, we demonstrate that our proposed algorithm significantly outperforms an existing method, particularly in scenarios with sparse discrete price points. + +# 1 Introduction + +Contextual dynamic pricing involves updating product prices over time based on contextual information such as customer features, product attributes, and market conditions. Given its importance and practical applications in revenue management, this topic has been extensively explored across statistics, machine learning, and operations research [9, 43, 34]. The primary objective of contextual dynamic pricing is to maximize the seller's revenue through determining optimal prices that account for both covariates and demand uncertainty. A key challenge in dynamic pricing is balancing exploration, which focuses on learning the underlying demand, with exploitation, which leverages current knowledge to set optimal prices. Striking this balance is essential for developing effective dynamic pricing strategies. + +A commonly studied framework in contextual dynamic pricing is the binary choice model, where the seller receives binary purchase feedback based on the posted prices [2, 24, 37, 44, 3, 30, 7, 10, 31]. Specifically, at each time $t = 1,\dots ,T$ , the seller observes a covariate $X_{t}\in \mathbb{R}^{d}$ that captures customer and product features. Based on the observed covariate and historical sales data, the seller determines a price $P_{t}$ for the product. The customer's valuation of the product, denoted as a random variable $V_{t}\in \mathbb{R}_{\geq 0}$ , is unknown to the seller. Following the posted price, the seller receives binary feedback $Y_{t}\in \{\bar{0},1\}$ , indicating whether a purchase occurred. The customer purchases the product if and only if their valuation $V_{t}$ exceeds the offered price $P_{t}$ , which can be expressed as $Y_{t} = \mathbb{1}\{V_{t} > P_{t}\}$ . Notably, $V_{t}$ is not directly observed, as it is censored by $P_{t}$ . In the statistical literature, such data + +structures are called case 1 interval-censored data, also known as current status data [18]. Case 1 interval-censored data has been extensively studied in survival analysis [11, 13, 28, 38, 25, 20, 19]. + +We consider a contextual pricing problem under the binary choice model. Let $F(v \mid X_{t}) = \mathbb{P}(V_{t} \leq v \mid X_{t})$ and $S(v \mid X_{t}) = 1 - F(v \mid X_{t})$ be the cumulative distribution function (c.d.f.) and complementary c.d.f. (or survival function) of $V_{t}$ given $X_{t}$ , respectively. The expected revenue from a posted price $p$ given the covariate $X_{t}$ is given as $\mathbb{E}(p \cdot Y_{t} \mid X_{t}) = p\mathbb{P}(V_{t} > p \mid X_{t}) = pS(p \mid X_{t})$ . Then an optimal price $P_{t}^{*}$ at time $t$ is defined as a price that maximizes the expected revenue: + +$$ +P _ {t} ^ {*} \in \underset {p} {\operatorname {a r g m a x}} p S (p \mid X _ {t}). \tag {1} +$$ + +The regret at time $t$ is the difference between the expected revenue generated by the optimal price $P_{t}^{*}$ and that from the posted price $P_{t}$ , given by $r(t) = P_{t}^{*}S(P_{t}^{*}\mid X_{t}) - P_{t}S(P_{t}\mid X_{t})$ . An important objective is to design a pricing policy that minimizes the cumulative regret over a given time horizon $T$ , defined as $R(T) = \sum_{t = 1}^{T}r(t)$ . + +As shown in (1), designing an effective pricing policy necessitates accurately estimating the complementary c.d.f. $S(\cdot \mid X_{t})$ . Thus, a wide range of contextual dynamic pricing algorithms have been developed using various models for the conditional distribution of $V_{t}$ given $X_{t}$ . Linear models [2, 36, 23, 17, 24, 3, 44, 6, 30, 33, 10, 31], where $F(v \mid X_{t}) = F_{0}(v - X_{t}^{\top}\beta)$ , and log-linear models [37], where $F(v \mid X_{t}) = F_{0}(v \cdot \exp (-X_{t}^{\top}\beta))$ , serve as key examples. Here, $F_{0}(v) = \mathbb{P}(V_{t} \leq v \mid X_{t} = 0)$ represents the baseline c.d.f., and $\beta \in \mathbb{R}^{d}$ captures the contextual effect. More recently, [7] proposed using the Cox proportional hazards (PH) model, in which the complementary c.d.f. is modeled as $S(v \mid X_{t}) = S_{0}(v)^{\exp (X_{t}^{\top}\beta)}$ . Here, $S_{0}(v) = 1 - F_{0}(v)$ represents the baseline complementary c.d.f. In particular, semi-parametric models, which assume that both the nonparametric baseline function $F_{0}$ (or $S_{0}$ ) and the parametric coefficient $\beta$ are unknown, have gained considerable attention recently due to their flexibility and interpretability [37, 30, 7, 10, 31]. + +In real-world applications, it is crucial to note that offered prices are often observed only on a discrete set. For instance, retailers commonly restrict prices to convenient values for ease of communication and consumer familiarity, and businesses often adhere to predefined discount levels or promotional price points [39, 22]. The significance of discrete price sets in revenue management has been widely recognized [12, 4, 32]. Much of the existing dynamic pricing literature, however, focuses on continuous price spaces [2, 36, 23, 17, 24, 37, 3, 44, 33, 6, 30, 7, 10, 31]. + +From a theoretical perspective, discrete price sets provide significant advantages for estimating model parameters. In a simple survival analysis setup with i.i.d. case 1 interval-censored observations, [41] demonstrated that the inferential performance of the underlying survival function can be improved by leveraging the fact that the monitoring time (the offered price in the dynamic pricing setting) is discretely supported. To be specific, if the monitoring time is continuous, the optimal convergence rate for estimating the unknown survival function with case 1 interval-censored data is known as $n^{-1/3}$ , where $n$ is the sample size [21]. On the other hand, in [41], the monitoring times are assumed to be supported on an equally spaced grid set. Then, they proved that the nonparametric maximum likelihood estimator (NPMLE) achieves a convergence rate of $n^{-(1 - \gamma)/2}$ for $\gamma < 1/3$ and $n^{-1/3}$ for $\gamma \geq 1/3$ , where $\gamma \in (0,1]$ represents the sparsity level of the grid relative to the sample size $n$ (a rigorous definition is provided in Section 2). In other words, one can achieve much faster convergence rates if the grid is sparse ( $\gamma < 1/3$ ). Moreover, they developed an inferential procedure, such as the construction of confidence intervals, that does not depend on the unknown quantity $\gamma$ , often referred to as an adaptive procedure. While the adaptive procedure in [41] is quite complicated, [5] demonstrated that a much simpler and more practical Bayes procedure is also adaptive, and that the corresponding Bayes estimator achieves the same convergence rate. Although these aforementioned theoretical results are based on a non-contextual setup and i.i.d. data, they suggest that incorporating the discrete support of the price may lead to a pricing policy with smaller cumulative regret compared to one that ignores this information. + +Motivated by these insights, we propose a novel semi-parametric contextual dynamic pricing algorithm, BayesCoxCP, based on a Bayesian approach to the Cox PH model with case 1 interval-censored data. The algorithm is specifically designed to exploit the discreteness of the offered price, leading to improved performance. Our theoretical contributions are threefold: + +- We derive the posterior convergence rate of the Bayes estimators of the semi-parametric Cox PH model under the i.i.d. setup. We assume that the offered price is supported on an equally + +Table 1: Existing regret bounds for contextual dynamic pricing algorithms based on semi-parametric models. Note that the optimal rates depend on the model and assumptions, such as the smoothness of $F_0$ . + +
METHODMODEL FOR VtREGRET UPPER BOUNDOPTIMALITY IN TADAPTATION TO DISCRETE SUPPORT
[10]LINEARO((Td)2m+1/4m-1)XX
[31]LINEARO(T2/3 + ||β - β*||1T)XX
[30]LINEARO(T2/3d2)XX
[30]LINEARO(T3/4d)XX
[37]LOG-LINEARO(T1/2d11/4)XX
[7]PHO(T2/3d)○X
OUR WORKPHO(T1+γ/2 + √dT) (γ<1/3)○○
O(T2/3 + √dT) (γ≥1/3)
+ +spaced grid set and prove that the posterior distribution converges at the optimal rate, which adapts to the grid sparsity. This result generalizes the work of [5], who studied the survival model without covariates, to the PH model. It is also worth noting that our prior for the baseline cumulative hazard differs from that of [5] and can achieve computational benefits. + +- We derive the regret upper bound of the proposed BayesCoxCP algorithm. Specifically, our algorithm achieves a regret upper bound of order $T^{\frac{1 + \gamma}{2}} + (dT)^{1/2}$ for $\gamma < 1/3$ and $T^{2/3} + (dT)^{1/2}$ for $\gamma \geq 1/3$ , up to a logarithmic factor, where $\gamma$ represents the grid sparsity relative to the time horizon $T$ . Notably, the BayesCoxCP algorithm does not rely on the value of $\gamma$ , i.e., our algorithm adapts to the sparsity level. A careful selection of the exploration parameter $\eta_l$ is crucial in the algorithm's design; see Section 4 for further details. +- We also establish a non-contextual minimax lower bound for the cumulative regret in the discrete pricing problem, as stated in Theorem 5.3. It turns out that our regret upper bound for the BayesCoxCP algorithm is optimal up to a logarithmic factor in terms of $T$ . + +Through extensive numerical experiments, we empirically demonstrate that the proposed pricing algorithm significantly outperforms the state-of-the-art method in [7] when prices are discretely supported. + +The remainder of this paper is organized as follows. In the following subsections, we introduce the notations used throughout the paper and provide a brief summary of related works. Section 2 introduces the basic setup for case 1 interval-censored data on a grid and describes the Cox PH model, along with the prior distributions employed. Section 3 establishes the convergence rate of the posterior distribution under the i.i.d. setup. Section 4 introduces the BayesCoxCP algorithm and Section 5 presents its regret analysis. Finally, Section 6 presents numerical experiments to evaluate the effectiveness of our proposed algorithm. + +# 1.1 Notation + +For two real numbers $a$ and $b$ , $a \vee b$ and $a \wedge b$ denote the maximum and minimum of $a$ and $b$ , respectively. For two densities $p$ and $q$ with dominating measure $\nu$ , let $\mathcal{D}_H(p, q) = (\int (p^{1/2} - q^{1/2})^2 d\nu)^{1/2}$ be the Hellinger distance and $K(p, q) = \int \log(p/q)p d\nu$ be Kullback-Leibler divergence. For a metric space $(\mathcal{F}, \mathcal{D})$ , the $\epsilon$ -covering and $\epsilon$ -bracketing numbers of $\mathcal{F}$ with respect to distance $\mathcal{D}$ are denoted as $N(\epsilon, \mathcal{F}, \mathcal{D})$ and $N_{[]}(\epsilon, \mathcal{F}, \mathcal{D})$ , respectively. We write $a = O(b)$ or $a \lesssim b$ if $a \leq Cb$ for some constant $C > 0$ , where $C$ is an absolute constant unless otherwise specified. In addition, we write $a = \Omega(b)$ or $a \gtrsim b$ if $a \geq Cb$ for some constant $C > 0$ . The notation $\widetilde{O}(\cdot)$ denotes the corresponding bound that ignore logarithmic factors. + +# 1.2 Related works + +The problem of contextual dynamic pricing has been extensively studied in the literature. Many recent works have focused on semi-parametric models where $F_{0}$ is unknown and nonparametric. For instance, [30, 31, 10] considered linear models with an unknown $F_{0}$ under certain smoothness + +assumptions. In [10], $F_{0}$ is assumed to be $m(\geq 2)$ th-order smooth, achieving a regret upper bound of $\widetilde{O}((Td)^{\frac{2m + 1}{4m - 1}})$ . [31] relaxed this assumption by assuming that $F_{0}$ is Lipschitz continuous and second-order smooth. They obtained a regret upper bound of $\widetilde{O}(T^{2/3} + \| \widehat{\beta} - \beta^{*} \|_1 T)$ , where $\| \widehat{\beta} - \beta^{*} \|_1$ represents the estimation error of $\beta^{*}$ . Similarly, [30] considered the same setting and achieved a regret upper bound of $\widetilde{O}(T^{2/3} d^2)$ under the Lipschitz and second-order smoothness assumptions on $F_{0}$ , while showing that under a weaker Lipschitz assumption alone, the regret upper bound increases to $\widetilde{O}(T^{3/4} d)$ . On the other hand, [37] used a log-linear model with a second-order smoothness assumption on $F_{0}$ , achieving a regret upper bound of $\widetilde{O}(T^{1/2} d^{11/4})$ but with suboptimal dependency on the dimension $d$ . Similar to our approach, [7] used the Cox PH model, assuming that $F_{0}$ is Lipschitz continuous. They derived a regret upper bound of $\widetilde{O}(T^{2/3} d)$ , which improves the dimensional dependency compared to [37], but their analysis is limited to continuous pricing settings. The overall comparison of regret bounds from these semi-parametric studies, along with our results, is summarized in Table 1. In addition to these works, earlier studies often assumed that $F_{0}$ is known and log-concave. For instance, [24] and [44] both considered linear models under these assumptions. [24] additionally analyzed the case where $F_{0}$ is unknown but belongs to a parametric log-concave family, deriving a regret upper bound of order $T^{1/2}$ . + +# 2 Preliminaries + +# 2.1 Basic setup + +In the current and next sections, we study the behavior of the posterior distribution from the PH model for analyzing case 1 interval-censored data on a grid under the i.i.d. regime. + +To set the scene, suppose that $(X_{t},P_{t},Y_{t}), t = 1,\dots,n$ , are i.i.d. copies of $(X,P,Y)$ . In particular, we assume that $P_{t}$ 's are supported on the grid set $\mathcal{G} = \{g_1,\ldots ,g_K\}$ within the (fixed) interval $[p_{\min},p_{\max}]$ , whose cardinality may depend on the sample size $n$ . The grid points are assumed to be uniform in the sense that $g_{k + 1} = g_k + \delta$ for every $k\geq 0$ , where $g_0 = p_{\mathrm{min}}$ , and $K$ is the largest integer such that $g_{K}\leq p_{\mathrm{max}}$ , that is, $K = \lfloor (p_{\mathrm{max}} - p_{\mathrm{min}}) / \delta \rfloor$ . We further assume that the grid resolution $\delta$ is controlled by two constants $\gamma \in (0,1]$ and $\kappa >0$ , according to the relation $\delta = \kappa n^{-\gamma}$ . Note that generalizations to nonuniform grids are discussed in Appendix D. + +Let $\mathbb{Q}(\cdot \mid X)$ denote the conditional distribution of $P$ given $X$ , with $q(\cdot \mid X)$ denoting the corresponding probability mass function. In addition, the marginal distribution of the price $P$ is denoted by $\mathbb{Q}(\cdot)$ , with its probability mass function given by $q(\cdot)$ . Let $\mathbb{P}_X$ and $p_X$ be the marginal distribution and the corresponding density of $X$ . + +# 2.2 Proportional hazards model for $V_{t}$ + +We consider the Cox PH model for the conditional distribution of $V_{t}$ given $X_{t}$ . Formally, the complementary c.d.f. $v \mapsto S(v \mid X_t)$ of $V_{t}$ is modeled as + +$$ +S (v \mid X _ {t}) = S _ {0} (v) ^ {\exp (X _ {t} ^ {\top} \beta)}, +$$ + +where $S_0(\cdot)$ is a baseline complementary c.d.f., and $\beta \in \mathbb{R}^d$ is a regression coefficient. We assume that $V_t$ is continuous. Let $F_0 = 1 - S_0$ , $\lambda_0 = F_0' / S_0$ and $\Lambda_0(v) = \int_0^v \lambda_0(u) du$ be the c.d.f., hazard and cumulative hazard functions, respectively, corresponding to $S_0$ , where $F_0'$ denotes the derivative of $F_0$ . + +We remark that the joint distribution of $(X_{t},P_{t},Y_{t})$ depends on the unknown parameters $(S_0,\beta ,p_X,q(\cdot |\cdot))$ . Among these, While $(S_0,\beta)$ are the parameters of interest, while $(p_X,q(\cdot |\cdot))$ are treated as nuisance parameters (at least in the current and next sections). Since $P_{t}$ is supported on $\mathcal{G}$ , the joint distribution of $(X_{t},P_{t},Y_{t})$ depends on $S_0$ only through the vector $\mathbf{S}_0 = (S_{0,1},\ldots ,S_{0,K})\in [0,1]^K$ , where $S_{0,k} = S_0(g_k)$ . Here, $\mathcal{X}$ represents the support of the covariate $X$ . The parameter space is defined as $\Theta = \{\theta = (\mathbf{S}_0,\beta)\in S_0\times \mathbb{R}^d\}$ , where $S_0 = \{\mathbf{S}_0 = (S_{0,1},\dots,S_{0,K}):1 > S_{0,1}\geq \dots \geq S_{0,K} > 0\}$ . + +# 2.3 Prior + +We note that $V_{t}$ is continuous while $P_{t}$ is discrete with support $\mathcal{G}$ . To reflect this structure, we model the baseline hazard function $\lambda_0(\cdot)$ as a left-continuous step function, where the jump points are located at grid points. Let $\pmb{\lambda}_0 = (\lambda_{0,1},\dots,\lambda_{0,K})$ , $\Lambda_0 = (\Lambda_{0,1},\dots,\Lambda_{0,K})$ and + +$$ +\lambda_ {0} (p) = \sum_ {k = 1} ^ {K} \lambda_ {0, k} \mathbb {1} \left\{p \in \left(g _ {k - 1}, g _ {k} \right] \right\}, \tag {2} +$$ + +where $\Lambda_{0,k} = -\log S_{0,k}$ . Since there is a one-to-one correspondence between $\mathbf{S}_0$ and $\lambda_0$ , one can impose a prior on $\mathbf{S}_0$ through $\lambda_0$ . We consider an independent prior for the unknown parameters $\lambda_0$ and $\beta$ , specified as $\Pi = \Pi_{\beta} \times \Pi_{\lambda_0}$ . Here, $\Pi_{\lambda_0}$ consists of independent gammas: + +$$ +\lambda_ {0, k} \sim \operatorname {G a m m a} \left(\alpha_ {k}, \rho\right), \quad k = 1, \dots , K, \tag {3} +$$ + +where $\mathrm{Gamma}(\alpha_k,\rho)$ denotes the gamma distribution with mean $\alpha_{k} / \rho$ and variance $\alpha_{k} / \rho^{2}$ . Gamma priors are commonly employed for $\lambda_0$ in Bayesian analyses of the PH model, as seen in [27, 47, 35, 29]. We further impose the following conditions on the prior: + +(P1) $\Pi_{\beta}$ has a continuous and positive Lebesgue density on $\mathbb{R}^d$ +(P2) There exist positive constants $\underline{\alpha} < \overline{\alpha}$ , such that $\underline{\alpha} \leq \alpha_k \leq \overline{\alpha}$ for $k = 1, \dots, K$ . + +# 3 Posterior convergence rate under i.i.d. setup + +To clarify notation, we use the superscript $*$ to denote the true parameter, e.g., $\Lambda_0^*$ , $\beta^*$ , and $\lambda_0^*$ . Suppose that there exists a true parameter $\theta^{*} = (\mathbf{S}_{0}^{*},\beta^{*})$ generating the data $\mathbf{D}_n = \{(X_t,P_t,Y_t)\}_{t = 1}^n$ . (We may regard $p_X$ and $q(\cdot |\cdot)$ as known parameters if we are only interested in inferring $\theta$ .) Given the data $\mathbf{D}_n$ , let $\Pi (\cdot \mid \mathbf{D}_n)$ be the joint posterior distribution of $\lambda_0$ and $\beta$ . + +Assumptions We will prove that $\Pi (\cdot \mid \mathbf{D}_n)$ concentrates around $\theta^{*}$ under the following assumptions: + +(A1) $\| \beta^{*}\|_{2}\leq B$ for some constant $B > 0$ +(A2) $\mathbb{P}_X(\mathcal{X}) = 1$ and $p_X$ is bounded away from zero on $\mathcal{X}$ , where $\mathcal{X} = \{x\in \mathbb{R}^d:\| x\| _2\leq L\}$ . +(A3) $\mathbb{P}(X^{\top}\beta_{1}\neq X^{\top}\beta_{2}) > 0$ for $\beta_{1}\neq \beta_{2}$ +(A4) For $x \in \mathcal{X}$ and $1 \leq k \leq K$ , suppose $\mathbb{Q}(\mathcal{G} \mid X = x) = 1$ , and $q(g_k \mid x) \gtrsim n^{-\frac{1 + \gamma}{2}} (\log n)^{\frac{1}{2}}$ if $\gamma < 1/3$ , or $q(g_k \mid x) \gtrsim n^{-\gamma - \frac{1}{3}} (\log n)^{\frac{1}{2}}$ otherwise. +(A5) The support of $F_0^*$ is $[v_{\min}, v_{\max}]$ , $v_{\min} < p_{\min} < p_{\max} < v_{\max}$ , and $S_0^*$ has a continuous and strictly negative derivative on $[v_{\min}, v_{\max}]$ . + +Assumptions (A1) and (A2) are commonly adopted in the stochastic contextual dynamic pricing literature. Assumption (A3) ensures the identifiability of the regression coefficient. Assumption (A4) requires that the conditional distribution $\mathbb{Q}(\cdot \mid x)$ maintains a certain level of uniformity over the grid set $\mathcal{G}$ . For instance, when $\mathbb{Q}(\cdot \mid x)$ follows a uniform distribution over $\mathcal{G}$ , (A4) is satisfied for any $\gamma \in [0,1]$ . In the contextual dynamic pricing problem, the function $q(\cdot \mid x)$ is parameterized by the pricing policy. Therefore, constructing a policy that satisfies (A4) is crucial. In Section 4, we explicitly design a policy that fulfills (A4). Assumption (A5) implies that $S_0^*$ is $L_0$ -Lipschitz on $[p_{\min}, p_{\max}]$ for some constant $L_0 > 0$ , and that $S_{0,1}$ and $S_{0,K}$ are bounded away from 0 and 1. + +We define the distance $\mathcal{D}_{\mathbb{Q}}$ on the parameter space $\Theta$ as + +$$ +\mathcal {D} _ {\mathbb {Q}} \left(\theta_ {1}, \theta_ {2}\right) = \| \mathbf {S} _ {0, 1} - \mathbf {S} _ {0, 2} \| _ {2, \mathbb {Q}} + \| \beta_ {1} - \beta_ {2} \| _ {2}, +$$ + +for any $\theta_{1} = (\mathbf{S}_{0,1},\beta_{1}),\theta_{2} = (\mathbf{S}_{0,2},\beta_{2})\in \Theta$ , where $\| \cdot \|_{2,\mathbb{Q}}$ denotes the $L_{2}(\mathbb{Q})$ norm with respect to a probability measure $\mathbb{Q}$ , that is, $\| \mathbf{S}_0\|_{2,\mathbb{Q}} = \left(\sum_{k = 1}^{K}(S_{0,k})^2 q(g_k)\right)^{1 / 2}$ . For a given parameter $\theta$ , let $\mathbb{P}_{\theta}^{n}$ denote the law of $\mathbf{D}_n$ under $\theta$ , and let $\mathbb{E}_{\theta}^{n}$ be the corresponding expectation. With these definitions in place, we now state two theorems that establish the convergence rates of the posterior distribution in two distinct cases: $\gamma < 1 / 3$ and $\gamma \geq 1 / 3$ + +Theorem 3.1 (Case $\gamma < 1/3$ ). Suppose that $\gamma < 1/3$ and assumptions (A1)-(A5) hold. Let $\epsilon_{n} = n^{-\frac{1 - \gamma}{2}}\sqrt{\log n} + \sqrt{\frac{d}{n}}\sqrt{\log(d \vee n)}$ . Then, there exist positive constants $C_{1}, \ldots, C_{4}$ , depending only on $L, B, p_{\min}, p_{\max}, \kappa, \underline{\alpha}, \overline{\alpha}, \rho$ , such that for $n \geq C_{4}$ , + +$$ +\Pi \left(\mathcal {D} _ {\mathbb {Q}} (\theta , \theta^ {*}) \geq C _ {1} \epsilon_ {n} \mid \mathbf {D} _ {n}\right) \leq C _ {2} \exp (- C _ {3} n \epsilon_ {n} ^ {2}), +$$ + +with $\mathbb{P}_{\theta^*}^{n}$ -probability at least $1 - \left(\exp (-C_3n\epsilon_n^2) + 1 / n\epsilon_n^2\right)$ . + +Theorem 3.2 (Case $\gamma \geq 1/3$ ). Suppose that $\gamma \geq 1/3$ and assumptions (A1)-(A5) hold. Let $\epsilon_n = \left(\frac{\log n}{n}\right)^{\frac{1}{3}} + \sqrt{\frac{d}{n}} \sqrt{\log(d \vee n)}$ . Then, there exist positive constants $C_1, \ldots, C_4$ , depending only on $L, B, p_{\min}, p_{\max}, \kappa, \underline{\alpha}, \overline{\alpha}, \rho$ , such that + +$$ +\Pi \left(\mathcal {D} _ {\mathbb {Q}} (\theta , \theta^ {*}) \geq C _ {1} \epsilon_ {n} \mid \mathbf {D} _ {n}\right) \leq C _ {2} \xi_ {n}, \quad n \geq C _ {4} +$$ + +with $\mathbb{P}_{\theta^*}^n$ -probability at least $1 - \left(\xi_n + 1 / n\epsilon_n^2\right)$ , where $\xi_{n} = \begin{cases} \exp (-C_{3}n\epsilon_{n}^{2}) & \text{if } \gamma < \frac{2}{3}, \\ \exp \left(-C_{3}n^{\frac{1}{3}}\right) & \text{if } \gamma \geq \frac{2}{3}. \end{cases}$ + +Theorems 3.1 and 3.2 show that the convergence rate of the posterior distribution adapts to the sparsity level $\gamma$ . Importantly, when $\gamma \geq 1/3$ , the posterior achieves the convergence rate of $n^{-1/3}$ , as in the continuous observation setting. In contrast, for $\gamma < 1/3$ , the posterior attains a faster rate of $n^{-\frac{1 - \gamma}{2}}$ , highlighting the advantage of discrete observations in sparse grids. This result generalizes the work of [5], which focused on the non-contextual case 1 interval-censored data, to the PH model. + +# 4 Proposed BayesCoxCP algorithm + +We now propose the contextual discrete pricing algorithm, namely BayesCoxCP, based on a Bayesian approach to the semi-parametric Cox PH model. Consider the discrete pricing setting introduced earlier. Assume that the support of $P_{t}$ is $\mathcal{G} = \{g_k : k = 1, \dots, K\}$ for every $t = 1, \dots, T$ , where $T$ denotes the time horizon. Here, $g_{k} = p_{\mathrm{min}} + k\delta$ for $k = 1, \dots, K$ , $K = \lfloor (p_{\mathrm{max}} - p_{\mathrm{min}}) / \delta \rfloor$ , and $\delta = \kappa T^{-\gamma}$ for two constants $\gamma \in (0,1]$ and $\kappa > 0$ . Under the PH assumption with the true pair $(S_0^*, \beta^*)$ , the optimal price $P_{t}^{*}$ at time $t$ can be defined as $P_{t}^{*} \in \operatorname{argmax}_{p \in \mathcal{G}} \left\{ p \cdot S_{0}^{*}(p)^{\exp(X_{t}^{\top}\beta^{*})} \right\}$ . Let $\mathbb{Q}^*$ denote the marginal distribution of $P_{t}^{*}$ , with its associated probability mass function denoted by $q^{*}$ . + +We employ an epoch-based design that divides the given horizon $T$ into multiple epochs and executes identical pricing policies on a per-epoch basis. Such a design was widely adopted in the literature [24, 44, 7]. Epochs are indexed by $l$ , and the length of the epoch $l$ is denoted by $n_l$ . The length increases geometrically with $l$ , given by $n_l = n_1 2^{l - 1}$ for $l \geq 1$ . The set of time indices for epoch $l$ is given by $\mathcal{E}_l = \{\sum_{s=0}^{l-1} n_s + 1, \ldots, \sum_{s=0}^l n_s\}$ , with $n_0 = 0$ , ensuring a sequential partitioning of the entire horizon. + +Posterior-based estimation Let $\mathbf{D}_l = \{(X_t, P_t, Y_t)\}_{t \in \mathcal{E}_l}$ denote the data collected during epoch $l \geq 1$ . We employ a consistent prior across all epochs, denoted as $\Pi = \Pi_{\beta} \times \Pi_{\lambda_0}$ , where $\Pi_{\lambda_0}$ consists of independent gamma distributions: + +$$ +\lambda_ {0, k} \sim \operatorname {G a m m a} \left(\alpha_ {k}, \rho\right), \quad k = 1, \dots , K, \tag {4} +$$ + +with $\alpha_{k} = \alpha$ for $k = 1,\dots ,K$ and a fixed constant $\alpha >0$ . The prior $\Pi_{\beta}$ on $\beta$ has a density with respect to the Lebesgue measure on $\mathbb{R}^d$ , bounded away from zero in a neighborhood of $\beta^{*}$ . Common choices for $\Pi_{\beta}$ include multivariate distributions such as the normal distribution. For each epoch $l$ , let $\Pi (\cdot \mid \mathbf{D}_{l - 1})$ denote the joint posterior distribution of $\lambda_0$ and $\beta$ based on the data from the previous epoch, $\mathbf{D}_{l - 1}$ . We denote the point estimator for the true parameter $\theta^{*}$ as $\widehat{\theta}^{l - 1} = (\widehat{\mathbf{S}}_0^{l - 1},\widehat{\beta}^{l - 1})$ , derived from the observations $\mathbf{D}_{l - 1}$ in the previous epoch. Specifically, the estimator $\widehat{\theta}^{l - 1}$ is obtained as the mean of truncated posterior distribution $\widetilde{\Pi} (\cdot \mid \mathbf{D}_{l - 1}) = \Pi (\cdot \mid \mathbf{D}_{l - 1}) / \Pi (\widetilde{\Theta}\mid \mathbf{D}_{l - 1})$ , where the truncated parameter space is defined by $\widetilde{\Theta} = S_0\times [a,b]^d$ for fixed constants $a$ and $b$ . + +Algorithm 1 Bayes Cox Contextual Pricing Algorithm (BayesCoxCP) +Input: $n_1$ : the length of the first epoch; $\eta_{1},\eta_{2}$ degree of exploration; $\Pi_{\lambda_0},\Pi_{\beta}$ prior; $a,b$ truncated range +1: For $t = 1,\dots ,n_1$ uniformly choose $P_{t}$ from $\mathcal{G}$ and get reward $Y_{t}$ +2: for epoch $l = 2,3,\ldots$ do +3: Obtain the estimator $\widehat{\theta}^{l - 1} = (\widehat{\mathbf{S}}_0^{l - 1},\widehat{\beta}^{l - 1})$ from $\Pi (\cdot \mid \mathbf{D}_{l - 1})$ +4: for time $t\in \mathcal{E}_l$ do +5: Observe $X_{t}$ and draw a binary number $R$ from Bernoulli(1- $\eta_l$ . +6: if $R = 1$ then $P_{t}\in \mathrm{argmax}_{p\in \mathcal{G}}\left\{p\cdot \widehat{S}_0^{l - 1}(p)^{\exp (X_t^\top \widehat{\beta}^{l - 1})}\right\}$ +7: else Uniformly choose $P_{t}$ from $\mathcal{G}$ +8: end if +9: Get reward $Y_{t}$ +10: end for +11: end for + +Pricing policy We denote the pricing policy for epoch $l$ as $\pi_l: \mathcal{X} \to \mathcal{P}(\mathcal{G})$ , where $\mathcal{P}(\mathcal{G})$ denotes the set of all probability distributions over the grid $\mathcal{G}$ . Specifically, given covariates $X_t$ for $t \in \mathcal{E}_l$ , the distribution $\pi_l(X_t)$ is defined as a mixture distribution given by + +$$ +\pi_ {l} \left(X _ {t}\right) (A) = \left(1 - \eta_ {l}\right) \cdot \delta_ {\hat {P} _ {t} ^ {l - 1}} (A) + \eta_ {l} \cdot \mathbb {U} _ {\mathcal {G}} (A) \tag {5} +$$ + +for any $A \subset \mathcal{G}$ , where $\widehat{P}_t^{l-1}$ is the myopic policy determined by the estimate $\widehat{\theta}^{l-1} = (\widehat{\mathbf{S}}_0^{l-1}, \widehat{\beta}^{l-1})$ as $\widehat{P}_t^{l-1} \in \operatorname{argmax}_{p \in \mathcal{G}} \left\{ p \cdot \widehat{S}_0^{l-1}(p)^{\exp(X_t^\top \widehat{\beta}^{l-1})} \right\}$ . Here, $\delta_P$ denotes the Dirac measure centered at $P$ , $\mathbb{U}_{\mathcal{G}}$ represents the discrete uniform distribution over $\mathcal{G}$ , and $\eta_l$ is an epoch-specific exploration parameter, defined as + +$$ +\eta_ {l} = \min \left\{\eta_ {1} \left(\eta_ {2} \sqrt {\left| \mathcal {G} \right| / 2 ^ {l - 1}} \wedge 2 ^ {- \frac {l - 1}{3}}\right) \sqrt {\log 2 ^ {l - 1}}, 1 \right\}, \tag {6} +$$ + +where $\eta_{1}$ and $\eta_{2}$ are global constants. The design in (6) reduces the need for uniform exploration when the grid is sparse, while increasing it as the grid becomes denser, effectively balancing exploration and exploitation across different epochs. The choice of $\eta_{l}$ directly ensures that assumption (A4) is satisfied, since $\eta_{l}$ controls the degree of uniform exploration over the grid, which is reflected in $q(\cdot \mid x)$ . This connection is rigorously established in Lemma C.4. In our numerical experiments, $\eta_{1}$ and $\eta_{2}$ are tuned to optimize the degree of exploration. + +in + +The pseudo-code for the proposed policy is presented in Algorithm 1. In this algorithm, for each time $t \in \mathcal{E}_l$ , the offered price $P_t$ solely relies on the observed covariate $X_t$ and the data from previous epochs, $\mathbf{D}_1, \ldots, \mathbf{D}_{l-1}$ , while the distribution of $V_t$ only depends on $X_t$ . Thus, Algorithm 1 ensures conditional independence between $V_t$ and $P_t$ given $X_t$ , i.e., $V_t \perp P_t | X_t$ for each $t \in \mathcal{E}_l$ . Moreover, given the data from previous epochs $1, \ldots, l-1$ , $\{(X_t, P_t, Y_t)\}_{t \in \mathcal{E}_l}$ are independent and identically distributed observations, which facilitates separate estimation of $\theta^*$ for each epoch. + +For the computation of $\widehat{\theta}^{l-1}$ for each epoch $l$ , we employ the variational Bayesian (VB) method for the PH model with case 1 interval-censored data. The VB approach has recently emerged as a computationally efficient alternative while maintaining estimation accuracy; see [29]. Alternatively, one may employ Markov chain Monte Carlo (MCMC) methods [27, 47, 35], which facilitate inference, such as constructing credible intervals for $\theta^*$ . + +# 5 Regret analysis + +In this section, we analyze the regret upper bound for the BayesCoxCP algorithm. Furthermore, we prove the regret lower bound for the discrete pricing problem. + +# 5.1 Regret upper bound + +We first introduce several technical assumptions and a key lemma that establishes the estimation error of the estimator $\widehat{\theta}^{l-1}$ for each epoch. + +We begin by assuming the following additional conditions: + +(B1) For any $x \in \mathcal{X}$ , there exists a unique maximizer of the map $p \mapsto pS_0^*(p)^{\exp(x^\top \beta^*)} : [p_{\min}, p_{\max}] \to \mathbb{R}$ . +(B2) The density of the unique maximizer of the map $p \mapsto p S_0^*(p)^{\exp(X^\top \beta^*)} : [p_{\min}, p_{\max}] \to \mathbb{R}$ is bounded away from zero on $[p_{\min}, p_{\max}]$ . + +The uniqueness condition in assumption (B1) is commonly adopted in the contextual dynamic pricing literature [24, 44, 10]. Additionally, assumption (B2) ensures that $q^{*}(p) \asymp \delta$ for all $p \in \mathcal{G}$ . + +We remark that the grid $\mathcal{G}$ remains unchanged across epochs in our setup, so the grid sparsity relative to the sample size $n_l$ differs for each epoch. For each $l = 1,2,\ldots$ , define $\gamma_l$ as the sparsity level in epoch $l$ , such that $K = \left\lfloor (p_{\max} - p_{\min}) / (\kappa n_l^{-\gamma_l})\right\rfloor$ . Therefore, for all $l = 1,2,\ldots$ , we have + +$$ +\left\lfloor \frac {p _ {\operatorname* {m a x}} - p _ {\operatorname* {m i n}}}{\kappa} n _ {\gamma} ^ {\gamma_ {\iota}} \right\rfloor = \left\lfloor \frac {p _ {\operatorname* {m a x}} - p _ {\operatorname* {m i n}}}{\kappa} T ^ {\gamma} \right\rfloor . \tag {7} +$$ + +Let $\mathbb{Q}_l(\cdot \mid X)$ and $q_{l}(\cdot \mid X)$ denote the conditional distribution and corresponding probability mass function of $P_{t}$ given $X$ (and $\mathbf{D}_1,\ldots ,\mathbf{D}_{l - 1})$ during epoch $l$ , respectively. The marginal distribution of $P_{t}$ during epoch $l$ is denoted by $\mathbb{Q}_l$ , with $q_{l}$ as its probability mass function. Then, $q_{l}(\cdot \mid x) = \pi_{l}(x)(\cdot)$ for $x\in \mathcal{X}$ . + +The following lemma provides an upper bound on the estimation error of the point estimator $\widehat{\theta}^{l-1}$ at epoch $l$ . + +Lemma 5.1. Let the prior $\Pi$ and policy $\pi_l$ be as described above (see (4) and (5)). Suppose that assumptions (A1)-(A3) and (A5) hold. Then, there exist positive constants $C_1, C_2, C_3$ and $C_4$ depending on $L, B, p_{\min}, p_{\max}, \kappa, \alpha, \rho, a, b$ and $n_1$ , such that for $l \geq C_4$ , + +$$ +\mathcal {D} _ {\mathbb {Q} _ {l - 1}} \left(\widehat {\theta} ^ {l - 1}, \theta^ {*}\right) \leq C _ {1} \epsilon_ {l - 1} +$$ + +with probability at least $1 - \zeta_{l-1} - 1 / (n_{l-1}\epsilon_{l-1}^2)$ , where + +$$ +\epsilon_ {l} = \left\{ \begin{array}{l l} n _ {l} ^ {- \frac {1 - \gamma_ {l}}{2}} \sqrt {\log n _ {l}} + \sqrt {\frac {d}{n _ {l}}} \sqrt {\log (d \lor n _ {l})} & i f \gamma_ {l} < \frac {1}{3} \\ \left(\frac {\log n _ {l}}{n _ {l}}\right) ^ {\frac {1}{3}} + \sqrt {\frac {d}{n _ {l}}} \sqrt {\log (d \lor n _ {l})} & i f \gamma_ {l} \geq \frac {1}{3} \end{array} \right. a n d \zeta_ {l} = \left\{ \begin{array}{l l} \exp (- C _ {2} n _ {l} \epsilon_ {l} ^ {2}) & i f \gamma_ {l} < \frac {1}{3}, \\ \exp (- C _ {3} n _ {l} ^ {\frac {1}{3}}) & i f \gamma_ {l} \geq \frac {1}{3}. \end{array} \right. +$$ + +Lemma 5.1 implies that $\widehat{\theta}^{l-1}$ achieves an error bound that is adaptive to the grid sparsity level $\gamma_l$ in epoch $l$ . By leveraging the consistency of $\widehat{\theta}^{l-1}$ and Hoeffding's inequality, the regret during epoch $l$ can be upper bounded by + +$$ +\sum_ {t \in \mathcal {E} _ {l}} r (t) \leq C _ {1} n _ {l} \mathcal {D} _ {\mathbb {Q} _ {l - 1}} \left(\widehat {\theta} ^ {l - 1}, \theta^ {*}\right) + C _ {2} n _ {l} \eta_ {l} +$$ + +with high probability, where $C_1$ and $C_2$ are positive constants which do not scale with $n_l$ and $d$ (see Lemma B.1 for details). This inequality shows how the estimation error of $\widehat{\theta}^{l-1}$ and the exploration parameter $\eta_l$ affect the regret upper bound. Combining Lemma 5.1 and (6), we now state the main results regarding the regret upper bound for the BayesCoxCP algorithm. + +Theorem 5.2. Under the same conditions as in Lemma 5.1, along with assumptions (B1) and (B2), there exist positive constants $C_1,\ldots ,C_7$ depending on $L,B,p_{\mathrm{min}},p_{\mathrm{max}},\kappa ,\alpha ,\rho ,a,b,\eta_1,\eta_2,\gamma$ and $n_1$ such that for $T\geq C_1$ + +$$ +R (T) \leq \left\{ \begin{array}{l l} C _ {2} \sqrt {d T \log (d \lor T)} + C _ {3} T ^ {\frac {\gamma + 1}{2}} \sqrt {\log T} & i f \gamma < \frac {1}{3}, \\ C _ {4} \sqrt {d T \log (d \lor T)} + C _ {5} T ^ {\frac {2}{3}} \sqrt {\log T} & i f \gamma \geq \frac {1}{3}, \end{array} \right. +$$ + +with probability at least $1 - \zeta$ , where $\zeta = \begin{cases} C_6 \log (T / n_1 + 1) / T^\gamma & \text{if } \gamma < \frac{1}{3}, \\ C_7 \log (T / n_1 + 1) / T^{2/9} & \text{if } \gamma \geq \frac{1}{3}. \end{cases}$ + +Theorem 5.2 shows that the BayesCoxCP algorithm achieves a regret upper bound that adapts to the unknown sparsity level $\gamma$ , ensuring efficient performance without prior knowledge of $\gamma$ . + +![](images/5d94fbb52d0223fac8d0203855f789a1a8e81789794e14fe780b18eb0d313c48.jpg) + +![](images/bbf0acc358af0250a20f62496f7a4eb40b59696b0c7cf76ecd08afa2ccb4b05a.jpg) +BayesCoxCP Choi et al. (2023) + +![](images/ffe0073f6b5bc1058bfe2ba2cebd87c3412564a46ed854eaf439e880d39485da.jpg) + +![](images/520f8dc30b49c297d8591bd44955c45b562495ce59b4da485b698d7983a99394.jpg) + +Figure 1: The cumulative regret curves compare the proposed algorithm (BayesCoxCP) with other method. Solid lines indicate averages, and bands show standard errors over replications. +![](images/85c17d08faba7ddc8ec12c7501013847d482c0e80ad4c8eeb2140f6fefc43d51.jpg) +(b) $\frac{3}{4}\mathrm{TN}(3.25,0.5^2,1,10) + \frac{1}{4}\mathrm{TN}(7.75,0.5^2,1,10)$ + +![](images/a0ff1ab213a87bce79cea3b14cfa3bea004c95108e21a38bc5b43999b411e58f.jpg) +(a) $\frac{1}{2}\mathrm{U}(1,4) + \frac{1}{2}\mathrm{U}(4,10)$ +BayesCoxCP Choi et al. (2023) + +![](images/a5f9a8f1ef9fbc043e90f3e24f5182644bd5278ab7bfbb0b6cb21c40cdf02d9b.jpg) + +![](images/287163a1531da4fa9cbd4c4d09dc8e2a333f0f96a3ca3f1642ccc7c262c04e3e.jpg) + +Compared to existing work in continuous pricing settings, our results underscore the theoretical advantage of utilizing the information that the price is discretely supported. For instance, [7] derived a regret upper bound of $\widetilde{O}(T^{2/3}d)$ under similar assumptions. While this bound is comparable to our result for $\gamma \geq 1/3$ , our algorithm achieves a strictly faster regret rate of $O(T^{\frac{1+\gamma}{2}}+(dT)^{1/2})$ for $\gamma < 1/3$ , which is a distinct advantage of the grid-based setting. For additional discussion on the possibility of replacing the Bayes estimator with the NPMLE and its effect on exploration parameter choice, please refer to Appendix G. + +# 5.2 Regret lower bound + +In this subsection, we establish a regret lower bound for the non-contextual pricing problem in the discrete pricing setting. The proof carefully incorporates ideas from [26] and [14], widely used for regret lower bounds in dynamic pricing and multi-armed bandit problems, with a focus on the discrete price setting. Specifically, for dense grids where $\gamma \geq 1/3$ , we partition the grid set $\mathcal{G}$ into $T^{1/3}$ segments to derive the lower bound. Further details of the proof are provided in Appendix B.3. + +Theorem 5.3. (Lower bound for non-contextual pricing) Consider a non-contextual pricing problem where the valuations are sampled independently and identically from a fixed unknown distribution satisfying the c.d.f. $F(v)$ is bounded away from 0 and 1 for $v \in [p_{\min}, p_{\max}]$ and at least one maximizer of the revenue curve $v \cdot (1 - F(v))$ lies over $\mathcal{G}$ . Then, for any $\eta > 0$ , no pricing policy (algorithm) can achieve expected regret $O(T^{\frac{1 + \gamma}{2} - \eta})$ if $\gamma < 1/3$ , and $O(T^{\frac{2}{3} - \eta})$ if $\gamma \geq 1/3$ . + +As in Theorem 5.3, the regret lower bound depends on the grid sparsity level $\gamma$ as well. Specifically, for $\gamma < 1/3$ , the regret lower bound scales as $\Omega(T^{\frac{1+\gamma}{2}})$ , while for $\gamma \geq 1/3$ , it scales as $\Omega(T^{2/3})$ . Comparing these results with Theorems 3.1 and 3.2, the regret upper bounds achieved by BayesCoxCP algorithm match the lower bounds in terms of $T$ , up to a logarithmic factor. Note that the dependency on the dimension $d$ is not addressed in this work, leaving it as a direction for future research. + +# 6 Numerical experiments + +In this section, we conduct numerical experiments to evaluate the performance of the BayesCoxCP algorithm. Since our objective is to highlight the benefits of leveraging discrete support information, we focus on a comparison with the algorithm proposed by [7]. For comparisons of the PH model-based algorithm with other approaches, such as linear and log-linear model-based algorithms, we refer to [7]. + +We consider the following experimental setup. The covariate $X_{t}$ is drawn from a $d$ -dimensional ball with a radius of $1/2$ under a uniform distribution, where $d = 5$ . The true regression coefficient $\beta^{*}$ is + +set as $\beta^{*} = \frac{4}{\sqrt{d}}\mathbf{1}_{d}$ , where $\mathbf{1}_d$ denotes a $d$ -dimensional vector of ones. For the true baseline distribution, we consider two different mixture distributions. The first is a uniform mixture distribution given by $\frac{1}{2}\mathrm{U}(1,4) + \frac{1}{2}\mathrm{U}(4,10)$ , where $\mathrm{U}(a,b)$ denotes the uniform distribution over $[a,b]$ . The second is a truncated normal mixture distribution given by $\frac{3}{4}\mathrm{TN}(3.25,0.5^2,1,10) + \frac{1}{4}\mathrm{TN}(7.75,0.5^2,1,10)$ , where $\mathrm{TN}(\mu ,\sigma^2,a,b)$ represents the truncated normal distribution with mean $\mu$ , variance $\sigma^2$ , and support $[a,b]$ . The grid set $\mathcal{G} = \{g_1,\dots,g_K\}$ is chosen from $[1,10]$ with four different values of $K\in \{10,100,1000,30000\}$ . The total time horizon is set to $T = 30000$ for all experiments. To conserve space, the detailed hyperparameter settings for all algorithms used in the experiments are provided in Appendix H. + +The cumulative regret results for different grid sizes, averaged over 20 replications, are shown in Figure 1. BayesCoxCP consistently achieves lower cumulative regret compared to the method proposed by [7], with the difference being particularly significant when $K$ is small. Notably, the performance gap gradually diminishes as $K$ increases. These findings empirically demonstrate that BayesCoxCP adapts effectively to varying grid resolutions, providing strong empirical support for its theoretical guarantees. + +# Acknowledgements + +This work was supported by the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (No. RS-2023-00240861, RS-2023-00252026), a Korea Institute for Advancement of Technology (KIAT) grant funded by the Korea Government (MOTIE) (RS-2024-00409092, 2024 HRD Program for Industrial Innovation), and Institute of Information & Communications Technology Planning & Evaluation(IITP)-Global Data-X Leader HRD program grant funded by the Korea government (MSIT) (IITP-2024-RS-2024-00441244). + +# References + +[1] Pierre Alquier and James Ridgway. Concentration of tempered posteriors and of their variational approximations. Annals of Statistics, 48(3):1475-1497, 2020. +[2] Kareem Amin, Afshin Rostamizadeh, and Umar Syed. Repeated contextual auctions with strategic buyers. In Proc. Neural Information Processing Systems, 2014. +[3] Gah-Yi Ban and N Bora Keskin. Personalized dynamic pricing with machine learning: High-dimensional features and heterogeneous elasticity. Management Science, 67(9):5549-5568, 2021. +[4] Omar Besbes and Assaf Zeevi. Blind network revenue management. Operations Research, 60(6):1537-1550, 2012. +[5] Minwoo Chae. Adaptive Bayesian inference for current status data on a grid. Bernoulli, 29(1):403-427, 2023. +[6] Xi Chen, David Simchi-Levi, and Yining Wang. Privacy-preserving dynamic personalized pricing with demand learning. Management Science, 68(7):4878-4898, 2022. +[7] Young-Geun Choi, Gi-Soo Kim, Yunseo Choi, Wooseong Cho, Myunghee Cho Paik, and Min-hwan Oh. Semi-parametric contextual pricing algorithm using Cox proportional hazards model. In Proc. International Conference on Machine Learning, pages 5771-5786, 2023. +[8] Gregory Cox. A generalized argmax theorem with applications. arXiv preprint arXiv:2209.08793, 2022. +[9] Arnoud V. den Boer. Dynamic pricing and learning: Historical origins, current research, and new directions. Surveys in Operations Research and Management Science, 20(1):1-18, 2015. +[10] Jianqing Fan, Yongyi Guo, and Mengxin Yu. Policy optimization using semiparametric models for dynamic pricing. Journal of the American Statistical Association, 119(545):552-564, 2024. +[11] Dianne M Finkelstein and Robert A Wolfe. A semiparametric model for regression analysis of interval-censored failure time data. Biometrics, 41(4):933-945, 1985. + +[12] Guillermo Gallego and Garrett Van Ryzin. Optimal dynamic pricing of inventories with stochastic demand over finite horizons. Management Science, 40(8):999-1020, 1994. +[13] Robert Gentleman and Charles J Geyer. Maximum likelihood for interval censored data: Consistency and computation. Biometrika, 81(3):618-623, 1994. +[14] Sébastien Gerchinovitz and Tor Lattimore. Refined lower bounds for adversarial bandits. In Proc. Neural Information Processing Systems, 2016. +[15] Subhashis Ghosal and Aad Van Der Vaart. Convergence rates of posterior distributions for noniid observations. Annals of Statistics, 35(1):192-223, 2007. +[16] Subhashis Ghosal and Aad Van der Vaart. Fundamentals of Nonparametric Bayesian Inference, volume 44. Cambridge University Press, 2017. +[17] Negin Golrezaei, Adel Javanmard, and Vahab Mirrokni. Dynamic incentive-aware learning: Robust pricing in contextual auctions. In Proc. Neural Information Processing Systems, 2019. +[18] Piet Groeneboom. Nonparametric maximum likelihood estimators for interval censoring and deconvolution. Technical report, Department of Statistics, Stanford University, Stanford, California, 1991. +[19] Piet Groeneboom and Kim Hendrickx. Current status linear regression. Annals of Statistics, 46(4):1415-1444, 2018. +[20] Piet Groeneboom, Marloes H Maathuis, and Jon A Wellner. Current status data with competing risks: Limiting distribution of the mle. Annals of Statistics, 36(3):1064-1089, 2008. +[21] Piet Groeneboom and Jon A Wellner. Information Bounds and Nonparametric Maximum Likelihood Estimation, volume 19. Birkhäuser, 2012. +[22] Pavithra Harsha, Shivaram Subramanian, and Joline Uichanco. Dynamic pricing of omnichannel inventories. Manufacturing & Service Operations Management, 21(1):47-65, 2019. +[23] Adel Javanmard. Perishability of data: Dynamic pricing under varying-coefficient models. Journal of Machine Learning Research, 18(53):1-31, 2017. +[24] Adel Javanmard and Hamid Nazerzadeh. Dynamic pricing in high-dimensions. Journal of Machine Learning Research, 20(9):1-49, 2019. +[25] Nicholas P Jewell and Mark van der Laan. Current status data: Review, recent developments and open problems. In Handbook of Statistics, volume 23, pages 625-642. Elsevier, 2003. +[26] Robert Kleinberg and Tom Leighton. The value of knowing a demand curve: Bounds on regret for online posted-price auctions. In Proc. IEEE Symposium on Foundations of Computer Science, pages 594-605, 2003. +[27] Xiaoyan Lin, Bo Cai, Lianming Wang, and Zhigang Zhang. A Bayesian proportional hazards model for general interval-censored data. *Lifetime Data Analysis*, 21:470–490, 2015. +[28] Jane C Lindsey and Louise M Ryan. Methods for interval-censored data. Statistics in Medicine, 17(2):219-238, 1998. +[29] Wenting Liu, Huiqiong Li, Niansheng Tang, and Jun Lyu. Variational Bayesian approach for analyzing interval-censored data under the proportional hazards model. Computational Statistics & Data Analysis, 195:107957, 2024. +[30] Yiyun Luo, Will Wei Sun, and Yufeng Liu. Contextual dynamic pricing with unknown noise: Explore-then-ucb strategy and improved regrets. In Proc. Neural Information Processing Systems, pages 37445–37457, 2022. +[31] Yiyun Luo, Will Wei Sun, and Yufeng Liu. Distribution-free contextual dynamic pricing. Mathematics of Operations Research, 49(1):599-618, 2024. + +[32] Chandrasekhar Manchiraju, Milind Dawande, and Ganesh Janakiraman. Multiproduct pricing with discrete price sets. Operations Research, 70(4):2185-2193, 2022. +[33] Sentao Miao, Xi Chen, Xiuli Chao, Jiaxi Liu, and Yidong Zhang. Context-based dynamic pricing with online clustering. Production and Operations Management, 31(9):3559-3575, 2022. +[34] Velibor V Mišić and Georgia Perakis. Data analytics in operations management: A review. Manufacturing & Service Operations Management, 22(1):158-169, 2020. +[35] Chun Pan, Bo Cai, and Lianming Wang. A Bayesian approach for analyzing partly interval-censored data under the proportional hazards model. Statistical Methods in Medical Research, 29(11):3192-3204, 2020. +[36] Sheng Qiang and Mohsen Bayati. Dynamic pricing with demand covariates. arXiv preprint arXiv:1604.07463, 2016. +[37] Virag Shah, Ramesh Johari, and Jose Blanchet. Semi-parametric dynamic contextual pricing. In Proc. Neural Information Processing Systems, 2019. +[38] Xiaotong Shen. Linear regression with current status data. Journal of the American Statistical Association, 95(451):842-852, 2000. +[39] Kalyan T Talluri and Garrett J Van Ryzin. The Theory and Practice of Revenue Management, volume 68. Springer Science & Business Media, 2006. +[40] Inder Jeet Taneja and Pranesh Kumar. Relative information of type s, csiszar's f-divergence, and information inequalities. Information Sciences, 166(1):105-125, 2004. +[41] Runlong Tang, Moulinath Banerjee, and Michael R Kosorok. Likelihood based inference for current status data on a grid: A boundary phenomenon and an adaptive inference procedure. Annals of Statistics, 40(1):45-72, 2012. +[42] Alexandre B. Tsybakov. Introduction to Nonparametric Estimation. Springer New York, NY, 2009. +[43] Mike Mingcheng Wei and Fuqiang Zhang. Recent research developments of strategic consumer behavior in operations management. Computers & Operations Research, 93:166-176, 2018. +[44] Jianyu Xu and Yu-Xiang Wang. Logarithmic regret in feature-based dynamic pricing. In Proc. Neural Information Processing Systems, pages 13898-13910, 2021. +[45] Yun Yang, Debdeep Pati, and Anirban Bhattacharya. $\alpha$ -variational inference with statistical guarantees. Annals of Statistics, 48(2):886-905, 2020. +[46] Fengshuo Zhang and Chao Gao. Convergence rates of variational posterior distributions. Annals of Statistics, 48(4):2180-2207, 2020. +[47] Haiming Zhou and Timothy Hanson. A unified framework for fitting Bayesian semiparametric models to arbitrarily censored survival data, including spatially referenced data. Journal of the American Statistical Association, 113(522):571-581, 2018. + +# NeurIPS Paper Checklist + +# 1. Claims + +Question: Do the main claims made in the abstract and introduction accurately reflect the paper's contributions and scope? + +Answer: [Yes] + +Justification: The abstract and introduction clearly state the contributions. These claims are supported by both theoretical analysis and numerical experiments. + +Guidelines: + +- The answer NA means that the abstract and introduction do not include the claims made in the paper. +- The abstract and/or introduction should clearly state the claims made, including the contributions made in the paper and important assumptions and limitations. A No or NA answer to this question will not be perceived well by the reviewers. +- The claims made should match theoretical and experimental results, and reflect how much the results can be expected to generalize to other settings. +- It is fine to include aspirational goals as motivation as long as it is clear that these goals are not attained by the paper. + +# 2. Limitations + +Question: Does the paper discuss the limitations of the work performed by the authors? + +Answer: [Yes] + +Justification: We discuss the limitation that regret lower bounds are driven only for the non-contextual pricing, and the dependence on the dimension $d$ is not addressed. The need for future work in this direction is mentioned in Section 5.2. + +Guidelines: + +- The answer NA means that the paper has no limitation while the answer No means that the paper has limitations, but those are not discussed in the paper. +- The authors are encouraged to create a separate "Limitations" section in their paper. +- The paper should point out any strong assumptions and how robust the results are to violations of these assumptions (e.g., independence assumptions, noiseless settings, model well-specification, asymptotic approximations only holding locally). The authors should reflect on how these assumptions might be violated in practice and what the implications would be. +- The authors should reflect on the scope of the claims made, e.g., if the approach was only tested on a few datasets or with a few runs. In general, empirical results often depend on implicit assumptions, which should be articulated. +- The authors should reflect on the factors that influence the performance of the approach. For example, a facial recognition algorithm may perform poorly when image resolution is low or images are taken in low lighting. Or a speech-to-text system might not be used reliably to provide closed captions for online lectures because it fails to handle technical jargon. +- The authors should discuss the computational efficiency of the proposed algorithms and how they scale with dataset size. +- If applicable, the authors should discuss possible limitations of their approach to address problems of privacy and fairness. +- While the authors might fear that complete honesty about limitations might be used by reviewers as grounds for rejection, a worse outcome might be that reviewers discover limitations that aren't acknowledged in the paper. The authors should use their best judgment and recognize that individual actions in favor of transparency play an important role in developing norms that preserve the integrity of the community. Reviewers will be specifically instructed to not penalize honesty concerning limitations. + +# 3. Theory assumptions and proofs + +Question: For each theoretical result, does the paper provide the full set of assumptions and a complete (and correct) proof? + +Answer: [Yes] + +Justification: All assumptions required for the theoretical results are explicitly stated in the paper. Most proofs are provided in the appendix, while the main theorems, key lemmas and insights are included in the main body. + +Guidelines: + +- The answer NA means that the paper does not include theoretical results. +- All the theorems, formulas, and proofs in the paper should be numbered and cross-referenced. + +- All assumptions should be clearly stated or referenced in the statement of any theorems. +- The proofs can either appear in the main paper or the supplemental material, but if they appear in the supplemental material, the authors are encouraged to provide a short proof sketch to provide intuition. +- Inversely, any informal proof provided in the core of the paper should be complemented by formal proofs provided in appendix or supplemental material. +- Theorems and Lemmas that the proof relies upon should be properly referenced. + +# 4. Experimental result reproducibility + +Question: Does the paper fully disclose all the information needed to reproduce the main experimental results of the paper to the extent that it affects the main claims and/or conclusions of the paper (regardless of whether the code and data are provided or not)? + +Answer: [Yes] + +Justification: The implementation details of the proposed algorithm are described in Section 4. The experimental setup and hyperparameters are provided in Section 6 and Appendix H. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- If the paper includes experiments, a No answer to this question will not be perceived well by the reviewers: Making the paper reproducible is important, regardless of whether the code and data are provided or not. +- If the contribution is a dataset and/or model, the authors should describe the steps taken to make their results reproducible or verifiable. +- Depending on the contribution, reproducibility can be accomplished in various ways. For example, if the contribution is a novel architecture, describing the architecture fully might suffice, or if the contribution is a specific model and empirical evaluation, it may be necessary to either make it possible for others to replicate the model with the same dataset, or provide access to the model. In general, releasing code and data is often one good way to accomplish this, but reproducibility can also be provided via detailed instructions for how to replicate the results, access to a hosted model (e.g., in the case of a large language model), releasing of a model checkpoint, or other means that are appropriate to the research performed. +- While NeurIPS does not require releasing code, the conference does require all submissions to provide some reasonable avenue for reproducibility, which may depend on the nature of the contribution. For example +(a) If the contribution is primarily a new algorithm, the paper should make it clear how to reproduce that algorithm. +(b) If the contribution is primarily a new model architecture, the paper should describe the architecture clearly and fully. +(c) If the contribution is a new model (e.g., a large language model), then there should either be a way to access this model for reproducing the results or a way to reproduce the model (e.g., with an open-source dataset or instructions for how to construct the dataset). +(d) We recognize that reproducibility may be tricky in some cases, in which case authors are welcome to describe the particular way they provide for reproducibility. In the case of closed-source models, it may be that access to the model is limited in some way (e.g., to registered users), but it should be possible for other researchers to have some path to reproducing or verifying the results. + +# 5. Open access to data and code + +Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material? + +Answer: [Yes] + +Justification: All code necessary to reproduce the numerical experiments is included in the supplemental material. + +Guidelines: + +- The answer NA means that paper does not include experiments requiring code. +- Please see the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details. +- While we encourage the release of code and data, we understand that this might not be possible, so "No" is an acceptable answer. Papers cannot be rejected simply for not including code, unless this is central to the contribution (e.g., for a new open-source benchmark). +- The instructions should contain the exact command and environment needed to run to reproduce the results. See the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details. +- The authors should provide instructions on data access and preparation, including how to access the raw data, preprocessed data, intermediate data, and generated data, etc. +- The authors should provide scripts to reproduce all experimental results for the new proposed method and baselines. If only a subset of experiments are reproducible, they should state which ones are omitted from the script and why. +- At submission time, to preserve anonymity, the authors should release anonymized versions (if applicable). +- Providing as much information as possible in supplemental material (appended to the paper) is recommended, but including URLs to data and code is permitted. + +# 6. Experimental setting/details + +Question: Does the paper specify all the training and test details (e.g., data splits, hyperparameters, how they were chosen, type of optimizer, etc.) necessary to understand the results? + +Answer: [Yes] + +Justification: We provide the detailed experimental settings including how cumulative regret is computed and the hyperparameter configurations in Section 6 and Appendix H. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The experimental setting should be presented in the core of the paper to a level of detail that is necessary to appreciate the results and make sense of them. +- The full details can be provided either with the code, in appendix, or as supplemental material. + +# 7. Experiment statistical significance + +Question: Does the paper report error bars suitably and correctly defined or other appropriate information about the statistical significance of the experiments? + +Answer: [Yes] + +Justification: The experimental results are obtained via repeated trials, and error bars are shown and explained in the figure and this caption. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The authors should answer "Yes" if the results are accompanied by error bars, confidence intervals, or statistical significance tests, at least for the experiments that support the main claims of the paper. +- The factors of variability that the error bars are capturing should be clearly stated (for example, train/test split, initialization, random drawing of some parameter, or overall run with given experimental conditions). +- The method for calculating the error bars should be explained (closed form formula, call to a library function, bootstrap, etc.) +- The assumptions made should be given (e.g., Normally distributed errors). +- It should be clear whether the error bar is the standard deviation or the standard error of the mean. + +- It is OK to report 1-sigma error bars, but one should state it. The authors should preferably report a 2-sigma error bar than state that they have a $96\%$ CI, if the hypothesis of Normality of errors is not verified. +- For asymmetric distributions, the authors should be careful not to show in tables or figures symmetric error bars that would yield results that are out of range (e.g. negative error rates). +- If error bars are reported in tables or plots, The authors should explain in the text how they were calculated and reference the corresponding figures or tables in the text. + +# 8. Experiments compute resources + +Question: For each experiment, does the paper provide sufficient information on the computer resources (type of compute workers, memory, time of execution) needed to reproduce the experiments? + +Answer: [Yes] + +Justification: We provide the computational resources used in our experiments in Appendix I. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The paper should indicate the type of compute workers CPU or GPU, internal cluster, or cloud provider, including relevant memory and storage. +- The paper should provide the amount of compute required for each of the individual experimental runs as well as estimate the total compute. +- The paper should disclose whether the full research project required more compute than the experiments reported in the paper (e.g., preliminary or failed experiments that didn't make it into the paper). + +# 9. Code of ethics + +Question: Does the research conducted in the paper conform, in every respect, with the NeurIPS Code of Ethics https://neurips.cc/public/EthicsGuidelines? + +Answer: [Yes] + +Justification: We have carefully reviewed the NeurIPS Code of Ethics and confirmed that our research does not violate any of its principles. All numerical experiments are conducted using simulated data, and thus this work do not involve any human subjects or data-related concerns. + +Guidelines: + +The answer NA means that the authors have not reviewed the NeurIPS Code of Ethics. +- If the authors answer No, they should explain the special circumstances that require a deviation from the Code of Ethics. +- The authors should make sure to preserve anonymity (e.g., if there is a special consideration due to laws or regulations in their jurisdiction). + +# 10. Broader impacts + +Question: Does the paper discuss both potential positive societal impacts and negative societal impacts of the work performed? + +Answer: [NA] + +Justification: This paper focuses on theoretical analysis of regret bounds in contextual pricing, and does not have any direct societal impact. + +Guidelines: + +- The answer NA means that there is no societal impact of the work performed. +- If the authors answer NA or No, they should explain why their work has no societal impact or why the paper does not address societal impact. +- Examples of negative societal impacts include potential malicious or unintended uses (e.g., disinformation, generating fake profiles, surveillance), fairness considerations (e.g., deployment of technologies that could make decisions that unfairly impact specific groups), privacy considerations, and security considerations. + +- The conference expects that many papers will be foundational research and not tied to particular applications, let alone deployments. However, if there is a direct path to any negative applications, the authors should point it out. For example, it is legitimate to point out that an improvement in the quality of generative models could be used to generate deepfakes for disinformation. On the other hand, it is not needed to point out that a generic algorithm for optimizing neural networks could enable people to train models that generate Deepfakes faster. +- The authors should consider possible harms that could arise when the technology is being used as intended and functioning correctly, harms that could arise when the technology is being used as intended but gives incorrect results, and harms following from (intentional or unintentional) misuse of the technology. +- If there are negative societal impacts, the authors could also discuss possible mitigation strategies (e.g., gated release of models, providing defenses in addition to attacks, mechanisms for monitoring misuse, mechanisms to monitor how a system learns from feedback over time, improving the efficiency and accessibility of ML). + +# 11. Safeguards + +Question: Does the paper describe safeguards that have been put in place for responsible release of data or models that have a high risk for misuse (e.g., pretrained language models, image generators, or scraped datasets)? + +Answer: [NA] + +Justification: This paper is theoretical and conducts experiments solely on simulated data. No high-risk models or real-world datasets are released. + +Guidelines: + +- The answer NA means that the paper poses no such risks. +- Released models that have a high risk for misuse or dual-use should be released with necessary safeguards to allow for controlled use of the model, for example by requiring that users adhere to usage guidelines or restrictions to access the model or implementing safety filters. +- Datasets that have been scraped from the Internet could pose safety risks. The authors should describe how they avoided releasing unsafe images. +- We recognize that providing effective safeguards is challenging, and many papers do not require this, but we encourage authors to take this into account and make a best faith effort. + +# 12. Licenses for existing assets + +Question: Are the creators or original owners of assets (e.g., code, data, models), used in the paper, properly credited and are the license and terms of use explicitly mentioned and properly respected? + +Answer: [NA] + +Justification: The paper does not use existing assets. + +Guidelines: + +- The answer NA means that the paper does not use existing assets. +- The authors should cite the original paper that produced the code package or dataset. +- The authors should state which version of the asset is used and, if possible, include a URL. +- The name of the license (e.g., CC-BY 4.0) should be included for each asset. +- For scraped data from a particular source (e.g., website), the copyright and terms of service of that source should be provided. +- If assets are released, the license, copyright information, and terms of use in the package should be provided. For popular datasets, paperswithcode.com/datasets has curated licenses for some datasets. Their licensing guide can help determine the license of a dataset. +- For existing datasets that are re-packaged, both the original license and the license of the derived asset (if it has changed) should be provided. + +- If this information is not available online, the authors are encouraged to reach out to the asset's creators. + +# 13. New assets + +Question: Are new assets introduced in the paper well documented and is the documentation provided alongside the assets? + +Answer: [NA] + +Justification: The paper does not release new assets. + +Guidelines: + +- The answer NA means that the paper does not release new assets. +- Researchers should communicate the details of the dataset/code/model as part of their submissions via structured templates. This includes details about training, license, limitations, etc. +- The paper should discuss whether and how consent was obtained from people whose asset is used. +- At submission time, remember to anonymize your assets (if applicable). You can either create an anonymized URL or include an anonymized zip file. + +# 14. Crowdsourcing and research with human subjects + +Question: For crowdsourcing experiments and research with human subjects, does the paper include the full text of instructions given to participants and screenshots, if applicable, as well as details about compensation (if any)? + +Answer: [NA] + +Justification: The paper does not involve crowdsourcing nor research with human subjects. + +Guidelines: + +- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. +- Including this information in the supplemental material is fine, but if the main contribution of the paper involves human subjects, then as much detail as possible should be included in the main paper. +- According to the NeurIPS Code of Ethics, workers involved in data collection, curation, or other labor should be paid at least the minimum wage in the country of the data collector. + +# 15. Institutional review board (IRB) approvals or equivalent for research with human subjects + +Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals (or an equivalent approval/review based on the requirements of your country or institution) were obtained? + +Answer: [NA] + +Justification: The paper does not involve crowdsourcing nor research with human subjects. + +Guidelines: + +- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. +- Depending on the country in which research is conducted, IRB approval (or equivalent) may be required for any human subjects research. If you obtained IRB approval, you should clearly state this in the paper. +- We recognize that the procedures for this may vary significantly between institutions and locations, and we expect authors to adhere to the NeurIPS Code of Ethics and the guidelines for their institution. +- For initial submissions, do not include any information that would break anonymity (if applicable), such as the institution conducting the review. + +# 16. Declaration of LLM usage + +Question: Does the paper describe the usage of LLMs if it is an important, original, or non-standard component of the core methods in this research? Note that if the LLM is used only for writing, editing, or formatting purposes and does not impact the core methodology, scientific rigorousness, or originality of the research, declaration is not required. + +# Answer: [NA] + +Justification: The core method development in this research does not involve LLMs as any important, original, or non-standard components. + +# Guidelines: + +- The answer NA means that the core method development in this research does not involve LLMs as any important, original, or non-standard components. +- Please refer to our LLM policy (https://neurips.cc/Conferences/2025/LLM) for what should or should not be described. + +# Appendix + +We begin this appendix with a proof roadmap that outlines how the main lemmas and theorems are logically connected. The roadmap provides a high-level overview of the argument structure, highlighting the key intermediate steps and how they contribute to the main convergence and regret results. This overview is intended to enhance clarity and to guide the reader through the subsequent detailed proofs. + +![](images/a1a755892946f2b46d00ec97f3cbf420ab75784d51ec4550d2e4194950bd6768.jpg) +Figure 2: Proof roadmap summarizing the logical connections among lemmas and theorems leading to the main results. + +# A Proofs for Section 3 + +In this section, we first establish the posterior consistency of the Cox PH model, which serves as a foundation for proving the main theorems in Section 3. + +Lemma A.1. (Posterior consistency) Suppose that the grid resolution satisfies $\delta = \kappa n^{-\gamma}$ for $\kappa > 0$ and $\gamma \in (0,1]$ , and assumptions (A1)-(A5) hold. If $\gamma < 2/3$ , then, for every $\epsilon > 0$ , there exist positive constants $C_1, C_2$ and $C_3$ depending on $(L, B, p_{\min}, p_{\max}, \kappa, \underline{\alpha}, \overline{\alpha}, \rho, \epsilon)$ such that + +$$ +\Pi \left(U ^ {c} \mid \mathbf {D} _ {n}\right) < C _ {2} \exp \left(- C _ {3} n\right), \quad n \geq C _ {1}, \tag {8} +$$ + +where + +$$ +U = \left\{\theta \in \Theta : \| \mathbf {S} _ {0} - \mathbf {S} _ {0} ^ {*} \| _ {\infty} \vee \| \beta - \beta^ {*} \| _ {2} < \epsilon \right\} +$$ + +with $\mathbb{P}_{\theta^*}^n$ probability at least $1 - \exp (-C_3n)$ + +If $\gamma \geq 2/3$ , then, for every $\epsilon > 0$ , there exist positive constants $C_4, C_5$ and $C_6$ depending on $(L, B, p_{\min}, p_{\max}, \kappa, \underline{\alpha}, \overline{\alpha}, \rho, \epsilon)$ such that + +$$ +\Pi (U ^ {c} \mid \mathbf {D} _ {n}) < C _ {5} \exp \left(- C _ {6} n ^ {\frac {1}{3}}\right), \quad n \geq C _ {4}, +$$ + +with $\mathbb{P}_{\theta^*}^n$ -probability at least $1 - \exp(-C_6 n^{1/3})$ . + +Remark A.2. As discussed in Section 4, the conditional distribution of posted prices $\mathbb{Q}(\cdot \mid X)$ is parameterized by the policy. By allowing uniform sampling at a rate of $\eta_{l}$ , defined in (6), the policy constructed in (5) satisfies (A4). Strengthening assumption (A4) to the more restrictive condition $q(g\mid x)\gtrsim n^{-1}(\log n)^{1 / 2}$ for $g\in \mathcal{G}$ and $x\in \mathcal{X}$ when $\gamma \geq 1 / 3$ yields the same results as in (8) for all $\gamma \in (0,1]$ . In such a case, $\eta_{l}$ can be adjusted accordingly to satisfy this restrictive condition. However, increasing $\eta_{l}$ leads to a higher regret due to increased exploration. Therefore, imposing a weak condition, as in (A4), is essential for achieving a tight regret upper bound. For further details, see the proof in Section B.2. + +To begin with, for a given parameter $\theta = (\mathbf{S}_0,\beta)$ , the joint density $p_{\theta}$ of $(X_{t},P_{t},Y_{t})$ is expressed as: + +$$ +p _ {\theta} (x, p, y) = \{S _ {\theta} (p | x) \} ^ {y} \left\{1 - S _ {\theta} (p | x) \right\} ^ {1 - y} q (p | x) p _ {X} (x), +$$ + +for $x \in \mathcal{X}, p \in \mathcal{G}$ and $y \in \{0,1\}$ , where $S_{\theta}(p \mid x) = S_0(p)^{\exp (x^\top \beta)}$ , and $\mathcal{X}$ denotes the support of the covariate $X$ . Here, we suppress the dependency of $p_{\theta}$ on the nuisance parameters, as they do not affect the inference of $\theta$ once an independent prior is used. The log-likelihood function corresponding to $\theta \in \Theta$ for the data $\mathbf{D}_n = \{(X_t,P_t,Y_t)\}_{t=1}^n$ is given by: + +$$ +\ell_ {n} (\theta) = \sum_ {t = 1} ^ {n} \log p _ {\theta} (X _ {t}, P _ {t}, Y _ {t}). +$$ + +# A.1 Proof of Lemma A.1 + +Lemma A.3. Under the conditions of Lemma A.1, there is an exponentially consistent sequence of tests for + +$$ +H _ {0}: \theta = (\mathbf {S} _ {0} ^ {*}, \beta^ {*}), +$$ + +$$ +H _ {1}: \theta \in \left\{\left(\mathbf {S} _ {0}, \beta\right) \in \mathcal {S} _ {0} \times \mathbb {R} ^ {d}: \| \beta - \beta^ {*} \| _ {2} \geq \eta \right\} +$$ + +for any $\eta > 0$ . + +Proof. Suppose that $\gamma < 1/3$ . Let $\mathcal{T}_n$ denote the set of every disjoint pair of index sets $I_1$ and $I_2$ such that $I_1 \cup I_2 = [K]$ . Given an index set $I \subseteq [K]$ , we denote the subset of $\mathcal{G}$ corresponding to $I$ by $\mathcal{G}(I) = \{g_k \in \mathcal{G} : k \in I\}$ . For each $(I_1, I_2) \in \mathcal{T}_n$ , define $\mathcal{S}_0(I_1, I_2) := \{\mathbf{S}_0 = (S_{0,1}, \ldots, S_{0,K}) \in \mathcal{S}_0 : S_{0,i} \geq S_{0,i}^*, S_{0,j} < S_{0,j}^*\text{ for } i \in I_1\text{ and } j \in I_2\}$ . We define the quadrant $Q_{\mathbf{e}} = \{\mathbf{z} \in \mathbb{R}^d : z_j e_j > 0, \forall j = 1, \ldots, d\}$ for each $\mathbf{e} = (e_1, \ldots, e_d) \in \{-1, 1\}^d$ . For $j = 1, \ldots, d$ , let $\mathbf{e}^{j, +}, \mathbf{e}^{j, -} \in \{-1, 1\}^d$ denote the vectors where $j$ -th element is positive and negative, respectively. Consider the following two groups of hypotheses for each $(I_1, I_2) \in \mathcal{T}_n$ and $\mathbf{e}^{j, +}, \mathbf{e}^{j, -} \in \{-1, 1\}^d$ with $j = 1, \ldots, d$ , + +$$ +H _ {0}: \theta = \left(\mathbf {S} _ {0} ^ {*}, \beta^ {*}\right), \quad H _ {1}: \theta \in \Theta_ {\mathbf {e} ^ {j, -}, I _ {1}, I _ {2}} \tag {9} +$$ + +$$ +H _ {0}: \theta = \left(\mathbf {S} _ {0} ^ {*}, \beta^ {*}\right), \quad H _ {1}: \theta \in \Theta_ {\mathbf {e} ^ {j, +}, I _ {1}, I _ {2}} \tag {10} +$$ + +where $\Theta_{\mathbf{e}^{j, - },I_1,I_2} = \mathcal{S}_0(I_1,I_2)\times \{\beta \in \mathbb{R}^d:\beta_j^*\geq \beta_j + \xi ,\beta -\beta^*\in Q_{\mathbf{e}^{j, - }}\}$ $\Theta_{\mathbf{e}^{j, + },I_1,I_2} =$ $\mathcal{S}_0(I_1,I_2)\times \{\beta \in \mathbb{R}^d:\beta_j\geq \beta_j^* +\xi ,\beta -\beta^*\in Q_{\mathbf{e}^{j, + }}\}$ and $\xi = \eta /\sqrt{d}$ + +Fix an arbitrary $j = 1,\ldots ,d$ $\mathbf{e}^{j, - }$ $\mathbf{e}^{j, + }\in \{-1,1\} ^d$ and $(I_{1},I_{2})\in \mathcal{T}_{n}$ . By Lemma C.4, take a constant $\epsilon >0$ such that $\mathbb{P}(|X_{t,j}| > \epsilon) > 0$ for all $j = 1,\dots ,d$ , given data $D_{t} = (X_{t},P_{t},Y_{t})$ For the first group of hypotheses (9), define a function $\phi_1 = \max \{\phi_{1,1},\phi_{1,2}\}$ , where $\phi_{1,1}(D_t) =$ $\mathbb{1}\{X_t\in Q_{-\mathbf{e}^{j, - }}|X_{t,j}| > \epsilon ,P_t\in \mathcal{G}(I_1),Y_t = 1\}$ and $\phi_{1,2}(D_t) = \mathbb{1}\{X_t\in Q_{\mathbf{e}^{j, - }}|X_{t,j}|>$ $\epsilon ,P_t\in \mathcal{G}(I_2),Y_t = 0\}$ . Under the event $\Omega_1 = \{X_t\in Q_{-\mathbf{e}^{j, - }}|X_{t,j}| > \epsilon ,P_t\in \mathcal{G}(I_1)\}$ for any $\theta = (\mathbf{S}_0,\beta)\in \Theta_{\mathbf{e}^{j, - },I_1,I_2}$ , we have $X_{t}^{\top}\beta < X_{t}^{\top}\beta^{*} - \epsilon \xi$ . This implies $\exp (X_t^\top \beta) <$ $\exp (X_t^\top \beta^*)\exp (-\epsilon \xi)$ . Then, on the event $\Omega_1$ , we have + +$$ +S _ {0} (P _ {t}) ^ {\exp (X _ {t} ^ {\top} \beta)} \geq S _ {0} ^ {*} (P _ {t}) ^ {\exp (X _ {t} ^ {\top} \beta)} > S _ {0} ^ {*} (P _ {t}) ^ {\exp (X _ {t} ^ {\top} \beta^ {*}) \exp (- \epsilon \xi)} > S _ {0} ^ {*} (P _ {t}) ^ {\exp (X _ {t} ^ {\top} \beta^ {*})} + \Delta_ {1}, +$$ + +where the last inequality holds because of the mean value theorem and assumptions (A1), (A2), and (A5), with a positive constant $\Delta_{1}$ depending on $M_1, M_2, L, B, \epsilon$ and $\xi$ . Let $q_n = n^{-(1 + \gamma)/2} (\log n)^{1/2}$ . By assumption (A4), we have $q(p \mid x) \gtrsim q_n$ for all $x \in \mathcal{X}$ and $p \in \mathcal{G}$ when $\gamma < 1/3$ . Then, for any $\theta = (\mathbf{S}_0, \beta) \in \Theta_{\mathbf{e}^{j, -, I_1, I_2}}$ , we have + +$$ +\begin{array}{l} \mathbb {E} _ {\theta} \left[ \phi_ {1, 1} (D _ {t}) \right] = \mathbb {E} _ {X _ {t}, P _ {t}} \left[ S _ {0} (P _ {t}) ^ {\exp (X _ {t} ^ {\top} \beta)} \mathbb {1} \{\Omega_ {1} \} \right] \\ > \mathbb {E} _ {X _ {t}, P _ {t}} \left[ S _ {0} ^ {*} (P _ {t}) ^ {\exp (X _ {t} ^ {\top} \beta^ {*})} \mathbb {1} \{\Omega_ {1} \} \right] + \mathbb {E} _ {X _ {t}, P _ {t}} \left[ \Delta_ {1} \mathbb {1} \{\Omega_ {1} \} \right] \\ \geq \mathbb {E} _ {\theta^ {*}} \left[ \phi_ {1, 1} (D _ {t}) \right] + C _ {1} \Delta_ {1} | I _ {1} | q _ {n}, \\ \end{array} +$$ + +where $C_1$ be a positive constant depending on $\mathbb{P}_X$ and $\epsilon$ . Similarly, under the event $\Omega_2 = \{X_t \in Q_{\mathbf{e}^{j}, -}, |X_{t,j}| > \epsilon, P_t \in \mathcal{G}(I_2)\}$ , for any $\theta = (\mathbf{S}_0, \beta) \in \Theta_{\mathbf{e}^{j, -}, I_1, I_2}$ , we have $\exp(X_t^\top \beta) > \exp(X_t^\top \beta^*) \exp(\epsilon\xi)$ . Then, on the event $\Omega_2$ , we have $S_0(P_t)^{\exp(X_t^\top \beta)} < S_0^*(P_t)^{\exp(X_t^\top \beta)} < S_0^*(P_t)^{\exp(X_t^\top \beta^*)} \exp(\epsilon\xi) < S_0^*(P_t)^{\exp(X_t^\top \beta^*)} - \Delta_2$ , where the last inequality holds because of the mean value theorem and assumptions (A1), (A2), and (A5), with a positive constant $\Delta_2$ depending on $M_1, M_2, L, B, \epsilon$ and $\xi$ . Then, for any $\theta = (\mathbf{S}_0, \beta) \in \Theta_{\mathbf{e}^{j, -}, I_1, I_2}$ , we have + +$$ +\begin{array}{l} \mathbb {E} _ {\theta} \left[ \phi_ {1, 2} (D _ {t}) \right] = \mathbb {E} _ {X _ {t}, P _ {t}} \left[ \left(1 - S _ {0} (P _ {t}) ^ {\exp (X _ {t} ^ {\top} \beta)}\right) \mathbb {1} \{\Omega_ {2} \} \right] \\ > \mathbb {E} _ {X _ {t}, P _ {t}} \left[ \left(1 - S _ {0} ^ {*} (P _ {t}) ^ {\exp (X _ {t} ^ {\top} \beta^ {*})}\right) \mathbb {1} \{\Omega_ {2} \} \right] + \mathbb {E} _ {X _ {t}, P _ {t}} \left[ \Delta_ {2} \mathbb {1} \{\Omega_ {2} \} \right] \\ \geq \mathbb {E} _ {\theta^ {*}} \left[ \phi_ {1, 2} (D _ {t}) \right] + C _ {2} \Delta_ {2} | I _ {2} | q _ {n}, \\ \end{array} +$$ + +where $C_2$ be a positive constant depending on $\mathbb{P}_X$ and $\epsilon$ . Combining the last two displays, we have + +$$ +\begin{array}{l} \mathbb {E} _ {\theta} \left[ \phi_ {1} (D _ {t}) \right] = \mathbb {E} _ {\theta} \left[ \phi_ {1, 1} (D _ {t}) \right] + \mathbb {E} _ {\theta} \left[ \phi_ {1, 1} (D _ {t}) \right] \\ > \mathbb {E} _ {\theta^ {*}} \left[ \phi_ {1, 1} (D _ {t}) \right] + \mathbb {E} _ {\theta^ {*}} \left[ \phi_ {1, 2} (D _ {t}) \right] + \min \left\{C _ {1} \Delta_ {1}, C _ {2} \Delta_ {2} \right\} \left(| I _ {1} | + | I _ {2} |\right) q _ {n} \tag {11} \\ > \mathbb {E} _ {\theta^ {*}} \left[ \phi_ {1} (D _ {t}) \right] + C _ {3} n ^ {\gamma} q _ {n}, \\ \end{array} +$$ + +where $C_3$ be a positive constant depending on $C_1, C_2, \Delta_1, \Delta_2, p_{\min}, p_{\max}$ and $\kappa$ . Define tests as follows: + +$$ +\Phi_ {\mathbf {e} ^ {j, -, I _ {1}, I _ {2}}} (\mathbf {D} _ {n}) := \mathbb {1} \left\{\sum_ {t = 1} ^ {n} \phi_ {1} (D _ {t}) > \sum_ {t = 1} ^ {n} \left(\mathbb {E} _ {\theta^ {*}} \left[ \phi_ {1} (D _ {t}) \right] + \mathbb {E} _ {\theta} \left[ \phi_ {1} (D _ {t}) \right]\right) / 2 \right\}. +$$ + +Then, we have + +$$ +\begin{array}{l} \mathbb {E} _ {\theta^ {*}} ^ {n} \left[ \Phi_ {\mathbf {e} ^ {j, -}, I _ {1}, I _ {2}} (\mathbf {D} _ {n}) \right] = \mathbb {P} _ {\theta^ {*}} ^ {n} \left(\sum_ {t = 1} ^ {n} \left(\phi_ {1} (D _ {t}) - \mathbb {E} _ {\theta^ {*}} \left[ \phi_ {1} (D _ {t}) \right]\right) > \sum_ {t = 1} ^ {n} \left(\mathbb {E} _ {\theta} \left[ \phi_ {1} (D _ {t}) \right] - \mathbb {E} _ {\theta^ {*}} \left[ \phi_ {1} (D _ {t}) \right]\right) / 2\right) \\ \leq \mathbb {P} _ {\theta^ {*}} ^ {n} \left(\sum_ {t = 1} ^ {n} \left(\phi_ {1} (D _ {t}) - \mathbb {E} _ {\theta^ {*}} \left[ \phi_ {1} (D _ {t}) \right]\right) > n \left(C _ {3} n ^ {\gamma} q _ {n}\right) / 2\right) \\ \leq \exp \left(- \frac {C _ {3} ^ {2} n ^ {1 + 2 \gamma} q _ {n} ^ {2}}{2}\right), \\ \end{array} +$$ + +where the first inequality holds by (11) and the last inequality holds by Hoeffding's inequality. On the other hand, applying Hoeffding's inequality to $1 - \phi_{1}(D_{t})$ , + +$$ +\begin{array}{l} \sup _ {\theta \in \Theta_ {\mathbf {e} ^ {j, -}, I _ {1}, I _ {2}}} \mathbb {E} _ {\theta} ^ {n} \left[ 1 - \Phi_ {\mathbf {e} ^ {j, -}, I _ {1}, I _ {2}} (\mathbf {D} _ {n}) \right] \\ = \sup _ {\theta \in \Theta_ {\mathbf {e} ^ {j, -}, I _ {1}, I _ {2}}} \mathbb {P} _ {\theta} ^ {n} \left(\sum_ {t = 1} ^ {n} \left(\left(1 - \phi_ {1} (D _ {t})\right) - \left(1 - \mathbb {E} _ {\theta} \left[ \phi_ {1} (D _ {t}) \right]\right)\right) \geq \sum_ {t = 1} ^ {n} \left(\mathbb {E} _ {\theta} \left[ \phi_ {1} (D _ {t}) \right] - \mathbb {E} _ {\theta^ {*}} \left[ \phi_ {1} (D _ {t}) \right]\right) / 2\right) \\ \leq \sup _ {\theta \in \Theta_ {\mathbf {e} ^ {j, -}, I _ {1}, I _ {2}}} \mathbb {P} _ {\theta} ^ {n} \left(\sum_ {t = 1} ^ {n} \left(\left(1 - \phi_ {1} (D _ {t})\right) - \left(1 - \mathbb {E} _ {\theta} \left[ \phi_ {1} (D _ {t}) \right]\right)\right) \geq n \left(C _ {3} n ^ {\gamma} q _ {n}\right) / 2\right) \\ \leq \exp \left(- \frac {C _ {3} ^ {2} n ^ {1 + 2 \gamma} q _ {n} ^ {2}}{2}\right), \\ \end{array} +$$ + +where the first inequality holds by (11). + +The construction of tests for the second group of hypotheses (10) is similar. Define the tests as follows: + +$$ +\Phi_ {\mathbf {e} ^ {j, +}, I _ {1}, I _ {2}} (\mathbf {D} _ {n}) := \mathbb {1} \left\{\sum_ {t = 1} ^ {n} \phi_ {2} (D _ {t}) > \sum_ {t = 1} ^ {n} \left(\mathbb {E} _ {\theta^ {*}} [ \phi_ {2} (D _ {t}) ] + \mathbb {E} _ {\theta} [ \phi_ {2} (D _ {t}) ]\right) / 2 \right\}, +$$ + +where $\phi_{2} = \max \{\phi_{2,1},\phi_{2,2}\}$ is a function with $\phi_{2,1}(D_t) = \mathbb{1}\{X_t\in Q_{-\mathbf{e}^{j, + }}|X_{t,j}| > \epsilon ,P_t\in$ $\mathcal{G}(I_1),Y_t = 1\}$ and $\phi_{2,2}(D_t) = \mathbb{1}\{X_t\in Q_{\mathbf{e}^{j, + }}|X_{t,j}| > \epsilon ,P_t\in \mathcal{G}(I_2),Y_t = 0\}$ . Similarly, we see that + +$$ +\mathbb {E} _ {\boldsymbol {\theta} ^ {*}} ^ {n} \left[ \Phi_ {\mathbf {e} ^ {j, +}, I _ {1}, I _ {2}} (\mathbf {D} _ {n}) \right] \leq \exp \left(- \frac {C _ {4} ^ {2} n ^ {1 + 2 \gamma} q _ {n} ^ {2}}{2}\right), +$$ + +$$ +\sup _ {\theta \in \Theta_ {\mathbf {e} ^ {j, -}, I _ {1}, I _ {2}}} \mathbb {E} _ {\theta} ^ {n} \left[ 1 - \Phi_ {\mathbf {e} ^ {j, -}, I _ {1}, I _ {2}} (\mathbf {D} _ {n}) \right] \leq \exp \left(- \frac {C _ {4} ^ {2} n ^ {1 + 2 \gamma} q _ {n} ^ {2}}{2}\right), +$$ + +where $C_4$ be a positive constant. + +Note that the union of the sets in the alternative hypotheses (9) and (10) for all $(I_1, I_2) \in \mathcal{T}_n$ and $\mathbf{e}^{j, +}, \mathbf{e}^{j, -} \in \{-1, 1\}^d$ with $j = 1, \ldots, d$ contains $\Theta_\eta := \{(\mathbf{S}_0, \beta) \in \mathcal{S}_0 \times \mathbb{R}^d : \| \beta - \beta^* \|_2 \geq \eta\}$ . We set $\Phi_n := \max_{(I_1, I_2) \in \mathcal{T}_n, \mathbf{e}^{j, -}, \mathbf{e}^{j, +} \in \{-1, 1\}^d, j \in [d]} \{\Phi_{\mathbf{e}^{j, -}, I_1, I_2} \vee \Phi_{\mathbf{e}^{j, +}, I_1, I_2}\}$ , then we have + +$$ +\begin{array}{l} \mathbb {E} _ {\boldsymbol {\theta} ^ {*}} ^ {n} \left[ \Phi_ {n} (\mathbf {D} _ {n}) \right] \leq d 2 ^ {d} 2 ^ {K} \exp \left(- C _ {5} n ^ {1 + 2 \gamma} q _ {n} ^ {2}\right) \\ \leq \exp \left(C _ {6} d \vee n ^ {\gamma} - C _ {5} n ^ {1 + 2 \gamma} q _ {n} ^ {2}\right), \\ \sup _ {\theta \in \Theta_ {\eta}} \mathbb {E} _ {\theta} ^ {n} [ 1 - \Phi_ {n} (\mathbf {D} _ {n}) ] \leq \exp \left(- C _ {5} n ^ {1 + 2 \gamma} q _ {n} ^ {2}\right), \\ \end{array} +$$ + +where $C_5 = \min \{C_3^2 / 2, C_4^2 / 2\}$ and $C_6$ be a positive constant depending on $p_{\mathrm{min}}, p_{\mathrm{max}}$ and $\kappa$ . Then, for fixed $d$ , by the definition of $q_n$ , we have $\mathbb{E}_{\theta^*}^n[\Phi_n(\mathbf{D}_n)] \to 0$ and $\sup_{\theta \in \Theta_\eta} \mathbb{E}_\theta^n [1 - \Phi_n(\mathbf{D}_n)] \to 0$ as $n \to \infty$ . By Lemma D.11 of [16], there exist tests $\Psi_n$ and a constant $C_7 > 0$ such that $\mathbb{E}_{\theta^*}^n[\Psi_n(\mathbf{D}_n)] \leq \exp(-C_7 n)$ and $\sup_{\theta \in \Theta_\eta} \mathbb{E}_\theta^n [1 - \Psi_n(\mathbf{D}_n)] \leq \exp(-C_7 n)$ . + +Suppose that $\gamma \geq 1/3$ . Recall the grid support $\mathcal{G} = \{g_k : k = 1, \dots, K\}$ , where each grid point $g_k$ is defined as $g_k = p_{\min} + k\delta$ with $\delta = \kappa n^{-\gamma}$ for some constant $\kappa > 0$ . Let $\epsilon_n = n^{-1/3}$ and $J = \lceil (p_{\max} - p_{\min}) / (\kappa \epsilon_n) \rceil$ . Define $(k_1, \dots, k_J)$ as a subsequence of $[K]$ such that $p_{\min} + (j - 1)\kappa \epsilon_n < g_{k_j} \leq p_{\min} + (j - 1)\kappa \epsilon_n + \delta$ for $j = 1, \dots, J - 1$ , and set $k_J = K$ . + +Let $\mathcal{T}_J$ denote the set of every disjoint pair of sets $I_1'$ and $I_2'$ such that $I_1' \cup I_2' = [J]$ . For each $(I_1', I_2') \in \mathcal{T}_J$ , define + +$\mathcal{S}_0(I_1',I_2') = \{\mathbf{S}_0 = (S_{0,1},\ldots ,S_{0,K})\in \mathcal{S}_0:S_{0,k_i}\geq S_{0,k_i}^*,S_{0,k_j} < S_{0,k_j}^*\text{for} i\in I_1'\text{and} j\in I_2'\}$ + +Consider the following two groups of hypotheses for each $(I_1', I_2') \in \mathcal{T}_J$ and $\mathbf{e}^{j, +}, \mathbf{e}^{j, -} \in \{-1, 1\}^d$ with $j = 1, \ldots, d$ , + +$$ +H _ {0}: \theta = \left(\mathbf {S} _ {0} ^ {*}, \beta^ {*}\right), \quad H _ {1}: \theta \in \Theta_ {\mathbf {e} ^ {j, -}, I _ {1} ^ {\prime}, I _ {2} ^ {\prime}} \tag {12} +$$ + +$$ +H _ {0}: \theta = \left(\mathbf {S} _ {0} ^ {*}, \beta^ {*}\right), \quad H _ {1}: \theta \in \Theta_ {\mathbf {e} ^ {j, +}, I _ {1} ^ {\prime}, I _ {2} ^ {\prime}}, \tag {13} +$$ + +where $\Theta_{\mathbf{e}^{j, - },I_1',I_2'} = \mathcal{S}_0(I_1',I_2')\times \{\beta \in \mathbb{R}^d:\beta_j^*\geq \beta_j + \xi ,\beta -\beta^*\in Q_{\mathbf{e}^{j, - }}\}$ $\Theta_{\mathbf{e}^{j, + },I_1',I_2'} =$ $S_0(I_1',I_2')\times \{\beta \in \mathbb{R}^d:\beta_j\geq \beta_j^* +\xi ,\beta -\beta^*\in Q_{\mathbf{e}^{j, + }}\}$ and $\xi = \eta /\sqrt{d}$ + +Fix $j = 1, \ldots, d$ , $\mathbf{e}^{j, -}, \mathbf{e}^{j, +} \in \{-1, 1\}^d$ and $(I_1', I_2') \in \mathcal{T}_J$ . Define the index set between $k_i$ and $k_{i+1}$ as $I_i = \{k \in [K] : k_i \leq k \leq k_{i+1}\}$ for $i = 1, \ldots, J-1$ . We define partitions $\mathcal{I}_1, \mathcal{I}_2, \mathcal{I}_3$ and $\mathcal{I}_4$ of set $\{I_1, \ldots, I_{J-1}\}$ by + +$$ +\mathcal {I} _ {1} = \left\{I _ {i}, i = 1, \dots , J - 1: S _ {0, k _ {i}} \geq S _ {0, k _ {i}} ^ {*}, S _ {0, k _ {i + 1}} \geq S _ {0, k _ {i + 1}} ^ {*} \right\}, +$$ + +$$ +\mathcal {I} _ {2} = \left\{I _ {i}, i = 1, \dots , J - 1: S _ {0, k _ {i}} < S _ {0, k _ {i}} ^ {*}, S _ {0, k _ {i + 1}} < S _ {0, k _ {i + 1}} ^ {*} \right\}, +$$ + +$$ +\mathcal {I} _ {3} = \left\{I _ {i}, i = 1, \dots , J - 1: S _ {0, k _ {i}} < S _ {0, k _ {i}} ^ {*}, S _ {0, k _ {i + 1}} \geq S _ {0, k _ {i + 1}} ^ {*} \right\}, +$$ + +$$ +\mathcal {I} _ {4} = \left\{I _ {i}, i = 1, \dots , J - 1: S _ {0, k _ {i}} \geq S _ {0, k _ {i}} ^ {*}, S _ {0, k _ {i + 1}} < S _ {0, k _ {i + 1}} ^ {*} \right\}. +$$ + +Note that for any $I \in \mathcal{I}_4$ , there exists a unique $k' \in I$ such that $S_{0,k'} \geq S_{0,k'}^*$ and $S_{0,k'+1} < S_{0,k'+1}^*$ . Thus, given $I \in \mathcal{I}_4$ , we can define $\overline{I} = \{k \in I : k \leq k'\}$ and $\underline{I} = \{k \in I : k > k'\}$ . For the first group of hypotheses (12), we define a function $\phi_3 = \max \{\phi_{3,1}, \phi_{3,2}, \phi_{3,3}, \phi_{3,4}, \phi_{3,5}\}$ , where + +$$ +\phi_ {3, 1} (D _ {t}) = \mathbb {1} \left\{X _ {t} \in Q _ {- \mathbf {e} ^ {j, -}}, | X _ {t, j} | > \epsilon , P _ {t} \in \bigcup_ {I \in \mathcal {I} _ {1}} \mathcal {G} (I), Y _ {t} = 1 \right\}, +$$ + +$$ +\phi_ {3, 2} (D _ {t}) = \mathbb {1} \{X _ {t} \in Q _ {\mathbf {e} ^ {j, -}}, | X _ {t, j} | > \epsilon , P _ {t} \in \bigcup_ {I \in \mathcal {I} _ {2}} \mathcal {G} (I), Y _ {t} = 0 \}, +$$ + +$$ +\phi_ {3, 3} (D _ {t}) = \mathbb {1} \{X _ {t} \in Q _ {- {\bf e} ^ {j, -}}, | X _ {t, j} | > \epsilon , P _ {t} \in \bigcup_ {I \in \mathcal {I} _ {3}} \mathcal {G} (I), Y _ {t} = 1 \}, +$$ + +$$ +\phi_ {3, 4} (D _ {t}) = \mathbb {1} \{X _ {t} \in Q _ {- \mathbf {e} ^ {j, -}}, | X _ {t, j} | > \epsilon , P _ {t} \in \bigcup_ {I \in \mathcal {I} _ {4}} \mathcal {G} (\bar {I}), Y _ {t} = 1 \}, +$$ + +$$ +\phi_ {3, 5} (D _ {t}) = \mathbb {1} \{X _ {t} \in Q _ {\mathbf {e} ^ {j, -}}, | X _ {t, j} | > \epsilon , P _ {t} \in \bigcup_ {I \in \mathcal {I} _ {4}} \mathcal {G} (\underline {{I}}), Y _ {t} = 0 \}. +$$ + +Note that under the event $\Omega_{3,1} \coloneqq \{X_t \in Q_{-\mathbf{e}^{j, -}}, |X_{t,j}| > \epsilon, P_t \in \bigcup_{I \in \mathcal{I}_1} \mathcal{G}(I)\}$ , for $\theta = (\mathbf{S}_0, \beta) \in \Theta_{\mathbf{e}^{j, -}, I_1', I_2'}$ , we have $\exp(X_t^\top \beta) < \exp(X_t^\top \beta^*) \exp(-\epsilon\xi)$ . For any $I_i \in \mathcal{I}_1$ and $k \in I_i$ , we have + +$$ +\begin{array}{l} S _ {0, k} - S _ {0, k} ^ {*} \geq S _ {0, k _ {i + 1}} - S _ {0, k _ {i + 1}} ^ {*} + S _ {0, k _ {i + 1}} ^ {*} - S _ {0, k} ^ {*} \\ \geq - L _ {0} \left(\kappa \epsilon_ {n} + \delta\right) \\ \geq - 2 L _ {0} \kappa \epsilon_ {n}, \\ \end{array} +$$ + +where the second inequality holds by the definition of $\mathcal{I}_1$ and $I_{i}$ , and the last inequality holds because $\delta \leq \epsilon_{n}$ for $\gamma \geq 1 / 3$ . Then, on the event $\Omega_{3,1}$ , we have + +$$ +\begin{array}{l} S _ {0} \left(P _ {t}\right) ^ {\exp \left(X _ {t} ^ {T} \beta\right)} > S _ {0} \left(P _ {t}\right) ^ {\exp \left(X _ {t} ^ {T} \beta^ {*}\right) \exp (- \epsilon \xi)} \\ \geq \left(S _ {0} ^ {*} (P _ {t}) - 2 L _ {0} \kappa \epsilon_ {n}\right) ^ {\exp (X _ {t} ^ {T} \beta^ {*}) \exp (- \epsilon \xi)} \\ > S _ {0} ^ {*} (P _ {t}) ^ {\exp (X _ {t} ^ {T} \beta^ {*}) \exp (- \epsilon \xi)} - C _ {8} \epsilon_ {n} \\ > S _ {0} ^ {*} \left(P _ {t}\right) ^ {\exp \left(X _ {t} ^ {T} \beta^ {*}\right)} + \Delta_ {1} - C _ {8} \epsilon_ {n} \\ > S _ {0} ^ {*} (P _ {t}) ^ {\exp \left(X _ {t} ^ {T} \beta^ {*}\right)} + \Delta_ {1} / 2, \\ \end{array} +$$ + +where the first inequality holds because $\exp (X_t^\top \beta) < \exp (X_t^\top \beta^*)\exp (-\epsilon \xi)$ , the second inequality holds by the preceding display, the third and fourth inequality holds because of the mean value theorem and assumptions (A1), (A2), and (A5), and the last inequality holds for sufficiently large $n$ such that $\epsilon_{n} < \Delta_{1} / (2C_{8})$ . Let $q_n' = n^{-\gamma -1 / 3}(\log n)^{1 / 2}$ . By assumption (A4), we have $q(p\mid x)\gtrsim q_n'$ for all $x\in \mathcal{X}$ and $p\in \mathcal{G}$ when $\gamma \geq 1 / 3$ . + +Then, for any $\theta = (\mathbf{S}_0,\beta)\in \Theta_{\mathbf{e}^{j, - },I_1',I_2'}$ , we have + +$$ +\begin{array}{l} \mathbb {E} _ {\theta} [ \phi_ {3, 1} (D _ {t}) ] = \mathbb {E} _ {X _ {t}, P _ {t}} \left[ S _ {0} (P _ {t}) ^ {\exp (X _ {t} ^ {T} \beta)} \mathbb {1} \{\Omega_ {3, 1} \} \right] \\ > \mathbb {E} _ {X _ {t}, P _ {t}} \left[ S _ {0} ^ {*} (P _ {t}) ^ {\exp (X _ {t} ^ {T} \beta^ {*})} \mathbb {1} \{\Omega_ {3, 1} \} \right] + \Delta_ {1} / 2 \cdot \mathbb {E} _ {X _ {t}, P _ {t}} [ \mathbb {1} \{\Omega_ {3, 1} \} ] \\ \geq \mathbb {E} _ {\theta^ {*}} \left[ \phi_ {3, 1} \left(D _ {t}\right) \right] + C _ {9} | \mathcal {I} _ {1} | K q _ {n} ^ {\prime} / J, \\ \end{array} +$$ + +where the second inequality holds by the preceding display, and the last inequality holds with a positive constant $C_9$ because $|I_i| \geq K / J$ for all $i = 1, \dots, J - 1$ . Similarly, there exist positive constants $C_{10}, C_{11}, C_{12}$ and $C_{13}$ such that + +$$ +\mathbb {E} _ {\theta} \left[ \phi_ {3, 2} (D _ {t}) \right] > \mathbb {E} _ {\theta^ {*}} \left[ \phi_ {3, 2} (D _ {t}) \right] + C _ {1 0} | \mathcal {I} _ {2} | K q _ {n} ^ {\prime} / J, +$$ + +$$ +\mathbb {E} _ {\theta} \left[ \phi_ {3, 3} (D _ {t}) \right] > \mathbb {E} _ {\theta^ {*}} \left[ \phi_ {3, 3} (D _ {t}) \right] + C _ {1 1} \left| \mathcal {I} _ {3} \right| K q _ {n} ^ {\prime} / J, +$$ + +$$ +\mathbb {E} _ {\theta} \left[ \phi_ {3, 4} (D _ {t}) \right] > \mathbb {E} _ {\theta^ {*}} \left[ \phi_ {3, 4} (D _ {t}) \right] + C _ {1 2} \sum_ {I \in \mathcal {I} _ {4}} | \bar {I} | q _ {n} ^ {\prime}, +$$ + +$$ +\mathbb {E} _ {\theta} \left[ \phi_ {3, 5} (D _ {t}) \right] > \mathbb {E} _ {\theta^ {*}} \left[ \phi_ {3, 5} (D _ {t}) \right] + C _ {1 3} \sum_ {I \in \mathcal {I} _ {4}} | \underline {{I}} | q _ {n} ^ {\prime}. +$$ + +Combining the last two displays, we have + +$$ +\begin{array}{l} \mathbb {E} _ {\theta} \left[ \phi_ {3} \left(D _ {t}\right) \right] = \sum_ {s = 1} ^ {5} \mathbb {E} _ {\theta} \left[ \phi_ {3, s} \left(D _ {t}\right) \right] \\ > \sum_ {s = 1} ^ {5} \mathbb {E} _ {\theta^ {*}} [ \phi_ {3, s} (D _ {t}) ] + C _ {1 4} \left(\left(| \mathcal {I} _ {1} | + | \mathcal {I} _ {2} | + | \mathcal {I} _ {3} |\right) K q _ {n} ^ {\prime} / J + \sum_ {I \in \mathcal {I} _ {4}} \left(| \bar {I} | + | \underline {{I}} |\right) q _ {n} ^ {\prime}\right) \\ > \sum_ {s = 1} ^ {5} \mathbb {E} _ {\theta^ {*}} \left[ \phi_ {3, s} (D _ {t}) \right] + C _ {1 4} \left(| \mathcal {I} _ {1} | + | \mathcal {I} _ {2} | + | \mathcal {I} _ {3} | + | \mathcal {I} _ {4} |\right) K q _ {n} ^ {\prime} / J \\ > \mathbb {E} _ {\theta^ {*}} \left[ \phi_ {3} (D _ {t}) \right] + C _ {1 5} n ^ {\gamma} q _ {n} ^ {\prime}, \tag {14} \\ \end{array} +$$ + +where the second inequality holds because $|\overline{I}| + |\underline{I}| = |I| \geq K / J$ for all $I \in \mathcal{I}_4$ , and the last inequality holds because $|\mathcal{I}_1| + |\mathcal{I}_2| + |\mathcal{I}_3| + |\mathcal{I}_4| = J$ . Define tests as follows: + +$$ +\Phi_ {\mathbf {e} ^ {j, -}, I _ {1} ^ {\prime}, I _ {2} ^ {\prime}} (\mathbf {D} _ {n}) := \mathbb {1} \left\{\sum_ {t = 1} ^ {n} \phi_ {3} (D _ {t}) > \sum_ {t = 1} ^ {n} (\mathbb {E} _ {\theta^ {*}} [ \phi_ {3} (D _ {t}) ] + \mathbb {E} _ {\theta} [ \phi_ {3} (D _ {t}) ]) / 2 \right\}. +$$ + +By Hoeffding's inequality and (14), we have + +$$ +\begin{array}{l} \mathbb {E} _ {\theta^ {*}} ^ {n} \left[ \Phi_ {\mathbf {e} ^ {j, -}, I _ {1} ^ {\prime}, I _ {2} ^ {\prime}} (\mathbf {D} _ {n}) \right] \leq \mathbb {P} _ {\theta^ {*}} ^ {n} \left(\sum_ {t = 1} ^ {n} \left(\phi_ {3} (D _ {t}) - \mathbb {E} _ {\theta^ {*}} \left[ \phi_ {3} (D _ {t}) \right]\right) > n \left(C _ {1 5} n ^ {\gamma} q _ {n} ^ {\prime}\right) / 2\right) \\ \leq \exp \left(- \frac {C _ {1 5} ^ {\prime 2} n ^ {1 + 2 \gamma} q _ {n} ^ {\prime 2}}{2}\right), \\ \sup _ {\theta \in \Theta_ {\mathbf {e} ^ {j, -}, I _ {1} ^ {\prime}, I _ {2} ^ {\prime}}} \mathbb {E} _ {\theta^ {*}} ^ {n} \left[ 1 - \Phi_ {\mathbf {e} ^ {j, -}, I _ {1} ^ {\prime}, I _ {2} ^ {\prime}} (\mathbf {D} _ {n}) \right] \leq \exp \left(- \frac {C _ {1 5} ^ {2} n ^ {1 + 2 \gamma} q _ {n} ^ {\prime 2}}{2}\right). \\ \end{array} +$$ + +The construction of tests for the second group of hypotheses (13) is similar. Define the tests as follows: + +$$ +\Phi_ {\mathbf {e} ^ {j, +}, I _ {1} ^ {\prime}, I _ {2} ^ {\prime}} (\mathbf {D} _ {n}) := \mathbb {1} \left\{\sum_ {t = 1} ^ {n} \phi_ {4} (D _ {t}) > \sum_ {t = 1} ^ {n} (\mathbb {E} _ {\theta^ {*}} [ \phi_ {4} (D _ {t}) ] + \mathbb {E} _ {\theta} [ \phi_ {4} (D _ {t}) ]) / 2 \right\}, +$$ + +where $\phi_4 = \max \{\phi_{4,1},\phi_{4,2},\phi_{4,3},\phi_{4,4},\phi_{4,5}\}$ is a function with + +$$ +\phi_ {4, 1} (D _ {t}) = \mathbb {1} \{X _ {t} \in Q _ {- \mathbf {e} ^ {j, +}}, | X _ {t, j} | > \epsilon , P _ {t} \in \bigcup_ {I \in \mathcal {I} _ {1}} \mathcal {G} (I), Y _ {t} = 1 \}, +$$ + +$$ +\phi_ {4, 2} (D _ {t}) = \mathbb {1} \{X _ {t} \in Q _ {\mathbf {e} ^ {j, +}}, | X _ {t, j} | > \epsilon , P _ {t} \in \bigcup_ {I \in \mathcal {I} _ {2}} \mathcal {G} (I), Y _ {t} = 0 \}, +$$ + +$$ +\phi_ {4, 3} (D _ {t}) = \mathbb {1} \{X _ {t} \in Q _ {- \mathbf {e} ^ {j, +}}, | X _ {t, j} | > \epsilon , P _ {t} \in \bigcup_ {I \in \mathcal {I} _ {3}} \mathcal {G} (I), Y _ {t} = 1 \}, +$$ + +$$ +\phi_ {4, 4} (D _ {t}) = \mathbb {1} \{X _ {t} \in Q _ {- \mathbf {e} ^ {j, +}}, | X _ {t, j} | > \epsilon , P _ {t} \in \bigcup_ {I \in \mathcal {I} _ {4}} \mathcal {G} (\bar {I}), Y _ {t} = 1 \}, +$$ + +$$ +\phi_ {4, 5} (D _ {t}) = \mathbb {1} \{X _ {t} \in Q _ {\mathbf {e} ^ {j, +}}, | X _ {t, j} | > \epsilon , P _ {t} \in \bigcup_ {I \in \mathcal {I} _ {4}} \mathcal {G} (\underline {{I}}), Y _ {t} = 0 \}. +$$ + +Similarly, we see that + +$$ +\mathbb {E} _ {\boldsymbol {\theta} ^ {*}} ^ {n} \left[ \Phi_ {\mathbf {e} ^ {j, +}, I _ {1} ^ {\prime}, I _ {2} ^ {\prime}} (\mathbf {D} _ {n}) \right] \leq \exp \left(- \frac {C _ {1 6} ^ {2} n ^ {1 + 2 \gamma} q _ {n} ^ {\prime 2}}{2}\right), +$$ + +$$ +\sup _ {\theta \in \Theta_ {\mathbf {e} ^ {j, +}, I _ {1} ^ {\prime}, I _ {2} ^ {\prime}}} \mathbb {E} _ {\theta} ^ {n} \left[ 1 - \Phi_ {\mathbf {e} ^ {j, +}, I _ {1} ^ {\prime}, I _ {2} ^ {\prime}} (\mathbf {D} _ {n}) \right] \leq \exp \left(- \frac {C _ {1 6} ^ {2} n ^ {1 + 2 \gamma} q _ {n} ^ {\prime 2}}{2}\right), +$$ + +where $C_{16}$ be a positive constant. + +Note that the union of the sets in the alternative hypotheses (12) and (13) for all $(I_1', I_2') \in \mathcal{T}_J$ and $\mathbf{e}^{j, +}, \mathbf{e}^{j, -} \in \{-1, 1\}^d$ with $j = 1, \ldots, d$ contains $\Theta_\eta := \{(\mathbf{S}_0, \beta) \in S_0 \times \mathbb{R}^d : \| \beta - \beta^* \|_2 \geq \eta\}$ . We set $\Phi_n' := \max_{(I_1', I_2') \in \mathcal{T}_J, \mathbf{e}^{j, -}, \mathbf{e}^{j, +} \in \{-1, 1\}^d, j \in [d]} \{\Phi_{\mathbf{e}^{j, -}, I_1', I_2'} \vee \Phi_{\mathbf{e}^{j, +}, I_1', I_2'}\}$ , then we have + +$$ +\begin{array}{l} \mathbb {E} _ {\theta^ {*}} ^ {n} \left[ \Phi_ {n} ^ {\prime} \left(\mathbf {D} _ {n}\right) \right] \leq d 2 ^ {d} 2 ^ {J} \exp \left(- C _ {1 7} n ^ {1 + 2 \gamma} q _ {n} ^ {\prime 2}\right) \\ \leq \exp \left(C _ {1 8} d \vee n ^ {\frac {1}{3}} - C _ {1 7} n ^ {1 + 2 \gamma} q _ {n} ^ {\prime 2}\right), \\ \sup _ {\theta \in \Theta_ {\eta}} \mathbb {E} _ {\theta} ^ {n} [ 1 - \Phi_ {n} ^ {\prime} (\mathbf {D} _ {n}) ] \leq \exp \left(- C _ {1 7} n ^ {1 + 2 \gamma} q _ {n} ^ {\prime 2}\right), \\ \end{array} +$$ + +where $C_{17} = \min \{C_{15}^2 / 2, C_{16}^2 / 2\}$ and $C_{18}$ be a positive constant depending on $p_{\mathrm{min}}, p_{\mathrm{max}}$ and $\kappa$ . Then, for fixed $d$ , by the definition of $q_n'$ , we have $\mathbb{E}_{\theta^*}^n[\Phi_n'(\mathbf{D}_n)] \to 0$ and $\sup_{\theta \in \Theta_\eta} \mathbb{E}_\theta^n [1 - \Phi_n'(\mathbf{D}_n)] \to 0$ as $n \to \infty$ . By Lemma D.11 of [16], there exist tests $\Psi_n'$ and a constant $C_{19} > 0$ such that $\mathbb{E}_{\theta^*}^n[\Psi_n'(\mathbf{D}_n)] \leq \exp(-C_{19} n)$ and $\sup_{\theta \in \Theta_\eta} \mathbb{E}_\theta^n [1 - \Psi_n'(\mathbf{D}_n)] \leq \exp(-C_{19} n)$ . The proof is then complete. + +Lemma A.4. Suppose that the grid resolution satisfies $\delta = \kappa n^{-\gamma}$ for $\kappa > 0$ and $\gamma \in (0,2/3)$ , and assumptions (A1)-(A5) hold. Then, there is an exponentially consistent sequence of tests for + +$$ +H _ {0}: \theta = (\mathbf {S} _ {0} ^ {*}, \beta^ {*}), +$$ + +$$ +H _ {1}: \theta \in \left\{(\mathbf {S} _ {0}, \beta) \in \mathcal {S} _ {0} \times \mathbb {R} ^ {d}: \| \mathbf {S} _ {0} - \mathbf {S} _ {0} ^ {*} \| _ {\infty} \geq \eta_ {1}, \| \beta - \beta^ {*} \| _ {2} < \eta_ {2} \right\} +$$ + +for any $\eta_1 > 0$ and sufficiently small $\eta_2 > 0$ . + +Proof. There exist constants $M_1, M_2 \in (0, 1)$ such that $M_1 \leq S_0^*(v) \leq M_2$ for any $v \in [p_{\min}, p_{\max}]$ under assumption (A5). We choose $\eta_1$ to be less than $\min \{1 - M_2, M_1\}$ to ensure that $\{\mathbf{S}_0 \in \mathcal{S}_0 : \| \mathbf{S}_0 - \mathbf{S}_0^* \|_\infty \geq \eta_1\} \neq \emptyset$ . Consider the following two groups of hypotheses for each $k \in [K]$ , + +$$ +H _ {0}: \theta = \left(\mathbf {S} _ {0} ^ {*}, \beta^ {*}\right), \quad H _ {1}: \theta \in \Theta_ {k, 1} \tag {15} +$$ + +$$ +H _ {0}: \theta = \left(\mathbf {S} _ {0} ^ {*}, \beta^ {*}\right), \quad H _ {1}: \theta \in \Theta_ {k, 2} \tag {16} +$$ + +where $\Theta_{k,1} = \{(\mathbf{S}_0,\beta)\in \mathcal{S}_0\times \mathbb{R}^d:S_{0,k}\geq S_{0,k}^* +\eta_1,\| \beta -\beta^*\| _2 < \eta_2\}$ and $\Theta_{k,2} = \{(\mathbf{S}_0,\beta)\in$ $\mathcal{S}_0\times \mathbb{R}^d:S_{0,k}\leq S_{0,k}^* -\eta_1,\| \beta -\beta^*\| _2 < \eta_2\}$ + +Fix an arbitrary $k \in [K]$ . For the first group of hypotheses (15), define a function $\phi_1(D_t) = \mathbb{1}\{P_t = g_k, Y_t = 1\}$ . For any $\beta$ such that $\| \beta - \beta^* \|_2 < \eta_2$ , by the Cauchy-Schwartz inequality and the assumption (A2), $|X_t^\top (\beta - \beta^*)| \leq \| X_t \|_2 \| \beta - \beta^* \|_2 < L\eta_2$ almost surely. This implies $\exp(X_t^\top \beta) < \exp(X_t^\top \beta^*) \exp(L\eta_2)$ . Then, for any $\theta = (\mathbf{S}_0, \beta) \in \Theta_{k,1}$ , we have + +$$ +S _ {0, k} ^ {\exp (X _ {t} ^ {\top} \beta)} > \left(S _ {0, k} ^ {*} + \eta_ {1}\right) ^ {\exp (L \eta_ {2}) \exp (X _ {t} ^ {\top} \beta^ {*})}. +$$ + +It is easy to show that there exists a positive constant $C_1$ depending on $M_1, M_2, L$ and $\eta_1$ such that $C_1 \leq \log \log((S_{0,k}^* + \eta_1/2)^{-1}) - \log \log((S_{0,k}^* + \eta_1)^{-1})$ for any $S_{0,k}^* \in [M_1, M_2]$ . If we choose a sufficiently small $\eta_2$ such that $L\eta_2 \leq C_1$ , we have $(S_{0,k}^* + \eta_1)^{\exp(L\eta_2)} \geq S_{0,k}^* + \eta_1/2$ . Combining this with the previous display, + +$$ +\begin{array}{l} S _ {0, k} ^ {\exp (X _ {t} ^ {\top} \beta)} > \left(S _ {0, k} ^ {*} + \frac {\eta_ {1}}{2}\right) ^ {\exp (X _ {t} ^ {\top} \beta^ {*})} \\ \geq S _ {0, k} ^ {*} \stackrel {\exp (X _ {t} ^ {\top} \beta^ {*})} {=} + C _ {2}, \\ \end{array} +$$ + +where the last inequality holds with a positive constant $C_2$ depending on $M_1, M_2, L, B$ and $\eta_1$ by assumptions (A1), (A2), (A5) and the mean value theorem. Then, for any $\theta = (\mathbf{S}_0, \beta) \in \Theta_{k,1}$ , we have + +$$ +\begin{array}{l} \mathbb {E} _ {\theta} \left[ \phi_ {1} (D _ {t}) \right] = \mathbb {E} _ {X _ {t}, P _ {t}} \left[ S _ {0} (P _ {t}) ^ {\exp \left(X _ {t} ^ {\top} \beta\right)} \mathbb {1} \{P _ {t} = g _ {k} \} \right] \\ > \mathbb {E} _ {X _ {t}, P _ {t}} \left[ S _ {0} ^ {*} (P _ {t}) ^ {\exp \left(X _ {t} ^ {\top} \beta^ {*}\right)} \mathbb {1} \left\{P _ {t} = g _ {k} \right\} \right] + C _ {2} q \left(g _ {k}\right) \tag {17} \\ = \mathbb {E} _ {\theta^ {*}} \left[ \phi_ {1} (D _ {t}) \right] + C _ {2} q (g _ {k}). \\ \end{array} +$$ + +In addition, for either $\theta \in \Theta_{k,1}$ or $\theta = \theta^{*}$ , we have + +$$ +\begin{array}{l} \operatorname {V a r} _ {\theta} \left(\phi_ {1} (D _ {t}) - \mathbb {E} _ {\theta} \left(\phi_ {1} (D _ {t})\right)\right) = \mathbb {E} _ {\theta} \left[ \phi_ {1} (D _ {t}) \right] \left(1 - \mathbb {E} _ {\theta} \left[ \phi_ {1} (D _ {t}) \right]\right) \\ \leq \mathbb {E} _ {\theta} [ \phi_ {1} (D _ {t}) ] \\ = \mathbb {E} _ {X _ {t}, P _ {t}} \left[ S _ {0} \left(P _ {t}\right) ^ {\exp \left(X _ {t} ^ {\top} \beta\right)} \mathbb {1} \left\{P _ {t} = g _ {k} \right\} \right] \tag {18} \\ < q (g _ {k}), \\ \end{array} +$$ + +where $\mathrm{Var}_{\theta}$ is the variance with respect to the distribution $\mathbb{P}_{\theta}$ . Define tests as follows: + +$$ +\Phi_ {k, 1} (\mathbf {D} _ {n}) := \mathbb {1} \left\{\sum_ {t = 1} ^ {n} \phi_ {1} (D _ {t}) > \sum_ {t = 1} ^ {n} \left(\mathbb {E} _ {\theta^ {*}} \left[ \phi_ {1} (D _ {t}) \right] + \mathbb {E} _ {\theta} \left[ \phi_ {1} (D _ {t}) \right]\right) / 2 \right\}. +$$ + +Then, we have + +$$ +\begin{array}{l} \mathbb {E} _ {\theta^ {*}} ^ {n} \left[ \Phi_ {k, 1} (\mathbf {D} _ {n}) \right] = \mathbb {P} _ {\theta^ {*}} ^ {n} \left(\sum_ {t = 1} ^ {n} \left(\phi_ {1} \left(D _ {t}\right) - \mathbb {E} _ {\theta^ {*}} \left[ \phi_ {1} \left(D _ {t}\right) \right]\right) > \sum_ {t = 1} ^ {n} \left(\mathbb {E} _ {\theta} \left[ \phi_ {1} \left(D _ {t}\right) \right] - \mathbb {E} _ {\theta^ {*}} \left[ \phi_ {1} \left(D _ {t}\right) \right]\right) / 2\right) \\ \leq \mathbb {P} _ {\theta^ {*}} ^ {n} \left(\sum_ {t = 1} ^ {n} \left(\phi_ {1} (D _ {t}) - \mathbb {E} _ {\theta^ {*}} \left[ \phi_ {1} (D _ {t}) \right]\right) > \frac {C _ {2}}{2} n q \left(g _ {k}\right)\right) \\ \leq \exp \left(- \frac {\left(C _ {2} ^ {2} / 8\right) n ^ {2} q \left(g _ {k}\right) ^ {2}}{\sum_ {t = 1} ^ {n} \operatorname {V a r} _ {\theta^ {*}} \left(\phi_ {1} \left(D _ {t}\right) - \mathbb {E} _ {\theta^ {*}} \left[ \phi_ {1} \left(D _ {t}\right) \right]\right) + \left(C _ {2} / 6\right) n q \left(g _ {k}\right)}\right) \\ \leq \exp (- C _ {3} n q (g _ {k})) \\ \leq \exp \left(- C _ {4} n q _ {n}\right), \tag {19} \\ \end{array} +$$ + +where the first inequality holds by (17), the second inequality holds by Bernstein inequality, the third inequality holds by (18) with a positive constant $C_3 = C_2^2 / (8(1 + C_2 / 6))$ , and the last inequality holds because $q(p) \gtrsim q_n$ with + +$$ +q _ {n} = \left\{ \begin{array}{l l} n ^ {- \frac {1 + \gamma}{2}} (\log n) ^ {\frac {1}{2}} & \quad \text {i f} \gamma < \frac {1}{3}, \\ n ^ {- \gamma - \frac {1}{3}} (\log n) ^ {\frac {1}{2}} & \quad \text {i f} \gamma \geq \frac {1}{3}, \end{array} \right. +$$ + +under the assumption (A4). On the other hand, applying (17), (18) and Bernstein inequality to $1 - \phi_{1}(D_{t})$ , we have + +$$ +\begin{array}{l} \sup _ {\theta \in \Theta_ {k, 1}} \mathbb {E} _ {\theta} ^ {n} \left[ 1 - \Phi_ {k, 1} \left(\mathbf {D} _ {n}\right) \right] \\ = \sup _ {\theta \in \Theta_ {k, 1}} \mathbb {P} _ {\theta} ^ {n} \left(\sum_ {t = 1} ^ {n} \left(\left(1 - \phi_ {1} (D _ {t})\right) - \left(1 - \mathbb {E} _ {\theta} [ \phi_ {1} (D _ {t}) ]\right)\right) \geq \sum_ {t = 1} ^ {n} \left(\mathbb {E} _ {\theta} [ \phi_ {1} (D _ {t}) ] - \mathbb {E} _ {\theta^ {*}} [ \phi_ {1} (D _ {t}) ]\right) / 2\right) \\ \leq \sup _ {\theta \in \Theta_ {k, 1}} \mathbb {P} _ {\theta} ^ {n} \left(\sum_ {t = 1} ^ {n} \left(\left(1 - \phi_ {1} (D _ {t})\right) - \left(1 - \mathbb {E} _ {\theta} \left[ \phi_ {1} (D _ {t}) \right]\right)\right) \geq \frac {C _ {2}}{2} n q \left(g _ {k}\right)\right) \\ \leq \sup _ {\theta \in \Theta_ {k, 1}} \exp \left(- \frac {\left(C _ {2} ^ {2} / 8\right) n ^ {2} q \left(g _ {k}\right) ^ {2}}{\sum_ {t = 1} ^ {n} \operatorname {V a r} _ {\theta} \left(\mathbb {E} _ {\theta} \left[ \phi_ {1} \left(D _ {t}\right) \right] - \phi_ {1} \left(D _ {t}\right)\right) + \left(C _ {2} / 6\right) n q \left(g _ {k}\right)}\right) \\ \leq \exp (- C _ {3} n q (g _ {k})) \\ \leq \exp \left(- C _ {4} n q _ {n}\right). \tag {20} \\ \end{array} +$$ + +The construction of tests for the second group of hypotheses (16) is similar. Define the tests as follows: + +$$ +\Phi_ {k, 2} (\mathbf {D} _ {n}) := \mathbb {1} \left\{\sum_ {t = 1} ^ {n} \phi_ {2} (D _ {t}) > \sum_ {t = 1} ^ {n} \left(\mathbb {E} _ {\theta^ {*}} \left[ \phi_ {2} (D _ {t}) \right] + \mathbb {E} _ {\theta} \left[ \phi_ {2} (D _ {t}) \right]\right) / 2 \right\}, +$$ + +where $\phi_2(D_t) = \mathbb{1}\{P_t = g_k, Y_t = 0\}$ is a function. Similarly, we see that there exists a positive constant $C_5$ depending on $M_1, M_2, L, B, \eta_1$ such that + +$$ +\mathbb {E} _ {\theta^ {*}} ^ {n} \left[ \Phi_ {k, 2} (\mathbf {D} _ {n}) \right] \leq \exp (- C _ {5} n q _ {n}), \quad \sup _ {\theta \in \Theta_ {k, 2}} \mathbb {E} _ {\theta} ^ {n} \left[ 1 - \Phi_ {k, 2} (\mathbf {D} _ {n}) \right] \leq \exp (- C _ {5} n q _ {n}). \tag {21} +$$ + +Note that the union of the sets in the alternative hypotheses (15) and (16) for all $k = 1,\ldots ,K$ contains $\Theta_{\eta_1,\eta_2}\coloneqq \{(\mathbf{S}_0,\beta)\in \mathcal{S}_0\times \mathbb{R}^d:\| \mathbf{S}_0 - \mathbf{S}_0^*\|_\infty \geq \eta_1,\| \beta -\beta^*\| _2 < \eta_2\}$ . We set $\Phi_n\coloneqq$ $\max_{k\in [K]}\{\Phi_{k,1}\vee \Phi_{k,2}\}$ . Combining (19), (20) and (21), we have + +$$ +\mathbb {E} _ {\boldsymbol {\theta} ^ {*}} ^ {n} \left[ \Phi_ {n} (\mathbf {D} _ {n}) \right] \leq K \exp \left(- C _ {6} n q _ {n}\right) +$$ + +$$ += \exp \left(\log K - C _ {6} n q _ {n}\right), +$$ + +$$ +\sup _ {\theta \in \Theta_ {\eta_ {1}, \eta_ {2}}} \mathbb {E} _ {\theta} ^ {n} \left[ 1 - \Phi_ {n} (\mathbf {D} _ {n}) \right] \leq \exp \left(- C _ {6} n q _ {n}\right), +$$ + +where $C_6 = \min \{C_4,C_5\}$ . By the definition of $q_{n}$ , we have $\mathbb{E}_{\theta^{*}}^{n}[\Phi_{n}(\mathbf{D}_{n})]\to 0$ and $\sup_{\theta \in \Theta_{\eta_1,\eta_2}}\mathbb{E}_\theta^n [1 - \Phi_n(\mathbf{D}_n)]\to 0$ as $n\to \infty$ when $\gamma < 2 / 3$ . By Lemma D.11 of [16], there exist tests $\Psi_{n}$ and a constant $C_7 > 0$ such that $\mathbb{E}_{\theta^*}^n [\Psi_n(\mathbf{D}_n)]\leq \exp (-C_7n)$ and $\sup_{\theta \in \Theta_{\eta_1,\eta_2}}\mathbb{E}_\theta^n [1 - \Psi_n(\mathbf{D}_n)]\leq \exp (-C_7n)$ . The proof is then complete. + +Lemma A.5. Suppose that the grid resolution satisfies $\delta = \kappa n^{-\gamma}$ for $\kappa > 0$ and $\gamma \in [1/3, 1]$ , and assumptions (A1)-(A5) hold. Let $\epsilon = n^{-1/3}$ and $J = \lceil (p_{\max} - p_{\min}) / (\kappa \epsilon) \rceil$ . Define $(k_1, \ldots, k_J)$ as a subsequence of $[K]$ such that $p_{\min} + (j - 1)\kappa \epsilon < g_{k_j} \leq p_{\min} + (j - 1)\kappa \epsilon + \delta$ for $j = 1, \ldots, J - 1$ , and set $k_J = K$ . Then, there is an exponentially consistent sequence of tests for + +$$ +H _ {0}: \theta = (\mathbf {S} _ {0} ^ {*}, \beta^ {*}), +$$ + +$$ +H _ {1}: \theta \in \left\{\left(\mathbf {S} _ {0}, \beta\right) \in \mathcal {S} _ {0} \times \mathbb {R} ^ {d}: \max _ {2 \leq j \leq J - 1} \left| S _ {0, k _ {j}} - S _ {0, k _ {j}} ^ {*} \right| \geq \eta_ {1}, \| \beta - \beta^ {*} \| _ {2} < \eta_ {2} \right\} +$$ + +for any $\eta_1 > 0$ and sufficiently small $\eta_2 > 0$ . + +Proof. We consider the following two groups of hypotheses for each $j = 2, \ldots, J - 1$ , + +$$ +H _ {0}: \theta = \left(\mathbf {S} _ {0} ^ {*}, \beta^ {*}\right), \quad H _ {1}: \theta \in \Theta_ {j, 1} \tag {22} +$$ + +$$ +H _ {0}: \theta = \left(\mathbf {S} _ {0} ^ {*}, \beta^ {*}\right), \quad H _ {1}: \theta \in \Theta_ {j, 2} \tag {23} +$$ + +where $\Theta_{j,1} = \{(\mathbf{S}_0,\beta)\in \mathcal{S}_0\times \mathbb{R}^d:S_{0,k_j}\geq S_{0,k_j}^*\} +\eta_1,\| \beta -\beta^*\| _2 < \eta_2\}$ and $\Theta_{j,2} = \{(\mathbf{S}_0,\beta)\in$ $\mathcal{S}_0\times \mathbb{R}^d:S_{0,k_j}\leq S_{0,k_j}^*\} .$ $\begin{array}{r}I_{j} = \{k\in [K]:k_{j}\leq k\leq k_{j + 1}\} \end{array}$ for $j = 1,\dots ,J - 1$ . Given an index set $I\subseteq [K]$ , we denote the subset of $\mathcal{G}$ corresponding to $I$ by $\mathcal{G}(I) = \{g_k\in \mathcal{G}:k\in I\}$ + +Fix $j = 2, \ldots, J - 1$ . For the first group of hypotheses (22), define a function $\phi_1(D_t) = \mathbb{1}\{P_t \in \mathcal{G}(I_{j-1}), Y_t = 1\}$ . For any $\theta \in \Theta_{j,1}$ and $k \in I_{j-1}$ , we have + +$$ +\begin{array}{l} S _ {0, k} - S _ {0, k} ^ {*} \geq S _ {0, k _ {j}} - S _ {0, k _ {j}} ^ {*} + S _ {0, k _ {j}} ^ {*} - S _ {0, k} ^ {*} \\ \geq \eta_ {1} - L _ {0} | g _ {k _ {j}} - g _ {k} | \\ \geq \eta_ {1} - L _ {0} (\kappa \epsilon + \delta) \\ \geq \eta_ {1} - 2 L _ {0} \kappa \epsilon \\ \geq \frac {\eta_ {1}}{2}, \\ \end{array} +$$ + +where the second inequality holds because $\theta \in \Theta_{j,1}$ and $S_0^*$ is $L_0$ -Lipschitz continuous under the assumption (A5), the third inequality holds by the definition of $(k_1,\dots ,k_J)$ , the fourth inequality holds because $\delta \leq \kappa \epsilon$ when $\gamma \geq 1 / 3$ , and the last inequality holds for sufficiently large $n$ such that $\epsilon \leq \eta_1 / (4L_0\kappa)$ . By a similar argument as the proof in Lemma A.4, for a sufficiently small $\eta_{2}$ , there exists a positive constant $C_1$ depending on $M_1,M_2,L,B$ and $\eta_{1}$ such that for $\theta = (\mathbf{S}_0,\beta)\in \Theta_{j,1}$ and $k\in I_{j - 1}$ + +$$ +S _ {0, k} ^ {\exp (X _ {t} ^ {\top} \beta)} > S _ {0, k} ^ {*} ^ {\exp (X _ {t} ^ {\top} \beta^ {*})} + C _ {1}. +$$ + +Then, for any $\theta = (\mathbf{S}_0,\beta)\in \Theta_{j,1}$ , we have + +$$ +\begin{array}{l} \mathbb {E} _ {\theta} \left[ \phi_ {1} (D _ {t}) \right] = \mathbb {E} _ {X _ {t}, P _ {t}} \left[ S _ {0} (P _ {t}) ^ {\exp \left(X _ {t} ^ {\top} \beta\right)} \mathbb {1} \{P _ {t} \in \mathcal {G} (I _ {j - 1}) \} \right] \\ > \mathbb {E} _ {X _ {t}, P _ {t}} \left[ S _ {0} ^ {*} (P _ {t}) ^ {\exp \left(X _ {t} ^ {\top} \beta^ {*}\right)} \mathbb {1} \left\{P _ {t} \in \mathcal {G} \left(I _ {j - 1}\right) \right\} \right] + C _ {1} \sum_ {k \in I _ {j - 1}} q \left(g _ {k}\right) \\ = \mathbb {E} _ {\theta^ {*}} \left[ \phi_ {1} (D _ {t}) \right] + C _ {1} \sum_ {k \in I _ {j - 1}} q (g _ {k}), \\ \end{array} +$$ + +In addition, for either $\theta \in \Theta_{j,1}$ or $\theta = \theta^{*}$ , we have + +$$ +\begin{array}{l} \operatorname {V a r} _ {\theta} \left(\phi_ {1} (D _ {t}) - \mathbb {E} _ {\theta} \left(\phi_ {1} (D _ {t})\right)\right) = \mathbb {E} _ {\theta} \left[ \phi_ {1} (D _ {t}) \right] \left(1 - \mathbb {E} _ {\theta} \left[ \phi_ {1} (D _ {t}) \right]\right) \\ \leq \mathbb {E} _ {\theta} [ \phi_ {1} (D _ {t}) ] \\ = \mathbb {E} _ {X _ {t}, P _ {t}} \left[ S _ {0} (P _ {t}) ^ {\exp \left(X _ {t} ^ {\top} \beta\right)} \mathbb {1} \left\{P _ {t} \in \mathcal {G} \left(I _ {j - 1}\right) \right\} \right] \\ < \sum_ {k \in I _ {j - 1}} q (g _ {k}). \\ \end{array} +$$ + +Define tests as follows: + +$$ +\Phi_ {j, 1} (\mathbf {D} _ {n}) := \mathbb {1} \left\{\sum_ {t = 1} ^ {n} \phi_ {1} (D _ {t}) > \sum_ {t = 1} ^ {n} \left(\mathbb {E} _ {\theta^ {*}} \left[ \phi_ {1} (D _ {t}) \right] + \mathbb {E} _ {\theta} \left[ \phi_ {1} (D _ {t}) \right]\right) / 2 \right\}. +$$ + +Combining the last three displays, by Bernstein inequality, we have + +$$ +\begin{array}{l} \mathbb {E} _ {\theta^ {*}} ^ {n} \left[ \Phi_ {j, 1} (\mathbf {D} _ {n}) \right] \leq \mathbb {P} _ {\theta^ {*}} ^ {n} \left(\sum_ {t = 1} ^ {n} \left(\phi_ {1} (D _ {t}) - \mathbb {E} _ {\theta^ {*}} \left[ \phi_ {1} (D _ {t}) \right]\right) > \frac {C _ {1}}{2} n \sum_ {k \in I _ {j - 1}} q (g _ {k})\right) \\ \leq \exp \left(- \frac {\left(C _ {1} ^ {2} / 8\right) n ^ {2} \left(\sum_ {k \in I _ {j - 1}} q (g _ {k})\right) ^ {2}}{\sum_ {t = 1} ^ {n} \operatorname {V a r} _ {\theta^ {*}} \left(\phi_ {1} \left(D _ {t}\right) - \mathbb {E} _ {\theta^ {*}} \left[ \phi_ {1} \left(D _ {t}\right) \right]\right) + \left(C _ {1} / 6\right) n \sum_ {k \in I _ {j - 1}} q (g _ {k})}\right) \\ \leq \exp \left(- C _ {2} n \sum_ {k \in I _ {j - 1}} q \left(g _ {k}\right)\right) \\ \leq \exp \left(- C _ {3} n | I _ {j - 1} | q _ {n}\right) \\ \leq \exp \left(- C _ {3} n ^ {\gamma + \frac {2}{3}} q _ {n}\right), \tag {24} \\ \end{array} +$$ + +where $C_2 = C_1^2 / (8(1 + C_1 / 6))$ be a positive constant, the fourth inequality holds with a positive constant $C_3$ depending on $C_2$ and $q(\cdot)$ because $q(p) \gtrsim q_n$ for any $p \in \mathcal{G}$ with $q_n = n^{-\gamma - 1/3} (\log n)^{1/2}$ under the assumption (A4), and the last inequality holds because $|I_j| \geq K / J \geq n^{\gamma - 1/3}$ for any $j = 1, \ldots, J - 1$ . Similarly, we have + +$$ +\sup _ {\theta \in \Theta_ {j, 1}} \mathbb {E} _ {\theta} ^ {n} \left[ 1 - \Phi_ {j, 1} (\mathbf {D} _ {n}) \right] \leq \exp \left(- C _ {3} n ^ {\gamma + \frac {2}{3}} q _ {n}\right), \tag {25} +$$ + +The construction of tests for the second group of hypotheses (23) is similar. Define the tests as follows: + +$$ +\Phi_ {j, 2} \left(\mathbf {D} _ {n}\right) := \mathbb {1} \left\{\sum_ {t = 1} ^ {n} \phi_ {2} (D _ {t}) > \sum_ {t = 1} ^ {n} \left(\mathbb {E} _ {\theta^ {*}} \left[ \phi_ {2} (D _ {t}) \right] + \mathbb {E} _ {\theta} \left[ \phi_ {2} (D _ {t}) \right]\right) / 2 \right\}, \tag {26} +$$ + +where $\phi_2(D_t) = \mathbb{1}\{P_t\in \mathcal{G}(I_j),Y_t = 0\}$ is a function. By a similar argument as the preceding, for any $\theta \in \Theta_{j,2}$ , $k\in I_j$ , and for sufficiently large $n$ such that $\epsilon \leq \eta_1 / (4L_0\kappa)$ , we have + +$$ +S _ {0, k} ^ {*} - S _ {0, k} \geq S _ {0, k} ^ {*} - S _ {0, k _ {j}} ^ {*} + S _ {0, k _ {j}} ^ {*} - S _ {0, k _ {j}} \geq - L _ {0} | g _ {k} - g _ {k _ {j}} | + \eta_ {1} \geq \frac {\eta_ {1}}{2}. +$$ + +By a similar argument as the proof in Lemma A.4, for a sufficiently small $\eta_{2}$ , there exists a positive constant $C_4$ depending on $M_1, M_2, L, B$ and $\eta_{1}$ such that for any $\theta \in \Theta_{j,2}$ , + +$$ +\mathbb {E} _ {\theta} \left[ \phi_ {2} (D _ {t}) \right] > \mathbb {E} _ {\theta *} \left[ \phi_ {2} (D _ {t}) \right] + C _ {4} \sum_ {k \in I _ {j}} q (g _ {k}). +$$ + +In addition, for either $\theta \in \Theta_{j,2}$ or $\theta \in \theta^{*}$ , we have + +$$ +\begin{array}{l} \operatorname {V a r} _ {\theta} \left( \right.\phi_ {2} \left(D _ {t}\right) - \mathbb {E} _ {\theta} \left(\phi_ {2} \left(D _ {t}\right)\right) \leq \mathbb {E} _ {\theta} \left[ \phi_ {2} \left(D _ {t}\right)\right] \\ = \mathbb {E} _ {X _ {t}, P _ {t}} \left[ \left(1 - S _ {0} \left(P _ {t}\right) ^ {\exp \left(X _ {t} ^ {\top} \beta\right)}\right) \mathbb {1} \left\{P _ {t} \in \mathcal {G} (I _ {j}) \right\} \right] \\ < \sum_ {k \in I _ {j}} q (g _ {k}). \\ \end{array} +$$ + +Combining the last two displays, by Bernstein inequality, there exists a positive constant $C_5$ depending on $C_4$ and $q(\cdot)$ such that + +$$ +\mathbb {E} _ {\theta^ {*}} ^ {n} \left[ \Phi_ {j, 2} (\mathbf {D} _ {n}) \right] \leq \exp \left(- C _ {5} n ^ {\gamma + \frac {2}{3}} q _ {n}\right), \quad \sup _ {\theta \in \Theta_ {j, 2}} \mathbb {E} _ {\theta} ^ {n} \left[ 1 - \Phi_ {j, 2} (\mathbf {D} _ {n}) \right] \leq \exp \left(- C _ {5} n ^ {\gamma + \frac {2}{3}} q _ {n}\right). \tag {27} +$$ + +Note that the union of the sets in the alternative hypotheses (22) and (23) for all $j = 2,\dots ,J - 1$ contains $\Theta_{\eta_1,\eta_2}\coloneqq \{(\mathbf{S}_0,\beta)\in \mathcal{S}_0\times \mathbb{R}^d:\max_{2\leq j\leq J - 1}|S_{0,k_j} - S_{0,k_j}^* |\geq \eta_1,\| \beta -\beta^*\| _2 < \eta_2\}$ . We set $\Phi_n\coloneqq \max_{2\leq j\leq J - 1}\{\Phi_{j,1}\vee \Phi_{j,2}\}$ . Combining (24), (25) and (27), we have + +$$ +\begin{array}{l} \mathbb {E} _ {\theta^ {*}} ^ {n} \left[ \Phi_ {n} (\mathbf {D} _ {n}) \right] \leq J \exp \left(- C _ {6} n ^ {\gamma + \frac {2}{3}} q _ {n}\right) \\ = \exp \left(\log J - C _ {6} n ^ {\gamma + \frac {2}{3}} q _ {n}\right), \\ \end{array} +$$ + +$$ +\sup _ {\theta \in \Theta_ {\eta_ {1}, \eta_ {2}}} \mathbb {E} _ {\theta} ^ {n} \left[ 1 - \Phi_ {n} (\mathbf {D} _ {n}) \right] \leq \exp \left(- C _ {6} n ^ {\gamma + \frac {2}{3}} q _ {n}\right), +$$ + +where $C_6 = \min \{C_3,C_5\}$ . By the definition of $q_{n}$ , we have $\mathbb{E}_{\theta^{*}}^{n}[\Phi_{n}(\mathbf{D}_{n})]\to 0$ and $\sup_{\theta \in \Theta_{\eta_1,\eta_2}}\mathbb{E}_\theta^n [1 - \Phi_n(\mathbf{D}_n)]\to 0$ as $n\to \infty$ when $\gamma \geq 1 / 3$ . By Lemma D.11 of [16], there exist tests $\Psi_{n}$ and a constant $C_7 > 0$ such that $\mathbb{E}_{\theta^*}^n [\Psi_n(\mathbf{D}_n)]\leq \exp (-C_7n)$ and $\sup_{\theta \in \Theta_{\eta_1,\eta_2}}\mathbb{E}_\theta^n [1 - \Psi_n(\mathbf{D}_n)]\leq \exp (-C_7n)$ . The proof is then complete. + +![](images/22e7715bc3c65573724d132e42f325d717b2b06290ef52b8c1d00372fbdd8c9e.jpg) + +Proof of Lemma A.1. We proceed with the proof by considering two separate cases: $\gamma < 2/3$ and $\gamma \geq 2/3$ . First, we suppose that $\gamma < 2/3$ . Let $\epsilon_0 > 0$ be a constant to be chosen later, and define $\Theta_{\epsilon_0} = \{(\mathbf{S}_0,\beta) \in \Theta : \| \boldsymbol{\Lambda}_0 - \boldsymbol{\Lambda}_0^*\|_\infty \vee \| \beta - \beta^*\|_2 \leq \epsilon_0\}$ . Here, $\boldsymbol{\Lambda}_0 = (\Lambda_{0,1},\dots,\Lambda_{0,K})$ and $\boldsymbol{\Lambda}_0^* = (\Lambda_{0,1}^*,\dots,\Lambda_{0,K}^*)$ are $K$ -dimensional vectors corresponding to $\mathbf{S}_0$ and $\mathbf{S}_0^*$ , respectively, such that $\Lambda_{0,k} = -\log S_{0,k}$ and $\Lambda_{0,k}^* = -\log S_{0,k}^*$ for $k = 1,\dots,K$ . The log-likelihood ratio satisfies + +$$ +\begin{array}{l} \log \frac {p _ {\theta^ {*}}}{p _ {\theta}} (x, p, y) = y \log \frac {H _ {\theta^ {*}} (x , p)}{H _ {\theta} (x , p)} + (1 - y) \log \frac {1 - H _ {\theta^ {*}} (x , p)}{1 - H _ {\theta} (x , p)} \\ \leq \max \left\{\log \frac {H _ {\theta^ {*}} (x , p)}{H _ {\theta} (x , p)}, \log \frac {1 - H _ {\theta^ {*}} (x , p)}{1 - H _ {\theta} (x , p)} \right\}. \\ \end{array} +$$ + +By assumption (A5), there exist constants $M_1$ and $M_2$ such that $0 < M_1 \leq S_0^*(p_{\mathrm{max}}) < S_0^*(p_{\mathrm{min}}) \leq M_2 < 1$ . Note that for $\theta \in \Theta_{\epsilon_0}$ , where $\epsilon_0 \leq (M_1 \wedge (1 - M_2)/2) \wedge B$ , we have $S_0(v) \in [M_1/2, (1 + M_2)/2]$ for any $v \in [p_{\mathrm{min}}, p_{\mathrm{max}}]$ , and $\|\beta\| \leq 2B$ under assumptions (A1) and (A5). Furthermore, by assumption (A2), both $H_{\theta^*}(x,p)$ and $H_{\theta}(x,p)$ are bounded away from 0 and 1 for any $x \in \mathcal{X}$ , $p \in \mathcal{G}$ and $\theta \in \Theta_{\epsilon_0}$ . Since $|\log p - \log q| \leq |p - q| \max\{p^{-1}, q^{-1}\}$ for any $0 < p, q < 1$ , we have + +$$ +\left\| \log \frac {p _ {\theta^ {*}}}{p _ {\theta}} \right\| _ {\infty} \leq C _ {0} \| H _ {\theta^ {*}} - H _ {\theta} \| _ {\infty}, \tag {28} +$$ + +where $C_0$ is a positive constant depending on $M_1, M_2, L$ and $B$ . In addition, by Lemma C.2, there exist positive constants $c_1$ and $c_2$ , depending on $M_1, M_2, L$ and $B$ , such that for any $x \in \mathcal{X}$ and $p \in \mathcal{G}$ , + +$$ +\begin{array}{l} \left| H _ {\theta^ {*}} (x, p) - H _ {\theta} (x, p) \right| \leq c _ {1} \left\| \mathbf {S} _ {0} - \mathbf {S} _ {0} ^ {*} \right\| _ {\infty} + c _ {2} \| \beta - \beta^ {*} \| _ {2} \\ \leq c _ {1} \| \boldsymbol {\Lambda} _ {0} - \boldsymbol {\Lambda} _ {0} ^ {*} \| _ {\infty} + c _ {2} \| \beta - \beta^ {*} \| _ {2}, \\ \end{array} +$$ + +where the last inequality holds because $\| \mathbf{S}_0 - \mathbf{S}_0^*\|_{\infty}\leq \| \pmb {\Lambda}_0 - \pmb {\Lambda}_0^*\|_{\infty}$ . Combining the last two displays, for $\theta \in \Theta_{\epsilon_0}$ , we have $K(p_{\theta^{*}},p_{\theta})\leq \| \log (p_{\theta^{*}} / p_{\theta})\|_{\infty} < C_1\epsilon_0$ , where $C_1 = C_0(c_1 + c_2)$ is a positive constant. Then, we obtain + +$$ +\Theta_ {\epsilon_ {0}} \subseteq \left\{\theta \in \Theta : K \left(p _ {\theta^ {*}}, p _ {\theta}\right) < C _ {1} \epsilon_ {0} \right\}. \tag {29} +$$ + +We denote the renormalized restriction of $\Pi$ to $\Theta_{\epsilon_0}$ by $\Pi_{\epsilon_0}$ . We note that + +$$ +\begin{array}{l} \int_ {\Theta} \prod_ {t = 1} ^ {n} \frac {p _ {\theta}}{p _ {\theta^ {*}}} \left(D _ {t}\right) d \Pi (\theta) \geq \Pi \left(\Theta_ {\epsilon_ {0}}\right) \int_ {\Theta_ {\epsilon_ {0}}} \prod_ {t = 1} ^ {n} \frac {p _ {\theta}}{p _ {\theta^ {*}}} \left(D _ {t}\right) d \Pi_ {\epsilon_ {0}} (\theta) \tag {30} \\ \geq \Pi (\Theta_ {\epsilon_ {0}}) \exp \left(- \int_ {\Theta_ {\epsilon_ {0}}} \sum_ {t = 1} ^ {n} \log \left(\frac {p _ {\theta^ {*}}}{p _ {\theta}}\right) (D _ {t}) d \Pi_ {\epsilon_ {0}} (\theta)\right), \\ \end{array} +$$ + +where the last inequality holds by Jensen's inequality. Since the log-likelihood ratio is bounded from (28), by Hoeffding's inequality, we have + +$$ +\mathbb {P} _ {\theta^ {*}} ^ {n} \left(\sum_ {t = 1} ^ {n} \log \left(\frac {p _ {\theta^ {*}}}{p _ {\theta}}\right) (D _ {t}) - n K \left(p _ {\theta^ {*}}, p _ {\theta}\right) < \epsilon_ {0} n\right) > 1 - \exp \left(- \frac {\epsilon_ {0} ^ {2}}{2 C _ {0} ^ {2}} n\right). \tag {31} +$$ + +Let $\Omega_1$ be the event in the left-hand side of the last display. Thus, on the event $\Omega_1$ , we have + +$$ +\begin{array}{l} - \int_ {\Theta_ {\epsilon_ {0}}} \sum_ {t = 1} ^ {n} \log \left(\frac {p _ {\theta^ {*}}}{p _ {\theta}}\right) (D _ {t}) d \Pi_ {\epsilon_ {0}} (\theta) > - \int_ {\Theta_ {\epsilon_ {0}}} n K \left(p _ {\theta^ {*}}, p _ {\theta}\right) d \Pi_ {\epsilon_ {0}} (\theta) - \epsilon_ {0} n \\ > - (C _ {1} + 1) \epsilon_ {0} n, \\ \end{array} +$$ + +where the last inequality holds by (29). Combining this with (30), on the event $\Omega_1$ , we have + +$$ +\int_ {\Theta} \prod_ {t = 1} ^ {n} \frac {p _ {\theta}}{p _ {\theta^ {*}}} \left(D _ {t}\right) d \Pi (\theta) > \Pi \left(\Theta_ {\epsilon_ {0}}\right) \exp \left(- \left(C _ {1} + 1\right) \epsilon_ {0} n\right). \tag {32} +$$ + +Let $\epsilon_{1} = \epsilon_{0}n^{-\gamma}$ . By Lemma C.3, with the specified prior (3) and the hyperparameter condition (P2), there exist positive constants $c_{3}, c_{4}$ and $c_{5}$ depending on $p_{\mathrm{min}}, p_{\mathrm{max}}, M_{1}, M_{2}, \underline{\alpha}, \overline{\alpha}, \rho$ and $\epsilon_{0}$ , such that + +$$ +\Pi \left(\left\| \boldsymbol {\Lambda} _ {0} - \boldsymbol {\Lambda} _ {0} ^ {*} \right\| _ {\infty} \leq \epsilon_ {1}\right) \geq c _ {3} \exp \left(- c _ {4} K - c _ {5} K \log_ {-} \epsilon_ {1}\right). +$$ + +In addition, under the prior condition (P1), we have $\Pi (\| \beta -\beta^{*}\|_{2}\leq \epsilon_{0})\geq C_{2}$ , where $C_2$ is a positive constant depending on $d$ , $\epsilon_0$ and the lower bound of the prior on a neighborhood of $\beta^*$ . Then, we have + +$$ +\begin{array}{l} \Pi \left(\Theta_ {\epsilon_ {0}}\right) \geq \Pi \left(\| \boldsymbol {\Lambda} _ {0} - \boldsymbol {\Lambda} _ {0} ^ {*} \| _ {\infty} \leq \epsilon_ {0}\right) \cdot \Pi \left(\| \beta - \beta^ {*} \| _ {2} \leq \epsilon_ {0}\right) \\ \geq C _ {2} c _ {3} \exp \left(- c _ {4} K - c _ {5} \log_ {-} \epsilon_ {1} K\right), \\ \end{array} +$$ + +where the last inequality follows from the previous display and the fact that $\epsilon_1 \leq \epsilon_0$ for $n \geq 1$ . Combining this with (32), on the event $\Omega_1$ , we have + +$$ +\begin{array}{l} \int_ {\Theta} \prod_ {t = 1} ^ {n} \frac {p _ {\theta}}{p _ {\theta^ {*}}} (D _ {t}) d \Pi (\theta) > C _ {2} c _ {3} \exp \left(- c _ {4} K - c _ {5} \log_ {-} \epsilon_ {1} K - (C _ {1} + 1) \epsilon_ {0} n\right) \\ > C _ {2} c _ {3} \exp (- C _ {3} K \log n - (C _ {1} + 1) \epsilon_ {0} n), \tag {33} \\ \end{array} +$$ + +where the last inequality holds by $C_3 = c_4 + c_5(\log_{-}\epsilon_0 + 1)$ because $\log_{-}\epsilon_{1} < \log_{-}\epsilon_{0} + \log n$ . By Lemma A.3 and A.4, there exist tests $\Phi_n$ such that + +$$ +\mathbb {E} _ {\theta^ {*}} ^ {n} \left[ \Phi_ {n} \right] \leq \exp (- C _ {4} n), \quad \sup _ {\theta \in U ^ {c}} \mathbb {E} _ {\theta} ^ {n} \left[ 1 - \Phi_ {n} \right] \leq \exp (- C _ {4} n), \tag {34} +$$ + +where $C_4$ is a positive constant depending on $M_{1}, M_{2}, L, B, p_{\min}, p_{\max}, \kappa$ and $\epsilon$ . Then, we have + +$$ +\begin{array}{l} \mathbb {E} _ {\theta^ {*}} ^ {n} \left[ \Pi (U ^ {c} | \mathbf {D} _ {n}) \right] \leq \mathbb {E} _ {\theta^ {*}} ^ {n} \left[ \Phi_ {n} \right] + \mathbb {E} _ {\theta^ {*}} ^ {n} \left[ (1 - \Phi_ {n}) \Pi (U ^ {c} | \mathbf {D} _ {n}) \mathbb {1} \left\{\Omega_ {1} \right\} \right] + \mathbb {P} _ {\theta^ {*}} ^ {n} \left(\Omega_ {1} ^ {c}\right) \\ \leq \mathbb {E} _ {\theta^ {*}} ^ {n} \left[ \Phi_ {n} \right] + \left(C _ {2} c _ {3}\right) ^ {- 1} \exp \left(C _ {3} K \log n + (C _ {1} + 1) \epsilon_ {0} n\right) \sup _ {\theta \in U ^ {c}} \mathbb {E} _ {\theta} ^ {n} \left[ 1 - \Phi_ {n} \right] + \exp \left(- \frac {\epsilon_ {0} ^ {2}}{2 C _ {0} ^ {2}} n\right) \\ \leq \exp (- C _ {4} n) + C _ {5} \exp \left(C _ {3} C _ {p} n ^ {\frac {2}{3}} \log n - (C _ {4} - (C _ {1} + 1) \epsilon_ {0}) n\right) + \exp (- C _ {6} n), \\ \end{array} +$$ + +where the second inequality holds by (31) and (33), and the last inequality follows from (34), with $C_5 = (C_2c_3)^{-1}$ and $C_6 = \epsilon_0^2 / (2C_0^2)$ , and $K \leq C_p n^{2/3}$ for $\gamma < 2/3$ , where $C_p$ is a positive constant depending on $p_{\min}, p_{\max}$ and $\kappa$ . We choose $\epsilon_0 = (C_4 / (3(C_1 + 1))) \wedge ((M_1 \wedge (1 - M_2) / 2) \wedge B)$ to ensure that the second term on the right-hand side of the previous display is bounded by $C_5 \exp(-(C_4 / 3)n)$ , provided that $n^{1/3} (\log n)^{-1} \geq 3C_3 C_p / C_4$ . Then, we have + +$$ +\begin{array}{l} \mathbb {E} _ {\theta^ {*}} ^ {n} \left[ \Pi (U ^ {c} | \mathbf {D} _ {n}) \right] \leq \exp (- C _ {4} n) + C _ {5} \exp (- (C _ {4} / 3) n) + \exp (- C _ {6} n) \\ \leq C _ {7} \exp (- C _ {8} n), \\ \end{array} +$$ + +where $C_7 = C_5 + 2$ and $C_8 = (C_4 / 3)\wedge C_6$ . By the Markov inequality, for $n\geq (3C_3C_p / C_4)^3$ + +$$ +\mathbb {P} _ {\theta^ {*}} ^ {n} \left(\Pi (U ^ {c} | \mathbf {D} _ {n}) \geq C _ {7} \exp (- C _ {9} n)\right) < \exp (- C _ {9} n), +$$ + +where $C_9 = C_8 / 2$ . This concludes the proof for the case where $\gamma < 2 / 3$ . + +Now, we suppose that $\gamma \geq 2/3$ . Let $\epsilon_2 = n^{-1/3}$ and $J = \lceil (p_{\max} - p_{\min}) / (\kappa \epsilon_2) \rceil$ . Define $(k_1, \ldots, k_J)$ as a subsequence of $[K]$ such that $p_{\min} + (j-1)\kappa \epsilon_2 < g_{k_j} \leq p_{\min} + (j-1)\kappa \epsilon_2 + \delta$ for $j = 1, \ldots, J-1$ , and set $k_J = K$ . + +Suppose that $|S_{0,k_j} - S_{0,k_j}^*| < \epsilon / 2$ for every $j = 1, \ldots, J$ . Then, for any $k \in [K]$ with $k_j \leq k < k_{j+1}$ for $j = 1, \ldots, J-2$ , we have + +$$ +\begin{array}{l} S _ {0, k} ^ {*} - S _ {0, k} \leq S _ {0, k} ^ {*} - S _ {0, k _ {j + 1}} \\ \leq | S _ {0, k _ {j + 1}} ^ {*} - S _ {0, k _ {j + 1}} | + | S _ {0, k} ^ {*} - S _ {0, k _ {j + 1}} ^ {*} | \\ < \epsilon / 2 + L _ {0} \left| g _ {k} - g _ {k _ {j + 1}} \right| \\ \leq \epsilon / 2 + L _ {0} (\kappa \epsilon_ {2} + \delta) \\ \leq \epsilon / 2 + 2 L _ {0} \kappa \epsilon_ {2} \\ \leq \epsilon , \\ \end{array} +$$ + +where the third inequality holds by our assumption and $L_0$ -Lipschitz continuity of $S_0^*$ , with $L_0$ being a positive constant because (A5) is assumed, the fourth inequality holds by the definition of $k_j$ , the fifth inequality holds because $\delta \leq \kappa \epsilon_2$ for $\gamma \geq 2/3$ , and the last inequality holds for sufficiently large $n$ so that $\epsilon_2 \leq \epsilon/(4L_0\kappa)$ . Note that $|g_{k_{J-1}} - g_{k_J}| < 2\kappa \epsilon_2$ by the definition of $J$ . Then, for any $k \in [K]$ with $k_{J-1} \leq k \leq k_J$ , we have + +$$ +S _ {0, k} ^ {*} - S _ {0, k} < \epsilon . +$$ + +Combining the preceding two displays, we have $S_{0,k}^{*} - S_{0,k} < \epsilon$ for any $k \in [K]$ . Similarly, for any $k \in [K]$ , we have $S_{0,k} - S_{0,k}^{*} < \epsilon$ . Therefore, for $n \geq (4L_0\kappa /\epsilon)^3$ , we have + +$$ +\left\{\mathbf {S} _ {0} \in \mathcal {S} _ {0}: \left\| \mathbf {S} _ {0} - \mathbf {S} _ {0} ^ {*} \right\| _ {\infty} \geq \epsilon \right\} \subset \left\{\mathbf {S} _ {0} \in \mathcal {S} _ {0}: \left\| \mathbf {S} _ {0} - \mathbf {S} _ {0} ^ {*} \right\| _ {\infty , J} \geq \epsilon / 2 \right\}. \tag {35} +$$ + +Then, we can decompose + +$$ +\begin{array}{l} \mathbb {E} _ {\theta^ {*}} ^ {n} \left[ \Pi \left(U ^ {c} \mid \mathbf {D} _ {n}\right) \right] \leq \mathbb {E} _ {\theta^ {*}} ^ {n} \left[ \Pi \left(\left\{\theta \in \Theta : \| \mathbf {S} _ {0} - \mathbf {S} _ {0} ^ {*} \| _ {\infty , J} \geq \epsilon / 2 \text {o r} \| \beta - \beta^ {*} \| _ {2} \geq \epsilon \right\} \mid \mathbf {D} _ {n}\right) \right] \\ \leq \underbrace {\mathbb {E} _ {\theta^ {*}} ^ {n} \left[ \Pi \left(U _ {1} \mid \mathbf {D} _ {n}\right) \right]} _ {\text {(i)}} + \underbrace {\mathbb {E} _ {\theta^ {*}} ^ {n} \left[ \Pi \left(U _ {2} \mid \mathbf {D} _ {n}\right) \right]} _ {\text {(ii)}} + \underbrace {\mathbb {E} _ {\theta^ {*}} ^ {n} \left[ \Pi \left(U _ {3} \mid \mathbf {D} _ {n}\right) \right]} _ {\text {(iii)}}, \tag {36} \\ \end{array} +$$ + +where + +$$ +U _ {1} = \left\{\theta \in \Theta : \max _ {2 \leq j \leq J - 1} \left| S _ {0, k _ {j}} - S _ {0, k _ {j}} ^ {*} \right| \geq \epsilon / 4 \text {o r} \| \beta - \beta^ {*} \| _ {2} \geq \epsilon \right\}, +$$ + +$$ +U _ {2} = \{\theta \in \Theta : | S _ {0, k _ {1}} - S _ {0, k _ {1}} ^ {*} | \geq \epsilon / 2, \max _ {2 \leq j \leq J - 1} | S _ {0, k _ {j}} - S _ {0, k _ {j}} ^ {*} | < \epsilon / 4 \}, +$$ + +$$ +U _ {3} = \{\theta \in \Theta : | S _ {0, k _ {J}} - S _ {0, k _ {J}} ^ {*} | \geq \epsilon / 2, \max _ {2 \leq j \leq J - 1} | S _ {0, k _ {j}} - S _ {0, k _ {j}} ^ {*} | < \epsilon / 4 \}. +$$ + +The proof for (i) is similar to that of $\gamma < 2/3$ . Let $\epsilon_3 > 0$ be a constant to be chosen later. Similarly as in (35), we have + +$$ +\left\{\mathbf {S} _ {0} \in \mathcal {S} _ {0}: \| \boldsymbol {\Lambda} _ {0} - \boldsymbol {\Lambda} _ {0} ^ {*} \| _ {\infty} \geq \epsilon_ {3} n ^ {- \frac {1}{3}} \right\} \subset \left\{\mathbf {S} _ {0} \in \mathcal {S} _ {0}: \| \boldsymbol {\Lambda} _ {0} - \boldsymbol {\Lambda} _ {0} ^ {*} \| _ {\infty , J} \geq C _ {1 0} \epsilon_ {3} n ^ {- \frac {1}{3}} \right\}, \tag {37} +$$ + +where $C_{10}$ is a positive constant depending on $L_0$ and $\kappa$ . Then, we have + +$$ +\begin{array}{l} \Pi \left(\Theta_ {\epsilon_ {3}}\right) \geq \Pi \left(\left\| \boldsymbol {\Lambda} _ {0} - \boldsymbol {\Lambda} _ {0} ^ {*} \right\| _ {\infty} \leq \epsilon_ {3}\right) \cdot \Pi \left(\left\| \beta - \beta^ {*} \right\| _ {2} \leq \epsilon_ {3}\right) \\ \geq C _ {1 1} \Pi \left(\| \boldsymbol {\Lambda} _ {0} - \boldsymbol {\Lambda} _ {0} ^ {*} \| _ {\infty , J} \leq C _ {1 0} \epsilon_ {3} n ^ {- \frac {1}{3}}\right) \\ \geq C _ {1 1} c _ {6} \exp \left(- c _ {7} J - c _ {8} \log_ {-} (C _ {1 0} \epsilon_ {3} n ^ {- \frac {1}{3}}) J\right), \\ \end{array} +$$ + +where the second inequality holds by a positive constant $C_{11}$ depending on $d$ , $\epsilon_3$ and the prior's lower bound near $\beta^{*}$ , and the last inequality follows from constants $c_{6}, c_{7}$ and $c_{8}$ in Lemma C.3, depending on $p_{\mathrm{min}}, p_{\mathrm{max}}, M_1, M_2, \underline{\alpha}, \overline{\alpha}, \rho, C_{10}$ and $\epsilon_3$ . Similarly as in (33), there exists the event $\Omega_2$ such that $\mathbb{P}_{\theta^*}^n(\Omega_2^c) \leq \exp(-\epsilon_3^2/(2C_0^2)n)$ , and on the event $\Omega_2$ , we have + +$$ +\begin{array}{l} \int_ {\Theta} \prod_ {t = 1} ^ {n} \frac {p _ {\theta}}{p _ {\theta^ {*}}} (D _ {t}) d \Pi (\theta) > C _ {1 1} c _ {6} \exp \left(- c _ {7} J - c _ {8} \log_ {-} (C _ {1 0} \epsilon_ {3} n ^ {- \frac {1}{3}}) J - (C _ {1} + 1) \epsilon_ {3} n\right) \\ > C _ {1 1} c _ {6} \exp (- C _ {1 2} J \log n - (C _ {1} + 1) \epsilon_ {3} n) \\ \geq C _ {1 1} c _ {6} \exp \left(- C _ {1 2} C _ {p} n ^ {\frac {1}{3}} \log n - (C _ {1} + 1) \epsilon_ {3} n\right), \\ \end{array} +$$ + +where the second inequality holds by $C_{12} = c_7 + c_8(\log_{-}(C_{10}\epsilon_3) + 1)$ , and the last inequality holds because $J \leq C_p n^{1/3}$ by the definition of $J$ . By Lemma A.5, there exist tests $\Phi_{n,1}$ such that + +$$ +\mathbb {E} _ {\theta^ {*}} ^ {n} \left[ \Phi_ {n, 1} \right] \leq \exp (- C _ {1 3} n), \quad \sup _ {\theta \in U _ {1}} \mathbb {E} _ {\theta} ^ {n} \left[ 1 - \Phi_ {n, 1} \right] \leq \exp (- C _ {1 3} n), +$$ + +where $C_{13}$ is a positive constant depending on $M_1, M_2, L, B, p_{\min}, p_{\max}, \kappa$ and $\epsilon$ . We choose $\epsilon_3 = (C_{13} / (3(C_1 + 1))) \wedge ((M_1 \wedge (1 - M_2) / 2) \wedge B)$ . Combining the last two displays, for $n \geq (3C_{12}C_p / C_{13})^{3/2}$ , we have + +$$ +\mathbb {E} _ {\theta^ {*}} ^ {n} \left[ \Pi \left(U _ {1} \mid \mathbf {D} _ {n}\right) \right] \leq C _ {1 4} \exp \left(- C _ {1 5} n\right), \tag {38} +$$ + +where $C_{14} = (C_{11}c_6)^{-1} + 2$ and $C_{15} = (C_{13} / 3)\wedge (\epsilon_3^2 /(2C_0^2))$ are positive constants. + +We now consider the term (ii). We split $U_{2}$ in $U_{2, -}$ and $U_{2, +}$ , where + +$$ +U _ {2, -} = \left\{\theta \in \Theta : S _ {0, k _ {1}} - S _ {0, k _ {1}} ^ {*} \leq - \epsilon / 2, \max _ {2 \leq j \leq J - 1} \left| S _ {0, k _ {j}} - S _ {0, k _ {j}} ^ {*} \right| < \epsilon / 4 \right\}, +$$ + +$$ +U _ {2, +} = \left\{\theta \in \Theta : S _ {0, k _ {1}} - S _ {0, k _ {1}} ^ {*} \geq \epsilon / 2, \max _ {2 \leq j \leq J - 1} | S _ {0, k _ {j}} - S _ {0, k _ {j}} ^ {*} | < \epsilon / 4 \right\}. +$$ + +Note that $U_{2, - } \subset U_{2, - }^{1} \cup U_{2, - }^{2}$ , where + +$$ +U _ {2, -} ^ {1} = \left\{\theta \in \Theta : \| \beta - \beta^ {*} \| _ {2} \geq \eta \right\}, +$$ + +$$ +U _ {2, -} ^ {2} = \left\{\theta \in \Theta : S _ {0, k _ {1}} - S _ {0, k _ {1}} ^ {*} \leq - \epsilon / 2, \| \beta - \beta^ {*} \| _ {2} < \eta \right\}, +$$ + +for some sufficiently small positive constant $\eta$ depending on $\epsilon$ . By Lemma A.3, there exist exponentially consistent tests $\Phi_{n,2,1}$ for testing $H_0: \theta = \theta^*$ , $H_1: \theta \in U_{2, -}^1$ . Similarly as (26) in the proof of Lemma A.5, we construct the tests $\Psi_{n,2,2}$ for $U_{2, -}^2$ by + +$$ +\Psi_ {n, 2, 2} = \mathbb {1} \left\{\sum_ {t = 1} ^ {n} \phi_ {2} (D _ {t}) > \sum_ {t = 1} ^ {n} \left(\mathbb {E} _ {\theta^ {*}} \left[ \phi_ {2} (D _ {t}) \right] + \mathbb {E} _ {\theta} \left[ \phi_ {2} (D _ {t}) \right]\right) / 2 \right\}, +$$ + +where $\phi_2(D_t) = \mathbb{1}\{P_t\in \{g_{k_1},\ldots ,g_{k_2}\} ,Y_t = 0\}$ . By a similar argument as the proof in Lemma A.5, we can show that $\mathbb{E}_{\theta^*}^n [\Psi_{n,2,2}(\mathbf{D}_n)]\to 0$ and $\sup_{\theta \in U_{2, - }^{2}}\mathbb{E}_{\theta}^{n}[1 - \Psi_{n,2,2}(\mathbf{D}_{n})]\to 0$ as $n\to \infty$ . By Lemma D.11 of [16], there exist exponentially consistent tests $\Phi_{n,2,2}$ for testing $H_0:\theta = \theta^*,H_1:\theta \in U_{2, - }^{2}$ . Let $\Phi_{n,2} = \Phi_{n,2,1}\lor \Phi_{n,2,2}$ . Then, there exists a positive constant $C_{16}$ depending on $M_1$ , $M_2$ , $L$ , $B$ , $p_{\mathrm{min}}$ , $p_{\mathrm{max}}$ , $\kappa$ and $\epsilon$ such that + +$$ +\mathbb {E} _ {\theta^ {*}} ^ {n} \left[ \Phi_ {n, 2} \right] \leq \exp (- C _ {1 6} n), \quad \sup _ {\theta \in U _ {2, -}} \mathbb {E} _ {\theta} ^ {n} \left[ 1 - \Phi_ {n, 2} \right] \leq \exp (- C _ {1 6} n). +$$ + +Then, by a similar argument as the preceding, for $n \geq (3C_{12}C_p / C_{16})^{3/2}$ , it holds that + +$$ +\mathbb {E} _ {\theta^ {*}} ^ {n} \left[ \Pi \left(U _ {2, -} \mid \mathbf {D} _ {n}\right) \right] \leq C _ {1 4} \exp (- C _ {1 7} n), \tag {39} +$$ + +where $C_{17} = (C_{16} / 3)\wedge (\epsilon_3^2 /(2C_0^2))$ is a positive constant. + +We restrict ourselves to vectors $\mathbf{S}_0$ such that $\max_{2\leq j\leq J - 1}|S_{0,k_j} - S_{0,k_j}^*| < \epsilon /4$ . Suppose that $S_{0,k_1} - S_{0,k_2} < \epsilon /4$ . Then, we have + +$$ +\begin{array}{l} S _ {0, k _ {1}} - S _ {0, k _ {1}} ^ {*} = S _ {0, k _ {1}} - S _ {0, k _ {2}} + S _ {0, k _ {2}} - S _ {0, k _ {1}} ^ {*} \\ \leq S _ {0, k _ {1}} - S _ {0, k _ {2}} + S _ {0, k _ {2}} - S _ {0, k _ {2}} ^ {*} \\ < \epsilon / 4 + \epsilon / 4 \\ = \epsilon / 2, \\ \end{array} +$$ + +where the first inequality holds by the monotonicity of $S_0^*$ , and the last inequality follows from our assumption and the fact that $|S_{0,k_2} - S_{0,k_2}^*| < \epsilon / 4$ . Thus, it holds that + +$$ +\begin{array}{l} U _ {2, +} \subset \left\{\theta \in \Theta : S _ {0, k _ {1}} - S _ {0, k _ {2}} \geq \epsilon / 4, \max _ {2 \leq j \leq J - 1} \left| S _ {0, k _ {j}} - S _ {0, k _ {j}} ^ {*} \right| < \epsilon / 4 \right\} \\ \subset \{\theta \in \Theta : S _ {0, k _ {1}} - S _ {0, k _ {2}} \geq \epsilon / 4 \}. \\ \end{array} +$$ + +Then, it is sufficient to show that $\mathbb{E}_{\theta^*}^n\left[\Pi (U_{2, + }^\prime \mid \mathbf{D}_n)\right]\to 0$ as $n\to \infty$ , where $U_{2, + }^{\prime} = \{\theta \in \Theta : S_{0,k_1} - S_{0,k_2}\geq \epsilon /4\}$ . Let $\epsilon_4 = \epsilon_5n^{-1 / 3}$ , where $\epsilon_5 > 0$ is a sufficiently small constant to be chosen later. Similarly as in (32), there exists the event $\Omega_3$ such that $\mathbb{P}_{\theta^*}^n (\Omega_3^c)\leq \exp (-\epsilon_4^2 /(2C_0^2)n)$ , and on the event $\Omega_3$ , we have + +$$ +\int_ {\Theta} \prod_ {t = 1} ^ {n} \frac {p _ {\theta}}{p _ {\theta^ {*}}} (D _ {t}) d \Pi (\theta) > \Pi \left(\Theta_ {\epsilon_ {4}}\right) \exp \left(- \left(C _ {1} + 1\right) \epsilon_ {4} n\right). +$$ + +Furthermore, we have + +$$ +\begin{array}{l} \Pi \left(\Theta_ {\epsilon_ {4}}\right) \geq \Pi \left(\| \boldsymbol {\Lambda} _ {0} - \boldsymbol {\Lambda} _ {0} ^ {*} \| _ {\infty} \leq \epsilon_ {4}\right) \cdot \Pi \left(\| \beta - \beta^ {*} \| _ {2} \leq \epsilon_ {4}\right) \\ \geq c _ {6} \exp (- c _ {7} J - c _ {8} \log_ {-} (C _ {1 0} \epsilon_ {4}) J) \cdot \Pi (\| \beta - \beta^ {*} \| _ {2} \leq \epsilon_ {4}) \\ \geq C _ {1 8} c _ {6} \exp \left(- c _ {7} J - c _ {8} \log_ {-} (C _ {1 0} \epsilon_ {4}) J\right) \cdot \epsilon_ {4} ^ {d} \\ = C _ {1 8} c _ {6} \exp \left(- c _ {7} J - c _ {8} \log_ {-} \left(C _ {1 0} \epsilon_ {4}\right) J + d \log \epsilon_ {4}\right), \\ \end{array} +$$ + +where the second inequality holds by (37) and Lemma C.3, and the last inequality holds because $\Pi (\| \beta -\beta^{*}\|_{2}\leq \epsilon_{4})\geq C_{18}\epsilon_{4}^{d}$ with a positive constant $C_{18}$ depending on $d$ and the prior's lower bound near $\beta^{*}$ . Combining the last two displays, on the event $\Omega_3$ , we have + +$$ +\begin{array}{l} \int_ {\Theta} \prod_ {t = 1} ^ {n} \frac {p _ {\theta}}{p _ {\theta^ {*}}} (D _ {t}) d \Pi (\theta) > C _ {1 8} c _ {6} \exp \left(- c _ {7} J - c _ {8} \log_ {-} (C _ {1 0} \epsilon_ {4}) J + d \log \epsilon_ {4} - (C _ {1} + 1) \epsilon_ {4} n\right) \\ \geq C _ {1 8} c _ {6} \exp \left(- C _ {1 9} n ^ {\frac {1}{3}} \log n - (C _ {1} + 1) \epsilon_ {5} n ^ {\frac {2}{3}}\right), \\ \end{array} +$$ + +where the last inequality holds by a positive constant $C_{19} = C_p(c_7 + c_8(\log_{-}(C_{10}\epsilon_5) + 1)) + d\log_{-}\epsilon_5$ . This implies that + +$$ +\begin{array}{l} \mathbb {E} _ {\theta^ {*}} ^ {n} \left[ \Pi \left(U _ {2, +} ^ {\prime} \mid \mathbf {D} _ {n}\right) \right] \leq \mathbb {E} _ {\theta^ {*}} ^ {n} \left[ \Pi \left(U _ {2, +} ^ {\prime} \mid \mathbf {D} _ {n}\right) \mathbb {1} \left\{\Omega_ {3} \right\} \right] + \mathbb {P} _ {\theta^ {*}} ^ {n} \left(\Omega_ {3} ^ {c}\right) \\ \leq \left(C _ {1 8} c _ {6}\right) ^ {- 1} \exp \left(C _ {1 9} n ^ {\frac {1}{3}} \log n + \left(C _ {1} + 1\right) \epsilon_ {5} n ^ {\frac {2}{3}}\right) \Pi \left(U _ {2, +} ^ {\prime}\right) + \exp \left(- \frac {\epsilon_ {5} ^ {2}}{2 C _ {0} ^ {2}} n ^ {\frac {1}{3}}\right). \tag {40} \\ \end{array} +$$ + +We now prove that the prior mass of $U_{2, + }^{\prime}$ is exponentially small. Note that + +$$ +\begin{array}{l} S _ {0, k _ {1}} - S _ {0, k _ {2}} = \exp (- \Lambda_ {0, k _ {1}}) - \exp (- \Lambda_ {0, k _ {2}}) \\ \leq \Lambda_ {0, k _ {2}} - \Lambda_ {0, k _ {1}} \\ = \delta \sum_ {k = k _ {1} + 1} ^ {k _ {2}} \lambda_ {0, k}. \\ \end{array} +$$ + +Let $\overline{\lambda} = \sum_{k=k_1+1}^{k_2} \lambda_{0,k}$ . By (3), $\overline{\lambda}$ is gamma distributed with parameters $\alpha_0$ and $\rho$ , where $\alpha_0 = \sum_{k=k_1+1}^{k_2} \alpha_k$ . Then, we have + +$$ +\begin{array}{l} \Pi \left(U _ {2, +} ^ {\prime}\right) \leq \Pi \left(\delta \overline {{\lambda}} \geq \frac {\epsilon}{4}\right) \\ \leq \Pi \left(\bar {\lambda} \geq \frac {\epsilon}{4 C _ {p}} K\right) \\ \leq 2 ^ {\alpha_ {0}} \exp \left(- \frac {\rho \epsilon}{8 C _ {p}} K\right) \\ \leq \exp \left(C _ {2 0} \log 2 \cdot n ^ {\gamma - \frac {1}{3}} - C _ {2 1} n ^ {\gamma}\right) \\ \leq \exp \left(- C _ {2 1} / 2 \cdot n ^ {\gamma}\right), \\ \end{array} +$$ + +where the second inequality holds because $K \leq C_p \delta^{-1}$ , the third inequality follows from Chernoff bounds. Here, the fourth inequality holds because $K \geq C_p' n^\gamma$ and $\alpha_0 \leq K / J \cdot \overline{\alpha} \leq C_{20} \cdot n^{\gamma - 1/3}$ under (P2) with positive constants $C_{20}$ depending on $\overline{\alpha}$ , $p_{\min}$ , $p_{\max}$ and $\kappa$ , and $C_{21} = \rho \epsilon C_p' / (8C_p)$ . The last inequality holds for $n \geq (2C_{20} \log 2 / C_{21})^3$ . Combining this with (40), we have + +$$ +\mathbb {E} _ {\theta^ {*}} ^ {n} \left[ \Pi (U _ {2, +} ^ {\prime} \mid \mathbf {D} _ {n}) \right] \leq C _ {2 2} \exp \left(C _ {1 9} n ^ {\frac {1}{3}} \log n + (C _ {1} + 1) \epsilon_ {5} n ^ {\frac {2}{3}} - \frac {C _ {2 1}}{2} n ^ {\frac {2}{3}}\right) + \exp \left(- \frac {\epsilon_ {5} ^ {2}}{2 C _ {0} ^ {2}} n ^ {\frac {1}{3}}\right), +$$ + +where the inequality holds because $\gamma \geq 2/3$ with a positive constant $C_{22} = (C_{18}c_6)^{-1}$ . We choose $\epsilon_5 = (C_{21} / (6(C_1 + 1)))\wedge ((M_1\wedge (1 - M_2) / 2)\wedge B)$ . Then, for $n\geq (3C_{19} / C_{21})^3\vee (2C_{20}\log 2 / C_{21})^3$ , we have + +$$ +\begin{array}{l} \mathbb {E} _ {\theta^ {*}} ^ {n} \left[ \Pi \left(U _ {2, +} ^ {\prime} \mid \mathbf {D} _ {n}\right) \right] \leq C _ {2 2} \exp \left(- \frac {C _ {2 1}}{3} n ^ {\frac {2}{3}}\right) + \exp \left(- \frac {\epsilon_ {5} ^ {2}}{2 C _ {0} ^ {2}} n ^ {\frac {1}{3}}\right) \\ \leq C _ {2 3} \exp \left(- C _ {2 4} n ^ {\frac {1}{3}}\right), \tag {41} \\ \end{array} +$$ + +where the last inequality holds by positive constants $C_{23} = C_{22} + 1$ and $C_{24} = (C_{21} / 3)\wedge (\epsilon_5^2 /(2C_0^2))$ . Combining (39) and (41), for $n\geq (3C_{12}C_p / C_{16})^{3 / 2}\vee (3C_{19} / C_{21})^3\vee (2C_{20}\log 2 / C_{21})^3$ , we have + +$$ +\begin{array}{l} \mathbb {E} _ {\theta^ {*}} ^ {n} \left[ \Pi \left(U _ {2} \mid \mathbf {D} _ {n}\right) \right] \leq C _ {1 4} \exp \left(- C _ {1 7} n\right) + C _ {2 3} \exp \left(- C _ {2 4} n ^ {\frac {1}{3}}\right) \\ \leq C _ {2 5} \exp \left(- C _ {2 6} n ^ {\frac {1}{3}}\right), \tag {42} \\ \end{array} +$$ + +where $C_{25} = C_{14} + C_{23}$ and $C_{26} = C_{17} \wedge C_{24}$ are positive constants. + +By a similar argument as (ii), there exist positive constants $C_{27}$ and $C_{28}$ such that + +$$ +\left(\text {i i i}\right) \leq C _ {2 7} \exp \left(- C _ {2 8} n ^ {\frac {1}{3}}\right). \tag {43} +$$ + +Combining (36), (38), (42) and (43), we have + +$$ +\mathbb {E} _ {\theta^ {*}} ^ {n} \left[ \Pi \left(U ^ {c} \mid \mathbf {D} _ {n}\right) \right] \leq C _ {2 9} \exp \left(- C _ {3 0} n ^ {\frac {1}{3}}\right), +$$ + +where $C_{29} = C_{14} + C_{25} + C_{27}$ and $C_{30} = C_{15} \wedge C_{26} \wedge C_{28}$ are positive constants. By the Markov inequality, we have + +$$ +\mathbb {P} _ {\theta^ {*}} ^ {n} \left(\Pi \left(U ^ {c} \mid \mathbf {D} _ {n}\right) \geq C _ {2 9} \exp \left(- C _ {3 1} n ^ {\frac {1}{3}}\right)\right) < \exp \left(- C _ {3 1} n ^ {\frac {1}{3}}\right), +$$ + +where $C_{31} = C_{30} / 2$ . This concludes the proof for the case where $\gamma \geq 2 / 3$ + +![](images/60b3276def5669df4b330c46102ab43bc5911cbafcb228be4446c971575d4d35.jpg) + +# A.2 Proof of Theorem 3.1 + +Lemma A.6. Let $\Theta' = \{(\mathbf{S}_0, \beta) \in \Theta : S_{0,K} \geq M_1, S_{0,1} \leq M_2, \| \beta \|_2 \leq D\}$ , where $M_1, M_2$ and $D$ are some positive constants such that $0 < M_1 < M_2 < 1$ , and let $\mathcal{P}' = \{p_\theta : \theta \in \Theta'\}$ . Under the assumption (A2), there exist positive constants $C_1$ and $C_2$ depending only on $M_1, M_2, L$ and $D$ such that for every $\epsilon > 0$ , it holds that + +$$ +N (\epsilon , \mathcal {P} ^ {\prime}, \mathcal {D} _ {H}) \leq (C _ {1} / \epsilon + K) ^ {K} (C _ {2} / \epsilon) ^ {d}. +$$ + +Proof. Let $S_0' = \{S_0 = (S_{0,1}, \dots, S_{0,K}) : M_2 \geq S_{0,1} \geq \dots \geq S_{0,K} \geq M_1\}$ and $\mathcal{H}_0' = \{\Lambda_0 = (\Lambda_{0,1}, \dots, \Lambda_{0,K}) : \lambda_2 \leq \Lambda_{0,1} \leq \dots \leq \Lambda_{0,K} \leq \lambda_1\}$ , where $\lambda_2 = -\log M_2$ and $\lambda_1 = -\log M_1$ . Then, for any $\Lambda_0$ corresponding to the vector $S_0 \in S_0'$ , $\Lambda_0$ belongs to $\mathcal{H}_0'$ since $\Lambda_{0,k} = -\log S_{0,k} \in [\lambda_2, \lambda_1]$ for any $k = 1, \dots, K$ . For $\epsilon > 0$ , let + +$$ +\mathcal {H} _ {0, \epsilon} ^ {\prime} = \left\{\Lambda_ {0} \in \mathcal {H} _ {0} ^ {\prime}: \left(\Lambda_ {0, 1}, \dots , \Lambda_ {0, K}\right) = \left(m _ {1} \epsilon , \dots , m _ {K} \epsilon\right) \right. +$$ + +for some positive integers $m_{1},\ldots ,m_{K}$ satisfying $m_{1}\leq \dots \leq m_{K}\}$ + +Then, it is not difficult to show that $\mathcal{H}_{0,\epsilon}^{\prime}$ is an $\epsilon$ -cover of $\mathcal{H}_0^\prime$ with respect to $\| \cdot \|_{\infty}$ . Note that the cardinality of $\mathcal{H}_{0,\epsilon}^{\prime}$ is the number of $K$ -tuples of integers $(m_1,\ldots ,m_K)$ satisfying $\lfloor \lambda_1 / \epsilon \rfloor \leq m_1\leq \dots \leq m_K\leq \lfloor \lambda_2 / \epsilon \rfloor$ , which is given as $\binom{\lfloor\lambda_2/\epsilon\rfloor-\lfloor\lambda_1/\epsilon\rfloor+K}{K}$ based on simple combinatorics. Hence, we have $N(\epsilon ,\mathcal{H}_0^\prime ,\| \cdot \|_\infty)\leq \binom {\lfloor\lambda_2 / \epsilon \rfloor +K}{K}\leq (\lambda_2 / \epsilon +K)^K$ . Therefore, we have + +$$ +N \left(\epsilon , \mathcal {H} _ {0} ^ {\prime}, \| \cdot \| _ {\infty}\right) \leq \left(\lambda_ {2} / \epsilon + K\right) ^ {K}. \tag {44} +$$ + +Take any two parameters $\theta = (\mathbf{S}_0,\beta),\theta^{\prime} = (\mathbf{S}_{0}^{\prime},\beta^{\prime})\in \Theta^{\prime}$ . By Lemma C.1 and C.2, there exist positive constants $c_{1},c_{2}$ and $c_{3}$ , depending on $M_1,M_2,L$ and $D$ , such that for any $x\in \mathcal{X}$ and $p\in \mathcal{G}$ + +$$ +\begin{array}{l} \mathcal {D} _ {H} \left(p _ {\theta}, p _ {\theta^ {\prime}}\right) \leq c _ {1} \left\| H _ {\theta} - H _ {\theta^ {\prime}} \right\| _ {\infty} \\ \leq C _ {1} \left\| \boldsymbol {\Lambda} _ {0} - \boldsymbol {\Lambda} _ {0} ^ {\prime} \right\| _ {\infty} + C _ {2} \left\| \beta - \beta^ {\prime} \right\| _ {2}, \tag {45} \\ \end{array} +$$ + +where $C_1 = c_1c_2$ and $C_2 = c_1c_3$ . + +Let $m \coloneqq N(\epsilon/(2C_1), \mathcal{H}_0', \| \cdot \|_{\infty})$ and $l \coloneqq N(\epsilon/(2C_2), \mathcal{B}', \| \cdot \|_2)$ , where $\mathcal{B}' = \{\beta \in \mathbb{R}^d : \| \beta \|_2 \leq D\}$ . This definition implies that there exist $\Lambda_{0,1}, \ldots, \Lambda_{0,m} \in \mathcal{H}_0'$ such that for every $\Lambda_0 \in \mathcal{H}_0'$ , the inequality $\| \Lambda_0 - \Lambda_{0,i} \|_{\infty} < \epsilon/(2C_1)$ holds for some $1 \leq i \leq m$ . Similarly, there exist $\beta_1, \ldots, \beta_l \in \mathcal{B}'$ such that for every $\beta \in \mathcal{B}'$ , $\| \beta - \beta_j \|_2 < \epsilon/(2C_2)$ holds for some $1 \leq j \leq l$ . Let $\theta_{ij} = (\mathbf{S}_{0,i}, \beta_j) \in \Theta'$ , where $\mathbf{S}_{0,i}$ be the vector corresponding to $\Lambda_{0,i}$ for $i = 1, \ldots, m$ and $j = 1, \ldots, l$ . By (45), for any $\theta = (\mathbf{S}_0, \beta) \in \Theta'$ , there exists $\theta_{ij}$ for some $1 \leq i \leq m$ and $1 \leq j \leq l$ such that + +$$ +\mathcal {D} _ {H} \left(p _ {\theta}, p _ {\theta_ {i j}}\right) \leq C _ {1} \left\| \Lambda_ {0} - \Lambda_ {0, i} \right\| _ {\infty} + C _ {2} \left\| \beta - \beta_ {j} \right\| _ {2} \leq \epsilon . +$$ + +Consequently, the covering number $N(\epsilon, \mathcal{P}', \mathcal{D}_H)$ is of order $ml$ . Note that $m \leq (2C_1\lambda_2 / \epsilon + K)^K$ by (44). Furthermore, by Proposition C.2 of [16], $l \leq (6DC_2 / \epsilon)^d$ . Therefore, we have + +$$ +N (\epsilon , \mathcal {P} ^ {\prime}, \mathcal {D} _ {H}) \leq (C _ {3} / \epsilon + K) ^ {K} (C _ {4} / \epsilon) ^ {d}, +$$ + +where $C_3 = 2C_1\lambda_2$ and $C_4 = 6DC_2$ are positive constants depending only on $M_1, M_2, L$ and $D$ . + +![](images/94e3568601a3d397c7d8aac05b3ec8101ee498d8c8ca2e74218e412bed19c841.jpg) + +Proof of Theorem 3.1. First, we define the square Kullback-Leibler variation as $V_0(p,q) = \int (\log (p / q) - K(p / q))^2 dP$ . For every $\epsilon > 0$ , we define neighborhoods of $\theta^*$ by + +$$ +B \left(\theta^ {*}, \epsilon\right) = \left\{\theta \in \Theta : K \left(p _ {\theta^ {*}}, p _ {\theta}\right) \leq \epsilon^ {2}, V _ {0} \left(p _ {\theta^ {*}}, p _ {\theta}\right) \leq \epsilon^ {2} \right\}. +$$ + +We begin by checking the prior mass condition. Note that there exist constants $M_{1}$ and $M_2$ such that $0 < M_{1} \leq S_{0}^{*}(v) \leq M_{2} < 1$ for any $v \in [p_{\min}, p_{\max}]$ by assumption (A5). Let $U = \{(\mathbf{S}_0, \beta) \in \Theta : \| \mathbf{S}_0 - \mathbf{S}_0^* \|_{\infty} \vee \| \beta - \beta^* \|_2 < \epsilon_0\}$ be a neighborhood of $\theta^*$ , where $\epsilon_0$ is a positive constant that can be chosen as $\epsilon_0 = ((M_1 \wedge (1 - M_2)) / 2) \wedge B$ to ensure that $U \subseteq \Theta$ . By (28) in the proof of Lemma A.1, there exists a positive constant $C_0$ depending on $M_1, M_2, L$ and $B$ such that for any $\theta \in U$ , + +$$ +\left\| \log \frac {p _ {\theta^ {*}}}{p _ {\theta}} \right\| _ {\infty} \leq C _ {0}. +$$ + +By Lemma B.2 in [16], the uniformly bounded likelihood ratio implies that + +$$ +K \left(p _ {\theta^ {*}}, p _ {\theta}\right) \leq c _ {1} \mathcal {D} _ {H} ^ {2} \left(p _ {\theta^ {*}}, p _ {\theta}\right) \left\| \frac {p _ {\theta^ {*}}}{p _ {\theta}} \right\| _ {\infty} \leq C _ {1} \mathcal {D} _ {H} ^ {2} \left(p _ {\theta^ {*}}, p _ {\theta}\right), \tag {46} +$$ + +$$ +V _ {0} \left(p _ {\theta^ {*}}, p _ {\theta}\right) \leq c _ {2} \mathcal {D} _ {H} ^ {2} \left(p _ {\theta^ {*}}, p _ {\theta}\right) \left\| \frac {p _ {\theta^ {*}}}{p _ {\theta}} \right\| _ {\infty} \leq C _ {2} \mathcal {D} _ {H} ^ {2} \left(p _ {\theta^ {*}}, p _ {\theta}\right), +$$ + +where $C_1 = c_1 \exp(C_0)$ and $C_2 = c_2 \exp(C_0)$ for universal constants $c_1$ and $c_2$ . By Lemma C.1 and Lemma C.2, there exist positive constants $c_3, c_4$ and $c_5$ , depending on $M_1, M_2, L$ and $B$ such that for any $\theta \in U$ , + +$$ +\begin{array}{l} \mathcal {D} _ {H} \left(p _ {\theta^ {*}}, p _ {\theta}\right) \leq c _ {3} \| H _ {\theta^ {*}} - H _ {\theta} \| _ {\infty} \\ \leq C _ {3} \left\| \boldsymbol {\Lambda} _ {0} - \boldsymbol {\Lambda} _ {0} ^ {*} \right\| _ {\infty} + C _ {4} \left\| \beta - \beta^ {*} \right\| _ {2}, \\ \end{array} +$$ + +where $C_3 = c_3c_4$ and $C_4 = c_3c_5$ . Let $\Theta_n = \{\theta \in \Theta : \| \pmb{\Lambda}_0 - \pmb{\Lambda}_0^*\|_\infty \leq C_5\epsilon_n, \| \beta - \beta^*\|_2 \leq C_6\epsilon_n\}$ , where $C_5 = 1 / (2C_3\sqrt{C_1 \vee C_2})$ and $C_6 = 1 / (2C_4\sqrt{C_1 \vee C_2})$ . Combining the last two displays, we have + +$$ +\Theta_ {n} \cap U \subseteq B \left(\theta^ {*}, \epsilon_ {n}\right) \cap U. +$$ + +Since $\| \mathbf{S}_0 - \mathbf{S}_0^*\|_{\infty}\leq \| \pmb {\Lambda}_0 - \pmb {\Lambda}_0^*\|_{\infty}$ , for sufficiently large $n$ such that $\epsilon_{n} < \epsilon_{0} / (C_{5}\lor C_{6})$ , it follows that $\Theta_n\subset U$ , implying $\Theta_{n}\cap U = \Theta_{n}$ . Thus, we see that + +$$ +\begin{array}{l} \Pi \left(B \left(\theta^ {*}, \epsilon_ {n}\right)\right) \geq \Pi \left(B \left(\theta^ {*}, \epsilon_ {n}\right) \cap U\right) \\ \geq \Pi \left(\left\| \boldsymbol {\Lambda} _ {0} - \boldsymbol {\Lambda} _ {0} ^ {*} \right\| _ {\infty} \leq C _ {5} \epsilon_ {n}\right) \cdot \Pi \left(\left\| \beta - \beta^ {*} \right\| _ {2} \leq C _ {6} \epsilon_ {n}\right). \tag {47} \\ \end{array} +$$ + +By Lemma C.3 with the specified prior (3), the first term in the right side of the last display is bounded below by $C_7\exp (-C_8K - C_9K\log_{-}(C_5\epsilon_n))$ , where $C_7,C_8$ and $C_9$ are positive constants depending on $p_{\mathrm{min}}$ , $p_{\mathrm{max}}$ , $M_{1}$ , $M_2$ , $\underline{\alpha}$ , $\overline{\alpha}$ and $\rho$ . Let $V_{d}(R)$ denote the volume of a $d$ -dimensional $L^2$ norm ball of radius $R > 0$ . The closed form of $V_{d}(R)$ is given by $V_{d}(R) = \pi^{d / 2} / \Gamma (\frac{d}{2} +1)\cdot R^{d}$ where $\Gamma$ is the gamma function. Note that $\Gamma (\frac{d}{2} +1)\leq \Gamma (d + 1) = d!\leq d^d$ for $d\geq 1$ . Then, the second term in the right side of the last display is bounded below by $C_{10}(\sqrt{\pi} /d)^{d}(C_{6}\epsilon_{n})^{d}$ where $C_{10}$ is the lower bound of the prior on a neighborhood of $\beta^{*}$ . Therefore, we have + +$$ +\begin{array}{l} \Pi \left(B \left(\theta^ {*}, \epsilon_ {n}\right)\right) \geq C _ {7} C _ {1 0} \exp \left(- C _ {8} K - C _ {9} K \log_ {-} \left(C _ {5} \epsilon_ {n}\right)\right) \cdot (\sqrt {\pi} / d) ^ {d} \left(C _ {6} \epsilon_ {n}\right) ^ {d} \\ \geq C _ {7} C _ {1 0} \exp \left(- C _ {8} K - C _ {9} K \log_ {-} \left(C _ {5} \epsilon_ {n}\right) - d \log d - d \log_ {-} \left(C _ {6} \epsilon_ {n}\right)\right) \\ \geq C _ {7} C _ {1 0} \exp \left(- C _ {8} C _ {p} n ^ {\gamma} - C _ {9} C _ {p} n ^ {\gamma} \log_ {-} \left(C _ {5} \epsilon_ {n}\right) - d \log d - d \log_ {-} \left(C _ {6} \epsilon_ {n}\right)\right) \\ \geq \exp (- C _ {1 1} n \epsilon_ {n} ^ {2}), \\ \end{array} +$$ + +where the third inequality holds because $K \geq C_p n^\gamma$ with a positive constant $C_p$ depending on $p_{\min}$ , $p_{\max}$ and $\kappa$ as defined by $K$ , and the last inequality holds by a positive constant $C_{11} = |\log (C_7C_{10})| + C_8C_p + C_9C_p(|\log C_5| + 1) + |\log C_6| + 2$ because $\log_-(C\epsilon_n) \leq |\log C| + \log n$ holds for any $C > 0$ and $n^\gamma \log n \leq n\epsilon_n^2$ . Thus, by Lemma 10 of [15], there exists an event $\Omega_n$ such that $\mathbb{P}_{\theta^*}^n(\Omega_n) \geq 1 - 1 / (n\epsilon_n^2)$ , and in $\Omega_n$ , + +$$ +\int \exp \left(\ell_ {n} (\theta) - \ell_ {n} \left(\theta^ {*}\right)\right) d \Pi (\theta) \geq \exp \left(- \left(C _ {1 1} + 2\right) n \epsilon_ {n} ^ {2}\right). \tag {48} +$$ + +By the Kullback-Leibler inequality, note that $\mathbb{E}_{\theta^{*}}[\ell (\theta)]$ is maximized at $\theta = \theta^{*}$ , meaning its first derivative at $\theta^{*}$ is equal to 0. Additionally, for $\theta = (\mathbf{S}_0,\beta)\in U$ , note that $\mathbf{S}_0$ is uniformly bounded away from 0 and 1, and $\beta$ is bounded by assumptions (A1) and (A5). Since the covariate has bounded support by assumption (A2), by a Taylor expansion, there exists a positive constant $c_{0}$ depending on $M_1,M_2,L$ and $B$ such that for any $\theta \in U$ , we have + +$$ +c _ {0} \mathcal {D} _ {\mathbb {Q}} ^ {2} (\theta , \theta^ {*}) \leq \mathbb {E} _ {\theta^ {*}} \left[ \ell (\theta^ {*}) \right] - \mathbb {E} _ {\theta^ {*}} \left[ \ell (\theta) \right] = K \left(p _ {\theta^ {*}}, p _ {\theta}\right). +$$ + +Combining this with (46), we have + +$$ +C _ {H} \mathcal {D} _ {\mathbb {Q}} \left(\theta , \theta^ {*}\right) \leq D _ {H} \left(p _ {\theta^ {*}}, p _ {\theta}\right), \tag {49} +$$ + +where $C_H = \sqrt{c_0 / C_1}$ . Then, we have + +$$ +\begin{array}{l} \mathbb {E} _ {\theta^ {*}} ^ {n} \left[ \Pi (\mathcal {D} _ {\mathbb {Q}} (\theta , \theta^ {*}) \geq M J \epsilon_ {n} \mid \mathbf {D} _ {n}) \mathbb {1} \{\Omega_ {n} \} \right] \\ \leq \mathbb {E} _ {\theta^ {*}} ^ {n} [ \Pi (\left\{\mathcal {D} _ {\mathbb {Q}} (\theta , \theta^ {*}) \geq M J \epsilon_ {n} \right\} \cap U \mid \mathbf {D} _ {n}) \mathbb {1} \left\{\Omega_ {n} \right\} ] + \mathbb {E} _ {\theta^ {*}} ^ {n} [ \Pi (U ^ {c} \mid \mathbf {D} _ {n}) ] \\ \leq \mathbb {E} _ {\theta^ {*}} ^ {n} \left[ \Pi \left(\Gamma_ {n} \mid \mathbf {D} _ {n}\right) \mathbb {1} \left\{\Omega_ {n} \right\} \right] + c _ {7} \exp \left(- c _ {8} n\right), \tag {50} \\ \end{array} +$$ + +where $\Gamma_{n} = \{\theta \in U:\mathcal{D}_{H}(p_{\theta^{*}},p_{\theta})\geq C_{H}MJ\epsilon_{n}\}$ for large constants $M$ and $J$ to be chosen later, and the last inequality holds for $n\geq c_6$ by Lemma A.1 with positive constants $c_{6},c_{7}$ and $c_{8}$ depending on $M_1,M_2,L,B,p_{\mathrm{min}},p_{\mathrm{max}},\underline{\alpha},\overline{\alpha},\rho$ and $\kappa$ + +Define $\mathcal{P}_U = \{p_\theta : \theta \in U\}$ and $N_n^* = \sup_{\epsilon > \epsilon_n} N(\epsilon / 36, \{p_\theta \in \mathcal{P}_U : \theta \in \Gamma_n\}, \mathcal{D}_H)$ . By Lemma A.6, there exist positive constants $C_{12}$ and $C_{13}$ depending on $M_1, M_2, L$ , and $B$ such that + +$$ +\begin{array}{l} N _ {n} ^ {*} \leq N \left(\epsilon_ {n} / 3 6, \mathcal {P} _ {n} ^ {\text {S i e v e}}, \mathcal {D} _ {H}\right) \\ \leq \left(3 6 C _ {1 2} / \epsilon_ {n} + K\right) ^ {K} \left(3 6 C _ {1 3} / \epsilon_ {n}\right) ^ {d} \\ \leq \exp (K \cdot \log (3 6 C _ {1 2} / \epsilon_ {n} + K) + d \cdot \log (3 6 C _ {1 3} / \epsilon_ {n})) \\ \leq \exp \left(C _ {1 4} n \epsilon_ {n} ^ {2}\right), \tag {51} \\ \end{array} +$$ + +where the last inequality holds by a positive constant $C_{14}$ depending on $C_{12}, C_{13}, p_{\mathrm{min}}, p_{\mathrm{max}}$ and $\kappa$ . In addition, by Lemma 2 of [15] and Lemma 9 of [15], applied with $\epsilon = C_H M \epsilon_n$ , where $C_H M \geq 2$ , there exist tests $\phi_n$ that satisfy + +$$ +\mathbb {E} _ {\theta^ {*}} ^ {n} \phi_ {n} \leq N _ {n} ^ {*} \exp \left(- \frac {1}{2} C _ {H} ^ {2} M ^ {2} n \epsilon_ {n} ^ {2}\right) \frac {1}{1 - \exp \left(- \frac {1}{2} C _ {H} ^ {2} M ^ {2} n \epsilon_ {n} ^ {2}\right)}, \tag {52} +$$ + +$$ +\sup _ {\theta \in \Gamma_ {n}} \mathbb {E} _ {\theta} ^ {n} (1 - \phi_ {n}) \leq \exp \left(- \frac {1}{2} C _ {H} ^ {2} M ^ {2} J ^ {2} n \epsilon_ {n} ^ {2}\right), +$$ + +for any $J \geq 1$ . Then, by (48), the first term of (50) is upper bounded by + +$$ +\mathbb {E} _ {\theta^ {*}} ^ {n} \left[ \Pi \left(\Gamma_ {n} \mid \mathbf {D} _ {n}\right) \mathbb {1} \left\{\Omega_ {n} \right\} \right] \leq \mathbb {E} _ {\theta^ {*}} ^ {n} \phi_ {n} + \exp \left(\left(C _ {1 1} + 2\right) n \epsilon_ {n} ^ {2}\right) \sup _ {\theta \in \Gamma_ {n}} \mathbb {E} _ {\theta} ^ {n} \left(1 - \phi_ {n}\right). \tag {53} +$$ + +If $M$ is sufficiently large to ensure that $C_H^2 M^2 / 2 - C_{14} > C_H^2 M^2 / 4$ , by combining (51) and the first line of (52), we have + +$$ +\begin{array}{l} \mathbb {E} _ {\theta^ {*}} ^ {n} \phi_ {n} \leq \exp \left(\left(C _ {1 4} - \frac {1}{2} C _ {H} ^ {2} M ^ {2}\right) n \epsilon_ {n} ^ {2}\right) \frac {1}{1 - \exp \left(- \frac {1}{2} C _ {H} ^ {2} M ^ {2} n \epsilon_ {n} ^ {2}\right)} \\ \leq C _ {1 5} \exp \left(- \frac {1}{4} C _ {H} ^ {2} M ^ {2} n \epsilon_ {n} ^ {2}\right), \\ \end{array} +$$ + +where $C_{15} = (1 - \exp (-2C_{14}))^{-1}$ is a positive constant. If we set $J = 1$ and choose $M$ to be sufficiently large such that $C_H^2 M^2 /2 - (C_{11} + 2) > C_H^2 M^2 /4$ , by the second line of (52), the second term in the right hand side of (53) is bounded by + +$$ +\exp \left(- \frac {1}{4} C _ {H} ^ {2} M ^ {2} n \epsilon_ {n} ^ {2}\right). +$$ + +Therefore, if we choose $M$ to be sufficiently large such that $M \geq 2\sqrt{(C_{11} + 2) \vee C_{14}} / C_H$ , by combining the preceding two displays, (50) and (53), we have + +$$ +\begin{array}{l} \mathbb {E} _ {\theta^ {*}} ^ {n} \left[ \Pi \left(\mathcal {D} _ {\mathbb {Q}} (\theta , \theta^ {*}) \geq M \epsilon_ {n} \mid \mathbf {D} _ {n}\right) \mathbb {I} \left\{\Omega_ {n} \right\} \right] \leq \left(C _ {1 5} + 1\right) \exp \left(- C _ {1 6} n \epsilon_ {n} ^ {2}\right) + c _ {7} \exp (- c _ {8} n) \\ \leq C _ {1 7} \exp \left(- \left(C _ {1 6} \wedge c _ {8}\right) n \epsilon_ {n} ^ {2}\right), \\ \end{array} +$$ + +where $C_{17} = C_{15} + c_7 + 1$ and $C_{16} = (C_{11} + 2) \lor C_{14}$ are positive constants. An application of the Markov inequality yields that + +$$ +\mathbb {P} _ {\theta^ {*}} ^ {n} \left(\Pi (\mathcal {D} _ {\mathbb {Q}} (\theta , \theta^ {*}) \geq M \epsilon_ {n} \mid \mathbf {D} _ {n}) \mathbb {1} \{\Omega_ {n} \} \geq C _ {1 7} \exp \left(- C _ {1 8} n \epsilon_ {n} ^ {2}\right)\right) \leq \exp \left(- C _ {1 8} n \epsilon_ {n} ^ {2}\right), +$$ + +where $C_{18} = (C_{16} \wedge c_8) / 2$ be a positive constant. Note that it is easy to show that + +$$ +\begin{array}{l} \mathbb {P} _ {\theta^ {*}} ^ {n} \left(\left\{\Pi \left(\mathcal {D} _ {\mathbb {Q}} (\theta , \theta^ {*}) \geq M \epsilon_ {n} \mid \mathbf {D} _ {n}\right) \geq C _ {1 7} \exp \left(- C _ {1 8} n \epsilon_ {n} ^ {2}\right) \right\} \cap \Omega_ {n}\right) \\ \leq \mathbb {P} _ {\theta^ {*}} ^ {n} \left(\Pi (\mathcal {D} _ {\mathbb {Q}} (\theta , \theta^ {*}) \geq M \epsilon_ {n} \mid \mathbf {D} _ {n}) \mathbb {1} \{\Omega_ {n} \} \geq C _ {1 7} \exp \left(- C _ {1 8} n \epsilon_ {n} ^ {2}\right)\right). \\ \end{array} +$$ + +Combining the last two displays and (48), we have + +$$ +\begin{array}{l} \mathbb {P} _ {\theta^ {*}} ^ {n} \left(\left\{\Pi (\mathcal {D} _ {\mathbb {Q}} (\theta , \theta^ {*}) \geq M \epsilon_ {n} | \mathbf {D} _ {n}) \geq C _ {1 7} \exp (- C _ {1 8} n \epsilon_ {n} ^ {2}) \right\}\right) \\ \leq \mathbb {P} _ {\theta^ {*}} ^ {n} \left(\left\{\Pi (\mathcal {D} _ {\mathbb {Q}} (\theta , \theta^ {*}) \geq M \epsilon_ {n} | \mathbf {D} _ {n}) \geq C _ {1 7} \exp (- C _ {1 8} n \epsilon_ {n} ^ {2}) \right\} \cap \Omega_ {n}\right) + \mathbb {P} _ {\theta^ {*}} ^ {n} (\Omega_ {n}) \\ \leq \mathbb {P} _ {\theta^ {*}} ^ {n} \left(\Pi (\mathcal {D} _ {\mathbb {Q}} (\theta , \theta^ {*}) \geq M \epsilon_ {n} | \mathbf {D} _ {n}) \mathbb {1} \left\{\Omega_ {n} \right\} \geq C _ {1 7} \exp \left(- C _ {1 8} n \epsilon_ {n} ^ {2}\right)\right) + \mathbb {P} _ {\theta^ {*}} ^ {n} (\Omega_ {n}) \\ \leq \exp \left(- C _ {1 8} n \epsilon_ {n} ^ {2}\right) + \frac {1}{n \epsilon_ {n} ^ {2}}. \\ \end{array} +$$ + +Thus, we have that with probability at least $1 - \left(\exp \left(-C_{18} n \epsilon_n^2\right) + 1 / n \epsilon_n^2\right)$ , + +$$ +\Pi (\mathcal {D} _ {\mathbb {Q}} (\theta , \theta^ {*}) \geq M \epsilon_ {n} | \mathbf {D} _ {n}) < C _ {1 7} \exp \left(- C _ {1 8} n \epsilon_ {n} ^ {2}\right). +$$ + +If we fix $M = \lceil 2\sqrt{(C_{11} + 2)\lor C_{14}} /C_H\rceil$ , then the proof is complete. + +![](images/4e89f9d3c2517c606a88ac01316ceac9eaa569ef59aab2e9c0ac94f28d3e8a9d.jpg) + +# A.3 Proof of Theorem 3.2 + +Lemma A.7. Let $\Theta' = \{(\mathbf{S}_0, \beta) \in \Theta : S_{0,K} \geq M_1, S_{0,1} \leq M_2, \| \beta \|_2 \leq D\}$ , where $M_1, M_2$ and $D$ are some positive constants such that $0 < M_1 < M_2 < 1$ , and let $\mathcal{P}' = \{p_\theta : \theta \in \Theta'\}$ . Under the assumption (A2), there exist positive constants $C_1$ and $C_2$ depending only on $M_1, M_2, L$ and $D$ such that for every $\epsilon > 0$ , it holds that + +$$ +N (\epsilon , \mathcal {P} ^ {\prime}, \mathcal {D} _ {H}) \leq \exp (C _ {1} / \epsilon) (C _ {2} / \epsilon) ^ {d}. +$$ + +Proof. Let $\mathcal{F}_0$ be the collection of monotone functions $f:(p_{\mathrm{min}},p_{\mathrm{max}})\to [\lambda_2,\lambda_1]$ , where $\lambda_{1} = -\log M_{1}$ and $\lambda_{2} = -\log M_{2}$ . Additionally, let $\Lambda_0$ denote the cumulative hazard functions with respect to the baseline complementary c.d.f. $S_{0}$ . Then, for any $S_{0}$ corresponding to the vector $\mathbf{S}_0\in \{\mathbf{S}_0\in S_0:S_{0,K}\geq M_1,S_{0,1}\leq M_2\}$ , $\Lambda_0 = -\log S_0$ belongs to $\mathcal{F}_0$ . + +Take any two parameters $\theta = (\mathbf{S}_0, \beta)$ and $\theta' = (\mathbf{S}_0', \beta') \in \Theta'$ . By Lemma C.1 and Lemma C.2, we have + +$$ +\begin{array}{l} \mathcal {D} _ {H} \left(p _ {\theta}, p _ {\theta^ {\prime}}\right) \leq C _ {0} \left[ \mathbb {E} _ {X, P} \left| H _ {\theta} (X, P) - H _ {\theta^ {\prime}} (X, P) \right| ^ {2} \right] ^ {1 / 2} \\ \leq C _ {1} \| \mathbf {S} _ {0} - \mathbf {S} _ {0} ^ {\prime} \| _ {2, \mathbb {Q}} + C _ {2} \| \beta - \beta^ {\prime} \| _ {2} \\ \leq C _ {1} \| \boldsymbol {\Lambda} _ {0} - \boldsymbol {\Lambda} _ {0} ^ {\prime} \| _ {2, \mathbb {Q}} + C _ {2} \| \beta - \beta^ {\prime} \| _ {2}, \tag {54} \\ \end{array} +$$ + +where $\mathbb{E}_{X,P}$ denotes the expectation with respect to the covariate $X$ and the price $P$ , and $C_0, C_1$ and $C_2$ are positive constants depending on $M_1, M_2, L$ and $D$ . + +Let $m \coloneqq N(\epsilon/(2C_1), \mathcal{F}_0, \| \cdot \|_{2,\mathbb{Q}})$ and $l \coloneqq N(\epsilon/(2C_2), \mathcal{B}', \| \cdot \|_2)$ , where $\mathcal{B}' = \{\beta \in \mathbb{R}^d : \| \beta \|_2 \leq D\}$ . This definition implies that there exist $\Lambda_{0,1}, \ldots, \Lambda_{0,m} \in \mathcal{F}_0$ such that for every $\Lambda_0 \in \mathcal{F}_0$ , the inequality $\| \Lambda_0 - \Lambda_{0,i} \|_{2,\mathbb{Q}} < \epsilon/(2C_1)$ holds for some $1 \leq i \leq m$ . Similarly, there exist $\beta_1, \ldots, \beta_l \in \mathcal{B}'$ such that for every $\beta \in \mathcal{B}'$ , $\| \beta - \beta_j \|_2 < \epsilon/(2C_2)$ holds for some $1 \leq j \leq l$ . Let $\mathbf{S}_{0,i} = (S_{0,i}(g_1), \ldots, S_{0,i}(g_K))$ where $S_{0,i} = \exp(-\Lambda_{0,i})$ and $\theta_{ij} = (\mathbf{S}_{0,i}, \beta_j) \in \Theta'$ for $i = 1, \ldots, m$ and $j = 1, \ldots, l$ . By (54), for any $\theta = (\mathbf{S}_0, \beta) \in \Theta'$ , there exists $\theta_{ij}$ for some $i \in \{1, \ldots, m\}$ and $j \in \{1, \ldots, l\}$ such that + +$$ +\mathcal {D} _ {H} \left(p _ {\theta}, p _ {\theta_ {i j}}\right) \leq C _ {1} \| \Lambda_ {0} - \Lambda_ {0, i} \| _ {2, \mathbb {Q}} + C _ {2} \| \beta - \beta_ {j} \| _ {2} \leq \epsilon . +$$ + +Consequently, the covering number $N(\epsilon, \mathcal{P}', \mathcal{D}_H)$ is of order $ml$ . By Proposition C.8 of [16], note that $m \leq N_{[l]}(\epsilon/(2C_1), \mathcal{F}_0, \| \cdot \|_{2,\mathbb{Q}}) \leq \exp(2C_1C_3\lambda_1/\epsilon)$ , where $C_3$ is a universal constant. Furthermore, by Proposition C.2 of [16], $l \leq (6DC_2/\epsilon)^d$ . Therefore, we have + +$$ +N (\epsilon , \mathcal {P} ^ {\prime}, \mathcal {D} _ {H}) \leq \exp (C _ {4} / \epsilon) (C _ {5} / \epsilon) ^ {d}, +$$ + +where $C_4 = 2C_1C_3\lambda_1$ and $C_5 = 6DC_2$ are positive constants depending only on $M_1, M_2, L$ and $D$ . + +Proof of Theorem 3.2. Recall the grid support $\mathcal{G} = \{g_k : k = 1, \dots, K\}$ , where each grid point $g_k$ is defined as $g_k = p_{\min} + k\delta$ with $\delta = \kappa n^{-\gamma}$ for some constant $\kappa > 0$ . Let $J = \lceil (p_{\max} - p_{\min}) / (\kappa \epsilon_n) \rceil$ and define $(k_1, \ldots, k_J)$ as a subsequence of $[K]$ such that $p_{\min} + (j - 1)\kappa \epsilon_n < g_{k_j} \leq p_{\min} + (j - 1)\kappa \epsilon_n + \delta$ for $j = 1, \ldots, J - 1$ , and set $k_J = K$ . Suppose that $|S_{0,k_j} - S_{0,k_j}^*| < \epsilon_n$ for every $j = 1, \ldots, J$ . Then, for any $k \in [K]$ with $k_j \leq k < k_{j+1}$ for $j = 1, \ldots, J-2$ , we have + +$$ +\begin{array}{l} S _ {0, k} ^ {*} - S _ {0, k} \leq S _ {0, k} ^ {*} - S _ {0, k _ {j + 1}} \\ \leq \left| S _ {0, k _ {j + 1}} ^ {*} - S _ {0, k _ {j + 1}} \right| + \left| S _ {0, k} ^ {*} - S _ {0, k _ {j + 1}} ^ {*} \right| \\ < \epsilon_ {n} + L _ {0} | g _ {k} - g _ {k _ {j + 1}} | \\ \leq \epsilon_ {n} + L _ {0} \left| g _ {k _ {j}} - g _ {k _ {j + 1}} \right| \\ \leq \epsilon_ {n} + L _ {0} (\kappa \epsilon_ {n} + \delta) \\ \leq (2 L _ {0} \kappa + 1) \epsilon_ {n}, \\ \end{array} +$$ + +where the third inequality holds by our assumption and $L_{0}$ -Lipschitz continuity of $S_0^*$ , with $L_{0}$ being a positive constant because (A5) is assumed, the fifth inequality holds by the definition of $k_{j}$ , and the last inequality holds because $\delta \leq \kappa \epsilon_{n}$ . Note that $|g_{k_{J - 1}} - g_{k_J}| < 2\kappa \epsilon_n$ by the definition of $J$ . Then, for any $k\in [K]$ with $k_{J - 1}\leq k\leq k_J$ , we have + +$$ +S _ {0, k} ^ {*} - S _ {0, k} < (2 L _ {0} \kappa + 1) \epsilon_ {n}. +$$ + +Combining the preceding two displays, there exists a positive constant $C_0 = 2L_0\kappa +1$ such that $S_{0,k}^{*} - S_{0,k} < C_{0}\epsilon_{n}$ for any $k\in [K]$ . Similarly, for any $k\in [K]$ , we have $S_{0,k} - S_{0,k}^{*} < C_{0}\epsilon_{n}$ . Therefore, we have + +$$ +\Pi \left(\left\| \mathbf {S} _ {0} - \mathbf {S} _ {0} ^ {*} \right\| _ {\infty} \leq C _ {0} \epsilon_ {n}\right) \geq \Pi \left(\left\| \mathbf {S} _ {0} - \mathbf {S} _ {0} ^ {*} \right\| _ {\infty , J} \leq \epsilon_ {n}\right), \tag {55} +$$ + +where $\| \mathbf{S}_0 - \mathbf{S}_0^*\|_{\infty ,J} = \max_{j = 1,\ldots ,J}|S_{0,k_j} - S_{0,k_j}^* |.$ + +Note that there exist constants $M_1$ and $M_2$ such that $0 < M_1 \leq S_0^*(v) \leq M_2 < 1$ for any $v \in [p_{\min}, p_{\max}]$ by assumption (A5). Let $U = \{(\mathbf{S}_0, \beta) \in \Theta : \| \mathbf{S}_0 - \mathbf{S}_0^* \|_\infty \vee \| \beta - \beta^* \|_2 < \epsilon_0\}$ be a neighborhood of $\theta^*$ , where $\epsilon_0$ is a positive constant that can be chosen as $\epsilon_0 = ((M_1 \wedge (1 - M_2)) / 2) \wedge B$ to ensure that $U \subseteq \Theta$ . Given in (47) of Theorem 3.1, there exist positive constants $C_1$ and $C_2$ depending on $M_1, M_2, L$ and $B$ such that for sufficiently large $n$ with $\epsilon_n < \epsilon_0 / (C_1 \vee C_2)$ , + +$$ +\Pi \left(B \left(\theta^ {*}, \epsilon_ {n}\right)\right) \geq \Pi \left(\| \boldsymbol {\Lambda} _ {0} - \boldsymbol {\Lambda} _ {0} ^ {*} \| _ {\infty} \leq C _ {1} \epsilon_ {n}\right) \cdot \Pi \left(\| \beta - \beta^ {*} \| _ {2} \leq C _ {2} \epsilon_ {n}\right). +$$ + +Note that for $\theta \in U$ , $\| \mathbf{S}_0 - \mathbf{S}_0^*\|_{\infty} \asymp \| \boldsymbol{\Lambda}_0 - \boldsymbol{\Lambda}_0^*\|_{\infty}$ and $\| \mathbf{S}_0 - \mathbf{S}_0^*\|_{\infty,J} \asymp \| \boldsymbol{\Lambda}_0 - \boldsymbol{\Lambda}_0^*\|_{\infty,J}$ , where constants in $\asymp$ depend on $M_1, M_2, L$ and $B$ . Thus, the inequality (55) implies that + +$$ +\Pi \left(\left\| \boldsymbol {\Lambda} _ {0} - \boldsymbol {\Lambda} _ {0} ^ {*} \right\| _ {\infty} \leq C _ {0} ^ {\prime} C _ {0} \epsilon_ {n}\right) \geq \Pi \left(\left\| \boldsymbol {\Lambda} _ {0} - \boldsymbol {\Lambda} _ {0} ^ {*} \right\| _ {\infty , J} \leq \epsilon_ {n}\right), +$$ + +where $C_0'$ is a positive constant depending on $M_1, M_2, L$ and $B$ . Combining the preceding two displays, for sufficiently large $n$ such that $\epsilon_n < \epsilon_0 / (C_1 \vee C_2)$ and $\delta \leq \kappa (C_0'C_0)^{-1}C_1\epsilon_n$ , we have + +$$ +\Pi \left(B \left(\theta^ {*}, \epsilon_ {n}\right)\right) \geq \Pi \left(\left\| \boldsymbol {\Lambda} _ {0} - \boldsymbol {\Lambda} _ {0} ^ {*} \right\| _ {\infty , J} \leq \left(C _ {0} ^ {\prime} C _ {0}\right) ^ {- 1} C _ {1} \epsilon_ {n}\right) \cdot \Pi \left(\left\| \beta - \beta^ {*} \right\| _ {2} \leq C _ {2} \epsilon_ {n}\right). +$$ + +Therefore, we have + +$$ +\begin{array}{l} \Pi \left(B \left(\theta^ {*}, \epsilon_ {n}\right)\right) \geq C _ {3} C _ {6} \exp \left(- C _ {4} J - C _ {5} J \log_ {-} \left(\left(C _ {0} ^ {\prime} C _ {0}\right) ^ {- 1} C _ {1} \epsilon_ {n}\right) - d \log d - d \log_ {-} \left(C _ {2} \epsilon_ {n}\right)\right) \\ \geq C _ {3} C _ {6} \exp \left(- C _ {4} C _ {7} \epsilon_ {n} ^ {- 1} - C _ {5} C _ {7} \epsilon_ {n} ^ {- 1} \log_ {-} \left(\left(C _ {0} ^ {\prime} C _ {0}\right) ^ {- 1} C _ {1} \epsilon_ {n}\right) - d \log d - d \log_ {-} \left(C _ {2} \epsilon_ {n}\right)\right) \\ \geq \exp (- C _ {8} n \epsilon_ {n} ^ {2}), \\ \end{array} +$$ + +where the first inequality holds by Lemma C.3 with positive constants $C_3$ , $C_4$ , $C_5$ depending on $p_{\mathrm{min}}$ , $p_{\mathrm{max}}$ , $M_1$ , $M_2$ , $\underline{\alpha}$ , $\overline{\alpha}$ and $\rho$ , and $C_6$ serving as the lower bound of the prior on a neighborhood of $\beta^*$ , the second inequality holds because $J \leq C_7 \epsilon_n^{-1}$ holds by the definition of $J$ with a positive constant $C_7$ depending on $p_{\mathrm{min}}$ , $p_{\mathrm{max}}$ and $\kappa$ , and the third inequality holds by a positive constant $C_8 = |\log (C_3 C_6)| + C_4 C_7 + C_5 C_7(|\log ((C_0' C_0)^{-1} C_1)| + 1) + |\log C_2| + 2$ because $\epsilon_n^{-1} \log_-(\epsilon_n) \leq n \epsilon_n^2$ and $d \log_-(\epsilon_n) \leq n \epsilon_n^2$ . Thus, by Lemma 10 of [15], there exists an event $\Omega_n$ such that $\mathbb{P}_{\theta^*}^n(\Omega_n) \geq 1 - 1 / n \epsilon_n^2$ , and in $\Omega_n$ , + +$$ +\int \exp \left(\ell_ {n} (\theta) - \ell_ {n} \left(\theta^ {*}\right)\right) d \Pi (\theta) \geq \exp \left(- \left(C _ {8} + 2\right) n \epsilon_ {n} ^ {2}\right). \tag {56} +$$ + +By (49) in the proof of Theorem 3.1, there exists a positive constant $C_H$ depending on $M_1, M_2, L$ and $B$ such that $C_H \mathcal{D}_{\mathbb{Q}}(\theta, \theta^*) \leq \mathcal{D}_H(p_{\theta^*}, p_\theta)$ for $\theta \in U$ . Then, we have + +$$ +\begin{array}{l} \mathbb {E} _ {\theta^ {*}} ^ {n} \left[ \Pi \left(\mathcal {D} _ {\mathbb {Q}} (\theta , \theta^ {*}) \geq M J \epsilon_ {n} \mid \mathbf {D} _ {n}\right) \mathbb {1} \left\{\Omega_ {n} \right\} \right] \\ \leq \mathbb {E} _ {\theta^ {*}} ^ {n} \left[ \Pi \left(\left\{\mathcal {D} _ {\mathbb {Q}} (\theta , \theta^ {*}) \geq M J \epsilon_ {n} \right\} \cap U \mid \mathbf {D} _ {n}\right) \mathbb {1} \left\{\Omega_ {n} \right\} \right] + \mathbb {E} _ {\theta^ {*}} ^ {n} \left[ \Pi \left(U ^ {c} \mid \mathbf {D} _ {n}\right) \right] \\ \leq \mathbb {E} _ {\theta^ {*}} ^ {n} \left[ \Pi \left(\Gamma_ {n} \mid \mathbf {D} _ {n}\right) \mathbb {1} \left\{\Omega_ {n} \right\} \right] + \mathbb {E} _ {\theta^ {*}} ^ {n} \left[ \Pi \left(U ^ {c} \mid \mathbf {D} _ {n}\right) \right], \tag {57} \\ \end{array} +$$ + +where $\Gamma_{n} = \{\theta \in U:\mathcal{D}_{H}(p_{\theta^{*}},p_{\theta})\geq C_{H}MJ\epsilon_{n}\}$ for large constants $M$ and $J$ to be chosen later. + +Define $\mathcal{P}_U = \{p_\theta : \theta \in U\}$ and $N_n^* = \sup_{\epsilon > \epsilon_n} N(\epsilon / 36, \{p_\theta \in \mathcal{P}_U : \theta \in \Gamma_n\}, \mathcal{D}_H)$ . By Lemma A.7, there exist positive constants $C_9$ and $C_{10}$ depending on $M_1, M_2, L$ and $B$ such that + +$$ +\begin{array}{l} N _ {n} ^ {*} \leq N \left(\epsilon_ {n} / 3 6, \mathcal {P} _ {n} ^ {\text {S i e v e}}, \mathcal {D} _ {H}\right) \\ \leq \exp (3 6 C _ {9} \epsilon_ {n} ^ {- 1}) (3 6 C _ {1 0} \epsilon_ {n} ^ {- 1}) ^ {d} \\ \leq \exp \left(3 6 C _ {9} n \epsilon_ {n} ^ {2} + d \log \left(3 6 C _ {1 0} n \epsilon_ {n} ^ {2}\right)\right) \\ \leq \exp \left(C _ {1 1} n \epsilon_ {n} ^ {2}\right), \tag {58} \\ \end{array} +$$ + +where the third inequality holds because $\epsilon_n^{-1} \leq n\epsilon_n^2$ , and the last inequality holds by $C_{11} = 36C_9 + |\log(36C_{10})| + 2$ because $d\log(n\epsilon_n^2) \leq 2n\epsilon_n^2$ . In addition, by Lemma 2 of [15] and Lemma 9 of [15], applied with $\epsilon = C_H M\epsilon_n$ , where $C_H M \geq 2$ , there exist tests $\phi_n$ that satisfy + +$$ +\mathbb {E} _ {\theta^ {*}} ^ {n} \phi_ {n} \leq N _ {n} ^ {*} \exp \left(- \frac {1}{2} C _ {H} ^ {2} M ^ {2} n \epsilon_ {n} ^ {2}\right) \frac {1}{1 - \exp \left(- \frac {1}{2} C _ {H} ^ {2} M ^ {2} n \epsilon_ {n} ^ {2}\right)}, \tag {59} +$$ + +$$ +\sup _ {\theta \in \Gamma_ {n}} \mathbb {E} _ {\theta} ^ {n} (1 - \phi_ {n}) \leq \exp \left(- \frac {1}{2} C _ {H} ^ {2} M ^ {2} J ^ {2} n \epsilon_ {n} ^ {2}\right), +$$ + +for any $J \geq 1$ . Then, by (56), the first term of (57) is upper bounded by + +$$ +\mathbb {E} _ {\theta^ {*}} ^ {n} \left[ \Pi \left(\Gamma_ {n} \mid \mathbf {D} _ {n}\right) \mathbb {1} \left\{\Omega_ {n} \right\} \right] \leq \mathbb {E} _ {\theta^ {*}} ^ {n} \phi_ {n} + \exp \left(\left(C _ {8} + 2\right) n \epsilon_ {n} ^ {2}\right) \sup _ {\theta \in \Gamma_ {n}} \mathbb {E} _ {\theta} ^ {n} \left(1 - \phi_ {n}\right). \tag {60} +$$ + +If $M$ is sufficiently large to ensure that $C_H^2 M^2 / 2 - C_{11} > C_H^2 M^2 / 4$ , by combining (58) and the first line of (59), we have + +$$ +\begin{array}{l} \mathbb {E} _ {\theta^ {*}} ^ {n} \phi_ {n} \leq \exp \left(\left(C _ {1 1} - \frac {1}{2} C _ {H} ^ {2} M ^ {2}\right) n \epsilon_ {n} ^ {2}\right) \frac {1}{1 - \exp \left(- \frac {1}{2} C _ {H} ^ {2} M ^ {2} n \epsilon_ {n} ^ {2}\right)} \\ \leq C _ {1 2} \exp \left(- \frac {1}{4} C _ {H} ^ {2} M ^ {2} n \epsilon_ {n} ^ {2}\right), \\ \end{array} +$$ + +where $C_{12} = (1 - \exp (-2C_{11}))^{-1}$ is a positive constant. If we set $J = 1$ and choose $M$ to be sufficiently large such that $C_H^2 M^2 /2 - (2 + C_8) > C_H^2 M^2 /4$ , by the second line of (59), the second term in the right hand side of (60) is bounded by + +$$ +\exp \left(- \frac {1}{4} C _ {H} ^ {2} M ^ {2} n \epsilon_ {n} ^ {2}\right). +$$ + +Therefore, if we choose $M$ to be sufficiently large such that $M \geq 2\sqrt{(2 + C_8)\vee C_{11}} /C_H$ , by combining the preceding two displays, (57) and (60), we have + +$$ +\mathbb {E} _ {\theta^ {*}} ^ {n} \left[ \Pi \left(\mathcal {D} _ {\mathbb {Q}} (\theta , \theta^ {*}) \geq M \epsilon_ {n} \mid \mathbf {D} _ {n}\right) \mathbb {1} \left\{\Omega_ {n} \right\} \right] \leq \left(C _ {1 2} + 1\right) \exp \left(- C _ {1 3} n \epsilon_ {n} ^ {2}\right) + \mathbb {E} _ {\theta^ {*}} ^ {n} \left[ \Pi \left(U ^ {c} \mid \mathbf {D} _ {n}\right) \right], +$$ + +where $C_{13} = (C_8 + 2) \vee C_{11}$ is a positive constant. By Lemma A.1, if $\gamma < 2/3$ , the second term on the right-hand side of the last display is bounded by $c_2 \exp(-c_3 n)$ for $n \geq c_1$ , and if $\gamma \geq 2/3$ , it is bounded by $c_5 \exp(-c_6 n^{1/3})$ for $n \geq c_4$ , where $c_1, \ldots, c_6$ are positive constants depending on $M_1, M_2, L, B, p_{\min}, p_{\max}, \kappa, \underline{\alpha}, \overline{\alpha}$ and $\rho$ . Then, we have + +$$ +\mathbb{E}_{\theta^{*}}^{n}[\Pi (\mathcal{D}_{\mathbb{Q}}(\theta ,\theta^{*})\geq M\epsilon_{n}\mid \mathbf{D}_{n})\mathbb{1}\{\Omega_{n}\} ]\leq \left\{ \begin{array}{ll}C_{14}\exp \left(-C_{15}n\epsilon_{n}^{2}\right), & \quad \text{if $\gamma < \frac{2}{3}$},\\ C_{16}\exp \left(-C_{17}n^{\frac{1}{3}}\right), & \quad \text{if $\gamma \geq \frac{2}{3}$}, \end{array} \right. +$$ + +where $C_{14} = C_{12} + 1 + c_2$ , $C_{15} = C_{13} \wedge c_3$ , $C_{16} = C_{12} + 1 + c_5$ and $C_{17} = C_{13} \wedge c_6$ . By the Markov inequality and (56), we have that if $\gamma < 2/3$ , + +$$ +\Pi \left(\mathcal {D} _ {\mathbb {Q}} \left(\theta , \theta^ {*}\right) \geq M \epsilon_ {n} \mid \mathbf {D} _ {n}\right) \leq C _ {1 4} \exp \left(- C _ {1 8} n \epsilon_ {n} ^ {2}\right), +$$ + +with probability at least $1 - (\exp (-C_{18}n\epsilon_n^2) + 1 / n\epsilon_n^2)$ , where $C_{18} = C_{15} / 2$ , and if $\gamma \geq 2 / 3$ + +$$ +\Pi (\mathcal {D} _ {\mathbb {Q}} (\theta , \theta^ {*}) \geq M \epsilon_ {n} \mid \mathbf {D} _ {n}) \leq C _ {1 6} \exp \left(- C _ {1 9} n ^ {\frac {1}{3}}\right), +$$ + +with probability at least $1 - \left( \exp \left( -C_{19} n^{1/3} \right) + 1 / n \epsilon_n^2 \right)$ , where $C_{19} = C_{17} / 2$ . If we fix $M = \lceil 2\sqrt{(C_8 + 2) \vee C_{11}} / C_H \rceil$ , then the proof is complete. + +![](images/47f3bd420f679920b74342953a7691dd13a4c28339d6dea4bce7d5c826722de8.jpg) + +# B Proofs for Section 5 + +# B.1 Proof of Lemma 5.1 + +Proof. We first consider the epoch $l - 1$ under the condition $\gamma_{l-1} < 1/3$ . Let $q_l(\cdot \mid x)$ be the conditional probability mass function of $P_t$ given $X_t = x$ for $t \in \mathcal{E}_l$ in epoch $l$ . By Lemma C.5, $q_l(\cdot \mid x)$ satisfies the assumption (A4) for every epoch $l$ . Then, by Theorem 3.1, there exist positive constants $c_1, c_2, c_3$ and $c_4$ depending on $L, B, p_{\min}, p_{\max}, \kappa, \alpha, \rho$ such that for $l \geq \lceil \log_2(c_4 / n_1) \rceil + 1$ , + +$$ +\Pi \left(\mathcal {D} _ {\mathbb {Q} _ {l - 1}} \left(\theta , \theta^ {*}\right) \geq c _ {1} \epsilon_ {l - 1} \mid \mathbf {D} _ {l - 1}\right) \leq c _ {2} \exp \left(- c _ {3} n _ {l - 1} \epsilon_ {l - 1} ^ {2}\right) \tag {61} +$$ + +with probability at least $1 - \exp (-c_{3}n_{l - 1}\epsilon_{l - 1}^{2}) - 1 / (n_{l - 1}\epsilon_{l - 1}^{2})$ . We partition the parameter space $\widetilde{\Theta}$ into two subsets $\widetilde{\Theta}_{l - 1,1} = \{\theta \in \widetilde{\Theta}:\mathcal{D}_{\mathbb{Q}_{l - 1}}(\theta ,\theta^{*}) < c_{1}\epsilon_{l - 1}\}$ and $\widetilde{\Theta}_{l - 1,2} = \{\theta \in \widetilde{\Theta}:\mathcal{D}_{\mathbb{Q}_{l - 1}}(\theta ,\theta^{*})\geq c_{1}\epsilon_{l - 1}\}$ . Then, we can decompose $\widehat{\theta}^{l - 1}$ as + +$$ +\begin{array}{l} \widehat {\theta} ^ {l - 1} = \int_ {\widetilde {\Theta}} \theta d \widetilde {\Pi} (\theta \mid \mathbf {D} _ {l - 1}) \\ = \int_ {\widetilde {\Theta} _ {l - 1, 1}} \theta d \widetilde {\Pi} (\theta \mid \mathbf {D} _ {l - 1}) + \int_ {\widetilde {\Theta} _ {l - 1, 2}} \theta d \widetilde {\Pi} (\theta \mid \mathbf {D} _ {l - 1}) \\ = \left(1 - \tau_ {l - 1}\right) \widehat {\theta} _ {1} ^ {l - 1} + \tau_ {l - 1} \widehat {\theta} _ {2} ^ {l - 1}, \tag {62} \\ \end{array} +$$ + +where $\tau_{l-1} = \widetilde{\Pi}(\widetilde{\Theta}_{l-1,2} \mid \mathbf{D}_{l-1})$ . Here, $\widehat{\theta}_1^{l-1}$ and $\widehat{\theta}_2^{l-1}$ are the mean estimates of the probability measures resulting from the restriction and normalization of the truncated posterior distribution on the sets $\widetilde{\Theta}_{l-1,1}$ and $\widetilde{\Theta}_{l-1,2}$ , respectively. It is easy to check that the function $\theta \mapsto \mathcal{D}_{\mathbb{Q}_{l-1}}(\theta, \theta^*)$ is convex and bounded over the domain $\widetilde{\Theta}$ . By Jensen's inequality, we have + +$$ +\begin{array}{l} \mathcal {D} _ {\mathbb {Q} _ {l - 1}} \left(\widehat {\theta} _ {1} ^ {l - 1}, \theta^ {*}\right) \leq \int_ {\widetilde {\Theta} _ {l - 1, 1}} \mathcal {D} _ {\mathbb {Q} _ {l - 1}} \left(\theta , \theta^ {*}\right) d \widetilde {\Pi} _ {1} \left(\theta \mid \mathbf {D} _ {l - 1}\right) \\ < c _ {1} \epsilon_ {l - 1}, \tag {63} \\ \end{array} +$$ + +where $\widetilde{\Pi}_1(\cdot \mid \mathbf{D}_{l - 1})$ be the probability measure obtained by restricting and renormalizing $\widetilde{\Pi} (\cdot \mid \mathbf{D}_{l - 1})$ to $\widetilde{\Theta}_{l - 1,1}$ , and the last inequality holds by the definition of $\widetilde{\Theta}_{l - 1,1}$ . On the event that the inequality (61) holds, we have that with probability at least $1 - \exp (-c_3n_{l - 1}\epsilon_{l - 1}^2) - 1 / (n_{l - 1}\epsilon_{l - 1}^2)$ , for $l\geq \left[\log_2(c_4 / n_1)\right] + 1$ , it follows that + +$$ +\begin{array}{l} \mathcal {D} _ {\mathbb {Q} _ {l - 1}} \left(\widehat {\theta} ^ {l - 1}, \theta^ {*}\right) \leq \left(1 - \tau_ {l - 1}\right) \mathcal {D} _ {\mathbb {Q} _ {l - 1}} \left(\widehat {\theta} _ {1} ^ {l - 1}, \theta^ {*}\right) + \tau_ {l - 1} \mathcal {D} _ {\mathbb {Q} _ {l - 1}} \left(\widehat {\theta} _ {2} ^ {l - 1}, \theta^ {*}\right) \\ < c _ {1} \epsilon_ {l - 1} + \frac {\Pi (\widetilde {\Theta} _ {l - 1 , 2} \mid \mathbf {D} _ {l - 1})}{\Pi (\widetilde {\Theta} \mid \mathbf {D} _ {l - 1})} \mathcal {D} _ {\mathbb {Q} _ {l - 1}} \left(\widehat {\theta} _ {2} ^ {l - 1}, \theta^ {*}\right) \\ \leq c _ {1} \epsilon_ {l - 1} + \frac {c _ {2} \exp (- c _ {3} n _ {l - 1} \epsilon_ {l - 1} ^ {2})}{1 - c _ {2} \exp (- c _ {3} n _ {l - 1} \epsilon_ {l - 1} ^ {2})} \cdot (1 + \sqrt {d} (a \lor b) + B) \\ \leq C _ {1} \epsilon_ {l - 1}, \\ \end{array} +$$ + +where the first inequality holds because of the convexity of the function $\theta \mapsto \mathcal{D}_{\mathbb{Q}_{l - 1}}(\theta ,\theta^{*})$ and (62), and the second inequality holds by (63) and the definition of $\tau_{l - 1}$ . The third inequality follows from $\Pi (\widetilde{\Theta}\mid \mathbf{D}_{l - 1})\geq 1 - \Pi (\widetilde{\Theta}_{l - 1,2}\mid \mathbf{D}_{l - 1})$ , combined with inequality (61) and the boundedness of $\mathcal{D}_{\mathbb{Q}_{l - 1}}$ over $\widetilde{\Theta}$ under the assumption (A1). The last inequality holds with a positive constant $C_1$ depending on $c_{1},c_{2},c_{3},a,b$ and $B$ , since $\sqrt{d}\exp (-c_3n_{l - 1}\epsilon_{l - 1}^2) / (1 - c_2\exp (-c_3n_{l - 1}\epsilon_{l - 1}^2))\lesssim \epsilon_{l - 1}$ . + +By similar arguments as before, for the epoch $l - 1$ under the condition $\gamma_{l-1} \geq 1/3$ , by Theorem 3.2, there exist positive constants $c_5, c_6, c_7$ and $c_8$ depending on $L, B, p_{\min}, p_{\max}, \kappa, \alpha$ and $\rho$ such that for $l \geq \left\lceil \log_2(c_8 / n_1) \right\rceil + 1$ , + +$$ +\Pi \left(\mathcal {D} _ {\mathbb {Q} _ {l - 1}} \left(\theta , \theta^ {*}\right) \geq c _ {5} \epsilon_ {l - 1} \mid \mathbf {D} _ {l - 1}\right) \leq c _ {6} \xi_ {k - 1}, +$$ + +where + +$$ +\xi_ {k - 1} = \left\{ \begin{array}{l l} \exp (- c _ {7} n _ {l - 1} \epsilon_ {l - 1} ^ {2}) & \quad \text {i f \frac {1}{3} \leq \gamma_ {l - 1} < \frac {2}{3}}, \\ \exp (- c _ {7} n _ {l - 1} ^ {1 / 3}) & \quad \text {i f \gamma_ {l - 1} \geq \frac {2}{3}}, \end{array} \right. +$$ + +with probability at least $1 - \xi_{k-1} - 1 / (n_{l-1}\epsilon_{l-1}^2)$ . Since $\exp(-c_7n_{l-1}\epsilon_{l-1}^2) \leq \exp(-c_7n_{l-1}^{1/3})$ , we unify the cases where $\gamma_{l-1}$ is either greater than or less than $2/3$ and obtain the bound + +$$ +\Pi \left(\mathcal {D} _ {\mathbb {Q} _ {l - 1}} \left(\theta , \theta^ {*}\right) \geq c _ {5} \epsilon_ {l - 1} \mid \mathbf {D} _ {l - 1}\right) \leq c _ {6} \exp \left(- c _ {7} n _ {l - 1} ^ {1 / 3}\right), \tag {64} +$$ + +with probability at least $1 - \exp (-c_7n_{l - 1}^{1 / 3}) - 1 / (n_{l - 1}\epsilon_{l - 1}^2)$ . Similarly, for $l\geq \lceil \log_2(c_8 / n_1)\rceil +1$ we have + +$$ +\mathcal {D} _ {\mathbb {Q} _ {l - 1}} \left(\widehat {\theta} ^ {l - 1}, \theta^ {*}\right) \leq C _ {2} \epsilon_ {l - 1}, +$$ + +with probability at least $1 - \exp (-c_7n_{l - 1}^{1 / 3}) - 1 / (n_{l - 1}\epsilon_{l - 1}^2)$ , where $C_2$ is a positive constant depending on $c_{5},c_{6},c_{7},a,b$ and $B$ . The proof is then complete. + +# B.2 Proof of Theorem 5.2 + +Lemma B.1. Suppose that assumptions (A1)-(A3), (A5) and (B1)-(B2) hold. Suppose that the prior distribution $\Pi$ is specified as in (4), and the policy $\pi_l$ for each epoch $l$ is defined by (5). Then, there exist positive constants $C_1$ , $C_2$ , $C_3$ and $C_4$ depending on $L$ , $B$ , $p_{\min}$ , $p_{\max}$ , $\kappa$ , $\alpha$ , $\rho$ , $a$ , $b$ , $\eta_1$ , $\eta_2$ and $n_1$ such that for $l \geq C_1$ , + +$$ +\sum_ {t \in \mathcal {E} _ {l}} r (t) \leq C _ {2} n _ {l} \mathcal {D} _ {\mathbb {Q} _ {l - 1}} (\widehat {\theta} ^ {l - 1}, \theta^ {*}) + C _ {3} \left(n _ {l} ^ {\frac {1 + \gamma_ {l}}{2}} \wedge n _ {l} ^ {\frac {2}{3}}\right) (\log n _ {l}) ^ {\frac {1}{2}} +$$ + +with probability at least $1 - (\exp (-C_4n_l^{1 / 3}) + 3 / n_l^2)$ + +Proof. The regret in epoch $l$ is decomposed and upper bounded by + +$$ +\begin{array}{l} \sum_ {t \in \mathcal {E} _ {l}} r (t) = \sum_ {t \in \mathcal {E} _ {l}} \left(P _ {t} ^ {*} H _ {\theta^ {*}} \left(X _ {t}, P _ {t} ^ {*}\right) - P _ {t} H _ {\theta^ {*}} \left(X _ {t}, P _ {t}\right)\right) \\ = \sum_ {t \in \mathcal {E} _ {l}} \left\{\left(P _ {t} ^ {*} H _ {\theta^ {*}} \left(X _ {t}, P _ {t} ^ {*}\right) - P _ {t} ^ {*} H _ {\widehat {\theta} ^ {l - 1}} \left(X _ {t}, P _ {t} ^ {*}\right)\right) + \left(P _ {t} ^ {*} H _ {\widehat {\theta} ^ {l - 1}} \left(X _ {t}, P _ {t} ^ {*}\right) - P _ {t} H _ {\widehat {\theta} ^ {l - 1}} \left(X _ {t}, P _ {t}\right)\right) \right. \\ \left. + \left(P _ {t} H _ {\hat {\theta} ^ {l - 1}} \left(X _ {t}, P _ {t}\right) - P _ {t} H _ {\theta^ {*}} \left(X _ {t}, P _ {t}\right)\right) \right\} \\ \leq \underbrace{\sum_{t\in\mathcal{E}_{l}}\left(P_{t}^{*}H_{\widehat{\theta}^{l - 1}}(X_{t},P_{t}^{*}) - P_{t}H_{\widehat{\theta}^{l - 1}}(X_{t},P_{t})\right)}_{\text{(i)}} + p_{\max}\underbrace{\sum_{t\in\mathcal{E}_{l}}\left|H_{\widehat{\theta}^{l - 1}}(X_{t},P_{t}^{*}) - H_{\theta^{*}}(X_{t},P_{t}^{*})\right|}_{\text{(ii)}} \\ + p _ {\max } \underbrace {\sum_ {t \in \mathcal {E} _ {l}} \left| H _ {\widehat {\theta} ^ {l - 1}} \left(X _ {t} , P _ {t}\right) - H _ {\theta^ {*}} \left(X _ {t} , P _ {t}\right) \right|} _ {\text {(i i i)}} \tag {65} \\ \end{array} +$$ + +where the last inequality holds because any $P_{t}$ and $P_{t}^{*}$ lie in $\mathcal{G} \subset [p_{\min}, p_{\max}]$ almost surely. Note that $\{(X_{t}, P_{t}, P_{t}^{*})\}_{t \in \mathcal{E}_{t}}$ is an i.i.d. sample of joint distribution which satisfies $P_{t} \sim \mathbb{Q}_{l}, P_{t}^{*} \sim \mathbb{Q}^{*}$ and $X_{t} \sim \mathbb{P}_{X}$ . Since $P_{t}^{*}H_{\widehat{\theta}^{l-1}}(X_{t}, P_{t}^{*}) - P_{t}H_{\widehat{\theta}^{l-1}}(X_{t}, P_{t}) \in [-p_{\max}, p_{\max}]$ , by Hoeffding's inequality, it holds that + +$$ +\left. \left(\mathrm {i}\right) < 2 p _ {\max } n _ {l} ^ {\frac {1}{2}} \left(\log n _ {l}\right) ^ {\frac {1}{2}} + \mathbb {E} \left[ \sum_ {t \in \mathcal {E} _ {l}} \left(P _ {t} ^ {*} H _ {\widehat {\theta} ^ {l - 1}} \left(X _ {t}, P _ {t} ^ {*}\right) - P _ {t} H _ {\widehat {\theta} ^ {l - 1}} \left(X _ {t}, P _ {t}\right)\right) \right], \right. \tag {66} +$$ + +with probability at least $1 - 1 / n_l^2$ . Let $\widehat{P}_t \in \mathrm{argmax}_{p \in \mathcal{G}} p H_{\widehat{\theta}^{l-1}}(X_t, p)$ , and let $U$ denote a uniform random variable over $\mathcal{G}$ . By the design of Algorithm 1, we have $P_t = R \widehat{P}_t + (1 - R)U$ , where $R$ is a Bernoulli random variable with success probability $1 - \eta_l$ . By the law of total expectation, we have + +$$ +\mathbb {E} \left[ P _ {t} H _ {\widehat {\theta} ^ {l - 1}} \left(X _ {t}, P _ {t}\right) \right] = (1 - \eta_ {l}) \mathbb {E} \left[ \widehat {P} _ {t} H _ {\widehat {\theta} ^ {l - 1}} \left(X _ {t}, \widehat {P} _ {t}\right) \right] + \eta_ {l} \mathbb {E} \left[ U H _ {\widehat {\theta} ^ {l - 1}} \left(X _ {t}, U\right) \right]. +$$ + +By substituting this in (66), the second term of (66) is bounded by + +$$ +\begin{array}{l} \mathbb {E} \left[ \sum_ {t \in \mathcal {E} _ {l}} \left(P _ {t} ^ {*} H _ {\widehat {\theta} ^ {l - 1}} \left(X _ {t}, P _ {t} ^ {*}\right) - P _ {t} H _ {\widehat {\theta} ^ {l - 1}} \left(X _ {t}, P _ {t}\right)\right) \right] = (1 - \eta_ {l}) \sum_ {t \in \mathcal {E} _ {l}} \mathbb {E} \left[ P _ {t} ^ {*} H _ {\widehat {\theta} ^ {l - 1}} \left(X _ {t}, P _ {t} ^ {*}\right) - \widehat {P} _ {t} H _ {\widehat {\theta} ^ {l - 1}} \left(X _ {t}, \widehat {P} _ {t}\right) \right] \\ + \eta_ {l} \sum_ {t \in \mathcal {E} _ {l}} \mathbb {E} \left[ P _ {t} ^ {*} H _ {\widehat {\theta} ^ {l - 1}} \left(X _ {t}, P _ {t} ^ {*}\right) - U H _ {\widehat {\theta} ^ {l - 1}} \left(X _ {t}, U\right) \right] \\ \leq \eta_ {l} \sum_ {t \in \mathcal {E} _ {l}} p _ {\max } \\ \leq C _ {0} \left(n _ {l} ^ {\frac {1 + \gamma_ {l}}{2}} \wedge n _ {l} ^ {\frac {2}{3}}\right) (\log n _ {l}) ^ {\frac {1}{2}} \\ \end{array} +$$ + +where the first inequality holds because $P_{t}^{*}H_{\widehat{\theta}^{l - 1}}(X_{t},P_{t}^{*}) - \widehat{P}_{t}H_{\widehat{\theta}^{l - 1}}(X_{t},\widehat{P}_{t})\leq 0$ by the definition of $\widehat{P}_t$ , and the last inequality holds by the definition of $\eta_l$ (6) and (7) with a positive constant $C_0$ depending on $p_{\mathrm{min}},p_{\mathrm{max}},\kappa ,\eta_1$ and $\eta_2$ . Combining this with (66), we have + +$$ +\left(\mathrm {i}\right) < C _ {1} \left(n _ {l} ^ {\frac {1 + \gamma_ {l}}{2}} \wedge n _ {l} ^ {\frac {2}{3}}\right) \left(\log n _ {l}\right) ^ {\frac {1}{2}}, \tag {67} +$$ + +with probability at least $1 - 1 / n_{l}^{2}$ , where $C_1 = 2p_{\mathrm{max}} + C_0$ is a positive constant. + +For (ii) and (iii), by Lemma C.6, for every $\epsilon > 0$ , there exist positive constants $c_{1}$ and $c_{2}$ depending on $L, B, p_{\min}, p_{\max}, \kappa, \alpha, \rho, a, b, n_{1}$ and $\epsilon$ such that for large $l \geq c_{1}$ , we have + +$$ +\| \widehat {\mathbf {S}} _ {0} ^ {l - 1} - \mathbf {S} _ {0} ^ {*} \| _ {\infty} + \| \widehat {\boldsymbol {\beta}} ^ {l - 1} - \boldsymbol {\beta} ^ {*} \| _ {2} < \epsilon , +$$ + +with probability at least $1 - \exp(-c_2 n_{l-1}^{1/3})$ . Note that there exist constants $M_1$ and $M_2$ such that $0 < M_1 \leq S_0^*(p_{\max}) < S_0^*(p_{\min}) \leq M_2 < 1$ by assumption (A5). Take $\epsilon = C_2$ , where $C_2 = ((M_1 \wedge (1 - M_2)) / 2 \wedge B)$ . On the event that the preceding inequality holds, we have + +$$ +\| \widehat {\mathbf {S}} _ {0} ^ {l - 1} - \mathbf {S} _ {0} ^ {*} \| _ {\infty} + \| \widehat {\boldsymbol {\beta}} ^ {l - 1} - \boldsymbol {\beta} ^ {*} \| _ {2} < C _ {2}. +$$ + +This implies that $\widehat{S}_{0,1}^{l - 1} > M_1 / 2 > 0$ , $\widehat{S}_{0,K}^{l - 1} < (1 + M_2) / 2 < 1$ and $\| \widehat{\beta}^{l - 1}\| _2 < 2B$ . Then, by Lemma C.2, for any $p\in \mathcal{G}$ and $l\geq c_{1}$ , with probability at least $1 - \exp (-c_2n_{l - 1}^{1 / 3})$ , we have + +$$ +\left| H _ {\widehat {\theta} ^ {l - 1}} \left(X _ {t}, p\right) - H _ {\theta^ {*}} \left(X _ {t}, p\right) \right| \leq C _ {3} \left| \widehat {S} _ {0} ^ {l - 1} (p) - S _ {0} ^ {*} (p) \right| + C _ {4} \| \widehat {\beta} ^ {l - 1} - \beta^ {*} \| _ {2}, \tag {68} +$$ + +where $C_3$ and $C_4$ are positive constants depending on $M_1, M_2, L$ and $B$ . + +Let $\Omega_1$ be the event that (68) holds. For (ii), under the event $\Omega_1$ , we have + +$$ +\begin{array}{l} \left(\mathrm {i i}\right) < C _ {3} \sum_ {t \in \mathcal {E} _ {l}} \left| \widehat {S} _ {0} ^ {l - 1} \left(P _ {t} ^ {*}\right) - S _ {0} ^ {*} \left(P _ {t} ^ {*}\right) \right| + C _ {4} \sum_ {t \in \mathcal {E} _ {l}} \| \widehat {\beta} ^ {l - 1} - \beta^ {*} \| _ {2} \\ = C _ {3} \sum_ {t \in \mathcal {E} _ {l}} | \widehat {S} _ {0} ^ {l - 1} (P _ {t} ^ {*}) - S _ {0} ^ {*} (P _ {t} ^ {*}) | + C _ {4} n _ {l} \| \widehat {\beta} ^ {l - 1} - \beta^ {*} \| _ {2}. \\ \end{array} +$$ + +Since $|\widehat{S}_0^{l-1}(P_t^*) - S_0^*(P_t^*)| \leq 1$ , by Hoeffding's inequality, there exists an event $\Omega_2$ such that $\mathbb{P}(\Omega_2) \geq 1 - 1/n_l^2$ , and in $\Omega_2$ , + +$$ +\sum_ {t \in \mathcal {E} _ {l}} | \widehat {S} _ {0} ^ {l - 1} (P _ {t} ^ {*}) - S _ {0} ^ {*} (P _ {t} ^ {*}) | \leq n _ {l} \mathbb {E} \left[ | \widehat {S} _ {0} ^ {l - 1} (P _ {t} ^ {*}) - S _ {0} ^ {*} (P _ {t} ^ {*}) | \right] + n _ {l} ^ {\frac {1}{2}} (\log n _ {l}) ^ {\frac {1}{2}}. +$$ + +Recall the definition of $P_{c}$ from (87). It is easy to see that if $P_{t}^{*} = p$ for some $p \in \mathcal{G}$ , then $P_{c} \in (p - \delta, p + \delta)$ . Thus, we have $\mathbb{P}(P_{c} \in (p - \delta, p + \delta)) \geq \mathbb{P}(P_{t}^{*} = p)$ . Let $P_{l}$ be a random + +variable distributed from $\mathbb{Q}_l$ in epoch $l$ . By Lemma C.7 and Portmanteau theorem, we obtain $\lim_{l\to \infty}\mathbb{P}(P_l\in (p - \delta ,p + \delta)) = \mathbb{P}(P_c\in (p - \delta ,p + \delta))$ . Combining these results, we have $\lim_{l\to \infty}q_l(p) = \lim_{l\to \infty}\mathbb{P}(P_l = p)\geq \mathbb{P}(P_t^* = p) = q^* (p)$ . Then, for sufficiently large $l$ , we have + +$$ +\begin{array}{l} \mathbb {E} \left[ | \widehat {S} _ {0} ^ {l - 1} \left(P _ {t} ^ {*}\right) - S _ {0} ^ {*} \left(P _ {t} ^ {*}\right) | \right] = \sum_ {p \in \mathcal {G}} | \widehat {S} _ {0} ^ {l - 1} (p) - S _ {0} ^ {*} (p) | q ^ {*} (p) \\ = \sum_ {p \in \mathcal {G}} | \widehat {S} _ {0} ^ {l - 1} (p) - S _ {0} ^ {*} (p) | \frac {q ^ {*} (p)}{q _ {k - 1} (p)} q _ {k - 1} (p) \\ \leq C _ {5} \sum_ {p \in \mathcal {G}} | \widehat {S} _ {0} ^ {l - 1} (p) - S _ {0} ^ {*} (p) | q _ {k - 1} (p) \\ \leq C _ {5} \left(\sum_ {p \in \mathcal {G}} | \widehat {S} _ {0} ^ {l - 1} (p) - S _ {0} ^ {*} (p) | ^ {2} q _ {k - 1} (p)\right) ^ {\frac {1}{2}} \\ = C _ {5} \| \widehat {\mathbf {S}} _ {0} ^ {l - 1} - \mathbf {S} _ {0} ^ {*} \| _ {2, \mathbb {Q} _ {l - 1}}, \\ \end{array} +$$ + +where the last inequality holds by Jensen's inequality, and $C_5$ be a positive constant. Combining the last three displays, under the event $\Omega_1 \cap \Omega_2$ , we have + +$$ +\begin{array}{l} \left(\mathrm {i i}\right) < C _ {6} n _ {l} \left(\| \widehat {\mathbf {S}} _ {0} ^ {l - 1} - \mathbf {S} _ {0} ^ {*} \| _ {2, \mathbb {Q} _ {l - 1}} + \| \widehat {\beta} ^ {l - 1} - \beta^ {*} \| _ {2}\right) + C _ {3} n _ {l} ^ {\frac {1}{2}} \left(\log n _ {l}\right) ^ {\frac {1}{2}} \\ = C _ {6} n _ {l} \mathcal {D} _ {\mathbb {Q} _ {l - 1}} \left(\left(\widehat {\mathbf {S}} _ {0} ^ {l - 1}, \widehat {\beta} ^ {l - 1}\right), \left(\mathbf {S} _ {0} ^ {*}, \beta^ {*}\right)\right) + C _ {3} n _ {l} ^ {\frac {1}{2}} \left(\log n _ {l}\right) ^ {\frac {1}{2}} \tag {69} \\ \end{array} +$$ + +where $C_6 = C_3C_5\lor C_4$ be a positive constant. + +Similarly, for (iii), under the event $\Omega_1$ , we have + +$$ +\left(\text {i i i}\right) < C _ {3} \sum_ {t \in \mathcal {E} _ {l}} \left| \widehat {S} _ {0} ^ {l - 1} \left(P _ {t}\right) - S _ {0} ^ {*} \left(P _ {t}\right) \right| + C _ {4} n _ {l} \| \widehat {\beta} ^ {l - 1} - \beta^ {*} \| _ {2}. +$$ + +By Hoeffding's inequality, there exists an event $\Omega_3$ such that $\mathbb{P}(\Omega_3) \geq 1 - 1/n_l^2$ , and in $\Omega_3$ , + +$$ +\sum_ {t \in \mathcal {E} _ {l}} | \widehat {S} _ {0} ^ {l - 1} (P _ {t}) - S _ {0} ^ {*} (P _ {t}) | \leq n _ {l} \mathbb {E} \left[ | \widehat {S} _ {0} ^ {l - 1} (P _ {t}) - S _ {0} ^ {*} (P _ {t}) | \right] + n _ {l} ^ {\frac {1}{2}} (\log n _ {l}) ^ {\frac {1}{2}}. +$$ + +Note that if $P_{c} \in (p - \delta, p + \delta)$ for some $p \in \mathcal{G}$ , then $P_{t}^{*} \in \{p - \delta, p, p + \delta\}$ . By Lemma C.7 and Portmanteau theorem, we have $\lim_{l \to \infty} q_{l}(p) = \lim_{l \to \infty} \mathbb{P}(P_{l} \in (p - \delta, p + \delta)) = \mathbb{P}(P_{c} \in (p - \delta, p + \delta)) \leq \mathbb{P}(P_{t}^{*} = p - \delta) + \mathbb{P}(P_{t}^{*} = p) + \mathbb{P}(P_{t}^{*} = p + \delta) \lesssim \delta$ , where the last inequality holds by Assumption (B2). Then, for sufficiently large $l$ , we have + +$$ +\begin{array}{l} \mathbb {E} \left[ | \widehat {S} _ {0} ^ {l - 1} (P _ {t}) - S _ {0} ^ {*} (P _ {t}) | \right] = \sum_ {p \in \mathcal {G}} | \widehat {S} _ {0} ^ {l - 1} (p) - S _ {0} ^ {*} (p) | q _ {l} (p) \\ = \sum_ {p \in \mathcal {G}} | \widehat {S} _ {0} ^ {l - 1} (p) - S _ {0} ^ {*} (p) | \frac {q _ {l} (p)}{q ^ {*} (p)} \frac {q ^ {*} (p)}{q _ {k - 1} (p)} q _ {k - 1} (p) \\ \leq C _ {7} \sum_ {p \in \mathcal {G}} | \widehat {S} _ {0} ^ {l - 1} (p) - S _ {0} ^ {*} (p) | q _ {k - 1} (p) \\ \leq C _ {7} \left(\sum_ {p \in \mathcal {G}} | \widehat {S} _ {0} ^ {l - 1} (p) - S _ {0} ^ {*} (p) | ^ {2} q _ {k - 1} (p)\right) ^ {\frac {1}{2}} \\ = C _ {7} \| \hat {\mathbf {S}} _ {0} ^ {l - 1} - \mathbf {S} _ {0} ^ {*} \| _ {2, \mathbb {Q} _ {l - 1}}. \\ \end{array} +$$ + +Combining the last three displays, under the event $\Omega_1 \cap \Omega_3$ , we have + +$$ +< C _ {8} n _ {l} \mathcal {D} _ {\mathbb {Q} _ {l - 1}} \left(\left(\widehat {\mathbf {S}} _ {0} ^ {l - 1}, \widehat {\beta} ^ {l - 1}\right), \left(\mathbf {S} _ {0} ^ {*}, \boldsymbol {\beta} ^ {*}\right)\right) + C _ {3} n _ {l} ^ {\frac {1}{2}} (\log n _ {l}) ^ {\frac {1}{2}}, \tag {70} +$$ + +where $C_8 = C_3C_7\lor C_4$ be a positive constant. + +From (65), (67), (69) and (70), for sufficiently large $l$ , with probability at least $1 - (\exp(-c_2 / 2^{1/3} \cdot n_l^{1/3}) + 3 / n_l^2)$ , it holds that + +$$ +\begin{array}{l} \sum_ {t \in \mathcal {E} _ {l}} r (t) \leq C _ {1} \left(n _ {l} ^ {\frac {1 + \gamma_ {l}}{2}} \wedge n _ {l} ^ {\frac {2}{3}}\right) (\log n _ {l}) ^ {\frac {1}{2}} + p _ {\max } \left(C _ {6} n _ {l} \mathcal {D} _ {\mathbb {Q} _ {l - 1}} \left((\widehat {\mathbf {S}} _ {0} ^ {l - 1}, \widehat {\beta} ^ {l - 1}), (\mathbf {S} _ {0} ^ {*}, \beta^ {*})\right) + C _ {3} n _ {l} ^ {\frac {1}{2}} (\log n _ {l}) ^ {\frac {1}{2}}\right) \\ + p _ {\max } \left(C _ {8} n _ {l} \mathcal {D} _ {\mathbb {Q} _ {l - 1}} \left(\left(\widehat {\mathbf {S}} _ {0} ^ {l - 1}, \widehat {\beta} ^ {l - 1}\right), \left(\mathbf {S} _ {0} ^ {*}, \boldsymbol {\beta} ^ {*}\right)\right) + C _ {3} n _ {l} ^ {\frac {1}{2}} (\log n _ {l}) ^ {\frac {1}{2}}\right) \\ = C _ {9} n _ {l} \mathcal {D} _ {\mathbb {Q} _ {l - 1}} \left((\widehat {\mathbf {S}} _ {0} ^ {l - 1}, \widehat {\boldsymbol {\beta}} ^ {l - 1}), (\mathbf {S} _ {0} ^ {*}, \boldsymbol {\beta} ^ {*})\right) + C _ {1 0} \left(n _ {l} ^ {\frac {1 + \gamma_ {l}}{2}} \wedge n _ {l} ^ {\frac {2}{3}}\right) (\log n _ {l}) ^ {\frac {1}{2}}, \\ \end{array} +$$ + +where $C_9 = p_{\max}(C_6 + C_8)$ and $C_{10} = C_1 + 2p_{\max}C_3$ are positive constants. Then, the proof is complete. + +![](images/2d7e32086725bdbf056065a0b7b8228280bb97453ec2040b6fa45d25ce6a5a55.jpg) + +Lemma B.2. Suppose that assumptions (A1)-(A3), (A5) and (B1)-(B2) hold. Suppose that the prior distribution $\Pi$ is specified as in (4), and the policy $\pi_l$ for each epoch $l$ is defined by (5). Then, there exist positive constants $C_1, \ldots, C_5$ depending on $L$ , $B$ , $p_{\min}$ , $p_{\max}$ , $\kappa$ , $\alpha$ , $\rho$ , $a$ , $b$ , $\eta_1$ , $\eta_2$ and $n_1$ such that for $l \geq C_1$ , + +$$ +\sum_ {t \in \mathcal {E} _ {l}} r (t) \leq \left\{ \begin{array}{l l} C _ {2} d ^ {\frac {1}{2}} n _ {l} ^ {\frac {1}{2}} (\log (d \vee n _ {l})) ^ {\frac {1}{2}} + C _ {3} n _ {l} ^ {\frac {1 + \gamma_ {l}}{2}} (\log n _ {l}) ^ {\frac {1}{2}} & \quad i f \gamma_ {l - 1} < \frac {1}{3} \\ C _ {2} d ^ {\frac {1}{2}} n _ {l} ^ {\frac {1}{2}} (\log (d \vee n _ {l})) ^ {\frac {1}{2}} + C _ {3} n _ {l} ^ {\frac {2}{3}} (\log n _ {l}) ^ {\frac {1}{2}} & \quad i f \gamma_ {l - 1} \geq \frac {1}{3}, \end{array} \right. +$$ + +with probability at least $1 - \zeta_{l}$ , where + +$$ +\zeta_ {l} = \left\{ \begin{array}{l l} C _ {4} / n _ {l - 1} ^ {\gamma_ {l - 1}} & \quad i f \gamma_ {l - 1} < \frac {1}{3} \\ C _ {5} / n _ {l - 1} ^ {\frac {1}{3}} & \quad i f \gamma_ {l - 1} \geq \frac {1}{3}. \end{array} \right. +$$ + +Proof. By Lemma B.1, there exist positive constants $c_{1}, c_{2}, c_{3}$ and $c_{4}$ depending on $L, B, p_{\min}, p_{\max}, \kappa, \alpha, \rho, a, b, \eta_{1}, \eta_{2}$ and $n_{1}$ such that for $l \geq c_{1}$ , + +$$ +\sum_ {t \in \mathcal {E} _ {l}} r (t) \leq c _ {2} n _ {l} \mathcal {D} _ {\mathbb {Q} _ {l - 1}} \left(\widehat {\theta} ^ {l - 1}, \theta^ {*}\right) + c _ {3} \left(n _ {l} ^ {\frac {1 + \gamma_ {l}}{2}} \wedge n _ {l} ^ {\frac {2}{3}}\right) \left(\log n _ {l}\right) ^ {\frac {1}{2}} \tag {71} +$$ + +with probability at least $1 - \exp (-c_4n_l^{1 / 3}) - 3 / n_l^2$ . In addition, by Lemma 5.1, there exist positive constants $c_{5}, c_{6}, c_{7}$ and $c_{8}$ depending on $L, B, p_{\mathrm{min}}, p_{\mathrm{max}}, \kappa, \alpha, \rho, a, b$ and $n_1$ such that for $l \geq c_5$ , + +$$ +\mathcal {D} _ {\mathbb {Q} _ {l - 1}} \left(\widehat {\theta} ^ {l - 1}, \theta^ {*}\right) \leq c _ {6} \epsilon_ {l - 1} \tag {72} +$$ + +with probability at least $1 - \zeta_{l-1} - 1 / (n_{l-1}\epsilon_{l-1}^2)$ , where + +$$ +\epsilon_ {l} = \left\{ \begin{array}{l l} \sqrt {\frac {d}{n _ {l}}} \sqrt {\log (d \lor n _ {l})} + n _ {l} ^ {- \frac {1 - \gamma_ {l}}{2}} \sqrt {\log n _ {l}} & \quad \text {i f} \gamma_ {l} < \frac {1}{3}, \\ \sqrt {\frac {d}{n _ {l}}} \sqrt {\log (d \lor n _ {l})} + \left(\frac {\log n _ {l}}{n _ {l}}\right) ^ {\frac {1}{3}} & \quad \text {i f} \gamma_ {l} \geq \frac {1}{3}. \end{array} \right. +$$ + +and + +$$ +\zeta_ {l} = \left\{ \begin{array}{l l} \exp (- c _ {7} n _ {l} \epsilon_ {l} ^ {2}) & \quad \text {i f \gamma_ {l} < \frac {1}{3}}, \\ \exp (- c _ {8} n _ {l} ^ {\frac {1}{3}}) & \quad \text {i f \gamma_ {l} \geq \frac {1}{3}}. \end{array} \right. +$$ + +Consider the epoch $l - 1$ satisfying $\gamma_{l-1} < 1/3$ . On the event both (71) and (72) hold, for epoch $l \geq c_1 \lor c_5$ , we have + +$$ +\begin{array}{l} \sum_ {t \in \mathcal {E} _ {l}} r (t) \leq c _ {2} c _ {6} n _ {l} \epsilon_ {l - 1} + c _ {3} \left(n _ {l} ^ {\frac {1 + \gamma_ {l}}{2}} \wedge n _ {l} ^ {\frac {2}{3}}\right) (\log n _ {l}) ^ {\frac {1}{2}} \\ \leq 2 c _ {2} c _ {6} \left(\sqrt {n _ {l - 1} d} \sqrt {\log (d \vee n _ {l - 1})} + n _ {l - 1} ^ {\frac {1 + \gamma_ {l - 1}}{2}} \sqrt {\log n _ {l - 1}}\right) + c _ {3} \left(n _ {l} ^ {\frac {1 + \gamma_ {l}}{2}} \wedge n _ {l} ^ {\frac {2}{3}}\right) (\log n _ {l}) ^ {\frac {1}{2}} \\ \leq C _ {1} d ^ {\frac {1}{2}} n _ {l} ^ {\frac {1}{2}} \big (\log (d \vee n _ {l}) \big) ^ {\frac {1}{2}} + C _ {2} n _ {l} ^ {\frac {1 + \gamma_ {l}}{2}} (\log n _ {l}) ^ {\frac {1}{2}}, \\ \end{array} +$$ + +where the second inequality holds by substituting $n_l = 2n_{l-1}$ and $\epsilon_{l-1}$ , and the last inequality holds with positive constants $C_1 = 2c_2c_6$ and $C_2 = 2c_2c_6c_p + c_3$ because it holds that $n_{l-1}^{\gamma_{l-1}} \asymp n_l^{\gamma_l}$ due to (7), where $c_p$ is a positive constant depending on $p_{\min}, p_{\max}$ and $\kappa$ . Similarly, for epoch $l \geq c_1 \lor c_5$ satisfying $\gamma_{l-1} \geq 1/3$ , we have + +$$ +\begin{array}{l} \sum_ {t \in \mathcal {E} _ {l}} r (t) \leq c _ {2} c _ {6} n _ {l} \epsilon_ {l - 1} + c _ {3} \left(n _ {l} ^ {\frac {1 + \gamma_ {l}}{2}} \wedge n _ {l} ^ {\frac {2}{3}}\right) (\log n _ {l}) ^ {\frac {1}{2}} \\ \leq 2 c _ {2} c _ {6} \left(\sqrt {n _ {l - 1} d} \sqrt {\log (d \vee n _ {l - 1})} + n _ {l - 1} ^ {\frac {2}{3}} (\log n _ {l - 1}) ^ {\frac {1}{3}}\right) + c _ {3} \left(n _ {l} ^ {\frac {1 + \gamma_ {l}}{2}} \wedge n _ {l} ^ {\frac {2}{3}}\right) (\log n _ {l}) ^ {\frac {1}{2}} \\ \leq C _ {1} d ^ {\frac {1}{2}} n _ {l} ^ {\frac {1}{2}} (\log (d \lor n _ {l})) ^ {\frac {1}{2}} + C _ {2} n _ {l} ^ {\frac {2}{3}} (\log n _ {l}) ^ {\frac {1}{2}}. \\ \end{array} +$$ + +Let $\Omega_1$ and $\Omega_2$ denote the events where inequalities (71) and (72) hold, respectively. Then, for epoch $l - 1$ satisfying $\gamma_{l - 1} < 1 / 3$ , we obtain + +$$ +\begin{array}{l} \mathbb {P} (\Omega_ {1} ^ {c} \cup \Omega_ {2} ^ {c}) \leq \exp (- c _ {4} n _ {l} ^ {1 / 3}) + 3 / n _ {l} ^ {2} + \exp (- c _ {7} n _ {l - 1} \epsilon_ {l - 1} ^ {2}) + 1 / (n _ {l - 1} \epsilon_ {l - 1} ^ {2}) \\ \leq 1 / \left(2 ^ {1 / 3} c _ {4} n _ {l - 1} ^ {1 / 3}\right) + 3 / \left(4 n _ {l - 1} ^ {2}\right) + 1 / \left(c _ {7} n _ {l - 1} ^ {\gamma_ {l - 1}}\right) + 1 / n _ {l - 1} ^ {\gamma_ {l - 1}} \\ \leq C _ {3} / n _ {l - 1} ^ {\gamma_ {- 1}}, \\ \end{array} +$$ + +where the second inequality holds because $\exp(-x) \leq 1/x$ for any $x > 0$ and $n_l \epsilon_l^2 \geq n_l^{\gamma_l}$ , and the last inequality holds since $\gamma_{l-1} < 1/3$ , and $C_3$ is a positive constant defined as $C_3 = 1/(2^{1/3} c_4) + 1/c_7 + 7/4$ . Similarly, for epoch $l - 1$ satisfying $\gamma_{l-1} \geq 1/3$ , we have + +$$ +\begin{array}{l} \mathbb {P} \left(\Omega_ {1} ^ {c} \cup \Omega_ {2} ^ {c}\right) \leq \exp \left(- c _ {4} n _ {l} ^ {1 / 3}\right) + 3 / n _ {l} ^ {2} + \exp \left(- c _ {8} n _ {l - 1} ^ {1 / 3}\right) + 1 / \left(n _ {l - 1} \epsilon_ {l - 1} ^ {2}\right) \\ \leq 1 / (2 ^ {1 / 3} c _ {4} n _ {l - 1} ^ {1 / 3}) + 3 / (4 n _ {l - 1} ^ {2}) + 1 / (c _ {8} n _ {l - 1} ^ {1 / 3}) + 1 / n _ {l - 1} ^ {1 / 3} \\ \leq C _ {4} / n _ {l - 1} ^ {1 / 3}, \\ \end{array} +$$ + +where the second inequality holds because $n_l \epsilon_l^2 \geq n_l^{1/3}$ , and the last inequality holds by a positive constant $C_4 = 1 / (2^{1/3} c_4) + 1 / c_8 + 7/4$ . Then, the proof is complete. + +Proof of Theorem 5.2. Before proceeding, we may without loss of generality assume that the last epoch is complete (i.e., $T = n_{1}(2^{N} - 1)$ for some integer $N \geq 1$ ). If not (i.e., $n_{1}(2^{N - 1} - 1) < T < n_{1}(2^{N} - 1)$ ), the regret associated with the incomplete last epoch will be no greater than if it were completed. Thus, the number of epochs $N$ and $T$ satisfies $T = n_{1}(2^{N} - 1)$ , equivalently $N = \log_2(T / n_1 + 1)$ . + +We first consider the case where $\gamma < 1/3$ . We define $N_{\gamma_0} \coloneqq \left\lfloor \log_2(T^{\gamma_0} / n_1) \right\rfloor + 2$ , where $\gamma_0 \in (0,1)$ is a constant to be chosen later. Note that $N_{\gamma_0} < N$ for a sufficiently large $T \geq 2^{2/(1-\gamma_0)}$ . For epoch $l \leq N_{\gamma_0}$ , we have + +$$ +l \leq \left\lfloor \log_ {2} \left(T ^ {\gamma_ {0}} / n _ {1}\right) \right\rfloor + 2 \leq \log_ {2} \left(T ^ {\gamma_ {0}} / n _ {1}\right) + 2, +$$ + +and hence $n_{l - 1} = n_12^{l - 2}\leq T^{\gamma_0}$ . Therefore, by the equation (7), we have + +$$ +K _ {k - 1} = \left\lfloor \frac {p _ {\operatorname* {m a x}} - p _ {\operatorname* {m i n}}}{\kappa} T ^ {\gamma} \right\rfloor \geq \left\lfloor \frac {p _ {\operatorname* {m a x}} - p _ {\operatorname* {m i n}}}{\kappa} n _ {l - 1} ^ {\gamma / \gamma_ {0}} \right\rfloor , +$$ + +for all $l \leq N_{\gamma_0}$ . If we set $\gamma_0 = 3\gamma$ , then we have + +$$ +\left\lfloor \frac {p _ {\operatorname* {m a x}} - p _ {\operatorname* {m i n}}}{\kappa} n _ {l - 1} ^ {\gamma_ {l - 1}} \right\rfloor \geq \left\lfloor \frac {p _ {\operatorname* {m a x}} - p _ {\operatorname* {m i n}}}{\kappa} n _ {l - 1} ^ {1 / 3} \right\rfloor . +$$ + +Then, the condition $\gamma_{l-1} \geq 1/3$ is sufficient to hold the preceding inequality. On the other hand, for epoch $l > N_{\gamma_0}$ , since $l$ is an integer value, we have $l \geq \left\lfloor \log_2(T^{\gamma_0} / n_1) \right\rfloor + 3 > \log_2(T^{\gamma_0} / n_1) + 2$ and hence $n_{l-1} = n_1 2^{l-2} > T^{\gamma_0}$ . Therefore, by the equation (7) and setting $\gamma_0 = 3\gamma$ , we have + +$$ +\left\lfloor \frac {p _ {\mathrm {m a x}} - p _ {\mathrm {m i n}}}{\kappa} n _ {l - 1} ^ {\gamma_ {l - 1}} \right\rfloor < \left\lfloor \frac {p _ {\mathrm {m a x}} - p _ {\mathrm {m i n}}}{\kappa} n _ {l - 1} ^ {1 / 3} \right\rfloor +$$ + +for all $l > N_{\gamma_0}$ , and the condition $\gamma_{l-1} < 1/3$ is sufficient to hold this inequality. By Lemma B.2, there exist positive constants $c_1, \ldots, c_5$ depending on $L, B, p_{\min}, p_{\max}, \kappa, \alpha, \rho, a, b, \eta_1, \eta_2$ and $n_1$ such that for $l \geq c_1$ , + +$$ +\sum_ {t \in \mathcal {E} _ {l}} r (t) \leq \left\{ \begin{array}{l l} c _ {2} d ^ {\frac {1}{2}} n _ {l} ^ {\frac {1}{2}} \left(\log \left(d \vee n _ {l}\right)\right) ^ {\frac {1}{2}} + c _ {3} n _ {l} ^ {\frac {1 + \gamma_ {l}}{2}} \left(\log n _ {l}\right) ^ {\frac {1}{2}} & \text {i f} l > N _ {3 \gamma} \\ c _ {2} d ^ {\frac {1}{2}} n _ {l} ^ {\frac {1}{2}} \left(\log \left(d \vee n _ {l}\right)\right) ^ {\frac {1}{2}} + c _ {3} n _ {l} ^ {\frac {2}{3}} \left(\log n _ {l}\right) ^ {\frac {1}{2}} & \text {i f} l \leq N _ {3 \gamma}, \end{array} \right. \tag {73} +$$ + +with probability at least $1 - \zeta_{l}$ , where + +$$ +\zeta_ {l} = \left\{ \begin{array}{l l} c _ {4} / n _ {l - 1} ^ {\gamma_ {l - 1}} & \quad \text {i f} l > N _ {3 \gamma} \\ c _ {5} / n _ {l - 1} ^ {\frac {1}{3}} & \quad \text {i f} l \leq N _ {3 \gamma}. \end{array} \right. +$$ + +For unity of notation, we denote $N_0 \coloneqq \lceil c_1 \rceil - 1$ . Note that $N_0 < N_{3\gamma}$ for sufficiently large $T \geq (2^{N_0 + 1}n_1)^{1 / (3\gamma)}$ . Let $\Omega_{1,l}$ and $\Omega_{2,l}$ denote the events where the first and second inequalities in (73) are satisfied for each epoch $l$ , respectively. Then, we have + +$$ +\begin{array}{l} \mathbb {P} \left[ \left\{\bigcap_ {l = N _ {0} + 1} ^ {N _ {3 \gamma}} \Omega_ {1, l} \right\} \cap \left\{\bigcap_ {l = N _ {3 \gamma} + 1} ^ {N} \Omega_ {2, l} \right\} \right] \geq 1 - \sum_ {l = N _ {0} + 1} ^ {N _ {3 \gamma}} c _ {5} n _ {l - 1} ^ {- \frac {1}{3}} - \sum_ {l = N _ {3 \gamma} + 1} ^ {N} c _ {4} n _ {l - 1} ^ {- \gamma_ {l - 1}} \\ > 1 - c _ {4} \vee c _ {5} \cdot \sum_ {l = N _ {0} + 1} ^ {N} n _ {l - 1} ^ {- \gamma_ {l - 1}} \\ \geq 1 - \left(c _ {4} \vee c _ {5}\right) c _ {p} ^ {- 1} \cdot \log (T / n _ {1} + 1) T ^ {- \gamma}, \tag {74} \\ \end{array} +$$ + +where the second inequality holds because $\gamma_{l-1} < 1/3$ for $l > N_{3\gamma}$ , and the last inequality follows from $n_{l-1}^{\gamma_{l-1}} \geq c_p T^\gamma$ by (7). Here, $c_p$ is a positive constant depending on $p_{\min}, p_{\max}$ and $\kappa$ . + +Now, we decompose the cumulative regret as + +$$ +R (T) = \underbrace {\sum_ {l = 1} ^ {N _ {0}} \sum_ {t \in \mathcal {E} _ {l}} r (t)} _ {\text {(i)}} + \underbrace {\sum_ {l = N _ {0} + 1} ^ {N _ {3 \gamma}} \sum_ {t \in \mathcal {E} _ {l}} r (t)} _ {\text {(i i)}} + \underbrace {\sum_ {l = N _ {3 \gamma} + 1} ^ {N} \sum_ {t \in \mathcal {E} _ {l}} r (t)} _ {\text {(i i i)}}. +$$ + +For (i), note that $pS_0^*(p)^{\exp(\mathbf{x}^\top \beta^*)}$ is upper bounded by a positive constant $C_1 := \max \{pS_0^*(p)^{\exp(v)} : p \in [p_{\min}, p_{\max}], v \in [-BL, BL]\}$ depending on $p_{\min}, p_{\max}, B$ and $L$ . Then, we have + +$$ +\left(\mathrm {i}\right) \leq \sum_ {l = 1} ^ {N _ {0}} \sum_ {t \in \mathcal {E} _ {l}} C _ {1} = C _ {2}, +$$ + +where $C_2 = n_1(2^{N_0} - 1)C_1$ be a constant and does not depend on $T$ . Let $\Omega$ be the event in the probability notation in the display (74). For (ii), under the event $\Omega$ , by (73), we have + +$$ +\begin{array}{l} \left(\mathrm {i i}\right) \leq \sum_ {l = N _ {0} + 1} ^ {N _ {3 \gamma}} \left\{c _ {2} d ^ {\frac {1}{2}} n _ {l} ^ {\frac {1}{2}} \left(\log (d \vee n _ {l})\right) ^ {\frac {1}{2}} + c _ {3} n _ {l} ^ {\frac {2}{3}} \left(\log n _ {l}\right) ^ {\frac {1}{2}} \right\} \\ \leq C _ {3} \cdot d ^ {\frac {1}{2}} T ^ {\frac {3}{2} \gamma} (\log (d \vee T)) ^ {\frac {1}{2}} + C _ {4} \cdot T ^ {2 \gamma} (\log T) ^ {\frac {1}{2}}, \\ \end{array} +$$ + +where $C_3 = 2^{\frac{5}{2}}c_2$ and $C_4 = 2^3 c_3$ are positive constants. Similarly, for (iii), under the event $\Omega$ , we obtain + +$$ +\begin{array}{l} \left(\mathrm {i i i}\right) \leq \sum_ {l = N _ {3 \gamma} + 1} ^ {N} \left\{c _ {2} d ^ {\frac {1}{2}} n _ {l} ^ {\frac {1}{2}} \left(\log (d \vee n _ {l})\right) ^ {\frac {1}{2}} + c _ {3} n _ {l} ^ {\frac {1 + \gamma_ {l}}{2}} \left(\log n _ {l}\right) ^ {\frac {1}{2}} \right\} \\ \leq C _ {5} \cdot d ^ {\frac {1}{2}} T ^ {\frac {1}{2}} (\log (d \vee T)) ^ {\frac {1}{2}} + c _ {3} \sum_ {l = 1} ^ {N} n _ {l} ^ {\frac {\gamma_ {l}}{2}} n _ {l} ^ {\frac {1}{2}} (\log n _ {l}) ^ {\frac {1}{2}} \\ \leq C _ {5} \cdot d ^ {\frac {1}{2}} T ^ {\frac {1}{2}} (\log (d \vee T)) ^ {\frac {1}{2}} + c _ {3} C _ {p} ^ {\frac {1}{2}} \cdot T ^ {\frac {\gamma}{2}} \sum_ {l = 1} ^ {N} n _ {l} ^ {\frac {1}{2}} (\log n _ {l}) ^ {\frac {1}{2}} \\ \leq C _ {5} \cdot d ^ {\frac {1}{2}} T ^ {\frac {1}{2}} (\log (d \lor T)) ^ {\frac {1}{2}} + C _ {6} \cdot T ^ {\frac {\gamma + 1}{2}} (\log T) ^ {\frac {1}{2}}, \\ \end{array} +$$ + +where the first inequality follows from (73), the second inequality holds for a positive constant $C_5 = 2^2 c_2$ , the third inequality holds because $n_l^{\gamma l} \leq C_p T^\gamma$ for any $l = 1, \ldots, N$ by (7) with a positive constant $C_p$ depending on $p_{\mathrm{min}}$ , $p_{\mathrm{max}}$ and $\kappa$ , and the last inequality holds for a positive constant $C_6 = 2^2 c_3 C_p^{\frac{1}{2}}$ . Combining the last five displays, for sufficiently large $T \geq C_7$ , it holds that + +$$ +\begin{array}{l} R (T) \leq C _ {2} + \left(C _ {3} + C _ {5}\right) \cdot d ^ {\frac {1}{2}} T ^ {\frac {1}{2}} (\log (d \vee T)) ^ {\frac {1}{2}} + C _ {4} \cdot T ^ {2 \gamma} (\log T) ^ {\frac {1}{2}} + C _ {6} \cdot T ^ {\frac {\gamma + 1}{2}} (\log T) ^ {\frac {1}{2}} \\ \leq C _ {8} d ^ {\frac {1}{2}} T ^ {\frac {1}{2}} \left(\log (d \vee T)\right) ^ {\frac {1}{2}} + C _ {9} T ^ {\frac {\gamma + 1}{2}} \left(\log T\right) ^ {\frac {1}{2}}, \\ \end{array} +$$ + +with probability at least $1 - C_{10}\log (T / n_1 + 1) / T^{\gamma}$ , where $C_7 = (2^{2 / (1 - 3\gamma)})\vee ((2^{N_0 + 1}n_1)^{1 / (3\gamma)})$ $C_8 = C_2 + C_3 + C_5$ , $C_9 = C_4 + C_6$ and $C_{10} = (c_4\lor c_5)c_p^{-1}$ are positive constants, and the last inequality holds because $T^{2\gamma}\leq T^{\frac{\gamma + 1}{2}}$ for $\gamma < 1 / 3$ . + +Next, we consider the case where $\gamma \geq 1/3$ . For any epoch $l \leq N$ , we have $2^{l-2} < 2^l - 1 \leq T / n_1$ and hence $n_{l-1} = n_1 2^{l-2} < T$ . Therefore, by the equation (7), we have + +$$ +\left\lfloor \frac {p _ {\operatorname* {m a x}} - p _ {\operatorname* {m i n}}}{\kappa} n _ {l - 1} ^ {\gamma_ {l - 1}} \right\rfloor = \left\lfloor \frac {p _ {\operatorname* {m a x}} - p _ {\operatorname* {m i n}}}{\kappa} T ^ {\gamma} \right\rfloor \geq \left\lfloor \frac {p _ {\operatorname* {m a x}} - p _ {\operatorname* {m i n}}}{\kappa} n _ {l - 1} ^ {\gamma} \right\rfloor . +$$ + +Then, the condition $\gamma_{l-1} \geq \gamma \geq 1/3$ is sufficient to hold the last display. By Lemma B.2, the event $\Omega_{2,l}$ in (73) holds for all $l > N_0$ , and we have + +$$ +\mathbb {P} (\Omega_ {2, l}) \geq 1 - c _ {5} / n _ {l - 1} ^ {\frac {1}{3}} \quad \text {f o r} l > N _ {0}. +$$ + +We define $N_{1} \coloneqq \lfloor \log_{2}(T^{2 / 3} / n_{1})\rfloor +1$ . Note that $N_0 < N_1$ for sufficiently large $T > (2^{N_0}n_1)^{3 / 2}$ . Then, by the preceding display, we obtain + +$$ +\begin{array}{l} \mathbb {P} \left[ \left\{\bigcap_ {l = N _ {1} + 1} ^ {N} \Omega_ {2, l} \right\} \right] \geq 1 - \sum_ {l = N _ {1} + 1} ^ {N} c _ {5} n _ {l - 1} ^ {- \frac {1}{3}} \\ > 1 - 2 ^ {\frac {1}{3}} c _ {5} \sum_ {l = N _ {1} + 1} ^ {N} T ^ {- \frac {2}{9}} \\ \geq 1 - 2 ^ {\frac {1}{3}} c _ {5} \cdot \log (T / n _ {1} + 1) T ^ {- \frac {2}{9}}, \\ \end{array} +$$ + +where the second inequality holds because $n_{l-1} > 2^{-1}T^{2/3}$ for $l \geq N_1 + 1$ . Let $\Omega'$ be the event in the probability notation in the preceding display. + +Now, we decompose the cumulative regret as + +$$ +R (T) = \underbrace {\sum_ {l = 1} ^ {N _ {1}} \sum_ {t \in \mathcal {E} _ {l}} r (t)} _ {\text {(I)}} + \underbrace {\sum_ {l = N _ {1} + 1} ^ {N} \sum_ {t \in \mathcal {E} _ {l}} r (t)} _ {\text {(I I)}}. +$$ + +For (I), we have + +$$ +(\mathbf {I}) \leq \sum_ {l = 1} ^ {N _ {1}} \sum_ {t \in \mathcal {E} _ {l}} C _ {1} \leq C _ {1} \sum_ {l = 1} ^ {N _ {1}} n _ {l} \leq 2 C _ {1} \cdot T ^ {\frac {2}{3}}. +$$ + +For (II), under the event $\Omega^{\prime}$ , we obtain + +$$ +\begin{array}{l} (\mathrm {I I}) \leq \sum_ {l = N _ {1} + 1} ^ {N} \left\{c _ {2} d ^ {\frac {1}{2}} n _ {l} ^ {\frac {1}{2}} \left(\log (d \vee n _ {l})\right) ^ {\frac {1}{2}} + c _ {3} n _ {l} ^ {\frac {2}{3}} (\log n _ {l}) ^ {\frac {1}{2}} \right\} \\ \leq C _ {1 1} \cdot d ^ {\frac {1}{2}} T ^ {\frac {1}{2}} (\log (d \vee T)) ^ {\frac {1}{2}} + C _ {1 2} \cdot T ^ {\frac {2}{3}} (\log T) ^ {\frac {1}{2}}, \\ \end{array} +$$ + +where $C_{11} = 2^{2}c_{2}$ and $C_{12} = 2^{\frac{7}{3}}c_{3}$ are positive constants. Combining the last four displays, for sufficiently large $T \geq C_{13}$ , it holds that + +$$ +R (T) \leq C _ {1 1} d ^ {\frac {1}{2}} T ^ {\frac {1}{2}} (\log (d \lor T)) ^ {\frac {1}{2}} + C _ {1 4} T ^ {\frac {2}{3}} (\log T) ^ {\frac {1}{2}}, +$$ + +with probability at least $1 - C_{15}\log (T / n_1 + 1) / T^{2 / 9}$ , where $C_{13} = (2^{N_0}n_1)^{3 / 2} + 1$ , $C_{14} = 2C_1 + C_{12}$ and $C_{15} = 2^{1 / 3}c_5$ are positive constants. This completes the proof. + +# B.3 Proof of Theorem 5.3 + +Proof. Recall the grid support $\mathcal{G} = \{g_i : i = 1, \dots, K\}$ , where $g_i = p_{\min} + i\delta$ , $\delta = \kappa T^{-\gamma}$ for some $\kappa > 0$ , and $K = \left\lfloor (p_{\max} - p_{\min}) / \delta \right\rfloor$ . Let $p_t, r_t$ and $y_t$ be the offered price, the revenue and the feedback at time $t$ , respectively. Let $h_t = (p_1, r_1, p_2, r_2, \dots, p_t, r_t)$ be a history over $t$ times. Note that $h_t$ can be induced by $(p_1, y_1, \dots, p_t, y_t)$ since $r_t = p_t y_t$ . Define a policy $\pi = (\pi_t)_{t=1}^T$ , where $\pi_t$ is a conditional distribution of price $p_t$ given $h_{t-1}$ supported on $\mathcal{G}$ . We denote the conditional distribution of revenue $r_t$ given $p_t$ with respect to the complementary c.d.f. $S(p) = 1 - F(p)$ by $P_{p_t}^S$ . With abuse of notation we view $\pi_t : \mathcal{G} \to [0,1]$ and $P_{p_t}^S : \{0, p_t\} \to [0,1]$ as probability mass function. In addition, we abuse notation by writing $P_i^S = P_{p_t}^S$ if $p_t = g_i$ for some $g_i \in \mathcal{G}$ . For given $S(\cdot)$ , note that $r_t = p_t y_t$ where $y_t \sim \mathrm{Bin}(1, S(p_t))$ . Then, we have + +$$ +\begin{array}{l} P _ {p _ {t}} ^ {S} (r _ {t}) = S (p _ {t}) ^ {\frac {r _ {t}}{p _ {t}}} (1 - S (p _ {t})) ^ {1 - \frac {r _ {t}}{p _ {t}}} \\ = S \left(p _ {t}\right) ^ {y _ {t}} \left(1 - S \left(p _ {t}\right)\right) ^ {1 - y _ {t}}. \tag {75} \\ \end{array} +$$ + +For given $S$ , let $v_{S} = (P_{1}^{S}, P_{2}^{S}, \ldots, P_{K}^{S})$ be the reward distributions associated with a $K$ -armed bandit. For given policy $\pi$ and bandit $v_{S}$ , we denote the joint distribution of $(p_{1}, r_{1}, \ldots, p_{T}, r_{T})$ by $P_{v_{S}\pi}$ . Then, the probability of obtaining a fixed configuration $(p_{1}, r_{1}, \ldots, p_{T}, r_{T})$ is given by + +$$ +P _ {v _ {S} \pi} \left(p _ {1}, r _ {1}, \dots , p _ {T}, r _ {T}\right) = \prod_ {t = 1} ^ {T} \pi_ {t} \left(p _ {t} \mid h _ {t - 1}\right) P _ {p _ {t}} ^ {S} \left(r _ {t}\right). \tag {76} +$$ + +Based on this, we define the expected regret by + +$$ +R (T, S) := \mathbb {E} _ {v _ {S} \pi} \left[ \sum_ {t = 1} ^ {T} r \left(p ^ {*}, S\right) - r \left(p _ {t}, S\right) \right], +$$ + +where $r(p, S) \coloneqq pS(p)$ be the expected revenue with respect to $S$ for $p \in \mathcal{G}$ , $p^* = \operatorname{argmax}_{p \in \mathcal{G}} \{pS(p)\}$ be the optimal price and $\mathbb{E}_{v_S\pi}$ denotes the expectation under $P_{v_S\pi}$ . Further, we define the suboptimality gap of index $i$ by $\Delta_i^S \coloneqq r(p^*, S) - r(g_i, S)$ for $i = 1, \dots, K$ . For the simplicity of notation we use $P_S$ and $\mathbb{E}_S$ in place of $P_{v_S\pi}$ and $\mathbb{E}_{v_S\pi}$ , respectively, for a fixed policy $\pi$ . + +Now, we first construct two bandits $v_{S_1}$ and $v_{S_1'}$ for $S_1, S_1' \in S := \{S : \mathcal{G} \to [0,1] \mid 1 > M_2 > S(g_1) \geq \dots \geq S(g_K) > M_1 > 0 \text{ for some } 0 < M_1 < M_2 < 1\}$ in the following description. Fix a policy $\pi$ and suppose that $\gamma < 1/3$ . Let $\epsilon > 0$ be some constant to be chosen later. We define a bandit $v_{S_1} = (P_1^{S_1}, \ldots, P_K^{S_1})$ such that for some $j_1 \in [K]$ , + +$$ +\left\{ \begin{array}{l l} S _ {1} \left(g _ {i}\right) = (c + \epsilon) \cdot g _ {i} ^ {- 1} & \text {i f} i = j _ {1} \\ S _ {1} \left(g _ {i}\right) = c \cdot g _ {i} ^ {- 1} & \text {o t h e r w i s e ,} \end{array} \right. \tag {77} +$$ + +where $c > 0$ be a constant so that $S_{1} \in \mathcal{S}$ . Note that $c$ only depends on $M_{1}, M_{2}, p_{\min}$ and $p_{\max}$ . For $i = 1, \ldots, K$ , let $N_{i}(t) := \sum_{s=1}^{t} \mathbb{1}\{p_{s} = g_{i}\}$ be the number of times price $g_{i}$ was chosen by the policy over $t$ times, and $j_{1}' = \operatorname*{argmin}_{i \neq j_{1}} \mathbb{E}_{S_{1}}[N_{i}(T)]$ . Since $\sum_{i=1}^{K} \mathbb{E}_{S_{1}}[N_{i}(T)] = T$ , it holds that $\mathbb{E}_{S_{1}}[N_{j_{1}'}(T)] \leq \frac{T}{K-1}$ . + +The second bandit $v_{S_1'} = (P_1^{S_1'}, \ldots, P_K^{S_1'})$ is defined by + +$$ +\left\{ \begin{array}{l l} S _ {1} ^ {\prime} (g _ {i}) = (c + \epsilon) \cdot g _ {i} ^ {- 1} & \text {i f} i = j _ {1} \\ S _ {1} ^ {\prime} (g _ {i}) = (c + 2 \epsilon) \cdot g _ {i} ^ {- 1} & \text {i f} i = j _ {1} ^ {\prime} \\ S _ {1} ^ {\prime} (g _ {i}) = c \cdot g _ {i} ^ {- 1} & \text {o t h e r w i s e .} \end{array} \right. \tag {78} +$$ + +Therefore, $r(g_{i},S_{1}) = r(g_{i},S_{1}^{\prime})$ except at index $j_1^\prime$ and the optimal price in $v_{S_1}$ is $g_{j_1}$ , while in $v_{S_1'}$ , $g_{j_1'}$ is the optimal price. Then, we have + +$$ +\begin{array}{l} R (T, S _ {1}) = \mathbb {E} _ {S _ {1}} \left[ \sum_ {t = 1} ^ {T} r \left(g _ {j _ {1}}, S _ {1}\right) - r \left(p _ {t}, S _ {1}\right) \right] \\ = \sum_ {i = 1} ^ {K} \Delta_ {i} ^ {S _ {1}} \mathbb {E} _ {S _ {1}} [ N _ {i} (T) ] \\ = \sum_ {i \in [ K ], i \neq j _ {1}} \epsilon \cdot \mathbb {E} _ {S _ {1}} \left[ N _ {i} (T) \right] \\ = \epsilon (T - \mathbb {E} _ {S _ {1}} [ N _ {j _ {1}} (T) ]) \\ \geq \frac {T \epsilon}{2} \cdot P _ {S _ {1}} \left(N _ {j _ {1}} (T) \leq \frac {T}{2}\right), \\ \end{array} +$$ + +where the second equality holds by the regret decomposition and the third equality holds because $\Delta_i^{S_1} = (c + \epsilon) - c = \epsilon$ for $i\neq j_{1}$ . Similarly, we have + +$$ +\begin{array}{l} R (T, S _ {1} ^ {\prime}) = \sum_ {i = 1} ^ {K} \Delta_ {i} ^ {S _ {1} ^ {\prime}} \mathbb {E} _ {S _ {1} ^ {\prime}} [ N _ {i} (T) ] \\ > \epsilon \cdot \mathbb {E} _ {S _ {1} ^ {\prime}} \left[ N _ {j _ {1}} (T) \right] \\ > \frac {T \epsilon}{2} \cdot P _ {S _ {1} ^ {\prime}} \left(N _ {j _ {1}} (T) > \frac {T}{2}\right). \\ \end{array} +$$ + +Combining the last two displays and Lemma 2.6 in [42], we have + +$$ +\begin{array}{l} R \left(T, S _ {1}\right) + R \left(T, S _ {1} ^ {\prime}\right) > \frac {T \epsilon}{2} \left(P _ {S _ {1}} \left(N _ {j _ {1}} (T) \leq \frac {T}{2}\right) + P _ {S _ {1} ^ {\prime}} \left(N _ {j _ {1}} (T) > \frac {T}{2}\right)\right) \\ \geq \frac {T \epsilon}{4} \exp \left(- K \left(P _ {S _ {1}}, P _ {S _ {1} ^ {\prime}}\right)\right). \\ \end{array} +$$ + +By Lemma 1 in [14], the KL divergence $K(P_{S_1}, P_{S_1'})$ is bounded by + +$$ +\begin{array}{l} K \left(P _ {S _ {1}}, P _ {S _ {1} ^ {\prime}}\right) = \sum_ {i = 1} ^ {K} \mathbb {E} _ {S _ {1}} \left[ N _ {i} (T) \right] K \left(P _ {i} ^ {S _ {1}}, P _ {i} ^ {S _ {1} ^ {\prime}}\right) \\ = \mathbb {E} _ {S _ {1}} \left[ N _ {j _ {1} ^ {\prime}} (T) \right] K (P _ {j _ {1} ^ {\prime}} ^ {S _ {1}}, P _ {j _ {1} ^ {\prime}} ^ {S _ {1} ^ {\prime}}) \\ \leq \frac {T}{K - 1} K (P _ {j _ {1} ^ {\prime}} ^ {S _ {1}}, P _ {j _ {1} ^ {\prime}} ^ {S _ {1} ^ {\prime}}), \\ \end{array} +$$ + +where the first equality holds by (76), the second equality holds by the definition of $S_{1}$ and $S_{1}^{\prime}$ , and the first inequality holds by the definition of index $j_{1}^{\prime}$ . Note that (75) implies that $P_{j_1'}^{S_1}$ and $P_{j_1'}^{S_1'}$ are distributions of Bernoulli random variables with parameters $S_{1}(g_{j_{1}^{\prime}})$ and $S_{1}^{\prime}(g_{j_{1}^{\prime}})$ , respectively. Therefore, by Corollary 3.1 in [40], we have + +$$ +\begin{array}{l} K \left(P _ {j _ {1} ^ {\prime}} ^ {S _ {1}}, P _ {j _ {1} ^ {\prime}} ^ {S _ {1} ^ {\prime}}\right) \leq \frac {\left(S _ {1} \left(g _ {j _ {1} ^ {\prime}}\right) - S _ {1} ^ {\prime} \left(g _ {j _ {1} ^ {\prime}}\right)\right) ^ {2}}{S _ {1} ^ {\prime} \left(g _ {j _ {1} ^ {\prime}}\right) \left(1 - S _ {1} ^ {\prime} \left(g _ {j _ {1} ^ {\prime}}\right)\right)} \\ < \frac {(2 \epsilon g _ {j _ {1} ^ {\prime}} ^ {- 1}) ^ {2}}{M _ {1} M _ {2}} \\ \leq \frac {4}{p _ {\mathrm {m i n}} ^ {2} M _ {1} M _ {2}} \epsilon^ {2}, \\ \end{array} +$$ + +where the second inequality holds by the definition of $S_{1}$ and $S_{1}^{\prime}$ , and the last inequality holds because $g_{i}\in [p_{\mathrm{min}},p_{\mathrm{max}}]$ for any $i\in [K]$ . + +Now, it remains to choose $\epsilon$ . Due to the monotonicity of the distribution functions $S_{1}$ and $S_{1}^{\prime}$ , there are additional constraints in choosing $\epsilon$ . Specifically, by the definition of $S_{1}$ (77), the following + +condition must hold: $(c + \epsilon) \cdot g_{j_1}^{-1} \leq c \cdot g_{j_1 - 1}^{-1}$ . By the direct calculations, we obtain $\epsilon \leq \frac{c}{g_{j_1 - 1}}\delta$ . Since $g_i \in [p_{\min}, p_{\max}]$ for any $i \in [K]$ , it is sufficient to choose $\epsilon \leq \frac{c}{p_{\max}}\delta$ to satisfy this condition. Similarly, we consider the monotonicity of $S_1'$ , but before doing so, we divide it into two cases: (a) $j_1 < j_1'$ and (b) $j_1 > j_1'$ . For the case (a), it is necessary that $S_1'(g_{j_1}') \leq S_1'(g_{j_1}) \leq S_1'(g_{j_1 - 1})$ holds, and for the case (b), $S_1'(g_{j_1}') \leq S_1'(g_{j_1' - 1})$ must hold. By the simple calculations, $\epsilon \leq \frac{c}{2p_{\max}}\delta$ is sufficient to satisfy the above conditions. Since $\sqrt{K / T} \ll \delta$ if $\gamma < 1 / 3$ , it is sufficient to choose $\epsilon = C\sqrt{K / T}$ for a small enough constant $C > 0$ . Combining this with the three preceding displays, there exists a constant $C_1 > 0$ such that + +$$ +\begin{array}{l} R (T, S _ {1}) + R (T, S _ {1} ^ {\prime}) \geq C _ {1} \sqrt {K T} \\ \gtrsim T ^ {\frac {1 + \gamma}{2}}. \\ \end{array} +$$ + +It completes the proof for the case $\gamma < 1 / 3$ . + +In the second case that $\gamma \geq 1/3$ , we construct another pair of bandits $v_{S_2}$ and $v_{S_2'}$ for $S_2, S_2' \in \mathcal{S}$ in the following description. Let $\epsilon_2 = \kappa T^{-\frac{1}{3}}$ and $i_1, \ldots, i_J$ be positive integers such that $p_{\min} + j\epsilon_2 - \delta < g_{i_j} \leq p_{\min} + j\epsilon_2$ for $j = 1, \ldots, J$ , where $J = \left\lfloor (p_{\max} - p_{\min}) / \epsilon_2 \right\rfloor$ . We define partitions $I_j$ of index set $[K]$ by $I_j = \{i \in [K] : i_{j-1} < i \leq i_j\}$ for $j = 1, \ldots, J$ , where $i_0 = 0$ with $g_0 = p_{\min}$ . Then, we define a bandit $v_{S_2} = (P_1^{S_2}, \ldots, P_K^{S_2})$ such that for some $j_2 \in [J]$ , + +$$ +\left\{ \begin{array}{l l} S _ {2} \left(g _ {i}\right) = \left(c + \epsilon_ {2}\right) \cdot g _ {i _ {j _ {2}}} ^ {- 1} & \text {i f} i \in I _ {j _ {2}} \\ S _ {2} \left(g _ {i}\right) = c \cdot g _ {i _ {j}} ^ {- 1} & \text {i f} i \in I _ {j} \text {f o r} j \in [ J ] \text {e x c e p t a t} j _ {2}. \end{array} \right. \tag {79} +$$ + +Let $M_{j}(t) \coloneqq \sum_{i \in I_{j}} N_{i}(t)$ for $j = 1, \ldots, J$ , and $j_{2}' = \operatorname{argmin}_{j \neq j_{2}} \mathbb{E}_{S_{2}}[M_{j}(T)]$ . Since $\sum_{j=1}^{J} \mathbb{E}_{S_{2}}[M_{j}(T)] = T$ , it holds that $\mathbb{E}_{S_{2}}[M_{j_{2}'}(T)] \leq \frac{T}{J-1}$ . Then, the second bandit $v_{S_{2}'} = (P_{1}^{S_{2}'}', \ldots, P_{K}^{S_{2}'}')$ is defined by + +$$ +\left\{ \begin{array}{l l} S _ {2} ^ {\prime} (g _ {i}) = (c + \epsilon_ {2}) \cdot g _ {i _ {j _ {2}}} ^ {- 1} & \text {i f} i \in I _ {j _ {2}} \\ S _ {2} ^ {\prime} (g _ {i}) = (c + 2 \epsilon_ {2}) \cdot g _ {i _ {j _ {2} ^ {\prime}}} ^ {- 1} & \text {i f} i \in I _ {j _ {2} ^ {\prime}} \\ S _ {2} ^ {\prime} (g _ {i}) = c \cdot g _ {i _ {j}} ^ {- 1} & \text {i f} i \in I _ {j} \text {f o r} j \in [ J ] \text {e x c e p t a t} j _ {2} \text {a n d} j _ {2} ^ {\prime}. \end{array} \right. \tag {80} +$$ + +Therefore, $r(g_{i}, S_{2}) = r(g_{i}, S_{2}^{\prime})$ except at index $i \in I_{j_2^{\prime}}$ and the optimal price in $v_{S_2}$ is $g_{i_{j_2}}$ , while in $v_{S_2^{\prime}}$ , $g_{i_{j_2^{\prime}}}$ is the optimal price. For $j = 1, \dots, J$ except at $j_2$ , note that $\Delta_i^{S_2} \geq \epsilon_2$ for $i \in I_j$ since $r(g_{i_{j_2}}, S_2) = c + \epsilon_2$ and $r(g_{i}, S_2) \leq c$ by the definition of $S_2$ . Then, we have + +$$ +\begin{array}{l} R (T, S _ {2}) = \sum_ {i = 1} ^ {K} \Delta_ {i} ^ {S _ {2}} \mathbb {E} _ {S _ {2}} [ N _ {i} (T) ] \\ = \sum_ {j \in [ J ], j \neq j _ {2}} \sum_ {i \in I _ {j}} \Delta_ {i} ^ {S _ {2}} \mathbb {E} _ {S _ {2}} \left[ N _ {i} (T) \right] \\ \geq \sum_ {j \in [ J ], j \neq j _ {2}} \epsilon_ {2} \cdot \mathbb {E} _ {S _ {2}} [ M _ {j} (T) ] \\ = \epsilon_ {2} \left(T - \mathbb {E} _ {S _ {2}} \left[ M _ {j _ {2}} (T) \right]\right) \\ \geq \frac {T \epsilon}{2} \cdot P _ {S _ {2}} \left(M _ {j _ {2}} (T) \leq \frac {T}{2}\right), \\ \end{array} +$$ + +where the first inequality holds by the definition of $M_{j}(T)$ . Similarly, we have + +$$ +\begin{array}{l} R (T, S _ {2} ^ {\prime}) = \sum_ {j \in [ J ], j \neq j _ {2} ^ {\prime}} \sum_ {i \in I _ {j}} \Delta_ {i} ^ {S _ {2} ^ {\prime}} \mathbb {E} _ {S _ {2} ^ {\prime}} \left[ N _ {i} (T) \right] \\ > \epsilon_ {2} \cdot \mathbb {E} _ {S _ {2} ^ {\prime}} [ M _ {j 2} (T) ] \\ > \frac {T \epsilon_ {2}}{2} \cdot P _ {S _ {2} ^ {\prime}} \left(M _ {j _ {2}} (T) > \frac {T}{2}\right). \\ \end{array} +$$ + +Combining the two preceding displays and Lemma 2.6 in [42], we have + +$$ +\begin{array}{l} R \left(T, S _ {2}\right) + R \left(T, S _ {2} ^ {\prime}\right) > \frac {T \epsilon_ {2}}{2} \left(P _ {S _ {2}} \left(M _ {j _ {2}} (T) \leq \frac {T}{2}\right) + P _ {S _ {2} ^ {\prime}} \left(M _ {j _ {2}} (T) > \frac {T}{2}\right)\right) \tag {81} \\ \geq \frac {T \epsilon_ {2}}{4} \exp \left(- K \left(P _ {S _ {2}}, P _ {S _ {2} ^ {\prime}}\right)\right). \\ \end{array} +$$ + +By Lemma 1 in [14], we can decompose the KL divergence $K(P_{S_2}, P_{S_2'})$ as + +$$ +\begin{array}{l} K \left(P _ {S _ {2}}, P _ {S _ {2} ^ {\prime}}\right) = \sum_ {j = 1} ^ {J} \sum_ {i \in I _ {j}} \mathbb {E} _ {S _ {2}} \left[ N _ {i} (T) \right] K \left(P _ {i} ^ {S _ {2}}, P _ {i} ^ {S _ {2} ^ {\prime}}\right) \\ = \sum_ {i \in I _ {j _ {2} ^ {\prime}}} \mathbb {E} _ {S _ {2}} \left[ N _ {i} (T) \right] K (P _ {i} ^ {S _ {2}}, P _ {i} ^ {S _ {2} ^ {\prime}}). \\ \end{array} +$$ + +By Corollary 3.1 in [40] and the definition of $S_{2}, S_{2}^{\prime}$ , we have + +$$ +\begin{array}{l} K \left(P _ {i} ^ {S _ {2}}, P _ {i} ^ {S _ {2} ^ {\prime}}\right) \leq \frac {\left(S _ {2} \left(g _ {i}\right) - S _ {2} ^ {\prime} \left(g _ {i}\right)\right) ^ {2}}{S _ {2} ^ {\prime} \left(g _ {i}\right) \left(1 - S _ {2} ^ {\prime} \left(g _ {i}\right)\right)} \\ < \frac {(2 \epsilon_ {2} g _ {i _ {j _ {2}} ^ {\prime}} ^ {- 1}) ^ {2}}{M _ {1} M _ {2}} \\ \leq \frac {4}{p _ {\mathrm {m i n}} ^ {2} M _ {1} M _ {2}} \epsilon_ {2} ^ {2} \\ \end{array} +$$ + +for any $i\in I_{j_2'}$ . Then, by combining the two preceding displays, we have + +$$ +\begin{array}{l} K (P _ {S _ {2}}, P _ {S _ {2} ^ {\prime}}) = \sum_ {i \in I _ {j _ {2} ^ {\prime}}} \mathbb {E} _ {S _ {2}} \left[ N _ {i} (T) \right] K (P _ {i} ^ {S _ {2}}, P _ {i} ^ {S _ {2} ^ {\prime}}). \\ < c _ {2} \epsilon_ {2} ^ {2} \cdot \mathbb {E} _ {S _ {2}} \left[ M _ {j _ {2} ^ {\prime}} (T) \right] \tag {82} \\ \leq c _ {2} \frac {T}{J - 1} \epsilon_ {2} ^ {2}, \\ \end{array} +$$ + +where $c_{2} = \frac{4}{p_{\mathrm{min}}^{2}M_{1}M_{2}}$ . It is easy to check that $\epsilon_{2} = \kappa T^{-\frac{1}{3}}$ is sufficient to satisfy the monotonicity constraints of $S_{2}$ and $S_{2}'$ . Therefore, by combining (81), (82) and $J \asymp \epsilon_{2}^{-1}$ , there exists a constant $C_{2} > 0$ such that + +$$ +\begin{array}{l} R (T, S _ {2}) + R (T, S _ {2} ^ {\prime}) \geq C _ {2} T \cdot T ^ {- \frac {1}{3}} \\ \gtrsim T ^ {\frac {2}{3}}. \\ \end{array} +$$ + +It completes the proof for the case $\gamma \geq 1 / 3$ . + +![](images/2a8b9d92e08eb7040b104b247ff4869f5d24b9fdc22b4154f5998d3d141c5fc3.jpg) + +# C Technical Lemmas + +Lemma C.1. Let $\Theta' = \{(\mathbf{S}_0, \beta) \in \Theta : S_{0,K} \geq M_1, S_{0,1} \leq M_2, \| \beta \|_2 \leq D\}$ , where $M_1, M_2$ and $D$ are some positive constants such that $0 < M_1 < M_2 < 1$ . Suppose that the assumption (A2) holds. Then, it holds that for any $\theta_1, \theta_2 \in \Theta'$ , + +$$ +\mathcal {D} _ {H} \left(p _ {\theta_ {1}}, p _ {\theta_ {2}}\right) \asymp \left(\int_ {x \in \mathcal {X}} \sum_ {p \in \mathcal {G}} \left| H _ {\theta_ {1}} (x, p) - H _ {\theta_ {2}} (x, p) \right| ^ {2} q (p | x) p _ {X} (x) d x\right) ^ {\frac {1}{2}}, +$$ + +where constants in $\asymp$ depend only on $M_1, M_2, D$ and $L$ . + +Proof. By the assumption (A2), there exist positive constants $H_{1}$ and $H_{2}$ , depending on $M_{1}, M_{2}, D$ and $L$ , such that $0 < H_{1} < H_{\theta}(x,p) < H_{2} < 1$ for all $x \in \mathcal{X}, p \in \mathcal{G}$ and $\theta \in \Theta'$ . Then, for any $\theta_{1}, \theta_{2} \in \Theta'$ , we have + +$$ +\begin{array}{l} \mathcal {D} _ {H} ^ {2} \left(p _ {\theta_ {1}}, p _ {\theta_ {2}}\right) = \int_ {x \in \mathcal {X}} \sum_ {p \in \mathcal {G}} \sum_ {y = 0, 1} \left(\sqrt {p _ {\theta_ {1}} (x , p , y)} - \sqrt {p _ {\theta_ {2}} (x , p , y)}\right) ^ {2} d x \\ = \int_ {x \in \mathcal {X}} \sum_ {p \in \mathcal {G}} \left[ \left(\sqrt {H _ {\theta_ {1}} (x , p)} - \sqrt {H _ {\theta_ {2}} (x , p)}\right) ^ {2} \right. \\ \left. + \left(\sqrt {1 - H _ {\theta_ {1}} (x , p)} - \sqrt {1 - H _ {\theta_ {2}} (x , p)}\right) ^ {2} \right] q (p | x) p _ {X} (x) d x \\ \asymp \int_ {x \in \mathcal {X}} \sum_ {p \in \mathcal {G}} \left| H _ {\theta_ {1}} (x, p) - H _ {\theta_ {2}} (x, p) \right| ^ {2} q (p | x) p _ {X} (x) d x, \\ \end{array} +$$ + +where the third identity holds because the derivative of the map $t \mapsto \sqrt{t}$ is bounded below and above by positive constants on the interval $[H_1, H_2]$ . + +![](images/01d8bf1da81c023add6c0726ac76c78606cc7a7a195941dfea659ed78d42e1a5.jpg) + +Lemma C.2. Let $\Theta' = \{(\mathbf{S}_0, \beta) \in \Theta : S_{0,K} \geq M_1, S_{0,1} \leq M_2, \| \beta \|_2 \leq D\}$ , where $M_1, M_2$ and $D$ are some positive constants such that $0 < M_1 < M_2 < 1$ . Suppose that the assumption (A2) holds. Then, there exist positive constants $C_1$ and $C_2$ depending on $M_1, M_2, D$ and $L$ such that for any $\theta_1 = (\mathbf{S}_{0,1}, \beta_1)$ , $\theta_2 = (\mathbf{S}_{0,2}, \beta_2) \in \Theta'$ and $p \in \mathcal{G}$ , it holds that + +$$ +\left| H _ {\theta_ {1}} (X, p) - H _ {\theta_ {2}} (X, p) \right| \leq C _ {1} \left| S _ {0, 1} (p) - S _ {0, 2} (p) \right| + C _ {2} \| \beta_ {1} - \beta_ {2} \| _ {2} +$$ + +almost surely. + +Proof. We decompose the term $|H_{\theta_1}(X,p) - H_{\theta_2}(X,p)|$ as + +$$ +\begin{array}{l} | H _ {\theta_ {1}} (X, p) - H _ {\theta_ {2}} (X, p) | = | S _ {0, 1} (p) ^ {\exp (X ^ {\top} \beta_ {1})} - S _ {0, 2} (p) ^ {\exp (X ^ {\top} \beta_ {2})} | \\ \leq \left| S _ {0, 1} (p) ^ {\exp \left(X ^ {\top} \beta_ {1}\right)} - S _ {0, 2} (p) ^ {\exp \left(X ^ {\top} \beta_ {1}\right)} \right| + \left| S _ {0, 2} (p) ^ {\exp \left(X ^ {\top} \beta_ {1}\right)} - S _ {0, 2} (p) ^ {\exp \left(X ^ {\top} \beta_ {2}\right)} \right|. \tag {83} \\ \end{array} +$$ + +For the first term of the preceding display, the mean value theorem on a map $t \mapsto t^c$ ( $c > 0$ a constant) yields + +$$ +| S _ {0, 1} (p) ^ {\exp (X ^ {\top} \beta_ {1})} - S _ {0, 2} (p) ^ {\exp (X ^ {\top} \beta_ {1})} | = \exp (X ^ {\top} \beta_ {1}) \overline {{S}} _ {0} (p) ^ {\exp (X ^ {\top} \beta_ {1}) - 1} | S _ {0, 1} (p) - S _ {0, 2} (p) |, +$$ + +for some $\overline{S}_0(p)$ in $(S_{0,1}(p), S_{0,2}(p))$ . Under the assumption (A2), by the Cauchy-Schwart inequality and the boundedness of $\beta_1$ , we have $|X^\top \beta_1| \leq \| X\|_2 \| \beta_1\|_2 \leq LD$ almost surely. Furthermore, $\overline{S}_0(p)$ is bounded away from 0 and 1. Then, there exists a positive constant $C_1$ , depending on $M_1$ , $M_2$ , $D$ and $L$ , such that $\exp(X^\top \beta_1) \overline{S}_0(p)^{\exp(X^\top \beta_1) - 1} < C_1$ . Therefore, the first term of (83) is bounded by $C_1 |S_{0,1}(p) - S_{0,2}(p)|$ . Similarly, for the second term of (83), applying the mean value theorem to the map $t \mapsto c'^{\exp(t)}$ ( $c' > 0$ a constant) gives + +$$ +\left| S _ {0, 2} (p) ^ {\exp \left(X ^ {\top} \beta_ {1}\right)} - S _ {0, 2} (p) ^ {\exp \left(X ^ {\top} \beta_ {2}\right)} \right| = \left| \log S _ {0, 2} (p) \right| S _ {0, 2} (p) ^ {\exp \left(X ^ {\top} \bar {\beta}\right)} \exp \left(X ^ {\top} \bar {\beta}\right) \left| X ^ {\top} \left(\beta_ {1} - \beta_ {2}\right) \right|, +$$ + +for some $\bar{\beta}$ between $\beta_{1}$ and $\beta_{2}$ . Note that by the Cauchy-Schwartz inequality, $|X^{\top}(\beta_{1} - \beta_{2})| \leq \|X\|_{2}\|\beta_{1} - \beta_{2}\|_{2}$ . By assumption (A2) and the boundedness of $S_{0,2}$ and $\bar{\beta}$ , there exists a positive constant $C_{2}$ , depending on $M_{1}$ , $M_{2}$ , $D$ and $L$ , such that $|\log S_{0,2}(p)|S_{0,2}(p)^{\exp(X^{\top}\bar{\beta})}\exp(X^{\top}\bar{\beta})\|X\|_{2} < C_{2}$ almost surely. Combining these results with (83), we have + +$$ +\left| H _ {\theta_ {1}} (X, p) - H _ {\theta_ {2}} (X, p) \right| < C _ {1} \left| S _ {0, 1} (p) - S _ {0, 2} (p) \right| + C _ {2} \| \beta_ {1} - \beta_ {2} \| _ {2} +$$ + +almost surely for any $p\in \mathcal{G}$ + +![](images/8d3f3cb4a35c108d46a718e0557cf56ae814db45e2de80ab1312f67396f75483.jpg) + +Lemma C.3. Suppose that the assumption (A5) holds. If $\lambda_{0,k} \sim \mathrm{Gamma}(\alpha_k, \rho)$ are independent for $k = 1, \ldots, K$ , where $A\epsilon^b \leq \alpha_k \leq M$ , and $K\epsilon \leq N$ for some positive constants $A, \epsilon, b, M, N$ and $\rho$ , then there exist positive constants $C_1, C_2$ and $C_3$ depending only on $p_{\min}, p_{\max}, A, b, M, N$ and $\rho$ such that + +$$ +\Pi \left(\| \boldsymbol {\Lambda} _ {0} - \boldsymbol {\Lambda} _ {0} ^ {*} \| _ {\infty} \leq \epsilon\right) \geq C _ {1} \exp \left(- C _ {2} K - C _ {3} K \log_ {-} \epsilon\right). +$$ + +Proof. First, we assume that $M = 1$ , so that $\alpha_{k} \leq 1$ for all $k = 1, \ldots, K$ . Fix $k \in \{1, \ldots, K\}$ . In the model (2), we can represent $\Lambda_{0}(g_{k})$ as $\delta \sum_{s=1}^{k} \lambda_{0,s}$ . Similarly, $\Lambda_{0}^{*}(g_{k})$ is given by $\delta \sum_{s=1}^{k} \Delta_{0,s}^{*}$ where $\Delta_{0,1}^{*} = \Lambda_{0}^{*}(g_{1}) / \delta$ and $\Delta_{0,s}^{*} = (\Lambda_{0}^{*}(g_{s}) - \Lambda_{0}^{*}(g_{s-1})) / \delta$ for $s = 2, \ldots, K$ . By combining this and the preceding display, we have + +$$ +\begin{array}{l} \| \boldsymbol {\Lambda} _ {0} - \boldsymbol {\Lambda} _ {0} ^ {*} \| _ {\infty} = \max _ {1 \leq k \leq K} | \Lambda_ {0} (g _ {k}) - \Lambda_ {0} ^ {*} (g _ {k}) | \\ \leq \max _ {1 \leq k \leq K} \delta \sum_ {s = 1} ^ {k} \left| \lambda_ {0, s} - \Delta_ {0, s} ^ {*} \right| \\ = \delta \sum_ {k = 1} ^ {K} \left| \lambda_ {0, k} - \Delta_ {0, k} ^ {*} \right|. \\ \end{array} +$$ + +Therefore, the probability on the left side of the lemma is bounded below by + +$$ +\begin{array}{l} \Pi \left(\| \boldsymbol {\Lambda} _ {0} - \boldsymbol {\Lambda} _ {0} ^ {*} \| _ {\infty} \leq \epsilon\right) \geq \Pi \left(\delta \sum_ {k = 1} ^ {K} \left| \lambda_ {0, k} - \Delta_ {0, k} ^ {*} \right| \leq \epsilon\right) \\ \geq \prod_ {k = 1} ^ {K} \Pi \left(\left| \lambda_ {0, k} - \Delta_ {0, k} ^ {*} \right| \leq C _ {p} \epsilon\right), \\ \end{array} +$$ + +where $C_p \coloneqq (K\delta)^{-1}$ be a positive constant depending only on $p_{\mathrm{min}}$ and $p_{\mathrm{max}}$ . Since $\lambda_{0,1}, \ldots, \lambda_{0,K}$ are independent variables with $\lambda_{0,k} \sim \mathrm{Gamma}(\alpha_k, \rho)$ , we have + +$$ +\begin{array}{l} \Pi \left(\left\| \boldsymbol {\Lambda} _ {0} - \boldsymbol {\Lambda} _ {0} ^ {*} \right\| _ {\infty} \leq \epsilon\right) \\ \geq \frac {\rho^ {\sum_ {k = 1} ^ {K} \alpha_ {k}}}{\prod_ {k = 1} ^ {K} \Gamma (\alpha_ {k})} \int_ {\max \left(\Delta_ {0, K} ^ {*} - C _ {p} \epsilon , 0\right)} ^ {\Delta_ {0, K} ^ {*} + C _ {p} \epsilon} \dots \int_ {\max \left(\Delta_ {0, 1} ^ {*} - C _ {p} \epsilon , 0\right)} ^ {\Delta_ {0, 1} ^ {*} + C _ {p} \epsilon} \prod_ {k = 1} ^ {K} u _ {k} ^ {\alpha_ {k} - 1} \exp (- \rho \sum_ {k = 1} ^ {K} u _ {k}) d u _ {1} \dots d u _ {K}. \\ \end{array} +$$ + +By assumption (A5), there exist constants $M_{1}$ and $M_2$ such that $0 < M_1\leq S_0^* (p_{\mathrm{max}}) < S_0^* (p_{\mathrm{min}})\leq$ $M_2 < 1$ , and it holds that $N_{2}\le \Lambda_{0}(v)\le N_{1}$ for all $v\in [p_{\mathrm{min}},p_{\mathrm{max}}]$ , where $N_{1} = -\log M_{1}$ and $N_{2} = -\log M_{2}$ . Note that within the interval of integration, $\sum_{k = 1}^{K}u_{k}\leq \Lambda_{0}^{*}(g_{K}) / \delta +C_{p}K\epsilon <$ $N_{1}C_{p}K + C_{p}K\epsilon$ . Furthermore, for any $0 < \alpha_{k}\leq 1$ , it holds that $\alpha_{k}\Gamma (\alpha_{k}) = \Gamma (\alpha_{k} + 1)\leq 1$ Therefore, the right side of the preceding display is bounded below by + +$$ +\rho^ {\sum_ {k = 1} ^ {K} \alpha_ {k}} \exp (- \rho C _ {p} K (N _ {1} + \epsilon)) \prod_ {k = 1} ^ {K} \left\{(\Delta_ {0, k} ^ {*} + C _ {p} \epsilon) ^ {\alpha_ {k}} - (\max (\Delta_ {0, k} ^ {*} - C _ {p} \epsilon , 0)) ^ {\alpha_ {k}} \right\}. +$$ + +By the mean value theorem, the terms of the product in the preceding display is bounded below by $\alpha_{k}(\overline{\Delta}_{0,k}^{*})^{\alpha_{k}-1}C_{p}\epsilon$ for some $\overline{\Delta}_{0,k}^{*} \in (\max(\Delta_{0,k}^{*} - C_{p}\epsilon, 0), \Delta_{0,k}^{*} + C_{p}\epsilon)$ . Since $\overline{\Delta}_{0,k}^{*} < N_{1}C_{p}K + C_{p}\epsilon$ and $\alpha_{k} - 1 \leq 0$ for all $k = 1, \ldots, K$ , by combining the last two displays, we have + +$$ +\begin{array}{l} \Pi \left(\left\| \boldsymbol {\Lambda} _ {0} - \boldsymbol {\Lambda} _ {0} ^ {*} \right\| _ {\infty} \leq \epsilon\right) \\ \geq \rho^ {\sum_ {k = 1} ^ {K} \alpha_ {k}} \exp (- \rho C _ {p} N _ {1} K)) \exp (- \rho C _ {p} K \epsilon) \cdot (C _ {p} \epsilon) ^ {K} (N _ {1} C _ {p} K + C _ {p} \epsilon) ^ {\sum_ {k = 1} ^ {K} \alpha_ {k} - K} \prod_ {k = 1} ^ {K} \alpha_ {k}. \\ \end{array} +$$ + +Note that $N_{1}C_{p}K + C_{p}\epsilon = 1 / \epsilon \cdot (N_{1}C_{p}K\epsilon + C_{p}\epsilon^{2}) \leq 1 / (C^{\prime}\epsilon)$ where $C^{\prime} := 1 / (N_{1}NC_{p} + C_{p}(A^{2 / b})^{-1})$ is a positive constant, as $K\epsilon \leq N$ and $A\epsilon^{b} \leq 1$ by assumption. Therefore, we have + +$$ +\begin{array}{l} \Pi \left(\left\| \boldsymbol {\Lambda} _ {0} - \boldsymbol {\Lambda} _ {0} ^ {*} \right\| _ {\infty} \leq \epsilon\right) \\ \geq \rho^ {\sum_ {k = 1} ^ {K} \alpha_ {k}} \exp (- \rho C _ {p} N _ {1} K)) \exp (- \rho C _ {p} N) (C _ {p} \epsilon) ^ {K} (1 / (C ^ {\prime} \epsilon)) ^ {\sum_ {k = 1} ^ {K} \alpha_ {k} - K} (A \epsilon^ {b}) ^ {K} \\ \geq C _ {1} \exp \left(- C _ {2} K - C _ {3} K \log_ {-} \epsilon\right), \\ \end{array} +$$ + +for positive constants $C_1 \coloneqq \exp(-\rho C_p N)$ , $C_2 \coloneqq \rho C_p N_1 + \log_- \rho + \log_- A + \log_- C_p + \log_- C'$ and $C_3 \coloneqq b + 2$ , where the first inequality holds because $K\epsilon \leq N$ and $A\epsilon^b \leq \alpha_k \leq 1$ by assumption, the second inequality holds because $\log x \geq -\log_- x$ for any $x > 0$ , where $\log_-$ denotes the negative parts of logarithm. This concludes the proof in the case that $M = 1$ . + +We may assume without loss of generality that a $M$ is an positive integer. For each $k = 1, \ldots, K$ , we can represent the $\lambda_{0,k}$ as the sum of independent random variables $(\lambda_{0,k,m} : m = 1, \ldots, M)$ , where $\lambda_{0,k,m}$ is distributed from Gamma distribution with parameters $\alpha_{k,m} = \alpha_k / M$ and $\rho$ for $m = 1, \ldots, M$ . Then, it satisfies the conditions of the lemma in the case of $M = 1$ , with $K$ and $A$ being adjusted to $MK$ and $A / M$ , respectively. The proof is then complete. + +![](images/8014ff116608b42731a171e40b2586af911cb271f7a5edc8e7c3f5676f960323.jpg) + +Lemma C.4. Under the assumption (A3), for a random sample $X_{t} = (X_{t,1},\ldots ,X_{t,d})$ , there is a constant $\epsilon >0$ such that $\mathbb{P}(|X_{t,j}| > \epsilon) > 0$ for all $j = 1,\dots ,d$ . + +Proof. Suppose that for any $\epsilon > 0$ , there exists $j' \in \{1, \ldots, d\}$ such that $\mathbb{P}(|X_{t,j'}| > \epsilon) = 0$ . It follows that $\mathbb{P}(\|X_t\|_\infty > \epsilon) = 0$ . Since $\epsilon > 0$ is an arbitrary number, we have $\mathbb{P}(\|X_t\|_\infty = 0) = 1$ . Take $j^* = \operatorname{argmax}_{j=1,\dots,d} |X_{t,j}|$ and $\beta_1, \beta_2$ such that $\beta_{1,j^*} \neq \beta_{2,j^*}$ and $\beta_{1,j} = \beta_{2,j}$ for $j \in [d] \setminus \{j^*\}$ . Note that $X_t^\top (\beta_1 - \beta_2) = X_{t,j^*} (\beta_{1,j^*} - \beta_{1,j^*})$ . Since $\mathbb{P}(|X_{t,j^*}| = 0) = 1$ , we have $\mathbb{P}(X_t^\top (\beta_1 - \beta_2) = 0) = 1$ . This contradicts the fact that $\mathbb{P}(X_t^\top (\beta_1 - \beta_2) \neq 0) > 0$ from the assumption (A3), and therefore the proof is complete. + +![](images/2cddcae0098811cb200f20f96b8d5deaac2d4eabd9dd7173776c90c8a0510fb3.jpg) + +Lemma C.5. If the pricing policy $\pi_l$ for each epoch $l$ is specified as in (5), then the assumption (A4) is satisfied. + +Proof. Let $q_{l}(\cdot \mid x)$ be the conditional probability mass function of $P_{t}$ given $X_{t} = x$ for $t \in \mathcal{E}_l$ . Note that for any $x \in \mathcal{X}$ and $p \in \mathcal{G}$ , we have + +$$ +\begin{array}{l} q _ {l} (p \mid x) = \pi_ {l} (x) (\{p \}) \\ \geq \eta_ {l} / K \\ \gtrsim \eta \cdot n _ {l} ^ {- \gamma} \\ \gtrsim \left\{ \begin{array}{l l} n _ {l} ^ {- \frac {1 + \gamma_ {l}}{2}} (\log n _ {l}) ^ {\frac {1}{2}} & \quad \text {i f} \gamma_ {l} < \frac {1}{3}, \\ n _ {l} ^ {- \gamma_ {l} - \frac {1}{3}} (\log n _ {l}) ^ {\frac {1}{2}} & \quad \text {i f} \gamma_ {l} \geq \frac {1}{3}, \end{array} \right. \\ \end{array} +$$ + +where the first inequality holds by (5), the second inequality holds by (7), and the last inequality holds by (6). Thus, the conditional probability mass function $q_{l}(\cdot \mid x)$ , parameterized by the policy $\pi_{l}$ , satisfies the assumption (A4). + +![](images/c0768cb6a3040618df68bb139fbc35f1793629f75100082ea014907991d04f5e.jpg) + +Lemma C.6. Suppose that the prior distribution $\Pi$ is specified as in (4), and the policy $\pi_l$ for each epoch $l$ is defined by (5). Suppose that assumptions (A1)-(A3), (A5) hold. Then, for every $\epsilon > 0$ , there exist positive constants $C_1$ and $C_2$ depending on $L, B, p_{\min}, p_{\max}, \kappa, \alpha, \rho, a, b, n_1$ and $\epsilon$ , such that for $l \geq C_1$ , + +$$ +\| \widehat {\mathbf {S}} _ {0} ^ {l - 1} - \mathbf {S} _ {0} ^ {*} \| _ {\infty} + \| \widehat {\boldsymbol {\beta}} ^ {l - 1} - \boldsymbol {\beta} ^ {*} \| _ {2} \leq \epsilon +$$ + +with probability at least $1 - \exp (-C_2n_{l - 1}^{1 / 3})$ + +Proof. We define the distance $\mathcal{D}_{\infty}(\theta_1,\theta_2)$ between $\theta_{1} = (\mathbf{S}_{0,1},\beta_{1})$ and $\theta_{2} = (\mathbf{S}_{0,2},\beta_{2})$ on $\Theta$ as + +$$ +\mathcal {D} _ {\infty} \left(\theta_ {1}, \theta_ {2}\right) = \left\| \mathbf {S} _ {0, 1} - \mathbf {S} _ {0, 2} \right\| _ {\infty} + \left\| \beta_ {1} - \beta_ {2} \right\| _ {2}. +$$ + +Let $q_{l}(\cdot \mid x)$ be the conditional probability mass function of $P_{t}$ given $X_{t} = x$ for $t \in \mathcal{E}_l$ in epoch $l$ . By Lemma C.5, $q_{l}(\cdot \mid x)$ satisfies the assumption (A4) for every epoch $l$ . Then, by Lemma A.1, for every $\epsilon > 0$ and $\gamma_{l-1} \in (0,1]$ , there exist positive constants $c_{1}, c_{2}$ and $c_{3}$ depending on $L, B, p_{\min}, p_{\max}, \kappa, \alpha, \rho$ and $\epsilon$ such that for $l \geq \lceil \log_2(c_1 / n_1) \rceil + 1$ , + +$$ +\Pi \left(\mathcal {D} _ {\infty} \left(\theta , \theta^ {*}\right) \geq \epsilon / 2 \mid \mathbf {D} _ {l - 1}\right) < c _ {2} \exp \left(- c _ {3} n _ {l - 1} ^ {\frac {1}{3}}\right) \tag {84} +$$ + +with probability at least $1 - \exp (-c_3n_{l - 1}^{1 / 3})$ + +We partition the parameter space $\widetilde{\Theta}$ into two subsets + +$$ +\widetilde {\Theta} _ {1} = \left\{\theta \in \widetilde {\Theta}: \mathcal {D} _ {\infty} \left(\theta , \theta^ {*}\right) < \epsilon / 2 \right\}, +$$ + +$$ +\widetilde {\Theta} _ {2} = \{\theta \in \widetilde {\Theta}: \mathcal {D} _ {\infty} (\theta , \theta^ {*}) \geq \epsilon / 2 \}. +$$ + +Then, we can decompose $\widehat{\theta}^{l - 1}$ as + +$$ +\begin{array}{l} \widehat {\theta} ^ {l - 1} = \int_ {\widetilde {\Theta}} \theta d \widetilde {\Pi} (\theta \mid \mathbf {D} _ {l - 1}) \\ = \int_ {\widetilde {\Theta} _ {1}} \theta d \widetilde {\Pi} (\boldsymbol {\theta} \mid \mathbf {D} _ {l - 1}) + \int_ {\widetilde {\Theta} _ {2}} \theta d \widetilde {\Pi} (\boldsymbol {\theta} \mid \mathbf {D} _ {l - 1}) \\ = \left(1 - \tau_ {l - 1}\right) \widehat {\theta} _ {1} ^ {l - 1} + \tau_ {l - 1} \widehat {\theta} _ {2} ^ {l - 1}, \tag {85} \\ \end{array} +$$ + +where $\tau_{l-1} = \widetilde{\Pi}(\widetilde{\Theta}_2 \mid \mathbf{D}_{l-1})$ . Here, $\widehat{\theta}_1^{l-1}$ and $\widehat{\theta}_2^{l-1}$ are the mean estimates of the probability measures resulting from the restriction and normalization of the truncated posterior distribution on the sets $\widetilde{\Theta}_1$ and $\widetilde{\Theta}_2$ , respectively. It is easy to check that the function $\theta \mapsto \mathcal{D}_{\infty}(\theta, \theta^*)$ is convex and bounded over the domain $\widetilde{\Theta}$ . By Jensen's inequality, we have + +$$ +\begin{array}{l} \mathcal {D} _ {\infty} \left(\widehat {\theta} _ {1} ^ {l - 1}, \theta^ {*}\right) \leq \int_ {\widetilde {\Theta} _ {1}} \mathcal {D} _ {\infty} \left(\theta , \theta^ {*}\right) d \widetilde {\Pi} _ {1} \left(\theta \mid \mathbf {D} _ {l - 1}\right) \\ < \epsilon / 2, \tag {86} \\ \end{array} +$$ + +where $\widetilde{\Pi}_1(\cdot \mid \mathbf{D}_{l - 1})$ be the probability measure obtained by restricting and renormalizing $\widetilde{\Pi} (\cdot \mid \mathbf{D}_{l - 1})$ to $\widetilde{\Theta}_1$ , and the last inequality holds by the definition of $\widetilde{\Theta}_1$ . On the event that the inequality (84) holds, for $l\geq \lceil \log_2(c_1 / n_1)\rceil +1$ , we have + +$$ +\begin{array}{l} \mathcal {D} _ {\infty} (\widehat {\boldsymbol {\theta}} ^ {l - 1}, \boldsymbol {\theta} ^ {*}) \leq (1 - \tau_ {l - 1}) \mathcal {D} _ {\infty} (\widehat {\boldsymbol {\theta}} _ {1} ^ {l - 1}, \boldsymbol {\theta} ^ {*}) + \tau_ {l - 1} \mathcal {D} _ {\infty} (\widehat {\boldsymbol {\theta}} _ {2} ^ {l - 1}, \boldsymbol {\theta} ^ {*}) \\ < \frac {\epsilon}{2} + \frac {\Pi (\widetilde {\Theta} _ {2} \mid \mathbf {D} _ {l - 1})}{\Pi (\widetilde {\Theta} \mid \mathbf {D} _ {l - 1})} \mathcal {D} _ {\infty} (\widehat {\theta} _ {2} ^ {l - 1}, \theta^ {*}) \\ \leq \frac {\epsilon}{2} + \frac {c _ {2} \exp \left(- c _ {3} n _ {l - 1} ^ {1 / 3}\right)}{1 - c _ {2} \exp \left(- c _ {3} n _ {l - 1} ^ {1 / 3}\right)} \cdot (1 + \sqrt {d} (a \vee b) + B), \\ \end{array} +$$ + +where the first inequality holds because of the convexity of the function $\theta \mapsto \mathcal{D}_{\infty}(\theta, \theta^{*})$ and (85), and the second inequality holds by (86) and the definition of $\tau_{l-1}$ . The last inequality follows from $\Pi(\widetilde{\Theta} \mid \mathbf{D}_{l-1}) \geq 1 - \Pi(\widetilde{\Theta}_{l-1,2} \mid \mathbf{D}_{l-1})$ , combined with inequality (84) and the boundedness of $\mathcal{D}_{\infty}$ over $\widetilde{\Theta}$ under the assumption (A1). Note that the second term on the right of the preceding display is upper bounded by $\epsilon/2$ for $n_{l-1} \geq (\log(c_2(1 + C_1)/C_1)/c_3)^3$ , where $C_1 = \epsilon/(2(1 + \sqrt{d}(a \vee b) + B))$ . Combining this result with the preceding display, we conclude that for $l \geq (\lceil \log_2(c_1/n_1) \rceil + 1) \vee (\lceil \log_2(C_2/n_1) \rceil + 1)$ , + +$$ +\mathcal {D} _ {\infty} (\widehat {\theta} ^ {l - 1}, \theta^ {*}) < \epsilon , +$$ + +with probability at least $1 - \exp(-c_3 n_{l-1}^{1/3})$ , where $C_2 = (\log(c_2(1 + C_1) / C_1) / c_3)^3$ . The proof is then complete. + +Lemma C.7. Suppose that assumptions (A1)-(A3), (A5) and (B1) hold. Let observations $\mathbf{D}_l = \{(X_t, P_t, Y_t)\}_{t \in \mathcal{E}_l}$ be i.i.d. copies of $(X, P_l, Y)$ , where $P_l$ is a random variable distributed from $\mathbb{Q}_l$ as specified in Algorithm 1. We consider the collection of random variables $\{\mathbb{M}(p) : p \in (p_{\min}, p_{\max})\}$ , where $\mathbb{M}(p) := p S_0^*(p)^{\exp(X^\top \beta^*)}$ . Let $P_c$ denote the points of maximum of $\mathbb{M}(p)$ over $(p_{\min}, p_{\max})$ such that + +$$ +P _ {c} \in \underset {p \in \left(p _ {\min }, p _ {\max }\right)} {\operatorname {a r g m a x}} \mathbb {M} (p). \tag {87} +$$ + +Then, $P_{l}$ converges to $P_{c}$ in distribution as $l\to \infty$ + +Proof. We consider the collection $\{\mathbb{M}_k(p):p\in (p_{\min},p_{\max})\}$ of random variables, where $\mathbb{M}_k(p)\coloneqq p\widehat{S}_0^{l - 1}(p)^{\exp (X^\top \widehat{\beta}^{l - 1})}$ . For each $l$ , define the point of maximum of $\mathbb{M}_k(p)$ over $p\in \mathcal{G}$ by + +$$ +\widehat{P}_{l}\in \operatorname *{argmax}_{p\in \mathcal{G}}\mathbb{M}_{k}(p). +$$ + +We first show that $\widehat{P_l}$ converges weakly to $P_{c}$ . To see this, we need to verify the conditions of Theorem 1 of [8]. We say that $\mathcal{G}$ Painlevé-Kuratowski (PK) converges to $(p_{\min}, p_{\max})$ if + +$$ +\{p \in (p _ {\min }, p _ {\max }): \lim _ {n \to \infty} \inf _ {g \in \mathcal {G}} | p - g | = 0 \} = \{p \in (p _ {\min }, p _ {\max }): \lim _ {n \to \infty} \sup _ {g \in \mathcal {G}} \inf _ {| p - g | = 0} | p - g | = 0 \} = (p _ {\min }, p _ {\max }) +$$ + +Let $N$ be the number of epochs for a given horizon $T$ , satisfying $N = \log_2(T / n_1 + 1)$ . As $l \to \infty$ implies $T \to \infty$ , it is easy to see that the grid set $\mathcal{G}$ PK converges to the continuous interval $(p_{\min}, p_{\max})$ . + +We denote the conditional distribution of $P_{l}$ given $X$ by $\mathbb{Q}_l(\cdot \mid X)$ , where $\mathbb{Q}_l(\cdot \mid X) = \pi_l(X)(\cdot)$ as defined in Algorithm 1. By the design of Algorithm 1 and the definition of $\eta_{l}$ , the conditional distribution $\mathbb{Q}_l(\cdot \mid X)$ satisfies the assumption (A4) for all $l$ . Then, by Lemma A.1 and Theorem 6.8 of [16], for $\epsilon > 0$ , there exist positive constants $c_{1}, c_{2}$ and $c_{3}$ such that for sufficiently large $l$ with $n_{l - 1} \geq c_1$ and for any $\gamma$ , we have + +$$ +\left\| \widehat {\mathbf {S}} _ {0} ^ {l - 1} - \mathbf {S} _ {0} ^ {*} \right\| _ {\infty} + \left\| \widehat {\beta} ^ {l - 1} - \beta^ {*} \right\| _ {2} < \epsilon + c _ {2} \exp \left(- c _ {3} n _ {l - 1} ^ {\frac {1}{3}}\right), \tag {88} +$$ + +with probability at least $1 - \exp (-c_3n_{l - 1}^{1 / 3})$ . Note that there exist constants $M_{1}$ and $M_2$ such that $0 < M_1\leq S_0^* (p_{\mathrm{max}}) < S_0^* (p_{\mathrm{min}})\leq M_2 < 1$ by assumption (A5). Let $C_0\coloneqq ((M_1\wedge (1 - M_2)) / 2\wedge B)$ and take $\epsilon < C_0 / 2$ . For large $l$ such that $n_{l - 1}\geq c_1\vee (c_3^{-1}\log (2c_2 / C_0))^3$ , by (88), we have + +$$ +\left\| \widehat {\mathbf {S}} _ {0} ^ {l - 1} - \mathbf {S} _ {0} ^ {*} \right\| _ {\infty} + \left\| \widehat {\boldsymbol {\beta}} ^ {l - 1} - \boldsymbol {\beta} ^ {*} \right\| _ {2} < C _ {0}, +$$ + +with probability at least $1 - \exp (-c_3n_{l - 1}^{1 / 3})$ . This implies that $\widehat{S}_{0,1}^{l - 1} > M_1 / 2 > 0$ , $\widehat{S}_{0,K}^{l - 1} < (1 + M_2) / 2 < 1$ and $\| \widehat{\beta}^{l - 1}\| _2 < 2B$ . Then, by Lemma C.2, for any $p\in (p_{\mathrm{min}},p_{\mathrm{max}})$ , we can decompose as + +$$ +\begin{array}{l} | \mathbb {M} _ {k} (p) - \mathbb {M} (p) | = p | \widehat {S} _ {0} ^ {l - 1} (p) ^ {\exp (X ^ {\top} \widehat {\beta} ^ {l - 1})} - S _ {0} ^ {*} (p) ^ {\exp (X ^ {\top} \beta^ {*})} | \\ \leq C _ {1} \left| \widehat {S} _ {0} ^ {l - 1} (p) - S _ {0} ^ {*} (p) \right| + C _ {2} \| \widehat {\beta} ^ {l - 1} - \beta^ {*} \| _ {2}, \tag {89} \\ \end{array} +$$ + +where $C_1$ and $C_2$ are positive constants depending on $M_1, M_2, L, B$ and $p_{\mathrm{max}}$ , and the inequality holds almost surely. For each $p \in (p_{\mathrm{min}}, p_{\mathrm{max}})$ , there exists $s \in \{2, \dots, K\}$ such that $p \in [g_{s-1}, g_s]$ . Then, we have + +$$ +\begin{array}{l} \widehat {S} _ {0} ^ {l - 1} (p) - S _ {0} ^ {*} (p) \leq \widehat {S} _ {0, s - 1} ^ {l - 1} - S _ {0, s - 1} ^ {*} + S _ {0, s - 1} ^ {*} - S _ {0} ^ {*} (p) \\ \leq \left| \widehat {S} _ {0, s - 1} ^ {l - 1} - S _ {0, s - 1} ^ {*} \right| + L _ {0} \delta , \\ \end{array} +$$ + +where the first inequality holds by the monotonicity of $\widehat{S}_0^{l-1}$ , and the last inequality holds by $L_0$ -Lipschitz continuity of $S_0^*$ under the assumption (A5). Similarly, we have $S_0^*(p) - \widehat{S}_0^{l-1}(p) \leq |\widehat{S}_{0,s}^{l-1} - S_{0,s}^*| + L_0\delta$ . By combining this with (89), + +$$ +\left| \mathbb {M} _ {k} (p) - \mathbb {M} (p) \right| \leq C _ {3} \left(\left\| \widehat {\mathbf {S}} _ {0} ^ {l - 1} - \mathbf {S} _ {0} ^ {*} \right\| _ {\infty} + \left\| \widehat {\boldsymbol {\beta}} ^ {l - 1} - \boldsymbol {\beta} ^ {*} \right\| _ {2}\right) + C _ {1} L _ {0} \delta , +$$ + +where $C_3 = C_1 \vee C_2$ is a positive constant. Combining this with (88), for sufficiently large $l$ and $T$ , we have + +$$ +\left| \mathbb {M} _ {k} (p) - \mathbb {M} (p) \right| \leq \left(C _ {3} + 1\right) \epsilon , +$$ + +with probability at least $1 - \exp (-c_3n_{l - 1}^{1 / 3})$ . Thus, for each $p\in (p_{\mathrm{min}},p_{\mathrm{max}})$ , $\mathbb{M}_k(p)\to \mathbb{M}(p)$ as $l\rightarrow \infty$ in probability, implying convergence in distribution. Since $p$ is arbitrary, $\mathbb{M}_k$ converges weakly $\mathbb{M}$ in $\ell^{\infty}(A)$ for every compact $A\subset (p_{\mathrm{min}},p_{\mathrm{max}})$ , where $\ell^{\infty}(A)$ denote the space of real-valued bounded functions on $A$ . By the assumption (B1) and Theorem 1 of [8], we conclude that $\widehat{P}_l$ converges weekly to $P_{c}$ . + +By the design of policy $\pi_l$ in Algorithm 1, the random variable $P_{l}$ is defined as $P_{l} = R\widehat{P}_{l} + (1 - R)U$ where $R$ is Bernoulli distributed with success probability $1 - \eta_{l}$ . The variable $U$ is uniformly distributed on $\mathcal{G}$ . Let $f:(p_{\mathrm{min}},p_{\mathrm{max}})\to \mathbb{R}$ be any bounded $L_{1}$ -Lipschitz continuous function for some positive constant $L_{1}$ . Then, we have + +$$ +\begin{array}{l} \left| \mathbb {E} \left[ f \left(P _ {l}\right) \right] - \mathbb {E} \left[ f \left(P _ {c}\right) \right] \right| = \left| \mathbb {E} \left[ f \left(P _ {l}\right) \right] - \mathbb {E} \left[ f \left(\widehat {P} _ {l}\right) \right] + \mathbb {E} \left[ f \left(\widehat {P} _ {l}\right) \right] - \mathbb {E} \left[ f \left(P _ {c}\right) \right] \right| \\ \leq \mathbb {E} [ L _ {1} | (R - 1) \widehat {P} _ {l} + (1 - R) U | ] + | \mathbb {E} [ f (\widehat {P} _ {l}) ] - \mathbb {E} [ f (P _ {c}) ] | \\ \leq 2 L _ {1} \eta_ {l} p _ {\max } + | \mathbb {E} [ f (\widehat {P} _ {l}) ] - \mathbb {E} [ f (P _ {c}) ] |, \\ \end{array} +$$ + +where the first inequality holds because $f$ is $L_{1}$ -Lipschitz function, and the last inequality holds because $\widehat{P_l}$ and $U$ are less than $p_{\mathrm{max}}$ almost surely. By the definition of $\eta_l$ and Portmanteau theorem, the right-hand side of the preceding display converges to 0 as $n \to \infty$ . Then, we apply the Portmanteau theorem to conclude that $P_{l}$ converges weekly to $P_{c}$ . + +![](images/a82b9dcfde4a54b25828161106af5548746ae651ed61d74c15a97fc432e44d98.jpg) + +# D Extension to nonuniform grids + +The assumption of equally spaced prices is made solely for technical convenience in developing the theory. However, with some additional technical work, our results can be readily extended to more general discrete price sets. For instance, one may consider a nonuniform grid $\mathcal{G} = \{g_1,\dots ,g_K\}$ satisfying + +$$ +a \delta \leq \left| g _ {k + 1} - g _ {k} \right| \leq b \delta \quad \text {f o r a l l} k = 1, \dots , K, +$$ + +where $\delta \asymp T^{-\gamma}$ and $a, b > 0$ are constants. This more general setting implies that our theoretical findings can be extended to more practical settings. + +Importantly, since the regret rate in our analysis depends on the discrete set only through the sparsity level $\gamma$ , this generalization does not fundamentally change the regret behavior. Therefore, while such an extension increases technical complexity, it does not yield additional theoretical insights. + +# E Discussion on the Cox PH model assumption + +We here provide additional discussion on our choice of the Cox PH model, addressing both its suitability and potential concerns about model misspecification. + +The key distinction between the Cox PH model and standard linear demand models lies in the use of the hazard function, which is a central concept in survival analysis. Unlike linear models, which model the conditional mean of a random variable, the Cox PH model focuses on modeling the hazard rate, a quantity that fully characterizes the survival distribution and is particularly well-suited to censored data settings. A key advantage of the Cox PH model is that it permits separate analysis of $\lambda_0$ and $\beta$ , enabling theoretical development under minimal assumptions on the functional form of $\lambda_0$ . This makes the Cox PH model an appropriate and principled choice for contextual dynamic pricing. + +At the same time, we note that every model carries some risk of misspecification. Models that directly target the mean (e.g., linear or log-linear) are often highly sensitive to tail behavior and thus more vulnerable to misspecification. By contrast, models focusing on the hazard rate, such as the Cox PH model, tend to be more robust in these settings. A fully distribution-free approach might be preferable in order to avoid the risk of model misspecification. However, such an approach does not appear to be suitable in our context, as interval-censored data contains very limited information about the valuation distribution. + +# F Discussion on the variational Bayes estimator + +In our theoretical analysis, the regret bounds are derived under the assumption that the estimator $\hat{\theta}^{l-1}$ corresponds to the posterior mean of the true Bayesian posterior, which contracts to the ground truth at the rates established in Theorems 3.1 and 3.2. In practice, we employ a variational Bayes (VB) approximation to obtain this estimator due to its computational efficiency in high-dimensional and nonparametric settings. The variational family used in our implementation is sufficiently expressive so that the VB posterior mean closely approximates the true posterior mean. As empirically demonstrated in [29], the considered VB approach performs comparably to, or better than, traditional MCMC in terms of estimation accuracy. + +From a theoretical perspective, the regret bound depends directly on the convergence rate of $\hat{\theta}^{l-1}$ . Therefore, if the VB posterior attains the same contraction rate as the true posterior, the regret guarantees remain valid. Recent advances in the theoretical study of VB methods (e.g., [46, 1, 45]) provide sufficient conditions under which the VB posterior achieves the same contraction rate as the full posterior. Although a rigorous regret analysis for VB-based estimation in our specific setting remains open, these results indicate that our regret guarantees continue to hold under appropriate contraction assumptions. + +# G Additional discussion on estimator replacement + +Although we do not provide a formal proof, the Bayes estimator in our proposed algorithm could potentially be replaced by the NPMLE. However, even if so, careful selection of the exploration parameter $\eta_l$ is crucial for designing an optimal pricing policy. As empirically demonstrated in Section 6, our choice of the exploration parameter (6) substantially improves cumulative regret compared to the parameter choice in [7] (denoted as $\alpha_k$ in their notation), which employed the NPMLE. + +# H Details of the experimental setup + +We use a Gamma prior with $\alpha_{1} = \dots = \alpha_{K} = \rho = 10^{-5}$ , For a prior of $\beta$ , we use a multivariate normal distribution, $N(\mathbf{0},\mathbf{I}_d)$ , where $\mathbf{I}_d$ denotes the $d\times d$ identity matrix. The truncated point estimator is computed within a parameter space of $\beta$ truncated to $[-10,10]^d$ . The proposed algorithm involves three hyperparameters: the first-epoch size $n_1$ , the exploration parameters $\eta_{1}$ and $\eta_{2}$ . These are tuned through grid search over the ranges $n_1\in \{64,128,256\}$ , $\eta_{1}\in \{2^{-4 / 3},2^{-3 / 3},2^{-2 / 3},2^{-1 / 3},2^{0}\}$ and $\eta_{2}\in \{2^{-12 / 2},2^{-11 / 2},2^{-10 / 2},2^{-9 / 2}\}$ with an initial period of $T_{0} = 3000$ for each combination. + +For a fair comparison, the hyperparameters of [7] are also tuned using the grid search procedure. We use the CoxCP algorithm with the $\epsilon$ -greedy heuristic, as described in their experiments. This algorithm involves two hyperparameters: the first-epoch size $\tau_{1}$ and the degree of exploration parameter $\tau$ . The hyperparameters are tuned over the ranges $\tau_{1} \in \{64, 128, 256, 512, 1024\}$ and $\tau \in \{2^{-4}, 2^{-3}, 2^{-2}, 2^{-1}, 2^{0}\}$ , following the procedure outlined in their work. + +# I Computational resources used + +All experiments in this paper were conducted using a machine equipped with an Intel(R) Core(TM) i9-10900X CPU. No GPU was used. \ No newline at end of file diff --git a/NeurIPS/2025/A Bayesian Approach to Contextual Dynamic Pricing using the Proportional Hazards Model with Discrete Price Data/images.zip b/NeurIPS/2025/A Bayesian Approach to Contextual Dynamic Pricing using the Proportional Hazards Model with Discrete Price Data/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..80239d68711d6f81bba00b420e8e43e32bce04b2 --- /dev/null +++ b/NeurIPS/2025/A Bayesian Approach to Contextual Dynamic Pricing using the Proportional Hazards Model with Discrete Price Data/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2f274f896155ce2af8ee24f0ed3175dad81bce0bc5f164d45a7966f042f92a26 +size 3411540 diff --git a/NeurIPS/2025/A Bayesian Approach to Contextual Dynamic Pricing using the Proportional Hazards Model with Discrete Price Data/layout.json b/NeurIPS/2025/A Bayesian Approach to Contextual Dynamic Pricing using the Proportional Hazards Model with Discrete Price Data/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..45541e4b855c341909a6e5d8bba7914eabf32707 --- /dev/null +++ b/NeurIPS/2025/A Bayesian Approach to Contextual Dynamic Pricing using the Proportional Hazards Model with Discrete Price Data/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:018d57add16893d56fab671880aa27e9decfa8bf7b7776261bc13401a78c4675 +size 3596869 diff --git a/NeurIPS/2025/A Bayesian Fast-Slow Framework to Mitigate Interference in Non-Stationary Reinforcement Learning/d8827cd3-70a8-4258-8907-e8c82f233487_content_list.json b/NeurIPS/2025/A Bayesian Fast-Slow Framework to Mitigate Interference in Non-Stationary Reinforcement Learning/d8827cd3-70a8-4258-8907-e8c82f233487_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..768e507cd6154dc9c782fee4160c6208853cbb6a --- /dev/null +++ b/NeurIPS/2025/A Bayesian Fast-Slow Framework to Mitigate Interference in Non-Stationary Reinforcement Learning/d8827cd3-70a8-4258-8907-e8c82f233487_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:34a7aaf4d41f80de5da33cacadcc203bf1624cd0f2ddcdadc3bd5c5e5be30e9d +size 147382 diff --git a/NeurIPS/2025/A Bayesian Fast-Slow Framework to Mitigate Interference in Non-Stationary Reinforcement Learning/d8827cd3-70a8-4258-8907-e8c82f233487_model.json b/NeurIPS/2025/A Bayesian Fast-Slow Framework to Mitigate Interference in Non-Stationary Reinforcement Learning/d8827cd3-70a8-4258-8907-e8c82f233487_model.json new file mode 100644 index 0000000000000000000000000000000000000000..d2825a8bd49cdc739b34a75980703d0421ce8411 --- /dev/null +++ b/NeurIPS/2025/A Bayesian Fast-Slow Framework to Mitigate Interference in Non-Stationary Reinforcement Learning/d8827cd3-70a8-4258-8907-e8c82f233487_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e0d3fd9574b45d8c63546ada9835d2f0e7ad529c89bf374e14a8a20e18fea71a +size 186461 diff --git a/NeurIPS/2025/A Bayesian Fast-Slow Framework to Mitigate Interference in Non-Stationary Reinforcement Learning/d8827cd3-70a8-4258-8907-e8c82f233487_origin.pdf b/NeurIPS/2025/A Bayesian Fast-Slow Framework to Mitigate Interference in Non-Stationary Reinforcement Learning/d8827cd3-70a8-4258-8907-e8c82f233487_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..06d5ba50b3381863071e8ab779f2091610b12309 --- /dev/null +++ b/NeurIPS/2025/A Bayesian Fast-Slow Framework to Mitigate Interference in Non-Stationary Reinforcement Learning/d8827cd3-70a8-4258-8907-e8c82f233487_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2f45452dac412f2e7c7febddcdf560034d702d015db5c01e33ac3f5e0d778e76 +size 3177624 diff --git a/NeurIPS/2025/A Bayesian Fast-Slow Framework to Mitigate Interference in Non-Stationary Reinforcement Learning/full.md b/NeurIPS/2025/A Bayesian Fast-Slow Framework to Mitigate Interference in Non-Stationary Reinforcement Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..a6b7c51b234759948bd3e9df0fbbe90fe1557790 --- /dev/null +++ b/NeurIPS/2025/A Bayesian Fast-Slow Framework to Mitigate Interference in Non-Stationary Reinforcement Learning/full.md @@ -0,0 +1,740 @@ +# A Bayesian Fast-Slow Framework to Mitigate Interference in Non-Stationary Reinforcement Learning + +Yihuan Mao + +Institute for Interdisciplinary Information Sciences + +Tsinghua University + +maoyh1024@gmail.com + +Chongjie Zhang + +Department of Computer Science & Engineering + +Washington University in St. Louis + +chongjie@wustl.edu + +# Abstract + +Given the ever-changing nature of the world and its inhabitants, agents must possess the ability to adapt and evolve over time. Recent research in Given the ever-changing nature of the world and its inhabitants, agents must possess the ability to adapt and evolve over time. Recent research in non-stationary MDPs has focused on addressing this challenge, providing algorithms inspired by task inference techniques. However, these methods ignore the detrimental effects of interference, which particularly harm performance in contradictory tasks, leading to low efficiency in some environments. To address this issue, we propose a Bayesian Fast-Slow Framework (BFSF) that tackles both cross-task generalization and resistance to cross-task interference. Our framework consists of two components: a 'fast' policy, learned from recent data, and a 'slow' policy, learned through meta-reinforcement learning (meta-RL) using data from all previous tasks. A Bayesian estimation mechanism determines the current choice of 'fast' or 'slow' policy, balancing exploration and exploitation. Additionally, in the 'fast' policy, we introduce a dual-reset mechanism and a data relabeling technique to further accelerate convergence when encountering new tasks. Experiments demonstrate that our algorithm effectively mitigates interference and outperforms baseline approaches. Code is available at https://github.com/cedesu/BFSF. + +Reinforcement Learning (RL) in non-stationary environments has long attracted significant attention, leading to the emergence of research areas such as continual RL [30, 19] and non-stationary MDPs [7, 24, 31]. These areas approach challenges from different perspectives. For example, catastrophic forgetting, a well-known issue in continual RL, arises due to limited memory storage. In contrast, research on non-stationary MDPs focuses on understanding the underlying dynamics of the environment to facilitate better adaptation across varying contexts. + +One critical challenge in non-stationary environments is interference, where the learning process is negatively impacted by experiences from previous tasks, as illustrated in Figure 1. This interference arises primarily because task boundaries are either unknown or absent in the streaming task setting. Many real-world problems exhibit such non-stationarity and suffer from interference. For instance, a UAV must adapt its behavior under varying weather conditions [27, 26]. Similarly, the evolving regime of the stock market can be viewed as a time-varying environment, where an effective strategy + +![](images/cd12543a085edac6c2bafc1815652114b8d6c85250118b70e2b80e2b06c8df44.jpg) +(a) The diagram of interference in non-stationary MDPs. + +![](images/aff08efb1f630686f04f27a2d94bdf394ec76c3fc51e324f477150480fefbde8.jpg) +(b) The learning curve of how BFSF alleviates the problem of interference. +Figure 1: Figure 1(a) illustrates the interference problem in non-stationary MDPs. The agent learns to perform well on the current task, but the changepoint is unknown. As a result, when the agent begins learning a new task (e.g., task 2), experience from previous tasks (e.g., task 1) can hinder performance. This interference phenomenon also occurs across consecutive tasks. To address this, we propose BFSF, which incorporates a 'fast' policy that learns from recent data to mitigate interference, alongside a 'slow' policy using meta-RL to learn a context-based policy from all previous tasks. Figure 1(b) demonstrates BFSF's ability to resist interference in the Cheetah-Dir task, which involves two contradictory tasks: moving forward and backward. In the second phase, the learning curve shows less disruption compared to the 'slow' policy only. The highest performance throughout the non-stationary MDP process is close to the upper bound, which represents the scenario where the two tasks are trained separately. + +must stay adaptive while leveraging past experience [13, 3]. These real-world scenarios underscore the urgent need for a framework that can operate efficiently in non-stationary MDPs with interference. + +Despite its significant negative impact on performance, the issue of interference has been largely overlooked in the literature. Some existing works [18] address interference from the perspective of representation, while others [20] discuss the inverse interference of current tasks on previously learned tasks, a phenomenon referred to as catastrophic forgetting in continual RL. This lack of attention to cross-task interference is concerning, as it can severely degrade performance in successive tasks. In this work, we specifically analyze the effects of interference and propose effective strategies to mitigate its detrimental impacts. + +To address the interference problem, we introduce the Bayesian Fast-Slow Framework (BFSF). This framework dynamically selects between two learning strategies: a 'fast' policy, which quickly adapts to new tasks using recent data, and a 'slow' policy, which is learned through meta-RL and captures knowledge from historical data. Unlike previous approaches that focus solely on latter, often leading to severe interference in the face of sudden task changes, our framework not only mitigates interference but also preserves the advantages of meta-RL. A Bayesian estimation mechanism is employed in each epoch to decide which policy, fast or slow, is more promising based on recent history. Only the most recent returns are used to update the Bayesian estimates, ensuring that outdated data does not influence the decision. + +We also identify that the 'fast' policy can sometimes underperform. One reason is that neural networks often experience performance degradation when trained on data from different distributions, a common issue in non-stationary environments. To address this, we introduce a dual-reset mechanism that periodically reinitializes one of the dual networks to prevent degradation, while alternating between the networks to ensure stable performance. Another challenge is that learning from scratch typically requires extensive online interaction. To mitigate this, we propose data relabeling, utilizing historical data from previous tasks to enhance learning efficiency and improve performance in few-shot settings when facing new tasks. + +In summary, our contributions are twofold: i) We analyze the interference problem in non-stationary MDPs. ii) We propose the Bayesian Fast-Slow Framework (BFSF), which combines a fast policy, enhanced by a dual-reset mechanism and data relabeling, to efficiently handle recent tasks, and a slow policy for cross-task generalization. Through Bayesian estimation, we effectively address interference and improve overall performance. Experimental results demonstrate BFSF's superiority in resisting interference and outperforming baseline methods across various non-stationary environments. + +# 1 Preliminaries + +Notations and problem definition A Markov Decision Process (MDP) is defined as $M = \langle S, A, P, R \rangle$ , where $S$ and $A$ represent the state and action spaces, respectively. The transition function of the environment is denoted as $P$ , and $R$ represents the reward function. The expected return of a policy is given by $\mathbb{E}[\sum_{t=0}^{\infty} R_t]$ . In non-stationary MDPs, the underlying MDP evolves over time. These changes can occur sequentially, such as $M_1, M_2, \dots$ , or gradually over time. The objective is to maximize the expected return, $\mathbb{E}[\sum_{t=0}^{\infty} R_t]$ with the evolving dynamics of the MDP. + +Context-based policy In a standard MDP, the policy function is defined as $\pi(a|s)$ , which determines the probability of selecting action $a$ given state $s$ . In non-stationary settings, a context-based policy is introduced to adapt to varying environments. This policy is denoted as $\pi(a|s,c)$ , where $c$ is the context, a set of trajectories related to the current environment. A trajectory consists of the sequence of states, actions, and rewards at each timestep, expressed as $s_0, a_0, r_0, s_1, a_1, r_1, s_2, a_2, r_2, \ldots$ . In implementation, the context is collected from recent interactions, reflecting the underlying MDP. + +The technique of learning the context-based policy has been extensively studied in the field of meta-reinforcement learning (meta-RL) [24, 37]. However, meta-RL differs from the non-stationary MDPs setting in this work. In meta-RL, task information is explicitly available, and there is no continuous adaptation process. Context-based meta-RL methods typically map the contextual information, often represented as transition data, into a latent space $\mathcal{Z}$ . By assigning a latent variable $z\in \mathcal{Z}$ to represent the task, this approach effectively frames the problem as a partially observable MDP (POMDP) [16], where $z$ constitutes the unobserved portion of the state. In meta RL, PEARL learns the posterior distribution $q(z|c)$ , which means the posterior latent variable distribution given the context $c$ , and uses posterior sampling to sample $z$ to integrate these latent variables with off-policy RL algorithms. + +Algorithm 1 Bayesian fast-slow framework (BFSF) +1: Input: A 'fast' policy $\pi_{\text{fast}}$ (including the dual policies $\pi_{\text{fast}}^{(1)}, \pi_{\text{fast}}^{(2)}$ ), a 'slow' policy $\pi_{\text{slow}}$ , the number of epochs $E$ , the window of recent data $w$ +2: Initialize return list $\{R_i\}_{i \in 1 \dots E}$ and choice list $\{\text{choice}_j\}_{i \in 1 \dots E}$ +3: for epoch $e = 1 \dots E$ do +4: # Bayesian inference of the expected return +5: $\hat{R}_{\text{fast}} = \text{Posterior}(\{R_i \in [e - w, e] | \text{choice}_i = \text{fast}\})$ +6: $\hat{R}_{\text{slow}} = \text{Posterior}(\{R_i \in [e - w, e] | \text{choice}_i = \text{slow}\})$ +7: # Online interaction +8: if $\hat{R}_{\text{fast}} > \hat{R}_{\text{slow}}$ then +9: $\text{choice}_e := \text{fast}$ +10: Collect data using $\pi_{\text{fast}}$ +11: else +12: $\text{choice}_e := \text{slow}$ +13: Collect data using $\pi_{\text{slow}}$ +14: end if +15: # Training +16: Update $\pi_{\text{fast}}$ by Algorithm 2 +17: Update $\pi_{\text{slow}}$ by the meta-RL algorithm +18: end for + +# 2 Bayesian Fast-Slow Framework (BFSF) + +The Bayesian Fast-Slow Framework (BFSF) is designed to mitigate interference by dynamically deploying either a 'fast' policy, which learns from recent data, or a 'slow' policy, trained using meta-RL principles. The term 'fast' arises from its 'fast-adaptation' ability to learn directly and efficiently from recent data. In contrast, the 'slow' policy enables cross-task understanding and generalization, which may hinder training speed, especially when the number of observed tasks is limited in early phase. The decision is made based on Bayesian estimation of current expected return. + +As illustrated in Algorithm 1, during each epoch, Bayesian inference is applied to estimate the posterior expected return using the recent return history, for both the fast and slow policies. Let $R_{i}$ denote the return obtained in the $i$ -th epoch, and $\text{choice}_i$ indicate whether the 'fast' or 'slow' policy was selected during that epoch. During the online interaction phase, the policy with the higher estimated posterior value is selected, aiming to generate higher-quality experience. At the end of each epoch, both the fast and slow policies are updated using their respective replay buffers. The detailed computation of the Bayesian posterior, described in lines 5 and 6 of Algorithm 1, is elaborated in Section 2.1. In addition to the online interaction and Bayesian estimation, the training process for the 'fast' policy is detailed in Algorithm 2, while the 'slow' policy is trained according to the context-based meta-RL algorithm PEARL [24], which is one of the first context-based methods and serves as the baseline for numerous subsequent works. The visualization of choosing the 'fast' or 'slow' policy is provided in Appendix E. + +# 2.1 Bayesian Inference + +The detailed update rule for Bayesian posterior estimation is outlined below. For simplicity, assume that the recent returns of a given policy, $R_{i_1}, R_{i_2}, \ldots$ , are approximately drawn from a normal distribution $\mathcal{N}(\mu, 1 / \phi)$ , where $\phi$ is a constant. While this assumption is commonly used, other distributional forms could also be considered depending on the context. The prior distribution for the parameter $\mu$ is assumed to follow $\mu \sim \mathcal{N}(\mu_0, 1 / \phi_0)$ . The posterior estimation of $\mu$ then follows the standard derivation below, as detailed in Appendix D. + +$$ +p (\mu | \{R _ {i _ {1}}, R _ {i _ {2}}, \dots \}, \mu_ {0}, \phi_ {0}) \sim \mathcal {N} (\mu_ {1}, 1 / \sigma_ {1} ^ {2}), +$$ + +$$ +\text {w h e r e} \mu_ {1} = \frac {\phi_ {0} \mu_ {0} + n \phi \bar {R}}{\phi_ {0} + n \phi}, \sigma_ {1} ^ {2} = \frac {1}{\phi_ {0} + n \phi}. \tag {1} +$$ + +To better interpret the result of Bayesian inference, note that the posterior mean $\mu_{1}$ can be decomposed: + +$$ +\mu_ {1} = \frac {\phi_ {0} \mu_ {0} + n \phi \bar {R}}{\phi_ {0} + n \phi} = \frac {\phi_ {0}}{\phi_ {0} + n \phi} \mu_ {0} + \frac {n \phi}{\phi_ {0} + n \phi} \bar {R}. \tag {2} +$$ + +It shows that the posterior mean $\mu_{1}$ is a weighted average of the prior mean $\mu_0$ and the sample average $\overline{R}$ . As more samples are collected, the weight shifts toward trusting the sample average $\overline{R}$ . Conversely, when only a few samples are available, the posterior relies more heavily on the prior $\mu_0$ . + +# 2.2 'Fast' Policy Learning + +The 'fast' policy, learned from recent data, is critical for mitigating interference in non-stationary MDPs. However, the standard learning paradigm often encounters challenges under these conditions. To address these issues, we propose specific structural designs that significantly enhance the efficiency and adaptability of the 'fast' policy. + +Algorithm 2 Training process of the 'fast' policy. +1: Input: The 'fast' policy $\pi_{\text{fast}}$ (including the dual policies $\pi_{\text{fast}}^{(1)}, \pi_{\text{fast}}^{(2)}$ ), a contextual dynamics model $M$ , current epoch $e$ , reset frequency $\nu$ +2: Output: The updated $\pi_{\text{fast}}$ +3: # Dual-reset mechanism +4: if $e \bmod \nu = 0$ then +5: $\pi_{\text{fast}}^{(1)}, \pi_{\text{fast}}^{(2)} = \pi_{\text{fast}}^{(2)}$ , Init $(\pi_{\text{fast}}^{(1)})$ +6: end if +7: # Data relabeling +8: Relabel the recent data by $M$ using the recent trajectories as the context. +9: # Training process +10: Train $\pi_{\text{fast}}^{(1)}, \pi_{\text{fast}}^{(2)}$ using the relabeled data + +Dual-reset Mechanism A key challenge in continual learning is the performance degradation of neural networks when trained on successive tasks. One common observation is that the learning curve for the second task often struggles to converge to an optimal point, even when task difficulty is comparable (as detailed in Appendix C.2). This phenomenon is also noted and studied in ITER [14]. + +To address this, we propose the dual-reset mechanism, as outlined in Algorithm 2, which mitigates performance degradation by periodically reinitializing the model. However, to avoid the inferior performance typically observed immediately after reinitialization, we introduce a dual-model system. This ensures that during any interaction phase, even directly after initialization, a fully trained model is always available for deployment. + +Data Relabeling Another challenge lies in the limited recent data available for the 'fast' policy, which is necessary for rapid adaptation in non-stationary environments. This small dataset size may not support learning a robust policy, especially over prolonged training periods, in contrast to the 'slow' policy that can utilize the entire historical dataset. As a result, the 'fast' policy tends to perform significantly worse, as shown in Figure 5. To address this limitation, we incorporate data relabeling, which significantly enhances the amount of usable data, enabling the learning of a stronger policy. + +Specifically, the data relabeling process relies on maintaining a context-based dynamics model. This model takes the context, state, and action as input, and outputs the relabeled next state and reward. Leveraging the context-based property, it becomes possible to relabel historical data from other tasks into the context of the current task. By combining relabeled data with the original dataset, the learning efficiency of the 'fast' policy is substantially improved. Further details can be found in Appendix C.1. + +# 2.3 Theoretical Analysis + +In this section, we provide a theoretical analysis of the sub-optimality bound of the Bayesian Fast-Slow Framework (BFSF) and present Theorem 2.1. + +# Theorem 2.1. + +$$ +\begin{array}{l} \text {S u b o p t i m a l i t y} (B F S F) \\ \leq \left[ \mathcal {D} _ {\ell_ {1}} \left(p _ {M}, p _ {M ^ {\prime}}\right) \left(r _ {\max } + V _ {\max }\right) + r _ {\text {d i f f}} \right] H \tag {3} \\ + \left| U _ {r, r ^ {\prime}} \left(\pi_ {M ^ {\prime}} ^ {*}\right) \right| + \frac {1}{2} V _ {m a x} \mathcal {D} _ {\ell_ {1}} \left(p _ {M ^ {\prime}} (s, a), p _ {M} (s, a)\right), \\ \end{array} +$$ + +where $M, M'$ denote the original and relabeled MDPs, and $p_M, p_{M'}$ are their transition functions. $H$ is the horizon, $r_{\text{max}}, V_{\text{max}}$ are the maximum reward and value, $r_{\text{diff}}$ represents the maximum reward gap between $M, M'$ , $U_{r,r'}(\pi)$ is defined as $\mathbb{E}_{(s,a) \sim \rho_M^\pi}[r'(s,a) - r(s,a)]$ , and $\mathcal{D}_{\ell_1}$ is the $L1$ distance. + +The following provides a proof sketch and interpretation of Theorem 2.1. First, the suboptimality is defined as the minimum between the fast and slow policies, with the Bayesian estimation serving as an unbiased estimate. We focus on the suboptimality of the 'fast' policy, $|\eta_{M}(\pi_{M}^{*}) - \eta_{M}(\pi_{M^{\prime}}^{*})|$ . This suboptimality can then be decomposed into several components as detailed in Appendix A.1. For ease of comprehension, the first term represents the gap in optimal expected return between the relabeled and original MDPs. The second and third terms arise from the performance difference of the same policy under different dynamics. A further analysis of the bound that incorporates optimization error is provided in Appendix A.2. + +# 3 Experiments + +The Bayesian Fast-Slow Framework (BFSF) is designed to address the complexities of non-stationary MDPs, specifically tackling the interference issue while ensuring generalization across experiences from different tasks. In this section, we focus on two main questions: i) How does BFSF overcome interference in non-stationary environments? ii) How does BFSF perform in real-world scenarios? iii) How do the individual components of BFSF contribute to improving performance? + +To answer the first question, we provide experimental comparisons with baselines in Section 3.1, demonstrating the superiority of BFSF in mitigating interference. For the second question, we design an infinite-world simulation to better approximate real-world conditions and evaluate the performance + +![](images/d22171ada80b80b0271bd75027ea11efa93ff0869877015e7e0709d5c8ce6650.jpg) + +![](images/8ccd1d0fe9f7aaa8be3bca3c72322853fe02cccbfa57361761bb41e88b801e23.jpg) + +![](images/62ec2edcaa8f427ea74dcd7cfb2a61f94831ec6573ba2618094250ae21d85550.jpg) + +![](images/f2c15b7d75aafe3ed8c589cec3d5f13f1d2d668682997587723b9917cac7d32c.jpg) +Figure 2: The learning curve of BFSF and other baselines, on the non-stationary MDPs based on 5 MuJoCo locomotion environments and 1 Meta-World environment. For clarity, we only display the curves for BFSF, LILAC, and ITER. Additional curves for CEMRL and CoMPs, along with implementation details, can be found in Appendix B. + +![](images/937f40d0424622fee94aaf7f6e3554e0c0da3917004dc80b8941fda5cf80108a.jpg) + +![](images/ea1f8f78e9f3bcdbe45773113d1887c23be11572bb19cb320ce291583b723775.jpg) + +
CHEETAH-DIRCHEETAH-VELANT-DIRANT-GOAL
BFSF1209.0 ± 24.0-99.1 ± 2.8671.0 ± 37.7-204.9 ± 9.3
LILAC757.7 ± 94.9-154.6 ± 2.5318.5 ± 66.2-426.6 ± 4.6
ITER813.4 ± 19.0-102.0 ± 2.8-156.8 ± 25.7-519.2 ± 50.0
CEMRL852.6 ± 36.1-196.0 ± 5.2289.6 ± 18-547.1 ± 67.6
CoMPS277.3 ± 40.2-117.8 ± 6.1-142.1 ± 6.7-602.0 ± 72.2
WALKERREACHANT-DIR-INFANT-CIR-INF
BFSF530.9 ± 59.81543.4 ± 93.5153.3 ± 4.2165.2 ± 6.8
LILAC448.8 ± 279.11341.0 ± 47.785.5 ± 18.769.5 ± 1.6
ITER65.6 ± 3.81112.1 ± 168.5-338.5 ± 64.1-337.0 ± 69.8
CEMRL577.9 ± 68.1977.5 ± 514.3-11.8 ± 17.428.0 ± 13.7
CoMPS80.0 ± 34.4526.9 ± 380.0-151.0 ± 19.4-121.1 ± 4.8
+ +Table 1: The average return throughout the training process, comparing all the baselines. 'Walker' and 'Reach' are abbreviations for Walker-Rand-Params and Meta-World Reach, respectively. + +of both BFSF and the baselines. Finally, for the third question, we conduct detailed ablation studies on the modules within BFSF, offering evidence of their effectiveness. + +# 3.1 Mains Results + +We evaluate the Bayesian Fast-Slow Framework (BFSF) on five MuJoCo environments and one Meta-World environment. The MuJoCo environments [28] focus on robotic locomotion and are based on the MuJoCo simulator, while Meta-World [35] is a benchmark designed for Multi-Task and meta-RL, specifically with robot manipulation tasks. These environments require adaptation across different reward functions (e.g., walking direction for Cheetah-Dir and Ant-Dir, target velocity for Cheetah-Vel, and goal location for Ant-Goal and Meta-World Reach), or across different dynamics (e.g., environment parameters for Walker-Rand-Params). These meta-RL environments are widely used in the meta-RL literature and are well-suited for non-stationary MDPs as well. We set the switching frequency of the underlying task to 2000 episodes. + +We compare BFSF with four reproduced baselines. LILAC [31] is an algorithm for non-stationary MDPs that uses meta-RL techniques. It learns a latent variable to discriminate between tasks based on experiences. ITER [14] proposes an iterative approach to relearn the neural network, aiming to overcome non-stationarity. CEMRL [5] learns a task encoder from the gradients of a decoder + +and provides the task encoding to downstream RL. CoMPS [4] continuously alternates between two subroutines: learning a new task using RL and performing completely offline meta-learning to prepare for subsequent task learning. + +As shown in Figure 2, BFSF outperforms the baselines in non-stationary environments. The task switching is indicated by the gray dashed lines for clarity. While we conduct experiments with a fixed switching interval, it is important to note that our algorithms are designed for the general setting where the task distribution and switching timing are completely unknown to the agent. For comprehensiveness, we also experiment with an unfixed switching interval in Section 3.3. + +In general, all algorithms show gradual performance improvements within a single task phase but experience a sudden performance drop immediately after task switches. This decay is expected, as the new task is unfamiliar to the agent. However, we observe that the learning curve in the second phase does not increase as quickly as in the first, which we refer to as the interference phenomenon. + +BFSF effectively mitigates interference, leading to better overall performance across continual phases. LILAC, based on meta-RL methods, provides good cross-task generalization. However, it fails to address interference, resulting in slower learning during the second task. ITER's iterative relearning approach is a solid defense against interference, but relying solely on this approach leads to the learning of elementary policies, which hinders further generalization and improvement. For clarity, we only compare the curves of two baselines in the given figure, while a full comparison of average return is provided in Table 1. Full experiment results on the learning curve is provided in Appendix B.2. + +# 3.2 Infinite-World Simulation + +While the main results in Section 3.1 demonstrate the superiority of BFSF in mitigating interference and achieving cross-task generalization, it remains unclear how such methods perform in more realistic scenarios. In the real world, there are typically no explicit task boundaries, no fixed task initializations, and no finite set of predefined tasks. To better approximate these characteristics, we introduce an Infinite-World Simulation in MuJoCo, an environment where the agent operates in a non-episodic, continuous manner without resets. + +Unlike traditional MuJoCo benchmarks, where each episode ends and resets after a fixed number of steps (e.g., every 1000 steps), our infinite-world environment allows the agent to move seamlessly through a boundless plane without ever being reinitialized. This design leads to a non-episodic interaction flow, closely mimicking the persistent nature of real-world settings. To support this, we implement a dynamic terrain loading module that handles environment generation on the fly, avoiding memory overload while preserving the illusion of an endless space. + +We design two signature environments to evaluate BFSF and baselines under this setup. Ant-Dir-Inf is a non-episodic, infinite-world variant of the standard Ant-Direction environment, where the agent is required to walk alternately left and right across the plane; each time a directional goal is reached, the target direction flips. Ant-Goal-Inf is derived from Ant-Goal, where the target moves along a circular trajectory of infinite radius, requiring the agent to constantly adjust and track it over time. + +The corresponding results are reported in Table 1. In addition, Figure 3 visualizes the trajectories of different methods. BFSF consistently follow the evolving goal direction, while baseline methods react more slowly. This highlights BFSF's effectiveness in non-episodic and task-free environments. + +# 3.3 Ablation Studies + +The ablation studies on each module of BFSF are conducted to answer the question: How do the individual components of BFSF contribute to improving performance? The results show that the 'fast' policy, as a whole, alleviates interference and enhances overall performance in non-stationary MDPs. Additionally, the dual-reset mechanism and relabeling further support the 'fast' policy by enabling more effective learning. + +Unfixed Switching Interval We conducted experiments with an unfixed switching interval, as shown in Figure 4(a). The overall performance exhibits a pattern similar to that observed in the main experiments with a fixed switching interval: the interference issue is mitigated by BFSF, and the agent's performance continues to improve as training progresses. + +![](images/ed798cd1efa60b5f4d0d7e12a9aa23f3a1f759e76e427215b36f89a0344fc942.jpg) + +![](images/3ec095b05c637d295a9994bf2afe7cc184ca1182a4ade8a576160258e239c707.jpg) + +![](images/22d604320dd2ec8b592ac075994bb89786a8faf19581f23ce1c47a46b835f3aa.jpg) +Figure 3: Visualized trajectories for Ant-Dir-Inf and Ant-Cir-Inf are shown. In Ant-Dir-Inf (left), only the BFSF algorithm successfully adapts quickly to the alternating goals between left and right directions. In Ant-Cir-Inf (right), only BFSF demonstrates rapid adaptation to the continuously moving goal along a circular path. +(a) BFSF with an unfixed switching interval. +Figure 4: Ablation studies about the unfixed switching interval and other modules. + +![](images/03893f6973f615c496ba5a6757aed5ae2a8e94965148d08773448475e9057151.jpg) +(b) The ablation study of BFSF. + +'Fast' Policy Ablation The presence of the 'fast' policy, which learns from recent data, enables the agent to better adapt to changes in non-stationary MDPs, as shown in Figure 4(b). The curve labeled 'without fast policy' is identical to the baseline LILAC, as introduced in Section 3.1. While LILAC can generalize across tasks, it struggles to efficiently learn the optimal policy in the second task due to the interference problem. In contrast, BFSF addresses this issue, learning the second task at nearly the same speed as the first task, effectively overcoming interference. This trend is also observed in ongoing tasks, where interference does not negatively impact the performance of BFSF. + +Dual-Reset Ablation The dual-reset mechanism ensures the effectiveness of the 'fast' policy. As shown in Figure 4(b), without the dual-reset, the learning curve exhibits lower performance due to a suboptimal 'fast' policy. + +Relabeling Ablation As seen in Figure 4(b), relabeling significantly improves the learning efficiency of the 'fast' policy. With relabeling, the 'fast' policy can continuously improve its performance, even on later tasks. In contrast, BFSF without relabeling, as shown in Figure 5, struggles to achieve similar improvement in later tasks. + +Bayesian Inference Ablation As introduced in Section 2.1, Bayesian inference provides a suitable estimate of the expected return for both the 'fast' and 'slow' policies. Without it, the estimation becomes less effective, and proper hyperparameter tuning may be required. As illustrated in Figure 4(b), the performance and resistance to interference deteriorate in the absence of Bayesian inference. + +# 4 Related works + +Non-Stationary MDPs Research on non-stationary MDPs primarily focuses on the challenge of recognizing potential tasks, as understanding the task transforms the non-stationary MDP into a fixed MDP. LILAC [31] first employs latent variable models to learn environment representations based on + +![](images/073375081b6377a62bad0c11a80ca78eaacfa8626f81214816e0b25355eec92d.jpg) +Figure 5: A comparison illustrating the performance of the 'slow' and 'fast' policies. The main difference is that with relabeling, the performance of the 'fast' policy remains higher throughout the phases, rather than significantly dropping after the initial phases. + +![](images/0a233dd9e135ea61554fb058bb0484abd20f1827a8ad00999e827e998d064cdc.jpg) + +current and past experiences, drawing inspiration from online learning and probabilistic inference. Subsequent works identified shortcomings in LILAC and proposed solutions. For instance, FANS-RL [10] models non-stationarity in terms of individual latent change factors and causal graphs. ITER [14] highlights the impact of non-stationarity on latent representations, a form of interference similar to the one discussed in our work, leading to the proposal of Iterated Relearning (ITER). Additionally, several theoretical works have also focused on non-stationary MDPs [1, 11, 2, 7]. However, none of these prior works simultaneously address both cross-task generalization and the interference problem. + +Besides, continual RL shares similarities with non-stationary MDPs but focuses on different challenges, particularly catastrophic forgetting [25]. Although non-stationary MDPs and continual RL are often treated as distinct problems, their focus differs, primarily due to the assumption that task boundaries are known to agents in continual RL. As a result, research in continual RL focuses on designing submodules within the overall algorithm [30], such as replay buffers [6], network architecture [23], representation [22], or optimization strategies [21], rather than addressing the non-stationarity itself. + +Meta-RL Meta-RL aims to enable agents to adapt more quickly to new tasks by leveraging prior experience from multiple tasks. It bears strong resemblance to non-stationary MDPs, but assumes that agents have full access to all tasks, thus ignoring interference effects. $\mathrm{RL}^2$ [9] learns an agent's learning algorithm, enabling it to adapt quickly to new tasks by adjusting its internal update rule based on prior experience. MAML [12], a gradient-based meta-RL algorithm, seeks a set of model parameters that can be quickly adapted to new tasks with minimal gradient updates. In contrast, context-based meta-RL methods, such as PEARL [24] and VariBAD [37], leverage contextual information to enable more efficient adaptation. While traditional meta-RL assumes access to all tasks during training, recent research has explored meta-learning in the continual task setting [4, 5], which is closely related to our work on non-stationary MDPs. + +Inference-related Works The issue of interference has been widely explored in related areas. In multi-task RL, task interference has been observed, and specialized network architectures have been proposed to mitigate this challenge [17, 8]. However, the source of interference in these works differs from that in non-stationary MDPs, where interference arises from unknown, streaming tasks. In continual RL, which focuses on maintaining high performance across incremental tasks, interference is also recognized and investigated at the representation level [18]. + +# 5 Conclusion + +Non-stationarity poses a significant challenge when deploying RL agents in real-world environments. In addition to cross-task generalization through context-based algorithms, a challenge that has been thoroughly explored in previous works, we have identified that interference can severely hinder performance, especially when tasks conflict with each other. To address this issue, we introduce the Bayesian Fast-Slow Framework (BFSF), which incorporates a 'fast' policy that learns from recent history to prevent interference from previous tasks, and a 'slow' policy that maintains strong cross-task generalization. The use of Bayesian estimation ensures an effective and unbiased selection between the fast and slow policies, enhancing the framework's adaptability and robustness. We also introduce a dual-reset mechanism and data relabeling to further enhance efficiency. Experimental + +results demonstrate BFSF's effectiveness in resisting interference and show that it outperforms baseline methods in various non-stationary environments. + +Although BFSF improves adaptability and efficiency in non-stationary MDPs, the current experiments are limited to a small set of environments. Other types of non-stationarity, such as blurred boundaries or stochastic non-stationary MDPs, have not yet been tested. Additionally, more realistic scenarios are needed to assess its applicability in real-world situations. In future work, we aim to develop more realistic benchmarks that align closely with real-world applications of non-stationary MDPs, while further testing BFSF and exploring new challenges. + +# 6 Acknowledgment + +This work was supported by AF Office of Scientific Research (AFOSR) under grant FA9550-25-1-0318 and with the generous support of the Amazon Research Award program. + +# References + +[1] Alekh Agarwal, Sham Kakade, Mikael Henaff, and Wen Sun. Pc-pg: policy cover directed exploration for provable policy gradient learning. In Proceedings of the 34th International Conference on Neural Information Processing Systems, NIPS '20, Red Hook, NY, USA, 2020. Curran Associates Inc. +[2] Reda Alami, Mohammed Mahfoud, and Eric Moulines. Restarted bayesian online change-point detection for non-stationary markov decision processes. In Sarath Chandar, Razvan Pascanu, Hanie Sedghi, and Doina Precup, editors, Proceedings of The 2nd Conference on Lifelong Learning Agents, volume 232 of Proceedings of Machine Learning Research, pages 715-744. PMLR, 22-25 Aug 2023. +[3] Andrew Ang and Allan Timmermann. Regime changes and financial markets. Annual Review of Financial Economics, 4(Volume 4, 2012):313-337, 2012. +[4] Glen Berseth, Zhiwei Zhang, Grace Zhang, Chelsea Finn, and Sergey Levine. CoMPS: Continual meta policy search. In Deep RL Workshop NeurIPS 2021, 2021. +[5] Zhenshan Bing, David Lerch, Kai Huang, and Alois Knoll. Meta-reinforcement learning in non-stationary and dynamic environments. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(3):3476-3491, 2023. +[6] Lucas Caccia, Rahaf Aljundi, Nader Asadi, Tinne Tuytelaars, Joelle Pineau, and Eugene Belilovsky. New insights on reducing abrupt representation change in online continual learning. ArXiv, abs/2203.03798, 2021. +[7] Wang Chi Cheung, David Simchi-Levi, and Ruihao Zhu. Reinforcement learning for nonstationary Markov decision processes: The blessing of (More) optimism. In Hal Daumé III and Aarti Singh, editors, Proceedings of the 37th International Conference on Machine Learning, volume 119 of Proceedings of Machine Learning Research, pages 1843-1854. PMLR, 13-18 Jul 2020. +[8] Chuntao Ding, Zhichao Lu, Shangguang Wang, Ran Cheng, and Vishnu Naresh Boddeti. Mitigating task interference in multi-task learning via explicit task routing with non-learnable primitives. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7756–7765, June 2023. +[9] Yan Duan, John Schulman, Xi Chen, Peter L. Bartlett, Ilya Sutskever, and P. Abbeel. $\mathsf{Rl}^2$ : Fast reinforcement learning via slow reinforcement learning. ArXiv, abs/1611.02779, 2016. +[10] Fan Feng, Biwei Huang, Kun Zhang, and Sara Magliacane. Factored adaptation for nonstationary reinforcement learning. In S. Koyejo, S. Mohamed, A. Agarwal, D. Belgrave, K. Cho, and A. Oh, editors, Advances in Neural Information Processing Systems, volume 35, pages 31957-31971. Curran Associates, Inc., 2022. + +[11] Songtao Feng, Ming Yin, Ruiquan Huang, Yu-Xiang Wang, Jing Yang, and Yingbin Liang. Non-stationary reinforcement learning under general function approximation. In Proceedings of the 40th International Conference on Machine Learning, ICML'23. JMLR.org, 2023. +[12] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In Proceedings of the 34th International Conference on Machine Learning - Volume 70, ICML'17, pages 1126-1135. JMLR.org, 2017. +[13] Weiyu Guo and Mark E. Wohar. Identifying regime changes in market volatility. Journal of Financial Research, 29(1):79–93, 2006. +[14] Maximilian Igl, Gregory Farquhar, Jelena Luketina, Wendelin Boehmer, and Shimon Whiteson. Transient non-stationarity and generalisation in deep reinforcement learning. In International Conference on Learning Representations, 2021. +[15] Michael Janner, Justin Fu, Marvin Zhang, and Sergey Levine. When to trust your model: model-based policy optimization. Curran Associates Inc., Red Hook, NY, USA, 2019. +[16] Leslie P Kaelbling, Michael L. Littman, and Anthony R. Cassandra. Planning and acting in partially observable stochastic domains. Technical report, USA, 1996. +[17] Menelaos Kanakis, David Bruggemann, Suman Saha, Stamatios Georgoulis, Anton Obukhov, and Luc Van Gool. Reparameterizing convolutions for incremental multi-task learning without task interference. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XX 16, pages 689-707. Springer, 2020. +[18] Samuel Kessler, Jack Parker-Holder, Philip Ball, Stefan Zohren, and Stephen J. Roberts. Same state, different task: Continual reinforcement learning without interference. 36:7143-7151, Jun. 2022. +[19] Khimya Khetarpal, Matthew Riemer, Irina Rish, and Doina Precup. Towards continual reinforcement learning: A review and perspectives. J. Artif. Intell. Res., 75:1401-1476, 2020. +[20] Vincent Liu, Han Wang, Luo Yu Tao, Khurram Javed, Adam White, and Martha White. Measuring and mitigating interference in reinforcement learning. In Sarath Chandar, Razvan Pascanu, Hanie Sedghi, and Doina Precup, editors, Proceedings of The 2nd Conference on Lifelong Learning Agents, volume 232 of Proceedings of Machine Learning Research, pages 781-795. PMLR, 22-25 Aug 2023. +[21] David Lopez-Paz and Marc'Aurelio Ranzato. Gradient episodic memory for continual learning. In Proceedings of the 31st International Conference on Neural Information Processing Systems, NIPS'17, pages 6470-6479, Red Hook, NY, USA, 2017. Curran Associates Inc. +[22] Divyam Madaan, Jaehong Yoon, Yuanchun Li, Yunxin Liu, and Sung Ju Hwang. Representational continuity for unsupervised continual learning. In International Conference on Learning Representations, 2022. +[23] Arun Mallya, Dillon Davis, and Svetlana Lazebnik. Piggyback: Adapting a single network to multiple tasks by learning to mask weights. In Vittorio Ferrari, Martial Hebert, Cristian Sminchisescu, and Yair Weiss, editors, Computer Vision – ECCV 2018, pages 72–88, Cham, 2018. Springer International Publishing. +[24] Kate Rakelly, Aurick Zhou, Chelsea Finn, Sergey Levine, and Deirdre Quillen. Efficient off-policy meta-reinforcement learning via probabilistic context variables. In Kamalika Chaudhuri and Ruslan Salakhutdinov, editors, Proceedings of the 36th International Conference on Machine Learning, volume 97 of Proceedings of Machine Learning Research, pages 5331-5340. PMLR, 09-15 Jun 2019. +[25] ANTHONY ROBINS. Catastrophic forgetting, rehearsal and pseudorehearsal. Connection Science, 7(2):123-146, 1995. +[26] Amila Thibbotuwawa, Grzegorz Bocewicz, Grzegorz Radzki, Peter Nielsen, and Zbigniew Banaszak. Uav mission planning resistant to weather uncertainty. Sensors, 20(2), 2020. + +[27] Amila Thibbotuwawa, Grzegorz Bocewicz, Banaszak Zbigniew, and Peter Nielsen. A solution approach for uav fleet mission planning in changing weather conditions. Applied Sciences, 9(19), 2019. +[28] Emanuel Todorov, Tom Erez, and Yuval Tassa. Mujoco: A physics engine for model-based control. In 2012 IEEE/RSJ International Conference on Intelligent Robots and Systems, pages 5026-5033, 2012. +[29] Michael Wan, Jian Peng, and Tanmay Gangwani. Hindsight foresight relabeling for meta-reinforcement learning. In International Conference on Learning Representations, 2022. +[30] Liyuan Wang, Xingxing Zhang, Hang Su, and Jun Zhu. A comprehensive survey of continual learning: Theory, method and application. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(8):5362-5383, 2024. +[31] Annie Xie, James Harrison, and Chelsea Finn. Deep reinforcement learning amidst continual structured non-stationarity. In Marina Meila and Tong Zhang, editors, Proceedings of the 38th International Conference on Machine Learning, volume 139 of Proceedings of Machine Learning Research, pages 11393-11403. PMLR, 18-24 Jul 2021. +[32] Huazhe Xu, Yuanzhi Li, Yuandong Tian, Trevor Darrell, and Tengyu Ma. Algorithmic framework for model-based reinforcement learning with theoretical guarantees. CoRR, abs/1807.03858, 2018. +[33] Tengye Xu, Zihao Li, and Qinyuan Ren. Meta-reinforcement learning robust to distributional shift via performing lifelong in-context learning. In Proceedings of the 41st International Conference on Machine Learning, ICML'24. JMLR.org, 2024. +[34] Zhuoran Yang, Yuchen Xie, and Zhaoran Wang. A theoretical analysis of deep q-learning, 2020. +[35] Tianhe Yu, Deirdre Quillen, Zhanpeng He, Ryan Julian, Karol Hausman, Chelsea Finn, and Sergey Levine. Meta-world: A benchmark and evaluation for multi-task and meta reinforcement learning. In Conference on Robot Learning (CoRL), 2019. +[36] Tianhe Yu, Garrett Thomas, Lantao Yu, Stefano Ermon, James Zou, Sergey Levine, Chelsea Finn, and Tengyu Ma. Mopo: Model-based offline policy optimization, 2020. +[37] Luisa Zintgraf, Sebastian Schulze, Cong Lu, Leo Feng, Maximilian Igl, Kyriacos Shiarlis, Yarin Gal, Katja Hofmann, and Shimon Whiteson. Varibad: variational bayes-adaptive deep rl via meta-learning. J. Mach. Learn. Res., 22(1), January 2021. + +# NeurIPS Paper Checklist + +# 1. Claims + +Question: Do the main claims made in the abstract and introduction accurately reflect the paper's contributions and scope? + +Answer: [Yes] + +Justification: The introduction mentions all the parts in the text. + +Guidelines: + +- The answer NA means that the abstract and introduction do not include the claims made in the paper. +- The abstract and/or introduction should clearly state the claims made, including the contributions made in the paper and important assumptions and limitations. A No or NA answer to this question will not be perceived well by the reviewers. +- The claims made should match theoretical and experimental results, and reflect how much the results can be expected to generalize to other settings. +- It is fine to include aspirational goals as motivation as long as it is clear that these goals are not attained by the paper. + +# 2. Limitations + +Question: Does the paper discuss the limitations of the work performed by the authors? + +Answer: [Yes] + +Justification: The limitations are included in the conclusion. + +Guidelines: + +- The answer NA means that the paper has no limitation while the answer No means that the paper has limitations, but those are not discussed in the paper. +- The authors are encouraged to create a separate "Limitations" section in their paper. +- The paper should point out any strong assumptions and how robust the results are to violations of these assumptions (e.g., independence assumptions, noiseless settings, model well-specification, asymptotic approximations only holding locally). The authors should reflect on how these assumptions might be violated in practice and what the implications would be. +- The authors should reflect on the scope of the claims made, e.g., if the approach was only tested on a few datasets or with a few runs. In general, empirical results often depend on implicit assumptions, which should be articulated. +- The authors should reflect on the factors that influence the performance of the approach. For example, a facial recognition algorithm may perform poorly when image resolution is low or images are taken in low lighting. Or a speech-to-text system might not be used reliably to provide closed captions for online lectures because it fails to handle technical jargon. +- The authors should discuss the computational efficiency of the proposed algorithms and how they scale with dataset size. +- If applicable, the authors should discuss possible limitations of their approach to address problems of privacy and fairness. +- While the authors might fear that complete honesty about limitations might be used by reviewers as grounds for rejection, a worse outcome might be that reviewers discover limitations that aren't acknowledged in the paper. The authors should use their best judgment and recognize that individual actions in favor of transparency play an important role in developing norms that preserve the integrity of the community. Reviewers will be specifically instructed to not penalize honesty concerning limitations. + +# 3. Theory assumptions and proofs + +Question: For each theoretical result, does the paper provide the full set of assumptions and a complete (and correct) proof? + +Answer: [Yes] + +Justification: The assumptions and proof are in the main text and Appendix A.1. + +# Guidelines: + +- The answer NA means that the paper does not include theoretical results. +- All the theorems, formulas, and proofs in the paper should be numbered and cross-referenced. +- All assumptions should be clearly stated or referenced in the statement of any theorems. +- The proofs can either appear in the main paper or the supplemental material, but if they appear in the supplemental material, the authors are encouraged to provide a short proof sketch to provide intuition. +- Inversely, any informal proof provided in the core of the paper should be complemented by formal proofs provided in appendix or supplemental material. +- Theorems and Lemmas that the proof relies upon should be properly referenced. + +# 4. Experimental result reproducibility + +Question: Does the paper fully disclose all the information needed to reproduce the main experimental results of the paper to the extent that it affects the main claims and/or conclusions of the paper (regardless of whether the code and data are provided or not)? + +Answer: [Yes] + +Justification: The implementation details can be found in the code. + +# Guidelines: + +- The answer NA means that the paper does not include experiments. +- If the paper includes experiments, a No answer to this question will not be perceived well by the reviewers: Making the paper reproducible is important, regardless of whether the code and data are provided or not. +- If the contribution is a dataset and/or model, the authors should describe the steps taken to make their results reproducible or verifiable. +- Depending on the contribution, reproducibility can be accomplished in various ways. For example, if the contribution is a novel architecture, describing the architecture fully might suffice, or if the contribution is a specific model and empirical evaluation, it may be necessary to either make it possible for others to replicate the model with the same dataset, or provide access to the model. In general, releasing code and data is often one good way to accomplish this, but reproducibility can also be provided via detailed instructions for how to replicate the results, access to a hosted model (e.g., in the case of a large language model), releasing of a model checkpoint, or other means that are appropriate to the research performed. +- While NeurIPS does not require releasing code, the conference does require all submissions to provide some reasonable avenue for reproducibility, which may depend on the nature of the contribution. For example +(a) If the contribution is primarily a new algorithm, the paper should make it clear how to reproduce that algorithm. +(b) If the contribution is primarily a new model architecture, the paper should describe the architecture clearly and fully. +(c) If the contribution is a new model (e.g., a large language model), then there should either be a way to access this model for reproducing the results or a way to reproduce the model (e.g., with an open-source dataset or instructions for how to construct the dataset). +(d) We recognize that reproducibility may be tricky in some cases, in which case authors are welcome to describe the particular way they provide for reproducibility. In the case of closed-source models, it may be that access to the model is limited in some way (e.g., to registered users), but it should be possible for other researchers to have some path to reproducing or verifying the results. + +# 5. Open access to data and code + +Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material? + +# Answer: [Yes] + +Justification: The code is in the supplementary materials. + +# Guidelines: + +- The answer NA means that paper does not include experiments requiring code. +- Please see the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details. +- While we encourage the release of code and data, we understand that this might not be possible, so "No" is an acceptable answer. Papers cannot be rejected simply for not including code, unless this is central to the contribution (e.g., for a new open-source benchmark). +- The instructions should contain the exact command and environment needed to run to reproduce the results. See the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details. +- The authors should provide instructions on data access and preparation, including how to access the raw data, preprocessed data, intermediate data, and generated data, etc. +- The authors should provide scripts to reproduce all experimental results for the new proposed method and baselines. If only a subset of experiments are reproducible, they should state which ones are omitted from the script and why. +- At submission time, to preserve anonymity, the authors should release anonymized versions (if applicable). +- Providing as much information as possible in supplemental material (appended to the paper) is recommended, but including URLs to data and code is permitted. + +# 6. Experimental setting/details + +Question: Does the paper specify all the training and test details (e.g., data splits, hyperparameters, how they were chosen, type of optimizer, etc.) necessary to understand the results? + +# Answer: [Yes] + +Justification: The implementation details can be found in the code. + +# Guidelines: + +- The answer NA means that the paper does not include experiments. +- The experimental setting should be presented in the core of the paper to a level of detail that is necessary to appreciate the results and make sense of them. +- The full details can be provided either with the code, in appendix, or as supplemental material. + +# 7. Experiment statistical significance + +Question: Does the paper report error bars suitably and correctly defined or other appropriate information about the statistical significance of the experiments? + +# Answer: [Yes] + +Justification: The error bars are provided in the Table 1. + +# Guidelines: + +- The answer NA means that the paper does not include experiments. +- The authors should answer "Yes" if the results are accompanied by error bars, confidence intervals, or statistical significance tests, at least for the experiments that support the main claims of the paper. +- The factors of variability that the error bars are capturing should be clearly stated (for example, train/test split, initialization, random drawing of some parameter, or overall run with given experimental conditions). +- The method for calculating the error bars should be explained (closed form formula, call to a library function, bootstrap, etc.) +- The assumptions made should be given (e.g., Normally distributed errors). +- It should be clear whether the error bar is the standard deviation or the standard error of the mean. + +- It is OK to report 1-sigma error bars, but one should state it. The authors should preferably report a 2-sigma error bar than state that they have a $96\%$ CI, if the hypothesis of Normality of errors is not verified. +- For asymmetric distributions, the authors should be careful not to show in tables or figures symmetric error bars that would yield results that are out of range (e.g. negative error rates). +- If error bars are reported in tables or plots, The authors should explain in the text how they were calculated and reference the corresponding figures or tables in the text. + +# 8. Experiments compute resources + +Question: For each experiment, does the paper provide sufficient information on the computer resources (type of compute workers, memory, time of execution) needed to reproduce the experiments? + +Answer: [Yes] + +Justification: The implementation details can be found in the code. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The paper should indicate the type of compute workers CPU or GPU, internal cluster, or cloud provider, including relevant memory and storage. +- The paper should provide the amount of compute required for each of the individual experimental runs as well as estimate the total compute. +- The paper should disclose whether the full research project required more compute than the experiments reported in the paper (e.g., preliminary or failed experiments that didn't make it into the paper). + +# 9. Code of ethics + +Question: Does the research conducted in the paper conform, in every respect, with the NeurIPS Code of Ethics https://neurips.cc/public/EthicsGuidelines? + +Answer: [Yes] + +Justification: The research satisfies the NeurIPS Code of Ethics. + +Guidelines: + +The answer NA means that the authors have not reviewed the NeurIPS Code of Ethics. +- If the authors answer No, they should explain the special circumstances that require a deviation from the Code of Ethics. +- The authors should make sure to preserve anonymity (e.g., if there is a special consideration due to laws or regulations in their jurisdiction). + +# 10. Broader impacts + +Question: Does the paper discuss both potential positive societal impacts and negative societal impacts of the work performed? + +Answer: [NA] + +Justification: + +Guidelines: + +- The answer NA means that there is no societal impact of the work performed. +- If the authors answer NA or No, they should explain why their work has no societal impact or why the paper does not address societal impact. +- Examples of negative societal impacts include potential malicious or unintended uses (e.g., disinformation, generating fake profiles, surveillance), fairness considerations (e.g., deployment of technologies that could make decisions that unfairly impact specific groups), privacy considerations, and security considerations. +- The conference expects that many papers will be foundational research and not tied to particular applications, let alone deployments. However, if there is a direct path to any negative applications, the authors should point it out. For example, it is legitimate to point out that an improvement in the quality of generative models could be used to + +generate deepfakes for disinformation. On the other hand, it is not needed to point out that a generic algorithm for optimizing neural networks could enable people to train models that generate Deepfakes faster. + +- The authors should consider possible harms that could arise when the technology is being used as intended and functioning correctly, harms that could arise when the technology is being used as intended but gives incorrect results, and harms following from (intentional or unintentional) misuse of the technology. +- If there are negative societal impacts, the authors could also discuss possible mitigation strategies (e.g., gated release of models, providing defenses in addition to attacks, mechanisms for monitoring misuse, mechanisms to monitor how a system learns from feedback over time, improving the efficiency and accessibility of ML). + +# 11. Safeguards + +Question: Does the paper describe safeguards that have been put in place for responsible release of data or models that have a high risk for misuse (e.g., pretrained language models, image generators, or scraped datasets)? + +Answer: [NA] + +Justification: + +Guidelines: + +- The answer NA means that the paper poses no such risks. +- Released models that have a high risk for misuse or dual-use should be released with necessary safeguards to allow for controlled use of the model, for example by requiring that users adhere to usage guidelines or restrictions to access the model or implementing safety filters. +- Datasets that have been scraped from the Internet could pose safety risks. The authors should describe how they avoided releasing unsafe images. +- We recognize that providing effective safeguards is challenging, and many papers do not require this, but we encourage authors to take this into account and make a best faith effort. + +# 12. Licenses for existing assets + +Question: Are the creators or original owners of assets (e.g., code, data, models), used in the paper, properly credited and are the license and terms of use explicitly mentioned and properly respected? + +Answer: [Yes] + +Justification: The creator of the code is mentioned in the code. + +Guidelines: + +- The answer NA means that the paper does not use existing assets. +- The authors should cite the original paper that produced the code package or dataset. +- The authors should state which version of the asset is used and, if possible, include a URL. +- The name of the license (e.g., CC-BY 4.0) should be included for each asset. +- For scraped data from a particular source (e.g., website), the copyright and terms of service of that source should be provided. +- If assets are released, the license, copyright information, and terms of use in the package should be provided. For popular datasets, paperswithcode.com/datasets has curated licenses for some datasets. Their licensing guide can help determine the license of a dataset. +- For existing datasets that are re-packaged, both the original license and the license of the derived asset (if it has changed) should be provided. +- If this information is not available online, the authors are encouraged to reach out to the asset's creators. + +# 13. New assets + +Question: Are new assets introduced in the paper well documented and is the documentation provided alongside the assets? + +# Answer: [NA] + +Justification: + +Guidelines: + +- The answer NA means that the paper does not release new assets. +- Researchers should communicate the details of the dataset/code/model as part of their submissions via structured templates. This includes details about training, license, limitations, etc. +- The paper should discuss whether and how consent was obtained from people whose asset is used. +- At submission time, remember to anonymize your assets (if applicable). You can either create an anonymized URL or include an anonymized zip file. + +# 14. Crowdsourcing and research with human subjects + +Question: For crowdsourcing experiments and research with human subjects, does the paper include the full text of instructions given to participants and screenshots, if applicable, as well as details about compensation (if any)? + +Answer: [NA] + +Justification: + +Guidelines: + +- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. +- Including this information in the supplemental material is fine, but if the main contribution of the paper involves human subjects, then as much detail as possible should be included in the main paper. +- According to the NeurIPS Code of Ethics, workers involved in data collection, curation, or other labor should be paid at least the minimum wage in the country of the data collector. + +# 15. Institutional review board (IRB) approvals or equivalent for research with human subjects + +Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals (or an equivalent approval/review based on the requirements of your country or institution) were obtained? + +Answer: [NA] + +Justification: + +Guidelines: + +- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. +- Depending on the country in which research is conducted, IRB approval (or equivalent) may be required for any human subjects research. If you obtained IRB approval, you should clearly state this in the paper. +- We recognize that the procedures for this may vary significantly between institutions and locations, and we expect authors to adhere to the NeurIPS Code of Ethics and the guidelines for their institution. +- For initial submissions, do not include any information that would break anonymity (if applicable), such as the institution conducting the review. + +# 16. Declaration of LLM usage + +Question: Does the paper describe the usage of LLMs if it is an important, original, or non-standard component of the core methods in this research? Note that if the LLM is used only for writing, editing, or formatting purposes and does not impact the core methodology, scientific rigorousness, or originality of the research, declaration is not required. + +Answer: [NA] + +# Justification: + +# Guidelines: + +- The answer NA means that the core method development in this research does not involve LLMs as any important, original, or non-standard components. +- Please refer to our LLM policy (https://neurips.cc/Conferences/2025/LLM) for what should or should not be described. + +# Technical Appendices and Supplementary Material + +# A About Theorem 2.1 + +# A.1 Proof of Theorem 2.1 + +Proof. First, since Bayesian estimation provides an unbiased estimate, the suboptimality of BFSF is bounded by the better performance between the fast and slow policies. + +$$ +\text {S u b o p t i m a l i t y} (B F S F) \leq \min (\text {S u b o p t i m a l i t y} (\text {F a s t}), \text {S u b o p t i m a l i t y} (\text {S l o w})) \tag {4} +$$ + +We analyze the suboptimality of the 'fast' policy $|\eta_{M}(\pi_{M}^{*}) - \eta_{M}(\pi_{M^{\prime}}^{*})|$ , in the following proof. + +Let MDP $M$ have the dynamic function $p$ and reward function $r$ . Similarly, let MDP $M'$ have the dynamic function $p'$ and reward function $r'$ . We denote $M_{p,r}$ as the MDP with dynamics function $p$ and reward function $r$ . In Equation 5, we decompose $\left|\eta_M(\pi_M^*) - \eta_M(\pi_{M'}^*)\right|$ and use Theorem A.1, Lemma A.2, A.3 to complete the proof. + +$$ +\begin{array}{l} | \eta_ {M} (\pi_ {M} ^ {*}) - \eta_ {M} (\pi_ {M ^ {\prime}} ^ {*}) | \\ \leq \left| \eta_ {M} \left(\pi_ {M} ^ {*}\right) - \eta_ {M ^ {\prime}} \left(\pi_ {M ^ {\prime}} ^ {*}\right) \right| + \left| \eta_ {M ^ {\prime}} \left(\pi_ {M ^ {\prime}} ^ {*}\right) - \eta_ {M} \left(\pi_ {M ^ {\prime}} ^ {*}\right) \right| \\ = | \eta_ {M _ {p, r}} (\pi_ {M _ {p, r}} ^ {*}) - \eta_ {M _ {p ^ {\prime}, r ^ {\prime}}} (\pi_ {M _ {p ^ {\prime}, r ^ {\prime}}} ^ {*}) | + | \eta_ {M _ {p ^ {\prime}, r ^ {\prime}}} (\pi_ {M _ {p ^ {\prime}, r ^ {\prime}}} ^ {*}) - \eta_ {M _ {p, r}} (\pi_ {M _ {p ^ {\prime}, r ^ {\prime}}} ^ {*}) | \\ \leq \left| \eta_ {M _ {p, r}} \left(\pi_ {M _ {p, r}} ^ {*}\right) - \eta_ {M _ {p ^ {\prime}, r ^ {\prime}}} \left(\pi_ {M _ {p ^ {\prime}, r ^ {\prime}}} ^ {*}\right) \right| \\ + | \eta_ {M _ {p ^ {\prime}}, r ^ {\prime}} (\pi_ {M _ {p ^ {\prime}, r ^ {\prime}}} ^ {*}) - \eta_ {M _ {p, r ^ {\prime}}} (\pi_ {M _ {p ^ {\prime}, r ^ {\prime}}} ^ {*}) | + | \eta_ {M _ {p, r ^ {\prime}}} (\pi_ {M _ {p ^ {\prime}, r ^ {\prime}}} ^ {*}) - \eta_ {M _ {p, r}} (\pi_ {M _ {p ^ {\prime}, r ^ {\prime}}} ^ {*}) | \\ \leq \left[ \mathcal {D} _ {\ell_ {1}} \left(p _ {M _ {1}}, p _ {M _ {2}}\right) \left(r _ {m a x} + V _ {m a x}\right) + r _ {d i f f} \right] H + \left| U _ {r _ {1}, r _ {2}} \left(\pi_ {M _ {2}} ^ {*}\right) \right| + \frac {1}{2} V _ {m a x} \mathcal {D} _ {\ell_ {1}} \left(p _ {M _ {2}} (s, a), p _ {M _ {1}} (s, a)\right) \tag {5} \\ \end{array} +$$ + +![](images/95f3119a859667ab17471b02fe78616ede83f811417e61f6a7db86f98896aea9.jpg) + +Theorem A.1. (Relabeling gap 1) Let $M_1, M_2$ be two finite-horizon MDPs with the same reward function $r$ . Then the distance of $\eta_{M_1}(\pi_{M_1}^*)$ and $\eta_{M_2}(\pi_{M_2}^*)$ is bounded by + +$$ +\left| \eta_ {M _ {1}} \left(\pi_ {M _ {1}} ^ {*}\right) - \eta_ {M _ {2}} \left(\pi_ {M _ {2}} ^ {*}\right) \right| \leq \epsilon_ {h}, \tag {6} +$$ + +where $\epsilon_{h} = [\mathcal{D}_{\ell_{1}}(p_{M_{1}},p_{M_{2}})(r_{max} + V_{max}) + r_{diff}](H - h),\forall h\in [H],s\in S.$ + +Proof. Since $\pi_{M_1}^*$ is the optimal policy of MDP $M_1$ and $\pi_{M_2}^*$ is the optimal policy of MDP $M_2$ , $\eta_{M_1}(\pi_{M_1}^*) = V_{M_1}^*$ , $\eta_{M_2}(\pi_{M_2}^*) = V_{M_2}^*$ . + +We begin our proof from the final horizon, $h = H$ , and use the closeness at horizon $h$ to establish the closeness at horizon $h - 1$ . + +For the final horizon $h = H$ , we have $V_{M,H}^{*}(s) = 0 \forall M$ since it is the terminal state. Therefore, $||V_{M_1,H}^* (s) - V_{M_2,H}^* ||_\infty \leq 0 = \epsilon_H$ . Suppose + +$$ +\forall s \in \mathcal {S} _ {h}, \left\| V _ {M _ {1}, h} ^ {*} (s) - V _ {M _ {2}, h} ^ {*} \right\| _ {\infty} \leq \epsilon_ {h}. \tag {7} +$$ + +We need to prove + +$$ +\forall s \in \mathcal {S} _ {h - 1}, \| V _ {M _ {1}, h - 1} ^ {*} (s) - V _ {M _ {2}, h - 1} ^ {*} \| _ {\infty} \leq \epsilon_ {h - 1}. \tag {8} +$$ + +It is equivalent to prove + +$$ +- \epsilon_ {h - 1} \leq V _ {M _ {2}, h - 1} ^ {*} (s) - V _ {M _ {1}, h - 1} ^ {*} (s) \leq \epsilon_ {h - 1}. \tag {9} +$$ + +For simplicity, but without loss of generality, we will prove the inequality on the right-hand side of the above equation. + +$$ +\begin{array}{l} L H S = \max _ {a \in \mathcal {A}} \left\{\sum_ {s ^ {\prime} \in \mathcal {S} _ {h}} p _ {M _ {2}} \left(s ^ {\prime} \mid s, a\right) \left(r _ {2} (s, a) + \gamma V _ {M _ {2}, h} ^ {*} \left(s ^ {\prime}\right)\right) \right\} \\ - \max _ {a \in \mathcal {A}} \left\{\sum_ {s ^ {\prime} \in \mathcal {S} _ {h}} p _ {M _ {1}} \left(s ^ {\prime} \mid s, a\right) \left(r _ {1} (s, a) + \gamma V _ {M _ {1}, h} ^ {*} \left(s ^ {\prime}\right)\right) \right\} \\ = \max _ {a \in \mathcal {A}} \left\{\sum_ {s ^ {\prime} \in \mathcal {S} _ {h}} \left[ p _ {M _ {1}} \left(s ^ {\prime} \mid s, a\right) \left(r (s, a) + \gamma V _ {M _ {1}, h} ^ {*} \left(s ^ {\prime}\right)\right) \right. \right. \\ + \left(p _ {M _ {2}} \left(s ^ {\prime} \mid s, a\right) r _ {2} (s, a) - p _ {M _ {1}} \left(s ^ {\prime} \mid s, a\right) r _ {1} (s, a)\right) + p _ {M _ {2}} \left(s ^ {\prime} \mid s, a\right) V _ {M _ {2}, h} ^ {*} \left(s ^ {\prime}\right) \\ \left. - p _ {M _ {1}} \left(s ^ {\prime} \mid s, a\right) V _ {M _ {1}, h} ^ {*} \left(s ^ {\prime}\right) \right] \} - \max _ {a \in \mathcal {A}} \left\{\sum_ {s ^ {\prime} \in \mathcal {S} _ {h}} p _ {M _ {1}} \left(s ^ {\prime} \mid s, a\right) \left(r (s, a) + \gamma V _ {M _ {1}, h} ^ {*} \left(s ^ {\prime}\right)\right) \right\} \\ \leq \max _ {a \in \mathcal {A}} \left\{\sum_ {s ^ {\prime} \in \mathcal {S} _ {h}} p _ {M _ {1}} \left(s ^ {\prime} | s, a\right) \left(r (s, a) + \gamma V _ {M _ {1}, h} ^ {*} \left(s ^ {\prime}\right)\right) \right\} \\ + \max _ {a \in \mathcal {A}} \left\{\sum_ {s ^ {\prime} \in \mathcal {S} _ {h}} \left[ \left(p _ {M _ {2}} \left(s ^ {\prime} \mid s, a\right) r _ {2} (s, a) - p _ {M _ {1}} \left(s ^ {\prime} \mid s, a\right) r _ {1} (s, a)\right) \right. \right. \\ \left. + p _ {M _ {2}} \left(s ^ {\prime} \mid s, a\right) V _ {M _ {2}, h} ^ {*} \left(s ^ {\prime}\right) - p _ {M _ {1}} \left(s ^ {\prime} \mid s, a\right) V _ {M _ {1}, h} ^ {*} \left(s ^ {\prime}\right) \rbrack \right\} \\ - \max _ {a \in \mathcal {A}} \left\{\sum_ {s ^ {\prime} \in \mathcal {S} _ {h}} p _ {M _ {1}} \left(s ^ {\prime} \mid s, a\right) \left(r (s, a) + \gamma V _ {M _ {1}, h} ^ {*} \left(s ^ {\prime}\right)\right) \right\} \\ = \max _ {a \in \mathcal {A}} \left\{\sum_ {s ^ {\prime} \in \mathcal {S} _ {h}} \left[ \left(p _ {M _ {2}} \left(s ^ {\prime} \mid s, a\right) r _ {2} (s, a) - p _ {M _ {1}} \left(s ^ {\prime} \mid s, a\right) r _ {1} (s, a)\right) \right. \right. \\ \left. + p _ {M _ {2}} \left(s ^ {\prime} \mid s, a\right) V _ {M _ {2}, h} ^ {*} \left(s ^ {\prime}\right) - p _ {M _ {1}} \left(s ^ {\prime} \mid s, a\right) V _ {M _ {1}, h} ^ {*} \left(s ^ {\prime}\right) \rbrack \right\} \\ = \max _ {a \in \mathcal {A}} \left\{\sum_ {s ^ {\prime} \in \mathcal {S} _ {h}} [ (p _ {M _ {2}} (s ^ {\prime} | s, a) - p _ {M _ {1}} (s ^ {\prime} | s, a)) r _ {2} (s, a) + \sum_ {s ^ {\prime} \in \mathcal {S} _ {h}} p _ {M _ {1}} (s ^ {\prime} | s, a) (r _ {2} (s, a) - r _ {1} (s, a)) \right. \\ \left. + p _ {M _ {2}} \left(s ^ {\prime} \mid s, a\right) V _ {M _ {2}, h} ^ {*} \left(s ^ {\prime}\right) - p _ {M _ {1}} \left(s ^ {\prime} \mid s, a\right) V _ {M _ {1}, h} ^ {*} \left(s ^ {\prime}\right) \rbrack \right\} \\ \leq \max _ {a \in \mathcal {A}} \left\{\sum_ {s ^ {\prime} \in \mathcal {S} _ {h}} | p _ {M _ {2}} (s ^ {\prime} | s, a) - p _ {M _ {1}} (s ^ {\prime} | s, a) | \right\} r _ {m a x} + \max _ {a \in \mathcal {A}} \left\{\sum_ {s ^ {\prime} \in \mathcal {S} _ {h}} p _ {M _ {1}} (s ^ {\prime} | s, a) r _ {d i f f} \right\} \\ + \max _ {a \in \mathcal {A}} \left\{\sum_ {s ^ {\prime} \in \mathcal {S} _ {h}} \left(p _ {M _ {2}} \left(s ^ {\prime} | s, a\right) - p _ {M _ {1}} \left(s ^ {\prime} | s, a\right)\right) V _ {M _ {2}, h} ^ {*} \left(s ^ {\prime}\right) + p _ {M _ {1}} \left(s ^ {\prime} | s, a\right) \left(V _ {M _ {2}, h} ^ {*} \left(s ^ {\prime}\right) - V _ {M _ {1}, h} ^ {*} \left(s ^ {\prime}\right)\right) \right\} \\ \leq \left\{D _ {\ell_ {1}} \left(p _ {M _ {1}} (\cdot | s, a), p _ {M _ {2}} (\cdot | s, a)\right) \right\} \left(r _ {m a x} + V _ {m a x}\right) + r _ {d i f f} + 1 \cdot \left(V _ {M _ {2}, h} ^ {*} (s ^ {\prime}) - V _ {M _ {1}, h} ^ {*} (s ^ {\prime})\right) \\ \leq \left[ D _ {\ell_ {1}} \left(p _ {M _ {1}}, p _ {M _ {2}}\right) \left(r _ {\max } + V _ {\max }\right) + r _ {d i f f} \right] + 1 \cdot \epsilon (h) \\ = \left[ D _ {\ell_ {1}} \left(p _ {M _ {1}}, p _ {M _ {2}}\right) \left(r _ {\max } + V _ {\max }\right) + D _ {\ell_ {1}} \left(p _ {M _ {1}}, p _ {M _ {2}}\right) \left(r _ {\max } + V _ {\max }\right) + r _ {\text {d i f f}} \right] (H - h) \\ = \left[ D _ {\ell_ {1}} \left(p _ {M _ {1}}, p _ {M _ {2}}\right) \left(r _ {\max } + V _ {\max }\right) + r _ {\text {d i f f}} \right] (H - h + 1). \tag {10} \\ \end{array} +$$ + +Lemma A.2. Let $M_1, M_2$ be two MDPs with the same dynamics function, but different reward functions $r_1, r_2$ . Define $U_{r_1,r_2}(\pi) = \mathbb{E}_{(s,a)\sim \rho_{M_1}^{\pi}}[r_2(s,a) - r_1(s,a)]$ , which characterizes how erroneous the model is along trajectories induced by $\pi$ . Then + +$$ +\eta_ {M _ {2}} (\pi) - \eta_ {M _ {1}} (\pi) = U _ {r _ {1}, r _ {2}} (\pi) \tag {11} +$$ + +Proof. We know that $\tilde{M}$ and $\hat{M}$ shares the same transition dynamics $p$ , but different reward functions $\tilde{r}(s, a) = \hat{r}(s, a) - \lambda u(s, a)$ . Therefore, + +$$ +\begin{array}{l} \eta_ {M _ {2}} (\pi) = \mathbb {E} _ {(s, a) \sim \rho_ {M _ {1}} ^ {\pi}} [ r _ {2} (s, a) ] \\ = \mathbb {E} _ {(s, a) \sim \rho_ {M _ {1}} ^ {\pi}} [ r _ {1} (s, a) + (r _ {2} (s, a) - r _ {1} (s, a)) ] \tag {12} \\ = \mathbb {E} _ {(s, a) \sim \rho_ {M _ {1}} ^ {\pi}} r _ {1} (s, a) - \mathbb {E} _ {(s, a) \sim \rho_ {M _ {1}} ^ {\pi}} \left(r _ {2} (s, a) - r _ {1} (s, a)\right) \\ = \eta_ {M _ {1}} (\pi) + U _ {r _ {1}, r _ {2}} (\pi). \\ \end{array} +$$ + +Lemma A.3. (Telescoping lemma) [36, 32]. Let $M_1$ and $M_2$ be two MDPs with the same reward $r(s, a)$ , but different dynamics $p_{M_1}$ and $p_{M_2}$ respectively. Let + +$$ +\begin{array}{r l} & G _ {M _ {2}} ^ {\pi} (s, a) := \\ & \quad \mathbb {E} _ {s ^ {\prime} \sim p _ {M _ {2}} (s, a)} \left[ V _ {M _ {1}} ^ {\pi} \left(s ^ {\prime}\right) \right] - \mathbb {E} _ {s ^ {\prime} \sim p _ {M _ {1}} (s, a)} \left[ V _ {M _ {1}} ^ {\pi} \left(s ^ {\prime}\right) \right], \end{array} \tag {13} +$$ + +Then, + +$$ +\eta_ {M _ {2}} (\pi) - \eta_ {M _ {1}} (\pi) = \gamma \mathbb {E} _ {(s, a) \sim \rho_ {M _ {2}} ^ {\pi}} [ G _ {M _ {2}} ^ {\pi} (s, a) ]. \tag {14} +$$ + +For each $s \in \mathcal{S}, a \in \mathcal{A}, a \ell_1$ -based bound of $|G_{M_2}^{\pi}(s, a)|$ is + +$$ +\left| G _ {M _ {2}} ^ {\pi} (s, a) \right| \leq \frac {1}{2} V _ {\max } \delta_ {\ell_ {1}} \left(p _ {M _ {2}} (s, a), p _ {M _ {1}} (s, a)\right). \tag {15} +$$ + +# A.2 Bound Considering the Optimization Error + +Theorem 2.1 accounts for the error introduced by the relabeling process. To maintain consistency with prior work, we explicitly incorporate optimization suboptimality, following the approach in [34], by considering the policy obtained after $K$ policy-update iterations $\pi_{M'}^{K}$ , rather than the idealized optimal policy $\pi_{M'}^*$ . + +Total Suboptimality + +$$ +\begin{array}{l} \leq \left| \eta_ {M} \left(\pi_ {M} ^ {*}\right) - \eta_ {M} \left(\pi_ {M ^ {\prime}} ^ {K}\right) \right| \\ \leq \left| \eta_ {M} \left(\pi_ {M} ^ {*}\right) - \eta_ {M ^ {\prime}} \left(\pi_ {M ^ {\prime}} ^ {*}\right) \right| + \left| \eta_ {M ^ {\prime}} \left(\pi_ {M ^ {\prime}} ^ {*}\right) - \eta_ {M ^ {\prime}} \left(\pi_ {M ^ {\prime}} ^ {K}\right) \right| + \left| \eta_ {M ^ {\prime}} \left(\pi_ {M ^ {\prime}} ^ {K}\right) - \eta_ {M} \left(\pi_ {M ^ {\prime}} ^ {K}\right) \right| \\ \leq \left[ D _ {1} \left(p _ {M}, p _ {M ^ {\prime}}\right) \left(r _ {\max } + V _ {\max }\right) + r _ {\text {d i f f}} \right] H + \left| U _ {r, r ^ {\prime}} \left(\pi_ {M ^ {\prime}} ^ {*}\right) \right| + \frac {1}{2} V _ {\max } D _ {1} \left(p _ {M ^ {\prime}} (s, a), p _ {M} (s, a)\right) \\ + C \cdot \frac {\phi_ {\mu , \sigma} \cdot \gamma}{(1 - \gamma) ^ {2}} \cdot | A | \cdot (\log n) ^ {1 + 2 \xi^ {*}} \cdot n ^ {(\alpha^ {*} - 1) / 2} + \frac {4 \gamma^ {K + 1}}{(1 - \gamma) ^ {2}} \cdot R _ {\max }, \\ \end{array} +$$ + +where each term explicitly captures different sources of error: + +- The first 3 terms (matching our original Theorem 3.1) represent the suboptimality caused by the relabeling process, quantifying the error introduced due to differences in dynamics and reward functions between the original MDP $M$ and the relabeled MDP $M'$ . +- The newly introduced 4th and 5th terms explicitly quantify the optimization suboptimality, representing errors arising from finite-sample approximations and iterative optimization. + +# B Experiments + +# B.1 Implementation Details + +The experiments are repeated three times, with the mean and standard deviation shown in the curves and table. + +The common hyperparameters are consistent with the original PEARL implementation. Additionally, the context consists of 200 episodes, and the relabeling percentage is set to $50\%$ (i.e., half of the used batch is relabeled). The window of recent data $w = 100$ episodes. The reset frequency $\nu = 50$ episodes. The discount factor $\gamma = 0.99$ . + +# B.2 Full Learning Curves + +A complete comparison of the learning curves for all four baselines is provided in Figure 6. + +# B.3 Sensitive Studies and Other Experiments + +Sensitive study on the relabeling percentage We performed an sensitive study on the relabeling percentage in the Table 2. + +![](images/8e216b5b0b6769c2616e7cef7197739816ed3487135aa9dd6b7afa264d9338fe.jpg) +Figure 6: The learning curve of BFSF with all the baselines including LILAC, ITER, CEMRL and CoMPs. + +![](images/6510ef206dd4b2d3959ac42be78d9433ba562fd80d08cac53b1d49a8f774c10a.jpg) + +![](images/a0a31b8685f225d127314e8374983782f2cb6a024f3c20aa33507536075fe7d0.jpg) + +
Relabel Percentage (%)01020304050
Average Return1055.5954.91046.31097.51186.41039.4
Relabel Percentage (%)60708090100
Average Return1217.71200.91148.11027.61129.1
+ +Sensitive study on the window of recent data $w$ and the reset frequency $\nu$ is shown in Table 3. For reference, the average return of the baseline (LILAC) is 757.8. The results indicate that performance of our method is relatively insensitive to the choice of window size $w$ and reset frequency $\nu$ . + +The experiment in gradually changing environments To validate our method in dynamically changing environments, we conducted supplementary experiments in Table 4 in a 180-task gradually changing Ant-Goal environment, with goals evenly distributed around a circle. + +The experiment in stochastically changing environments As to settings with highly stochastic task boundaries, we conduct experiments on a Cheetah-Dir environment in Table 5. This environment consists of two tasks: moving forward and moving backward. A boundary period (one-third of the task length) exists during task switching, with each task having a probability of $50\%$ . + +# C Algorithm Details + +# C.1 Data Relabeling + +To elaborate, relabeling is accomplished by a learned dynamics and reward model $s', r = f(s, a)$ [15], which estimates the next state $s'$ and reward $r$ after taking action $a$ in the state $s$ . We can + +Table 2: Sensitive study on the relabeling percentage + +
w=10w=50w=100w=150w=200
BFSF1163.31273.61209.01121.1945.5
ν=10ν=25ν=50ν=75ν=100
BFSF1250.61169.01209.01132.11116.6
+ +Table 3: Sensitive study on the window of recent data $w$ and the reset frequency $\nu$ . + +
BFSFBaseline: LILAC
Gradual Ant-Goal-328.5 ± 2.1-619.2 ± 3.7
+ +Table 4: The experiment results in a gradual Ant-Goal environment. BFSF outperforms the baseline, showing its ability to adapt continuously to a changing environment. + +
BFSFBaseline: LILAC
Stochastic Ant-Goal887.8 ± 21.6615.4 ± 5.7
+ +substitute the original next state and reward in the experience replay, even from different tasks, with those predicted by the model, represented as $s_{relabel}'$ , $r_{relabel} = f(s, a)$ . + +Relabeling is widely used in the RL community [29, 33]. The underlying principle is to maximize data reuse for sample efficiency. In our work, given that the environment evolves over time, we leverage a context-based dynamics model $f(s,a,c)$ , which provides different dynamics depending on context $c$ . + +As discussed in [29, 33] and confirmed by our ablation studies, data relabeling substantially enhances sample efficiency. In our work, we adopt data relabeling to mitigate the issue of performance degradation in a non-stationary environment, because the increased amount of relabeled data allow the 'fast' policy to better and faster adapt to these tasks, which is verified by our ablation results. + +# C.2 Motivation for the Dual-Reset Mechanism + +The dual-reset mechanism was introduced upon observing that RL algorithms tend to encounter performance degradation when alternating between tasks, a trend corroborated by prior research [4]. This mechanism effectively addresses the issue. We present the phase performance for SAC in the alternating Cheetah-Dir environment (where the two tasks are moving forward and backward) in Table 6. + +# C.3 Bayesian Fast-Slow Framework + +We only utilize data within a recent window, which keeps the effective sample size small, and we set the prior value to a dynamically updated upper bound to encourage exploration. As a result, if a policy has not been sufficiently selected, the prior strongly influences the Bayesian estimate, leading to a large posterior value that naturally encourages exploration of that policy. + +Table 5: The experiment results in a stochastic Ant-Goal environment. BFSF significantly outperforms baseline LILAC in such an environment. + +
Phase performanceTask 1Task 2Task 1Task 2Task 1
Without dual-reset954.6117.892.08.6-27.4
With dual-reset759.7625.6697.5409.8782.7
+ +Table 6: In the first phase (Task 1), SAC without the dual-reset mechanism performs well, even outpacing the SAC with dual-reset. However, during the alternating tasks, the performance degrades significantly. + +# D Posterior Calculation of Normal Distributions + +Assume $\phi$ is known. + +$$ +\begin{array}{l} p (\mu | \{R _ {i _ {1}}, R _ {i _ {2}}, \dots \}, 1 / \phi) \propto p (\mu) p (\{R _ {i _ {1}}, R _ {i _ {2}}, \dots \} | \mu , 1 / \phi) \\ \propto \exp \left\{- \frac {\phi_ {0}}{2} (\mu - \mu_ {0}) ^ {2} \right\} \times \exp \left\{- \frac {n \phi}{2} (\mu - \overline {{y}}) ^ {2} \right\} \\ \propto \exp \left\{- \frac {1}{2} (\phi_ {0} + n \phi) \mu^ {2} + \frac {1}{2} (2 \mu_ {0} \phi_ {0} + 2 n \phi \overline {{y}}) \mu \right\} \\ \propto \exp \left\{- \frac {1}{2} \left(\phi_ {0} + n \phi\right) \left(\mu - \frac {\phi_ {0} \mu_ {0} + n \phi \bar {y}}{\phi_ {0} + n \phi}\right) ^ {2} \right\} \tag {17} \\ \sim \operatorname {N o r m a l} (\mu_ {1}, \sigma_ {1} ^ {2}), \\ \end{array} +$$ + +$$ +\mathrm {w h e r e} \mu_ {1} = \frac {\phi_ {0} \mu_ {0} + n \phi \overline {{R}}}{\phi_ {0} + n \phi}, \sigma_ {1} ^ {2} = \frac {1}{\phi_ {0} + n \phi}. +$$ + +# E Visualization of Choosing Slow/Fast Policies + +We presented visualization in left sub-figure of Figure 5 in our paper (attached here in Figure 7). In the left sub-figure, the 'fast' policy predominates in the selection during the initial phases, while the 'slow' policy shows competitive performance as the amount of accumulated data from different tasks increases. Additionally, in the latter part of a single phase, the 'fast' policy surpasses the 'slow' policy after learning from relabeled recent data. + +![](images/4841385c628651e66f3efc1d21f2928b63cbc31692e96612d37a11b3fa87647f.jpg) +Figure 7: A comparison illustrating the performance of the 'slow' and 'fast' policies. The main difference is that with relabeling, the performance of the 'fast' policy remains higher throughout the phases, rather than significantly dropping after the initial phases. + +![](images/875084a46c8e12fcc0de075af93fcbf32dc612b6c4493d16a92e9f1b493d3cd1.jpg) \ No newline at end of file diff --git a/NeurIPS/2025/A Bayesian Fast-Slow Framework to Mitigate Interference in Non-Stationary Reinforcement Learning/images.zip b/NeurIPS/2025/A Bayesian Fast-Slow Framework to Mitigate Interference in Non-Stationary Reinforcement Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..c381373aad5dee9bbb764cd50dc777aeaaaccfb2 --- /dev/null +++ b/NeurIPS/2025/A Bayesian Fast-Slow Framework to Mitigate Interference in Non-Stationary Reinforcement Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c51cfad2656506f153be83e3a914c9f3e5b468018f5afd839e1ae9db7592bbdb +size 902521 diff --git a/NeurIPS/2025/A Bayesian Fast-Slow Framework to Mitigate Interference in Non-Stationary Reinforcement Learning/layout.json b/NeurIPS/2025/A Bayesian Fast-Slow Framework to Mitigate Interference in Non-Stationary Reinforcement Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..647b8c61dd1b7a48eb1b76765cb3e36aca837afc --- /dev/null +++ b/NeurIPS/2025/A Bayesian Fast-Slow Framework to Mitigate Interference in Non-Stationary Reinforcement Learning/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:07fe24c3fc2ae94d8777941f838d3bacdfdb87c8c2f3ec5f844067b5aa2465f2 +size 766217 diff --git a/NeurIPS/2025/A Beyond-Worst-Case Analysis of Greedy k-means++/068dae68-5bc2-4a14-85da-60460fc405fb_content_list.json b/NeurIPS/2025/A Beyond-Worst-Case Analysis of Greedy k-means++/068dae68-5bc2-4a14-85da-60460fc405fb_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..0ba83a40f28cb099a680819d36081cad1ee86e1b --- /dev/null +++ b/NeurIPS/2025/A Beyond-Worst-Case Analysis of Greedy k-means++/068dae68-5bc2-4a14-85da-60460fc405fb_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b7b46dab15e34d1f6e0a4c3af903e4fe875f99771c44f93e56af312cf414bddb +size 190479 diff --git a/NeurIPS/2025/A Beyond-Worst-Case Analysis of Greedy k-means++/068dae68-5bc2-4a14-85da-60460fc405fb_model.json b/NeurIPS/2025/A Beyond-Worst-Case Analysis of Greedy k-means++/068dae68-5bc2-4a14-85da-60460fc405fb_model.json new file mode 100644 index 0000000000000000000000000000000000000000..754169793813c6c99bb736d30f4d0543914483ca --- /dev/null +++ b/NeurIPS/2025/A Beyond-Worst-Case Analysis of Greedy k-means++/068dae68-5bc2-4a14-85da-60460fc405fb_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d4696f44f2aa1533af4ae7675049367c0a7663b639161d5b344eb0bad5c1e9da +size 239037 diff --git a/NeurIPS/2025/A Beyond-Worst-Case Analysis of Greedy k-means++/068dae68-5bc2-4a14-85da-60460fc405fb_origin.pdf b/NeurIPS/2025/A Beyond-Worst-Case Analysis of Greedy k-means++/068dae68-5bc2-4a14-85da-60460fc405fb_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..f25d2136715d89304538081531b4a1155fe2377b --- /dev/null +++ b/NeurIPS/2025/A Beyond-Worst-Case Analysis of Greedy k-means++/068dae68-5bc2-4a14-85da-60460fc405fb_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:847c4566796af14433a594656687573126496390e2eadac16bb26eb57f3f5c17 +size 767749 diff --git a/NeurIPS/2025/A Beyond-Worst-Case Analysis of Greedy k-means++/full.md b/NeurIPS/2025/A Beyond-Worst-Case Analysis of Greedy k-means++/full.md new file mode 100644 index 0000000000000000000000000000000000000000..742aae4dce5382e94fff22b6dbd4dc7692ffb171 --- /dev/null +++ b/NeurIPS/2025/A Beyond-Worst-Case Analysis of Greedy k-means++/full.md @@ -0,0 +1,919 @@ +# A Beyond-Worst-Case Analysis of Greedy k-means++* + +Qingyun Chen + +UC Santa Cruz + +qchen161@ucsc.edu + +Sungjin Im + +UC Santa Cruz + +sim9@ucsc.edu + +Ryan Milstrey + +UC Merced + +rmilstrey@ucmerced.edu + +Benjamin Moseley + +Carnegie Mellon University + +moseleyb@andrew.cmu.edu + +Chenyang Xu + +East China Normal University + +cyxu@sei.ecnu.edu.cn + +Ruilong Zhang + +Technical University of Munich + +ruilong.zhang@tum.de + +# Abstract + +$k$ -means++ and the related greedy $k$ -means++ algorithm are celebrated algorithms that efficiently compute seeds for Lloyd's algorithm. Greedy $k$ -means++ is a generalization of $k$ -means++ where, in each iteration, a new seed is greedily chosen among multiple $\ell \geq 2$ points sampled, as opposed to a single seed being sampled in $k$ -means++. While empirical studies consistently show the superior performance of greedy $k$ -means++, making it a preferred method in practice, a discrepancy exists between theory and practice. No theoretical justification currently explains this improved performance. Indeed, the prevailing theory suggests that greedy $k$ -means++ exhibits worse performance than $k$ -means++ in worst-case scenarios. + +This paper presents an analysis demonstrating the outperformance of the greedy algorithm compared to $k$ -means++ for a natural class of well-separated instances with exponentially decaying distributions, such as Gaussian, specifically when $\ell = \ln k + \Theta(1)$ , a common parameter setting in practical applications. + +# 1 Introduction + +Clustering of $k$ means is the most widely used method for data analytics. In the problem, given a set of points in a high-dimensional space along with a parameter $k > 0$ as input, we are asked to find a set of $k$ centers in the space that minimizes the total squared distance of the points to the center set. This problem is known to be NP-hard [Aloise et al., 2009, Mahajan et al., 2012] and does not admit an approximation arbitrarily close the optimum unless $\mathrm{P} = \mathrm{NP}$ [Awasthi et al., 2015, Lee et al., 2017]. + +While several constant-factor approximation algorithms have been found [Kanungo et al., 2002, Ahmadian et al., 2019, Grandoni et al., 2022, Cohen-Addad et al., 2022], Lloyd's heuristic [Lloyd, 1982] remains the most popular in practice due to its fast running time, simplicity, and easy adaptation to parallel and distributed settings. Lloyd's heuristic starts with $k$ initial seeds and iteratively updates the $k$ clusters by alternating between assigning points to their closest centroid and recomputing the centroid of each cluster. + +It is well known that the quality of Lloyd's $k$ -means clustering critically depends on the initial centroids (seeds). In fact, poorly chosen seeds can lead to significantly worse clustering results, even an unbounded cost compared to the optimum even for fixed $k$ and $n$ values [Arthur et al., 2007]. + +To address this issue, [Arthur et al., 2007] proposed a very simple seeding method called $k$ -means++. It begins by choosing the first seed uniformly at random from the data points and subsequently samples a seed with probability proportional to the squared distance of that point to the already chosen seeds. They proved that $k$ -means++ achieves an $O(\ln k)$ -approximation in expectation. $k$ -means++ has been commonly used in conjunction with Lloyd's heuristic because the seeding is fast and it can make the solution returned by Lloyd's significantly better. + +Interestingly, what is more commonly implemented is a variant of $k$ -means++, which is called greedy $k$ -means++. The greedy variant of the algorithm differs in that in each iteration, it first samples $\ell \geq 2$ candidates and chooses the candidate from them as a seed that decreases the cost function the most. This greedy algorithm is also discussed in the paper that introduced $k$ -means++ [Arthur et al., 2007] and is implemented in popular libraries such as Scikit-learn library [Pedregosa et al., 2011] because of its superior experimental performance over $k$ -means++. + +The expectation had been that the greedy variant would yield a seeding at least as good as, if not better than, $k$ -means++ theoretically in the worst case. Surprisingly, the greedy $k$ -means++ was recently shown to be $\Omega(\ell \log k)$ -approximate [Bhattacharya et al., 2020]. That is, strictly worse than the standard $k$ -means++ for certain instances, which contradicts empirical findings. + +The lower bound was further improved to be $\Omega (\ell^3\log^3 k / \log^2 (\ell \log k))$ [Grunau et al., 2023]. The paper additionally showed that it is $O(\ell^3\log^3 k)$ -approximate on the positive side. These results do not explain why greedy outperforms k-means++ in practice. The gap between theory and practice motivates this paper. + +# 1.1 Our Results and Contributions + +In this paper, we present the first beyond-worst-case analysis of Greedy $k$ -means++ for a natural class of instances to bridge the gap between theory and practice. Our analysis demonstrates that the greedy $k$ -means++ algorithm indeed outperforms $k$ -means++ for such instances. + +Theoretical challenge. In the line of research focused on beyond-worst-case analysis of the $k$ -means problem, a well-explored class of instances is well-separable inputs, where an optimal $k$ -means clustering partitions $X$ into $k$ clusters and the distance between any two cluster centroids is sufficiently large compared to the variance of each cluster. Well-separable instances were widely considered in the literature [Jaiswal and Garg, 2012, Ackermann and Blömer, 2010, Braverman et al., 2011, Shechner et al., 2020]. However, even with such a simple point set, it has been theoretically proven that greedy $k$ -means++ performs worse than $k$ -means++ in [Bhattacharya et al., 2020]. + +Main theoretical result. We demonstrate that Greedy $k$ -means++ outperforms $k$ -means++ for a natural class of instances, which we call regular $^2$ , with the following properties: + +1. Points of each cluster follow an exponentially decaying distribution, such as Gaussian [Bishop, 2007, Aggarwal and Reddy, 2014]. Further, each cluster is assumed to be symmetric since $k$ -means clustering is not well suited for highly asymmetric clusters. +2. Clusters have a similar number of points within a constant factor. This is justified by the fact that while $k$ -means doesn't explicitly assume approximately equal-sized clusters, it is known to struggle with clusters of significantly different size [Bishop, 2013]. +3. The distances between clusters are within a constant factor of $k^{\theta}$ where $\theta \in (0,1/2]$ , assuming that each cluster's radius (the average distance of points to the center) is $\Theta(1)$ . This property is also obtained naturally: the distance between $k$ centers sampled uniformly at random from $[0,1]^k$ is $\Theta(\sqrt{k})$ with a high probability. + +In essence, the above assumptions characterize the instances where the $k$ -means clustering method can be ideally used. For these $k$ -means clustering-friendly instances, we show that Greedy $k$ -means++ with $\ell = \ln k + \Theta(1)$ is $O((\ln \ln k)^2)$ -approximate for the above regular instances, as opposed to $k$ -means++ that is shown to be $\Omega(\ln k)$ -approximate. Provided that $\ell = \Theta(\ln k)$ is a common choice in practice—for example, the Scikit-learn library [Pedregosa et al., 2011] sets $\ell = \lceil \ln k \rceil + 2$ , our result implies that the commonly used Greedy has a better theoretical guarantee over $k$ -means++, even asymptotically. + +Analysis overview. The high-level analysis idea is as follows. Let us say that a cluster is covered if a point within the cluster has been chosen as a seed. Since clusters are sufficiently far from each other and have similar sizes, it is advantageous to cover all clusters. + +For simplicity, suppose that we are at the beginning of the $\kappa$ -th iteration and have successfully covered the $\kappa - 1$ clusters without wasting any previous seeds. Further, assume that the chosen seeds are close to the cluster centers, which is ensured by the fact that clusters have exponentially decaying tails in distance, and therefore most sampled points are not too far from their centers. Here, it is used that if a point is conditioned on being sampled from a specific cluster, the sample more or less follows a uniform sampling from the cluster since the clusters are sufficiently far from one another. + +Then, we can show that given multiple sampled candidates, the greedy algorithm chooses one from an uncovered cluster to decrease the $k$ -means objective the most. Thus, as long as the greedy algorithm finds at least one candidate point from an uncovered cluster out of $\ell = \Theta (\ln k)$ candidates, it successfully covers an additional cluster in the current $\kappa$ -th iteration. + +The greedy algorithm may select a candidate point further from the centers when given multiple such candidate points, since it may prefer a point that is close to many clusters simultaneously. However, since the clusters have exponentially decaying tails, we can show that even the worst candidate point has a distance of at most $O(\ln \ln k)$ from its respective center in expectation. In contrast, we show that for regular instances, $k$ -means++ is highly likely to fail to cover many clusters because it only samples one point in each iteration. + +When clusters are very far apart from each other. Finally, we briefly discuss when the regular instance has large distances between the cluster centers, more formally when $\theta \geq 1 / 2$ . In this case, it becomes hard to show that the greedy algorithm performs better in terms of the $k$ -means objective. This is not surprising. Intuitively, when clusters are very far away from each other, each cluster can essentially be seen as a single point, and therefore there is very little room for $k$ -means++ to make mistakes. However, we can still show that the greedy algorithm has a higher chance of covering all clusters than $k$ -means++. Intuitively, the initial seeds having covered all clusters will likely result in a good clustering by a subsequent run of Lloyd's algorithm. Thus, even in this case, we indirectly show the advantage of the greedy algorithm over $k$ -means++. + +Experiments. Since it is well known that the greedy algorithm outperforms $k$ -means++ in practice and empirical studies, we rather focus on our experiments on tracking how the algorithm makes choices over iterations towards a better seeding. We create synthetic data sets using various distributions and study how the algorithm makes progress in terms of covering new clusters over iterations, not only tracking how the objective changes. The experiments show that the greedy algorithm outperforms $k$ -means++ in both decreasing the objective and covering new clusters. Thus, the experiments further corroborate the greedy algorithm's better performance, together with the theoretical analysis. + +# 1.2 Other Related Work + +While $k$ -means clustering is NP-hard to solve optimally, for any $\epsilon > 0$ , the problem admits a $(1 + \epsilon)$ -approximation when either $k$ or $d$ (the number of dimensions) is a constant. Feldman et al. [2007], Kumar et al. [2004]. Recent works show that $k$ -means++ can also be used with a small number of steps of a local search algorithm Lattanzi and Sohler [2019], Choo et al. [2020] to yield $O(1)$ -approximations. This result is further improved to $(9 + \epsilon)$ by Beretta et al. [2023], which matches the best approximation for the local search algorithm Kanungo et al. [2002]. Sketching can be used to compress the input data into a compact subset of points, called a coreset. This allows for faster clustering by running the algorithm on the coreset instead of the original data (e.g., Har-Peled and Mazumdar [2004], Chen [2009]). + +There is currently no theoretical analysis of Greedy $k$ -means++ beyond the works by Bhattacharya et al. [2020], Grunau et al. [2023]. For a comparative study of seeding methods, see Celebi et al. [2013]. They recommend a value of $\ell$ (number of candidates sampled per iteration) proportional to the logarithm of $k$ (number of clusters). The Scikit-learn library specifically sets $\ell = \lceil \ln k + 2 \rceil$ Pedregosa et al. [2011]. + +Except for the classical Greedy $k$ -means++ and $k$ -means++, several variants have also been studied. Aggarwal et al. [2009] show that $k$ -means++ is $O(1)$ -approximation with constant probability if it allows selecting $O(k)$ centers. This bicriteria approximation is further improved by Makarychev et al. + +[2020], Wei [2016]. Balcan et al. [2018] suggest seeding the initial centers via $D^{\alpha}$ -sampling, which generalizes the $k$ -means++ algorithm ( $\alpha = 2$ ). Bamas et al. [2024] analyze the new seeding method and show that $D^{\alpha}$ -sampling admits better approximation than $D^{2}$ -sampling under specific instances. + +The $k$ -means clustering has also been studied in various settings, including distributed Bahmani et al. [2012], streaming Ailon et al. [2009], and dynamic environments Bhattacharya et al. [2024]. In particular, Bahmani et al. [2012] extends $k$ -means++ to a distributed setting. + +# 1.3 Organization + +In the following section, we recall the greedy $k$ -means++ and $k$ -means++ algorithms and formally define the instances we will consider throughout this paper. To make our presentation transparent, we will only show our results for more restricted instances requiring fewer parameters, deferring the analysis of the more general instances to the appendix. Then, we analyze the greedy and $k$ -means++ when the distances between clusters are not too large in Section 3, 4. The other case is handled in Section 5. After presenting experiments in Section 6, we conclude the paper. + +# 2 Preliminaries + +This section formally defines the $k$ -means clustering problem, along with notations and background. + +$k$ -Means Clustering. Consider an $m$ -dimensional Euclidean space $\mathbb{R}^m$ . For a point $x \in \mathbb{R}^m$ , its connection cost to a point set $C \subseteq \mathbb{R}^m$ is defined as the squared distance of $x$ to its closest point in $C$ , i.e., $\varphi(x, C) := \min_{c \in C} ||x - c||_2^2$ . In the $k$ -means problem, we are given a set of $n$ points $X \subseteq \mathbb{R}^m$ as well as a parameter $k \in \mathbb{N}_{>0}$ , and the goal is to find a set of $k$ centers $S \subseteq \mathbb{R}^d$ that minimizes the total connection cost of points in $X$ to $S$ , i.e., $\varphi(X, S) := \sum_{x \in X} \varphi(x, S)$ . + +For a given point set $X$ , let $\{C_i\}_{i \in [k]}$ denote the $k$ clusters in the optimal solution, and let $\{\mu_i\}_{i \in [k]}$ represent the corresponding cluster centers. We define $C_i(r)$ as the set of points in $C_i$ that are at a distance $r$ from $\mu_i$ . + +We first formally present the statements of greedy $k$ -means++ in Algorithm 1. The $k$ -means++ algorithm is a special case of Algorithm 1 when $\ell = 1$ . Both of the two algorithms run iteratively. In each iteration, $k$ -means++ samples one candidate from the probability distribution $\{\varphi(x, S) / \varphi(X, S)\}$ while greedy $k$ -means++ samples $\ell > 1$ candidates and pick the one with the minimum connection cost. Since Greedy $k$ -means++ commonly uses $\ell = \ln k + \Theta(1)$ Pedregosa et al. [2011], we will assume $\ell = \ln k$ and $k$ is sufficiently large unless stated otherwise. + +Algorithm 1 Greedy $k$ -means++ Initialization Arthur et al. [2007] +Input: A point set $X\subseteq \mathbb{R}^m$ and parameters $k > 0,\ell = \ln k^3$ +Output: A center set $S$ that serves as the initial centers of Lloyd's heuristic. +1: Independently and uniformly sample $\ell$ points $x_{1},\ldots ,x_{\ell}\in X$ +2: Greedily pick $x\coloneqq \arg \min_{x_i\in \{x_1,\dots,x_\ell \}}\varphi (X,\{x_i\})$ and set $S\gets \{x\}$ +3: for $t = 2,\ldots ,k$ do +4: Sample $\ell$ points $x_{1},\ldots ,x_{\ell}\in X$ independently (with replacement) with probability $\frac{\varphi(x,S)}{\varphi(X,S)}$ +5: Greedily pick $x\coloneqq \arg \min_{x_i\in \{x_1,\dots,x_\ell \}}\varphi (X,S\cup \{x_i\})$ , breaking ties arbitrarily. +6: Set $S\gets S\cup \{x\}$ +7: end for +8: return $S$ + +The paper analyzes these two algorithms on the input point set $X$ satisfying the three properties: + +- Exponentially Distributed: For each cluster $C_i$ , the density of points decreases exponentially as the distance from the center increases. Specifically, $\frac{|C_i(r)|}{|C_i|} = \frac{1}{b} \cdot e^{-r / b}$ , for a constant $b > 0$ . +- Well Separable: The minimum distance $d$ between any two centers of the optimal clusters be sufficiently large relative to the cluster distribution parameter $b$ , i.e., $d = k^{\theta} \cdot b$ , for a constant $\theta > 0$ . + +- Well Spread: The optimal clusters are roughly homogeneous, with the number of points in each cluster and the distances between clusters differing by at most a constant factor. + +Define a point set that satisfies the above properties as an EWW point set. We remark that our analysis applies to any subexponential-tailed distribution, such as Gaussian and sub-Gaussian distributions, with a mild assumption. The details are deferred to Appendix C. For simplicity and readability, we assume throughout the analysis that each cluster has the same size and equal pairwise distances. We remark that the analysis can be easily extended to the case where these quantities differ by at most a constant factor. These extensions can also be found in Appendix C. + +We now present several useful observations concerning the structure of EWW point sets, which can be readily derived through straightforward mathematical calculations. + +Observation 1. The optimal total connection cost is achieved by selecting each cluster center $\mu_{i}$ , resulting in an objective: $\mathsf{OPT} = n\int_{0}^{\infty}\frac{r^{2}}{b}\cdot e^{-r / b}dr = 2b^{2}n$ , where $n$ is the number of points. + +Observation 2. Consider a cluster $C_i$ with its center $\mu_i$ . For a point $x$ located at distance $h$ from $\mu_i$ , the total connection cost of all points in $C_i$ to $x$ is $(2b^2 + h^2) \cdot |C_i|$ . + +# 3 Approximation Guarantee of Greedy $k$ -Means++ + +In this section, we analyze the performance of the greedy k-means++ algorithm on EWW point sets and aim to show the following: + +Theorem 1. Given any EWW point set $X$ , the greedy $k$ -means++ algorithm admits an expected approximation ratio of $O((\ln \ln k)^2)$ . + +Proof Outlines. Due to the exponentially distributed property, we can show that, with very high probability, the points sampled by the algorithm are located close to the optimal centers $\{\mu_i\}_{i\in [k]}$ . In the following, we define a concentration ball for each cluster and prove that we may assume the algorithm never samples points outside these balls. In particular, the chance the algorithm samples a point outside these balls is very low and, due to this, the total contribution to the expected objective is small in this case. Under this concentration assumption, we classify the clusters $\{C_i\}_{i\in [k]}$ into two types at each iteration: covered clusters (those for which the algorithm has already selected a point within the corresponding concentration ball), and uncovered clusters (the rest). Due to the greedy nature of the algorithm, it always prefers candidate points from uncovered clusters over those from covered clusters. Leveraging the well-separability and well-spreadness, we show that with high probability, the algorithm covers a new cluster in most iterations, thereby achieving the desired approximation ratio. + +We remark that, while the above captures the high-level strategy of our analysis, applying only these techniques can only give an approximation ratio of $O((\ln k)^2)$ . To improve the ratio to $O((\ln \ln k)^2)$ , we further exploit the independence among clusters induced by well separability and conduct a more careful analysis. We begin by introducing the definition of a concentration ball and the corresponding lemma. + +Definition 1 (Concentration Ball). We define the concentration radius $\delta \coloneqq 4b(1 + \theta)\ln k$ , where $\theta$ and $b$ are input parameters. For each cluster $C_i$ with center $\mu_i$ , we define its concentration ball $\sigma(C_i)$ as the set of all points in $C_i$ whose distance to $\mu_i$ is at most $\delta$ . + +Lemma 1 (Concentration Lemma). Let $\mathcal{A}$ denote the event that greedy $k$ -means++ samples at least one candidate point outside the concentration balls during any iteration. Given any EWW point set, the probability that $\mathcal{A}$ occurs is at most $1 / k$ . Furthermore, the contribution of this event to the expected objective can be bounded: $\operatorname*{Pr}[\mathcal{A}] \cdot \mathbb{E}[\mathsf{OBJ} \mid \mathcal{A}] \leq \frac{n}{k}$ , where $\mathbb{E}[\mathsf{OBJ} \mid \mathcal{A}]$ denotes the expected objective value (i.e., total connection cost) conditioned on the occurrence of $\mathcal{A}$ . Furthermore, this upper bound holds as well when $\mathcal{A}$ is $k$ -means++. + +As $\mathrm{OPT} = \Theta(n)$ (see Observation 1), the above lemma implies that the total contribution of points outside the concentration balls is modest compared to the optimal cost. Hence, throughout the remainder of this paper, we adopt the concentration assumption, under which no point outside the concentration balls is sampled by the algorithm, whether it is greedy $k$ -means++ or $k$ -means++. + +We say that a feasible solution covers a cluster $C_i$ if it includes at least one point from its concentration ball $\sigma(C_i)$ . We claim the following: + +Lemma 2. Under the concentration assumption, for each cluster $C_i$ , we have: + +(2a) If greedy $k$ -means++ does not cover this cluster, then its total connection cost is $\Omega (k^{2\theta} \cdot |C_i|)$ . +(2b) If greedy $k$ -means++ covers this cluster, then the total connection cost for $C_i$ does not exceed $O((\ln k)^2 \cdot |C_i|)$ , and further, its expectation is $O((\ln \ln k)^2 \cdot |C_i|)$ and $\Omega(|C_i|)$ . + +As established in Lemma 2, under the well-spread assumption, achieving the approximation guarantee stated in Theorem 1 requires that the number of uncovered clusters is at most $O(k^{1 - 2\theta} \cdot (\ln \ln k)^2)$ . Otherwise, the total connection cost contributed by the uncovered clusters would exceed $(\ln \ln k)^2 \cdot OPT$ , violating the desired bound. The next lemma establishes the probability of covering a new cluster, which will later be used to determine the number of uncovered clusters. + +Lemma 3. In each iteration $t \leq k - O(k^{1 - 2\theta} \cdot (\ln \ln k)^2)$ , the greedy $k$ -means++ algorithm covers a new cluster with probability at least $1 - 1 / k^{2\theta + 2}$ . + +Proof. We first leverage (2a) of Lemma 2 to show that, with high probability, the greedy $k$ -means++ algorithm covers a new cluster in each iteration up to iteration $k - O(k^{1 - 2\theta} \cdot (\ln k)^2)$ . Then, by applying (2b) of Lemma 2 along with the Chernoff bound, we demonstrate that the algorithm continues to cover a new cluster in each iteration, with high probability, even during the range of iterations from $k - O(k^{1 - 2\theta} \cdot (\ln k)^2)$ to $k - O(k^{1 - 2\theta} \cdot (\ln \ln k)^2)$ . Intuitively, the reason we cannot directly apply the Chernoff bound is that, when doing so, we require the connection cost upper bound of each individual covered cluster to be significantly smaller than the expected total connection cost of the covered clusters, which only holds when the greedy is at at a later iteration, say for $t \geq k / 2$ . + +First, consider an iteration $k \leq t_0 \coloneqq k - 2^{2\theta + 5} \cdot k^{1 - 2\theta} \cdot (\ln k)^2$ . Lemma 2 implies that the greedy comparison step always prefers points from uncovered concentration balls over those from already covered ones. This is because if the greedy selects a point from an uncovered concentration ball (by (2a)), it decreases the objective by $\Omega(n / k)k^{2\theta}$ while if it selects from a covered concentration ball, it only decreases the objective by $O(n / k)(\ln k)^2$ (by (2b)). Thus, it follows that, in iteration $t$ , if at least one of the $\ln k$ sampled candidates is from an uncovered concentration ball, then the algorithm is guaranteed to cover a new cluster in that iteration. Suppose that $p$ clusters have already been covered by $S_{t-1}$ at the beginning of iteration $t$ (so $p < t$ ). By (2b) of Lemma 2 and the well-spread property, the total connection cost contributed by the covered clusters is at most $\frac{n}{k} \cdot p \cdot (\ln k)^2$ , while the total connection cost from the uncovered clusters is at least $\frac{n}{k} \cdot (k - p) \cdot k^{2\theta}$ , where we omit constant factors for simplicity—such constants do not affect our analysis. Therefore, the probability4 that the algorithm fails to sample any point from the uncovered concentration balls is at most + +$$ +\left(\frac {p \cdot (\ln k) ^ {2}}{p \cdot (\ln k) ^ {2} + (k - p) \cdot k ^ {2 \theta}}\right) ^ {\ln k} \leq 1 / k ^ {2 \theta + 5} \text {w h e n} p \leq t _ {0}. +$$ + +Next, we analyze the time period between iteration $t_0$ and $t_1 \coloneqq k - 2^{2\theta + 5} \cdot k^{1 - 2\theta} \cdot (\ln \ln k)^2$ . From the earlier analysis and by applying a union bound, we know that with probability at least $1 - 1 / k^{2\theta + 4}$ , the algorithm has already covered $t_0$ clusters within the first $t_0$ iterations. Then, according to Lemma 2, conditioned on this event, the expected connection cost of these covered clusters is lower bounded by $\Omega \left( \frac{n}{k} \cdot t_0 \right)$ , which is asymptotically much larger than the upper bound on the connection cost for any single cluster, $O\left( \frac{n}{k} \cdot (\ln k)^2 \right)$ . This suggests the connection cost of the covered clusters is well-concentrated. + +We shall upper bound the total connection cost of the covered clusters in the first $t_0$ iterations via the concentration bound. Note that each cluster can be approximately treated as being sampled uniformly based on the analysis in the proof of Lemma 2, which enables us to use concentration inequalities. + +Applying the concentration bound, we get the following claim; the proof is deferred to the appendix. + +Claim 1. With probability at least $1 - \exp \left(\Theta (1)\cdot (-k)\cdot \left(\frac{\ln\ln k}{\ln k}\right)^2\right)$ , the total connection cost of the clusters covered in the first $t_0$ iterations is at most $2b^{2}\cdot \frac{n}{k}\cdot t_{0}\cdot (\ln \ln k)^{2}$ . + +Consider any iteration $t \in (t_0, t_1]$ . The algorithm can cover at most $t_1 - t_0$ new clusters during the interval from iteration $t_0$ to $t_1$ , and each newly covered cluster contributes at most $O\left(\frac{n}{k} \cdot (\ln k)^2\right)$ to the connection cost. Thus, with a probability of at least + +$$ +\left(1 - \frac {1}{k ^ {2 \theta + 4}}\right) \left(1 - \exp \left(\Theta (1) \cdot (- k) \cdot \left(\frac {\ln \ln k}{\ln k}\right) ^ {2}\right)\right), +$$ + +the total connection cost of the already covered clusters at the beginning of iteration $t$ is at most + +$$ +A := 2 \cdot \frac {n}{k} \cdot \left(b ^ {2} \cdot t _ {0} \cdot (\ln \ln k) ^ {2} + (t _ {1} - t _ {0}) \cdot (\ln k) ^ {2}\right). +$$ + +Note that the total connection cost from the uncovered clusters is at least $B \coloneqq \frac{n}{k} \cdot (k - t_0) \cdot k^{2\theta}$ . Similar to the analysis in the previous case, the probability that the algorithm fails to sample any point from the uncovered concentration balls is at most $(A / (A + B))^{\ln k}$ . This implies that the probability that greedy $k$ -means++ fails to cover a new cluster is at most $1 / k^{2\theta + 2}$ . + +Proof of Theorem 1. By Lemma 3 and a union bound, we have that, with probability at least $1 - 1 / k^{2\theta +1}$ , the greedy $k$ -means++ algorithm covers $k - O\left(k^{1 - 2\theta}\cdot (\ln \ln k)^{2}\right)$ clusters and achieves an approximation ratio of $O((\ln \ln k)^2)$ . If this case does not occur, we can simply use an upper bound on the objective of $O(n\cdot k^{2\theta})$ . Taking expectation over both cases, we obtain an expected approximation ratio of $O((\ln \ln k)^2)$ . + +# 4 Approximation Lower Bound of $k$ -Means++ + +This section analyzes the $k$ -means++ algorithm and establishes a lower bound $\Omega(\ln k)$ on its approximation ratio. This lower bound highlights a gap between the performance of the greedy $k$ -means++ and standard $k$ -means++ algorithms: while the latter suffers from an $\Omega(\ln k)$ lower bound, the former achieves an $O((\ln \ln k)^2)$ approximation ratio on the same instances, demonstrating the theoretical advantage. + +Theorem 2. Given any EWW point set $X$ with parameter $\theta \in (0,1 / 2]$ , the $k$ -means++ algorithm has an expected approximation ratio of $\Omega (\ln k)$ . + +Our proof strategy mirrors that used for greedy $k$ -means++, where we analyze the expected number of uncovered clusters to derive bounds on the approximation ratio. Specifically, to establish a lower bound for the algorithm, we aim to show that with non-negligible probability, $k$ -means++ selects points from already covered concentration balls in certain iterations (recall that $k$ -means++ samples only one point per iteration, whereas greedy $k$ -means++ samples $\ln k$ candidates). To this end, we require a lemma symmetric to Lemma 2, which provides a lower bound on the probability that the algorithm samples from an already covered concentration ball. + +Lemma 4. Under the concentration assumption, for each cluster $C_i$ , we have: + +(4a) If $k$ -means++ does not cover this cluster, then its total connection cost is $\Theta (k^{2\theta}\cdot |C_i|)$ . +(4b) If $k$ -means++ covers this cluster using exactly one center—that is, the final solution includes exactly one point from $\sigma(C_i)$ —then the total connection cost for $C_i$ is $\Omega(|C_i|)$ . + +Proof of Lemma 4: This proof follows a similar analysis to that in Lemma 2. By Observation 2, the total connection cost of a cluster $C_i$ is $(2b^2 + h^2) \cdot |C_i|$ , where $h$ denotes the distance from the cluster center $\mu_i$ to the current set of selected centers $S$ . When the cluster is not yet covered, $h \in [d - \delta, d + 2\delta]$ , which is of order $\Theta(k^\theta)$ ; whereas once the cluster is covered, $h$ can be as small as 0. This completes the proof of the lemma. + +Proof of Theorem 2. Lemma 4 shows that if the number of uncovered clusters is $p$ , then the objective value is at least $\frac{n}{k} \cdot p \cdot k^{2\theta}$ , omitting constant factors for simplicity. Therefore, to establish a lower bound of $\Omega(\ln k)$ on the approximation ratio, it suffices to show that, in expectation, $k$ -means++ leaves $\Omega(\ln k \cdot k^{1 - 2\theta})$ clusters uncovered. This also explains why we require $\theta \in (0, 1/2]$ : otherwise, $\ln k \cdot k^{1 - 2\theta}$ would be subconstant, rendering the argument meaningless. + +We partition all possible outcomes of the algorithm into two cases based on the number of uncovered clusters: (1) the final number of uncovered clusters is at least $\Delta$ , and (2) the final number of uncovered clusters is less than $\Delta$ , where $\Delta = \ln k \cdot k^{1 - 2\theta}$ . Clearly, in all outcomes falling into the first case, the approximation ratio is $\Omega(\ln k)$ . Next, we analyze the second case and show that, in expectation, the number of uncovered clusters remains $\Omega(\Delta)$ . + +Consider an arbitrary iteration $t$ . In the second case, where the final number of uncovered clusters is less than $\Delta$ , the number of clusters already covered by the solution $S_{t-1}$ at the beginning of this iteration must be at least $t - \Delta$ . Otherwise, even if every subsequent iteration covers a new cluster, the final number of uncovered clusters would exceed $\Delta$ , contradicting the assumption. Then, by the pigeonhole principle, at least $t - 2\Delta$ clusters must be covered by exactly one center—that is, $S_{t-1}$ contains exactly one point from each of these clusters. By Lemma 4 and the well-spread property, the total connection cost of the covered clusters is at least $\frac{n}{k} \cdot (t - 2\Delta)$ , while the total connection cost of the uncovered clusters is at most $\frac{n}{k} \cdot (k - t + \Delta) \cdot k^{2\theta}$ . Therefore, in each iteration $t > 2\Delta$ , the probability that the $k$ -means++ algorithm fails to cover a new cluster is at least $\frac{t - 2\Delta}{(t - 2\Delta) + (k - t + \Delta) \cdot k^{2\theta}}$ . + +We compute the expected number of uncovered clusters by summing the failure probabilities across all iterations (conditioned on the second case). Specifically, we have: + +$$ +\begin{array}{l} \mathbb {E} [ \text {n u m b e r o f u n c o v e r e d c l u s t e r s} ] \geq \sum_ {t > 2 \Delta} ^ {k} \frac {t - 2 \Delta}{(t - 2 \Delta) + (k - t + \Delta) \cdot k ^ {2 \theta}} \\ \geq \sum_ {t \geq k / 2} ^ {k} \frac {t - 2 \Delta}{(t - 2 \Delta) + (k - t + \Delta) \cdot k ^ {2 \theta}} \geq \sum_ {t \geq k / 2} ^ {k} \frac {k / 4}{k / 4 + (k - t + \Delta) \cdot k ^ {2 \theta}} \quad (\Delta = o (k)) \\ = k ^ {1 - 2 \theta} \cdot \sum_ {t \geq k / 2} ^ {k} \frac {1}{k ^ {- 2 \theta} + 4 (k - t + \Delta)} \geq k ^ {1 - 2 \theta} \ln \left(\frac {k / 2 + \Delta}{\Delta}\right) \\ \geq k ^ {1 - 2 \theta} \ln \left(\frac {k ^ {2 \theta}}{2 \ln k}\right) = \Omega (k ^ {1 - 2 \theta} \ln k), \\ \end{array} +$$ + +which implies that the expected number of uncovered clusters is $\Omega (\Delta)$ and completes the proof. + +![](images/c70e5d2b0f713b38b0b8d9542d660b14cacfaffcb036730fb4169ba2ed178afb.jpg) + +# 5 Analysis of Covering Probability + +Theorem 1 and Theorem 2 demonstrate that, on the EWW point set with parameter $\theta \in (0,1 / 2]$ , the greedy $k$ -means++ algorithm achieves a better approximation ratio than the standard $k$ -means++ algorithm. The intuition is that when $\theta \in (0,1 / 2]$ , the optimal clusters are not yet well-separated, so the probability that $k$ -means++ fails to cover a new cluster in a given iteration remains relatively high. In contrast, greedy $k$ -means++ can exponentially reduce this failure probability through multiple samples per iteration. + +As $\theta$ increases further and the optimal clusters become more widely separated, the failure probability for standard $k$ -means++ correspondingly decreases, reducing the approximation gap between the two algorithms. This section formally addresses such cases, showing that even in these settings, greedy $k$ -means++ remains theoretically superior to standard $k$ -means++ from a certain perspective. + +Theorem 3. Given any EWW point set with parameter $\theta > 1/2$ , the probability that greedy $k$ -means++ covers all optimal clusters is greater than that of $k$ -means++. + +One might find this theorem intuitively trivial. Since greedy $k$ -means++ performs multiple samples per iteration, it should naturally have a higher probability of covering a new cluster than $k$ -means++, which would suggest the theorem's correctness. However, this reasoning strictly holds only when both algorithms share the same set of selected centers $S$ , which we cannot guarantee. In fact, during the execution of greedy $k$ -means++ and $k$ -means++, the distribution over all possible center sets $S_{t}$ at each iteration $t$ may differ significantly, which makes the proof of the theorem non-trivial. + +Observe that if an algorithm fails to cover all optimal clusters, it must have selected at least two points from the same optimal cluster. Therefore, the probability that an algorithm covers all optimal clusters is equal to 1 minus the probability that there exists at least one iteration in which the algorithm selects a point from an already covered cluster. + +To prove Theorem 3, we analyze the probabilities that greedy $k$ -means++ and $k$ -means++ encounter such a bad event. Specifically, we establish a lower bound on that probability for $k$ -means++ (Lemma 5) and an upper bound for greedy $k$ -means++ (Lemma 6), and finally show that the former is greater than the latter. Let event $\mathcal{B}$ denote the bad event that the algorithm selects a point from an already covered cluster. We claim the following. + +Lemma 5. The probability that $k$ -means++ encounters $\mathcal{B}$ is at least $\frac{k - 1}{k - 1 + k^{2\theta}}$ . + +Proof of Lemma 5: We partition the bad event into two sub-events based on the time at which $\mathcal{B}$ first occurs: (1) $k$ -means++ encounters $\mathcal{B}$ before the last iteration $k$ , and (2) $k$ -means++ first encounters $\mathcal{B}$ at the last iteration $k$ . We denote these two sub-events as $\mathcal{P}$ and $\mathcal{Q}$ , respectively. By expanding the conditional probability of the second sub-event, we derive a lower bound on the probability that $k$ -means++ encounters $\mathcal{B}$ : + +$$ +\Pr [ k \text {- m e a n s + +} \text {e n c o u n t e r s} \mathcal {B} ] = \Pr [ \mathcal {P} ] + \Pr [ \mathcal {Q} ] = \Pr [ \mathcal {P} ] + \Pr [ \neg \mathcal {P} ] \cdot \Pr [ \mathcal {Q} | \neg \mathcal {P} ] \geq \Pr [ \mathcal {Q} | \neg \mathcal {P} ]. +$$ + +Conditioned on $\neg \mathcal{P}$ , the notion of "first" in event $\mathcal{Q}$ is not essential— $\operatorname*{Pr}[\mathcal{Q} \mid \neg \mathcal{P}]$ simply equals the probability that $k$ -means++ samples a point from one of the $k-1$ already covered clusters. By Lemma 4, the total connection cost of the already covered clusters is at least $\frac{n}{k} \cdot (k-1)$ , while the total connection cost of the last uncovered cluster is at most $\frac{n}{k} \cdot k^{2\theta}$ , omitting constant factors for simplicity. Thus, the probability that $k$ -means++ encounters event $\mathcal{B}$ can be lower bounded by + +$$ +\frac {k - 1}{k - 1 + k ^ {2 \theta}}. +$$ + +Lemma 6. The probability that greedy $k$ -means++ encounters $\mathcal{B}$ is at most $\left(\frac{e \cdot (k - 1) \cdot (\ln k)^2}{(k - 1) \cdot (\ln k)^2 + k^{2\theta}}\right)^{\ln k}$ . + +Proof of Lemma 6: To upper bound the probability for greedy $k$ -means++, we partition the bad event into $k$ sub-events $\{\mathcal{P}_t\}_{t \in [k]}$ , where $\mathcal{P}_t$ denotes the event that greedy $k$ -means++ encounters $\mathcal{B}$ for the first time in iteration $t$ . Similarly, we then expand the probability of $\mathcal{P}_t$ using conditional probability: + +$$ +\begin{array}{l} \Pr [ \text {g r e e d y} k \text {- m e a n s + +} \text {e n c o u n t e r s} \mathcal {B} ] = \sum_ {t \in [ k ]} \Pr [ \mathcal {P} _ {t} ] \\ = \sum_ {t \in [ k ]} \Pr \left[ \neg \left(\mathcal {P} _ {1} \vee \dots \vee \mathcal {P} _ {t - 1}\right) \right] \cdot \Pr \left[ \mathcal {P} _ {t} \mid \neg \left(\mathcal {P} _ {1} \vee \dots \vee \mathcal {P} _ {t - 1}\right) \right] \leq k \cdot \Pr \left[ \mathcal {P} _ {k} \mid \neg \left(\mathcal {P} _ {1} \vee \dots \vee \mathcal {P} _ {k - 1}\right) \right] \\ \end{array} +$$ + +where the last inequality uses the fact that $\operatorname*{Pr}[\mathcal{P}_t\mid \neg (\mathcal{P}_1\lor \dots \lor \mathcal{P}_{t - 1})]$ attains its maximum at $t = k$ + +The term $\operatorname*{Pr}[\mathcal{P}_k\mid \neg (\mathcal{P}_1\lor \dots \lor \mathcal{P}_{k - 1})]$ simply equals the probability that greedy $k$ -means++ samples all $\ln k$ candidates from one of the $k - 1$ already uncovered clusters. By Lemma 2, the total connection cost of the already covered cluster is at most $\frac{n}{k}\cdot (k - 1)\cdot (\ln k)^2$ , while the total connection cost of the last uncovered cluster is at most $\frac{n}{k}\cdot k^{2\theta}$ , omitting constant factors for simplicity. Thus, the probability that greedy $k$ -means++ encounters $\mathcal{B}$ can be upper bounded by + +$$ +k \cdot \left(\frac {(k - 1) \cdot (\ln k) ^ {2}}{(k - 1) \cdot (\ln k) ^ {2} + k ^ {2 \theta}}\right) ^ {\ln k} = \left(\frac {2 \cdot (k - 1) \cdot (\ln k) ^ {2}}{(k - 1) \cdot (\ln k) ^ {2} + k ^ {2 \theta}}\right) ^ {\ln k}. +$$ + +Proof of Theorem 3. Lemma 5 and Lemma 6, through a series of mathematical calculations, directly establish the theorem. More specifically, when $\theta > 1/2$ , the lower bound on the failure probability for $k$ -means++ is $\Theta(k^{1-2\theta})$ , while the upper bound for greedy $k$ -means++ is $\Theta\left(k^{e\ln \ln k + (1-2\theta)\ln k}\right)$ . As the former asymptotically dominates the latter, we can conclude that greedy $k$ -means++ has a higher probability of covering all optimal clusters than $k$ -means++. + +# 6 Experiments + +Our experiments are conducted on a machine with Processor 11th Gen Intel(R) Core(TM) i5-1135G7 2.40GHz, 1382 Mhz, 4 Core(s), 8 Logical Processor(s) and 12 GB RAM. We evaluate the performance of greedy $k$ -means++ and $k$ -means++ on 3 datasets. + +Our goal is to demonstrate that the theory is predictive of practice. Our experiments correspond to our theoretical model and remark that there is extensive empirical work on both $k$ -means++ and the greedy variant previously known. + +Input Data. We conduct experiments on synthetic datasets. To generate a dataset, we first fix a distribution. In the experiment, we use the exponential, half-normal (the absolute variant of the Gaussian), and Lomax (heavy-tail sub-exponential) distributions for different datasets. We first sample $k$ centers in $\mathbb{R}^k$ uniformly at random from a unit hypercube. Then we sample the radius for each cluster uniformly at random from $(0,2)$ . The number of points of each cluster is uniformly sampled from [64, 256]. To generate the points of a cluster, we choose uniform random points whose distance to the center follows the fixed distribution. Since the centers are sampled from a unit hypercube, it is well-separable because the distances between every two centers are $\sqrt{k}$ . The generation of the points guarantees that it is well-spread. + +Experiments. We consider two different metrics - the $k$ -means objective and coverage probability. For greedy $k$ -means++, we sample $\lceil \ln k \rceil + 2$ candidates in every iteration. The experiment involves 100 repetitions. We compare the average objective and coverage probability in every iteration. The average objective in an iteration is the average $k$ -means objective using the currently chosen centers in 100 repetitions. Similarly, the average coverage probability is the probability that a new center is covered in 100 repetitions. We show the results over iterations for $k = 16$ in Figure 1 (refer to Appendix D for different $k$ ). We can see the superior performance of greedy $k$ -means++ over $k$ -means++ under two different measures, which validates our theory that greedy $k$ -means++ outperforms $k$ -means++. + +![](images/f00b2cf45b85a68fc364fdf425651e346c1d80b49c69099593b4f2dcebcb51b4.jpg) +Figure 1: Coverage probability and $k$ -means objective over iterations for $k = 16$ . + +![](images/7f179963714aff49ae86416ad384409c2e38656779df252510c666c2dc99beb9.jpg) + +# 7 Conclusions + +In this paper, we presented the first beyond-worst-case analysis of the greedy $k$ -means++ algorithm. We conclude the paper with some open problems. Our analysis assumes that the greedy algorithm samples $\ln k + \Theta(1)$ candidate points per iteration. While this is commonly used in practice, sampling a constant number of candidates could still place the greedy ahead of $k$ -means++. Our current analysis falls short of showing this and studying the greedy's performance when $\ell = o(\ln k)$ could be interesting. Also, it could be plausible that one can prove the greedy algorithm has a better approximation ratio than $O((\ln \ln k)^2)$ for the EWW instances. Finally, it would be very interesting to discover new algorithms that improve the greedy algorithm now that we have a theoretical understanding of it in a beyond-worst-case setting. + +# Acknowledgements + +Chenyang Xu was supported by the National Natural Science Foundation of China (No. 62302166) and the National Key Research Project of China under Grant No. 2023YFA1009402. Qingyun Chen, Sungjin Im, and Ryan Milstrey were supported in part by NSF grants CCF-2535599, CCF-2537126, and CCF-2423106. Benjamin Moseley was supported in part by a Google Research Award and NSF grants CCF-2121744 and CCF-1845146. + +# References + +Marcel R Ackermann and Johannes Blömer. Bregman clustering for separable instances. In Scandinavian Workshop on Algorithm Theory, pages 212-223. Springer, 2010. +Ankit Aggarwal, Amit Deshpande, and Ravi Kannan. Adaptive sampling for k-means clustering. In International Workshop on Approximation Algorithms for Combinatorial Optimization, pages 15-28. Springer, 2009. +Charu C. Aggarwal and Chandan K. Reddy, editors. Data Clustering: Algorithms and Applications. CRC Press, 2014. ISBN 978-1-46-655821-2. URL http://www.crcpress.com/product/isbn/9781466558212. +Sara Ahmadian, Ashkan Norouzi-Fard, Ola Svensson, and Justin Ward. Better guarantees for k-means and euclidean k-median by primal-dual algorithms. SIAM Journal on Computing, 49(4): FOCS17-97, 2019. +Nir Ailon, Ragesh Jaiswal, and Claire Monteleoni. Streaming k-means approximation. In Y. Bengio, D. Schuurmans, J. Lafferty, C. Williams, and A. Culotta, editors, Advances in Neural Information Processing Systems, volume 22. Curran Associates, Inc., 2009. URL https://proceedings.neurips.cc/paper_files/paper/2009/file/4f16c818875d9fcb6867c7bdc89be7eb-Paper.pdf. +Daniel Aloise, Amit Deshpande, Pierre Hansen, and Preyas Popat. Np-hardness of euclidean sum-of-squares clustering. Machine learning, 75:245-248, 2009. +David Arthur, Sergei Vassilvitskii, et al. k-means++: The advantages of careful seeding. In Soda, volume 7, pages 1027-1035, 2007. +Pranjal Awasthi, Moses Charikar, Ravishankar Krishnaswamy, and Ali Kemal Sinop. The hardness of approximation of euclidean k-means. arXiv preprint arXiv:1502.03316, 2015. +Bahman Bahmani, Benjamin Moseley, Andrea Vattani, Ravi Kumar, and Sergei Vassilvitskii. Scalable k-means++. Proc. VLDB Endow., 5(7):622-633, 2012. doi: 10.14778/2180912.2180915. URL http://vldb.org/pvldb/vol1/p622_bahmanbahmani_vldb2012.pdf. +Maria-Florina F Balcan, Travis Dick, and Colin White. Data-driven clustering via parameterized lloyd's families. Advances in neural information processing systems, 31, 2018. +Etienne Bamas, Sai Ganesh Nagarajan, and Ola Svensson. Analyzing $d^{\alpha}$ seeding for $k$ -means. In *Forty-first International Conference on Machine Learning*, 2024. +Lorenzo Beretta, Vincent Cohen-Addad, Silvio Lattanzi, and Nikos Parotsidis. Multi-swap k-means++. Advances in Neural Information Processing Systems, 36:26069–26091, 2023. +Anup Bhattacharya, Jan Eube, Heiko Röglin, and Melanie Schmidt. Noisy, Greedy and Not so Greedy k-Means++. In Fabrizio Grandoni, Grzegorz Herman, and Peter Sanders, editors, 28th Annual European Symposium on Algorithms (ESA 2020), volume 173 of Leibniz International Proceedings in Informatics (LIPics), pages 18:1-18:21, Dagstuhl, Germany, 2020. Schloss Dagstuhl – Leibniz-Zentrum für Informatik. ISBN 978-3-95977-162-7. doi: 10.4230/LIPicsESA.2020.18. URL https://drops.dagstuhl.de/entities/document/10.4230/LIPicsESA.2020.18. +Sayan Bhattacharya, Martin Costa, Silvio Lattanzi, and Nikos Parotsidis. Fully dynamic $k$ -clustering in $\tilde{O}(k)$ update time. Advances in Neural Information Processing Systems, 36, 2024. + +Christopher M. Bishop. Pattern recognition and machine learning, 5th Edition. Information science and statistics. Springer, 2007. ISBN 9780387310732. URL https://www.worldcat.org/oclc/71008143. +Christopher M Bishop. Model-based machine learning. Philosophical Transactions of the Royal Society A: Mathematical, Physical and Engineering Sciences, 371(1984):20120222, 2013. +Vladimir Braverman, Adam Meyerson, Rafail Ostrovsky, Alan Roytman, Michael Shindler, and Brian Tagiku. Streaming k-means on well-clusterable data. In Proceedings of the twenty-second annual ACM-SIAM symposium on Discrete Algorithms, pages 26-40. SIAM, 2011. +M Emre Celebi, Hassan A Kingravi, and Patricio A Vela. A comparative study of efficient initialization methods for the k-means clustering algorithm. Expert systems with applications, 40(1):200-210, 2013. +Ke Chen. On coresets for k-median and k-means clustering in metric and euclidean spaces and their applications. SIAM Journal on Computing, 39(3):923-947, 2009. +Davin Choo, Christoph Grunau, Julian Portmann, and Václav Rozhon. k-means++: few more steps yield constant approximation. In International Conference on Machine Learning, pages 1909-1917. PMLR, 2020. +Vincent Cohen-Addad, Hossein Esfandiari, Vahab Mirrokni, and Shyam Narayanan. Improved approximations for euclidean k-means and k-median, via nested quasi-independent sets. In Proceedings of the 54th Annual ACM SIGACT Symposium on Theory of Computing, pages 1621–1628, 2022. +Dan Feldman, Morteza Monemizadeh, and Christian Sohler. A ptas for k-means clustering based on weak coresets. In Proceedings of the twenty-third annual symposium on Computational geometry, pages 11-18, 2007. +Fabrizio Grandoni, Rafail Ostrovsky, Yuval Rabani, Leonard J Schulman, and Rakesh Venkat. A refined approximation for euclidean k-means. Information Processing Letters, 176:106251, 2022. +Christoph Grunau, Ahmet Alper Özüdogru, Václav Rozhón, and Jakub Tětek. A nearly tight analysis of greedy k-means++. In Proceedings of the 2023 Annual ACM-SIAM Symposium on Discrete Algorithms (SODA), pages 1012-1070. SIAM, 2023. +Sariel Har-Peled and Soham Mazumdar. On coresets for k-means and k-median clustering. In Proceedings of the thirty-sixth annual ACM symposium on Theory of computing, pages 291-300, 2004. +Ragesh Jaiswal and Nitin Garg. Analysis of k-means++ for separable data. In International Workshop on Approximation Algorithms for Combinatorial Optimization, pages 591-602. Springer, 2012. +Tapas Kanungo, David M Mount, Nathan S Netanyahuu, Christine D Piatko, Ruth Silverman, and Angela Y Wu. A local search approximation algorithm for k-means clustering. In Proceedings of the eighteenth annual symposium on Computational geometry, pages 10-18, 2002. +Amit Kumar, Yogish Sabharwal, and Sandeep Sen. A simple linear time $(1 + \epsilon)$ -approximation algorithm for k-means clustering in any dimensions. In Proceedings of the 45th Annual IEEE Symposium on Foundations of Computer Science (FOCS'04), pages 332-341. IEEE, 2004. +Silvio Lattanzi and Christian Sohler. A better k-means++ algorithm via local search. In International Conference on Machine Learning, pages 3662-3671, 2019. +Euiwoong Lee, Melanie Schmidt, and John Wright. Improved and simplified inapproximability for k-means. Information Processing Letters, 120:40-43, 2017. +Stuart Lloyd. Least squares quantization in pmc. IEEE transactions on information theory, 28(2): 129-137, 1982. +Meena Mahajan, Prajakta Nimbhorkar, and Kasturi Varadarajan. The planar k-means problem is np-hard. Theoretical Computer Science, 442:13-21, 2012. + +Konstantin Makarychev, Aravind Reddy, and Liren Shan. Improved guarantees for k-means++ and k-means++ parallel. Advances in Neural Information Processing Systems, 33:16142-16152, 2020. +Fabian Pedregosa, Gael Varoquaux, Alexandre Gramfort, Vincent Michel, Bertrand Thirion, Olivier Grisel, Mathieu Blondel, Peter Prettenhofer, Ron Weiss, Vincent Dubourg, et al. Scikit-learn: Machine learning in python. the Journal of machine learning research, 12:2825-2830, 2011. +Moshe Shechner, Or Sheffet, and Uri Stemmer. Private k-means clustering with stability assumptions. In International Conference on Artificial Intelligence and Statistics, pages 2518-2528. PMLR, 2020. +Dennis Wei. A constant-factor bi-criteria approximation guarantee for k-means++. Advances in neural information processing systems, 29, 2016. + +# NeurIPS Paper Checklist + +# 1. Claims + +Question: Do the main claims made in the abstract and introduction accurately reflect the paper's contributions and scope? + +Answer: [Yes] + +Justification: The main claims made in the abstract and introduction accurately reflect the paper's contribution. + +Guidelines: + +- The answer NA means that the abstract and introduction do not include the claims made in the paper. +- The abstract and/or introduction should clearly state the claims made, including the contributions made in the paper and important assumptions and limitations. A No or NA answer to this question will not be perceived well by the reviewers. +- The claims made should match theoretical and experimental results, and reflect how much the results can be expected to generalize to other settings. +- It is fine to include aspirational goals as motivation as long as it is clear that these goals are not attained by the paper. + +# 2. Limitations + +Question: Does the paper discuss the limitations of the work performed by the authors? + +Answer: [Yes] + +Justification: The paper aims to seek an instance that separates the greedy k-means++ and k-means++. Thus, there are some assumptions for the instance, but the instance assumption captures the instance in reality. + +Guidelines: + +- The answer NA means that the paper has no limitation while the answer No means that the paper has limitations, but those are not discussed in the paper. +- The authors are encouraged to create a separate "Limitations" section in their paper. +- The paper should point out any strong assumptions and how robust the results are to violations of these assumptions (e.g., independence assumptions, noiseless settings, model well-specification, asymptotic approximations only holding locally). The authors should reflect on how these assumptions might be violated in practice and what the implications would be. +- The authors should reflect on the scope of the claims made, e.g., if the approach was only tested on a few datasets or with a few runs. In general, empirical results often depend on implicit assumptions, which should be articulated. +- The authors should reflect on the factors that influence the performance of the approach. For example, a facial recognition algorithm may perform poorly when image resolution is low or images are taken in low lighting. Or a speech-to-text system might not be used reliably to provide closed captions for online lectures because it fails to handle technical jargon. +- The authors should discuss the computational efficiency of the proposed algorithms and how they scale with dataset size. +- If applicable, the authors should discuss possible limitations of their approach to address problems of privacy and fairness. +- While the authors might fear that complete honesty about limitations might be used by reviewers as grounds for rejection, a worse outcome might be that reviewers discover limitations that aren't acknowledged in the paper. The authors should use their best judgment and recognize that individual actions in favor of transparency play an important role in developing norms that preserve the integrity of the community. Reviewers will be specifically instructed to not penalize honesty concerning limitations. + +# 3. Theory assumptions and proofs + +Question: For each theoretical result, does the paper provide the full set of assumptions and a complete (and correct) proof? + +# Answer: [Yes] + +Justification: The main contribution of this paper is theoretical. All proofs and assumptions are explicitly described in the paper. + +# Guidelines: + +- The answer NA means that the paper does not include theoretical results. +- All the theorems, formulas, and proofs in the paper should be numbered and cross-referenced. +- All assumptions should be clearly stated or referenced in the statement of any theorems. +- The proofs can either appear in the main paper or the supplemental material, but if they appear in the supplemental material, the authors are encouraged to provide a short proof sketch to provide intuition. +- Inversely, any informal proof provided in the core of the paper should be complemented by formal proofs provided in appendix or supplemental material. +- Theorems and Lemmas that the proof relies upon should be properly referenced. + +# 4. Experimental result reproducibility + +Question: Does the paper fully disclose all the information needed to reproduce the main experimental results of the paper to the extent that it affects the main claims and/or conclusions of the paper (regardless of whether the code and data are provided or not)? + +# Answer: [Yes] + +Justification: The paper includes the experiments. The experiment aims to verify the theoretical findings of the paper. All required information for reproducibility is provided in the main text and supplemental material. + +# Guidelines: + +- The answer NA means that the paper does not include experiments. +- If the paper includes experiments, a No answer to this question will not be perceived well by the reviewers: Making the paper reproducible is important, regardless of whether the code and data are provided or not. +- If the contribution is a dataset and/or model, the authors should describe the steps taken to make their results reproducible or verifiable. +- Depending on the contribution, reproducibility can be accomplished in various ways. For example, if the contribution is a novel architecture, describing the architecture fully might suffice, or if the contribution is a specific model and empirical evaluation, it may be necessary to either make it possible for others to replicate the model with the same dataset, or provide access to the model. In general, releasing code and data is often one good way to accomplish this, but reproducibility can also be provided via detailed instructions for how to replicate the results, access to a hosted model (e.g., in the case of a large language model), releasing of a model checkpoint, or other means that are appropriate to the research performed. +- While NeurIPS does not require releasing code, the conference does require all submissions to provide some reasonable avenue for reproducibility, which may depend on the nature of the contribution. For example +(a) If the contribution is primarily a new algorithm, the paper should make it clear how to reproduce that algorithm. +(b) If the contribution is primarily a new model architecture, the paper should describe the architecture clearly and fully. +(c) If the contribution is a new model (e.g., a large language model), then there should either be a way to access this model for reproducing the results or a way to reproduce the model (e.g., with an open-source dataset or instructions for how to construct the dataset). +(d) We recognize that reproducibility may be tricky in some cases, in which case authors are welcome to describe the particular way they provide for reproducibility. In the case of closed-source models, it may be that access to the model is limited in some way (e.g., to registered users), but it should be possible for other researchers to have some path to reproducing or verifying the results. + +# 5. Open access to data and code + +Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material? + +Answer: [Yes] + +Justification: The submission includes the source code of the experiment in supplemental material. + +Guidelines: + +- The answer NA means that paper does not include experiments requiring code. +- Please see the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details. +- While we encourage the release of code and data, we understand that this might not be possible, so "No" is an acceptable answer. Papers cannot be rejected simply for not including code, unless this is central to the contribution (e.g., for a new open-source benchmark). +- The instructions should contain the exact command and environment needed to run to reproduce the results. See the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details. +- The authors should provide instructions on data access and preparation, including how to access the raw data, preprocessed data, intermediate data, and generated data, etc. +- The authors should provide scripts to reproduce all experimental results for the new proposed method and baselines. If only a subset of experiments are reproducible, they should state which ones are omitted from the script and why. +- At submission time, to preserve anonymity, the authors should release anonymized versions (if applicable). +- Providing as much information as possible in supplemental material (appended to the paper) is recommended, but including URLs to data and code is permitted. + +# 6. Experimental setting/details + +Question: Does the paper specify all the training and test details (e.g., data splits, hyperparameters, how they were chosen, type of optimizer, etc.) necessary to understand the results? + +Answer: [Yes] + +Justification: The paper includes the experiments. All settings of the experiment are explicitly described in the experiment section. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The experimental setting should be presented in the core of the paper to a level of detail that is necessary to appreciate the results and make sense of them. +- The full details can be provided either with the code, in appendix, or as supplemental material. + +# 7. Experiment statistical significance + +Question: Does the paper report error bars suitably and correctly defined or other appropriate information about the statistical significance of the experiments? + +Answer: [Yes] + +Justification: The experiment result includes different values of parameters. The statistical result is shown in a figure. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The authors should answer "Yes" if the results are accompanied by error bars, confidence intervals, or statistical significance tests, at least for the experiments that support the main claims of the paper. + +- The factors of variability that the error bars are capturing should be clearly stated (for example, train/test split, initialization, random drawing of some parameter, or overall run with given experimental conditions). +- The method for calculating the error bars should be explained (closed form formula call to a library function, bootstrap, etc.) +- The assumptions made should be given (e.g., Normally distributed errors). +- It should be clear whether the error bar is the standard deviation or the standard error of the mean. +- It is OK to report 1-sigma error bars, but one should state it. The authors should preferably report a 2-sigma error bar than state that they have a $96\%$ CI, if the hypothesis of Normality of errors is not verified. +- For asymmetric distributions, the authors should be careful not to show in tables or figures symmetric error bars that would yield results that are out of range (e.g. negative error rates). +- If error bars are reported in tables or plots, The authors should explain in the text how they were calculated and reference the corresponding figures or tables in the text. + +# 8. Experiments compute resources + +Question: For each experiment, does the paper provide sufficient information on the computer resources (type of compute workers, memory, time of execution) needed to reproduce the experiments? + +Answer: [Yes] + +Justification: All experimental details are described in the Experiment section. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The paper should indicate the type of compute workers CPU or GPU, internal cluster or cloud provider, including relevant memory and storage. +- The paper should provide the amount of compute required for each of the individual experimental runs as well as estimate the total compute. +- The paper should disclose whether the full research project required more compute than the experiments reported in the paper (e.g., preliminary or failed experiments that didn't make it into the paper). + +# 9. Code of ethics + +Question: Does the research conducted in the paper conform, in every respect, with the NeurIPS Code of Ethics https://neurips.cc/public/EthicsGuidelines? + +Answer: [Yes] + +Justification: The submission respects the NeurIPS code of ethics. The submission is theoretical work and there is no ethics issue. + +Guidelines: + +The answer NA means that the authors have not reviewed the NeurIPS Code of Ethics +- If the authors answer No, they should explain the special circumstances that require a deviation from the Code of Ethics. +- The authors should make sure to preserve anonymity (e.g., if there is a special consideration due to laws or regulations in their jurisdiction). + +# 10. Broader impacts + +Question: Does the paper discuss both potential positive societal impacts and negative societal impacts of the work performed? + +Answer: [NA] + +Justification: This is a theoretical work and there is no societal impact. + +Guidelines: + +- The answer NA means that there is no societal impact of the work performed. + +- If the authors answer NA or No, they should explain why their work has no societal impact or why the paper does not address societal impact. +- Examples of negative societal impacts include potential malicious or unintended uses (e.g., disinformation, generating fake profiles, surveillance), fairness considerations (e.g., deployment of technologies that could make decisions that unfairly impact specific groups), privacy considerations, and security considerations. +- The conference expects that many papers will be foundational research and not tied to particular applications, let alone deployments. However, if there is a direct path to any negative applications, the authors should point it out. For example, it is legitimate to point out that an improvement in the quality of generative models could be used to generate deepfakes for disinformation. On the other hand, it is not needed to point out that a generic algorithm for optimizing neural networks could enable people to train models that generate Deepfakes faster. +- The authors should consider possible harms that could arise when the technology is being used as intended and functioning correctly, harms that could arise when the technology is being used as intended but gives incorrect results, and harms following from (intentional or unintentional) misuse of the technology. +- If there are negative societal impacts, the authors could also discuss possible mitigation strategies (e.g., gated release of models, providing defenses in addition to attacks, mechanisms for monitoring misuse, mechanisms to monitor how a system learns from feedback over time, improving the efficiency and accessibility of ML). + +# 11. Safeguards + +Question: Does the paper describe safeguards that have been put in place for responsible release of data or models that have a high risk for misuse (e.g., pretrained language models, image generators, or scraped datasets)? + +Answer: [NA] + +Justification: This is a theoretical submission, and there is no safeguard issue. + +Guidelines: + +- The answer NA means that the paper poses no such risks. +- Released models that have a high risk for misuse or dual-use should be released with necessary safeguards to allow for controlled use of the model, for example by requiring that users adhere to usage guidelines or restrictions to access the model or implementing safety filters. +- Datasets that have been scraped from the Internet could pose safety risks. The authors should describe how they avoided releasing unsafe images. +- We recognize that providing effective safeguards is challenging, and many papers do not require this, but we encourage authors to take this into account and make a best faith effort. + +# 12. Licenses for existing assets + +Question: Are the creators or original owners of assets (e.g., code, data, models), used in the paper, properly credited and are the license and terms of use explicitly mentioned and properly respected? + +Answer: [NA] + +Justification: The submission does not use existing assets. + +Guidelines: + +- The answer NA means that the paper does not use existing assets. +- The authors should cite the original paper that produced the code package or dataset. +- The authors should state which version of the asset is used and, if possible, include a URI. +- The name of the license (e.g., CC-BY 4.0) should be included for each asset. +- For scraped data from a particular source (e.g., website), the copyright and terms of service of that source should be provided. + +- If assets are released, the license, copyright information, and terms of use in the package should be provided. For popular datasets, paperswithcode.com/datasets has curated licenses for some datasets. Their licensing guide can help determine the license of a dataset. +- For existing datasets that are re-packaged, both the original license and the license of the derived asset (if it has changed) should be provided. +- If this information is not available online, the authors are encouraged to reach out to the asset's creators. + +# 13. New assets + +Question: Are new assets introduced in the paper well documented and is the documentation provided alongside the assets? + +Answer: [NA] + +Justification: The submission does not release new assets. + +Guidelines: + +- The answer NA means that the paper does not release new assets. +- Researchers should communicate the details of the dataset/code/model as part of their submissions via structured templates. This includes details about training, license, limitations, etc. +- The paper should discuss whether and how consent was obtained from people whose asset is used. +- At submission time, remember to anonymize your assets (if applicable). You can either create an anonymized URL or include an anonymized zip file. + +# 14. Crowdsourcing and research with human subjects + +Question: For crowdsourcing experiments and research with human subjects, does the paper include the full text of instructions given to participants and screenshots, if applicable, as well as details about compensation (if any)? + +Answer: [NA] + +Justification: The submission does not involve crowdsourcing or research with human subjects. + +Guidelines: + +- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. +- Including this information in the supplemental material is fine, but if the main contribution of the paper involves human subjects, then as much detail as possible should be included in the main paper. +- According to the NeurIPS Code of Ethics, workers involved in data collection, curation, or other labor should be paid at least the minimum wage in the country of the data collector. + +# 15. Institutional review board (IRB) approvals or equivalent for research with human subjects + +Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals (or an equivalent approval/review based on the requirements of your country or institution) were obtained? + +Answer: [NA] + +Justification: The submission does not involve crowdsourcing nor research with human subjects. + +Guidelines: + +- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. + +- We recognize that the procedures for this may vary significantly between institutions and locations, and we expect authors to adhere to the NeurIPS Code of Ethics and the guidelines for their institution. +- For initial submissions, do not include any information that would break anonymity (if applicable), such as the institution conducting the review. + +# 16. Declaration of LLM usage + +Question: Does the paper describe the usage of LLMs if it is an important, original, or non-standard component of the core methods in this research? Note that if the LLM is used only for writing, editing, or formatting purposes and does not impact the core methodology, scientific rigorousness, or originality of the research, declaration is not required. + +Answer: [NA] + +Justification: The submission does not use any LLMs. + +Guidelines: + +- The answer NA means that the core method development in this research does not involve LLMs as any important, original, or non-standard components. +- Please refer to our LLM policy (https://neurips.cc/Conferences/2025/LLM) for what should or should not be described. + +# A Missing Proofs in Section 3 + +Proof of Lemma 1: To show the lemma for both cases when $\mathcal{A}$ is the greedy and $k$ -means++, assume that the greedy samples $\ell$ candidate points per iteration, where $\ell \in \{1,2,\dots,\ln k\}$ . Clearly, it captures both the greedy that samples $\ell$ candidates per iteration and $k$ -means++ that samples exactly one candidate per iteration. For an arbitrary $r \geq \delta$ , let $\mathcal{A}(r)$ denote the subevent of $\mathcal{A}$ in which the farthest distance between any sampled candidate and its corresponding cluster center in greedy $k$ -means++ is exactly $r$ . Clearly, $\operatorname*{Pr}[\mathcal{A}] = \int_{4b(1 + \theta)\ln k}^{\infty}\operatorname*{Pr}[\mathcal{A}(r)]\mathrm{d}r$ . + +We now analyze $\operatorname*{Pr}[\mathcal{A}(r)]$ . To this end, we partition $\mathcal{A}(r)$ into $k$ subevents $\{\mathcal{A}^{(t)}(r)\}_{t\in [k]}$ , based on the iteration in which a candidate point at distance $r$ from its cluster center is first sampled. Specifically, $\mathcal{A}^{(t)}(r)$ denotes the subevent in which such a point is sampled for the first time in iteration $t$ . + +We next prove an upper bound on each $\operatorname*{Pr}[\mathcal{A}^{(t)}(r)]$ . By the definition of the event, all centers selected by the algorithm before iteration $t$ , i.e., the points in $S_{t-1}$ , must lie within a distance less than $r$ from their respective cluster centers. Consider all points in $\bigcup_{i \in [k]} C_i(r)$ . By the triangle inequality, the distance from any such point to any selected center in $S_{t-1}$ is at most $d + 2r$ . Therefore, the total connection cost of these points to $S_{t-1}$ is at most + +$$ +(d + 2 r) ^ {2} \cdot \frac {1}{b} \cdot e ^ {- r / b} \cdot n. +$$ + +According to Observation 1, the total connection cost of all points to $S_{t-1}$ is at least $2b^{2} \cdot n$ . Since $\ell \leq \ln k$ , from union bounds, we have + +$$ +\Pr \left[ \mathcal {A} ^ {(t)} (r) \right] \leq \ln k \cdot \frac {(d + 2 r) ^ {2} \cdot e ^ {- r / b}}{2 b ^ {3}}. +$$ + +Therefore, + +$$ +\begin{array}{l} \Pr [ \mathcal {A} ] = \int_ {4 b (1 + \theta) \ln k} ^ {\infty} \Pr [ \mathcal {A} (r) ] d r \\ \leq \int_ {4 b (1 + \theta) \ln k} ^ {\infty} k \ln k \cdot \frac {(d + 2 r) ^ {2} \cdot e ^ {- r / b}}{2 b ^ {3}} d r \\ \end{array} +$$ + +(Sum of the upper bounds over the $k$ subevents) + +$$ +\begin{array}{l} \leq \int_ {4 b (1 + \theta) \ln k} ^ {\infty} k ^ {1 + 2 \theta} \ln k \cdot \frac {1 6 r ^ {2} \cdot e ^ {- r / b}}{b} d r \\ \leq 1 6 \cdot k ^ {1 + 2 \theta} \ln k \cdot (2 b + 4 b (1 + \theta) \ln k) ^ {2} \cdot e ^ {- 4 (1 + \theta) \ln k}, \\ \end{array} +$$ + +which is $o(1 / k)$ asymptotically. Similarly, for the contribution of event $\mathcal{A}$ to the expected objective, we have + +$$ +\begin{array}{l} \Pr [ \mathcal {A} ] \cdot \mathbb {E} [ \mathrm {O B J} \mid \mathcal {A} ] = \int_ {4 b (1 + \theta) \ln k} ^ {\infty} \Pr [ \mathcal {A} (r) ] \cdot \mathbb {E} [ \mathrm {O B J} \mid \mathcal {A} (r) ] \mathrm {d} r \\ \leq \int_ {4 b (1 + \theta) \ln k} ^ {\infty} \Pr [ \mathcal {A} (r) ] \cdot n \cdot \left(2 b ^ {2} + (d + 2 r) ^ {2}\right) d r \quad (\text {O b s e r v a t i o n 2}) \\ \leq \int_ {4 b (1 + \theta) \ln k} ^ {\infty} n \cdot (2 b ^ {2} + (d + r) ^ {2}) \cdot k \ln k \cdot \frac {(d + r) ^ {2} \cdot e ^ {- r / b}}{2 b ^ {3}} \mathrm {d} r, \\ \end{array} +$$ + +which is bounded by $o(n / k)$ asymptotically. + +Proof of Lemma 2: The first argument (2a) follows directly. By the concentration assumption, for any uncovered cluster, the distance from its center $\mu_{i}$ to solution $S$ is $\Omega(k^{\theta})$ , which yields a connection cost of $\Omega(k^{2\theta} \cdot |C_i|)$ by Observation 2. Similarly, since the distance from $\mu_{i}$ to solution $S$ is $O(\ln k)$ for covered clusters, the first part of the second argument follows. It remains to show the second part. + +Consider the first iteration $t$ in which the algorithm covers cluster $C_i$ . In this iteration, the algorithm samples a set of candidate points from $\sigma(C_i)$ . Suppose that $p$ such points are sampled, denoted by $P = \{x_1, \ldots, x_p\}$ , where $1 \leq p \leq \ln k$ . We will show that the expected maximum connection cost from the cluster center $\mu_i$ to the sampled candidates is $O((\ln p)^2)$ . Then, by Observation 2, the expected total connection cost for cluster $C_i$ is also bounded by $O((\ln p)^2)$ . Since $p \leq \ln k$ , this implies a bound of $O((\ln \ln k)^2)$ . + +As $t$ is the first iteration in which $C_i$ is covered, all centers selected before this iteration, i.e., the points in $S_{t-1}$ , are far from the points in $\sigma(C_i)$ , with distances lying in the range $[d - 2\delta, d + 2\delta]$ . Recalling that $d = \Theta(k^\theta)$ and $\delta = \Theta(\ln k)$ , we have for any point $x \in \sigma(C_i)$ , the connection cost $\varphi(x, S_{t-1})$ differs from that of other points in $\sigma(C_i)$ by at most a small constant factor, especially when $k$ is large. Consequently, up to a constant-factor loss in expectation, we may assume that the algorithm samples $p$ points uniformly at random from $\sigma(C_i)$ . For notational simplicity, we omit the conditioning on the event that exactly $p$ candidates are sampled from $\sigma(C_i)$ in the expectation below. We have + +$$ +\begin{array}{l} \mathbb {E} \left[ \max _ {x \in P} \| x - \mu_ {i} \| _ {2} ^ {2} \right] \leq \int_ {0} ^ {\infty} \Pr \left[ \max _ {x \in P} \| x - \mu_ {i} \| _ {2} ^ {2} \geq r \right] d r \\ = \int_ {0} ^ {\infty} \Pr \left[ \max _ {x \in P} \| x - \mu_ {i} \| _ {2} \geq \sqrt {r} \right] d r \\ = \int_ {0} ^ {\infty} 2 h \cdot \Pr \left[ \max _ {x \in P} \| x - \mu_ {i} \| _ {2} \geq h \right] d h \\ = \int_ {0} ^ {\infty} 2 h \cdot \Pr [ \exists x \in P \text {s u c h t h a t} \| x - \mu_ {i} \| _ {2} \geq h ] d h \\ = \int_ {0} ^ {\infty} 2 h \cdot \left(1 - (1 - e ^ {- h / b}) ^ {p}\right) d h. \\ \end{array} +$$ + +A standard calculation can show that the expression above is asymptotically bounded by $O((\ln p)^2)$ . From the fact of the exponential distribution, it is straightforward to see that $\mathbb{E}\left[\max_{x\in P}\| x - \mu_i\| _2^2\right] = \Omega (1)$ . Thus, we have established the upper and lower bounds on $C_i$ 's expected connected cost when covered as claimed. This completes the proof of (2b). + +Proof of Claim 1: We shall use the following concentration bound: + +Theorem 4 (Hoeffding's inequality). Let $X_{1}, X_{2}, \ldots, X_{n}$ be independent random variables such that $a_{i} \leq X_{i} \leq b_{i}$ . Let $X = \sum_{i \in [n]} X_{i}$ and $\mu = \mathbb{E}[X]$ . For all $t > 0$ , we have + +$$ +\operatorname * {P r} [ X - \mu \geq t ] \leq \exp \Big (- \frac {2 t ^ {2}}{\sum_ {i \in [ n ]} (b _ {i} - a _ {i}) ^ {2}} \Big) +$$ + +Let random variable $Y_{t}$ denote the connection cost of a cluster covered in the $t$ -th iteration with $t \in \{1,2,\dots ,t_0\}$ . By (2b) of Lemma 2, we know that $Y_{t} \leq \rho_{1} \cdot (\ln k)^{2} \cdot \frac{n}{k}$ and $\mathbb{E}[\sum_{t \in [t_0]} Y_t] \leq \rho_2 \cdot (\ln \ln k)^2 \cdot \frac{n}{k} \cdot t_0$ where we are allowed to let $\rho_{1} = (4b(1 + \theta))^{2}$ , and $\rho_{2} = b^{2}$ . For simplicity, we denote $\mathbb{E}[\sum_{t \in [t_0]} Y_t]$ by $\mu$ and its upper bound $\rho_{2} \cdot (\ln \ln k)^{2} \cdot \frac{n}{k} \cdot t_{0}$ by $C$ . Applying Chernoff-Hoeffding's inequality (Theorem 4 in the appendix), we have + +$$ +\begin{array}{l} \Pr \left[ \sum_ {t \in [ t _ {0} ]} Y _ {t} \geq 2 \cdot C \right] \leq \Pr \left[ \sum_ {t \in [ t _ {0} ]} Y _ {t} \geq \mu + C \right] \leq \exp \left(- \frac {2 C ^ {2}}{\sum_ {t \in [ t _ {0} ]} \left(\rho_ {1} \cdot (\ln k) ^ {2} \cdot \frac {n}{k}\right) ^ {2}}\right) \\ \leq \exp \left(- 2 k \cdot \left(\frac {\rho_ {2} \ln \ln k}{\rho_ {1} \ln k}\right) ^ {2}\right) \\ \end{array} +$$ + +# B Missing Proofs in Section 4 + +Proof of Lemma 4: This proof follows a similar analysis to that in Lemma 2. By Observation 2, the total connection cost of a cluster $C_i$ is $(2b^2 + h^2) \cdot |C_i|$ , where $h$ denotes the distance from + +the cluster center $\mu_{i}$ to the current set of selected centers $S$ . When the cluster is not yet covered, $h \in [d - \delta, d + 2\delta]$ , which is of order $\Theta(k^{\theta})$ ; whereas once the cluster is covered, $b$ is still constant that implies the total connection cost is $\Omega(|C_i|)$ . This completes the proof of the lemma. + +# C Distributions with Sub-Exponential Tails + +In the main body, we present the results and detailed proofs under the assumption that each cluster follows an exponential distribution. However, this assumption is not essential—in fact, the same analysis applies to any subexponential-tailed distributions with a mild assumption. In an extreme case, all points in a cluster share the same location as the center. $k$ -means++ or the greedy can always pick a new uncovered cluster in every iteration. To avoid points being overly concentrated around the centers, we assume that the distributions satisfy $\operatorname{Pr}_{x \sim \mathcal{D}}[x \geq \epsilon] = \Omega(1)$ . Clearly, the exponential distribution has this property. Below, we provide three representative examples: Gaussian, sub-Gaussian, and sub-exponential distributions. Each of them exhibits tail probabilities that decay exponentially with distance from the center. + +# Gaussian Distribution + +- Parameters: Mean vector $b \in \mathbb{R}$ ,variance ${\sigma }^{2}$ . +- Tail bound: For $X \sim \mathcal{N}(b, \sigma^2)$ , + +$$ +\operatorname * {P r} \left(X - b \geq r\right) \leq \exp \left(- \frac {r ^ {2}}{2 \sigma^ {2}}\right), +$$ + +which decays exponentially in $r^2$ . + +# Sub-Gaussian Distribution + +Parameter: Mean $b\in \mathbb{R}$ , variance proxy parameter $\sigma^2$ +- Tail bound: For a sub-Gaussian random vector $X$ , + +$$ +\Pr \left(X - b \geq r\right) \leq \exp \left(- \frac {c r ^ {2}}{\sigma^ {2}}\right) +$$ + +for some absolute constant $c > 0$ , showing exponential decay in $r^2$ . + +# Sub-Exponential Distribution + +- Parameters: Mean $b \in \mathbb{R}$ ,sub-exponential parameters $\left( {\nu ,\alpha }\right)$ ,where ${\nu }^{2}$ reflects the variance-like behavior,and $\alpha$ controls the tail heaviness. +- Tail bound: For a sub-exponential random variable $X$ , + +$$ +\operatorname * {P r} \left(X - b \geq r\right) \leq \left\{ \begin{array}{l l} \exp \left(- \frac {r ^ {2}}{2 \nu^ {2}}\right), & \text {f o r} 0 \leq r \leq \frac {\nu^ {2}}{\alpha}, \\ \exp \left(- \frac {r}{2 \alpha}\right), & \text {f o r} r > \frac {\nu^ {2}}{\alpha}, \end{array} \right. +$$ + +showing sub-Gaussian decay for small $t$ and exponential decay for large $t$ . + +The sub-exponential distribution is a well-known generalization of the sub-Gaussian and the Gaussian distribution. For completeness, we will show the results in the main body under the sub-exponential distributions with a mild assumption. First, we restate our assumption. We denote the sub-exponential distribution by $\mathcal{D} = \mathrm{subE}(\nu, \alpha)$ . Each cluster has a different distribution $\mathcal{D}_i = \mathrm{subE}(\nu_i, \alpha_i)$ with a mean $b_i$ . Let $\lambda \geq 1$ be a universal constant which is the upper bound of all constant parameters. The input point set $X$ satisfying the three properties: + +- Sub-Exponentially Distributed: For each cluster $C_i$ , the density of points decreases subexponentially as the distance from the center increases. Specifically, + +$$ +\frac {\left| C _ {i} (r) \right|}{\left| C _ {i} \right|} \sim \operatorname {s u b E} \left(\nu_ {i}, \alpha_ {i}\right) +$$ + +- Well Separable: The minimum distance $d$ between any two centers of the optimal clusters is sufficiently large, i.e., $d = k^{\theta} \cdot \lambda$ , for a constant $\theta > 0$ . +- Well Spread: The optimal clusters are roughly homogeneous, with the number of points in each cluster and the distances between clusters differing by at most a constant factor. For the cluster $C_i$ , we denote the number of points by $n_i$ . Further, we have + +$$ +\operatorname * {P r} _ {x \sim \mathcal {D} _ {i}} \left[ x \geq \epsilon_ {i} \right] \geq p _ {i} \quad \forall i \in [ k ] +$$ + +where $\epsilon_0, p_0$ are some constant. + +The universal constant controls the effects from the heterogeneous parameters where $1 / \lambda \leq \nu_{i},\alpha_{i},b_{i},n_{i} / (n / k),\epsilon_{i},p_{i}\leq \lambda$ + +Theorem 5 (Corresponding to Theorem 1). Given any general EWW point set $X$ , the greedy $k$ -means++ algorithm admits an expected approximation ratio of $O((\log \log k)^2)$ . + +Similarly, we define the concentration balls as follows. + +Definition 2 (Corresponding to Definition 1). We define the concentration radius $\delta \coloneqq 4\lambda (1 + \theta)\ln k$ where $\theta$ and $b$ are input parameters. For each cluster $C_i$ with center $\mu_{i}$ , we define its concentration ball $\sigma (C_i)$ as the set of all points in $C_i$ whose distance to $\mu_{i}$ is at most $\delta$ . + +Lemma 7 (Corresponding to Lemma 1). Let $\mathcal{A}$ denote the event that greedy $k$ -means++ samples at least one candidate point outside the concentration balls during any iteration. Given any general EWW point set $X$ , the probability that $\mathcal{A}$ occurs is at most $1 / k$ . Furthermore, the contribution of this event to the expected objective can be bounded: + +$$ +\Pr [ \mathcal {A} ] \cdot \mathbb {E} [ \mathrm {O B J} \mid \mathcal {A} ] \leq \frac {n}{k}, +$$ + +where $\mathbb{E}[\mathrm{OBJ}|\mathcal{A}]$ denotes the expected objective value (i.e., total connection cost) conditioned on the occurrence of $\mathcal{A}$ . + +Proof. To show the lemma for both cases when $\mathcal{A}$ is the greedy and $k$ -means++, assume that the greedy samples $\ell$ candidate points per iteration where $\ell \in [1, \log k]$ . Clearly, it captures both the greedy that samples $\ell$ candidates per iteration and $k$ -means++ that samples exactly one candidate per iteration. For an arbitrary $r \geq \delta$ , let $\mathcal{A}(r)$ denote the subevent of $\mathcal{A}$ in which the farthest distance between any sampled candidate and its corresponding cluster center in greedy $k$ -means++ is exactly $r$ . Clearly, $\operatorname*{Pr}[\mathcal{A}] = \int_{4\lambda(1+\theta)\ln k}^{\infty} \operatorname*{Pr}[\mathcal{A}(r)] \mathrm{d}r$ . + +We now analyze $\operatorname*{Pr}[\mathcal{A}(r)]$ . To this end, we partition $\mathcal{A}(r)$ into $k$ subevents $\{\mathcal{A}^{(t)}(r)\}_{t\in [k]}$ , based on the iteration in which a candidate point at distance $r$ from its cluster center is first sampled. Specifically, $\mathcal{A}^{(t)}(r)$ denotes the subevent in which such a point is sampled for the first time in iteration $t$ . + +We next prove an upper bound on each $\operatorname*{Pr}[\mathcal{A}^{(t)}(r)]$ . By the definition of the event, all centers selected by the algorithm before iteration $t$ , i.e., the points in $S_{t - 1}$ , must lie within a distance less than $r$ from their respective cluster centers. Consider all points in $\bigcup_{i\in [k]}C_i(r)$ . By the triangle inequality, the distance from any such point to any selected center in $S_{t - 1}$ is at most $d + 2r$ . Therefore, the total connection cost of these points to $S_{t - 1}$ for a cluster $i$ is at most + +$$ +(d + 2 r) ^ {2} \cdot e ^ {- r / (2 \alpha_ {i})} \cdot n _ {i} \leq (d + 2 r) ^ {2} \cdot e ^ {- r / (2 \lambda)} \cdot \lambda n / k. +$$ + +From the well-spread property of the dataset, the total connection cost of all points to $S_{t-1}$ for a cluster $i$ is at least $e_i^2 p_i \cdot n_i \geq (n/k)/\lambda^4$ . Since $\ell \leq \log k$ , from union bounds, we have + +$$ +\Pr [ \mathcal {A} ^ {(t)} (r) ] \leq \log k \cdot \frac {(d + 2 r) ^ {2} \cdot e ^ {- r / (2 \lambda)} \cdot \lambda}{1 / \lambda^ {4}}. +$$ + +Therefore, + +$$ +\Pr [ \mathcal {A} ] = \int_ {4 \lambda (1 + \theta) \ln k} ^ {\infty} \Pr [ \mathcal {A} (r) ] d r +$$ + +$$ +\leq \int_ {4 \lambda (1 + \theta) \ln k} ^ {\infty} k \log k \cdot \lambda^ {5} (d + 2 r) ^ {2} \cdot e ^ {- r / (2 \lambda)} d r +$$ + +(Sum of the upper bounds over the $k$ subevents) + +$$ +\begin{array}{l} \leq \int_ {4 \lambda (1 + \theta) \ln k} ^ {\infty} 1 6 \lambda^ {7} k ^ {1 + 2 \theta} \log k \cdot r ^ {2} \cdot e ^ {- r / (2 \lambda)} d r \\ \leq 1 2 8 \cdot \lambda^ {1 0} k ^ {1 + 2 \theta} \log k \cdot (2 \lambda + 4 \lambda (1 + \theta) \ln k) ^ {2} \cdot e ^ {- 4 \lambda (1 + \theta) \ln k}, \\ \end{array} +$$ + +which is $o(1 / k)$ asymptotically. Similarly, for the contribution of event $\mathcal{A}$ to the expected objective, we have + +$$ +\begin{array}{l} \Pr [ \mathcal {A} ] \cdot \mathbb {E} [ \mathrm {O B J} \mid \mathcal {A} ] = \int_ {4 \lambda (1 + \theta) \ln k} ^ {\infty} \Pr [ \mathcal {A} (r) ] \cdot \mathbb {E} [ \mathrm {O B J} \mid \mathcal {A} (r) ] \mathrm {d} r \\ \leq \int_ {4 \lambda (1 + \theta) \ln k} ^ {\infty} \Pr [ \mathcal {A} (r) ] \cdot n \cdot \left(2 \lambda^ {2} + (d + 2 r) ^ {2}\right) d r \quad (\text {O b s e r v a t i o n 2}) \\ \leq \int_ {4 \lambda (1 + \theta) \ln k} ^ {\infty} n \cdot (2 \lambda^ {2} + (d + r) ^ {2}) \cdot k \log k \cdot \lambda^ {5} (d + 2 r) ^ {2} \cdot e ^ {- r / (2 \lambda)} d r, \\ \end{array} +$$ + +which is bounded by $o(n / k)$ asymptotically. + +![](images/6fd2fc4eba04bf547cafb26c4aff4fa288d007bdc3bfa16be8abfbbe52b9b7be.jpg) + +Lemma 8 (Corresponding to Lemma 2). Given any general EWW point set $X$ , under the concentration assumption, for each cluster $C_i$ , we have: + +(8a) If greedy $k$ -means++ does not cover this cluster, then its total connection cost is $\Omega (k^{2\theta} \cdot |C_i|)$ . +(8b) If greedy $k$ -means++ covers this cluster, then the total connection cost for $C_i$ does not exceed $O((\ln k)^2 \cdot |C_i|)$ , and further, its expectation is $O((\ln \ln k)^2 \cdot |C_i|)$ and $\Omega(|C_i|)$ . + +Proof. The first argument (8a) follows directly. By the concentration assumption, for any uncovered cluster, the distance from its center $\mu_{i}$ to solution $S$ is $\Omega(k^{\theta})$ , which yields a connection cost of $\Omega(k^{2\theta} \cdot |C_i|)$ by Observation 2. Similarly, since the distance from $\mu_{i}$ to solution $S$ is $O(\ln k)$ for covered clusters, the first part of the second argument follows. It remains to show the second part. + +We apply the same strategy in Lemma 2. As $t$ is the first iteration in which $C_i$ is covered, all centers selected before this iteration, i.e., the points in $S_{t-1}$ , are far from the points in $\sigma(C_i)$ , with distances lying in the range $[d - 2\delta, d + 2\delta]$ . Recalling that $d = \Theta(k^\theta)$ and $\delta = 4\lambda(1 + \theta)\ln k$ , we have for any point $x \in \sigma(C_i)$ , the connection cost $\varphi(x, S_{t-1})$ differs from that of other points in $\sigma(C_i)$ by at most a small constant factor, especially when $k$ is large. Consequently, up to a constant-factor loss in expectation, we may assume that the algorithm samples $p$ points uniformly at random from $\sigma(C_i)$ . For notational simplicity, we omit the conditioning on the event that exactly $p$ candidates are sampled from $\sigma(C_i)$ in the expectation below. We have + +$$ +\begin{array}{l} \mathbb {E} \left[ \max _ {x \in P} \| x - \mu_ {i} \| _ {2} ^ {2} \right] \leq \int_ {0} ^ {\infty} \Pr \left[ \max _ {x \in P} \| x - \mu_ {i} \| _ {2} ^ {2} \geq r \right] \mathrm {d} r \\ = \int_ {0} ^ {\infty} \Pr \left[ \max _ {x \in P} \| x - \mu_ {i} \| _ {2} \geq \sqrt {r} \right] d r \\ \leq \int_ {0} ^ {(\lambda \ln p) ^ {2}} 1 d h + \int_ {(\lambda \ln p) ^ {2}} ^ {\infty} 2 h \cdot \Pr \left[ \max _ {x \in P} \| x - \mu_ {i} \| _ {2} \geq h \right] d h \\ = (\lambda \ln p) ^ {2} + \int_ {(\lambda \ln p) ^ {2}} ^ {\infty} 2 h \cdot \Pr [ \exists x \in P \text {s u c h t h a t} \| x - \mu_ {i} \| _ {2} \geq h ] d h \\ \leq (\lambda \ln p) ^ {2} + \int_ {(\lambda \ln p) ^ {2}} ^ {\infty} 2 h \cdot \ln p \cdot e ^ {- h / (2 \lambda)} d h \\ \end{array} +$$ + +A standard calculation can show that the expression above is asymptotically bounded by $O((\ln p)^2)$ . From the well-spread property, it is straightforward to see that $\mathbb{E}\left[\max_{x\in P}\| x - \mu_i\| _2^2\right] = \Omega (1)$ . Thus, we have established the upper and lower bounds on $C_i$ 's expected connected cost when covered as claimed. This completes the proof. + +Lemma 9 (Corresponding to Lemma 3). Given any general EWW point set $X$ , in each iteration $t \leq k - O(k^{1 - 2\theta} \cdot (\log \log k)^2)$ , the greedy $k$ -means++ algorithm covers a new cluster with probability at least $1 - 1 / k^{2\theta + 2}$ . + +Proof. Since we have Lemma 8 for sub-exponential distributions, using the same proof for Lemma 2, we could have the lemma. For the sake of completeness, we provide the following proof with a slightly different constant factor. We first leverage (8a) of Lemma 8 to show that, with high probability, the greedy $k$ -means++ algorithm covers a new cluster in each iteration up to iteration $k - O(k^{1 - 2\theta} \cdot (\ln k)^2)$ . Then, by applying (8b) of Lemma 8 along with the Chernoff bound, we demonstrate that the algorithm continues to cover a new cluster in each iteration, with high probability, even during the range of iterations from $k - O(k^{1 - 2\theta} \cdot (\ln k)^2)$ to $k - O(k^{1 - 2\theta} \cdot (\ln \ln k)^2)$ . + +First, consider an iteration $k \leq t_0 \coloneqq k - 2^{2\theta + 5} \cdot k^{1 - 2\theta} \cdot (\ln k)^2$ . Lemma 8 implies that the greedy comparison step always prefers points from uncovered concentration balls over those from already covered ones. This is because if the greedy selects a point from an uncovered concentration ball (by (8a)), it decreases the objective by $\Omega(n / k)k^{2\theta}$ while if it selects from a covered concentration ball, it only decreases the objective by $O(n / k)(\ln k)^2$ (by (8b)). Thus, it follows that, in iteration $t$ , if at least one of the $\ln k$ sampled candidates is from an uncovered concentration ball, then the algorithm is guaranteed to cover a new cluster in that iteration. Suppose that $p$ clusters have already been covered by $S_{t-1}$ at the beginning of iteration $t$ (so $p < t$ ). By (8b) of Lemma 8 and the well-spread property, the total connection cost contributed by the covered clusters is at most $\lambda \cdot \frac{n}{k} \cdot p \cdot (4\lambda(1 + \theta)\ln k)^2$ , while the total connection cost from the uncovered clusters is at least $1 / \lambda \cdot \frac{n}{k} \cdot (k - p) \cdot \lambda^2 k^{2\theta}$ , therefore, the probability5 that the algorithm fails to sample any point from the uncovered concentration balls is at most + +$$ +\left(\frac {\lambda p \cdot (4 \lambda (1 + \theta) \ln k) ^ {2}}{\lambda p \cdot (4 \lambda (1 + \theta) \ln k) ^ {2} + (k - p) \cdot \lambda k ^ {2 \theta}}\right) ^ {\ln k} \leq 1 / k ^ {2 \theta + 5} \text {w h e n} p \leq t _ {0}. +$$ + +Next, we analyze the time period between iteration $t_0$ and $t_1 \coloneqq k - 2^{2\theta + 5} \cdot k^{1 - 2\theta} \cdot (\ln \ln k)^2$ . From the earlier analysis and by applying a union bound, we know that with probability at least $1 - 1 / k^{2\theta + 4}$ , the algorithm has already covered $t_0$ clusters within the first $t_0$ iterations. Then, according to Lemma 8, conditioned on this event, the expected connection cost of these covered clusters is lower bounded by $\Omega \left( \frac{n}{k} \cdot t_0 \right)$ , which is asymptotically much larger than the upper bound on the connection cost for any single cluster, $O \left( \frac{n}{k} \cdot (\ln k)^2 \right)$ . This suggests the connection cost of the covered clusters is well-concentrated. + +We shall upper bound the total connection cost of the covered clusters in the first $t_0$ iterations via the concentration bound. Note that each cluster can be approximately treated as being sampled uniformly based on the analysis in the proof of Lemma 8, which enables us to use concentration inequalities. + +Let random variable $Y_{t}$ denote the connection cost of a cluster covered in the $t$ -th iteration with $t \in \{1, 2, \dots, t_0\}$ . By (8b) of Lemma 8, we know that $Y_{t} \leq \rho_{1} \cdot (\ln k)^{2} \cdot \frac{n}{k}$ and $\mathbb{E}[\sum_{t \in [t_0]} Y_t] \leq \rho_2 \cdot (\ln \ln k)^2 \cdot \frac{n}{k} \cdot t_0$ where we are allowed to let $\rho_1 = \lambda (4\lambda (1 + \theta))^2$ , and $\rho_2 = 4\lambda^3$ . For simplicity, we denote $\mathbb{E}[\sum_{t \in [t_0]} Y_t]$ by $\mu$ and its upper bound $\rho_2 \cdot (\ln \ln k)^2 \cdot \frac{n}{k} \cdot t_0$ by $C$ . Applying Chernof-Hoeffding's inequality (Theorem 4 in the appendix), we have + +$$ +\begin{array}{l} \Pr \left[ \sum_ {t \in [ t _ {0} ]} Y _ {t} \geq 2 \cdot C \right] \leq \Pr \left[ \sum_ {t \in [ t _ {0} ]} Y _ {t} \geq \mu + C \right] \leq \exp \left(- \frac {2 C ^ {2}}{\sum_ {t \in [ t _ {0} ]} (\rho_ {1} \cdot (\ln k) ^ {2} \cdot \frac {n}{k}) ^ {2}}\right) \\ \leq \exp \left(- 2 k \cdot \left(\frac {\rho_ {2} \ln \ln k}{\rho_ {1} \ln k}\right) ^ {2}\right) \\ \end{array} +$$ + +Thus, with probability at least $1 - \exp \left(\Theta (1)\cdot (-k)\cdot \left(\frac{\ln\ln k}{\ln k}\right)^2\right)$ , the total connection cost of the clusters covered in the first $t_0$ iterations is at most $4\lambda^3\cdot \frac{n}{k}\cdot t_0\cdot (\ln \ln k)^2$ . + +Consider any iteration $t \in (t_0, t_1]$ . The algorithm can cover at most $t_1 - t_0$ new clusters during the interval from iteration $t_0$ to $t_1$ , and each newly covered cluster contributes at most $O\left(\frac{n}{k} \cdot (\ln k)^2\right)$ to + +the connection cost. Thus, with a probability of at least + +$$ +\left. \left(1 - \frac {1}{k ^ {2 \theta + 4}}\right) \left(1 - \exp \left(\Theta (1) \cdot (- k) \cdot \left(\frac {\ln \ln k}{\ln k}\right) ^ {2}\right)\right), \right. +$$ + +the total connection cost of the already covered clusters at the beginning of iteration $t$ is at most + +$$ +A := 1 6 \lambda^ {3} (1 + \theta) ^ {2} \cdot \frac {n}{k} \cdot \left(b ^ {2} \cdot t _ {0} \cdot (\ln \ln k) ^ {2} + (t _ {1} - t _ {0}) \cdot (\ln k) ^ {2}\right). +$$ + +Note that the total connection cost from the uncovered clusters is at least $B \coloneqq \lambda_{k}^{n} \cdot (k - t_{0}) \cdot k^{2\theta}$ . Similar to the analysis in the previous case, the probability that the algorithm fails to sample any point from the uncovered concentration balls is at most $(A / (A + B))^{\ln k}$ . This implies that the probability that greedy $k$ -means++ fails to cover a new cluster is at most $1 / k^{2\theta + 2}$ . + +Proof of Theorem 5. By Lemma 9 and a union bound, we have that, with probability at least $1 - 1 / k^{2\theta +1}$ , the greedy $k$ -means++ algorithm covers $k - O\left(k^{1 - 2\theta}\cdot (\ln \ln k)^{2}\right)$ clusters and achieves an approximation ratio of $O((\ln \ln k)^2)$ . If this case does not occur, we can simply upper bound the objective by $O(n\cdot k^{2\theta})$ . Taking expectation over both cases, we obtain an expected approximation ratio of $O((\ln \ln k)^2)$ . + +Theorem 6 (Corresponding to Theorem 2). Given any general EWW point set $X$ with parameter $\theta \in (0,1/2]$ , the $k$ -means++ algorithm admits an expected approximation ratio of $\Omega (\log k)$ . + +Lemma 10 (Corresponding to Lemma 4). Given any general EWW point set $X$ , under the concentration assumption, for each cluster $C_i$ , we have: + +- If $k$ -means++ does not cover this cluster, then its total connection cost is $\Theta (k^{2\theta} \cdot |C_i|)$ . +- If $k$ -means++ covers this cluster using exactly one center—that is, the final solution includes exactly one point from $\sigma(C_i)$ —then the total connection cost for $C_i$ is $\Omega(|C_i|)$ . + +Proof. By Observation 2, the total connection cost of a cluster $C_i$ is $(2\lambda^2 + h^2) \cdot |C_i|$ , where $h$ denotes the distance from the cluster center $\mu_i$ to the current set of selected centers $S$ . When the cluster is not yet covered, $h \in [d - \delta, d + 2\delta]$ , here $d = \lambda k^\theta$ , which is of order $\Theta(k^\theta)$ ; whereas once the cluster is covered, the well-spread property, where $\mathrm{Pr}_{x \sim \mathcal{D}_i}[x \geq \epsilon_i] \geq p_i \forall i \in [k]$ , implies the total connection cost is at least $|C_i| / \lambda^2 = \Omega(|C_i|)$ . This completes the proof of the lemma. + +Proof of Theorem 6. Lemma 10 shows that if the number of uncovered clusters is $p$ , then the objective value is at least $\lambda \frac{n}{k} \cdot p \cdot k^{2\theta}$ , omitting constant factors for simplicity. Therefore, to establish a lower bound of $\Omega(\ln k)$ on the approximation ratio, it suffices to show that, in expectation, $k$ -means++ leaves $\Omega(\ln k \cdot k^{1 - 2\theta})$ clusters uncovered. This also explains why we require $\theta \in (0, 1/2]$ : otherwise, $\ln k \cdot k^{1 - 2\theta}$ would be subconstant, rendering the argument meaningless. + +We partition all possible outcomes of the algorithm into two cases based on the number of uncovered clusters: (1) the final number of uncovered clusters is at least $\Delta$ , and (2) the final number of uncovered clusters is less than $\Delta$ , where $\Delta = \ln k \cdot k^{1 - 2\theta}$ . Clearly, in all outcomes falling into the first case, the approximation ratio is $\Omega(\ln k)$ . Next, we analyze the second case and show that, in expectation, the number of uncovered clusters remains $\Omega(\Delta)$ . + +Consider an arbitrary iteration $t$ . In the second case, where the final number of uncovered clusters is less than $\Delta$ , the number of clusters already covered by the solution $S_{t-1}$ at the beginning of this iteration must be at least $t - \Delta$ . Otherwise, even if every subsequent iteration covers a new cluster, the final number of uncovered clusters would exceed $\Delta$ , contradicting the assumption. Then, by the pigeonhole principle, at least $t - 2\Delta$ clusters must be covered by exactly one center—that is, $S_{t-1}$ contains exactly one point from each of these clusters. By Lemma 10 and the well-spread property, the total connection cost of the covered clusters is at least $1 / \lambda^3 \cdot \frac{n}{k} \cdot (t - 2\Delta)$ , while the total connection cost of the uncovered clusters is at most $\lambda^3 \cdot \frac{n}{k} \cdot (k - t + \Delta) \cdot k^{2\theta}$ . Therefore, in each iteration $t > 2\Delta$ , the probability that the $k$ -means++ algorithm fails to cover a new cluster is at least $1 / \lambda^6 \cdot \frac{t - 2\Delta}{(t - 2\Delta) + (k - t + \Delta) \cdot k^{2\theta}}$ . + +We compute the expected number of uncovered clusters by summing the failure probabilities across all iterations (conditioned on the second case). Specifically, we have: + +$$ +\begin{array}{l} \mathbb {E} [ \text {n u m b e r o f u n c o v e r e d c l u s t e r s} ] \geq 1 / \lambda^ {6} \cdot \sum_ {t > 2 \Delta} ^ {k} \frac {t - 2 \Delta}{(t - 2 \Delta) + (k - t + \Delta) \cdot k ^ {2 \theta}} \\ \geq 1 / \lambda^ {6} \cdot \sum_ {t \geq k / 2} ^ {k} \frac {t - 2 \Delta}{(t - 2 \Delta) + (k - t + \Delta) \cdot k ^ {2 \theta}} \geq 1 / \lambda^ {6} \cdot \sum_ {t \geq k / 2} ^ {k} \frac {k / 4}{k / 4 + (k - t + \Delta) \cdot k ^ {2 \theta}} (\Delta = o (k)) \\ = 1 / \lambda^ {6} \cdot k ^ {1 - 2 \theta} \cdot \sum_ {t \geq k / 2} ^ {k} \frac {1}{k ^ {- 2 \theta} + 4 (k - t + \Delta)} \geq 1 / \lambda^ {6} \cdot k ^ {1 - 2 \theta} \ln \left(\frac {k / 2 + \Delta}{\Delta}\right) \\ \geq 1 / \lambda^ {6} \cdot k ^ {1 - 2 \theta} \ln \left(\frac {k ^ {2 \theta}}{2 \ln k}\right) = \Omega (k ^ {1 - 2 \theta} \ln k), \\ \end{array} +$$ + +which implies that the expected number of uncovered clusters is $\Omega (\Delta)$ and completes the proof. + +Theorem 7 (Corresponding to Theorem 3). Given any general EWW point set with parameter $\theta > 1/2$ , the probability that greedy $k$ -means++ covers all optimal clusters is greater than that of $k$ -means++. + +Let event $\mathcal{B}$ denote the bad event that the algorithm selects a point from an already covered cluster. + +Lemma 11 (Corresponding to Lemma 5). Given any general EWW point set $X$ , the probability that $k$ -means++ encounters $\mathcal{B}$ is at least $\frac{k - 1}{k - 1 + k^{2\theta}}$ . + +Proof. We partition the bad event into two sub-events based on the time at which $\mathcal{B}$ first occurs: (1) $k$ -means++ encounters $\mathcal{B}$ before the last iteration $k$ , and (2) $k$ -means++ first encounters $\mathcal{B}$ at the last iteration $k$ . We denote these two sub-events as $\mathcal{P}$ and $\mathcal{Q}$ , respectively. By expanding the conditional probability of the second sub-event, we derive a lower bound on the probability that $k$ -means++ encounters $\mathcal{B}$ : + +$$ +\Pr [ k \text {- m e a n s + +} \text {e n c o u n t e r s} \mathcal {B} ] = \Pr [ \mathcal {P} ] + \Pr [ \mathcal {Q} ] = \Pr [ \mathcal {P} ] + \Pr [ \neg \mathcal {P} ] \cdot \Pr [ \mathcal {Q} | \neg \mathcal {P} ] \geq \Pr [ \mathcal {Q} | \neg \mathcal {P} ]. +$$ + +Conditioned on $\neg \mathcal{P}$ , the notion of "first" in event $\mathcal{Q}$ is not essential— $\operatorname*{Pr}[\mathcal{Q} \mid \neg \mathcal{P}]$ simply equals the probability that $k$ -means++ samples a point from one of the $k-1$ already covered clusters. By Lemma 10, the total connection cost of the already covered clusters is at least $\frac{n}{k} \cdot (k-1) / \lambda^3$ , while the total connection cost of the last uncovered cluster is at most $\frac{n}{k} \cdot \lambda^3 k^{2\theta}$ . Thus, the probability that $k$ -means++ encounters event $\mathcal{B}$ can be lower bounded by $1 / \lambda^6 \cdot \frac{k-1}{k-1 + k^{2\theta}}$ . + +Lemma 12 (Corresponding to Lemma 6). Given any general EWW point set $X$ , the probability that greedy $k$ -means++ encounters $\mathcal{B}$ is at most $\left(\frac{16e(1 + \theta)^2 \cdot (k - 1) \cdot (\log k)^2}{(k - 1) \cdot (\log k)^2 + k^{2\theta}}\right)^{\log k}$ . + +Proof. To upper bound the probability for greedy $k$ -means++, we partition the bad event into $k$ sub-events $\{\mathcal{P}_t\}_{t \in [k]}$ , where $\mathcal{P}_t$ denotes the event that greedy $k$ -means++ encounters $\mathcal{B}$ for the first time in iteration $t$ . Similarly, we then expand the probability of $\mathcal{P}_t$ using conditional probability: + +$$ +\begin{array}{l} \Pr [ \text {g r e e d y} k \text {- m e a n s + + e n c o u n t e r s} \mathcal {B} ] = \sum_ {t \in [ k ]} \Pr [ \mathcal {P} _ {t} ] \\ = \sum_ {t \in [ k ]} \Pr \left[ \neg \left(\mathcal {P} _ {1} \vee \dots \vee \mathcal {P} _ {t - 1}\right) \right] \cdot \Pr \left[ \mathcal {P} _ {t} \mid \neg \left(\mathcal {P} _ {1} \vee \dots \vee \mathcal {P} _ {t - 1}\right) \right] \leq k \cdot \Pr \left[ \mathcal {P} _ {k} \mid \neg \left(\mathcal {P} _ {1} \vee \dots \vee \mathcal {P} _ {k - 1}\right) \right] \\ \end{array} +$$ + +where the last inequality uses the fact that $\operatorname*{Pr}[\mathcal{P}_t\mid \neg (\mathcal{P}_1\lor \dots \lor \mathcal{P}_{t - 1})]$ attains its maximum at $t = k$ + +The term $\operatorname*{Pr}[\mathcal{P}_k\mid \neg (\mathcal{P}_1\lor \dots \lor \mathcal{P}_{k - 1})]$ simply equals the probability that greedy $k$ -means++ samples all $\ln k$ candidates from one of the $k - 1$ already uncovered clusters. By Lemma 8, the total connection cost of the already covered cluster is at most $\lambda \cdot \frac{n}{k}\cdot (k - 1)\cdot (4\lambda (1 + \theta)\ln k)^2$ , while the total connection + +cost of the last uncovered cluster is at most $\lambda^3 \cdot \frac{n}{k} \cdot k^{2\theta}$ , omitting constant factors for simplicity. Thus, the probability that greedy $k$ -means++ encounters $\mathcal{B}$ can be upper bounded by + +$$ +k \cdot \left(\frac {(4 (1 + \theta)) ^ {2} (k - 1) \cdot (\ln k) ^ {2}}{(k - 1) \cdot (\ln k) ^ {2} + k ^ {2 \theta}}\right) ^ {\ln k} = \left(\frac {1 6 e (1 + \theta) ^ {2} \cdot (k - 1) \cdot (\ln k) ^ {2}}{(k - 1) \cdot (\ln k) ^ {2} + k ^ {2 \theta}}\right) ^ {\ln k}. +$$ + +![](images/ca3325411a53c7443a85cefdc641b4c1e8297ab76c4927d72a0a42bb78dce4e8.jpg) + +Proof of Theorem 7. Lemma 11 and Lemma 12, through a series of mathematical calculations, directly establish the theorem. More specifically, when $\theta > 1/2$ , the lower bound on the failure probability for $k$ -means++ is $\Theta(k^{1 - 2\theta})$ , while the upper bound for greedy $k$ -means++ is $\Theta\left(k^{16e\ln \ln k + (1 - 2\theta)\ln k}\right)$ . As the former asymptotically dominates the latter, we can conclude that greedy $k$ -means++ has a higher probability of covering all optimal clusters than $k$ -means++. + +# D Additional Experiments + +In this section, we present more experimental results on varying $k$ , which validates our theory works for different $k$ . + +![](images/0e32c6237e816c855452757c672ec4f28a1a03fef262080c5ad29eb74a2a363a.jpg) +Figure 2: Coverage probability and $k$ -means objective over iterations for $k = 8$ + +![](images/61a6b6a04b4eee54faffe978d4659972ddc5e6828fb3a7153e829572eaee66bd.jpg) + +![](images/90dbc0373b44104d6d4e1f0b08accbf381805cfc5412500b4491d2679cdba4bc.jpg) +Figure 3: Coverage probability and $k$ -means objective over iterations for $k = 32$ + +![](images/77d377a0d85f8e55af47409b30f03be736abbc1723007025d93912de9f8b3610.jpg) + +![](images/5f54ca78f6a06b94f5ded62d3b54d813f008e12a4e6aeb6851cfe602bfb66389.jpg) +Figure 4: Coverage probability and $k$ -means objective over iterations for $k = 64$ + +![](images/9c3f58857ecbbb3a08afe98f90ee1bcbbaa1398fd16908c754a0183d6edc8750.jpg) \ No newline at end of file diff --git a/NeurIPS/2025/A Beyond-Worst-Case Analysis of Greedy k-means++/images.zip b/NeurIPS/2025/A Beyond-Worst-Case Analysis of Greedy k-means++/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..a390dcd500f636d5f51747d5b8735b0cae429b79 --- /dev/null +++ b/NeurIPS/2025/A Beyond-Worst-Case Analysis of Greedy k-means++/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:878eee13677e439bf8ceba1f23b8cddc80c6a0a2bee2395b40e1548be8d50c95 +size 652455 diff --git a/NeurIPS/2025/A Beyond-Worst-Case Analysis of Greedy k-means++/layout.json b/NeurIPS/2025/A Beyond-Worst-Case Analysis of Greedy k-means++/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..6dab22cec4106c3d944b2cc0e4360c663170e6bd --- /dev/null +++ b/NeurIPS/2025/A Beyond-Worst-Case Analysis of Greedy k-means++/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c202a9974eb49a2738aea882ade9445e83ca68c9767920c34ba693bcf95c3814 +size 1482222 diff --git a/NeurIPS/2025/A Black-Box Debiasing Framework for Conditional Sampling/625b2af5-88a8-4509-ba78-f0bbeecd61e6_content_list.json b/NeurIPS/2025/A Black-Box Debiasing Framework for Conditional Sampling/625b2af5-88a8-4509-ba78-f0bbeecd61e6_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..24cebd39089361007f700b02fbf63ebdb5f12185 --- /dev/null +++ b/NeurIPS/2025/A Black-Box Debiasing Framework for Conditional Sampling/625b2af5-88a8-4509-ba78-f0bbeecd61e6_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:424e7b339f9d12781a0f9ea54778d2e69500c40a23dcd1bc4e1e6374f47b5019 +size 208608 diff --git a/NeurIPS/2025/A Black-Box Debiasing Framework for Conditional Sampling/625b2af5-88a8-4509-ba78-f0bbeecd61e6_model.json b/NeurIPS/2025/A Black-Box Debiasing Framework for Conditional Sampling/625b2af5-88a8-4509-ba78-f0bbeecd61e6_model.json new file mode 100644 index 0000000000000000000000000000000000000000..1301457693905cec2cc67bf2a420fe83899b66d5 --- /dev/null +++ b/NeurIPS/2025/A Black-Box Debiasing Framework for Conditional Sampling/625b2af5-88a8-4509-ba78-f0bbeecd61e6_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:245c6d8a4831b8305432161b16335561f0564c73ee30a388a353fb9203a52219 +size 253572 diff --git a/NeurIPS/2025/A Black-Box Debiasing Framework for Conditional Sampling/625b2af5-88a8-4509-ba78-f0bbeecd61e6_origin.pdf b/NeurIPS/2025/A Black-Box Debiasing Framework for Conditional Sampling/625b2af5-88a8-4509-ba78-f0bbeecd61e6_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..94f65f7921eeb49ea12865ff554002c515099fb2 --- /dev/null +++ b/NeurIPS/2025/A Black-Box Debiasing Framework for Conditional Sampling/625b2af5-88a8-4509-ba78-f0bbeecd61e6_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7f5efe89e324450108ec87ad462d9ea7ac695029c72fb6f4a354bb911958ff45 +size 605304 diff --git a/NeurIPS/2025/A Black-Box Debiasing Framework for Conditional Sampling/full.md b/NeurIPS/2025/A Black-Box Debiasing Framework for Conditional Sampling/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d29cc05a8f20df6ec5667c3925b99fa41ca6cd45 --- /dev/null +++ b/NeurIPS/2025/A Black-Box Debiasing Framework for Conditional Sampling/full.md @@ -0,0 +1,1165 @@ +# A Black-Box Debiasing Framework for Conditional Sampling + +Han Cui + +University of Illinois at Urbana-Champaign + +Champaign, IL + +hancui5@illinois.edu + +Jingbo Liu + +University of Illinois at Urbana-Champaign + +Champaign, IL + +jingbol@illinois.edu + +# Abstract + +Conditional sampling is a fundamental task in Bayesian statistics and generative modeling. Consider the problem of sampling from the posterior distribution $P_{X|Y = y^*}$ for some observation $y^*$ , where the likelihood $P_{Y|X}$ is known, and we are given $n$ i.i.d. samples $D = \{X_i\}_{i=1}^n$ drawn from an unknown prior distribution $\pi_X$ . Suppose that $f(\hat{\pi}_{X^n})$ is the distribution of a posterior sample generated by an algorithm (e.g., a conditional generative model or the Bayes rule) when $\hat{\pi}_{X^n}$ is the empirical distribution of the training data. Although averaging over the randomness of the training data $D$ , we have $\mathbb{E}_D(\hat{\pi}_{X^n}) = \pi_X$ , we do not have $\mathbb{E}_D\{f(\hat{\pi}_{X^n})\} = f(\pi_X)$ due to the nonlinearity of $f$ , leading to a bias. In this paper we propose a black-box debiasing scheme that improves the accuracy of such a naive plug-in approach. For any integer $k$ and under boundedness of the likelihood and smoothness of $f$ , we generate samples $\hat{X}^{(1)}, \ldots, \hat{X}^{(k)}$ and weights $w_1, \ldots, w_k$ such that $\sum_{i=1}^k w_i P_{\hat{X}^{(i)}}$ is a $k$ -th order approximation of $f(\pi_X)$ , where the generation process treats $f$ as a black-box. Our generation process achieves higher accuracy when averaged over the randomness of the training data, without degrading the variance, which can be interpreted as improving memorization without compromising generalization in generative models. + +# 1 Introduction + +Conditional sampling is a major task in Bayesian statistics and generative modeling. Given an observation $y^{*}$ , the objective is to draw samples from the posterior distribution $P_{X|Y = y^*}$ , where the likelihood $P_{Y|X}$ is known but the prior distribution $\pi_{X}$ is unknown. Instead, we are provided with a dataset $D = \{X_i\}_{i=1}^n$ consisting of $n$ i.i.d. samples drawn from $\pi_{X}$ . + +The setting is common in a wide range of applications, including inpainting and image deblurring [9, 5] (where $X$ is an image and $Y|X$ is a noisy linear transform), text-conditioned image generation [7, 13](where $X$ is an image and $Y$ is a natural language prompt), simulating biomedical structures with desired properties, and trajectory simulations for self-driving cars. Moreover, conditional sampling is equally vital in high-impact machine learning and Bayesian statistical methods, particularly under distribution shift, such as in transfer learning. For instance, conditional sampling has enabled diffusion models to generate trajectories under updated policies, achieving state-of-the-art performance in offline reinforcement learning [8, 1, 26]. Pseudo-labeling, a key technique for unsupervised pretraining [10] and transfer learning calibration [20], relies on generating conditional labels. Additionally, conditional diffusion models seamlessly integrate with likelihood-free inference [6, 18, 27]. Existing approaches often use generative models such as VAEs or Diffusion models to generate samples by learning $P_{X|Y=y^*}$ implicitly from the data. + +Our work focuses on approximating the true posterior $P_{X|Y=y^*}$ using the observed samples $D = X^n = (X_1, \ldots, X_n)$ and the new observation $y^*$ , but without the knowledge of the prior. Denote by $P_{\hat{X}|Y=y^*,D}$ the approximating distribution. We can distinguish two kinds of approximations: First, $P_{\hat{X}|Y=y^*,D} \approx P_{X|Y=y^*}$ with high probability over $D$ , which captures the generalization ability since the model must learn the distribution from the training samples. This criterion is commonly adopted in estimation theory and has also been examined in the convergence analysis of generative models [16, 28, 26, 22]. Second, $\mathbb{E}_D(P_{\hat{X}|Y=y^*,D}) \approx P_{X|Y=y^*}$ is a weaker condition since it only requires approximation when averaged over the randomness of the training data, but is still useful in some sampling and generative tasks, e.g. generating samples for bootstrapping or Monte Carlo estimates of function expectations. The second condition captures the ability to memorize or imitate training sample distribution. It is interesting to note that in the unconditional setting (i.e., without distribution shift), a permutation sampler can perfectly imitate the unknown training data distribution, even if $n = 1$ , so the problem is trivial from the sample complexity perspective. However, in the conditional setting, it is impossible to get such a perfect imitation with finite training data, as a simple binary distribution example in Section 3.2 illustrates. It naturally leads to the following question: + +How fast can the posterior approximation converge to the true posterior as $n \to \infty$ , and is there a sampling scheme achieving this convergence rate? + +Contribution. We address the question above by proposing a novel debiasing framework for posterior approximation. Our main contributions can be summarized as follows: + +- Debiasing framework for posterior approximation. We introduce a novel debiasing framework for posterior approximation with an unknown prior. Our method leverages the known likelihood $P_{Y|X}$ and the observed data to construct an improved approximate posterior $\widetilde{P}_{X^n}(x|y^*)$ with provably reduced bias. In particular, let $f(\hat{\pi}_{X^n})$ represent the distribution of a posterior sample generated by an algorithm $f$ when $\hat{\pi}_{X^n}$ is the empirical distribution of the training data. Then for any integer $k$ , assuming that the likelihood function $P_{Y|X}$ is bounded and $f$ is sufficiently smooth, we generate samples $\hat{X}^{(1)},\dots,\hat{X}^{(k)}$ from $f$ based on multiple resampled empirical distributions. These are then combined with designed (possibility negative) weights $w_{1},\ldots ,w_{k}$ to construct an approximate posterior: + +$$ +\widetilde {P} _ {X ^ {n}} \left(\cdot | y ^ {*}\right) = \sum_ {i = 1} ^ {k} w _ {i} P _ {\hat {X} ^ {(i)}} +$$ + +which is a $k$ -th order approximation of $f(\pi_X)$ , treating the generation process $f$ as a black-box. Our generation process achieves higher accuracy when averaged over the randomness of the training data, but not conditionally on the training data, which highlights the trade-off between memorization and generalization in generative models. Specifically, we do not assume any parametric form for the prior and our method can achieve a bias rate of $\mathcal{O}(n^{-k})$ for any prescribed integer $k$ and a variance rate of $\mathcal{O}(n^{-1})$ . + +- Theoretical bias and variance guarantees. We establish theoretical guarantees on both bias and variance for the Bayes-optimal sampler under continuous prior setting and for a broad class of samplers $f$ with a continuous $2k$ -th derivative, as specified in Assumption 2, under the discrete prior setting. The proposed debiasing framework can also be applied in a black-box manner (see Remark 2 for the intuition), making it applicable to a broad class of state-of-the-art conditional samplers, such as diffusion models and conditional VAE. Based on this perspective, we treat the generative model $f$ as a black box that can output posterior samples given resampled empirical distributions. Applying $f$ to multiple recursive resampled versions of the training data and combining the outputs with polynomial weights, we obtain a bias-corrected approximation of the posterior. The procedure is described in Algorithm 1. + +Our approach is also related to importance sampling. Since the true posterior $P_{X|Y}$ is intractable to compute, we can use expectations under the debiased posterior $\widetilde{P}_{X^n}(x|y^*)$ to approximate the expectations under the true posterior $P_{X|Y = y^*}$ . For a test function $h$ , we estimate $\mathbb{E}_{P_{X|Y = y^*}}\{h(X)\}$ by + +$$ +\mathbb {E} _ {\tilde {P} _ {X ^ {n}} (x \mid y ^ {*})} \left\{h (X) \right\} \approx \frac {1}{N} \sum_ {i = 1} ^ {N} h \left(\tilde {X} _ {j}\right) \frac {\widetilde {P} _ {X ^ {n}} \left(\tilde {X} _ {j} \mid y ^ {*}\right)}{q \left(\tilde {X} _ {j} \mid y ^ {*}\right)}, \tag {1} +$$ + +Algorithm 1 Posterior Approximation via Debiasing Framework + +Input: Observation $y^{*}$ , likelihood $P_{Y|X}$ , data $X^n = (X_1, \ldots, X_n)$ , number of steps $k$ , a black-box conditional sampler $f$ (i.e., a map from a prior distribution to a posterior distribution) + +Output: $\hat{X}^{(j)}, j = 1, \ldots, k$ such that $\sum_{j=0}^{k-1} \binom{k}{j+1} (-1)^j P_{\hat{X}^{(j+1)}}$ is a high-order approximation of the posterior $P_{X|Y=y^*}$ + +1: Initialize $\hat{p}^{(1)}\coloneqq \hat{\pi}_{X^n}$ +2: for $\ell = 2$ to $k$ do +3: Generate $n$ i.i.d. samples from $\hat{p}^{(\ell -1)}$ +4: Let $\hat{p}^{(\ell)}$ be the empirical distribution of the sampled data +5: end for +6: for $j = 1$ to $k$ do +7: Generate samples $\hat{X}^{(j)}\sim f(\hat{p}^{(j)})$ +8: end for +9: Return $\hat{X}^{(j)}, j = 1, \dots, k$ + +where $\tilde{X}_j\sim q(x|y^*)$ for a chosen proposal distribution $q$ . This resembles our method, in which we approximate the true posterior by a weighted combination $\sum_{i = 1}^{k}w_{i}P_{\hat{X}^{(i)}}$ . And in (1), the term $\widetilde{P}_{X^n}(\tilde{X}_j|y^*) / q(\tilde{X}_j|y^*)$ can be interpreted as a weight assigned to each sample, analogous to the weights $w_{i}$ in our framework. Therefore, we expect that Algorithm 1 can be broadly applied to Monte Carlo estimates of function expectations, similar to the standard importance sampling technique. + +# 2 Related work + +Jackknife Technique. Our work is related to the jackknife technique [17], a classical method for bias reduction in statistical estimation that linearly combines estimators computed on subsampled datasets. Specifically, the jackknife technique generates leave-one-out (or more generally, leave- $s$ -out where $s \geq 1$ ) versions of an estimator, and then forms a weighted combination to cancel lower-order bias terms. Recently, Nowozin [14] applied the jackknife to the importance-weighted autoencoder (IWAE) bound $\hat{\mathcal{L}}_n$ , which estimates the marginal likelihood $\log \pi(x)$ using $n$ samples. While $\hat{\mathcal{L}}_n$ is proven to be an estimator with bias of order $\mathcal{O}(n^{-1})$ , the jackknife correction produces a new estimator with reduced bias of order $\mathcal{O}(n^{-m})$ . Our paper introduces a debiaisng framework based on the similar idea that using a linear combination of multiple approximations to approximate the posterior. + +Conditional Generative Models. Conditional generative models have become influential and have been extensively studied for their ability to generate samples from the conditional data distribution $P(\cdot | y)$ where $y$ is the conditional information. This framework is widely applied in vision generation tasks such as text-to-image synthesis [13, 24, 2] where $y$ is an input text prompt, and image inpainting [11, 21] where $y$ corresponds to the known part of an image. We expect that our proposed debiasing framework could work for a broad class of conditional generative models to construct a high order approximation of the posterior $P(\cdot | y)$ . + +Memorization in Generative Models. The trade-off between memorization and generalization has been a focus of research in recent years. In problems where generating new structures or preserving privacy of training data is of high priority, generalization is preferred over memorization. For example, a study by Carlini et al. [4] demonstrates that diffusion models can unintentionally memorize specific images from their training data and reproduce them when generating new samples. To reduce the memorization of the training data, Somepalli et al. [19] applies randomization and augmentation techniques to the training image captions. Additionally, Yoon et al. [25] investigates the connection between generalization and memorization, proposing that these two aspects are mutually exclusive. Their experiments suggest that diffusion models are more likely to generalize when they fail to memorize the training data. On the other hand, memorizing and imitating the training data may be intentionally exploited, if the goal is Monte Carlo sampling for evaluations of expected values, or if the task does not involve privacy issues, e.g. image inpainting and reconstruction. In these applications, the ability to imitate or memorize the empirical distribution of the training data becomes essential, especially when generalization is unattainable due to the insufficient data. Our work focuses + +on memorization phase and shows that it is possible to construct posterior approximations with provably reduced bias by exploiting the empirical distribution. + +Mixture-Based Approximation of Target Distributions. Sampling from a mixture of distributions $a_1P_{X_1} + a_2P_{X_2} + \dots +a_kP_{X_k}$ to approximate a target distribution $P^{*}$ is commonly used in Bayesian statistics, machine learning, and statistical physics, especially when individual samples or proposals are poor approximations, but their ensemble is accurate. Traditional importance sampling methods often rely on positive weights, but recent work has expanded the landscape to include more flexible and powerful strategies, including the use of signed weights and gradient information. For example, Oates et al. [15] uses importance sampling and control functional estimators to construct a linear combination of estimators with weights $a_{k}$ to form a variance-reduced estimator for an expectation under a target distribution $P^{*}$ . Liu and Lee [12] select the weights $a_{k}$ by minimizing the empirical version of the kernelized Stein discrepancy (KSD), which often results in negative weights. + +# 3 Problem setup and notation + +Consider a dataset $\{X_{i}\}_{i = 1}^{n}$ consisting of $n$ independent and identically distributed (i.i.d.) samples, where $X_{i}\in \mathcal{X}$ is drawn from an unknown prior distribution $\pi_{X}$ and the conditional distribution $P_{Y|X}$ is assumed to be known. In the Bayesian framework, the posterior distribution of $X$ given $Y$ is given by + +$$ +P _ {X | Y} (d x | y) = \frac {P _ {Y | X} (y | x) \pi_ {X} (d x)}{\int P _ {Y | X} (y | x) \pi_ {X} (d x)}. +$$ + +Given the observed data $X^n = (X_1, \dots, X_n)$ and the new observation $y^*$ , our goal is to approximate the true posterior $P_{X|Y = y^*}$ . + +# 3.1 Naive plug-in approximation + +A natural approach is to replace the unknown prior $\pi_{X}$ with its empirical counterpart + +$$ +\hat {\pi} _ {X ^ {n}} = n ^ {- 1} \sum_ {i = 1} ^ {n} \delta_ {X _ {i}} +$$ + +in the Bayes' rule to compute an approximate posterior which yields the plug-in posterior + +$$ +\widehat {P} _ {X \mid Y} (d x \mid y ^ {*}) = \frac {P _ {Y \mid X} \left(y ^ {*} \mid x\right) \widehat {\pi} _ {X ^ {n}} (d x)}{\int P _ {Y \mid X} \left(y ^ {*} \mid x\right) \widehat {\pi} _ {X ^ {n}} (d x)}. \tag {2} +$$ + +Note that even though $\mathbb{E}_D(\hat{\pi}_{X^n}) = \pi_X$ , the nonlinearity of Bayes' rule makes the resulting posterior (2) still biased, that is, $\mathbb{E}_D\left\{\widehat{P}_{X|Y}(\cdot |y^*)\right\} \neq P_{X|Y}(\cdot |y^*)$ . If the denominator in (2) were replaced with $\int P_{Y|X}(y^{*}|x)\pi_{X}(dx)$ , then averaging the R.H.S. of (2) over the randomness in $X^n$ would yield the true posterior $P_{X|Y}(dx|y^*) = P_{Y|X}(y^* |x)\pi_X(dx) / \int P_{Y|X}(y^* |x)\pi_X(dx)$ exactly. + +For typical choices of $P_{Y|X}$ which have nice conditional density (e.g., the additive Gaussian noise channel), $\int P_{Y|X}(y^* |x)\hat{\pi}_{X^n}(dx)$ converges at the rate of $n^{-1 / 2}$ , by the central limit theorem. Consequently, $\mathbb{E}_D(\widehat{P}_{X|Y = y^*})$ converges to the true posterior at the rate $\tilde{\mathcal{O}}(n^{-1 / 2})$ in the $\infty$ -Renyi divergence metric regardless of the smoothness of $\pi_X$ . Under appropriate regularity conditions, we can in fact show that $\mathbb{E}_D(\widehat{P}_{X|Y = y^*})$ converges at the rate of $\tilde{\mathcal{O}}(n^{-1})$ , which comes from the variance term in the Taylor expansion. Naturally, we come to an essential question: can we eliminate the bias entirely? That is, is it possible that $\mathbb{E}_D\{\widehat{P}_{X|Y}(\cdot |y^*)\} = P_{X|Y}(\cdot |y^*)$ ? + +# 3.2 Impossibility of exact unbiasedness + +Exact unbiasedness is, in general, unattainable. Consider the simple case where $X$ is binary, that is, $X \sim \operatorname{Bern}(q)$ for some unknown parameter $q \in (0,1)$ . Define the likelihood ratio $\alpha = \alpha(y^{*}) :=$ + +$P_{Y|X}(y^{*}|1) / P_{Y|X}(y^{*}|0)$ . Then the true posterior is + +$$ +X | Y = y ^ {*} \sim \mathrm {B e r n} \left(\frac {\alpha q}{\alpha q + 1 - q}\right). +$$ + +On the other hand, if we approximate the posterior distribution as $\hat{P}_{X|Y}(1|y = y^*) = \mathrm{Bern}(p(k))$ upon seeing $k$ outcomes equal to 1, then + +$$ +\mathbb {E} _ {D} \left\{\widehat {P} _ {X \mid Y} \left(1 \mid y ^ {*}\right) \right\} = \sum_ {k = 0} ^ {n} p (k) \binom {n} {k} q ^ {k} (1 - q) ^ {n - k}, \tag {3} +$$ + +which is a polynomial function of $q$ , and hence cannot equal the rational function $\alpha q / (\alpha q + 1 - q)$ for all $q$ . This implies that an exact imitation, in the sense that $\mathbb{E}_D\{\widehat{P}_{X|Y}(\cdot |y^*)\} = P_{X|Y}(\cdot |y^*), \forall \pi_X$ , is impossible. However, since a rational function can be approximated arbitrarily well by polynomials, this does not rule out the possibility that a better sampler achieving convergence faster than, say, the $\tilde{\mathcal{O}}(n^{-1/2})$ rate of the naive plug-in method. Indeed, in this paper we propose a black-box method that can achieve convergence rates as fast as $\mathcal{O}(n^{-k})$ for any fixed $k > 0$ . + +# 3.3 Objective and notation + +Since the bias in the plug-in approximation arises from the nonlinearity of Bayes' rule, we aim to investigate whether a faster convergence rate can be achieved. Our objective is to construct an approximation $\widetilde{P}_{X^n}(x|y = y^*)$ that improves the plug-in approximation by reducing the bias. Specifically, the debiased approximation satisfies the following condition: + +$$ +\left| \mathbb {E} _ {X ^ {n}} \left\{\widetilde {P} _ {X ^ {n}} (x | y = y ^ {*}) \right\} - P _ {X | Y} (x | y ^ {*}) \right| < \left| \mathbb {E} _ {X ^ {n}} \left\{\widehat {P} _ {X | Y} (x | y ^ {*}) \right\} - P _ {X | Y} (x | y ^ {*}) \right|. +$$ + +More generally, we can replace the Bayes rule by an arbitrary map $f$ from a prior to a posterior distribution (e.g. by a generative model), and the goal is a construct a debiased map $\tilde{f}$ such that + +$$ +\left\| \mathbb {E} _ {X ^ {n}} \tilde {f} \left(\widehat {\pi} _ {X ^ {n}}\right) - f (\pi) \right\| _ {\mathrm {T V}} < \left\| \mathbb {E} _ {X ^ {n}} f \left(\widehat {\pi} _ {X ^ {n}}\right) - f (\pi) \right\| _ {\mathrm {T V}}. +$$ + +Notation. Let $\delta_{x}$ denote the Dirac measure, $\| \cdot \|_{\mathrm{TV}}$ denote the total variation norm. For any positive integer $m$ , denote $[m] = \{1,\dots ,m\}$ as the set of all positive integers smaller than all equal to $m$ . Write $b_{n} = \mathcal{O}(a_{n})$ if $b_{n} / a_{n}$ is bounded as $n\to \infty$ . Write $b_{n} = \mathcal{O}_{s}(a_{n})$ if $b_{n} / a_{n}$ is bounded by $C(s)$ as $n\to \infty$ for some constant $C(s)$ that depends only on $s$ . We use the notation $a\lesssim b$ to indicate that there exists a constant $C > 0$ such that $a\leq Cb$ . Similarly, $a\lesssim k$ $b$ means that there exists a constant $C(k) > 0$ that depends only on $k$ such that $a\leq C(k)b$ . Furthermore, for notational simplicity, we will use $\pi$ to denote the true prior $\pi_X$ and $\hat{\pi}$ to denote the empirical prior $\hat{\pi}_{X^n}$ in the rest of the paper. + +# 4 Main result + +# 4.1 Debiased posterior approximation under continuous prior + +Let $\Delta_{\mathcal{X}}$ denote the space of probability measures on $\mathcal{X}$ . Define the likelihood function $\ell(x) \coloneqq P_{Y|X}(y^*|x)$ , which represents the probability of observing the data $y^*$ given $x$ . Let $f: \Delta_{\mathcal{X}} \to \Delta_{\mathcal{X}}$ be a map from the prior measure to the posterior measure, conditioned on the observed data $y^*$ . Let $B_n$ be the operator such that for any function $f: \Delta_{\mathcal{X}} \to \Delta_{\mathcal{X}}$ , + +$$ +B _ {n} f (p) = \mathbb {E} \left\{f (\hat {p}) \right\}, \tag {4} +$$ + +where $\hat{p}$ denotes the empirical measure of $n$ i.i.d. samples from measure $p$ . + +We consider the case that $f$ represents a mapping corresponding to the Bayes posterior distribution. Using Bayes' theorem, for any measure $\pi \in \Delta_{\mathcal{X}}$ and any measurable set $A \subset \mathcal{X}$ , the posterior measure $f(\pi)$ is expressed as + +$$ +f (\pi) (A) = \frac {\int_ {A} \ell (x) \pi (d x)}{\int_ {\mathcal {X}} \ell (x) \pi (d x)}. +$$ + +As discussed in Section 3, the equality $B_{n}f(\pi) = f(\pi)$ is not possible due to the nonlinearity of $f$ . However, we can achieve substantial improvements over the plug-in method by using polynomial approximation techniques analogous to those from prior statistical work by Cai and Low [3] and Wu and Yang [23]. For $k \geq 1$ , we define the operator $D_{n,k}$ as a linear combination of the iterated operators $B_{n}^{j}$ for $j = 0, \dots, k - 1$ : + +$$ +D _ {n, k} = \sum_ {j = 0} ^ {k - 1} \binom {k} {j + 1} (- 1) ^ {j} B _ {n} ^ {j}. +$$ + +Assumption 1. The likelihood function $\ell$ is bounded, i.e., there exists $0 < L_{1} \leq L_{2}$ such that $L_{1} \leq \ell(x) \leq L_{2}$ . + +The following theorem provides a systematic method for constructing an approximation of $f(\pi)$ with an approximation error of order $\mathcal{O}(n^{-k})$ for any desired integer $k$ . + +Theorem 1. Under Assumption 1, for any measurable set $A \subset \mathcal{X}$ and any $k \in \mathbb{N}^+$ , we have + +$$ +\left\| \mathbb {E} _ {X ^ {n}} \left\{D _ {n, k} f (\hat {\pi}) \right\} - f (\pi) \right\| _ {\mathrm {T V}} = \mathcal {O} _ {L _ {1}, L _ {2}, k} \left(n ^ {- k}\right), \tag {5} +$$ + +$$ +\operatorname {V a r} _ {X ^ {n}} \left\{D _ {n, k} f (\hat {\pi}) (A) \right\} = \mathcal {O} _ {L _ {1}, L _ {2}, k} \left(n ^ {- 1}\right). \tag {6} +$$ + +Remark 1. $D_{n,k}f(\hat{\pi}) = \sum_{j=0}^{k-1} \binom{k}{j+1} (-1)^j B_n^j f(\hat{\pi})$ in (5) can be interpreted as a weighted average of the distribution of some samples. Specifically, if we treat the coefficient $\binom{k}{j+1} (-1)^j$ as the weight $w_j$ and $B_n^j f(\hat{\pi})$ as the distribution of some sample $\hat{X}^{(j)}$ , then $D_{n,k}f(\hat{\pi}) = \sum_{j=0}^{k-1} w_j P_{\hat{X}^{(j)}}$ . + +Remark 2. Recall the binary case discussed in Section 3, (3) illustrates that we cannot get an exact approximation for the true posterior. But from (5), we demonstrate that even if $\| \mathbb{E}_{X^n}\{D_{n,k}f(\hat{\pi})\} - f(\pi)\|_{\mathrm{TV}} = 0$ is impossible, it can be arbitrarily small. Although the theoretical guarantees are derived for the Bayes-optimal sampler, (5) is expected to hold for general sampler $f$ such as diffusion models. Here we give the intuition for this conjecture. We view the operator $B_n f(\pi) \coloneqq \mathbb{E}\{f(\hat{\pi})\}$ as a good approximation of $f(\pi)$ , i.e., $B_n \approx I$ , where $I$ is the identity operator. This implies that the error operator $E \coloneqq I - B_n$ is a "small" operator. Under this heuristic, if $Ef(\pi) = \mathcal{O}(n^{-1})$ , intuitively we have $E^k f(\pi) = \mathcal{O}(n^{-k})$ . Using the binomial expansion of $E^k = (I - B_n)^k$ , we have $E^k f(\pi) = f(\pi) - \sum_{j=1}^{k} \binom{k}{j} (-1)^{j-1} B_n^j f(\pi) = f(\pi) - \mathbb{E}\{\sum_{j=1}^{k} \binom{k}{j} (-1)^{j-1} B_n^{j-1} f(\hat{\pi})\} = f(\pi) - \mathbb{E}\{D_{n,k}f(\hat{\pi})\}$ . This representation motivates the specific form of $D_{n,k}$ . + +Remark 3. In general, the curse of dimensionality may arise and depends on the specific distribution of $X$ and the likelihood function $\ell$ . There is no universal relationship between $n$ and the dimension $d$ . However, to build intuition, we give an example that illustrates how $n$ and $d$ may relate. Suppose that $Y = (Y(1),\ldots ,Y(d))$ and $X = (X(1),\ldots ,X(d))$ have i.i.d. components, and $L_{1}\leq P\big(Y(i)|X(i)\big)\leq L_{2}$ for $1\leq i\leq d$ . Then we have $\ell (X)\coloneqq P(Y|X)\in [L_1^d,L_2^d ]$ . Note that $\mathcal{O}_{L_1,L_2,k}(n^{-k})$ in (5) can be bounded by $C(k)(L_2^d /L_1^d)^{2k}n^{-k}$ for some constant $C(k)$ depending only on $k$ . To ensure that our debiasing method improves over the baseline method without debiasing in the case of growing dimensions, it suffices to let $n$ and $d$ satisfy that $(L_2^d /L_1^d)^{2k}n^{-k}\ll n^{-1}$ when $k\geq 2$ , which is equivalent to $kd\ll \log (n)$ . + +Sketch proof for Theorem 1. First let $\mu = \int_{\mathcal{X}}\ell (x)\pi (dx)$ and $\mu_{A} = \int_{A}\ell (x)\pi (dx)$ and introduce a new operator + +$$ +C _ {n, k} := \sum_ {j = 1} ^ {k} \binom {k} {j} (- 1) ^ {j - 1} B _ {n} ^ {j}, +$$ + +then we have $B_{n}D_{n,k} = C_{n,k}$ . By the definition of $B_{n}$ , it suffices to show that + +$$ +C _ {n, k} f (\pi) (A) - f (\pi) (A) = \mathcal {O} _ {L _ {1}, L _ {2}, k} \left(n ^ {- k}\right). +$$ + +The first step is to express $B_n^j f(\pi)$ with the recursive resampled versions of the training data. Specifically, let $\hat{\pi}^{(0)} = \pi, \hat{\pi}^{(1)} = \hat{\pi}$ and set $(X_1^{(0)},\ldots ,X_n^{(0)}) \equiv (X_1,\ldots ,X_n)$ . For $j = 1,\dots ,k$ , + +we define $\hat{\pi}^{(j)}$ as the empirical measure of $n$ i.i.d. samples $(X_{1}^{(j - 1)},\ldots ,X_{n}^{(j - 1)})$ drawn from the measure $\hat{\pi}^{(j - 1)}$ . Additionally, let + +$$ +e _ {n} ^ {(j)} = n ^ {- 1} \sum_ {i = 1} ^ {n} \left\{\ell \left(X _ {i} ^ {(j)}\right) - \mu \right\} \quad \text {a n d} \quad \mu_ {A} ^ {(j)} = n ^ {- 1} \sum_ {i = 1} ^ {n} \ell \left(X _ {i} ^ {(j)}\right) \delta_ {X _ {i} ^ {(j)}} (A). +$$ + +Then we have + +$$ +C _ {n, k} f (\pi) (A) = \sum_ {j = 1} ^ {k} \binom {k} {j} (- 1) ^ {j - 1} B _ {n} ^ {j} f (\pi) (A) = \sum_ {j = 1} ^ {k} \binom {k} {j} (- 1) ^ {j - 1} \mathbb {E} \left(\frac {\mu_ {A} ^ {(j - 1)}}{e _ {n} ^ {(j - 1)} + \mu}\right). \tag {7} +$$ + +The second step is to rewrite (7) with Taylor expansion of $\mu_A^{(j - 1)} / (e_n^{(j - 1)} + \mu)$ with respect to $e_n^{(j - 1)}$ up to order $2k - 1$ . $L_{1}\leq l(X_{i}^{(j - 1)})\leq L_{2}$ and Hoeffding's inequality implies that the expectation of the residual term $\mathbb{E}\{(e_n^{(j - 1)})^{2k} / \xi^{2k + 1}\}$ for some $\xi$ between $e_n^{(j - 1)} + \mu$ and $\mu$ is $\mathcal{O}_{L_1,L_2,k}(n^{-k})$ . Now we instead to show that + +$$ +B _ {k, r} := \sum_ {j = 1} ^ {k} \binom {k} {j} (- 1) ^ {j - 1} \mathbb {E} \left\{\mu_ {A} ^ {(j - 1)} (e _ {n} ^ {(j - 1)}) ^ {r} \right\} = \mathcal {O} _ {L _ {1}, L _ {2}, k} (n ^ {- k}), +$$ + +since (7) is equal to $\mu_A / \mu +\sum_{r = 1}^{2k - 1}(-1)^r\mu^{-r - 1}B_{k,r} + \mathcal{O}_{L_1,L_2,k}(n^{-k})$ + +Define a new operator $B: h \mapsto \mathbb{E}[h(\hat{\pi})]$ for any $h: \Delta_{\mathcal{X}} \to \mathbb{R}$ and let $h_s(\pi) = \{\int_A \ell(x)\pi(dx)\} \{\int \ell(x)\pi(dx)\}^s$ . Then + +$$ +B _ {k, r} = \sum_ {s = 0} ^ {r} \binom {r} {s} (- 1) ^ {(r - s)} \mu^ {r - s} \sum_ {j = 1} ^ {k} \binom {k} {j} (- 1) ^ {j - 1} B ^ {j} h _ {s} (\pi). +$$ + +The last step is to prove + +$$ +(I - B) ^ {k} h _ {s} (\pi) = \mathcal {O} _ {L _ {1}, L _ {2}, s} (n ^ {- k}), \tag {8} +$$ + +since (8) is equivalent to $\sum_{j=1}^{k} \binom{k}{j} (-1)^{j-1} B^j h_s(\pi) = h_s(\pi) + \mathcal{O}_{L_1, L_2, s}(n^{-k})$ . Finally (8) follows from the fact that $(I-B)^k h_s(\pi)$ can be expressed as a finite sum of the terms which have the following form: + +$$ +\alpha_ {\mathbf {a}, \mathbf {s}, v} \left\{\int_ {A} \ell^ {v} (x) \pi (d x) \right\} \prod_ {i} \left\{\int \ell^ {a _ {i}} (x) \pi (d x) \right\} ^ {s _ {i}}, +$$ + +where $|\alpha_{\mathbf{a},\mathbf{s},v}|\leq C_k(s)n^{-k}$ for some constnat $C_k(s)$ (see our Lemma 2). + +![](images/ccab9c57e2ab6cea16258613a5d98910b0894885f981960d8e2b8120f21e1c7b.jpg) + +# 4.2 Debiased posterior approximation under discrete prior + +In this section, we consider the case where $X$ follows a discrete distribution. As mentioned in Remark 2, the result in Theorem 1 is expected to hold in a broader class of samplers $f$ under smoothness, extending beyond just the Bayes-optimal sampler $f$ . The assumption of finite $\mathcal{X}$ in this section allows us to simplify some technical aspects in the proof. + +Let the support of $X$ be denoted as $\mathcal{X} = \{u_1, u_2, \ldots, u_m\}$ . Assume that $|\mathcal{X}| = m$ is finite, and $X$ is distributed according to an unknown prior distribution $\pi(x)$ such that the probability of $X$ taking the value $u_i$ is given by $\pi(X = u_i) = q_i$ for $i = 1, 2, \ldots, m$ . Here, the probabilities $q_i$ are unknown and satisfy the usual constraints that $q_i \geq 0$ for all $i$ and $\sum_{i=1}^{m} q_i = 1$ . + +Let $\mathbf{q} = (q_{1},\dots ,q_{m})^{\top}$ represent the true prior probability vector associated with the probability distribution $\pi (x)$ . Let $\mathbf{g}$ be a map from a prior probability vector to a posterior probability vector. Then $\mathbf{g}(\mathbf{q}) = (g_1(\mathbf{q}),\dots ,g_m(\mathbf{q}))^\top$ is the probability vector associated with the posterior. Let $\mathbf{T} = (T_{1},\dots ,T_{m})^{\top}$ where $T_{j} = \sum_{i = 1}^{n}\mathbb{1}_{X_{i} = u_{j}}$ for $j = 1,\dots ,m$ . In such setting, by the definition (4) of operator $B_{n}$ , we can rewrite the operator $B_{n}$ as + +$$ +B _ {n} g _ {s} (\mathbf {q}) = \mathbb {E} \left\{g _ {s} (\mathbf {T} / n) \right\} = \sum_ {\boldsymbol {\nu} \in \bar {\Delta} _ {m}} g _ {s} (\frac {\boldsymbol {\nu}}{n}) \binom {n} {\boldsymbol {\nu}} \mathbf {q} ^ {\boldsymbol {\nu}}, +$$ + +where $\bar{\Delta}_m = \{\pmb {\nu}\in \mathbb{N}^m:\sum_{j = 1}^m\pmb {\nu}_j = n\}$ and + +$$ +\left( \begin{array}{c} n \\ \boldsymbol {\nu} \end{array} \right) = \frac {n !}{\nu_ {1} ! \cdots \nu_ {m} !}, \quad \mathbf {q} ^ {\boldsymbol {\nu}} = q _ {1} ^ {\nu_ {1}} \dots q _ {m} ^ {\nu_ {m}} . +$$ + +Additionally, let $\Delta_m = \{\mathbf{q} \in \mathbb{R}^m : q_j \geq 0, \sum_{j=1}^m q_j = 1\}$ and let $\| \cdot \|_{C^k(\Delta_m)}$ denote the $C^k(\Delta_m)$ -norm which is defined as $\| f \|_{C^k(\Delta_m)} = \sum_{\|\alpha\|_1 \leq k} \| \partial^\alpha f \|_\infty$ for any $f \in C^k(\Delta_m)$ . + +Assumption 2. $|\mathcal{X}| = m$ is finite, and $\max_{s\in [m]}\| g_s\|_{C^{2k}(\Delta_m)}\leq G$ for some constant $G$ + +The following theorem provides a systematic method for constructing an approximation of $g_{s}(\mathbf{q})$ with an error of order $\mathcal{O}(n^{-k})$ for any desired integer $k$ . + +Theorem 2. If $|\mathcal{X}| = m$ , let $\mathbf{q} = (q_1, \dots, q_m)^\top$ be the true prior probability vector associated with a discrete probability distribution and $\mathbf{T} = (T_1, \dots, T_m)^\top$ where $T_j = \sum_{i=1}^{n} \mathbb{1}_{X_i = u_j}$ for $j = 1, \dots, m$ . Under Assumption 2, the following holds for any $s \in \{1, \dots, m\}$ and any $k \in \mathbb{N}^+$ : + +$$ +\begin{array}{l} \mathbb {E} _ {X ^ {n}} \left\{D _ {n, k} \left(g _ {s}\right) \left(\mathbf {T} / n\right) \right\} - g _ {s} (\mathbf {q}) = \mathcal {O} _ {k, m, G} \left(n ^ {- k}\right), \\ \mathrm {V a r} _ {X ^ {n}} \left\{D _ {n, k} (g _ {s}) (\mathbf {T} / n) \right\} = \mathcal {O} _ {k, m, G} (n ^ {- 1}). \\ \end{array} +$$ + +Theorem 2 follows directly from the following lemma, which provides the key approximation result. + +Lemma 1. For any integers $k, m \in \mathbb{N}^+$ and any function $f \in C^{k}(\Delta_{m})$ , we have + +$$ +\| C _ {n, \lceil k / 2 \rceil} (f) - f \| _ {\infty} = \| (B _ {n} - I) ^ {\lceil k / 2 \rceil} (f) \| _ {\infty} \lesssim_ {k, m} \| f \| _ {C ^ {k} (\Delta_ {m})} n ^ {- k / 2}. +$$ + +Note that Theorem 2 holds for all mappings $\mathbf{g}$ that satisfy Assumption 2. When $\mathbf{g}$ represents a mapping corresponding to the Bayes posterior distribution, we know the exact form of $\mathbf{g}(\mathbf{q})$ . Hence, we can explore sampling schemes for Bayes-optimal mapping $\mathbf{g}$ . + +We claim that Bayes-optimal mapping $\mathbf{g}$ satisfies Assumption 2. In fact, let $\ell_s = \ell(u_s) := P_{Y|X}(y^* | u_s)$ . Using Bayes' theorem, the posterior probability of $X = u_s$ given $y^*$ is given by + +$$ +P _ {X | Y} (u _ {s} | y ^ {*}) = \frac {\ell_ {s} q _ {s}}{\sum_ {j = 1} ^ {m} \ell_ {j} q _ {j}}. +$$ + +In this case, $g_{s}(\mathbf{q}) \coloneqq \ell_{s} q_{s} / \sum_{j=1}^{m} \ell_{j} q_{j}$ for $s = 1, \dots, m$ . Since $|\mathcal{X}| = m$ is finite, we know that there exists a constant $c_{1}, c_{2} > 0$ such that $c_{1} \leq l_{j} \leq c_{2}$ for all $1 \leq j \leq m$ , which implies that $\max_{s \in [m]} \| g_{s} \|_{C^{2k}(\Delta_{m})} \leq G$ for some constant $G$ based on $k$ . + +Moreover, estimating $g_{s}(\mathbf{q})$ based on the observations of $X^{n} = (X_{1},\dots ,X_{n})$ and $y^{*}$ is sufficient to generate samples from the posterior distribution $P_{X|Y}(u_s|y^*)$ for $s = 1,\dots ,m$ . Since the exact form of $g_{s}$ is known, if we let $\widetilde{P}_{X^n}(x = u_s|y = y^*) = D_{n,k}(g_s)(\mathbf{T} / n)$ where $\mathbf{T} / n$ denotes the empirical of the training set, we obtain the following theorem. + +Theorem 3. Under Assumption 2, for any $k \in \mathbb{N}^{+}$ , if $|\mathcal{X}| = m$ is finite, then there exists an approximate posterior $\widetilde{P}_{X^n}(x|y = y^*)$ satisfies that for any $s \in \{1, \dots, m\}$ , + +$$ +\begin{array}{l} \mathbb {E} _ {X ^ {n}} \left\{\widetilde {P} _ {X ^ {n}} (x = u _ {s} | y = y ^ {*}) \right\} - P _ {X | Y} (u _ {s} | y ^ {*}) = \mathcal {O} _ {k, m, G} (n ^ {- k}), \\ \operatorname {V a r} _ {X ^ {n}} \left\{\widetilde {P} _ {X ^ {n}} (x = u _ {s} | y = y ^ {*}) \right\} = \mathcal {O} _ {k, m, G} (n ^ {- 1}). \\ \end{array} +$$ + +The proposed sampling scheme in Algorithm 1 generates $k$ samples and a linear combination of whose distributions approximates the posterior. In applications where it is desired to still generate one sample (rather than using a linear combination), we may consider a rejection sampling algorithm based on Theorem 3 to sample from $\widetilde{P}_{X^n}(x|y = y^*)$ . Let $\mathbf{T} = (T_1, \dots, T_m)^\top$ where $T_j = \sum_{i=1}^{n} \mathbb{1}_{X_i = u_j}$ for $j = 1, \dots, m$ . Then $\left(g_1(\mathbf{T}/n), \dots, g_m(\mathbf{T}/n)\right)^\top$ is the posterior probability vector associated with the plug-in posterior $\widehat{P}_{X^n}(x|y = y^*)$ and $\left(D_{n,k}(g_1)(\mathbf{T}/n), \dots, D_{n,k}(g_m)(\mathbf{T}/n)\right)^\top$ is the posterior probability vector associated with the debiased posterior $\widetilde{P}_{X^n}(x|y = y^*)$ . The rejection sampling is described in Algorithm 2. + +Algorithm 2 Rejection Sampling for Debiased Posterior $\widetilde{P}_{X^n}(x\mid y = y^*)$ +Input: Plug-in posterior $\widehat{P}_{X^n}(x\mid y = y^*)$ , debiased posterior $\widetilde{P}_{X^n}(x\mid y = y^*)$ , large enough constant $M > 0$ +Output: Sample from the debiased posterior $\widetilde{P}_{X^n}(x\mid y = y^*)$ +1: repeat +2: Sample $x^{\prime}\sim \widehat{P}_{X^{n}}(x\mid y = y^{*})$ +3: Sample $u\sim$ Uniform(0,M) +4: until $u < \frac{\widetilde{P}_{X^n}(x'|\mathcal{Y} = \mathcal{Y}^*)}{\widehat{P}_{X^n}(x'|\mathcal{Y} = \mathcal{Y}^*)}$ +5: return $x^{\prime}$ + +In Algorithm 2, + +$$ +M = \max _ {x \in \mathcal {X}} \frac {\widetilde {P} _ {X ^ {n}} (x | y = y ^ {*})}{\widehat {P} _ {X ^ {n}} (x | y = y ^ {*})} = \max _ {j} \left\{\frac {D _ {n , k} (g _ {j}) (\mathbf {T} / n)}{g _ {j} (\mathbf {T} / n)} \right\} +$$ + +is the upper bound of the ratio of the debiased posterior to the plug-in posterior. + +# 5 Experiments + +In this section, we provide numerical experiments to illustrate the debiasing framework for posterior approximation under the binary prior case and the Gaussian mixture prior case. + +Binary prior case. Suppose that $\mathcal{X} = \{0,1\}$ and $X\sim \mathrm{Bern}(q)$ for some unknown prior $q\in (0,1)$ . Let $\alpha = \alpha (y^{*})\coloneqq P_{Y|X}(y^{*}|1) / P_{Y|X}(y^{*}|0)$ be the likelihood ratio. Then the posterior distribution is given by $X|Y\sim \mathrm{Bern}\bigl (\alpha q / (\alpha q + 1 - q)\bigr)$ . We estimate $g(q)\coloneqq \alpha q / (\alpha q + 1 - q)$ based on the observations of $X^n$ and $y^{*}$ . + +Proposition 1 provides a debiased approximation as a special case of Theorem 2 when $|\mathcal{X}| = 2$ . + +Proposition 1. Let $T = \sum_{i=1}^{n} X_i$ . For $k = 1,2,3,4$ , we have + +$$ +\mathbb {E} _ {X ^ {n}} \left\{D _ {n, k} g (T / n) \right\} - g (q) = \mathcal {O} \left(n ^ {- k}\right), +$$ + +where $D_{n,k} = \sum_{j=0}^{k-1} \binom{k}{j+1} (-1)^j B_n^j$ and $B_n(g)(x) = \sum_{k=0}^{n} g\left(\frac{k}{n}\right)\binom{n}{k} x^k (1-x)^{n-k}$ is the Bernstein polynomial approximation of $g$ . + +In the proof of Theorem 2, we notice that for any $k \in \mathbb{N}^{+}$ , $\mathbb{E}_{X^n}\{D_{n,k}g(T / n)\} = C_{n,k}g(q)$ , which allows Proposition 1 to be verified in closed form. To validate this result numerically, we consider two parameter settings: in the first experiment we set $q = 0.4$ , $y^{*} = 2$ , and $Y|X \sim \mathcal{N}(X,1)$ , while in the second we set $q = 3 / 11$ , $y^{*} = 1$ , and $Y|X \sim \mathcal{N}(X,1 / 4)$ . + +For both settings, we examine the convergence rate of the debiased estimators $D_{n,k}g(T / n)$ for $k = 1,2,3,4$ . The results are shown in log-log plots in Figure 1, where the vertical axis represents the logarithm of the absolute error and the horizontal axis represents the logarithm of the sample size $n$ . Reference lines with slopes corresponding to $n^{-1}, n^{-2}, n^{-3}$ , and $n^{-4}$ are included for comparison. + +Gaussian mixture prior case. Suppose that $X \sim \frac{1}{2}\mathcal{N}(0,1) + \frac{1}{2}\mathcal{N}(1,1)$ and $Y = X + \xi$ where $\xi \sim \mathcal{N}(0,1/16)$ . Additionally, let $y^{*} = 0.8$ and $A = \{x : x \geq 0.5\}$ . In this case, we validate the theoretical convergence rate + +$$ +\left| \mathbb {E} _ {X ^ {n}} \left\{D _ {n, k} f (\hat {\pi}) (A) \right\} - f (\pi) (A) \right| = \mathcal {O} (n ^ {- k}). +$$ + +Since $\mathbb{E}_{X^n}\{D_{n,k}f(\hat{\pi})(A)\}$ does not have a closed-form expression, we approximate it using Monte Carlo simulation. To ensure that the Monte Carlo error is negligible compared to the bias $\mathcal{O}(n^{-k})$ , we select the number of Monte Carlo samples $N$ such that $N \gg n^{2k-1}$ . In practice, we run simulations for $k = 1$ and $k = 2$ and set $N = n^3$ for $k = 1$ and $N = n^4$ for $k = 2$ . + +The results are shown in Figure 2. The figure presents log-log plots where the vertical axis represents the logarithm of the absolute error or of the variance and the horizontal axis represents the logarithm of the sample size $n$ . For both $k = 1$ and $k = 2$ , the observed convergence rates align closely with the theoretical predictions. + +![](images/a430ce170b8a13e8f3060c07dfdc6e8f215c605047125cf7250d9e1831231f62.jpg) +(a) $q = 0.4,y^{*} = 2,Y|X\sim \mathcal{N}(X,1)$ + +![](images/a0b749a95d3d10c35a4a9c09af6ab4648c3cf07a62f4a48aae5679f90fb51402.jpg) +(b) $q = 3 / 11,y^{*} = 1,Y|X\sim \mathcal{N}(X,1 / 4)$ + +![](images/f64041a286142056bf8b35c6c66df8e671e380f4335298c90a050b6980814627.jpg) +Figure 1: Convergence of plug-in and debiased estimators in the binary prior case. The plot compares the approximation error of $\bar{D}_{n,k}g(T / n)$ ( $k = 1,2,3,4$ ) against $n$ . Reference lines with slopes corresponding to $n^{-1}, n^{-2}, n^{-3}$ , and $n^{-4}$ are included to highlight the convergence rates. +(a) Bias convergence rate +Figure 2: Convergence of debiased estimators in the Gaussian mixture prior case with $X \sim \frac{1}{2}\mathcal{N}(0,1) + \frac{1}{2}\mathcal{N}(1,1)$ , $Y = X + \xi$ , $\xi \sim \mathcal{N}(0,1/16)$ , $y^{*} = 0.8$ , and $A = \{x : x \geq 0.5\}$ . (a) shows the bias decay of $D_{n,k}f(\hat{\pi})(A)$ for $k = 1, 2$ , with reference lines of slopes corresponding to $n^{-1}$ and $n^{-2}$ included for comparison. (b) shows the corresponding variance decay, alongside a reference slope corresponding to $n^{-1}$ . + +![](images/bcb45048e2e74ee193c7664bd3a2853055e0f349b65c87ca37de6414f028f669.jpg) +(b) Variance convergence rate + +# 6 Conclusion + +We introduced a general framework for constructing a debiased posterior approximation through observed samples $D$ and the known likelihood $P_{Y|X}$ when the prior distribution is unknown. Here, a naive strategy that directly plugs the empirical distribution into the Bayes formula or a generative model has a bias, because the likelihood is nonconstant, inducing a distribution shift, and the map from the prior to posterior is nonlinear. It can be shown that the plug-in approach generates $\hat{X}$ with bias $\| \mathbb{E}_D(P_{\hat{X} |Y = y^*,D}) - P_{X|Y = y^*}\|_{\mathrm{TV}} = \mathcal{O}(n^{-1})$ and variance $\operatorname{Var}_D(P_{\hat{X} |Y = y^*,D}) = \mathcal{O}(n^{-1})$ . In contrast, our proposed debiasing framework achieves arbitrarily high-order bias rate of $\mathcal{O}(n^{-k})$ for any integer $k$ , while maintaining the order of magnitude of the variance. Our framework is black-box in the sense that we only need to resample the training data and feed it into a given black-box conditional generative model. In particular, we provide a rigorous proof for the Bayes-optimal sampler $f$ under the continuous prior setting and for a broad class of samplers $f$ with a continuous $2k$ -th derivative, as specified in Assumption 2, under the discrete prior setting. We expect that the proposed debiasing framework could work for general $f$ and will support future developments in bias-corrected posterior estimation and conditional sampling. + +# Acknowledgments + +This research was supported in part by NSF Grant DMS-2515510. + +# References + +[1] Anurag Ajay, Yilun Du, Abhi Gupta, Joshua B. Tenenbaum, Tommi S. Jaakkola, and Pulkit Agrawal. Is conditional generative modeling all you need for decision making? In The Eleventh International Conference on Learning Representations, 2023. +[2] Kevin Black, Michael Janner, Yilun Du, Ilya Kostrikov, and Sergey Levine. Training diffusion models with reinforcement learning. In The Twelfth International Conference on Learning Representations, 2024. +[3] T. Tony Cai and Mark G. Low. Testing composite hypotheses, hermite polynomials and optimal estimation of a nonsmooth functional. The Annals of Statistics, 39(2):1012-1041, 2011. +[4] Nicholas Carlini, Jamie Hayes, Milad Nasr, Matthew Jagielski, Vikash Sehwag, Florian Tramère, Borja Balle, Daphne Ippolito, and Eric Wallace. Extracting training data from diffusion models. In USENIX Security Symposium, pages 5253-5270, 2023. +[5] Hyungjin Chung, Jeongsol Kim, Michael Thompson McCann, Marc Louis Klasky, and Jong Chul Ye. Diffusion posterior sampling for general noisy inverse problems. In The Eleventh International Conference on Learning Representations, 2023. +[6] Kyle Cranmer, Johann Brehmer, and Gilles Louppe. The frontier of simulation-based inference. Proceedings of the National Academy of Sciences, 117(48):30055-30062, 2020. +[7] Yaru Hao, Zewen Chi, Li Dong, and Furu Wei. Optimizing prompts for text-to-image generation. In Advances in Neural Information Processing Systems, 2023. +[8] Michael Janner, Yilun Du, Joshua Tenenbaum, and Sergey Levine. Planning with diffusion for flexible behavior synthesis. In Proceedings of the 39th International Conference on Machine Learning, volume 162 of Proceedings of Machine Learning Research, pages 9902–9915, 2022. +[9] Bahjat Kawar, Michael Elad, Stefano Ermon, and Jiaming Song. Denoising diffusion restoration models. In Advances in Neural Information Processing Systems, 2022. +[10] Dong-Hyun Lee. Pseudo-label : The simple and efficient semi-supervised learning method for deep neural networks. In Workshop on Challenges in Representation Learning, ICML, 2013. +[11] Haoying Li, Yifan Yang, Meng Chang, Shiqi Chen, Huajun Feng, Zhihai Xu, Qi Li, and Yueting Chen. Srdiff: Single image super-resolution with diffusion probabilistic models. Neurocomputing, 479:47-59, 2022. +[12] Qiang Liu and Jason Lee. Black-box Importance Sampling. In Proceedings of the 20th International Conference on Artificial Intelligence and Statistics, volume 54, pages 952-961, 2017. +[13] Alexander Quinn Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. GLIDE: Towards photorealistic image generation and editing with text-guided diffusion models. In Proceedings of the 39th International Conference on Machine Learning, 2022. +[14] Sebastian Nowozin. Debiasing evidence approximations: On importance-weighted autoencoders and jackknife variational inference. In International Conference on Learning Representations, 2018. +[15] Chris J. Oates, Mark Girolami, and Nicolas Chopin. Control Functionals for Monte Carlo Integration. Journal of the Royal Statistical Society Series B: Statistical Methodology, 79(3): 695-718, 05 2016. + +[16] Kazusato Oko, Shunta Akiyama, and Taiji Suzuki. Diffusion models are minimax optimal distribution estimators. In International Conference on Machine Learning, pages 26517-26582, 2023. +[17] Maurice Henry Quenouille. Notes on bias in estimation. Biometrika, 43(3-4):353-360, 1956. +[18] Louis Sharrock, Jack Simons, Song Liu, and Mark Beaumont. Sequential neural score estimation: Likelihood-free inference with conditional score based diffusion models. In Proceedings of the 41st International Conference on Machine Learning, volume 235, pages 44565-44602, 2024. +[19] Gowthami Somepalli, Vasu Singla, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Understanding and mitigating copying in diffusion models. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. +[20] Kaizheng Wang. Pseudo-labeling for kernel ridge regression under covariate shift. arXiv preprint arXiv:2302.10160, 2024. +[21] Qixun Wang, Xu Bai, Haofan Wang, Zekui Qin, Anthony Chen, Huaxia Li, Xu Tang, and Yao Hu. Instantid: Zero-shot identity-preserving generation in seconds. arXiv preprint arXiv:2401.07519, 2024. +[22] Andre Wibisono, Yihong Wu, and Kaylee Yingxi Yang. Optimal score estimation via empirical bayes smoothing. In The Thirty Seventh Annual Conference on Learning Theory, pages 4958-4991, 2024. +[23] Yihong Wu and Pengkun Yang. Polynomial methods in statistical inference: Theory and practice. Foundations and Trends® in Communications and Information Theory, 17(4):402-586, 2020. +[24] Ling Yang, Zhaochen Yu, Chenlin Meng, Minkai Xu, Stefano Ermon, and Bin Cui. Mastering text-to-image diffusion: Recaptioning, planning, and generating with multimodal llms. In *Forty-first International Conference on Machine Learning*, 2024. +[25] TaeHo Yoon, Joo Young Choi, Sehyun Kwon, and Ernest K. Ryu. Diffusion probabilistic models generalize when they fail to memorize. In ICML 2023 Workshop on Structured Probabilistic Inference & Generative Modeling, 2023. +[26] Hui Yuan, Kaixuan Huang, Chengzhuo Ni, Minshuo Chen, and Mengdi Wang. Reward-directed conditional diffusion: Provable distribution estimation and reward improvement. In Advances in Neural Information Processing Systems, volume 36, pages 60599–60635, 2023. +[27] Andrew Zammit-Mangion, Matthew Sainsbury-Dale, and Raphaël Huser. Neural methods for amortized inference. Annual Review of Statistics and Its Application, 12, 2024. +[28] Kaihong Zhang, Heqi Yin, Feng Liang, and Jingbo Liu. Minimax optimality of score-based diffusion models: Beyond the density lower bound assumptions. In International Conference on Machine Learning, 2024. + +# NeurIPS Paper Checklist + +# 1. Claims + +Question: Do the main claims made in the abstract and introduction accurately reflect the paper's contributions and scope? + +Answer: [Yes] + +Justification: The model setting and the assumptions of our claim are clearly stated in abstract and introduction. Our main contributions are stated in introduction. + +Guidelines: + +- The answer NA means that the abstract and introduction do not include the claims made in the paper. +- The abstract and/or introduction should clearly state the claims made, including the contributions made in the paper and important assumptions and limitations. A No or NA answer to this question will not be perceived well by the reviewers. +- The claims made should match theoretical and experimental results, and reflect how much the results can be expected to generalize to other settings. +- It is fine to include aspirational goals as motivation as long as it is clear that these goals are not attained by the paper. + +# 2. Limitations + +Question: Does the paper discuss the limitations of the work performed by the authors? + +Answer: [Yes] + +Justification: We provide a black-box debiasng framework since we only provide the rigorous proof for the Bayes-optimal sampler under continuous prior setting and a broad class of samplers with a continuous $2k$ -th derivative under discrete prior setting. We expect the framework could work for general samplers $f$ . + +Guidelines: + +- The answer NA means that the paper has no limitation while the answer No means that the paper has limitations, but those are not discussed in the paper. +- The authors are encouraged to create a separate "Limitations" section in their paper. +- The paper should point out any strong assumptions and how robust the results are to violations of these assumptions (e.g., independence assumptions, noiseless settings, model well-specification, asymptotic approximations only holding locally). The authors should reflect on how these assumptions might be violated in practice and what the implications would be. +- The authors should reflect on the scope of the claims made, e.g., if the approach was only tested on a few datasets or with a few runs. In general, empirical results often depend on implicit assumptions, which should be articulated. +- The authors should reflect on the factors that influence the performance of the approach. For example, a facial recognition algorithm may perform poorly when image resolution is low or images are taken in low lighting. Or a speech-to-text system might not be used reliably to provide closed captions for online lectures because it fails to handle technical jargon. +- The authors should discuss the computational efficiency of the proposed algorithms and how they scale with dataset size. +- If applicable, the authors should discuss possible limitations of their approach to address problems of privacy and fairness. +- While the authors might fear that complete honesty about limitations might be used by reviewers as grounds for rejection, a worse outcome might be that reviewers discover limitations that aren't acknowledged in the paper. The authors should use their best judgment and recognize that individual actions in favor of transparency play an important role in developing norms that preserve the integrity of the community. Reviewers will be specifically instructed to not penalize honesty concerning limitations. + +# 3. Theory assumptions and proofs + +Question: For each theoretical result, does the paper provide the full set of assumptions and a complete (and correct) proof? + +Answer: [Yes] + +Justification: All assumptions are clearly stated or referenced in the statement of our theorems and lemmas. The proofs appear in the main paper and the supplemental material. + +Guidelines: + +- The answer NA means that the paper does not include theoretical results. +- All the theorems, formulas, and proofs in the paper should be numbered and cross-referenced. +- All assumptions should be clearly stated or referenced in the statement of any theorems. +- The proofs can either appear in the main paper or the supplemental material, but if they appear in the supplemental material, the authors are encouraged to provide a short proof sketch to provide intuition. +- Inversely, any informal proof provided in the core of the paper should be complemented by formal proofs provided in appendix or supplemental material. +- Theorems and Lemmas that the proof relies upon should be properly referenced. + +# 4. Experimental result reproducibility + +Question: Does the paper fully disclose all the information needed to reproduce the main experimental results of the paper to the extent that it affects the main claims and/or conclusions of the paper (regardless of whether the code and data are provided or not)? + +Answer: [Yes] + +Justification: We describe our experiments in the Experiment Section in detail. + +Guidelines: + +- The answer NA means that the paper does not include experiments. + +- If the paper includes experiments, a No answer to this question will not be perceived well by the reviewers: Making the paper reproducible is important, regardless of whether the code and data are provided or not. + +- If the contribution is a dataset and/or model, the authors should describe the steps taken to make their results reproducible or verifiable. + +- Depending on the contribution, reproducibility can be accomplished in various ways. For example, if the contribution is a novel architecture, describing the architecture fully might suffice, or if the contribution is a specific model and empirical evaluation, it may be necessary to either make it possible for others to replicate the model with the same dataset, or provide access to the model. In general, releasing code and data is often one good way to accomplish this, but reproducibility can also be provided via detailed instructions for how to replicate the results, access to a hosted model (e.g., in the case of a large language model), releasing of a model checkpoint, or other means that are appropriate to the research performed. + +- While NeurIPS does not require releasing code, the conference does require all submissions to provide some reasonable avenue for reproducibility, which may depend on the nature of the contribution. For example + +(a) If the contribution is primarily a new algorithm, the paper should make it clear how to reproduce that algorithm. +(b) If the contribution is primarily a new model architecture, the paper should describe the architecture clearly and fully. +(c) If the contribution is a new model (e.g., a large language model), then there should either be a way to access this model for reproducing the results or a way to reproduce the model (e.g., with an open-source dataset or instructions for how to construct the dataset). +(d) We recognize that reproducibility may be tricky in some cases, in which case authors are welcome to describe the particular way they provide for reproducibility. In the case of closed-source models, it may be that access to the model is limited in some way (e.g., to registered users), but it should be possible for other researchers to have some path to reproducing or verifying the results. + +# 5. Open access to data and code + +Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material? + +Answer: [Yes] + +Justification: We used the simulation data in the experiment. + +Guidelines: + +- The answer NA means that paper does not include experiments requiring code. +- Please see the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details. +- While we encourage the release of code and data, we understand that this might not be possible, so "No" is an acceptable answer. Papers cannot be rejected simply for not including code, unless this is central to the contribution (e.g., for a new open-source benchmark). +- The instructions should contain the exact command and environment needed to run to reproduce the results. See the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details. +- The authors should provide instructions on data access and preparation, including how to access the raw data, preprocessed data, intermediate data, and generated data, etc. +- The authors should provide scripts to reproduce all experimental results for the new proposed method and baselines. If only a subset of experiments are reproducible, they should state which ones are omitted from the script and why. +- At submission time, to preserve anonymity, the authors should release anonymized versions (if applicable). +- Providing as much information as possible in supplemental material (appended to the paper) is recommended, but including URLs to data and code is permitted. + +# 6. Experimental setting/details + +Question: Does the paper specify all the training and test details (e.g., data splits, hyperparameters, how they were chosen, type of optimizer, etc.) necessary to understand the results? + +Answer: [Yes] + +Justification: We provide the setting of our simulation in the Experiment section. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The experimental setting should be presented in the core of the paper to a level of detail that is necessary to appreciate the results and make sense of them. +- The full details can be provided either with the code, in appendix, or as supplemental material. + +# 7. Experiment statistical significance + +Question: Does the paper report error bars suitably and correctly defined or other appropriate information about the statistical significance of the experiments? + +Answer: [No] + +Justification: Our experiments do not include any error bars. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The authors should answer "Yes" if the results are accompanied by error bars, confidence intervals, or statistical significance tests, at least for the experiments that support the main claims of the paper. +- The factors of variability that the error bars are capturing should be clearly stated (for example, train/test split, initialization, random drawing of some parameter, or overall run with given experimental conditions). + +- The method for calculating the error bars should be explained (closed form formula, call to a library function, bootstrap, etc.) +- The assumptions made should be given (e.g., Normally distributed errors). +- It should be clear whether the error bar is the standard deviation or the standard error of the mean. +- It is OK to report 1-sigma error bars, but one should state it. The authors should preferably report a 2-sigma error bar than state that they have a $96\%$ CI, if the hypothesis of Normality of errors is not verified. +- For asymmetric distributions, the authors should be careful not to show in tables or figures symmetric error bars that would yield results that are out of range (e.g. negative error rates). +- If error bars are reported in tables or plots, The authors should explain in the text how they were calculated and reference the corresponding figures or tables in the text. + +# 8. Experiments compute resources + +Question: For each experiment, does the paper provide sufficient information on the computer resources (type of compute workers, memory, time of execution) needed to reproduce the experiments? + +Answer: [No] + +Justification: Our experiment is a simple simulation for binary case. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The paper should indicate the type of compute workers CPU or GPU, internal cluster, or cloud provider, including relevant memory and storage. +- The paper should provide the amount of compute required for each of the individual experimental runs as well as estimate the total compute. +- The paper should disclose whether the full research project required more compute than the experiments reported in the paper (e.g., preliminary or failed experiments that didn't make it into the paper). + +# 9. Code of ethics + +Question: Does the research conducted in the paper conform, in every respect, with the NeurIPS Code of Ethics https://neurips.cc/public/EthicsGuidelines? + +Answer: [Yes] + +Justification: The research conducted in the paper conforms with the NeurIPS Code of Ethics. + +Guidelines: + +- The answer NA means that the authors have not reviewed the NeurIPS Code of Ethics. +- If the authors answer No, they should explain the special circumstances that require a deviation from the Code of Ethics. +- The authors should make sure to preserve anonymity (e.g., if there is a special consideration due to laws or regulations in their jurisdiction). + +# 10. Broader impacts + +Question: Does the paper discuss both potential positive societal impacts and negative societal impacts of the work performed? + +Answer: [NA] + +Justification: Our main contribution is constructing a debiased approximation of the posterior distribution which does not have immediate societal impact. + +Guidelines: + +- The answer NA means that there is no societal impact of the work performed. +- If the authors answer NA or No, they should explain why their work has no societal impact or why the paper does not address societal impact. + +- Examples of negative societal impacts include potential malicious or unintended uses (e.g., disinformation, generating fake profiles, surveillance), fairness considerations (e.g., deployment of technologies that could make decisions that unfairly impact specific groups), privacy considerations, and security considerations. +- The conference expects that many papers will be foundational research and not tied to particular applications, let alone deployments. However, if there is a direct path to any negative applications, the authors should point it out. For example, it is legitimate to point out that an improvement in the quality of generative models could be used to generate deepfakes for disinformation. On the other hand, it is not needed to point out that a generic algorithm for optimizing neural networks could enable people to train models that generate Deepfakes faster. +- The authors should consider possible harms that could arise when the technology is being used as intended and functioning correctly, harms that could arise when the technology is being used as intended but gives incorrect results, and harms following from (intentional or unintentional) misuse of the technology. +- If there are negative societal impacts, the authors could also discuss possible mitigation strategies (e.g., gated release of models, providing defenses in addition to attacks, mechanisms for monitoring misuse, mechanisms to monitor how a system learns from feedback over time, improving the efficiency and accessibility of ML). + +# 11. Safeguards + +Question: Does the paper describe safeguards that have been put in place for responsible release of data or models that have a high risk for misuse (e.g., pretrained language models, image generators, or scraped datasets)? + +Answer: [NA] + +Justification: + +Guidelines: + +- The answer NA means that the paper poses no such risks. +- Released models that have a high risk for misuse or dual-use should be released with necessary safeguards to allow for controlled use of the model, for example by requiring that users adhere to usage guidelines or restrictions to access the model or implementing safety filters. +- Datasets that have been scraped from the Internet could pose safety risks. The authors should describe how they avoided releasing unsafe images. +- We recognize that providing effective safeguards is challenging, and many papers do not require this, but we encourage authors to take this into account and make a best faith effort. + +# 12. Licenses for existing assets + +Question: Are the creators or original owners of assets (e.g., code, data, models), used in the paper, properly credited and are the license and terms of use explicitly mentioned and properly respected? + +Answer: [NA] + +Justification: + +Guidelines: + +- The answer NA means that the paper does not use existing assets. +- The authors should cite the original paper that produced the code package or dataset. +- The authors should state which version of the asset is used and, if possible, include a URI. +- The name of the license (e.g., CC-BY 4.0) should be included for each asset. +- For scraped data from a particular source (e.g., website), the copyright and terms of service of that source should be provided. +- If assets are released, the license, copyright information, and terms of use in the package should be provided. For popular datasets, paperswithcode.com/datasets has curated licenses for some datasets. Their licensing guide can help determine the license of a dataset. + +- For existing datasets that are re-packaged, both the original license and the license of the derived asset (if it has changed) should be provided. +- If this information is not available online, the authors are encouraged to reach out to the asset's creators. + +# 13. New assets + +Question: Are new assets introduced in the paper well documented and is the documentation provided alongside the assets? + +Answer: [NA] + +Justification: + +Guidelines: + +- The answer NA means that the paper does not release new assets. +- Researchers should communicate the details of the dataset/code/model as part of their submissions via structured templates. This includes details about training, license, limitations, etc. +- The paper should discuss whether and how consent was obtained from people whose asset is used. +- At submission time, remember to anonymize your assets (if applicable). You can either create an anonymized URL or include an anonymized zip file. + +# 14. Crowdsourcing and research with human subjects + +Question: For crowdsourcing experiments and research with human subjects, does the paper include the full text of instructions given to participants and screenshots, if applicable, as well as details about compensation (if any)? + +Answer: [NA] + +Justification: + +Guidelines: + +- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. +- Including this information in the supplemental material is fine, but if the main contribution of the paper involves human subjects, then as much detail as possible should be included in the main paper. +- According to the NeurIPS Code of Ethics, workers involved in data collection, curation, or other labor should be paid at least the minimum wage in the country of the data collector. + +# 15. Institutional review board (IRB) approvals or equivalent for research with human subjects + +Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals (or an equivalent approval/review based on the requirements of your country or institution) were obtained? + +Answer: [NA] + +Justification: + +Guidelines: + +- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. +- Depending on the country in which research is conducted, IRB approval (or equivalent) may be required for any human subjects research. If you obtained IRB approval, you should clearly state this in the paper. +- We recognize that the procedures for this may vary significantly between institutions and locations, and we expect authors to adhere to the NeurIPS Code of Ethics and the guidelines for their institution. +- For initial submissions, do not include any information that would break anonymity (if applicable), such as the institution conducting the review. + +# 16. Declaration of LLM usage + +Question: Does the paper describe the usage of LLMs if it is an important, original, or non-standard component of the core methods in this research? Note that if the LLM is used only for writing, editing, or formatting purposes and does not impact the core methodology, scientific rigorousness, or originality of the research, declaration is not required. + +Answer: [NA] + +Justification: + +Guidelines: + +- The answer NA means that the core method development in this research does not involve LLMs as any important, original, or non-standard components. +- Please refer to our LLM policy (https://neurips.cc/Conferences/2025/LLM) for what should or should not be described. + +# A Proof of Theorem 1 + +Proof of Theorem 1. We begin by introducing notations that facilitates the analysis. Define + +$$ +\mu = \int_ {\mathcal {X}} \ell (x) \pi (d x), \quad \mu_ {A} = \int_ {A} \ell (x) \pi (d x). +$$ + +Let $\hat{\pi}^{(0)} = \pi, \hat{\pi}^{(1)} = \hat{\pi}$ and set $(X_{1}^{(0)},\ldots,X_{n}^{(0)}) \equiv (X_{1},\ldots,X_{n})$ . For $j = 1,\dots,k$ , we define $\hat{\pi}^{(j)}$ as the empirical measure of $n$ i.i.d. samples $(X_{1}^{(j-1)},\ldots,X_{n}^{(j-1)})$ drawn from the measure $\hat{\pi}^{(j-1)}$ . Furthermore, for each $j = 0,\ldots,k$ , define + +$$ +e _ {n} ^ {(j)} = n ^ {- 1} \sum_ {i = 1} ^ {n} \Bigl \{\ell (X _ {i} ^ {(j)}) - \mu \Bigr \}, \quad \mu_ {A} ^ {(j)} = n ^ {- 1} \sum_ {i = 1} ^ {n} \ell (X _ {i} ^ {(j)}) \delta_ {X _ {i} ^ {(j)}} (A). +$$ + +Let + +$$ +C _ {n, k} = \sum_ {j = 1} ^ {k} \binom {k} {j} (- 1) ^ {j - 1} B _ {n} ^ {j}, +$$ + +so that it suffices to show that + +$$ +C _ {n, k} f (\pi) (A) - f (\pi) (A) = \mathcal {O} _ {L _ {1}, L _ {2}, k} \left(n ^ {- k}\right) \tag {9} +$$ + +since $B_{n}D_{n,k} = C_{n,k}$ + +The Radon-Nikodym derivative of $f(\pi)$ with respect to $\pi$ is + +$$ +\frac {d f (\pi)}{d \pi} (x) = \frac {\ell (x)}{\int_ {\mathcal {X}} \ell (x) \pi (d x)}. +$$ + +For the empirical measure $\hat{\pi}$ , the corresponding Radon-Nikodym derivative of $f(\hat{\pi})$ with respect to $\hat{\pi}$ takes the form + +$$ +\begin{array}{l} \frac {d f (\hat {\pi})}{d \hat {\pi}} (x) = \left\{ \begin{array}{l l} \frac {\ell (x)}{\int_ {\mathcal {X}} \ell (x) \hat {\pi} (d x)}, & \text {i f} x \in \left\{X _ {1} ^ {(0)}, \ldots , X _ {n} ^ {(0)} \right\}, \\ 0, & \text {o t h e r w i s e}, \end{array} \right. \\ = \left\{ \begin{array}{l l} \frac {\ell (x)}{n ^ {- 1} \sum_ {i = 1} ^ {n} \ell (X _ {i} ^ {(0)})}, & \text {i f} x \in \left\{X _ {1} ^ {(0)}, \ldots , X _ {n} ^ {(0)} \right\}, \\ 0, & \text {o t h e r w i s e}. \end{array} \right. \\ \end{array} +$$ + +Consequently, + +$$ +\begin{array}{l} B _ {n} f (\pi) (A) = \mathbb {E} \left\{f (\hat {\pi}) (A) \right\} \\ = \mathbb {E} \left\{\int_ {A} \frac {d f (\hat {\pi})}{d \hat {\pi}} (x) \hat {\pi} (d x) \right\} \\ = \mathbb {E} \left\{\int_ {A} \frac {\ell (x)}{n ^ {- 1} \sum_ {i = 1} ^ {n} l \left(X _ {i} ^ {(0)}\right)} \hat {\pi} (d x) \right\} \\ = \mathbb {E} \left\{\frac {n ^ {- 1} \sum_ {i = 1} ^ {n} \ell (X _ {i} ^ {(0)}) \delta_ {X _ {i} ^ {(0)}} (A)}{n ^ {- 1} \sum_ {i = 1} ^ {n} \ell (X _ {i} ^ {(0)})} \right\}. \\ \end{array} +$$ + +Moreover, by the definition of $B_{n}$ and iterated conditioning, we have $\mathbb{E}\left\{f(\hat{\pi}^{(j)})(A)\right\} = \mathbb{E}\left[\mathbb{E}\left\{f(\hat{\pi}^{(j)})(A)|\hat{\pi}^{(j - 1)}\right\} \right] = \mathbb{E}\left\{B_{n}f(\hat{\pi}^{(j - 1)})(A)\right\} = \dots = \mathbb{E}\left\{B_{n}^{j - 1}f(\hat{\pi}^{(1)})(A)\right\} = B_{n}^{j}f(\pi)(A)$ . + +By the same logic, for any $j = 2,\ldots ,k$ , we have + +$$ +\mathbb {E} \left\{f (\hat {\pi} ^ {(j)}) (A) \right\} = \mathbb {E} \left\{\frac {n ^ {- 1} \sum_ {i = 1} ^ {n} \ell \left(X _ {i} ^ {(j - 1)}\right) \delta_ {X _ {i} ^ {(j - 1)}} (A)}{n ^ {- 1} \sum_ {i = 1} ^ {n} \ell \left(X _ {i} ^ {(j - 1)}\right)} \right\}. +$$ + +Thus, + +$$ +\begin{array}{l} C _ {n, k} f (\pi) (A) = \sum_ {j = 1} ^ {k} \binom {k} {j} (- 1) ^ {j - 1} B _ {n} ^ {j} f (\pi) (A) \\ = \sum_ {j = 1} ^ {k} \binom {k} {j} (- 1) ^ {j - 1} \mathbb {E} \left\{f (\hat {\pi} ^ {(j)}) (A) \right\} \\ = \sum_ {j = 1} ^ {k} {\binom {k} {j}} (- 1) ^ {j - 1} \mathbb {E} \left\{\frac {n ^ {- 1} \sum_ {i = 1} ^ {n} \ell (X _ {i} ^ {(j - 1)}) \delta_ {X _ {i} ^ {(j - 1)}} (A)}{n ^ {- 1} \sum_ {i = 1} ^ {n} \ell (X _ {i} ^ {(j - 1)})} \right\} \\ = \sum_ {j = 1} ^ {k} {\binom {k} {j}} (- 1) ^ {j - 1} \mathbb {E} \left(\frac {\mu_ {A} ^ {(j - 1)}}{e _ {n} ^ {(j - 1)} + \mu}\right). \\ \end{array} +$$ + +Then (9) holds if + +$$ +\sum_ {j = 1} ^ {k} \binom {k} {j} (- 1) ^ {j - 1} \mathbb {E} \left(\frac {\mu_ {A} ^ {(j - 1)}}{e _ {n} ^ {(j - 1)} + \mu}\right) = \frac {\mu_ {A}}{\mu} + \mathcal {O} _ {L _ {1}, L _ {2}, k} \left(n ^ {- k}\right). \tag {10} +$$ + +Now we show that (10) holds. By using the Taylor expansion of $1 / (e_n^{(j - 1)} + \mu)$ , we have + +$$ +\frac {1}{e _ {n} ^ {(j - 1)} + \mu} = \frac {1}{\mu} + \sum_ {r = 1} ^ {2 k - 1} \frac {(- 1) ^ {r}}{\mu^ {r + 1}} (e _ {n} ^ {(j - 1)}) ^ {r} + \frac {(e _ {n} ^ {(j - 1)}) ^ {2 k}}{\xi^ {2 k + 1}}, +$$ + +where $\xi$ lies between $e_n^{(j - 1)} + \mu$ and $\mu$ . + +Since $\min \{e_n^{(j - 1)} + \mu ,\mu \} \geq L_1$ , we have $1 / \xi^{2k + 1}\leq L_1^{-2k - 1}$ . Additionally, $L_{1}\leq l(X_{i}^{(j - 1)})\leq L_{2}$ and Hoeffding's inequality implies that + +$$ +\mathbb {P} (| n e _ {n} ^ {(j - 1)} | > t) \leq 2 \exp \left\{- \frac {2 t ^ {2}}{n (L _ {2} - L _ {1}) ^ {2}} \right\} +$$ + +for all $t > 0$ , which is equivalent to + +$$ +\mathbb {P} (| e _ {n} ^ {(j - 1)} | > t) \leq 2 \exp \left\{- \frac {2 n t ^ {2}}{(L _ {2} - L _ {1}) ^ {2}} \right\}. +$$ + +Then + +$$ +\begin{array}{l} \mathbb {E} \left(\left| e _ {n} ^ {(j - 1)} \right| ^ {2 k}\right) = \int_ {0} ^ {\infty} \mathbb {P} \left(\left| e _ {n} ^ {(j - 1)} \right| ^ {2 k} > t\right) d t \\ = \int_ {0} ^ {\infty} \mathbb {P} \left(\left| e _ {n} ^ {(j - 1)} \right| > t ^ {1 / 2 k}\right) d t \\ \leq \int_ {0} ^ {\infty} 2 k u ^ {k - 1} \exp \left\{- \frac {2 n u}{(L _ {2} - L _ {1}) ^ {2}} \right\} d u \\ = 2 k n ^ {- k} \int_ {0} ^ {\infty} \exp \left\{- \frac {2 v}{(L _ {2} - L _ {1}) ^ {2}} \right\} v ^ {k - 1} d v \\ = \mathcal {O} _ {L _ {1}, L _ {2}, k} (n ^ {- k}). \\ \end{array} +$$ + +Therefore, we have + +$$ +\mathbb {E} \left(\frac {\mu_ {A} ^ {(j - 1)}}{e _ {n} ^ {(j - 1)} + \mu}\right) = \frac {\mu_ {A}}{\mu} + \mathbb {E} \left\{\sum_ {r = 1} ^ {2 k - 1} \frac {(- 1) ^ {r}}{\mu^ {r + 1}} \mu_ {A} ^ {(j - 1)} (e _ {n} ^ {(j - 1)}) ^ {r} \right\} + \mathcal {O} _ {L _ {1}, L _ {2}, k} (n ^ {- k}), +$$ + +which implies that the L.H.S. of (10) can be written as + +$$ +\frac {\mu_ {A}}{\mu} + \sum_ {r = 1} ^ {2 k - 1} \frac {(- 1) ^ {r}}{\mu^ {r + 1}} \left[ \sum_ {j = 1} ^ {k} \binom {k} {j} (- 1) ^ {j - 1} \mathbb {E} \left\{\mu_ {A} ^ {(j - 1)} (e _ {n} ^ {(j - 1)}) ^ {r} \right\} \right] + \mathcal {O} _ {L _ {1}, L _ {2}, k} (n ^ {- k}). +$$ + +Thus to prove the bound on the bias, it remains to show that for any given $k$ and $1 \leq r \leq 2k - 1$ , + +$$ +B _ {k, r} := \sum_ {j = 1} ^ {k} \binom {k} {j} (- 1) ^ {j - 1} \mathbb {E} \left\{\mu_ {A} ^ {(j - 1)} (e _ {n} ^ {(j - 1)}) ^ {r} \right\} = \mathcal {O} _ {L _ {1}, L _ {2}, k} (n ^ {- k}). +$$ + +Define a new operator $B:h\mapsto \mathbb{E}\{h(\hat{\pi})\}$ for any $h:\Delta_{\mathcal{X}}\to \mathbb{R}$ and let + +$$ +h _ {s} (\pi) = \left\{\int_ {A} \ell (x) \pi (d x) \right\} \left\{\int \ell (x) \pi (d x) \right\} ^ {s}. +$$ + +Since $B^{j}h_{s}(\pi) = \mathbb{E}\left\{h_{s}(\hat{\pi}^{(j)})\right\}$ , we have + +$$ +\begin{array}{l} B _ {k, r} = \sum_ {j = 1} ^ {k} \binom {k} {j} (- 1) ^ {j - 1} \sum_ {s = 0} ^ {r} \binom {r} {s} B ^ {j} h _ {s} (\pi) (- 1) ^ {(r - s)} \mu^ {r - s} \\ = \sum_ {s = 0} ^ {r} {\binom {r} {s}} (- 1) ^ {r - s} \mu^ {r - s} \sum_ {j = 1} ^ {k} {\binom {k} {j}} (- 1) ^ {j - 1} B ^ {j} h _ {s} (\pi). \\ \end{array} +$$ + +We claim that $B_{k,r} = \mathcal{O}_{L_1,L_2,k}(n^{-k})$ holds if for any $0\leq s\leq r\leq 2k - 1$ + +$$ +(I - B) ^ {k} h _ {s} (\pi) = \mathcal {O} _ {L _ {1}, L _ {2}, s} (n ^ {- k}). \tag {11} +$$ + +Indeed, $(I - B)^{k}h_{s}(\pi) = \mathcal{O}_{L_{1},L_{2},s}(n^{-k})$ is equivalent to $\sum_{j = 1}^{k}\binom{k}{j}(-1)^{j - 1}B^{j}h_{s}(\pi) = h_{s}(\pi) + \mathcal{O}_{L_{1},L_{2},s}(n^{-k})$ . Therefore (11) implies + +$$ +\begin{array}{l} B _ {k, r} = \sum_ {s = 0} ^ {r} \binom {r} {s} (- 1) ^ {r - s} \mu^ {r - s} \left\{h _ {s} (\pi) + \mathcal {O} _ {L _ {1}, L _ {2}, s} (n ^ {- k}) \right\} \\ = \sum_ {s = 0} ^ {r} \left\{\binom {r} {s} (- 1) ^ {r - s} \mu_ {A} \mu^ {r} + \mathcal {O} _ {L _ {1}, L _ {2}, s} (n ^ {- k}) \right\} \\ = \mathcal {O} _ {L _ {1}, L _ {2}, k} (n ^ {- k}). \\ \end{array} +$$ + +Now, to prove the bound on the bias we only need to show that (11) holds. For any $k \in \mathbb{N}$ and $s \in \mathbb{N}^+$ , let + +$$ +\mathfrak {J} _ {s} := \left\{(\mathbf {a}, \mathbf {s}, v) \colon \mathbf {a} = (a _ {1}, a _ {2}, \dots), \mathbf {s} = (s _ {1}, s _ {2}, \dots), a _ {i}, s _ {i} \in \mathbb {N} ^ {+}, v \in \mathbb {N}, a _ {1} > a _ {2} > \dots \geq 1, \sum_ {i} a _ {i} s _ {i} + v = s \right\} +$$ + +and + +$$ +\mathcal {A} _ {s} ^ {k} := \left\{\sum_ {(\mathbf {a}, \mathbf {s}, v) \in \mathfrak {J} _ {s}} \alpha_ {\mathbf {a}, \mathbf {s}, v} \left\{\int_ {A} \ell^ {v} (x) \pi (d x) \right\} \prod_ {i} \left\{\int \ell^ {a _ {i}} (x) \pi (d x) \right\} ^ {s _ {i}}: | \alpha_ {\mathbf {a}, \mathbf {s}, v} | \leq C _ {k} (s) n ^ {- k} \right\}, +$$ + +where $C_0(s), C_1(s), \ldots$ are constants from Lemma 2. Since $h_s(\pi) \in \mathcal{A}_{s+1}^0$ , Lemma 2 implies that $(I-B)^k h_s(\pi) \in \mathcal{A}_{s+1}^k$ . Therefore, $(I-B)^k h_s(\pi) = \mathcal{O}_{L_1,L_2,s}(n^{-k})$ , finishing the proof for the bias bound. + +Finally, to prove the bound on the variance, consider the function $F(x,y) = x / y$ . By construction, + +$$ +f (\hat {\pi} ^ {(j)}) (A) = F \left(\mu_ {A} ^ {(j - 1)}, e _ {n} ^ {(j - 1)} + \mu\right). +$$ + +Applying the Taylor expansion of $F(x,y)$ yields + +$$ +f (\hat {\pi} ^ {(j)}) (A) = f (\pi) (A) + \frac {1}{\mu_ {A}} (\mu_ {A} ^ {(j - 1)} - \mu_ {A}) - \frac {\mu_ {A}}{\mu^ {2}} e _ {n} ^ {(j - 1)} - \frac {1}{\xi_ {y} ^ {2}} (\mu_ {A} ^ {(j - 1)} - \mu_ {A}) e _ {n} ^ {(j - 1)} + \frac {\xi_ {x}}{\xi_ {y} ^ {3}} \left(e _ {n} ^ {(j - 1)}\right) ^ {2}, +$$ + +for some $\xi_{x}$ lying between $\mu_{A}$ and $\mu_A^{(j - 1)}$ , and $\xi_{y}$ lying between $\mu$ and $e_n^{(j - 1)} + \mu$ . Since $L_{1} \leq \mu_{A}, \mu_{A}^{(j - 1)}, \mu, e_n^{(j - 1)} + \mu \leq L_{2}$ implies that $|1 / \xi_y^2|$ and $|\xi_x / \xi_y^3|$ are bounded by some constant depending on $L_{1}$ and $L_{2}$ . + +Moreover, since + +$$ +\begin{array}{l} \left| \mathbb {E} \left\{\left(\mu_ {A} ^ {(j - 1)} - \mu_ {A}\right) e _ {n} ^ {(j - 1)} \right\} \right| = \left| \operatorname {C o v} \left(\mu_ {A} ^ {(j - 1)}, e _ {n} ^ {(j - 1)} + \mu\right) \right| \\ \leq \left\{\operatorname {V a r} \left(\mu_ {A} ^ {(j - 1)}\right) \right\} ^ {1 / 2} \left\{\operatorname {V a r} \left(e _ {n} ^ {(j - 1)} + \mu\right) \right\} ^ {1 / 2} \\ = \left[ \operatorname {V a r} \left\{n ^ {- 1} \sum_ {i = 1} ^ {n} \ell \left(X _ {i} ^ {(j)}\right) \delta_ {X _ {i} ^ {(j - 1)}} (A) \right\} \right] ^ {1 / 2} \left[ \operatorname {V a r} \left\{n ^ {- 1} \sum_ {i = 1} ^ {n} \ell \left(X _ {i} ^ {(j - 1)}\right) \right\} \right] ^ {1 / 2} \\ = \left[ \frac {1}{n} \operatorname {V a r} \left\{\ell \left(X _ {i} ^ {(j)}\right) \delta_ {X _ {i} ^ {(j - 1)}} (A) \right\} \right] ^ {1 / 2} \left[ \frac {1}{n} \operatorname {V a r} \left\{\ell \left(X _ {i} ^ {(j - 1)}\right) \right\} \right] ^ {1 / 2} \\ = \mathcal {O} _ {L _ {1}, L _ {2}} (n ^ {- 1}), \\ \end{array} +$$ + +and + +$$ +\mathbb {E} \left(e _ {n} ^ {(j - 1)}\right) ^ {2} = \frac {1}{n} \mathrm {V a r} \left\{\ell (X _ {i} ^ {(j - 1)}) \right\} = \mathcal {O} _ {L _ {1}, L _ {2}} (n ^ {- 1}). +$$ + +Combining these bounds with the Taylor expansion, we conclude that for any $j \geq 1$ , + +$$ +B _ {n} ^ {j} f (\pi) (A) = \mathbb {E} \left\{f (\hat {\pi} ^ {(j)} (A) \right\} = f (\pi) (A) + \mathcal {O} _ {L _ {1}, L _ {2}} (n ^ {- 1}). +$$ + +By the same logic, we also have $B_{n}\{f(\pi)(A)\}^{2} = \{f(\pi)(A)\}^{2} + \mathcal{O}_{L_{1},L_{2}}(n^{-1})$ + +Therefore, + +$$ +\begin{array}{l} D_{n,k}f(\pi)(A) = \sum_{j = 0}^{k - 1}\binom {k}{j + 1}(-1)^{j}B_{n}^{j}f(\pi)(A) \\ = \sum_ {j = 0} ^ {k - 1} \binom {k} {j + 1} (- 1) ^ {j} \left\{f (\pi) (A) + \mathcal {O} _ {L _ {1}, L _ {2}} (n ^ {- 1}) \right\} \\ = f (\pi) (A) + \mathcal {O} _ {L _ {1}, L _ {2}, k} (n ^ {- 1}), \\ \end{array} +$$ + +and + +$$ +\begin{array}{l} \operatorname {V a r} _ {X ^ {n}} \left\{D _ {n, k} f (\hat {\pi}) (A) \right\} = \mathbb {E} \left[ \left\{D _ {n, k} f (\hat {\pi}) (A) \right\} ^ {2} \right] - \left[ \mathbb {E} \left\{D _ {n, k} f (\hat {\pi}) (A) \right\} \right] ^ {2} \\ = B _ {n} \left\{D _ {n, k} f (\pi) (A) \right\} ^ {2} - \left\{f (\pi) (A) + \mathcal {O} _ {L _ {1}, L _ {2}, k} \left(n ^ {- k}\right) \right\} ^ {2} \\ = B _ {n} \left\{f (\pi) (A) + \mathcal {O} _ {L _ {1}, L _ {2}, k} \left(n ^ {- 1}\right) \right\} ^ {2} - \left\{f (\pi) (A) + \mathcal {O} _ {L _ {1}, L _ {2}, k} \left(n ^ {- k}\right) \right\} ^ {2} \\ = B _ {n} \left\{f (\pi) (A) \right\} ^ {2} + \mathcal {O} _ {L _ {1}, L _ {2}, k} \left(n ^ {- 1}\right) - \left\{f (\pi) (A) \right\} ^ {2} \\ = \mathcal {O} _ {L _ {1}, L _ {2}, k} (n ^ {- 1}). \\ \end{array} +$$ + +![](images/78a170c2dd2d02eb7ac21d0276e4ec9bafa5df6b8714cb9c5b64974636d19527.jpg) + +Lemma 2. There exist constants $C_0(s), C_1(s), C_2(s), \ldots$ , such that the following holds. + +For any $k\in \mathbb{N}$ and $s,n\in \mathbb{N}^+$ , let + +$$ +\mathfrak {J} _ {s} := \left\{(\mathbf {a}, \mathbf {s}, v) \colon \mathbf {a} = (a _ {1}, a _ {2}, \dots), \mathbf {s} = (s _ {1}, s _ {2}, \dots), a _ {i}, s _ {i} \in \mathbb {N} ^ {+}, v \in \mathbb {N}, a _ {1} > a _ {2} > \dots \geq 1, \sum_ {i} a _ {i} s _ {i} + v = s \right\} +$$ + +and + +$$ +\mathcal {A} _ {s} ^ {k} := \left\{\sum_ {(\mathbf {a}, \mathbf {s}, v) \in \mathfrak {I} _ {s}} \alpha_ {\mathbf {a}, \mathbf {s}, v} \left\{\int_ {A} \ell^ {v} (x) \pi (d x) \right\} \prod_ {i} \left\{\int \ell^ {a _ {i}} (x) \pi (d x) \right\} ^ {s _ {i}}: | \alpha_ {\mathbf {a}, \mathbf {s}, v} | \leq C _ {k} (s) n ^ {- k} \right\}. +$$ + +If $h(\pi) \in \mathcal{A}_s^0$ , then for any $k \in \mathbb{N}$ , we have + +$$ +(I - B) ^ {k} h (\pi) \in \mathcal {A} _ {s} ^ {k}, \tag {12} +$$ + +where $B$ is an operator defined as $Bh(\pi) = \mathbb{E}\{h(\hat{\pi})\}$ where $\hat{\pi}$ is the empirical distribution of $X_{1},X_{2},\ldots ,X_{n}\stackrel {\mathrm{i.i.d.}}{\sim}\pi$ . + +Proof of Lemma 2. We begin by proving that $(I - B)h(\pi)\in \mathcal{A}_s^1$ . Since $h(\pi)\in \mathcal{A}_s^0$ , let + +$$ +h (\pi) = \sum_ {(\mathbf {a}, \mathbf {s}, v) \in \mathfrak {J} _ {s}} \alpha_ {\mathbf {a}, \mathbf {s}, v} \left\{\int_ {A} \ell^ {v} (x) \pi (d x) \right\} \prod_ {i} \left\{\int \ell^ {a _ {i}} (x) \pi (d x) \right\} ^ {s _ {i}}. +$$ + +Note that $|\mathfrak{I}_s|$ does not depend on $n$ and $|\alpha_{\mathbf{a},\mathbf{s},v}|\leq C_0(s)$ . It suffices to verify that each individual term in the sum satisfies + +$$ +(I - B) \left[ \left\{\int_ {A} \ell^ {v} (x) \pi (d x) \right\} \prod_ {i} \left\{\int \ell^ {a _ {i}} (x) \pi (d x) \right\} ^ {s _ {i}} \right] \in \mathcal {A} _ {s} ^ {1}. +$$ + +Without loss of generality, let $\mathbf{a} = (a_1,\dots ,a_p)$ and $\mathbf{s} = (s_1,\ldots ,s_p), s' = \sum_{i=1}^{p} s_i$ . Then we have $\sum_{i}^{p} a_i s_i + v = s$ and + +$$ +\begin{array}{l} B \left[ \left\{\int_ {A} \ell^ {v} (x) \pi (d x) \right\} \prod_ {i = 1} ^ {p} \left\{\int \ell^ {a _ {i}} (x) \pi (d x) \right\} ^ {s _ {i}} \right] \\ = \mathbb {E} \left[ \left\{\int_ {A} \ell^ {v} (x) \hat {\pi} (d x) \right\} \prod_ {i = 1} ^ {p} \left\{\int \ell^ {a _ {i}} (x) \hat {\pi} (d x) \right\} ^ {s _ {i}} \right] \\ = \frac {1}{n ^ {s ^ {\prime} + 1}} \mathbb {E} \left[ \left\{\sum_ {j = 1} ^ {n} \ell^ {v} \left(X _ {j}\right) \delta_ {X _ {j}} (A) \right\} \prod_ {i = 1} ^ {p} \left\{\sum_ {j = 1} ^ {n} \ell^ {a _ {i}} \left(X _ {j}\right) \right\} ^ {s _ {i}} \right]. \\ \end{array} +$$ + +For the term $\prod_{i=1}^{p}\left\{\sum_{j=1}^{n}\ell^{a_i}(X_j)\right\}^{s_i}$ , let $m_j^{(i)}$ denote the times $X_j$ appears with powers $a_i$ , then we have $\sum_{j=1}^{n}m_j^{(i)} = s_i$ for $1 \leq i \leq p$ . Define + +$$ +\mathcal {I} _ {\mathbf {s}} = \left\{\mathbf {m} = \left(m _ {j} ^ {(i)}\right) _ {j \in [ n ], i \in [ p ]}: \sum_ {j = 1} ^ {n} m _ {j} ^ {(i)} = s _ {i} \mathrm {f o r a l l} i \in [ p ] \right\}. +$$ + +Therefore, + +$$ +\left\{\sum_ {j = 1} ^ {n} \ell^ {v} \left(X _ {j}\right) \delta_ {X _ {j}} (A) \right\} \prod_ {i = 1} ^ {p} \left\{\sum_ {j = 1} ^ {n} \ell^ {a _ {i}} \left(X _ {j}\right) \right\} ^ {s _ {i}} = \left\{\sum_ {j = 1} ^ {n} \ell^ {v} \left(X _ {j}\right) \delta_ {X _ {j}} (A) \right\} \sum_ {\mathbf {m} \in \mathcal {I} _ {\mathrm {s}}} c _ {\mathbf {s}, \mathbf {m}} \prod_ {j = 1} ^ {n} \ell^ {\sum_ {i = 1} ^ {p} a _ {i} m _ {j} ^ {(i)}} \left(X _ {j}\right), \tag {13} +$$ + +where + +$$ +c _ {\mathbf {s}, \mathbf {m}} = \prod_ {i = 1} ^ {p} \frac {s _ {i} !}{\prod_ {j = 1} ^ {n} m _ {j} ^ {(i)} !}. +$$ + +Note that $c_{\mathbf{s},\mathbf{m}}$ does not depend on $n$ . Now we expand R.H.S. of (13) based on the number of distinct variables $X_{j}$ appear, i.e., $\sum_{j = 1}^{n}\mathbb{1}_{\sum_{i = 1}^{p}a_{i}m_{j}^{(i)} > 0}$ , which is equal to $\sum_{j = 1}^{n}\mathbb{1}_{\sum_{i = 1}^{p}m_{j}^{(i)} > 0}$ . Define + +$$ +\mathcal {J} _ {\mathbf {m}} = \left\{j \in [ n ] \colon \sum_ {i = 1} ^ {p} m _ {j} ^ {(i)} > 0 \right\}, +$$ + +then we have $1 \leq |\mathcal{J}_{\mathbf{m}}| \leq s'$ . + +Hence, + +$$ +\begin{array}{l} \mathbb {E} \left[ \left\{\sum_ {j = 1} ^ {n} \ell^ {v} \left(X _ {j}\right) \delta_ {X _ {j}} (A) \right\} \prod_ {i = 1} ^ {p} \left\{\sum_ {j = 1} ^ {n} \ell^ {a _ {i}} \left(X _ {j}\right) \right\} ^ {s _ {i}} \right] \\ = \mathbb{E}\left[\left\{\sum_{j = 1}^{n}\ell^{v}(X_{j})\delta_{X_{j}}(A)\right\} \sum_{m = 1}^{s^{\prime}}\sum_{\substack{\mathbf{m}\in \mathcal{I}_{s}\\ |\mathcal{I}_{\mathbf{m}}| = m}}c_{\mathbf{s},\mathbf{m}}\prod_{j = 1}^{n}\ell^{\sum_{i = 1}^{p}a_{i}m_{j}^{(i)}}(X_{j})\right] \\ = \mathbb{E}\left\{\sum_{m = 1}^{s^{\prime}}\sum_{\substack{\mathbf{m}\in \mathcal{I}_{\mathbf{s}}\\ |\mathcal{J}_{\mathbf{m}}| = m}}c_{\mathbf{s},\mathbf{m}}\sum_{t = 1}^{n}\ell^{v}(X_{t})\delta_{X_{t}}(A)\prod_{j = 1}^{n}\ell^{\sum_{i = 1}^{p}a_{i}m_{j}^{(i)}}(X_{j})\right\} \\ = n (n - 1) \dots \left(n - s ^ {\prime}\right) c _ {\mathbf {s}, \mathbf {m} ^ {*}} \left\{\int_ {A} \ell^ {v} (x) \pi (d x) \right\} \prod_ {i = 1} ^ {p} \left[ \mathbb {E} \left\{\ell^ {a _ {i}} (X) \right\} \right] ^ {s _ {i}} \\ +\mathbb{E}\left\{\sum_{\substack{\mathbf{m}\in \mathcal{I}_{\mathbf{s}}\\ |\mathcal{J}_{\mathbf{m}}| = s^{\prime}}}c_{\mathbf{s},\mathbf{m}}\sum_{t\in \mathcal{J}_{\mathbf{m}}}\ell^{v}(X_{t})\delta_{X_{t}}(A)\prod_{j = 1}^{n}\ell^{\sum_{i = 1}^{p}a_{i}m_{j}^{(i)}}(X_{j})\right\} \\ + \mathbb {E} \left\{\sum_ {m = 1} ^ {s ^ {\prime} - 1} \sum_ {\substack {\mathbf {m} \in \mathcal {I} _ {\mathbf {s}} \\ | \mathcal {J} _ {\mathbf {m}} | = m}} c _ {\mathbf {s}, \mathbf {m}} \sum_ {t = 1} ^ {n} \ell^ {v} \left(X _ {t}\right) \delta_ {X _ {t}} (A) \prod_ {j = 1} ^ {n} \ell^ {\sum_ {i = 1} ^ {p} a _ {i} m _ {j} ^ {(i)}} \left(X _ {j}\right) \right\}, \tag{14} \\ \end{array} +$$ + +where $c_{\mathbf{s},\mathbf{m}^*} = \prod_{i = 1}^{p}s_i!$ + +The three terms in (14) are interpreted as follows: we can expand $\{\sum_{t=1}^{n} \ell^v(X_t) \delta_{X_t}(A)\} \prod_{i=1}^{p} \left\{ \sum_{j=1}^{n} \ell^{a_i}(X_j) \right\}^{s_i}$ as the sum of many product terms of the form $\ell^v(X_t) \prod_{i=1}^{p} \prod_{l=1}^{s_i} \ell^{a_i}(X_{j_i,l})$ . The first term in (14) corresponds to the partial sum of terms in which all of $X_t, (X_{j_i,l})_{i,l}$ are distinct. The second term in (14) corresponds to the partial sum of terms in which $X_t$ is identical to one of $(X_{j_i,l})_{i,l}$ while the latter are distinct. The third term corresponds to the partial sum of terms in which at least two of $(X_{j_i,l})_{i,l}$ are identical. The last two term in (14) are at least $\mathcal{O}(n^{-1})$ factor smaller than the first (due to fewer terms involved in the sum because of the constraint of having identical terms), while the first term will cancel with $I \cdot h(\pi)$ when applying $I - B$ to $h$ . + +Let $\mathcal{P}(b_1, \ldots, b_m)$ denote the set of all distinct permutations of the vector consisting of $m$ nonzero elements $b_1, \ldots, b_m$ and $n - m$ zeros. Note that even the values of $b_i$ may be the same, we still treat the $b_i$ s are distinguishable. Then since 0s are identical, we have $|\mathcal{P}(b_1, \ldots, b_m)| = n(n - 1) \cdots (n - m + 1) = \mathcal{O}(n^m)$ . Additionally, for any $\mathbf{a}$ and $\mathbf{m}$ , we define + +$$ +\Psi (\mathbf {a}, \mathbf {m}) = \left(\sum_ {i = 1} ^ {p} a _ {i} m _ {1} ^ {(i)}, \ldots , \sum_ {i = 1} ^ {p} a _ {i} m _ {n} ^ {(i)}\right). +$$ + +Now we can write (14) as + +$$ +n (n - 1) \dots (n - s ^ {\prime}) c _ {\mathbf {s}, \mathbf {m} ^ {*}} \left\{\int_ {A} \ell^ {v} (x) \pi (d x) \right\} \prod_ {i = 1} ^ {p} [ \mathbb {E} \{\ell^ {a _ {i}} (X) \} ] ^ {s _ {i}} +$$ + +$$ +\begin{array}{l} + \sum_ {b _ {k}: \sum_ {k = 1} ^ {s ^ {\prime}} b _ {k} = s - v} \sum_ {\mathbf {m}: \Psi (\mathbf {a}, \mathbf {m}) \in \mathcal {P} \left(b _ {1}, \dots , b _ {s ^ {\prime}}\right)} c _ {\mathbf {s}, \mathbf {m}} \sum_ {t = 1} ^ {m} \left[ \prod_ {i \neq t} \mathbb {E} \left\{\ell^ {b _ {i}} (X) \right\} \right] \int_ {A} \ell^ {b _ {t} + v} (x) \pi (d x) \\ + \sum_ {m = 1} ^ {s ^ {\prime} - 1} \sum_ {b _ {k}: \sum_ {k = 1} ^ {m} b _ {k} = s - v} \sum_ {\mathbf {m}: \Psi (\mathbf {a}, \mathbf {m}) \in \mathcal {P} (b _ {1}, \dots , b _ {m})} c _ {\mathbf {s}, \mathbf {m}} \mathbb {E} \left\{\sum_ {t = 1} ^ {n} \ell^ {v} \left(X _ {t}\right) \delta_ {X _ {t}} (A) \prod_ {i = 1} ^ {m} \ell^ {b _ {i}} \left(X _ {i}\right) \right\} \\ = n (n - 1) \dots \left(n - s ^ {\prime}\right) c _ {\mathbf {s}, \mathbf {m} ^ {*}} \left\{\int_ {A} \ell^ {v} (x) \pi (d x) \right\} \prod_ {i = 1} ^ {p} \left[ \mathbb {E} \left\{\ell^ {a _ {i}} (X) \right\} \right] ^ {s _ {i}} \\ + \sum_ {b _ {k}: \sum_ {k = 1} ^ {s ^ {\prime}} b _ {k} = s - v} \mathcal {O} \left(n ^ {s ^ {\prime}}\right) c _ {\mathbf {s}, \mathbf {m}} \sum_ {t = 1} ^ {m} \left[ \prod_ {i \neq t} \mathbb {E} \left\{\ell^ {b _ {i}} (X) \right\} \right] \int_ {A} \ell^ {b _ {t} + v} (x) \pi (d x) \\ + \sum_ {m = 1} ^ {s ^ {\prime} - 1} \sum_ {b _ {k}: \sum_ {k = 1} ^ {m} b _ {k} = s - v} \mathcal {O} (n ^ {m}) c _ {\mathbf {s}, \mathbf {m}} \mathbb {E} \left\{\sum_ {t = 1} ^ {n} \ell^ {v} (X _ {t}) \delta_ {X _ {t}} (A) \prod_ {i = 1} ^ {m} \ell^ {b _ {i}} (X _ {i}) \right\}. \\ \end{array} +$$ + +Therefore, + +$$ +\begin{array}{l} (I - B) \left[ \left\{\int_ {A} \ell^ {v} (x) \pi (d x) \right\} \prod_ {i} \left\{\int \ell^ {a _ {i}} (x) \pi (d x) \right\} ^ {s _ {i}} \right] \\ = \frac {n ^ {s ^ {\prime} + 1} - n (n - 1) \cdots (n - s ^ {\prime})}{n ^ {s ^ {\prime} + 1}} c _ {\mathbf {s}, \mathbf {m} ^ {*}} \left\{\int_ {A} \ell^ {v} (x) \pi (d x) \right\} \prod_ {i = 1} ^ {p} \left[ \mathbb {E} \left\{\ell^ {a _ {i}} (X) \right\} \right] ^ {s _ {i}} \\ - \sum_ {b _ {k}: \sum_ {k = 1} ^ {s ^ {\prime}} b _ {k} = s - v} \mathcal {O} (n ^ {- 1}) c _ {\mathbf {s}, \mathbf {m}} \sum_ {t = 1} ^ {m} \left[ \prod_ {i \neq t} \mathbb {E} \{\ell^ {b _ {i}} (X) \} \right] \int_ {A} \ell^ {b _ {t} + v} (x) \pi (d x) \\ - \sum_ {m = 1} ^ {s ^ {\prime} - 1} \sum_ {b _ {k}: \sum_ {k = 1} ^ {m} b _ {k} = s - v} \mathcal {O} (n ^ {m - s ^ {\prime} - 1}) c _ {\mathbf {s}, \mathbf {m}} \mathbb {E} \left\{\sum_ {t = 1} ^ {n} \ell^ {v} \left(X _ {t}\right) \delta_ {X _ {t}} (A) \prod_ {i = 1} ^ {m} \ell^ {b _ {i}} \left(X _ {i}\right) \right\} \\ = \frac {n ^ {s ^ {\prime} + 1} - n (n - 1) \cdots (n - s ^ {\prime})}{n ^ {s ^ {\prime} + 1}} c _ {\mathbf {s}, \mathbf {m} ^ {*}} \left\{\int_ {A} \ell^ {v} (x) \pi (d x) \right\} \prod_ {i = 1} ^ {p} \left[ \mathbb {E} \left\{\ell^ {a _ {i}} (X) \right\} \right] ^ {s _ {i}} \\ - \sum_ {b _ {k}: \sum_ {k = 1} ^ {s ^ {\prime}} b _ {k} = s - v} \mathcal {O} (n ^ {- 1}) c _ {\mathbf {s}, \mathbf {m}} \sum_ {t = 1} ^ {m} \left[ \prod_ {i \neq t} \mathbb {E} \left\{\ell^ {b _ {i}} (X) \right\} \right] \int_ {A} \ell^ {b _ {t} + v} (x) \pi (d x) \\ - \sum_ {m = 1} ^ {s ^ {\prime} - 1} \sum_ {b _ {k}: \sum_ {k = 1} ^ {m} b _ {k} = s - v} \mathcal {O} \left(n ^ {m - s ^ {\prime} - 1}\right) c _ {\mathbf {s}, \mathbf {m}} \sum_ {t = 1} ^ {m} \left[ \prod_ {i \neq t} \mathbb {E} \left\{\ell^ {b _ {i}} (X) \right\} \right] \int_ {A} \ell^ {b _ {t} + v} (x) \pi (d x) \\ - \sum_ {m = 1} ^ {s ^ {\prime} - 1} \sum_ {b _ {k}: \sum_ {k = 1} ^ {m} b _ {k} = s - v} \mathcal {O} \left(n ^ {m - s ^ {\prime}}\right) c _ {\mathbf {s}, \mathbf {m}} \left\{\int_ {A} \ell^ {v} (x) \pi (d x) \right\} \prod_ {i = 1} ^ {m} \mathbb {E} \left\{\ell^ {b _ {i}} \left(X _ {i}\right) \right\} \\ \end{array} +$$ + +$$ +\in \mathcal {A} _ {s} ^ {1}. +$$ + +The last inclusion follows from that fact that $\{n^{s' + 1} - n(n - 1)\dots (n - s')\} /n^{s' + 1} = \mathcal{O}(n^{-1})$ and the number of solutions to $\sum_{k = 1}^{m}b_{k} = s - v$ does not depend on $n$ but depends on $s$ . + +Now we suppose (12) holds for $k$ . Then we can set + +$$ +(I - B) ^ {k} h (\pi) = \sum_ {(\mathbf {a}, \mathbf {s}, v) \in \mathfrak {I} _ {s}} \alpha_ {\mathbf {a}, \mathbf {s}, v} ^ {\prime} \left\{\int_ {A} \ell^ {v} (x) \pi (d x) \right\} \prod_ {i} \left\{\int \ell^ {a _ {i}} (x) \pi (d x) \right\} ^ {s _ {i}}, +$$ + +where $|\alpha_{\mathbf{a},\mathbf{s},v}^{\prime}| \leq C_k(s)n^{-k}$ . Then for $k + 1$ , we have + +$$ +(I - B) ^ {k + 1} h (\pi) = \sum_ {(\mathbf {a}, \mathbf {s}, v) \in \mathfrak {J} _ {s}} \alpha_ {\mathbf {a}, \mathbf {s}, v} ^ {\prime} (I - B) \left\{\int_ {A} \ell^ {v} (x) \pi (d x) \right\} \prod_ {i} \left\{\int \ell^ {a _ {i}} (x) \pi (d x) \right\} ^ {s _ {i}}. +$$ + +Since for all $\mathbf{a},\mathbf{s},v$ such that $(\mathbf{a},\mathbf{s},v)\in \mathfrak{J}_s,\left\{\int_A\ell^v (x)\pi (dx)\right\} \prod_i\left\{\int \ell^{a_i}(x)\pi (dx)\right\}^{s_i}\in \mathcal{A}_s^0$ we have $(I - B)\left\{\int_{A}\ell^{v}(x)\pi (dx)\right\} \prod_{i}\left\{\int \ell^{a_{i}}(x)\pi (dx)\right\}^{s_{i}}\in \mathcal{A}_{s}^{1}$ , namely, + +$$ +\begin{array}{l} (I - B) \left\{\int_ {A} \ell^ {v} (x) \pi (d x) \right\} \prod_ {i} \left\{\int \ell^ {a _ {i}} (x) \pi (d x) \right\} ^ {s _ {i}} \\ = \sum_ {(\mathbf {b}, \mathbf {t}, u) \in \mathfrak {J} _ {s}} \alpha_ {\mathbf {b}, \mathbf {t}, u} (\mathbf {a}, \mathbf {s}, v) \left\{\int_ {A} \ell^ {u} (x) \pi (d x) \right\} \prod_ {i} \left\{\int \ell^ {b _ {i}} (x) \pi (d x) \right\} ^ {t _ {i}}, \\ \end{array} +$$ + +where $|\alpha_{\mathbf{b},\mathbf{t},u}(\mathbf{a},\mathbf{s},v)| \leq C_0(s)n^{-1}$ . Therefore, + +$$ +\begin{array}{l} (I - B) ^ {k + 1} h (\pi) \\ = \sum_ {(\mathbf {a}, \mathbf {s}, v) \in \mathfrak {J} _ {s}} \alpha_ {\mathbf {a}, \mathbf {s}, v} ^ {\prime} \sum_ {(\mathbf {b}, \mathbf {t}, u) \in \mathfrak {J} _ {s}} \alpha_ {\mathbf {b}, \mathbf {t}, u} (\mathbf {a}, \mathbf {s}, v) \left\{\int_ {A} \ell^ {u} (x) \pi (d x) \right\} \prod_ {i} \left\{\int \ell^ {b _ {i}} (x) \pi (d x) \right\} ^ {t _ {i}} \\ \in \mathcal {A} _ {s} ^ {k + 1}. \\ \end{array} +$$ + +![](images/cbb3741b09b635eea1c301f26e54cec2e0fe80d3fcdc7b92fb6e0376fda7e755.jpg) + +# B Proof of Theorem 2 + +In order to prove Theorem 2, we first make some preliminary observations. + +Let function $f$ defined on the simplex $\Delta_m = \{\mathbf{q} \in \mathbb{R}^m : q_j \geq 0, \sum_{j=1}^m q_j = 1\}$ . Define the generalized Bernstein basis polynomials of degree $n$ as + +$$ +b _ {n, \nu} (\mathbf {q}) = \left( \begin{array}{c} n \\ \nu \end{array} \right) \mathbf {q} ^ {\nu}. +$$ + +Lemma 3. $\left|\sum_{\nu \in \bar{\Delta}_m}(\nu /n - \mathbf{q})^\alpha b_{n,\nu}(\mathbf{q})\right| \lesssim n^{-\| \boldsymbol {\alpha}\| _1 / 2}.$ + +Proof of Lemma 3. It suffices to show that $\left|\sum_{\nu \in \bar{\Delta}_m}(\nu - n\mathbf{q})^\alpha b_{n,\nu}(\mathbf{q})\right| \lesssim n^{\|\alpha\|_1/2}$ . Since $\mathbf{q} \in \Delta_m$ , we treat $T_{n,\alpha} \equiv \sum_{\nu \in \bar{\Delta}_m}(\nu - n\mathbf{q})^\alpha b_{n,\nu}(\mathbf{q})$ as a function of the variables $q_1, \dots, q_{m-1}$ . For any $\beta \in \mathbb{N}^{m-1}$ such that $\| \beta \|_1 = 1$ , we let $\gamma = \gamma(\beta) \equiv (\beta^\top, 0)^\top$ . Additionally, let $\pmb{\theta} = (0, \dots, 0, 1)^\top \in \mathbb{N}^m$ . Since + +$$ +\partial^ {\beta} (\boldsymbol {\nu} - n \mathbf {q}) ^ {\alpha} = - n \alpha^ {\gamma} (\boldsymbol {\nu} - n \mathbf {q}) ^ {\alpha - \gamma} + n \alpha^ {\theta} (\boldsymbol {\nu} - n \mathbf {q}) ^ {\alpha - \theta}, +$$ + +and + +$$ +\begin{array}{l} \partial^ {\beta} b _ {n, \nu} (\mathbf {q}) = \left( \begin{array}{c} n \\ \nu \end{array} \right) (\nu^ {\gamma} \mathbf {q} ^ {\nu - \gamma} - \nu^ {\theta} \mathbf {q} ^ {\nu - \theta}) \\ = b _ {n, \nu} (\mathbf {q}) \left\{\frac {1}{\mathbf {q} ^ {\gamma}} (\nu - n \mathbf {q}) ^ {\gamma} - \frac {1}{\mathbf {q} ^ {\theta}} (\nu - n \mathbf {q}) ^ {\theta} \right\}, \\ \end{array} +$$ + +we have + +$$ +\begin{array}{l} \partial^ {\boldsymbol {\beta}} T _ {n, \boldsymbol {\alpha}} = \sum_ {\boldsymbol {\nu} \in \bar {\Delta} _ {m}} \partial^ {\boldsymbol {\beta}} (\boldsymbol {\nu} - n \mathbf {q}) ^ {\boldsymbol {\alpha}} b _ {n, \boldsymbol {\nu}} (\mathbf {q}) + \sum_ {\boldsymbol {\nu} \in \bar {\Delta} _ {m}} (\boldsymbol {\nu} - n \mathbf {q}) ^ {\boldsymbol {\alpha}} \partial^ {\boldsymbol {\beta}} b _ {n, \boldsymbol {\nu}} (\mathbf {q}) \\ = - n \boldsymbol {\alpha} ^ {\gamma} T _ {n, \boldsymbol {\alpha} - \gamma} + n \boldsymbol {\alpha} ^ {\theta} T _ {n, \boldsymbol {\alpha} - \theta} + \frac {1}{\mathbf {q} ^ {\gamma}} T _ {n, \boldsymbol {\alpha} + \gamma} - \frac {1}{\mathbf {q} ^ {\theta}} T _ {n, \boldsymbol {\alpha} + \theta}, \\ \end{array} +$$ + +i.e., + +$$ +\mathbf {q} ^ {\gamma} \partial^ {\beta} T _ {n, \alpha} = - n \alpha^ {\gamma} \mathbf {q} ^ {\gamma} T _ {n, \alpha - \gamma} + n \alpha^ {\theta} \mathbf {q} ^ {\gamma} T _ {n, \alpha - \theta} + T _ {n, \alpha + \gamma} - \frac {\mathbf {q} ^ {\gamma}}{\mathbf {q} ^ {\theta}} T _ {n, \alpha + \theta}. +$$ + +By summing the above equation over $\beta \in \mathbb{N}^{m - 1}$ such that $\| \beta \| _1 = 1$ , we have + +$$ +\begin{array}{l} \sum_ {\| \boldsymbol {\beta} \| _ {1} = 1} \mathbf {q} ^ {\gamma} \partial^ {\beta} T _ {n, \boldsymbol {\alpha}} \\ = - n \sum_ {\| \boldsymbol {\beta} \| _ {1} = 1} \alpha^ {\gamma} \mathbf {q} ^ {\gamma} T _ {n, \boldsymbol {\alpha} - \boldsymbol {\gamma}} + n \alpha^ {\theta} \sum_ {\| \boldsymbol {\beta} \| _ {1} = 1} \mathbf {q} ^ {\gamma} T _ {n, \boldsymbol {\alpha} - \boldsymbol {\theta}} + \sum_ {\| \boldsymbol {\beta} \| _ {1} = 1} T _ {n, \boldsymbol {\alpha} + \boldsymbol {\gamma}} - \frac {1 - \mathbf {q} ^ {\theta}}{\mathbf {q} ^ {\theta}} T _ {n, \boldsymbol {\alpha} + \boldsymbol {\theta}} \\ = - n \sum_ {\| \boldsymbol {\beta} \| _ {1} = 1} \alpha^ {\gamma} \mathbf {q} ^ {\gamma} T _ {n, \boldsymbol {\alpha} - \gamma} + n \boldsymbol {\alpha} ^ {\theta} (1 - \mathbf {q} ^ {\theta}) T _ {n, \boldsymbol {\alpha} - \theta} + \sum_ {\| \boldsymbol {\beta} \| _ {1} = 1} T _ {n, \boldsymbol {\alpha} + \gamma} - \frac {1 - \mathbf {q} ^ {\theta}}{\mathbf {q} ^ {\theta}} T _ {n, \boldsymbol {\alpha} + \theta} \\ = - n \sum_ {\| \boldsymbol {\beta} \| _ {1} = 1} \boldsymbol {\alpha} ^ {\gamma} \mathbf {q} ^ {\gamma} T _ {n, \boldsymbol {\alpha} - \boldsymbol {\gamma}} + n \boldsymbol {\alpha} ^ {\theta} (1 - \mathbf {q} ^ {\theta}) T _ {n, \boldsymbol {\alpha} - \boldsymbol {\theta}} - \frac {1}{\mathbf {q} ^ {\theta}} T _ {n, \boldsymbol {\alpha} + \boldsymbol {\theta}}, \\ \end{array} +$$ + +where the last equality follows from the fact that $\sum_{\| \pmb {\beta}\| _1 = 1}T_{n,\pmb {\alpha} + \pmb{\gamma}} + T_{n,\pmb {\alpha} + \pmb{\theta}} = 0$ + +Therefore, we have the following recurrence formula: + +$$ +T _ {n, \alpha + \theta} = - n \mathbf {q} ^ {\theta} \sum_ {\| \boldsymbol {\beta} \| _ {1} = 1} \boldsymbol {\alpha} ^ {\gamma} \mathbf {q} ^ {\gamma} T _ {n, \alpha - \gamma} + n \boldsymbol {\alpha} ^ {\theta} \mathbf {q} ^ {\theta} (1 - \mathbf {q} ^ {\theta}) T _ {n, \alpha - \theta} - \mathbf {q} ^ {\theta} \sum_ {\| \boldsymbol {\beta} \| _ {1} = 1} \mathbf {q} ^ {\gamma} \partial^ {\beta} T _ {n, \alpha}, \tag {15} +$$ + +$$ +T _ {n, \alpha + \gamma} = \frac {\mathbf {q} ^ {\gamma}}{\mathbf {q} ^ {\theta}} T _ {n, \alpha + \theta} + n \boldsymbol {\alpha} ^ {\gamma} \mathbf {q} ^ {\gamma} T _ {n, \alpha - \gamma} - n \boldsymbol {\alpha} ^ {\theta} \mathbf {q} ^ {\gamma} T _ {n, \alpha - \theta} + \mathbf {q} ^ {\gamma} \partial^ {\beta} T _ {n, \alpha}. \tag {16} +$$ + +Using the recurrence formula (15), (16) and the fact that $T_{n,(1,0,\dots,0)^{\top}} = 0$ , $T_{n,(2,0,\dots,0)^{\top}} = nq_{1}(1 - q_{1})$ , $T_{n,(1,1,0,\dots,0)^{\top}} = -nq_{1}q_{2}$ , $T_{n,\alpha}$ has the following form by using induction: + +$$ +T _ {n, \alpha} = \sum_ {j = 1} ^ {\lfloor \| \boldsymbol {\alpha} \| _ {1} / 2 \rfloor} n ^ {j} \left(\sum_ {\eta \leq \alpha} c _ {j, \eta} \mathbf {q} ^ {\eta}\right), \tag {17} +$$ + +where $c_{j,\eta}$ is independent of $n$ . Then we can conclude that $|T_{n,\alpha}| \lesssim n^{\lfloor \| \alpha \|_1 / 2 \rfloor} \lesssim n^{\|\alpha\|_1 / 2}$ . + +![](images/e8c23014a60db5f760b26e8d3fe1b3fd6301ba5d2057863c37ef896dbc418365.jpg) + +Proof of Lemma 1. We prove the theorem by induction on $k$ . + +For $k = 1$ , by Taylor's expansion, there exists $\xi \in \Delta_m$ such that + +$$ +f \left(\frac {\boldsymbol {\nu}}{n}\right) = f (\mathbf {q}) + \sum_ {\| \boldsymbol {\alpha} \| _ {1} = 1} \partial^ {\boldsymbol {\alpha}} f (\boldsymbol {\xi}) \left(\frac {\boldsymbol {\nu}}{n} - \mathbf {q}\right) ^ {\boldsymbol {\alpha}}. +$$ + +Then we have + +$$ +\begin{array}{l} | B _ {n} (f) (\mathbf {q}) - f (\mathbf {q}) | = \left| \sum_ {\nu \in \bar {\Delta} _ {m}} \left\{f \left(\frac {\nu}{n}\right) - f (\mathbf {q}) \right\} b _ {n, \nu} (\mathbf {q}) \right| \\ = \left| \sum_ {\boldsymbol {\nu} \in \bar {\Delta} _ {m}} \left\{\sum_ {\| \boldsymbol {\alpha} \| _ {1} = 1} \partial^ {\boldsymbol {\alpha}} f (\boldsymbol {\xi}) \left(\frac {\boldsymbol {\nu}}{n} - \mathbf {q}\right) ^ {\boldsymbol {\alpha}} \right\} b _ {n, \boldsymbol {\nu}} (\mathbf {q}) \right| \\ \leq \| f \| _ {C ^ {1} \left(\Delta_ {m}\right)} \sum_ {\| \boldsymbol {\alpha} \| _ {1} = 1} \sum_ {\boldsymbol {\nu} \in \bar {\Delta} _ {m}} \left| \left(\frac {\boldsymbol {\nu}}{n} - \mathbf {q}\right) ^ {\boldsymbol {\alpha}} \right| b _ {n, \boldsymbol {\nu}} (\mathbf {q}) \\ \leq \| f \| _ {C ^ {1} \left(\Delta_ {m}\right)} \sum_ {\| \boldsymbol {\alpha} \| _ {1} = 1} \left\{\sum_ {\boldsymbol {\nu} \in \bar {\Delta} _ {m}} \left(\frac {\boldsymbol {\nu}}{n} - \mathbf {q}\right) ^ {2 \boldsymbol {\alpha}} b _ {n, \boldsymbol {\nu}} (\mathbf {q}) \right\} ^ {1 / 2} \\ \lesssim \| f \| _ {C ^ {1} \left(\Delta_ {m}\right)} \sum_ {\| \boldsymbol {\alpha} \| _ {1} = 1} n ^ {- 1 / 2} \\ \lesssim_ {m} \| f \| _ {C ^ {1} \left(\Delta_ {m}\right)} n ^ {- 1 / 2}, \\ \end{array} +$$ + +where the second inequality follows from Cauchy-Schwarz inequality, and the third inequality follows from Lemma 3. + +Suppose the theorem holds up through $k$ . Now we prove the theorem for $k + 1$ . For $k + 1$ , by Taylor's expansion, there exists $\pmb{\xi} \in \Delta_{m}$ such that + +$$ +f \left(\frac {\nu}{n}\right) = f (\mathbf {q}) + \sum_ {\| \boldsymbol {\alpha} \| _ {1} = 1} ^ {k} \frac {\partial^ {\boldsymbol {\alpha}} f (\mathbf {q})}{\boldsymbol {\alpha} !} \left(\frac {\nu}{n} - \mathbf {q}\right) ^ {\boldsymbol {\alpha}} + \sum_ {\| \boldsymbol {\alpha} \| _ {1} = k + 1} \frac {\partial^ {\boldsymbol {\alpha}} f (\boldsymbol {\xi})}{\boldsymbol {\alpha} !} \left(\frac {\nu}{n} - \mathbf {q}\right) ^ {\boldsymbol {\alpha}}. +$$ + +Then we have + +$$ +\begin{array}{l} B _ {n} (f) (\mathbf {q}) - f (\mathbf {q}) = \sum_ {\nu \in \bar {\Delta} _ {m}} \left(f (\frac {\nu}{n}) - f (\mathbf {q})\right) b _ {n, \nu} (\mathbf {q}) \\ = \sum_ {\boldsymbol {\nu} \in \bar {\Delta} _ {m}} \left\{\sum_ {\| \boldsymbol {\alpha} \| _ {1} = 1} ^ {k} \frac {\partial^ {\boldsymbol {\alpha}} f (\mathbf {q})}{\boldsymbol {\alpha} !} \left(\frac {\boldsymbol {\nu}}{n} - \mathbf {q}\right) ^ {\boldsymbol {\alpha}} + \sum_ {\| \boldsymbol {\alpha} \| _ {1} = k + 1} \frac {\partial^ {\boldsymbol {\alpha}} f (\boldsymbol {\xi})}{\boldsymbol {\alpha} !} \left(\frac {\boldsymbol {\nu}}{n} - \mathbf {q}\right) ^ {\boldsymbol {\alpha}} \right\} b _ {n, \boldsymbol {\nu}} (\mathbf {q}) \\ = \sum_ {\| \boldsymbol {\alpha} \| _ {1} = 1} ^ {k} \frac {\partial^ {\boldsymbol {\alpha}} f (\mathbf {q})}{\boldsymbol {\alpha} !} \left\{\sum_ {\boldsymbol {\nu} \in \bar {\Delta} _ {m}} \left(\frac {\boldsymbol {\nu}}{n} - \mathbf {q}\right) ^ {\boldsymbol {\alpha}} b _ {n, \boldsymbol {\nu}} (\mathbf {q}) \right\} \\ + \sum_ {\| \boldsymbol {\alpha} \| _ {1} = k + 1} \frac {\partial^ {\boldsymbol {\alpha}} f (\boldsymbol {\xi})}{\boldsymbol {\alpha} !} \left\{\sum_ {\boldsymbol {\nu} \in \bar {\Delta} _ {m}} \left(\frac {\boldsymbol {\nu}}{n} - \mathbf {q}\right) ^ {\boldsymbol {\alpha}} b _ {n, \boldsymbol {\nu}} (\mathbf {q}) \right\}. \\ \end{array} +$$ + +Therefore, + +$$ +\begin{array}{l} \left(B _ {n} - I\right) ^ {\lceil (k + 1) / 2 \rceil} (f) (\mathbf {q}) = \sum_ {\| \boldsymbol {\alpha} \| _ {1} = 1} ^ {k} \left(B _ {n} - I\right) ^ {\lceil (k + 1) / 2 \rceil - 1} \left\{\frac {\partial^ {\boldsymbol {\alpha}} f (\mathbf {q})}{\boldsymbol {\alpha} !} \sum_ {\boldsymbol {\nu} \in \bar {\Delta} _ {m}} \left(\frac {\boldsymbol {\nu}}{n} - \mathbf {q}\right) ^ {\boldsymbol {\alpha}} b _ {n, \boldsymbol {\nu}} (\mathbf {q}) \right\} \\ + \sum_ {\left\| \boldsymbol {\alpha} \right\| _ {1} = k + 1} \left(B _ {n} - I\right) ^ {\lceil (k + 1) / 2 \rceil - 1} \left\{\frac {\partial^ {\alpha} f (\boldsymbol {\xi})}{\alpha !} \sum_ {\boldsymbol {\nu} \in \bar {\Delta} _ {m}} \left(\frac {\boldsymbol {\nu}}{n} - \mathbf {q}\right) ^ {\alpha} b _ {n, \boldsymbol {\nu}} (\mathbf {q}) \right\}. \tag {18} \\ \end{array} +$$ + +First, we consider the first term of the right-hand side of (18). We know that $(\alpha!)^{-1}\partial^{\alpha}f(\mathbf{q})\sum_{\nu \in \bar{\Delta}_m}(\nu /n - \mathbf{q})^\alpha b_{n,\nu}(\mathbf{q})\in C^{k + 1 - \| \boldsymbol {\alpha}\| _1}(\Delta_m)|$ since $f\in C^{k + 1}(\Delta_m)$ . By the induction hypothesis, we have + +$$ +\begin{array}{l} \left\| \left(B _ {n} - I\right) ^ {\lceil (k + 1 - \| \boldsymbol {\alpha} \| _ {1}) / 2 \rceil} \left\{\frac {\partial^ {\boldsymbol {\alpha}} f (\mathbf {q})}{\boldsymbol {\alpha} !} \sum_ {\boldsymbol {\nu} \in \Delta_ {m}} \left(\frac {\boldsymbol {\nu}}{n} - \mathbf {q}\right) ^ {\boldsymbol {\alpha}} b _ {n, \boldsymbol {\nu}} (\mathbf {q}) \right\} \right\| _ {\infty} \\ \lesssim_ {k + 1 - \| \boldsymbol {\alpha} \| _ {1}, m} \left\| \frac {\partial^ {\boldsymbol {\alpha}} f (\mathbf {q})}{\boldsymbol {\alpha} !} \sum_ {\boldsymbol {\nu} \in \bar {\Delta} _ {m}} \left(\frac {\boldsymbol {\nu}}{n} - \mathbf {q}\right) ^ {\boldsymbol {\alpha}} b _ {n, \boldsymbol {\nu}} (\mathbf {q}) \right\| _ {C ^ {k + 1 - \| \boldsymbol {\alpha} \| _ {1}} \left(\Delta_ {m}\right)} n ^ {- (k + 1 - \| \boldsymbol {\alpha} \| _ {1}) / 2}. \\ \end{array} +$$ + +Let + +$$ +g _ {\boldsymbol {\alpha}} (\mathbf {q}) = \frac {\partial^ {\boldsymbol {\alpha}} f (\mathbf {q})}{\boldsymbol {\alpha} !} \sum_ {\boldsymbol {\nu} \in \bar {\Delta} _ {m}} \left(\frac {\boldsymbol {\nu}}{n} - \mathbf {q}\right) ^ {\boldsymbol {\alpha}} b _ {n, \boldsymbol {\nu}} (\mathbf {q}), +$$ + +For any $|\beta |\leq k + 1 - \| \pmb {\alpha}\| _1$ , we have + +$$ +\begin{array}{l} \| \partial^ {\beta} g _ {\boldsymbol {\alpha}} (\mathbf {q}) \| _ {\infty} = \left\| \frac {1}{\boldsymbol {\alpha} !} \sum_ {0 \leq \gamma \leq \beta} \binom {\beta} {\gamma} \partial^ {\alpha + \gamma} f (\mathbf {q}) \partial^ {\beta - \gamma} \left\{\sum_ {\nu \in \bar {\Delta} _ {m}} \left(\frac {\nu}{n} - \mathbf {q}\right) ^ {\boldsymbol {\alpha}} b _ {n, \nu} (\mathbf {q}) \right\} \right\| _ {\infty} \\ \lesssim_ {k + 1} \| f \| _ {C ^ {k + 1} \left(\Delta_ {m}\right)} \sum_ {0 \leq \gamma \leq \beta} \binom {\beta} {\gamma} \left\| \partial^ {\beta - \gamma} \left\{\sum_ {\nu \in \bar {\Delta} _ {m}} (\frac {\nu}{n} - \mathbf {q}) ^ {\alpha} b _ {n, \nu} (\mathbf {q}) \right\} \right\| _ {\infty} \\ \end{array} +$$ + +$$ +\lesssim_ {k + 1} \left\| f \right\| _ {C ^ {k + 1} \left(\Delta_ {m}\right)} n ^ {- \left\| \boldsymbol {\alpha} \right\| _ {1} / 2}, +$$ + +where the last inequality follows from the fact that $\left\| \partial^{\beta - \gamma} \left\{ \sum_{\pmb{\nu} \in \bar{\Delta}_m} (\pmb{\nu} / n - \mathbf{q})^\alpha b_{n,\pmb{\nu}}(\mathbf{q}) \right\} \right\|_\infty \lesssim n^{-\|\pmb{\alpha}\|_1/2}$ which can be derived by using the form of $T_{n,\pmb{\alpha}}$ in (17). + +Therefore, we have + +$$ +\begin{array}{l} \left\| \left(B _ {n} - I\right) ^ {\lceil (k + 1 - \| \boldsymbol {\alpha} \| _ {1}) / 2 \rceil} \left\{\frac {\partial^ {\boldsymbol {\alpha}} f (\mathbf {q})}{\boldsymbol {\alpha} !} \sum_ {\boldsymbol {\nu} \in \bar {\Delta} _ {m}} \left(\frac {\boldsymbol {\nu}}{n} - \mathbf {q}\right) ^ {\boldsymbol {\alpha}} b _ {n, \boldsymbol {\nu}} (\mathbf {q}) \right\} \right\| _ {\infty} \\ \lesssim_ {k + 1} \| f \| _ {C ^ {k + 1} (\Delta_ {m})} n ^ {- (k + 1) / 2}. \\ \end{array} +$$ + +Then we consider the second term of the right-hand side of (18). + +$$ +\begin{array}{l} \left\| \sum_ {\left\| \boldsymbol {\alpha} \right\| _ {1} = k + 1} (B _ {n} - I) ^ {\lceil (k + 1) / 2 \rceil - 1} \left\{\frac {\partial^ {\boldsymbol {\alpha}} f (\boldsymbol {\xi})}{\boldsymbol {\alpha} !} \sum_ {\boldsymbol {\nu} \in \bar {\Delta} _ {m}} (\frac {\boldsymbol {\nu}}{n} - \mathbf {q}) ^ {\boldsymbol {\alpha}} b _ {n, \boldsymbol {\nu}} (\mathbf {q}) \right\} \right\| _ {\infty} \\ \lesssim_ {k + 1} \| \left(B _ {n} - I\right) ^ {\lceil (k + 1) / 2 \rceil - 1} \| _ {\infty} \| f \| _ {C ^ {k + 1} \left(\Delta_ {m}\right)} \sum_ {\| \boldsymbol {\alpha} \| _ {1} = k + 1} \sum_ {\boldsymbol {\nu} \in \bar {\Delta} _ {m}} \left| \left(\frac {\boldsymbol {\nu}}{n} - \mathbf {q}\right) ^ {\boldsymbol {\alpha}} \right| b _ {n, \boldsymbol {\nu}} (\mathbf {q}) \\ \lesssim_ {k + 1} \| \left(B _ {n} - I\right) ^ {\lceil (k + 1) / 2 \rceil - 1} \| _ {\infty} \| f \| _ {C ^ {k + 1} \left(\Delta_ {m}\right)} \sum_ {\| \boldsymbol {\alpha} \| _ {1} = k + 1} \left\{\sum_ {\boldsymbol {\nu} \in \bar {\Delta} _ {m}} \left(\frac {\boldsymbol {\nu}}{n} - \mathbf {q}\right) ^ {2 \boldsymbol {\alpha}} b _ {n, \boldsymbol {\nu}} (\mathbf {q}) \right\} ^ {1 / 2} \\ \lesssim_ {k + 1, m} \| (B _ {n} - I) ^ {\lceil (k + 1) / 2 \rceil - 1} \| _ {\infty} \| f \| _ {C ^ {k + 1} (\Delta_ {m})} n ^ {- (k + 1) / 2}. \\ \end{array} +$$ + +Finally, we have + +$$ +\begin{array}{l} \| (B _ {n} - I) ^ {\lceil (k + 1) / 2 \rceil} (f) (\mathbf {q}) \| _ {\infty} \leq \sum_ {\| \boldsymbol {\alpha} \| _ {1} = 1} ^ {k} \left\| (B _ {n} - I) ^ {\lceil (k + 1) / 2 \rceil - 1} \left\{\frac {\partial^ {\boldsymbol {\alpha}} f (\mathbf {q})}{\boldsymbol {\alpha} !} \sum_ {\boldsymbol {\nu} \in \bar {\Delta} _ {m}} (\frac {\boldsymbol {\nu}}{n} - \mathbf {q}) ^ {\boldsymbol {\alpha}} b _ {n, \boldsymbol {\nu}} (\mathbf {q}) \right\} \right\| _ {\infty} \\ + \left\| \sum_ {\| \boldsymbol {\alpha} \| _ {1} = k + 1} (B _ {n} - I) ^ {\lceil (k + 1) / 2 \rceil - 1} \left\{\frac {\partial^ {\boldsymbol {\alpha}} f (\boldsymbol {\xi})}{\boldsymbol {\alpha} !} \sum_ {\boldsymbol {\nu} \in \bar {\Delta} _ {m}} (\frac {\boldsymbol {\nu}}{n} - \mathbf {q}) ^ {\boldsymbol {\alpha}} b _ {n, \boldsymbol {\nu}} (\mathbf {q}) \right\} \right\| _ {\infty} \\ \lesssim_ {k + 1, m} \sum_ {\| \boldsymbol {\alpha} \| _ {1} = 1} ^ {k} \| (B _ {n} - I) ^ {\lceil \| \boldsymbol {\alpha} \| _ {1} / 2 \rceil - 1} \| _ {\infty} \| f \| _ {C ^ {k + 1} (\Delta_ {m})} n ^ {- (k + 1) / 2} \\ + \| \left(B _ {n} - I\right) ^ {\lceil (k + 1) / 2 \rceil - 1} \| _ {\infty} \| f \| _ {C ^ {k + 1} \left(\Delta_ {m}\right)} n ^ {- (k + 1) / 2} \\ \lesssim_ {k + 1, m} \left\| f \right\| _ {C ^ {k + 1} \left(\Delta_ {m}\right)} n ^ {- (k + 1) / 2}. \\ \end{array} +$$ + +The last inequality holds when $\| B_n - I \|_{\infty}$ is bounded by a constant independent of $n$ . In fact, $\| (B_n - I)f \|_{\infty} = \sup_{\mathbf{q} \in \Delta_m} |(B_n - I)f(\mathbf{q})| = \sup_{\mathbf{q} \in \Delta_m} \left| \sum_{\nu \in \bar{\Delta}_m} \{f(\nu / n) - f(\mathbf{q})\} b_{n,\nu}(\mathbf{q}) \right| \leq 2 \| f \|_{\infty}$ and $\| B_n - I \|_{\infty} = \sup_{f \in C^k(\Delta_m), \| f \|_{\infty} \leq 1} \| (B_n - I)f \|_{\infty} \leq \sup_{f \in C^k(\Delta_m), \| f \|_{\infty} \leq 1} 2 \| f \|_{\infty} \leq 2$ . + +Proof of Theorem 2. The first claim follows from Lemma 1 and the fact that $\max_{s\in [m]}\| g_s\|_{C^{2k}(\Delta_m)}\leq G$ and $\mathbb{E}_{X^n}\{D_{n,k}(g_s)(\mathbf{T} / n)\} = C_{n,k}(g_s)(\mathbf{q})$ . + +Additionally, we have + +$$ +\begin{array}{l} D _ {n, k} (g _ {s}) (\mathbf {q}) = \sum_ {j = 0} ^ {k - 1} \binom {k} {j + 1} (- 1) ^ {j} B _ {n} ^ {j} (g _ {s}) (\mathbf {q}) \\ = \sum_ {j = 0} ^ {k - 1} \binom {k} {j + 1} (- 1) ^ {j} \left\{g _ {s} (\mathbf {q}) + \mathcal {O} _ {k, m, G} (n ^ {- 1}) \right\} \\ = g _ {s} (\mathbf {q}) + \mathcal {O} _ {k, m, G} (n ^ {- 1}), \\ \end{array} +$$ + +and + +$$ +\begin{array}{l} \mathbb {E} _ {X ^ {n}} \left[ \left\{D _ {n, k} \left(g _ {s}\right) (\mathbf {T} / n) \right\} ^ {2} \right] = \sum_ {\boldsymbol {\nu} \in \bar {\Delta} _ {m}} \left\{D _ {n, k} \left(g _ {s}\right) (\boldsymbol {\nu} / n) \right\} ^ {2} b _ {n, \boldsymbol {\nu}} (\mathbf {q}) \\ = B _ {n} \left[ \left\{D _ {n, k} (g _ {s}) \right\} ^ {2} \right] (\mathbf {q}) \\ = \left\{D _ {n, k} \left(g _ {s}\right) (\mathbf {q}) \right\} ^ {2} + \mathcal {O} _ {k, m, G} \left(n ^ {- 1}\right) \\ = \left\{g _ {s} (\mathbf {q}) + \mathcal {O} _ {k, m, G} (n ^ {- 1}) \right\} ^ {2} + \mathcal {O} _ {k, m, G} (n ^ {- 1}) \\ = g _ {s} ^ {2} (\mathbf {q}) + \mathcal {O} _ {k, m, G} (n ^ {- 1}). \\ \end{array} +$$ + +Therefore, + +$$ +\begin{array}{l} \operatorname {V a r} _ {X ^ {n}} \left\{D _ {n, k} \left(g _ {s}\right) (\mathbf {T} / n) \right\} = \mathbb {E} _ {X ^ {n}} \left[ \left\{D _ {n, k} \left(g _ {s}\right) (\mathbf {T} / n) \right\} ^ {2} \right] - \left[ \mathbb {E} _ {X ^ {n}} \left\{D _ {n, k} \left(g _ {s}\right) (\mathbf {T} / n) \right\} \right] ^ {2} \\ = g _ {s} ^ {2} (\mathbf {q}) + \mathcal {O} _ {k, m, G} (n ^ {- 1}) - \left\{g _ {s} (\mathbf {q}) + \mathcal {O} _ {k, m} (n ^ {- k}) \right\} ^ {2} \\ = \mathcal {O} _ {k, m, G} (n ^ {- 1}). \\ \end{array} +$$ + +![](images/63ebc0cff6b2956f07f9313fc2806cd68e7a5fffc8886533fd2a1cfe55757009.jpg) + +# C Proof of Theorem 3 + +Proof of Theorem 3. By Theorem 2, it suffices to let $\widetilde{P}_{X^n}(x = u_s|y = y^*) = D_{n,k}(g_s)(\mathbf{T} / n)$ . Moreover, we have $\sum_{s=1}^{m}D_{n,k}(g_s)(\mathbf{T} / n) = 1$ . \ No newline at end of file diff --git a/NeurIPS/2025/A Black-Box Debiasing Framework for Conditional Sampling/images.zip b/NeurIPS/2025/A Black-Box Debiasing Framework for Conditional Sampling/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..90a3a088758e72051e57f2a93c47c02ea2d877ce --- /dev/null +++ b/NeurIPS/2025/A Black-Box Debiasing Framework for Conditional Sampling/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:27b3ed650f7bb4feda66444a399f356e7760bd80383543661f4ca977921ded87 +size 1477204 diff --git a/NeurIPS/2025/A Black-Box Debiasing Framework for Conditional Sampling/layout.json b/NeurIPS/2025/A Black-Box Debiasing Framework for Conditional Sampling/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b2d23c9d8fcdf189148f08035e1bb25064be2799 --- /dev/null +++ b/NeurIPS/2025/A Black-Box Debiasing Framework for Conditional Sampling/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2d23bae83b18a9c9e8775f626d869980df1f4c7d56bc1879fffdc5b13342fae8 +size 1342041 diff --git a/NeurIPS/2025/A Cautionary Tale on Integrating Studies with Disparate Outcome Measures for Causal Inference/c8832069-6602-4780-8eac-0fbeccda8859_content_list.json b/NeurIPS/2025/A Cautionary Tale on Integrating Studies with Disparate Outcome Measures for Causal Inference/c8832069-6602-4780-8eac-0fbeccda8859_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..84415e3f51f15e63547988fe5a65e54a563bf46e --- /dev/null +++ b/NeurIPS/2025/A Cautionary Tale on Integrating Studies with Disparate Outcome Measures for Causal Inference/c8832069-6602-4780-8eac-0fbeccda8859_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1cbfd145be2817881a7be8179273a4cc4e0ac5819f88317813ca26139b359219 +size 118887 diff --git a/NeurIPS/2025/A Cautionary Tale on Integrating Studies with Disparate Outcome Measures for Causal Inference/c8832069-6602-4780-8eac-0fbeccda8859_model.json b/NeurIPS/2025/A Cautionary Tale on Integrating Studies with Disparate Outcome Measures for Causal Inference/c8832069-6602-4780-8eac-0fbeccda8859_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b9c82e9ce201aefd96335491f477c4fc60e80e37 --- /dev/null +++ b/NeurIPS/2025/A Cautionary Tale on Integrating Studies with Disparate Outcome Measures for Causal Inference/c8832069-6602-4780-8eac-0fbeccda8859_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:aa4acfdab28c50fed1c84b8b2df01cdd9dd896320aad23e55f5080b52e2c55af +size 139801 diff --git a/NeurIPS/2025/A Cautionary Tale on Integrating Studies with Disparate Outcome Measures for Causal Inference/c8832069-6602-4780-8eac-0fbeccda8859_origin.pdf b/NeurIPS/2025/A Cautionary Tale on Integrating Studies with Disparate Outcome Measures for Causal Inference/c8832069-6602-4780-8eac-0fbeccda8859_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..247fe35ffcc0721a839824978610de37856e5346 --- /dev/null +++ b/NeurIPS/2025/A Cautionary Tale on Integrating Studies with Disparate Outcome Measures for Causal Inference/c8832069-6602-4780-8eac-0fbeccda8859_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4cb79dd16b2c1229088dfb373a44637fb69f4545f583d4d452d752690bc68694 +size 4371452 diff --git a/NeurIPS/2025/A Cautionary Tale on Integrating Studies with Disparate Outcome Measures for Causal Inference/full.md b/NeurIPS/2025/A Cautionary Tale on Integrating Studies with Disparate Outcome Measures for Causal Inference/full.md new file mode 100644 index 0000000000000000000000000000000000000000..2c455b78b07de70b806e8aeca0e44da9813d6570 --- /dev/null +++ b/NeurIPS/2025/A Cautionary Tale on Integrating Studies with Disparate Outcome Measures for Causal Inference/full.md @@ -0,0 +1,546 @@ +# A Cautionary Tale on Integrating Studies with Disparate Outcome Measures for Causal Inference + +Harsh Parikh + +Yale University + +harsh.parikh@yale.edu + +Trang Quynh Nguyen + +Johns Hopkins University + +trang.nguyen@jhu.edu + +Elizabeth A. Stuart + +Johns Hopkins University + +estuart@jhu.edu + +Kara Rudolph + +Columbia University + +kr2854@cumc.columbia.edu + +Caleb Miles + +Columbia University + +cm3825@cumc.columbia.edu + +# Abstract + +Data integration approaches are increasingly used to enhance the efficiency and generalizability of studies. However, a key limitation of these methods is the assumption that outcome measures are identical across datasets – an assumption that often does not hold in practice. Consider the following opioid use disorder (OUD) studies: the XBOT trial and the POAT study, both evaluating the effect of medications for OUD on withdrawal symptom severity (not the primary outcome of either trial). While XBOT measures withdrawal severity using the subjective opiate withdrawal scale, POAT uses the clinical opiate withdrawal scale. We analyze this realistic yet challenging setting where outcome measures differ across studies and where neither study records both types of outcomes. Our paper studies whether and when integrating studies with disparate outcome measures leads to efficiency gains. We introduce three sets of assumptions – with varying degrees of strength – linking both outcome measures. Our theoretical and empirical results highlight a cautionary tale: integration can improve asymptotic efficiency only under the strongest assumption linking the outcomes. However, misspecification of this assumption leads to bias. In contrast, a milder assumption may yield finite-sample efficiency gains, yet these benefits diminish as sample size increases. We illustrate these trade-offs via a case study integrating the XBOT and POAT datasets to estimate the comparative effect of two medications for opioid use disorder on withdrawal symptoms. By systematically varying the assumptions linking the SOW and COW scales, we show potential efficiency gains and the risks of bias. Our findings emphasize the need for careful assumption selection when fusing datasets with differing outcome measures, offering guidance for researchers navigating this common challenge in modern data integration. + +# 1 Introduction + +Robust decision-making increasingly depends on integrating information from diverse sources – a practice commonly referred to as data integration. By harnessing complementary datasets, researchers can improve the accuracy, generalizability, and efficiency of statistical inference (Bareinboim and Pearl, 2016). In the realm of causal inference, data integration has emerged as a central focus, recently cited among the top ten priorities for advancing the field (Mitra et al., 2022). This surge of interest reflects its wide-ranging utility: from generalizing or transporting evidence (Degtiar and Rose, 2023; Parikh et al., 2024; Huang and Parikh, 2024), to heterogeneous causal effect estimation (Brantner + +et al., 2023), boosting statistical efficiency (Rosenman et al., 2023), and mitigating bias (Kallus et al., 2018). + +However, in many real-world scenarios, various data sources may capture outcomes that, while related, are not identical to those measured in the trial. For example, in studies on medications for opioid use disorder (MOUD), the intensity of withdrawal symptoms can be measured using two different scales: the Clinical Opiate Withdrawal Scale (COWS) and the Subjective Opiate Withdrawal Scale (SOWS) (Wesson and Ling, 2003; Handelsman et al., 1987). In the XBOT trial that compared the effectiveness of injection naltrexone to sublingual buprenorphine in terms of reducing risks of returning to regular opioid use, withdrawal symptoms were measured using SOWS (Lee et al., 2018). However, the POATS study, which compared the effectiveness of adding counseling to sublingual buprenorphine treatment, used COWS to measure the strength of withdrawal symptoms (Weiss et al., 2010). Despite the differences in outcome measures, researchers might wish to leverage the POATS study to improve the precision of treatment effect estimates in the XBOT trial (or vice versa). This raises an important question: when can integration of primary study and auxiliary data with disparate outcome measures yield efficiency gains for causal effect estimates if neither study has observation of both outcome measures on the same group of individuals? + +Contributions. Our paper addresses this question by examining scenarios in which neither the trial nor the auxiliary data records both outcome measures on the same set of individuals. + +- We formulate a principal assumption that connects the primary outcome in the trial with the auxiliary outcome in external data, offering a conceptual "license" to borrow strength from auxiliary sources. We present three versions of this assumption - ranging from strong to weak - thereby providing a flexible framework that reflects varying degrees of identifiability. +- We characterize the conditions under which integrating studies can improve semiparametric efficiency as well as finite sample gains. We show that asymptotic gains are only possible under the strongest assumptions (albeit at a risk of some bias). However, under milder (and perhaps more realistic) conditions, finite-sample improvements may be realized, although these benefits diminish as sample sizes grow. +- We illustrate these insights through simulation studies and a real-world case study from the MOUD trial. Our findings underscore both the promise and the limitations of using auxiliary data with non-overlapping outcomes. Importantly, we provide practical guidance for researchers aiming to navigate these tradeoffs in applied causal inference settings. + +In a nutshell, this paper presents a cautionary framework for data integration in the presence of disparate outcomes, showing that while such integration may yield marginal gains under ideal conditions, it carries a significant risk of bias when assumptions are violated – as illustrated by our case study. To the best of our knowledge, we present the first formal quantification of this tradeoff, emphasizing the need for scrutiny before applying such methods in practice. + +The paper is organized as follows. Section 3 introduces the notation, setup, and standard assumptions. Section 4 presents the key structural assumption linking primary and auxiliary outcomes, along with three scenarios that reflect varying degrees of prior knowledge about this relationship. Sections 4.2-4.4 contains our main theoretical contributions: semiparametric efficiency bounds under each scenario, as well as worst-case bounds on finite-sample estimation errors. In Section 5, we apply these methods to estimate the causal effect of medications for opioid use disorder (MOUD) on withdrawal severity, using SOWS (from the XBOT trial) and COWS (from the POAT study). Section 6 concludes with a summary of key findings, limitations, and directions for future research. Appendix A presents simulation results evaluating estimator performance across varying sample sizes and dimensions. Additional theoretical discussion and proofs are provided in Appendices B, C, and D. + +# 2 Relevant Literature + +We briefly review four bodies of literature related to our work: (i) data integration in causal inference, (ii) meta-analysis, (iii) data harmonization, and (iv) surrogate outcomes. + +Data Integration for Causal Inference. Data integration has emerged as a central focus in causal inference, recently cited among the top priorities for advancing the field (Mitra et al., 2022). It + +supports a wide range of goals, including generalizing evidence across populations (Degtiar and Rose, 2023; Pearl, 2015; Parikh et al., 2024; Huang and Parikh, 2024), estimating heterogeneous effects (Brantner et al., 2023), boosting efficiency (Li and Luedtke, 2023), and mitigating bias (Kallus et al., 2018; Parikh et al., 2023b). Recent methods improve efficiency by combining auxiliary datasets while controlling bias, such as James-Stein shrinkage (Rosenman et al., 2023), semiparametric estimators (Yang et al., 2020), bias correction (Kallus et al., 2018; Yang and Ding, 2020), and Bayesian borrowing (Lin et al., 2024). However, these approaches typically assume consistent measures – including outcomes – across datasets. + +Meta-Analysis and Evidence Synthesis. When outcomes differ, naïve pooling can induce substantial bias (Van Cleave et al., 2011). Early evidence synthesis methods, such as standardizing outcomes (Murad et al., 2019; Deeks et al., 2019), rely on strong equivalence assumptions. Traditional meta-analyses, as in (Deeks et al., 2019), uses heuristics like dichotomization or normalization, assuming commensurability across studies (Murad et al., 2019). More sophisticated approaches jointly model multiple outcomes, using multivariate Bayesian methods (Bujkiewicz et al., 2016) or multi-task learning analogs (Zhang and Yang, 2018). These frameworks exploit known outcome dependencies or co-measurement of outcomes to synthesize information while allowing outcome-specific variation. + +Data Harmonization. Data harmonization methods are a set of tools that aim to equate measures across data sources to facilitate data integration. These methods typically align heterogeneous outcomes through co-calibration (Nance et al., 2017) or latent constructs (Snively et al., 2014). Bridge studies, where multiple outcomes are measured on the same set of individuals, can estimate mappings between outcome measures, while latent variable models treat observed outcomes as noisy indicators of a shared construct. These approaches typically require both outcome measurements for the same individual and introduce additional modeling assumptions. + +Leveraging Surrogate Outcomes. Another relevant literature is on data integration methods leveraging studies with surrogate outcomes. For instance, Athey et al. (2019) and Ghassami et al. (2022) combine experimental data with short-term outcome measures with an observational study where long-term outcome is measured to yield a consistent estimate of the long-term treatment effect. Surrogate indices that aggregate multiple proxies can substantially improve efficiency (Ghassami et al., 2022), but rely on strong structural assumptions about the proxy–outcome relationship. Existing methods generally require at least one dataset with measurement of primary and surrogate outcomes on the same set of individuals – an assumption often violated in practice and one that motivates our work. + +# 3 Preliminaries + +Setup and Notations. We consider two studies: a primary study ( $S = 0$ ) and an auxiliary study ( $S = 1$ ). The primary study observes the outcome of interest $Y$ , while the auxiliary study observes a related but distinct outcome $W$ . Crucially, $Y$ and $W$ are never observed for the same individual. In both studies, we observe treatment $T \in \{0,1\}$ and covariates $X$ . Let $Y(t)$ and $W(t)$ denote the potential outcomes under treatment $T = t$ . To unify notation, define the observed outcome as $V \coloneqq (1 - S)Y + SW$ , and the observed data as $O \coloneqq (X,S,T,V)$ . We let $\mathcal{S}_n = \{O_1,\dots,O_n\}$ denote a sample of $n$ units, with $n_0$ and $n_1$ representing the number of units in the primary and auxiliary studies, respectively. + +For any (random) function $f$ , let $\mathbb{E}[f(A)]$ denote the expectation, $\mathcal{P}_n(f(A)) = \frac{1}{n}\sum_{i=1}^{n}f(A_i)$ the empirical average, and $\mathcal{P}(f(A)) = \int f(a)dP(a)$ the population average treating $f$ as fixed. Note that $\mathbb{E}[f(A)]$ integrates over randomness in both $A$ and $f$ , while $\mathcal{P}(f(A))$ treats $f$ as fixed. We also define the $L^q(P)$ norm as $\| f\|_q = \left(\int |f(o)|^q dP(o)\right)^{1/q}$ . Further, for compactness, we write $\mu_A(B = b) := \mathbb{E}[A \mid B = b]$ to denote the conditional expectation of $A$ given $B = b$ , and $\nu_A^t(B = b) := \mathbb{E}[A(t) \mid B = b]$ for the conditional mean of the potential outcome $A(t)$ . + +Our goal is to estimate the conditional average treatment effect (CATE): $\tau_0(x) \coloneqq \nu_Y^1(X = x) - \nu_Y^0(\bar{X} = x)$ , and the average treatment effect (ATE): $\tau_0 \coloneqq \nu_Y^1 - \nu_Y^0$ , both defined with respect to the primary outcome $Y$ . + +Assumptions & Identification We make the following assumptions: + +A.1. (S-ignorability) $\forall x, \quad Y(t), W(t) \perp S \mid X = x$ . +A.2. (Treatment Positivity) $\epsilon < P(T = t \mid X, S = 0) < 1 - \epsilon$ , for all $t \in \{0, 1\}$ . +A.3. (Sampling Positivity) $\epsilon < P(S = 0 \mid X) < 1 - \epsilon$ . +A.4. (Conditional Ignorability) $\forall x,s,\quad Y(t),W(t)\perp T\mid X = x,S = s.$ + +We assume the following structural models for the potential outcomes: + +$$ +Y (t) = \theta (X) t + g (X) + \gamma , \quad \gamma \sim \mathcal {N} \left(0, \sigma_ {Y} ^ {2}\right), \tag {1} +$$ + +$$ +W (t) = \phi (X) t + f (X) + \delta , \quad \delta \sim \mathcal {N} \left(0, \sigma_ {W} ^ {2}\right), \tag {2} +$$ + +where $\theta(X)$ and $\phi(X)$ are the treatment effect functions for $Y$ and $W$ , respectively. These formulations is commonly used in the causal inference literature (Robinson, 1988; Chernozhukov et al., 2018; Hahn et al., 2020; Rudolph et al., 2025). From Equation (1), it follows that the CATE in the primary population is: $\tau_0(x) = \mathbb{E}[Y(1) - Y(0) \mid X = x] = \theta(x)$ . By Assumptions A.2. and A.4., the potential outcome means $\nu_Y^t(X = x)$ are identified by observed data as: $\nu_Y^t(X = x) = \mu_Y(X = x, T = t)$ . Hence, the CATE is identified as: $\tau_0(x) = \mu_Y(X = x, T = 1) - \mu_Y(X = x, T = 0)$ . + +Influence Function. Let $\eta$ denote the collection of nuisance parameters, specifically $\eta = \{\mu_V(X,S),\mu_T(X,S),\mu_S(X)\}$ . Let $\theta_0$ be the true parameter of interest, and $\eta_0$ the true nuisance parameters governing the data-generating process. For any regular, consistent, and asymptotically linear estimator $\hat{\theta}$ of $\theta_0$ , there exists a function $\psi$ - called the influence function - such that we decompose the estimation error $\hat{\theta} -\theta_0$ using the von Mises expansion as: + +$$ +\hat {\theta} - \theta_ {0} = \underbrace {(\mathcal {P} _ {n} - \mathcal {P}) \psi (O ; \theta_ {0} , \eta_ {0})} _ {M _ {1}} - \underbrace {\mathcal {P} [ \psi (O ; \theta_ {0} , \hat {\eta}) - \psi (O ; \theta_ {0} , \eta_ {0}) ]} _ {M _ {2} (\hat {\eta})} + M _ {3} (\hat {\eta}) +$$ + +where $\mathcal{P}[\psi (O;\theta_0,\eta_0)] = 0$ (Tsiatis, 2006; Kennedy, 2016). Here, (i) the first term, $M_{1}$ , represents sampling variability resulting in asymptotic variance - capturing the first-order behavior of $\hat{\theta}$ and reflects its asymptotic linearity (Ichimura and Newey, 2022; Kennedy, 2016); (ii) the second term, $M_2(\hat{\eta})$ , captures bias due to finite sample estimation of nuisance functions; (iii) the third term, $M_3(\hat{\eta})$ , accounts for remaining higher-order approximation error that converges to 0 in probability at rate faster than $\sqrt{n}$ . + +By the Central Limit Theorem and the Slutsky theorem, the estimator is asymptotically normal (provided Donsker condition holds or sample splitting is used): + +$$ +\sqrt {n} (\hat {\theta} - \theta_ {0}) \rightsquigarrow \mathcal {N} \left(0, \mathbb {E} [ \psi (O; \theta_ {0}, \eta_ {0}) \psi (O; \theta_ {0}, \eta_ {0}) ^ {T} ]\right). +$$ + +The asymptotic variance of $\hat{\theta}$ is thus determined by the variance of the influence function and can be consistently estimated via the empirical variance of the estimated influence function, $\hat{\psi}$ . + +Efficient Influence Function. The influence function depends on the values of $\theta$ and $\eta$ , although we suppress this dependency for notational convenience. Among all influence functions corresponding to regular, asymptotically linear estimators of $\theta_0$ , the efficient influence function (EIF), denoted $\psi^*$ , achieves the smallest possible asymptotic variance. This minimal variance – known as the semiparametric efficiency bound – is given by: $\mathbb{E}[\psi^{*}(O;\theta_{0},\eta_{0})\psi^{*}(O;\theta_{0},\eta_{0})^{T}]$ , and represents the best achievable precision for unbiased estimation (Tsiatis, 2006; Newey, 1990). + +**Procedure to Derive EIF.** Consider the log-likelihood $\mathcal{L}(O; \theta, \eta)$ of observed data with parameter of interest $\theta$ and nuisance parameters $\eta$ , maximized at $(\theta_0, \eta_0)$ . We define the score functions with respect to $\theta$ and $\eta$ as $R_{\theta}(O; \theta_0, \eta_0) = \frac{\partial \mathcal{L}}{\partial \theta} \big|_{\theta_0, \eta_0}$ and $R_{\eta}(O; \theta_0, \eta_0) = \frac{\partial \mathcal{L}}{\partial \eta} \big|_{\theta_0, \eta_0}$ , respectively. $R_{\theta}$ reflects the sensitivity of the likelihood to $\theta$ . However, it may also be sensitive to $\eta$ . Projecting it orthogonally to the space spanned by $R_{\eta}$ isolates the component of information unique to $\theta$ . The efficient score function is the residual of $R_{\theta}$ after projecting out components in the linear span of $R_{\eta}$ : + +$$ +R ^ {*} (O; \theta_ {0}, \eta_ {0}) = R _ {\theta} - \Pi \left(R _ {\theta} | \Lambda_ {\eta}\right), +$$ + +where $\Pi(R_{\theta}, |, \Lambda_{\eta}) = \mathbb{E}\left[R_{\theta}R_{\eta}^{T}\right]\left\{\mathbb{E}\left[R_{\eta}R_{\eta}^{T}\right]\right\}^{-1}R_{\eta}$ and arguments $(O; \theta_0, \eta_0)$ are suppressed for brevity. The efficient influence function is given by $\psi^{*} = \left\{\mathbb{E}\left[R^{*}(R^{*})^{T}\right]\right\}^{-1}R^{*}$ . This influence function achieves the semiparametric efficiency bound and serves as the optimal estimating function for $\theta$ under the given model. For further discussion and derivation, we refer readers to Tsiatis (2006). + +# 4 Data Integration with Disparate Outcome Measures + +To leverage auxiliary data for estimating treatment effects on the primary outcome $Y$ , we must establish a relationship between $Y$ and the auxiliary outcome $W$ . We posit the following structural assumption that provides a foundation – or “license” – for incorporating $W$ into the analysis: + +A.5. (Outcome Link Assumption) For all $x$ and $t$ , there exist functions $\alpha$ and $\beta$ of pre-treatment covariates such that $\nu_{Y}^{t}(x) = \alpha(x)\nu_{W}^{t}(x) + \beta(x)$ + +Remark 1 (On Assumption A.5.). The assumption allows for flexible and heterogeneous relationships between primary and auxiliary outcomes across units with different values of $X$ . However, this assumption also imposes structural restrictions on the relationship: the primary outcome is a partially linear function of the auxiliary outcome $W$ , with the scaling factor $\alpha(X)$ and shift $\beta(X)$ modulated by pre-treatment covariates $X$ . + +This assumption is plausible in settings where $W$ serves as a meaningful proxy for $Y$ . For instance, in biomedical studies, $W$ might represent a surrogate endpoint (e.g., a biomarker) that reflects the underlying disease progression captured by $Y$ (Weir and Walley, 2006). In such cases, prior studies or mechanistic understanding can inform how changes in $W$ relate to changes in $Y$ . + +# 4.1 Assumption Sets on $\alpha(X)$ and $\beta(X)$ + +To explore the range of identifiability and efficiency in leveraging auxiliary data, we consider three increasingly weaker assumptions about prior knowledge about $\alpha(X)$ and $\beta(X)$ : + +A.5(a) Fully Known Link: Both $\alpha(X)$ and $\beta(X)$ are known from prior domain knowledge. +A.5(b) Partially Known Link: Only $\beta (X)$ is known; $\alpha (X)$ is unknown. +A.5(c) Unknown Link: Neither $\alpha (X)$ nor $\beta (X)$ are known. + +Assumption A.5(a) represents the strongest assumption, and is most tenable in domains with well-characterized mechanistic knowledge – such as certain areas of biology, pharmacology, or engineering – where $\alpha(X)$ and $\beta(X)$ are grounded in empirical studies or physical theory (Puniya et al., 2018; Parikh et al., 2023a). In such contexts, auxiliary outcomes can be confidently incorporated using known mappings to the primary outcome. Assumption A.5(b) relaxes this requirement by assuming only the baseline shift $\beta(X)$ is known. This is common in applications where historical data or expert knowledge informs baseline trends, but the strength of association (i.e., scaling) between $Y$ and $W$ varies across populations or settings. Such partial knowledge arises frequently in social sciences or public health (Handelsman et al., 1987). Assumption A.5(c)) is the most general and aligns with many real-world scenarios where no prior information is available about the relationship between $Y$ and $W$ . This assumption allows maximum flexibility, but also introduces the greatest challenge in using an auxiliary study. + +These assumptions represent a spectrum of tradeoffs between realism and statistical precision. Stronger assumptions enable tighter and more efficient estimation but rely more heavily on prior knowledge. We make this tradeoff explicit in Sections 4.2, 4.3 and 4.4. Ultimately, the appropriate assumption set depends on the context and credibility of available domain knowledge. + +Function Class Complexity Assumption. We make additional assumptions about the complexity of functions in A.5.: + +A.6. There exists positive constants $\varepsilon > 0$ such that $\mathcal{A}$ and $\mathcal{B}$ satisfies the covering number bound: $\log N(\varepsilon, \mathcal{A}, \| \cdot \|) = O(\varepsilon^{-\omega_{\alpha}})$ , and $\log N(\varepsilon, \mathcal{B}, \| \cdot \|) = O(\varepsilon^{-\omega_{\beta}})$ . Further, we assume that the function class for $\mu_{Y}$ and $\mu_{W} - \mathcal{M}$ - satisfies covering number bounds: $\log N(\varepsilon, \mathcal{M}, \| \cdot \|) = O(\varepsilon^{-\omega})$ , with $\omega_{\alpha} + \omega_{\beta} \leq \omega$ . + +This assumption imposes regularity conditions on the function classes involved in the decomposition of $\mu_Y(X)$ into $\alpha (X)$ and $\beta (X)$ . Specifically, it ensures that the combined complexity of $\alpha$ and $\beta$ , measured via covering number bounds, does not exceed that of $\mu_Y$ . This is a mild and natural requirement: if the auxiliary outcome $W$ is informative about $Y$ , then the residual mapping captured by $\alpha (X)$ is expected to be simpler than modeling $\mu_Y(X)$ directly. In this sense, Assumption A.6 reflects a form of functional regularization, where using a predictive surrogate reduces the effective complexity of the learning task. + +# 4.2 Semiparametric Efficiency Bounds + +Now, we derive the efficient bounds under each of the following three assumptions A.5(a), A.5(b), and A.5(c), and investigate if and when data integration yields semiparametric efficiency gains. Throughout this section, we assume that assumptions A.1. to A.4. hold. + +Recall $\psi^{*}(O;\theta_{0},\eta_{0}) = \left\{\mathbb{E}[R^{*}(O;\theta_{0},\eta_{0})R^{*}(O;\theta_{0},\eta_{0})^{T}]\right\}^{-1}R^{*}(O;\theta_{0},\eta_{0})$ and the semiparametrically efficient asymptotic variance (i.e., efficiency bound) is equal to $\left\{\mathbb{E}[R^{*}(O;\theta_{0},\eta_{0})R^{*}(O;\theta_{0},\eta_{0})^{T}]\right\}^{-1}$ , we only present the efficient score function $R^{*}$ instead of the EIF $\psi^{*}$ . However, note that deriving the EIF from $R^{*}$ is straightforward in our context. In our case, $P(O = o;\theta ,\eta) = P(X = x)P(S = s\mid X = x)P(T = t\mid X = x,S = s)P(V = v\mid T = t,X = x,S = s)$ and $\mathcal{L}(O;\theta ,\eta) = \log P(O = o;\theta ,\eta)$ . + +First, we derive the efficiency bound for the semiparametrically efficient estimator that only uses the primary study. We use this result as a base case to compare the efficiency bounds for the data integration-based estimators. This efficiency bound is akin to the one derived in (Robinson, 1988). + +Theorem 1 (Efficiency bound using only primary data). Under assumptions A.1.-A.4., the efficient score function using only the primary study $(S = 0)$ is $R_0^*(O; \theta_0, \eta_0) = (1 - S) \cdot \Delta_0$ . The corresponding asymptotic variance is $\mathbb{V}_0^\theta(X) = \left( \mathbb{E} \left[ \Delta_0^2 \mid S = 0, X \right] p(S = 0 \mid X) \right)^{-1}$ , where $\Delta_0 = \left( (V - \mu_Y(X, 0) - \theta(X)(T - \mu_T(X, 0))) \cdot \frac{T - \mu_T(X, 0)}{\sigma_Y^2} \right)$ . + +Now, we derive the efficiency bound that leverages auxiliary data under A.5(a). + +Theorem 2 (Efficiency bound under known $\alpha(X)$ and $\beta(X)$ ). Under assumptions A.1.-A.4. and A.5(a), the efficient score function is: + +$$ +R _ {a} ^ {*} (O; \theta_ {0}, \eta_ {0}) = S \cdot \Delta_ {1} + (1 - S) \cdot \Delta_ {0}, +$$ + +where + +$$ +\Delta_ {1} = (\alpha (X) (V - \mu_ {W} (X, 1)) - \theta (X) (T - \mu_ {T} (X, 1))) \cdot \frac {T - \mu_ {T} (X , 1)}{\alpha^ {2} (X) \sigma_ {W} ^ {2}}. +$$ + +The asymptotic variance is: + +$$ +\mathbb {V} _ {a} ^ {\theta} (X) = \left(\mathbb {E} \left[ \Delta_ {0} ^ {2} \mid S = 0, X \right] p (S = 0 \mid X) + \mathbb {E} \left[ \Delta_ {1} ^ {2} \mid S = 1, X \right] p (S = 1 \mid X)\right) ^ {- 1}. +$$ + +Corollary 1. Integrating primary and auxiliary data under assumption A.5(a) yields efficiency gain i.e. $\mathbb{V}_a^\theta (X)\leq \mathbb{V}_0^\theta (X)$ + +Next, we derive the efficiency score and the efficiency bound for a case when A.5(b) holds. + +Theorem 3 (Efficiency bound under known $\beta(X)$ only). Under assumptions A.1.-A.4. and A.5(b), the efficient score function for $\theta(X)$ and $\alpha(X)$ is: + +$$ +R _ {b} ^ {*} (O; \theta_ {0}, \alpha_ {0}, \eta_ {0}) = \left( \begin{array}{c} S \cdot \Delta_ {1} + (1 - S) \cdot \Delta_ {0} \\ S \cdot \left(\frac {\theta (X) (T - \mu_ {T} (X , 1)) - \alpha (X) (V - \mu_ {W} (X , 1))}{\alpha^ {2} (X) \sigma_ {W} ^ {2}} \cdot \frac {\theta (X) (T - \mu_ {T} (X , 1))}{\alpha (X)}\right) \end{array} \right). +$$ + +The corresponding asymptotic variance-covariance matrix is: + +$$ +\boldsymbol {\Sigma} _ {b} (X) := \left( \begin{array}{c c} \mathbb {V} _ {b} ^ {\theta} (X) & \operatorname {C o v} _ {b} ^ {\theta , \alpha} (X) \\ \operatorname {C o v} _ {b} ^ {\theta , \alpha} (X) & \mathbb {V} _ {b} ^ {\alpha} (X) \end{array} \right), w i t h \mathbb {V} _ {b} ^ {\theta} (X) = \left(\mathbb {E} [ \Delta_ {0} ^ {2} \mid S = 0, X ] P (S = 0 \mid X)\right) ^ {- 1}. +$$ + +Corollary 2. The asymptotic variance of the efficient estimator of $\theta(X)$ under assumption A.5(b) is equal to that under using primary data only: $\mathbb{V}_b^\theta(X) = \mathbb{V}_0^\theta(X)$ . Thus, when $\alpha(X)$ is unknown, incorporating auxiliary data provides no efficiency gain. + +Theorem 4 (Efficiency bound under unknown $\alpha(X)$ and $\beta(X)$ ). Under assumptions A.1.-A.4. and A.5(c), the efficient score function is identical to that in Theorem 3: $R_{c}^{*}(O; \theta_{0}, \alpha_{0}, \eta_{0}) = R_{b}^{*}(O; \theta_{0}, \alpha_{0}, \eta_{0})$ . Therefore, the asymptotic variance for estimating $\theta(X)$ remains: $\mathbb{V}_{c}^{\theta}(X) = \mathbb{V}_{b}^{\theta}(X) = \mathbb{V}_{0}^{\theta}(X)$ . + +Corollary 3. If both $\alpha(X)$ and $\beta(X)$ are unknown, there are no efficiency gains from using auxiliary data compared to using only primary data. + +The proofs and results in Theorems 1 to 4 are a direct consequence of following the procedure to derive EIF described in Section 3 and are provided in Appendix B. + +# 4.3 ATE Estimation under A.5(a) + +Now, we present the ATE estimation under assumption A.5(a). We use the efficient score $R_{a}^{*}$ to guide the estimation of $\theta_0$ using the property that $\mathbb{E}[R_a^* (O;\theta_0,\eta_0)] = 0$ . Recall, that $R_{a}^{*}(O;\theta_{0},\eta_{0}) = S\Delta_{1} + (1 - S)\Delta_{0}$ . Assuming an unbiased and consistent estimate of the nuisance parameter $\hat{\eta}$ , a solution to $\frac{1}{n_0}\sum_iR_a^* (O_i;\theta ,\hat{\eta})$ - denoted by $\hat{\theta}_a$ - is an unbiased and consistent estimate of $\theta_0$ . Let $r_A(B)\coloneqq A - \mu_A(B)$ denote the residual of random variable $A$ after regressing $A$ on $B$ . Then, the estimator $\hat{\theta}_a$ is given by: + +$$ +\hat {\theta} _ {a} = \frac {\sum_ {i} \left((1 - S _ {i}) \frac {\hat {r} _ {Y} (X _ {i} , 0) \hat {r} _ {T} (X _ {i} , 0)}{\hat {\sigma} _ {Y} ^ {2}} + S _ {i} \frac {\hat {r} _ {W} (X _ {i} , 1) \hat {r} _ {T} (X _ {i} , 1)}{\alpha (X _ {i}) \hat {\sigma} _ {W} ^ {2}}\right)}{\sum_ {i} \left((1 - S _ {i}) \frac {\hat {r} _ {T} ^ {2} (X _ {i} , 0)}{\hat {\sigma} _ {Y} ^ {2}} + S _ {i} \frac {\hat {r} _ {T} ^ {2} (X _ {i} , 1)}{\alpha^ {2} (X _ {i}) \hat {\sigma} _ {W} ^ {2}}\right)} +$$ + +Misspecification Bias under A.5(a): We showed the efficiency bound under three varied assumptions and our results highlighted that efficiency gain is only feasible under the strongest assumption. Now, we investigate the cost of making the wrong assumption i.e. what happens if we assume A.5(a) but $\alpha$ is misspecified. Let $\alpha^{\star}$ denote the true $\alpha$ and $\alpha_{mis}$ denote a misspecified $\alpha$ . + +Theorem 5 (Misspecification Bias). Under assumptions A.1. - A.4. and a misspecified $A.5(a)$ , the estimator, $\hat{\theta}_a$ , is biased where the bias is equal to $\mathbb{E}\left[B(X) \mid S = 1\right]$ , where + +$$ +B (X) := \mathbb {E} \left[ \left(\frac {\left(\alpha_ {m i s} (X) - \alpha^ {\star} (X)\right)}{\alpha^ {\star} (X)}\right) \theta (X) \mid S = 1, X \right] +$$ + +# 4.4 Estimation under A.5(b) and A.5(c): Finite-Sample Gains + +In cases when $\alpha$ is unknown (i.e. A.5(b) and A.5(c)), it is not feasible to yield efficiency gains by leveraging auxiliary data. However, consider the estimator for ATE only using primary data + +$$ +\hat {\theta} _ {0} = \frac {\sum_ {i} (1 - S _ {i}) [ (\hat {r} _ {Y} (X _ {i} , 0)) (\hat {r} _ {T} (X _ {i} , 0)) ]}{\sum_ {i} (1 - S _ {i}) [ (\hat {r} _ {T} (X _ {i} , 0)) ^ {2} ]}. +$$ + +This estimator can be modified, under A.5., to use the auxiliary data to potentially have finite sample benefits. One natural approach to leveraging auxiliary data is the following two-stage estimator: in the first stage, we estimate the auxiliary regression $\mu_W(X,1) = \mathbb{E}[W|X,S = 1]$ using the auxiliary data and we then use this estimated function to predict $\hat{\mu}_W(X,0)$ for units in the primary data. In the second stage, we estimate $\mu_Y$ as: $\hat{\mu}_{Y,b}(X,0) = \hat{\alpha} (X)\hat{\mu}_W(X,0) + \hat{\beta} (X)$ where $\hat{\alpha},\hat{\beta}\in \left[\arg \min_{\alpha ,\beta \in \mathcal{A},\mathcal{B}}\frac{1}{n_0}\sum_i(1 - S_i)(Y_i - \alpha (X_i)\hat{\mu}_W(X_i,0) - \beta (X_i))^2\right]$ . The resulting fitted function $\hat{\mu}_{Y,b}(X,0)$ combines both sources of information and provides a data-adaptive estimator of the conditional mean outcome in the primary population. This approach is akin to adjusting for the prognostic or benefit score along with the vector of covariates (Liao et al., 2025). Thus, the resulting estimator leveraging the auxiliary data is given as: + +$$ +\hat {\theta} _ {b} = \frac {\sum_ {i} (1 - S _ {i}) [ (\hat {r} _ {Y , b} (X _ {i} , 0)) (\hat {r} _ {T} (X _ {i} , 0)) ]}{\sum_ {i} (1 - S _ {i}) [ (\hat {r} _ {T} (X _ {i} , 0)) ^ {2} ]}. +$$ + +Quantifying Finite-Sample Risk. Asymptotically, if nuisance estimators $\hat{\eta}$ belong to a Donsker class or are fit using sample splitting, the $M_2$ and $M_3$ vanish asymptotically at a rate + +faster than $\sqrt{n}$ . However, in finite samples, they contribute non-negligibly to estimation error. We focus on $M_2(\hat{\eta}) = \mathcal{P}(\psi (O;\theta_0,\hat{\eta}) - \psi (O;\theta_0,\eta_0))$ , which depends on the accuracy of nuisance function estimates. For cross-fitted estimators, we have: $|M_2(\hat{\eta})| = o_p\left(n^{-1 / 2}\left(\| \mu_Y - \hat{\mu}_Y\| \cdot \| \mu_T - \hat{\mu}_T\| +\theta_0\| \mu_T - \hat{\mu}_T\| ^2\right)\right)$ . Since the only difference between $\hat{\theta}_0$ and $\hat{\theta}_b$ lies in the choice of outcome regression, smaller $\| \mu_Y - \hat{\mu}_Y\|$ directly translates into precise estimates. + +Theorem 6 (Error bound for $\hat{\mu}_Y$ ). Given assumptions A.1.-A.4., A.5(b), and A.6., the empirical errors for $\hat{\mu}_{Y,0}$ and the two-stage estimator $\hat{\mu}_{Y,b}$ are + +$$ +\| \hat {\mu} _ {Y, 0} - \mu_ {Y} \| = o _ {p} \left(n _ {0} ^ {- \frac {1}{2 + \omega}}\right), a n d \| \hat {\mu} _ {Y, b} - \mu_ {Y} \| = o _ {p} \left(n _ {0} ^ {- \frac {1}{2 + \omega}} \left(n _ {0} ^ {\frac {1}{2 + \omega} - \frac {1}{2 + \omega_ {\alpha}}} + (n _ {1} / n _ {0}) ^ {- \frac {1}{2 + \omega}}\right)\right). +$$ + +Characterizing Finite Sample Gains. Theorem 6 demonstrates that one may not even achieve finite sample gains when leveraging auxiliary data. The two-stage estimator $\hat{\mu}_{Y,b}$ can outperform the direct regression estimator $\hat{\mu}_{Y,0}$ only when certain structural and sample size conditions are met. Qualitatively, gains arise when leveraging the auxiliary data allow for decomposition of $\mu_{Y}(X)$ into less complex functions. Additionally, leveraging auxiliary data helps only if $\mu_W(X)$ can be estimated accurately - that is, when the auxiliary sample size $n_1$ is sufficiently large relative to the primary sample size $n_0$ . Importantly, when the auxiliary outcome $W$ is highly predictive of the primary outcome $Y$ - that is, when $\operatorname{Cov}(Y,W|X)$ is large - the function $\mu_Y(X)$ can be well-approximated by $\mu_W(X)$ then the function $\alpha (X)$ captures only residual structure and tends to be significantly simpler than $\mu_Y(X)$ itself, implying that the entropy exponent $\omega_{\alpha}$ is relatively much smaller than $\omega$ . Quantitatively, finite-sample improvement occurs when $n_0^{\frac{1}{2 + \omega} -\frac{1}{2 + \omega_\alpha}} + (n_0 / n_1)^{\frac{1}{2 + \omega}} < 1$ . The first term captures the gain from replacing the full function class $\mathcal{M}_Y$ with a lower-complexity class $\mathcal{A}$ , and the second term reflects the accuracy of estimating $\mu_W(X)$ from the auxiliary data. Gains are most pronounced when $\omega_{\alpha}\ll \omega$ (i.e., $\alpha (X)$ is much simpler than $\mu_Y(X)$ ) and when $n_1\gg n_0$ (i.e., we have ample auxiliary data). Our characterization formally supports the intuition that structural assumptions and additional data may result in finite sample gains. + +In summary, finite-sample efficiency gains from incorporating auxiliary data arise when the function $\alpha(X)$ , capturing the dependence between Y and W, is simpler to estimate than $\mu_Y(X)$ . In such cases, one may first estimate $\mu_W(X)$ using auxiliary data, and then use the combined data to estimate $\alpha(X)$ . However, the existence and extent of these gains depend on the relative complexity of the function classes and the sample sizes involved. As $n_0$ increases, the benefit of this two-stage strategy diminishes, and asymptotically, both the direct and the auxiliary-based estimators converge to the same efficiency. + +# 5 Medication for Opioid Use Disorder and Withdrawal Symptoms + +We apply our framework to compare the effectiveness of extended-release naltrexone (XR-NTX) and buprenorphine-naloxone (BUP-NX) in reducing opioid withdrawal symptoms between 10 and 12 weeks after treatment initiation. We begin by describing the primary and auxiliary datasets and the causal quantity of interest, followed by estimates obtained under three approaches: (i) using only primary data, (ii) incorporating auxiliary data with known outcome linkage (Assumption A.5(a)), and (iii) incorporating auxiliary data under partial knowledge of the link (Assumption A.5(b)). + +# 5.1 Data Description + +Primary Study: XBOT Trial. The NIDA CTN-0051 (XBOT) trial was a multisite study comparing extended-release naltrexone (XR-NTX) and buprenorphine-naloxone (BUP-NX) for opioid use disorder treatment (Lee et al., 2018). A total of 540 patients were randomized 1:1 to receive either treatments over 24 weeks. We focus on the most severe withdrawal symptoms in the $4^{th}$ week, measured by the Subjective Opiate Withdrawal Scale (SOWS) – a 16-item self-report instrument where patients rate each symptom from 0 to 4, reflecting subjective withdrawal experiences. + +Auxiliary Study: POAT Study. The NIDA CTN-0030 (POATS) trial enrolled individuals dependent on prescription opioids for outpatient treatment using BUP-NX (Weiss et al., 2011). Withdrawal symptoms were assessed using the Clinical Opiate Withdrawal Scale (COWS), an 11-item clinician-administered tool capturing objective signs of withdrawal. We use POATS as + +auxiliary data to improve the estimation of withdrawal severity under BUP-NX in the XBOT trial, leveraging the worst COWS scores in the $4^{th}$ week. Although the XT-NTX arm is absent in the auxiliary dataset, the BUP-NX treatment is shared across both studies. We aim to evaluate if and when auxiliary information on BUP-NX can be used to improve the efficiency of estimating the outcome under BUP-NX in the primary population, while carefully considering the assumptions required for valid data fusion. + +# 5.2 Analysis + +We evaluate the comparative effectiveness of XR-NTX $(T = 1)$ versus BUP-NX $(T = 0)$ in reducing withdrawal symptom severity, as measured by the worst SOWS score in the fourth week $(Y)$ , among participants in the XBOT trial. In the auxiliary POAT study, withdrawal severity is measured on the COWS scale during the same period $(W)$ . We use a common set of covariates assessed in both trials $(X)$ . Further, we only considered patients for whom we observed the outcomes - our study excluded individuals for whom treatment was not initiated or who dropped out before our outcome window. + +To harmonize the two scales, we derive the transformation coefficient $\alpha$ from published clinical thresholds. According to Wesson and Ling (2003), COWS ranges of 5-12, 13-24, 25-36, and $>36$ correspond to mild, moderate, moderately severe, and severe withdrawal, respectively. Similarly, Handelsman et al. (1987) defines SOWS ranges of 1-10, 11-15, 16-20, and 21-30 for the same categories. Assuming both scales share a zero point (no withdrawal), we align category midpoints and estimate a linear mapping $Y = \alpha W + \varepsilon$ , yielding $\alpha = 0.61$ and intercept $\beta = 0$ . Figure 1(a) visualizes this relationship. We assume $\alpha$ is constant across covariate values $X$ , and interpret lower values of both $Y$ and $W$ as indicating better outcomes. We then apply the three estimators introduced in Section 4.3 and 4.4: $\hat{\theta}_0$ (primary data only), $\hat{\theta}_b$ (auxiliary data, unknown $\alpha$ ), and $\hat{\theta}_a$ (auxiliary data, known $\alpha$ ). As shown in Figure 1(b), $\hat{\theta}_0$ and $\hat{\theta}_b$ suggest that XR-NTX and BUP-NX are almost equally effective. However, $\hat{\theta}_a$ suggests BUP-NX is marginally more effective in lowering withdrawal symptoms compared to XR-NTX. Specifically: (i) $\hat{\theta}_0 = -0.18$ (95% CI width: 1.52), (ii) $\hat{\theta}_b = 0.42$ (95% CI width: 0.65), and (iii) $\hat{\theta}_a = -1.08$ (95% CI width: 1.41). While $\hat{\theta}_a$ achieves a statistically significant result, it relies on the correctness of the assumed $\alpha$ . To assess robustness, we conduct a sensitivity analysis by varying $\alpha$ within $\pm 50\%$ of the estimated value, i.e., $\alpha \in [0.31, 0.92]$ , assuming the linear form remains valid. Figure 1(c) displays the resulting $\hat{\theta}_a$ estimates across this range. Although the point estimates vary – from 0.50 to 0.33 – they consistently favor BUP-NX over XR-NTX. However, for $\alpha > 0.75$ , the 95% confidence intervals include zero. + +Takeaways. In our case study, we assess whether XT-NTX is more effective than BUP-NX in reducing withdrawal symptom severity. The point estimates from the primary study are slightly negative, suggesting a marginal advantage for XT-NTX, but the $95\%$ confidence interval includes zero, indicating no statistically significant difference between the two treatments (see Figure 1(b); estimate $\hat{\theta}_0$ ). To improve estimation precision, we explore leveraging auxiliary data. Under the strong assumption, our combined analysis yields a statistically significant result favoring BUP-NX over XT-NTX (see Figure 1(b); estimate $\hat{\theta}_a$ ). While such findings may appear actionable, they hinge critically on an untestable assumption linking outcomes Y and W. If this assumption is violated, the resulting estimates may be misleadingly precise. Our case study thus serves as a cautionary example: although integrating auxiliary data can improve precision, it must be done with scrutiny of the underlying assumptions, which—if invalid—can lead to confidently incorrect conclusions. + +# 6 Discussion & Conclusion + +Summary. This paper presents a principled framework for integrating primary and auxiliary datasets with non-overlapping, disparate outcomes to improve efficiency in causal effect estimation. We focus on settings where the primary outcome is never jointly observed with the auxiliary outcome, and we introduce a structural assumption that links the two. Building on this, we define three scenarios reflecting varying levels of prior knowledge about outcome relationship and derive semiparametric efficiency bounds under each. Our findings show that efficiency gains are guaranteed only under the strongest assumptions, when the linking equation is fully known. In contrast, under weaker assumptions, asymptotic efficiency is not ensured. However, finite-sample improvements are still possible, particularly when the auxiliary outcome is highly predictive of the primary outcome. These + +![](images/fc11002b6a00fce88855f3a4c5d70c09dcccad0f2900ced5fc41adb41d6756a5.jpg) +(a) + +![](images/a7e01609bf83bb351460994d77026a32b7989c915c2d8aa514f57a9cef19e1f9.jpg) +(b) +Figure 1: MOUD Results. (a) Scatter plot showing the relationship between SOWS and COWS. (b) Treatment effect estimates of MOUD on withdrawal symptoms. Point estimates and corresponding $95\%$ confidence intervals for $\hat{\theta}_0, \hat{\theta}_a$ and $\hat{\theta}_b$ . (c) Assessing the sensitivity of $\hat{\theta}_a$ to the different values of $\alpha$ in the range $50\%$ above and below the original guess of $\alpha = 0.61$ . + +![](images/88ecd4067ec1ad664401f1dd97e6a748104413cfa917ce60f20912046e3c121b.jpg) +(c) + +benefits taper off as the primary sample size increases, highlighting the limitations of auxiliary data in isolation. We support our theoretical results with both simulations and a case study estimating the effect of medications for opioid use disorder (MOUD) on withdrawal severity. Here, we combine data from the XBOT trial (SOWS scale) and the POAT study (COWS scale), demonstrating the framework's practical utility. + +Limitations & Future works. Our analysis and results in this paper depend on the structural assumption between $Y$ and $W$ . While moving ahead, we will focus on making our results more general by relaxing this assumption; it is important to note that the lack of efficiency gains in a restrictive context would imply a similar conclusion in a more complex context. Further, we will focus on incorporating a third "bridge" dataset, where both outcomes are observed, which could help relax strong assumptions and expand the conditions under which efficiency gains are possible. We will also explore relaxing the assumption of conditional study exchangeability, extending the framework to accommodate discordance in treatments and covariates across studies. Further, our framework requires that at least one treatment arm be shared across the primary and auxiliary studies. When there is no treatment overlap between datasets and the outcomes vary from one study to another, it becomes impossible to link the potential outcome distributions. This makes it difficult to use auxiliary data to enhance efficiency gains. This situation reveals a significant structural limitation in data fusion settings. In the future, we plan to explore data fusion in such scenarios by imposing a distance metric to the treatment space, which will enable us to compare different treatments. + +# Acknowledgments + +The authors would like to thank the reviewers, the area chair, and the program chair of NeurIPS 2025 for their constructive input to help improve the paper. Further, Harsh Parikh, Kara Rudolph, and Elizabeth Stuart would like to acknowledge that this work was funded by NIH NIDA R01DA056407, and Caleb Miles and Kara Rudolph would like to acknowledge that this work was funded by NIH NIDA R01DA059824. Trang Nguyen and Elizabeth Stuart were funded by NIH NIMH R01MH126856. + +# References + +Athey, S., Chetty, R., Imbens, G. W., and Kang, H. (2019). The surrogate index: Combining short-term proxies to estimate long-term treatment effects more rapidly and precisely. Technical report, National Bureau of Economic Research. +Bareinboim, E. and Pearl, J. (2016). Causal inference and the data-fusion problem. Proceedings of the National Academy of Sciences, 113(27):7345-7352. +Brantner, C. L., Chang, T.-H., Nguyen, T. Q., Hong, H., Di Stefano, L., and Stuart, E. A. (2023). Methods for integrating trials and non-experimental data to examine treatment effect heterogeneity. Statistical science: a review journal of the Institute of Mathematical Statistics, 38(4):640. + +Bujkiewicz, S., Thompson, J. R., Riley, R. D., and Abrams, K. R. (2016). Bayesian meta-analytical methods to incorporate multiple surrogate endpoints in drug development process. *Statistics in medicine*, 35(7):1063-1089. +Chernozhukov, V., Chetverikov, D., Demirer, M., Duflo, E., Hansen, C., Newey, W., and Robins, J. (2018). Double/debiased machine learning for treatment and structural parameters. +Deeks, J. J., Higgins, J. P., Altman, D. G., and Group, C. S. M. (2019). Analysing data and undertaking meta-analyses. Cochrane handbook for systematic reviews of interventions, pages 241-284. +Degtiar, I. and Rose, S. (2023). A review of generalizability and transportability. Annual Review of Statistics and Its Application, 10:501-524. +Ghassami, A., Yang, A., Richardson, D., Shpitser, I., and Tchetgen, E. T. (2022). Combining experimental and observational data for identification and estimation of long-term causal effects. arXiv preprint arXiv:2201.10743. +Györfi, L., Kohler, M., Krzyzak, A., and Walk, H. (2006). A distribution-free theory of nonparametric regression. Springer Science & Business Media. +Hahn, P. R., Murray, J. S., and Carvalho, C. M. (2020). Bayesian regression tree models for causal inference: Regularization, confounding, and heterogeneous effects (with discussion). *Bayesian Analysis*, 15(3):965–1056. +Handelsman, L., Cochrane, K. J., Aronson, M. J., Ness, R., Rubinstein, K. J., and Kanof, P. D. (1987). Two new rating scales for opiate withdrawal. The American journal of drug and alcohol abuse, 13(3):293-308. +Huang, M. Y. and Parikh, H. (2024). Towards generalizing inferences from trials to target populations. Harvard Data Science Review, 6(4). +Ichimura, H. and Newey, W. K. (2022). The influence function of semiparametric estimators. Quantitative Economics, 13(1):29-61. +Kallus, N., Puli, A. M., and Shalit, U. (2018). Removing hidden confounding by experimental grounding. Advances in Neural Information Processing Systems, 31. +Kennedy, E. H. (2016). Semiparametric theory and empirical processes in causal inference. Statistical causal inferences and their applications in public health research, pages 141-167. +Lee, J. D., Nunes, E. V., Novo, P., Bachrach, K., Bailey, G. L., Bhatt, S., Farkas, S., Fishman, M., Gauthier, P., Hodgkins, C. C., et al. (2018). Comparative effectiveness of extended-release naltrexone versus buprenorphine-naloxone for opioid relapse prevention (x: Bot): a multicentre, open-label, randomised controlled trial. The Lancet, 391(10118):309-318. +Li, S. and Luedtke, A. (2023). Efficient estimation under data fusion. Biometrika, 110(4):1041-1054. +Liao, L. D., Højbjerre-Frandsen, E., Hubbard, A. E., and Schuler, A. (2025). Prognostic adjustment with efficient estimators to unbiasedly leverage historical data in randomized trials. The International Journal of Biostatistics. +Lin, X., Tarp, J. M., and Evans, R. J. (2024). Combine experimental and observational data through a power likelihood. arXiv preprint arXiv:2304.02339. +Mitra, N., Roy, J., and Small, D. (2022). The future of causal inference. American journal of epidemiology, 191(10):1671-1676. +Murad, M. H., Wang, Z., Chu, H., and Lin, L. (2019). When continuous outcomes are measured using different scales: guide for meta-analysis and interpretation. *Bmj*, 364. +Nance, R. M., Delaney, J. C., Golin, C. E., Wechsberg, W. M., Cunningham, C., Altice, F., Christopoulos, K., Knight, K., Quan, V., Gordon, M. S., et al. (2017). Co-calibration of two self-reported measures of adherence to antiretroviral therapy. AIDS care, 29(4):464-468. + +Newey, W. K. (1990). Semiparametric efficiency bounds. Journal of applied econometrics, 5(2):99-135. +Parikh, H., Hoffman, K., Sun, H., Zafar, S. F., Ge, W., Jing, J., Liu, L., Sun, J., Struck, A., Volfovsky, A., et al. (2023a). Effects of epileptiform activity on discharge outcome in critically ill patients in the usa: a retrospective cross-sectional study. The Lancet Digital Health, 5(8):e495–e502. +Parikh, H., Morucci, M., Orlandi, V., Roy, S., Rudin, C., and Volfovsky, A. (2023b). A double machine learning approach to combining experimental and observational data. arXiv preprint arXiv:2307.01449. +Parikh, H., Ross, R., Stuart, E., and Rudolph, K. (2024). Who are we missing? a principled approach to characterizing the underrepresented population. arXiv preprint arXiv:2401.14512. +Pearl, J. (2015). Generalizing experimental findings. Journal of Causal Inference, 3(2):259-266. +Puniya, B. L., Todd, R. G., Mohammed, A., Brown, D. M., Barberis, M., and Helikar, T. (2018). A mechanistic computational model reveals that plasticity of $\mathrm{cd4 + }$ t cell differentiation is a function of cytokine composition and dosage. Frontiers in physiology, 9:878. +Robinson, P. M. (1988). Root-n-consistent semiparametric regression. *Econometrica: Journal of the Econometric Society*, pages 931-954. +Rosenman, E. T., Basse, G., Owen, A. B., and Baiocchi, M. (2023). Combining observational and experimental datasets using shrinkage estimators. Biometrics, 79(4):2961-2973. +Rudolph, K. E., Williams, N. T., Stuart, E. A., and Diaz, I. (2025). Improving efficiency in transporting average treatment effects. Biometrika. +Snavely, A. C., Harrington, D. P., and Li, Y. (2014). A latent variable transformation model approach for exploring dysphagia. Statistics in medicine, 33(25):4337-4352. +Tsiatis, A. A. (2006). Semiparametric theory and missing data, volume 4. Springer. +Tsybakov, A. B. and Tsybakov, A. B. (2009). Lower bounds on the minimax risk. Introduction to Nonparametric Estimation, pages 77-135. +Van Cleave, J. H., Egleston, B. L., Bourbonniere, M., and McCorkle, R. (2011). Combining extant datasets with differing outcome measures across studies of older adults after cancer surgery. Research in gerontological nursing, 4(1):36-45. +Weir, C. J. and Walley, R. J. (2006). Statistical evaluation of biomarkers as surrogate endpoints: a literature review. Statistics in medicine, 25(2):183-203. +Weiss, R. D., Potter, J. S., Fiellin, D. A., Byrne, M., Connery, H. S., Dickinson, W., Gardin, J., Griffin, M. L., Gourevitch, M. N., Haller, D. L., et al. (2011). Adjunctive counseling during brief and extended buprenorphine-naloxone treatment for prescription opioid dependence: a 2-phase randomized controlled trial. Archives of general psychiatry, 68(12):1238-1246. +Weiss, R. D., Potter, J. S., Provost, S. E., Huang, Z., Jacobs, P., Hasson, A., Lindblad, R., Connery, H. S., Prather, K., and Ling, W. (2010). A multi-site, two-phase, prescription opioid addiction treatment study (poats): rationale, design, and methodology. Contemporary clinical trials, 31(2):189-199. +Wesson, D. R. and Ling, W. (2003). The clinical opiate withdrawal scale (cows). Journal of psychoactive drugs, 35(2):253-259. +Yang, S. and Ding, P. (2020). Combining multiple observational data sources to estimate causal effects. Journal of the American Statistical Association, 115:1540-1554. +Yang, S., Zeng, D., and Wang, X. (2020). Improved inference for heterogeneous treatment effects using real-world data subject to hidden confounding. arXiv preprint arXiv:2007.12922. +Zhang, Y. and Yang, Q. (2018). An overview of multi-task learning. National Science Review, 5(1):30-43. + +# Appendix A Synthetic Data Study and Results + +In this section, we are interested in understanding the performance of estimators under various sets of assumptions (A.5(a) to A.5(c)). In particular, we are interested in understanding the potential gains as (i) total number of units, $n$ , increases, (ii) the dimensionality of $X$ , denoted as $p$ , increases and (iii) the log of the ratio of number of units in the auxiliary to primary dataset, $\log \left( \frac{P(S=1)}{P(S=0)} \right)$ , increases. First, we discuss our data generative procedures, and then we present and discuss our results. + +Data Generative Procedure. The data generation procedure (DGP) in this study is designed to simulate a complex causal structure. We begin by generating covariates $X = (X_{1},X_{2},\ldots ,X_{p})$ from a multivariate normal distribution with zero mean and identity covariance matrix where $p$ is the number of covariates. The binary study indicator $S$ is then generated as a Bernoulli random variable, where the probability of assignment to the auxiliary study (i.e., $S = 1$ ) is $\operatorname*{Pr}(S = 1|X) = \exp (a_0 + a_1X_1 + a_2X_2)$ , where $\exp (x) = \frac{1}{1 + e^{-x}}$ . The treatment assignment $T$ is also generated as a study and covariate-dependent Bernoulli variable $\operatorname*{Pr}(T = 1|X,S) = (1 - S)\times 0.5 + S\times \exp (\zeta_1X_1)$ . The auxiliary outcome $W$ , observed only in the auxiliary study $(S = 1)$ , is defined as follows: + +$$ +W = \mu_ {W} (X, T, S) + \delta = (\gamma_ {0} + \gamma_ {1} X _ {1} + \gamma_ {2} X _ {2}) \cdot T + \beta_ {1} X _ {1} + \beta_ {2} X _ {2} + \beta_ {3} X _ {3} + \beta_ {0} + \delta , +$$ + +where vectors $\gamma$ and $\beta$ define the treatment and baseline effects on $W$ . This equation includes both linear and interaction terms, capturing treatment-covariate dependencies. In the primary study (where $S = 0$ ), the primary outcome $Y$ is modeled as: $Y = \alpha(X) \cdot \mu_W(X, T, S) + \gamma$ where $\alpha(X) = \rho_1 X_1 + \rho_0$ . This outcome depends on the treatment effect modulated by covariate-driven heterogeneity in $\alpha(X)$ , capturing treatment-mediated effects of covariates on $Y$ . + +Here the true ATE in primary is given as $\theta_0 = \mathbb{E}\left[(\rho_0 + \rho_1X_1)\cdot (\gamma_0 + \gamma_1X_1 + \gamma_2X_2)\mid S = 0\right]$ + +Analysis and Results. We use mean-squared error (MSE) to compare the performance of the following three estimators: (i) efficient estimator only using primary data $(\hat{\theta}_0)$ , (ii) efficient estimator augmented with auxiliary score $(\hat{\theta}_b)$ and (iii) efficient estimator with known $\alpha$ integrating auxiliary data $(\hat{\theta}_a)$ . The simulation results are compiled in Figure 2. As expected, the performance of all three estimators improves as $n$ increases and deteriorates as $p$ increases. Further, $\hat{\theta}_a$ dominates $\hat{\theta}_0$ and $\hat{\theta}_b$ especially for scenarios with large $p$ and/or large log $\left(\frac{P(S=1)}{P(S=0)}\right)$ – indicating that in scenarios where the primary study is relatively small and the problem is high-dimensional leveraging auxiliary data yields more benefits. This aligns with our theoretical results showing that knowing $\alpha$ can yield efficiency gains. For, $\hat{\theta}_b$ (which uses auxiliary data), we observe that it yields benefits relative to $\hat{\theta}_0$ in small $n$ scenarios especially when $p$ and $\log \left(\frac{P(S=1)}{P(S=0)}\right)$ are large. However, these benefits diminish relative to $\hat{\theta}_0$ as $n$ grows. This is consistent with our theoretical result showing that there are no asymptotic benefits if $\alpha$ is unknown. However, there are some finite sample benefits of using the auxiliary score even when $\alpha$ is unknown. + +# Appendix B Efficiency Score Functions Derivation (Theorems 1-4) + +Following the above mentioned procedure we derive the EIFs and corresponding efficiency bounds under the three sets of assumptions. As $\psi^{*}(O;\theta_{0},\eta_{0}) = \{\mathbb{E}[R^{*}(O;\theta_{0},\eta_{0})R^{*}(O;\theta_{0},\eta_{0})^{T}]\}^{-1}R^{*}(O;\theta_{0},\eta_{0})$ , and the semiparametrically efficient asymptotic variance (i.e., efficiency bound) is equal to $\{\mathbb{E}[R^{*}(O;\theta_{0},\eta_{0})R^{*}(O;\theta_{0},\eta_{0})^{T}]\}^{-1}$ , we only present the efficient score function $R^{*}$ instead of the EIF $\psi^{*}$ . However, note that deriving the EIF from $R^{*}$ is straightforward in our context. + +![](images/79018837d93f6bbd4601c2787b0063039a280292abc34a735ccf31f462f59850.jpg) + +![](images/f66a7fce804d88c91c78deddb2603f0f19f566744b48e026ebf2bd90789ffc76.jpg) + +![](images/a57cd8e5aca4ad131dc604fca91ba1fadb8c9b237181a4b2886465b43702cc68.jpg) + +![](images/19c78f30eaced6499b3320f94c5fe23eb75c734f4bb406652950eaa1559daae3.jpg) + +![](images/1b0d28a9ff5b649b224aadbbfd0fe433e39b53e80af5aeedea5f0289525c77fe.jpg) + +![](images/e946beb53171845a3c079d40008cb9c4e3bd1fa4a752f172efdc47d6c5a91f66.jpg) + +![](images/8823629658e0a68dbbf359e985f481bbcd8569fcf9b2654dfad165a284a42710.jpg) +Figure 2: Simulation Study Results. Mean squared error rates for three different estimators $\hat{\theta}_0$ , $\hat{\theta}_a$ and $\hat{\theta}_b$ based on $R_0^*$ , $R_a^*$ , and $R_b^*$ + +![](images/ccdc0d3cc5ac1eaf7274fe228f4832c06deaf24a8e9d309a7b78dc5614a6c0ee.jpg) + +![](images/37ee36ee2827a8b098f73f59763bfc533cb77a7381599de7e48c0b35885464ed.jpg) + +In our case, $O = (X,S,T,V)$ , $P(O = o;\theta ,\eta) = P(X = x)P(S = s\mid X = x)P(T = t\mid X = x,S = s)P(V = v\mid T = t,X = x,S = s)$ and $\mathcal{L}(O;\theta ,\eta) = \log P(O = o;\theta ,\eta)$ . Thus, + +$$ +\begin{array}{l} \mathcal {L} (O; \theta , \eta) = \log P (X = x) + \log P (S = s \mid X = x) + \log P (T = t \mid S = s, X = x) \\ + \log P (V = v \mid T = t, S = s, X = x) \\ = \log P (X = x) + \log \left(S \mu_ {S} (x) + (1 - S) \left(1 - \mu_ {S} (x)\right)\right) \\ + \log (T \mu_ {T} (x, s) + (1 - T) (1 - \mu_ {T} (x, s))) \\ + \log (P (V = v \mid T = t, S = s, X = x)), \\ \end{array} +$$ + +We know that $V = SW + (1 - S)Y$ and $Y = \alpha(X)W + \beta(X) + \varepsilon$ . + +Thus, + +$$ +P (V = v \mid T = t, S = s, X = x) = P (S (Y - \beta (X) - \varepsilon) + \alpha (X) (1 - S) Y = \alpha (X) v \mid T = t, S = s, X = x). +$$ + +. Simplifying it further, + +$$ +P (V = v \mid T = t, S = s, X = x) = P ((S + \alpha (X) (1 - S)) Y - S \varepsilon = \alpha (X) v + \beta (X) S \mid T = t, S = s, X = x). +$$ + +Substituting $Y$ with $\theta(X)T + g(X) + \gamma$ : + +$$ +\begin{array}{l} P (V = v \mid T = t, S = s, X = x) \\ = P \left(\left(S + \alpha (X) (1 - S)\right) (\theta (X) T + g (X) + \gamma) - S \varepsilon = \alpha (X) v + \beta (X) S \mid T = t, S = s, X = x\right) \\ = s P (\gamma - \varepsilon = \alpha (X) (v - \mu_ {W} (X, 1)) - \theta (X) (T - \mu_ {T} (X, 1)) \mid T = t, S = s, X = x) \\ + (1 - s) P (\gamma = (v - \mu_ {Y} (X, 0)) - \theta (X) (T - \mu_ {T} (X, 0)) \mid T = t, S = s, X = x) \\ = s P (\alpha (X) \delta = \alpha (X) (v - \mu_ {W} (X, 1)) - \theta (X) (T - \mu_ {T} (X, 1)) \mid T = t, S = s, X = x) \\ + (1 - s) P (\gamma = (v - \mu_ {Y} (X, 0)) - \theta (X) (T - \mu_ {T} (X, 0)) \mid T = t, S = s, X = x) \\ = s P (\delta = (v - \mu_ {W} (X, 1)) - \frac {\theta (X)}{\alpha (X)} (T - \mu_ {T} (X, 1)) \mid T = t, S = s, X = x) \\ + (1 - s) P (\gamma = (v - \mu_ {Y} (X, 0)) - \theta (X) (T - \mu_ {T} (X, 0)) \mid T = t, S = s, X = x). \\ \end{array} +$$ + +Assuming $\gamma$ and $\delta$ are normally distributed with mean 0 and homoskedastic variances $\sigma_{\gamma}^{2}$ and $\sigma_{\delta}^{2}$ respectively, + +$$ +\begin{array}{l} \log P (V = v \mid T = t, S = s, X = x) \\ = l o g \left( \begin{array}{c} s \exp \left(- \frac {((v - \mu_ {W} (x , 1)) - \frac {\theta (x)}{\alpha (x)} (t - \mu_ {T} (x , 1))) ^ {2}}{2 \sigma_ {\delta} ^ {2}}\right) \\ + (1 - s) \exp \left(- \frac {((v - \mu_ {Y} (x , 0)) - \theta (x) (t - \mu_ {T} (X , 0))) ^ {2}}{2 \sigma_ {\gamma} ^ {2}}\right) \end{array} \right) \\ \end{array} +$$ + +Efficient score function using only primary data (Result of Theorem 1). Now, we first show the efficient score function for the case that only uses primary data. This works as a baseline case for us and the subsequent efficiency bounds are compared with this case. The efficient score function under assumptions A.2. and A.4. is given as: + +$$ +R _ {0} ^ {*} (O; \theta_ {0}, \eta_ {0}) = (1 - S) \cdot \left(\left((V - \mu_ {Y} (X, 0)) - \theta (X) (T - \mu_ {T} (X, 0))\right) \cdot \frac {(T - \mu_ {T} (X , 0))}{\sigma_ {\gamma} ^ {2}}\right). +$$ + +Let $\Delta_0 = \left((V - \mu_Y(X,0)) - \theta (X)(T - \mu_T(X,0))\right)\cdot \frac{(T - \mu_T(X,0))}{\sigma_\gamma^2}$ then $\mathbb{E}\left[(R_0^* (O;\theta_0,\eta_0))(R_0^* (O;\theta_0,\eta_0))^T\right] = \mathbb{E}\left[(1 - S)^2\Delta_0^2\right]$ , and the asymptotic variance + +$$ +\mathbb {V} _ {0} ^ {\theta} (X) := \left(\mathbb {E} \left[ \left(R _ {0} ^ {*} (O; \theta_ {0}, \eta_ {0})\right) \left(R _ {0} ^ {*} (O; \theta_ {0}, \eta_ {0})\right) ^ {T} \right]\right) ^ {- 1} = \frac {1}{\mathbb {E} \left[ \Delta_ {0} ^ {2} \mid S = 0 , X \right] p (S = 0 \mid X)}. +$$ + +Efficient score function under assumptions A.1.-A.4. and A.5(a) (Result of Theorem 2). Under this assumption, only $\theta$ is unknown and would be estimated using the data while $\alpha$ and $\beta$ are known a priori. Thus, the efficient score function under A.5(a) is: + +$$ +\begin{array}{r l r} {R _ {a} ^ {*} (O; \theta_ {0}, \eta_ {0})} & = & {\left( \begin{array}{c} S \cdot \Big ((\alpha (X) (V - \mu_ {W} (X, 1)) - \theta (X) (T - \mu_ {T} (X, 1))) \cdot \frac {(T - \mu_ {T} (X , 1))}{\alpha^ {2} (X) \sigma_ {\delta} ^ {2}} \Big) \\ + (1 - S) \cdot \Big ((V - \mu_ {Y} (X, 0)) - \theta (X) (T - \mu_ {T} (X, 0))) \cdot \frac {(T - \mu_ {T} (X , 0))}{\sigma_ {\gamma} ^ {2}} \Big) \end{array} \right)} \end{array} +$$ + +Let $\Delta_1 = \left((\alpha(X)(V - \mu_W(X, 1)) - \theta(X)(T - \mu_T(X, 1))) \cdot \frac{(T - \mu_T(X, 1))}{\alpha^2(X)\sigma_\delta^2}\right)$ . Then, the asymptotic variance + +$$ +\mathbb {V} _ {a} ^ {\theta} (X) := \left(\mathbb {E} [ R _ {a} ^ {*} (O; \theta_ {0}, \eta_ {0}) (R _ {a} ^ {*} (O; \theta_ {0}, \eta_ {0})) ^ {T} | X ]\right) ^ {- 1} = \left( \begin{array}{c} \mathbb {E} \left[ \Delta_ {0} ^ {2} | S = 0, X \right] p (S = 0 | X) \\ + \mathbb {E} \left[ \Delta_ {1} ^ {2} | S = 1, X \right] p (S = 1 | X) \end{array} \right) ^ {- 1}. +$$ + +$\mathbb{V}_a^\theta (X)$ is always smaller than or equal to $\mathbb{V}_0^\theta (X)$ because $\mathbb{E}\left[\Delta_1^2\mid S = 1,X\right]p(S = 1\mid X)$ is non-negative. + +Efficient score function under assumptions A.1.-A.4. and A.5(b) (Result of Theorem 3). Here, along with $\theta$ , $\alpha$ is an unknown parameter. Thus, the efficient score function is given as: + +$$ +R _ {b} ^ {*} (O; \{\theta_ {0}, \alpha_ {0} \}, \eta_ {0}) = \left( \right.\left(\begin{array}{c}S \cdot \left((\alpha (X) (V - \mu_ {W} (X, 1)) - \theta (X) (T - \mu_ {T} (X, 1))) \cdot \frac {(T - \mu_ {T} (X , 1))}{\alpha^ {2} (X) \sigma_ {\delta} ^ {2}}\right)\\+ (1 - S) \cdot \left(((V - \mu_ {Y} (X, 0)) - \theta (X) (T - \mu_ {T} (X, 0))) \cdot \frac {(T - \mu_ {T} (X , 0))}{\sigma_ {\gamma} ^ {2}}\right)\\S \cdot \left(\frac {(\theta (X) (T - \mu_ {T} (X , 1)) - \alpha (X) (V - \mu_ {W} (X , 1)))}{\alpha^ {2} (X) \sigma_ {\delta} ^ {2}} \cdot \frac {\theta (X) (t - \mu_ {T} (X , 1))}{\alpha (X)}\right)\end{array}\right). +$$ + +$$ +\begin{array}{l} \mathbb {E} [ (R _ {b} ^ {*} (R _ {b} ^ {*}) ^ {T}) \mid X ] = \mathbb {E} \left[ \left( \begin{array}{c} S \Delta_ {1} + (1 - S) \Delta_ {0} \\ \frac {\theta (X)}{\alpha (X)} S \Delta_ {1} \end{array} \right) \left( \begin{array}{c c} S \Delta_ {1} + (1 - S) \Delta_ {0} & \frac {\theta (X)}{\alpha (X)} S \Delta_ {1} \end{array} \right) \mid X \right] \\ = \left( \begin{array}{cc}\mathbb{E}[\Delta_{1}^{2}\mid X,S = 1]P(S = 1\mid X) + \mathbb{E}[\Delta_{0}^{2}\mid X,S = 0]P(S = 0\mid X) & \frac{\theta(X)}{\alpha(X)}\mathbb{E}[\Delta_{1}^{2}\mid X,S = 1]P(S = 1\mid X)\\ \frac{\theta(X)}{\alpha(X)}\mathbb{E}[\Delta_{1}^{2}\mid X,S = 1]P(S = 1\mid X) & \frac{\theta^{2}(X)}{\alpha^{2}(X)}\mathbb{E}[\Delta_{1}^{2}\mid X,S = 1]P(S = 1\mid X) \end{array} \right) \\ \end{array} +$$ + +The asymptotic variance-covariance is then + +$$ +\begin{array}{l} \boldsymbol {\Sigma} _ {b} (X) := \left( \begin{array}{c c} \mathbb {V} _ {b} ^ {\theta} (X) & C o v _ {b} ^ {\theta , \alpha} (X) \\ C o v _ {b} ^ {\theta , \alpha} (X) & \mathbb {V} _ {b} ^ {\alpha} (X) \end{array} \right) := \left(\mathbb {E} [ (R _ {b} ^ {*} (R _ {b} ^ {*}) ^ {T}) | X ]\right) ^ {- 1} \\ = \left( \begin{array}{c c} \frac {1}{\mathbb {E} [ \Delta_ {0} ^ {2} | X , S = 0 ] P (S = 0 | X)} & - \frac {\alpha_ {0} (X)}{\theta_ {0} (X)} \frac {1}{\mathbb {E} [ \Delta_ {0} ^ {2} | X , S = 0 ] P (S = 0 | X)} \\ - \frac {\alpha_ {0} (X)}{\theta_ {0} (X)} \frac {1}{\mathbb {E} [ \Delta_ {0} ^ {2} | X , S = 0 ] P (S = 0 | X)} & \left(\frac {\alpha_ {0} (X)}{\theta_ {0} (X)}\right) ^ {2} \left(\frac {1}{\mathbb {E} [ \Delta_ {0} ^ {2} | X , S = 0 ] P (S = 0 | X)} + \frac {1}{\mathbb {E} [ \Delta_ {1} ^ {2} | X , S = 1 ] P (S = 1 | X)}\right) \end{array} \right) \\ \end{array} +$$ + +From this, we see that the asymptotic variance for the efficient estimator of $\theta$ is $\mathbb{V}_b^\theta (X) = \frac{1}{\mathbb{E}[\Delta_0^2|X,S = 0]P(S = 0|X)}$ . Note, that this asymptotic variance $\mathbb{V}_b^\theta (X) = \mathbb{V}_0^\theta (X)$ . This highlights that under assumption $A.5(b)$ there are no efficiency gains from leveraging auxiliary data compared to the baseline which only uses the primary study. + +Efficient score function under assumptions A.1.-A.4. and A.5(c) (Result of Theorem 4). As the likelihood is agnostic of $\beta$ , the efficient score function under A.5(c) is identical to that of A.5(b), i.e., + +$$ +R _ {c} ^ {*} (O; \{\theta_ {0}, \alpha_ {0} \}, \eta_ {0}) = R _ {b} ^ {*} (O; \{\theta_ {0}, \alpha_ {0} \}, \eta_ {0}). +$$ + +As the score functions are identical under assumptions A.5(b) and A.5(c), the asymptotic variance is also identical. This indicates that there are no efficiency gains from leveraging auxiliary data compared to the baseline that uses only the primary study. + +# Appendix C Proof of Theorem 5 (Misspecification Bias) + +Proof of Theorem 5. We begin by defining $\hat{\theta}_a$ as the estimator solving the empirical moment condition $\mathcal{P}_n R_a^*(O; \hat{\theta}_a, \hat{\eta}) = 0$ . In the population, $\theta_0$ solves $\mathbb{E}[R_a^*(O; \theta_0, \eta_0)] = 0$ only under the correct specification of $\alpha = \alpha^\star$ . We now investigate what happens when the analyst assumes $\alpha = \alpha_{\mathrm{mis}}$ where $\alpha_{\mathrm{mis}} \neq \alpha^\star$ . Recall, $\hat{\theta}_a$ is + +$$ +\frac {\sum_ {i} \left((1 - S _ {i}) \frac {\hat {r} _ {Y} (X _ {i} , 0) \hat {r} _ {T} (X _ {i} , 0)}{\hat {\sigma} _ {Y} ^ {2}} + S _ {i} \frac {\hat {r} _ {W} (X _ {i} , 1) \hat {r} _ {T} (X _ {i} , 1)}{\alpha (X _ {i}) \hat {\sigma} _ {W} ^ {2}}\right)}{\sum_ {i} \left((1 - S _ {i}) \frac {\hat {r} _ {T} ^ {2} (X _ {i} , 0)}{\hat {\sigma} _ {Y} ^ {2}} + S _ {i} \frac {\hat {r} _ {T} ^ {2} (X _ {i} , 1)}{\alpha^ {2} (X _ {i}) \hat {\sigma} _ {W} ^ {2}}\right)} +$$ + +Thus, $\mathbb{E}[\hat{\theta}_a(\alpha_{mis}) - \hat{\theta}_a(\alpha^\star)] = \mathbb{E}[\hat{\theta}_a(\alpha_{mis}) - \hat{\theta}_a(\alpha^\star) \mid S = 0]P(S = 0) + \mathbb{E}[\hat{\theta}_a(\alpha_{mis}) - \hat{\theta}_a(\alpha^\star) \mid S = 1]P(S = 1)$ . In the estimator, terms with $(1 - S)$ do not interact with $\alpha$ . Thus, $\mathbb{E}[\hat{\theta}_a(\alpha_{mis}) - \hat{\theta}_a(\alpha^\star) \mid S = 0]P(S = 0) = 0$ . Now, consider $\mathbb{E}[\hat{\theta}_a(\alpha_{mis}) - \hat{\theta}_a(\alpha^\star) \mid S = 1]P(S = 1)$ . + +$$ +\mathbb {E} \left[ \hat {\theta} _ {a} \left(\alpha_ {m i s}\right) - \hat {\theta} _ {a} \left(\alpha^ {\star}\right) \mid S = 1 \right] = \mathbb {E} \left[ \mathbb {E} \left[ \hat {\theta} _ {a} \left(\alpha_ {m i s}\right) - \hat {\theta} _ {a} \left(\alpha^ {\star}\right) \mid X, S = 1 \right] \mid S = 1 \right] +$$ + +$$ +\begin{array}{l} \mathbb {E} [ \hat {\theta} _ {a} (\alpha_ {m i s}) - \hat {\theta} _ {a} (\alpha^ {\star}) \mid X, S = 1 ] = \mathbb {E} \left[ \frac {(\alpha_ {m i s} (X) - \alpha^ {\star} (X)) \mathbb {E} [ \hat {r} _ {W} (X , 1) \hat {r} _ {T} (X , 1) ]}{(\mathbb {E} [ \hat {r} _ {T} ^ {2} (X , 1) ])} \mid X, S = 1 \right] \\ = \mathbb {E} \left[ \frac {\left(\alpha_ {m i s} (X) - \alpha^ {\star} (X)\right)}{\alpha^ {\star} (X)} \frac {\mathbb {E} \left[ \hat {r} _ {Y} (X , 1) \hat {r} _ {T} (X , 1) \right]}{\left(\mathbb {E} \left[ \hat {r} _ {T} ^ {2} (X , 1) \right]\right)} \mid X, S = 1 \right] \\ = \mathbb {E} \left[ \frac {\left(\alpha_ {m i s} (X) - \alpha^ {\star} (X)\right)}{\alpha^ {\star} (X)} \theta (X) \mid X, S = 1 \right] \\ \end{array} +$$ + +![](images/d558f04e9499f509a3513f9cf358342a4bfd54320c55cd90383f2e2c74d2b0ce.jpg) + +# Appendix D Proof of Theorem 6 (Error bound for $\hat{\mu}_Y$ ) + +Proof. We analyze the estimation error for both $\hat{\mu}_{Y,0}$ and $\hat{\mu}_{Y,b}$ under the given metric entropy assumptions. + +(i) One-stage estimator $\hat{\mu}_{Y,0}$ . By assumption A.6., $\mu_{Y} \in \mathcal{M}$ , and the metric entropy of $\mathcal{M}$ satisfies + +$$ +\log N (\varepsilon , \mathcal {M}, \| \cdot \|) \leq C \varepsilon^ {- \omega}. +$$ + +From standard results in empirical process theory and nonparametric regression (e.g., Györfi et al. (2006) and Tsybakov and Tsybakov (2009)), it follows that the least-squares estimator $\hat{\mu}_{Y,0}$ satisfies + +$$ +\left\| \hat {\mu} _ {Y, 0} - \mu_ {Y} \right\| = o _ {p} \left(n _ {0} ^ {- 1 / (2 + \omega)}\right). +$$ + +(ii) Two-stage estimator $\hat{\mu}_{Y,b}$ . By assumption A.5., $\mu_{Y}(X) = \alpha(X)\mu_{W}(X) + \beta(X)$ . We estimate $\hat{\mu}_{W}(X)$ from $n_1$ auxiliary samples. Let $\hat{\mu}_{W}$ be an estimator satisfying + +$$ +\left\| \hat {\mu} _ {W} - \mu_ {W} \right\| = o _ {p} \left(n _ {1} ^ {- 1 / (2 + \omega)}\right), +$$ + +under the assumption that $\mu_W\in \mathcal{M}$ and satisfies the same entropy bound as $\mu_Y$ + +The two-stage estimator is defined as: + +$$ +\hat {\mu} _ {Y, b} (X) = \hat {\alpha} (X) \cdot \hat {\mu} _ {W} (X, 0) + \hat {\beta} (X), +$$ + +where $(\hat{\alpha},\hat{\beta})$ minimize the squared error loss over the primary sample: + +$$ +(\hat {\alpha}, \hat {\beta}) = \arg \min _ {\alpha \in \mathcal {A}, \beta \in \mathcal {B}} \frac {1}{n _ {0}} \sum_ {i = 1} ^ {n _ {0}} \left(Y _ {i} - \alpha (X _ {i}) \hat {\mu} _ {W} (X _ {i}, 0) - \beta (X _ {i})\right) ^ {2}. +$$ + +We now decompose the error: + +$$ +\left\| \hat {\mu} _ {Y, b} - \mu_ {Y} \right\| = \left\| \hat {\alpha} \hat {\mu} _ {W} + \hat {\beta} - \alpha \mu_ {W} - \beta \right\|. +$$ + +Adding and subtracting intermediate terms: + +$$ += \| (\hat {\alpha} - \alpha) \hat {\mu} _ {W} + \alpha (\hat {\mu} _ {W} - \mu_ {W}) + (\hat {\beta} - \beta) \|. +$$ + +Applying triangle inequality: + +$$ +\| \hat {\mu} _ {Y, b} - \mu_ {Y} \| \leq \| (\hat {\alpha} - \alpha) \| \| \hat {\mu} _ {W} \| _ {\infty} + \| \alpha \| _ {\infty} \| (\hat {\mu} _ {W} - \mu_ {W}) \| + \| \hat {\beta} - \beta \|. +$$ + +Under the assumption that $\hat{\mu}_W$ is uniformly bounded (which holds if $\mu_W$ and $\hat{\mu}_W$ are bounded and consistent), and using the entropy conditions on $\mathcal{A}$ and $\mathcal{B}$ : + +$$ +\| \hat {\alpha} - \alpha \| = o _ {p} \left(n _ {0} ^ {- 1 / \left(2 + \omega_ {\alpha}\right)}\right) \quad (\text {g i v e n A . 6 .}), \text {a n d ,} \| \hat {\beta} - \beta \| = 0 \quad (\text {g i v e n A . 5 (b)}). +$$ + +Combining all the pieces, we obtain: + +$$ +\left\| \hat {\mu} _ {Y, b} - \mu_ {Y} \right\| = o _ {p} \left(n _ {0} ^ {- 1 / (2 + \omega)} \left(n _ {0} ^ {\frac {1}{2 + \omega} - \frac {1}{2 + \omega_ {\alpha}}} + \left(\frac {n _ {1}}{n _ {0}}\right) ^ {- 1 / (2 + \omega)}\right)\right), +$$ + +where the first term reflects the complexity reduction from modeling $\mu_{Y}$ via $\alpha (X)$ , and the second term reflects the error propagated from estimating $\mu_W$ using auxiliary data. \ No newline at end of file diff --git a/NeurIPS/2025/A Cautionary Tale on Integrating Studies with Disparate Outcome Measures for Causal Inference/images.zip b/NeurIPS/2025/A Cautionary Tale on Integrating Studies with Disparate Outcome Measures for Causal Inference/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..0b33e74a2c7def480c753417f7e80e299fee94c0 --- /dev/null +++ b/NeurIPS/2025/A Cautionary Tale on Integrating Studies with Disparate Outcome Measures for Causal Inference/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e2afa3f5e20ceac0dbd9fecb0f0898bc5d004091e15c1f6d3fa8e17c33cd72af +size 635086 diff --git a/NeurIPS/2025/A Cautionary Tale on Integrating Studies with Disparate Outcome Measures for Causal Inference/layout.json b/NeurIPS/2025/A Cautionary Tale on Integrating Studies with Disparate Outcome Measures for Causal Inference/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c1607c8a5780455073849d4523418a6db9dab694 --- /dev/null +++ b/NeurIPS/2025/A Cautionary Tale on Integrating Studies with Disparate Outcome Measures for Causal Inference/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ebb6ebd95923f42929ed903a11c820fdc75810a03de1a401b27805bfdf20f22d +size 781805 diff --git a/NeurIPS/2025/A Circular Argument_ Does RoPE need to be Equivariant for Vision_/38fc9860-1759-4a83-a66d-cd4b47bf2317_content_list.json b/NeurIPS/2025/A Circular Argument_ Does RoPE need to be Equivariant for Vision_/38fc9860-1759-4a83-a66d-cd4b47bf2317_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..07e2ad4c51ea8a6b30ae369cc4f2abb1d563838a --- /dev/null +++ b/NeurIPS/2025/A Circular Argument_ Does RoPE need to be Equivariant for Vision_/38fc9860-1759-4a83-a66d-cd4b47bf2317_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0ac8789d63bfa18d91a342643f07bf272b02dcf7826895bb63959efa54b91782 +size 231636 diff --git a/NeurIPS/2025/A Circular Argument_ Does RoPE need to be Equivariant for Vision_/38fc9860-1759-4a83-a66d-cd4b47bf2317_model.json b/NeurIPS/2025/A Circular Argument_ Does RoPE need to be Equivariant for Vision_/38fc9860-1759-4a83-a66d-cd4b47bf2317_model.json new file mode 100644 index 0000000000000000000000000000000000000000..044281fc1bcd0d855644d1589badc926dd250354 --- /dev/null +++ b/NeurIPS/2025/A Circular Argument_ Does RoPE need to be Equivariant for Vision_/38fc9860-1759-4a83-a66d-cd4b47bf2317_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0b518b1dcf92116390ce3fcbaa920ec6185feabb311bfe44848a22762269410f +size 289080 diff --git a/NeurIPS/2025/A Circular Argument_ Does RoPE need to be Equivariant for Vision_/38fc9860-1759-4a83-a66d-cd4b47bf2317_origin.pdf b/NeurIPS/2025/A Circular Argument_ Does RoPE need to be Equivariant for Vision_/38fc9860-1759-4a83-a66d-cd4b47bf2317_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..e2bd8c7f9e508f3edda9416234151b96f807c7b3 --- /dev/null +++ b/NeurIPS/2025/A Circular Argument_ Does RoPE need to be Equivariant for Vision_/38fc9860-1759-4a83-a66d-cd4b47bf2317_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cca3e17acfba40f1bfb23b7bd3507a9e21e5e9b09ff19151d840eb757ba19ca9 +size 1575519 diff --git a/NeurIPS/2025/A Circular Argument_ Does RoPE need to be Equivariant for Vision_/full.md b/NeurIPS/2025/A Circular Argument_ Does RoPE need to be Equivariant for Vision_/full.md new file mode 100644 index 0000000000000000000000000000000000000000..678b1016c9a79119cfceb89d32969cde60857b5a --- /dev/null +++ b/NeurIPS/2025/A Circular Argument_ Does RoPE need to be Equivariant for Vision_/full.md @@ -0,0 +1,1251 @@ +# A Circular Argument: Does RoPE need to be Equivariant for Vision? + +Chase van de Geijn $^{1,\dagger}$ , Timo Lüddecke $^{1}$ , Polina Turishcheva $^{1}$ Alexander S. Ecker $^{1,2,*}$ + +$^{1}$ Institute of Computer Science and Campus Institute Data Science, University of Göttingen + +$^{2}$ Max Planck Institute for Dynamics and Self-Organization, Göttingen, Germany + +†chase.geijn@uni-goettingen.de + +* ecker@cs.uni-goettingen.de + +# Abstract + +Rotary Positional Encodings (RoPE) have emerged as a highly effective technique for one-dimensional sequences in Natural Language Processing spurring recent progress towards generalizing RoPE to higher-dimensional data such as images and videos. The success of RoPE has been thought to be due to its positional equivariance, i.e. its status as a relative positional encoding. In this paper, we mathematically show RoPE to be one of the most general solutions for equivariant positional embedding in one-dimensional data. Moreover, we show Mixed RoPE to be the analogously general solution for $M$ -dimensional data, if we require commutative generators - a property necessary for RoPE's equivariance. However, we question whether strict equivariance plays a large role in RoPE's performance. We propose Spherical RoPE, a method analogous to Mixed RoPE, but assumes non-commutative generators. Empirically, we find Spherical RoPE to have the equivalent or better learning behavior compared to its equivariant analogues. This suggests that relative positional embeddings are not as important as is commonly believed, at least within computer vision. We expect this discovery to facilitate future work in positional encodings for vision that can be faster and generalize better by removing the preconception that they must be relative. + +# 1 Introduction + +Deep learning is in the age of transformers [73]. At their core, transformers are built on attention [4, 61], which is a permutation-invariant operation [74], making them agnostic to word or token position within a corpus. To break this symmetry, tokens must be modified with position embeddings [25, 83]. Recently, Rotational Positional Encodings (RoPE) [69] have gained popularity, touting an emphasis on the relative position between two tokens rather than their absolute positions [22, 26, 43, 46]. However, some of the original claims of RoPE have been called into question leading to confusion as to why it works: Su et al. [69] claimed the attention scores to decay with distance between tokens. This was found to be true only for attention with the same query and key [5]. Moreover, transformers with causal masking have been shown to require no positional encodings to be capable of recovering absolute position [24], making RoPE's relative (shift-equivariant) claim questionable. However, many new methods continue to be motivated by RoPE's benefit from shift-equivariance [27, 60, 85]. To guide future research in positional encodings, it is important to discover whether shift-equivariance truly makes RoPE successful and needs to be preserved when extending it. + +Both transformer and RoPE were originally designed for one-dimensional sequences such as language. RoPE encodes position by pairing dimensions within the query and key vectors within a transformer and rotating the paired dimensions. Transformers have become the current staple across all AI fields [17, 29, 46, 47, 63]. Naturally, RoPE's recent popularity in NLP has also spread to Vision + +Transformers (ViT), where the data is two- or three-dimensional, corresponding to images and videos. How to extend RoPE to other modalities is nontrivial and assumptions must be made to maintain equivariance [44, 60]. The most commonly used approach for extending RoPE to ViTs is through Axial RoPE [15, 26, 77], partitioning the embedding dimensions into dimensions rotated independently either by the horizontal or the vertical position of the tokens. However, this approach does not allow for diagonal attention patterns where horizontal and vertical information "mix", which have been hypothesized to enhance generalization; consequently, learned Mixed RoPE was proposed [26]. Even more recently, LieRE [52] generalized Mixed RoPE from pair-wise rotations to higher dimensional rotations using learned skew-symmetric Lie algebras. If one defines rotations to be special orthogonal transformation, LieRE is the most general form of rotation encoding. However, while general, LieRE does not guarantee equivariance. + +In this paper, we investigate the relationships across these different forms of positional encoding. In Section 3, we mathematically show RoPE with parameterized rotation speeds to be equivalent to LieRE for one-dimensional data. When the number of positional dimensions is higher dimensional, LieRE is not guaranteed to be equivariant unless constraints are placed on the Lie algebras. Using this insight, we derive Axial RoPE by imposing a "mutual exclusivity" constraint on the eigenvalues of LieRE's generators. Further, we will show that if one loosens this constraint – requiring the Lie algebra to be commutative between the generator – then one arrives at Mixed RoPE. To be a relative encoding, this commutativity property is necessary [44, 60], thus making Mixed RoPE the most general form of LieRE which maintains equivariance. However, it has been noted that requiring the positional embedding to be relative is an inductive bias whose necessity to RoPE's success is unclear [1, 5, 24]. + +The perceived necessity of equivariance has led to a circular argument where positional embeddings are assumed to perform well because they are relative, and all new embeddings must be relative because relative embeddings perform well. To break this cycle, we believe that it is imperative to establish the importance of equivariance embeddings for multi-dimensional RoPE. In Section 4, we propose alternative methods to establish a cause-effect experiment to evaluate whether equivariances is a predominant contributor to RoPE's faster training dynamics and generalization. To this end, we propose Spherical RoPE which takes a non-commutative assumption, thus breaking equivariance, and Uniform RoPE, which maintains equivariance, but has only a single shared rotation speed. + +In Section 5, we find that Spherical RoPE has the same training behaviors as its equivariant analogues and we find that Uniform RoPE outperforms the standard learned encodings, while performing worse than other RoPE methods. We conclude that our evidence suggests that the performance of RoPE over traditional embeddings is not explained by equivariance. + +# 2 Background + +In this section, we review concepts and notation from previous work on rotary positional embeddings. We introduce the methods in both historical and progressively general order which we will use to prove in Section 3 that Mixed RoPE is the most general $M$ -D rotary embedding with equivariance. For a broader literature review on positional embeddings see Appendix C. For a compact overview of symbols, see Appendix D. + +# 2.1 Attention + +We use the standard attention mechanism from Vaswani et al. [73], given by + +$$ +\mathbf {Z} = \operatorname {A t t e n t i o n} (\mathbf {Q}, \mathbf {K}, \mathbf {V}) = \operatorname {s o f t m a x} \left(\frac {\mathbf {Q} \mathbf {K} ^ {\top}}{\sqrt {d _ {k}}}\right) \mathbf {V}. \tag {1} +$$ + +We consider only single-headed attention to simplify notation, so here $\mathbf{Q}$ , $\mathbf{K}$ , and $\mathbf{V}$ are elements of $\mathbb{R}^{T\times N}$ , where $T$ is the number of tokens and $N$ is the network's latent dimension. We will primarily use index notation, where the above equation is expressed as, $\mathbf{z}_i = \sum_{j=1}^{\top} a(\mathbf{q}_i, \mathbf{k}_j) \mathbf{v}_j$ . We define the attention mechanism, $a(\mathbf{q}_i, \mathbf{k}_j)$ , as + +$$ +a \left(\mathbf {q} _ {i}, \mathbf {k} _ {j}\right) = \frac {e ^ {\alpha \left(\mathbf {q} _ {i} , \mathbf {k} _ {j}\right)}}{\sum_ {j = 1} ^ {T} e ^ {\alpha \left(\mathbf {q} _ {i} , \mathbf {k} _ {j} ^ {\prime}\right)}}, \tag {2} +$$ + +where what we refer to as the attention score is given by + +$$ +\alpha (\mathbf {q}, \mathbf {k}) = \mathbf {q} ^ {\top} \mathbf {k}. \tag {3} +$$ + +This formulation of attention is equivariant to permutations of the token order. To break this symmetry, the position of the tokens must be "coded" into the attention scores. Thus, we re-express the attention score as a function of the content of the query token $\mathbf{x}_i \in \mathbb{R}^N$ and key token $\mathbf{x}_j \in \mathbb{R}^N$ , and their positions $p_i, p_j \in \mathbb{R}$ , + +$$ +\alpha_ {i j} := \alpha \left(\mathbf {q} _ {i}, \mathbf {k} _ {j}\right) := \alpha \left(\left(\mathbf {x} _ {i}, p _ {i}\right), \left(\mathbf {x} _ {j}, p _ {j}\right)\right) := \alpha \left(\mathbf {x} _ {i}, \mathbf {x} _ {j}, p _ {i}, p _ {j}\right). \tag {4} +$$ + +Throughout this paper, we will abuse the notation of $\alpha$ and use these expressions interchangeably for ease of notation. If the position affects the query and key directly, as in RoPE, we will introduce the notation $\alpha (\varphi (x_i,p_i),\varphi (x_j,p_j))$ for positional encoding function $\varphi$ . + +# 2.2 Absolute and Relative Positional Encoding + +Absolute Positional Encoding (APE) is a common way of embedding token positions in transformers by adding position-dependent vectors, i.e. $\varphi (\mathbf{x},p)\coloneqq \mathbf{x} + \mathrm{PE}(p)$ , where $\mathbf{x}$ is a token embedding, $p$ is its position, and $\mathrm{PE}:\mathbb{Z}\to \mathbb{R}^N$ . Previous work has suggested learning a per-position token as PE [17, 20]. However, this restricts the network to fixed context length, removing the ability to extrapolate to different sequence lengths. The alternative is to add a deterministic function to the embedding. Vaswani et al. [73] proposed to add Fourier modes, + +$$ +P E _ {n} (p) = \left\{ \begin{array}{l l} \sin \left(p \omega_ {\frac {n}{2}}\right), & \text {i f n m o d 2 = 0} \\ \cos \left(p \omega_ {\lfloor \frac {n}{2} \rfloor}\right), & \text {i f n m o d 2 = 1}, \end{array} \right. \tag {5} +$$ + +where $n$ is a dimension within the positional embedding vector and $\omega_{n}$ is a frequency term which increases with dimension. Note that this pairs elements in the embedding vector with each pair being transformed by the same frequency. + +For ease in future notation, we will use $D \coloneqq N / 2$ as the number of pairs and interpret the embedded token as a $D \times 2$ tensor. One can also interpret this tensor as representing the coefficients of a complex number, the first representing the real and second representing the complex part. Then we can succinctly write this form of positional encoding as + +$$ +\varphi (\bar {\mathbf {x}}, p) = \bar {\mathbf {x}} + e ^ {i \omega p}, \tag {6} +$$ + +where we use $\bar{\cdot}$ to indicate complex-valued vectors, $\bar{\mathbf{x}}\in \mathbb{C}^{D}$ . For this notation, we should also adjust the attention score for complex numbers, + +$$ +\alpha (\bar {\mathbf {q}}, \bar {\mathbf {k}}) = \operatorname {R e} \left[ \bar {\mathbf {q}} ^ {\top} \bar {\mathbf {k}} \right], \tag {7} +$$ + +where $\bar{\mathbf{q}} = \bar{\mathbf{W}}_q\varphi (\bar{\mathbf{x}},p)$ and $\bar{\mathbf{k}} = \bar{\mathbf{W}}_k\varphi (\bar{\mathbf{x}},p)$ , with $\bar{\mathbf{W}}_q,\bar{\mathbf{W}}_k\in \mathbb{C}^{D\times D}$ , and $\top$ is assumed to be the Hermitian transpose. With Eq. 7 implied, we will continue with the notation in Eq. 3. + +Relative Positional Encodings Positional embeddings rely on being able to assign position values to each token. However, how one assigns positions can often be arbitrary. One could just as correctly assign the first token the value zero and consider natural numbers, or assign the middle token of a corpus zero and consider integers. We can relax the assumption of a canonical way of labeling positions in APE by relying on relative distances between tokens, resulting in $\alpha_{ij} = \alpha (\mathbf{x}_i,\mathbf{x}_j,p_i - p_j)$ . This is called relative positional encoding. We refer to this property as embeddings having a relative positional bias, or equivalently, having shift-equivariance (see Appendix E.1 for discussion of the equivalence). In this manuscript, we will simply use the term equivariance with the implication that the attention score is invariant to shifts in the query and key. + +# 2.3 Rotary Positional Encodings (RoPE) + +There are four common properties that are often preferred for positional embeddings: equivariance, key-query separability, linearity, and locality. For further details and why one may want these properties see Appendix E. + +From the properties, Rotary Positional Embeddings (RoPE) were derived by Su et al. [69]. Rather than adding a positional embedding to the patch embedding, RoPE proposed to modify the queries + +and keys by rotating them in pairs. By interpreting queries and keys as complex vectors, we can express this rotation as + +$$ +\varphi \left(\bar {q} _ {d}, p\right) = e ^ {i \omega_ {d} p} \bar {q} _ {d} \quad \varphi \left(\bar {k} _ {d}, p\right) = e ^ {i \omega_ {d} p} \bar {k} _ {d}. \tag {8} +$$ + +Since we assume the same operation is applied to the queries and keys, from now on, we will use $\mathbf{z}$ to refer to operations which act on both. In matrix form, this is given by $e^{i\omega_{d}p}$ can be represented as + +$$ +e ^ {i \omega_ {d} p} \equiv \left[ \begin{array}{c c} \cos (\omega_ {d} p) & - \sin (\omega_ {d} p) \\ \sin (\omega_ {d} p) & \cos (\omega_ {d} p) \end{array} \right] = \mathbf {R} _ {\omega_ {d} p}, \tag {9} +$$ + +where $\mathbf{R}_{\omega_q p_t}$ is a rotation matrix. While the rotation matrix is more intuitive, the complex exponential form will be useful for the mathematics in Section 3, so we will alternate between the two. Recall, we use the convention that real valued queries and keys will have dimension $N$ and the complex interpretation will have dimension $D$ . + +One can represent the effect of RoPE as the application of a block diagonal of rotation matrices, + +$$ +R o P E (\mathbf {z}, p) = \mathbf {R} _ {p} \mathbf {z} = \left[ \begin{array}{c c c} \mathbf {R} _ {p \omega_ {1}} & \mathbf {0} & \dots \\ \mathbf {0} & \mathbf {R} _ {p \omega_ {2}} & \dots \\ \vdots & \vdots & \ddots \end{array} \right] \left( \begin{array}{c} \mathbf {z} _ {1} \\ \vdots \\ \mathbf {z} _ {D} \end{array} \right) \equiv \left[ \begin{array}{c c c} e ^ {i p \omega_ {1}} & 0 & \dots \\ 0 & e ^ {i p \omega_ {2}} & \dots \\ \vdots & \vdots & \ddots \end{array} \right] \bar {\mathbf {z}}, \tag {10} +$$ + +where $\mathbf{z}_d$ is a query pair. We introduce this block-diagonal form as it was the notation used in Su et al. [69]. However, we will primarily stick to the index notation in Eq. 8. + +# 2.4 2D RoPE Embeddings + +RoPE is constrained to operate on (1D) sequences. Motivated by the success of RoPE in language modeling, there have been growing efforts to extend it to multi-dimensional positions [10, 15], which we outline below. We will use $M$ to refer to the dimensionality of the position, but will primarily focus on images, where $M = 2$ and $p_i, p_j \in \mathbb{R}^2$ . + +Trivial 2D RoPE. One could trivially encode $\mathbf{p} = (p_x, p_y)$ using rotation matrices $\mathbf{R}_{\omega_d p_x}$ , $\mathbf{R}_{\omega_d p_y}$ : + +$$ +\varphi (\mathbf {z} _ {d}, \mathbf {p}) = \mathbf {R} _ {\omega_ {d} p _ {x}} \mathbf {R} _ {\omega_ {d} p _ {y}} \mathbf {q} _ {d} = \mathbf {R} _ {\omega_ {d} \left(p _ {x} + p _ {y}\right)} \mathbf {z} _ {d}. \tag {11} +$$ + +However, in this case all positions with $p_x + p_y = c$ would get the same positional encoding $\mathbf{R}_{\omega_d c}$ . + +Axial RoPE. More practically, RoPE is extended to multiple dimensions by letting $x$ and $y$ act on different dimensions, + +$$ +\varphi \left(\mathbf {z} _ {d}, \mathbf {p}\right) = \left[ \begin{array}{c c} \mathbf {R} _ {p _ {x} \omega_ {d}} & \mathbf {0} \\ \mathbf {0} & \mathbf {R} _ {p _ {y} \omega_ {d}} \end{array} \right] \mathbf {z} _ {d}, \tag {12} +$$ + +where queries and keys are now split into four-dimensional vectors, $\mathbf{z}_d^\top = \left[z_1^{(x)},z_2^{(x)},z_1^{(y)},z_2^{(y)}\right]$ . The block-diagonal matrix can once again be viewed as a tensor of shape $N / 2M\times M\times 2$ , where $M$ is once again the dimensionality of position - in this case $M = 2$ for horizontal and vertical position. This gives the index notation + +$$ +\varphi \left(\mathbf {z} _ {d, m}, p _ {m}\right) = \mathbf {R} _ {\omega_ {d, m} p _ {m}} \mathbf {z} _ {d, m}, \tag {13} +$$ + +for $m \in \{x, y\}$ . From a programming perspective, one can interpret this as a form of batched matrix multiplication. + +While this method eliminates the symmetry, it treats $x$ and $y$ as independent. The result is a separable attention score of the form + +$$ +\alpha_ {i, j} = \alpha_ {i j} ^ {(x)} + \alpha_ {i j} ^ {(y)}, \tag {14} +$$ + +where $\alpha_{ij}^{(x)}$ and $\alpha_{ij}^{(y)}$ are components of the attention score which depend only on $p_x$ and $p_y$ , respectively. The frequencies are restricted to the axes, hence it is called Axial RoPE. This overemphasizes horizontal and vertical relationships at the expense of oblique directions creating gridded patterns shown in Figure 1. To represent oblique patterns, the rotations would have to be performed along directions that contain both an $x$ and a $y$ component, i.e. frequencies that are not aligned on the axis in Figure1. These frequencies have been referred to as "mixed frequencies" [26]. + +![](images/b02e864660ab4951e5f912e0a84921405d96f4ab71b658f16a45cc8ff827adc8.jpg) +Figure 1: The attention patterns of Axial and Mixed RoPE. A. Each dimension pair in the query and key vectors is rotated based on the position creating an attention pattern. The pixel value of the attention pattern is $\alpha (\mathbf{x}_q,\mathbf{x}_k,\mathbf{p},\mathbf{0})$ , where $\mathbf{p} = (i,j) -$ the pixel location. On the left, the attention pattern of individual component-pairs in the embedding vector is shown and, on the right, the components are combined into the overall attention pattern for a randomly sampled query and key vector. B. Location of the rotations frequencies in 2D frequency space. Axial RoPE can only represent frequencies that lie on an axis resulting in the grid-like attention patterns. Unlike Axial RoPE, Mixed RoPE can assign different directions to each component-pair (A Bottom). When Axial RoPE uses fixed frequencies, the frequencies are spread exponentially. However, they can be implemented as learnable parameters. For Uniform RoPE, all frequencies are fixed to a single value for each axis. + +![](images/3e7a8f5895b602c7adda916d46de6c5bc5173e4ac355f899e08ded5fe388f43a.jpg) + +Mixed RoPE: Learned mixed frequencies. The inclusion of mixed frequencies has empirically been shown to positively impact learning and generalization [26]. The naive approach in Eq. 11 is only a problem when $x$ and $y$ rotate by the same frequency in every dimension. One could instead parameterize the frequencies with two separate frequencies in each dimension, + +$$ +\varphi (\mathbf {z} _ {d}) = \mathbf {R} _ {\omega_ {d x} x} \mathbf {R} _ {\omega_ {d y} y} \mathbf {q} _ {d} = \mathbf {R} _ {\omega_ {d x} x + \omega_ {d y} y} \mathbf {q} _ {d}. \tag {15} +$$ + +By making the $\omega_{x}$ and $\omega_{y}$ parameters learnable, the attention pattern can learn mixed-frequency patterns by constructing a superposition of different diagonal patterns, as shown in Figure 1. + +LieRE. Recently, RoPE has been interpreted through the lens of Lie algebras [44, 52, 60]. For an intuitive introduction to how Lie algebras appear, see Appendix E.3. Lie Rotary Position Encodings (LieRE) [52] extend Mixed RoPE by applying $N$ -dimensional rotation matrices, rather than $2 \times 2$ matrices applied to pairs, using a linear combination of learned skew-symmetric Lie algebras, + +$$ +\varphi (\mathbf {z}, p) = \exp \left(\mathcal {A} _ {x} p _ {x} + \mathcal {A} _ {y} p _ {y}\right) \mathbf {z}, \tag {16} +$$ + +where $\exp$ is the matrix exponential, and the $\mathcal{A}$ terms are $N\times N$ skew-symmetric matrices - which are Lie group generators of a subgroup of $SO(N)$ . Mathematically, LieRE is the most general rotary-based embedding method as skew-symmetric matrices are the generators of any $N$ -D rotation. However, unlike the other two methods, LieRE is not guaranteed to be equivariant. + +# 3 The Generality of Learned RoPE and Mixed RoPE + +While LieRE is motivated as generalizing RoPE to $M$ -D rotations, in this section we will show that LieRE in one dimension can be learned by implementing RoPE with parameterized frequencies. For $M$ -D positions, LieRE is not equivariant unless the generators commute. If the generators are required to commute, we show that LieRE can be re-expressed as Mixed RoPE. Thus, we conclude Mixed RoPE to be a general solution for $M$ -D equivariant rotary embeddings. In this section, we will give informal proofs focused on high-level insights. + +# 3.1 1D-LieRE is equivalent to 1D RoPE with learned frequencies + +In this section, we prove that any one-dimensional LieRE can be expressed as RoPE with parameterized rotation frequencies. Thus, we conclude RoPE to be a computationally efficient way of expressing a $D$ -dimensional rotation, i.e., 1D-LieRE. + +To see why Proposition 1 holds, suppose we have a 1D-LieRE embedding with a learned generator $\mathcal{A}$ . By formulation, $\mathcal{A}$ is skew-symmetric, $\mathcal{A}^{\top} = -\mathcal{A}$ . The positionally encoded attention between query $\mathbf{q} = \mathbf{W}_{\mathbf{q}}\mathbf{x}$ and $\mathbf{k} = \mathbf{W}_{\mathbf{k}}\mathbf{x}$ is + +$$ +\alpha \left(x _ {i}, x _ {j}, p _ {i}, p _ {j}\right) = \left(\exp \left(\mathcal {A} p _ {q}\right) \mathbf {q}\right) ^ {\top} \exp \left(\mathcal {A} p _ {k}\right) \mathbf {k}. \tag {17} +$$ + +Any skew-symmetric matrix has an eigenvalue decomposition $\mathcal{A} = \mathbf{U}\Lambda_{\mathcal{I}}\mathbf{U}^{\top}$ where $\Lambda_{\mathcal{I}}$ is a diagonal matrix of purely imaginary (or zero) eigenvalues and $\mathbf{U}$ is a unitary matrix, $\mathbf{U}^{\top}\mathbf{U} = \mathbb{I}$ . Moreover, the matrix-exponential of an eigenvalue decomposition simplifies to $\exp (\mathbf{U}\Lambda_{\mathcal{I}}\mathbf{U}^{\top}) = \mathbf{U}\exp (\Lambda_{\mathcal{I}})\mathbf{U}^{\top}$ . This allows us to express attention as + +$$ +\begin{array}{l} \alpha \left(x _ {i}, x _ {j}, p _ {i}, p _ {j}\right) = \mathbf {q} ^ {\top} \mathbf {U} \exp \left(- p _ {q} \boldsymbol {\Lambda} _ {\mathcal {I}}\right) \mathbf {U} ^ {\top} \mathbf {U} \exp \left(p _ {k} \boldsymbol {\Lambda} _ {\mathcal {I}}\right) \mathbf {U} ^ {\top} \mathbf {k} (18) \\ = \mathbf {q} ^ {\prime \top} \exp \left(- p _ {q} \boldsymbol {\Lambda} _ {\mathcal {I}}\right) \exp \left(p _ {k} \boldsymbol {\Lambda} _ {\mathcal {I}}\right) \mathbf {k} ^ {\prime}, (19) \\ \end{array} +$$ + +where $\mathbf{q}' = \mathbf{W}_q'\mathbf{x}$ with $\mathbf{W}_q' = \mathbf{U}\mathbf{W}_q$ , and $\mathbf{k}' = \mathbf{W}_k'\mathbf{x}$ with $\mathbf{W}_k' = \mathbf{U}^\top \mathbf{W}_k$ . Because the eigenvalue matrix is diagonal, the exponential is given by + +$$ +\exp (p \boldsymbol {\Lambda} _ {\mathcal {I}}) = \left[ \begin{array}{c c c c} e ^ {i \lambda_ {0} p} & 0 & \dots & 0 \\ 0 & e ^ {i \lambda_ {1} p} & \ddots & \vdots \\ \vdots & \ddots & \ddots & 0 \\ 0 & \dots & 0 & e ^ {i \lambda_ {N - 1} p} \end{array} \right]. \tag {20} +$$ + +Notice that this is the same as the complex formulation of RoPE defined in Eq. 10, where the eigenvalues of the generator correspond to the rotation frequencies of the rotation matrices. Thus, any 1D-LieRE can be expressed as RoPE with learnable frequencies by absorbing the matrix of eigenvectors of $\mathcal{A}$ into the weight matrices $\mathbf{W}_{\mathbf{q}}$ and $\mathbf{W}_{\mathbf{k}}$ . Since 1D-LieRE learns a rotation in $SO(D)$ , RoPE can be seen as an efficient way to represent a rotation in $\mathbb{R}^D$ . + +# 3.2 Extending RoPE to more than one dimension + +While this proof works for 1D positions, it does not generalize to $M$ -D without introducing extra inductive biases or giving up equivariance. By imposing constraints on $\mathcal{A}_x$ and $\mathcal{A}_y$ , we can categorize the other RoPE methods based on the assumptions made. + +Generators rotate independent subspaces. For example, one can impose the assumption that $p_x$ and $p_y$ rotate independent subspaces in $\mathbb{R}^N$ . Mathematically, this assumption would imply that + +$$ +\forall d \in [ 1, D ]: \lambda_ {d} ^ {(x)} = 0 \text {o r} \lambda_ {d} ^ {(y)} = 0, \tag {21} +$$ + +where $\lambda_d^{(x)}$ and $\lambda_d^{(y)}$ are the eigenvalues of $\mathcal{A}_x$ and $\mathcal{A}_y$ , respectively. This is equivalent to rotating independent components of the query/key as done by Axial-RoPE. + +Commutative generators. For LieRE to be equivariant, we only need to ensure that the generators commute. If we make this assumption, then we arrive at Mixed RoPE. + +Proposition 2. Any $M$ -dimensional LieRE with commutative generators can be parameterized by Mixed RoPE. + +To see why Proposition 2 holds, suppose we can diagonalize $\mathcal{A}_x = \mathbf{U}_x\pmb{\Lambda}_x\mathbf{U}_x^\top$ and $\mathcal{A}_y = \mathbf{U}_y\pmb{\Lambda}_y\mathbf{U}_y^\top$ . If we take the assumption that $\mathcal{A}_x$ and $\mathcal{A}_y$ commute, + +$$ +\mathcal {A} _ {x} \mathcal {A} _ {y} = \mathcal {A} _ {y} \mathcal {A} _ {x} \quad \Longrightarrow \quad [ \mathcal {A} _ {x}, \mathcal {A} _ {y} ] = [ \mathcal {A} _ {y}, \mathcal {A} _ {x} ] = 0. \tag {22} +$$ + +$\left[\mathcal{A}_x,\mathcal{A}_y\right] = \mathcal{A}_x\mathcal{A}_y - \mathcal{A}_y\mathcal{A}_x$ is the Lie bracket. This implies that $\mathcal{A}_x$ and $\mathcal{A}_y$ are simultaneously diagonalizable (Lemma 2). Thus, commutativity implies that $\mathbf{U}_x = \mathbf{U}_y\coloneqq \mathbf{U}$ . We can write + +$$ +\mathbf {A} = \exp \left(\mathcal {A} _ {x} p _ {x} + \mathcal {A} _ {y} p _ {y}\right) = \mathbf {U} \exp \left(\boldsymbol {\Lambda} _ {x} p _ {x} + \boldsymbol {\Lambda} _ {y} p _ {y}\right) \mathbf {U} ^ {\top}, \tag {23} +$$ + +![](images/cdb68ee00c24f0249a744b9a4a9912c04066600082dc9fa545f1b1ba4e3ed4f7.jpg) +RoPE + +![](images/bbdffea642a7eda5866323f9234b28a79ab6a4f850c4b681940b06c254eac7d2.jpg) +Axial RoPE + +![](images/6c0463594af51c58368cc260b03d6781365f20805d4460e0d6b78f26e3668268.jpg) +Mixed RoPE + +![](images/5185707b5a2c13f77e51c2e2f8927f91f00c068933ce33fef3d3911aee173ac6.jpg) +Figure 2: Diagram of each rotary embedding's effect on the subvector, $\mathbf{z}_d$ . While Mixed RoPE affects 2D vector pairs, Spherical RoPE affects 3D vector triplets. Axial RoPE rotates independent dimensions for $p_x$ , thus containing pairs of pairs, or effectively quadruples. Each $\mathbf{z}$ contains $D$ sub-vectors rotating at different frequencies. While the order in which the rotations are applied does not matter for Axial or Mixed RoPE, order matters for Spherical RoPE. Explicitly, the triplet is first rotated around the axis associated with $p_x$ and then rotated around the axis associated with $p_y$ . + +![](images/24fbcba846027b0ed5373b1100cca576df62a98d66fa5223bbe9d9a9b81174b3.jpg) +Spherical RoPE + +![](images/cec36111aa91add83d8e18f51ff44dd5b5fdb6e6aeaa3eff76070ad6f0aa04db.jpg) + +which leads to Eq. 15. Thus, Mixed RoPE forms a general solution for assumptions of commutativity, which is necessary for LieRE to be relative. This also mathematically shows Mixed RoPE to be the strict generalization of Axial RoPE. + +In summary, learned frequency RoPE embeddings represent an efficient way of learning a much more general set of $SO(D)$ than is commonly believed. However, in order to generalize to higher dimensions while retaining their status as relative positional encodings, assumptions must be made. Mixed RoPE generalizes this to $M$ -D positions, spanning the entire solution class for relative LieRE. Thus, LieRE-like methods with commutative $\mathcal{A}_x$ and $\mathcal{A}_y$ – such as in Schenck et al. [60] – are not more expressive than Mixed RoPE and any empirical differences in performance must be attributed to the learning dynamics due to different parameterizations. However, it remains unclear whether equivariance is the real reason for RoPE's success. + +# 4 Experiments + +When extending RoPE to more than one dimension, we must either constrain ourselves to commuting Lie algebras or give up relativity. We therefore ask the question: Why does RoPE work? Which properties should be preserved for generalizing RoPE to vision? To explore this question, we propose two new RoPE variants: Spherical RoPE, which takes a non-commutative assumption, and Uniform-Frequency RoPE, which uses a single fixed rotation frequency across all dimensions. Below we provide a high level outline the different embeddings. We compare the existing positional embedding methods APE [17], Axial RoPE [15], Mixed RoPE [10], and LieRE [52] to these two new variants to understand whether equivariance, oblique directions or a variety of spatial frequencies are important features of PEs for vision. + +Spherical RoPE. We propose Spherical RoPE as a method between Mixed RoPE and LieRE that minimally changes 2D RoPE to break equivariance. Spherical RoPE embeds position as + +$$ +\varphi \left(\mathbf {z} _ {d}, \mathbf {p}\right) = \mathcal {Y} _ {\omega_ {d x} x} \mathcal {R} _ {\omega_ {d y} y} \mathbf {q} _ {d}, \tag {24} +$$ + +where $\mathbf{q}_d\in \mathbb{R}^3$ is now a triplet instead of a pair, and $\mathcal{V}$ is a block diagonal of $3\times 3$ yaw matrices and $\mathcal{R}$ is a block diagonal of roll matrices. + +$$ +\mathcal {Y} _ {\omega_ {d x} x} = \left[ \begin{array}{c c c} \cos (\omega_ {d x} x) & - \sin (\omega_ {d x} x) & 0 \\ \sin (\omega_ {d x} x) & \cos (\omega_ {d x} x) & 0 \\ 0 & 0 & 1 \end{array} \right] \quad \mathcal {R} _ {\omega_ {d y} y} = \left[ \begin{array}{c c c} 1 & 0 & 0 \\ 0 & \cos (\omega_ {d y} y) & - \sin (\omega_ {d y} y) \\ 0 & \sin (\omega_ {d y} y) & \cos (\omega_ {d y} y) \end{array} \right]. \tag {25} +$$ + +Intuitively, rather than RoPE rotating around a circle, Spherical RoPE rotates around a sphere using Euler angles. + +Importantly, spherical rotations like LieRE are non-commutative making them not equivariant. In fact, their generators are strictly noncommutative, $\mathcal{A}_x\mathcal{A}_y \neq \mathcal{A}_y\mathcal{A}_x$ . While non-commutativity does not mean Spherical RoPE is incapable of learning or approximating equivariance throughout the network, it is the component of LieRE removed by Mixed RoPE and works which enforce commutativity such as Yu et al. [85] and Schenck et al. + +Table 1: Table listing the properties of each of the rotary-based methods. + +
Positional EncodingVisionStrictly +EquivariantOblique +DirectionsRequires +Learning
Rotary (RoPE) [69]X✓N/AX
Axial RoPE [68]✓✓XX
Mixed RoPE [26]✓✓✓✓
LieRE [52]✓X✓✓
Spherical RoPE✓X✓X
Uniform RoPE✓✓XX
+ +We hypothesized Spherical RoPE to have a number of advantages. While Axial RoPE is unable to express oblique directions, Spherical RoPE can. Like Axial RoPE, Spherical RoPE can use fixed frequencies making it computationally cheaper than LieRE and Mixed RoPE since sines and cosines of the frequencies can be precomputed. However, our main interest is that Spherical RoPE is comparable in terms of expressivity to Mixed and Axial RoPE while being non-equivariant. + +Uniform-Frequency RoPE. For an initial evaluation on the impact of relative position, we propose Uniform-Frequency RoPE. For this method, we perform Axial RoPE with a single frequency shared across all rotation matrices. While still being relative, this serves as a more restricted version of RoPE. If this method performs significantly worse than other methods, it indicates the importance of having a range of frequencies. We implement uniform frequencies for Axial RoPE to gauge against relative importance of equivariance. + +In one extreme, the rotation frequency could be zero resulting in no changes to the queries and keys. In the other extreme, the frequency could be set very high resulting in large changes to the queries and keys. As a note, it is the frequency relative to the resolution of the image that is important. Frequencies higher than the sampling rate are equivalent to low frequencies. To ensure every position has a unique encoding, we fix the frequency to perform one rotation cycle across the entire image. + +Datasets and architecture. We test the different PEs on CIFAR100 [39] and ImageNet [58] using a standard Vision Transformer – the ViT-S implementation from the timm [79] library. For Learned APE, we use the baseline ViT-S which uses learned positional encodings rather than sinusoidal. We follow much of the DeiT-III training procedure proposed in Touvron et al. [72]. However, for ImageNet, we do not use dropout, MixUp, or CutMix as we observed that they significantly increase the number of epochs necessary for convergence. For ImageNet, we evaluate models trained after 200 epochs and 400 epochs for CIFAR100. We evaluate without any hyperparameter tuning directly on the validation sets. For further details on hyperparameters and experimental setup, see Appendix H. Error bars were created using three models with different random seeds. + +Generalization to larger image sizes. We also perform an experiment to test how well different PEs generalize across image sizes. Our approach to this experiment follows prior research [26, 52]. The learned embeddings in Learned APE cannot be extrapolated, so we interpolate new embeddings when changing the number of patches. For RoPE embeddings, we take square dimensions and parameterize position such that the top-left corner of the image corresponds to $p_x = p_y = -\pi$ and the bottom-right corner correspond to $p_x = p_y = \pi$ with all other positions are evenly spread between the two for training. When increasing the image size, we extrapolate by scaling the range by the ratio of the new image size to the training image size while keeping the patch size constant. + +Additional Evaluations Additional evaluations can be found in Appendix I including method speeds, experiments with smaller data splits, a segmentation task, and evaluation of the learned weights. + +Table 2: Performance comparison (top-1 accuracy) across datasets and methods. + +
Fixed EncodingTop-1 Accuracy (%)
CIFAR100ImageNet
Learned APE64.2±0.972.7 ±0.1
Axial RoPE72.1±0.675.6 ±0.2
Uniform RoPE (Our Ablation)70.5±0.274.9 ±0.3
Spherical RoPE (Our Ablation)73.2±0.476.4 ±0.3
Learned Encoding
Learned Axial RoPE72.9±0.675.7 ±0.4
LieRE73.1±0.2
Mixed RoPE74.7±0.377.4 ±0.1
Learned Spherical RoPE (Our Ablation)74.1±0.477.4 ±0.2
+ +![](images/ad04e9feb09017d2a0b5d2f3690259234fe7cfa72e24043b7f69c5faf5080819.jpg) +Figure 3: Dependence of accuracy on image resolution for ViT-S with various positional embedding methods on ImageNet-1k. Error bars reflect the standard deviation across three models trained with different seeds. + +# 5 Results + +To evaluate the importance of different properties of positional embeddings in vision transformers, we trained the same ViT with different positional embeddings on CIFAR100 and ImageNet-1K. We start by evaluating the models on images of the same resolution as during training. If equivariance is important, we would see Axial and Mixed RoPE to perform better than Spherical RoPE, which lacks strict equivariance. On the other hand, if oblique frequencies are important, then we would observe Mixed and Spherical RoPE to do better than Axial RoPE, which does not capture oblique directions. We find that the lack of equivariance does not hinder Spherical RoPE. It outperforms Axial RoPE and performs comparably to Mixed RoPE. Moreover, we would expect equivariant methods to be especially effective in the low data regime. However, in Appendix I we observe Learned Spherical RoPE performs the best despite its lack of inductive bias. This suggests that the benefits to performance and generalization on ImageNet for Mixed RoPE may be due to its extra parameters. However, Axial RoPE and Uniform RoPE perform significantly worse suggesting oblique directions to be more important than equivariance. + +When comparing with absolute positional encodings, we observe that all forms of RoPE perform better than learned APE (Table 2). This includes Uniform RoPE, the variant that uses only a single frequency. Moreover, all forms of RoPE using diverse frequencies outperform Uniform RoPE and have similar performance (whether they are learned or not), suggesting that diversity of frequencies is important. Spherical RoPE adheres much closer to the vectorized implementation of other RoPE methods than LieRE. As our goal was primarily to identify the most impactful properties of $M$ -D RoPE and not maximize accuracy, none of our conclusions depend on precise performance numbers for LieRE. + +Last, we asked how well different PEs generalize across image sizes. Equivariance is often thought to aid model generalization. However, when evaluating each model using higher resolutions images, i.e. increasing the number of patches, we found Spherical RoPE to be the most effective method (Figure 3), suggesting equivariance may not be the reason for RoPE's generalization. + +# 6 Discussion + +Because we see very little variation between Spherical RoPE and Mixed RoPE, we conclude that equivariance is only a minor contributor to the increased performance seen by RoPE for vision. In fact, Spherical RoPE appeared to extrapolate to higher resolutions better than Axial RoPE. This could suggest that oblique frequencies are important for extrapolation. However, extrapolation is only done on short length scales, so this may not hold in language. + +There are two important differences between vision and language transformers: context length and patch variation. Where LLMs have on the order of 128K context windows [22], vision transformers + +only have $16 \times 16$ patches. Moreover, patches have more variation as tokens than language tokens, thus allowing the content embeddings to store information about the relevance of oblique directions. Because the context size is small, we hypothesize that there could be methods that perform better than Mixed RoPE and are more general than LieRE for vision. While LieRE was proposed with skew-symmetric generators to generalize RoPE to $N$ -D rotations, Lie algebras do not have to be skew-symmetric. The skew-symmetry is important for maintaining numerical stability over long contexts [69]. However, skew-symmetry also results in Proposition 3.1 of Barbero et al. [5], which proves RoPE to be non-local. Since the context size is small for images, numerical stability is likely not an issue, thus freeing the space of Lie algebras available to us - including Lie algebras that encourage locality. + +We observe a decrease in generalization when using uniform frequencies. This finding qualifies Barbero et al. [5]'s hypothesis that the various semantic lengths contribute to RoPE's performance. However, Uniform RoPE outperformed learned APE, suggesting the reason why RoPE performs well is not among the properties we tested. We speculate this could be a flaw in additive positional embeddings themselves; additive methods create a trade-off between the magnitude of position and content – forcing tokens that vary significantly with position to have lower magnitude to be closer to the origin. + +# 7 Conclusion + +We conclude that Mixed RoPE is a very general solution for $M$ -D data if equivariance is a necessity. However, we see little evidence that strict relative positional bias is impactful for vision transformers. However, RoPE methods have still been found to greatly improve performance in ViTs. Thus, we conclude that evidence suggests that RoPE does not need strict equivariance constraints to boost performance over APE methods. + +# 8 Acknowledgments + +This project has received funding from the European Research Council (ERC) under the European Union's Horizon Europe research and innovation programme (Grant agreement No. 101041669). We gratefully acknowledge the computing time granted by the Resource Allocation Board and provided on the supercomputer Emmy/Grete at NHR-Nord@Göttingen as part of the NHR infrastructure (project nim00012). + +# References + +[1] Josh Abramson, Jonas Adler, Jack Dunger, Richard Evans, Tim Green, Alexander Pritzel, Olaf Ronneberger, Lindsay Willmore, Andrew J Ballard, Joshua Bambrick, et al. Accurate structure prediction of biomolecular interactions with alphafold 3. Nature, 630(8016):493-500, 2024. +[2] Meta AI. Llama 4: Multimodal language models, 2025. URL https://ai.meta.com/blog/llama-4-multimodal-intelligence/. Accessed: 2025-05-15. +[3] Giorgio Angelotti. Hype: Attention with hyperbolic biases for relative positional encoding. arXiv preprint arXiv:2310.19676, 2023. +[4] Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. Neural machine translation by jointly learning to align and translate. arXiv preprint arXiv:1409.0473, 2014. +[5] Federico Barbero, Alex Vitvitskyi, Christos Perivolaropoulos, Razvan Pascanu, and Petar Velicković. Round and round we go! what makes rotary positional encodings useful? arXiv preprint arXiv:2410.06205, 2024. +[6] Erik J Bekkers, Sharvaree Vadgama, Rob D Hesselink, Putri A Van der Linden, and David W Romero. Fast, expressive se $(n)$ equivariant networks through weight-sharing in position-orientation space. arXiv preprint arXiv:2310.02970, 2023. +[7] Stella Biderman, Sid Black, Charles Foster, Leo Gao, Eric Hallahan, Horace He, Ben Wang, and Phil Wang. Rotary embeddings: A relative revolution. blog .eleuther.ai/, 2021. [Online; accessed ]. + +[8] Johannes Brandstetter, Rob Hesselink, Elise van der Pol, Erik J Bekkers, and Max Welling. Geometric and physical quantities improve e (3) equivariant message passing. arXiv preprint arXiv:2110.02905, 2021. +[9] Johann Brehmer, Sonke Behrends, Pim de Haan, and Taco Cohen. Does equivariance matter at scale? arXiv preprint arXiv:2410.23179, 2024. +[10] Yiting Chen and Junchi Yan. What rotary position embedding can tell us: Identifying query and key weights corresponding to basic syntactic or high-level semantic information. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. +[11] Ta-Chung Chi, Ting-Han Fan, Peter J Ramadge, and Alexander Rudnicky. Kerple: Kernelized relative positional embedding for length extrapolation. Advances in Neural Information Processing Systems, 35:8386-8399, 2022. +[12] Ta-Chung Chi, Ting-Han Fan, Alexander I Rudnicky, and Peter J Ramadge. Dissecting transformer length extrapolation via the lens of receptive field analysis. arXiv preprint arXiv:2212.10356, 2022. +[13] Krzysztof Choromanski, Valerii Likhosherstov, David Dohan, Xingyou Song, Andreea Gane, Tamas Sarlos, Peter Hawkins, Jared Davis, Afroz Mohiuddin, Lukasz Kaiser, et al. Rethinking attention with performers. arXiv preprint arXiv:2009.14794, 2020. +[14] Xiangxiang Chu, Zhi Tian, Bo Zhang, Xinlong Wang, and Chunhua Shen. Conditional positional encodings for vision transformers. arXiv preprint arXiv:2102.10882, 2021. +[15] Xiangxiang Chu, Jianlin Su, Bo Zhang, and Chunhua Shen. Visionllama: A unified llama interface for vision tasks. arXiv e-prints, pages arXiv-2403, 2024. +[16] Tri Dao, Dan Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. Flashattention: Fast and memory-efficient exact attention with io-awareness. Advances in neural information processing systems, 35:16344-16359, 2022. +[17] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. +[18] Gamealeldin Elsayed, Prajit Ramachandran, Jonathon Shlens, and Simon Kornblith. Revisiting spatial invariance with low-rank local connectivity. In International Conference on Machine Learning, pages 2868-2879. PMLR, 2020. +[19] Mark Everingham, S. M. Ali Eslami, Luc Van Gool, Christopher K. I. Williams, John Winn, and Andrew Zisserman. The Pascal visual object classes challenge: A retrospective. International Journal of Computer Vision, 111(1):98-136, 2015. doi: 10.1007/s11263-014-0733-5. +[20] Jonas Gehring, Michael Auli, David Grangier, Denis Yarats, and Yann N. Dauphin. Convolutional sequence to sequence learning. CoRR, abs/1705.03122, 2017. URL http://arxiv.org/abs/1705.03122. +[21] Justin Gilmer, Samuel S Schoenholz, Patrick F Riley, Oriol Vinyals, and George E Dahl. Neural message passing for quantum chemistry. In International conference on machine learning, pages 1263-1272. PMLR, 2017. +[22] Aaron Grattafori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. +[23] D.J. Griffiths. Introduction to Quantum Mechanics. CUP, 2018. +[24] Adi Haviv, Ori Ram, Ofir Press, Peter Izsak, and Omer Levy. Transformer language models without positional encodings still learn positional information. arXiv preprint arXiv:2203.16634, 2022. +[25] Yu He, Cristian Bodnar, and Pietro Lio. Sheaf-based positional encodings for graph neural networks. In NeurIPS 2023 Workshop on Symmetry and Geometry in Neural Representations, volume 9, 2023. +[26] Byeongho Heo, Song Park, Dongyoon Han, and Sangdoo Yun. Rotary position embedding for vision transformer. In European Conference on Computer Vision, pages 289-305. Springer, 2024. + +[27] Mohammad Mohaiminul Islam, Rishabh Anand, David R Wessels, Friso de Kruiff, Thijs P Kuipers, Rex Ying, Clara I Sánchez, Sharvaree Vadgama, Georg Bökman, and Erik J Bekkers. Platonic transformers: A solid choice for equivariance. arXiv preprint arXiv:2510.03511, 2025. +[28] Albert Q Jiang, Alexandre Sablayrolles, Antoine Roux, Arthur Mensch, Blanche Savary, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Emma Bou Hanna, Florian Bressand, et al. Mixtral of experts. arXiv preprint arXiv:2401.04088, 2024. +[29] John Jumper, Richard Evans, Alexander Pritzel, Tim Green, Michael Figurnov, Olaf Ronneberger, Kathryn Tunyasuvunakool, Russ Bates, Augustin Žídek, Anna Potapenko, et al. Highly accurate protein structure prediction with alphafold. nature, 596(7873):583-589, 2021. +[30] Pentti Kanerva. Hyperdimensional computing: An introduction to computing in distributed representation with high-dimensional random vectors. Cognitive computation, 1:139-159, 2009. +[31] Pentti Kanerva. Hyperdimensional computing: An algebra for computing with vectors. Advances in Semiconductor Technologies: Selected Topics Beyond Conventional CMOS, pages 25-42, 2022. +[32] Angelos Katharopoulos, Apoorv Vyas, Nikolaos Pappas, and François Fleuret. Transformers are rnns: Fast autoregressive transformers with linear attention. In International conference on machine learning, pages 5156-5165. PMLR, 2020. +[33] Amirhossein Kazemnejad, Inkit Padhi, Karthikeyan Natesan Ramamurthy, Payel Das, and Siva Reddy. The impact of positional encoding on length generalization in transformers. Advances in Neural Information Processing Systems, 36:24892-24928, 2023. +[34] T Anderson Keller and Max Welling. Topographic vaes learn equivariant capsules. Advances in Neural Information Processing Systems, 34:28585-28597, 2021. +[35] T Anderson Keller and Max Welling. Neural wave machines: learning spatiotemporally structured representations with locally coupled oscillatory recurrent neural networks. In International Conference on Machine Learning, pages 16168-16189. PMLR, 2023. +[36] David Knigge, David Wessels, Riccardo Valperga, Samuele Papa, Jan-Jakob Sonke, Erik Bekkers, and Efstratios Gavves. Space-time continuous pde forecasting using equivariant neural fields. Advances in Neural Information Processing Systems, 37:76553-76577, 2024. +[37] Miltiadis Kofinas, Naveen Nagaraja, and Efstratios Gavves. Roto-translated local coordinate frames for interacting dynamical systems. Advances in Neural Information Processing Systems, 34:6417-6429, 2021. +[38] Devin Kreuzer, Dominique Beaini, Will Hamilton, Vincent Létourneau, and Prudencio Tossou. Rethinking graph transformers with spectral attention. Advances in Neural Information Processing Systems, 34:21618-21629, 2021. +[39] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images, 2009. +[40] Christopher J Kymn, Denis Kleyko, E Paxon Frady, Connor Bybee, Pentti Kanerva, Friedrich T Sommer, and Bruno A Olshausen. Computing with residue numbers in high-dimensional representation. Neural Computation, 37(1):1-37, 2024. +[41] James Lee-Thorp, Joshua Ainslie, Ilya Eckstein, and Santiago Ontanon. Fnet: Mixing tokens with fourier transforms. arXiv preprint arXiv:2105.03824, 2021. +[42] Pan Li, Yanbang Wang, Hongwei Wang, and Jure Leskovec. Distance encoding: Design provably more powerful neural networks for graph representation learning. Advances in Neural Information Processing Systems, 33:4465-4478, 2020. +[43] Aixin Liu, Bei Feng, Bing Xue, Bingxuan Wang, Bochao Wu, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, et al. Deepseek-v3 technical report. arXiv preprint arXiv:2412.19437, 2024. +[44] Haiping Liu and Hongpeng Zhou. Rethinking rope: A mathematical blueprint for n-dimensional positional encoding. arXiv preprint arXiv:2504.06308, 2025. +[45] Xuanqing Liu, Hsiang-Fu Yu, Inderjit S. Dhillon, and Cho-Jui Hsieh. Learning to encode position for transformer with continuous dynamical model. CoRR, abs/2003.09229, 2020. URL https://arxiv.org/abs/2003.09229. + +[46] Yunwu Liu, Ruisheng Zhang, Tongfeng Li, Jing Jiang, Jun Ma, and Ping Wang. Molrope-bert: An enhanced molecular representation with rotary position embedding for molecular property prediction. Journal of Molecular Graphics and Modelling, 118:108344, 2023. +[47] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. +[48] Zikang Liu, Longteng Guo, Yepeng Tang, Junxian Cai, Kai Ma, Xi Chen, and Jing Liu. Vrope: Rotary position embedding for video large language models. arXiv preprint arXiv:2502.11664, 2025. +[49] Sindy Löwe, Phillip Lippe, Maja Rudolph, and Max Welling. Complex-valued autoencoders for object discovery. arXiv preprint arXiv:2204.02075, 2022. +[50] Sindy Löwe, Phillip Lippe, Francesco Locatello, and Max Welling. Rotating features for object discovery. Advances in Neural Information Processing Systems, 36:59606-59635, 2023. +[51] Takeru Miyato, Sindy Löwe, Andreas Geiger, and Max Welling. Artificial kuramoto oscillatory neurons. arXiv preprint arXiv:2410.13821, 2024. +[52] Sophie Ostmeier, Brian Axelrod, Michael E Moseley, Akshay Chaudhari, and Curtis Langlotz. Liere: Generalizing rotary position encodings. arXiv preprint arXiv:2406.10322, 2024. +[53] Wonpyo Park, Woonggi Chang, Donggeon Lee, Juntae Kim, and Seung-won Hwang. Grpe: Relative positional encoding for graph transformer. arXiv preprint arXiv:2201.12787, 2022. +[54] Ngoc-Quan Pham, Thanh-Le Ha, Tuan-Nam Nguyen, Thai-Son Nguyen, Elizabeth Salesky, Sebastian Stüker, Jan Niehues, and Alexander Waibel. Relative positional encoding for speech recognition and direct translation. arXiv preprint arXiv:2005.09940, 2020. +[55] Ofir Press, Noah A Smith, and Mike Lewis. Train short, test long: Attention with linear biases enables input length extrapolation. arXiv preprint arXiv:2108.12409, 2021. +[56] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research, 21(140):1-67, 2020. +[57] Ali Rahimi and Benjamin Recht. Random features for large-scale kernel machines. Advances in neural information processing systems, 20, 2007. +[58] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115:211-252, 2015. +[59] Victor Garcia Satorras, Emiel Hoogeboom, and Max Welling. E (n) equivariant graph neural networks. In International conference on machine learning, pages 9323-9332. PMLR, 2021. +[60] Connor Schenck, Isaac Reid, Mithun George Jacob, Alex Bewley, Joshua Ainslie, David Rendleman, Deepali Jain, Mohit Sharma, Avinava Dubey, Ayzaan Wahid, et al. Learning the ropes: Better 2d and 3d position encodings with string. arXiv preprint arXiv:2502.02562, 2025. +[61] Jürgen Schmidhuber. Learning to control fast-weight memories: An alternative to dynamic recurrent networks. Neural Computation, 4(1):131-139, 1992. +[62] Kristof Schütt, Oliver Unke, and Michael Gastegger. Equivariant message passing for the prediction of tensorial properties and molecular spectra. In International Conference on Machine Learning, pages 9377-9388. PMLR, 2021. +[63] Fahad Shamshad, Salman Khan, Syed Waqas Zamir, Muhammad Haris Khan, Munawar Hayat, Fahad Shahbaz Khan, and Huazhu Fu. Transformers in medical imaging: A survey. Medical image analysis, 88:102802, 2023. +[64] Peter Shaw, Jakob Uszkoreit, and Ashish Vaswani. Self-attention with relative position representations. arXiv preprint arXiv:1803.02155, 2018. +[65] Vincent Sitzmann, Julien N. P. Martel, Alexander W. Bergman, David B. Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. In Advances in Neural Information Processing Systems (NeurIPS). Curran Associates, Inc., 2020. URL http://arxiv.org/abs/2006.09661v1. + +[66] Ben Sorscher, Gabriel C Mel, Samuel A Ocko, Lisa M Giocomo, and Surya Ganguli. A unified theory for the computational and mechanistic origins of grid cells. Neuron, 111(1):121-137, 2023. +[67] Steven H Strogatz and Ian Stewart. Coupled oscillators and biological synchronization. Scientific american, 269(6):102-109, 1993. +[68] Jianlin Su. Transformer upgrade road: 4. rotating position coding of two-dimensional position, May 2021. URL https://kexue.fm/archives/8397. Accessed: 2025-04-28. +[69] Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding. Neurocomputing, 568:127063, 2024. +[70] Matthew Tancik, Pratul Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. Advances in neural information processing systems, 33:7537-7547, 2020. +[71] Gemma Team, Thomas Mesnard, Cassidy Hardin, Robert Dadashi, Surya Bhupatiraju, Shreya Pathak, Laurent Sifre, Morgane Riviere, Mihir Sanjay Kale, Juliette Love, et al. Gemma: Open models based on gemini research and technology. arXiv preprint arXiv:2403.08295, 2024. +[72] Hugo Touvron, Matthieu Cord, and Hervé Jégou. Deit iii: Revenge of the vit. In European conference on computer vision, pages 516-533. Springer, 2022. +[73] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. +[74] Petar Velickovic, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Lio, and Yoshua Bengio. Graph attention networks. arXiv preprint arXiv:1710.10903, 2017. +[75] Jie Wang, Tao Ji, Yuanbin Wu, Hang Yan, Tao Gui, Qi Zhang, Xuanjing Huang, and Xiaoling Wang. Length generalization of causal transformers without position encoding. arXiv preprint arXiv:2404.12224, 2024. +[76] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-v1: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. +[77] Xilin Wei, Xiaoran Liu, Yuhang Zang, Xiaoyi Dong, Pan Zhang, Yuhang Cao, Jian Tong, Haodong Duan, Qipeng Guo, Jiaqi Wang, et al. Videorope: What makes for good video rotary position embedding? arXiv preprint arXiv:2502.05173, 2025. +[78] David R Wessels, David M Knigge, Samuele Papa, Riccardo Valperga, Sharvaree Vadgama, Efstratios Gavves, and Erik J Bekkers. Grounding continuous representations in geometry: Equivariant neural fields. arXiv preprint arXiv:2406.05753, 2024. +[79] Ross Wightman. Pytorch image models. https://github.com/rwrightman/pytorch-image-models, 2019. +[80] Yiheng Xie, Towaki Takikawa, Shunsuke Saito, Or Litany, Shiqin Yan, Numair Khan, Federico Tombari, James Tompkin, Vincent Sitzmann, and Srinath Sridhar. Neural fields in visual computing and beyond. Computer Graphics Forum, 2022. ISSN 1467-8659. doi: 10.1111/cgf.14505. +[81] Keyulu Xu, Weihua Hu, Jure Leskovec, and Stefanie Jegelka. How powerful are graph neural networks? arXiv preprint arXiv:1810.00826, 2018. +[82] Bowen Yang, Bharat Venkitesh, Dwarak Talupuru, Hangyu Lin, David Cairuz, Phil Blunsom, and Acyr Locatelli. Rope to nope and back again: A new hybrid attention strategy. arXiv preprint arXiv:2501.18795, 2025. +[83] Jiaxuan You, Rex Ying, and Jure Leskovec. Position-aware graph neural networks. In International conference on machine learning, pages 7134-7143. PMLR, 2019. +[84] Yang You, Jing Li, Sashank Reddi, Jonathan Hseu, Sanjiv Kumar, Srinadh Bhojanapalli, Xiaodan Song, James Demmel, Kurt Keutzer, and Cho-Jui Hsieh. Large batch optimization for deep learning: Training bert in 76 minutes. arXiv preprint arXiv:1904.00962, 2019. + +[85] Hao Yu, Tangyu Jiang, Shuning Jia, Shannan Yan, Shunning Liu, Haolong Qian, Guanghao Li, Shuting Dong, Huaisong Zhang, and Chun Yuan. Comrope: Scalable and robust rotary position embedding parameterized by trainable commuting angle matrices, 2025. URL https://arxiv.org/abs/2506.03737. +[86] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6023-6032, 2019. +[87] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412, 2017. + +# A Broader Impact + +This work is fundamental research. While this work could lead to the discovery of better positional encodings and higher performing visual foundation models, the positivity or negativity of this impact is determined by the downstream task and not this work. + +# B Limitations + +While our results do not show relative embeddings to be detrimental, we believe them to be evidence that equivariance is not the reason for RoPE's success. However, our experiments were performed in Vision where the number of tokens is limited compared to the long context lengths of NLP. Moreover, the datasets are not what many believe to be "at scale". While Spherical RoPE and LieRE would intuitively be favored at scale over Axial RoPE, as they have less inductive bias, it is unclear whether inductive bias and equivariance is favored at scale [9]. + +It has also been shown that vision is not a purely equivariant task and benefits from relaxed equivariance [18]. Our results do not show that equivariance is not useful in tasks that are grounded in physics and obey strict symmetries. + +# C Literature Review + +# C.1 Natural Language Processing + +In natural language, positional encoding has been used to break the permutation, "bag of words", symmetry [73]. Although this could be done by learning a vector per position, this is both memory-expensive for large context sizes making it practical to apply to only the first layer. Moreover, it does not allow for extrapolation at test time to context sizes beyond training. Thus, it is favorable to perform positional embeddings with a predictable deterministic function. One way of doing this is to make the attention relative with local receptive fields, as is done implicitly in convolutional neural networks [12]. Sinusoidal positional embeddings were proposed due to approximate local and shift-invariant properties of Random Fourier Features [57]. Since sinusoidal, other methods have been proposed to get guaranteed shift invariance by explicitly parameterizing based on distance [64, 54, 55]. However, these methods require a positional embedding for every pair of positions which is not supported by many of the efficient attention optimizations such as Flash attention [16] [3]. + +Rotary Positional Embeddings (RoPE) have become the staple in NLP having recently been adopted by many of the large language models [76, 22, 71, 43, 28]. However, these methods also use causal masking, which has been shown to allow models with no positional embedding to recover absolute position [24, 82, 75, 33]. This has led to questions on the importance of relative position [5]. + +In language, there has also been extensions to RoPE proposed through NTKs and kernel methods [11]. However, these methods have not, to our knowledge, seen use in vision. + +# C.2 Vision and Video + +Vision transformers were introduced in Dosovitskiy et al. [17] and, though they tried sinusoidal position encodings, found learnable position encodings to perform best. For convolution-esque models such as SWin transformers, relative positional encodings have been popular [47, 14]. More recently, RoPE has been shown to be an efficient and simple way to have relative embeddings and has been extended to 2D using Axial and Mixed RoPE. Going beyond 2D to Video data, Axial RoPE has become increasingly popular. The extension was first attributed to Wang et al. [76] as 3D-RoPE or M-RoPE, leading to two separate Video-RoPE papers from Wei et al. [77] and Liu et al. [48]. Both of these focus on the order of the position enumeration and interleaving positions. However, this should not be a problem if frequencies are not deterministic, or if frequencies are indexed by both $d$ and modality $m$ as done in Eq 13. We highly recommend using either Mixed RoPE or LieRE which extend naturally for videos. + +LieRE embeddings have thus far been the most general form of RoPE to $N$ -D. However, Schenck et al. [60] has claimed the method to have a large memory footprint and proposed STRING. This + +paper, a preprint released concurrently with the writing of this manuscript, follows much of the same math as this paper. However, they did not recognize that an orthogonal matrix is implicitly learned by the query and key matrix. Moreover, their method relies on commuting Lie algebras. From our insights in Section 3, their method can likely be viewed as a slower implementation of $N$ -D Mixed-RoPE. + +It is also worth noting that positional encodings have also been explored within vision through the area of Neural Fields [80]. Traditional coordinate MLPs have been found to be biased toward low-frequency functions [70] leading to more advanced positional encodings such as Random Fourier Features [57] or sinusoidal activation functions [65]. These implicit functions have been used to encode attention and message passing in graph neural networks with recent work being put in to make these functions equivariant to symmetry transformations [59, 8, 36]. + +# C.3 Graphs and AI in Science + +Positional encodings are well studied within graph neural networks [42, 53]. Graphs are limited in their expressivity up to the Weisfeiler-Lehman (WL) graph isomorphism test [81], so positional encodings can break the isomorphism symmetry [25, 83]. Within this community, they propose spectral attention and graph Laplacians for positional encoding [38]. These methods seem extremely close to our analysis of RoPE, but from a very different perspective. We show that the frequencies of RoPE can be interpreted as the eigenvalues of an orthogonal transformation by taking the spectral decomposition. + +In an overlapping vein, relative position encodings have been studied in terms of equivariant graph neural networks, often for scientific disciplines such as molecular physics [8, 62] or drug discovery [29]. One method to achieve equivariance is through defining relative coordinate frames [37]. This corresponds to the learned relative positional method described in Shaw et al. [64], but can be generalized to higher dimensions and different transformation using bi-invariant distance functions [6, 36, 78]. The message-passing functions of these works correspond to a generalization of attention scores [21]. + +However, even in these tasks with physics-grounded symmetries, the need for equivariance is hotly debated. While AlphaFold [29] was originally touted as the example of the success of equivariant inductive biases in science, AlphaFold 3 [1] explicitly stated that they benefited from removing this inductive bias at scale. However, while the harm of inductive bias at scale is the prevalent zeitgeist, it is not an established fact [9]. + +# C.4 Computational Neuroscience + +Coupled oscillators have become a growing area of interest within computational neuroscience [34, 35, 67]. By observing the projection of the RoPE circles onto the real axis, one can interpret RoPE as time progression in $D$ uncoupled, undamped harmonic oscillators. This perspective naturally connects RoPE to Löwe et al. [49]'s series of papers on complex autoencoders and their extensions [50, 51]. + +In another, vein of research, there has been some work in hyper-dimensional computing[30, 31] in Phasor and Residue VSAs [40] which represent concepts as rotations around unit circles in high-dimensional spaces. These representations have strong connections with RoPE. Additionally, progress has been made in hypothesizing how biological neural networks encode positional knowledge with hexagonal grid cells, which can be represented as a discrete sum of three periodic functions oriented at the cubic roots of unity[66]. + +# C.5 Generality of RoPE + +The generality of RoPE has been found by others. Schenck et al. [60], Su [68], and Liu and Zhou [44] all propose proofs similar to Proposition 1. However, Schenck et al. [60] miss that the orthogonal transformation can be incorporated into key matrix. Liu and Zhou [44] and Su [68] take the assumption of reversibility, which leads to the independent eigenvalue assumptions of Axial RoPE. All three works take the assumption of an abelian subgroup - i.e. commutative generators, - but miss the generality of Mixed RoPE. While Su [68] propose quaternions - i.e. spherical rotations - as a direction, they immediately dismiss it as a no-go because they lack equivariance. This exemplifies the + +"circular argument," where equivariance is assumed to be necessary because work will not investigate non-equivariant positional encodings because equivariance is necessary. + +Because our derivation was found independently of these works and the previous works are, to our knowledge, not published, we have left in Proposition 1. We would like to acknowledge their work, but retain the flow of this paper. + +# D Notation + +
Symbol / TermDimensionMeaningNotes
xiRDPatch/token/content vector of token iRaw input embedding
xiXAbstract content of token iRaw input embedding
piRM or PPosition of token i, can be M-D or abstract PScalar (1D) or vector (2D)
mZModality indexe.g., x, y, time
MZNumber, or space, of Modalities
DZHidden dimensionNumber of pairs/triples/quadruples
TZNumber of Tokens
Wq, Wk, WvRX×DQuery, Key, Value Matrices
qRNqi = WqxiQuery vector
kRNkj = WkxjKey vector
vRNvj = Wv, xjValue vector
Q, K, VRT×NQuery, Key, ValuesT tokens, D latent dimensions
φ(x, p)X × P → RDPositional Encoding function
ZRT×NOutput of AttentionZ = Attention(Q, K, V)
a(i, j)RAttention weightSoftmax of attention scores
α(q, k)RAttention scoreInner product q' k
ωd/λdRRotation frequency for dimension dEquivalent to eigenvalue of generator
qdR2/3/4Query pair/triple/quadrule at dimension dAfter RoPE or LieRE applied
RωdpR2×22 × 2 rotation matrixRotation based on frequency and position
+ +Table 3: Summary of Notations and Key Concepts + +
Positional EncodingVisionLearnedExtrapolationQK SeparableRelativeLinear FlowUsed In
Absolute (Sinusoidal)X✓/X✓✓✓XTransformer[73]
Absolute (Learned)✓✓X✓✓XBERT, GPT, ViT[17]
Absolute (Random-Fourier)XX✓✓X✓FNet[41], Performer [13]
Relative (Learned)X✓XXXXTransformer-XL, T5 [56]
ALiBiX✓/X✓✓✓✓LLaMA 2 [22], ALiBi [55]
NoPEX*X✓*✓*✓*✓*LLaMA 4 [2]
Rotary (RoPE)XX✓✓✓✓Contemporary LLMs [76, 22, 71, 28]
Axial RoPE✓✓/X✓✓✓✓VisionLLaMA[15], Qwen2[76], VideoRoPE[77]
Mixed RoPE✓✓✓✓✓✓Heo et al. [26]
LieRE✓✓✓✓X✓[52]
Spherical RoPE✓✓/X✓✓X✓Ours
Uniform RoPE✓✓/X✓✓✓✓Ours
+ +Table 4: Comparison of positional encoding methods in transformer models. *NoPE makes some properties trivially true. + +# E Positional Encoding Properties + +Rotary positional embeddings were derived in Su et al. [69] by drawing equations from assumed properties. While these appear as arithmetic assumptions and equations in their work, we formalize what properties these assumptions imply and why we may choose these assumptions in this section. In their paper, to derive their equations, they use equivariance (relativity), query-key separability of the positional encoding, linearity and incompressability, locality, and query-key symmetry. + +1. Equivariance/Relativity: Attention score should be affected only by the relative position of two tokens, i.e. have the form + +$$ +\alpha \left(x _ {i}, x _ {j}, p _ {i}, p _ {j}\right) = \hat {\alpha} \left(x _ {i}, x _ {j}, p _ {i} - p _ {j}\right). \tag {26} +$$ + +2. Key-query seperability: The positional encoding, $\varphi$ , of the query should not depend on the position of the key + +$$ +\alpha \left(x _ {i}, x _ {j}, p _ {i}, p _ {j}\right) = \bar {\alpha} \left(\varphi \left(x _ {i}, p _ {i}\right), \varphi \left(x _ {j}, p _ {j}\right)\right) \tag {27} +$$ + +3. Linearity: The positional encoding should be a linear flow, see Appendix E.3. Namely, + +$$ +\varphi \left(\varphi \left(x, p _ {i}\right), p _ {j}\right) = \varphi \left(x, p _ {i} + p _ {j}\right). \tag {28} +$$ + +4. Locality: The attention score between two tokens should decay with distance + +$$ +\lim _ {\left| p _ {i} - p _ {j} \right|\rightarrow \infty} \alpha \left(x _ {i}, x _ {j}, p _ {i}, p _ {j}\right) = 0 \tag {29} +$$ + +# E.1 Relativity and Equivariant + +We use the term equivariant interchangeably with relative. Strictly speaking, one should specify the transformation or group you would like to be relative to, e.g. shift/rotation or $SO(2)$ . As previous literature always refers to relative positional bias in terms of shifts/translations, in the main text, this is what we mean. We use the term equivariance to be the generalization of relativity beyond language because we would like to refrain from using the term "relativity" to describe the property of being a relative PE too often due to its connotation within theoretical physics. First, we define relative in the case of positional encodings in language as + +$$ +\alpha \left(x _ {i}, x _ {j}, p _ {i}, p _ {j}\right) = \hat {\alpha} \left(x _ {i}, x _ {j}, p _ {i} - p _ {j}\right). \tag {30} +$$ + +In the rest of this section, we mathematically explore where this equation comes from. + +The behavior we are trying to capture is that if we renumber the words in the sentence, it should not affect the attentions score. Intuitively, if a text is padded with spaces at the beginning, that will not have a significant effect on the meaning of the sentences. We can ensure this by colloquially saying that the attention between two words should depend on the distance between them. Notice, that strictly speaking this is not a proper distance, since it can be negative; it is, instead, a signed distance function. Though this may seem pedantic in one dimension, in two dimensions defining a distance function is less unique. For example, one may choose $\mathbb{L}_1$ or $\mathbb{L}_2$ distance metrics. Because distance functions are more nebulous, it makes more sense to define relative in terms of the transformations that we would like our attention score to be independent of. + +$$ +\alpha \left(x _ {i}, x _ {j}, p _ {i}, p _ {j}\right) = \alpha \left(x _ {i}, x _ {j}, T \left(p _ {i}\right), T \left(p _ {j}\right)\right). \tag {31} +$$ + +These transformations can be combined to generate a set of transformations which leave the attention score unchanged, or symmetric. This set has the mathematical properties of a group and is known as a symmetry group. We can index transformations by elements in the symmetry group, $g \in G$ , and let the elements act on + +$$ +\alpha \left(x _ {i}, x _ {j}, p _ {i}, p _ {j}\right) = \alpha \left(x _ {i}, x _ {j}, g. p _ {i}, g. p _ {j}\right). \tag {32} +$$ + +As an example, $g$ could represent an angle, $\theta$ , and it may act on a vector $\mathbf{p}$ as a rotation $g.\mathbf{p} = \mathbf{R}_{\theta}\mathbf{p}$ . + +Connecting everything back to Eq. 30, Noether's theorem states that any continuous symmetry can be expressed as a conservation law. This allows us to introduce bi-invariant function [36, 78], or "Noether charge", $\beta(p_i, p_j)$ , that is invariant under the group action, + +$$ +\beta \left(p _ {i}, p _ {j}\right) = \beta \left(g. p _ {i}, g. p _ {j}\right) \Longrightarrow \beta \left(p _ {i}, p _ {j}\right) - \beta \left(g. p _ {i}, g. p _ {j}\right) = 0. \tag {33} +$$ + +Thus, we can express our symmetry group through isodistances of $\beta$ , + +$$ +\alpha \left(x _ {i}, x _ {j} p _ {i}, p _ {j}\right) := \hat {\alpha} \left(x _ {i}, x _ {j}, \beta \left(p _ {i}, p _ {j}\right)\right). \tag {34} +$$ + +For example, we can pick the function + +$$ +\beta \left(p _ {i}, p _ {j}\right) = p _ {i} - p _ {j} = \left(p _ {i} - p _ {0}\right) - \left(p _ {j} - p _ {0}\right) = \beta \left(p _ {i} - o, p _ {j} - p _ {0}\right) \tag {35} +$$ + +If we were to define $\beta(p_i, p_j) = |p_i - p_j|$ , then we would additionally be equivariant to reflection of the order of tokens in a sentence. If we trivially define $\beta(p_i, p_j) = C$ , then we arrive at bag of words, or no positional encoding (NoPE). For a list of common transformations and their corresponding bi-invariants see Theorem 1 of Bekkers et al. [6]. + +# E.2 Query-Key Separability + +Query and key separability is important for efficiency reasons. If we can decompose our positional encoded attention score as, + +$$ +\alpha \left(x _ {i}, x _ {j}, p _ {i}, p _ {j}\right) = \alpha \left(\varphi \left(x _ {i}, p _ {i}\right), \varphi \left(x _ {j}, p _ {j}\right)\right) \tag {36} +$$ + +then we can pre-compute the positional encoding for the queries and keys on time making the computation $O(T)$ . If the positional encoding is not separable, then it will need to be computed for every pair, $(i,j)[47,56,64]$ . Although there are many symmetries that can be exploited to make this not a quadratic computation, it removes the symmetries exploited by efficient attention mechanisms [7, 13, 32]. + +# E.3 Linear Flow Property + +The property of being a "flow" was first proposed in Liu et al. [45], however it is not often discussed. It is a property inherently present in RoPE[69], LieRE[52] and ALiBi [55] embeddings, specifically as a linear flow. + +We use the term linear flow for this property because the embedding can be found by repeated application of a linear function. However, the term "linear" this is a small misnomer because it is only locally linear. We define a flow as function + +$$ +\varphi : \mathbb {R} ^ {N} \times \mathbb {R} \rightarrow \mathbb {R} ^ {N} \tag {37} +$$ + +such that for all $x \in X$ and $p_1, p_2 \in \mathbb{R}$ , the following conditions hold: + +1. Initial condition (identity at time zero): + +$$ +\varphi (0, x) = x \tag {38} +$$ + +2. Group property (flow property): + +$$ +\varphi \left(\varphi (\mathbf {x}, p _ {1}), p _ {2})\right) = \varphi (x, p _ {1} + p _ {2}) \tag {39} +$$ + +3. Continuity (or differentiability): $\varphi$ is continuous with respect to its variables, depending on the context + +Strictly speaking, continuity is not necessary for positional encodings as positions tend to be integer values. What we really wish to capture with this property is for the positional encoding to be recursively defined. It may be strange to wish to apply the positional encoding multiple times; however, by having the positional encoding as an endomorphism it can allow for more predictable behavior when extrapolating to larger contexts, which we suspect helps the model train. + +We define a position embedding to be a linear flow if the flow has the form: + +$$ +\varphi (\mathbf {x}, \Delta p) = \mathbf {A} \mathbf {x}, \tag {40} +$$ + +for $\mathbf{A} \in \mathbb{R}^{N \times N}$ and $\mathbf{x} \in \mathbb{R}^N$ , where $\Delta p$ is the increment rate for position. By Eq. 39, any position $p \coloneqq p_0 \Delta p$ can then be attained by, + +$$ +\varphi (\mathbf {x}, p) = \mathbf {A} ^ {p _ {0}} \mathbf {x}. \tag {41} +$$ + +This can be seen as a geometric series if $\mathbf{A}$ is a scalar as seen in Press et al. [55]. If we let $\Delta t$ become infinitesimal, then we can express the recurrence relationship as the ODE, + +$$ +\frac {\partial \varphi}{\partial t} = \mathcal {A} \varphi \tag {42} +$$ + +which we can integrate to get, + +$$ +\varphi (\mathbf {x}, p) = \exp (\mathcal {A} p) \mathbf {x} \tag {43} +$$ + +This $\mathcal{A}$ is our generator of the flow, which is also a generator for a matrix Lie algebra, which we focus on in the main text. The matrix exponential, $\exp : \mathbb{R}^{N \times N} \to \mathbb{R}^{N \times N}$ , can be unstable for long contexts; similar to the scalar exponential function $e^{xp}$ , the function can quickly become large for high values of $x$ . However, this can be stable value $x = 0$ , since it always results in one. Similarly, the matrix exponential can be stable if the divergence of the flow - trace of the generator - is zero. We call flow "incompressible" or "divergence-free" if the trace of $\mathcal{A}$ is zero, making the determinant of $\mathbf{A}$ unit. If fluid dynamics, this is called incompressibility. For fluids, this implies that the flow conserves mass. + +If there are more than one generator of the Lie group, $\mathcal{A}_1$ and $\mathcal{A}_2$ , then Eq. 39 must be modified to, + +$$ +\varphi \left(\varphi \left(\mathbf {x}, \mathbf {p} _ {1}\right), \mathbf {p} _ {2}\right)\left. \right) = \varphi \left(\mathbf {x}, \mathbf {p} _ {1} \circ \mathbf {p} _ {2}\right), \tag {44} +$$ + +where $\circ$ is the group product. By the Baker-Campbell-Hausdorff formula, $\exp \mathcal{A}_1p_1\exp \mathcal{A}_2p_2 = \exp \mathcal{A}_1p_1 + \mathcal{A}_2p_2$ iff the commutator of $\mathcal{A}_1p_1$ and $\mathcal{A}_2p_2$ is zero, i.e. the matrices commute. If they do commute, then + +$$ +\varphi \left(\varphi \left(\mathbf {x}, \mathbf {p} _ {1}\right), \mathbf {p} _ {2}\right)\left. \right) = \varphi \left(\varphi \left(\mathbf {x}, \mathbf {p} _ {2}\right), \mathbf {p} _ {1}\right)) \Longrightarrow \varphi \left(\mathbf {x}, \mathbf {p} _ {1} \circ \mathbf {p} _ {2}\right) = \varphi \left(\mathbf {x}, \mathbf {p} _ {2} \circ \mathbf {p} _ {1}\right) \tag {45} +$$ + +thus making $\circ$ commutative and having the same properties as addition, $\circ := "$ +, and Eq. 39 will hold. In this case, the group/flow is known as an abelian Lie group, or abelian flow. However, if they do not commute, then $\circ$ will not commute and they are known as non-abelian. This also makes the flow non-integrable. + +# E.4 Locality + +Locality is often conflated with relativity. The general idea is that tokens far from each other should be independent of one another – i.e. attention should decay as distance grows. This often motivates the definition + +$$ +\lim _ {\left| p _ {i} - p _ {j} \right|\rightarrow \infty} \alpha \left(x _ {i}, x _ {j}, p _ {i}, p _ {j}\right) = 0 \tag {46} +$$ + +for $p_i, p_j \in \mathbb{R}$ and $x_i, x_j \in \mathbb{R}^D$ . However, this definition is both relative and local. We instead define local as, + +$$ +\lim _ {\left| p _ {i} - p _ {0} \right|\rightarrow \infty} \alpha \left(x _ {i}, x _ {j}, p _ {i}, p _ {0}\right) = 0. \tag {47} +$$ + +The difference being that $p_0$ is the origin position. If an embedding is relative, then the origin is arbitrary and can be defined as $p_i$ or $p_j$ . In Press et al. [55], they define the origin vector as the next word. However, they can only do this because of the causal mask. + +In general, the most natural way to measure locality is through the concept of the quantum mechanical concept of the variance of an operator. We will simply use exponential decay, but we point interested readers to Chapter 3 of Griffiths [23]. This formalism works for RoPE as it is a linear transformation and the attention mechanism defines a Hilbert space. + +To be clear, RoPE and LieRE are not local embeddings. This was shown for RoPE in Barbero et al. [5]. Because they are orthogonal matrices, they have unit determinant, which naturally precludes locality. + +# E.5 Other properties + +For completeness, there are two additional assumptions that are common. + +Adjoint symmetry of the Positional Encoding We implicitly assume that the positional encoding is symmetric for the query and key. That is, we assume that the query and key are from the same domain, so the positional encoding has the same representation. More generally, the positional encoding can act differently on the query and key, + +$$ +\alpha \left(\bar {\varphi} \left(x _ {i}, p _ {i}\right), \varphi \left(x _ {j}, p _ {j}\right)\right) = \alpha \left(\varphi \left(x _ {i}, p _ {i}\right), \varphi \left(x _ {j}, p _ {j}\right)\right), \tag {48} +$$ + +where $\bar{\varphi}$ is the positional encoding function for queries. More generally, we can have a relative embedding by letting $\bar{\varphi}$ act on queries differently from the keys. For example, if we let + +$$ +\varphi (x, p) = \exp (\Lambda p) \quad \bar {\varphi} (x, p) = \exp (- \Lambda p), \tag {49} +$$ + +where $\Lambda$ is a diagonal matrix. We end up with, + +$$ +\alpha \left(\bar {\varphi} \left(x _ {i}, p _ {i}\right), \varphi \left(x _ {j}, p _ {j}\right)\right) = \mathbf {q} _ {i} ^ {\top} \exp \left(\Lambda \left(p _ {j} - p _ {i}\right)\right) \mathbf {k} _ {j}, \tag {50} +$$ + +where RoPE can be interpreted as a simple harmonic oscillator, by weakening the symmetry requirement, one could incorporate damping. This can also be used to incorporate graph Laplacian positional encodings into the framework. + +Reversibility Reversibility means that the positional encoding is an injective map – that is, every coordinate is mapped to a unique rotation, thus position can be recovered. This property is important in Liu and Zhou [44] and Su [68] to derive Axial RoPE. While it prevents Eq. 11, it is necessary only for the $D = 1$ case. More generally, Mixed RoPE can learn an injective map for large $D$ . Moreover, while having a “lossless” positional encoding is nice mathematically, its practical utility has yet to be soundly justified, especially if the positional encoding is learnable. + +# F Fast Implementation + +We follow a vectorized implementation for Spherical RoPE similar to the "fast implementation" proposed in Su et al. [69]. + +First, apply the rotation directly on after the other: + +$$ +z _ {d} [ 1 ] = \cos \left(\omega_ {y} p _ {y}\right) z _ {d} [ 1 ] - \sin \left(\omega_ {y} p _ {y}\right) z _ {d} [ 3 ] \tag {51} +$$ + +$$ +z _ {d} [ 3 ] = \sin \left(\omega_ {y} y\right) z _ {d} [ 1 ] + \cos \left(\omega_ {y}\right) z _ {d} [ 3 ], \tag {52} +$$ + +then + +$$ +z _ {d} [ 2 ] = \cos \left(\omega_ {y} p _ {x}\right) z _ {d} [ 2 ] - \sin \left(\omega_ {x} p _ {x}\right) z _ {i} [ 3 ] \tag {53} +$$ + +$$ +z _ {d} [ 3 ] = \sin \left(\omega_ {x} p _ {x}\right) z _ {d} [ 2 ] + \cos \left(\omega_ {x} p _ {x}\right) z _ {d} [ 3 ], \tag {54} +$$ + +where steps 51 and 52 happen simultaneously, and steps 53 and 54 occur at the same time. + +# G Experimental Setup + +Models We use the ViT-S backbone from the timm library [79]. The network always has a depth of 12. We keep $N$ as close to constant across models as we can. For CIFAR100, the embedding dimensions are changed from $64 \times N_{\mathrm{heads}}$ to $60 \times N_{\mathrm{heads}}$ to be compatible with pairs, triplets and quadruples. For ImageNet, we make the embedding dimension $63 \times N_{\mathrm{heads}}$ for Spherical RoPE and $64 \times N_{\mathrm{heads}}$ for other methods. For classification, we use a class token to pool the tokens and predict. Unlike the patch tokens, the class token is not affected by any positional encoding. + +CIFAR100 All experiments on CIFAR100 were performed on one A100 GPUs with a batch size 256. We use a patch size of $4 \times 4$ on the original image size $32 \times 32$ . The training uses heavy regularization and augmentations including dropout, MixUp [87] and CutMix [86]. The models are trained for 400 epochs, taking $\sim 40$ seconds per training loop. + +ImageNet All experiments on ImageNet-1k were performed on four A100 GPUs with a batch size 256. We used cosine learning rate with a learning rate of $3e - 3$ for 200 epochs with 5 epochs of linear warm-up. We used a patch size of $16 \times 16$ on the cropped and resized $224 \times 224$ image after applying 3-Augment [72]. We use the LAMB [84] optimizer. All experiments took $\sim 20$ hrs with $\sim 5$ to 8 minutes to complete a training loop depending on method. + +Positional Encodings For testing with different resolutions, the images from ImageNet's validation set were normalized, resized and cropped. On training, the patches were assigned position $[- \pi, \pi]$ and for evaluation, the patch positions were extrapolated to the range $[- \frac{P}{P_0} \pi, \frac{P}{P_0} \pi]$ . For Learned APE, the positional embeddings are instead interpolated. The fixed frequencies were given by $\omega_d = 1 / 100^{2d / D}$ , where $d$ is the index of the pair/tuple/quadruple. One frequency is shared between both $x$ and $y$ in our implementation of Axial RoPE. + +# H Hyperparameters + +Table 5: Hyperparameters for ImageNet-1K Training + +
CategorySetting
Model Architecture
Patch Size16x16
Heads6
Latent Dimension64 (63 for Spherical) × Heads
Depth12
Pooling[CLS]
Stochastic DepthNo
DropoutNo
LayerScale1
Optimization
OptimizerLAMB [84]
Base Learning Rate4e-3
Weight Decay0.05
Learning Rate ScheduleCosine Decay
Warmup ScheduleLinear
Warmup Epochs5
Epochs200
Batch Size512
Gradient Clipping✓
Precision and Backend
PrecisionMixed (bfloat16)
Backendtorch.autocast
Data Augmentation - Train
CropRandomResizedCrop (192→224)
Flip✓
3-Augment✓
Color Jitter(0.3, 0.3, 0.3, 0.0)
Mixup [87]×
Cutmix [86]×
NormalizationImageNet-1K Statistics
Data Augmentation - Test
ResizeResize → Resolution
CropCenterCrop
NormalizeImageNet-1K Statistics
+ +Table 6: Hyperparameters for CIFAR100 Training + +
CategorySetting
Model Architecture
Patch Size16x16
Heads12
Latent Dimension60 × Heads
Depth12
Pooling[CLS]
Stochastic Depth0.1
Dropout0.1
LayerScale✓
Optimization
OptimizerLAMB [84]
Base Learning Rate4e-3
Weight Decay0.05
Learning Rate ScheduleCosine Decay
Warmup ScheduleLinear
Warmup Epochs5
Epochs400
Batch Size1024
Gradient Clipping✓
Precision and Backend
PrecisionMixed (bfloat16)
Backendtorch.autocast
Data Augmentation - Train
CropRandomResizedCrop (32)
Flip✓
3-Augment✓
Color Jitter(0.3, 0.3, 0.3, 0.0)
Mixup [87]0.8
Cutmix [86]1.0
NormalizationCIFAR Statistics
Data Augmentation - Test
NormalizeCIFAR Statistics
+ +# I Additional Evaluations + +In this section, we include extra evaluations including, basic data scaling, segmentation and speed. We also include additional experiments on the effect of rotation frequencies on Uniform RoPE. + +# I.1 Data Scaling + +Below we evaluate the data scaling of each method. We partition the CIFAR100 dataset into smaller subsets. The number of epochs is scaled, so that the number of training steps is matched on the data subsets. Each model is trained only once on each data split. This experiment tests whether a commutative constraint is beneficial in smaller data regimes as an inductive bias. + +Table 7: Performance on different portions of CIFAR100. + +
Dataset SizeSpherical (Learned)Axial (Learned)MixedUniformAPE
0.256.04 (57.2)55.3 (56.6)56.952.8245.9
0.463.6 (65.34)63.3 (62.5)64.459.753.4
0.667.6 (69.8)66.0 (66.78)70.064.157.7
0.869.8 (72.6)69.9 (69.1)71.665.859.0
+ +Equivalence, in theory, should provide better performance at small scales due to its inductive bias. However, we observe that learned Spherical RoPE performs on-par or better than Mixed RoPE with less parameters. The small gap + +# I.2 Segmentation + +Below we include rudimentary experiments on segmentation to show that the equivalent performance of Spherical RoPE is not caused by the simplicity of classification as a task. For these experiments, we use the models trained on ImageNet-1k as pretrained backbones and fine-tune for Pascal VOC Segmentation [19]. The heads of the models are replaced with a single MLP which is used to get patch logits for each of . Bilinear interpolation is used to create individual pixel logits. + +Table 8: Segmentation results (IoU) on VOC with and without augmentation. + +
SphericalAxial (Learned)MixedUniform
VOC (No Aug.)0.45(0.46)0.42 (0.43)0.440.41
VOC (Simple Aug.)0.498±.007 (0.502±.012)0.474±.011 (0.468±.010)0.502±.0080.461±.012
+ +# I.3 Wall Clock Time + +Below we include the wall clock time for each method. Beyond vectorization as described in Appendix F, no optimizations were made for speed. LieRE was implemented following the pseudo-code in Ostmeier et al. [52]. + +Table 9: Time comparison across different positional encodings + +
Time comparisonSpherical (Learned)Axial (Learned)MixedLieREAPEUniform
Without torch.autocast16.6s (16.6s)16.5s (16.7s)15.7s27.4s13.1s16.5s
With torch.autocast6.7s (5.8s)6.5s (5.7s)5.2s13.6s3.9s6.6s
+ +The experiment was performed by running a dummy input of dimension $(\mathrm{B} = 256, \mathrm{C} = 3, \mathrm{H} = 224, \mathrm{W} = 224)$ 100 times with a ViT backbone on one A100gpu. This is simulated training time, so the rotation matrices were recalculated with every pass for learnable methods. + +Note, Mixed RoPE is faster due to naive the use of naive vector partitioning operations and broadcasting. The main conclusion is that learning parameters and Spherical RoPE cause negligible computational overhead. + +# I.4 Learned Frequencies + +When the frequencies of Spherical RoPE are learned, it is possible for the model to learn equivariance in a particular layer. Like Mixed RoPE, if the rotation frequencies in a layer is set to zero, then the + +![](images/fbec28f7bc10f7bd757197ed88d92a699c9182ac4305d7d814e482668ea271c5.jpg) +Figure 4: The distribution of learned frequencies in each layer of the ViT. Every method tends to learn low frequency positional encodings in the later layers of the network, meaning representations in the later layers are more invariant to position. + +![](images/321551df99e9155805dfa80ddb31e327ab9ad9ecd7c0804da41363d526e0dc71.jpg) + +![](images/ae471b7bfe301589a4f1f849a7906887c2b1d97cddf27cf3f253cee676a1b6c2.jpg) + +![](images/5e953692ed5e1913030807148323359781dc78e7b301fc6e134f1b265f9594d4.jpg) +Figure 5: The scatterplot of learned $\omega_{x}$ and $\omega_{y}$ . Note, though Axial RoPE is plotting $\omega_{x}$ and $\omega_{y}$ together, the rotations will always be axial, so there is no importance to the pairing. + +attention score is position invariant. If one of the rotation frequencies is set to zero, then Spherical RoPE will become trivially equivariant in the remaining direction. This makes it interesting to observe what weights the model learns. Below we show the learned frequencies in each layer of the network after being trained on ImageNet-1k. + +Because frequencies progressively trend toward the axis in deeper layers of the network which makes the positional encodings equivariant in that direction, one could argue that Spherical RoPE learns an equivariant representation in its later layers. However, this same trend can be seen in Mixed RoPE and more notably in Axial RoPE. Because Axial RoPE assumes mutual exclusivity, the frequency pairing is arbitrary. Since we still see the trend toward the axes, the observation that later layers use lower frequencies could be an artifact of backpropagation rather than a necessity for the model to learn an equivariant representation. + +Interestingly, every method has notable clusters at zero frequencies. This suggests that much of the information in the images may be position agnostic. This further explains why setting low frequencies to zero in traditional RoPE improves performance as observed in Barbero et al. [5]. An additional cluster can be observed most notably in the later layers of Spherical and Axial RoPE. We hypothesize this frequency corresponds to some information about the resolution of the image, i.e. the spacing of the grid. Some insight on how to generalize to higher resolutions may come from how this frequency corresponds to training data. + +# J Proofs and Lemmas + +# Axial RoPE Separability + +Proposition 3. Axial RoPE is separable in $x$ and $y$ , that is, the attention score can be decomposed into, + +$$ +\alpha (\mathbf {x} _ {i}, \mathbf {x} _ {j}, \mathbf {p} _ {i}, \mathbf {p} _ {j}) = \alpha_ {i j} ^ {(x)} + \alpha_ {i j} ^ {(y)} +$$ + +Proof. Suppose we define the dot-product attention score as + +$$ +\alpha (\mathbf {q}, \mathbf {k}) = \mathbf {q} ^ {\top} \mathbf {k}. +$$ + +We incorporate Axial Rotary Positional Embeddings by rotating each 2-dimensional subvector of the query (and likewise the key). Concretely, if the hidden dimension is $2n$ , we partition + +$$ +\mathbf {q} = \left[ \mathbf {q} _ {x, 1}, \mathbf {q} _ {y, 1}, \dots , \mathbf {q} _ {x, n}, \mathbf {q} _ {y, n} \right] ^ {\top}, \quad \mathbf {k} = \left[ \mathbf {k} _ {x, 1}, \mathbf {k} _ {y, 1}, \dots , \mathbf {k} _ {x, n}, \mathbf {k} _ {y, n} \right] ^ {\top}, \tag {55} +$$ + +where each $\mathbf{q}_{x,d},\mathbf{q}_{y,d},\mathbf{k}_{x,d},\mathbf{k}_{y,d}\in \mathbb{R}^2$ . At spatial location $\mathbf{p} = (p_x,p_y)$ , we apply rotations + +$$ +\mathbf {q} _ {x, d} ^ {\prime} = \mathbf {R} \left(\omega_ {d} p _ {x}\right) \mathbf {q} _ {x, d}, \quad \mathbf {q} _ {y, d} ^ {\prime} = \mathbf {R} \left(\omega_ {d} p _ {y}\right) \mathbf {q} _ {y, d}, +$$ + +and similarly for $\mathbf{k}$ . Here $\mathbf{R}(\theta) \in \mathbb{R}^{2 \times 2}$ is the planar rotation by angle $\theta$ . + +For tokens at positions $\mathbf{p}_i = (p_{i,x},p_{i,y})$ and $\mathbf{p}_j = (p_{j,x},p_{j,y})$ , their rotated queries and keys yield + +$$ +\alpha_ {i j} = \sum_ {d = 1} ^ {n} \left[ \left(\mathbf {q} _ {x, d}\right) ^ {\top} \mathbf {R} \left(\omega_ {d} \left(p _ {j, x} - p _ {i, x}\right)\right) \mathbf {k} _ {x, d} + \left(\mathbf {q} _ {y, d}\right) ^ {\top} \mathbf {R} \left(\omega_ {d} \left(p _ {j, y} - p _ {i, y}\right)\right) \mathbf {k} _ {y, d} \right]. +$$ + +Define the horizontal and vertical components by + +$$ +\alpha_ {i j} ^ {(x)} := \sum_ {d = 1} ^ {n} (\mathbf {q} _ {x, d}) ^ {\top} \mathbf {R} \left(\omega_ {d} \left(p _ {j, x} - p _ {i, x}\right)\right) \mathbf {k} _ {x, d}, \quad \alpha_ {i j} ^ {(y)} := \sum_ {d = 1} ^ {n} (\mathbf {q} _ {y, d}) ^ {\top} \mathbf {R} \left(\omega_ {d} \left(p _ {j, y} - p _ {i, y}\right)\right) \mathbf {k} _ {y, d}. +$$ + +Hence the total attention decomposes additively: + +$$ +\alpha_ {i j} = \alpha_ {i j} ^ {(x)} + \alpha_ {i j} ^ {(y)}, +$$ + +demonstrating that axial rotary embeddings factorize the positional dependence along each axis. + +![](images/a7a472f58339f9cfe5fa1c9aca42800a318281d37d1617c4a6e74ff551dbbc69.jpg) + +Matrix Exponentiation Computing the matrix exponential by exponentiating the eigenvalues is a common result in linear algebra and numerics, however we provide it here for those unfamiliar. + +Lemma 1. Let $\mathbf{A}$ be a diagonalizable matrix $\mathbf{A} = \mathbf{U}\boldsymbol{\Lambda}\mathbf{U}^{-1}$ , then the matrix exponential of $\mathbf{A}$ is given by + +$$ +\exp (\mathbf {A}) = \mathbf {U} \exp (\boldsymbol {\Lambda}) \mathbf {U} ^ {- 1} +$$ + +Proof. + +Recall the power-series definition of the matrix exponential: + +$$ +\exp (\mathbf {A}) = \sum_ {k = 0} ^ {\infty} \frac {1}{k !} \mathbf {A} ^ {k}. \tag {56} +$$ + +Since $\mathbf{A}$ is diagonalizable, + +$$ +\mathbf {A} ^ {k} = \left(\mathbf {U} \boldsymbol {\Lambda} \mathbf {U} ^ {- 1}\right) ^ {k} = \mathbf {U} \boldsymbol {\Lambda} ^ {k} \mathbf {U} ^ {- 1}. \tag {57} +$$ + +Substituting into the series gives + +$$ +\exp (\mathbf {A}) = \sum_ {k = 0} ^ {\infty} \frac {1}{k !} \left(\mathbf {U} \boldsymbol {\Lambda} ^ {k} \mathbf {U} ^ {- 1}\right) = \mathbf {U} \left(\sum_ {k = 0} ^ {\infty} \frac {1}{k !} \boldsymbol {\Lambda} ^ {k}\right) \mathbf {U} ^ {- 1}. \tag {58} +$$ + +Because $\pmb{\Lambda}$ is diagonal, the series $\sum_{k=0}^{\infty} \frac{1}{k!} \pmb{\Lambda}^{k}$ is itself the diagonal matrix of scalar exponentials, + +$$ +\exp (\boldsymbol {\Lambda}) = \operatorname {d i a g} \left(e ^ {\lambda_ {1}}, \dots , e ^ {\lambda_ {n}}\right). \tag {59} +$$ + +Hence is well defined, and + +$$ +\exp (\mathbf {A}) = \mathbf {U} \exp (\boldsymbol {\Lambda}) \mathbf {U} ^ {- 1}. \tag {60} +$$ + +![](images/ca8d935b21e9100a1479efef137ccd77074725f92c52f62fb5385817f72e8650.jpg) + +Simultaneous-Diagonalizability The proof that two (diagonalizable) matrixes are simultaneous-diagonalizability if and only if they are commutative is also a standard result. However, we once again provide it here: + +Lemma 2. Let $\mathcal{A}_x$ and $\mathcal{A}_y$ be skew-symmetric. Then $\mathcal{A}_x$ and $\mathcal{A}_y$ are simultaneously diagonalizable if and only if $\mathcal{A}_x\mathcal{A}_y = \mathcal{A}_y\mathcal{A}_x$ . + +# Proof. + +Suppose $\mathcal{A}_x$ and $\mathcal{A}_y$ are simultaneously diagonalizable. Then, because they are skew-symmetric, there exists a unitary matrix $\mathbf{U}$ such that + +$$ +\mathbf {U} \Lambda_ {x} \mathbf {U} ^ {\top} = \mathcal {A} _ {x} \quad \text {a n d} \quad \mathbf {U} \Lambda_ {y} \mathbf {U} ^ {\top} = \mathcal {A} _ {y}, \tag {61} +$$ + +where $\Lambda_{x}$ and $\Lambda_{y}$ are diagonal matrices. + +Then, + +$$ +\mathcal {A} _ {x} \mathcal {A} _ {y} = \mathbf {U} \Lambda_ {x} \mathbf {U} ^ {\top} \mathbf {U} \Lambda_ {y} \mathbf {U} ^ {\top} = \mathbf {U} \Lambda_ {x} \Lambda_ {y} \mathbf {U} ^ {\top} = \mathbf {U} \Lambda_ {y} \Lambda_ {x} \mathbf {U} ^ {\top} = \mathcal {A} _ {x} \mathcal {A} _ {y} \tag {62} +$$ + +Hence, $\mathcal{A}_x$ and $\mathcal{A}_y$ commute. + +Now suppose $\mathcal{A}_x$ and $\mathcal{A}_y$ commute, $\mathcal{A}_x\mathcal{A}_y = \mathcal{A}_y\mathcal{A}_x$ . Since $\mathcal{A}_x$ and $\mathcal{A}_y$ are skew-symmetric, they are diagonalizable in $\mathbb{C}^{D_xD}$ , thus there exists a basis of eigenvectors of $\mathcal{A}_x$ . Because $\mathcal{A}_y$ commutes with $\mathcal{A}_x$ , the eigenspaces of $\mathcal{A}_x$ are invariant under $\mathcal{A}_y$ . That is, for any eigenvalue $\lambda$ of $\mathcal{A}_x$ , the corresponding eigenspace + +$$ +E _ {\lambda} = \{v \in \mathbb {C} ^ {D}: \mathcal {A} _ {x} v = \lambda v \} \tag {63} +$$ + +is $\mathcal{A}_y$ -invariant: if $v \in E_{\lambda}$ , then + +$$ +\mathcal {A} _ {x} \left(\mathcal {A} _ {y} v\right) = \mathcal {A} _ {y} \left(\mathcal {A} _ {x} v\right) = \mathcal {A} _ {y} (\lambda v) = \lambda \mathcal {A} _ {y} v \Rightarrow \mathcal {A} _ {y} v \in E _ {\lambda}. \tag {64} +$$ + +Now, restrict $\mathcal{A}_x$ to each eigenspace $E_{\lambda}$ . Since $\mathbb{C}$ is algebraically closed and $\mathcal{A}_y|_{E_\lambda}$ is a linear operator on a finite-dimensional space, $\mathcal{A}_y$ is diagonalizable on $E_{\lambda}$ . Thus, we can choose a basis of eigenvectors for $\mathcal{A}_y$ in each $E_{\lambda}$ . + +Putting these together, we get a basis for $\mathbb{C}^N$ consisting of vectors that are eigenvectors for both $\mathcal{A}_x$ and $\mathcal{A}_y$ . Therefore, $\mathcal{A}_x$ and $\mathcal{A}_y$ are simultaneously diagonalizable. + +![](images/06f33c5e86293468db96adb621def35b913c0b15924a805a36ace8431e56f565.jpg) + +1-D LieRE is equivalent to RoPE In this section, we will more formally prove that the traditional RoPE with learned rotation frequencies is equivalent to 1-D RoPE as proposed in Section 3. + +Proposition 1. Any $D$ -dimensional rotation can be parameterized by RoPE with learned frequencies. + +# Proof. + +We define a rotation to be an orthogonal matrix with positive determinant; that is, it is an element of $\mathbf{R} \in \mathrm{SO}(N)$ . We can write any element of $\mathrm{SO}(N)$ via the exponential map $\mathbf{R} = e^{\mathcal{A}}$ where $\mathcal{A} \in \mathfrak{so}(N)$ , i.e. $\mathcal{A}$ is a skew-symmetric matrix. It is well-known that the eigenvalues of a real, skew-symmetric matrix are purely imaginary (or zero), and such a matrix is unitarily (i.e. orthogonally) diagonalizable over $\mathbb{C}$ , resulting in a spectral decomposition with a purely imaginary eigenvalue matrix. Thus, + +$$ +\mathcal {A} = \mathbf {U} \boldsymbol {\Lambda} i \mathbf {U} ^ {\dagger} \tag {65} +$$ + +and, by Lemma 1, + +$$ +\exp (\mathcal {A}) = \mathbf {U} \exp (\boldsymbol {\Lambda} i) \mathbf {U} ^ {\dagger}. \tag {66} +$$ + +where, because $\Lambda$ is diagonal, $\exp (\Lambda)$ is simply the scalar-exponential of each element. The positional encoding of a token to a query can be written as, + +$$ +\varphi (\mathbf {x}, p) = \exp (\mathcal {A} p) \mathbf {W} _ {q} \mathbf {x} = \mathbf {U} \exp (\boldsymbol {\Lambda} i p) \mathbf {W} _ {\mathbf {q}} ^ {\prime} \mathbf {x} \tag {67} +$$ + +where $\mathbf{W}_q^{\prime} = \mathbf{W}_q\mathbf{U}$ . We assume the same encoding for the key with a different matrix, $\mathbf{W}_k^{\prime}$ and the same generator, $\mathcal{A}$ . This equation can be rewritten as $\varphi (\mathbf{x},p) = \mathbf{U}R o P E(\mathbf{x},p)$ by Eq.10. If the attention score is given by $\alpha (\mathbf{q},\mathbf{k}) = \mathbf{q}^{\dagger}\mathbf{k}$ , where $\dagger$ denotes the Hermitian transpose, then the attention score can be expanded into, + +$$ +\begin{array}{l} \alpha \left(\mathbf {x} _ {i}, \mathbf {x} _ {j}, p _ {i}, p _ {j}\right) = R o P E \left(\mathbf {x} _ {i}, p _ {i}\right) ^ {\dagger} \mathbf {U} ^ {\dagger} \mathbf {U} R o P E \left(\mathbf {x} _ {j}, p _ {j}\right) (68) \\ = R o P E \left(\mathbf {x} _ {i}, p _ {i}\right) ^ {\dagger} R o P E \left(\mathbf {x} _ {j}, p _ {j}\right). (69) \\ \end{array} +$$ + +Hence, any LieRE of one generator can be expressed as RoPE with learned rotation frequencies. + +Any commutative LieRE is equivalent to Mixed RoPE We now prove that multi-dimensional LieRE with commutative generators generalizes directly to Mixed RoPE. + +Proposition 2. Any $M$ -dimensional LieRE with commutative generators can be parameterized by Mixed RoPE. + +# Proof. + +Let $\mathcal{A}_1, \ldots, \mathcal{A}_M \subset \mathfrak{so}(N)$ be skew-symmetric generators such that $[\mathcal{A}_m, \mathcal{A}_n] = \mathbf{0}$ for all $m, n$ . By Lemma 2, commuting normal matrices are simultaneously unitarily diagonalizable. Thus, there exists a unitary $\mathbf{U}$ and diagonal matrices $\Lambda_1, \ldots, \Lambda_M$ such that + +$$ +\mathcal {A} _ {m} = \mathbf {U} \boldsymbol {\Lambda} _ {m} i \mathbf {U} ^ {\dagger} \quad \text {f o r a l l} m = 1, \dots , M. \tag {70} +$$ + +For a position vector $\mathbf{p} = (p_1, \ldots, p_M) \in \mathbb{R}^M$ , the LieRE positional encoding is + +$$ +\operatorname {L i e R E} (\mathbf {x}, \mathbf {p}) = \exp \left(\sum_ {m = 1} ^ {M} \mathcal {A} _ {m} p _ {m}\right) \mathbf {W} q \mathbf {x}, \tag {71} +$$ + +which, using Lemmas 1 and 2, can be written as + +$$ +\operatorname {L i e R E} (\mathbf {x}, \mathbf {p}) = \mathbf {U} \exp \left(\sum_ {m = 1} ^ {M} \boldsymbol {\Lambda} _ {m} i, p _ {m}\right) \mathbf {U} ^ {\dagger} \mathbf {W} _ {q} \mathbf {x}. \tag {72} +$$ + +Let $\mathbf{W}_q^{\prime} = \mathbf{U}^{\dagger}\mathbf{W}_{q}$ . Then + +$$ +\operatorname {L i e R E} (\mathbf {x}, \mathbf {p}) = \operatorname {U M i x e d R o P E} (\mathbf {x}, \mathbf {p}), \tag {73} +$$ + +where MixedRoPE applies elementwise complex rotations + +$$ +e ^ {i \left(\lambda_ {1} ^ {(k)} p _ {1} + \dots + \lambda_ {M} ^ {(k)} p _ {M}\right)} \tag {74} +$$ + +to each channel $k$ , with frequencies $\lambda_m^{(k)}$ learned from $\Lambda_{m}$ . + +If the attention score is given by $\alpha (\mathbf{q},\mathbf{k}) = \mathbf{q}^{\dagger}\mathbf{k}$ , then + +$$ +\begin{array}{l} \alpha \left(\mathbf {x} _ {i}, \mathbf {x} _ {j}, \mathbf {p} _ {i}, \mathbf {p} _ {j}\right) = \operatorname {M i x e d R o P E} \left(\mathbf {x} _ {i}, \mathbf {p} _ {i}\right) ^ {\dagger} \mathbf {U} ^ {\dagger} \mathbf {U M i x e d R o P E} \left(\mathbf {x} _ {j}, \mathbf {p} _ {j}\right) (75) \\ = \operatorname {M i x e d R o P E} \left(\mathbf {x} _ {i}, \mathbf {p} _ {i}\right) ^ {\dagger} \operatorname {M i x e d R o P E} \left(\mathbf {x} _ {j} \mathbf {p} _ {j}\right). (76) \\ \end{array} +$$ + +Hence, any $M$ -dimensional LieRE with commutative generators is equivalent to a Mixed RoPE parameterization with learned rotation frequencies. $\square$ + +# NeurIPS Paper Checklist + +# 1. Claims + +Question: Do the main claims made in the abstract and introduction accurately reflect the paper's contributions and scope? + +Answer: [Yes] + +Justification: We provide proofs and theoretical evidence on benchmarks. + +Guidelines: + +- The answer NA means that the abstract and introduction do not include the claims made in the paper. +- The abstract and/or introduction should clearly state the claims made, including the contributions made in the paper and important assumptions and limitations. A No or NA answer to this question will not be perceived well by the reviewers. +- The claims made should match theoretical and experimental results, and reflect how much the results can be expected to generalize to other settings. +- It is fine to include aspirational goals as motivation as long as it is clear that these goals are not attained by the paper. + +# 2. Limitations + +Question: Does the paper discuss the limitations of the work performed by the authors? + +Answer: [Yes] + +Justification: We emphasize that our conclusions are limited to vision and have a limitations section in the Appendix. + +Guidelines: + +- The answer NA means that the paper has no limitation while the answer No means that the paper has limitations, but those are not discussed in the paper. +- The authors are encouraged to create a separate "Limitations" section in their paper. +- The paper should point out any strong assumptions and how robust the results are to violations of these assumptions (e.g., independence assumptions, noiseless settings, model well-specification, asymptotic approximations only holding locally). The authors should reflect on how these assumptions might be violated in practice and what the implications would be. +- The authors should reflect on the scope of the claims made, e.g., if the approach was only tested on a few datasets or with a few runs. In general, empirical results often depend on implicit assumptions, which should be articulated. +- The authors should reflect on the factors that influence the performance of the approach. For example, a facial recognition algorithm may perform poorly when image resolution is low or images are taken in low lighting. Or a speech-to-text system might not be used reliably to provide closed captions for online lectures because it fails to handle technical jargon. +- The authors should discuss the computational efficiency of the proposed algorithms and how they scale with dataset size. +- If applicable, the authors should discuss possible limitations of their approach to address problems of privacy and fairness. +- While the authors might fear that complete honesty about limitations might be used by reviewers as grounds for rejection, a worse outcome might be that reviewers discover limitations that aren't acknowledged in the paper. The authors should use their best judgment and recognize that individual actions in favor of transparency play an important role in developing norms that preserve the integrity of the community. Reviewers will be specifically instructed to not penalize honesty concerning limitations. + +# 3. Theory assumptions and proofs + +Question: For each theoretical result, does the paper provide the full set of assumptions and a complete (and correct) proof? + +Answer: [Yes] + +Justification: While we do not formally list the assumptions, we implicitly make assumptions on positional encoding through assuming $N$ -D LieRE. + +# Guidelines: + +- The answer NA means that the paper does not include theoretical results. +- All the theorems, formulas, and proofs in the paper should be numbered and cross-referenced. +- All assumptions should be clearly stated or referenced in the statement of any theorems. +- The proofs can either appear in the main paper or the supplemental material, but if they appear in the supplemental material, the authors are encouraged to provide a short proof sketch to provide intuition. +- Inversely, any informal proof provided in the core of the paper should be complemented by formal proofs provided in appendix or supplemental material. +- Theorems and Lemmas that the proof relies upon should be properly referenced. + +# 4. Experimental result reproducibility + +Question: Does the paper fully disclose all the information needed to reproduce the main experimental results of the paper to the extent that it affects the main claims and/or conclusions of the paper (regardless of whether the code and data are provided or not)? + +Answer: [Yes] + +Justification: We do our best to provide hyper-parameters for reproducing our results. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- If the paper includes experiments, a No answer to this question will not be perceived well by the reviewers: Making the paper reproducible is important, regardless of whether the code and data are provided or not. +- If the contribution is a dataset and/or model, the authors should describe the steps taken to make their results reproducible or verifiable. +- Depending on the contribution, reproducibility can be accomplished in various ways. For example, if the contribution is a novel architecture, describing the architecture fully might suffice, or if the contribution is a specific model and empirical evaluation, it may be necessary to either make it possible for others to replicate the model with the same dataset, or provide access to the model. In general, releasing code and data is often one good way to accomplish this, but reproducibility can also be provided via detailed instructions for how to replicate the results, access to a hosted model (e.g., in the case of a large language model), releasing of a model checkpoint, or other means that are appropriate to the research performed. +- While NeurIPS does not require releasing code, the conference does require all submissions to provide some reasonable avenue for reproducibility, which may depend on the nature of the contribution. For example + +(a) If the contribution is primarily a new algorithm, the paper should make it clear how to reproduce that algorithm. +(b) If the contribution is primarily a new model architecture, the paper should describe the architecture clearly and fully. +(c) If the contribution is a new model (e.g., a large language model), then there should either be a way to access this model for reproducing the results or a way to reproduce the model (e.g., with an open-source dataset or instructions for how to construct the dataset). +(d) We recognize that reproducibility may be tricky in some cases, in which case authors are welcome to describe the particular way they provide for reproducibility. In the case of closed-source models, it may be that access to the model is limited in some way (e.g., to registered users), but it should be possible for other researchers to have some path to reproducing or verifying the results. + +# 5. Open access to data and code + +Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material? + +Answer: [Yes] + +Justification: We intend to make the code public. + +Guidelines: + +- The answer NA means that paper does not include experiments requiring code. +- Please see the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details. +- While we encourage the release of code and data, we understand that this might not be possible, so "No" is an acceptable answer. Papers cannot be rejected simply for not including code, unless this is central to the contribution (e.g., for a new open-source benchmark). +- The instructions should contain the exact command and environment needed to run to reproduce the results. See the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details. +- The authors should provide instructions on data access and preparation, including how to access the raw data, preprocessed data, intermediate data, and generated data, etc. +- The authors should provide scripts to reproduce all experimental results for the new proposed method and baselines. If only a subset of experiments are reproducible, they should state which ones are omitted from the script and why. +- At submission time, to preserve anonymity, the authors should release anonymized versions (if applicable). +- Providing as much information as possible in supplemental material (appended to the paper) is recommended, but including URLs to data and code is permitted. + +# 6. Experimental setting/details + +Question: Does the paper specify all the training and test details (e.g., data splits, hyperparameters, how they were chosen, type of optimizer, etc.) necessary to understand the results? + +Answer: [Yes] + +Justification: We provide them to the best of our ability. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The experimental setting should be presented in the core of the paper to a level of detail that is necessary to appreciate the results and make sense of them. +- The full details can be provided either with the code, in appendix, or as supplemental material. + +# 7. Experiment statistical significance + +Question: Does the paper report error bars suitably and correctly defined or other appropriate information about the statistical significance of the experiments? + +Answer: [Yes] + +Justification: We trained models from several random seeds. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The authors should answer "Yes" if the results are accompanied by error bars, confidence intervals, or statistical significance tests, at least for the experiments that support the main claims of the paper. +- The factors of variability that the error bars are capturing should be clearly stated (for example, train/test split, initialization, random drawing of some parameter, or overall run with given experimental conditions). +- The method for calculating the error bars should be explained (closed form formula, call to a library function, bootstrap, etc.) +- The assumptions made should be given (e.g., Normally distributed errors). +- It should be clear whether the error bar is the standard deviation or the standard error of the mean. + +- It is OK to report 1-sigma error bars, but one should state it. The authors should preferably report a 2-sigma error bar than state that they have a $96\%$ CI, if the hypothesis of Normality of errors is not verified. +- For asymmetric distributions, the authors should be careful not to show in tables or figures symmetric error bars that would yield results that are out of range (e.g. negative error rates). +- If error bars are reported in tables or plots, The authors should explain in the text how they were calculated and reference the corresponding figures or tables in the text. + +# 8. Experiments compute resources + +Question: For each experiment, does the paper provide sufficient information on the computer resources (type of compute workers, memory, time of execution) needed to reproduce the experiments? + +Answer: [Yes] + +Justification: We provide basic information about the GPUs used. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The paper should indicate the type of compute workers CPU or GPU, internal cluster, or cloud provider, including relevant memory and storage. +- The paper should provide the amount of compute required for each of the individual experimental runs as well as estimate the total compute. +- The paper should disclose whether the full research project required more compute than the experiments reported in the paper (e.g., preliminary or failed experiments that didn't make it into the paper). + +# 9. Code of ethics + +Question: Does the research conducted in the paper conform, in every respect, with the NeurIPS Code of Ethics https://neurips.cc/public/EthicsGuidelines? + +Answer: [Yes] + +Justification: We do not believe there is any violations. + +Guidelines: + +- The answer NA means that the authors have not reviewed the NeurIPS Code of Ethics. +- If the authors answer No, they should explain the special circumstances that require a deviation from the Code of Ethics. +- The authors should make sure to preserve anonymity (e.g., if there is a special consideration due to laws or regulations in their jurisdiction). + +# 10. Broader impacts + +Question: Does the paper discuss both potential positive societal impacts and negative societal impacts of the work performed? + +Answer: [Yes] + +Justification: We include a section in the appendix, however, it is mostly not applicable for our paper. + +Guidelines: + +- The answer NA means that there is no societal impact of the work performed. +- If the authors answer NA or No, they should explain why their work has no societal impact or why the paper does not address societal impact. +- Examples of negative societal impacts include potential malicious or unintended uses (e.g., disinformation, generating fake profiles, surveillance), fairness considerations (e.g., deployment of technologies that could make decisions that unfairly impact specific groups), privacy considerations, and security considerations. + +- The conference expects that many papers will be foundational research and not tied to particular applications, let alone deployments. However, if there is a direct path to any negative applications, the authors should point it out. For example, it is legitimate to point out that an improvement in the quality of generative models could be used to generate deepfakes for disinformation. On the other hand, it is not needed to point out that a generic algorithm for optimizing neural networks could enable people to train models that generate Deepfakes faster. +- The authors should consider possible harms that could arise when the technology is being used as intended and functioning correctly, harms that could arise when the technology is being used as intended but gives incorrect results, and harms following from (intentional or unintentional) misuse of the technology. +- If there are negative societal impacts, the authors could also discuss possible mitigation strategies (e.g., gated release of models, providing defenses in addition to attacks, mechanisms for monitoring misuse, mechanisms to monitor how a system learns from feedback over time, improving the efficiency and accessibility of ML). + +# 11. Safeguards + +Question: Does the paper describe safeguards that have been put in place for responsible release of data or models that have a high risk for misuse (e.g., pretrained language models, image generators, or scraped datasets)? + +Answer: [NA] + +Justification: Our paper is more theoretical. + +Guidelines: + +- The answer NA means that the paper poses no such risks. +- Released models that have a high risk for misuse or dual-use should be released with necessary safeguards to allow for controlled use of the model, for example by requiring that users adhere to usage guidelines or restrictions to access the model or implementing safety filters. +- Datasets that have been scraped from the Internet could pose safety risks. The authors should describe how they avoided releasing unsafe images. +- We recognize that providing effective safeguards is challenging, and many papers do not require this, but we encourage authors to take this into account and make a best faith effort. + +# 12. Licenses for existing assets + +Question: Are the creators or original owners of assets (e.g., code, data, models), used in the paper, properly credited and are the license and terms of use explicitly mentioned and properly respected? + +Answer: [NA] + +Justification: We cite libraries used and datasets, however they are standard libraries and benchmarks. There are no other specialized assets used. + +Guidelines: + +- The answer NA means that the paper does not use existing assets. +- The authors should cite the original paper that produced the code package or dataset. +- The authors should state which version of the asset is used and, if possible, include a URL. +- The name of the license (e.g., CC-BY 4.0) should be included for each asset. +- For scraped data from a particular source (e.g., website), the copyright and terms of service of that source should be provided. +- If assets are released, the license, copyright information, and terms of use in the package should be provided. For popular datasets, paperswithcode.com/datasets has curated licenses for some datasets. Their licensing guide can help determine the license of a dataset. +- For existing datasets that are re-packaged, both the original license and the license of the derived asset (if it has changed) should be provided. + +- If this information is not available online, the authors are encouraged to reach out to the asset's creators. + +# 13. New assets + +Question: Are new assets introduced in the paper well documented and is the documentation provided alongside the assets? + +Answer: [NA] + +Justification: + +Guidelines: + +- The answer NA means that the paper does not release new assets. +- Researchers should communicate the details of the dataset/code/model as part of their submissions via structured templates. This includes details about training, license, limitations, etc. +- The paper should discuss whether and how consent was obtained from people whose asset is used. +- At submission time, remember to anonymize your assets (if applicable). You can either create an anonymized URL or include an anonymized zip file. + +# 14. Crowdsourcing and research with human subjects + +Question: For crowdsourcing experiments and research with human subjects, does the paper include the full text of instructions given to participants and screenshots, if applicable, as well as details about compensation (if any)? + +Answer: [NA] + +Justification: + +Guidelines: + +- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. +- Including this information in the supplemental material is fine, but if the main contribution of the paper involves human subjects, then as much detail as possible should be included in the main paper. +- According to the NeurIPS Code of Ethics, workers involved in data collection, curation, or other labor should be paid at least the minimum wage in the country of the data collector. + +# 15. Institutional review board (IRB) approvals or equivalent for research with human subjects + +Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals (or an equivalent approval/review based on the requirements of your country or institution) were obtained? + +Answer: [NA] + +Justification: + +Guidelines: + +- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. +- Depending on the country in which research is conducted, IRB approval (or equivalent) may be required for any human subjects research. If you obtained IRB approval, you should clearly state this in the paper. +- We recognize that the procedures for this may vary significantly between institutions and locations, and we expect authors to adhere to the NeurIPS Code of Ethics and the guidelines for their institution. +- For initial submissions, do not include any information that would break anonymity (if applicable), such as the institution conducting the review. + +# 16. Declaration of LLM usage + +Question: Does the paper describe the usage of LLMs if it is an important, original, or non-standard component of the core methods in this research? Note that if the LLM is used only for writing, editing, or formatting purposes and does not impact the core methodology, scientific rigorousness, or originality of the research, declaration is not required. + +Answer: [NA] + +Justification: + +Guidelines: + +- The answer NA means that the core method development in this research does not involve LLMs as any important, original, or non-standard components. +- Please refer to our LLM policy (https://neurips.cc/Conferences/2025/LLM) for what should or should not be described. \ No newline at end of file diff --git a/NeurIPS/2025/A Circular Argument_ Does RoPE need to be Equivariant for Vision_/images.zip b/NeurIPS/2025/A Circular Argument_ Does RoPE need to be Equivariant for Vision_/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..0dce0403c472baea6a6418dfd65990775bfb74bc --- /dev/null +++ b/NeurIPS/2025/A Circular Argument_ Does RoPE need to be Equivariant for Vision_/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3eb1ea176aee04f809ea76c562e46233b82fcd0683d86154ccf96158c5689cf8 +size 1066256 diff --git a/NeurIPS/2025/A Circular Argument_ Does RoPE need to be Equivariant for Vision_/layout.json b/NeurIPS/2025/A Circular Argument_ Does RoPE need to be Equivariant for Vision_/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..edce273210c03fa51b885e83aba967413a22b1ce --- /dev/null +++ b/NeurIPS/2025/A Circular Argument_ Does RoPE need to be Equivariant for Vision_/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6d6fad449c6ec79196987193dce340d66fa9d8553a40c64e5422e3b647e449fd +size 1246389 diff --git a/NeurIPS/2025/A Clean Slate for Offline Reinforcement Learning/ed648d0c-14ca-4d08-893f-aaf5c791f438_content_list.json b/NeurIPS/2025/A Clean Slate for Offline Reinforcement Learning/ed648d0c-14ca-4d08-893f-aaf5c791f438_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..14c01c037cde5c098fe29dae23766776685d7d54 --- /dev/null +++ b/NeurIPS/2025/A Clean Slate for Offline Reinforcement Learning/ed648d0c-14ca-4d08-893f-aaf5c791f438_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:da457d4bd3b59c47b6c91a7a7fb3d9813a5b2cc6045031e3f683390bb23a19da +size 178308 diff --git a/NeurIPS/2025/A Clean Slate for Offline Reinforcement Learning/ed648d0c-14ca-4d08-893f-aaf5c791f438_model.json b/NeurIPS/2025/A Clean Slate for Offline Reinforcement Learning/ed648d0c-14ca-4d08-893f-aaf5c791f438_model.json new file mode 100644 index 0000000000000000000000000000000000000000..47d5439c7e77843347180bc5df7575c701465080 --- /dev/null +++ b/NeurIPS/2025/A Clean Slate for Offline Reinforcement Learning/ed648d0c-14ca-4d08-893f-aaf5c791f438_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:366fa46309feb961f1337c7b468fc8d18770c60d01fbdc658806d46adebe92e5 +size 219178 diff --git a/NeurIPS/2025/A Clean Slate for Offline Reinforcement Learning/ed648d0c-14ca-4d08-893f-aaf5c791f438_origin.pdf b/NeurIPS/2025/A Clean Slate for Offline Reinforcement Learning/ed648d0c-14ca-4d08-893f-aaf5c791f438_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..cffbc493fa950a6e61063cd04b02a216d383ab0b --- /dev/null +++ b/NeurIPS/2025/A Clean Slate for Offline Reinforcement Learning/ed648d0c-14ca-4d08-893f-aaf5c791f438_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2b8ecdaa8c8880e86a2231afea3889614ec4334afb6133937c95bccda12a3592 +size 11648667 diff --git a/NeurIPS/2025/A Clean Slate for Offline Reinforcement Learning/full.md b/NeurIPS/2025/A Clean Slate for Offline Reinforcement Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..9e5ce8bbff742fed30e05aa56100cf93f4ae96c2 --- /dev/null +++ b/NeurIPS/2025/A Clean Slate for Offline Reinforcement Learning/full.md @@ -0,0 +1,853 @@ +# A Clean Slate for Offline RL + +Matthew T. Jackson* Uljad Berdica* Jarek Liesen* + +Shimon Whiteson Jakob N. Foerster + +University of Oxford + +{jackson,uljadb,jarek}@robots.ox.ac.uk + +# Abstract + +Progress in offline reinforcement learning (RL) has been impeded by ambiguous problem definitions and entangled algorithmic designs, resulting in inconsistent implementations, insufficient ablations, and unfair evaluations. Although offline RL explicitly avoids environment interaction, prior methods frequently employ extensive, undocumented online evaluation for hyperparameter tuning, complicating method comparisons. Moreover, existing reference implementations differ significantly in boilerplate code, obscuring their core algorithmic contributions. We address these challenges by first introducing a rigorous taxonomy and a transparent evaluation procedure that explicitly quantifies online tuning budgets. To resolve opaque algorithmic design, we provide clean, minimalistic, single-file implementations of various model-free and model-based offline RL methods, significantly enhancing clarity and achieving substantial speed-ups. Leveraging these streamlined implementations, we propose Unifloral, a unified algorithm that encapsulates diverse prior approaches within a single, comprehensive hyperparameter space, enabling algorithm development in a shared hyperparameter space. Using Unifloral with our rigorous evaluation procedure, we develop two novel algorithms—TD3-AWR (model-free) and MoBRAC (model-based)—which substantially outperform established baselines. All code for this project can be found in our public codebase. + +# 1 Introduction + +Offline reinforcement learning (RL)—the task of learning effective policies from pre-collected, static datasets—is critical for applying RL in real-world settings where online experimentation is expensive or risky. Despite significant interest [1-5], the field has struggled to converge on clear, actionable insights. Algorithms and methods proliferate rapidly but no broadly agreed-upon conclusions or standardized benchmarks have emerged [6]. This undermines both practical application and theoretical progress. In this work, we identify and address two primary problems that contribute to stagnation and confusion in offline RL research: an ambiguous problem setting and opaque algorithmic design. + +Problem 1: Ambiguous Problem Setting Recent work in offline RL has lacked a rigorously articulated definition or standardized evaluation procedure. The broad mission statement, learning from a static dataset without direct environment interaction, is prone to misinterpretation that skews proposed methods towards impractical evaluation practices. Existing literature implicitly relaxes various definitions concerning critical details such as hyperparameter tuning allowances [4, 7], the extent of post-deployment policy adaptation [8], and the specifics of evaluation procedures [9]. Consequently, comparisons between methods are confounded as each study might assume fundamentally different + +experimental conditions. While some approaches restrict tuning based on related dataset performance [10], most approaches extensively tune hyperparameters on the target environment [11, 5, 12]. Using the target environment to tune hyperparameters needs a large number of online evaluations, which is in conflict with the basic premise of offline RL. + +Solution 1: A Novel Taxonomy and Evaluation Procedure We first introduce a rigorous and explicit taxonomy of offline RL evaluation variants (Section 3.1) and specify the one we find to be implicitly adopted by most prior research. To facilitate consistent and transparent evaluation, we propose a rigorous procedure for this setting (Section 3.2) that evaluates algorithmic performance using a fixed hyperparameter range across multiple datasets. This procedure explicitly quantifies performance at various permissible levels of online hyperparameter tuning, i.e., interactions with the target environment, thus providing clarity about the practical deployment requirements of each method. To ensure ease of adoption and reproducibility, we release a straightforward software interface for performing this evaluation procedure, thereby empowering future work to evaluate offline RL algorithms robustly and transparently. + +Problem 2: Opaque Algorithmic Design Offline RL methods are often presented as intricate bundles with intertwining algorithmic components, implementation-specific details, and unclear tuning procedures. Researchers compare proposed methods to baseline performance quoted directly from prior publications [6], inadvertently propagating these methodological issues. As a result, it is difficult to isolate the impact of individual methodological choices. Thus, the state-of-the-art remains ambiguous, with no method demonstrating uniformly strong performance across all datasets [13-15]. + +Solution 2: Consistent Reimplementations and a Unified Algorithm We first dissect the novel components of prior algorithms by defining a phylogenetic tree based on their compositional structure (Section 4.1). We use this representation to provide single-file reimplementations of a wide range of offline RL methods. These minimal implementations eliminate extraneous code differences and highlight fundamental components, as well as achieving average training speedups of $131.5 \times$ and $74.8 \times$ against OfflineRL-Kit [16] and CORL [17], two leading offline RL libraries. Furthermore, we propose a unified offline RL algorithm (Unifloral, Section 4.2) that integrates core components from various prior methods into one coherent framework. Crucially, Unifloral provides a single, unified hyperparameter space containing all of these algorithms. + +Leveraging Unifloral with our evaluation procedure, we introduce two novel offline RL methods: a model-free approach (TD3-AWR, Section 5.1) and a model-based one (MoBRAC, Section 5.2). These methods demonstrate substantial performance improvements over established baselines, validating both our unified methodology and rigorous evaluation framework. + +![](images/3e367ca1a95e318b6b808a6b3186389f3f5f42a38cf1d301a79126a0f6f3888a.jpg) +Figure 1: Formalizing the variants of offline RL—we define a range of offline RL variants (Section 3.1), with policy performance being measured post-deployment. Pre-deployment policy selection (2a) and post-deployment policy selection (2b) use a policy-selection bandit after offline training, whilst (3) uses unrestricted policy updates. + +# Setting + +# 1. Zero-shot + +Example: Autonomous search and rescue + +# 2a. Pre-deployment policy selection + +Example: Autonomous vehicles with safety driver testing + +# 2b. Post-deployment policy selection + +Example: Autonomous search and rescue + +# 3. Offline-to-online + +Example: Multi-step language reasoning models + +# 2 Preliminaries + +# 2.1 Reinforcement Learning + +We apply RL to a finite-horizon Markov Decision Process (MDP) defined by the tuple $\langle S_0, S, \mathcal{A}, D, R, T \rangle$ . Here $S$ is the state space, $\mathcal{A}$ is the action space, and $T$ is the horizon. $D: S \times \mathcal{A} \to \Delta(S)$ is the transition dynamics, defining how the state changes given a state and the action taken on that state. $\Delta(S)$ is the set of all possible distributions over $s$ . The scalar reward function is $R: S \times \mathcal{A} \to \mathbb{R}$ . The environments in this paper are all fully observable as the Markov state is directly observed at each timestep. + +A policy $\pi$ maps a state in $\mathcal{S}$ to an action distribution over $\mathcal{A}$ . The policy is trained to maximize the expected return $J_{M}^{\pi}$ for a given MDP $M$ with trajectory length $T$ : + +$$ +J _ {M} ^ {\pi} := \mathbb {E} _ {a _ {0: T} \sim \pi , s _ {0} \sim \mathcal {S} _ {0, s _ {1: T}} \sim D} \left[ \sum_ {t = 0} ^ {T} r _ {t} \right]. \tag {1} +$$ + +# 2.2 Offline Reinforcement Learning + +Offline RL methods use a pre-collected dataset to optimize a target policy to maximize $J^{\pi}$ by, without online interactions in the environment. This dataset consists of transitions $(s_i, a_i, r_i, s_{i+1}, a_{i+1})$ for $i = 1, \dots, N$ , where $s_i, s_{i+1} \in S, a_i \in \mathcal{A}, r_i \in \mathbb{R}$ are the current and next states, action, and reward, respectively. Here, initial states are drawn from the distribution $s_0 \sim S_0$ and trajectories are gathered through a behaviour policy $\pi_b$ interacting with the environment. Since $\pi_b$ may be suboptimal, the resulting dataset might not contain sufficient coverage of the environment's state space to learn an effective policy. + +An effective offline RL method must learn policies that generalize from this limited dataset to perform reliably when deployed in their environment. Typically, these methods require significant regularization to avoid overestimation bias. For model-free methods, this is commonly done with critic ensembles, where the minimum state value estimated by the ensemble is used for policy optimization. Model-based methods generalize by training a dynamics model $\hat{D}(s,a)$ to predict future states and rewards. This can be used to generate synthetic rollouts from the target policy, allowing for direct optimization of its performance. + +# 3 Refining Evaluation in Offline RL + +This section describes our taxonomy of offline RL, illustrated in Figure 1, which motivates our evaluation procedure in Figure 2. We also outline the procedure in detail and use it to analyze the performance of a set of model-free and model-based algorithms in multiple environments. + +# 3.1 Variants of Offline RL + +The goal of offline RL is to train an agent using solely offline data, with the objective of maximizing performance from deployment, i.e., the point where the agent is evaluated online. In this setting, deployment marks a strict separation between the offline training phase and the online evaluation phase. However, some methods may relax this strict separation in two ways. Firstly, pre-deployment interaction allows the agent to take limited interactions with the environment before deployment to improve post-deployment performance. For instance, to tune hyperparameters before selecting a policy for deployment. Secondly, post-deployment adaptation allows the agent to continue learning after deployment, and the performance metric includes all returns collected after deployment. Examples include dataset aggregation from multiple online episodes [18], selection from a set of policies trained offline [7, 9], and fine-tuning a single policy [8], all of which can be performed both before and after deployment. While any combination of these is possible, we identify four key settings. + +# A Taxonomy of Offline RL + +1. ZERO-SHOT OFFLINE RL + +- Train one policy offline, then deploy online with no further adaptation. +- No pre-deployment interaction, no post-deployment adaptation. + +2a. OFFLINE RL WITH PRE-DEPLOYMENT ONLINE POLICY SELECTION + +- Train a set of policies offline, select the best policy based on $N$ online evaluations before deployment. +- Limited pre-deployment interaction, no post-deployment adaptation. + +2b. OFFLINE RL WITH POST-DEPLOYMENT ONLINE POLICY SELECTION + +- Train a set of policies offline, then deploy online, adaptively selecting a policy every episode based on online performance. +- No pre-deployment interaction, post-deployment adaptation via policy selection. + +3. OFFLINE-TO-ONLINE RL + +- Train one policy offline, then deploy online and fine-tune the policy on online data. +- Limited pre-deployment interaction and post-deployment adaptation via finetuning. + +Many offline RL papers implicitly perform pre-deployment policy selection (Setting 2a), as they report final performance after extensive hyperparameter tuning involving online evaluation [11, 5]. However, due to differences in the number of hyperparameters or computational resources, this tuning process varies in scope across studies. As a result, reported performances are often not directly comparable since they reflect not only algorithmic quality but also differences in tuning budgets. Furthermore, these procedures typically assume low-variance estimates of each policy's performance, determined by an indefinite number of online evaluations. This is rarely made explicit as hyperparameter tuning is often considered a technical detail and not part of the method, even though it can dramatically affect performance (Section 3.3). + +Finally, much prior work has blurred the line between algorithms and hyperparameters in offline RL, proposing different hyperparameter values or ranges for each task. This ambiguity enables the same "method" to have dramatically different behaviour across tasks, + +undermining the assumption of limited interactions by essentially proposing a different method for each task. To resolve this, we define an offline RL method to include a fixed hyperparameter range, which remains constant across datasets (see A Definition of Offline RL Methods). + +# A Definition of Offline RL Methods + +A method in offline RL consists of an algorithm and a fixed sampling range for each hyperparameter. + +# 3.2 Proposed Evaluation Procedure + +We now propose a rigorous and practical evaluation procedure for offline RL with pre-deployment policy selection (Setting 2a), as it implicitly is the standard setting for evaluating offline RL methods (see Section 3.1). Our goal is to evaluate offline RL algorithms under a fixed budget of $N$ pre-deployment environment interactions used for tuning. We measure this budget in terms of the number of evaluation episodes, reflecting practical deployment constraints where each online interaction can be costly. Whilst the tuning algorithm may be defined as part of the method, most research focuses on offline policy optimization prior to tuning. Therefore, we provide an upper confidence bound (UCB) bandit [19] in our implementation as the default tuning algorithm. + +Furthermore, to reflect real-world limitations, we assume that the expected return of each policy is not directly observable, with each pull from the bandit sampling a single episodic return from that policy's return distribution. This models the high-variance, sample-limited setting typical in real deployments, where evaluating a policy's performance requires interacting with the environment and yields only noisy, episodic feedback. The importance of this is demonstrated by the emergence of distractor policies, as discussed in Section 3.3. + +In essence, our evaluation procedure repeatedly simulates hyperparameter tuning with a fixed online budget, using a bandit to select a single policy for final deployment. This procedure (Figure 2) has two steps: score collection and bandit evaluation. + +![](images/a3458b291d9fc687b61de79de0452690d6ebc78f23e58787cf82f2e58ad56a6d.jpg) +Figure 2: Overview of our evaluation procedure. Left: We sample hyperparameters, train the corresponding policies, and collect their final evaluation scores. Right: We simulate hyperparameter tuning using the collected scores by subsampling $K$ policy scores and recording the best-arm performance of a UCB tuning bandit operating over them. + +Step 1: Train Policies and Collect Scores Firstly, we collect a dataset of episodic evaluation scores from policies trained by the target algorithm. To do this, we sample $P$ hyperparameter settings (with replacement and random seeds) from the range defined by the method, and then train $P$ corresponding policies. These policies are evaluated online for a large number of episodes $R$ and their episodic scores recorded. Following this, the policies may be discarded as only their episodic scores are required for bandit evaluation. + +Step 2: Run Bootstrapped Tuning Bandit Using our collected episodic evaluation dataset, we then repeatedly simulate hyperparameter tuning to measure algorithm performance at different tuning budgets. This is performed by subsampling $K$ policies (i.e., their corresponding episodic scores) and running a multi-armed bandit over them. In this bandit, each arm corresponds to a policy, with each pull sampling one episode's return from the corresponding policy. At each number of pulls $N$ , we evaluate the performance of the algorithm by selecting the policy estimated to have the highest return by the bandit, and taking its true average return. We repeat this process $B$ times to obtain a bootstrapped estimate of algorithm performance. + +Recommended Datasets It is essential to evaluate methods on a diverse distribution of tasks to ensure generality. Alarmingly, the majority of offline RL methods considered in this work were evaluated only on MuJoCo and Adroit tasks from the D4RL suite [20]. While computational budgets may be limited, we argue that they would be better spent considering a wider range of tasks and behaviour policies. In order to make environment selection consistent, we recommend starting with the following environments, where algorithms currently obtain non-trivial performance: hopper-medium, halfcheetah-medium-expert, and walker2d-medium-replay, as a representative subset of MuJoCo locomotion; pen-human, pen-cloned, and pen-expert, as algorithms often achieve zero or perfect performance on other Adroit environments; kitchen-mixed, maze2d-large, and antmaze-large-diverse, to provide diversity in the evaluated environments. + +# 3.3 Results + +In Figure 3, we evaluate a range of prior algorithms (list in Appendix A). For this, we uniformly sample from the hyperparameter tuning ranges specified in each algorithm's original paper or the union of ranges when multiple are provided. Generally, an algorithm performs better if its curve is closer to the top left corner of a plot, representing strong performance after few online interactions. Prior work has typically reported performance after unlimited online tuning, which is the limit of the score with an increasing number of policy evaluations, i.e., the top right corner. + +Inconsistent Algorithm Performance No algorithm consistently performs well across all datasets. However, ReBRAC and IQL are competitive for the overall best performing algorithm, with ReBRAC achieving top performance at some number of evaluations on 5 out of 9 datasets and IQL on 4 out of 9 datasets. Even though both of these algorithms are worse than competing baselines on other datasets, we believe them to be the clearest baselines for future method development, as done in Section 5.1. + +![](images/7fd0079e4d77c3dee26080c6504ce78ec2f0bc4035f5a6fef4b97425b49cb923.jpg) + +![](images/36871d657fdeff8a325f2c59fc8eea980b0f69306dbf206e22c0416b45eecc3f.jpg) + +![](images/56408a852523988c84610c33e652baa13ff57776338e7865a50243fb87329532.jpg) + +![](images/e943362de33b0ef70a7c6e87ae39e46a52dc36afff3a892860de7421c8fdd676.jpg) + +![](images/2ed93996d5081ab75f402952c3dfb9ba87103a1270fc1b6fb9618a442fa01615.jpg) + +![](images/311383057c510ac69359a1106e10e71db8785b0630dc84fc1ea0a6e11b3bd3b4.jpg) +Number of policy evaluations + +![](images/00b516806e780231aeaecc98f156e13aba349032644089f3d7f9471730db37ed.jpg) + +![](images/a60a8c208919237a4043645d5b7d15ce88e92974541a2467039a089da46d7b0c.jpg) + +![](images/9ddc5d94ccaba6c7bb5e36b260958f91d0d2bd11987ef9b976266384209a2c22.jpg) +Figure 3: Evaluation of prior algorithms—mean and $95\%$ CI over 500 bandit rollouts, with $K = 8$ policy arms subsampled from 20 trained policies each rollout. The $x$ -axis denotes the number of bandit pulls, whilst the $y$ -axis denotes the true expected score of the estimated best arm after $x$ pulls. The full evaluation results in Appendix B. + +Overfit Model-Based Methods The model-based algorithms we evaluate—MOPO, MOReL, and COMBO (Appendix A.2)—achieve notably poor performance on all non-locomotion datasets, ranking no higher than $6^{\text{th}}$ out of the 10 evaluated algorithms (and failing to beat BC) at any number of policy evaluations. While these results are surprising, we emphasize that our implementation successfully reproduces reference results with the specialized hyperparameters for each dataset (Appendix G). Instead, these results suggest that these methods are deeply overfit to the locomotion datasets they were originally evaluated on (Appendix C), providing a sobering reflection of the field. + +Distractor Policy Phenomenon While performance typically improves as more bandit arms are pulled, certain performance curves exhibit distinctive dips—temporary decreases in measured performance despite additional policy evaluations. To better understand this, we examine the ranked performance distribution of numerous ReBRAC policies trained on hopper-medium (Figure 4). This analysis reveals a notable cluster of policies that exhibit suboptimal average performance but possess a higher maximum performance compared to consistently better-performing policies. We refer to these anomalous policies as distractor policies. + +To demonstrate their impact on evaluation, we simulate the initial phase of a bandit rollout over these policies, i.e., when the bandit enumerates all arms (Figure 9a). Over this phase, we observe a clear increase in the probability of preferring a distractor policy, explaining the initial decrease in evaluation + +performance. This phenomenon runs counter to the expectation that increasing policy evaluations would monotonically reduce estimator variance and underscores the need to directly consider environment interactions in evaluation, a crucial distinction from prior evaluation methodologies [9]. Further analysis of distractor policies is provided in Appendix D. + +![](images/bb0e857cf953bef426827e5ca9701984fe3aa0b094ec7bd550aed580216a6b6e.jpg) +Figure 4: Ranked ReBRAC performance—blue shaded area, solid and dashed lines representing the standard deviation, mean, min, and max episodic return, respectively. + +# 4 Elucidating Algorithm Design in Offline RL + +In this section, we seek to simplify algorithm design in offline RL. Firstly, we present a genealogy of prior algorithms, using it to propose and implement a set of compositional reimplementations. Following this, we propose a unified algorithm, Unifloral, capable of expressing these methods—as well as any combination of their components—in a single hyperparameter space. + +# 4.1 Disentangling Prior Methods + +New offline RL methods are typically derived from preceding ones by adding or editing individual components of the agent's objective or architecture. Despite this, methods typically suffer from a range of unnecessary implementation differences, making it difficult for researchers to identify their contribution or fairly compare methods. Even in popular single-file implementations, we observe significant code differences between "parent" and "child" algorithms, which should require only the individual components to be edited. This encourages researchers to compare entire algorithms rather than abating components. We discuss this and how it informs our code philosophy in Appendix E. + +![](images/ee194e163babba60fee8e0ae7dad045e8ff1881128039d3bc770e47b98fe7b10.jpg) +Figure 5: Speed up from our JAX reimplementations — algorithms trained for 1M update steps on HalfCheetah-medium-expert using a single L40S GPU. Our library, Unifloral, is the fastest across the board. Full details can be found in Appendix F. + +As a solution, we provide single-file implementations of a range of existing model-free (BC, TD3-BC, ReBRAC, IQL, SAC-N, LB-SAC, EDAC, CQL, DT) and model-based (MOPO, MOReL, COMBO) methods. Our implementation has a number of advantages. Firstly, we focus on code clarity and minimal code edits between algorithms, leading to a dramatic reduc + +tion in code differences between algorithms. Secondly, we implement our algorithms in end-to-end compiled JAX, leading to major speed-ups against competing implementations (Figure 5). We believe these implementations will lead to better algorithm understanding and fairer evaluation, as well as enabling powerful experiments on low compute budgets. We verify the correctness of our implementations in Appendix G. + +# 4.2 A Unified Hyperparameter Space for Offline RL + +Implementation inconsistency and missing ablations are common flaws of offline RL research. The plethora of design decisions in each algorithm obfuscates evaluating how each feature contributes to the performance. To address this, we combine all components from a range of model-free and model-based algorithms (Appendix A) into a unified algorithm and single-file implementation, which we name Unifloral. We start by compiling a minimal subspace of components covering the model-free and model-based offline RL algorithms examined in this work (Appendix H). This has a range of hyperparameters in each of four broad design categories, which we identify from prior algorithms: model design, critic objective, actor objective, and dynamics modelling. A more detailed description of design category is in Appendix I. + +Model Design The choice of neural network architecture and optimizer is consistent across most offline RL research, with proposed algorithms commonly using multi-layer perceptrons and the Adam optimizer. However, the hyperparameters of these components commonly vary between algorithms. Regarding the model architecture, this includes the number of layers, layer width, and usage of observation and layer normalization. For optimization, this includes the learning rate (shared and actor-specific), learning rate schedule, discount factor, batch size, and Polyak averaging step size. The actor and critic networks can also have different structures, such as the size of the critic ensemble and whether the policy stochasticity. + +Critic Objective The core contribution of offline RL research is often a novel critic objective [12, 13]. However, many of the components in the proposed objectives are shared with prior work. We define the critic objective as the weighted sum of those components, or a selection between them if mutually exclusive, in order to include all referenced methods (except CQL, which we omit due to its substandard performance and high complexity). More detail in Appendix I.1. + +Actor Objective We define the unified actor loss as the weighted sum of three terms: + +$$ +\mathcal {L} _ {\text {a c t o r}} = \beta_ {q} \cdot \mathcal {L} _ {q} + \beta_ {\mathrm {B C}} \cdot \mathcal {L} _ {\mathrm {B C}} - \beta_ {\mathcal {H}} \cdot \mathcal {H} (\pi (\cdot | s _ {t})). \tag {2} +$$ + +This consists of $q$ loss $\mathcal{L}_q$ , behaviour cloning loss $\mathcal{L}_{\mathrm{BC}}$ , and policy entropy $\mathcal{H}(\cdot)$ , with coefficients $\beta_q, \beta_{\mathrm{BC}}, \beta_{\mathcal{H}} \in \mathbb{R}$ controlling the weight of these terms. More detail in Appendix I.2. + +Dynamics Modelling We include optional dynamics model training and sampling, broadening Unifloral's coverage to include model-based methods. As is standard, we use an ensemble of dynamics models $\hat{D}_{\theta} = \{\hat{D}_{\theta}^{1},\hat{D}_{\theta}^{2},\dots,\hat{D}_{\theta}^{M}\}$ , where each $\hat{D}_{\theta}^{i}$ is trained to predict state transitions and rewards. Following MOPO, we penalize the agent for going to states where the ensemble disagreement is high as measured by the standard deviation of the model's predictions. More in Appendix I.3. + +# 5 Novel Methods Research with Unifloral + +Our unified algorithm and hyperparameter space enable researchers to combine different components and search through algorithm designs by only modifying the configuration of the unified implementation. To demonstrate the avenues our work opens up and encourage further research, we provide two "mini-papers" completed entirely by specifying configurations of the unified implementation, without any code changes. We examine a model-free and a model-based improvement. + +# 5.1 TD3 with Advantage Weighted Regression + +Hypothesis In Section 3.3, we show that one of two methods consistently outperformed existing baselines: ReBRAC [5] and IQL [21]. ReBRAC is derived from TD3-BC, meaning it optimizes its actor using TD3 value loss in combination with a BC loss term for regularization. In contrast, IQL uses only a BC loss but performs advantage weighted regression (AWR) by weighting the BC loss of each action by its estimated advantage. We hypothesise that substituting the BC term in ReBRAC with AWR, a method we name TD3-AWR, would combine the strengths of these methods and lead to improved performance overall. + +Evaluation We define TD3-AWR in Unifloral by using the AWR hyperparameters from IQL and the ReBRAC hyperparameters elsewhere. In Figure 6, we show that TD3-AWR's performance curve strictly dominates ReBRAC on 6 out of 9 datasets and is dominated by ReBRAC in only 1. Interestingly, TD3-AWR achieves superior performance to ReBRAC under few policy evaluations—such as in halfcheetah-medium-expert and pen-expert—despite searching over a wider range of hyperparameters. Similarly, TD3-AWR strictly dominates IQL on 7 datasets, thereby outperforming both of its source algorithms. + +![](images/81a78c92f1f501027367ced741a2df28b61bef153b8230e8067bbc4ca6b50e05.jpg) + +![](images/6d2d09077f83a4e8cfdd9aa412573d611d1e09246532e9d78e50328b69cd9d7c.jpg) +Number of policy evaluations + +![](images/b0a5ba1085e8a2adbc2347c7291722a06dc1adc2ef22a74fbb03b39d583ca64f.jpg) + +![](images/87a489992812d81b63934f48d6be45d7138ab988b97172e75f155dffe26741c6.jpg) + +![](images/519202b96b3f799e19e092a0c1a961df492f9e35f9c9a8b9225039ff4237423f.jpg) +Figure 6: TD3-AWR evaluation against ReBRAC and IQL (full results in Appendix J). + +# 5.2 Improving Policy Optimization for Model-Based Offline RL + +Hypothesis In Section 3.3, we demonstrate the poor performance of model-based methods on non-locomotion environments. Whilst this is partially due to overfit hyperparameters, the design space of policy optimizers in model-based methods is underexplored, with all considered methods using SAC-N or CQL (Figure 5). Given the performance improvements from recent methods, we posit that these methods would be more competitive with an alternative policy optimizer. We therefore propose using ReBRAC with synthetic rollouts generated from a MOPO world model, which we name Model-based Behaviour Regularized Actor-Critic, or MoBRAC. + +Evaluation We implement MoBRAC in Unifloral, using the MOPO hyperparameters for dynamics model training and sampling, then using the ReBRAC hyperparameters elsewhere. Figure 7 shows how MoBRAC outperforms other model-based methods for all datasets, except for MOPO in maze2d-large-v1. Under a transparent evaluation budget, we find that MoBRAC outperforms the other model-based methods in 6 out of 9 datasets and is tied with MOPO for 3 others (Appendix K). + +![](images/d43d5f543bccf5ad0677eac455403f2807a00d17c4cceeb0597380e8b225612f.jpg) + +![](images/47921e790f31c7a16e6350e09a4e83de1ae8304efd72e978bfc5322bed99be00.jpg) +Number of policy evaluations + +![](images/adbe6db3a6ddddd3fa35fe7760722b8d336a4df13acfc13c55bf268dceaa92a8.jpg) + +![](images/5c4703b41d05c655703a2bd169ce08600cc44cd8587fa8192079a292cd0a95cd.jpg) + +![](images/2d39f6309d4c6ab755c58e36ca4e21735c2afd518c01022f148905dc9b51cd83.jpg) +Figure 7: MoBRAC evaluation against prior model-based algorithms (full results in Appendix K). + +# 6 Related Work + +Our work builds upon several foundational aspects of offline RL, including evaluation strategies, open-source implementations, and algorithmic unification. Existing evaluation regimes primarily address hyperparameter tuning either through limited online interactions [8, 9] or by estimating policy performance offline [4, 10, 22]. In contrast, our approach introduces an evaluation procedure that requires neither reference policies nor additional hyperparameters, offering broader applicability across the entire D4RL benchmark suite. Furthermore, our single-file implementations draw inspiration from projects such as CORL [17, 23] and CleanRL [24], whilst our unified algorithm, Unifloral, is informed by prior unification attempts [25, 26, 15, 27]. For a comprehensive review, see Appendix L. + +# 7 Conclusion + +In this work, we addressed critical challenges in problem formulation, evaluation, and algorithm unification in offline RL. We introduced a taxonomy that clearly distinguishes between offline RL variants—spanning zero-shot deployment to approaches with limited pre-deployment tuning or post-deployment adaptation. This categorization exposes the hidden online interactions, such as hyperparameter tuning, that have long confounded fair evaluation and reproducibility. To overcome these issues, we proposed a rigorous evaluation procedure that transparently quantifies the cost of online interactions via noisy, single-episode feedback. Additionally, by dissecting components of existing offline RL algorithms, we developed Unifloral, a novel unified offline RL algorithm that combines improvements of many previous methods, enabling seamless ablation of algorithmic components. We demonstrate this with two novel algorithms inside Unifloral, TD3-AWR and MoBRAC, which integrate the strengths of existing methods to achieve superior performance over a wide range of tasks. Collectively, our contributions set a new standard for addressing ambiguity in offline RL, promoting rigorous evaluation, and driving reproducible, impactful research in the field. + +# Acknowledgements + +The authors thank Michael Beukman, Cong Lu, Jack Parker-Holder, Hugh Bishop, and Nathan Monette for their valuable feedback on the paper. MJ, UB, and JL are funded by the EPSRC Centre for Doctoral Training in Autonomous Intelligent Machines and Systems. MJ is also funded by Amazon Web Services, UB is also funded by the Rhodes Scholarship and JL is funded by Sony Interactive Entertainment Europe Ltd. + +# References + +[1] Sergey Levine, Aviral Kumar, George Tucker, and Justin Fu. Offline Reinforcement Learning: Tutorial, Review, and Perspectives on Open Problems, November 2020. URL http://arxiv.org/abs/2005.01643. arXiv:2005.01643 [cs, stat]. +[2] Tianhe Yu, Garrett Thomas, Lantao Yu, Stefano Ermon, James Zou, Sergey Levine, Chelsea Finn, and Tengyu Ma. MOPO: Model-based Offline Policy Optimization, November 2020. URL http://arxiv.org/abs/2005.13239. arXiv:2005.13239 [cs, stat]. +[3] Scott Fujimoto and Shixiang Shane Gu. A Minimalist Approach to Offline Reinforcement Learning, December 2021. URL http://arxiv.org/abs/2106.06860. arXiv:2106.06860 [cs, stat]. +[4] Tom Le Paine, Cosmin Paduraru, Andrea Michi, Caglar Gulcehre, Konrad Zolna, Alexander Novikov, Ziyu Wang, and Nando de Freitas. Hyperparameter Selection for Offline Reinforcement Learning, July 2020. URL http://arxiv.org/abs/2007.09055. arXiv:2007.09055 [cs]. +[5] Denis Tarasov, Vladislav Kurenkov, Alexander Nikulin, and Sergey Kolesnikov. Revisiting the Minimalist Approach to Offline Reinforcement Learning, October 2023. URL http://arxiv.org/abs/2305.09836. arXiv:2305.09836 [cs]. +[6] Rafael Figueiredo Prudencio, Marcos ROA Maximo, and Esther Luna Colombini. A survey on offline reinforcement learning: Taxonomy, review, and open problems. IEEE Transactions on Neural Networks and Learning Systems, 2023. +[7] Ksenia Konyushova, Yutian Chen, Thomas Paine, Caglar Gulcehre, Cosmin Paduraru, Daniel J Mankowitz, Misha Denil, and Nando de Freitas. Active offline policy selection. Advances in Neural Information Processing Systems, 34:24631-24644, 2021. +[8] Tatsuya Matsushima, Hiroki Furuta, Yutaka Matsuo, Ofir Nachum, and Shixiang Gu. Deployment-efficient reinforcement learning via model-based offline optimization. arXiv preprint arXiv:2006.03647, 2020. +[9] Vladislav Kurenikov and Sergey Kolesnikov. Showing Your Offline Reinforcement Learning Work: Online Evaluation Budget Matters, June 2022. URL http://arxiv.org/abs/2110.04156. arXiv:2110.04156 [cs]. +[10] Matthew Smith, Lucas Maystre, Zhenwen Dai, and Kamil Ciosek. A strong baseline for batch imitation learning. arXiv preprint arXiv:2302.02788, 2023. +[11] Rahul Kidambi, Aravind Rajeswaran, Praneeth Netrapalli, and Thorsten Joachims. MOREL: Model-Based Offline Reinforcement Learning, March 2021. URL http://arxiv.org/abs/2005.05951. arXiv:2005.05951 [cs, stat]. +[12] Aviral Kumar, Aurick Zhou, George Tucker, and Sergey Levine. Conservative Q-Learning for Offline Reinforcement Learning, August 2020. URL http://arxiv.org/abs/2006.04779.arXiv:2006.04779 [cs, stat]. +[13] Gaon An, Seungyong Moon, Jang-Hyun Kim, and Hyun Oh Song. Uncertainty-Based Offline Reinforcement Learning with Diversified Q-Ensemble, October 2021. URL http://arxiv.org/abs/2110.01548. arXiv:2110.01548 [cs]. + +[14] Tianhe Yu, Aviral Kumar, Rafael Rafailov, Aravind Rajeswaran, Sergey Levine, and Chelsea Finn. COMBO: Conservative Offline Model-Based Policy Optimization, January 2022. URL http://arxiv.org/abs/2102.08363.arXiv:2102.08363 [cs]. +[15] Cong Lu, Philip J. Ball, Jack Parker-Holder, Michael A. Osborne, and Stephen J. Roberts. Revisiting Design Choices in Offline Model-Based Reinforcement Learning, March 2022. URL http://arxiv.org/abs/2110.04135.arXiv:2110.04135 [cs]. +[16] Yihao Sun. Offlinerl-kit: An elegant pytorch offline reinforcement learning library. https://github.com/yihaosun1124/OfflineRL-Kit, 2023. +[17] Denis Tarasov, Alexander Nikulin, Dmitry Akimov, Vladislav Kurenkov, and Sergey Kolesnikov. CORL: Research-oriented deep offline reinforcement learning library. In 3rd Offline RL Workshop: Offline RL as a "Launchpad", 2022. URL https://openreview.net/forum?id=SyAS49bBcv. +[18] Stéphane Ross, Geoffrey Gordon, and Drew Bagnell. A reduction of imitation learning and structured prediction to no-regret online learning. In Proceedings of the fourteenth international conference on artificial intelligence and statistics, pages 627-635. JMLR Workshop and Conference Proceedings, 2011. +[19] Peter Auer, Nicolo Cesa-Bianchi, and Paul Fischer. Finite-time analysis of the multiarmed bandit problem. Machine learning, 47:235-256, 2002. +[20] Justin Fu, Aviral Kumar, Ofir Nachum, George Tucker, and Sergey Levine. D4rl: Datasets for deep data-driven reinforcement learning. arXiv preprint arXiv:2004.07219, 2020. +[21] Ilya Kostrikov, Ashvin Nair, and Sergey Levine. Offline Reinforcement Learning with Implicit Q-Learning, October 2021. URL http://arxiv.org/abs/2110.06169. arXiv:2110.06169 [cs]. +[22] Aviral Kumar, Anikait Singh, Stephen Tian, Chelsea Finn, and Sergey Levine. A Workflow for Offline Model-Free Robotic Reinforcement Learning, September 2021. URL http://arxiv.org/abs/2109.10813. arXiv:2109.10813 [cs]. +[23] Soichiro Nishimori. Jax-crl: Clean single-file implementations of offline rl algorithms in jax. 2024. URL https://github.com/nissymori/JAX-CORL. +[24] Shengyi Huang, Rousslan Fernand Julien Dossa, Chang Ye, Jeff Braga, Dipam Chakraborty, Kinal Mehta, and JoAGo GM AraA\$jo. Cleanrl: High-quality single-file implementations of deep reinforcement learning algorithms. Journal of Machine Learning Research, 23(274):1-18, 2022. +[25] Matteo Hessel, Joseph Modayil, Hado Van Hasselt, Tom Schaul, Georg Ostrovski, Will Dabney, Dan Horgan, Bilal Piot, Mohammad Azar, and David Silver. Rainbow: Combining improvements in deep reinforcement learning. In Proceedings of the AAAI conference on artificial intelligence, volume 32, 2018. +[26] Matteo Hessel, Ivo Danihelka, Fabio Viola, Arthur Guez, Simon Schmitt, Laurent Sifre, Theophane Weber, David Silver, and Hado Van Hasselt. Muesli: Combining improvements in policy optimization. In International conference on machine learning, pages 4214-4226. PMLR, 2021. +[27] Harshit Sikchi, Qinqing Zheng, Amy Zhang, and Scott Niekum. Dual rl: Unification and new methods for reinforcement and imitation learning. arXiv preprint arXiv:2302.08560, 2023. +[28] Tuomas Haarnoja, Aurick Zhou, Pieter Abbeel, and Sergey Levine. Soft actor-critic: Off-policy maximum entropy deep reinforcement learning with a stochastic actor. In International conference on machine learning, pages 1861-1870. Pmlr, 2018. +[29] Scott Fujimoto, Herke Hoof, and David Meger. Addressing Function Approximation Error in Actor-Critic Methods. In Proceedings of the 35th International Conference on Machine Learning, pages 1587-1596. PMLR, July 2018. URL https://proceedings.mlr.press/v80/fujimoto18a.html. ISSN: 2640-3498. + +[30] Xue Bin Peng, Aviral Kumar, Grace Zhang, and Sergey Levine. Advantage-weighted regression: Simple and scalable off-policy reinforcement learning. arXiv preprint arXiv:1910.00177, 2019. +[31] Dean A Pomerleau. Alvinn: An autonomous land vehicle in a neural network. Advances in neural information processing systems, 1, 1988. +[32] Yifan Wu, George Tucker, and Ofir Nachum. Behavior regularized offline reinforcement learning. arXiv preprint arXiv:1911.11361, 2019. +[33] Lili Chen, Kevin Lu, Aravind Rajeswaran, Kimin Lee, Aditya Grover, Misha Laskin, Pieter Abbeel, Aravind Srinivas, and Igor Mordatch. Decision transformer: Reinforcement learning via sequence modeling. Advances in neural information processing systems, 34:15084-15097, 2021. +[34] Ishita Mediratta, Qingfei You, Minqi Jiang, and Roberta Raileanu. The Generalization Gap in Offline Reinforcement Learning, March 2024. URL http://arxiv.org/abs/2312.05742.arXiv:2312.05742 [cs]. +[35] Chris Lu, Jakub Kuba, Alistair Letcher, Luke Metz, Christian Schroeder de Witt, and Jakob Foerster. Discovered policy optimisation. Advances in Neural Information Processing Systems, 35:16455-16468, 2022. +[36] Han Wang, Archit Sakhadeo, Adam White, James Bell, Vincent Liu, Xutong Zhao, Puer Liu, Tadashi Kozuno, Alona Fyshe, and Martha White. No more pesky hyperparameters: Offline hyperparameter tuning for rl. arXiv preprint arXiv:2205.08716, 2022. +[37] Takuma Seno. d3rlpy: An offline deep reinforcement library. https://github.com/takuseno/d3rlpy, 2020. +[38] Antonin Raffin, Ashley Hill, Adam Gleave, Anssi Kanervisto, Maximilian Ernestus, and Noah Dormann. Stable-baselines3: Reliable reinforcement learning implementations. Journal of machine learning research, 22(268):1-8, 2021. +[39] Joshua Achiam. Spinning Up in Deep Reinforcement Learning, 2018. URL https://spinningup.openai.com/en/latest/index.html. +[40] Jarek Liesen, Chris Lu, and Robert Lange. rejax, 2024. URL https://github.com/ keraJLi/rejax. + +# NeurIPS Paper Checklist + +# 1. Claims + +Question: Do the main claims made in the abstract and introduction accurately reflect the paper's contributions and scope? + +Answer: [Yes] + +Justification: We point out problems in offline RL, describe an informative taxonomy Figure 1 and propose an evaluation procedure, the results of which support our claims Section 3.3. We then provide extensive explanation on our implementations' design and performance (Section 4.1). The unified implementations (Appendix I) and the evaluation procedure (Section 3.2 are used to discover two new algorithms that we describe and test in Section 5. + +Guidelines: + +- The answer NA means that the abstract and introduction do not include the claims made in the paper. +- The abstract and/or introduction should clearly state the claims made, including the contributions made in the paper and important assumptions and limitations. A No or NA answer to this question will not be perceived well by the reviewers. +- The claims made should match theoretical and experimental results, and reflect how much the results can be expected to generalize to other settings. +- It is fine to include aspirational goals as motivation as long as it is clear that these goals are not attained by the paper. + +# 2. Limitations + +Question: Does the paper discuss the limitations of the work performed by the authors? + +Answer: [Yes] + +Justification: We address lack of algorithmic clarity which is inherent in the field we are aiming to unify Section 1 and we acknowledge the limited number of evaluation environments Appendix C. In our full results for the two novel algorithms, Figure 15 and Figure 14, we include challenging environments where our algorithms underperform. However, these results themselves are further evidence of the robustness and reliability of our evaluation procedure. + +Guidelines: + +- The answer NA means that the paper has no limitation while the answer No means that the paper has limitations, but those are not discussed in the paper. +- The authors are encouraged to create a separate "Limitations" section in their paper. +- The paper should point out any strong assumptions and how robust the results are to violations of these assumptions (e.g., independence assumptions, noiseless settings, model well-specification, asymptotic approximations only holding locally). The authors should reflect on how these assumptions might be violated in practice and what the implications would be. +- The authors should reflect on the scope of the claims made, e.g., if the approach was only tested on a few datasets or with a few runs. In general, empirical results often depend on implicit assumptions, which should be articulated. +- The authors should reflect on the factors that influence the performance of the approach. For example, a facial recognition algorithm may perform poorly when image resolution is low or images are taken in low lighting. Or a speech-to-text system might not be used reliably to provide closed captions for online lectures because it fails to handle technical jargon. +- The authors should discuss the computational efficiency of the proposed algorithms and how they scale with dataset size. +- If applicable, the authors should discuss possible limitations of their approach to address problems of privacy and fairness. + +- While the authors might fear that complete honesty about limitations might be used by reviewers as grounds for rejection, a worse outcome might be that reviewers discover limitations that aren't acknowledged in the paper. The authors should use their best judgment and recognize that individual actions in favor of transparency play an important role in developing norms that preserve the integrity of the community. Reviewers will be specifically instructed to not penalize honesty concerning limitations. + +# 3. Theory assumptions and proofs + +Question: For each theoretical result, does the paper provide the full set of assumptions and a complete (and correct) proof? + +Answer: [NA]. + +Justification: This paper does not have theoretical results. + +# Guidelines: + +- The answer NA means that the paper does not include theoretical results. +- All the theorems, formulas, and proofs in the paper should be numbered and cross-referenced. +- All assumptions should be clearly stated or referenced in the statement of any theorems. +- The proofs can either appear in the main paper or the supplemental material, but if they appear in the supplemental material, the authors are encouraged to provide a short proof sketch to provide intuition. +- Inversely, any informal proof provided in the core of the paper should be complemented by formal proofs provided in appendix or supplemental material. +- Theorems and Lemmas that the proof relies upon should be properly referenced. + +# 4. Experimental result reproducibility + +Question: Does the paper fully disclose all the information needed to reproduce the main experimental results of the paper to the extent that it affects the main claims and/or conclusions of the paper (regardless of whether the code and data are provided or not)? + +# Answer: [Yes] + +Justification: We plan to open-source our library. Regardless, we have provided with exhaustive detail on algorithm similarity Section 4 and our code philosophy Appendix E. To encourage transparent and reliable scientific practices, we also provide snippets from the actual code to demonstrate our clean and consistent implementations Figure 12 and Figure 13. + +# Guidelines: + +- The answer NA means that the paper does not include experiments. +- If the paper includes experiments, a No answer to this question will not be perceived well by the reviewers: Making the paper reproducible is important, regardless of whether the code and data are provided or not. +- If the contribution is a dataset and/or model, the authors should describe the steps taken to make their results reproducible or verifiable. +- Depending on the contribution, reproducibility can be accomplished in various ways. For example, if the contribution is a novel architecture, describing the architecture fully might suffice, or if the contribution is a specific model and empirical evaluation, it may be necessary to either make it possible for others to replicate the model with the same dataset, or provide access to the model. In general, releasing code and data is often one good way to accomplish this, but reproducibility can also be provided via detailed instructions for how to replicate the results, access to a hosted model (e.g., in the case of a large language model), releasing of a model checkpoint, or other means that are appropriate to the research performed. +- While NeurIPS does not require releasing code, the conference does require all submissions to provide some reasonable avenue for reproducibility, which may depend on the nature of the contribution. For example +(a) If the contribution is primarily a new algorithm, the paper should make it clear how to reproduce that algorithm. + +(b) If the contribution is primarily a new model architecture, the paper should describe the architecture clearly and fully. +(c) If the contribution is a new model (e.g., a large language model), then there should either be a way to access this model for reproducing the results or a way to reproduce the model (e.g., with an open-source dataset or instructions for how to construct the dataset). +(d) We recognize that reproducibility may be tricky in some cases, in which case authors are welcome to describe the particular way they provide for reproducibility. In the case of closed-source models, it may be that access to the model is limited in some way (e.g., to registered users), but it should be possible for other researchers to have some path to reproducing or verifying the results. + +# 5. Open access to data and code + +Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material? + +Answer: [Yes] + +Justification: Yes, the paper features an anonymous repository and is built to optimize for transparent ablations, reproducibility and evaluations. We use standard D4RL [20] datasets available through their API. + +# Guidelines: + +- The answer NA means that paper does not include experiments requiring code. +- Please see the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details. +- While we encourage the release of code and data, we understand that this might not be possible, so "No" is an acceptable answer. Papers cannot be rejected simply for not including code, unless this is central to the contribution (e.g., for a new open-source benchmark). +- The instructions should contain the exact command and environment needed to run to reproduce the results. See the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details. +- The authors should provide instructions on data access and preparation, including how to access the raw data, preprocessed data, intermediate data, and generated data, etc. +- The authors should provide scripts to reproduce all experimental results for the new proposed method and baselines. If only a subset of experiments are reproducible, they should state which ones are omitted from the script and why. +- At submission time, to preserve anonymity, the authors should release anonymized versions (if applicable). +- Providing as much information as possible in supplemental material (appended to the paper) is recommended, but including URLs to data and code is permitted. + +# 6. Experimental setting/details + +Question: Does the paper specify all the training and test details (e.g., data splits, hyperparameters, how they were chosen, type of optimizer, etc.) necessary to understand the results? + +Answer: [Yes] + +Justification: Yes, we have a unified hyperparameter space for all the algorithms Table 5 and describe every design decision in Section 4.2 and Appendix I. + +# Guidelines: + +- The answer NA means that the paper does not include experiments. +- The experimental setting should be presented in the core of the paper to a level of detail that is necessary to appreciate the results and make sense of them. +- The full details can be provided either with the code, in appendix, or as supplemental material. + +# 7. Experiment statistical significance + +Question: Does the paper report error bars suitably and correctly defined or other appropriate information about the statistical significance of the experiments? + +Answer: [Yes] + +Justification: As described in Section 3.2, we provide a bootstrapped estimate of each algorithm's performance as more online evaluations are allowed to be used for hyperparameter selection. We plot mean and $95\%$ confidence intervals everywhere and add mean and standard deviation when relevant, like Figure 4). + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The authors should answer "Yes" if the results are accompanied by error bars, confidence intervals, or statistical significance tests, at least for the experiments that support the main claims of the paper. +- The factors of variability that the error bars are capturing should be clearly stated (for example, train/test split, initialization, random drawing of some parameter, or overall run with given experimental conditions). +- The method for calculating the error bars should be explained (closed form formula, call to a library function, bootstrap, etc.) +- The assumptions made should be given (e.g., Normally distributed errors). +- It should be clear whether the error bar is the standard deviation or the standard error of the mean. +- It is OK to report 1-sigma error bars, but one should state it. The authors should preferably report a 2-sigma error bar than state that they have a $96\%$ CI, if the hypothesis of Normality of errors is not verified. +- For asymmetric distributions, the authors should be careful not to show in tables or figures symmetric error bars that would yield results that are out of range (e.g. negative error rates). +- If error bars are reported in tables or plots, The authors should explain in the text how they were calculated and reference the corresponding figures or tables in the text. + +# 8. Experiments compute resources + +Question: For each experiment, does the paper provide sufficient information on the computer resources (type of compute workers, memory, time of execution) needed to reproduce the experiments? + +Answer: [Yes] + +Justification: Yes, in Table 2 and Figure 5 we specify the number of steps used for the runtime measurements and the hardware, namely a single Nvidia-L40S GPU. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The paper should indicate the type of compute workers CPU or GPU, internal cluster, or cloud provider, including relevant memory and storage. +- The paper should provide the amount of compute required for each of the individual experimental runs as well as estimate the total compute. +- The paper should disclose whether the full research project required more compute than the experiments reported in the paper (e.g., preliminary or failed experiments that didn't make it into the paper). + +# 9. Code of ethics + +Question: Does the research conducted in the paper conform, in every respect, with the NeurIPS Code of Ethics https://neurips.cc/public/EthicsGuidelines? + +Answer: [Yes] + +Justification: There are no experiments in this work requiring human participants. All the datasets are in the public domain and have been used in accordance with their respective licenses. + +Guidelines: + +- The answer NA means that the authors have not reviewed the NeurIPS Code of Ethics. +- If the authors answer No, they should explain the special circumstances that require a deviation from the Code of Ethics. +- The authors should make sure to preserve anonymity (e.g., if there is a special consideration due to laws or regulations in their jurisdiction). + +# 10. Broader impacts + +Question: Does the paper discuss both potential positive societal impacts and negative societal impacts of the work performed? + +Answer: [NA]. + +Justification: There are no societal impacts beyond encouraging the researchers and practitioners to report results in a more unified and transparent way. We provide the tools, baselines and demonstrate the effectiveness of our work by discovering two new algorithms. + +Guidelines: + +- The answer NA means that there is no societal impact of the work performed. +- If the authors answer NA or No, they should explain why their work has no societal impact or why the paper does not address societal impact. +- Examples of negative societal impacts include potential malicious or unintended uses (e.g., disinformation, generating fake profiles, surveillance), fairness considerations (e.g., deployment of technologies that could make decisions that unfairly impact specific groups), privacy considerations, and security considerations. +- The conference expects that many papers will be foundational research and not tied to particular applications, let alone deployments. However, if there is a direct path to any negative applications, the authors should point it out. For example, it is legitimate to point out that an improvement in the quality of generative models could be used to generate deepfakes for disinformation. On the other hand, it is not needed to point out that a generic algorithm for optimizing neural networks could enable people to train models that generate Deepfakes faster. +- The authors should consider possible harms that could arise when the technology is being used as intended and functioning correctly, harms that could arise when the technology is being used as intended but gives incorrect results, and harms following from (intentional or unintentional) misuse of the technology. +- If there are negative societal impacts, the authors could also discuss possible mitigation strategies (e.g., gated release of models, providing defenses in addition to attacks, mechanisms for monitoring misuse, mechanisms to monitor how a system learns from feedback over time, improving the efficiency and accessibility of ML). + +# 11. Safeguards + +Question: Does the paper describe safeguards that have been put in place for responsible release of data or models that have a high risk for misuse (e.g., pretrained language models, image generators, or scraped datasets)? + +Answer: [NA]. + +Justification: This is not relevant for the paper. The justification is similar to the above. + +Guidelines: + +- The answer NA means that the paper poses no such risks. +- Released models that have a high risk for misuse or dual-use should be released with necessary safeguards to allow for controlled use of the model, for example by requiring that users adhere to usage guidelines or restrictions to access the model or implementing safety filters. +- Datasets that have been scraped from the Internet could pose safety risks. The authors should describe how they avoided releasing unsafe images. +- We recognize that providing effective safeguards is challenging, and many papers do not require this, but we encourage authors to take this into account and make a best faith effort. + +# 12. Licenses for existing assets + +Question: Are the creators or original owners of assets (e.g., code, data, models), used in the paper, properly credited and are the license and terms of use explicitly mentioned and properly respected? + +Answer: [Yes] + +Justification: We give credit and cite every piece of code that we use or were inspired by. We emphasize that our paper does not rely on extensive amounts of boilerplate code, which is also one of our key contributions (Appendix E). Every dataset we used in accordance with its license and standard community practices. + +Guidelines: + +- The answer NA means that the paper does not use existing assets. +- The authors should cite the original paper that produced the code package or dataset. +- The authors should state which version of the asset is used and, if possible, include a URL. +- The name of the license (e.g., CC-BY 4.0) should be included for each asset. +- For scraped data from a particular source (e.g., website), the copyright and terms of service of that source should be provided. +- If assets are released, the license, copyright information, and terms of use in the package should be provided. For popular datasets, paperswithcode.com/datasets has curated licenses for some datasets. Their licensing guide can help determine the license of a dataset. +- For existing datasets that are re-packaged, both the original license and the license of the derived asset (if it has changed) should be provided. +- If this information is not available online, the authors are encouraged to reach out to the asset's creators. + +# 13. New assets + +Question: Are new assets introduced in the paper well documented and is the documentation provided alongside the assets? + +Answer: or [NA]. + +Justification: We do not provide any assets beyond the implementation, configuration files and package requirements. + +Guidelines: + +- The answer NA means that the paper does not release new assets. +- Researchers should communicate the details of the dataset/code/model as part of their submissions via structured templates. This includes details about training, license, limitations, etc. +- The paper should discuss whether and how consent was obtained from people whose asset is used. +- At submission time, remember to anonymize your assets (if applicable). You can either create an anonymized URL or include an anonymized zip file. + +# 14. Crowdsourcing and research with human subjects + +Question: For crowdsourcing experiments and research with human subjects, does the paper include the full text of instructions given to participants and screenshots, if applicable, as well as details about compensation (if any)? + +Answer: [NA]. + +Justification: This paper does not contain such experiments. + +Guidelines: + +- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. +- Including this information in the supplemental material is fine, but if the main contribution of the paper involves human subjects, then as much detail as possible should be included in the main paper. + +- According to the NeurIPS Code of Ethics, workers involved in data collection, curation, or other labor should be paid at least the minimum wage in the country of the data collector. + +# 15. Institutional review board (IRB) approvals or equivalent for research with human subjects + +Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals (or an equivalent approval/review based on the requirements of your country or institution) were obtained? + +Answer: [NA]. + +Justification: This paper does not involve crowd-sourcing. + +# Guidelines: + +- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. +- Depending on the country in which research is conducted, IRB approval (or equivalent) may be required for any human subjects research. If you obtained IRB approval, you should clearly state this in the paper. +- We recognize that the procedures for this may vary significantly between institutions and locations, and we expect authors to adhere to the NeurIPS Code of Ethics and the guidelines for their institution. +- For initial submissions, do not include any information that would break anonymity (if applicable), such as the institution conducting the review. + +# 16. Declaration of LLM usage + +Question: Does the paper describe the usage of LLMs if it is an important, original, or non-standard component of the core methods in this research? Note that if the LLM is used only for writing, editing, or formatting purposes and does not impact the core methodology, scientific rigorousness, or originality of the research, declaration is not required. + +Answer: [NA] + +Justification: The contributions and reproduction of this work does not require any LLMs, neither did its ideation and development. + +# Guidelines: + +- The answer NA means that the core method development in this research does not involve LLMs as any important, original, or non-standard components. +- Please refer to our LLM policy (https://neurips.cc/Conferences/2025/LLM) for what should or should not be described. + +# A Algorithm Implementations in Unifloral + +# A.1 Model-Free Offline RL + +SAC Soft Actor-Critic (SAC) by Haarnoja et al. [28] is a $Q$ -learning method with a stochastic actor. The authors use two independently optimized $Q$ -functions and take their minimum for the value function gradient to reduce positive bias in the policy improvements. SAC uses function approximators for both the policy and value functions. + +EDAC Function approximators do not operate well out-of-distribution (OOD), which poses a significant challenge for offline RL methods that rely on a fixed dataset of logged trajectories. An et al. [13] propose increasing the size of the $Q$ -function ensemble. They find that SAC requires a large ensemble to avoid optimistic value estimations for OOD actions as the cosine similarity of the gradients increases. To minimize this similarity within the ensemble, the authors propose the Ensemble-Diversified Actor-Critic (EDAC), which adds an ensemble similarity penalty to the $Q$ -function loss in SAC. We refer to SAC with more than two members in the ensemble as SAC-N. + +CQL Optimistic value estimations when bootstrapping from OOD actions is a persisting issue in offline RL. Kumar et al. [12] propose learning a conservative $Q$ -function that lower bounds the true value. They perform SAC updates to the $Q$ -function with an additional minimization term that uses the value of randomly sampled actions. Their Conservative $Q$ -Learning (CQL) algorithm is also implemented on top of a SAC-N policy update similar to EDAC. + +TD3-BC Fujimoto et al. [29] formulate the Twin-Delayed Policy Deep Deterministic policy gradient algorithm (TD3) to address the value estimation pathology in online RL where the ensemble of $Q$ -networks is updated at a higher frequency than the actor. TD3 also takes the minimum over the critics ensemble as in CQL, SAC-N, and EDAC. Follow-up work by Fujimoto and Gu [3] adapts the method for the offline paradigm by adding a behaviour cloning (BC) regularization term to the actor's updates. This augmented algorithm is commonly referred to as TD3-BC. Not having to update two networks in every training step brings significant speed-ups while still matching the highest scores across all D4RL [20] locomotion tasks at an increased stability. + +IQL Implicit $Q$ -learning by Kostrikov et al. [21] is a computationally efficient algorithm that avoids querying out-of-sample actions altogether by using expectile regression. The $Q$ -function is updated using a mean squared error loss on state-action pairs from the dataset. This approximation of the optimal $Q$ -function is used to extract the policy through advantage-weighted regression [30], where each action is weighted according to the exponentiated advantage with an inverse temperature hyperparameter that directs the policy towards higher $Q$ -values when increased and approximates behavior cloning [31] when decreased. + +ReBRAC Tarasov et al. [5] use the Behavior Regularized Actor-Critic (BRAC) framework [32] and the behavior cloning term from TD3-BC [3] to propose the Revisited BRAC algorithm (ReBRAC). Specifically, they decouple the BC penalty coefficient in the critic and the actor objectives, thus requiring additional hyperparameters to the benefit of higher scores and faster convergence on D4RL. In addition, ReBRAC [5] proposes several improvements, like using deeper networks, training with larger batches, adding layer norms to the critic network, and changing the $\gamma$ hyperparameter for tasks with different reward sparsity. However, these design decisions add new hyperparameters with tuning overheads since they are reportedly different for each D4RL dataset. + +# A.2 Model-Based Offline RL + +MOPO In Model-Based Offline Policy Optimization (MOPO), Yu et al. [2] argue that offline RL algorithms should be able to go beyond the behaviors in the data manifold to avert sub-optimalities in the dataset and generalize to new tasks to deliver on the promises of real-world deployment. MOPO provides several bounds and theoretical guarantees on behavior policy improvement. The model is implemented through an ensemble of multiple dynamics models trained via maximum likelihood. For every policy step during training, the maximum standard deviation of the learned models' prediction at that step is subtracted from the reward. The highest results are obtained on short truncated rollouts + +that are $0.5\%$ to $1\%$ of the real environment's episode length. The model predictions are used to form the batch for the SAC [28] policy update step. + +MOReL The model-based offline RL algorithm (MOReL) by Kidambi et al. [11] claims to not require severely truncated rollouts due to learning a pessimistic MDP (P-MDP) that is implemented in a similar way to the MOPO dynamics model with an additional early termination condition in the event of high ensemble disagreement. This scalar halting threshold is calculated by taking the maximum distance between the predictions of any two models of the ensemble for every state and action pair in the dataset. Even for academic demonstration datasets like D4RL, this poses a major overhead in addition to model and policy training. The reported rollout length approximating $50\%$ of the original episode length is only achievable through extensive tuning of the pessimism coefficient that scales the discrepancy threshold. + +COMBO Conservative Offline Model-Based Policy Optimization (COMBO) by Yu et al. [14] is implemented on top of MOPO [2] with more policy improvement guarantees. They use a CQL [12] policy update step with an added loss term using transitions from the dataset to penalize $Q$ -values on likely out-of-support state-actions while increasing $Q$ -values on trustworthy pairs. There are many similarities across model-based methods, and many of their algorithmic contributions like the P-MDP from MOReL, uncertainty penalties from MOPO, and the policy update from COMBO can be combined through our framework. + +# A.3 Imitation Learning + +This section examines methods that operate outside traditional RL paradigms. These methods use identical offline RL datasets and have achieved scores comparable to other offline RL methods when evaluated under the same conditions. + +BC Behavioral cloning (BC), originally formalized by Pomerleau [31], directly optimizes the actor by learning the transitions from the dataset in a supervised manner, thus making the final online performance fully reliant on the quality of the dataset. Recent work by Kurenkov and Kolesnikov [9] further points out the effectiveness of BC under restricted budgets. + +DT Introduced by Chen et al. [33], Decision Transformers (DT) have shown remarkable generalization [34] in ORL. DT bypasses the need for traditional RL algorithms to use discounted rewards and bootstrapping for long-term credit assignment by using the logged environment interactions as a sequence modelling objective. Instead of sampling from a policy conditioned on the current states, the trained transformer autoregressively generates the next action based on a fixed intra-episode context of previous interaction and a target cumulative return. This target return can be a hyperparameter that significantly increases the tuning overhead if its value is unknown or a way to obtain optimal performance results when the target return is known. + +The reward at each step is decremented from the target return, which is referred to as return-to-go at time $t$ . Formally, $\hat{R}_t = \sum_{t' = t}^T r_{t'}$ where $r_{t'}$ are the observed rewards. Rather than directly modelling the reward function $R$ , the model is conditioned on the return-to-go values to enable generation based on desired future returns. + +The trajectory representation $\tau$ is structured as an ordered sequence of return-to-go values, states, and actions: + +$$ +\tau = \left(\hat {R} _ {1}, s _ {1}, a _ {1}, \hat {R} _ {2}, s _ {2}, a _ {2}, \dots , \hat {R} _ {T}, s _ {T}, a _ {T}\right), \tag {3} +$$ + +where $(s_t, a_t) \in S \times \mathcal{A}$ for all timesteps $t$ . + +During online evaluation, the model is initialized with a desired target return and an initial state $s_0 \sim S_0$ . After executing action $a_t$ , the received reward is subtracted from the target: $\hat{R}_{t+1} = \hat{R}_t - r_t$ . + +# B Full Results from the Proposed Evaluation Procedure + +![](images/617c93f9d05fca91095058c7ff23a9fccb25175912d10bdfa1e7078ff8cea428.jpg) + +![](images/f20f2abfb250db9d10f5ad3d1d08384bd9c97862d16a6a2c8d756f87a71fadb2.jpg) + +![](images/f5a0ea9dfc16fd1329e255ab2433b4d87f0d4700097f26ed3e5f3ca6b9cb6c4b.jpg) + +![](images/05a56e8634f3b13fa6ecb103818fdddc13e359f2595bfa36eaa691c1ce77d2e7.jpg) + +![](images/6ce6d71800cf00b29b8b2fb3ec303a4423201b72a3af589d9d1ffe629967fe25.jpg) + +![](images/4d864adb8c01997b034fcc78f0b0778fddb6533d208803393f7328527d27fff1.jpg) + +![](images/db7f5fdbf4ae94747541d2fea1d6986d6ca32f6c3772e7e5d866d5ef2862185e.jpg) + +![](images/012db8c27842846e445c928db67207a4308ee03ce9b4d96a2a1cdddc9c12f23f.jpg) + +![](images/28935a27a53c21efef7155827120bd3e3f9de2aea18861683cc1f286699dfec7.jpg) + +![](images/0f19485885fd452b7f2eddea24db2ed4962c20df34e7ead2b3726586baf2e904.jpg) +Figure 8: Evaluation of prior algorithms—mean and $95\%$ CI over 500 bandit rollouts, with $K = 8$ policy arms subsampled from 20 trained policies each rollout. The $x$ -axis denotes the number of bandit pulls, whilst the $y$ -axis denotes the true expected score of the estimated best arm after $x$ pulls. + +# C Evaluation Benchmarks in Prior Work + +Table 1: Evaluations performed in the papers introducing the offline RL algorithms we consider. A "√" indicates complete evaluation, "~" indicates a partial evaluation, and "-" indicates that the domain was not evaluated. MuJoCo locomotion is the most widely studied domain, although random and expert datasets are often omitted. Atari experiments are limited to only 5 datasets (Breakout, Qbert, Pong, Seaquest, and Asterix). Notably, the model-based offline RL works referenced here only evaluate on locomotion tasks, which may explain their dramatic performance collapse on non-locomotion tasks. + +
AlgorithmD4RL Fu et al. [20]
LocomotionAdroitKitchenMaze2dAntMazeMinigridCarlaFlowAtari
CQL [12]~✓✓-✓---~
DT [33]~-------~
EDAC [13]✓✓-------
IQL [21]~✓✓-✓----
ReBRAC [5]✓✓✓✓✓----
SAC-N [13]✓✓-------
TD3-BC [3]✓---~----
COMBO [14]~--------
MOPO [2]~--------
MOReL [11]~--------
+ +# D Distractor Policy Phenomenon + +Here, we show additional observations from the analysis of distractor policies in Section 3.3. + +![](images/9c8585266f9a8e4ccd23201bc53a74b252d5fa11f1f7a7c5e53640908fad5ccb.jpg) +(a) Probability of preferring a distractor policy (inside dashed orange lines in Figure 4) against the number of pulls (mean over 100K random policy orderings). The probability of preferring an unstable policy increases over time. + +![](images/5d38e44844f085c90f9f5dd7ea57204f40ab828cd329010fcc74fe08b12ff960.jpg) +(b) The number of subsampled policies influences evaluation behaviour—as the number of policies increases, we observe a greater dip in selected-policy performance from our UCB bandit. This is due to the presence of distractor policies (Figure 4), which achieve higher peak performance with a lower mean. + +# E Code Philosophy + +# E.1 Single-file + +We follow the community's preference for single-file algorithm implementations with integrated loggers and evaluations [17, 23, 35, 24]. All of our model-free algorithm implementations are self-contained, with every object necessary to set the hyperparameters, run the training loop, and evaluate the policy included in a single file. As model-based methods typically run sequential dynamics and policy training phases, we implement a single-file dynamics training script that saves trained model checkpoints. These can then be imported by any of the policy training scripts for the model-based algorithms. + +# E.2 Consistent + +Even within the same library, algorithm implementations often differ in boilerplate code. We change the minimum number of lines between implementations to control for implementation differences and help developers. Guided by the design genealogy illustrated in Figure 10a, we first ensure the single file implementation of the base algorithms like BC and SAC-N is clear and concise (Figure 10b) and then make minimal differences from their algorithmic ancestors (Figure 10c). + +Figure 11a shows the minimal differences between clean implementations of each algorithm, and Figure 11b shows the line differences from CQL. We acknowledge that prior implementations do not directly seek to minimize the differences between single-file implementations, but we believe it to be a beneficial feature for research. See Figure 12 and Figure 13 for a more complete illustration of the full code. + +![](images/6878996047cead0247323cd20b0ff5ba99e2d58b8cb6c04475bccbc6b9f63be3.jpg) +(a) Genealogy of algorithms. + +![](images/b6ea598e0fa6bc5cdf894759b1ad131f36b8997a8ac4ffd4e87026b9298d791f.jpg) +(b) Length of root algorithms. + +![](images/2d716346a3538de6e0fd6b437c091c224be5a2e9fa564fe2989312655bae54ef.jpg) +(c) Length of diff from parent. + +```python +algorithm: str = "sac_n" +algorithm: str = "edac" +# --- Update critics --- +@jax.value_and_grad +@partial(jax.value_and_grad, has(aux=True) +def_q_loss_fn(param): + q_pred = q_apply_fn(param, batch_obs, batch.action) + return jnp(square((q_pred - jnpexpand_dims(target, -1))).sum(-1).mean()) + value_loss = jnp(square((q_pred - jnpexpand_dims(target, -1))) + value_loss = value_loss.sum(-1).mean() + diversity_loss =jax.vmap_diversity_loss_fn)(batch_obs, batch.action) + diversity_loss = (1 / (args.num criterics - 1)) * diversity_loss.mean() + critic_loss = value_loss + args.eta * diversity_loss + return critic_loss, (value_loss, diversity_loss) + critic_loss, critic_grad = _q_loss_fn(agent_state.VEC_q.params) + (critic_loss, (value_loss, diversity_loss)), critic_grad = _q_loss_fn( agent_state.VEC_q.params) +updated_q = agent_state.VEC_q.apply_gradients(grades=crtic_grad) +agent_state = agent_state.replace(vec_q=updated_q) +``` + +(a) Using command line tool diff on our implementations of SAC-N and EDAC. + +![](images/d03e571a182acdfafa1372fc273fce94be1543f0587575e1a8dc07b0973cc8a2.jpg) +Figure 10: We provide clean and consistent single-file implementations, as demonstrated by compact implementations and minimal differences between algorithms. +Figure 11: Analysis of algorithmic differences between offline RL implementations. + +(b) Implementation length difference of each algorithm from CQL in their respective repository. + +![](images/5922564d1acabc7083e5c046d6bb286c479bd91058b2e45b14abe0473798d32b.jpg) +Figure 12: All code edits across implementations, from left to right: SAC-N, CQL, and EDAC. + +![](images/16b19e1d142e480a3c0274495cb6ba4269f1d5f1d7d914ff63e703e0cc1c37b4.jpg) +Figure 13: Full code difference for SAC-N, EDAC, and CQL from left to right. The code for the final evaluation loop is omitted to illustrate the consistency of the algorithm implementations. + +# F Reimplementation Training Time + +Table 2: Speed up from our JAX implementations, training time in minutes. Algorithms trained for 1M update steps on HalfCheetah—medium—expert using a single L40S GPU. Our library, Unifloral, is the fastest across the board. + +
AlgorithmOfflineRL-KitCORLJAX-CORLUnifloral
BC19.815.0—1.7
TD3-BC56.142.56.93.1
IQL79.765.05.24.0
ReBRAC—8.7—6.8
SAC-N107.598.8—7.7
CQL203.9180.320.79.8
EDAC127.1113.0—20.8
MOPO168.1——14.0
MORel———14.0
COMBO289.6——22.0
+ +# G Results Reproduction + +Table 3: Performance of our algorithm reimplementations over 5 training seeds, Mean±Std. + +
Env.DatasetBCCOMBOCQLEDACIQLMOPOMORELReBRACSAC-NTD3-BC
HalfCheetahExpert93.0 ± 0.489.5 ± 9.33.3 ± 1.32.3 ± 0.096.3 ± 0.362.7 ± 19.143.0 ± 27.2106.3 ± 0.998.8 ± 2.898.0 ± 0.8
Medium42.5 ± 0.272.2 ± 1.563.9 ± 1.152.2 ± 28.048.5 ± 0.472.8 ± 0.972.1 ± 1.665.6 ± 1.365.2 ± 1.448.6 ± 0.3
Medium-Expert59.4 ± 10.993.6 ± 4.766.1 ± 8.3102.8 ± 1.192.3 ± 3.180.9 ± 19.263.2 ± 6.8104.5 ± 2.3103.4 ± 5.692.9 ± 3.5
Medium-Replay37.3 ± 2.054.4 ± 13.655.2 ± 1.155.8 ± 1.043.8 ± 0.569.0 ± 1.565.4 ± 3.549.1 ± 0.857.4 ± 1.344.8 ± 0.5
Random2.2 ± 0.034.1 ± 1.630.7 ± 1.116.8 ± 13.312.5 ± 3.030.5 ± 1.031.8 ± 3.016.9 ± 17.826.6 ± 1.012.0 ± 1.6
HopperExpert109.5 ± 3.312.5 ± 15.31.4 ± 0.34.9 ± 0.2105.5 ± 4.52.2 ± 0.810.6 ± 6.8108.2 ± 4.393.8 ± 12.2109.4 ± 3.1
Medium55.7 ± 4.83.1 ± 0.47.6 ± 0.4100.8 ± 1.764.7 ± 5.646.6 ± 51.127.0 ± 10.4101.8 ± 0.875.2 ± 36.062.3 ± 4.9
Medium-Expert53.6 ± 4.42.8 ± 0.512.2 ± 3.0109.9 ± 0.3108.4 ± 4.925.2 ± 47.277.0 ± 44.4108.0 ± 3.490.5 ± 22.1105.2 ± 9.3
Medium-Replay25.0 ± 5.328.1 ± 26.7103.0 ± 0.3101.2 ± 0.473.5 ± 7.586.3 ± 28.447.4 ± 13.884.4 ± 26.8101.9 ± 0.451.1 ± 24.0
Random4.9 ± 4.827.0 ± 8.622.0 ± 12.822.6 ± 15.27.3 ± 0.131.4 ± 0.021.9 ± 13.07.8 ± 1.226.6 ± 10.58.4 ± 0.7
Walker2dExpert108.5 ± 0.222.6 ± 24.02.4 ± 2.479.0 ± 45.3112.7 ± 0.555.5 ± 10.719.4 ± 21.3112.4 ± 0.13.2 ± 2.2110.3 ± 0.3
Medium63.8 ± 9.884.5 ± 0.487.9 ± 0.675.1 ± 40.984.0 ± 2.081.3 ± 2.616.4 ± 36.984.3 ± 2.387.9 ± 0.684.5 ± 0.7
Medium-Expert108.1 ± 0.4101.2 ± 0.988.9 ± 36.3112.9 ± 0.7111.8 ± 0.3110.0 ± 1.521.7 ± 48.8111.6 ± 0.5114.8 ± 0.7110.1 ± 0.5
Medium-Replay23.8 ± 11.376.5 ± 2.079.1 ± 1.686.9 ± 1.582.8 ± 3.911.7 ± 3.3-0.2 ± 0.082.7 ± 5.382.3 ± 1.678.4 ± 4.0
Random0.9 ± 0.43.4 ± 2.69.1 ± 4.92.0 ± 0.04.4 ± 0.84.3 ± 6.30.3 ± 0.317.8 ± 8.920.7 ± 1.20.3 ± 0.4
+ +Table 3 presents the results achieved by our method reimplementations on locomotion datasets, matching the performance of prior implementations [17]. For our library's completeness, we also implement the Decision Transformer (DT) [33] using the hyperparameters from CORL [17]. +Table 4: Performance of our Decision Transformer implementation over 5 training seeds, Mean±Std. + +
Env.DatasetDecision Transformer
HalfCheetahExpert92.9 ± 0.1
Medium42.8 ± 0.5
Medium-Expert92.5 ± 0.2
Medium-Replay37.8 ± 1.3
HopperExpert110.2 ± 1.5
Medium61.3 ± 5.4
Medium-Expert111.3 ± 0.5
Medium-Replay25.5 ± 8.8
Walker2dExpert108.4 ± 0.2
Medium71.4 ± 6.03
Medium-Expert108.1 ± 0.3
Medium-Replay53.24± 13.7
+ +# H Unifloral Hyperparameters + +Table 5: Hyperparameters of prior algorithms in Unifloral—light gray values indicate inactive settings. + +
HyperparameterIQLSAC-NEDACTD3-BCReBRAC
Batch size2562562562561024
Actor learning rate3e-43e-43e-43e-41e-3
Critic learning rate3e-43e-43e-43e-41e-3
Learning rate schedulecosineconstantconstantconstantconstant
Discount factor γ0.990.990.990.990.99
Polyak step size0.0050.0050.0050.0050.005
Normalize observationsTrueFalseFalseTrueFalse
Actor layers23323
Actor hidden size256256256256256
Actor layer normalizationFalseFalseFalseFalseTrue
Deterministic policyFalseFalseFalseTrueTrue
Deterministic evalTrueFalseFalseFalseFalse
Apply tanh to meanTrueFalseFalseTrueTrue
Learn action stdTrueFalseFalseFalseFalse
Log std min-20.0-5.0-5.0-5.0-5.0
Log std max2.02.02.02.02.0
# of critics2[5-200][10-50]22
Critic layers23323
Critic hidden size256256256256256
Critic layer normalizationFalseFalseFalseFalseTrue
Actor BC coefficient1.00.00.01.0[5e-4-1.0]
Actor Q coefficient0.01.01.0[1.0-4.0]1.0
Use Q target in actorFalseFalseFalseFalseFalse
Normalize Q lossFalseFalseFalseTrueTrue
Q aggregation methodminminminfirstmin
Use AWRTrueFalseFalseFalseFalse
AWR temperature[0.5-10.0]1.01.01.01.0
AWR advantage clip100.0100.0100.0100.0100.0
Critic BC coefficient0.00.00.00.0[0-0.1]
# of critic updates per step11122
Diversity coefficient0.00.0[0.0-1e3]0.00.0
Policy noise0.00.00.00.20.2
Noise clip0.00.00.00.50.5
Use target actorFalseFalseFalseTrueTrue
Use entropy lossFalseTrueTrueFalseFalse
Actor entropy coefficient0.01.01.00.00.0
Critic entropy coefficient0.01.01.00.00.0
Use value targetFalseFalseFalseFalseFalse
Value expectile[0.5-0.9]0.80.80.80.8
+ +# I Unified Algorithm Details + +In this section we write out the different design decisions in a unified notation. + +# I.1 Critic Objective + +First, we compute the value target using one of two methods, selectable via the method configuration: + +$$ +v _ {t + 1} = \left\{ \begin{array}{l} v \left(s _ {t + 1}\right) \\ \min _ {n = 1} ^ {N} q _ {n} ^ {\prime} \left(s _ {t + 1}, \operatorname {c l i p} \left(\hat {a} _ {t + 1} + \operatorname {c l i p} \left(\epsilon , \epsilon_ {\min }, \epsilon_ {\max }\right), a _ {\min }, a _ {\max }\right)\right) \end{array} , \right. \tag {4} +$$ + +where $N$ is the number of ensemble members, $v$ is a value function trained with expectile regression (as in IQL [21]), $\hat{a}_{t + 1} \sim \pi(a_{t + 1}|s_{t + 1})$ is an action sampled from $\pi$ (or a Polyak averaged target policy), $\epsilon \sim \mathcal{N}(0,\sigma^2)$ is random action noise with standard deviation $\sigma$ , and $\epsilon_{\min}$ , $\epsilon_{\max}$ , $a_{\min}$ , and $a_{\max}$ are clipping ranges. The value target is then augmented with behaviour cloning and entropy terms (coefficients $\alpha_{\mathrm{BC}}$ and $\alpha_{\mathcal{H}}$ ), defined as + +$$ +\hat {v} _ {t + 1} = v _ {t + 1} + \alpha_ {\mathrm {B C}} \cdot \left(\tilde {a} _ {t + 1} - a _ {t + 1}\right) + \alpha_ {\mathcal {H}} \cdot \mathcal {H} (\pi (\cdot | s _ {t + 1})), \tag {5} +$$ + +which is then used to compute the value loss, + +$$ +\mathcal {L} _ {v} = \sum_ {n = 1} ^ {N} \left(q _ {n} \left(s _ {t}, a _ {t}\right) - \left(r + (1 - d) \cdot \gamma \cdot \hat {v} _ {t + 1}\right)\right) ^ {2}. \tag {6} +$$ + +Finally, we add the critic diversity loss term from EDAC [13] with coefficient $\alpha_{\mathrm{div}}$ , giving the final critic loss + +$$ +\mathcal {L} _ {\text {c r i t i c}} = \mathcal {L} _ {v} + \frac {\alpha_ {\text {d i v}}}{N - 1} \cdot \sum_ {1 \leq i \neq j \leq N} \left\langle \nabla_ {a _ {t}} q _ {i} \left(s _ {t}, a _ {t}\right), \nabla_ {a _ {t}} q _ {j} \left(s _ {t}, a _ {t}\right) \right\rangle . \tag {7} +$$ + +# I.2 Actor Objective + +We write the actor loss as: + +$$ +\mathcal {L} _ {\text {a c t o r}} = \beta_ {q} \cdot \mathcal {L} _ {q} + \beta_ {\mathrm {B C}} \cdot \mathcal {L} _ {\mathrm {B C}} - \beta_ {\mathcal {H}} \cdot \mathcal {H} (\pi (\cdot | s _ {t})). \tag {8} +$$ + +This consists of $q$ loss $\mathcal{L}_q$ , behaviour cloning loss $\mathcal{L}_{\mathrm{BC}}$ , and policy entropy $\mathcal{H}(\cdot)$ , with coefficients $\beta_q, \beta_{\mathrm{BC}}, \beta_{\mathcal{H}} \in \mathbb{R}$ controlling the weight of these terms. + +The first term, $\mathcal{L}_q$ is defined simply by a selectable aggregation function over the $q$ -network ensemble, with the minimum being the most common choice, + +$$ +\mathcal {L} _ {q} = \left\{ \begin{array}{l} - \min _ {n = 1} ^ {N} \left(q _ {n} \left(s _ {t}, a _ {t}\right)\right) \\ - \frac {1}{N} \sum_ {n = 1} ^ {N} q _ {n} \left(s _ {t}, a _ {t}\right) \\ - q _ {0} \left(s _ {t}, a _ {t}\right) \end{array} . \right. \tag {9} +$$ + +This term may also be normalized across the batch in order to stabilize learning. The second term, $\mathcal{L}_{\mathrm{BC}}$ , is most commonly defined as the distance $d$ between the target policy and dataset action, being the mean squared error for deterministic policies or log-probability for stochastic policies. However, some methods use advantage weighted regularization (AWR), which further weights this loss by the clipped and exponentiated advantage of the behaviour policy action in order to clone only positive actions within the dataset. Therefore, this term has the following variants: + +$$ +d = \left\{ \begin{array}{l} {\left(a _ {t} - \hat {a} _ {t}\right) ^ {2}} \\ - \log \pi \left(a _ {t} \mid s _ {t}\right) \end{array} , \quad \mathcal {L} _ {\mathrm {B C}} = \left\{ \begin{array}{l} d \\ d \cdot \min \left(A _ {\max }, e ^ {\eta \cdot \left(q \left(s _ {t}, a _ {t}\right) - V \left(s _ {t}\right)\right)}\right) \end{array} , \right. \right. \tag {10} +$$ + +where $\eta$ and $A_{\mathrm{max}}$ are the temperature and maximum value for exponential advantage. + +# I.3 Dynamics Modelling + +We use an ensemble of dynamics models $\hat{D}_{\theta} = \{\hat{D}_{\theta}^{1},\hat{D}_{\theta}^{2},\dots,\hat{D}_{\theta}^{M}\}$ , where each $\hat{D}_{\theta}^{i}$ is trained to predict state transitions and rewards. Following MOPO, we penalize the agent for going to states where the ensemble disagreement is higher, as measured by the standard deviation of the model's predictions. This is used to penalize the reward during policy optimization with a pessimism coefficient $\eta$ , + +$$ +\hat {R} \left(s _ {t}, a _ {t}\right) = \frac {1}{M} \sum_ {m = 1} ^ {M} R _ {\theta} ^ {m} \left(s _ {t}, a _ {t}\right) - \eta \cdot \sigma \left(\hat {\boldsymbol {D}} _ {\theta} ^ {\Delta s} \left(s _ {t}, a _ {t}\right)\right), \tag {11} +$$ + +where $\sigma (\hat{D}_{\theta}^{\Delta s}(s_t,a_t))$ represents the standard deviation across the models' state-change predictions and $R_{\theta}^{m}(s_{t},a_{t})$ is reward prediction of the $m$ -th ensemble member. + +# J Complete TD3-AWR Results + +![](images/33702e209a30a5ebae84af3c50b6d7060baf768ccd89570deccc20f3e12e54b7.jpg) + +![](images/dd7e15467acd88a2870449b541fcf7ed8690c4c6800e310bfc009e13baec1f4e.jpg) + +![](images/803509b74e36b8f5af8596a83d549181e343f574aa695892e0707701d6664099.jpg) + +![](images/de094ba0e19c7a40b50f8dfa01ad9c7c06d2b7203ca9800eb927664709826e59.jpg) + +![](images/7ff79cefecd3ce40ff32ea70c95f2878b51e274fac44cd01dc2e299edbf08c95.jpg) + +![](images/8170f7a628f4d54a0d6007ce6b521a14a728a32f2e5b3ab9593f312641db4bef.jpg) + +![](images/059c4a5c32d445786682b5345a6b41321e360e0553886898ae6fd363e1762154.jpg) + +![](images/5468fc10b8ce8bbc4091f3c4688d807f3bcceaad3968ebb2502581687fb08bfb.jpg) + +![](images/1aa62d96d9d722890258e3adfac15ac9e1c9a55524440c86441d61c86c55ce5e.jpg) + +![](images/fbae809030277cc1b4807af20c24f9fc7295e0491db5daae9d8fc1a76e614e47.jpg) +Figure 14: Full comparison of TD3-AWR to prior model-based methods across all datasets. + +# K Complete MoBRAC Results + +![](images/ada65a2782e1b4ac1c399f4b8b8eaea2c6da08d5e8d97b26750cb7de450f0eb0.jpg) + +![](images/8323996b3f9e5424303580a9d16f19c9802c4daae843e4ad8aecd2ab4b4e749b.jpg) + +![](images/508b5d3b2a9e2e541bcc7c1eb1721f3cb98c2a3126bbc2161ba20411eaff0977.jpg) + +![](images/d49462e4f0615164ab14622d34041c8754b2d8180e7d426ba986c6ea2f46fea2.jpg) + +![](images/c973a7f082083241058cb2bb4476cb7c3578620c7b797fc95404478c86ba59ab.jpg) + +![](images/5081c540788d6d91df982085193528d22bf2c00abf1ede6a803790cb9724baed.jpg) + +![](images/1f577c935e4d105e877b08880b9a0c101b65721678eceac5fc52d4db0dc52d3a.jpg) + +![](images/e3a5fe1c2dd71b099ac0a0908644cdaeb58098462686abd1a6652b48b410fbb2.jpg) +Number of policy evaluations + +![](images/4b5b56ec58f0b6cf6595dd3ffb05e85b8cce1f6807f2b49d99ff1b93fa7bd5ad.jpg) + +![](images/e1fecf4dfc0ebd5ea1b27a30fdf69b6494580b7258a12fd63de4455f659e3e56.jpg) +Figure 15: Full comparison of MoBRAC to prior model-based methods across all datasets. + +# L Full Related Work + +In this section, we describe the prior work related to our evaluation procedure, implementation, and unified algorithm. We implement a comprehensive selection of offline RL algorithms, for which more information can be found in Appendix A. + +# L.1 Evaluation Regimes for Offline RL + +The challenge of hyperparameter tuning in RL spans various domains. Wang et al. [36] discuss offline tuning and the practical risks of deploying policies of unknown quality in the real world, whilst Paine et al. [4] directly tackle this issue, estimating the zero-shot performance of offline-trained policies without any prior online interactions. Their evaluation is limited to behavioural cloning [31, BC] and two critic-based methods, which have since been outperformed by modern algorithms. Konyushova et al. [7] extend this procedure with an online phase, using a UCB-based bandit to investigate policy selection over multiple online evaluations. Further highlighting these challenges, Smith et al. [10] propose a procedure where offline evaluation methods are first calibrated using policies of known quality, evaluating on D4RL [20] locomotion tasks. Unlike their work, we evaluate across the D4RL suite and introduce a procedure that eliminates the need for reference policies or additional hyperparameters. Matsushima et al. [8] present a variant of offline RL that uses a limited number of online deployments to update the dataset and iteratively train offline to match the performance of online methods, introducing an online deployment frequency hyperparameter. Kurenkov and Kolesnikov [9] address the practice of unreported online evaluations for hyperparameter tuning, demonstrating how the performance of each algorithm changes with the number of online evaluations. Unlike our procedure, they assume a low-variance estimate of a policy's true performance each evaluation but still conclude that BC outperforms all baselines. + +# L.2 Open-Source Implementations + +Offline RL Inspiring our implementation, Clean Offline RL [17, CORL] provides single-file implementations of model-free offline RL methods in PyTorch. JAX-CORL [23] is a JAX-based port of CORL, albeit with a limited range of only model-free algorithms, slower training time than our implementations, and lacking our evaluation procedure and code consistency. OfflineRLKit [16] and d3rlpy [37] implement a range of offline RL methods and feature both model-based and model-free methods. Although the repository has transparent class inheritance and polymorphism, it lacks any further attempt at algorithmic unification. + +Online RL StableBaselines3 [38] is a set of reliable RL algorithm implementations in PyTorch with the aim of abstracting away training and deployment through an object-oriented interface. SpinningUp [39] is a similar, education-oriented effort of jointly implementing various online RL algorithms. CleanRL [24] follows a different design philosophy with method-focused, single-file implementations of online RL algorithms in PyTorch and JAX. PureJaxRL [35] also follows the single-file approach and is implemented in JAX. Rejax [40] is a popular multi-file JAX-based implementation of PureJaxRL with extensive logging integration and a selection of SOTA methods. + +CleanRL, CORL, and JAX-CORL provide clear and accessible logs of their final runs, a standard of reproducibility we plan to uphold throughout every release of our work. + +# L.3 Method Unification + +Our unified algorithm, Unifloral, is heavily inspired by prior work that also seeks to ablate and unify a range of methods. Lu et al. [15] investigate the key components of model-based offline RL algorithms to find an optimized algorithm that outperforms all model-based baselines. Sikchi et al. [27] cast multiple offline RL methods in the same dual optimization framework and use this unification to categorize them in regularized policy learning and pessimistic value learning. Prudencio et al. [6] provide a survey of offline RL, focused on elucidating the taxonomy and disambiguating the contributions of each algorithm. In online RL, Hessel et al. [25] combine independent components of Deep $Q$ -network algorithms into a unified algorithm, Rainbow, reaching SOTA in the Atari 2600 benchmark. Muesli [26] examines the combination of policy optimization and model-based methods. \ No newline at end of file diff --git a/NeurIPS/2025/A Clean Slate for Offline Reinforcement Learning/images.zip b/NeurIPS/2025/A Clean Slate for Offline Reinforcement Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..3d28715fbeaa4015290350565bf9725acc571f55 --- /dev/null +++ b/NeurIPS/2025/A Clean Slate for Offline Reinforcement Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7e17c2135c3f2ca14ffe3824f581b0ede7cf0fb23571593292272137867d5e8a +size 1289368 diff --git a/NeurIPS/2025/A Clean Slate for Offline Reinforcement Learning/layout.json b/NeurIPS/2025/A Clean Slate for Offline Reinforcement Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..73e60566bb7b8da9200f8ec7cfc177992826a6bf --- /dev/null +++ b/NeurIPS/2025/A Clean Slate for Offline Reinforcement Learning/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:60f9c29e5bf0c3ad2964ca8f3afbe9d69250fbcc76a29f337c9fd45c22c8117b +size 876060 diff --git a/NeurIPS/2025/A Closed-Form Solution for Fast and Reliable Adaptive Testing/96193993-4132-41b2-b62e-f648615fafc7_content_list.json b/NeurIPS/2025/A Closed-Form Solution for Fast and Reliable Adaptive Testing/96193993-4132-41b2-b62e-f648615fafc7_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..3c04aae4bc2a8f51d9e3ac90d299b19474d6babd --- /dev/null +++ b/NeurIPS/2025/A Closed-Form Solution for Fast and Reliable Adaptive Testing/96193993-4132-41b2-b62e-f648615fafc7_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:255c5a6fad4fed220c5e9880c32f80a65188346e5bbfdde6bbee3e36705d5dfd +size 163650 diff --git a/NeurIPS/2025/A Closed-Form Solution for Fast and Reliable Adaptive Testing/96193993-4132-41b2-b62e-f648615fafc7_model.json b/NeurIPS/2025/A Closed-Form Solution for Fast and Reliable Adaptive Testing/96193993-4132-41b2-b62e-f648615fafc7_model.json new file mode 100644 index 0000000000000000000000000000000000000000..23a1bab95e413c532f1e7cd0041c3e6d1b6de235 --- /dev/null +++ b/NeurIPS/2025/A Closed-Form Solution for Fast and Reliable Adaptive Testing/96193993-4132-41b2-b62e-f648615fafc7_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f736ed236d48f92613ff971a4b7699d5e745ac140d2efd2a13d0ecdab9561a1b +size 205063 diff --git a/NeurIPS/2025/A Closed-Form Solution for Fast and Reliable Adaptive Testing/96193993-4132-41b2-b62e-f648615fafc7_origin.pdf b/NeurIPS/2025/A Closed-Form Solution for Fast and Reliable Adaptive Testing/96193993-4132-41b2-b62e-f648615fafc7_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..62d82e041f213cdbdf0d64d2aff0847ed1a00b8d --- /dev/null +++ b/NeurIPS/2025/A Closed-Form Solution for Fast and Reliable Adaptive Testing/96193993-4132-41b2-b62e-f648615fafc7_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b7db5c3e3eac8e7c8bac26b99181598d21657ae3328b497a41278dfd286b030f +size 746721 diff --git a/NeurIPS/2025/A Closed-Form Solution for Fast and Reliable Adaptive Testing/full.md b/NeurIPS/2025/A Closed-Form Solution for Fast and Reliable Adaptive Testing/full.md new file mode 100644 index 0000000000000000000000000000000000000000..8dcaec58eecee77c29aeff767fd46eb72dc84983 --- /dev/null +++ b/NeurIPS/2025/A Closed-Form Solution for Fast and Reliable Adaptive Testing/full.md @@ -0,0 +1,808 @@ +# A Closed-Form Solution for Fast and Reliable Adaptive Testing + +Yan Zhuang $^{1}$ , Chenye Ke $^{2}$ , Zirui Liu $^{1}$ , Qi Liu $^{1,3}$ , Yuting Ning $^{4}$ , Zhenya Huang $^{1,3}$ , Weizhe Huang $^{1}$ , Qingyang Mao $^{1}$ , Shijin Wang $^{5}$ + +1: State Key Laboratory of Cognitive Intelligence, University of Science and Technology of China +2: Anhui University 3: Institute of Artificial Intelligence, Hefei Comprehensive National Science Center +4: Ohio State University 5: iFLYTEK Co., Ltd +{zykb, liuzirui, hwz871982879, maoqy0503}@mail.ustc.edu.cn, kecy@stu.ahu.edu.cn, ning.151@osu.edu, {qiliuql, huangzhy}@ustc.edu.cn, sjwang3@iflytek.com + +# Abstract + +Human ability estimation is essential for educational assessment, career advancement, and professional certification. Adaptive Testing systems can improve estimation efficiency by selecting fewer, targeted questions, and are widely used in exams, e.g., GRE, GMAT, and Duolingo English Test. However, selecting an optimal subset of questions remains a challenging nested optimization problem. Existing methods rely on costly approximations or data-intensive training, making them unsuitable for today's large-scale and complex testing environments. Thus, we propose a Closed-Form solution for question subset selection in Adaptive Testing. It directly minimizes ability estimation error by reducing ability parameter's gradient bias while maintaining Hessian stability, which enables a simple greedy algorithm for question selection. Moreover, it can quantify the impact of human behavioral perturbations on ability estimation. Extensive experiments on large-scale educational datasets demonstrate that it reduces the number of required questions by $10\%$ compared to SOTA methods, while maintaining the same estimation accuracy. + +# 1 Introduction + +Accurate assessment of human abilities plays a crucial role in education, career advancement, and professional certification, directly influencing future opportunities. As a result, the demand for effective and efficient assessment methodologies has grown significantly [1, 2, 3]. Traditional paper-and-pencil tests require examinees to answer a large number of questions, leading to cognitive load and inefficiency. In contrast, Adaptive Testing has emerged as a highly efficient ability estimation approach and has been widely adopted in education systems, and has been successfully integrated into various standardized testing systems [4, 5]. + +The effectiveness of adaptive testing lies in a key insight: not all questions are equally valuable for estimating ability. To achieve efficiency while maintaining accuracy, an adaptive testing system relies on two key components: 1) Question selection algorithm - Identifying and selecting the most informative subset of questions from the full question pool; 2) Item Response Theory (IRT) - A psychometric framework [6] that models the relationship between an examinee's latent ability $\theta$ and their observed responses (correct/incorrect). IRT serves as the "user model" for estimating ability based on response data to the selected questions. + +From a machine learning perspective, the overall adaptive testing process can be formulated as a subset selection problem that seeks to minimize the error of ability estimation [7, 8]: Given a large question pool $V$ , selecting a question subset $S \subseteq V$ for an examinee to answer such that the ability estimate $\theta_{S}$ (inferred from responses to $S$ ) is as close as possible to the true (or optimal) ability $\theta^{*}$ : + +$$ +\min _ {S \subseteq V} \| \theta_ {S} - \theta^ {*} \|, \quad \text {s . t .} \quad \theta_ {S} = \arg \min _ {\theta \in \Theta} \sum_ {i \in S} \ell_ {i} (\theta), \tag {1} +$$ + +where $\ell_i(\theta)$ denotes the cross-entropy loss associated with the response to question $i$ , and $\theta$ represents the ability parameter modeled by IRT. Obviously, adaptive testing is a complex nested optimization w.r.t. the subset variable $S$ , requiring iterative updates: the outer loop selects the optimal subset $S$ (often represented as a sparse selection vector [8]), while the inner loop estimates the ability parameter via supervised learning. + +Given its complexity, recent works often rely on data-driven meta-learning [9], or reinforcement learning [10, 2] to derive the question selection policy. However, these approaches introduce significant computational overhead and may amplify biases present in the data [9]. Even latest heuristic algorithms [7, 11] still require approximating and matching gradients across the entire ability parameter space $\Theta$ , leading to prohibitively high complexity. These limitations are especially critical in real-world online assessments, e.g., the Duolingo English Test, GRE Online, and remote certifications, which involve massive item pools, diverse examinees, and complex user behavior. These settings demand adaptive testing systems that are interpretable, robust, and efficient enough for real-time operations [12, 13]. + +To address these, this paper proposes a fundamental shift in the optimization paradigm of adaptive testing. For the first time, we derive a closed-form solution for the unknown subset variable $S$ , referred to as CFAT (Closed-Form expression for Adaptive Testing). It allows us to directly solve for the optimal subset without iterative sampling or complex nested optimization. Specifically, we successfully quantify the ability estimation error and demonstrate that it can be interpreted as minimizing the gradient bias while maintaining a stable Hessian structure. Furthermore, we prove that the objective function exhibits approximate submodularity, enabling the use of a simple greedy algorithm to efficiently select the subset. + +Beyond improving question selection, such closed-form formulation allows us to quantify the impact of human behavioral perturbations (e.g., guessing and slipping) on ability estimation. CFAT ultimately enables a bias correction mechanism for more reliable assessments. By fundamentally shifting the optimization paradigm of adaptive testing, CFAT uses statistical learning principles for efficient, direct computation. Experiments on three educational datasets demonstrate that our method reduces the number of required test questions by $10\%$ compared to the best baseline, under the same estimation accuracy. Moreover, CFAT achieves at least a $12\times$ improvement in selection efficiency (computation time) over latest methods. It can also exhibit higher robustness in high-noise scenarios, accurately recovering ability estimates and improving prediction reliability. + +# 2 Background and Related Works + +Adaptive testing has been widely adopted in human ability assessment especially in education, and has gradually been incorporated into high-stakes examinations. To achieve both accuracy and efficiency, adaptive testing typically consists of two key components: IRT and question selection algorithms: + +(1) Item Response Theory (IRT). IRT serves as a user model that captures the relationship between an examinee's ability and their responses [4]. Widely used in various large-scale assessments such as OECD/PISA, a common example is the two-parameter logistic (2PL) model, which defines the probability of a correct response to question $i$ as: $p(\text{correct}) = \sigma (\alpha_i(\theta -\beta_i))$ , where $\alpha_{i}$ and $\beta_{i}$ represent the discrimination and difficulty parameters, respectively. These question parameters are pre-calibrated [14], while the examinee's ability $\theta$ is estimated during testing. IRT models are interpretable: higher ability implies higher probability of success on items of fixed difficulty. Extensions include multidimensional IRT [15] and neural cognitive diagnosis models [16, 17, 18], which capture more complex interactions. All these methods rely on maximum likelihood estimation (minimizing cross-entropy loss) to estimate ability parameters from observed response data. + +(2) Selection Algorithm. This is the core of achieving efficient assessment, as it determines a valuable subset for estimating examinee ability in IRT. Traditional algorithms rely on statistical heuristics + +based on information measures, such as Fisher information [14], KL information [19] and various improved information metrics [20, 21, 22], to guide selection. Alternatively, active learning methods select informative questions based on question diversity and uncertainty [23]. Recently, to directly solve the nested optimizations, researchers have increasingly adopted data-driven approaches, e.g., reinforcement learning and meta-learning, to optimize subset selection [10, 9, 2, 8]. These methods iteratively train a policy (often represented as a neural network) from large-scale response data. + +In this work, we aim to bypass the nested optimization by deriving a closed-form expression for the estimation error w.r.t. the selected subset. It allows us to determine the optimal subset directly. Compared to data-driven/neural network-based approaches, this statistical method eliminates the need for extensive training. Compared to the latest gradient-based heuristic algorithms [7, 11], it incorporates second-order gradient (Hessian matrix) information, meanwhile, mitigating the impact of guessing and mistakes on ability estimation. Furthermore, CFAT is up to $12 \times$ more efficient than these SOTA heuristic methods, making it highly practical for real-time human assessments. + +# 3 Method + +Adaptive testing estimates ability efficiently by selecting a small, informative subset $S \subseteq V$ from a larger question pool $V$ . It can reduce test length while maintaining accuracy. + +Problem Statement. Formally, an examinee responds to the selected subset $S$ , producing $\{(q_1, y_1), \dots, (q_{|S|}, y_{|S|})\}$ , where $S = \{q_i\}_{i=1}^{|S|} \subseteq V$ is the question set selected by the adaptive selection algorithm, and $y_i \in \{0, 1\}$ denotes the response label, with 1 representing a correct response and 0 otherwise. The examinee's ability is then estimated by minimizing cross-entropy loss $\ell$ over $S$ . + +$$ +\theta_ {S} = \arg \min _ {\theta} \sum_ {i \in S} \ell_ {i} (\theta) = \arg \min _ {\theta} \sum_ {i \in S} - \log p _ {\theta} \left(q _ {i}, y _ {i}\right), \tag {2} +$$ + +where $p_{\theta}(q_i, y_i)$ represents the probability of observing response $(q_i, y_i)$ from an examinee with ability $\theta$ . The precise form of $p_{\theta}$ depends on the IRT model. Assuming an examinee's true latent ability is denoted as $\theta^*$ , one can theoretically approximate it by minimizing the expected cross-entropy loss over the entire question pool: $\theta^* = \arg \min_{\theta} \sum_{i \in V} \ell_i(\theta)$ [7]. The objective of adaptive testing is to ensure that the estimated ability $\theta_S$ from the subset is as close as possible to $\theta^*$ (Figure 1): + +Definition 1 (Definition of Adaptive Testing). Given a fixed test length $T$ , the task is to select an optimal subset $S \subseteq V$ such that the ability estimate $\theta_S$ approximates the estimate $\theta^*$ . The adaptive testing task can be formulated to a nested optimization problem as follows: + +$$ +\min _ {| S | = T} \| \theta_ {S} - \theta^ {*} \|, \quad s. t. \quad \theta_ {S} = \arg \min _ {\theta} \sum_ {i \in S} \ell_ {i} (\theta). \tag {3} +$$ + +In the outer loop, the subset $S$ can be generated using a selection policy $\pi$ [10, 9, 2], or it can be treated as a trainable indicator or sparse selection vector that determines question selection [8]. In the inner loop, a base optimization algorithm estimates $\theta_{S}$ using the responses on the selected $S$ , following standard supervised learning principles. + +While reinforcement/meta-learning methods have shown promise in adaptive testing [1], they are often computationally intensive due to multi-step gradient descent and repeated backpropagation. This raises: Can we directly formulate $\|\theta_S - \theta^*\|$ and optimize $S$ without resorting to iterative meta-optimization? If the effect of question selection on ability estimation can be explicitly modeled, more efficient selection strategies may be possible. + +# 3.1 Avoid the Nested Optimization Trap + +The key challenge in reformulating is to establish a direct link between $\theta_S - \theta^*$ and $S$ , without relying on an inner-loop optimization (arg min). To achieve this, we can simplify the problem by framing it as an issue of parameter estimation under data reduction: Consider the pool $V$ as the full dataset for estimation, while $S$ is its selected subset. The problem then becomes: analyzing how removing a subset $Z$ (where $S = V \setminus Z$ ) affects the estimated ability. + +![](images/9df7d132cf1f55ca7a5a0c44284a277df75214720205fa59d869747d619a20d3.jpg) +Figure 1: Illustration of subset selection in adaptive testing. The full question pool $V$ is divided into a selected subset $S$ and a remainder $Z$ . The estimation error is approximated via first-order (gradient) and second-order (Hessian) terms, capturing $S$ 's representativeness and informativeness, respectively. + +Measuring the Impact of Question Reduction on the Ability Estimator. Obviously, the most direct approach would be to recompute/retrain the parameter estimate from scratch for each choice of $S$ , as done in the inner loop's minimization in Eq.(2). But that is computationally prohibitive. Thus, instead of outright removing questions from $V$ , we down-weight their influence in the ability estimation process. This leads to the definition of a perturbed estimator: + +$$ +\theta_ {S} ^ {\gamma} = \arg \min _ {\theta} \frac {1}{| V |} \sum_ {i \in V} \ell_ {i} (\theta) - \gamma \sum_ {i \in Z} \ell_ {i} (\theta), \tag {4} +$$ + +where $Z$ is the set of down-weighted (or "removed") questions, and $\gamma \in [0,1 / |V|]$ . This formulation reduces the contribution of response to $Z$ to the total loss, thereby approximating the effect of excluding them from estimation. For a first-order approximation, we expand the gradient of the loss function evaluated at $\theta_S^\gamma$ using a Taylor expansion of Eq.(4) around $\theta^*$ . Since $\theta_S^\gamma$ is a minimizer, its gradient is approximately zero: + +$$ +0 \approx \frac {1}{| V |} \sum_ {i \in V} \nabla \ell_ {i} \left(\theta^ {*}\right) - \gamma \sum_ {i \in Z} \nabla \ell_ {i} \left(\theta^ {*}\right) + \left(\frac {1}{| V |} \sum_ {i \in V} \nabla^ {2} \ell_ {i} \left(\theta^ {*}\right) - \gamma \sum_ {i \in Z} \nabla^ {2} \ell_ {i} \left(\theta^ {*}\right)\right) \left(\theta_ {S} ^ {\gamma} - \theta^ {*}\right). \tag {5} +$$ + +In particular, when we set $Z = V \setminus S$ and choose $\gamma = 1 / |V|$ , the perturbed estimator exactly recovers the ability estimate based on the subset $S$ , i.e., $\theta_S^\gamma = \theta_S$ . Since $\theta^*$ satisfies the optimality condition $\sum_{i \in V} \nabla \ell_i(\theta^*) = 0$ , we obtain: + +$$ +\theta_ {S} ^ {\gamma} - \theta^ {*} = \theta_ {S} - \theta^ {*} \approx - \mathcal {H} ^ {- 1} (S, \theta^ {*}) \sum_ {i \in S} \nabla \ell_ {i} (\theta^ {*}), \tag {6} +$$ + +where $\mathcal{H}(S,\theta^{*}) = \sum_{i\in S}\nabla^{2}\ell_{i}(\theta^{*})$ denotes the Hessian of the loss function for ability estimation, and $\mathcal{H}^{-1}$ denotes its inverse. Here, $\mathcal{H}(S,\theta^{*})$ is invertible, which holds under standard regularity conditions in IRT [8, 11]. For complex neural network-based models, where computing the exact Hessian is often intractable, we adopt a quasi-Newton approximation (details are provided in Appendix A). This result can be viewed as an extension of influence function theory [24, 25, 26], which originated in statistics in the 1970s. It characterizes how perturbations in the data affect an estimator. Here, we approximate the effect of selecting a subset $S$ without resorting to explicit re-optimization. + +Closed-Form Expression of Estimation Error. The key takeaway is that the error in ability estimation based on the selected subset $S$ admits a closed-form expression (Figure 1): + +Lemma 1. Let the true ability parameter be $\theta^{*}$ . When using IRT for ability estimation, the estimation error based on any subset $S \subseteq V$ can be directly computed as in Eq.(6). This allows for directly optimizing the selection of $S$ to minimize the estimation error without the need to recompute $\theta_{S}$ : + +$$ +\min _ {| S | = T} \| \theta_ {S} - \theta^ {*} \| \Rightarrow \min _ {| S | = T} \left\| \underbrace {\mathcal {H} ^ {- 1} (S , \theta^ {*})} _ {\text {s e c o n d - o r d e r}} \overbrace {\sum_ {i \in S} \nabla \ell_ {i} (\theta^ {*})} ^ {\text {f i r s t - o r d e r}} \right\|. \tag {7} +$$ + +This reformulated objective directly quantifies the influence of the selected subset $S$ on the estimation error, bypassing the need for re-optimization of $\theta_{S}$ in previous nested optimizations. The selection of $S$ balances two critical factors: simultaneously managing both bias minimization (first-order stability) and conditioning of the Hessian (second-order stability): + +Factor 1: First-Order Gradient Alignment. The term $\sum_{i\in S}\nabla \ell_i(\theta^*)$ captures the aggregate first-order (gradient) contribution of the selected questions. If this sum deviates significantly from zero, it introduces directional bias into the estimated parameter $\theta_S$ . Intuitively, the goal is to find a subset whose gradients "agree" with those of the full question pool. This ensures that the subset is representative of the entire pool in terms of gradient information, and does not skew the estimation. + +Factor 2: Second-Order Information Control. The Hessian inverse, $\mathcal{H}^{-1}(S,\theta^{*})$ , controls how the subset's curvature information influences estimation stability. The optimal subset must ensure that the Hessian remains well-conditioned while retaining crucial second-order information to stabilize parameter updates. Consider the case of IRT: the expected Hessian can be approximated by the Fisher information $\mathcal{I}$ , i.e., $\mathbb{E}[\mathcal{H}(S,\theta)] \approx -\sum_{i \in S} \mathcal{I}_i(\theta) = -\sum_{i \in S} \alpha_i^2 \cdot p_\theta(q_i,0) \cdot p_\theta(q_i,1)$ . This suggests that it tries to find informative questions with high discrimination ( $\alpha$ ) and maximum response uncertainty, e.g., $p(q_i,1) \approx 0.5^2$ + +Thus, the best subset is both diverse and informative—minimizing gradient bias while maintaining a stable Hessian structure—leading to efficient and reliable estimation. + +# 3.2 Approximate Optimization for Subset Selection + +Based on the above reformulated objective, we aim to select a subset $S$ that minimizes the set function: $\min f(S) = \min \left\| \mathcal{H}^{-1}(S,\theta^{*})\sum_{i\in S}\nabla \ell_{i}(\theta^{*})\right\|$ . This problem is combinatorial and generally NP-hard. Exhaustively searching for the optimal subset is computationally infeasible for large pool due to the exponential number of possible combinations. + +Fortunately, we observe that this objective function exhibits a diminishing marginal gain property, which aligns with the concept of submodularity [29]. Submodularity is a useful concept in combinatorial optimization problems that plays a crucial role in designing efficient approximation [30, 31, 32], e.g., greedy algorithm. More precisely, the objective function $f(S)$ is approximately submodular, a property referred to as $\epsilon$ -submodularity. This property implies that the incremental benefit of adding an element $x$ decreases as the set grows: + +Theorem 1 ( $\epsilon$ -Submodularity of the Set Function). Estimating the ability $\theta$ using $IRT$ , the loss function $\ell(\theta)$ is $\mu$ -strongly convex. Assume that the gradient norm and Hessian's spectral norm are bounded: $\|\nabla_{\theta}\ell_{i}(\theta)\| \leq G$ and $\|\nabla_{\theta}^{2}\ell_{i}(\theta)\| \leq H$ . The objective $f(S) = \left\|\mathcal{H}^{-1}(S,\theta^{*})\sum_{i\in S}\nabla \ell_{i}(\theta^{*})\right\|$ is $\epsilon$ -submodular, and $\epsilon = \frac{2G(\mu + H)}{\mu^2|A|} + \frac{2HG}{\mu^2|A|^2}$ , i.e., for any subsets $A \subseteq B \subseteq V$ : + +$$ +f (A \cup \{x \}) - f (A) \geq f (B \cup \{x \}) - f (B) - \left(\frac {2 G (\mu + H)}{\mu^ {2} | A |} + \frac {2 H G}{\mu^ {2} | A | ^ {2}}\right). \tag {8} +$$ + +The proofs can be found in Appendix B. The $\epsilon = \frac{2G(\mu + H)}{\mu^2|A|} +\frac{2HG}{\mu^2|A|^2}$ bound decreases as $|A|$ increases. This means the function becomes more submodular as the subset grows, which is intuitive—the marginal benefit of adding a new question becomes more stable as more questions are already selected. This bound provides theoretical justification for using greedy methods: if $\epsilon$ is small (e.g., due to large $|A|$ ), greedy selection will be near-optimal even though the function is not strictly submodular. + +Greedy Question Selection. Given that the objective $f(S)$ is $\epsilon$ -submodular, we can use a greedy algorithm to iteratively construct an optimal subset $S$ . The approximate submodularity property ensures that a greedy selection achieves a near-optimal solution with theoretically bounded suboptimality. + +For size-constrained minimization of $f(S)$ , a simple reverse greedy algorithm can be adopted. It sequentially selects elements that yield the smallest marginal increase in $f(S)$ . Specifically, it starts with an empty subset $S_0 = \emptyset$ . At each step $t$ , the question that minimizes the marginal gain is selected, formally given by $q_t = \arg \min_{q \in V \setminus S_{t-1}} (f(S_{t-1} \cup \{(q, y)\}) - f(S_{t-1}))$ . After selecting $q_t$ , the subset is updated as: $S_t = S_{t-1} \cup \{(q_t, y_t)\}$ . + +In practice, the parameter $\theta^{*}$ is unknown and we use the estimate $\theta^t$ obtained from $S_{t}$ . The objective function can be approximated: $f(S\mid \theta^t) = \left\| \mathcal{H}^{-1}(S,\theta^t)\sum_{i\in S}\nabla \ell_i(\theta^t)\right\|$ . Meanwhile, the true + +labels $y$ are also unobserved, we take the expectation over $y$ , selecting the next question: + +$$ +q _ {t} = \underset {q \in V \backslash S _ {t - 1}} {\arg \min } \mathbb {E} _ {y} [ f (S _ {t - 1} \cup \{q, y \} \mid \theta^ {t - 1}) ]. \tag {9} +$$ + +The sequential selection process continues until the selected questions reach a predefined maximum size $T$ , corresponding to the termination condition of the test. Based on the asymptotic unbiasedness of MLE, we find an upper bound on the approximation error when substituting $\theta^t$ for $\theta^*$ . + +Lemma 2 (Approximation Error). When using IRT for ability estimation, the function $f(S)$ is Lipschitz continuous w.r.t. $\theta$ . With probability at least $1 - \delta$ , the approximation error incurred by using $\theta^t$ satisfies the upper bound: $|f(S|\theta^t) - f(S)|\leq \left(\frac{H}{\mu_1} +\frac{MG}{\mu_1\mu_2}\right)\frac{C(\delta)}{\sqrt{|S_t|}}$ , where $\mu_{1},\mu_{2},M,H, G$ , and $C$ are model-dependent constants characterizing the properties of the objective function. + +The proofs can be found in Appendix C. The substitution of $\theta^{*}$ with $\theta^t$ is justified due to the consistency and asymptotic normality of estimators. According to this bounded approximation error in Lemma 2, the error introduced by estimating $\theta^{*}$ diminishes at a rate of $O(|S_t|^{-1/2})$ , ensuring the robustness of the adaptive selection process. + +# 3.3 Bias Correction in Ability Estimation: Guessing and Slipping + +The idealized ability estimation above assumes that an examinee's responses accurately reflect their true ability. However, in practical testing, the observed response $y$ may not correspond perfectly to true ability due to guessing and slipping [1]. 1) Guessing: An examinee correctly answers a question they should not have been able to solve purely by chance. For example, if a multiple-choice question has three options, random guessing yields a $33.3\%$ success probability; 2) Slipping: An examinee fails to answer a question correctly despite having the ability to do so. It arises due to carelessness, misreading, or other lapses. + +Both factors induce label flipping in the observed responses $y \in \{0,1\}$ , leading to biased ability estimates $\theta_S$ that contain unpredictable noise. This distortion can be explicitly quantified within this CFAT framework: Consider a response $(q_m, y_m)$ affected by label flipping, resulting in the incorrect label $(q_m, 1 - y_m)$ . The corresponding loss becomes: $\widetilde{\ell}_m(\theta) = -(1 - y_m)\log p_\theta(q_m, 1) - y_m\log p_\theta(q_m, 0)$ . After incorporating the flipped response, the new ability estimate on $S$ is: $\theta_{S(m)}^{\gamma} = \arg \min_{\theta}\frac{1}{|S|}\sum_{i\in S}\ell_i(\theta) - \gamma \ell_m(\theta) + \gamma \widetilde{\ell}_m(\theta)$ . Note that this also applies a weighted adjustment rather than physically replacing the affected data. When $\gamma = \frac{1}{|S|}$ , the correction becomes equivalent to a full replacement of the original response. + +Applying a Taylor expansion around $\theta_{S}$ , similar to the derivations in Section 3.1, we approximate: + +$$ +\theta_ {S (m)} = \theta_ {S} + \left[ \mathcal {H} (S \backslash q _ {m}, \theta_ {S}) + \widetilde {\mathcal {H}} (q _ {m}, \theta_ {S}) \right] ^ {- 1} (1 - 2 y _ {m}) \nabla \log \frac {p _ {\theta} (q _ {m} , 1)}{p _ {\theta} (q _ {m} , 0)}, \tag {10} +$$ + +where $\widetilde{\mathcal{H}}(q_m, \theta_S) = \nabla^2 \widetilde{\ell}_m(\theta_S)$ . The term $\Delta \theta_{S(m)} = \theta_{S(m)} - \theta_S$ provides a quantitative measure of how a flipped response skews the estimate. Even if we cannot pinpoint the specific flipped samples, analyzing the expected effect enables us to understand the direction and magnitude of the systematic bias caused by response errors. See Appendix D for a detailed derivation of Eq.(10). + +Thus, instead of relying on the potentially biased estimator $\theta_{S}$ , we introduce a bias-corrected ability estimate by subtracting the expected distortion: $\theta_{S} - \mathbb{E}[\Delta \theta] = \theta_{S} - \sum_{m\in S}\left[\pi_{g}(1 - y_{m}) + \pi_{s}y_{m}\right]\Delta \theta_{S(m)}$ , where $\pi_{g}$ is the guessing probability, capturing the likelihood of obtaining a correct response by chance, and $\pi_{s}$ is the slipping probability, representing the likelihood of incorrect responses despite having the requisite ability. The complete CFAT framework is shown in Algorithm 1 + +# 4 Experiments + +Evaluation Tasks. To assess the efficiency of question selection algorithms in adaptive testing, we evaluate the accuracy of ability estimation under a fixed test length. Specifically, we compare the final estimated ability $\theta_{S}$ , where $S$ represents the selected question subset chosen by different selection algorithms. The evaluation is conducted across two primary tasks [9, 7]: 1) Student Performance + +Algorithm 1: The proposed framework CFAT +```txt +Require: $V$ - Question pool, $p_{\theta}$ - Parameterized probability model (IRT or neural network), $\pi_g$ - Guessing probability, $\pi_s$ - Slipping probability +``` + +Initialize: Initialize the ability estimate $\theta^0$ and responses data $S_0\gets \emptyset$ + +1 for $t = 1$ to $T$ do +```txt +2 Select the next question $q_{t}$ by minimizing the set function: $q_{t} = \arg \min_{q\in V\setminus S_{t - 1}}\mathbb{E}_{y}f(S_{t - 1}\cup \{q,y\} \mid \theta^{t - 1})$ +3 Obtain the examinee's response label $y_{t}$ .. $S_{t}\gets S_{t - 1}\cup \{(q_{t},y_{t})\}$ +4 Update examinee's ability estimate: $\theta^t\gets \arg \min_{\theta \in \Theta}\sum_{i\in S_t}\ell_i(\theta)$ +5 Apply bias correction to adjust for response errors: $\theta^t\gets \theta^t -\sum_{m\in S_t}\left[\pi_g(1 - y_m) + \pi_s y_m\right]\Delta \theta_{S_t(m)}$ +``` + +Output: The examinee's ability estimate $\theta_S = \theta^T$ using the responses on the selected $S$ . + +Prediction: Using the estimated $\theta_{S}$ , predicting students' responses (correct/incorrect) on a held-out test set and measure predictive performance using Accuracy and AUC; 2) Ability Estimation Error: In a simulation setting, the ground-truth ability $\theta^{*}$ is constructed and simulate students' response behavior during testing. We then compute the estimation error using the Mean Squared Error (MSE) $\mathbb{E}||\theta_S - \theta^* ||^2$ + +Experimental Implementation Details. We set the maximum test length to $|S| = T = 20$ , consistent with typical adaptive tests. All methods are implemented in PyTorch and trained on a Tesla V100 GPU. Hyperparameters are tuned via grid search, with batch size 64, learning rate 0.001, and behavioral noise parameters $\pi_g = 0.002$ , $\pi_s = 0.001$ . Optimization is performed using Adam. + +Following [9, 1], we split examinees into $70\%$ training, $20\%$ validation, and $10\%$ testing. The training set is used to estimate question parameters and train some data-driven models. During validation and testing, we simulate adaptive testing: Specifically, for the student performance prediction task, each examinee's responses are divided into a candidate set $V_{i}$ (for question selection and ability estimation) and a meta set $M_{i}$ (for evaluation via Accuracy/AUC). At each step, a question is selected from $V_{i}$ , ability is updated, and performance is evaluated on $M_{i}$ . For ability estimation error, ground-truth abilities $\theta^{*}$ are estimated from full responses, allowing simulated examinees to answer any question in $V$ for direct error computation. + +Datasets. We conduct experiments on three widely used educational testing benchmark datasets: ASSIST, NIPS-EDU, and EXAM: ASSIST [33] is derived from the online educational platform ASSISTments and contains examinees' practice logs on mathematics; NeurIPS-EDU [34] originates from the NeurIPS 2020 Education Challenge, comprising a large-scale dataset collected from examinees' responses to questions on Eedi, an educational platform. EXAM is a dataset from iFLYTEK Co., Ltd. that records junior high school students' performances on mathematical exams. The implementation code is available on: https://github.com/54zy/CFAT. + +Compared Approaches. For ability estimation, we consider both classical IRT model and neural network-based approaches: NeuralCDM [35], a flexible framework that generalizes various IRT and cognitive diagnosis models, e.g., MIRT [36] and Matrix Factorization [37, 38]. The objective of our experiments is to compare the proposed selection algorithm against existing selection methods in terms of their impact on ability estimation. Thus, we evaluate the following SOTA algorithms as baselines: Random Selection serves as a benchmark by selecting questions uniformly at random, providing a reference for assessing the improvements achieved by other algorithms; Fisher Information [14] and KL Information [19] are classical methods that prioritize questions based on their informativeness; MAAT [23] uses active learning to balance uncertainty and diversity. BOBCAT [9] and UATS [8] apply meta-learning to solve the nested selection problem via cross-entropy minimization. NCAT [10] and GMOCAT [2] frame selection as reinforcement learning, leveraging transformers and GNNs to train a data-driven selection policy. BECAT [7] uses a greedy heuristic based on first-order gradient approximation, between the selected subset and the entire question pool. + +Table 1: The performances on Student Performance Prediction. It reports ACC and AUC at 5th, 10th, and 20th step (subset size). Panel 1 presents results based on the IRT model for ability estimation, while Panel 2 uses a neural network-based model (NeuralCDM). Note that information/uncertainty-based methods (e.g., Fisher) are not applicable to deep models. Bold values indicate statistically significant improvements ( $p$ -value $< 0.01$ ) over the best baseline. + +
MethodASSIST (ACC/AUC)NIPS-EDU (ACC/AUC)EXAM (ACC/AUC)
@5@10@20@5@10@20@5@10@20
Random70.89/70.7871.99/71.8473.02/72.4566.57/69.0268.11/71.4270.00/73.9077.58/70.3477.22/71.8380.33/74.09
Fisher71.87/71.2272.63/72.3073.11/73.5667.70/70.6270.59/73.5171.23/76.3377.35/70.5179.75/72.2583.03/75.90
KL71.95/71.3172.68/72.5073.13/73.5767.09/69.7169.29/73.3070.41/75.7377.37/70.5879.22/72.1183.01/75.73
MAAT72.11/71.2472.03/72.3873.20/73.0566.44/69.3169.10/71.1269.27/73.4075.27/70.3277.99/72.1280.12/73.67
BOBCAT72.33/71.7272.56/72.1873.78/73.3169.55/74.4170.99/75.6671.71/76.4480.61/68.2983.81/72.0283.44/72.82
NCAT72.22/71.6672.52/72.3873.83/73.5167.30/72.1170.68/75.8071.91/76.6680.92/70.7283.96/72.6783.88/74.19
UATS72.29/72.8272.04/72.7474.14/74.8467.58/73.3370.50/74.8271.84/76.5779.17/70.2282.33/73.2984.91/75.24
BECAT71.92/71.3473.01/72.7373.96/73.6366.98/73.1571.61/75.8572.00/76.8280.93/70.7483.80/72.8884.20/75.03
CFAT72.86/73.4873.37/73.2674.29/75.2269.62/74.5572.25/76.2273.87/78.0381.11/71.0384.13/73.8086.05/77.83
+ +
MethodASSIST (ACC/AUC)NIPS-EDU (ACC/AUC)EXAM (ACC/AUC)
@5@10@20@5@10@20@5@10@20
Random71.21/71.0272.53/72.0872.51/72.8367.13/69.3968.42/71.5170.59/74.9379.80/72.4878.33/74.5279.31/78.22
MAAT72.09/70.7472.31/72.0371.75/72.2967.83/70.0070.42/72.5870.63/75.8582.87/70.2282.55/74.2983.72/79.36
BOBCAT72.64/71.4672.77/72.7373.80/72.8271.02/76.1272.46/77.8273.42/79.0678.13/78.2878.13/81.4578.04/79.53
NCAT72.29/71.6472.62/72.3473.92/73.5670.43/74.1272.84/77.9273.44/79.0982.33/78.5483.13/81.4681.44/79.35
UATS73.02/72.3272.92/73.0573.16/72.7371.87/75.1373.13/78.1274.14/79.7081.26/77.1282.46/80.9283.79/80.82
BECAT72.30/71.6173.11/72.8774.13/73.7071.33/76.3173.07/78.2473.58/79.2682.84/78.7583.22/81.4984.77/79.70
CFAT74.13/72.9273.45/73.9874.53/74.3871.20/76.1974.43/78.3874.72/81.7783.33/80.9884.12/82.8785.12/81.66
+ +# 4.1 Experimental Results + +We evaluate the proposed CFAT framework on two core tasks (i.e., student performance prediction and ability estimation error) across three benchmark datasets. + +Task 1: Student Performance Prediction: This task assesses the efficiency of ability estimation in adaptive testing. Specifically, we compare the prediction accuracy of response labels (correct/incorrect) under different question selection strategies, where each algorithm selects the same number of questions. As shown in Table 1, we report the AUC and ACC scores for each method at the 5th, 10th, and 20th steps. The proposed CFAT consistently achieves the highest prediction accuracy under limited question settings. Notably, simple greedy algorithm CFAT outperforms neural network methods, e.g., reinforcement learning (NCAT) and meta-learning approaches (BOBCAT, UATS), by an average margin of $2\%$ in AUC. + +These results support our central claim: formulating the subset selection problem with a closed-form objective yields better performance than complex nested paradigm. Furthermore, CFAT outperforms the gradient-based BECAT. It highlights the advantage of incorporating second-order information (i.e., the Hessian matrix) over relying solely on first-order gradients. This superiority is observed both theoretically and empirically at scale. Although our theoretical derivation is grounded in the classical IRT model, the CFAT framework also demonstrates strong performance when applied to neural network ability estimation models (NeuralCDM). This suggests that our subset selection formulation and its approximation are generalizable and extensible across different modeling paradigms. + +Task 2: Ability Estimation Error: To evaluate the accuracy of ability estimation, we adopt a widely used simulation protocol in adaptive testing. Specifically, we treat the ability estimate derived from an examinee's full response data, denoted as $\theta^{*}$ , as the ground truth. During the testing process, this ground-truth ability allows us to simulate response labels for any question, while the tested algorithms only have access to observed response data and not the true ability. Figure 2 illustrates the estimation error $\| \theta^t - \theta^*\|$ over the testing process, where $\theta^t$ denotes the estimated ability at step $t$ . The results show that our proposed CFAT achieves comparable estimation accuracy using only $30\% - 45\%$ of the questions required by random selection. Compared to recent SOTA methods (e.g., UATS), CFAT reduces the number of required questions by at least $15\%$ . + +Although CFAT exhibits a relatively slower start in the early stages of testing, its estimation error decreases rapidly as more questions are selected. This initial lag is somewhat less favorable compared to other data-driven methods (e.g., meta learning) that are good at mitigating the cold-start problem [39]. These empirical observations align well with the analysis in Theorem 1, which demonstrates that the submodular nature guarantees the near-optimality of the greedy question selection algorithm as the selected subset grows. + +![](images/dc73c8c842a491b0bd311d8441bd6039d07af4611ea16bf791e53f3787be515b.jpg) +Figure 2: (a) Simulation results for ability estimation: MSE of ability estimation, $\mathbb{E}\| \theta^t -\theta_0\|^2$ under different subset sizes (steps) for five representative question selection algorithms. Results are averaged over 10 repetitions, with error bars indicating the standard deviation. (b) Average time required to select a single question for each method. + +![](images/7c518ef588e8a55571c0f33af30f4060f6e39676e2a37f2377bdc66cccb901fd.jpg) + +![](images/b767501f064a539942d55ab8a94a02813de5118c57e51454c979f9df2890e1a9.jpg) + +![](images/b59a72bc335efb01781c040d0c73e0067e9a98f15c13b268c18f816f2489a105.jpg) + +Figure 3: Characteristics of selected questions across different methods. We randomly sample 10 examinees and compare the question subsets selected by CFAT with several SOTA baselines. The distributions of question difficulty and discrimination parameters are visualized. +![](images/194a3ada9516bef46fe15ce5aa499c629f2ccbf697d006174fb2b8c2c76c6757.jpg) +All Questions ★ Selected Questions ——Examinee Ability (8) + +![](images/968f9321937b214c332b52e0cbcc1b5d3f61fe7fc3a578bfd9c5320d8c909563.jpg) + +![](images/3988592a7983dd61011bc67a2b248c5fb2746364118d1bdf0dd4716b89ad2ff0.jpg) + +![](images/4bbdfdf3c29f49be14ed3d3f8cd024db6ad20b90e8365960700dfe0a3c1ebcd3.jpg) + +Analysis of Computational Efficiency and Subset Characteristics. We compare the computational efficiency of different question selection algorithms to assess their practical applicability in large-scale testing (note: no acceleration techniques or engineering optimizations are applied). Specifically, in Figure 2(b), we report the average time required to select a single question. CFAT demonstrates higher efficiency compared to SOTA methods e.g., MAAT and BECAT-achieving approximately $12 \times$ speedup over BECAT. Notably, CFAT matches the speed of the classical Fisher information method, while simultaneously delivering at least a $20\%$ improvement in estimation accuracy, as evidenced in Figure 2(a). Meanwhile, Figure 3 illustrates the characteristics of question subsets selected by different methods, along with the true ability estimates $\theta^{*}$ of 10 randomly sampled examinees. As shown, the questions selected by CFAT tend to have higher discrimination and are well-aligned with the examinees' ability levels (i.e., question difficulty closely matches ability). In contrast, other methods often prioritize diverse questions, many of which are "outliers"-either too easy or too difficult for the examinees. Such low-discrimination or mismatched questions tend to be less informative and may hinder accurate ability estimation [40]. + +# Reliability under Guessing and Slipping Noise. + +In real-world scenarios, examinee's responses may be affected by guessing (label flipping $0\rightarrow 1$ ) or slipping (label flipping $1\to 0$ [41]. To model this, label noise is introduced in above simulation by flipping response labels with a certain probability. Table 2 illustrates the estimation error across different algorithms as the flipping probability increases. Previous approaches exhibit significant performance degradation under noise. In contrast, CFAT consistently maintains lower estimation error, outperforming its ablated version (CFAT w/o correction), which lacks the bias correction term. Notably, under high noise levels (e.g., $20\%$ label flipping), CFAT still achieves stable and accurate ability estimates. These results empirically validate our theoretical analysis in Section 3.3, highlighting the effectiveness of incorporating a correction term when estimating the ability $\theta_{S}$ . + +Table 2: MSE for different selection algorithms in ASSIST under varying levels of label perturbation (Step=20). Perturbation is applied to the examinee's response label. 'No Pert.' denotes MSE without any label noise. + +
MethodNo Pert.5% Pert.10% Pert.20% Pert.
Random0.37650.3827 (+0.0062)0.4936 (+0.1171)0.6681 (+0.2316)
KL0.35990.3638 (+0.0039)0.4744 (+0.1145)0.5869 (+0.2270)
BECAT0.36970.3741 (+0.0044)0.4814 (+0.1117)0.6005 (+0.2308)
GMOCAT0.23220.2375 (+0.0053)0.2956 (+0.0634)0.3478 (+0.1156)
CFAT (w/o correction)0.19620.2024 (+0.0062)0.3121 (+0.1159)0.4324 (+0.2362)
CFAT0.17380.1778 (+0.0040)0.2082 (+0.0344)0.2770 (+0.1032)
+ +# 5 Conclusion + +This paper addresses the subset selection problem in ability estimation: how to select a small question subset such that the estimated ability closely approximates the true ability. Instead of relying on the traditional nested optimization paradigm, we derive a closed-form objective that allows for direct optimization. It shows that a simple greedy algorithm can effectively solve this problem, and partially correct the bias of the ability estimator. Extensive experiments demonstrate that it is computationally efficient, yields more accurate ability estimates, better adapts to individuals, and remains robust under high-noise conditions. + +# Acknowledgements + +This research was supported by grants from the National Key Research and Development Program of China (Grant No. 2024YFC3308200, 62477044), the National Natural Science Foundation of China (62525606), the Key Technologies R & D Program of Anhui Province (No. 202423k09020039), the National Education Science Planning Project (Grant No. ZSA240466), and the Fundamental Research Funds for the Central Universities. + +# References + +[1] Qi Liu, Yan Zhuang, Haoyang Bi, Zhenya Huang, Weizhe Huang, Jiatong Li, Junhao Yu, Zirui Liu, Zirui Hu, Yuting Hong, et al. Survey of computerized adaptive testing: A machine learning perspective. arXiv preprint arXiv:2404.00712, 2024. +[2] Hangyu Wang, Ting Long, Liang Yin, Weinan Zhang, Wei Xia, Qichen Hong, Dingyin Xia, Ruiming Tang, and Yong Yu. Gmocat: A graph-enhanced multi-objective method for computerized adaptive testing. In Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 2279–2289, 2023. +[3] Chris Piech, Jonathan Bassen, Jonathan Huang, Surya Ganguli, Mehran Sahami, Leonidas J Guibas, and Jascha Sohl-Dickstein. Deep knowledge tracing. Advances in neural information processing systems, 28, 2015. +[4] Wim J Van der Linden and Peter J Pashley. Item selection and ability estimation in adaptive testing. In Computerized adaptive testing: Theory and practice, pages 1-25. Springer, 2000. +[5] Xiaoshan Yu, Chuan Qin, Dazhong Shen, Haiping Ma, Le Zhang, Xingyi Zhang, Hengshu Zhu, and Hui Xiong. Rdgt: enhancing group cognitive diagnosis with relation-guided dual-side graph transformer. IEEE Transactions on Knowledge and Data Engineering, 2024. +[6] Hua-Hua Chang. Psychometrics behind computerized adaptive testing. Psychometrika, 80(1):1-20, 2015. +[7] Yan Zhuang, Qi Liu, GuanHao Zhao, Zhenya Huang, Weizhe Huang, Zachary Pardos, Enhong Chen, Jinze Wu, and Xin Li. A bounded ability estimation for computerized adaptive testing. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. +[8] Junhao Yu, Yan Zhuang, Zhenya Huang, Qi Liu, Xin Li, LI Rui, and Enhong Chen. A unified adaptive testing system enabled by hierarchical structure search. In *Forty-first International Conference on Machine Learning*, 2024. +[9] Aritra Ghosh and Andrew Lan. Bobcat: Bilevel optimization-based computerized adaptive testing. In Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence, IJCAI-21, pages 2410–2417. International Joint Conferences on Artificial Intelligence Organization, 8 2021. +[10] Yan Zhuang, Qi Liu, Zhenya Huang, Zhi Li, Shuanghong Shen, and Haiping Ma. Fully adaptive framework: Neural computerized adaptive testing for online education. Proceedings of the AAAI Conference on Artificial Intelligence, 36(4):4734-4742, Jun. 2022. + +[11] Zirui Liu, Yan Zhuang, Qi Liu, Jiatong Li, Yuren Zhang, Zhenya Huang, Jinze Wu, and Shijin Wang. Computerized adaptive testing via collaborative ranking. Advances in Neural Information Processing Systems, 37:95488-95514, 2024. +[12] Micheline Chalhoub-Deville and Craig Deville. Computer adaptive testing in second language contexts. Annual Review of Applied Linguistics, 19:273–299, 1999. +[13] Yan Zhuang, Qi Liu, Zhenya Huang, Zhi Li, Binbin Jin, Haoyang Bi, Enhong Chen, and Shijin Wang. A robust computerized adaptive testing approach in educational question retrieval. In Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 416-426, 2022. +[14] Frederic M Lord. Applications of item response theory to practical testing problems. Routledge, 2012. +[15] Giles Hooker, Matthew Finkelman, and Armin Schwartzman. Paradoxical results in multidimensional item response theory. Psychometrika, 74(3):419-442, 2009. +[16] Fei Wang, Qi Liu, Enhong Chen, Zhenya Huang, Yuying Chen, Yu Yin, Zai Huang, and Shijin Wang. Neural cognitive diagnosis for intelligent education systems. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 6153-6161, 2020. +[17] Weibo Gao, Qi Liu, Linan Yue, Fangzhou Yao, Hao Wang, Yin Gu, et al. Collaborative cognitive diagnosis with disentangled representation learning for learner modeling. Advances in Neural Information Processing Systems, 37:562-588, 2024. +[18] Wei Song, Qi Liu, Qingyang Mao, Yiyan Wang, Weibo Gao, Zhenya Huang, Shijin Wang, Enhong Chen, et al. Towards accurate and fair cognitive diagnosis via monotonic data augmentation. Advances in Neural Information Processing Systems, 37:47767-47789, 2024. +[19] Hua-Hua Chang and Zhiliang Ying. A global information approach to computerized adaptive testing. Applied Psychological Measurement, 20(3):213-229, 1996. +[20] Lawrence M Rudner. An examination of decision-theory adaptive testing procedures. In annual meeting of the American Educational Research Association, 2002. +[21] Wim J van der Linden. Bayesian item selection criteria for adaptive testing. Psychometrika, 63(2):201-216, 1998. +[22] Wim JJ Veerkamp and Martijn PF Berger. Some new item selection criteria for adaptive testing. Journal of Educational and Behavioral Statistics, 22(2):203-226, 1997. +[23] Haoyang Bi, Haiping Ma, Zhenya Huang, Yu Yin, Qi Liu, Enhong Chen, Yu Su, and Shijin Wang. Quality meets diversity: A model-agnostic framework for computerized adaptive testing. In 2020 IEEE International Conference on Data Mining (ICDM), pages 42–51. IEEE, 2020. +[24] Frank R. Hampel. The influence curve and its role in robust estimation. Journal of the American Statistical Association, 69(346):383-393, 1974. +[25] L. A. Jaeckel. The infinitesimal jackknife. Unpublished memorandum, Bell Telephone Laboratories, Murray Hill, NJ, 1972. +[26] Pang Wei Koh and Percy Liang. Understanding black-box predictions via influence functions. In Doina Precup and Yee Whye Teh, editors, Proceedings of the 34th International Conference on Machine Learning, volume 70 of Proceedings of Machine Learning Research, pages 1885-1894. PMLR, 06-11 Aug 2017. +[27] Zhilei Wang, Pranjal Awasthi, Christoph Dann, Ayush Sekhari, and Claudio Gentile. Neural active learning with performance guarantees. Advances in Neural Information Processing Systems, 34:7510-7521, 2021. +[28] Alex Tamkin, Dat Nguyen, Salil Deshpande, Jesse Mu, and Noah Goodman. Active learning helps pretrained models learn the intended task. Advances in Neural Information Processing Systems, 35:28140-28153, 2022. + +[29] Abhimanyu Das and David Kempe. Approximate submodularity and its applications: Subset selection, sparse approximation and dictionary selection. Journal of Machine Learning Research, 19(3):1-34, 2018. +[30] Baharan Mirzasoleiman, Jeff Bilmes, and Jure Leskovec. Coresets for data-efficient training of machine learning models. In Hal Daumé III and Aarti Singh, editors, Proceedings of the 37th International Conference on Machine Learning, volume 119 of Proceedings of Machine Learning Research, pages 6950-6960. PMLR, 13-18 Jul 2020. +[31] Artin Tajdini, Lalit Jain, and Kevin G Jamieson. Nearly minimax optimal submodular maximization with bandit feedback. Advances in Neural Information Processing Systems, 37:96254-96281, 2024. +[32] Abir De and Soumen Chakrabarti. Neural estimation of submodular functions with applications to differentiable subset selection. Advances in Neural Information Processing Systems, 35:19537-19552, 2022. +[33] Zachary A Pardos, Ryan SJD Baker, Maria OCZ San Pedro, Sujith M Gowda, and Supreeth M Gowda. Affective states and state tests: Investigating how affect throughout the school year predicts end of year learning outcomes. In Proceedings of the third international conference on learning analytics and knowledge, pages 117–124, 2013. +[34] Zichao Wang, Angus Lamb, Evgeny Saveliev, Pashmina Cameron, Yordan Zaykov, José Miguel Hernández-Lobato, Richard E Turner, Richard G Baraniuk, Craig Barton, Simon Peyton Jones, Simon Woodhead, and Cheng Zhang. Diagnostic questions: The neurips 2020 education challenge. arXiv preprint arXiv:2007.12061, 2020. +[35] Fei Wang, Qi Liu, Enhong Chen, Zhenya Huang, Yu Yin, Shijin Wang, and Yu Su. Neuralcd: a general framework for cognitive diagnosis. IEEE Transactions on Knowledge and Data Engineering, 35(8):8312-8327, 2022. +[36] Mark D. Reckase. Multidimensional Item Response Theory Models, pages 79-112. Springer New York, New York, NY, 2009. +[37] Andreas Tscher. Collaborative filtering applied to educational data mining. Journal of Machine Learning Research, 2010. +[38] Michel Desmarais. Mapping question items to skills with nonnegative matrix factorization. Sigkdd Explorations, 13:30-36, 05 2012. +[39] Manasi Vartak, Arvind Thiagarajan, Conrado Miranda, Jeshua Bratman, and Hugo Larochelle. A meta-learning perspective on cold-start recommendations for items. Advances in neural information processing systems, 30, 2017. +[40] Pedro Rodriguez, Joe Barrow, Alexander Miserlis Hoyle, John P Lalor, Robin Jia, and Jordan Boyd-Graber. Evaluation examples are not equally informative: How should that change nlp leaderboards? In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 4486–4503, 2021. +[41] Sujith M Gowda, Jonathan P Rowe, Ryan Shaun Joazeiro de Baker, Min Chi, and Kenneth R Koedinger. Improving models of slipping, guessing, and moment-by-moment learning with estimates of skill difficulty. *EDM*, 2011:199-208, 2011. +[42] John E Dennis, Jr and Jorge J Moré. Quasi-newton methods, motivation and theory. SIAM review, 19(1):46-89, 1977. +[43] J. Shermen and WJ Morrison. Adjustment of an inverse matrix corresponding to changes in the elements of a given column or a given row of the original matrix. Annual Mathematical Statistics, 20:621-625, 1949. +[44] Ronald Aylmer Fisher. Theory of statistical estimation. In Mathematical proceedings of the Cambridge philosophical society, volume 22, pages 700-725. Cambridge University Press, 1925. + +# NeurIPS Paper Checklist + +# 1. Claims + +Question: Do the main claims made in the abstract and introduction accurately reflect the paper's contributions and scope? + +Answer: [Yes] + +Justification: The abstract and introduction clearly state the main contributions of the paper, including the formulation of a closed-form objective for subset selection in ability estimation and the development of an efficient greedy algorithm. These claims are supported by both theoretical analysis and extensive empirical results, as detailed in Sections 3 and 4. + +Guidelines: + +- The answer NA means that the abstract and introduction do not include the claims made in the paper. +- The abstract and/or introduction should clearly state the claims made, including the contributions made in the paper and important assumptions and limitations. A No or NA answer to this question will not be perceived well by the reviewers. +- The claims made should match theoretical and experimental results, and reflect how much the results can be expected to generalize to other settings. +- It is fine to include aspirational goals as motivation as long as it is clear that these goals are not attained by the paper. + +# 2. Limitations + +Question: Does the paper discuss the limitations of the work performed by the authors? + +Answer: [Yes] + +Justification: As discussed in Section 3.1, the proposed theoretical results are derived under the assumptions of IRT models, and may not directly hold for complex neural networks. However, we address this limitation by employing quasi-Newton approximations (Appendix A). Additionally, as noted in Section 4.1, while the computational cost of our method is not best, it achieves the best overall accuracy among the compared approaches. + +Guidelines: + +- The answer NA means that the paper has no limitation while the answer No means that the paper has limitations, but those are not discussed in the paper. +- The authors are encouraged to create a separate "Limitations" section in their paper. +- The paper should point out any strong assumptions and how robust the results are to violations of these assumptions (e.g., independence assumptions, noiseless settings, model well-specification, asymptotic approximations only holding locally). The authors should reflect on how these assumptions might be violated in practice and what the implications would be. +- The authors should reflect on the scope of the claims made, e.g., if the approach was only tested on a few datasets or with a few runs. In general, empirical results often depend on implicit assumptions, which should be articulated. +- The authors should reflect on the factors that influence the performance of the approach. For example, a facial recognition algorithm may perform poorly when image resolution is low or images are taken in low lighting. Or a speech-to-text system might not be used reliably to provide closed captions for online lectures because it fails to handle technical jargon. +- The authors should discuss the computational efficiency of the proposed algorithms and how they scale with dataset size. +- If applicable, the authors should discuss possible limitations of their approach to address problems of privacy and fairness. +- While the authors might fear that complete honesty about limitations might be used by reviewers as grounds for rejection, a worse outcome might be that reviewers discover limitations that aren't acknowledged in the paper. The authors should use their best judgment and recognize that individual actions in favor of transparency play an important role in developing norms that preserve the integrity of the community. Reviewers will be specifically instructed to not penalize honesty concerning limitations. + +# 3. Theory assumptions and proofs + +Question: For each theoretical result, does the paper provide the full set of assumptions and a complete (and correct) proof? + +Answer: [Yes] + +Justification: See Appendix B and C + +Guidelines: + +- The answer NA means that the paper does not include theoretical results. +- All the theorems, formulas, and proofs in the paper should be numbered and cross-referenced. +- All assumptions should be clearly stated or referenced in the statement of any theorems. +- The proofs can either appear in the main paper or the supplemental material, but if they appear in the supplemental material, the authors are encouraged to provide a short proof sketch to provide intuition. +- Inversely, any informal proof provided in the core of the paper should be complemented by formal proofs provided in appendix or supplemental material. +- Theorems and Lemmas that the proof relies upon should be properly referenced. + +# 4. Experimental result reproducibility + +Question: Does the paper fully disclose all the information needed to reproduce the main experimental results of the paper to the extent that it affects the main claims and/or conclusions of the paper (regardless of whether the code and data are provided or not)? + +Answer: [Yes] + +Justification: See Section 4. It provides all the details. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- If the paper includes experiments, a No answer to this question will not be perceived well by the reviewers: Making the paper reproducible is important, regardless of whether the code and data are provided or not. +- If the contribution is a dataset and/or model, the authors should describe the steps taken to make their results reproducible or verifiable. +- Depending on the contribution, reproducibility can be accomplished in various ways. For example, if the contribution is a novel architecture, describing the architecture fully might suffice, or if the contribution is a specific model and empirical evaluation, it may be necessary to either make it possible for others to replicate the model with the same dataset, or provide access to the model. In general, releasing code and data is often one good way to accomplish this, but reproducibility can also be provided via detailed instructions for how to replicate the results, access to a hosted model (e.g., in the case of a large language model), releasing of a model checkpoint, or other means that are appropriate to the research performed. +- While NeurIPS does not require releasing code, the conference does require all submissions to provide some reasonable avenue for reproducibility, which may depend on the nature of the contribution. For example +(a) If the contribution is primarily a new algorithm, the paper should make it clear how to reproduce that algorithm. +(b) If the contribution is primarily a new model architecture, the paper should describe the architecture clearly and fully. +(c) If the contribution is a new model (e.g., a large language model), then there should either be a way to access this model for reproducing the results or a way to reproduce the model (e.g., with an open-source dataset or instructions for how to construct the dataset). +(d) We recognize that reproducibility may be tricky in some cases, in which case authors are welcome to describe the particular way they provide for reproducibility. In the case of closed-source models, it may be that access to the model is limited in some way (e.g., to registered users), but it should be possible for other researchers to have some path to reproducing or verifying the results. + +# 5. Open access to data and code + +Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material? + +Answer: [Yes] + +Justification: We have uploaded the code to the anonymous link https://github.com/54zy/CFAT (See Section 4). + +Guidelines: + +- The answer NA means that paper does not include experiments requiring code. +- Please see the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details. +- While we encourage the release of code and data, we understand that this might not be possible, so "No" is an acceptable answer. Papers cannot be rejected simply for not including code, unless this is central to the contribution (e.g., for a new open-source benchmark). +- The instructions should contain the exact command and environment needed to run to reproduce the results. See the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details. +- The authors should provide instructions on data access and preparation, including how to access the raw data, preprocessed data, intermediate data, and generated data, etc. +- The authors should provide scripts to reproduce all experimental results for the new proposed method and baselines. If only a subset of experiments are reproducible, they should state which ones are omitted from the script and why. +- At submission time, to preserve anonymity, the authors should release anonymized versions (if applicable). +- Providing as much information as possible in supplemental material (appended to the paper) is recommended, but including URLs to data and code is permitted. + +# 6. Experimental setting/details + +Question: Does the paper specify all the training and test details (e.g., data splits, hyperparameters, how they were chosen, type of optimizer, etc.) necessary to understand the results? + +Answer: [Yes] + +Justification: See Section 4 + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The experimental setting should be presented in the core of the paper to a level of detail that is necessary to appreciate the results and make sense of them. +- The full details can be provided either with the code, in appendix, or as supplemental material. + +# 7. Experiment statistical significance + +Question: Does the paper report error bars suitably and correctly defined or other appropriate information about the statistical significance of the experiments? + +Answer: [Yes] + +Justification: The main results (Figure 2(a)) in Section 4.1 report the deviation over 10 repetitions. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The authors should answer "Yes" if the results are accompanied by error bars, confidence intervals, or statistical significance tests, at least for the experiments that support the main claims of the paper. + +- The factors of variability that the error bars are capturing should be clearly stated (for example, train/test split, initialization, random drawing of some parameter, or overall run with given experimental conditions). +- The method for calculating the error bars should be explained (closed form formula call to a library function, bootstrap, etc.) +- The assumptions made should be given (e.g., Normally distributed errors). +- It should be clear whether the error bar is the standard deviation or the standard error of the mean. +- It is OK to report 1-sigma error bars, but one should state it. The authors should preferably report a 2-sigma error bar than state that they have a $96\%$ CI, if the hypothesis of Normality of errors is not verified. +- For asymmetric distributions, the authors should be careful not to show in tables or figures symmetric error bars that would yield results that are out of range (e.g. negative error rates). +- If error bars are reported in tables or plots, The authors should explain in the text how they were calculated and reference the corresponding figures or tables in the text. + +# 8. Experiments compute resources + +Question: For each experiment, does the paper provide sufficient information on the computer resources (type of compute workers, memory, time of execution) needed to reproduce the experiments? + +Answer: [Yes] + +Justification: Figure 2(b) reports the time cost of each method. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The paper should indicate the type of compute workers CPU or GPU, internal cluster, or cloud provider, including relevant memory and storage. +- The paper should provide the amount of compute required for each of the individual experimental runs as well as estimate the total compute. +- The paper should disclose whether the full research project required more compute than the experiments reported in the paper (e.g., preliminary or failed experiments that didn't make it into the paper). + +# 9. Code of ethics + +Question: Does the research conducted in the paper conform, in every respect, with the NeurIPS Code of Ethics https://neurips.cc/public/EthicsGuidelines? + +Answer: [Yes] + +Justification: This work adheres to the NeurIPS Code of Ethics in all respects. + +Guidelines: + +The answer NA means that the authors have not reviewed the NeurIPS Code of Ethics. +- If the authors answer No, they should explain the special circumstances that require a deviation from the Code of Ethics. +- The authors should make sure to preserve anonymity (e.g., if there is a special consideration due to laws or regulations in their jurisdiction). + +# 10. Broader impacts + +Question: Does the paper discuss both potential positive societal impacts and negative societal impacts of the work performed? + +Answer: [Yes] + +Justification: The item selection process in adaptive testing is inherently personalized, and societal impacts such as fairness constitute a separate line of research within the field. Due to the page limitation, we put it into Appendix E + +Guidelines: + +- The answer NA means that there is no societal impact of the work performed. + +- If the authors answer NA or No, they should explain why their work has no societal impact or why the paper does not address societal impact. +- Examples of negative societal impacts include potential malicious or unintended uses (e.g., disinformation, generating fake profiles, surveillance), fairness considerations (e.g., deployment of technologies that could make decisions that unfairly impact specific groups), privacy considerations, and security considerations. +- The conference expects that many papers will be foundational research and not tied to particular applications, let alone deployments. However, if there is a direct path to any negative applications, the authors should point it out. For example, it is legitimate to point out that an improvement in the quality of generative models could be used to generate deepfakes for disinformation. On the other hand, it is not needed to point out that a generic algorithm for optimizing neural networks could enable people to train models that generate Deepfakes faster. +- The authors should consider possible harms that could arise when the technology is being used as intended and functioning correctly, harms that could arise when the technology is being used as intended but gives incorrect results, and harms following from (intentional or unintentional) misuse of the technology. +- If there are negative societal impacts, the authors could also discuss possible mitigation strategies (e.g., gated release of models, providing defenses in addition to attacks, mechanisms for monitoring misuse, mechanisms to monitor how a system learns from feedback over time, improving the efficiency and accessibility of ML). + +# 11. Safeguards + +Question: Does the paper describe safeguards that have been put in place for responsible release of data or models that have a high risk for misuse (e.g., pretrained language models, image generators, or scraped datasets)? + +Answer: [NA] + +Justification: The paper poses no such risks. + +Guidelines: + +- The answer NA means that the paper poses no such risks. +- Released models that have a high risk for misuse or dual-use should be released with necessary safeguards to allow for controlled use of the model, for example by requiring that users adhere to usage guidelines or restrictions to access the model or implementing safety filters. +- Datasets that have been scraped from the Internet could pose safety risks. The authors should describe how they avoided releasing unsafe images. +- We recognize that providing effective safeguards is challenging, and many papers do not require this, but we encourage authors to take this into account and make a best faith effort. + +# 12. Licenses for existing assets + +Question: Are the creators or original owners of assets (e.g., code, data, models), used in the paper, properly credited and are the license and terms of use explicitly mentioned and properly respected? + +Answer: [Yes] + +Justification: We use the NIPS-EDU and ASSIST, which is publicly available under the CC BY 4.0, MIT License. We have properly cited the original source in the paper and included the version and URL where applicable. + +Guidelines: + +- The answer NA means that the paper does not use existing assets. +- The authors should cite the original paper that produced the code package or dataset. +- The authors should state which version of the asset is used and, if possible, include a URL. +- The name of the license (e.g., CC-BY 4.0) should be included for each asset. + +- For scraped data from a particular source (e.g., website), the copyright and terms of service of that source should be provided. +- If assets are released, the license, copyright information, and terms of use in the package should be provided. For popular datasets, paperswithcode.com/datasets has curated licenses for some datasets. Their licensing guide can help determine the license of a dataset. +- For existing datasets that are re-packaged, both the original license and the license of the derived asset (if it has changed) should be provided. +- If this information is not available online, the authors are encouraged to reach out to the asset's creators. + +# 13. New assets + +Question: Are new assets introduced in the paper well documented and is the documentation provided alongside the assets? + +Answer: [Yes] + +Justification: We introduce a new EXAM as part of this work. At submission time, all links and files have been anonymized to preserve double-blind review. + +Guidelines: + +- The answer NA means that the paper does not release new assets. +- Researchers should communicate the details of the dataset/code/model as part of their submissions via structured templates. This includes details about training, license, limitations, etc. +- The paper should discuss whether and how consent was obtained from people whose asset is used. +- At submission time, remember to anonymize your assets (if applicable). You can either create an anonymized URL or include an anonymized zip file. + +# 14. Crowdsourcing and research with human subjects + +Question: For crowdsourcing experiments and research with human subjects, does the paper include the full text of instructions given to participants and screenshots, if applicable, as well as details about compensation (if any)? + +Answer: [NA] + +Justification: This paper does not involve crowdsourcing nor research with human subjects. + +Guidelines: + +- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. +- Including this information in the supplemental material is fine, but if the main contribution of the paper involves human subjects, then as much detail as possible should be included in the main paper. +- According to the NeurIPS Code of Ethics, workers involved in data collection, curation, or other labor should be paid at least the minimum wage in the country of the data collector. + +# 15. Institutional review board (IRB) approvals or equivalent for research with human subjects + +Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals (or an equivalent approval/review based on the requirements of your country or institution) were obtained? + +Answer: [NA] + +Justification: The paper does not involve crowdsourcing nor research with human subjects. + +Guidelines: + +- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. + +- We recognize that the procedures for this may vary significantly between institutions and locations, and we expect authors to adhere to the NeurIPS Code of Ethics and the guidelines for their institution. +- For initial submissions, do not include any information that would break anonymity (if applicable), such as the institution conducting the review. + +# 16. Declaration of LLM usage + +Question: Does the paper describe the usage of LLMs if it is an important, original, or non-standard component of the core methods in this research? Note that if the LLM is used only for writing, editing, or formatting purposes and does not impact the core methodology, scientific rigorousness, or originality of the research, declaration is not required. + +Answer: [NA] + +Justification: The core method development in this research does not involve LLMs as any important, original, or non-standard components. + +Guidelines: + +- The answer NA means that the core method development in this research does not involve LLMs as any important, original, or non-standard components. +- Please refer to our LLM policy (https://neurips.cc/Conferences/2025/LLM) for what should or should not be described. + +# A Complete Algorithmic Procedure of CFAT + +This section presents the complete optimization process of CFAT in practical adaptive testing. Algorithm 2 provides a detailed illustration of the gradient-based ability estimation procedure. However, during the actual question selection phase of CFAT, computing the inverse of the Hessian matrix is required. While this is tractable for traditional IRT models, it becomes computationally infeasible for neural network-based models due to the high dimensionality and complexity of their parameter spaces. To address this, Algorithm 3 introduces an efficient approximation of the Hessian inverse using a quasi-Newton method [42]. This enables the practical deployment of CFAT in neural network settings, and forms the basis of our complete CFAT algorithm tailored for deep learning adaptive testing systems. + +Algorithm 2: Full Procedure of CFAT +Require: $V$ - Question pool, $p_{\theta}$ - Parameterized probability model (IRT or neural network), $\pi_g$ - Guessing probability, $\pi_s$ - Slipping probability. Initialize: Initialize the ability estimate $\theta^0$ and responses data $S_0\gets \emptyset$ +1 for $t = 1$ to $T$ do Select the next question $q_{t}$ by minimizing the set function: $q_{t} = \arg \min_{q\in V\backslash S_{t - 1}}\mathbb{E}_{y}f(S_{t - 1}\cup \{q,y\} \mid \theta^{t - 1}).$ +3 Obtain the examinee's response label $y_{t}\colon S_{t}\leftarrow S_{t - 1}\cup \{(q_{t},y_{t})\}$ . Initialize examinee's ability estimate $\theta_0^t\gets \theta_K^{t - 1}$ Update examinee's ability estimate: for $k = 1$ to $K$ do Update $\theta_k^t:\theta_k^t\gets \theta_{k - 1}^t -\alpha \nabla \ell_i(\theta_{k - 1}^t)$ Apply bias correction to adjust for response errors: $\theta_K^t\gets \theta_K^t -\sum_{m\in S_t}[\pi_g(1 - y_m) + \pi_s y_m]\Delta \theta_{S_t}(m)$ + +Output: The examinee's ability estimate $\theta_{S} = \theta_{K}^{T}$ using the responses on the selected $S$ . + +Algorithm 3: Full Procedure of CFAT (Approximate) +Require: $V$ - Question pool, $p_{\theta}$ - Parameterized probability model (IRT or neural network), $\pi_g$ Guessing probability, $\pi_s$ - Slipping probability, $\alpha$ - learning rate. Initialize: Initialize the ability estimate $\theta^0$ and responses data $S_0\gets \emptyset$ , the approximation of the inverse of Hessian matrix $\mathcal{H}_K^{-1^{(0)}}\gets I$ and the examinee's ability estimate $\theta_K^0$ . +for $t = 1$ to $T$ do +Let $\mathcal{H}^{-1}\leftarrow \mathcal{H}_K^{-1(t - 1)}$ and select the next question $q_{t}$ by minimizing the set function: $q_{t} = \arg \min_{q\in V\backslash S_{t - 1}}\mathbb{E}_{y}f(S_{t - 1}\cup \{q,y\} \mid \theta^{t - 1})$ +Obtain the examinee's response label $y_{t}\colon S_{t}\gets S_{t - 1}\cup \{(q_{t},y_{t})\}$ +Initialize examinee's ability estimate $\theta_0^t\gets \theta_K^{t - 1}$ +Update examinee's ability estimate: +for $k = 1$ to $K$ do Calculate the search direction: $d_k\gets -\mathcal{H}_{k - 1}^{-1}(t)\sum_{i\in S}\nabla \ell_i(\theta_{k - 1}^t)$ Update $\theta_k^t\colon \theta_k^t\gets \theta_{k - 1}^t +\alpha d_k$ Let $u_{k}\gets \theta_{k}^{t} - \theta_{k - 1}^{t}$ and $v_{k}\gets \sum_{i\in S}\nabla \ell_{i}(\theta_{k}^{t}) - \sum_{i\in S}\nabla \ell_{i}(\theta_{k - 1}^{t})$ Update the approximation of the inverse of Hessian matrix $H$ .. $\mathcal{H}_k^{-1(t)}\gets \mathcal{H}_{k - 1}^{-1}(t) + \frac{u_ku_k^{\mathrm{T}}}{u_k^{\mathrm{T}}v_k} -\frac{\mathcal{H}_{k - 1}^{-1}(t)v_kv_k^{\mathrm{T}}\mathcal{H}_{k - 1}^{-1}(t)}{v_k^{\mathrm{T}}\mathcal{H}_{k - 1}^{-1}(t)v_k}$ Apply bias correction to adjust for response errors: $\theta_K^t\gets \theta_K^t -\sum_{m\in S_t}[\pi_g(1 - y_m) + \pi_s y_m]\Delta \theta_{S_t}(m)$ + +Output: The examinee's ability estimate $\theta_{S} = \theta_{K}^{T}$ using the responses on the selected $S$ . + +# B Proofs of Theorem 1 + +Theorem 1 ( $\epsilon$ -Submodularity of the Set Function). Estimating the ability parameter $\theta$ using IRT, the loss function $\ell(\theta)$ is assumed to be $\mu$ -strongly convex [7]. Assume that the gradient norm and Hessian's spectral norm are bounded by $\|\nabla_{\theta}\ell_i(\theta)\| \leq G$ and $\|\nabla_{\theta}^2\ell_i(\theta)\| \leq H$ . The subset selection objective $f(S) = \left\|\mathcal{H}^{-1}(S,\theta^{*})\sum_{i \in S}\nabla\ell_i(\theta^{*})\right\|$ is $\epsilon$ -submodular, and $\epsilon = \frac{2G(\mu + H)}{\mu^2|A|} + \frac{2HG}{\mu^2|A|^2}$ , i.e., for any subsets $A \subseteq B \subseteq V$ : + +$$ +f (A \cup \{x \}) - f (A) \geq f (B \cup \{x \}) - f (B) - \left(\frac {2 G (\mu + H)}{\mu^ {2} | A |} + \frac {2 H G}{\mu^ {2} | A | ^ {2}}\right). \tag {11} +$$ + +Proof. Based on Lemma 1, the objective function can be formulated as: + +$$ +f (S) = \left\| \mathcal {H} ^ {- 1} (S, \theta^ {*}) \sum_ {i \in S} \nabla \ell_ {i} (\theta^ {*}) \right\|, +$$ + +where $\mathcal{H}(S,\theta) = \sum_{i\in S}\nabla^2\ell_i(\theta)$ is Hessian matrix. + +We assume the following boundedness conditions on the loss function $\ell_i(\theta)$ over the set $V$ : 1) The gradient norm is upper-bounded: $\| \nabla_{\theta}l_{i}(\theta)\| \leq G$ ; 2) The spectral norm of the Hessian is also bounded: $\| \nabla^2\ell_i(\theta)\| \leq H$ . + +For IRT-based ability estimation, the loss function $\ell(\theta)$ is known to be $\mu$ -strongly convex [7]. As a result, the Hessian matrix satisfies: $\mathcal{H}(x,\theta) \succeq \mu I_n$ . This implies that all eigenvalues of $\mathcal{H}(x,\theta)$ satisfy $\lambda_{\min}(\mathcal{H}(x,\theta)) \geq \mu$ . Considering the inverse Hessian matrix $\mathcal{H}(x,\theta)^{-1}$ , we have $\lambda (\mathcal{H}(x,\theta)^{-1}) = \frac{1}{\lambda(\mathcal{H}(x,\theta))}$ . Thus, the largest eigenvalue of $\mathcal{H}(x,\theta)^{-1}$ satisfies: + +$$ +\lambda_ {\max } \left(\mathcal {H} (x, \theta) ^ {- 1}\right) = \frac {1}{\lambda_ {\min } \left(\mathcal {H} (x , \theta)\right)} \leq \frac {1}{\mu}. \tag {12} +$$ + +Since the spectral norm (2-norm) of a symmetric matrix is equal to its largest eigenvalue, we conclude: + +$$ +\left\| \mathcal {H} (x, \theta) ^ {- 1} \right\| \leq \frac {1}{\mu}. \tag {13} +$$ + +Define: $\mathcal{H}_S = \mathcal{H}(S,\theta^*)$ as the Hessian of the current subset $S$ . $g_{S} = \sum_{i\in S}\nabla \ell_{i}(\theta^{*})$ as the cumulative gradient for the current subset. When we add a new element $x$ to $S$ , the function gain is given by: + +$$ +\Delta (x, S) = f (S \cup \{x \}) - f (S) = \left\| \left(\mathcal {H} _ {S} + \nabla_ {\theta} ^ {2} \ell_ {x} \left(\theta^ {*}\right)\right) ^ {- 1} \left(g _ {S} + \nabla_ {\theta} \ell_ {x} \left(\theta^ {*}\right)\right) \right\| - \left\| \mathcal {H} _ {S} ^ {- 1} g _ {S} \right\|. \tag {14} +$$ + +To prove that the function is $\epsilon$ -submodular, we must show that for any subsets $A \subseteq B \subseteq V$ , the following inequality holds: + +$$ +\Delta (x, A) \geq \Delta (x, B) - \epsilon , \quad \text {w h e r e} \epsilon > 0. \tag {15} +$$ + +For simplicity, we define $\Delta \mathcal{H} = \nabla_{\theta}^{2}\ell_{x}(\theta^{*})$ and $\Delta g = \nabla_{\theta}\ell_{x}(\theta^{*})$ . Now, applying the first-order approximation of the inverse matrix [43]: + +$$ +\left(\mathcal {H} _ {S} + \Delta \mathcal {H}\right) ^ {- 1} \approx \mathcal {H} _ {S} ^ {- 1} - \mathcal {H} _ {S} ^ {- 1} \Delta \mathcal {H} \mathcal {H} _ {S} ^ {- 1} + O \left(\| \Delta \mathcal {H} \| ^ {2}\right). \tag {16} +$$ + +Substituting this into Eq.(14) for $\Delta (x,S)$ : + +$$ +\Delta (x, S) \approx \left\| \mathcal {H} _ {S} ^ {- 1} g _ {S} + \mathcal {H} _ {S} ^ {- 1} \Delta g - \mathcal {H} _ {S} ^ {- 1} \Delta \mathcal {H} \mathcal {H} _ {S} ^ {- 1} g _ {S} - \mathcal {H} _ {S} ^ {- 1} \Delta \mathcal {H} \mathcal {H} _ {S} ^ {- 1} \Delta g \right\| - \left\| \mathcal {H} _ {S} ^ {- 1} g _ {S} \right\|. \tag {17} +$$ + +Using the triangle inequality, the gain associated with subset $A$ satisfies: + +$$ +\begin{array}{l} \Delta (x, A) \approx \left\| \mathcal {H} _ {A} ^ {- 1} g _ {A} + \mathcal {H} _ {A} ^ {- 1} \Delta g - \mathcal {H} _ {A} ^ {- 1} \Delta \mathcal {H} \mathcal {H} _ {A} ^ {- 1} g _ {A} - \mathcal {H} _ {A} ^ {- 1} \Delta \mathcal {H} \mathcal {H} _ {A} ^ {- 1} \Delta g \right\| - \left\| \mathcal {H} _ {A} ^ {- 1} g _ {A} \right\| \\ \geq \left\| \mathcal {H} _ {A} ^ {- 1} g _ {A} + \mathcal {H} _ {A} ^ {- 1} \Delta g \right\| - \left\| \mathcal {H} _ {A} ^ {- 1} \Delta \mathcal {H} \mathcal {H} _ {A} ^ {- 1} g _ {A} + \mathcal {H} _ {A} ^ {- 1} \Delta \mathcal {H} \mathcal {H} _ {A} ^ {- 1} \Delta g \right\| - \left\| \mathcal {H} _ {A} ^ {- 1} g _ {A} \right\| \\ \geq \left\| \mathcal {H} _ {A} ^ {- 1} g _ {A} \right\| - \left\| \mathcal {H} _ {A} ^ {- 1} \Delta g \right\| - \left\| \mathcal {H} _ {A} ^ {- 1} \Delta \mathcal {H} \mathcal {H} _ {A} ^ {- 1} g _ {A} \right\| - \left\| \mathcal {H} _ {A} ^ {- 1} \Delta \mathcal {H} \mathcal {H} _ {A} ^ {- 1} \Delta g \right\| - \left\| \mathcal {H} _ {A} ^ {- 1} g _ {A} \right\| \\ = - \left\| \mathcal {H} _ {A} ^ {- 1} \Delta g \right\| - \left\| \mathcal {H} _ {A} ^ {- 1} \Delta \mathcal {H} \mathcal {H} _ {A} ^ {- 1} g _ {A} \right\| - \left\| \mathcal {H} _ {A} ^ {- 1} \Delta \mathcal {H} \mathcal {H} _ {A} ^ {- 1} \Delta g \right\|. \tag {18} \\ \end{array} +$$ + +Using norm bounds, we can further estimate: + +$$ +\Delta (x, A) \geq - \frac {G}{\mu | A |} - \frac {H G}{\mu^ {2} | A |} - \frac {H G}{\mu^ {2} | A | ^ {2}}. \tag {19} +$$ + +Similarly, for subset $B$ , we obtain: + +$$ +\begin{array}{l} \Delta (x, B) \approx \left\| \mathcal {H} _ {B} ^ {- 1} g _ {B} + \mathcal {H} _ {B} ^ {- 1} \Delta g - \mathcal {H} _ {B} ^ {- 1} \Delta \mathcal {H} \mathcal {H} _ {B} ^ {- 1} g _ {B} - \mathcal {H} _ {B} ^ {- 1} \Delta \mathcal {H} \mathcal {H} _ {B} ^ {- 1} \Delta g \right\| - \left\| \mathcal {H} _ {B} ^ {- 1} g _ {B} \right\| \\ \leq \left\| \mathcal {H} _ {B} ^ {- 1} g _ {B} \right\| + \left\| \mathcal {H} _ {B} ^ {- 1} \Delta g \right\| + \left\| \mathcal {H} _ {B} ^ {- 1} \Delta \mathcal {H} \mathcal {H} _ {B} ^ {- 1} g _ {B} \right\| + \left\| \mathcal {H} _ {B} ^ {- 1} \Delta \mathcal {H} \mathcal {H} _ {B} ^ {- 1} \Delta g \right\| - \left\| \mathcal {H} _ {B} ^ {- 1} g _ {B} \right\| \\ \leq \frac {G}{\mu | B |} + \frac {H G}{\mu^ {2} | B |} + \frac {H G}{\mu^ {2} | B | ^ {2}}. \tag {20} \\ \end{array} +$$ + +Since $A \subseteq B$ , the difference: + +$$ +\Delta (x, A) - \Delta (x, B) \geq - \frac {2 G (\mu + H)}{\mu^ {2} | A |} - \frac {2 H G}{\mu^ {2} | A | ^ {2}}. \tag {21} +$$ + +Thus, the parameter $\epsilon = \frac{2G(\mu + H)}{\mu^2|A|} +\frac{2HG}{\mu^2|A|^2}$ . This completes the proof. + +![](images/d9f07ed4f5bc86dc2aa4481e0baddef36c13df874ccd3fafbc8b93926a843e13.jpg) + +# C Proofs of Lemma 2 + +Lemma 2. When using $IRT$ for ability estimation, the function $f(S)$ is Lipschitz continuous with respect to $\theta$ . Furthermore, with probability at least $1 - \delta$ , the approximation error incurred by using $\theta^t$ satisfies the upper bound: $|f(S|\theta^t) - f(S)| \leq \left(\frac{H}{\mu_1} + \frac{MG}{\mu_1\mu_2}\right)\frac{C(\delta)}{\sqrt{|S_t|}}$ , where $\mu_1, \mu_2, M, H, G,$ and $C$ are model-dependent constants characterizing the properties of the objective function. + +Proof. We first should prove that $f(S \mid \theta) = \left\| \left( \sum_{i \in S} \nabla_{\theta}^{2} \ell_{i}(\theta) \right)^{-1} \sum_{i \in S} \nabla_{\theta} \ell_{i}(\theta) \right\|$ is Lipschitz continuous with respect to $\theta$ , we analyze its sensitivity to small changes in $\theta$ . + +Since the gradient $\nabla_{\theta}\ell_{i}(\theta)$ is continuously differentiable in $\theta$ , the Mean Value Theorem guarantees the existence of some $\xi_{i}$ between $\theta_{1}$ and $\theta_{2}$ such that + +$$ +\nabla_ {\theta} \ell_ {i} \left(\theta_ {1}\right) - \nabla_ {\theta} \ell_ {i} \left(\theta_ {2}\right) = \nabla_ {\theta} ^ {2} \ell_ {i} \left(\xi_ {i}\right) \left(\theta_ {1} - \theta_ {2}\right). \tag {22} +$$ + +Assuming that $\| \nabla_{\theta}^{2}\ell_{i}(\theta)\| \leq H$ and taking the norm and summing over $i\in S$ gives + +$$ +\left\| \sum_ {i \in S} \nabla_ {\theta} \ell_ {i} \left(\theta_ {1}\right) - \sum_ {i \in S} \nabla_ {\theta} \ell_ {i} \left(\theta_ {2}\right) \right\| \leq \sum_ {i \in S} \left\| \nabla_ {\theta} ^ {2} \ell_ {i} \left(\xi_ {i}\right) \right\| \| \theta_ {1} - \theta_ {2} \| \leq H | S | \| \theta_ {1} - \theta_ {2} \|. \tag {23} +$$ + +Similarly, assuming $\| \nabla_{\theta}^{3}\ell_{i}(\theta)\| \leq M$ , there exists some $\eta_{i}$ between $\theta_{1}$ and $\theta_{2}$ such that + +$$ +\left\| \sum_ {i \in S} \nabla_ {\theta} ^ {2} \ell_ {i} \left(\theta_ {1}\right) - \sum_ {i \in S} \nabla_ {\theta} ^ {2} \ell_ {i} \left(\theta_ {2}\right) \right\| \leq \sum_ {i \in S} \left\| \nabla_ {\theta} ^ {3} \ell_ {i} \left(\eta_ {i}\right) \right\| \| \theta_ {1} - \theta_ {2} \| \leq M | S | \| \theta_ {1} - \theta_ {2} \|. \tag {24} +$$ + +Define: $\mathcal{H}(\theta) = \sum_{i\in S}\nabla_{\theta}^{2}\ell_{i}(\theta)$ and $g(\theta) = \sum_{i\in S}\nabla_{\theta}\ell_{i}(\theta)$ . We analyze + +$$ +\begin{array}{l} \left| f (S, \theta_ {1}) - f (S, \theta_ {2}) \right| = \left| \left\| \mathcal {H} (\theta_ {1}) ^ {- 1} g (\theta_ {1}) \right\| - \left\| \mathcal {H} (\theta_ {2}) ^ {- 1} g (\theta_ {2}) \right\| \right| \\ \leq \left\| \mathcal {H} \left(\theta_ {1}\right) ^ {- 1} g \left(\theta_ {1}\right) - \mathcal {H} \left(\theta_ {2}\right) ^ {- 1} g \left(\theta_ {2}\right) \right\| \\ = \| \mathcal {H} \left(\theta_ {1}\right) ^ {- 1} g \left(\theta_ {1}\right) - \mathcal {H} \left(\theta_ {1}\right) ^ {- 1} g \left(\theta_ {2}\right) + \mathcal {H} \left(\theta_ {1}\right) ^ {- 1} g \left(\theta_ {2}\right) - \mathcal {H} \left(\theta_ {2}\right) ^ {- 1} g \left(\theta_ {2}\right) \| \\ \leq \left\| \mathcal {H} \left(\theta_ {1}\right) ^ {- 1} \left(g \left(\theta_ {1}\right) - g \left(\theta_ {2}\right)\right) \right\| + \left\| \left(\mathcal {H} \left(\theta_ {1}\right) ^ {- 1} - \mathcal {H} \left(\theta_ {2}\right) ^ {- 1}\right) g \left(\theta_ {2}\right) \right\|. \tag {25} \\ \end{array} +$$ + +For the first term, using matrix norm properties: + +$$ +\left\| \mathcal {H} \left(\theta_ {1}\right) ^ {- 1} \left(g \left(\theta_ {1}\right) - g \left(\theta_ {2}\right)\right) \right\| \leq \left\| \mathcal {H} \left(\theta_ {1}\right) ^ {- 1} \right\| \cdot \left\| g \left(\theta_ {1}\right) - g \left(\theta_ {2}\right) \right\|. \tag {26} +$$ + +Assuming $\| \mathcal{H}(\theta_1)^{-1}\| \leq \frac{1}{|S|\mu_1}$ in a well-conditioned region (similar to Theorem 1), we obtain: + +$$ +\left\| \mathcal {H} \left(\theta_ {1}\right) ^ {- 1} \left(g \left(\theta_ {1}\right) - g \left(\theta_ {2}\right)\right) \right\| \leq \frac {H}{\mu_ {1}} \| \theta_ {1} - \theta_ {2} \|. \tag {27} +$$ + +For the second term, using: + +$$ +\mathcal {H} \left(\theta_ {1}\right) ^ {- 1} - \mathcal {H} \left(\theta_ {2}\right) ^ {- 1} = \mathcal {H} \left(\theta_ {1}\right) ^ {- 1} \left(\mathcal {H} \left(\theta_ {2}\right) - \mathcal {H} \left(\theta_ {1}\right)\right) \mathcal {H} \left(\theta_ {2}\right) ^ {- 1}, +$$ + +and assuming $\| \mathcal{H}(\theta_2)^{-1}\| \leq \frac{1}{|S|\mu_2}$ , we obtain: + +$$ +\left\| \left(\mathcal {H} \left(\theta_ {1}\right) ^ {- 1} - \mathcal {H} \left(\theta_ {2}\right) ^ {- 1}\right) g \left(\theta_ {2}\right) \right\| \leq \| \mathcal {H} \left(\theta_ {1}\right) ^ {- 1} \| \| \mathcal {H} \left(\theta_ {2}\right) - \mathcal {H} \left(\theta_ {1}\right) \| \| \mathcal {H} \left(\theta_ {2}\right) ^ {- 1} \| \| g \left(\theta_ {2}\right) \|. \tag {28} +$$ + +Using our previous bound $\| \mathcal{H}(\theta_2) - \mathcal{H}(\theta_1)\| \leq M|S\| \theta_1 - \theta_2\|$ and assuming $\| g(\theta)\| \leq |S|G$ in a bounded region, we get: + +$$ +\left\| \left(\mathcal {H} \left(\theta_ {1}\right) ^ {- 1} - \mathcal {H} \left(\theta_ {2}\right) ^ {- 1}\right) g \left(\theta_ {2}\right) \right\| \leq \frac {M G}{\mu_ {1} \mu_ {2}} \| \theta_ {1} - \theta_ {2} \|. \tag {29} +$$ + +Thus + +$$ +\left| f (S \mid \theta_ {1}) - f (S \mid \theta_ {2}) \right| \leq \left(\frac {H}{\mu_ {1}} + \frac {M G}{\mu_ {1} \mu_ {2}}\right) \| \theta_ {1} - \theta_ {2} \|. \tag {30} +$$ + +$\theta^t$ is obtained via Maximum Likelihood Estimation (MLE), we have: $\sqrt{|S_t|} (\theta^t -\theta^*)\stackrel {d}{\to}\mathcal{N}(0,I^{-1}(\theta^ {*}))$ , where $I(\theta^{*})$ is the Fisher information matrix. This follows the asymptotic normality of MLE [44]. This implies that, with high probability, the estimation error satisfies + +$$ +\left\| \theta^ {t} - \theta^ {*} \right\| \leq \frac {C (\delta)}{\sqrt {\left| S _ {t} \right|}} \quad \text {w i t h p r o b a b i l i t y a t l e a s t} 1 - \delta , \tag {31} +$$ + +for some constant $C(\delta)$ depending on the trace and spectral norm of $I^{-1}(\theta^{*})$ . Since $f(S|\theta)$ is Lipschitz continuous in $\theta$ , with probability at least $1 - \delta$ , + +$$ +\left| f (S \mid \theta^ {t}) - f (S) \right| \leq \left(\frac {H}{\mu_ {1}} + \frac {M G}{\mu_ {1} \mu_ {2}}\right) \| \theta^ {t} - \theta^ {*} \| \leq \left(\frac {H}{\mu_ {1}} + \frac {M G}{\mu_ {1} \mu_ {2}}\right) \frac {C (\delta)}{\sqrt {\left| S _ {t} \right|}} \tag {32} +$$ + +This guarantees that for sufficiently large $|S_t|$ , the error introduced by using $\theta^t$ in place of $\theta^*$ is small with high probability. + +![](images/2123874178d655fbfd919e6f1a48a5562aed21e90f4448826510610d079fc98a.jpg) + +# D Full Derivation of Bias-Corrected Estimate + +We begin by defining the perturbed objective: + +$$ +\theta_ {S (m)} ^ {\gamma} = \arg \min _ {\theta} \frac {1}{| S |} \sum_ {i \in S} \ell_ {i} (\theta) - \gamma \ell_ {m} (\theta) + \gamma \widetilde {\ell} _ {m} (\theta). \tag {33} +$$ + +Since $\theta_{S(m)}^{\gamma}$ minimizes the objective, it satisfies the first-order optimality condition: + +$$ +0 = \frac {1}{| S |} \sum_ {i \in S} \nabla \ell_ {i} \left(\theta_ {S (m)} ^ {\gamma}\right) - \gamma \nabla \ell_ {m} \left(\theta_ {S (m)} ^ {\gamma}\right) + \gamma \nabla \widetilde {\ell} _ {m} \left(\theta_ {S (m)} ^ {\gamma}\right). \tag {34} +$$ + +We now apply a first-order Taylor expansion of the gradient around $\theta_S$ : + +$$ +\begin{array}{l} 0 \approx \frac {1}{| S |} \sum_ {i \in S} \nabla \ell_ {i} (\theta_ {S}) - \gamma \nabla \ell_ {m} (\theta_ {S}) + \gamma \nabla \widetilde {\ell} _ {m} (\theta_ {S}) \\ + \left[ \frac {1}{| S |} \sum_ {i \in S} \nabla^ {2} \ell_ {i} \left(\theta_ {S}\right) - \gamma \nabla^ {2} \ell_ {m} \left(\theta_ {S}\right) + \gamma \nabla^ {2} \widetilde {\ell} _ {m} \left(\theta_ {S}\right) \right] \left(\theta_ {S (m)} ^ {\gamma} - \theta_ {S}\right). \tag {35} \\ \end{array} +$$ + +Choosing $\gamma = \frac{1}{|S|}$ , and noting that $\theta_S$ satisfies the original optimality condition $\sum_{i\in S}\nabla \ell_i(\theta_S) = 0$ , we simplify: + +$$ +\begin{array}{l} 0 \approx - \frac {1}{| S |} \nabla \ell_ {m} (\theta_ {S}) + \frac {1}{| S |} \nabla \widetilde {\ell} _ {m} (\theta_ {S}) \\ + \left[ \frac {1}{| S |} \sum_ {i \in S \backslash q _ {m}} \nabla^ {2} \ell_ {i} \left(\theta_ {S}\right) + \frac {1}{| S |} \nabla^ {2} \widetilde {\ell} _ {m} \left(\theta_ {S}\right) \right] \left(\theta_ {S (m)} - \theta_ {S}\right). \tag {36} \\ \end{array} +$$ + +Let $\mathcal{H}(S\setminus q_m,\theta_S) = \sum_{i\in S\setminus q_m}\nabla^2\ell_i(\theta_S)$ , and $\widetilde{\mathcal{H}} (q_m,\theta_S) = \nabla^2\widetilde{\ell}_m(\theta_S)$ , we obtain: + +$$ +\theta_ {S (m)} - \theta_ {S} \approx \left[ \mathcal {H} \left(S \backslash q _ {m}, \theta_ {S}\right) + \widetilde {\mathcal {H}} \left(q _ {m}, \theta_ {S}\right) \right] ^ {- 1} \left(\nabla \ell_ {m} \left(\theta_ {S}\right) - \nabla \widetilde {\ell} _ {m} \left(\theta_ {S}\right)\right). \tag {37} +$$ + +Now, recall the definitions of the original and flipped losses: + +$$ +\ell_ {m} (\theta) = - y _ {m} \log p _ {\theta} \left(q _ {m}, 1\right) - \left(1 - y _ {m}\right) \log p _ {\theta} \left(q _ {m}, 0\right), \tag {38} +$$ + +$$ +\widetilde {\ell} _ {m} (\theta) = - \left(1 - y _ {m}\right) \log p _ {\theta} \left(q _ {m}, 1\right) - y _ {m} \log p _ {\theta} \left(q _ {m}, 0\right). \tag {39} +$$ + +Taking the gradient difference: + +$$ +\begin{array}{l} \nabla \ell_ {m} (\theta_ {S}) - \nabla \widetilde {\ell} _ {m} (\theta_ {S}) = \nabla \left[ - y _ {m} \log p _ {\theta} (q _ {m}, 1) - (1 - y _ {m}) \log p _ {\theta} (q _ {m}, 0) \right] \\ - \nabla [ - (1 - y _ {m}) \log p _ {\theta} (q _ {m}, 1) - y _ {m} \log p _ {\theta} (q _ {m}, 0) ] \\ = (1 - 2 y _ {m}) \nabla \log \frac {p _ {\theta} \left(q _ {m} , 1\right)}{p _ {\theta} \left(q _ {m} , 0\right)}. \tag {40} \\ \end{array} +$$ + +Substituting back, we obtain the final expression: + +$$ +\theta_ {S (m)} \approx \theta_ {S} + \left[ \mathcal {H} \left(S \backslash q _ {m}, \theta_ {S}\right) + \widetilde {\mathcal {H}} \left(q _ {m}, \theta_ {S}\right) \right] ^ {- 1} (1 - 2 y _ {m}) \nabla \log \frac {p _ {\theta} \left(q _ {m} , 1\right)}{p _ {\theta} \left(q _ {m} , 0\right)}. \tag {41} +$$ + +# E Limitations and Broader Impact + +Despite the promising results of CFAT, several limitations remain that open avenues for future research. For example, while CFAT incorporates analytical corrections for guessing and slipping, it assumes these behavioral perturbations follow simple, predefined patterns. In practice, examinee behavior can be more complex and context-dependent. Future work could integrate richer cognitive models or leverage response time, clickstream, or eye-tracking data to better capture behavioral variability. + +CFAT offers a scalable, interpretable, and computationally efficient solution for adaptive testing, with potential for broad societal benefits: + +- Democratization of High-Quality Assessment: By reducing the number of required questions and computational overhead, CFAT can enable real-time, low-cost testing in low-resource settings, such as developing countries, where infrastructure is limited. +- Fairer and More Inclusive Testing: The bias-correction mechanism in CFAT helps mitigate the influence of irregular behaviors (e.g., guessing), potentially leading to fairer assessments across diverse populations. This is particularly important in high-stakes testing scenarios, where small inaccuracies can have significant consequences. +- Privacy: CFAT does not rely on large-scale user data or extensive training, reducing the need for data collection and storage. This not only preserves user privacy but also reduces the environmental footprint of deploying large-scale AI-driven assessment systems. \ No newline at end of file diff --git a/NeurIPS/2025/A Closed-Form Solution for Fast and Reliable Adaptive Testing/images.zip b/NeurIPS/2025/A Closed-Form Solution for Fast and Reliable Adaptive Testing/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..8f3abda10d3e86ca717d8d6b43b7a89de93dd245 --- /dev/null +++ b/NeurIPS/2025/A Closed-Form Solution for Fast and Reliable Adaptive Testing/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5532cc1d5517b5139868fadcfa4804f82ed634c2704b25ce5d3bf1c729f16aad +size 765088 diff --git a/NeurIPS/2025/A Closed-Form Solution for Fast and Reliable Adaptive Testing/layout.json b/NeurIPS/2025/A Closed-Form Solution for Fast and Reliable Adaptive Testing/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..817bd088b5b1424dbe2c58fe0bc017693a08e7e9 --- /dev/null +++ b/NeurIPS/2025/A Closed-Form Solution for Fast and Reliable Adaptive Testing/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:abcea80769f6e923e0eb01c0323a064da0cdb83dec6306477eccfcec44d7b763 +size 970495 diff --git a/NeurIPS/2025/A Closer Look at Graph Transformers_ Cross-Aggregation and Beyond/4a19158a-0a9c-418d-9eb4-d5f2789453f4_content_list.json b/NeurIPS/2025/A Closer Look at Graph Transformers_ Cross-Aggregation and Beyond/4a19158a-0a9c-418d-9eb4-d5f2789453f4_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..243529853a74806bb5fc22ff657c2ffadb9d8252 --- /dev/null +++ b/NeurIPS/2025/A Closer Look at Graph Transformers_ Cross-Aggregation and Beyond/4a19158a-0a9c-418d-9eb4-d5f2789453f4_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:562e2530e1ccae21786c937995295e0bf0c237ac9559c9b8e7a4a3116d9ddc94 +size 177487 diff --git a/NeurIPS/2025/A Closer Look at Graph Transformers_ Cross-Aggregation and Beyond/4a19158a-0a9c-418d-9eb4-d5f2789453f4_model.json b/NeurIPS/2025/A Closer Look at Graph Transformers_ Cross-Aggregation and Beyond/4a19158a-0a9c-418d-9eb4-d5f2789453f4_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7a5ea1ab1210ad64e24f1c4536ec979f5c4fa4cb --- /dev/null +++ b/NeurIPS/2025/A Closer Look at Graph Transformers_ Cross-Aggregation and Beyond/4a19158a-0a9c-418d-9eb4-d5f2789453f4_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:aa4f33a70b019ba3b7bba9b3c1e67d30dea8864f4072e1b4cf6627eadb695661 +size 230824 diff --git a/NeurIPS/2025/A Closer Look at Graph Transformers_ Cross-Aggregation and Beyond/4a19158a-0a9c-418d-9eb4-d5f2789453f4_origin.pdf b/NeurIPS/2025/A Closer Look at Graph Transformers_ Cross-Aggregation and Beyond/4a19158a-0a9c-418d-9eb4-d5f2789453f4_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..65cbefea6d6b18bd3a6b794ac4b25b8ab6926d1f --- /dev/null +++ b/NeurIPS/2025/A Closer Look at Graph Transformers_ Cross-Aggregation and Beyond/4a19158a-0a9c-418d-9eb4-d5f2789453f4_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ca74e51942f96e695e56012e3db260261273465406bd5ab2fffbf4f0bc2cc188 +size 1430955 diff --git a/NeurIPS/2025/A Closer Look at Graph Transformers_ Cross-Aggregation and Beyond/full.md b/NeurIPS/2025/A Closer Look at Graph Transformers_ Cross-Aggregation and Beyond/full.md new file mode 100644 index 0000000000000000000000000000000000000000..155602560bebdc2b06745805d903f8c511a7f8af --- /dev/null +++ b/NeurIPS/2025/A Closer Look at Graph Transformers_ Cross-Aggregation and Beyond/full.md @@ -0,0 +1,885 @@ +# A Closer Look at Graph Transformers: Cross-Aggregation and Beyond + +Jiaming Zhuo $^{1}$ , Ziyi Ma $^{1}$ , Yintong Lu $^{1}$ , Yuwei Liu $^{1}$ , Kun Fu $^{1}$ , Di Jin $^{2}$ , Chuan Wang $^{3}$ , Wenning Wu $^{4}$ , Zhen Wang $^{4}$ , Xiaochun Cao $^{5}$ , Liang Yang $^{1*}$ $^{1}$ Hebei Province Key Laboratory of Big Data Calculation, School of Artificial Intelligence, Hebei University of Technology, Tianjin, China + $^{2}$ College of Intelligence and Computing, Tianjin University, Tianjin, China + $^{3}$ School of Computer Science and Technology, Beijing JiaoTong University, Beijing, China + $^{4}$ School of Artificial Intelligence, OPtics and ElectroNics (iOPEN), School of Cybersecurity, Northwestern Polytechnical University, Xi'an, China + $^{5}$ School of Cyber Science and Technology, Shenzhen Campus of Sun Yat-sen University, Shenzhen, China +jiaming.zhuo@outlook.com, zyma@hebut.edu.cn, {202332803037, 202322802030}@stu.hebut.edu.cn, fukun@hebut.edu.cn, jindi@tju.edu.cn, wangchuan@iie.ac.cn, wuwenning@nwpu.edu.cn, w-zhen@nwpu.edu.cn, caoxiaochun@mail.sysu.edu.cn, yangliang@vip.qq.com + +# Abstract + +Graph Transformers (GTs), which effectively capture long-range dependencies and structural biases simultaneously, have recently emerged as promising alternatives to traditional Graph Neural Networks (GNNs). Advanced approaches for GTs to leverage topology information involve integrating GNN modules or modulating node attributes using positional encodings. Unfortunately, the underlying mechanism driving their effectiveness remains insufficiently understood. In this paper, we revisit these strategies and uncover a shared underlying mechanism—Cross Aggregation—that effectively captures the interaction between graph topology and node attributes. Building on this insight, we propose the Universal Graph Cross-attention Transformer (UGCFoerner), a universal GT framework with linear computational complexity. The idea is to interactively learn the representations of graph topology and node attributes through a linearized Dual Cross-attention (DCA) module. In theory, this module can adaptively capture interactions between these two types of graph information, thereby achieving effective aggregation. To alleviate overfitting arising from the dual-channel design, we introduce a consistency constraint that enforces representational alignment. Extensive evaluations on multiple benchmark datasets demonstrate the effectiveness and efficiency of UGCFoerner. + +# 1 Introduction + +Node classification, aimed at accurately predicting node categories based on the graph topology and node attributes, is a fundamental task in identifying the properties of individual nodes [12, 18, 14, 32, 11, 10]. As a powerful class of models for fusing topology and attribute information in graphs, Graph Neural Networks (GNNs) have achieved initial successes in this task [29, 5, 52, 24, 22, 25]. In general, they follow the graph-bound Message Passing (MP) paradigm [16]. While this paradigm endows GNNs with the localizing property, it also restricts their ability to capture long-range dependencies [9], resulting in well-known challenges such as over-smoothing [6, 59] and over-squashing [17]. + +Inspired by the remarkable success of Transformers in NLP [35], Graph Transformers (GTs) have emerged as powerful architectures for node classification tasks. The core component of Transformers is the Self-Attention (SA) module [46], which models full interactions among tokens within a sequence, thereby endowing the Transformers with globalizing properties. The initial success of GTs can be attributed to the strategic integration of discriminative graph topology into Transformer architectures, enabling the simultaneous capture of structural biases and long-range dependencies. To date, two primary strategies have achieved SOTA performance in existing GTs: (1) integrating GNN blocks [31, 50, 7, 61], and (2) modulating node attributes utilizing Positional Encodings (PEs) [2, 49, 44]. However, both strategies face inherent limitations. The first tends to inherit drawbacks from GNNs due to its reliance on them, whereas the second introduces additional computational complexity due to the use of PEs, thereby restricting the models' universality1 and scalability. + +This leads to a fundamental question: + +What underlying mechanism drives the effectiveness of diverse Graph Transformers? + +A thorough understanding of the underlying mechanisms can offer valuable insights for developing more advanced and efficient architectures. Following this line, this paper theoretically investigates the mechanism shared by the aforementioned types of GTs and, based on this insight, proposes a novel GT architecture. In particular, the unified cross-aggregation mechanism (as formally defined in Definition 1) is explored by analytically decoupling topology and attribute representations from node representations. Specifically, the GNN block in GTs can be interpreted as aggregating topology representations into attribute representations (in Theorem 1), indicating that this category of GTs inherently incorporates cross-aggregation. Furthermore, GTs employing PEs contain diverse forms of cross-aggregation between topology and attribute representations. Therefore, the shared underlying mechanism among these GTs is cross-aggregation between graph topology and node attributes. + +This understanding naturally leads to a key question: + +How can we design an effective and efficient GT architecture grounded in cross-aggregation? + +To this end, this paper proposes the Universal Graph Cross-attention Transformer (UGCFoerter), which implements the cross-aggregation mechanism via cross-attention. To be specific, it separately encodes graph topology and node attributes to obtain their initial representations. At its core lies a linearized Dual Cross-Attention (DCA) module that updates the topology and attribute representations by computing cross-attention scores among nodes and utilizing them for weighted aggregation. In theory, the DCA module adaptively captures both correlation and exclusion relationships between graph topology and node attributes, making it simple yet effective. Finally, the two representations are integrated to yield a comprehensive node representation. To prevent representation distortion, a consistency constraint is introduced to enforce mutual alignment between them. + +The main contributions of this work are summarized as follows: + +- Mechanism Revelation: We theoretically reveal a unified mechanism across typical Graph Transformers, namely cross-aggregation between graph topology and node attributes. +- Model Innovation: We propose UGCFoer, a GT architecture equipping with a linearized Dual Cross-Attention (DCA) module that implements the cross-aggregation mechanism. +- Comprehensive Evaluation: Extensive evaluations conducted on sixteen homophilic, heterophilic, and large-scale graphs demonstrate the universality and scalability of UGCFoermer. + +# 2 Preliminaries + +This section begins by presenting the notation used throughout this paper. Then, it introduces the concepts of Graph Neural Networks (GNNs) and Graph Transformers (GTs). + +# 2.1 Notations + +The subject of this paper is the widely-used undirected attribute graph, denoted as $\mathcal{G}(\mathcal{V},\mathcal{E})$ , where $\mathcal{V}$ and $\mathcal{E}$ represent the node set and edge set. $\mathcal{V}$ consists of $n$ node instances $\{(\mathbf{x}_v,\mathbf{y}_v)\}_{v\in \mathcal{V}}$ , where + +$\mathbf{x}_v \in \mathbb{R}^f$ and $\mathbf{y}_v \in \mathbb{R}^c$ denote the node attribute and label of node $v$ , respectively. $f$ is the dimension of attributes and $c$ is the dimension of labels. $\mathcal{E} = \{(v_i, v_j)\}$ terms the edge set. Typically, graph topology is described by the adjacency matrix $\mathbf{A} \in \mathbb{R}^{n \times n}$ where $a_{i,j} = 1$ only if $(v_i, v_j) \in \mathcal{E}$ , and $a_{i,j} = 0$ otherwise. In formal terms, the graph $\mathcal{G}$ can be redescribed as $\mathcal{G}(\mathbf{A}, \mathbf{X})$ . In the context of semi-supervised learning, the node labels are segmented into two sets: $\mathbf{Y}_L \in \mathbb{R}^{n_l \times c}$ for the labeled nodes and $\mathbf{Y}_U \in \mathbb{R}^{n_u \times c}$ for the unlabeled nodes. + +To verify the model's universality, this paper examines graphs with varying degrees of homophily. In homophilic graphs, edges are typically formed between nodes with similar labels. Conversely, in heterophilic graphs, edges tend to form between nodes with dissimilar labels [37, 6, 60, 62]. + +# 2.2 Graph Neural Networks + +Message Passing (MP)-based Graph Neural Networks (GNNs) follow an aggregation-combination strategy. Specifically, the representation of each node is iteratively updated by aggregating the features from its local neighbors and combining the aggregated features with its features, which is given by + +$$ +\mathbf {h} _ {v} ^ {l} \triangleq C O M ^ {l} \left(\mathbf {h} _ {v} ^ {l - 1}, A G G ^ {l} \left(\left\{\mathbf {h} _ {u} ^ {l - 1} \mid u \in \mathcal {N} (v) \right\}\right)\right), \tag {1} +$$ + +where $\mathcal{N}(v)$ denotes the set of neighboring nodes of node $v$ . For the functions $AGG(\cdot)$ and $COM(\cdot)$ , vanilla GNNs, e.g., GCN [29], adopt the sum function to implement them, that is, + +$$ +G C N (\mathbf {A}, \mathbf {H}): \mathbf {H} ^ {l + 1} = \sigma (\tilde {\mathbf {A}} \mathbf {H} ^ {l} \mathbf {W}), \mathbf {H} ^ {0} = \mathbf {X}, \tag {2} +$$ + +where $\sigma (\cdot)$ stands for the nonlinear activation functions, and $\tilde{\mathbf{A}} = \hat{\mathbf{D}}^{-\frac{1}{2}}\hat{\mathbf{A}}\hat{\mathbf{D}}^{-\frac{1}{2}}$ is the normalized adjacency matrix with $\hat{\mathbf{A}} = \mathbf{A} + \mathbf{I}$ . $\mathbf{W}$ denotes the trainable projection parameters. + +# 2.3 Transformers + +Inspired by the success of Transformers in NLP [46], numerous variant models have been designed for multiple fields, including CV [21] and Graph Learning. They typically consist of four functional components: attention module, feed-forward network, residual connection, and normalization. + +Self-attention Module. This is a core component of the vanilla Transformer to model intra-sequence relationships among all tokens [46]. Given a sequence containing $n$ tokens $\mathbf{H} = [\mathbf{h}_i]_{i=0}^{n-1} \in \mathbb{R}^{n \times d}$ , the module first projects $\mathbf{H}$ into Query $q(\mathbf{H})$ , Key $k(\mathbf{H})$ , and Value $k(\mathbf{H})$ . It then employs the attention scores calculated from all Query-Key pairs to perform a weighted sum of the Value vectors. + +A general formulation of the Self-Attention (SA) module is given by + +$$ +S A (\mathbf {H}): \hat {\mathbf {H}} _ {S A} = S o f t m a x \left(\frac {q (\mathbf {H}) k (\mathbf {H}) ^ {\top}}{\sqrt {d}}\right) v (\mathbf {H}), \tag {3} +$$ + +where $q(\cdot), k(\cdot)$ , and $v(\cdot)$ generate the Query, Key, and Value via MLPs [41] with learnable parameters $\mathbf{W}$ . The attention score $\text{Softmax}(q(\mathbf{H})k(\mathbf{H})^\top/\sqrt{d}) \in \mathbb{R}^{n \times n}$ is computed via the scaled dot product of full-token pairs, resulting in a quadratic computational complexity. + +Graph Transformers (GTs). Most existing models [51, 1, 39, 36, 4, 53, 61, 3] build upon the SA module. GTs differ from traditional Transformers in how they leverage topology information to capture structural biases. As discussed in the Introduction, two main strategies for incorporating topology information have achieved SOTA performance on node-level tasks: (1) integrating GNN blocks, and (2) modulating node attributes utilizing Positional Encodings (PEs). + +Cross-Attention Module. Unlike self-attention, which models the intra-source relationships, cross-attention captures the interactions between two distinct sources. For the features from two different sources $\mathbf{H} \in \mathbb{R}^{n_1 \times d}$ and $\mathbf{Z} \in \mathbb{R}^{n_2 \times d}$ , the Cross-Attention (CA) module can be expressed as + +$$ +C A (\mathbf {Z}, \mathbf {H}): \hat {\mathbf {H}} _ {C A} = S o f t m a x \left(\frac {q (\mathbf {Z}) k (\mathbf {H}) ^ {\top}}{\sqrt {d}}\right) v (\mathbf {H}). \tag {4} +$$ + +After the representation $\hat{\mathbf{H}}_{CA}$ is obtained, it is typically used as the cross-source representation to update $\mathbf{Z}$ . Due to its exceptional capacity for modeling inter-source relationships, this module has been applied in diverse domains, e.g., NLP [15] and CV [26]. However, it has received little attention in Graph Learning, largely due to the lack of motivation and well-defined applied target. Moreover, similar to the self-attention (Eq. 3), its computational complexity is quadratic, i.e., $O(n_1n_2)$ . + +![](images/fc73d50029853c2e914473c1ee03bbf75a7232ad80e8cf90748e3354f58ea430.jpg) +Figure 1: Overview of the proposed GT architecture UGCFormaler and its linear attention module. (a) The pipeline of UGCFormaler, which incorporates a dual cross-attention (DCA) module. First, two basic elements of graphs (i.e., graph topology and node attributes) are independently processed in their respective spaces utilizing distinct projection layers $f_{A}(\cdot)$ and $f_{X}(\cdot)$ . Next, the dual cross-attention (DCA) module with residual connections operates across the topology and attribute spaces, updating each representation by integrating correlated features from the other space. Finally, the two representations are combined to produce the final output representation. (b) Illustration of the proposed efficient cross-attention module, where parameters are shared between the query (Q) and key (K), and the representations are computed using linearized attention, given by $\mathbf{Q}(\mathbf{K}^{\top}\mathbf{V})$ . + +# 3 Methodology + +This section starts by theoretically exploring the functional mechanism shared by Graph Transformers (GTs) that use Graph Neural Network (GNN) blocks and GTs that utilize Positional Encodings (PEs). Inspired by this mechanism, it introduces UGCFoer, a simple yet universal graph cross-attention Transformer with linear complexity. Finally, it gives a comprehensive analysis of UGCFoer. + +# 3.1 Motivations + +As previously discussed, the underlying mechanism behind the effectiveness of typical GTs remains insufficiently explored. To address this issue, this subsection proposes a cross-aggregation mechanism and theoretically examines how it is manifested in the two types of SOTA GTs. + +The cross-aggregation mechanism is formally defined as follows. + +Definition 1. (Cross-aggregation mechanism) Given two representations $\mathbf{B} \in \mathbb{R}^{n_1 \times d_1}$ and $\mathbf{Z} \in \mathbb{R}^{n_2 \times d_2}$ from different modalities (sources), which share at least one same dimension, i.e., $n_1 = n_2$ or $d_1 = d_2$ . A general formula for two types of cross-aggregations can be expressed as + +$$ +\hat {\mathbf {Z}} \triangleq \left\{ \begin{array}{l l} S i m (\mathbf {Z}, \mathbf {B}) \mathbf {B}, & \text {i f} d _ {1} = d _ {2}, \\ \mathbf {B} S i m (\mathbf {B}, \mathbf {Z}), & \text {i f} n _ {1} = n _ {2}, \end{array} \right. \tag {5} +$$ + +where $Sim(\mathbf{Z},\mathbf{B})$ denotes a similarity function between $\mathbf{Z}$ and $\mathbf{B}$ , such as cosine similarity. + +The first case corresponds to sample (node)-level aggregation, e.g., cross-attention (Eq. 4), while the second corresponds to dimension (feature)-level aggregation [56, 57, 61]. When $\mathbf{Z} = \mathbf{B}$ , Eq. 5 reduces to a self-aggregation (e.g., self-attention). Accordingly, two theorems are presented. + +Theorem 1. In typical Graph Transformers, the diffusion matrix of GNN blocks can be expressed via eigendecomposition as $\mathbf{S} = \mathbf{U}\boldsymbol{\Lambda}\mathbf{U}^{\top}$ , where $\mathbf{U}$ and $\boldsymbol{\Lambda} = \operatorname{diag}([\lambda_1,\dots,\lambda_n])$ represents eigenvectors and eigenvalues, respectively (in descending order). Accordingly, the GNN block can be viewed as a cross-aggregation between attribute representations $\mathbf{XW}$ and topology representations $\mathbf{U}\sqrt{\boldsymbol{\Lambda}}$ . + +Theorem 2. Given the modulated node attributes using any PE, i.e., $\hat{\mathbf{X}} = [\mathbf{X};\mathbf{P}]$ , where $\mathbf{P}\in \mathbb{R}^{n\times k}$ represents the PE and $[\cdot ]$ denotes concatenation operator. PE-based GTs (Eq. 3) inherently contain a cross-aggregation between attribute representations $\mathbf{XW}$ and topology representations $\mathbf{PW}$ . + +The proofs for Theorems 1 and 2 are provided in Sections B and C, respectively. In short, the key mechanism of GTs using GNNs and PEs is Cross-Aggregation between topology and attributes. + +# 3.2 UGCFoermer + +Motivated by the cross-aggregation mechanism explored in the previous subsection, this subsection introduces UGCFoer, a simple yet universal GT. At its core, UGCFoer employs a linearized cross-attention module that implements the cross-aggregation mechanism to capture interactions between graph topology and node attributes. UGCFoer consists of four modules, each of which is described below. The detailed implementation is provided in Algorithm 1. + +Initial Representation Layer. Two different projection layers are utilized to independently generate initial representations for the two types of graph information. For simplicity, MLPs are used to process the adjacency matrix $\mathbf{A} \in \mathbb{R}^{n \times n}$ and the attribute matrix $\mathbf{X} \in \mathbb{R}^{n \times f}$ , producing the corresponding initial representations $\mathbf{Z}$ and $\mathbf{B}$ , that is, + +$$ +\mathbf {Z} ^ {0} = M L P _ {A} (\mathbf {A}), \quad \mathbf {B} ^ {0} = M L P _ {X} (\mathbf {X}) \in \mathbb {R} ^ {n \times d}, \tag {6} +$$ + +where $MLP_{A}(\cdot)$ and $MLP_{X}(\cdot)$ term the MLPs for processing topology and attributes, respectively. + +Dual Cross-attention Module. As an implementation of the cross-aggregation (in Definition 1), this module is designed to capture the interactions between these two types of graph information. However, directly employing the cross-attention (Eq. 4) may result in two issues: (1) unacceptable quadratic computational complexity due to the calculation of dot products for all node pairs, and (2) an increased number of parameters and overfitting risk due to the use of two separate channels. + +To alleviate these drawbacks, the proposed Dual Cross-Attention (DCA) module adopts two strategies: (1) linearized attention computation [50] and (2) parameter sharing, as shown in Fig. 1(b). Firstly, through approximating or replacing the Softmax attention utilizing separate kernel functions, the computation order in the SA module can be reordered from the standard $(\mathsf{Query}\times \mathsf{Key})\times \mathsf{Value}$ (Eq. 3) to the more efficient $\mathsf{Query}\times (\mathsf{Key}\times \mathsf{Value})$ format [28]. However, this strategy is unsuitable for the cross-attention module. Specifically, the attention score $k(\mathbf{B})^{\top}v(\mathbf{B})\in \mathbb{R}^{d\times d}$ computes the similarity between features within the same space, rather than across different spaces. To ensure cross-space interaction, DCA sets the Key to originate from the same space as the Query. Moreover, DCA shares parameters between the Query and Key to reduce the number of parameters. + +To streamline the description of the update process for topology representations and attribute representations, two abstract representations $\mathbf{H}_1$ and $\mathbf{H}_2$ are introduced. For clarity, layer indices are omitted. The general formulation of the DCA module is given as follows: + +$$ +D C A \left(\mathbf {H} _ {1}, \mathbf {H} _ {2}\right): \quad \mathbf {Q} = q \left(\mathbf {H} _ {1}\right), \mathbf {K} = k \left(\mathbf {H} _ {1}\right), \mathbf {V} = v \left(\mathbf {H} _ {2}\right), \tag {7} +$$ + +$$ +\tilde {\mathbf {Q}} = \frac {\mathbf {Q}}{\| \mathbf {Q} \| _ {\mathcal {F}}}, \quad \tilde {\mathbf {K}} = \frac {\mathbf {K}}{\| \mathbf {K} \| _ {\mathcal {F}}}, \tag {8} +$$ + +$$ +\mathbf {H} _ {1} ^ {*} = \mathbf {D} ^ {- 1} \left(\mathbf {V} + \frac {1}{n} \tilde {\mathbf {Q}} (\tilde {\mathbf {K}} ^ {\top} \mathbf {V})\right), \tag {9} +$$ + +where $q(\cdot)$ and $k(\cdot)$ stand for the Query and Key functions, respectively, with $q(\cdot) = k(\cdot)$ . And $v_{A}(\cdot)$ represents the Value function. These functions are implemented as MLPs. $\| \cdot \|_{\mathcal{F}}$ denotes the Frobenius norm. $\mathbf{D} = \text{Diag}\left(\mathbf{1} + \frac{1}{n}\tilde{\mathbf{Q}} (\tilde{\mathbf{K}}^{\top}\mathbf{1})\right)$ stands for a diagonal matrix and $\mathbf{1}$ is an all-one vector. The topology-related attribute representations can be obtained as $\mathbf{B}_{DCA}^{*} = DCA(\mathbf{B},\mathbf{Z})$ . Then, the topology representations are updated via + +$$ +\hat {\mathbf {Z}} = (1 - \tilde {\lambda}) \tilde {\mathbf {A}} \mathbf {V} + \tilde {\lambda} \mathbf {B} _ {D C A} ^ {*}, \tag {10} +$$ + +where $\tilde{\mathbf{A}}$ denotes the normalized adjacency matrix. The first term denotes the topology representation updated purely from the topology space, which can be viewed as being obtained via spectral clustering [48, 54] (see Theorem 3). $\tilde{\lambda} = \mathrm{Tanh}(\lambda)$ stands for a scalar to balance these two terms, where $\lambda$ is a learnable parameter. Combining these two terms allows for the fusion of topological details alongside the topology-related attribute information into the final topology representations. + +Similarly, the attribute representations are updated by incorporating relevant information from the topology space, that is, $\mathbf{Z}_{DCA}^{*} = DCA(\mathbf{Z},\mathbf{B})$ , with their representations. This can be expressed as + +$$ +\hat {\mathbf {B}} = (1 - \tilde {\gamma}) \mathbf {B} ^ {0} + \tilde {\gamma} \mathbf {Z} _ {D C A} ^ {*}, \tag {11} +$$ + +where $\tilde{\gamma} = \mathrm{Tanh}(\gamma)$ denotes a scalar to trade off the two terms with $\gamma$ denotes a learnable parameter. Note that DCA requires two separate sets of network parameters to generate the attribute and topology + +# Algorithm 1: UGCFoermer + +Input: Graph $\mathcal{G}(\mathbf{A},\mathbf{X})$ with labels $\mathbf{Y}$ , hyperparameters $\alpha$ , $\beta$ and $\tau$ . + +Output: Trained network parameters $\Theta^{*}$ . + +Initialization: Network parameters $\Theta$ + +# while not converged do + +1. Generate two initial node representations $\mathbf{Z}^0$ and $\mathbf{B}^0$ via Eq. 6; +2. Get two updated node representations $\hat{\mathbf{Z}}$ and $\hat{\mathbf{B}}$ via Eqs. 10 and 11; +3. Obtain the final predictions $\hat{\mathbf{Y}}$ via Eq. 12; +4. Calculate the overall loss $\mathcal{L}_{final}$ via Eqs. 13 and 14; +5. Optimize the parameters via $\Theta^{*}\gets \mathrm{Adam}(\mathcal{L},\Theta)$ + +# end + +# return Parameters $\Theta^{*}$ + +representations, e.g., $q_{A}(\cdot)$ and $q_{X}(\cdot)$ represent the Query for graph topology and node attributes, respectively, as shown in Fig. 1. + +Prediction Layer. After obtaining the topology representations $\hat{\mathbf{Z}}$ and attribute representations $\hat{\mathbf{B}}$ through $l$ layers, the final node representations can be generated by weight combining them. Next, the predictions are generated via an MLP network and nonlinearities (i.e., Softmax(\cdot)), that is, + +$$ +\hat {\mathbf {Y}} = \operatorname {S o f t m a x} \left(M L P \left((1 - \alpha) \hat {\mathbf {Z}} + \alpha \hat {\mathbf {B}}\right)\right), \tag {12} +$$ + +where $\alpha$ denotes a scalar that adjusts attention to topology and attribute representations. $\hat{\mathbf{Y}}\in \mathbb{R}^{n\times c}$ represents the predictions, indicating the estimated outcomes for each of the $n$ nodes across $c$ classes. + +Objective Function. Note that the proposed DCA module, with a large number of parameters across two distinct spaces, is susceptible to representation distortion caused by overfitting [8], especially when the number of training nodes is limited. Thus, a consistency constraint is introduced to align the two representations $\hat{\mathbf{Z}}$ and $\hat{\mathbf{B}}$ . First, pseudo-labels are derived by averaging the two representations. For a node $v$ , its pseudo-label can be computed as $\mathbf{y}_v = \frac{1}{2} (\hat{\mathbf{z}}_v + \hat{\mathbf{b}}_v)$ . Next, low-entropy pseudo-labels are obtained through a sharpening technique that controls the sharpness of the distribution. This can be formulated as $\bar{y}_{i,j} = y_{i,j}^{\frac{1}{\tau}} / \sum_{k=0}^{c-1} y_{i,j}$ , $(0 \leq j \leq c-1)$ , where $\tau \in (0,1]$ denotes a scaling factor that controls the sharpness of the distribution. + +Once the pseudo-label is obtained, the next step is to calculate the squared Euclidean distance between it and the two representations, which is given by + +$$ +\mathcal {L} _ {\text {c o n}} (\hat {\mathbf {Z}}, \hat {\mathbf {B}}) = \frac {1}{2} \sum_ {i} ^ {n - 1} \left(\| \bar {\mathbf {y}} _ {i} - \hat {\mathbf {z}} _ {i} \| _ {2} ^ {2} + \| \bar {\mathbf {y}} _ {i} - \hat {\mathbf {b}} _ {i} \| _ {2} ^ {2}\right). \tag {13} +$$ + +The overall objective of UGCFormer is to minimize the weighted sum of the cross-entropy loss and the consistency loss, defined as follows: + +$$ +\mathcal {L} _ {\text {o v e r a l l}} = \mathcal {L} _ {c e} + \beta \mathcal {L} _ {\text {c o n}}, \tag {14} +$$ + +where $\mathcal{L}_{ce} = -\sum_{v\in \mathcal{V}_L}\mathbf{y}_v\log \hat{\mathbf{y}}_v$ and $\beta$ stands for a balance hyperparameter. + +# 3.3 Model Analysis + +This subsection provides a comprehensive analysis of UGCFoerner. First, the computational complexity of UGCFoerner is analyzed. Then, the simplicity of UGCFoerner is examined through architectural comparison with existing GTs. Finally, the effectiveness of UGCFoerner is theoretically justified. + +Complexity Analysis. UGCFormaler operates with linear time complexity. The time complexity for generating initial representations through the projection layer is $O(md + nd^2)$ as the adjacency matrix is sparse, where $m$ represents the number of edges. Secondly, owing to the linearized cross-attention module, the aggregation operator incurs a computational overhead of $O(nd^2)$ . Finally, obtaining the predictions involves feature mapping and element-wise operations, resulting in a complexity of $O(nd)$ . UGCFormaler operates with linear space complexity. The space required to store the input topology + +Table 1: Accuracy (ACC) or ROC-AUC in percentage $(\mathrm{mean}_{\pm \mathrm{std}})$ over 10 trials of the node classification task on homophilic graphs. Best and runner-up models are in bold and underlined, respectively. + +
Model MetricCora ACC ↑CiteSeer ACC ↑PubMed ACC ↑Photo ACC ↑CS ACC ↑Physics ACC ↑Questions ROC-AUC ↑Avg ↑Rank ↓
GCN81.60±0.4071.60±0.4078.80±0.6092.70±0.2092.92±0.1296.18±0.0776.28±0.6484.3013.29
GAT83.00±0.7072.10±1.1079.00±0.4093.87±0.1093.61±0.1496.17±0.0874.94±0.5684.6711.43
GraphSAGE82.68±0.4771.93±0.8579.41±0.5394.59±0.1493.91±0.1396.49±0.0676.44±0.6285.069.71
APPNP83.30±0.5071.80±0.5080.10±0.2094.32±0.1494.49±0.0796.54±0.0775.51±0.2385.158.14
GPR-GNN84.20±0.5071.60±0.8080.07±0.9294.49±0.1695.13±0.0996.85±0.0867.15±1.9284.218.71
LINKX77.95±0.1268.25±0.2477.36±0.4291.97±0.1994.77±0.1996.29±0.1375.71±1.4083.1913.14
GloGNN82.17±0.2971.74±0.8880.37±0.9595.10±0.2095.00±0.1096.97±0.1567.15±1.9284.078.43
GraphGPS82.84±1.0372.73±1.2379.94±0.2695.06±0.1393.93±0.1297.12±0.1971.73±1.4784.768.29
NodeFormer82.20±0.9072.50±1.1079.90±1.0093.46±0.3595.64±0.2296.24±0.2474.27±1.4684.899.57
NAGphormer82.12±1.1871.47±1.3079.73±0.2895.49±0.1195.75±0.0997.34±0.0374.98±0.6385.277.71
Exphormer82.77±1.3871.63±1.1979.46±0.3595.35±0.2294.93±0.0196.89±0.0974.67±0.7985.108.86
GOAT83.18±1.2771.99±1.2679.13±0.3892.96±1.4894.21±0.3896.45±0.2875.76±1.6684.8110.00
SGFormer84.50±0.8072.60±0.2080.30±0.6095.10±0.4794.78±0.2096.60±0.1872.15±1.3185.146.57
Polynormer83.25±0.9372.31±0.7879.24±0.4396.46±0.2695.53±0.1697.27±0.0876.91±1.6385.854.71
Gradformer82.95±0.7372.80±0.5980.14±0.4895.76±0.2894.21±0.2997.06±0.1674.71±1.0785.386.14
UGCFomer84.94±0.4373.41±0.2781.79±0.8196.21±0.3195.91±0.2397.35±0.1777.02±0.7686.661.14
+ +and attributes is $O(m + nd)$ , where $m$ corresponds to the number of edges and $nd$ accounts for the feature matrix. The aggregated and updated representations each require $O(nd)$ space, since their dimensions do not exceed those of the input feature matrix. In the linearized attention computation (Fig. 1(b)), the attention matrix contributes an additional $O(d^2)$ space overhead. + +Components. To leverage discriminative graph topology and capture structural biases, existing GTs often resort to auxiliary components that compromise their efficiency and effectiveness. Specifically, the positional or structural encodings (e.g., Laplacian eigenvector encodings) used in GraphGPS [39], NAGphormer [2], Exphormer [44], and GOAT [31] as well as augmented training losses (e.g., edge regularization loss) in NodeFormer, often necessitate cubic computational complexity and quadratic space consumption. Moreover, the GNN module tends to generate representations that are susceptible to issues caused by the limited message passing. In contrast, the proposed UGCFformer features a streamlined and efficient design that relies solely on a linear cross-attention module. + +Theoretical Justification. Though designed to be simple and intuitive, the proposed UGCFoer is theoretically guaranteed to be effective from a graph optimization perspective [55, 58]. + +Theorem 3. Let $\mathbf{Z}$ and $\mathbf{B}$ denote the topology representations and attribute representations, respectively. The representation update in the dual cross-attention module DCA (Eq. 10 and Eq. 11) is equivalent to solving an optimization problem with the objective function: + +$$ +\underset {\mathbf {Z}, \mathbf {B}} {\arg \min } \lambda \operatorname {T r} \left(\mathbf {Z} ^ {\top} \tilde {\mathbf {L}} \mathbf {Z}\right) + \| \mathbf {B} - M L P (\mathbf {X}) \| _ {F} ^ {2} - \eta \| \mathbf {Z} ^ {\top} \mathbf {B} \| _ {F} ^ {2}, \tag {15} +$$ + +where $\tilde{\mathbf{L}}$ terms the Laplacian matrix of $\tilde{\mathbf{A}}$ , $\lambda$ and $\eta$ are the scalars used to balance these three terms. + +In Eq. 15, the first term stands for a relaxed optimization problem widely used in spectral clustering [48]. Thus, the DCA seeks to generate topology representations that capture mesoscopic community structures. The second term measures the distance between the attribute representation $\mathbf{B}$ and its initial representation $MLP(\mathbf{X})$ . The third term denotes the statistical dependence measure, approximated by the Hilbert-Schmidt Independence Criterion (HSIC) [19], that is, $\mathrm{HSIC}(\mathbf{Z},\mathbf{B})\approx \mathrm{Tr}(\mathbf{ZZ}^{\top}\mathbf{BB}^{\top}) = \|\mathbf{Z}^{\top}\mathbf{B}\|_{F}^{2}$ , which reflects the dependence between topology and attribute representations. Therefore, the interaction, whether mutual correlation (positive weights) or exclusion (negative weights), can be modulated by the parameters. In summary, Theorem 3 indicates that UGCFoer focuses on learning representations by mining the interactions of two basic graph information. + +# 4 Experiments + +This section evaluates the effectiveness and universality of the proposed UGCFormer by comparing its performances against various diverse graph learning models on the node classification task. Moreover, + +Table 2: Accuracy (ACC) in percentage (mean±std) over 10 trials of the node classification task on heterophilic graphs. Best and runner-up models are in bold and underlined, respectively. + +
Model MetricCornell ACC ↑Texas ACC ↑Wisconsin ACC ↑Actor ACC ↑Chameleon ACC ↑Squirrel ACC ↑Ratings ACC ↑Avg ↑Rank ↓
GCN58.41±3.2865.61±4.8061.28±5.8730.63±0.6243.43±1.9241.30±0.9447.77±0.6949.7811.57
GAT58.29±3.5260.73±6.2063.64±6.1830.36±0.9440.14±1.5735.09±0.7047.95±0.5348.0314.57
GraphSAGE75.95±5.3182.43±6.0781.18±4.5634.23±1.0739.11±5.0536.46±2.1653.11±0.5457.5011.00
APPNP73.68±3.9774.57±2.4870.61±3.4735.18±1.2139.42±3.8738.13±2.6749.78±0.7254.4912.71
GPR-GNN78.11±6.5581.35±5.3282.55±6.2335.16±0.8539.93±3.3038.95±1.9943.90±0.4857.1411.86
LINKX77.84±5.8174.60±8.3775.49±5.7236.10±1.5540.02±2.3539.88±2.5351.36±0.4756.4711.00
GloGNN83.51±4.2684.32±4.1587.06±3.5337.35±1.3038.43±3.7430.30±1.9237.28±0.6656.898.14
GraphGPS82.06±5.7382.21±6.1485.36±4.2436.18±1.2740.79±4.0339.67±2.8453.10±0.4259.917.29
NodeFormer82.15±6.7281.68±4.6583.41±5.5136.28±1.2543.09±2.8140.61±1.2550.12±0.6459.627.43
NAGphormer79.97±6.0780.18±4.5782.97±2.9834.36±0.7544.61±3.1041.27±1.0952.51±0.8359.417.86
Exphormer83.07±4.3182.81±3.5283.90±4.3136.82±1.9541.63±3.1240.32±1.5952.08±0.8160.066.00
GOAT83.18±1.2771.99±1.2679.13±0.3836.55±1.1942.56±3.1740.81±0.5449.68±0.5057.708.53
SGFormer81.64±3.8884.29±5.6783.59±5.4237.79±1.8944.93±3.9141.80±2.2748.01±0.4960.294.86
Polynormer81.90±4.1782.57±5.1183.95±2.9837.01±1.1041.97±3.1840.87±1.9653.29±0.2360.225.14
Gradformer83.06±5.1682.19±5.2484.26±2.2436.58±0.7140.73±3.6940.29±1.8853.11±0.2960.036.43
UGCFomer85.14±5.8384.59±4.6987.36±3.3037.41±0.7943.28±2.1741.56±2.0153.48±0.1461.831.71
+ +it provides additional analysis experiments to enhance the understanding of UGCFormer. Refer to Section E for details on the datasets, baselines, and experimental setups. + +# 4.1 Experimental Results + +Homophilic Graphs. The experiment results for node classification on homophilic graphs are shown in Tab. 1, from which three key observations can be made. Firstly, the performance of the backbone GNNs (e.g., GCN and GAT) lags behind that of GTs. To be specific, on six of the seven homophilic graphs, the models that rank in the top two positions are GTs. This is primarily because most GTs, such as NAGphormer, are built upon these backbone GNNs and specifically address the shortcomings of GNNs in capturing long-range dependencies. Secondly, the proposed UGCFoerner outperforms all baseline GTs across six of the seven datasets and achieves the optimal rank, demonstrating its consistent superior performance. In particular, on PubMed, UGCFoerner achieves a performance that is $2.55\%$ higher than the baseline Polynormer, which has an average rank of second, and its average rank is significantly lower. Thirdly, compared with the baseline LINKX, which also processes graph topology and node attributes separately and does not leverage message passing, UGCFoerner consistently achieves better results across all datasets. This can be attributed to its ability to capture the interactions between these two types of graph information and alleviate the representation distortion, which LINKX does not account for. This highlights the rationality of UGCFoerner's design. + +Heterophilic Graphs. Tab. 2 shows the results of the node classification task on seven heterophilic graphs, highlighting three key observations. Firstly, the baseline GTs perform slightly better than the baseline GNNs, but the difference is not substantial. In specific, the baseline GNNs, particularly GloGNN on Cornell, Texas, and Wisconsin, and GraphSAGE on the Ratings, achieve top-two results on five of the seven datasets. This can be attributed to the high complexity and large number of parameters in GTs, which make them prone to overfitting. Therefore, the baseline SGFormer, which linearly combines the local representation from the GNN module and the global representation from the GT module, achieves superior performance. This is evidenced by its ranking in the top two for three datasets. Secondly, the proposed UGCFoer outperforms the GT baselines on the majority of heterophilic graphs, proving its effectiveness. For example, on Cornell, UGCFoer exceeds the second-ranked GT, i.e., GOAT, by a significant margin of $1.96\%$ . Thirdly, UGCFoer consistently outperforms the baseline LINKX on all heterophilic datasets, highlighting the significance of capturing the relevance between graph topology and node attributes. Overall, UGCFoer achieves performance improvements on both homophilic and heterophilic graphs, demonstrating its universality. + +Scalability Study. To evaluate the scalability of the proposed UGCFoer, this experiment quantitatively changes the network size and records the running time and GPU memory usage. Specifically, it utilizes the ogbn-arxiv to randomly sample subsets of nodes, with the node numbers varying from 10K to 100K. As shown in Fig. 2, the running time and GPU memory usage of UGCFoer increase linearly with the size of the sampled graph. For example, the training time and memory usage with + +100k nodes are approximately five times higher than with 20k nodes. This indicates that UGCFoer exhibits linear time and space complexity, consistent with the conclusion in Section 3.3. + +![](images/cfce089f7f365d07581f5fe84f8c9c2d217d9d899222eadf09f4f580f5054844.jpg) +Figure 2: Training time and GPU memory usage of UGCFoermer. + +![](images/9edd18fa5801d72c24b72b3e2c201faf6b7548cbcae5e4718c9a9a519d5712e1.jpg) + +Node Property Prediction. This experiment seeks to evaluate the effectiveness and scalability of GTs by comparing them with GNNs on two large-scale benchmark datasets. Upon examining Tab. 3, which presents the results of the node property prediction task on these two datasets, two key conclusions can be drawn. Firstly, the backbone GTs generally outperform the backbone GNNs, which not only highlights the superiority of GTs but also underscores their scalability—a key challenge that GTs aim to address. This can be attributed to the integration of the GNN blocks in GTs, exemplified by SGFormer. These GTs generate the final prediction by combining the local representations from the GNN module with the global representations from the GT module. Secondly, the proposed UGCFormaler achieves optimal performance on these two datasets, indicating its effectiveness and scalability on large graphs. + +Table 3: Node property prediction performances on two large-scale graphs. + +
Model Metricogbn-proteins ROC-AUC ↑ogbn-arxiv ACC ↑
GCN72.51±0.3571.74±0.29
GAT72.02±0.4471.95±0.36
GPRGNN71.10±0.1271.10±0.12
LINKX66.18±0.3371.59±0.71
GraphGPS76.83±0.2670.97±0.41
NodeFormer77.45±1.1567.19±0.83
NAGphormer73.61±0.3370.13±0.55
Exphormer74.58±0.2672.44±0.28
GOAT74.84±1.1672.41±0.40
SGFormer79.53±0.3872.63±0.13
Polynormer78.97±0.4773.46±0.16
Gradformer77.64±0.5172.71±0.20
UGCFormaler79.95±0.7574.02±0.17
+ +# 4.2 Additional Analysis + +Ablation Study. This experiment evaluates the contributions of the proposed cross-attention module and the consistency constraint by comparing UGCFormaler with two variants lacking these components. Fig. 3 shows that these variants consistently underperform UGCFormaler across the four datasets. This illustrates that the efficacy of UGCFormaler stems from the collective contribution of all components. Besides, even without the consistency loss, the variant model $(\mathrm{w / o}\mathcal{L}_{con})$ still provides competitive performance compared to the baseline GTs, as seen in Table 1. This highlights the effectiveness of the cross-attention module and thereby reaffirms the rationality of the UGCFormaler architecture. + +![](images/a098dfd53ef615d6ff03df8e5ffe8a1d14eceb4b39083ced10f96e4e2c40ac50.jpg) +Figure 3: Impact of functional components (i.e., the CA and consistency constraint). + +![](images/c9928c4578fb2e8f7538014eebb2f6c6eed40cca6a74c2f7c07559e5aaf6975d.jpg) +Figure 4: Performance variations for varying $l$ . + +![](images/b43909dfaf89775f1c88c423832c0066d43b241daa8806c0718507e0ff1223d6.jpg) +Figure 5: Performance variations for varying $d$ . + +Parameter Sensitivity Analysis. These experiments aim to provide an intuitive understanding for the selection of hyperparameters. Performance changes due to varying the number of layers $(l)$ and layer dimensions $(d)$ are shown in Figs. 4 and 5, respectively. Number of Layers. Fig. 4 shows that UGCFoer has achieves stable performance across various layer numbers $\{1,2,3,4,5\}$ . Specifically, performance fluctuations are minimal, within $2.2\%$ on the Cora, $1.4\%$ on the CiteSeer, and $1.3\%$ on PubMed. This indicates that UGCFoer is relatively insensitive to the number of layers. Additionally, optimal performance is achieved with $\{3,4\}$ , likely due to the risk of over-smoothing in + +deeper models. Hidden Layer Dimension. As shown in Fig. 5, UGCFormaler maintains consistent performance across the hidden dimension range $\{64,128,256,512\}$ . For example, on the Cora, which shows the most significant performance variation, the difference is less than $2\%$ . This indicates that UGCFormaler is not sensitive to this parameter. Additionally, optimal performance on the three datasets corresponds to $d \in \{128,256\}$ , rather than the highest value of 512. This suggests that larger dimensions can lead to overfitting and distorted representations. Additional hyper-parameters (including $\alpha$ and $\beta$ ) are analyzed in Section E.4. + +# 5 Conclusions + +By revisiting two typical Graph Transformers (GTs), this study has uncovered a potential functional mechanism: cross-aggregation between graph topology and node attributes. To effectively implement this mechanism, this paper introduces UGCFoer, a linearized graph cross-attention Transformer. Extensive experiments on sixteen graph benchmarks demonstrate its effectiveness and efficiency. + +# 6 Acknowledgements + +This work was supported in part by the National Natural Science Foundation of China (No. 92570118, U22B2036, 62376088, 62272020, 62025604, 92370111, 62272340, 62261136549), in part by the Hebei Natural Science Foundation (No. F2024202047), in part by the National Science Fund for Distinguished Young Scholarship (No. 62025602), in part by the Hebei Yanzhao Golden Platform Talent Gathering Programme Core Talent Project (Education Platform) (HJZD202509), in part by the Post-graduate's Innovation Fund Project of Hebei Province (CXZZBS2025036), in part by the Tencent Foundation, and in part by the XPLORER PRIZE. + +# References + +[1] Dexiong Chen, Leslie O'Bray, and Karsten M. Borgwardt. Structure-aware transformer for graph representation learning. In ICML, volume 162, pages 3469-3489, 2022. +[2] Jinsong Chen, Kaiyuan Gao, Gaichao Li, and Kun He. Nagphormer: A tokenized graph transformer for node classification in large graphs. In ICLR, 2023. +[3] Jinsong Chen, Chenyang Li, Gaichao Li, John E. Hopcroft, and Kun He. Rethinking tokenized graph transformers for node classification. CoRR, abs/2502.08101, 2025. +[4] Jinsong Chen, Hanpeng Liu, John E. Hopcroft, and Kun He. Leveraging contrastive learning for enhanced node representations in tokenized graph transformers. In NeurIPS, 2024. +[5] Zhaoliang Chen, Zhihao Wu, Ylli Sadikaj, Claudia Plant, Hong-Ning Dai, Shiping Wang, Yu-Ming Cheung, and Wenzhong Guo. Adedgedrop: Adversarial edge dropping for robust graph neural networks. IEEE Transactions on Knowledge and Data Engineering, 37(9):4948-4961, 2025. +[6] Eli Chien, Jianhao Peng, Pan Li, and Olgica Milenkovic. Adaptive universal generalized pagerank graph neural network. In ICLR, 2021. +[7] Chenhui Deng, Zichao Yue, and Zhiru Zhang. Polynormer: Polynomial-expressive graph transformer in linear time. In ICLR, 2024. +[8] Claudio Filipi Goncalves dos Santos and João Paulo Papa. Avoiding overfitting: A survey on regularization methods for convolutional neural networks. ACM Comput. Surv., 54(10s):213:1-213:25, 2022. +[9] Vijay Prakash Dwivedi, Ladislav Rampasek, Michael Galkin, Ali Parviz, Guy Wolf, Anh Tuan Luu, and Dominique Beaini. Long range graph benchmark. In NeurIPS, 2022. +[10] Ruiyi Fang, Bingheng Li, Zhao Kang, Qiuhao Zeng, Nima Hosseini Dashtbayaz, Ruizhi Pu, Charles Ling, and Boyu Wang. On the benefits of attribute-driven graph domain adaptation. In ICLR, 2025. + +[11] Ruiyi Fang, Bingheng Li, Jingyu Zhao, Ruizhi Pu, Qiuhao Zeng, Gezheng Xu, Charles Ling, and Boyu Wang. Homophily enhanced graph domain adaptation. In ICML, 2025. +[12] Ruiyi Fang, Liangjian Wen, Zhao Kang, and Jianzhuang Liu. Structure-preserving graph representation learning. In ICDM, pages 927-932. IEEE, 2022. +[13] Matthias Fey and Jan Eric Lenssen. Fast graph representation learning with pytorch geometric. CoRR, abs/1903.02428, 2019. +[14] Chaofan Fu, Guanjie Zheng, Chao Huang, Yanwei Yu, and Junyu Dong. Multiplex heterogeneous graph neural network with behavior pattern modeling. In SIGKDD, pages 482-494, 2023. +[15] Mozhdeh Gheini, Xiang Ren, and Jonathan May. Cross-attention is all you need: Adapting pretrained transformers for machine translation. In EMNLP, pages 1754–1765, 2021. +[16] Justin Gilmer, Samuel S. Schoenholz, Patrick F. Riley, Oriol Vinyals, and George E. Dahl. Neural message passing for quantum chemistry. In ICML, 2017. +[17] Francesco Di Giovanni, T. Konstantin Rusch, Michael M. Bronstein, Andreea Deac, Marc Lackenby, Siddhartha Mishra, and Petar Velickovic. How does over-squashing affect the power of gnns? Trans. Mach. Learn. Res., 2024. +[18] Maoguo Gong, Hui Zhou, A. K. Qin, Wenfeng Liu, and Zhongying Zhao. Self-paced co-training of graph neural networks for semi-supervised node classification. IEEE Transactions on Neural Networks and Learning Systems, 34(11):9234-9247, 2023. +[19] Arthur Gretton, Olivier Bousquet, Alexander J. Smola, and Bernhard Scholkopf. Measuring statistical dependence with hilbert-schmidt norms. In ALT, volume 3734, pages 63-77, 2005. +[20] William L. Hamilton, Zhitao Ying, and Jure Leskovec. Inductive representation learning on large graphs. In NeurIPS, pages 1024-1034, 2017. +[21] Kai Han, Yunhe Wang, Hanting Chen, Xinghao Chen, Jianyuan Guo, Zhenhua Liu, Yehui Tang, An Xiao, Chunjing Xu, Yixing Xu, et al. A survey on vision transformer. IEEE transactions on pattern analysis and machine intelligence, 45(1):87-110, 2022. +[22] Dongxiao He, Yongqi Huang, Jitao Zhao, Xiaobao Wang, and Zhen Wang. Str-gcl: Structural commonsense driven graph contrastive learning. In WWW, pages 1129-1141, 2025. +[23] Weihua Hu, Matthias Fey, Marinka Zitnik, Yuxiao Dong, Hongyu Ren, Bowen Liu, Michele Catasta, and Jure Leskovec. Open graph benchmark: Datasets for machine learning on graphs. In NeurIPS, 2020. +[24] Yongqi Huang, Jitao Zhao, Dongxiao He, Di Jin, Yuxiao Huang, and Zhen Wang. Does gcl need a large number of negative samples? enhancing graph contrastive learning with effective and efficient negative sampling. In AAAI, volume 39, pages 17511-17518, 2025. +[25] Yongqi Huang, Jitao Zhao, Dongxiao He, Xiaobao Wang, Yawen Li, Yuxiao Huang, Di Jin, and Zhiyong Feng. One prompt fits all: Universal graph adaptation for pretrained models. arXiv preprint arXiv:2509.22416, 2025. +[26] Zilong Huang, Xinggang Wang, Yunchao Wei, Lichao Huang, Humphrey Shi, Wenyu Liu, and Thomas S. Huang. Ccnet: Criss-cross attention for semantic segmentation. IEEE Trans. Pattern Anal. Mach. Intell., 45(6):6896-6908, 2023. +[27] Md. Shamim Hussain, Mohammed J. Zaki, and Dharmashankar Subramanian. Global self-attention as a replacement for graph convolution. In KDD, pages 655–665. ACM, 2022. +[28] Angelos Katharopoulos, Apoorv Vyas, Nikolaos Pappas, and François Fleuret. Transformers are rnns: Fast autoregressive transformers with linear attention. In ICML, volume 119, pages 5156-5165, 2020. +[29] Thomas N. Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. In ICLR, 2017. + +[30] Johannes Klicpera, Aleksandar Bojchevski, and Stephan Gunnemann. Predict then propagate: Graph neural networks meet personalized pagerank. In ICLR, 2019. +[31] Kezhi Kong, Jiuhai Chen, John Kirchenbauer, Renkun Ni, C. Bayan Bruss, and Tom Goldstein. GOAT: A global transformer on large-scale graphs. In ICML, volume 202, pages 17375-17390, 2023. +[32] Xiang Li, Chaofan Fu, Zhongying Zhao, Guangjie Zheng, Chao Huang, Yanwei Yu, and Junyu Dong. Dual-channel multiplex graph neural networks for recommendation. IEEE Transactions on Knowledge and Data Engineering, 37(6):3327-3341, 2025. +[33] Xiang Li, Renyu Zhu, Yao Cheng, Caihua Shan, Siqiang Luo, Dongsheng Li, and Weining Qian. Finding global homophily in graph neural networks when meeting heterophily. In ICML, volume 162, pages 13242-13256, 2022. +[34] Derek Lim, Felix Hohne, Xiuyu Li, Sijia Linda Huang, Vaishnavi Gupta, Omkar Bhalerao, and Ser-Nam Lim. Large scale learning on non-homophilous graphs: New benchmarks and strong simple methods. In NeurIPS, pages 20887-20902, 2021. +[35] Tianyang Lin, Yuxin Wang, Xiangyang Liu, and Xipeng Qiu. A survey of transformers. AI Open, 3:111-132, 2022. +[36] Chuang Liu, Zelin Yao, Yibing Zhan, Xueqi Ma, Shirui Pan, and Wenbin Hu. Gradformer: Graph transformer with exponential decay. In IJCAI, pages 2171-2179, 2024. +[37] Hongbin Pei, Bingzhe Wei, Kevin Chen-Chuan Chang, Yu Lei, and Bo Yang. Geom-gcn: Geometric graph convolutional networks. In ICLR, 2020. +[38] Oleg Platonov, Denis Kuznedev, Michael Diskin, Artem Babenko, and Liudmila Prokhorenkova. A critical look at the evaluation of gnns under heterophily: Are we really making progress? In ICLR, 2023. +[39] Ladislav Rampásek, Michael Galkin, Vijay Prakash Dwivedi, Anh Tuan Luu, Guy Wolf, and Dominique Beini. Recipe for a general, powerful, scalable graph transformer. In NeurIPS, 2022. +[40] Benedek Rozemberczki, Carl Allen, and Rik Sarkar. Multi-scale attributed node embedding. J. Complex Networks, 9(2), 2021. +[41] David E Rumelhart, Geoffrey E Hinton, and Ronald J Williams. Learning representations by back-propagating errors. nature, 323(6088):533-536, 1986. +[42] Prithviraj Sen, Galileo Namata, Mustafa Bilgic, Lise Getoor, Brian Gallagher, and Tina Eliassi-Rad. Collective classification in network data. AI Mag., 29(3):93-106, 2008. +[43] Oleksandr Shchur, Maximilian Mumme, Aleksandar Bojchevski, and Stephan Gunnemann. Pitfalls of graph neural network evaluation. CoRR, abs/1811.05868, 2018. +[44] Hamed Shirzad, Ameya Velingker, Balaji Venkatachalam, Danica J. Sutherland, and Ali Kemal Sinop. Exphormer: Sparse transformers for graphs. In ICML, volume 202, pages 31613-31632. PMLR, 2023. +[45] Jie Tang, Jimeng Sun, Chi Wang, and Zi Yang. Social influence analysis in large-scale networks. In SIGKDD, pages 807-816, 2009. +[46] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NIPS, pages 5998-6008, 2017. +[47] Petar Velickovic, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Lio, and Yoshua Bengio. Graph attention networks. CoRR, abs/1710.10903, 2017. +[48] Ulrike von Luxburg. A tutorial on spectral clustering. Stat. Comput., 17(4):395-416, 2007. + +[49] Qitian Wu, Wentao Zhao, Zenan Li, David P. Wipf, and Junchi Yan. Nodeformer: A scalable graph structure learning transformer for node classification. In *NeurIPS*, 2022. +[50] Qitian Wu, Wentao Zhao, Chenxiao Yang, Hengrui Zhang, Fan Nie, Haitian Jiang, Yatao Bian, and Junchi Yan. Simplifying and empowering transformers for large-graph representations. In NeurIPS, 2023. +[51] Zhanghao Wu, Paras Jain, Matthew A. Wright, Azalia Mirhoseini, Joseph E. Gonzalez, and Ion Stoica. Representing long-range context for graph neural networks with global attention. In NeurIPS, pages 13266-13279, 2021. +[52] Zhihao Wu, Zhaoliang Chen, Shide Du, Sujia Huang, and Shiping Wang. Graph convolutional network with elastic topology. Pattern Recognition, 151:110364, 2024. +[53] Yujie Xing, Xiao Wang, Yibo Li, Hai Huang, and Chuan Shi. Less is more: on the overglobalizing problem in graph transformers. In ICML, 2024. +[54] Liang Yang, Zhenna Li, Jiaming Zhuo, Jing Liu, Ziyi Ma, Chuan Wang, Zhen Wang, and Xiaochun Cao. Graph contrastive learning with joint spectral augmentation of attribute and topology. In AAAI, pages 21983-21991, 2025. +[55] Liang Yang, Chuan Wang, Junhua Gu, Xiaochun Cao, and Bingxin Niu. Why do attributes propagate in graph convolutional neural networks? In AAAI, pages 4590-4598, 2021. +[56] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In CVPR, pages 5718-5729, 2022. +[57] Zhongying Zhao, Zhan Yang, Chao Li, Qingtian Zeng, Weili Guan, and Mengchu Zhou. Dual feature interaction-based graph convolutional network. IEEE Transactions on Knowledge and Data Engineering, 35(9):9019-9030, 2023. +[58] Meiqi Zhu, Xiao Wang, Chuan Shi, Houye Ji, and Peng Cui. Interpreting and unifying graph neural networks with an optimization framework. In WWW, pages 1215-1226, 2021. +[59] Jiaming Zhuo, Can Cui, Kun Fu, Bingxin Niu, Dongxiao He, Yuanfang Guo, Zhen Wang, Chuan Wang, Xiaochun Cao, and Liang Yang. Propagation is all you need: A new framework for representation learning and classifier training on graphs. In MM, pages 481-489, 2023. +[60] Jiaming Zhuo, Can Cui, Kun Fu, Bingxin Niu, Dongxiao He, Chuan Wang, Yuanfang Guo, Zhen Wang, Xiaochun Cao, and Liang Yang. Graph contrastive learning reimagined: Exploring universality. In WWW, pages 641-651, 2024. +[61] Jiaming Zhuo, Yuwei Liu, Yintong Lu, Ziyi Ma, Kun Fu, Chuan Wang, Yuanfang Guo, Zhen Wang, Xiaochun Cao, and Liang Yang. Dualformer: Dual graph transformer. In ICLR, 2025. +[62] Jiaming Zhuo, Feiyang Qin, Can Cui, Kun Fu, Bingxin Niu, Mengzhu Wang, Yuanfang Guo, Chuan Wang, Zhen Wang, Xiaochun Cao, and Liang Yang. Improving graph contrastive learning via adaptive positive sampling. In CVPR, pages 23179-23187, 2024. + +# A Algorithm Description + +A layer of the proposed dual cross-attention module DCA is depicted in Algorithm 2. + +Algorithm 2: PyTorch-style Code for DCA layer +```python +# N: instance number +# D: hidden dimension +# z: data embeddings sized [N, D] +# b: data embeddings sized [N, D] +# H: head number +# Wq, Wk, Wv: parameter matrices for feature transformation +q = Wq(z) # [N, H, D] +k = Wk(z) # [N, H, D] +v = Wv(b) # [N, H, D] +# numerator +kv = torch.einsum("lhm, lhd → hmd", k, v) +num = torch.einsum("nhm, hmd → nhd", q, kv) +num += N * v # [N, H, D] +# denominator +all Ones = torch Ones(N) +k_sum = torch.einsum("lhm, l → hm", k, all Ones) +den = torch.einsum("nhm, hm → nh", q, k_sum) # [N, H] +# aggregated results +den += torch.ones_like(len) * N +output = num / den.unsqueeze(2) # [N, H, D] +# head average +output = output.mean(dim=1) # [N, D] +``` + +# B Proof for Theorem 1 + +Proof. The proof unfolds in three stages: firstly, a concise description of the model to be validated is provided; subsequently, the model is decomposed into its topology and attribute components; and finally, it is verified that this structure aligns with the cross-aggregation form. + +For a single-layer GNN block, such as GCN [29] widely used in GTs [39, 2, 50], the feature update can be expressed as + +$$ +\mathbf {H} = \sigma (\tilde {\mathbf {A}} \mathbf {X} \mathbf {W}). \tag {16} +$$ + +where the diffusion matrix $\tilde{\mathbf{A}}$ denotes the normalized adjacency matrix. + +By omitting the nonlinear activation function and performing eigendecomposition on the adjacency matrix, the above equation can be reformulated as + +$$ +\begin{array}{l} \mathbf {H} = \tilde {\mathbf {A}} \mathbf {X} \mathbf {W} \\ = \mathbf {U} \boldsymbol {\Lambda} \mathbf {U} ^ {\top} \mathbf {X} \mathbf {W} \tag {17} \\ = \mathbf {U} \sqrt {\boldsymbol {\Lambda}} \sqrt {\boldsymbol {\Lambda}} \mathbf {U} ^ {\top} \mathbf {X} \mathbf {W}. \\ \end{array} +$$ + +Next, let $\mathbf{B} = \mathbf{U}\sqrt{\boldsymbol{\Lambda}}$ and $\mathbf{Z} = \mathbf{X}\mathbf{W}$ . The similarity function $\operatorname{Sim}(\cdot, \cdot)$ is implemented as matrix multiplication. With the definitions in Eq. 5, the GCN block can be interpreted as a cross-aggregation from the attribute space to the topology space. + +Extension to Multi-layer GNN Blocks. Following the discussion on single-layer GNNs, the solution for multi-layer GNNs is presented. Under the above assumptions, the node representations in the $l$ -th layer can be formulated as + +$$ +\begin{array}{l} \mathbf {H} ^ {l} = \left(\tilde {\mathbf {A}} ^ {1} \tilde {\mathbf {A}} ^ {2} \dots \tilde {\mathbf {A}} ^ {l}\right) \mathbf {X} \left(\mathbf {W} ^ {1} \mathbf {W} ^ {2} \dots \mathbf {W} ^ {l}\right) \tag {18} \\ = \mathcal {S} \mathbf {X} \mathcal {W}, \\ \end{array} +$$ + +where $\tilde{\mathbf{A}}^i$ and $\mathbf{W}^i$ denote the diffusion matrix and the parameter matrix, respectively, in the $i$ -th layer. $\mathcal{S} = \prod_{i=0}^{l} \mathbf{S}^l$ and $\mathcal{W} = \prod_{i=0}^{l} \mathbf{W}^l$ stand for the product of the diffusion matrices and the projection matrices. Given the properties of the diffusion matrix, the product diffusion matrix can be eigen-decomposed as $\mathcal{S} = \mathbf{U} \boldsymbol{\Lambda}^{(l)} \mathbf{U}^\top$ , where $\boldsymbol{\Lambda}^{(l)}$ represents the $l$ power of $\boldsymbol{\Lambda}$ . Therefore, the above conclusion still holds in the context of multi-layer GNNs. + +Remark. This interpretation holds under the assumption that the diffusion operators across layers share a common eigenspace (i.e., identical or mutually commutative $\tilde{\mathbf{A}}$ across layers). Otherwise, the equivalence serves as a first-order approximation of the aggregation process. + +# C Proof for Theorem 2 + +Proof. This proof first expands the model based on the feature and parameter matrices. Then, it identifies the representation updates of the topology and attributes within it. Finally, it establishes the relationship between these updated expressions and the cross-aggregation. + +Firstly, the function $\text{Softmax}(\cdot)$ can be approximated using Random Features mappings, that is, $\mathbf{H} = \text{Softmax}(\mathbf{Q}\mathbf{K}^{\top})\mathbf{V} \approx \phi(\mathbf{Q})\phi(\mathbf{K})^{\top}\mathbf{V}$ . Here, $\phi(\cdot)$ denotes a kernel-based feature mapping that linearizes the attention computation. In practice, the learnable projection matrix $\mathbf{W}$ applied to $\mathbf{X}$ can be viewed as a parametric approximation to this mapping. + +By expanding node attributes $\mathbf{X} = [\mathbf{X};\mathbf{P}]$ , where $\mathbf{X}\in \mathbb{R}^{n\times f}$ and $\mathbf{P}\in \mathbb{R}^{n\times k}$ , it can be obtained as + +$$ +\mathbf {H} = \left(\left[ \mathbf {X}; \mathbf {P} \right] \mathbf {W} ^ {q} \left(\mathbf {W} ^ {k}\right) ^ {\top} \left[ \begin{array}{l} \mathbf {X} ^ {\top} \\ \mathbf {P} ^ {\top} \end{array} \right]\right) \left[ \mathbf {X}; \mathbf {P} \right] \mathbf {W} ^ {v} \tag {19} +$$ + +Then, by expanding the Query (like $\mathbf{W}^q = \left[ \begin{array}{l}\mathbf{W}_1^q\\ \mathbf{W}_2^q \end{array} \right]$ ), and the Key and Value, it can be derived as + +$$ +\begin{array}{l} \mathbf {H} = \left(\left(\mathbf {X} \mathbf {W} _ {1} ^ {q} + \mathbf {P} \mathbf {W} _ {2} ^ {q}\right) \left(\left(\mathbf {W} _ {1} ^ {k}\right) ^ {\top} \mathbf {X} ^ {\top} + \left(\mathbf {W} _ {2} ^ {k}\right) ^ {\top} \mathbf {P} ^ {\top}\right)\right) \left(\mathbf {X} \mathbf {W} _ {1} ^ {v} + \mathbf {P} \mathbf {W} _ {2} ^ {v}\right) \\ = \left(\left(\mathbf {X} \mathbf {W} _ {1} ^ {q} + \mathbf {P} \mathbf {W} _ {2} ^ {q}\right) \left(\left(\mathbf {W} _ {1} ^ {k}\right) ^ {\top} \mathbf {X} ^ {\top} + \left(\mathbf {W} _ {2} ^ {k}\right) ^ {\top} \mathbf {P} ^ {\top}\right)\right) \mathbf {X} \mathbf {W} _ {1} ^ {v} \tag {20} \\ + \left(\left(\mathbf {X} \mathbf {W} _ {1} ^ {q} + \mathbf {P} \mathbf {W} _ {2} ^ {q}\right) \left(\left(\mathbf {W} _ {1} ^ {k}\right) ^ {\top} \mathbf {X} ^ {\top} + \left(\mathbf {W} _ {2} ^ {k}\right) ^ {\top} \mathbf {P} ^ {\top}\right)\right) \mathbf {P} \mathbf {W} _ {2} ^ {v}. \\ \end{array} +$$ + +It is evident that the equation includes several terms that describe the self-aggregation of topology and attributes. For instance, $\mathbf{PW}_2^q (\mathbf{PW}_2^k)^\top \mathbf{PW}_2^v$ and $\mathbf{XW}_1^q (\mathbf{XW}_1^k)^\top \mathbf{XW}_1^v$ stand for the self-aggregation of topology and attributes, respectively. Furthermore, this equation contains several terms that include cross-aggregation across topology and attributes. Exaggered by the term $\mathbf{XW}_1^q (\mathbf{PW}_2^k)^\top \mathbf{PW}_2^v$ , by setting $\mathbf{Z} = \mathbf{XW}_1^q$ and $\mathbf{B} = \mathbf{PW}_2$ , the cross-aggregation from the topology space to the attribute space can be obtained. Similarly, terms (e.g., $\mathbf{PW}_2^q (\mathbf{PW}_2^k)^\top \mathbf{XW}_1^v$ ) describing cross-aggregation from the attribute space to the topology space can be found. + +# D Proof for Theorem 3 + +Proof. This proof includes two main steps. First, we plan to derive the closed-form solutions for $\mathbf{Z}$ and $\mathbf{B}$ from the convex optimization objective (Eq. 15). These solutions are denoted as $\mathbf{Z}^*$ and $\mathbf{B}^*$ , respectively. Second, we aim to establish the equivalence between these derived solutions $\mathbf{Z}^*$ and $\mathbf{B}^*$ and the feature updates in Eq. 10 and Eq. 11, respectively. + +Let us denote the objective function (Eq. 15) as $\mathcal{O}(\mathbf{Z},\mathbf{B})$ , that is + +$$ +\begin{array}{l} \mathcal {O} (\mathbf {Z}, \mathbf {B}) = \lambda \operatorname {T r} \left(\mathbf {Z} ^ {\top} \tilde {\mathbf {L}} \mathbf {Z}\right) + \| \mathbf {B} - M L P (\mathbf {X}) \| _ {F} ^ {2} - \eta \| \mathbf {Z} ^ {\top} \mathbf {B} \| _ {F} ^ {2} \\ = \lambda \operatorname {T r} (\mathbf {Z} ^ {\top} (\mathbf {I} - \tilde {\mathbf {A}}) \mathbf {Z}) \\ + \operatorname {T r} \left(\left(\mathbf {B} - M L P (\mathbf {X})\right) ^ {\top} \left(\mathbf {B} - M L P (\mathbf {X})\right)\right) \tag {21} \\ - \eta \operatorname {T r} \left(\left(\mathbf {Z} ^ {\top} \mathbf {B}\right) ^ {\top} \left(\mathbf {Z} ^ {\top} \mathbf {B}\right)\right) \\ \end{array} +$$ + +Firstly, the partial derivatives of $\mathcal{O}(\mathbf{Z},\mathbf{B})$ with respect to $\mathbf{Z}$ can be calculated as + +$$ +\frac {\partial \mathcal {O} (\mathbf {Z} , \mathbf {B})}{\partial \mathbf {Z}} = 2 \lambda (\mathbf {I} - \tilde {\mathbf {A}}) \mathbf {Z} - 2 \eta \left(\mathbf {B} \mathbf {B} ^ {\top} \mathbf {Z}\right) \tag {22} +$$ + +The closed-form solution $\mathbf{Z}^*$ for the objective function $\mathcal{O}$ can be derived by setting $\frac{\partial\mathcal{O}(\mathbf{Z},\mathbf{B})}{\partial\mathbf{Z}} = 0$ that is + +$$ +2 \lambda (\mathbf {I} - \tilde {\mathbf {A}}) \mathbf {Z} - 2 \eta (\mathbf {B} \mathbf {B} ^ {\top} \mathbf {Z}) = 0 \tag {23} +$$ + +$$ +\Rightarrow \mathbf {Z} ^ {*} = \tilde {\mathbf {A}} \mathbf {Z} + \frac {\eta}{\lambda} \mathbf {B} \mathbf {B} ^ {\top} \mathbf {Z} \tag {24} +$$ + +Then, by setting $\omega_{1} = \frac{\eta}{\lambda + \eta}$ , we obtain + +$$ +\mathbf {Z} ^ {*} = \left(1 - \omega_ {1}\right) \tilde {\mathbf {A}} \mathbf {Z} + \omega_ {1} \mathbf {B} \mathbf {B} ^ {\top} \mathbf {Z} \tag {25} +$$ + +For the update of topology representations in the proposed DCA module, i.e., Eq. 10, the process can be rewritten as + +$$ +\hat {\mathbf {Z}} = (1 - \tilde {\lambda}) \tilde {\mathbf {A}} \mathbf {V} + \tilde {\lambda} \mathbf {S} \mathbf {V}, \tag {26} +$$ + +where $\mathbf{S} \in \mathbb{R}^{n \times n}$ stands for the cross-attention score matrix, with the scores $s_{i,i} = \frac{n + \mathbf{q}_{i,:} \mathbf{k}_{i,:}^{\top}}{n + \sum_{t} \mathbf{q}_{i,:} \mathbf{k}_{t,:}^{\top}}$ and $s_{i,j} = \frac{\mathbf{q}_{i,:} \mathbf{k}_{j,:}^{\top}}{n + \sum_{t} \mathbf{q}_{i,:} \mathbf{k}_{t,:}^{\top}}$ for $i \neq j$ . + +To demonstrate the equivalence between Eq. 25 and Eq. 26, the first step is to set $\omega_{1} = \tilde{\lambda}$ . With this condition met, the required proof is to derive a matrix $\mathbf{B}$ that satisfies the equation $\mathbf{BB}^{\top} = \mathbf{S}$ . + +Let us denote the diagonal matrix as $\mathbf{D}$ , where $d_{i,i} = \sqrt{\frac{n + 1}{n + \sum_t\mathbf{q}_{i,:}\mathbf{q}_{t,:}^\top}}$ , the construction of $\mathbf{B} = \mathbf{DQ}$ ensures that the above equation holds. + +For the diagonal elements of $\mathbf{BB}^{\top}$ , there is + +$$ +\left(\mathbf {B B} ^ {\top}\right) _ {i, i} = \sum_ {j = 0} ^ {d - 1} b _ {i, j} ^ {2} = \sum_ {j = 0} ^ {d - 1} \left(d _ {i, i} \cdot q _ {i, j}\right) ^ {2} = d _ {i, i} ^ {2} \sum_ {j = 0} ^ {d - 1} q _ {i, j} ^ {2} \tag {27} +$$ + +Given that the rows of $\mathbf{Q}$ are L2-normalized, we have $\sum_{j=0}^{d-1} q_{i,j}^2 = 1$ . Due to the same source and parameter sharing, it follows that $\mathbf{Q} = \mathbf{K}$ . Thus, we obtain $(\mathbf{BB}^\top)_{i,i} = d_{i,i}^2 = \frac{n+1}{n + \sum_t \mathbf{q}_{i,t} \cdot \mathbf{q}_{t,t}^\top}$ , which matches the definition of the score $s_{v,v}$ on the main diagonal. + +For the off-diagonal elements of $\mathbf{BB}^{\top}$ where $v\neq u$ , there is + +$$ +\begin{array}{l} \left(\mathbf {B B} ^ {\top}\right) _ {i, j} = \sum_ {t = 0} ^ {d - 1} b _ {i, t} \cdot b _ {j, t} = \sum_ {t = 0} ^ {d - 1} \left(d _ {i, i} \cdot q _ {i, t}\right) \left(d _ {j, j} \cdot q _ {j, t}\right) \tag {28} \\ = d _ {i, i} \cdot d _ {j, j} \sum_ {t = 0} ^ {d - 1} q _ {i, t} \cdot q _ {j, t} = c \frac {\mathbf {q} _ {i , :} \mathbf {q} _ {j , :} ^ {\top}}{n + \sum_ {k} \mathbf {q} _ {i , :} \mathbf {q} _ {k , :} ^ {\top}} \\ \end{array} +$$ + +Since $d_{i,i}$ and $d_{j,j}$ are the square roots of the denominators in the formula for $s_{i,i}$ and $s_{j,j}$ , respectively, and $\mathbf{q}_{i,:}\mathbf{q}_{j,:}^{\top}$ is the dot product of the $i$ -th and $j$ -th rows of $\mathbf{Q}$ , this matches the definition of $s_{i,j}$ . Therefore, the correct construction of $\mathbf{B}$ should be $\mathbf{B} = \mathbf{DQ}$ . + +Similarly, the closed-form solution $\mathbf{B}^*$ of the objective in Eq. 15 can be obtained by setting its derivative to 0 as + +$$ +\begin{array}{l} \frac {\partial \mathcal {O} (\mathbf {Z} , \mathbf {B})}{\partial \mathbf {B}} = 2 (\mathbf {B} - M L P (\mathbf {X})) + 2 \eta \left(\mathbf {Z} \mathbf {Z} ^ {\top} \mathbf {B}\right) = 0 (29) \\ \Rightarrow \mathbf {B} ^ {*} = M L P (\mathbf {X}) + \eta \mathbf {Z} \mathbf {Z} ^ {\top} \mathbf {B} (30) \\ \end{array} +$$ + +Then, by defining $\omega_{2} = \frac{1}{1 + \eta}$ , there is + +$$ +\mathbf {B} ^ {*} = \left(1 - \omega_ {2}\right) M L P (\mathbf {X}) + \omega_ {2} \mathbf {Z} \mathbf {Z} ^ {\top} \mathbf {B} \tag {31} +$$ + +The proposed dual cross-attention module for updating the attribute representations, i.e., Eq. 11, can be rephrased as + +$$ +\tilde {\mathbf {B}} = (1 - \tilde {\gamma}) M L P (\mathbf {X}) + \tilde {\gamma} \mathbf {S V} \tag {32} +$$ + +Similarly, considering that matrix $\mathbf{S}$ maintains the same structure as described in Eq. 26, the crucial step to ensure $\mathbf{B}^{*} = \hat{\mathbf{B}}$ is to establish the parameter $\omega_{2} = \tilde{\gamma}$ and to set $\mathbf{Z} = \mathbf{DQ}$ . This approach guarantees that the necessary conditions for the equivalence are met. + +# E Experimental Details + +# E.1 Datasets and Splitting + +Datasets. In the node classification experiments, sixteen publicly available benchmark datasets are utilized. These graphs can be classified into two categories based on whether their Edge Homophily [37] exceeds 0.5: seven graphs are tagged as homophilic graphs, including Cora [42], CiteSeer [42], PubMed [42], Photo [43], CS [43], Physics [43], and Questions [38]. The remaining seven graphs are marked as heterophilic graphs, containing Cornell [37], Texas [37], Wisconsin [37], Actor [45], Chameleon [40], Squirrel [40], and Ratings [38]. It is worth noting that the original Chameleon and Squirrel exhibit neighborhood overlap, and are thus filtered according to the study [38]. Additionally, two large-scale graph datasets, i.e., ogbn-arxiv [23] and ogbn-proteins [23], are employed for node property prediction experiment. Statistics are shown in Tab. 4. + +Table 4: Statistics of sixteen graph datasets. $\# h$ denotes the edge homophily shown in [37]. + +
DatasetNodesEdgesFeaturesClasses#h
Cora2,7085,2781,43370.81
CiteSeer3,3274,5523,70360.74
PubMed19,71744,32450030.80
Photo7,650238,16374580.83
CS18,33381,8946,805150.81
Physics34,493247,9628,41550.93
Questions48,921153,54030120.84
Cornell1832801,70350.30
Texas1832951,70350.11
Wisconsin2514661,70350.21
Actor7,60033,54493150.22
Chameleon8908,8542,32550.24
Squirrel2,22346,9982,08950.21
Ratings24,49293,05030050.38
ogbn-proteins132,53439,561,252820.38
ogbn-arxiv169,3431,157,7993001280.65
+ +Dataset Splitting. To ensure that the experimental results are credible and reproducible, this paper follows well-established dataset splitting strategies. For the Cora, CiteSeer, and PubMed, the public standard splitting described in [29] is adopted, with 20 nodes per class for training, 500 for validation, and 1000 for testing. The Photo, CS, and Physics are randomly divided into training, validation, and testing sets in a $60\%$ , $20\%$ , and $20\%$ ratio, respectively. For the heterophilic datasets Cornell, Texas, Wisconsin, Actor, and Chameleon, this paper employs 10 standard train/validation/test splits with a division ratio of $48\%$ , $32\%$ , and $20\%$ , respectively. Note that the Chameleon and Squirrel used here are duplicates-removed filtered versions as referenced in [38]. The Ratings, and Questions follow a $50\%/25\%/25\%$ train/Validation/test random split pattern. For the two datasets from the OGB [23], i.e., ogbn-arxiv and ogbn-proteins, the provided standard splits are utilized. + +# E.2 Introduction of Baselines + +The comparative analysis in the experiments involves seven Graph Neural Networks (GNNs) as well as seven Graph Transformers (GTs) as the baseline models. To be specific, the GNNs include four standard GNNs, i.e., GCN [29], GAT [47], GraphSAGE [20], and APPNP [30], and two universal GNNs for graphs with diverse homophily, i.e., GPR-GNN [6] and GloGNN [33], and a non-message-passing GNN with a separate topology and attribute design, i.e., LINKX [34]. Besides, the baseline GTs encompass eight state-of-the-art models, namely, GraphGPS [39], NAGphormer [2], Exphormer [44], GOAT [31], NodeFormer [49], SGFormer [50], Polynormer [7], and Gradformer [36]. These models are implemented following the released code of the original paper. + +# E.2.1 Graph Neural Networks (GNNs) + +The following specifies the GNN baselines employed in our comparative analysis. + +- GCN [29]: A seminal GNN that integrates graph topology and node attributes via graph convolution. +- GAT [47]: A classic graph attention network that weights propagation using an attention mechanism. +- GraphSAGE [20]: A scalable variant of GCN that employs neighbor sampling and diverse aggregation strategies. +- APPNP [30]: A variant of GCN that weights propagation based on personalized PageRank. +- GPR-GNN [6]: A universal variant of GCN that weights propagation using learnable layer coefficients. +- LINKX [34]: A non-message-passing GNN that directly combines representations of graph topology and node attributes. +- GloGNN [33]: A universal variant of GCN that obtains the propagation matrix from an optimization objective describing node relationships. + +# E.2.2 Graph Transformers (GTs) + +The following specifies the GT baselines utilized in our comparative analysis. + +- GraphGPS [39]: A general GT architecture incorporating positional encodings and local/global modules. +- NodeFormer [49]: A scalable GT architecture that learns layer-specific graph structures via a kernelized Gumbel-Softmax operator. +- NAGphormer [2]: A creative GT architecture constructing token vectors using neighborhood aggregation. +- Exphormer [44]: A general GT architecture combining local, extended, and virtual node-based global attention. +- GOAT [31]: A comprehensive GT architecture linearizing computational complexity based on the k-means algorithm. +- SGFormer [50]: A lightweight GT architecture featuring a single-layer self-attention module. +- Polynormer [7]: A polynomial-expressive GT architecture learning high-degree polynomials on input features. +- Gradformer [36]: An effective GT architecture dynamically modeling node relationships by exponentially diminishing values in the decay mask matrix. + +For the four GNN baselines, including GCN, GAT, GraphSAGE, and APPNP, we utilize the public library, PyTorch Geometric (PyG) [13], for their implementation. For the other three GNN baselines, we utilize their original code. The sources are outlined as + +GPR-GNN: https://github.com/jianhao2016/GPRGNN +LINKX: https://github.com/CUAI/Non-Homophily-Large-Scale +GloGNN: https://github.com/RecklessRonan/GloGNN + +For the GT baselines, including GraphGPS, NodeFormer, NAGphormer, Exphormer, GOAT, SG-Former, Polynormer, and Gradformer, we utilize their source code. The sources are detailed as + +GraphGPS: https://github.com/rampasek/GraphGPS +- NodeFormer: https://github.com/qitianwu/NodeFormer +- NAGphormer: https://github.com/JHL-HUST/NAGphormer +- Exphormer: https://github.com/hamed1375/Exphormer +- GOAT: https://github.com/devnkong/GOAT +- SGFormer: https://github.com/qitianwu/SGFormer +Polynormer: https://github.com/cornell-zhang/Polynormer +- Gradformer: https://github.com/LiuChuang0059/Gradformer + +# E.3 Experimental Setups + +Configurations. The experiment is performed on two Linux machines using a single GeForce RTX4090 24 GB GPU and a single NVIDIA A800 80GB GPU, respectively. The reported results are averaged over ten random trials. All models operate under a semi-supervised learning paradigm, where the results on validation sets are referenced to fine-tune hyperparameters. + +Table 5: Hyperparameters of UGCFormer per dataset. + +
Dataset# layers l# dimensions dlrαβwd
Cora42560.0010.515e-3
CiteSeer32560.0010.60.15e-3
PubMed31280.0010.50.11e-2
Photo45120.0010.30.15e-5
CS35120.0010.60.15e-4
Physics35120.0010.515e-4
Questions22560.0050.30.15e-4
Cornell52560.0010.90.0011e-2
Texas41280.0010.70.0011e-2
Wisconsin41280.0010.90.0011e-2
Actor55120.0010.90.011e-2
Chameleon25120.0010.20.015e-3
Squirrel45120.0010.20.015e-5
Ratings2640.0010.30.0010
ogbn-proteins11280.0010.70.015e-4
ogbn-arxiv25120.010.30.015e-4
+ +Hyper-parameters. The hyperparameters are selected via a grid search strategy. In the node classification task, models are trained employing an Adam optimizer with the learning rate among $\{0.001, 0.005, 0.01\}$ and the weight decay among $\{0, 1e - 5, 5e - 5, 1e - 4, 5e - 4, 1e - 3, 5e - 3, 1e - 2\}$ . The number of layers is selected from $\{1, 2, 3, 4, 5\}$ , and the dimension of hidden layers is chosen from $\{64, 128, 256, 512\}$ , and their impacts on model performance are analyzed in Section 4.2. For the node property prediction task, the hyperparameter selection follows the baseline [50]. For the unique hyperparameters in UGCFoermer, $\alpha$ is chosen from a range starting at 0.1 and increasing by increments of 0.1, up to 0.9, $\beta$ is selected from $\{0.001, 0.01, 0.1, 1\}$ , and $\tau$ is fixed to 0.5. Refer to Tab. 5 for the chosen parameters that correspond to the reported results. + +Table 6: Training time and GPU memory usage on three graphs. + +
MethodCiteSeerPubMedogbn-arxiv
Train/Epoch (ms)Mem. (MB)Train/Epoch (ms)Mem. (MB)Train/Epoch (ms)Mem. (MB)
GraphGPS16.8214046.99470166.108,102
NodeFormer10.2011011.1421884.002,066
NAGphormer10.8116617.65352760.501,962
Exphormer14.2015925.00696145.706,758
SGFormer5.80846.0714236.901,024
Polynormer9.2017415.20307170.074,729
UGCFourier8.8010610.6024669.602,050
+ +# E.4 Additional experiment results + +Running Time and Space Consumption. To further illustrate the efficiency and scalability of the proposed UGCFoer, this experiment compares it with other GTs in terms of runtime and GPU memory usage. Common hyperparameters are uniformly applied across all models to highlight the impact of their components, particularly the attention modules. As depicted in Table 6, UGCFoer consistently has the second-lowest running time and ranks among the top three in terms of lowest GPU memory usage across three datasets. Despite utilizing linearized attention mechanisms, most linearized GTs, including GraphGPS, NodeFormer, Exphormer, and Polynormer, perform worse than UGCFoer. This highlights the lightweight and efficient design of UGCFoer. + +![](images/834001823a00863d1aa4190ae92ae86f30a9c3238bae44df25ad3b173ff538e4.jpg) +Figure 6: Performance variations for varying $\alpha$ . + +![](images/1098bc49beb2d768c46c4a710357aef7e347d91c43d695ed0ff9b52c29771191.jpg) +Figure 7: Performance variations for varying $\beta$ . + +![](images/92591d22f28696fa42f70e27d0b30d47c8ba7c388ee181d4fd93b069b45ca965.jpg) + +![](images/e49f4811b575197d526d077eeb4d98999fcc03b05c5a37208206b94add553230.jpg) + +Hyperparameter $\alpha$ . As depicted in the Fig. 6, UGCFormaler exhibits stable performance within a specific parameter range for each dataset. For instance, on Cora, performance variation is minimal within the set $\{0.3, 0.4, 0.5, 0.6\}$ , indicating that the model is robust to changes in hyperparameter $\alpha$ . Similar observations are made for heterophilic graphs. + +Hyperparameter $\beta$ . From the Fig. 7, it can be observed that the performance remains stable across the selection range of $\beta$ , demonstrating that the model is not sensitive to this parameter. Even in the worst case of parameter selection, the model achieves performance that is comparable to the baseline. + +# F Discussion + +Comparison with Edge-augmented Graph Transformer (EGT). The proposed UGCFoer shares a conceptual connection with EGT [27], yet their core mechanisms differ substantially. EGT augments pairwise attribute-based attention between nodes using graph topology, whereas UGCFoer captures the interaction between topology and attributes through a cross-attention mechanism. Although EGT introduces an additional edge channel alongside the attribute channel, this design primarily aims to utilize edge features to modulate the attention process (via addition or gating) rather than to explicitly model the interaction between the two channels. Moreover, EGT requires large-scale edge features of size $O(n^{2}d)$ (where $n$ denotes the number of nodes and $d$ the feature dimension), which significantly increases computational complexity and limits scalability. In contrast, UGCFoer adopts a linear cross-attention module that efficiently models topology-attribute interactions with linear time and space complexity. + +Broader Relation to Other Categories of Graph Transformers. The proposed cross-aggregation mechanism can also provide a unified interpretation for other two types of Graph Transformers, that is, those based on context-node sampling or edge rewriting. Although both categories of GTs are implemented in different ways, their essence is to determine the subgraph for each node to perform local message passing (via graph convolution or self-attention). Thus, they can be uniformly expressed by obtaining the graph structure. As a result, the topology representation can be straightforwardly determined based on the eigenvalue decomposition of the adjacency matrix, as described in the manuscript. Ultimately, a formulation similar to that of cross-aggregation can be derived. Therefore, their underlying mechanism can be attributed to a cross-aggregation between topology and attribute representations. + +# G Limitations + +The proposed UGCFoermer, like other Graph Transformers (GTs), relies on a fixed set of hyperparameters, requiring significant tuning and prior knowledge to optimize results. This not only limits their applicability but also increases computational costs. Future work should explore adaptive learning mechanisms to automate hyperparameter adjustment based on input data characteristics, reducing manual intervention and enhancing generalizability. + +# NeurIPS Paper Checklist + +# 1. Claims + +Question: Do the main claims made in the abstract and introduction accurately reflect the paper's contributions and scope? + +Answer: [Yes] + +Justification: The abstract and introduction clearly outline the contributions of our paper, including the motivation and design of the proposed UGCFoermer. + +Guidelines: + +- The answer NA means that the abstract and introduction do not include the claims made in the paper. +- The abstract and/or introduction should clearly state the claims made, including the contributions made in the paper and important assumptions and limitations. A No or NA answer to this question will not be perceived well by the reviewers. +- The claims made should match theoretical and experimental results, and reflect how much the results can be expected to generalize to other settings. +- It is fine to include aspirational goals as motivation as long as it is clear that these goals are not attained by the paper. + +# 2. Limitations + +Question: Does the paper discuss the limitations of the work performed by the authors? + +Answer: [Yes] + +Justification: We have discussed the limitations in Section G, particularly regarding the large number of hyperparameters commonly found in GTs. We have outlined potential directions for future research to address this concern. + +Guidelines: + +- The answer NA means that the paper has no limitation while the answer No means that the paper has limitations, but those are not discussed in the paper. +- The authors are encouraged to create a separate "Limitations" section in their paper. +- The paper should point out any strong assumptions and how robust the results are to violations of these assumptions (e.g., independence assumptions, noiseless settings, model well-specification, asymptotic approximations only holding locally). The authors should reflect on how these assumptions might be violated in practice and what the implications would be. +- The authors should reflect on the scope of the claims made, e.g., if the approach was only tested on a few datasets or with a few runs. In general, empirical results often depend on implicit assumptions, which should be articulated. +- The authors should reflect on the factors that influence the performance of the approach. For example, a facial recognition algorithm may perform poorly when image resolution is low or images are taken in low lighting. Or a speech-to-text system might not be used reliably to provide closed captions for online lectures because it fails to handle technical jargon. +- The authors should discuss the computational efficiency of the proposed algorithms and how they scale with dataset size. +- If applicable, the authors should discuss possible limitations of their approach to address problems of privacy and fairness. +- While the authors might fear that complete honesty about limitations might be used by reviewers as grounds for rejection, a worse outcome might be that reviewers discover limitations that aren't acknowledged in the paper. The authors should use their best judgment and recognize that individual actions in favor of transparency play an important role in developing norms that preserve the integrity of the community. Reviewers will be specifically instructed to not penalize honesty concerning limitations. + +# 3. Theory assumptions and proofs + +Question: For each theoretical result, does the paper provide the full set of assumptions and a complete (and correct) proof? + +# Answer: [Yes] + +Justification: All theoretical results are accompanied by clearly stated assumptions and complete proofs, provided in the main paper and referenced appropriately. + +# Guidelines: + +- The answer NA means that the paper does not include theoretical results. +- All the theorems, formulas, and proofs in the paper should be numbered and cross-referenced. +- All assumptions should be clearly stated or referenced in the statement of any theorems. +- The proofs can either appear in the main paper or the supplemental material, but if they appear in the supplemental material, the authors are encouraged to provide a short proof sketch to provide intuition. +- Inversely, any informal proof provided in the core of the paper should be complemented by formal proofs provided in appendix or supplemental material. +- Theorems and Lemmas that the proof relies upon should be properly referenced. + +# 4. Experimental result reproducibility + +Question: Does the paper fully disclose all the information needed to reproduce the main experimental results of the paper to the extent that it affects the main claims and/or conclusions of the paper (regardless of whether the code and data are provided or not)? + +# Answer: [Yes] + +Justification: The supplemental material contains a file of our model's code, enabling the replication of the experiments. + +# Guidelines: + +- The answer NA means that the paper does not include experiments. +- If the paper includes experiments, a No answer to this question will not be perceived well by the reviewers: Making the paper reproducible is important, regardless of whether the code and data are provided or not. +- If the contribution is a dataset and/or model, the authors should describe the steps taken to make their results reproducible or verifiable. +- Depending on the contribution, reproducibility can be accomplished in various ways. For example, if the contribution is a novel architecture, describing the architecture fully might suffice, or if the contribution is a specific model and empirical evaluation, it may be necessary to either make it possible for others to replicate the model with the same dataset, or provide access to the model. In general, releasing code and data is often one good way to accomplish this, but reproducibility can also be provided via detailed instructions for how to replicate the results, access to a hosted model (e.g., in the case of a large language model), releasing of a model checkpoint, or other means that are appropriate to the research performed. +- While NeurIPS does not require releasing code, the conference does require all submissions to provide some reasonable avenue for reproducibility, which may depend on the nature of the contribution. For example + +(a) If the contribution is primarily a new algorithm, the paper should make it clear how to reproduce that algorithm. +(b) If the contribution is primarily a new model architecture, the paper should describe the architecture clearly and fully. +(c) If the contribution is a new model (e.g., a large language model), then there should either be a way to access this model for reproducing the results or a way to reproduce the model (e.g., with an open-source dataset or instructions for how to construct the dataset). +(d) We recognize that reproducibility may be tricky in some cases, in which case authors are welcome to describe the particular way they provide for reproducibility. In the case of closed-source models, it may be that access to the model is limited in some way (e.g., to registered users), but it should be possible for other researchers to have some path to reproducing or verifying the results. + +# 5. Open access to data and code + +Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material? + +Answer: [Yes] + +Justification: We have included complete and executable code within the supplemental material, ensuring the reproducibility of our results. + +Guidelines: + +- The answer NA means that paper does not include experiments requiring code. +- Please see the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details. +- While we encourage the release of code and data, we understand that this might not be possible, so "No" is an acceptable answer. Papers cannot be rejected simply for not including code, unless this is central to the contribution (e.g., for a new open-source benchmark). +- The instructions should contain the exact command and environment needed to run to reproduce the results. See the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details. +- The authors should provide instructions on data access and preparation, including how to access the raw data, preprocessed data, intermediate data, and generated data, etc. +- The authors should provide scripts to reproduce all experimental results for the new proposed method and baselines. If only a subset of experiments are reproducible, they should state which ones are omitted from the script and why. +- At submission time, to preserve anonymity, the authors should release anonymized versions (if applicable). +- Providing as much information as possible in supplemental material (appended to the paper) is recommended, but including URLs to data and code is permitted. + +# 6. Experimental setting/details + +Question: Does the paper specify all the training and test details (e.g., data splits, hyperparameters, how they were chosen, type of optimizer, etc.) necessary to understand the results? + +Answer: [Yes] + +Justification: We have provided detailed descriptions of our experimental setup in Section E, including data splits, hyperparameters, and optimizer, etc. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The experimental setting should be presented in the core of the paper to a level of detail that is necessary to appreciate the results and make sense of them. +- The full details can be provided either with the code, in appendix, or as supplemental material. + +# 7. Experiment statistical significance + +Question: Does the paper report error bars suitably and correctly defined or other appropriate information about the statistical significance of the experiments? + +Answer: [Yes] + +Justification: The experimental results are presented as the mean and standard deviation over 10 runs, as shown in Tabs. 1 and 2, as well as Figs. 4 and 5. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The authors should answer "Yes" if the results are accompanied by error bars, confidence intervals, or statistical significance tests, at least for the experiments that support the main claims of the paper. + +- The factors of variability that the error bars are capturing should be clearly stated (for example, train/test split, initialization, random drawing of some parameter, or overall run with given experimental conditions). +- The method for calculating the error bars should be explained (closed form formula call to a library function, bootstrap, etc.) +- The assumptions made should be given (e.g., Normally distributed errors). +- It should be clear whether the error bar is the standard deviation or the standard error of the mean. +- It is OK to report 1-sigma error bars, but one should state it. The authors should preferably report a 2-sigma error bar than state that they have a $96\%$ CI, if the hypothesis of Normality of errors is not verified. +- For asymmetric distributions, the authors should be careful not to show in tables or figures symmetric error bars that would yield results that are out of range (e.g. negative error rates). +- If error bars are reported in tables or plots, The authors should explain in the text how they were calculated and reference the corresponding figures or tables in the text. + +# 8. Experiments compute resources + +Question: For each experiment, does the paper provide sufficient information on the computer resources (type of compute workers, memory, time of execution) needed to reproduce the experiments? + +Answer: [Yes] + +Justification: We have provided the computational resources used for all experiments in Section E.3. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The paper should indicate the type of compute workers CPU or GPU, internal cluster or cloud provider, including relevant memory and storage. +- The paper should provide the amount of compute required for each of the individual experimental runs as well as estimate the total compute. +- The paper should disclose whether the full research project required more computer than the experiments reported in the paper (e.g., preliminary or failed experiments that didn't make it into the paper). + +# 9. Code of ethics + +Question: Does the research conducted in the paper conform, in every respect, with the NeurIPS Code of Ethics https://neurips.cc/public/EthicsGuidelines? + +Answer: [Yes] + +Justification: Our research adheres to the NeurIPS Code of Ethics, and we have ensured that all aspects of our work. + +Guidelines: + +The answer NA means that the authors have not reviewed the NeurIPS Code of Ethics +- If the authors answer No, they should explain the special circumstances that require a deviation from the Code of Ethics. +- The authors should make sure to preserve anonymity (e.g., if there is a special consideration due to laws or regulations in their jurisdiction). + +# 10. Broader impacts + +Question: Does the paper discuss both potential positive societal impacts and negative societal impacts of the work performed? + +Answer: [NA] + +Justification: Due to the nature of this work, there may be no potential negative social impact that is easily predictable. + +Guidelines: + +- The answer NA means that there is no societal impact of the work performed. +- If the authors answer NA or No, they should explain why their work has no societal impact or why the paper does not address societal impact. +- Examples of negative societal impacts include potential malicious or unintended uses (e.g., disinformation, generating fake profiles, surveillance), fairness considerations (e.g., deployment of technologies that could make decisions that unfairly impact specific groups), privacy considerations, and security considerations. +- The conference expects that many papers will be foundational research and not tied to particular applications, let alone deployments. However, if there is a direct path to any negative applications, the authors should point it out. For example, it is legitimate to point out that an improvement in the quality of generative models could be used to generate deepfakes for disinformation. On the other hand, it is not needed to point out that a generic algorithm for optimizing neural networks could enable people to train models that generate Deepfakes faster. +- The authors should consider possible harms that could arise when the technology is being used as intended and functioning correctly, harms that could arise when the technology is being used as intended but gives incorrect results, and harms following from (intentional or unintentional) misuse of the technology. +- If there are negative societal impacts, the authors could also discuss possible mitigation strategies (e.g., gated release of models, providing defenses in addition to attacks, mechanisms for monitoring misuse, mechanisms to monitor how a system learns from feedback over time, improving the efficiency and accessibility of ML). + +# 11. Safeguards + +Question: Does the paper describe safeguards that have been put in place for responsible release of data or models that have a high risk for misuse (e.g., pretrained language models, image generators, or scraped datasets)? + +Answer: [NA] + +Justification: Our work does not involve releasing data or models that pose a high risk for misuse, so no specific safeguards are necessary. + +Guidelines: + +- The answer NA means that the paper poses no such risks. +- Released models that have a high risk for misuse or dual-use should be released with necessary safeguards to allow for controlled use of the model, for example by requiring that users adhere to usage guidelines or restrictions to access the model or implementing safety filters. +- Datasets that have been scraped from the Internet could pose safety risks. The authors should describe how they avoided releasing unsafe images. +- We recognize that providing effective safeguards is challenging, and many papers do not require this, but we encourage authors to take this into account and make a best faith effort. + +# 12. Licenses for existing assets + +Question: Are the creators or original owners of assets (e.g., code, data, models), used in the paper, properly credited and are the license and terms of use explicitly mentioned and properly respected? + +Answer: [Yes] + +Justification: We have accurately credited the sources and provided URLs in Section E.2. + +Guidelines: + +- The answer NA means that the paper does not use existing assets. +- The authors should cite the original paper that produced the code package or dataset. +- The authors should state which version of the asset is used and, if possible, include a URL. +- The name of the license (e.g., CC-BY 4.0) should be included for each asset. + +- For scraped data from a particular source (e.g., website), the copyright and terms of service of that source should be provided. +- If assets are released, the license, copyright information, and terms of use in the package should be provided. For popular datasets, paperswithcode.com/datasets has curated licenses for some datasets. Their licensing guide can help determine the license of a dataset. +- For existing datasets that are re-packaged, both the original license and the license of the derived asset (if it has changed) should be provided. +- If this information is not available online, the authors are encouraged to reach out to the asset's creators. + +# 13. New assets + +Question: Are new assets introduced in the paper well documented and is the documentation provided alongside the assets? + +Answer: [Yes] + +Justification: The supplemental material includes the file of our model's code + +Guidelines: + +- The answer NA means that the paper does not release new assets. +- Researchers should communicate the details of the dataset/code/model as part of their submissions via structured templates. This includes details about training, license, limitations, etc. +- The paper should discuss whether and how consent was obtained from people whose asset is used. +- At submission time, remember to anonymize your assets (if applicable). You can either create an anonymized URL or include an anonymized zip file. + +# 14. Crowdsourcing and research with human subjects + +Question: For crowdsourcing experiments and research with human subjects, does the paper include the full text of instructions given to participants and screenshots, if applicable, as well as details about compensation (if any)? + +Answer: [NA] + +Justification: This paper does not involve crowdsourcing or research with human subjects, so this information is not applicable. + +Guidelines: + +- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. +- Including this information in the supplemental material is fine, but if the main contribution of the paper involves human subjects, then as much detail as possible should be included in the main paper. +- According to the NeurIPS Code of Ethics, workers involved in data collection, curation, or other labor should be paid at least the minimum wage in the country of the data collector. + +# 15. Institutional review board (IRB) approvals or equivalent for research with human subjects + +Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals (or an equivalent approval/review based on the requirements of your country or institution) were obtained? + +Answer: [NA] + +Justification: This research does not involve human subjects, so IRB approvals or equivalent reviews are not required. + +Guidelines: + +- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. + +- We recognize that the procedures for this may vary significantly between institutions and locations, and we expect authors to adhere to the NeurIPS Code of Ethics and the guidelines for their institution. +- For initial submissions, do not include any information that would break anonymity (if applicable), such as the institution conducting the review. + +# 16. Declaration of LLM usage + +Question: Does the paper describe the usage of LLMs if it is an important, original, or non-standard component of the core methods in this research? Note that if the LLM is used only for writing, editing, or formatting purposes and does not impact the core methodology, scientific rigorousness, or originality of the research, declaration is not required. + +Answer: [NA] + +Justification: The core method development in this research does not involve LLMs. + +Guidelines: + +- The answer NA means that the core method development in this research does not involve LLMs as any important, original, or non-standard components. +- Please refer to our LLM policy (https://neurips.cc/Conferences/2025/LLM) for what should or should not be described. \ No newline at end of file diff --git a/NeurIPS/2025/A Closer Look at Graph Transformers_ Cross-Aggregation and Beyond/images.zip b/NeurIPS/2025/A Closer Look at Graph Transformers_ Cross-Aggregation and Beyond/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..3bbc9ddc5d254747c47363398034f5b99a4c3339 --- /dev/null +++ b/NeurIPS/2025/A Closer Look at Graph Transformers_ Cross-Aggregation and Beyond/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b1a26190bf7f074be4a59e2896412296055e4c1f24a8a91945dcf9dae752d690 +size 904684 diff --git a/NeurIPS/2025/A Closer Look at Graph Transformers_ Cross-Aggregation and Beyond/layout.json b/NeurIPS/2025/A Closer Look at Graph Transformers_ Cross-Aggregation and Beyond/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..93083de5e88160f40577de82808b6d74b1ba61cf --- /dev/null +++ b/NeurIPS/2025/A Closer Look at Graph Transformers_ Cross-Aggregation and Beyond/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ece6cb799d1944ef4f7565085cc61b4e60bc8429e97972a1dbdc6df8e58de04 +size 1009483 diff --git a/NeurIPS/2025/A Closer Look at Model Collapse_ From a Generalization-to-Memorization Perspective/e204cff1-69c4-40b2-ab88-047ab025f74c_content_list.json b/NeurIPS/2025/A Closer Look at Model Collapse_ From a Generalization-to-Memorization Perspective/e204cff1-69c4-40b2-ab88-047ab025f74c_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..59423223a040e7c20d52c049c0a211f1ff4437f7 --- /dev/null +++ b/NeurIPS/2025/A Closer Look at Model Collapse_ From a Generalization-to-Memorization Perspective/e204cff1-69c4-40b2-ab88-047ab025f74c_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f07833e33266ff748707ba7f4c2b2532d1578b4b79840f5919d2d69e2dd0ab4c +size 176387 diff --git a/NeurIPS/2025/A Closer Look at Model Collapse_ From a Generalization-to-Memorization Perspective/e204cff1-69c4-40b2-ab88-047ab025f74c_model.json b/NeurIPS/2025/A Closer Look at Model Collapse_ From a Generalization-to-Memorization Perspective/e204cff1-69c4-40b2-ab88-047ab025f74c_model.json new file mode 100644 index 0000000000000000000000000000000000000000..3135cf3ff8d2f33e77cee6b74d023bccd4eeeba3 --- /dev/null +++ b/NeurIPS/2025/A Closer Look at Model Collapse_ From a Generalization-to-Memorization Perspective/e204cff1-69c4-40b2-ab88-047ab025f74c_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a64c28f2885b21f3d025595eb574c45f2b90fcc2045c46a2f59fb1f2481d0e59 +size 233663 diff --git a/NeurIPS/2025/A Closer Look at Model Collapse_ From a Generalization-to-Memorization Perspective/e204cff1-69c4-40b2-ab88-047ab025f74c_origin.pdf b/NeurIPS/2025/A Closer Look at Model Collapse_ From a Generalization-to-Memorization Perspective/e204cff1-69c4-40b2-ab88-047ab025f74c_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..ae0f82b76593e06eeb2dc22608b91313d2b75b22 --- /dev/null +++ b/NeurIPS/2025/A Closer Look at Model Collapse_ From a Generalization-to-Memorization Perspective/e204cff1-69c4-40b2-ab88-047ab025f74c_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d236179e115232a2db7e0ff66e18190647db34c0c906028d7d1fa5585de2ec60 +size 20132272 diff --git a/NeurIPS/2025/A Closer Look at Model Collapse_ From a Generalization-to-Memorization Perspective/full.md b/NeurIPS/2025/A Closer Look at Model Collapse_ From a Generalization-to-Memorization Perspective/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e0731ebc05a91c07175a6d13788509a57bdfa790 --- /dev/null +++ b/NeurIPS/2025/A Closer Look at Model Collapse_ From a Generalization-to-Memorization Perspective/full.md @@ -0,0 +1,880 @@ +# A Closer Look at Model Collapse: From a Generalization-to-Memorization Perspective + +Lianghe Shi* +University of Michigan +United States +lhshi@umich.edu + +Meng Wu* +University of Michigan +United States +wymond@umich.edu + +Huijie Zhang +University of Michigan +United States +huijiezh@umich.edu + +Zekai Zhang +University of Michigan +United States +zzekai@umich.edu + +Molei Tao +Georgia Institute of Technology +United States +mtao@gatech.edu + +Qing Qu +University of Michigan +United States +qingqu@umich.edu + +# Abstract + +The widespread use of diffusion models has led to an abundance of AI-generated data, raising concerns about model collapse—a phenomenon in which recursive iterations of training on synthetic data lead to performance degradation. Prior work primarily characterizes this collapse via variance shrinkage or distribution shift, but these perspectives miss practical manifestations of model collapse. This paper identifies a transition from generalization to memorization during model collapse in diffusion models, where models increasingly replicate training data instead of generating novel content during iterative training on synthetic samples. This transition is directly driven by the declining entropy of the synthetic training data produced in each training cycle, which serves as a clear indicator of model degradation. Motivated by this insight, we propose an entropy-based data selection strategy to mitigate the transition from generalization to memorization and alleviate model collapse. Empirical results show that our approach significantly enhances visual quality and diversity in recursive generation, effectively preventing collapse. The source code is available at https://github.com/shilianghe007/Model_Collapse.git + +# 1 Introduction + +As generative models, such as diffusion models, become widely used for image synthesis and video generation, a large volume of generated data has emerged on the Internet. Since state-of-the-art diffusion models can generate realistic content that even humans cannot easily distinguish, the training datasets for next-generation models will inevitably contain a significant proportion of synthetic data. Figure 1 illustrates this self-consuming loop, where at each iteration $^2$ , data generated by the current model is subsequently used to train the new model for the next iteration. Unfortunately, several recent studies have demonstrated that recursively training models on datasets contaminated by AI-generated data leads to performance degradation across these self-consuming iterations—even when synthetic data comprises only a small fraction of the dataset [1]. This phenomenon, termed model collapse in prior work, poses a significant threat to the future development and effectiveness of generative models. + +![](images/c8c7f1f5ab22ee584cafb8b3b2f3f20d7e61a5abbc9426ad93276bf3efe321e8.jpg) +Figure 1: High-level depiction of the self-consuming pipeline. Top: Collapse iteration represents the replace paradigm where models are trained solely on synthetic images generated by the previous diffusion model. Middle: In the mitigated iteration, original real data and previously generated data are added to train the next-generation model. Our proposed selection methods construct a training subset and can further mitigate collapse. Bottom Right: Evolution of the generated images. + +As surveyed by [2], recent studies have identified various collapse behaviors that depend on the performance metrics employed. A series of papers [3-6] reveal the model collapse phenomenon through the variance of the learned distribution. They empirically and theoretically show that the model continually loses information in the distribution tail, with variance tending towards 0. Another line of work [5, 7-10] investigates the issue from the perspective of population risk or distribution shifts. These studies observe that the generated distribution progressively deviates from the underlying distribution, causing the model's population risk to increase throughout the recursive process. Numerous studies [11-13] also report that models begin generating hallucinated data. Despite significant theoretical insights regarding variance dynamics, the reduction of variance to negligible levels typically occurs only after an extremely large number of iterations. As noted by [2, 4], the collapse progresses at such a slow pace that it is rarely a practical concern in real-world applications. In contrast, the visual quality and diversity of generated images deteriorate rapidly. Furthermore, although population risk or distribution shifts offer a holistic view of performance degradation, they do not adequately characterize specific collapse behaviors. + +Accordingly, this paper conducts an in-depth investigation into the collapse dynamics of diffusion models and identifies a generalization-to-memorization transition occurring across successive iterations. Specifically, during early iterations, the model demonstrates a strong capability to generate novel images distinct from those in the training set, but gradually shifts towards memorization in later iterations, merely replicating training images. This transition significantly reduces the diversity of generated content and results in higher FID scores. Moreover, directly reproducing images from training datasets may raise copyright concerns [14, 15]. Furthermore, we empirically reveal a strong linear correlation between the generalizability of the trained model and the entropy of its training dataset. As iterations progress, the entropy of the data distribution sharply decreases, directly resulting in a decline in the models generalizability, which illustrates a clear transition from generalization to memorization. Motivated by these empirical findings, we propose entropy-based selection methods to construct training subsets from candidate pools. Extensive experimental validation demonstrates that our proposed methods effectively identify high-entropy subsets, thereby decelerating the generalization-to-memorization transition. Additionally, our approach achieves superior image quality and lower FID scores in recursive training loops compared to baseline methods. + +Our Contributions. The contributions of this paper are summarized as follows: + +1. We identify the generalization-to-memorization transition within the self-consuming loop, providing a novel perspective for studying model collapse and highlighting critical practical issues arising from training on synthetic data. + +2. We investigate the key factor driving this transition and empirically demonstrate a strong correlation between the entropy of the training dataset and the generalization capability of the model. +3. Motivated by our empirical findings, we propose entropy-based data selection methods, whose effectiveness is validated through comprehensive experiments across various image datasets. + +# 2 Background + +In this work, we focus on the image generation task. Let $\mathcal{X}$ be the $d$ -dimensional image space, $\mathcal{X} \subseteq \mathbb{R}^d$ and let $P_0$ be a data distribution over the space $\mathcal{X}$ . We use bold letters to denote vectors in $\mathcal{X}$ . We assume the original training data $\mathcal{D}_{\mathrm{real}} = \{\pmb{x}_{\mathrm{real}}^{(1)}, \dots, \pmb{x}_{\mathrm{real}}^{(N_0)}\}$ are generated independently and identically distributed (i.i.d.) according to the underlying distribution $P_0$ , i.e., $\pmb{x}_{\mathrm{real}}^{(i)} \sim P_0$ . + +Diffusion Models. For a given data distribution, diffusion models do not directly learn the probability density function (pdf) of the distribution; instead, they define a forward process and a reverse process, and learn the score function utilized in the reverse process. Specifically, the forward process [16] progressively adds Gaussian noise to the image, and the conditional distribution of the noisy image is given by: $p(\pmb{x}_t|\pmb{x}_0) = \mathcal{N}(\pmb{x}_t;\sqrt{\bar{\alpha}_t}\pmb{x}_0,(1 - \bar{\alpha}_t)\pmb{I})$ , where $\bar{\alpha}_t$ is the scale schedule, $\pmb{x}_0$ is the clean image drawn from $P_0$ , and $\pmb{x}_t$ is the noisy image. This forward can also be described as a stochastic differential equation (SDE) [17]: $dx = f(\pmb{x},t)dt + g(t)d\pmb{w}$ , where $f(\cdot ,t):\mathbb{R}^d\to \mathbb{R}^d$ denotes the vector-valued drift coefficient, $g(t)\in \mathbb{R}$ is the diffusion coefficient, and $w$ is a standard Brownian motion. This SDE has a corresponding reverse SDE as $dx = [f(\pmb {x},t) - g^{2}(t)\nabla_{\pmb{x}}\log p_{t}(\pmb {x})]dt + g(t)d\pmb{w}$ , where $dt$ represents a negative infinitesimal time step, driving the process from $t = T$ to $t = 0$ . The reverse SDE enables us to gradually convert a Gaussian noise to a clean image $\pmb{x}\sim P_0$ . + +The score function $\nabla_{\pmb{x}}\log p_t$ is typically unknown and needs to be estimated using a neural network $s_\theta (\pmb {x},t)$ . The training objective can be formalized as + +$$ +\mathbb {E} _ {t \sim \mathcal {U} (0, T)} \mathbb {E} _ {p _ {t} (\boldsymbol {x})} \left[ \lambda (t) \| \nabla_ {\boldsymbol {x}} \log p _ {t} (\boldsymbol {x}) - \boldsymbol {s} _ {\theta} (\boldsymbol {x}, t) \| _ {2} ^ {2} \right], +$$ + +and can be efficiently optimized with score matching methods such as denoising score matching [18]. + +Self-Consuming Loop. Following the standard setting of model collapse [1, 4, 7, 11, 19-21], we denote the training dataset at the $n$ -th iteration as $\mathcal{D}_n$ . Let $\mathcal{A}(\cdot)$ denote the training algorithm that takes $\mathcal{D}_n$ as input and outputs a diffusion model characterized by the distribution $P_n$ , i.e., $P_n = \mathcal{A}(\mathcal{D}_n)$ . In this work, we train the diffusion model from scratch at each iteration. Subsequently, the diffusion model generates a synthetic dataset of size $N_n$ , denoted by $\mathcal{G}_n \sim P_n^{N_n}$ , which is used in subsequent iterations. + +Based on the specific way of constructing training datasets, previous studies [4, 11, 22] distinguish two distinct iterative paradigms: + +- The replaced training dataset. At each iteration, the training dataset consists solely of synthetic data generated by the previous diffusion model, i.e., $\mathcal{D}_n = \mathcal{G}_{n - 1}$ . Several studies [4, 8, 22] refer to this as the "replace" paradigm and have demonstrated that under this setting, the variance collapses to 0 or the population risk diverges to infinity. +- The accumulated training dataset. A more realistic paradigm [8, 11] is to maintain access to all previous data, thereby including both real images and all synthetic images generated thus far, i.e., $\mathcal{D}_n = (\cup_{j=1}^{n-1}\mathcal{G}_j) \cup \mathcal{D}_{\mathrm{real}}$ . However, continuously increasing the training dataset size quickly demands substantial computational resources. A practical compromise is to subsample a fixed-size subset from all candidate images, referred to as the "accumulate-subsample" paradigm in [4]. Under certain conditions, prior work [4, 8] have shown that accumulating real and synthetic data mitigates model degradation, preventing population risk from diverging. + +This work focuses on the replace and accumulate-subsample paradigms following prior studies of [3, 4, 6]. Please refer to the Appendix for additional related work. + +# 3 Model Collapses from Generalization to Memorization + +In this section, we empirically demonstrate the transition from generalization to memorization that occurs over recursive iterations, and investigate the underlying factors driving this transition. All experiments in this section are conducted on the CIFAR-10 dataset [23] using a UNet-based DDPM model [16], under the replace paradigm, where each model is trained solely on samples generated by the model from the previous iteration. However, in Appendix D, we extend the explorations to other datasets and paradigms, where the conclusion remains valid. + +Generalization Score. To quantify generalization ability, we adopt the generalization score [15, 24, 25], defined as the average distance between each generated image and its nearest training image: + +$$ +\operatorname {G S} (n) \triangleq \operatorname {D i s t} \left(\mathcal {D} _ {n}, \mathcal {G} _ {n}\right) = \frac {1}{\left| \mathcal {G} _ {n} \right|} \sum_ {\boldsymbol {x} \in \mathcal {G} _ {n}} \min _ {z \in \mathcal {D} _ {n}} \kappa (\boldsymbol {x}, z), \tag {1} +$$ + +where $\kappa (\cdot ,\cdot):\mathbb{R}^d\times \mathbb{R}^d\to \mathbb{R}$ denotes a distance metric between two data points. A higher generalization score $\mathbf{GS}(n)$ indicates that the model generates novel images rather than replicating training samples. + +Remark: [15] measure generalizability as the probability that the similarity between a generated image and its nearest training sample exceeds a threshold. [25] assesses memorization via a hypothesis test. While definitions of generalizability vary across studies, they are fundamentally similar, relying on nearest neighbor distances. + +Highlight of Observations: The generated data progressively collapses into numerous compact local clusters over model collapse iterations, as evidenced by both the sharp decline in entropy over iterations and visualizations. This localized concentration of data points then facilitates memorization in subsequent models, reducing their ability to generate novel images. Our claim is supported by the following three findings. + +# 3.1 Finding I: Generalization-to-Memorization Transition + +A generalization-to-memorization transition is revealed by experiments showing that the model initially generates novel images but gradually shifts to reproducing training samples in later iterations. We conduct iterative experiments on the CIFAR-10 benchmark [23] to illustrate this transition. Figure 2 visualizes generated samples alongside their nearest neighbors in the training set. With a relatively large sample size, i.e., 32,768 real samples as the starting training dataset, the model tends to generalize first and then memorize. At the early iterations, the diffusion model exhibits strong generalization in early iterations, producing high-quality novel images with little resemblance to training samples. However, its generalization ability deteriorates rapidly over successive iterations, and the model can only copy images from the training dataset after several iterations. + +To quantitatively validate the generalization-to-memorization transition, we track the generalization score (GS) introduced in Equation (1), which measures the similarity between the generated images $\mathcal{G}_i$ and the corresponding training images $\mathcal{D}_i$ at each iteration. We follow the protocol of [15] and construct six nested CIFAR-10 subsets of increasing size: $|\mathcal{D}_1| \in \{1,024; 2,048; 4,096; 8,192; 16,384; 32,768\}$ . These subsets span approximately $3\%$ to $65\%$ of the full training set, providing controlled start points that reflect varying degrees of memorization and generalization. As shown in Figure 2, GS drops almost exponentially with successive iterations, providing strong empirical evidence for the generalization-to-memorization transition. The decline is noticeably slower for larger training subsets, indicating that larger sample sizes preserve generalization longer and delay the onset of memorization. For the smallest dataset of 1,024 images, the model enters the memorization regime from the first iteration and remains there throughout. + +# 3.2 Finding II: The Entropy of the Training Set Shrinks Sharply over Iterations + +We identify entropy as the key evolving factor in the training data that drives the transition from generalization to memorization. Prior work [24] interprets generalization in diffusion models as a failure to memorize the entire training set. [15] further show that diffusion models tend to generalize + +![](images/4daf018998369b99222faeecc73bedab150b81a7be99f5d48392779dd02795a7.jpg) +Figure 2: The generalization-to-memorization transition. Left: visualization of the generated images $(\mathcal{G}_n)$ and their nearest neighbors in the training dataset $(\mathcal{D}_n)$ . As the iteration proceeds, the model can only copy images from the training dataset. Right: quantitative results of the generalization score of models over successive iterations. We use different colors to represent different dataset sizes. A smaller dataset has a larger decaying rate and even falls in the memorization regime at the start [15]. We use "iteration" to denote a full cycle of training and generation, rather than a gradient update. + +![](images/c805d465422f989a7fe72b5150ee0e535fae0e61f8ccc8c81f28ed5b47c70598.jpg) + +when trained on large datasets (e.g., $>2^{14}$ images in CIFAR-10) and to memorize when trained on small ones (e.g., $<2^{9}$ images). However, since we fix the size of the training dataset for every iteration, the previous conclusion [15] that a larger dataset leads to generalization cannot fully explain the phenomenon observed in Finding I. We hypothesize that although sample size remains constant, the amount of information it contains decreases over time, making it easier for the model to memorize. Based on this hypothesis, we adopt the differential entropy to quantify the information content and complexity of the continuous image distribution. + +Definition 3.1 (Differential Entropy [26]). Let $\mathbf{X}$ be a continuous random variable with probability density function $f$ supported on the set $\mathcal{X}$ . The differential entropy $H(\mathbf{X})$ is defined as + +$$ +H (\boldsymbol {X}) = \mathbb {E} [ - \log (f (\boldsymbol {X})) ] = - \int_ {\mathcal {X}} f (\boldsymbol {x}) \log f (\boldsymbol {x}) d \boldsymbol {x} +$$ + +Estimation. However, the density function $f$ of the image distribution is unknown. We use the following Kozachenko-Leonenko (KL) estimator proposed in a well-known paper [27] to empirically estimate $H(\mathbf{X})$ from a finite set of i.i.d. samples $\mathcal{D}$ drawn from the distribution $P$ : + +$$ +\hat {H} _ {\gamma} (\mathcal {D}) = \psi (| \mathcal {D} |) - \psi (\gamma) + \log c _ {d} + \frac {d}{| \mathcal {D} |} \sum_ {\boldsymbol {x} \in \mathcal {D}} \log \varepsilon_ {\gamma} (\boldsymbol {x}), \tag {2} +$$ + +where $\psi : \mathbb{N} \to \mathbb{R}$ is the digamma function, i.e., the derivative of the logarithm of the gamma function; $\gamma$ is any positive integer; $c_d$ denotes the volume of the unit ball in the $d$ -dimensional space; and $\varepsilon_{\gamma}(\pmb{x}) = \kappa(\pmb{x}, \pmb{x}_{\gamma})$ represents the $\gamma$ -nearest neighbor distance, where $\pmb{x}_{\gamma}$ is the $\gamma$ -th nearest neighbor of $\pmb{x}$ in the set $\mathcal{D}$ . Prior work [28] has shown that the KL estimator is asymptotically unbiased and consistent under broad conditions. + +We use the KL estimator with $\gamma = 1$ to measure the entropy of the image dataset at each iteration. As shown in Figure 3, the entropy of the generated image dataset—used as the training set in the next iteration—consistently decreases over iterations. With a fixed dataset size, the only dataset-dependent term in Equation (2) is the sum of nearest-neighbor distances $\varepsilon(\boldsymbol{x})$ indicating that samples in $\mathcal{D}$ become increasingly concentrated. This suggests the distribution is becoming spiky. Figure 3 further illustrates this trend by projecting high-dimensional images onto the subspace spanned by their top two eigenvectors. The visualization reveals that the generated images form numerous local clusters, while the overall support of the distribution remains relatively stable. + +# 3.3 Finding III: The Correlation between Entropy and Generalization Score + +We verify that the generalization score of the trained model is strongly correlated with the entropy of the training dataset. The similar collapsing trends of entropy and the generalization score motivate a + +![](images/1622583c19a5f3b0ea22f0f56b0bf43ca86f6f4b90e92a227788ea5727458c00.jpg) +Figure 3: Decreasing entropy and visualizations. Left: The evolving entropy of the training dataset over iterations. Under the replace paradigm, the training data is the generated data from the last iteration. Middle and Right: 2-D projection of data points onto the first two singular bases of the real dataset. The orange points represent the generated images at the 1-st and 21-st iterations, respectively. + +![](images/7ca7b4b44c0a05c79944283859d7b15233977499657b7c6b5bce837e1b3499f7.jpg) + +![](images/328e16861a73c02cc23c7c4d06db6569c552f0e0859dcd029aa86a8f40ce85d6.jpg) + +![](images/b95ec76c0c387928e085d5e62af3f31524adcfdc663fd68b2d3fad35e168aaf5.jpg) +(a) Generalization score versus estimated entropy. +Figure 4: Scatter plots of the generalization score and properties of the training dataset, i.e., entropy and variance. Each point denotes one iteration of training in the self-consuming loop. We use different colors to represent the results of different dataset sizes. + +![](images/a7be7cb8dffacf03563df77e96a22d84a25ede9077f0fd05a593e5e31331581a.jpg) +(b) Generalization score versus trace of covariance. + +deeper investigation into their relationship. In Figure 4a, we present a scatter plot of entropy versus generalization score across different dataset sizes and successive iterations, with the y-axis shown on a logarithmic scale. Notably, the entropy of the training dataset exhibits a significant linear relationship with the logarithm of the generalization score. The Pearson correlation coefficient is 0.91 with a $p$ -value near zero, quantitatively confirming the strength of this correlation. Furthermore, the scatter points corresponding to different dataset sizes are all approximately aligned along a single line, suggesting the generality of the relationship between entropy and generalization. Specifically, training datasets with higher entropy consistently yield better generalization in the trained model. We also observe that larger datasets typically have higher entropy and result in better generalization, aligning with the conclusion in [15] regarding the connection between dataset size and generalization. For comparison, Figure 4b shows the scatter plot of variance versus generalization score. The correlation appears substantially weaker than in Figure 4a, suggesting that variance in the training dataset may not directly influence the generalization performance of the trained model. + +Conclusion. Findings III collectively indicate that the generated data gradually collapses into compact clusters, as shown by declining entropy. This concentration promotes memorization in the model of the next iteration and reduces its ability to generalize. In Appendix D, we further extend these explorations to the FFHQ dataset and accumulate paradigm, demonstrating a more robust relation. + +# 4 Mitigating Model Collapse with Data Selection Methods + +In this section, we propose a sample selection method that selects a subset of images from the candidate pool to mitigate the model collapse. Motivated by the empirical finding in Section 3, the selected training images should have high entropy, characterized by large nearest-neighbor distances and greater diversity. This objective can be formalized as the following optimization problem: + +$$ +\max _ {\mathcal {D} \subset \mathcal {S}, | \mathcal {D} | = N} \hat {H} _ {1} (\mathcal {D}) \quad \Longleftrightarrow \quad \max _ {\mathcal {D} \subset \mathcal {S}, | \mathcal {D} | = N} \sum_ {\boldsymbol {x} \in \mathcal {D}} \log \min _ {\boldsymbol {y} \in \mathcal {D} \backslash \boldsymbol {x}} \kappa (\boldsymbol {x}, \boldsymbol {y}), +$$ + +where $S$ denotes the candidate pool. For the accumulate-subsample setting, $S$ is all the accumulated images so far. In the replace setting, $S$ consists of the images generated by the previous model, with size twice that of the target training set. + +Unfortunately, this non-convex max-min problem is hard to solve to a global solution efficiently. Alternatively, we approximate the solution through a greedy strategy inspired by farthest-point sampling, which aims to maximize the nearest-neighbor distance. + +Method I: Greedy Selection. The procedure iteratively constructs a subset $\mathcal{D} \subset S$ of size $n$ by adding the farthest point one at a time as follows: + +1. Initialization: Randomly select an initial point from the dataset $S$ and add it to the set $\mathcal{D}$ . +2. Iterative Selection: At each iteration, for every candidate point $\pmb{x} \in S \setminus \mathcal{D}$ , compute the minimum distance from $\pmb{x}$ to all points currently in $\mathcal{D}$ . Select the point with the maximum of these minimum distances and add it to $\mathcal{D}$ , i.e., $\pmb{x}_{\text{select}} = \arg \max_{\pmb{x} \in S \setminus \mathcal{D}} \min_{\pmb{y} \in \mathcal{D}} \kappa(\pmb{x}, \pmb{y})$ . +3. Termination: Repeat the selection process until $|\mathcal{D}| = N$ . + +The Greedy Selection method can efficiently and effectively extract a subset with a large entropy. However, this greedy method carries a risk of over-optimization, which may lead to an excessively expanded distribution and a progressively increasing variance in the selected samples. To mitigate this, we also provide the following Threshold Decay Filter, which can control the filtration strength by a decaying threshold. + +Methods II: Threshold Decay Filter. The procedure constructs a subset $\mathcal{D} \subset S$ of size $N$ by iteratively selecting samples that are sufficiently distant from the current set $\mathcal{D}$ . The algorithm proceeds as follows: + +1. Initialization: Set an initial threshold $\tau > 0$ . Randomly select one point from the dataset $\mathcal{S}$ and add it to the set $\mathcal{D}$ . +2. Threshold-based Selection: For each point $\pmb{x} \in S \setminus \mathcal{D}$ , compute the distance from $\pmb{x}$ to all points in $\mathcal{D}$ . If all distances are greater than the current threshold $\tau$ , add $\pmb{x}$ to $\mathcal{D}$ . +3. Threshold Decay: If $|\mathcal{D}| < N$ after a complete pass through $S \setminus \mathcal{D}$ , reduce the threshold $\tau$ by multiplying it with a decay factor $\alpha \in (0,1)$ , and repeat Step 2. +4. Termination: Repeat Steps 2-3 until $|\mathcal{D}| = N$ . + +Threshold Decay Filter is a soft variant of Greedy Selection that provides adjustable control over the selection strength. When the initial threshold is set sufficiently high and the decay factor is close to 1, the Threshold Decay Filter behaves similarly to Greedy Selection. Conversely, if both the initial threshold and decay factor are set to 0, the filter does not filter out any data point and reduces to the vanilla replace or accumulate-subsample paradigm. In practice, we first extract image features using a DINOv2 [29] model and compute distances in the feature space, i.e., $\kappa (\pmb {x},\pmb {y}) = \| h(\pmb {x}) - h(\pmb {y})\| _2$ where $h(\cdot)$ denotes the feature extractor. + +# 5 Experiments + +We empirically evaluate how our data selection strategies introduced in Section 4 interact with the two self-consuming paradigms: replace and accumulate-subsample. In both settings, our method serves as a plug-in component for selecting high-quality and diverse training samples from the candidate pool. We demonstrate that the proposed strategies effectively alleviate memorization and + +![](images/dfd6e04d616073133e957a815c3d02ada61387215c7aef4bd4b53fb284a5f060.jpg) +Figure 5: Generalization Score of the trained model over iterations. We indicate the settings on top of the subfigures. In each subfigure, three different lines are used to represent the vanilla paradigm and its variants augmented with the proposed selection methods. + +![](images/d18e67a873896600e9903f845ce81ce3e98fd0e43f40c2431a71eada2b3419c4.jpg) +Figure 6: FID of the generated images over iterations. We indicate the settings on top of the subfigures. In each subfigure, three different lines are used to represent the vanilla paradigm and its variants augmented with the proposed selection methods. + +reduce FID scores, thereby mitigating model collapse. Additionally, experiments on classifier-free guidance (CFG) [30] generation show that our method effectively mitigates diversity collapse of CFG. + +Datasets. We conduct experiments on three widely used image generation benchmarks. CIFAR-10 [23] consists of $32 \times 32$ color images in 10 classes. Due to computational constraints, we use a subset of 32,768 training images. Our goal is not to achieve state-of-the-art FID among large diffusion models but rather to demonstrate that our method mitigates memorization in the self-consuming loop. As shown in Section 3, this subset is sufficient to observe the transition from generalization to memorization. We also conduct experiments on subsets of FFFQ [31], downsampled to $32 \times 32$ resolution, and MNIST [32], using 8,192 and 12,000 training images, respectively. + +Model. For CIFAR-10 and FFHQ, we employ a UNet-based backbone [33] designed for $32 \times 32$ RGB images, which predicts noise residuals. The network contains approximately 16M parameters. For MNIST, we use a similar UNet-based architecture adapted for single-channel inputs, with a total of 19M parameters. Detailed network configurations are provided in the Appendix. + +Implementation. For iterative training and sampling, our implementation is based on the Hugging Face Diffusers codebase [34] of DDPM. We use a mixed-precision training of FP16 to train the models. We adopt an Adam optimizer with a learning rate of $10^{-4}$ and a weight decay of $10^{-6}$ . The batch size is 128. A 1000-step denoising process is used, with all other hyperparameters set to their default values. For Threshold Decay Filter, we use an initial threshold of 60 and a decay rate of 0.95. We show in the Appendix that our method is robust in a wide range of hyperparameters. + +Evaluation Metrics. We use the generalization score and entropy to evaluate the effectiveness of our method in mitigating memorization. Additionally, we adopt the Fréchet Inception Distance (FID) [35] as a metric to quantify the distributional divergence between generated images and real images. + +![](images/d38f18d8eb639c4e88cda4d1d95d4f23db795dace75764ca174bcf58b2dcbb14.jpg) +Figure 7: Estimated entropy of the training datasets over iterations. We indicate the settings on top of the subfigures. In each subfigure, three different lines are used to represent the vanilla paradigm and its variants augmented with our selection methods. + +![](images/5e37dd840f959672a4e8a828bea19cae13017092924a6f023019c320788c2c0c.jpg) +Figure 8: Proportion of the selected images from previous iterations or the real dataset. We use different colors to represent different sources. Particularly, the blue bars denote the proportion of the real images. The red line represents the $1/n$ curve that indicates the proportion of the real images if we evenly select the data subset from all available images (accumulate-subsample). We indicate the settings on top of the subfigures. + +![](images/b60b6b7ffbbd6590575afaf635fb3595bcfd534a755225b8508b5771c90de32f.jpg) + +![](images/ab85fc0099c36257732e7c50c4f9dde7ef4cca0dae15d7e5856ceba083ed4659.jpg) + +![](images/7efabdd0f511d591d51ebd2119b680c6ff3f62407ea30312de7d4c0fa38732ce.jpg) + +Results. To evaluate the efficacy of our selection methods in mitigating memorization within the self-consuming loop, we compare the generalization scores of two vanilla paradigms with those augmented by Greedy Selection or Threshold Decay Filter. As shown in Figure 5, Greedy Selection consistently improves generalization scores across all datasets and paradigms and is particularly effective in the accumulate-subsample setting. Importantly, our selection methods mitigate memorization without compromising FID performance; in fact, they can even slow down FID degradation. Figure 6 reports the FID of generated images over successive iterations. Under the accumulate-subsample paradigm, both methods yield notable improvements in FID, with Greedy Selection outperforming Threshold Decay Filter. For example, the vanilla accumulate paradigm reaches an FID of 75.7 at iteration 8, whereas Greedy Selection significantly reduces it to 44.7. On the FFHQ dataset under the replace paradigm, however, Threshold Decay Filter performs better, suggesting that adaptive selection strength may be beneficial in certain cases. + +Analysis for the Improvement. We present the estimated entropy of the training datasets in Figure 7. As shown in the figure, our selection methods effectively increase the entropy of the training data at each iteration, consistent with their design. These more diverse, higher-entropy datasets subsequently enable the next-iteration model to generalize better, as evidenced by the improved generalization scores in Figure 5. + +We further investigate which samples are selected by our methods under the accumulate-subsample paradigm, where the model has access to all prior synthetic images and the real images, but is trained on a subset of them. Figure 8 shows the proportion of selected samples originating from different sources, with the blue bar indicating the proportion of real images. As illustrated in the figure, both Greedy Selection and Threshold Decay Filter consistently select a significantly higher proportion of real images compared to the $1/n$ reference curve. For example, on the FFHQ dataset, Greedy Selection selects $65\%$ real images at the 8-th iteration, while vanilla subsampling includes only $12.5\%$ . This outcome arises because the image distribution progressively collapses into compact clusters over iterations, and our selection methods tend to prioritize boundary samples—namely, real images—by maximizing the nearest neighbor distance. + +![](images/e25854f2d6166a03a06ececfa23bb6579f8ff05517084c3a485753dff49deda2.jpg) +(a) Unconditional + +![](images/467fca542cfd4950db197d817a87207ad9404bd5ea36d54b7469cd4b39576690.jpg) +(b) CFG (Scale = 2) + +![](images/4cf8a495e638c34b61a557c32f9fdd83e3c86eca098a5cc6d2f652b3463b080f.jpg) +(c) CFG With Filtering +Figure 9: Comparisons of unconditional, CFG, and CFG augmented with filter method. (a)-(c): Generated samples on MNIST [32] at the 8-th iteration under the accumulate paradigm. The FIDs of images in (a)-(c) are 74.4, 66.2, and 22.4 respectively. (d): The estimated entropy of the generated dataset over iterations, which reflects the diversity of the generated images. + +![](images/287ea70fbbb63a8bc02e9a73e351fa19de79a28236b73d7015831909a4ab6880.jpg) +(d) Entropy versus iteration + +Diversity Improvement on Classifier-Free Diffusion Guidance (CFG) [30]. We further validate the effectiveness of our methods in the CFG setting and show that they substantially improve the diversity of generated images. CFG is a widely used conditional generation technique that consistently enhances perceptual quality but often sacrifices diversity [30, 36, 37]. Prior work by [38] identifies the CFG scale as a key factor influencing the rate of model collapse and suggests that setting a moderate scale can help mitigate model collapse. Experimentally, we observe that the CFG generation can indeed generate clearer images than the unconditional baseline, as compared in Figures 9a and 9b. However, the diversity of generated images collapses rapidly even with a modest guidance scale, with samples within each class soon becoming nearly identical. In contrast, Figures 9c and 9d demonstrate that augmenting CFG with our data selection methods significantly improves image diversity and yields substantially lower FID scores compared to the vanilla CFG paradigm. These results demonstrate that our approach effectively mitigates the diversity loss of CFG while preserving its quality advantage throughout the self-consuming loop. + +# 6 Conclusion and Discussion + +In this work, we reveal a generalization-to-memorization transition in diffusion models under recursive training, highlighting a serious practical concern and offering a new perspective on model collapse. We empirically demonstrate that the entropy of the training data decreases over iterations and is strongly correlated with the model's generalizability. Motivated by the findings, we propose an entropy-based data selection strategy that effectively alleviates the generalization-to-memorization transition and improves image quality, thus mitigating model collapse. + +Based on this work, we believe many future directions could be further investigated. This paper doesn't rigorously explain why the entropy is collapsing. The finite dataset size, training and sampling errors, and the model bias could all contribute to the collapsing entropy (or the information loss). We envision that a more formal theoretical analysis of collapse dynamics could be developed based on theoretical models such as the mixture of Gaussians [39-41]. Besides, this framework could be extended to the language modality, investigating the discrete diffusion model [42-44]. Additionally, a more efficient algorithm is needed, as the current greedy selection method is computationally expensive. + +# Acknowledgement + +LS, MW, HZ, ZZ, and QQ acknowledge funding support from NSF CCF-2212066, NSF CCF-2212326, NSF IIS 2402950, and ONR N000142512339, and the Google Research Scholar Award. MT is grateful for partial supports by NSF Grants DMS-1847802, DMS-2513699, DOE Grants NA0004261, SC0026274, and Richard Duke Fellowship. We also thank Prof. Peng Wang (University of Michigan and University of Macau), Mr. Xiang Li (University of Michigan), and Mr. Xiao Li (University of Michigan) for fruitful discussions and valuable feedback. + +# References + +[1] Elvis Dohmatob, Yunzhen Feng, and Julia Kempe. Strong model collapse. arXiv preprint arXiv:2410.04840, 2024. +[2] Rylan Schaeffer, Joshua Kazdan, Alvan Caleb Arulandu, and Sanmi Koyejo. Position: Model collapse does not mean what you think. arXiv preprint arXiv:2503.03150, 2025. +[3] Ilia Shumailov, Zakhar Shumaylov, Yiren Zhao, Nicolas Papernot, Ross J. Anderson, and Yarin Gal. AI models collapse when trained on recursively generated data. Nature, 631(8022):755-759, 2024. +[4] Joshua Kazdan, Ryan Schaeffer, Apratim Dey, Matthias Gerstgrasser, Rafael Rafailov, David L. Donoho, and Sanmi Koyejo. Collapse or thrive? perils and promises of synthetic data in a self-generating world. arxiv preprint arxiv:2410.16713, 2024. +[5] Quentin Bertrand, Avishek Joey Bose, Alexandre Duplessis, Marco Jiralerspong, and Gauthier Gidel. On the stability of iterative retraining of generative models on their own data. In ICLR, 2024. +[6] Ananda Theertha Suresh, Andrew Thangaraj, and Aditya Nanda Kishore Khandavally. Rate of model collapse in recursive training. arxiv preprint arxiv:2412.17646, 2024. +[7] Elvis Dohmatob, Yunzhen Feng, and Julia Kempe. Model collapse demystified: The case of regression. In NeurIPS, 2024. +[8] Matthias Gerstgrasser, Ryan Schaeffer, Apratim Dey, Rafael Rafailov, Henry Sleight, John Hughes, Tomasz Korbak, Rajashree Agrawal, Dhruv Pai, Andrey Gromov, Daniel A. Roberts, Diyi Yang, David L. Donoho, and Sanmi Koyejo. Is model collapse inevitable? breaking the curse of recursion by accumulating real and synthetic data. arXiv preprint arXiv:2404.01413, 2024. +[9] Elvis Dohmatob, Yunzhen Feng, Arjun Subramonian, and Julia Kempe. Strong model collapse. arxiv preprint arxiv:2410.04840, 2024. +[10] Shi Fu, Sen Zhang, Yingjie Wang, Xinmei Tian, and Dacheng Tao. Towards theoretical understandings of self-consuming generative models. In ICML, 2024. +[11] Sina Alemohammad, Josue Casco-Rodriguez, Lorenzo Luzi, Ahmed Imtiaz Humayun, Hossein Babaei, Daniel LeJeune, Ali Siahkoohi, and Richard G. Baraniuk. Self-consuming generative models go MAD. In ICLR, 2024. +[12] Matyas Bohacek and Hany Farid. Nepotistically trained generative-ai models collapse. arxiv preprint arxiv:2311.12202, 2023. +[13] Xulu Zhang, Xiaoyong Wei, Jinlin Wu, Jiaxin Wu, Zhaoxiang Zhang, Zhen Lei, and Qing Li. Generating on generated: An approach towards self-evolving diffusion models. arxiv preprint arxiv:2502.09963, 2025. +[14] Brendan Leigh Ross, Hamidreza Kamkari, Tongzi Wu, Rasa Hosseinzadeh, Zhaoyan Liu, George Stein, Jesse C. Cresswell, and Gabriel Loaiza-Ganem. A geometric framework for understanding memorization in generative models. arxiv preprint arxiv:2411.00113, 2024. +[15] Huijie Zhang, Jinfan Zhou, Yifu Lu, Minzhe Guo, Peng Wang, Liyue Shen, and Qing Qu. The emergence of reproducibility and consistency in diffusion models. In ICML, 2024. +[16] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. +[17] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. In NeurIPS, 2019. +[18] Pascal Vincent. A connection between score matching and denoising autoencoders. Neural Computing, 23:1661-1674, 2011. + +[19] Sina Alemohammad, Ahmed Imtiaz Humayun, Shruti Agarwal, John P. Collomosse, and Richard G. Baraniuk. Self-improving diffusion models with synthetic data. arxiv preprint arxiv:2408.16333, 2024. +[20] Elvis Dohmatob, Yunzhen Feng, Pu Yang, François Charton, and Julia Kempe. A tale of tails: Model collapse as a change of scaling laws. In ICML, 2024. +[21] Yunzhen Feng, Elvis Dohmatob, Pu Yang, Francois Charton, and Julia Kempe. Beyond model collapse: Scaling up with synthesized data requires verification. arXiv preprint arXiv:2406.07515, 2024. +[22] Apratim Dey and David L. Donoho. Universality of the $\pi^2 /6$ pathway in avoiding model collapse. arxiv preprint arxiv:2410.22812, 2024. +[23] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. +[24] TaeHo Yoon, Joo Young Choi, Sehyun Kwon, and Ernest K Ryu. Diffusion probabilistic models generalize when they fail to memorize. In workshop on structured probabilistic inference $\{\backslash \& \}$ generative modeling, ICML, 2023. +[25] Ahmed M. Alaa, Boris van Breugel, Evgeny S. Saveliev, and Mihaela van der Schaar. How faithful is your synthetic data? sample-level metrics for evaluating and auditing generative models. In ICML, 2022. +[26] Thomas M Cover. Elements of information theory. John Wiley & Sons, 1999. +[27] Leonenko Kozachenko. Sample estimate of the entropy of a random vector. Problems of Information Transmission, 23:9, 1987. +[28] Jan Kybic and Ivan Vnucko. Approximate all nearest neighbor search for high dimensional entropy estimation for image registration. Signal Processing, 92:1302-1316, 2012. +[29] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Mido Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Hervé Jégou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision. Transactions on Machine Learning Research, 2024, 2024. +[30] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arxiv preprint arxiv:2207.12598, 2022. +[31] Anoop Jain, Parag Sarda, and Jayant R. Haritsa. Providing diversity in k-nearest neighbor query results. In Honghua Dai, Ramakrishnan Srikant, and Chengqi Zhang, editors, PAKDD, 2004. +[32] Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278-2324, 1998. +[33] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, 2015. +[34] Patrick von Platen, Suraj Patil, Anton Lozhkov, Pedro Cuenca, Nathan Lambert, Kashif Rasul, Mishig Davaadorj, Dhruv Nair, Sayak Paul, William Berman, Yiyi Xu, Steven Liu, and Thomas Wolf. Diffusers: State-of-the-art diffusion models. https://github.com/huggingface/ diffusers, 2022. +[35] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In NeurIPS, 2017. +[36] Muthu Chidambaram, Khashayar Gatmiry, Sitan Chen, Holden Lee, and Jianfeng Lu. What does guidance do? A fine-grained analysis in a simple setting. In NeurIPS, 2024. + +[37] Xiang Li, Rongrong Wang, and Qing Qu. Towards understanding the mechanisms of classifier-free guidance. arXiv preprint arXiv:2505.19210, 2025. +[38] Youngseok Yoon, Dainong Hu, Iain Weissburg, Yao Qin, and Haewon Jeong. Model collapse in the self-consuming chain of diffusion finetuning: A novel perspective from quantitative trait modeling. arXiv preprint arXiv:2407.17493, 2024. +[39] Xiao Li, Zekai Zhang, Xiang Li, Siyi Chen, Zhihui Zhu, Peng Wang, and Qing Qu. Understanding representation dynamics of diffusion models via low-dimensional modeling. arXiv preprint arXiv:2502.05743, 2025. +[40] Peng Wang, Huijie Zhang, Zekai Zhang, Siyi Chen, Yi Ma, and Qing Qu. Diffusion models learn low-dimensional distributions via subspace clustering. arXiv preprint arXiv:2409.02426, 2024. +[41] Siyi Chen, Huijie Zhang, Minzhe Guo, Yifu Lu, Peng Wang, and Qing Qu. Exploring low-dimensional subspaces in diffusion models for controllable image editing. arXiv preprint arXiv:2409.02374, 2024. +[42] Aaron Lou, Chenlin Meng, and Stefano Ermon. Discrete diffusion modeling by estimating the ratios of the data distribution. In ICML, 2024. +[43] Kevin Rojas, Yuchen Zhu, Sichen Zhu, Felix X.-F. Ye, and Molei Tao. Diffuse everything: Multimodal diffusion models on arbitrary state spaces. arXiv preprint arXiv:2506.07903, 2025. +[44] Jaeyeon Kim, Kulin Shah, Vasilis Kontonis, Sham M. Kakade, and Sitan Chen. Train for the worst, plan for the best: Understanding token ordering in masked diffusions. arXiv preprint arXiv:2502.06768, 2025. +[45] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, Patrick Schramowski, Srivatsa Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. LAION-5B: an open large-scale dataset for training next generation image-text models. In NeurIPS, 2022. +[46] Sierra Calanda Wyllie, Ilia Shumailov, and Nicolas Papernot. Fairness feedback loops: Training on synthetic data amplifies bias. In ACM FAccT, pages 2113-2147, 2024. +[47] Mohamed El Amine Seddik, Suei-Wen Chen, Soufiane Hayou, Pierre Youssef, and Mérouane Debbah. How bad is training on synthetic data? A statistical analysis of language model collapse. arxiv preprint arxiv:2404.05090, 2024. +[48] Damien Ferbach, Quentin Bertrand, Avishek Joey Bose, and Gauthier Gidel. Self-consuming generative models with curated data provably optimize human preferences. In NeurIPS, 2024. +[49] Shi Fu, Yingjie Wang, Yuzhu Chen, Xinmei Tian, and Dacheng Tao. A theoretical perspective: How to prevent model collapse in self-consuming training loops. In ICLR, 2025. +[50] Huminhao Zhu, Fangyikang Wang, Tianyu Ding, Qing Qu, and Zhihui Zhu. Analyzing and mitigating model collapse in rectified flow models. arXiv preprint arXiv:2412.08175, 2024. +[51] Hugo Cui, Cengiz Pehlevan, and Yue M. Lu. A precise asymptotic analysis of learning diffusion models: theory and insights. arxiv preprint arxiv:2501.03937, 2025. +[52] Xuekai Zhu, Daixuan Cheng, Hengli Li, Kaiyan Zhang, Ermo Hua, Xingtai Lv, Ning Ding, Zhouhan Lin, Zilong Zheng, and Bowen Zhou. How to synthesize text data without model collapse? arxiv preprint arxiv:2412.14689, 2024. +[53] Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron C. Courville, and Yoshua Bengio. Generative adversarial nets. In NeurIPS, pages 2672–2680, 2014. +[54] Peng Wang, Huijie Zhang, Zekai Zhang, Siyi Chen, Yi Ma, and Qing Qu. Diffusion models learn low-dimensional distributions via subspace clustering. arXiv preprint arXiv:2409.02426, 2024. + +[55] Yuqing Wang, Ye He, and Molei Tao. Evaluating the design space of diffusion-based generative models. NeurIPS, 2024. +[56] Huijie Zhang, Zijian Huang, Siyi Chen, Jinfan Zhou, Zekai Zhang, Peng Wang, and Qing Qu. Understanding generalization in diffusion models via probability flow distance. arXiv preprint arXiv:2505.20123, 2025. +[57] Zahra Kadkhodaie, Florentin Guth, Eero P Simoncelli, and Stéphane Mallat. Generalization in diffusion models arises from geometry-adaptive harmonic representations. In ICLR, 2024. +[58] Jie An, De Wang, Pengsheng Guo, Jiebo Luo, and Alexander Schwing. On inductive biases that enable generalization of diffusion transformers. arXiv preprint arXiv:2410.21273, 2024. +[59] Binxu Wang and John J Vastola. The unreasonable effectiveness of gaussian score approximation for diffusion models and its applications. arXiv preprint arXiv:2412.09726, 2024. +[60] Xiang Li, Yixiang Dai, and Qing Qu. Understanding generalizability of diffusion models requires rethinking the hidden gaussian structure. In NeurIPS, 2024. +[61] Matthew Niedoba, Berend Zwartsenberg, Kevin Murphy, and Frank Wood. Towards a mechanistic explanation of diffusion model generalization. arXiv preprint arXiv:2411.19339, 2024. +[62] Mason Kamb and Surya Ganguli. An analytic theory of creativity in convolutional diffusion models. arXiv preprint arXiv:2412.20292, 2024. +[63] Xiangming Gu, Chao Du, Tianyu Pang, Chongxuan Li, Min Lin, and Ye Wang. On memorization in diffusion models. Transactions on Machine Learning Research, 2025. +[64] Nicholas Carlini, Jamie Hayes, Milad Nasr, Matthew Jagielski, Vikash Sehwag, Florian Tramér, Borja Balle, Daphne Ippolito, and Eric Wallace. Extracting training data from diffusion models. In USENIX Security Symposium, pages 5253-5270, 2023. +[65] Gowthami Somepalli, Vasu Singla, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Diffusion art or digital forgery? investigating data replication in diffusion models. In CVPR, pages 6048-6058, 2023. +[66] Gowthami Somepalli, Vasu Singla, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Understanding and mitigating copying in diffusion models. NeurIPS, 36:47783-47803, 2023. +[67] Yuxin Wen, Yuchen Liu, Chen Chen, and Lingjuan Lyu. Detecting, explaining, and mitigating memorization in diffusion models. In ICLR, 2024. +[68] Ben Sorscher, Robert Geirhos, Shashank Shekhar, Surya Ganguli, and Ari Morcos. Beyond neural scaling laws: beating power law scaling via data pruning. In NeurIPS, 2022. +[69] Germain Kolossov, Andrea Montanari, and Pulkit Tandon. Towards a statistical theory of data selection under weak supervision. In ICLR, 2024. +[70] Reyhane Askari Hemmat, Mohammad Pezeshki, Elvis Dohmatob, Florian Bordes, Pietro Astolfi, Melissa Hall, Jakob Verbeek, Michal Drozdal, and Adriana Romero-Soriano. Improving the scaling laws of synthetic data with deliberate practice. *arxiv preprint arxiv:2502.15588*, 2025. +[71] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jonathon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In CVPR, pages 2818-2826, 2016. +[72] Sofiane Abbar, Sihem Amer-Yahia, Piotr Indyk, Sepideh Mahabadi, and Kasturi R. Varadarajan. Diverse near neighbor problem. In SoCG, 2013. +[73] George Stein, Jesse C. Cresswell, Rasa Hosseinzadeh, Yi Sui, Brendan Leigh Ross, Valentin Villecloze, Zhaoyan Liu, Anthony L. Caterini, J. Eric T. Taylor, and Gabriel Loaiza-Ganem. Exposing flaws of generative model evaluation metrics and their unfair treatment of diffusion models. In NeurIPS, 2023. + +# NeurIPS Paper Checklist + +# 1. Claims + +Question: Do the main claims made in the abstract and introduction accurately reflect the paper's contributions and scope? + +Answer: [Yes] + +Justification: The abstract and introduction clearly state our motivations, empirical findings, proposed methods, experimental validation, and other contributions. + +Guidelines: + +- The answer NA means that the abstract and introduction do not include the claims made in the paper. +- The abstract and/or introduction should clearly state the claims made, including the contributions made in the paper and important assumptions and limitations. A No or NA answer to this question will not be perceived well by the reviewers. +- The claims made should match theoretical and experimental results, and reflect how much the results can be expected to generalize to other settings. +- It is fine to include aspirational goals as motivation as long as it is clear that these goals are not attained by the paper. + +# 2. Limitations + +Question: Does the paper discuss the limitations of the work performed by the authors? + +Answer: [Yes] + +Justification: We carefully discuss the problem setup and assumptions in Section 2. While our findings are empirical, exploring their theoretical foundations is an interesting avenue for future investigation. + +Guidelines: + +- The answer NA means that the paper has no limitation while the answer No means that the paper has limitations, but those are not discussed in the paper. +- The authors are encouraged to create a separate "Limitations" section in their paper. +- The paper should point out any strong assumptions and how robust the results are to violations of these assumptions (e.g., independence assumptions, noiseless settings, model well-specification, asymptotic approximations only holding locally). The authors should reflect on how these assumptions might be violated in practice and what the implications would be. +- The authors should reflect on the scope of the claims made, e.g., if the approach was only tested on a few datasets or with a few runs. In general, empirical results often depend on implicit assumptions, which should be articulated. +- The authors should reflect on the factors that influence the performance of the approach. For example, a facial recognition algorithm may perform poorly when image resolution is low or images are taken in low lighting. Or a speech-to-text system might not be used reliably to provide closed captions for online lectures because it fails to handle technical jargon. +- The authors should discuss the computational efficiency of the proposed algorithms and how they scale with dataset size. +- If applicable, the authors should discuss possible limitations of their approach to address problems of privacy and fairness. +- While the authors might fear that complete honesty about limitations might be used by reviewers as grounds for rejection, a worse outcome might be that reviewers discover limitations that aren't acknowledged in the paper. The authors should use their best judgment and recognize that individual actions in favor of transparency play an important role in developing norms that preserve the integrity of the community. Reviewers will be specifically instructed to not penalize honesty concerning limitations. + +# 3. Theory assumptions and proofs + +Question: For each theoretical result, does the paper provide the full set of assumptions and a complete (and correct) proof? + +# Answer: [NA] + +Justification: We don't have theoretical results that need assumptions or proofs. This work empirically focuses on the generalization-to-memorization transition in models happened during model collapse. + +# Guidelines: + +- The answer NA means that the paper does not include theoretical results. +- All the theorems, formulas, and proofs in the paper should be numbered and cross-referenced. +- All assumptions should be clearly stated or referenced in the statement of any theorems. +- The proofs can either appear in the main paper or the supplemental material, but if they appear in the supplemental material, the authors are encouraged to provide a short proof sketch to provide intuition. +- Inversely, any informal proof provided in the core of the paper should be complemented by formal proofs provided in appendix or supplemental material. +- Theorems and Lemmas that the proof relies upon should be properly referenced. + +# 4. Experimental result reproducibility + +Question: Does the paper fully disclose all the information needed to reproduce the main experimental results of the paper to the extent that it affects the main claims and/or conclusions of the paper (regardless of whether the code and data are provided or not)? + +# Answer: [Yes] + +Justification: Implementation details of our methods and experiments are provided in Section 5 and the Appendix. + +# Guidelines: + +- The answer NA means that the paper does not include experiments. +- If the paper includes experiments, a No answer to this question will not be perceived well by the reviewers: Making the paper reproducible is important, regardless of whether the code and data are provided or not. +- If the contribution is a dataset and/or model, the authors should describe the steps taken to make their results reproducible or verifiable. +- Depending on the contribution, reproducibility can be accomplished in various ways. For example, if the contribution is a novel architecture, describing the architecture fully might suffice, or if the contribution is a specific model and empirical evaluation, it may be necessary to either make it possible for others to replicate the model with the same dataset, or provide access to the model. In general, releasing code and data is often one good way to accomplish this, but reproducibility can also be provided via detailed instructions for how to replicate the results, access to a hosted model (e.g., in the case of a large language model), releasing of a model checkpoint, or other means that are appropriate to the research performed. +- While NeurIPS does not require releasing code, the conference does require all submissions to provide some reasonable avenue for reproducibility, which may depend on the nature of the contribution. For example +(a) If the contribution is primarily a new algorithm, the paper should make it clear how to reproduce that algorithm. +(b) If the contribution is primarily a new model architecture, the paper should describe the architecture clearly and fully. +(c) If the contribution is a new model (e.g., a large language model), then there should either be a way to access this model for reproducing the results or a way to reproduce the model (e.g., with an open-source dataset or instructions for how to construct the dataset). +(d) We recognize that reproducibility may be tricky in some cases, in which case authors are welcome to describe the particular way they provide for reproducibility. In the case of closed-source models, it may be that access to the model is limited in some way (e.g., to registered users), but it should be possible for other researchers to have some path to reproducing or verifying the results. + +# 5. Open access to data and code + +Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material? + +Answer: [Yes] + +Justification: We have clearly cited all datasets used in this paper, which are publicly accessible online. Our implementation is based on the Hugging Face Diffusers codebase [34] for DDPM. We will release the code upon acceptance of the paper. + +Guidelines: + +- The answer NA means that paper does not include experiments requiring code. +- Please see the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details. +- While we encourage the release of code and data, we understand that this might not be possible, so No is an acceptable answer. Papers cannot be rejected simply for not including code, unless this is central to the contribution (e.g., for a new open-source benchmark). +- The instructions should contain the exact command and environment needed to run to reproduce the results. See the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details. +- The authors should provide instructions on data access and preparation, including how to access the raw data, preprocessed data, intermediate data, and generated data, etc. +- The authors should provide scripts to reproduce all experimental results for the new proposed method and baselines. If only a subset of experiments are reproducible, they should state which ones are omitted from the script and why. +- At submission time, to preserve anonymity, the authors should release anonymized versions (if applicable). +- Providing as much information as possible in supplemental material (appended to the paper) is recommended, but including URLs to data and code is permitted. + +# 6. Experimental setting/details + +Question: Does the paper specify all the training and test details (e.g., data splits, hyperparameters, how they were chosen, type of optimizer, etc.) necessary to understand the results? + +Answer: [Yes] + +Justification: We specify all the needed training and test details in Sections 3 and 5. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The experimental setting should be presented in the core of the paper to a level of detail that is necessary to appreciate the results and make sense of them. +- The full details can be provided either with the code, in appendix, or as supplemental material. + +# 7. Experiment statistical significance + +Question: Does the paper report error bars suitably and correctly defined or other appropriate information about the statistical significance of the experiments? + +Answer: [Yes] + +Justification: We present the Pearson correlation coefficient in Section 3 to provide statistical significance of the correlation between entropy and generalization score. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The authors should answer "Yes" if the results are accompanied by error bars, confidence intervals, or statistical significance tests, at least for the experiments that support the main claims of the paper. + +- The factors of variability that the error bars are capturing should be clearly stated (for example, train/test split, initialization, random drawing of some parameter, or overall run with given experimental conditions). +- The method for calculating the error bars should be explained (closed form formula call to a library function, bootstrap, etc.) +- The assumptions made should be given (e.g., Normally distributed errors). +- It should be clear whether the error bar is the standard deviation or the standard error of the mean. +- It is OK to report 1-sigma error bars, but one should state it. The authors should preferably report a 2-sigma error bar than state that they have a $96\%$ CI, if the hypothesis of Normality of errors is not verified. +- For asymmetric distributions, the authors should be careful not to show in tables or figures symmetric error bars that would yield results that are out of range (e.g. negative error rates). +- If error bars are reported in tables or plots, The authors should explain in the text how they were calculated and reference the corresponding figures or tables in the text. + +# 8. Experiments compute resources + +Question: For each experiment, does the paper provide sufficient information on the computer resources (type of compute workers, memory, time of execution) needed to reproduce the experiments? + +Answer: [Yes] + +Justification: We provide the information on the computer resources in Section 5. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The paper should indicate the type of compute workers CPU or GPU, internal cluster, or cloud provider, including relevant memory and storage. +- The paper should provide the amount of compute required for each of the individual experimental runs as well as estimate the total compute. +- The paper should disclose whether the full research project required more compute than the experiments reported in the paper (e.g., preliminary or failed experiments that didn't make it into the paper). + +# 9. Code of ethics + +Question: Does the research conducted in the paper conform, in every respect, with the NeurIPS Code of Ethics https://neurips.cc/public/EthicsGuidelines? + +Answer: [Yes] + +Justification: Our research conform with the NeurIPS Code of Ethics in every respect. + +Guidelines: + +The answer NA means that the authors have not reviewed the NeurIPS Code of Ethics. +- If the authors answer No, they should explain the special circumstances that require a deviation from the Code of Ethics. +- The authors should make sure to preserve anonymity (e.g., if there is a special consideration due to laws or regulations in their jurisdiction). + +# 10. Broader impacts + +Question: Does the paper discuss both potential positive societal impacts and negative societal impacts of the work performed? + +Answer: [Yes] + +Justification: We discuss the potential impacts of our paper in the Appendix. + +Guidelines: + +- The answer NA means that there is no societal impact of the work performed. +- If the authors answer NA or No, they should explain why their work has no societal impact or why the paper does not address societal impact. + +- Examples of negative societal impacts include potential malicious or unintended uses (e.g., disinformation, generating fake profiles, surveillance), fairness considerations (e.g., deployment of technologies that could make decisions that unfairly impact specific groups), privacy considerations, and security considerations. +- The conference expects that many papers will be foundational research and not tied to particular applications, let alone deployments. However, if there is a direct path to any negative applications, the authors should point it out. For example, it is legitimate to point out that an improvement in the quality of generative models could be used to generate deepfakes for disinformation. On the other hand, it is not needed to point out that a generic algorithm for optimizing neural networks could enable people to train models that generate Deepfakes faster. +- The authors should consider possible harms that could arise when the technology is being used as intended and functioning correctly, harms that could arise when the technology is being used as intended but gives incorrect results, and harms following from (intentional or unintentional) misuse of the technology. +- If there are negative societal impacts, the authors could also discuss possible mitigation strategies (e.g., gated release of models, providing defenses in addition to attacks, mechanisms for monitoring misuse, mechanisms to monitor how a system learns from feedback over time, improving the efficiency and accessibility of ML). + +# 11. Safeguards + +Question: Does the paper describe safeguards that have been put in place for responsible release of data or models that have a high risk for misuse (e.g., pretrained language models, image generators, or scraped datasets)? + +Answer: [NA] + +Justification: This paper does not pose risks of misuse; rather, our methods help mitigate the risk of model collapse. + +Guidelines: + +- The answer NA means that the paper poses no such risks. +- Released models that have a high risk for misuse or dual-use should be released with necessary safeguards to allow for controlled use of the model, for example by requiring that users adhere to usage guidelines or restrictions to access the model or implementing safety filters. +- Datasets that have been scraped from the Internet could pose safety risks. The authors should describe how they avoided releasing unsafe images. +- We recognize that providing effective safeguards is challenging, and many papers do not require this, but we encourage authors to take this into account and make a best faith effort. + +# 12. Licenses for existing assets + +Question: Are the creators or original owners of assets (e.g., code, data, models), used in the paper, properly credited and are the license and terms of use explicitly mentioned and properly respected? + +Answer: [Yes] + +Justification: We cite all datasets, code, and models used, and comply with their licenses and usage terms. + +Guidelines: + +- The answer NA means that the paper does not use existing assets. +- The authors should cite the original paper that produced the code package or dataset. +- The authors should state which version of the asset is used and, if possible, include a URL. +- The name of the license (e.g., CC-BY 4.0) should be included for each asset. +- For scraped data from a particular source (e.g., website), the copyright and terms of service of that source should be provided. + +- If assets are released, the license, copyright information, and terms of use in the package should be provided. For popular datasets, paperswithcode.com/datasets has curated licenses for some datasets. Their licensing guide can help determine the license of a dataset. +- For existing datasets that are re-packaged, both the original license and the license of the derived asset (if it has changed) should be provided. +- If this information is not available online, the authors are encouraged to reach out to the asset's creators. + +# 13. New assets + +Question: Are new assets introduced in the paper well documented and is the documentation provided alongside the assets? + +Answer: [NA] + +Justification: This paper does not release new assets. + +Guidelines: + +- The answer NA means that the paper does not release new assets. +- Researchers should communicate the details of the dataset/code/model as part of their submissions via structured templates. This includes details about training, license, limitations, etc. +- The paper should discuss whether and how consent was obtained from people whose asset is used. +- At submission time, remember to anonymize your assets (if applicable). You can either create an anonymized URL or include an anonymized zip file. + +# 14. Crowdsourcing and research with human subjects + +Question: For crowdsourcing experiments and research with human subjects, does the paper include the full text of instructions given to participants and screenshots, if applicable, as well as details about compensation (if any)? + +Answer: [NA] + +Justification: This paper does not involve crowdsourcing nor research with human subjects. + +Guidelines: + +- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. +- Including this information in the supplemental material is fine, but if the main contribution of the paper involves human subjects, then as much detail as possible should be included in the main paper. +- According to the NeurIPS Code of Ethics, workers involved in data collection, curation, or other labor should be paid at least the minimum wage in the country of the data collector. + +# 15. Institutional review board (IRB) approvals or equivalent for research with human subjects + +Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals (or an equivalent approval/review based on the requirements of your country or institution) were obtained? + +Answer: [NA] + +Justification: This paper does not involve crowdsourcing nor research with human subjects. + +Guidelines: + +- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. +- Depending on the country in which research is conducted, IRB approval (or equivalent) may be required for any human subjects research. If you obtained IRB approval, you should clearly state this in the paper. + +We recognize that the procedures for this may vary significantly between institutions and locations, and we expect authors to adhere to the NeurIPS Code of Ethics and the guidelines for their institution. +- For initial submissions, do not include any information that would break anonymity (if applicable), such as the institution conducting the review. + +# 16. Declaration of LLM usage + +Question: Does the paper describe the usage of LLMs if it is an important, original, or non-standard component of the core methods in this research? Note that if the LLM is used only for writing, editing, or formatting purposes and does not impact the core methodology, scientific rigorousness, or originality of the research, declaration is not required. + +Answer: [NA] + +Justification: The core method development in this research does not involve LLMs as any important, original, or non-standard components. + +Guidelines: + +- The answer NA means that the core method development in this research does not involve LLMs as any important, original, or non-standard components. +- Please refer to our LLM policy (https://neurips.cc/Conferences/2025/LLM) for what should or should not be described. + +# A Related Work + +# A.1 Model Collapse + +As state-of-the-art generative models continue to improve in image quality, AI-generated images have become increasingly indistinguishable from real ones and are inevitably incorporated into the training datasets of next-generation models. In fact, [11] have shown that the LAION-5B dataset [45], which is used to train Stable Diffusion, indeed contains synthetic data. Unfortunately, recent studies [1, 4, 7, 11, 19–21] demonstrate that model performance deteriorates under such recursive training, potentially leading to model collapse, where the model generates increasingly homogeneous and meaningless content. The concept of model collapse was first introduced in [3], which also provides a theoretical framework based on Gaussian models. Their analysis shows that if a Gaussian model is recursively estimated using data generated by its predecessor, its variance converges to zero—collapsing the distribution into a delta function concentrated at a single point. + +Following this important line of work, substantial research has further explored the model collapse phenomenon. [6] extend the theoretical analysis from Gaussian to Bernoulli and Poisson models. [7] study recursive training under high-dimensional linear and ridge regression settings, providing a linear error rate and proposing optimal regularization to mitigate collapse. [20] argue that the conventional scaling laws for foundation models break down when synthetic data is incorporated into the training set. [46] demonstrate that model biases are amplified through recursive generation and proposes algorithmic reparation techniques to eliminate such biases and negative semantic shifts. [47] introduce a statistical model that characterizes the collapse process in language models and theoretically estimates the maximum allowable proportion of synthetic data before collapse occurs, validated empirically. [48] show that human preferences can be amplified in an iterative loop: modeling human curation as a reward process, the curated distribution $p_t$ converges to $p^*$ that maximizes the expected reward as $t \to \infty$ . In the work of [48], the reward $r(x)$ is a pointwise function over individual images, whereas our method can be viewed as a curation strategy that maximizes dataset-level entropy at each iteration. Whether the theory in [48] extends to dataset-wise rewards $r(S)$ , such as dataset entropy, remains an open question. Nonetheless, their intuition aligns with our results that entropy can be iteratively enhanced compared to vanilla methods. [49] provide the first generalization error bounds for Self-Consuming Training Loops (STLs), showing that mitigating model collapse requires maintaining a non-negligible portion of real data and ensuring recursive stability. In contrast, our work focuses on a specific collapse phenomenon—the generalization-to-memorization transition—and introduces an entropy-based data selection algorithm to mitigate this behavior. [50] provide a rigorous theoretical analysis for the model collapse in rectified flow models and then propose methods to mitigate it. Finally, [51] investigate collapse in diffusion models by analyzing the training dynamics of a two-layer autoencoder optimized via stochastic gradient descent, showing how network architecture shapes the model collapse. + +To mitigate model collapse, several strategies have been proposed in prior work. One common approach is to accumulate all previously generated samples along with real data into the training set—referred to as the accumulate paradigm in our paper. This strategy has been both empirically and theoretically validated. For example, [8] show that the accumulate paradigm prevents divergence of test error under a linear regression setup. [22] further establish the universality of the error upper bound across a broad class of canonical statistical models. In a related setting, [4] study an accumulate-subsample variant, confirms that the test error plateaus and examines interactions between real and synthetic data. [5] provide theoretical guarantees for stability in iterative training, assuming the initial model trained on real data is sufficiently accurate and the clean data proportion in each iteration remains high. Despite these findings, recent work [9] present a robust negative result, showing that model collapse persists even when real and synthetic data are mixed. Similarly, [11] argue that the accumulate paradigm merely delays, rather than prevents, collapse. Introducing fresh real data in each iteration may be necessary for long-term robustness. Other complementary approaches include verification [21] and re-editing [52]. For instance, [21] theoretically underscore the importance of data selection, though it does not propose a specific method. [52] target language models and introduces a token-level editing mechanism with theoretical guarantees. Compared to prior work, this paper proposes a novel entropy-based data selection method for diffusion models that improves both generalizability and image quality, thereby mitigating model collapse. + +Over the past few years, extensive research has explored model collapse from various perspectives and dimensions. An important study [2] makes a thorough survey about different definitions and + +patterns investigated in previous work. A prominent line of work [3-6] focuses on the phenomenon of variance collapse, both empirically and theoretically demonstrating that models progressively lose information in the distributions tail, with variance tending toward zero. Another series of studies [5, 7-10] investigates model collapse through the lens of population risk and distributional shift, observing that the generated distribution increasingly diverges from the true data distribution, leading to rising population risk across recursive training cycles. Moreover, several works [11-13] report that models begin to generate hallucinated or unrealistic data as collapse progresses. [20] further suggest that the inclusion of synthetic data alters the scaling laws of model performance. In addition, [51] study the progressive mode collapse [53] in diffusion models, where the number of modes in the generated distribution gradually decreases. In this paper, we introduce a novel perspective for analyzing model collapse in diffusion models by uncovering a generalization-to-memorization transition. We show that this transition is closely tied to the entropy of the training dataset, which serves as a key indicator of model generalizability. Our findings further motivate the development of entropy-based data selection strategies that effectively mitigate model collapse. + +# A.2 Generalization and Memorization + +Recent studies [15, 24] have identified two distinct learning regimes in diffusion models, depending on the size of the training dataset and the model's capacity: (1) Memorization regime, when models with sufficient capacity are trained on limited datasets, they tend to memorize the training data; and (2) Generalization regime, as the number of training samples increases, the model begins to approximate the underlying data distribution and generate novel samples. To investigate the transition between these regimes, [54] show that the number of training samples required for the transition from memorization to generalization scales linearly with the intrinsic dimension of the dataset. In addition, the analysis of training and generation accuracies in [55] provides a potential step toward quantifying generalization. [56] propose a theoretically grounded and computationally efficient metric, Probability Flow Distance (PFD), to measure the generalization ability of diffusion models. Specifically, PFD quantifies the distance between distributions by comparing their noise-to-data mappings induced by the probability flow ODE. Meanwhile, concurrent work also explores memorization and generalization separately. To understand generalization, studies such as [57, 58] attribute the generalization to implicit bias introduced by network architectures. Other works study the generalized distribution using Gaussian models [59, 60] and patch-wise optimal score functions [61, 62]. As for memorization, it is investigated in both unconditional and conditional [63], as well as the text-to-image diffusion models [64, 65]. Additionally, methods to mitigate memorization in diffusion models have been proposed in [66, 67]. Distinct from prior work, our study is the first to establish a connection between model collapse and the transition from generalization to memorization. This connection not only offers a novel perspective to understand model collapse but also provides insights to mitigate it by mitigating memorization. + +# A.3 Data Selection + +[68-70] focus on data pruning techniques, but not necessarily in the context of self-consuming loops. While both their approaches and ours demonstrate the benefits of data selection, there are several key differences: + +- Objective and evaluation: [68-70] primarily study how pruning improves scaling laws, achieving higher accuracy as dataset size varies. In contrast, our work examines how model performance evolves over an iterative training loop with a fixed dataset size, focusing on the generalization-to-memorization transition. +- Task and criteria: Prior works focus on classification tasks, where sample selection depends on label information. Our method targets generative modeling, where selection is based on dataset entropy rather than label-driven criteria. Our objectives also differ: prior work emphasizes classification accuracy, while we address model collapse from a generalization perspective, providing a different angle of analysis. +- Pruning strategy: Methods in [68-70] largely rely on per-sample evaluation, whereas our approach considers global dataset structure and relationships between samples. While this may lead to increased computational complexity, it opens up new possibilities for designing pruning criteria beyond per-sample evaluation. + +- Entropy definition: In [70], while the authors use a generative model to sample images, the goal is to improve the classification performance of a classifier. And the entropy used in their method is defined in the prediction space of a classifier, which is different from the entropy of a dataset measured in our work. + +# B Experimental Details + +# B.1 Network Structure + +For CIFAR-10 and FFHQ, we use a UNet-based backbone, taking RGB images as inputs and predicting noise residuals. Our implementation is based on the Hugging Face Diffusers base code [34]. The architecture hyperparameters of the neural network are listed as follows: + +- The numbers of in-channel and out-channel are 3. +- The number of groups for group normalization within Transformer blocks is 16. +- The number of layers per block is 2. +- The network contains 6 down-sampling blocks and 6 up-sampling blocks. +- The numbers of feature channels for the 6 blocks are 48, 48, 96, 96, 144, 144 respectively. + +For MNIST, we use a similar UNet-based backbone, taking single-channel images as inputs and predicting noise residuals. The architecture hyperparameters of the network are listed as follows: + +- The numbers of in-channel and out-channel are 1. +- The number of groups for group normalization within Transformer blocks is 32. +- The number of layers per block is 2. +- The network contains 4 down-sampling blocks and 4 up-sampling blocks. +- The numbers of feature channels for the 4 blocks are 64, 128, 256, 512 respectively. + +# B.2 Implementation Details + +In this paper, we use DDPM as our generative method. For efficiency, we use the FP-16 mixing precision to train our models, which is inherently implemented by the Hugging Face Diffusers codebase. The batch size of training all datasets is set to be 128. A 1000-step denoising process is used as the reverse process. For CIFAR-10, the epoch number is set to be 500; for FFHQ, the epoch number is set to be 1000. We use the Adam optimizer with a learning rate of $10^{-4}$ and a weight decay of $10^{-6}$ . Other experimental hyperparameters are exactly the default values in the original Hugging Face Diffusers codebase. We use the DINOv2 model [29] to extract features of images and then calculate the distance between two images in the feature space. We use the InceptionV3 model [71] to extract features of images to calculate the FID score. All experiments are conducted on a single NVIDIA A-100 GPU. + +# B.3 Pseudo-codes for the Algorithms + +We present the pseudo-code of the Greedy Selection and Threshold Decay Filter methods in Algorithms 1 and 2. + +# C Ablation Study + +# C.1 Training on More Samples + +In Section 5 of the main paper, we augment the vanilla replace paradigm with our data selection methods. Specifically, under the replace setting, the vanilla baseline generates $N$ images and trains the next-iteration model based solely on these $N$ images from the previous iteration. Instead, the data selection methods generate $2N$ images at each iteration, and select a subset of $N$ images from the $2N$ images for training the next-iteration model. One nature question is: what if we also generate $2N$ images and then train the next-iteration model on the entire $2N$ -image dataset without filtering. + +Algorithm 1 Greedy Selection +Require: Dataset $S$ target size $N$ distance function $\kappa (\cdot ,\cdot)$ +Ensure: Selected subset $\mathcal{D}$ of size $N$ +1: Initialize $\mathcal{D}\gets \{\mathrm{random~point~from} \mathcal{S}\}$ +2: while $|\mathcal{D}| < N$ do +3: for all $x\in S\setminus \mathcal{D}$ do +4: Compute $d(x)\gets \min_{y\in \mathcal{D}}\kappa (x,y)$ +5: end for +6: $x_{\mathrm{select}}\gets \arg \max_{x\in \mathcal{S}\setminus \mathcal{D}}d(x)$ +7: $\mathcal{D}\gets \mathcal{D}\cup \{x_{\mathrm{select}}\}$ +8: end while +9: return $\mathcal{D}$ + +Algorithm 2 Threshold Decay Filter +Require: Dataset $S$ target size $N$ initial threshold $\tau >0$ decay factor $\alpha \in (0,1)$ distance function $\kappa (\cdot ,\cdot)$ +Ensure: Selected subset $\mathcal{D}$ of size $N$ +1: Initialize $\mathcal{D}\gets \{\mathrm{random~point~from} \mathcal{S}\}$ +2: while $|\mathcal{D}| < N$ do +3: added $\leftarrow$ false +4: for all $x\in S\setminus \mathcal{D}$ do +5: Compute $d_{\mathrm{min}}(x)\gets \mathrm{min}_{y\in \mathcal{D}}\kappa (x,y)$ +6: if $d_{\mathrm{min}}(x) > \tau$ then +7: $\mathcal{D}\gets \mathcal{D}\cup \{x\}$ +8: added $\leftarrow$ true +9: end if +10: if $|\mathcal{D}| = N$ then +11: return $\mathcal{D}$ +12: end if +13: end for +14: if not added then +15: $\tau \leftarrow \alpha \cdot \tau$ ▷No point added $\Rightarrow$ decay threshold +16: end if +17: end while +18: return $\mathcal{D}$ + +Next, we show that when training on the whole $2N$ dataset, the performance (FID score) of the model is between the vanilla replace setting (generating $N$ images and training on those $N$ images) and the data selection method (generating $2N$ images and training on the selected $N$ -images subset). The results on CIFAR-10 are shown in Figure 10. + +Several conclusions can be drawn by comparing the results across these settings. + +- Incorporating more data into the training-sampling recursion can indeed mitigate the rate of model collapse. Compared to the vanilla replacement paradigm (i.e., the first line), using $2N$ images (second line) yields improved performance. This aligns with prior findings [4, 10] that sample size is a key factor influencing the collapse rate. +- Further augmenting the training data with our selection method leads to even better performance than training on the full set of $2N$ images. The results validate the effectiveness of our selection method, achieving a better FID performance while largely decreasing the training budget. + +In fact, there is a trade-off for the filter ratio. There are two clear extremes: + +- If the ratio is 1, all 2N generated images are used for training. As we show in Figure 10, it still degrades faster than filtering (ratio equals to $1/2$ ) + +![](images/5f95ce54c9b21d50c0863a7eb3b7eee07a6dca881c7564a934ae753762366289.jpg) +Figure 10: FID score of the generated images over iterations. We add an additional setting of generating $2N$ images and training the next model on the entire $2N$ -images data without filtering. + +
Ratio0.20.40.60.81.0
FID56.044.740.038.249.5
+ +Table 1: FID comparison of the models. We use various filter ratios to get the training subsets for those models. + +- Conversely, if the ratio approaches 0, too few images are selected for training, which detrimentally starves the model of data. + +We then use different filter ratios to get training subsets from the 2N generated images and then train models on those filtered datasets. The results in Table 1 show that an intermediate ratio yields the best FID performance. + +# C.2 Different Decay Rates + +The decay rate is one important hyperparameter for the Threshold Decay Filter. In this section, we use different decay rates and show that the filter is robust to a wide range of decay rates. + +Figure 11 shows the FID scores of generated CIFAR-10 images across iterations for different methods. As shown in the figure, the data selection methods consistently outperform the vanilla accumulate paradigm, indicating strong robustness of the hyperparameter. + +# D Additional Results + +# D.1 The Generalization-to-Memorization Collapse on Accumulate Paradigm + +This subsection presents additional results on the accumulate paradigm and FFHQ dataset as a supplement to the explorations in Section 3. Specifically, we show in Figure 12 that the generalization-to-memorization collapse also occurs on the accumulation paradigm. We note that model collapse has different definitions in previous studies. Here, we clarify that our results show the generalization score consistently decreases over early iterations. However, we cannot determine whether the score eventually collapses to zero or converges to a lower bound in the accumulate paradigm, as we only have fewer than 10 iterations—insufficient to draw conclusions about convergence. + +![](images/26339603022e21f256e64d3c50969198ae855e38a9411a4511b14f67a2a1711c.jpg) + +![](images/8405def9a8601088d136c35397c164fd9eb47200ce0d1d0702de304f1ac96777.jpg) +Figure 11: FID score of the generated images over iterations. We compare the results for different decay rates on CIFAR-10 dataset. + +![](images/9c54f3912437f55482f58df63828fc4052a371b6bc3baca66bdc9964d487f262.jpg) +Figure 12: The generalization-to-memorization transition on the accumulate paradigm. Similar to Figure 2, we visualize the generated images and their nearest neighbors in the training dataset. As the iteration proceeds, the model loses generalization ability, collapsing into the memorization regime. + +# D.2 The Robust Relationship between Generalization Score and Entropy + +In Figure 2, we plot the results on CIFAR-10 datasets, demonstrating the dataset size-independent relationship between generalization score and estimated entropy. This subsection further incorporates the results from FFHQ and the accumulate paradigm to show a more general relation. + +Specifically, we conduct the self-consuming loop for FFHQ with dataset sizes of 8, 192, 16, 384, and 32, 768. Besides, we also include the results from Appendix D.1. Combining those, we show a more robust relationship in Figure 13, where all the points align around the red dashed line. This result suggests that the entropy of the training dataset is a dataset-independent indicator for the generalization score. + +# D.3 Generated Images over Iterations + +This subsection provides additional generated images of the trained model across iterations and dataset sizes in a grid format. These images are generated from the vanilla replace paradigm. + +![](images/7a73f6ef9aafb7a84abe2c1efd24122f60ba4bdb256760747d904ccbbe8c821a.jpg) +Figure 13: Scatter plots of the generalization score and estimated entropy of the training data. Each point denotes one iteration of training in the self-consuming loop. In addition to the results in Figure 4a, we include more results from FFHQ and on the accumulate paradigm. We use different colors to denote the loops, different shapes (circles and triangles) to represent CIFAR and FFHQ, and solid versus hollow markers to distinguish the replace and accumulate paradigms. + +![](images/2ce3a6a7cfe0454f9ebf61e750351f19345890907e5a8c055598dbf31bd5519b.jpg) +Figure 14: Generated images of models trained on 1024 CIFAR-10 images over iterations. + +![](images/8d515e74772c0943116cb286b3f5a9c9f8e340bb0f9eeddd8ea4f272d92869ca.jpg) + +![](images/397524d9018e816c2e400040debfeb166f97f2af5954b66560ae7036eb4ed51b.jpg) + +Figures 14 to 19 present the generated images of the model trained on CIFAR-10 with various dataset sizes across successive iterations. As shown in Figures 14 and 15, the diffusion models tend to memorize small training datasets throughout the recursive process, exactly copying training images during generation. Because the generated samples are nearly identical to the training data, image quality does not noticeably degrade. However, duplicated generations emerge in later iterations, and diversity declines as the model gradually loses coverage over parts of the original images. + +On the contrary, on the large dataset of 32,768 images, the models generate novel images in the beginning. As the model collapses, the quality and diversity of the images gradually degrade over iterations. + +![](images/6b622835745d555c5fb53a2250b8be6efa8fac9d783cf92fe5eace45137d1cf2.jpg) +Figure 15: Generated images of models trained on 2048 CIFAR-10 images over iterations. + +![](images/0b946e12dcf17fd406c47061dac7fbfa5d8c9bb4bb3d49d93a20466d65197d96.jpg) + +![](images/ab8b923735f92350d0419e8b98425fce05232594f7dc917db83b2da24eab3ab9.jpg) + +# D.4 Generated Image Distribution Becomes Spiky + +In the main paper, we adopt the KL estimator of Equation (2) to evaluate the entropy given a finite image dataset. With a fixed dataset size, the only dataset-dependent term in Equation (2) is the sum of nearest-neighbor distances $\varepsilon(\boldsymbol{x})$ indicating that samples in $\mathcal{D}$ become increasingly concentrated. This suggests the distribution is becoming spiky. Figure 3 in the main paper also illustrates this trend by projecting high-dimensional images onto the subspace spanned by their top two eigenvectors. The visualization reveals that the generated images form numerous local clusters, while the overall support of the distribution remains relatively stable. + +To quantitatively measure the spiky degree of the empirical image distribution, we further adopt the Mean Nearest Neighbor Distance (MNND) [31, 72], which removes the data-independent constants and logarithm in the KL estimator: + +$$ +\operatorname {M N N D} \left(\mathcal {D} _ {t}\right) \triangleq \operatorname {D i s t} \left(\mathcal {D} _ {t}, \mathcal {D} _ {t}\right) = \frac {1}{\left| \mathcal {D} _ {t} \right|} \sum_ {\boldsymbol {x} \in \mathcal {D} _ {t}} \min _ {\boldsymbol {z} \in \mathcal {D} _ {t} \backslash \boldsymbol {x}} d (\boldsymbol {x}, \boldsymbol {z}). \tag {3} +$$ + +A lower MNND suggests a tightly clustered and spiky distribution, while a higher distance suggests dispersion and diversity. The KL estimator can be related to MNND through + +$$ +e ^ {\frac {\hat {H} _ {1} \left(\mathcal {D} _ {t}\right) - B}{d}} \leq \operatorname {M N N D} \left(\mathcal {D} _ {t}\right), \tag {4} +$$ + +where $\hat{H}_1(\mathcal{D}_t)$ represents the estimated entropy of $\mathcal{D}_t$ with $k = 1$ and $B$ is a constant offset given a fixed dataset size. + +We want to clarify the nuance between MNND and variance. A spiky distribution does not imply that all data points are concentrated in a single small region—this behavior has already been illustrated by the variance collapse behavior [3-5]. Specifically, Figure 20 shows that the variance of the generated dataset only slightly decreases along the successive iterations and is far from complete collapse. On the contrary, MNND decreases almost exponentially and reaches a small value after 10 iterations. Thus, the results align with the prior claim in [2, 4] that the collapse of variance progresses at such a slow pace that it is rarely a practical concern in real-world applications. In contrast, the collapse in MNND emerges at an early stage, highlighting a critical memorization issue caused by training on synthetic data. + +![](images/7a8fc0ebce1148b6ea96777f6ceda0e65c056cec8460e0deb089a57b8a4fad74.jpg) +Iteration 1 + +![](images/715f6742103e3892a4bdf6e5a0b385cd0727300d2a04187169fefa271d16589f.jpg) +Iteration 2 + +![](images/bfc918e9c3bddd427679df7cc01a731079518d2ecb7dfaad702266dde8875368.jpg) +Iteration 3 + +![](images/a99bdb8170eff8847a7b1d3609d72d1cefac5c74540b05e382d9b7319e26d8a0.jpg) +Iteration 4 +Figure 16: Generated images of models trained on 4096 CIFAR-10 images over iterations. + +![](images/909157fc7e2aeb177ee01460283d9d51e39adf2fed0ec14885e91964888a02f3.jpg) +Iteration 5 + +![](images/881107763d918aabf5fc40d3661650de6d69d7332e97f36c86959f20cab0f89c.jpg) +Iteration 6 + +
ModelABC
FID28.030.827.5
+ +Table 2: FID comparison of three models. + +# D.5 Self-consuming Loop with Fresh New Data + +[11] show that incorporating fresh real data can further mitigate model collapse. In this section, we conduct experiments to verify the effectiveness of our methods in this paradigm. The candidate data pool in each iteration is jointly composed of three sources: (1) the original real data, (2) synthetic data generated by the previous model, and (3) fresh real data that was not used in earlier iterations. This experiment is designed to simulate real-world deployment, where generative models are continuously updated with a mixture of prior synthetic data and incoming fresh real data, and the training budget also increases. Concretely, we first train Model A on 32,768 real images, which then generates 10,000 synthetic images. We construct a data poolsimulating an Internet-scale sourceby combining the 32,768 original real images, 10,000 synthetic images, and 10,000 additional fresh real images. From this pool, our entropy-based selection method chooses 40,000 images as the training dataset for Model C. For comparison, Model B is trained on 40,000 randomly selected images from the same pool. + +The FID scores of Models A, B, and C are summarized in Table 2. As the results indicate, our entropy-based selection method (Model C) enables the model to outperform the baseline (Model A), and far better than random sampling (Model B). This demonstrates that, in a practical scenario with evolving datasets, our approach not only mitigates model collapse but also delivers tangible performance gains. + +# D.6 Additional Metric + +Since our Greedy Section method is performed on the feature space extracted by the DINO model, we also use $\mathrm{FD}_{\mathrm{DINO}}$ [73] as an alternate for FID to measure the quality of the generated images. As + +![](images/e445ead3fd1995b499224e28b20a99a7b184c2da5efffc09bcf8dc1176d047ba.jpg) +Figure 17: Generated images of models trained on 8192 CIFAR-10 images over iterations. + +![](images/8f0bebb14d1dc48d7786960ffe773e1bf6b700e1a6f688e4dc40364528060056.jpg) + +![](images/e038be0f66c586d8f63d9d3353e7a49092bea25019486dca27a12cd0755b7467.jpg) + +we show in Figure 21, the Greedy Selection method can also improve the $\mathrm{FD_{DINO}}$ metric compared to the vanilla paradigms. + +# D.7 Additional Results for Section 5 + +We provide visualization of the generated FFHQ images with their nearest training neighbors to show the model's generalizability. Figure 22 shows some images for different training paradigms at 5-th iteration in grid format. With augmentation from the Greedy Selection method, the model generates images that deviate more from the training set compared to the vanilla accumulate paradigm, thereby enhancing its generalization ability. + +# E Impact Statement + +In this work, we investigate a critical failure mode of diffusion models known as model collapse, which occurs when models are recursively trained on synthetic data and gradually lose their generalization ability and generative diversity. As AI-generated data is unintentionally or deliberately incorporated into the training sets of next-generation models, understanding and mitigating model collapse is essential for ensuring long-term model reliability and performance. Our study identifies the generalization-to-memorization transition, demonstrates the relation between entropy of the training set and the generalizability of the trained model, and proposes practical solutions to mitigate model collapse through entropy-based data selection. + +We believe our findings will contribute to the responsible development and deployment of generative models, especially in scenarios where data source may be mixed or partially synthetic. While techniques for analyzing collapse may be misused to intentionally degrade generative models through poisoning attack, our intent is solely to build more robust, transparent, and self-aware AI systems. We encourage researchers in generative AI to use these results to mitigate model collapse and to build reliable models, even when training data contains AI-generated samples—a scenario that may become increasingly common in the future. + +![](images/9f384670aa9d5fb333cc2fac338fd337881d38efea545c1a0d2debcc772dcc76.jpg) +Iteration 1 + +![](images/b87c95565438b74bde5fea2b9e25f5dee6cebdec9ea4a33579578d31583506d5.jpg) +Iteration 2 + +![](images/bd6dabb2d54ce6c42b5ceadc540465299116e45fa217821f1ba715733905c490.jpg) +Iteration 3 + +![](images/c4566acad5de9a262769c471aef51b8aa7b8d657cbafb92cf857f5d317126736.jpg) +Iteration 4 +Figure 18: Generated images of models trained on 16384 CIFAR-10 images over iterations. + +![](images/ea6db226e3cb167ac737f5736e33ac9fc63d08c1a74585fa3e5b588336c11570.jpg) +Iteration 5 + +![](images/9cb1cb842b11dcaa53f420f47ca18477014a9f89c26b8994e1a4f4f2cd5651c8.jpg) +Iteration 6 + +![](images/a2a613dd5d2177ae5638dcb171be5405fa31a5ef4a6d44451687bd1af9b6cdf9.jpg) +Iteration 1 + +![](images/6246f5ffd7a6dbbf8297d58a171def6dbd4a4d5dad029541bf360295da812434.jpg) +Iteration 2 + +![](images/172a5bbee357f013a94219d6d03d411e10aabeafb6391616620e503509e377a4.jpg) +Iteration 3 + +![](images/5816545ca2dc242ea8fae4af12dcaf0567854f6e6948bb87c5174f1536c07c48.jpg) +Iteration 4 +Figure 19: Generated images of models trained on 32768 CIFAR-10 images over iterations. + +![](images/1f2d1e7601e660a1b3edff5b76fed451b9939adc229e85f84ca86e4471a1e5fd.jpg) +Iteration 5 + +![](images/0c08c48c757a46d349522e5258ac48dc064498d4f2b02e15312e5d4c775cdc85.jpg) +Iteration 6 + +![](images/57919c7d410bbff1f37c7132a502ca21fabb7cca28644d83a2d8009b987c8f47.jpg) +Figure 20: The MNND and the trace of the covariance matrix over iterations. + +![](images/91cd40f10970f73af4c178d06c36f252f2b255540095708681ef72823753a73a.jpg) + +![](images/233c228c2217822a6a03b0f227ae9aa8f25694c1bcde4333bdfd426472b0bea8.jpg) +Figure 21: $\mathrm{FD}_{\mathrm{DINO}}$ comparison of vanilla paradigms and our methods. + +![](images/48d292d2b02809094ecac34ed449ddd023af0e2f68feeb54e420849685a2017b.jpg) +Vanilla Accumulate Generalization score $= 29.6$ + +![](images/aaab8739a834df3854324edb13f47e983fde85a0cad9c31cd3c1e3dbf7e0fd4d.jpg) +With Greedy Selection (Generalization score $= 37.0$ +Figure 22: The visualization of the generated images and their nearest neighbors in the training dataset. Each pair of rows corresponds to one group: the top row shows the generated images, and the bottom row shows their nearest training images. \ No newline at end of file diff --git a/NeurIPS/2025/A Closer Look at Model Collapse_ From a Generalization-to-Memorization Perspective/images.zip b/NeurIPS/2025/A Closer Look at Model Collapse_ From a Generalization-to-Memorization Perspective/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..57917df424d8d711e7f22bcc62a2dfadd2ae9c18 --- /dev/null +++ b/NeurIPS/2025/A Closer Look at Model Collapse_ From a Generalization-to-Memorization Perspective/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:87a415ced81ff58658853259d232a578ca48ae7e894e5d828cfb33ed07a73960 +size 1751967 diff --git a/NeurIPS/2025/A Closer Look at Model Collapse_ From a Generalization-to-Memorization Perspective/layout.json b/NeurIPS/2025/A Closer Look at Model Collapse_ From a Generalization-to-Memorization Perspective/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..e01d8f4ccd0b0f3fd6e7564b71e357c2427e6138 --- /dev/null +++ b/NeurIPS/2025/A Closer Look at Model Collapse_ From a Generalization-to-Memorization Perspective/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4eafa01641734ebe542c8c37da1e3a0ddd6afc3a661e3fc3a5c340ddc94bd37d +size 1009005 diff --git a/NeurIPS/2025/A Closer Look at NTK Alignment_ Linking Phase Transitions in Deep Image Regression/18650913-3f43-4cf7-a499-d44c5588e8fd_content_list.json b/NeurIPS/2025/A Closer Look at NTK Alignment_ Linking Phase Transitions in Deep Image Regression/18650913-3f43-4cf7-a499-d44c5588e8fd_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..ea4d73bc90225318e6317325f3fd7f4552a917f1 --- /dev/null +++ b/NeurIPS/2025/A Closer Look at NTK Alignment_ Linking Phase Transitions in Deep Image Regression/18650913-3f43-4cf7-a499-d44c5588e8fd_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:296118d302bf0b55abc3fb91feba649d9bb32be052c475db3a48d2194d2d9399 +size 255188 diff --git a/NeurIPS/2025/A Closer Look at NTK Alignment_ Linking Phase Transitions in Deep Image Regression/18650913-3f43-4cf7-a499-d44c5588e8fd_model.json b/NeurIPS/2025/A Closer Look at NTK Alignment_ Linking Phase Transitions in Deep Image Regression/18650913-3f43-4cf7-a499-d44c5588e8fd_model.json new file mode 100644 index 0000000000000000000000000000000000000000..ab8d7ec9a2ae0225206aca97a76b398b0157deef --- /dev/null +++ b/NeurIPS/2025/A Closer Look at NTK Alignment_ Linking Phase Transitions in Deep Image Regression/18650913-3f43-4cf7-a499-d44c5588e8fd_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:41b1c6ad3b79ba59a5b83bd74888757dd5c02da71de4443b3047da49670d8c0c +size 306833 diff --git a/NeurIPS/2025/A Closer Look at NTK Alignment_ Linking Phase Transitions in Deep Image Regression/18650913-3f43-4cf7-a499-d44c5588e8fd_origin.pdf b/NeurIPS/2025/A Closer Look at NTK Alignment_ Linking Phase Transitions in Deep Image Regression/18650913-3f43-4cf7-a499-d44c5588e8fd_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..cf7e2ed8b51378a6d16c52516c65944b21f2b1c5 --- /dev/null +++ b/NeurIPS/2025/A Closer Look at NTK Alignment_ Linking Phase Transitions in Deep Image Regression/18650913-3f43-4cf7-a499-d44c5588e8fd_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:57c06a9dba3b2608ca46946fd08ee97e9410d37afaa755befdb317ed11aedc27 +size 3355557 diff --git a/NeurIPS/2025/A Closer Look at NTK Alignment_ Linking Phase Transitions in Deep Image Regression/full.md b/NeurIPS/2025/A Closer Look at NTK Alignment_ Linking Phase Transitions in Deep Image Regression/full.md new file mode 100644 index 0000000000000000000000000000000000000000..30bf8c67ff066c88deaee5dbd3c645208811e478 --- /dev/null +++ b/NeurIPS/2025/A Closer Look at NTK Alignment_ Linking Phase Transitions in Deep Image Regression/full.md @@ -0,0 +1,1422 @@ +# A Closer Look at NTK Alignment: Linking Phase Transitions in Deep Image Regression + +Giuseppe Castiglione +ool of Engineering and Informatics +University of Sussex +.castiglione@sussex.ac.uk + +Christopher Buckley +School of Engineering and Informatics +University of Sussex +c.l.buckley@sussex.ac.uk + +Ivor Simpson +School of Engineering and Informatics +University of Sussex +i.simpson@sussex.ac.uk + +# Abstract + +Deep neural networks trained with gradient descent exhibit varying rates of learning for different patterns. However, the complexity of fitting models to data makes direct elucidation of the dynamics of learned patterns challenging. To circumvent this, many works have opted to characterize phases of learning through summary statistics known as order parameters. In this work, we propose a unifying framework for constructing order parameters based on the Neural Tangent Kernel (NTK), in which the relationship with the data set is more transparent. In particular, we derive a local approximation of the NTK for a class of deep regression models (SIRENs) trained to reconstruct natural images. In so doing, we analytically connect three seemingly distinct phase transitions: the emergence of wave patterns in residuals (a novel observation), loss rate collapse, and NTK alignment. Our results provide a dynamical perspective on the observed biases of SIRENs, and deep image regression models more generally. + +# 1 Introduction + +Classical learning theory suggests that models with sufficient capacity - specifically, one whose parameters outnumber the training samples - tend to "memorise" individual examples rather than learn underlying patterns, leading to poor generalisation [1]. However, while Deep Neural Networks (DNNs) are typically over-parameterised, a growing body of research highlights the role of Gradient Descent (GD) in constraining their effective capacity [2, 3]. A recurring observation is that GD biases neural networks to prioritise learning simple patterns before more complex ones, resulting in distinct phases of learning [4, 5]. These phases are characterised by changes in the collective evolution of the network's weights, which can be quantified by statistics known as order parameters [6, 7, 8, 9]. Although numerous authors have independently proposed statistics to account for changes in convergence rate [10, 11] - and correspondingly, the memorisation [12] and over-fitting [13] of complex/noisy patterns - their interrelationships remain under-explored. More significantly, these existing approaches provide limited insight into the actual content being learned during each phase - and consequently, which patterns models systematically struggle to learn. Addressing these gaps is essential to developing a unified understanding of learning dynamics in DNNs. + +A major obstacle in understanding this inductive bias lies in the inherent complexity of GD itself. While conceptually GD can be viewed as a function mapping from the dataset, hyperparameters, and initial weights to the final learned weights, in practice, the thousands of iterations through high- + +dimensional parameter space obscure the relationship between order parameters and the underlying dataset characteristics. In recent years, the Neural Tangent Kernel (NTK) [14] has emerged as an alternative perspective on the dynamics of learning, recasting them in terms of the evolution of pointwise errors. Critically, in a phenomenon known as Neural Tangent Kernel Alignment (NTKA), the eigenspectra of the NTK undergo a sudden transition of their own, spontaneously aligning with the class structure of the dataset without direct supervision. NTKA has been widely documented and is suggested as a reason why real-world DNNs often outperform their infinite-width limit counterparts [15, 16, 17, 18, 19, 20]. However, despite repeated empirical demonstrations of NTKA, theoretical exploration of the phenomenon has been largely restricted to classification problems with toy models, such as two-layer neural networks [21, 22], and deep linear networks [22]. + +In this work, we move beyond these simple classification models and study NTKA in a considerably more complex setting: deep image regression using multi-layer SIRENs [23]. These Implicit Neural Representations (INRs) learn mappings from $\mathbb{R}^2\rightarrow \mathbb{R}$ , representing images as continuous functions, and find increasing application in tasks such as super-resolution. Despite the low input dimensionality, the depth and non-linear (sinusoidal) activations of these networks pose significant analytical challenges, exceeding the complexity of previously studied models. However, in addition to facilitating visualisation, this low-dimensionality permits us to leverage insights from computer vision to introspect the learning process. Our study is structured around three primary contributions: + +1. We derive novel approximations for the local structure of the SIREN NTK, allowing us to approximate: the principal eigenvector (3.3); order parameters such as the minimum value of the Cosine NTK (3.4); and the correlation lengthscale (3.2). In so doing, we theoretically establish connections between the onset of NTKA and other dynamical phase transitions. +2. We identify a novel learning phase in deep image regression, characterized by the appearance of diffusion-like wavecrests in the residuals, and relate this behaviour to the evolution of the NTK. +3. We experimentally verify that the critical points for these different phase transitions cluster in time. We also empirically investigate the impact of image complexity and SIREN hyperparameters on the occurrence and timing of phase transitions, and provide evidence that NTK alignment in image regression tasks occurs in response to difficulties in modelling edges. + +# 2 Preliminaries + +In this work, we consider 2D grayscale images, where pixel coordinates and their intensity form a dataset $\mathcal{D}$ of $N$ samples indexed with $i$ , $(x_{i},I(x_{i}))$ , where $x_{i}\in \mathbb{R}^{2}$ and $I:\mathbb{R}^2\mapsto \mathbb{R}$ . On this dataset, we fit SIREN models $f(x;\theta)$ of depth $N_{l}$ , defined recursively by: $h^{(0)} = x$ ; $h^{(l)} = \sin \omega_0(W^{(l)}h^{(l - 1)} + b^{(l)})$ ; $f(x;\theta) = W^{(N_l)}h^{(N_l - 1)} + b^{(N_l)}$ . Here $h^{(l)}$ denotes the output of the $l$ -th layer, $\theta = \{W^{(l)},b^{(l)}|l = 1,\dots ,N_l\}$ is the set of learnable parameters, and $\omega_0$ is a bandwidth hyperparameter. $\omega_0$ is generally chosen to ensure the sin function spans multiple periods (and thus frequencies) over the inputs. In the continuum limit, we assume the data is distributed uniformly $P_{data}(x) = \mathrm{Vol}(\mathcal{D})^{-1}$ . We identify two fields: the local residual field $r(x;\theta (t)) = I(x) - f(x;\theta (t))$ and gradient field $\nabla_{\theta}f(x;\theta (t))$ . Dynamics are induced by gradient flow $\dot{\theta} = -\nabla_{\theta}L$ on the mean square error: $L(\theta) = \frac{1}{2\mathrm{Vol}(\mathcal{D})}\int dx r(x;\theta)^2$ . Through the chain rule, the residuals evolve as follows: + +$$ +\begin{array}{l} \dot {r} (x; \theta (t)) = \nabla_ {\theta} r (x; \theta (t)) \cdot \dot {\theta} (1) \\ = - \frac {1}{\operatorname {V o l} (\mathcal {D})} \int d x ^ {\prime} r \left(x ^ {\prime}\right) \nabla_ {\theta} r (x; \theta (t)) \cdot \nabla_ {\theta} r \left(x ^ {\prime}; \theta (t)\right) (2) \\ = - \int d x ^ {\prime} r \left(x ^ {\prime}\right) \underbrace {\left(\frac {1}{\operatorname {V o l} (\mathcal {D})} \nabla_ {\theta} f (x ; \theta (t)) \cdot \nabla_ {\theta} f \left(x ^ {\prime} ; \theta (t)\right)\right)} _ {K _ {N T K} \left(x, x ^ {\prime}; \theta (t)\right)} (3) \\ \end{array} +$$ + +In the last line, we defined the NTK. Equation 3 is a linear dynamical system with a time-varying kernel. The eigenvectors $v_{k}(x,t)$ represent distinct normal modes of the dataset, each learning at a rate governed by its associated eigenvalue $\lambda_{k}(t)$ . This framework formalizes the intuitive notion that neural networks learn different patterns at different speeds. + +![](images/639354e5ebc394d98fafb94e4e4e0d8e1462e89e07666ed029585015c5fceaef.jpg) +(a) + +![](images/dc1a5e07f87438104b472f5a92e3f0191b527b52de761be2c4a39b331c9f0bc6.jpg) + +![](images/0f2b979da0ed62b351f75e71988c3e0b8eb58afa3d58294fefff1b693280e674.jpg) + +![](images/9bfc999671382b90d8e60aa366440457e8f928a56e607e2c484b09fc59166a96.jpg) + +![](images/34baeeca6581b5843a930665a2dbc18d8f84cf95847227d09dbe7c0575d9448d.jpg) + +![](images/5b3a2d9275fc8dc0d2449ecc8a4e40fe35b09ad6103a3d37a57d157e1a4a8cf2.jpg) + +![](images/b74d6874ff6bb4fc8071bfd6402c7a99d3f6dad5148cd651c0367c42f6cfc187.jpg) + +![](images/ff52382986fe0aba3d8d9350abbc6f9238fc78462d41eff4efd29ddb51dff852.jpg) + +![](images/15b1585f7faff854f6666e6d692250827f16dc5feb69ae4d03bc1682491cdfdd.jpg) + +![](images/596347c86d614a465e422c361d99d8b4d5d6ff545b2009e9ac3a140f3594ee1b.jpg) + +![](images/28753bfeb355eba5bcbd6d30a566efc9647e95926c9b57138ad8b43e7a763e0e.jpg) + +![](images/f20f55719937c1d8c86c206bb8bca112d43dffb2f24a49e008c8ced8e7d31678.jpg) +Figure 1: A Single Phase Transition Through Three Lenses: (a) The magnitude of the residuals over the training process. Near the critical point we see the formation of wavecrests. (b) Evolution of evaluation loss rate during training, which reaches a peak at the critical point. (c) Evolution of the principle eigenvector of the NTK, which reveals a sudden shift from disorder to structure. (d) Quantification of NTKA in terms of alignment between edges and the principal eigenvector. + +![](images/e811a10f4c60204f96f6c865dd50c9e8fb66f1a3b4ed73e663ad0f6372a3de34.jpg) +(d) + +![](images/f82ac52479cf5c63bbf3e6a93b8f864792e9d117ba9065c61e8075f6f12a449f.jpg) + +![](images/03751131758c4f23b89598d86b300d840fb496b1366ab5ae2a4f42bdf1f1eca8.jpg) + +Finally, for notational brevity, we will drop the explicit dependence on $\theta$ , and write $x' = x + u$ . We also define a kernel closely related to the NTK, the Cos NTK: + +$$ +C _ {N T K} (x, x + u) = \frac {1}{\operatorname {V o l} (\mathcal {D})} \frac {\nabla_ {\theta} f (x) \cdot \nabla_ {\theta} f (x + u)}{| | \nabla_ {\theta} f (x) | | | | \nabla_ {\theta} f (x + u) | |} \tag {4} +$$ + +# 3 Deriving Order Parameters from the NTK + +We illustrate the different phases of learning in Figure 1; we train a (five-layer, 256-unit wide, $\omega_0 = 60$ ) SIREN model on a $128 \times 128$ grayscale image using full-batch GD (learning rate $= 10^{-3}$ ). We evaluate the model on super-resolution at $256 \times 256$ . We examine the learning dynamics through three different lenses, each revealing a sudden shift. These shifts are quantitatively identified using statistics, known as order parameters. We demonstrate below how order paramters for each transition may be related to a common set of features, which control the local NTK structure. The three lenses are as follows: + +- Spatial Distribution of Residuals: Early in training, the loss decreases uniformly over the dataset (Drift Phase). However, at a critical point, we observe the formation of "wave-crests" corresponding to regions of low-loss, which propagate across the dataset (Diffusion Phase). To the best of our knowledge, we are the first to report this behaviour in SIREN models. We attribute this behaviour (in sec. 3.1) to changes in the equal-time correlation functions of the gradient field $\nabla_{\theta}f(x)$ , whose parameters we derive in Section 3.2. +- Principal Eigenvectors of the NTK: The principal eigenvector $v_{0}$ is initially static and appears highly-disordered (Disordered Phase). However, at a critical point, $v_{0}$ experiences a brief, sudden shift, in which it aligns with the edges of the image (Aligned Phase). Although NTKA has previously been studied in the context of classification problems [24, 25, 26, 27], there are additional subtleties to consider for a regression task such as INR training. To this end, we introduce a metric, $\mathrm{AUC}(|v_0|, \nabla I)$ in Section 3.3 to identify when alignment occurs. We also derive an approximation of $v_{0}$ based on the local structure of the NTK, as outlined in Sections 3.1 and 3.2. +- Training Curve Analysis: There is a rapid shift in the slope of the training curve, which we call the loss rate $\dot{L}$ . Learning is initially fast (high $\dot{L}$ ), but after a critical point, slows abruptly (low $\dot{L}$ ). Several works have studied this transition using order parameters, but in this work, we focus on + +![](images/fd0500ec4cf854d60adcafc162ea475f0b155ce25d190aeb0c627802af8a4ac2.jpg) +Figure 2: Spatiotemporal Evolution of the Cosine NTK ( $C_{NTK}$ ). Left: Global correlation function of the $C_{NTK}$ at different epochs. Dashed lines show fitted Gaussian approximation from equation equation 6, and error bars show variance across datasets. Right: Visualization of the $C_{NTK}$ around three points $x \in \{A, B, C\}$ for small separations, at the beginning and end of training. + +![](images/92abea5eb791bf1d1f7f0177f6b72ed4606bc9d1effa3985c7c1d5329f52e2c3.jpg) + +![](images/4288102767d604fa18540b60066580d934e5475a04d19f478ad6a5733124e763.jpg) + +![](images/d34eed51e93c880c06a28ee966b922dc6fa23cfa6a858a69ef32b48a8ecea4b6.jpg) + +![](images/0c5bbf811e38f97d0efb096ed2118aebaf61a47e0d34d0146a6e58efd5a749d1.jpg) + +the concept of gradient confusion, as described in [13], [11], [12]. In Section 3.4, we derive an approximation of this parameter based on the local structure of the NTK outlined in Section 3.2. + +# 3.1 Correlation Functions and the Onset of Diffusion + +The form of equation 3 is reminiscent of the linear response functions in statistical field theory [28, 29]: to find the rate of change of the residual field at a point $x$ , the kernel $K$ aggregates information about the residual at points $x + u$ . To quantify the range of these interactions, we examine the local, equal-time correlation functions for the gradients $\nabla_{\theta} f(x)$ separated by a distance $\epsilon$ : + +$$ +k (x, \epsilon) = \mathbb {E} _ {\phi} \left[ \nabla_ {\theta} f (x) \cdot \nabla_ {\theta} f \left(x + \epsilon \hat {e} _ {\phi}\right) \right] = \mathbb {E} _ {\phi} \left[ K _ {N T K} (x, x + \epsilon \hat {e} _ {\phi}) \right] \tag {5} +$$ + +Here, $\hat{e}_{\phi}$ denotes a unit vector in direction $\phi$ . Similarly, the global, equal-time correlation function is given by $k(\epsilon) = \mathbb{E}_x[k(x,\epsilon)]$ . We may define similar quantities for the $C_{NTk}$ , which we denote by $c(x,\epsilon)$ and $c(\epsilon)$ . We expect the range of these interactions to be short, as INRs are often carefully designed to ensure a diagonally dominant NTK[30, 31, 32]. To verify this, we group pairs of datapoints based on their distance, and then compute the mean $C_{NTk}$ value. An example of this empirical correlation function is shown in Figure 2. The observed structure motivates the following proposition, which we examine qualitatively in Figure 12 and validate numerically in Appendix F.1. + +Proposition 3.1. For SIREN models, the equal-time correlation functions of the NTK are well-approximated by Gaussians of the form: + +$$ +c (\epsilon) \approx \left(1 - c _ {\infty}\right) e ^ {- \epsilon^ {2} / 2 \xi_ {c o r r} ^ {2}} + c _ {\infty}, \tag {6} +$$ + +$$ +k (x, \epsilon) \approx \left\| \nabla_ {\theta} f (x) \right\| ^ {2} \left[ \left(1 - c _ {\infty} (x)\right) e ^ {- \epsilon^ {2} / 2 \xi (x) ^ {2}} + c _ {\infty} (x) \right]. \tag {7} +$$ + +Our approximation introduces two important order parameters: the first, the correlation length-scale $\xi_{corr}$ , controls the rate at which correlations decay with distance, defining the range of interactions. The second, the asymptotic value $c_{\infty}$ , describes the interactions between points at separations $\epsilon$ much greater than $\xi$ . Succeeding a consideration of anisotropic effects, the above correlation functions are consistent with the following form for the full NTK: + +Proposition 3.2. Gaussian Approximation of the SIREN NTK. The NTK may be approximated as a Gaussian kernel with spatially-varying amplitude $||\nabla_{\theta}f(x)||^{2}(1 - c_{\infty}(x))$ , bandwidth $\xi (x)$ and asymptotic value $||\nabla_{\theta}f(x)||^{2}c_{\infty}(x)$ : + +$$ +K _ {G a u s s} (x, x + u) \approx | | \nabla_ {\theta} f (x) | | ^ {2} \left(1 - c _ {\infty} (x)\right) \exp \left(- \| u \| ^ {2} / \xi^ {2} (x)\right) + | | \nabla_ {\theta} f (x) | | ^ {2} c _ {\infty} (x) \tag {8} +$$ + +Dynamically, we see from the left of Figure 2 that both $\xi$ and $c_{\infty}$ evolve during training, and we shall demonstrate that changes in these values account for the onset of diffusion. + +Theorem 3.1. Diffusive Evolution of the Residuals: Let the mean residual be denoted by $\mu_r \equiv \mathbb{E}_x[r]$ , and let $K_{\infty}$ denote the mean contribution to the $K_{NTK}$ at large distances $\|u\| \gg \xi(x)$ . Then, as $\mu_r K_{\infty} \to 0$ , assuming $K_{NTK} \approx K_{Gauss}$ , the residuals approximately evolve under the following diffusion equation: + +$$ +\frac {d}{d t} r (x, t) \approx - 2 \pi \xi^ {2} (x) | | \nabla_ {\theta} f (x) | | ^ {2} r (x, t) - \pi \xi^ {4} (x) | | \nabla_ {\theta} f (x) | | ^ {2} \Delta_ {x} ^ {2} r (x, t) \tag {9} +$$ + +Proof Sketch. (Full details in Appendix A.2). As $\mu_r K_\infty \to 0$ , local interactions dominate the background in determining the evolution of $r(x)$ in equation 3. Thus we may approximate: + +$$ +\frac {d r}{d t} \approx - \int d x r (x + u) K (x, x + u) \approx \int d x r (x + u) | | \nabla_ {\theta} f (x) | | ^ {2} \exp (- | | u | | ^ {2} / \xi^ {2} (x)) \tag {10} +$$ + +The exponentially-decaying NTK will suppress all contributions to $\dot{r} (x)$ from residuals $r(x + u)$ for which $\lvert\lvert u\rvert\rvert\gg\xi(x)$ . When $\xi(x)$ is small, we may perform a Taylor expansion around the point $x$ : + +$$ +r (x + u; \theta) \approx r (x; \theta) + u ^ {\top} \nabla_ {x} r (x; \theta) + \frac {1}{2} u ^ {\top} \nabla_ {x} ^ {2} r (x; \theta) u \tag {11} +$$ + +Inserting this into equation 10, the full integral may be solved using Gaussian integration. We obtain: + +$$ +\frac {d}{d t} r (x; \theta) = - 2 \pi \xi^ {2} (x) | | \nabla_ {\theta} f (x) | | ^ {2} r (x) - \pi \xi^ {4} (x) | | \nabla_ {\theta} f (x) | | ^ {2} \Delta_ {x} ^ {2} r, \tag {12} +$$ + +which resembles a standard diffusion equation + +![](images/1dab7359be4f0212d1fa320e52baab886606ab259a2b6a0f4aec7c825be975b0.jpg) + +# 3.2 Beyond the Isotropic Gaussian Approximation + +Though the Gaussian approximation of the NTK can account for the appearance of the diffusion wavecrests, it fails to capture other empirical properties. Namely, whereas the true NTK is anisotropic (see Figure 2) and can contain negative entries, the Gaussian kernel is isotropic, and satisfies $K_{\text{Gauss}}(x,x + u) > 0$ for all $u$ . To overcome these limitations, we present the following refinement: + +Theorem 3.2. Cauchy Approximation of the SIREN NTK: For small separations $u$ , the Cosine NTK locally takes the form of a Cauchy Distribution with structure parameters $a_x, D_x, H_x$ : + +$$ +C _ {N T K} (x, x + u) \approx \frac {2 a _ {x} ^ {2} + u ^ {\top} D _ {x}}{2 a _ {x} ^ {2} + u ^ {\top} D _ {x} + u ^ {\top} H _ {x} u}, \tag {13} +$$ + +These parameters are obtained from the model gradients as follows: + +$$ +a _ {x} = \left\| \nabla_ {\theta} f (x) \right\|; D _ {x} = \nabla_ {x} \left\| \nabla_ {\theta} f (x) \right\| ^ {2}; H _ {x} = \left(\nabla_ {x} \nabla_ {\theta} f (x)\right) \left(\nabla_ {x} \nabla_ {\theta} f (x)\right) ^ {\top} \tag {14} +$$ + +Proof Sketch. (Full details in Appendix A.3). Via the Law of Cosines, the $C_{NTK}$ satisfies: + +$$ +C _ {N T K} (x, x + u) = \frac {\left| \left| \nabla_ {\theta} f (x) \right| \right| ^ {2} + \left| \left| \nabla_ {\theta} f (x + u) \right| \right| ^ {2} - \left| \left| \nabla_ {\theta} f (x + u) - \nabla_ {\theta} f (x) \right| \right| ^ {2}}{2 \left| \left| \nabla_ {\theta} f (x) \right| \right| \left| \left| \nabla_ {\theta} f (x + u) \right| \right|} \tag {15} +$$ + +The result then follows by Taylor expanding the numerator and denominator to second order in $u$ . + +A benefit of this new approximation is that it can be used to predict the correlation length-scale: + +Corollary 3.2.1. The correlation length-scale for the NTK about a point $x$ may be constructed from its local structure parameters $a_x, D_x, H_x$ and asymptotic value $c_{\infty}(x)$ as follows: + +$$ +\xi^ {2} (x) \approx 2 \left(\frac {1 - c _ {\infty} (x)}{1 + c _ {\infty} (x)}\right) \frac {a _ {x} ^ {2}}{\sqrt {\det H _ {x}}} + \frac {1}{4} \left(\frac {1 - c _ {\infty} (x)}{1 + c _ {\infty} (x)}\right) ^ {2} \frac {D _ {x} ^ {\top} H _ {x} ^ {- 1} D _ {x}}{\sqrt {\det H _ {x}}} \tag {16} +$$ + +Proof Sketch. (Full details in Appendix A.4). Note that the levels sets of equation 13 correspond to ellipses. For a given value $c$ , the area of the level set can be shown to be: + +$$ +A _ {\text {e l l i p s e}} (x; c) = 2 \pi \left(\frac {1 - c}{c}\right) \frac {a _ {x} ^ {2}}{\sqrt {\det H _ {x}}} + \frac {\pi}{4} \left(\frac {1 - c}{c}\right) ^ {2} \frac {D _ {x} ^ {\top} H _ {x} ^ {- 1} D _ {x}}{\sqrt {\det H _ {x}}} \tag {17} +$$ + +The correlation lengthscale is then approximated as $\xi (x)\approx \sqrt{A_{ellipse}(x,c) / \pi}$ , where we choose $c = 1 / 2 + c_{\infty} / 2$ to account for the asymptotic value of the $C_{NTK}$ . + +# 3.3 Order Parameters for the Onset of NTK Alignment + +In the classification problems typically studied in the NTKA literature, the principle eigenvector $v_{0}(x)$ is seen to learn class-separating boundaries [24, 25]. Similarly, for our 2D image reconstruction task, we see the NTK learns information about the distribution of edges in the image (Figure 3). To quantify this alignment, we use a Canny Edge Detector [33] to estimate connected image edges. We then quantify the utility of $|v_{0}(x)|$ in predicting edges in terms of average recall, as measured by the area under the Receiver Operating Characteristic Curve (ROC AUC). We denote this measure $\mathrm{AUC}(|v_0|,\nabla I)$ , and it has the advantage of being insensitive to monotonic transformations of $|v_{0}|$ . This invariance is beneficial in two respects: (1) as a reliability measure for a binary predictor (edges), it obviates the need to specify a threshold, facilitating comparisons across datasets (see Section 4.1); and (2) empirically, it saturates during training, facilitating the identification of a critical point. + +Another hallmark of NTKA is early anisotropic growth of the NTK spectrum [25], as the NTK becomes stretched along a small number of directions correlated with the task. This is especially the case for the principal eigenvalue $\lambda_0$ , which grows orders of magnitude larger than the next leading eigenvalue. In Section 4.1, we will demonstrate empirically that this is also true for INRs. + +The divergence of $\lambda_0$ enables a particularly simple approximation of the principal eigenvector $v_{0}$ : + +Corollary 3.2.2. The principal eigenvector $v_{0}(x)$ of the NTK admits the following approximation in terms of the local asymptotic value $c_{\infty}(x)$ and the local correlation lengthscale $\xi (x)$ : + +$$ +v _ {0} (x) \approx a _ {x} ^ {2} \left[ c _ {\infty} (x) \operatorname {V o l} (\mathcal {D}) + 2 \pi \xi^ {2} (x) \left(1 - c _ {\infty} (x)\right) \right] \tag {18} +$$ + +Proof. Because the principal eigenvalue is so dominant, $K_{NTK}$ becomes effectively low-rank, and so power iterations converge quickly. Thus, choosing a vector of ones $v = 1$ as our initial vector, we expect $K1 / 1^{\top}1$ to have strong cosine alignment with the principal eigenvalue. In the continuum limit, this is simply given by: + +$$ +\begin{array}{l} K 1 / N \rightarrow \mathbb {E} _ {u} [ K (x, x + u) ] = \mathbb {E} _ {\epsilon} [ \mathbb {E} _ {u} [ K (x, x + u) | | u | | = \epsilon ] ] (19) \\ = \int_ {0} ^ {\epsilon_ {\max }} d \epsilon k (x, \epsilon) P (x, \epsilon) (20) \\ \end{array} +$$ + +Here, $P(x,\epsilon)$ denotes the density of points that are located a distance $\epsilon$ from the point $x$ , and $\epsilon_{max}$ is an upper bound on the distance that we assume is much greater than $\xi_{corr}$ . Close to this $x^{1}$ , $P(x,\epsilon)$ grows like $2\pi \epsilon$ . Thus, leveraging equations 7 and 14, we have: + +$$ +\begin{array}{l} v _ {0} (x) \approx 2 \pi a _ {x} ^ {2} \int_ {0} ^ {\epsilon_ {m a x}} d \epsilon \epsilon \left[ c _ {\infty} (x) + \left(1 - c _ {\infty} (x)\right) e ^ {- \epsilon^ {2} / 2 \xi^ {2} (x)} \right] (21) \\ = 2 \pi a _ {x} ^ {2} \left[ c _ {\infty} (x) \epsilon_ {m a x} ^ {2} + \xi^ {2} (x) \left(1 - c _ {\infty} (x)\right) \left(1 - e ^ {- \epsilon_ {m a x} ^ {2} / 2 \xi^ {2} (x)}\right) \right] (22) \\ \approx a _ {x} ^ {2} \Big [ c _ {\infty} (x) \mathrm {V o l} (\mathcal {D}) + 2 \pi \xi^ {2} (x) (1 - c _ {\infty} (x)) \Big ] \\ \end{array} +$$ + +We evaluate the fidelity of this approximation in Appendix F. As we approach the phase transition, the asymptotic values tend towards 0, and the second term dominates. Considering the approximation for the correlation length-scale $\xi$ in Corollary 3.2.1, we note that $v_{0}(x)$ grows as $\mathcal{O}(||\nabla_{\theta}f(x)||^{4})$ . This implies particular sensitivity to pixels in regions with substantial high-frequency information, such as edges and corners. As natural images tend to be piecewise smooth, pixels on boundaries have the strongest spatial gradients, and are therefore the greatest source of information, being poorly compressible due to the lack of smoothness, and accordingly disagreement in parameter gradients. Given the inability of models to accurately describe sharp discontinuities these edge pixels are considered influential datapoints, which accounts for their prominence within the principal eigenvector. We discuss other parallels between the moments of the NTK and traditional corners detection algorithms in Appendix E. In particular, we introduce another order parameter, termed MAG-Ma (Magnitude of the Average Gradient of the Log Gradient-Field Magnitudes), to monitor the breakdown of stationarity (ie local translation invariance) of the NTK. It is obtained as $\| \mathbb{E}_x[D_x / a_x^2 ]\| ^2$ + +# 3.4 Order Parameters for the Loss Rate Collapse + +In [13], [11], [12], and related works, the authors examine the role of gradient alignment statistics in determining the speed of learning under stochastic gradient descent. They note the emergence of negative alignments between batches correlates with a reduction in learning speed. Intuitively, when sample gradients become negatively aligned, the sum of the gradients approaches zero, resulting in a diminished learning signal. The minimum alignment is simply the minimum value of the $C_{NTK}$ , which we may obtain explicitly from Theorem 3.2 as follows: + +Corollary 3.2.3. The minimum value of the $C_{NTK}$ admits the following approximation in terms of the local structure parameters $a_x, D_x, H_x$ : + +$$ +\min _ {u} C _ {N T K} (x, x + u) = \frac {D _ {x} ^ {\top} H _ {x} ^ {- 1} D _ {x}}{D _ {x} ^ {\top} H _ {x} ^ {- 1} D _ {x} - 8 a _ {x} ^ {2}} \tag {23} +$$ + +Proof Sketch. (Full details in Appendix A.5). Setting $\partial_u C_{NTK}(x, x + u) = 0$ yields two solutions: $u = 0$ , corresponding to the maximum (1), and another corresponding to the minimum. $\square$ + +The min $C_{NTK}$ is then simply the minimum of 23 across the whole dataset. + +# 4 Experimental Results + +Setup: We fit SIREN models to a set of thirty $64 \times 64$ downsampled images and evaluate the MSE $L_{eval}$ on a super-resolution task (at $256 \times 256$ ). We used five random seeds and also varied the width, depth and bandwidth $\omega_0$ (ranges are given in Appendix B.2). We compute the eigenspectra of the NTK using Randomized SVD [34]. In addition to the order parameters described in Section 3, we examine three NTK-based order parameters from the literature: (1) The principal eigenvalue $\lambda_0$ of the NTK, which diverges at the critical point; (2) The variance of the gradients $\sigma_\theta^2$ , which peak during the Fast-Slow learning phases [10], and which may be connected (see Appendix A.6) to the trace of the NTK; (3) The Centred Kernel Alignment (CKA) between the NTK and a task kernel $K_Y$ . For INR regression, we use $K_Y(x,x + u) = \exp(-50||I(x) - I(x + u)||^2)$ . The similarity between kernels is measured using the normalized Hilbert-Schmidt Information Criterion (HSIC), as in [25, 26, 27]. Full experimental details may be found in Appendix B.2. + +# 4.1 Examining the Distribution of Critical Points + +Critical points cluster around run-specific times: The left-hand side of Figure 3 illustrates our procedure for identifying critical points in a given trial. We use a simple peak detector to identify the region of interest for the loss rate $\dot{L}_{eval}$ and the gradient variance $\sigma_{\theta}$ , using the FWHM to define a confidence region. For the min $C_{NTK}$ , we look for zero-crossings, with a confidence region constructed from the cumulative variance. For every other order parameter, we fit a sigmoid, where the inflection point marks the critical point, and the slope defines the confidence region (full details in Appendix B.2). The right side of Figure 3 demonstrates how frequently these confidence regions overlap across our experimental sweep2. Remarkably, the phase transitions described by the order parameters - despite being derived to measure different phenomena in the literature - consistently occur at the same time during training. + +Hyperparameters alter the timing of run-specific transitions: in Table 1 we observe that both depth and bandwidth $\omega_0$ have critical roles in controlling the shift in the loss rate $\dot{L}_{eval}$ . Generally, increasing depth and decreasing $\omega_0$ result in earlier transition times $t_{\mathrm{crit}}$ . However, these changes have opposite effects on the model performance: for fixed $\omega_0$ , deeper models to converge to better (lower $\dot{L}_{eval}$ ) solutions faster. However, it seems that lower values of $\omega_0$ cause models to converge prematurely. This may be deduced by studying the correlation between the final residuals (details in Appendix B.2). For equivalent depth (and therefore, equivalent traditional capacity), models with lower $\omega_0$ exhibit more correlations in the residuals. This is indicative of remaining structure in the residuals, and can be interpreted as evidence of under-fitting. + +![](images/00e8a2c52f613ce105d4e39bf23e9023eb70dee6c2c32ed1fcd2b6461acde251.jpg) +Figure 3: Alignment of Order Parameters. Left: Order parameter evolution and critical points during training of a SIREN model on the astro image. The red vertical lines denote the location of the critical points, and the green vertical lines denote confidence regions. Right: Heatmap showing the frequency of intersections between the confidence regions. Additional figures in Appendix G.1. + +![](images/a9a446e49246bdecb1d7b04844a335c9bdd1d9b07b222bc14bebdc740635e2bc.jpg) + +Table 1: Variation in model performance with hyperparameters: Comparing the dependence of transition times ( $t_{crit}$ ), super-resolution performance ( $\log_{10} L_{\mathrm{eval}}$ ), and residual correlation on depth and bandwidth. Also featuring expected correlation lengthscale ( $\mathbb{E}[\xi_{corr}(t)]$ ) and correlation between $\log ||\nabla_{\theta}f(x;\theta)||$ and $||\nabla_xI(x)||$ . Values are averaged over the same sweep defined in Section 4.1. + +
depth ω0E[ξcorr(t)]AUC(|v0|, I)Grad. Corr.log10tcritlog10LevalRes. Corr.
3/900.04 ± 0.000.59 ± 0.050.06 ± 0.102.83 ± 0.19-2.01 ± 0.310.35 ± 0.06
3/600.06 ± 0.000.60 ± 0.050.16 ± 0.112.84 ± 0.12-2.00 ± 0.320.40 ± 0.07
3/300.11 ± 0.000.60 ± 0.050.27 ± 0.082.44 ± 0.25-1.92 ± 0.300.44 ± 0.07
3/150.18 ± 0.010.60 ± 0.050.26 ± 0.072.03 ± 0.34-1.83 ± 0.290.48 ± 0.07
4/900.04 ± 0.000.67 ± 0.070.23 ± 0.132.66 ± 0.17-2.02 ± 0.320.35 ± 0.06
4/600.06 ± 0.000.69 ± 0.070.29 ± 0.122.61 ± 0.16-2.04 ± 0.330.39 ± 0.07
4/300.10 ± 0.000.71 ± 0.080.41 ± 0.102.22 ± 0.27-1.99 ± 0.310.41 ± 0.07
4/150.16 ± 0.010.68 ± 0.090.55 ± 0.081.87 ± 0.35-1.94 ± 0.300.43 ± 0.07
5/900.03 ± 0.000.68 ± 0.070.24 ± 0.142.54 ± 0.18-2.01 ± 0.320.34 ± 0.06
5/600.05 ± 0.000.71 ± 0.070.30 ± 0.132.42 ± 0.17-2.05 ± 0.330.39 ± 0.07
5/300.09 ± 0.000.73 ± 0.080.39 ± 0.122.15 ± 0.12-2.02 ± 0.320.40 ± 0.07
5/150.15 ± 0.010.72 ± 0.080.52 ± 0.091.85 ± 0.27-1.98 ± 0.310.41 ± 0.07
+ +# 4.2 Influence of Hyperparameters on Edge Alignment + +In the previous section, for fixed depth, we demonstrated that lower $\omega_0$ correlates with (1) earlier phase transitions, (2) higher validation loss, and (3) correlated residuals. Together, these observations suggest that models with lower $\omega_0$ converge prematurely, underutilizing their capacity. A natural question arises: which patterns do these models struggle to capture? Given the observed concurrence of loss rate collapse and NTK alignment, we analyse the NTK eigenspectrum to gain some insight. + +In Figure 4, we train four SIRENs on the sax dataset with $(\omega_0,\mathrm{depth})\in \{15,60\} \times \{3,5\}$ . We visualise both the log magnitudes of the parameter gradients, $\log ||\nabla_{\theta}(x)f||^2$ , and the principal eigenvector, $v_{0}(x)$ , at the end of training. Additional Figures may be found in Appendix C.2 + +![](images/df60001b6f27459d880a7bc3bfff7a7306c6164a79f3732ead87e306a5bab56e.jpg) + +![](images/74b436811432106cdc2fc5562fbacc0b726927ca4ae840000d79447ae0d4a089.jpg) + +![](images/0a331005c86036c1ef38545c4f1cf5f870259799180c2942f403e56f91a5713d.jpg) + +![](images/504db8873d1546edb1650c202b4c67eee70946d5bfb4c4e958ffce1bfa97d376.jpg) + +![](images/2d9b7e8d71d7dc643b587cc1df0387778f766f7a86ebd7b859fd6533114a696c.jpg) +Figure 4: Effect of Hyperparameters on Edge Alignment: Left (magma colormap): The norm of the parameter gradients $||\nabla_{\theta}f(x)||$ for $(\omega_0, \mathrm{depth}) \in \{15,60\} \times \{3,5\}$ , labelled with the Pearson correlation $\rho$ between the log of the norm and the spatial gradient $\nabla_xI$ of the target image. Right (viridis colormap): visualizing the principal eigenvector $v_0$ of the NTK for the same models, labelled with the edge alignment score $\mathrm{AUC}(|v_0|, \nabla I)$ . More images in Appendix C.2 + +![](images/c0e8bf5699d98ffa1afacc496e89e4c19f254a72502dade07eba733a0919716e.jpg) + +![](images/56b8195ad145db8f12912d0bb090dcd9ada303b6bde71ecbec76dd854612d768.jpg) + +![](images/6b8a8387eaf19905da628d6e8f7c6cc1c527fdefea3d732e8589aa304a88d99e.jpg) + +Generally, we observe that for low $\omega_0$ , $\log ||\nabla_\theta f(x;\theta)||^2$ swells and concentrates near image edges, becoming sparser, and more correlated with the spatial gradient magnitudes $||\nabla_x I(x)||$ (aggregated statistics may be seen in Table 1). This edge prominence in $v_0$ matches expectations, as per Corollary 3.2.2, $v_0(x) \sim \mathcal{O}(||\nabla_\theta f(x)||^4)$ . Overall, while edge alignment is seen across most settings, it is especially prominent for deeper models with lower values of $\omega_0$ . This indicates prioritization of these patterns by the NTK, and correspondingly, the patterns the model is most invested in. + +# 5 Discussion + +To explain the impact of $\omega_0$ , in Table 1 we track the expected correlation lengthscale $\mathbb{E}[\xi_{corr}(t)]$ over the course of training. In Figure 17 of the Supplementary Materials, we also examine the variance in $\xi_{corr}(0)$ with $\omega_0$ in all models. Consistently, we observe that lower $\omega_0$ is associated with larger values of the correlation lengthscale. This implies that the NTK integrates information across larger neighborhoods, implicitly averaging over high-frequency features. Consequently, we expect these SIRENs to struggle when modeling high-frequency patterns, which is consistent with other observations in the literature [32]. Following the discussion in Section 3.3, we expect that the difficulty of modeling edges is responsible for the swelling of the parameter gradients $||\nabla_{\theta}f(x)||^2$ . + +Intriguingly, we observe that the principal eigenvector becomes sparser as we increase depth, leading to stronger edge alignment, as seen in Table 1. Yet, this sparsification is associated with completely different generalization behavior: models achieve lower validation loss, with less correlated residuals, as we increase depth. We hypothesise a different mechanism underlies this sparsification in comparison with $\omega_0$ . Figure 4 demonstrates increasing $\omega_0$ increases the sensitivity of the gradient magnitudes (and hence the principal eigenvector) to noise in the images. For DNNs, gradient magnitudes decompose into a sum across layers, namely $||\nabla_{\theta}f||^2 = \sum_{l=1}^{\mathrm{depth}} ||\nabla_{\theta^{(l)}}f||^2$ . In effect, preference is given to points which are consistently confusing across layers, thus mitigating the effects of noise. + +# 6 Related Work + +Fast and Slow Phases of Neural Network Training: The literature highlights a dynamical phase transition in DNN training between fast to slow learning regimes [10]. The initial fast phase is characterized by large gradient norms and low fluctuations, yielding rapid loss reduction via broad agreement across examples. It is followed by a slow phase in which gradient fluctuations dominate + +and progress decelerates. This transition can be tracked by order parameters including gradient signal-to-noise [10], gradient confusion, [13, 11, 12], and correlation lengthscales [13]. Furthermore, it reflects a move from learning simple, shared patterns to fitting complex, idiosyncratic ones [4], and thus offers a window into feature learning. Related work studies the representation dynamics in ReLU networks [35]; here, we study them via the NTK in SIREN models. + +Neural Tangent Kernels for Implicit Neural Representations: Previous research has investigated the inductive biases of INRs using the Neural Tangent Kernel (NTK), focusing on aspects such as spectral properties [36] and dependencies on uniformly sampled data [30]. Furthermore, studies by [37] and [38] have analyzed the eigenfunctions of the empirical NTK to elucidate the approximation capabilities of INRs. These investigations, however, primarily examine static properties of the NTK at initialization, which do not account for feature learning dynamics. This is known to be a poor approximation [39]. In contrast, our work concentrates on the evolution of the NTK, aiming to deepen our understanding of how INRs learn to model images. + +Neural Tangent Kernel Alignment In practical settings, recent studies have shown that during training, the NTK dynamically aligns with a limited number of task-relevant directions [40, 41, 24, 21, 25, 22, 26, 27]. Concurrently, at the eigenfunction level, the modes increasingly reflect salient features of the dataset, such as class-separating boundaries [24, 25], and Fourier frequencies [25]. The widespread occurrence and influence of kernel alignment suggest its critical role in DNN feature learning, contributing to the superior performance of DNNs over models based on infinitewidth NTKs [26]. Direct optimisation of alignment measures has even been suggested as one way to enhance the convergence of GD and generalization of models[42, 43]. That said, theoretical investigation into spontaneous NTKA often focuses on shallow networks [21, 22], toy models [26, 25], and deep linear networks [22]. In contrast, the INRs we study are deep (3-6 layers), nonlinear models that see frequent use in Computer Vision problems. + +# 7 Conclusion + +We have developed new formulations that leverage the NTK to characterise the dynamics of feature learning in deep image regression models (SIRENs). By analytically deriving approximations for the local structure of SIREN NTKs - using Gaussian and Cauchy distributions - we were able to obtain approximate expressions for the correlation lengthscale, the minimum value of the $C_{NTK}$ , and the principal eigenvector. We related these expressions to order parameters for three phase transitions identified in different dynamical perspectives on learning: the appearance of diffusion wave-crests in residual evolution (first identified in this paper); the collapse of the loss rate; the onset of NTK alignment. We argued, based on these derivations and empirical demonstrations that critical points cluster in time, that these distinct phase transitions share a common, underlying mechanism. + +The following picture emerges from our analysis: as long range correlations between gradients decay, residuals only interact with their immediate neighbours (onset of diffusion), leading to increased gradient variance (loss rate collapse) and translational symmetry breaking. In parallel, the growth of the principal eigenvalue or the NTK leads the principal eigenvector to memorize the distribution of influential points, as measured by accumulating gradients. In images, one influential class of points are edges, leading to their prominence in the principal eigenvector (NTK alignment). + +In this study, we focused on SIREN models trained on a 2D super-resolution task using full-batch gradient descent. However, SIRENs are used in a variety of inverse problems, and it remains to be seen whether our observations extend to these settings. Future work may also explore the impact of different optimizers, such as ADAM [44], which adaptively adjusts learning rates and may influence the stability and divergence of the principal eigenvalue - a key factor in our study of NTK alignment. + +This work has demonstrated that the NTK provides a rich theoretical tool for deriving and relating order parameters to understand training dynamics. We provide new methodology to rigorously study the influence of inductive biases, such as model architectures and hyper-parameters, on the underlying learning process and may have practical utility in diagnosing causes of poor learning outcomes. + +# Acknowledgements + +Supported by the University of Sussex Be.AI doctoral scholarship, funded by the Leverhulme Trust. + +# References + +[1] Christopher M. Bishop. Pattern Recognition and Machine Learning (Information Science and Statistics). Springer, 1 edition, 2007. +[2] Boris Hanin and David Rolnick. Deep relu networks have surprisingly few activation patterns. In Neural Information Processing Systems, 2019. +[3] Maxwell Nye and Andrew M. Saxe. Are efficient deep representations learnable? ArXiv, abs/1807.06399, 2018. +[4] Devansh Arpit, Stanisław Jastrzebski, Nicolas Ballas, David Krueger, Emmanuel Bengio, Maxinder S. Kanwal, Tegan Maharaj, Asja Fischer, Aaron Courville, Yoshua Bengio, and Simon Lacoste-Julien. A closer look at memorization in deep networks. (arXiv:1706.05394), 2017. +[5] Xiao Zhang and Dongrui Wu. Rethink the connections among generalization, memorization, and the spectral bias of dnns. In International Joint Conference on Artificial Intelligence, 2020. +[6] Yu Feng and Yuhai Tu. Phases of learning dynamics in artificial neural networks in the absence or presence of mislabeled data. Machine Learning: Science and Technology, 2(4):043001, jul 2021. +[7] Cory Stephenson and Tyler Lee. When and how epochwise double descent happens. ArXiv, abs/2108.12006, 2021. +[8] Ziming Liu, Ouail Kitouni, Niklas Nolte, Eric J. Michaud, Max Tegmark, and Mike Williams. Towards understanding grokking: An effective theory of representation learning. *ArXiv*, abs/2205.10343, 2022. +[9] Liu Ziyin and Masakuni Ueda. Exact phase transitions in deep learning. ArXiv, abs/2205.12510, 2022. +[10] Ravid Shwartz-Ziv and Naftali Tishby. Opening the black box of deep neural networks via information. (arXiv:1703.00810), 2017. +[11] Karthik A. Sankararaman, Soham De, Zheng Xu, W. Ronny Huang, and Tom Goldstein. The impact of neural network overparameterization on gradient confusion and stochastic gradient descent. (arXiv:1904.06963), 2020. +[12] Yu Feng and Yuhai Tu. Phases of learning dynamics in artificial neural networks in the absence or presence of mislabeled data. 2(4):043001, 2021. Publisher: IOP Publishing. +[13] Stanislav Fort, Paweł Krzysztof Nowak, Stanislaw Jastrzebski, and Srini Narayanan. Stiffness: A new perspective on generalization in neural networks. (arXiv:1901.09491), 2020. +[14] Arthur Jacot, Franck Gabriel, and Clément Hongler. Neural tangent kernel: Convergence and generalization in neural networks. CoRR, abs/1806.07572, 2018. +[15] Boris Hanin and Mihai Nica. Finite depth and width corrections to the neural tangent kernel. (arXiv:1909.05989), 2019. +[16] Jiaoyang Huang and Horng-Tzer Yau. Dynamics of deep neural networks and neural tangent hierarchy. (arXiv:1909.08156), 2019. +[17] Laurence Aitchison. Why bigger is not always better: on finite and infinite neural networks. (arXiv:1910.08013), 2020. +[18] Lenaic Chizat, Edouard Oyallon, and Francis Bach. On lazy training in differentiable programming. (arXiv:1812.07956), 2020. +[19] Jaehoon Lee, Samuel S. Schoenholz, Jeffrey Pennington, Ben Adlam, Lechao Xiao, Roman Novak, and Jascha Sohl-Dickstein. Finite versus infinite neural networks: an empirical study. (arXiv:2007.15801), 2020. + +[20] Mariia Seleznova and Gitta Kutyniok. Neural tangent kernel beyond the infinite-width limit: Effects of depth and initialization. (arXiv:2202.00553), 2022. +[21] Jonas Paccolat, Leonardo Petrini, Mario Geiger, Kevin Tyloo, and Matthieu Wyart. Geometric compression of invariant manifolds in neural nets. 2021(4):044001. +[22] Alexander Atanasov, Blake Bordelon, and Cengiz Pehlevan. Neural networks as kernel learners: The silent alignment effect. (arXiv:2111.00034), 2021. +[23] Vincent Sitzmann, Julien N.P. Martel, Alexander W. Bergman, David B. Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. In Proc. NeurIPS, 2020. +[24] Dmitry Kopitkov and Vadim Indelman. Neural spectrum alignment: Empirical study. (arXiv:1910.08720), 2020. +[25] Aristide Baratin, Thomas George, César Laurent, R. Devon Hjelm, Guillaume Lajoie, Pascal Vincent, and Simon Lacoste-Julien. Implicit regularization via neural feature alignment. (arXiv:2008.00938), 2021. +[26] Haozhe Shan and Blake Bordelon. A theory of neural tangent kernel alignment and its influence on training. (arXiv:2105.14301), 2022. +[27] Abdulkadir Canatar and Cengiz Pehlevan. A kernel analysis of feature learning in deep neural networks. In 2022 58th Annual Allerton Conference on Communication, Control, and Computing (Allerton), pages 1-8. IEEE. +[28] James P. Sethna. Statistical mechanics: Entropy, order parameters, and complexity. Oxford University Press, 2021. +[29] Joseph W Goodman. Statistical optics. New York, Wiley-Interscience, 1985, 567 p., 1, 1985. +[30] Matthew Tancik, Pratul P. Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan T. Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. (arXiv:2006.10739), 2020. +[31] Filipe de Avila Belbute-Peres and J. Zico Kolter. Simple initialization and parametrization of sinusoidal networks via their kernel bandwidth, 2022. +[32] Zhen Liu, Hao Zhu, Qi Zhang, Jingde Fu, Weibing Deng, Zhan Ma, Yanwen Guo, and Xun Cao. Finer: Flexible spectral-bias tuning in implicit neural representation by variable-periodic activation functions, 2023. +[33] John Canny. A computational approach to edge detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, PAMI-8(6):679-698, 1986. +[34] Nathan Halko, Per-Gunnar Martinsson, and Joel A. Tropp. Finding structure with randomness: Probabilistic algorithms for constructing approximate matrix decompositions. SIAM Rev., 53:217-288, 2009. +[35] John Lazzari and Xiuwen Liu. Understanding the spectral bias of coordinate based MLPs via training dynamics. (arXiv:2301.05816), 2023. +[36] Zhemin Li, Hongxia Wang, and Deyu Meng. Regularize implicit neural representation by itself. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10280-10288. IEEE. +[37] Gizem Yüce, Guillermo Ortiz-Jiménez, Beril Besbinar, and Pascal Frossard. A structured dictionary perspective on implicit neural representations. (arXiv:2112.01917), 2022. +[38] Vishwanath Saragadam, Daniel LeJeune, Jasper Tan, Guha Balakrishnan, Ashok Veeraraghavan, and Richard G. Baraniuk. WIRE: Wavelet implicit neural representations. (arXiv:2301.05187), 2023. + +[39] Jeremy Vonderfecht and Feng Liu. Predicting the encoding error of sirens, 2024. +[40] Stanislav Fort, Gintare Karolina Dziugaite, Mansheej Paul, Sepideh Kharaghani, Daniel M. Roy, and Surya Ganguli. Deep learning versus kernel learning: an empirical study of loss landscape geometry and the time evolution of the neural tangent kernel. (arXiv:2010.15110), 2020. +[41] Mario Geiger, Stefano Spigler, Arthur Jacot, and Matthieu Wyart. Disentangling feature and lazy training in deep neural networks. 2020(11):113301. +[42] Shervin Khalafi, Saurabh Sihag, and Alejandro Ribeiro. Neural tangent kernels motivate cross-covariance graphs in neural networks. In Ruslan Salakhutdinov, Zico Kolter, Katherine Heller, Adrian Weller, Nuria Oliver, Jonathan Scarlett, and Felix Berkenkamp, editors, Proceedings of the 41st International Conference on Machine Learning, volume 235 of Proceedings of Machine Learning Research, pages 23577-23621. PMLR, 21-27 Jul 2024. +[43] Johannes Schwab, Bryan T. Kelly, Semyon Malamud, and Teng Andrea Xu. Training ntk to generalize with kare. ArXiv, abs/2505.11347, 2025. +[44] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. CoRR, abs/1412.6980, 2014. +[45] Stefan Van der Walt, Johannes L Schonberger, Juan Nunez-Iglesias, François Boulogne, Joshua D Warner, Neil Yager, Emmanuelle Gouillart, and Tony Yu. scikit-image: image processing in python. PeerJ, 2:e453, 2014. +[46] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255. IEEE, 2009. +[47] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Kopf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. Pytorch: An imperative style, high-performance deep learning library. In Advances in Neural Information Processing Systems 32, pages 8024–8035. Curran Associates, Inc., 2019. +[48] Richard Zou Horace He. functorch: Jax-like composable function transforms for pytorch. 2021. +[49] Fabian Pedregosa, Géel Varoquaux, Alexandre Gramfort, Vincent Michel, Bertrand Thirion, Olivier Grisel, Mathieu Blondel, Peter Prettenhofer, Ron Weiss, Vincent Dubourg, et al. Scikit-learn: Machine learning in python. Journal of machine learning research, 12(Oct):2825-2830, 2011. +[50] Pauli Virtanen, Ralf Gommers, Travis E. Oliphant, Matt Haberland, Tyler Reddy, David Cournapeau, Evgeni Burovski, PEARU Peterson, Warren Weckesser, Jonathan Bright, Stefan J. van der Walt, Matthew Brett, Joshua Wilson, K. Jarrod Millman, Nikolay Mayorov, Andrew R. J. Nelson, Eric Jones, Robert Kern, Eric Larson, C J Carey, Ilhan Polat, Yu Feng, Eric W. Moore, Jake VanderPlas, Denis Laxalde, Josef Perktold, Robert Cimrman, Ian Henriksen, E. A. Quintero, Charles R. Harris, Anne M. Archibald, Antonio H. Ribeiro, Fabian Pedregosa, Paul van Mulbregt, and SciPy 1.0 Contributors. SciPy 1.0: Fundamental Algorithms for Scientific Computing in Python. Nature Methods, 17:261-272, 2020. +[51] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis, 2020. +[52] C. Harris and M. Stephens. A combined corner and edge detector. In Proceedings of the Alvey Vision Conference, pages 23.1–23.6. Alvety Vision Club, 1988. doi:10.5244/C.2.23. + +# NeurIPS Paper Checklist + +The checklist is designed to encourage best practices for responsible machine learning research, addressing issues of reproducibility, transparency, research ethics, and societal impact. Do not remove the checklist: The papers not including the checklist will be desk rejected. The checklist should follow the references and follow the (optional) supplemental material. The checklist does NOT count towards the page limit. + +Please read the checklist guidelines carefully for information on how to answer these questions. For each question in the checklist: + +- You should answer [Yes], [No], or [NA]. +- [NA] means either that the question is Not Applicable for that particular paper or the relevant information is Not Available. +- Please provide a short (1–2 sentence) justification right after your answer (even for NA). + +The checklist answers are an integral part of your paper submission. They are visible to the reviewers, area chairs, senior area chairs, and ethics reviewers. You will be asked to also include it (after eventual revisions) with the final version of your paper, and its final version will be published with the paper. + +The reviewers of your paper will be asked to use the checklist as one of the factors in their evaluation. While "[Yes]" is generally preferable to "[No]", it is perfectly acceptable to answer "[No]" provided a proper justification is given (e.g., "error bars are not reported because it would be too computationally expensive" or "we were unable to find the license for the dataset we used"). In general, answering "[No]" or "[NA]" is not grounds for rejection. While the questions are phrased in a binary way, we acknowledge that the true answer is often more nuanced, so please just use your best judgment and write a justification to elaborate. All supporting evidence can appear either in the main paper or the supplemental material, provided in appendix. If you answer [Yes] to a question, in the justification please point to the section(s) where related material for the question can be found. + +# 1. Claims + +Question: Do the main claims made in the abstract and introduction accurately reflect the paper's contributions and scope? + +Answer: [Yes] + +Justification: Our main claims in the abstract are that (1) we derive a local approximation of the NTK for SIRENs (all of Section 3); (2) we construct order parameters from this approximation, namely, in Section 3.1 for diffusion wavecrests, loss rate collapse (Section 3.4), and NTK alignment (3.3). We discuss SIREN biases in section 4.1 and 4.2 and the Discussion (Section 5). Considering more general deep image regression, we compare against ReLU+PE in Appendix D. + +Guidelines: + +- The answer NA means that the abstract and introduction do not include the claims made in the paper. +- The abstract and/or introduction should clearly state the claims made, including the contributions made in the paper and important assumptions and limitations. A No or NA answer to this question will not be perceived well by the reviewers. +- The claims made should match theoretical and experimental results, and reflect how much the results can be expected to generalize to other settings. +- It is fine to include aspirational goals as motivation as long as it is clear that these goals are not attained by the paper. + +# 2. Limitations + +Question: Does the paper discuss the limitations of the work performed by the authors? + +Answer: [Yes] + +Justification: In addition to the conclusion (which clarifies we only consider full batch GD), we make numerous assumptions outlined in Section 3 that we only know to hold for + +SIRENs (which we validate in Appendix F). What's more, as outlined in the introduction, our analysis is only tractable because we are working in low dimensional domains (ie image regression). + +# Guidelines: + +- The answer NA means that the paper has no limitation while the answer No means that the paper has limitations, but those are not discussed in the paper. +- The authors are encouraged to create a separate "Limitations" section in their paper. +- The paper should point out any strong assumptions and how robust the results are to violations of these assumptions (e.g., independence assumptions, noiseless settings, model well-specification, asymptotic approximations only holding locally). The authors should reflect on how these assumptions might be violated in practice and what the implications would be. +- The authors should reflect on the scope of the claims made, e.g., if the approach was only tested on a few datasets or with a few runs. In general, empirical results often depend on implicit assumptions, which should be articulated. +- The authors should reflect on the factors that influence the performance of the approach. For example, a facial recognition algorithm may perform poorly when image resolution is low or images are taken in low lighting. Or a speech-to-text system might not be used reliably to provide closed captions for online lectures because it fails to handle technical jargon. +- The authors should discuss the computational efficiency of the proposed algorithms and how they scale with dataset size. +- If applicable, the authors should discuss possible limitations of their approach to address problems of privacy and fairness. +- While the authors might fear that complete honesty about limitations might be used by reviewers as grounds for rejection, a worse outcome might be that reviewers discover limitations that aren't acknowledged in the paper. The authors should use their best judgment and recognize that individual actions in favor of transparency play an important role in developing norms that preserve the integrity of the community. Reviewers will be specifically instructed to not penalize honesty concerning limitations. + +# 3. Theory assumptions and proofs + +Question: For each theoretical result, does the paper provide the full set of assumptions and a complete (and correct) proof? + +Answer: [Yes] + +Justification: The main paper, this is all in Section 3. Extra proofs found in Appendix A. + +Guidelines: + +- The answer NA means that the paper does not include theoretical results. +- All the theorems, formulas, and proofs in the paper should be numbered and cross-referenced. +- All assumptions should be clearly stated or referenced in the statement of any theorems. +- The proofs can either appear in the main paper or the supplemental material, but if they appear in the supplemental material, the authors are encouraged to provide a short proof sketch to provide intuition. +- Inversely, any informal proof provided in the core of the paper should be complemented by formal proofs provided in appendix or supplemental material. +- Theorems and Lemmas that the proof relies upon should be properly referenced. + +# 4. Experimental result reproducibility + +Question: Does the paper fully disclose all the information needed to reproduce the main experimental results of the paper to the extent that it affects the main claims and/or conclusions of the paper (regardless of whether the code and data are provided or not)? + +Answer: [Yes] + +Justification: Section 4 begins with a summary of our setup, and Appendix B gives the full details. Our actual experiments are very simple: training a variety of SIREN models and tracking summary statistics of the NTK. The bulk of the paper is derivations. + +# Guidelines: + +- The answer NA means that the paper does not include experiments. +- If the paper includes experiments, a No answer to this question will not be perceived well by the reviewers: Making the paper reproducible is important, regardless of whether the code and data are provided or not. +- If the contribution is a dataset and/or model, the authors should describe the steps taken to make their results reproducible or verifiable. +- Depending on the contribution, reproducibility can be accomplished in various ways. For example, if the contribution is a novel architecture, describing the architecture fully might suffice, or if the contribution is a specific model and empirical evaluation, it may be necessary to either make it possible for others to replicate the model with the same dataset, or provide access to the model. In general, releasing code and data is often one good way to accomplish this, but reproducibility can also be provided via detailed instructions for how to replicate the results, access to a hosted model (e.g., in the case of a large language model), releasing of a model checkpoint, or other means that are appropriate to the research performed. +- While NeurIPS does not require releasing code, the conference does require all submissions to provide some reasonable avenue for reproducibility, which may depend on the nature of the contribution. For example +(a) If the contribution is primarily a new algorithm, the paper should make it clear how to reproduce that algorithm. +(b) If the contribution is primarily a new model architecture, the paper should describe the architecture clearly and fully. +(c) If the contribution is a new model (e.g., a large language model), then there should either be a way to access this model for reproducing the results or a way to reproduce the model (e.g., with an open-source dataset or instructions for how to construct the dataset). +(d) We recognize that reproducibility may be tricky in some cases, in which case authors are welcome to describe the particular way they provide for reproducibility. In the case of closed-source models, it may be that access to the model is limited in some way (e.g., to registered users), but it should be possible for other researchers to have some path to reproducing or verifying the results. + +# 5. Open access to data and code + +Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material? + +Answer: [Yes] + +Justification: We will be releasing code as part of the supplementary materials, namely, our training code, monitoring code, and analysis code. + +# Guidelines: + +- The answer NA means that paper does not include experiments requiring code. +- Please see the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details. +- While we encourage the release of code and data, we understand that this might not be possible, so "No" is an acceptable answer. Papers cannot be rejected simply for not including code, unless this is central to the contribution (e.g., for a new open-source benchmark). +- The instructions should contain the exact command and environment needed to run to reproduce the results. See the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details. +- The authors should provide instructions on data access and preparation, including how to access the raw data, preprocessed data, intermediate data, and generated data, etc. +- The authors should provide scripts to reproduce all experimental results for the new proposed method and baselines. If only a subset of experiments are reproducible, they should state which ones are omitted from the script and why. + +- At submission time, to preserve anonymity, the authors should release anonymized versions (if applicable). +- Providing as much information as possible in supplemental material (appended to the paper) is recommended, but including URLs to data and code is permitted. + +# 6. Experimental setting/details + +Question: Does the paper specify all the training and test details (e.g., data splits, hyperparameters, how they were chosen, type of optimizer, etc.) necessary to understand the results? + +Answer: [Yes] + +Justification: We search over a range of hyperparameters summarized at the beginning of Section 4, and in detail in Appendix B. This is for exploration, rather than optimization, as this paper is about considering the inductive biases induced by different hyperparameters. Re optimizer we only use full batch GD (described in Section 2) with a small, constant learning rate described in Appendix B. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The experimental setting should be presented in the core of the paper to a level of detail that is necessary to appreciate the results and make sense of them. +- The full details can be provided either with the code, in appendix, or as supplemental material. + +# 7. Experiment statistical significance + +Question: Does the paper report error bars suitably and correctly defined or other appropriate information about the statistical significance of the experiments? + +Answer: [Yes] + +Justification: Table 1 summarizes results from across several runs (aggregated across architectures, datasets, and random seeds). The heatmap in Figure 3 is obtained through analysis of the confidence bounds we used to detect phase transitions. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The authors should answer "Yes" if the results are accompanied by error bars, confidence intervals, or statistical significance tests, at least for the experiments that support the main claims of the paper. +- The factors of variability that the error bars are capturing should be clearly stated (for example, train/test split, initialization, random drawing of some parameter, or overall run with given experimental conditions). +- The method for calculating the error bars should be explained (closed form formula call to a library function, bootstrap, etc.) +- The assumptions made should be given (e.g., Normally distributed errors). +- It should be clear whether the error bar is the standard deviation or the standard error of the mean. +- It is OK to report 1-sigma error bars, but one should state it. The authors should preferably report a 2-sigma error bar than state that they have a $96\%$ CI, if the hypothesis of Normality of errors is not verified. +- For asymmetric distributions, the authors should be careful not to show in tables or figures symmetric error bars that would yield results that are out of range (e.g. negative error rates). +- If error bars are reported in tables or plots, The authors should explain in the text how they were calculated and reference the corresponding figures or tables in the text. + +# 8. Experiments compute resources + +Question: For each experiment, does the paper provide sufficient information on the computer resources (type of compute workers, memory, time of execution) needed to reproduce the experiments? + +Answer: [Yes] + +Justification: This is in Appendix B.1. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The paper should indicate the type of compute workers CPU or GPU, internal cluster, or cloud provider, including relevant memory and storage. +- The paper should provide the amount of compute required for each of the individual experimental runs as well as estimate the total compute. +- The paper should disclose whether the full research project required more compute than the experiments reported in the paper (e.g., preliminary or failed experiments that didn't make it into the paper). + +# 9. Code of ethics + +Question: Does the research conducted in the paper conform, in every respect, with the NeurIPS Code of Ethics https://neurips.cc/public/EthicsGuidelines? + +Answer: [Yes] + +Justification: Our work doesn't involve any human subjects, and is fairly concentrated in its scope to theoretical aspects of learning dynamics in 2D image regression problems. Our research is not of a critical nature, though it can potentially be used to help interpret image regression models, and thus could be the foundation of tools to improve safety. + +Guidelines: + +- The answer NA means that the authors have not reviewed the NeurIPS Code of Ethics. +- If the authors answer No, they should explain the special circumstances that require a deviation from the Code of Ethics. +- The authors should make sure to preserve anonymity (e.g., if there is a special consideration due to laws or regulations in their jurisdiction). + +# 10. Broader impacts + +Question: Does the paper discuss both potential positive societal impacts and negative societal impacts of the work performed? + +Answer: [NA] + +Justification: Our work is of a theoretical nature specialized to very specific models, and we do not believe it has wider societal impact. + +Guidelines: + +- The answer NA means that there is no societal impact of the work performed. +- If the authors answer NA or No, they should explain why their work has no societal impact or why the paper does not address societal impact. +- Examples of negative societal impacts include potential malicious or unintended uses (e.g., disinformation, generating fake profiles, surveillance), fairness considerations (e.g., deployment of technologies that could make decisions that unfairly impact specific groups), privacy considerations, and security considerations. +- The conference expects that many papers will be foundational research and not tied to particular applications, let alone deployments. However, if there is a direct path to any negative applications, the authors should point it out. For example, it is legitimate to point out that an improvement in the quality of generative models could be used to generate deepfakes for disinformation. On the other hand, it is not needed to point out that a generic algorithm for optimizing neural networks could enable people to train models that generate Deepfakes faster. +- The authors should consider possible harms that could arise when the technology is being used as intended and functioning correctly, harms that could arise when the technology is being used as intended but gives incorrect results, and harms following from (intentional or unintentional) misuse of the technology. + +- If there are negative societal impacts, the authors could also discuss possible mitigation strategies (e.g., gated release of models, providing defenses in addition to attacks, mechanisms for monitoring misuse, mechanisms to monitor how a system learns from feedback over time, improving the efficiency and accessibility of ML). + +# 11. Safeguards + +Question: Does the paper describe safeguards that have been put in place for responsible release of data or models that have a high risk for misuse (e.g., pretrained language models, image generators, or scraped datasets)? + +Answer: [NA] + +Justification: We do not believe our research warrants this. It is a theoretical investigation on the dynamics of learning in a specific class of image regression models. We do not anticipate that our code or our derivations can be misused. + +Guidelines: + +- The answer NA means that the paper poses no such risks. +- Released models that have a high risk for misuse or dual-use should be released with necessary safeguards to allow for controlled use of the model, for example by requiring that users adhere to usage guidelines or restrictions to access the model or implementing safety filters. +- Datasets that have been scraped from the Internet could pose safety risks. The authors should describe how they avoided releasing unsafe images. +- We recognize that providing effective safeguards is challenging, and many papers do not require this, but we encourage authors to take this into account and make a best faith effort. + +# 12. Licenses for existing assets + +Question: Are the creators or original owners of assets (e.g., code, data, models), used in the paper, properly credited and are the license and terms of use explicitly mentioned and properly respected? + +Answer: [Yes] + +Justification: Appendix B outlines the images used from public datasets, and the means through which we obtained them. We make use of no other assets, beyond pytorch and the scipy ecosystem, which we cite. + +Guidelines: + +- The answer NA means that the paper does not use existing assets. +- The authors should cite the original paper that produced the code package or dataset. +- The authors should state which version of the asset is used and, if possible, include a URL. +- The name of the license (e.g., CC-BY 4.0) should be included for each asset. +- For scraped data from a particular source (e.g., website), the copyright and terms of service of that source should be provided. +- If assets are released, the license, copyright information, and terms of use in the package should be provided. For popular datasets, paperswithcode.com/datasets has curated licenses for some datasets. Their licensing guide can help determine the license of a dataset. +- For existing datasets that are re-packaged, both the original license and the license of the derived asset (if it has changed) should be provided. +- If this information is not available online, the authors are encouraged to reach out to the asset's creators. + +# 13. New assets + +Question: Are new assets introduced in the paper well documented and is the documentation provided alongside the assets? + +Answer: [NA] + +Justification: We do not introduce new assets. + +# Guidelines: + +- The answer NA means that the paper does not release new assets. +- Researchers should communicate the details of the dataset/code/model as part of their submissions via structured templates. This includes details about training, license, limitations, etc. +- The paper should discuss whether and how consent was obtained from people whose asset is used. +- At submission time, remember to anonymize your assets (if applicable). You can either create an anonymized URL or include an anonymized zip file. + +# 14. Crowdsourcing and research with human subjects + +Question: For crowdsourcing experiments and research with human subjects, does the paper include the full text of instructions given to participants and screenshots, if applicable, as well as details about compensation (if any)? + +Answer: [NA] + +Justification: We don't do any crowdsourcing or work with human subjects. + +# Guidelines: + +- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. +- Including this information in the supplemental material is fine, but if the main contribution of the paper involves human subjects, then as much detail as possible should be included in the main paper. +- According to the NeurIPS Code of Ethics, workers involved in data collection, curation, or other labor should be paid at least the minimum wage in the country of the data collector. + +# 15. Institutional review board (IRB) approvals or equivalent for research with human subjects + +Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals (or an equivalent approval/review based on the requirements of your country or institution) were obtained? + +Answer: [NA] + +Justification: There were no study participants. + +# Guidelines: + +- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. +- Depending on the country in which research is conducted, IRB approval (or equivalent) may be required for any human subjects research. If you obtained IRB approval, you should clearly state this in the paper. +- We recognize that the procedures for this may vary significantly between institutions and locations, and we expect authors to adhere to the NeurIPS Code of Ethics and the guidelines for their institution. +- For initial submissions, do not include any information that would break anonymity (if applicable), such as the institution conducting the review. + +# 16. Declaration of LLM usage + +Question: Does the paper describe the usage of LLMs if it is an important, original, or non-standard component of the core methods in this research? Note that if the LLM is used only for writing, editing, or formatting purposes and does not impact the core methodology, scientific rigorousness, or originality of the research, declaration is not required. + +Answer: [NA] + +Justification: LLMs are not used in this paper. + +# Guidelines: + +- The answer NA means that the core method development in this research does not involve LLMs as any important, original, or non-standard components. +- Please refer to our LLM policy (https://neurips.cc/Conferences/2025/LLM) for what should or should not be described. + +# Supplementary Materials + +# A Deferred Proofs 22 + +A.1 Decomposition of the NTK over layers 22 +A.2 Proof of Theorem 3.1: Diffusive Evolution of the Residuals 23 +A.3 Proof of Theorem 3.2: Local Cauchy Approximation of the $C_{NTK}$ 24 +A.4 Proof of Corollary 3.2.1: Obtaining the Correlation Lengthscale from the Cauchy Approximation 26 +A.5 Proof of Corollary 3.2.3: Minimum Value of $C_{NTK}$ 27 +A.6 Relating Loss Gradient Variance to the NTK 28 + +# B Experimental Details 28 + +B.1 Model Training 28 +B.2 Order Parameter Estimation 29 + +# C Occurrence Rates of Phase Transitions 30 + +C.1 Impact of Image Features 30 +C.2 Additional Figures: Impact of Hyperparameters 31 + +# D Comparison with ReLU Activations 32 + +# E Implications of Local Image Structure on Feature Learning 34 + +E.1 On the Relationship Between Structure Tensors and Tangent Kernels 34 +E.2 MAG-MA: Order Parameters From Translational Symmetry Breaking 36 + +# F Evaluating Fidelity of Approximation 36 + +F.1 Local Structure of the NTK 36 +F.2 Cauchy Approximation 37 + +# G Additional Experimental Results 38 + +G.1 Order Parameter Trajectories for Single Runs 38 +G.2 Influence of Hyperparameters on Order Parameter Trajectories 38 + +# A Deferred Proofs + +# A.1 Decomposition of the NTK over layers + +Consider a feedforward neural network, denoted by $f(x) = h^{(L)} \circ \ldots \circ h^{(1)}(x)$ . We furthermore define: + +$$ +z ^ {(l)} = \left[ W ^ {(l)} \right] ^ {\top} h ^ {(l - 1)} + b ^ {(l)} \tag {24} +$$ + +$$ +h ^ {(l)} = \sigma \left(z ^ {(l)}\right) \tag {25} +$$ + +In this way, we may calculate the parametric gradients as follows: + +$$ +\nabla_ {W ^ {(l)}} f = \frac {\partial f}{\partial z ^ {(l)}} [ h ^ {(l - 1)} ] ^ {\top} \tag {26} +$$ + +$$ +\operatorname {v e c} \left(\nabla_ {W ^ {(l)}} f\right) = \left(I \otimes \frac {\partial f}{\partial z ^ {(l)}}\right) h ^ {(l - 1)} \tag {27} +$$ + +$$ +\begin{array}{l} \operatorname {v e c} \left(\nabla_ {W ^ {(l)}} f (x _ {i})\right) ^ {\top} \operatorname {v e c} \left(\nabla_ {W ^ {(l)}} f (x _ {j})\right) = h ^ {(l - 1)} \left(x _ {i}\right) ^ {\top} \left(I \otimes \frac {\partial f \left(x _ {i}\right)}{\partial z ^ {(l)}} ^ {\top}\right) \left(I \otimes \frac {\partial f (x _ {j})}{\partial z ^ {(l)}}\right) h ^ {(l - 1)} \left(x _ {j}\right) (28) \\ = h ^ {(l - 1)} \left(x _ {i}\right) ^ {\top} \left(I \otimes \frac {\partial f \left(x _ {i}\right)}{\partial z ^ {(l)}} ^ {\top} \frac {\partial f \left(x _ {j}\right)}{\partial z ^ {(l)}}\right) h ^ {(l - 1)} \left(x _ {j}\right) (29) \\ = \left(\frac {\partial f \left(x _ {i}\right)}{\partial z ^ {(l)}} ^ {\top} \frac {\partial f \left(x _ {j}\right)}{\partial z ^ {(l)}}\right) \left(h ^ {(l - 1)} \left(x _ {i}\right) ^ {\top} h ^ {(l - 1)} \left(x _ {j}\right)\right) (30) \\ \end{array} +$$ + +The first term in this product defines functional similarity between points, while the second defines representational similarity. Thinking of each term as a separate kernel, the overall layer kernel - ie the product is defined via an AND operation. A similar formula holds for the other layers. For the biases, we have, more simply: + +$$ +\nabla_ {b ^ {(l)}} f = \frac {\partial f}{\partial z ^ {(l)}} \tag {31} +$$ + +$$ +\nabla_ {b ^ {(l)}} f \left(x _ {i}\right) ^ {\top} \nabla_ {b ^ {(l)}} f \left(x _ {j}\right) = \frac {\partial f \left(x _ {i}\right)}{\partial z ^ {(l)}} ^ {\top} \frac {\partial f \left(x _ {j}\right)}{\partial z ^ {(l)}} \tag {32} +$$ + +The full NTK is then given simply by: + +$$ +\begin{array}{l} K \left(x _ {i}, x _ {j}; \theta\right) = \sum_ {l = 1} ^ {N _ {l}} \left(\frac {\partial f \left(x _ {i}\right)}{\partial z ^ {(l)}} ^ {\top} \frac {\partial f \left(x _ {j}\right)}{\partial z ^ {(l)}}\right) \left(1 + h ^ {(l - 1)} \left(x _ {i}\right) ^ {\top} h ^ {(l - 1)} \left(x _ {j}\right)\right) (33) \\ \equiv \sum_ {l = 1} ^ {N _ {l}} K ^ {(l)} \left(x _ {i}, x _ {j}\right) (34) \\ \end{array} +$$ + +In particular, we have: + +$$ +K (x, x; \theta) = \sum_ {l = 1} ^ {N _ {l}} \left\| \frac {\partial f (x)}{\partial z ^ {(l)}} \right\| _ {2} ^ {2} \left(1 + \left\| h ^ {(l - 1)} (x) \right\| _ {2} ^ {2}\right) \tag {35} +$$ + +Following the same logic, the full NTK is defined as an OR over all the layers. For INRs, these layers tend to be frequency separated, so that lower layers correspond to lower frequencies. + +# A.2 Proof of Theorem 3.1: Diffusive Evolution of the Residuals + +The motivation for our ansatz in equation 10 is the empirical form of the correlation function in equation 7. Written fully, we have: + +$$ +K (x, x + u) \approx | | \nabla_ {\theta} f (x) | | ^ {2} \exp (- \| u \| ^ {2} / \xi^ {2} (x)) + | | \nabla_ {\theta} f (x) | | ^ {2} c _ {\infty} (x) \tag {36} +$$ + +Thus the residuals evolve according to: + +$$ +\begin{array}{l} \dot {r} (x) = - \int d u r (x + u) K (x, x + u) (37) \\ \approx - \left| | \nabla_ {\theta} f (x) | \right| ^ {2} \left[ \int d u \exp (- | | u | | ^ {2} / \xi^ {2} (x)) r (x + u) - c _ {\infty} (x) \int d u r (x + u) \right] (38) \\ = \left\| \nabla_ {\theta} f (x) \right\| ^ {2} \left[ \int d u \exp (- \| u \| ^ {2} / \xi^ {2} (x)) r (x + u) - \mu_ {r} c _ {\infty} (x) \operatorname {V o l} (\mathcal {D}) \right] (39) \\ \end{array} +$$ + +When $\mu_r \equiv \mathbb{E}[r]$ and $c_{\infty}$ decay to zero, the second, background term in the above equation becomes dominated by local interactions. Thus, in Section 4, we will track the following order parameter: + +$$ +\mu_ {r} K _ {\infty} \equiv | \mathbb {E} _ {x} [ r ] \mathbb {E} _ {x} [ \| \nabla_ {\theta} f (x) \| ^ {2} c _ {\infty} (x) ] | \tag {40} +$$ + +The order parameter is large in the Drift phase, and small in the Diffusion phase. In Section B, we overview the specifics of how we detect changes in the phase. For the remainder of this section, we analytically study kernels of the form: + +$$ +\begin{array}{l} K (x, x + u) = A (x) e ^ {- u ^ {2} / 2 \xi^ {2} (x)} (41) \\ = 2 \pi \xi^ {2} (x) A (x) \mathcal {N} (u; 0, \xi^ {2} (x) I) (42) \\ \end{array} +$$ + +That is, kernels without a background term. Here, $\mathcal{N}(u;\mu ,\Sigma)$ denotes the $d$ -dimensional; multivariate Gaussian Distribution: + +$$ +\mathcal {N} (u; \mu (x), \Sigma (x)) = \frac {1}{\sqrt {(2 \pi) ^ {d} \det \Sigma (x)}} \exp \left(- \frac {1}{2} (u - \mu (x)) ^ {\top} \Sigma^ {- 1} (x) (u - \mu (x))\right) \tag {43} +$$ + +For our case, $d = 2$ , and $\Sigma(x) = \xi^2(x)I$ . The determinant of the covariance is as follows: + +$$ +\det \Sigma (x) = (\xi^ {2}) ^ {2} \det I = \xi^ {4} (x) \tag {44} +$$ + +We now consider the integral of the following quadratic form: + +$$ +\begin{array}{l} \int d u \left(u ^ {\top} H u\right) e ^ {- u ^ {2} / 2 \xi^ {2} (x)} = 2 \pi \xi^ {2} (x) \int d u \left(u ^ {\top} H u\right) \mathcal {N} (u; 0, \xi^ {2} (x) I) (45) \\ = 2 \pi \xi^ {2} (x) \mathbb {E} _ {\mathcal {N} (u; 0, \xi^ {2} I)} \left[ u ^ {\top} H u \right] (46) \\ = 2 \pi \xi^ {2} (x) \operatorname {t r} \left(H \Sigma (x)\right) (47) \\ = 2 \pi \xi^ {4} (x) \operatorname {t r} (H) (48) \\ \end{array} +$$ + +Now, let's look at the following Taylor expansion: + +$$ +r (x + u) \approx r (x) + u ^ {\top} \nabla_ {x} r + \frac {1}{2} u ^ {\top} \left(\nabla_ {x} ^ {2} r\right) u \tag {49} +$$ + +When integrating the above in equation 3, the second term vanishes, because it involves a product of symmetric and anti-symmetric functions. Thus, we have + +$$ +\begin{array}{l} \int d u r (x + u) K (x, x + u) = A (x) \int d u r (x + u) e ^ {- u ^ {2} / 2 \xi^ {2} (x)} (50) \\ = A (x) \int d u \left[ r (x) e ^ {- u ^ {2} / 2 \xi^ {2} (x)} + \frac {1}{2} u ^ {\top} \left(\nabla_ {x} ^ {2} r\right) u e ^ {- u ^ {2} / 2 \xi^ {2} (x)} \right] (51) \\ \end{array} +$$ + +Leveraging our result for the quadratic term, we have, finally: + +$$ +\begin{array}{l} \int d u r (x + u) K (x, x + u) = 2 \pi \xi^ {2} (x) A (x) r (x) + \pi \xi^ {4} (x) A (x) \operatorname {t r} \left(\nabla_ {x} ^ {2} r\right) (52) \\ = 2 \pi \xi^ {2} (x) A (x) r (x) + \pi \xi^ {4} (x) A (x) \Delta^ {2} r (53) \\ \end{array} +$$ + +Thus, the diffusion equation becomes: + +$$ +\dot {r} = - 2 \pi \xi^ {2} (x) A (x) r (x) - \pi \xi^ {4} (x) A (x) \Delta^ {2} r +$$ + +# A.3 Proof of Theorem 3.2: Local Cauchy Approximation of the $C_{NTK}$ + +# A.3.1 Notation and Derivation + +We consider an arbitrary vector valued function $f(x)$ , and consider the cosine of the angle between $f(x)$ and $f(x + u)$ for small displacements $u$ . To ease notation, let us make use of the following shorthands: + +$$ +\begin{array}{l} a = f (x) (54) \\ b = f (x + u) (55) \\ c = b - a (56) \\ J = \nabla_ {x} a (57) \\ D = \nabla_ {x} \| a \| ^ {2} (58) \\ \end{array} +$$ + +To first order in $u$ , we have: + +$$ +b \approx a + u ^ {\top} J \tag {59} +$$ + +$$ +c \approx u ^ {\top} J \tag {60} +$$ + +$$ +\left| \left| b \right| \right| ^ {2} \approx \left| \left| a + u ^ {\top} J \right| \right| ^ {2} \tag {61} +$$ + +$$ += \left\| a \right\| ^ {2} + u ^ {\top} D + \left\| u ^ {\top} J \right\| ^ {2} \tag {62} +$$ + +Our goal is to discern the local behaviour of the cosine of the angle $\theta$ between $a$ and $b$ (as illustrated in Figure 5). To that end, our starting point is the law of cosines: + +![](images/294e2bf465d4857cb7abca70b5e4adb8f581f21c169eeaf38a9219c23a663b9e.jpg) +Figure 5: Triangle with vectors a, b, and b - a, inscribed in a circumcircle. + +$$ +\cos \phi = \frac {\left| \left| a \right| \right| ^ {2} + \left| \left| b \right| \right| ^ {2} - \left| \left| c \right| \right| ^ {2}}{2 \left| \left| a \right| \right| \left| \left| b \right| \right|} \tag {63} +$$ + +$$ +\approx \frac {2 \| a \| ^ {2} + u ^ {\top} D}{2 \| a \| ^ {2}} \left(1 + \frac {u ^ {\top} D}{\| a \| ^ {2}} + \frac {\| u ^ {\top} J \| ^ {2}}{\| a \| ^ {2}}\right) ^ {- \frac {1}{2}} \tag {64} +$$ + +To proceed, note that, for small scalar $\epsilon$ , we have the following identity: + +$$ +(1 + \epsilon) ^ {\frac {1}{2}} \approx 1 + \frac {\epsilon}{2} - \frac {\epsilon^ {2}}{8} \tag {65} +$$ + +Thus: + +$$ +\cos \phi \approx \frac {2 \| a \| ^ {2} + u ^ {\top} D}{2 \| a \| ^ {2} + u ^ {\top} D + \| u ^ {\top} J \| ^ {2} - \frac {1}{1 6 \| a \| ^ {2}} (u ^ {\top} D) ^ {2}} \tag {67} +$$ + +For the NTK, where we will have $a = \nabla_{\theta}f$ , $||a||$ is so large that we may ignore the term of order $||a||^{-2}$ . We illustrate our approximation in Figure 6. + +# A.3.2 Specialization for Feed Forward Neural Networks + +We want to consider the case where, per our previous derivation, $a = \nabla_{\theta}f$ . This procedure is straightforward for the biases. For the weights $W_{ij}^{(l)}$ , we have: + +$$ +\frac {\partial f (x ; \theta)}{\partial W _ {i j} ^ {(l)}} = \frac {\partial f (x ; \theta)}{\partial z _ {i} ^ {(l)}} h _ {j} ^ {(l - 1)} (x; \theta) \tag {68} +$$ + +![](images/ee712c9f9df3374958b50ce17670cdc1668d02f694f4e2b000c837771e48b044.jpg) +Figure 6: Cauchy Approximation of the Cosine NTK. Left: Sample image, and test point $x = A$ . Middle: visualization of $C_{NTK}(x, x + u)$ in the vicinity of the point $A$ for small separations $u$ . Right: the Cauchy approximation, capturing both the range, orientation, and the local minima of the true $C_{NTK}$ . + +![](images/f6a066ad846b678165ff81c23566e3b50d02c4504207f12b993d8b1d2be5d3c5.jpg) + +![](images/dc2564b9b749a6ea6ee8462130acad1038674b4904b4752d37e00b807c193291.jpg) + +Therefore: + +$$ +\begin{array}{l} \frac {\partial f ^ {2} (x ; \theta)}{\partial x _ {m} \partial W _ {i j} ^ {(l)}} = \frac {\partial^ {2} f (x ; \theta)}{\partial x _ {m} \partial z _ {i} ^ {(l)}} h _ {j} ^ {(l - 1)} (x; \theta) + \frac {\partial f (x ; \theta)}{\partial z _ {i} ^ {(l)}} \frac {\partial h _ {j} ^ {(l - 1)} (x ; \theta)}{\partial x _ {m}} (69) \\ \triangleq \left(J _ {z} ^ {(l)}\right) _ {i m} h _ {j} ^ {(l - 1)} + \left(J _ {h} ^ {(l - 1)}\right) _ {j m} \partial_ {z _ {i} ^ {(l)}} f (70) \\ \end{array} +$$ + +Before proceeding, let us note that the following holds: + +$$ +\sum_ {i} \left(J _ {z} ^ {(l)}\right) _ {i m} \left(\partial_ {z _ {i} ^ {(l)}} f\right) = \frac {1}{2} \partial_ {x _ {m}} \left| \left| \nabla_ {z ^ {(l)}} f \right| \right| ^ {2} \tag {71} +$$ + +$$ +\sum_ {i} \left(J _ {h} ^ {(l)}\right) _ {i m} h _ {i} ^ {(l)} = \frac {1}{2} \partial_ {x _ {m}} \left\| h ^ {(l)} \right\| ^ {2} \tag {72} +$$ + +The covariance matrix in our Gaussian approximation is thus given by: + +$$ +\begin{array}{l} H _ {W} ^ {(l)} = \sum_ {i, j} \frac {\partial f ^ {2}}{\partial x _ {m} \partial W _ {i j} ^ {(l)}} \frac {\partial f ^ {2}}{\partial x _ {n} \partial W _ {i j} ^ {(l)}} (73) \\ = \sum_ {i, j} \left(h _ {j} ^ {(l - 1)}\right) ^ {2} \left(J _ {z} ^ {(l)}\right) _ {i m} \left(J _ {z} ^ {(l)}\right) _ {i n} + \left(\partial_ {z _ {i} ^ {(l)}} f\right) ^ {2} \left(J _ {h} ^ {(l - 1)}\right) _ {j m} \left(J _ {h} ^ {(l - 1)}\right) _ {j n} (74) \\ + (J _ {z} ^ {(l)}) _ {i m} (\partial_ {z _ {i} ^ {(l)}} f) (J _ {h} ^ {(l - 1)}) _ {j n} h _ {j} ^ {(l - 1)} + (J _ {z} ^ {(l)}) _ {i n} (\partial_ {z _ {i} ^ {(l)}} f) (J _ {h} ^ {(l - 1)}) _ {j m} h _ {j} ^ {(l - 1)} \\ = \left\| h ^ {(l - 1)} \right\| ^ {2} J _ {z} ^ {(l)} J _ {z} ^ {(l) \top} + \left\| \nabla_ {z ^ {(l)}} f \right\| ^ {2} J _ {h} ^ {(l - 1)} \left(J _ {h} ^ {(l - 1)}\right) ^ {\top} (75) \\ + \frac {1}{4} \nabla_ {x} | | h ^ {(l - 1)} | | ^ {2} \otimes \nabla_ {x} | | \nabla_ {z ^ {(l)}} f | | ^ {2} + \frac {1}{4} \nabla_ {x} | | \nabla_ {z ^ {(l)}} f | | ^ {2} \otimes \nabla_ {x} | | h ^ {(l - 1)} | | ^ {2} \\ \end{array} +$$ + +The contribution from the bias is comparatively simple: + +$$ +H _ {b} ^ {(l)} = J _ {z} J _ {z} ^ {\top} \tag {76} +$$ + +# A.4 Proof of Corollary 3.2.1: Obtaining the Correlation Lengthscale from the Cauchy Approximation + +To determine the level sets of the Cauchy Approximation, we must solve: + +$$ +C _ {N T K} (x, x + u) = \frac {2 a _ {x} ^ {2} + u ^ {\top} D _ {x}}{2 a _ {x} ^ {2} + u _ {x} ^ {\top} D + u ^ {\top} H _ {x} u} = c \tag {77} +$$ + +Rearranging, and collecting terms, we have: + +$$ +2 a _ {x} ^ {2} + u ^ {\top} D _ {x} - c \left(2 a _ {x} ^ {2} + u _ {x} ^ {\top} D + u ^ {\top} H _ {x} u\right) = 0 \tag {78} +$$ + +$$ +\Rightarrow 2 (1 - c) a _ {x} ^ {2} - c \left(- \frac {1 - c}{c} u ^ {\top} D _ {x} + u ^ {\top} H _ {x} u\right) = 0 \tag {79} +$$ + +$$ +\Rightarrow u ^ {\top} H _ {x} u - \frac {1 - c}{c} u ^ {\top} D _ {x} = \frac {2 (1 - c)}{c} a _ {x} ^ {2} \tag {80} +$$ + +$$ +\begin{array}{l} \Rightarrow \left(u - \frac {1 - c}{2 c} H ^ {- 1} D\right) ^ {\top} H \left(u - \frac {1 - c}{2 c} H ^ {- 1} D\right) - \frac {(1 - c) ^ {2}}{4 c ^ {2}} D ^ {\top} H ^ {- 1} D = \frac {2 (1 - c)}{c} a _ {x} ^ {2} (81) \\ \Rightarrow \frac {\left(u - \frac {1 - c}{2 c} H ^ {- 1} D\right) ^ {\top} H \left(u - \frac {1 - c}{2 c} H ^ {- 1} D\right)}{\frac {2 (1 - c)}{c} a _ {x} ^ {2} + \frac {(1 - c) ^ {2}}{4 c ^ {2}} D ^ {\top} H ^ {- 1} D} = 1 (82) \\ \end{array} +$$ + +This is the equation of an ellipse centred at $u = \frac{1 - c}{2c} H^{-1}D$ , and with shape matrix: + +$$ +\Sigma_ {s h a p e} = \frac {H}{\frac {2 (1 - c)}{c} a _ {x} ^ {2} + \frac {(1 - c) ^ {2}}{4 c ^ {2}} D ^ {\top} H ^ {- 1} D} \tag {83} +$$ + +The area of this ellipse is (noting that $H$ is a 2x2 matrix): + +$$ +\begin{array}{l} A _ {\text {e l l i p s e}} = \frac {\pi}{\sqrt {\det \Sigma_ {\text {s h a p e}}}} (84) \\ = \frac {1}{\sqrt {\det H}} \left(\frac {2 (1 - c)}{c} a _ {x} ^ {2} + \frac {(1 - c) ^ {2}}{4 c ^ {2}} D ^ {\top} H ^ {- 1} D\right) (85) \\ \end{array} +$$ + +The correlation lengthscale is then obtained from: + +$$ +\xi = \sqrt {A _ {\text {e l l i p s e}} / \pi} \tag {86} +$$ + +# A.5 Proof of Corollary 3.2.3: Minimum Value of $C_{NTK}$ + +We consider minimizing the following function: + +$$ +f (u) = \frac {Q (u)}{P (u)} \tag {87} +$$ + +$$ +Q (u) = 2 a ^ {2} + u ^ {\top} D \tag {88} +$$ + +$$ +P (u) = Q (u) + u ^ {\top} H u \tag {89} +$$ + +Here, $H$ is non-degenerate and positive definite. Thus: + +$$ +\begin{array}{l} \frac {\partial f}{\partial u} = \frac {\partial_ {u} Q P - Q \partial_ {u} P}{P ^ {2}} = 0 (90) \\ \Longrightarrow \partial_ {u} Q P = Q \partial_ {u} P (91) \\ \end{array} +$$ + +Thus: + +$$ +(u ^ {\top} H u) D = \left(4 a ^ {2} + 2 u ^ {\top} D\right) H u \tag {92} +$$ + +Clearly $u = 0$ is a solution, and knowing that our expression locally approximates the cosine, we expect this to be a maximum. To find the other solution, which will be a minima, we take the dot product of both sides of the above equaton with $u$ . After simplifying, we obtain: + +$$ +u ^ {\top} D = - 4 a ^ {2} \tag {93} +$$ + +If we insert this into equation 92, we get: + +$$ +\begin{array}{l} (u ^ {\top} H u) D = - 4 a ^ {2} H u (94) \\ \Rightarrow \left(u ^ {\top} H u\right) H ^ {- 1} D = - 4 a ^ {2} u (95) \\ \Rightarrow (u ^ {\top} H u) \left(D ^ {\top} H ^ {- 1} D\right) = 1 6 a ^ {4} (96) \\ \Rightarrow u ^ {\top} H u = \frac {1 6 a ^ {4}}{D H ^ {- 1} D} (97) \\ \end{array} +$$ + +Armed with an expression for $u^{\top}D$ and $u^{\top}Hu$ , we derive the following formula for the min: + +$$ +\begin{array}{l} f _ {m i n} = \frac {2 a ^ {2} + u ^ {\top} D}{2 a ^ {2} + u ^ {\top} D + u ^ {\top} H u} \Bigg | _ {u = \operatorname {a r g m i n} f} (98) \\ = \frac {2 a ^ {2} - 4 a ^ {2}}{2 a ^ {2} - 4 a ^ {2} + \frac {1 6 a ^ {4}}{D H ^ {- 1} D}} (99) \\ = \frac {D H ^ {- 1} D}{D H ^ {- 1} D - 8 a ^ {2}} (100) \\ \end{array} +$$ + +# A.6 Relating Loss Gradient Variance to the NTK + +Our goal is to quantify the amount of noise in the gradients of the local loss $\mathcal{L}(x_i) = \frac{1}{2} r(x_i;\theta)^2$ . We have, in terms of the Jacobian $J_{ip} = \partial_{\theta_p}f(x_i)$ , the following sample matrix for the gradients: + +$$ +G = R J \tag {101} +$$ + +Here we have defined: + +$$ +R = \operatorname {d i a g} (r) \tag {102} +$$ + +For a dataset with $N$ samples, the sample mean and covariance are given by: + +$$ +\begin{array}{l} \mu = \frac {1}{N} G ^ {\top} 1 _ {N} (103) \\ = \frac {1}{N} J ^ {\top} r (104) \\ \end{array} +$$ + +$$ +C = \frac {1}{N} J ^ {\top} R ^ {2} J - \mu \mu^ {\top} \tag {105} +$$ + +From the cycle property of the trace, we have: + +$$ +\begin{array}{l} \operatorname {t r} \left(J ^ {\top} R ^ {2} J\right) = \operatorname {t r} \left(R ^ {2} J J ^ {\top}\right) (106) \\ = \operatorname {t r} \left(R ^ {2} K _ {N T K}\right). (107) \\ \end{array} +$$ + +We also have: + +$$ +\begin{array}{l} \operatorname {T r} \left(\mu \mu^ {\top}\right) = \left\| \mu \right\| ^ {2} (108) \\ = \frac {1}{N ^ {2}} r ^ {\top} J J ^ {\top} r (109) \\ = \frac {1}{N ^ {2}} r ^ {\top} K _ {N T K} r (110) \\ \end{array} +$$ + +Thus the variance of the loss gradients is given by: + +$$ +\sigma_ {\theta} ^ {2} = \frac {1}{N} \operatorname {T r} \left(R ^ {2} K _ {N T K}\right) - \frac {1}{N ^ {2}} r ^ {\top} K _ {N T K} r \tag {111} +$$ + +# B Experimental Details + +# B.1 Model Training + +All our SIREN models are trained on the images shown in Figure 7, which we obtain through the python package scikit-image [45], and the ImageNet dataset [46]. These images are down-sampled to a resolution of $64 \times 64$ for training, but as a validation task, we track the reconstruction error on the images downsampled to $256 \times 256$ resolution. Our SIREN models are implemented using Pytorch [47], and trained using NVIDIA RTX A6000 48GB GPUs for 10000 epochs, using full batch gradient descent with a learning rate of 1e-3. In our experimental sweeps, we consider the following ranges: + +- Random seeds from interval [0, 5]. +- Width from set $\{64, 128\}$ . +- Depth from set $\{3,4,5\}$ . +- $\omega_0$ from set $\{15, 30, 60, 90\}$ . + +![](images/eada0782b32fc43e61ce9c8d1f92a46e7c9ade10607bf30b7c13b4f610db335f.jpg) +Figure 7: Fifteen of the thirty images used for training INRs + +# B.2 Order Parameter Estimation + +Analytical Order Parameters: To compute the NTK, we use a manual implementation of backpropagation to compute the gradients $\nabla_{z^{(l)}}f(x)$ for each layer, along with the hidden activations $h_{(l)}(x)$ . The NTK is then constructed efficiently using the decomposition across layers outlined in Section A.1. To evaluate the local structure components $a$ and $D$ defined in Theorem 3.2, we obtain the spatial gradients using functorch [48]. We also assemble the $H$ defined in Theorem 3.2 in this way, except we leverage the decomposition outlined in Section A.3.2 to streamline this process, and occupy less memory. + +Empirical Order Parameters: Below we describe the estimation procedure for each of the empirical order parameters. + +- To estimate the correlation functions empirically, we group pairs of datapoints into 50 bins based on a uniform division of the range of distances. Based on the coordinate range, the minimum distance is 0, and the maximum distance is $2\sqrt{2}$ . Within each bin, we evaluate the mean of the $C_{NTK}$ , defining $c(\epsilon)$ . Based on these groups, we estimate our order parameters as follows: + +- To estimate the asymptotic value $c_{\infty}$ , we compute the mean value of $c(\epsilon)$ over the last ten bins (corresponding to points with the furthest separation). +- Given the asymptotic value, we rescale all $c(\epsilon) \to \tilde{c}(\epsilon) = \frac{c(\epsilon)}{1 - c_{\infty}}$ , and then use linear interpolation to find the value of $\epsilon$ for which $\tilde{c}(\epsilon) = 0.5$ , the FWHM. We then have $\xi_{corr} = \frac{\mathrm{FWHM}}{\sqrt{2 \ln 2}}$ . + +- As an additional measure of the correlation length-scale (which we will use in Appendix D), we may calculate the number of points $N_{C}$ for which $C_{NTK}$ is greater than some cutoff (we use $\frac{1}{2}(1 + c_{\infty})$ ). The effective correlation length-scale is then given by $\sqrt{N_{C}dA / \pi}$ , where $dA$ is the area of the coordinate grid cells. We denote this estimate $\xi_{FWHM}$ . + +- To estimate $\mathrm{AUC}(|v_0|, \nabla I)$ , the ground truth edges are identified using the Canny Edge Detector distributed through scikit-image [45]. We then evaluate the Area Under the Receiver Operating Characteristic Curve (ROC AUC) using the implementation in scikit-learn [49]. The principal eigenvector $v_0$ , and the principal eigenvalue $\lambda_0$ , are both computed using our + +own implementation of the Randomized Singular Value Decomposition built with pytorch [47], using 3 iterations and 10 oversamples. + +- To evaluate the Centred Kernel Alignment, in order to prevent zero modes from obscuring alignment, the following centred-variant of the normalized Hilbert-Schmidt Information Criterion (HSIC) is employed: + +$$ +\operatorname {C K A} \left(K, K ^ {\prime}\right) = \frac {\operatorname {T r} \left(K _ {c} K _ {c} ^ {\prime}\right)}{\sqrt {\operatorname {T r} \left(K _ {c} K _ {c}\right) \operatorname {T r} \left(K _ {c} ^ {\prime} K _ {c} ^ {\prime}\right)}} \tag {112} +$$ + +Here, $K_{c}$ denotes that a centering operation has been applied, and is defined as: + +$$ +K _ {c} = \left(I - \frac {1}{n} 1 1 ^ {\top}\right) K \left(I - \frac {1}{n} 1 1 ^ {\top}\right) \tag {113} +$$ + +For both $K_{X}$ and $K_{Y}$ , we use bandwidths $\kappa = 0.1$ . + +- To determine the residual correlations in Table 1, we randomly sample (and flatten) 15000 $15 \times 15$ patches from the validation residuals, and compute the pearson correlation matrix. We then record the mean correlation between all pixels in the patch and the patch centre. + +# Identifying Critical Points: + +- For the gradient variance $\sigma_{\theta}^{2}$ , the loss rate $\dot{L}_{\mathrm{eval}}$ , and the background contribution $\mu_r K_\infty$ the location, and confidence region, for the critical points are identified using the peak detection algorithm distributed through scipy.signal [50]. For the gradient variance, we filter for peaks with a prominence of 0.2, loss rate we use 0.4, and for the background we use 0.2. In the case where multiple peaks are found, we select the peaks which appear closest in time. Finally, for $\mu_r K_\infty$ , the phase transition occurs not at the peak itself, but after the signal decays to zero. Thus we use as confidence region the interval between the identified peak and the right-most boundary. +- For the min $C_{NTK}$ , we linearly interpolate to find the time $t$ where $\min C_{NTK}$ crosses 0. To compute the confidence interval, we also track the cumulative std of $\min C_{NTK}$ , denoted $\epsilon(t)$ . We then use the same linear interpolation strategy to find the times where $\min C_{NTK} = \epsilon(t)$ and $\min C_{NTK} = -\epsilon(t)$ . +- For all other parameters, we fit a sigmoid using the curve fitting function from scipy.optimize, with the default settings. The curve we fit has the form: + +$$ +f (x; A, B, \mu , w) = A + (B - A) \left(1 + e ^ {- (x - \mu) / w}\right) ^ {- 1} \tag {114} +$$ + +We identify the time $t = \mu$ with the critical point, with confidence region defined by $\mu \pm 2w$ . For MAG-MA, where the goal is to detect deviation from zero, we fit this sigmoid to the cumulative STD. + +# C Occurrence Rates of Phase Transitions + +# C.1 Impact of Image Features + +There are three main cases in which a critical point cannot be reliably identified in an order parameter trajectory: + +1. Peaks in the gradient variance $\sigma_{\theta}$ may be absent, or not prominent enough, to be detected using a standard peak detector. +2. A zero-crossing cannot be found for the min $C_{NTK}$ because, at initialization, it is already less than 0. +3. The order parameters do not saturate, and thus, are poorly represented as sigmoids. This is really only a problem for the edge alignment $\mathrm{AUC}(|v_0|, \nabla I)$ and the task alignment $\mathrm{CKA}(K_Y, K_{NTK})$ . In the latter case, in some trials we see CKA steadily decrease after the inflection point of the loss. Numerically, we omit runs where the mean squared error of the fitted sigmoid is greater than 0.01. + +Table 2: Proportion of runs with errors: Frequency at which runs were omitted in constructing Figure 3, as a function of depth and bandwidth $\omega_{9}$ . + +
depthω0AUC(|v0|, ∇I)CKA(KY, KNTK)σθmin CNTK
3900.5700.2930.0071.000
3600.4470.1770.0231.000
3300.5030.1570.0930.700
3150.6430.7730.5830.600
4900.1170.0670.0000.500
4600.0670.0130.0000.500
4300.0430.0570.0370.500
4150.2570.6930.1700.343
5900.0170.0170.0000.470
5600.0370.0030.0000.203
5300.0330.0600.0000.083
5150.0700.6430.0700.093
+ +The occurrence rates, as a function of the hyperparameters used, are shown in Table 2. It is important to note, phase transitions may still occur even during these failure modes - the shift in the order parameter may be simply too weak3 to be identified by the change detection algorithm outlined in Section B.2. It is for this reason that we employ multiple order parameters to identify the same transition (ex the min $C_{NTK}$ and the gradient variance $\theta$ ). Nevertheless, it is instructive to identify what properties of image datasets may be used to predict the aforementioned failure modes. To this end, for each experimental run, we determine if any of the previously mentioned failure modes has occurred, and then record the frequency of success for each image studied. In Figure 8, we see that these frequencies correlate with the complexity of the image, as measured the variance of the spatial gradient magnitudes $||\nabla_x I||$ . Namely, we see that more complex images result in sharper peaks of the parameter gradient variance $\sigma_\theta$ , but collapse of the kernel alignment as measured by $\mathrm{CKA}(K_Y, K_{NTK})$ . This is reflected in their strong negative/positive spearman correlations. These same properties correlate strongly with the best model performance achieved across all hyperparams (left of Figure 8). These correlations give additional support to the mechanism described in Section 3.3, whereby SIREN models struggle to fit edges as they have sharp gradients. Finally, we note that the image complexity seems to have little impact on the error rates for the edge alignment $\mathrm{AUC}(|v_0|, \nabla I)$ (spearman correlation -0.017) and the minimum value of the $C_{NTK}$ (spearman correlation -0.1320). By contrast, these parameters are more sensitive to the model architecture. + +# C.2 Additional Figures: Impact of Hyperparameters + +The broad effects of varying depth and $\omega_0$ on $\mathrm{AUC}(|v_0|, \nabla I)$ are summarized in Table 1. To gain deeper insight into how these parameters influence the principal eigenvector, we examine the two case studies illustrated in Figures 10-11. A lower $\omega_0$ , by broadening the correlation lengthscale, inducing a smoothing effect, retaining only the sharpest edges. Increasing depth also removes noise. + +![](images/6baacd26dc478b1b48ae6ad96b6c843d7bf4aaa63aef347c82399bc852063f9e.jpg) +Figure 8: Image Complexity Affects Detection of Phase Transitions. We measure the image complexity according to the standard deviation of the magnitude of the spatial gradients $(||\nabla_xI||)$ . Dashed red line indicates line of best fit. Legend records spearman correlation. Left: higher complexity images are positively correlated with higher losses (and therefore, worse performance). Middle: higher complexity images do not saturate the target kernel alignment, causing errors in our sigmoidal fits. Right: higher complexity images lead to sharper peaks in the parameter gradient variance, making their identification easier. + +![](images/28978101085c93844eb242bf414551aa115d736cedb750af0899c56d9d6cdc5f.jpg) + +![](images/528d3049d37a4a0c7d81bf509ac0662ebb9c2560fdc63ffc62e54281970705e9.jpg) + +![](images/6d8525b5bb79a8c0efe27e52aff5d61c52b759ba0f078aeeef2571d60a8b38e6.jpg) + +![](images/3bca1e00d8e5fd5203bb1a770e58ab8db229906f3765f303a2353b0ccc16da1e.jpg) + +![](images/41fafc1abc648a95ac9bf5a9d8210a765b2989f61252ab0bafff848e19e1530a.jpg) + +![](images/73c56831c425778129c223c009714ff0c74fa49bd3e20ed2db0a58e54ff4f1e0.jpg) + +![](images/813359f6a96e01aab472f1e76320dad2b0e0512e78bb6704502cf90ea831370c.jpg) + +![](images/8f60dfc46bd9fac9cb2193308169f058130614559f863cca4abffaabc78a751a.jpg) + +![](images/0e058a4f7d688b0702609a066324f43411126c74da0dd39463eda2ab1f796730.jpg) + +![](images/c6968632dc30e444a635b4fbc2179a80063f3d015d2c7989b9f66242b3e4d13e.jpg) + +![](images/ac1949550c23ffb03ebf771f0d8e28a632218980270fcf803c4a8779121c4b1c.jpg) +Figure 9: Variation in NTK Aliment with Hyperparameters (coffee). Principle eigenvectors of the NTK at the end of training. Best performing architecture highlighted in blue. + +![](images/5500f255d72b6111f73fdd43bcba17b0d8ab34171fdf41fa987367b79c9f6315.jpg) + +![](images/350b95fe133d4ee3d4752061d180ac0ad9a804429c04e0445c1c9001fc1539e1.jpg) + +![](images/dd50e2e7aa2a6c499debd0f6069556ae1ceb83b0979912c64fe2d75ea0dfc9e7.jpg) + +# D Comparison with ReLU Activations + +To justify our focus on sinusoidal neural networks, in this section we examine the learning dynamics of ReLU-MLPs, based on the positional encoding scheme used in [51]. The positional encoding layer is kept static, and we pre-compute the nyquist frequencies corresponding to our image size $(64 \times 64)$ , as is done in [23]. We denote this architecture ReLU-PE. All other architectural choices are identical to those described in Appendix B. We observe a number of differences between SIRENs and ReLU-PEs (visualized in Figure 13): + +![](images/d05b919dd7838a4af87c19d82dca22cf243d65b65e3908fec621625afc02ff30.jpg) + +![](images/498924b182960b57ad300ad8f0b03f51c3f1642b9fd7217be5a6334bebd55e29.jpg) + +![](images/f169e77723be138b33cbe94ad0d5f7d3c1cd1dc26968cfa2482358ab945dcce1.jpg) + +![](images/010c3c9737a76549a8d7cbfc82e94eca917e0c4051dd6f4674d3a062fc970645.jpg) + +![](images/c066a1b5004be6d1022e007a95f8ec7121d8e61431e7871b5df6f2ca6107a47b.jpg) +Figure 10: Effect of Hyperparameters on Edge Alignment: Reproduction of Figure 4 for the microphone dataset + +![](images/5198b3e59aa76d8c8ae3cfc74072dfde4e9b32a08c361deed8cea9430c3a8133.jpg) + +![](images/81acd02604e091ec7d99914da5defb0275faea2420e9bfa68615e4cb0ea50728.jpg) + +![](images/5fa68c2fe693e98f2d7382582f45595e3f2676d2c2cddf56b987550674ddaf1f.jpg) + +![](images/e24f7c2905365219729c8bcf50908093bba078f3e487e060613dfa25fb398509.jpg) + +![](images/e6e90d10241eea22b452150086143fe0f70df0eb1af154e9a573e143c77a2dbf.jpg) + +![](images/67c05b5548139d85471296b5e5e7582b2b3e2161860048ed725dc5286234d733.jpg) + +![](images/8d2280636e0c244d4d0fd0b736d632d93b2927a74e277ec4cd322bfd1a20b51e.jpg) + +![](images/5dd6942425629bead3d1dfcd15cd6a18db7695b00de14e4b6eef17d9ebd646e5.jpg) +Figure 11: Effect of Hyperparameters on Edge Alignment: Reproduction of Figure 4 for the coins dataset + +![](images/ff4da2e7e1793becfdf2e9824c01b2fdd9253fa79a24cdb9dd256d7c1228a976.jpg) + +![](images/c64bacc9f34848eb12b6a048dcbb8a3795ac3aff0230869dd19391d59b4e5b97.jpg) + +![](images/15244f692857266ba0e80a7bef162dc3462ff71349044a8d5c1187952cf0e10e.jpg) + +- Firstly, SIREN models exhibit strong locality: over the course of training, the asymptotic value of the $C_{NTK}$ decays to 0, whereas it grows in ReLU-PE models. What's more, the range of interaction as measured by $\xi_{FWHM}$ is larger in ReLU-PE models. An example comparing the correlation functions for both architectures is shown in Figure 12. +- Secondly, learning is much slower in ReLU-PE models than it is in SIRENs. One explanation for this is that there is more gradient confusion [11], that is, the minimum value of the $C_{NTK}$ is lower. In particular, $\min C_{NTK}$ is less than zero across all ReLU-PE runs, so that these models are always operating in the "slow" phase of learning. +- The principal eigenvalue $\lambda_0$ of the NTK grows to be orders of magnitude larger for SIREN models than for ReLU-PE models. That said, tangent kernel alignment still occurs in ReLU-PE models, it is just a much slower process. In Figure14, we train a 7-layer deep, 128-unit wide MLP full-batch with a learning rate of $1e - 3$ for 250k epochs, varying only the activation function. To reach the edge-alignment achieved by a SIREN model after 453 epochs, the ReLU-PE model must train for 239986 epochs. We also see that more of the edges are present in the principal eigenvector of the SIREN model's NTK. + +![](images/e6ae127cdc99b3e122a54a1d06f69ac3023b04149ef5ab410fa02ef1726d2a28.jpg) +Figure 12: Effect of Hyperparameters on Correlation Functions At Initialization: In ReLU-PE models, the Gaussian approximation of the $C_{NTK}$ correlation function is poor for all depths, due to high-variance, long range interactions. By contrast, for SIREN models, there is much less variance, and the range of the interactions shrinks for increasing $\omega_0$ . + +![](images/b28b7acb556ff4a77a1134101fb348f203481ac931a55a3347a64e99dcee0067.jpg) + +![](images/bbba8359b9e996cd822bb0ff6fcc47ceb00ab441344827ed4a0c1bae3bceeeaa.jpg) + +![](images/e5802b58e1b7ae390c1c40ecd0d0f65547f052757f8e6d9e5e8ef2e087655d7e.jpg) + +![](images/e9c801f6c7e2281d4a6cff83dc3d5578c07aebe3c9c4410dab9ab0bc04fe1601.jpg) + +![](images/86194f6abc476a2f1fffec74be2dd3276ee13f7d77af7502b9c35ab5cfe0aa91.jpg) + +![](images/4b8e2b4a627c8e40a9c537b68f760bc7bd6f615c2403f1b28d6920d91f8f5f8d.jpg) + +![](images/edafeaae3fe67794388af941164fb4108b900546052ea8d3a4c329d3b10dd8b2.jpg) + +![](images/2e49b1396759aca36cf247cca2c2405ca5ba98b232b559d866bac5e5cc2f3686.jpg) + +![](images/9e1f156c2aaae2f2a23687f086803b8980f20b4a2867153575840f7ada376fa1.jpg) + +![](images/b249a5b3dfa853dc66f799116355bcf54a8dbf6d19a8a8ed6fd2174179a2c1cf.jpg) + +![](images/418a86c6177a560db210ecb70516a2cfccce7bb1a80a054c0c7e41477acd707b.jpg) + +![](images/bf58550a3fd1207a5975b8df33001a25159a448fc6fe9d0cfcabd486ed8d01a9.jpg) +Figure 13: Learning Trajectories for SIREN and ReLU-PE models: Histograms visualizing the distribution of various order parameters throughout training. See Section B for full details on models and datasets used. + +![](images/ca755c7f664ad5a05e5a3e6f61987dddc010017d759098228545bb1da64b3ab8.jpg) + +![](images/7e1ecc17ef6d42d4ca273117347c2288f3e5560e32058849f010e53e0a29e73c.jpg) + +![](images/90469385fedc4b453cb21c7c61815fc24e0a5af5179109cdeef25316d21be89d.jpg) + +- At initialization, MAG-Ma is orders of magnitude lower for SIREN models than for Relu-PE models, indicating the latter are already operating in a phase where translational symmetry is broken. + +In summary, while ReLU-PE models exhibit Neural Tangent Kernel alignment, it is a much slower, non-local process, that does not coincide with loss-rate collapse or translational symmetry breaking. + +# E Implications of Local Image Structure on Feature Learning + +# E.1 On the Relationship Between Structure Tensors and Tangent Kernels + +We are now positioned to elucidate the features learned during NTK alignment. As proposed in Section 3.3, the local structure of the NTK adapts to the spatial variations in parameter gradients. In this section, we delve into the spectral consequences of this adaptation. We contend that the principal eigenvectors evolve into edge detectors, resembling the auto-correlation structure tensors + +![](images/d5695490a3d1947bb8db1bfd69b7848f6286f525100fcf65d68eb2c4cba30728.jpg) + +![](images/5f159036d148c46e3a1f09de03c0dfd9b255327a325c6926664fac516f9d3e2b.jpg) + +![](images/9e6799c36f33ef796d31129db11c3a74e104f776f187d41b062e59be79973e8b.jpg) + +![](images/92a709a229811573cbd1bf96ddb98de96de376b10bc8c4e3295e5fa372a86000.jpg) + +![](images/76ba6e3185fb5f2040f00567695b4be6903a53dda19e56cb3d6d7be2e9e99872.jpg) + +![](images/afb5e00462a40164f2e31c730b55d1fd1377c04f3f48aae37d47c8aa40a117e4.jpg) +Figure 14: NTK alignment in SIREN and ReLU-PE models: The principal eigenvectors of the NTK at the end of training. Final AUC $(|v_{1}|, \nabla I)$ for the ReLU-PE is 0.754, whereas AUC $(|v_{0}|, \nabla I)$ for the SIREN model is 0.804. The training time required to achieve an edge-alignment score greater than 0.75 for the SIREN model was 453 epochs, whereas for the ReLU-PE model it was 239986 epochs. + +![](images/58ef34bca0a402e3af1c3b22d6a19e40668b7bb50151d065efad577c78d3ab7b.jpg) + +![](images/42b26d9f3270cdde25e9cbbaf2172d64a1140f43c3c0486f25b965c418341023.jpg) + +![](images/c4cd93ce53e3e5a23239bcc650ecf03d6e04bc73390eeafa73f37e88bc14697a.jpg) + +commonly employed in traditional computer vision. This observation reinforces the concept of translation symmetry breaking: in computer vision, the utility of auto-correlation structure tensors stems from the premise that the most informative features are those that minimize redundancy. The auto-correlation function quantifies this through metrics of translational symmetry breaking. + +Per the discussion in Section 3.3, the principal eigenvector is closely related to the auto-correlation function. By leveraging the decomposition of the NTK in equation 30, we may relate to the features considered in computer vision. Let us define: + +$$ +w ^ {(l)} (u; x) = 1 + h ^ {(l - 1)} (x) ^ {\top} h ^ {(l - 1)} (x + u) \tag {115} +$$ + +so that the largest contribution comes from the immediate neighbourhood of $x$ . This motivates us to perform a Taylor expansion of the remaining terms as follows: + +$$ +\begin{array}{l} K _ {l} 1 = \sum_ {u} K _ {l} (x, x + u) (116) \\ = \sum_ {u} w _ {l} (u; x) \sum_ {d} \frac {\partial f (x)}{\partial z _ {l d}} \frac {\partial f (x + u)}{\partial z _ {l d}} (117) \\ \approx \sum_ {u} w _ {l} (u; x) \sum_ {d} \left(\frac {\partial f (x)}{\partial z _ {l d}} \frac {\partial f (x)}{\partial z _ {l d}} + h. o. t\right) (118) \\ = \operatorname {t r} \left(A _ {l} \left(x _ {i}\right)\right) + h. o. t (119) \\ \end{array} +$$ + +Here, $A_{l}$ denotes the structure tensor used in the Harris-Corner detector [52]. Accordingly, we see that $K1$ - and thus, the principle eigenvector - assess the extent of local translational symmetry disruption near a point $x$ . This principle underlies feature selection in computer vision, a concept mirrored in NTK feature learning, as evidenced by the principal eigenvectors that are predominantly maximized around dataset edges and corners. + +It is crucial to highlight that $A_{l}$ pertains to the structure tensor of a specific layer $l$ . Collectively, the entire DNN's NTK facilitates feature selection across a scale pyramid. + +![](images/5a940b78fdea53dd1ba09d9d7e28e959e496205b1ac3763462bb0cfed615e316.jpg) +Figure 15: Evolution of spatial variation of the parameter gradients. At initialization, there is a very small amount of variance (note the scale of the variations). As the variance grows, translational symmetry is broken, and a dynamical phase transition occurs. + +![](images/88b70891b035e531aaa7ff7da3a41e57d7ebc3af1474f8aa816d843f36df6416.jpg) + +![](images/5f24408c3ab47bb7bd6b2955ab8840057fcccf31a44a3df884e01e8492100f1a.jpg) + +![](images/7912df8de1144ae8a672b77ff6936a68ae1912047280d891b2d7f4b8aaec77d7.jpg) +(a) + +![](images/59a94ab904854b7c64b481124aaca1c5504bb344ad01e4679dd8dd876fe8379e.jpg) + +![](images/484c2e80a94d5f3c9eb5de9ebde37d53d375b4b0b4c3f89639264855d2907d1e.jpg) +Figure 16: Evolution of the Cosine NTK: We visualize $C_{NTK}(x, x + u)$ around three points $x \in \{A, B, C\}$ for small separations $u$ . At initialization, $C_{NTK}$ locally resembles an isotropic, translation-invariant RBF. However, as training progresses, these symmetries are broken. MAG-Ma (described in Section E.2) is an order-parameter that monitors the original symmetry, and changes at the critical point. + +![](images/3bc8fe3360db5bc6254aa92c769a37169507a86f4168d78fa01d2091ad8f50ba.jpg) + +![](images/9e6cf0f47e23deb43d0407dc2f98e9fc2873a8fb8a29ae0e44ca3616cb6a9013.jpg) +(b) + +![](images/46f24f423f2ccbd0ea9a67a317836b67775d9dcd675c21f61c62b8d88b0151e0.jpg) + +![](images/022bce3cacbb369d947d71866bd9cfe9b9b328a60e316a721113403de4b0db8d.jpg) + +![](images/53c5b7f219d00f8af3b351d4a4f82e00310fe86d13434e959452979c1c0859fb.jpg) + +![](images/f03990b11b63cfc90f8b3695e9816f616ed449f606b2b1a9d796f877cbe3c2cf.jpg) +(c) + +# E.2 MAG-MA: Order Parameters From Translational Symmetry Breaking + +While previous sections have focused on bottom-up construction of order parameters, this section adopts a top-down approach rooted in symmetry principles. In Sections 3.1-3.4, we expressed several order parameters in terms of the parameters $a$ , $D$ , $H$ , characterizing the local structure of the $C_{NTK}$ . Notably, each of these parameters is now a function of the spatial variation of the parameter gradients, whose evolution is showcased in Figure 15. This suggests it is a translation symmetry which is broken at the phase transition. Indeed, from Figure 16, we observe the $C_{NTK}$ is an approximately stationary, isotropic kernel - a desirable property for INRs [30]. As such, the Kernel exhibits no bias for location or direction. Over the course of training, we may monitor the emergence of such a bias with the following metric: + +$$ +\left| \right.\left| \right. \mathbb {E} _ {x} \left[ \right. \nabla_ {x} \log \left| \right.\left| \nabla_ {\theta} f \right| ^ {2} \left. \right]\left. \right|\left. \right| ^ {2} = \left|\left| \mathbb {E} _ {x} \left[ D _ {x} / a _ {x} ^ {2} \right]\right|\right| ^ {2} \tag {120} +$$ + +We refer to this statistic as MAG-Ma: the Magnitude of the Average Gradient of the Log Gradient-Field Magnitudes. Intuitively, this order parameter captures the statistical preference for a spatial direction in the dataset. The evolution of this quantity is plotted in Figure 16, and its alignment with the other order parameters is shown in Figure 3. We see that throughout the Fast Phase of training (before the peak in the loss rate $\dot{L}_{eval}$ ), the local structure of the $C_{NTK}$ is statistically translation invariant, and MAG-Ma is close to zero. However, just after the critical point, it grows rapidly - coinciding with the edge memorization described in Section 3.3. + +# F Evaluating Fidelity of Approximation + +# F.1 Local Structure of the NTK + +As described in Section 3.1, INRs are often carefully designed to ensure a diagonally dominant NTK [30, 31, 32]. In higher dimensions, diagonal dominance is equivalent to a bias towards local + +![](images/393f4f5162aab05b942d2efc0627915a18ba4ffbe3321d316cd156cd3daf0406.jpg) + +![](images/0810898420ea0c5c58b12b35a2d2de11686df2f926714dd85cb09214e72c8e06.jpg) + +![](images/f66143a367ee48e30a8f7457bce384b57f83dfa233a1d820fa11fdf2ef6e54af.jpg) +Figure 17: Hyperparameters Affect Local NTK Structure. Boxplots visualizing the distribution of structural parameters for the $C_{NTK}$ . Top row: variation in the initial correlation lengthscale $\xi_{corr}(0)$ . Bottom row: variation in the initial asymptotic value of the $C_{NTK}$ ( $C_{\infty}(0)$ ). + +![](images/8bf05b94fdaf089ae076f028024b054bd789d81f23b9cf270e5fdeb11918df3c.jpg) + +![](images/aca49cf64984fc97d2022c37813a5781b16434968189d0d06eee3706056e44f8.jpg) +Figure 18: Visualization showing the empirical correlation function for the normalized parameter gradients. On the left-hand side is the global correlation-function for the $C_{NTK}$ . On the right is the local-correlation function for the $K_{NTK}$ around a test point $x$ . Dashed lines show fitted Gaussian approximation, and error bars show variance across dataset. Over the course of training, both the global correlation lengthscale $\xi_{corr}$ , and the terminal value $c_{\infty}$ , evolve. + +![](images/ecf45d113ab4add9890c6f19608c068123158912a03c81f95e62b47f152da833.jpg) + +interactions. We see in Figure 17 the hyperparameters that most affect this local structure: we observe that, while depth has a small impact on the initial correlation lengthscale $\xi_{corr}(0)$ , higher values of $\omega_0$ cause the $C_{NTK}$ to become dramatically more localized. The converse is true for the asymptotic value $C_{\infty}(0)$ : $\omega_0$ has a minor effect, but increasing depth leads to stronger interactions across large distances. + +Beyond initialization, in Figure 18, we examine the evolution of the correlation function for a five-layer deep, 128-unit wide SIREN model on a $128 \times 128$ grayscale image of a cat, with bandwidth $\omega_0 = 30$ . Emprirically, we see that the Gaussian approximation described in Section 3.1 remains valid across training, with the asymptotic value $c_{\infty}$ of the $C_{NTK}$ decaying to zero. + +# F.2 Cauchy Approximation + +To ascertain the fidelity of the Cauchy Approximation, we estimate the Pearson correlation between the true values of the correlation lengthscale $\xi$ and $\min C_{NTK}$ , and the prediction based only on the local model. We choose this metric because the identification of critical points is insensitive to linear + +Table 3: Fidelity of Cauchy Approximation: Pearson correlation between the true order parameter and predictions using the local Cauchy Approximation. Mean and standard deviation are calculated over the spread of models and datasets described in Section B. + +
depthω0ξmin CNTKv0
3150.980 ± 0.0170.889 ± 0.0680.980 ± 0.011
300.924 ± 0.1100.909 ± 0.0630.985 ± 0.006
600.830 ± 0.2080.946 ± 0.0430.988 ± 0.004
900.856 ± 0.2200.967 ± 0.0200.986 ± 0.006
4150.983 ± 0.0150.925 ± 0.0590.982 ± 0.009
300.964 ± 0.0360.956 ± 0.0390.985 ± 0.007
600.920 ± 0.1520.955 ± 0.0360.986 ± 0.008
900.974 ± 0.0260.961 ± 0.0310.984 ± 0.009
5150.985 ± 0.0110.921 ± 0.0490.978 ± 0.010
300.969 ± 0.0230.953 ± 0.0380.983 ± 0.008
600.947 ± 0.0660.966 ± 0.0330.982 ± 0.028
900.959 ± 0.0370.974 ± 0.0260.982 ± 0.010
+ +transformations. The results are shown in Table 3. Similarly, we evaluate our approximation of the principle eigenvector $v_{0}$ , by looking at the absolute cosine distance between our approximation and the ground-truth. + +Finally, in Section 3.3, we approximated the principal eigenvector $v_{0}$ of the NTK $K$ with the row mean $K1 / 1^{\top}1$ . The median cosine alignment between the row mean and the true $v_{0}$ was found to be 0.99995 across all epochs surveyed, across all models and datasets. The IQR is 0.00446. The strength of this approximation is a testament to the extreme spectral gap of the NTK, which itself is a consequence of NTK alignment. + +# G Additional Experimental Results + +# G.1 Order Parameter Trajectories for Single Runs + +This section contains additional illustrations of the order parameter trajectories, and the corresponding confidence region estimates, similar to the left side of Figure 3. The results are shown in Figure 19. Each model is a 5 layer deep, 128-unit wide SIREN network, trained with full-batch gradient-descent with a learning rate of 1e-3. + +# G.2 Influence of Hyperparameters on Order Parameter Trajectories + +In this section, we perform an ablation study to understand the impact of different hyperparameters on the order parameter trajectories. The baseline model is a 5-layer 128-unit wide SIREN with $\omega_0 = 60$ . Figures 20-22 showcase the effect of depth. Figures 23-25 showcase the effect of the bandwidth parameter $\omega$ . + +When depth (and therefore model capacity) is decreased, we observe a corresponding increase in the validation error. In shallower models, the initial gradient confusion $(\min C_{NTK})$ is lower, delaying learning, and thus, the peak in the loss rate $\dot{L}_{eval}$ . While the location of the phase transition changes, the trajectory of the order parameters shapes remain consistent, and exhibit less variance with increased depth. By contrast, there is dramatic change in the shape of the trajectories as we vary $\omega_0$ . When $\omega_0$ is high, $\xi_{corr}$ starts very low, favouring interactions with immediate neighbours, leading to low overlap with the RBF. During training, the range broadens rapidly, causing $\mathrm{CKA}(K_X,K_{NTK})$ to grow sigmoidally at the critical point. Conversely, with low $\omega_0$ , the range starts large but shrinks during training. + +We additionally track the CKA between the NTK and a static RBF kernel with fixed bandwidth $K_{X}$ as described in 4.1. The evolution of this hyperparameter reflects the evolution of the correlation lengthscale $\xi_{corr}$ . When this value is large (as it is when $\omega_0$ is small), the NTK has a broad diagonal, + +![](images/976492c9df3964082cd910574b8b2a838433505c1603b9d85ac4ac0ed1b26e5c.jpg) +(a) aircraft_carrier + +![](images/7920cf4c01c8fb927a01410dfc5969dc190429a5ac368a09c5675738ab25e88a.jpg) + +![](images/7bdd2a1b073d8dca07ce6fe236b116d07fc05f0c41b5bbc10daa8c8a1f01c6ae.jpg) +(c) chain +Figure 19: Alignment of Order Parameters. Order parameter evolution and critical points during training of a SIREN model. The red vertical lines denote the location of the critical points, and the green vertical lines denote confidence regions. + +![](images/cb4f4bdd1823e69ea706d0db7d8fe97f41be5cee500cebdf332eb868d3ede40a.jpg) +(b)presso-maker +(d) violin + +and thus overlaps well with the RBF. Over the course of training, $\xi_{corr}$ shrinks, and thus, so does $\mathrm{CKA}(K_X,K_{NTK})$ . + +![](images/0c8e71a211e75c73e03032f23a8aad042e3ba9287bdbc2a6b4b12d328bc0c63e.jpg) +Figure 20: Effect of depth on Critical Behaviour (Microphone): Average MSEs, in order of ascending depth: $2.561e^{-2} \pm 9.355e^{-5}$ , $2.555e^{-2} \pm 8.970e^{-5}$ , $2.572e^{-2} \pm 7.209e^{-5}$ . Dashed vertical lines denote the location of the peak of the loss rate $\dot{L}_{\mathrm{eval}}$ , marking the phase transition. + +![](images/a156d93844212cab48a25284c6b4ab1d2b6c0ef09ea9258f06afc400adb1beff.jpg) + +![](images/4a8cd27395119e0cf4f190989d6d0e5f42f9c52c6d0b287476ef529b3fb18257.jpg) + +![](images/f45a33d8530fafa427f021ed8a457cc7bc64aa71704ec43471b96bf5758b6e72.jpg) +Figure 21: Effect of depth on Critical Behaviour (Sax): Average MSEs, in order of ascending depth: $1.628e^{-2} \pm 1.312e^{-4}$ , $1.513e^{-2} \pm 3.384e^{-5}$ , $1.494e^{-2} \pm 6.605e^{-5}$ . Dashed vertical lines denote the location of the peak of the loss rate $\dot{L}_{\mathrm{eval}}$ , marking the phase transition. + +![](images/e9d03b612458a115504ce77cd9d456561f2f9c21bec7b9fef5306d122485372c.jpg) + +![](images/2c03854547d8ed2e88831d50427a5beb67d701593c2300934bb6d170b53959cb.jpg) + +![](images/9dfee97b5fd7234e091da11ea512a85e1efb443da1a7e312c487140ed3f484a9.jpg) +Figure 22: Effect of depth on Critical Behaviour (Violin): Average MSEs, in order of ascending depth: $6.885e^{-3} \pm 1.677e^{-4}$ , $5.930e^{-3} \pm 5.016e^{-5}$ , $5.665e^{-3} \pm 3.640e^{-5}$ . Dashed vertical lines denote the location of the peak of the loss rate $\dot{L}_{\mathrm{eval}}$ , marking the phase transition. + +![](images/bd158f6d4e3251fabdcd38aede14f4582b57eb0b759d393c779dbef11409ca05.jpg) + +![](images/d8ca138ae68fc0dd2ad804de677b41781297675cd50c1a763b71665b1083a501.jpg) + +![](images/3a7223ae594a7d4faf9b344dddf45a8e52ce2846accdbb86ff264a73987937a9.jpg) +Figure 23: Effect of $\omega_0$ on Critical Behaviour (Microphone): Average MSEs, in order of ascending $\omega_0$ : $2.601e^{-2} \pm 1.804e^{-4}$ , $2.566e^{-2} \pm 1.327e^{-4}$ , $2.572e^{-2} \pm 7.209e^{-5}$ , $2.807e^{-2} \pm 7.688e^{-4}$ . Dashed vertical lines denote the location of the peak of the loss rate $\dot{L}_{\mathrm{eval}}$ , marking the phase transition. + +![](images/e515387005d51355046b589208fc98b26dd6c91d557bfb14fa292fd2087c8175.jpg) + +![](images/d682962cbc537bb323d95758e64e0156c3edeea3effd1db502aab59e450fb284.jpg) + +![](images/5a1aebbbba925a23424313a43d211ad4e63e7781790fbd3e5540ba456ec3baf5.jpg) +Figure 24: Effect of $\omega_0$ on Critical Behaviour (Sax): Average MSEs, in order of ascending $\omega_0$ : $1.680e^{-2} \pm 1.666e^{-4}$ , $1.561e^{-2} \pm 6.552e^{-5}$ , $1.494e^{-2} \pm 6.605e^{-5}$ , $1.639e^{-2} \pm 3.938e^{-4}$ . Dashed vertical lines denote the location of the peak of the loss rate $\dot{L}_{\mathrm{eval}}$ , marking the phase transition. + +![](images/41ed43cb39a27bb5f752045e72b23455f45c4b82bbbf54671d9ab240802f8de5.jpg) + +![](images/107a592585dd622ed266349e170cfe2d28ad773a58a97c0693a07593f734242b.jpg) + +![](images/911e12ce4507e81154725de15709a04004ef4f66ac08577f4ef0c3adfe214a5b.jpg) +Figure 25: Effect of $\omega_0$ on Critical Behaviour (Violin): Average MSEs, in order of ascending $\omega_0$ : $7.223e^{-3} \pm 1.503e^{-4}$ , $6.305e^{-3} \pm 3.139e^{-5}$ , $5.665e^{-3} \pm 3.640e^{-5}$ , $6.698e^{-3} \pm 3.359e^{-4}$ . Dashed vertical lines denote the location of the peak of the loss rate $\dot{L}_{\mathrm{eval}}$ , marking the phase transition. + +![](images/5a779508f420d54c42164af6470f3eb832978018b94743ade5277d0eac28f460.jpg) + +![](images/db12a5b0354c5be8d8c6a98ef45bf7db224899451afc7606c88ced2fb62db767.jpg) \ No newline at end of file diff --git a/NeurIPS/2025/A Closer Look at NTK Alignment_ Linking Phase Transitions in Deep Image Regression/images.zip b/NeurIPS/2025/A Closer Look at NTK Alignment_ Linking Phase Transitions in Deep Image Regression/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..501aa18470a66964e1cfbf3fd34e91293caca781 --- /dev/null +++ b/NeurIPS/2025/A Closer Look at NTK Alignment_ Linking Phase Transitions in Deep Image Regression/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:83744aa00728635c28068798b1da6682ebb37e41ccc6435eca2e41697232b5d4 +size 2657159 diff --git a/NeurIPS/2025/A Closer Look at NTK Alignment_ Linking Phase Transitions in Deep Image Regression/layout.json b/NeurIPS/2025/A Closer Look at NTK Alignment_ Linking Phase Transitions in Deep Image Regression/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..af954a44d7e6b9604c93bc4d1749680e1d9ecd36 --- /dev/null +++ b/NeurIPS/2025/A Closer Look at NTK Alignment_ Linking Phase Transitions in Deep Image Regression/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5ffdf206cbd12fc637abe2edf7c8092d64ae66c07f7fadd731ded98570d06cf7 +size 1511843 diff --git a/NeurIPS/2025/A Closer Look at TabPFN v2_ Understanding Its Strengths and Extending Its Capabilities/e53524b8-a557-4206-898a-de21155251ea_content_list.json b/NeurIPS/2025/A Closer Look at TabPFN v2_ Understanding Its Strengths and Extending Its Capabilities/e53524b8-a557-4206-898a-de21155251ea_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..aa6203c8b9885e07512c85153f4a838d6b19a94b --- /dev/null +++ b/NeurIPS/2025/A Closer Look at TabPFN v2_ Understanding Its Strengths and Extending Its Capabilities/e53524b8-a557-4206-898a-de21155251ea_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:edb7bf24afae8c3e45a1b74dae1a3b7bd170e2c3406f64d8c22f91ecfe2fbc19 +size 228397 diff --git a/NeurIPS/2025/A Closer Look at TabPFN v2_ Understanding Its Strengths and Extending Its Capabilities/e53524b8-a557-4206-898a-de21155251ea_model.json b/NeurIPS/2025/A Closer Look at TabPFN v2_ Understanding Its Strengths and Extending Its Capabilities/e53524b8-a557-4206-898a-de21155251ea_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7306665d7b0965633124524f09241c8ccf91e6c1 --- /dev/null +++ b/NeurIPS/2025/A Closer Look at TabPFN v2_ Understanding Its Strengths and Extending Its Capabilities/e53524b8-a557-4206-898a-de21155251ea_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9cc6d7ac151ebcdb70e73ac252a128eecd4e581f69c355442c0cf00d8e65fd4f +size 281493 diff --git a/NeurIPS/2025/A Closer Look at TabPFN v2_ Understanding Its Strengths and Extending Its Capabilities/e53524b8-a557-4206-898a-de21155251ea_origin.pdf b/NeurIPS/2025/A Closer Look at TabPFN v2_ Understanding Its Strengths and Extending Its Capabilities/e53524b8-a557-4206-898a-de21155251ea_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..5a556b1d5c74deba03cf0b5b53f064b2ea7fa84c --- /dev/null +++ b/NeurIPS/2025/A Closer Look at TabPFN v2_ Understanding Its Strengths and Extending Its Capabilities/e53524b8-a557-4206-898a-de21155251ea_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:27dc581883a5966a8bd16f757aecf81111e93b704028935432e3236f6799734e +size 2582853 diff --git a/NeurIPS/2025/A Closer Look at TabPFN v2_ Understanding Its Strengths and Extending Its Capabilities/full.md b/NeurIPS/2025/A Closer Look at TabPFN v2_ Understanding Its Strengths and Extending Its Capabilities/full.md new file mode 100644 index 0000000000000000000000000000000000000000..edf3265db67b52bae0bf77edf9611c65d165b569 --- /dev/null +++ b/NeurIPS/2025/A Closer Look at TabPFN v2_ Understanding Its Strengths and Extending Its Capabilities/full.md @@ -0,0 +1,904 @@ +# A Closer Look at TabPFN v2: Understanding Its Strengths and Extending Its Capabilities + +Han-Jia Ye & Si-Yang Liu + +School of Artificial Intelligence, Nanjing University + +National Key Laboratory for Novel Software Technology, Nanjing University + +{yehj, liusy}@lamda.nju.edu.cn + +Wei-Lun Chao + +The Ohio State University + +chao.209@osu.edu + +# Abstract + +Tabular datasets are inherently heterogeneous, presenting significant challenges for developing pre-trained foundation models. The recently introduced transformer-based Tabular Prior-data Fitted Network v2 (TabPFN v2) achieves unprecedented in-context learning performance across diverse downstream datasets, marking a pivotal advancement in tabular foundation models. In this paper, we take a closer look at TabPFN v2 to examine how it effectively handles heterogeneity and achieves high predictive accuracy, and to explore how its limitations in high-dimensional, many-category, and large-scale tasks can be mitigated. We find that TabPFN v2 can infer attribute relationships even when provided with randomized attribute token inputs, eliminating the need to explicitly learn dataset-specific attribute embeddings to address heterogeneity. We further show that TabPFN v2 can be transformed into a feature extractor, revealing its ability to construct a highly separable feature space for accurate predictions. Lastly, we demonstrate that TabPFN v2's limitations can be addressed through a test-time divide-and-conquer strategy, enabling scalable inference without requiring re-training. By uncovering the mechanisms behind TabPFN v2's success and introducing strategies to extend its applicability, this study offers key insights into the design of future tabular foundation models. + +# 1 Introduction + +Tabular data is ubiquitous across a wide range of applications, including healthcare [32], finance [43], and scientific research [33, 32]. In this format, each instance (e.g., a patient's record) is represented as a vector of attributes, and the goal of a machine learning model is to map these vectors to their corresponding labels [7]. Traditionally, tree-based models [56, 12] have dominated this domain, but recent advances in deep tabular models are increasingly closing the performance gap [22, 29, 76]. + +However, unlike vision and language domains, where pre-trained foundation models have driven significant progress [40, 81], tabular data is still desperately awaiting a similar breakthrough [63, 27, 54, 82, 77, 66]. A primary challenge arises from the inherent heterogeneity of tabular datasets, which often vary in dimensionality and attribute meanings, making the development of effective and versatile foundation models difficult. Additionally, there is an urgent need for such models, as many tabular datasets are small-scale—such as medical data with limited patient numbers. Training individual models from scratch for these datasets is highly sensitive to hyperparameter choices and often fails to generalize due to limited data [18, 24, 25]. + +Recently, the Tabular Prior-Fitted Network v2 (TabPFN v2) [28] has emerged as a significant step forward. Built on transformer architectures [67] and pre-trained on gigantic synthetic datasets [27, 28], TabPFN v2 can be directly applied to diverse downstream tasks without additional tuning. Specifically, TabPFN v2 takes both a labeled training set and an unlabeled test instance as input, predicting the test label in an "in-context learning" manner. When evaluated across both classification and regression tasks, TabPFN v2 consistently outperforms prior tabular methods, achieving state-of-the-art accuracy. + +Motivated by the remarkable performance of TabPFN v2, we aim to take a step further to understand the mechanisms behind its success1—specifically, how it effectively handles dataset heterogeneity and achieves high predictive accuracy. In addition, we investigate how to overcome its current limitations—namely, its suggested data regime of no more than 10,000 samples, 500 dimensions, and 10 classes [28]—ideally without requiring model re-training. We outline major insights as follows. + +1. TabPFN v2 internalizes attribute token learning to handle data heterogeneity. Given an instance with $d$ attributes, TabPFN v2 transforms it into a set of fixed-dimensional tokens and uses a transformer architecture to handle variability in $d$ , following [64, 22, 74]. In sharp contrast to prior methods that rely on known attribute semantics (e.g., word vectors) or learn dataset-specific attribute tokens, TabPFN v2 instead employs randomized attribute tokens—resampled at each inference. This design "syntactically" allows TabPFN v2 to be directly applied to new downstream datasets with varying dimensionalities and attribute meanings without additional tuning, but raises a fundamental question: how does it still make accurate predictions? Our analysis shows that, regardless of the randomness, TabPFN v2 can consistently infer attribute relationships through in-context learning, essentially integrating attribute token learning into the inference itself. In short, TabPFN v2 unifies representation learning and prediction within a single forward pass. +2. TabPFN v2 can be repurposed as a feature extractor for downstream tasks. The exceptional predictive performance of TabPFN v2 suggests that it produces instance-level feature representations that are highly discriminative. However, verifying this is non-trivial, as TabPFN v2's in-context learning mechanism assigns distinct roles to labeled training and unlabeled test instances, resulting in embeddings that are not directly comparable. To overcome this, we propose a leave-one-fold-out strategy that enables the extraction of instance features more closely aligned across training and test data. Our findings reveal that TabPFN v2 effectively maps tabular instances into a nearly linearly separable embedding space. Remarkably, training a linear model on these features yields accuracy comparable to that of TabPFN v2's in-context learner, highlighting its potential as a powerful feature encoder. This not only offers insights into TabPFN v2's inner workings but also opens the door to broader applications (e.g., visualization and error analysis). +3. Test-time divide-and-conquer effectively mitigates TabPFN v2's limitations. As noted in [28], TabPFN v2 faces challenges when applied to high-dimensional, many-category, or large-scale datasets. Rather than resorting to model re-training, we show that these limitations can be effectively addressed through carefully designed post-hoc divide-and-conquer strategies, reminiscent of test-time scaling techniques developed for large language models [70, 51]. Empirical results show significant accuracy gains across these challenging data regimes, highlighting the potential of advanced post-hoc methods to further extend the capabilities of tabular foundation models. + +Remark. This paper presents a timely and in-depth investigation into TabPFN v2, offering valuable insights for advancing tabular foundation models. While we do not propose a new architecture or training scheme, our contribution lies in the novel analysis and principled extension of TabPFN v2. This reflects a growing trend in foundation model research, where understanding, evaluating, and adapting powerful models is increasingly seen as being as impactful as designing new ones. + +# 2 Related Work + +Tabular foundation models. Pre-trained models have revolutionized the vision and language domains [40, 81], but their adoption in tabular data remains limited due to the substantial heterogeneity across datasets. Variations in attribute spaces, dimensionalities, and label distributions present significant challenges for joint training and transferability. One solution is to leverage the semantic meanings of attributes, as demonstrated by methods that convert tabular instances into textual descriptions and + +![](images/d977355ffbc6524810f14b13e7717660aa45a9f97d17b6a6916a4cdee54cd2a7.jpg) +Figure 1: Left: Illustration of TabPFN v2's mechanism for binary classification [28]. $\{A_1, \dots, A_d\}$ denote $d$ attributes of the task. Training examples and a test instance are combined into a tabular context and transformed into a $(N + 1) \times (d + 1) \times k$ tensor using a combination of learnable and randomized tokens. Two types of self-attention are applied alternately across rows (inter-sample) and columns (inter-feature). The output token corresponding to the (dummy) label of the test instance is processed through an MLP to generate a 10-class logit. Right: Wilcoxon-Holm test at a significance level of 0.05 over 273 small- to medium-scale datasets. We omit the 27 datasets used to select TabPFN v2's checkpoint from the 300 datasets in [75]. + +![](images/56e9d0b8152167b26095df9c7619aeeebb73752ee3820d841be0a5fb3d5d5f56.jpg) + +apply large language models for prediction [26, 79, 69, 71]. Alternatively, some approaches aim to improve transferability by pre-computing attribute tokens based on semantic embeddings [74, 39]. In practical domains such as healthcare or scientific measurement, the semantic meanings of attributes are often inaccessible due to privacy constraints, annotation costs, or a lack of descriability. To address this, [77] proposed representing each instance by its similarity profile to a fixed number of nearest neighbor examples, thereby mapping it into a consistent latent space with shared dimensional semantics. The TabPFN family [27, 28] leverages the in-context learning capabilities of transformers to directly predict labels by contextualizing test instances among training examples. This strategy inspired subsequent pre-trained tabular models such as [48, 15, 57]. While TabPFN v1 pads attribute vectors to a fixed dimension, TabPFN v2 introduces a specialized attribute tokenizer to handle heterogeneous input spaces. Meta-learning has also been explored to generate model weights tailored for downstream tabular tasks with limited data [34, 6, 50]. Other pre-trained models rely on lightweight fine-tuning to adapt to variations in attribute and label spaces [44, 80, 62, 82]. + +# 3 Background + +Learning with a single tabular dataset. A tabular dataset $\mathcal{D} = \{(x_i, y_i)\}_{i=1}^N$ contains $N$ training examples, corresponding to the rows in a table. Each instance $\boldsymbol{x}_i$ is characterized by $d$ features or attributes (i.e., columns in the table), where $d$ typically varies across datasets. Its label $y_i$ belongs to $[C] = \{1, \ldots, C\}$ for a classification task or is a numerical value for a regression task. We assume that all attributes of an instance are numerical (continuous). Categorical (discrete) attributes, if present, are transformed using ordinal or one-hot encoding beforehand. The goal of tabular machine learning is to learn a mapping $f$ from instances to their labels. Specifically, given an unseen instance $\boldsymbol{x}^* \in \mathbb{R}^d$ sampled from the same distribution as $\mathcal{D}$ , the learned mapping $f$ predicts its label as $\hat{y}^* = f(\boldsymbol{x}^* \mid \mathcal{D})$ . A smaller discrepancy between $\hat{y}^*$ and the true label $y^*$ indicates stronger generalizability of $f$ . + +TabPFN. The original TabPFN implements $f$ for classification using a transformer-like architecture [27]. Both training and test instances are first zero-padded to a fixed dimension $k'$ (e.g., 100). Then, $x_{i}$ and $y_{i}$ are linearly projected to $\tilde{x}_{i} \in \mathbb{R}^{k}$ and $\tilde{y}_i \in \mathbb{R}^k$ , respectively. TabPFN processes both a labeled training set and an unlabeled test instance jointly, predicting the test label in an in-context learning manner. The task context is defined as $\mathcal{C} = \{(\tilde{x}_1 + \tilde{y}_1),\dots ,(\tilde{x}_N + \tilde{y}_N),(\tilde{x}^*)\} \in \mathbb{R}^{(N + 1)\times k}$ , consisting of $N + 1$ tokens, each of dimension $k$ . These tokens are processed by multiple transformer layers, which accommodate variable-length inputs (i.e., variable $N$ ). The output token corresponding to the test instance is passed through a multi-layer perceptron (MLP) to produce a 10-class logit. + +TabPFN v2. The recently proposed variant [28] introduces several key modifications. First, each of the $d$ attributes in $\mathcal{D}$ is embedded into a $k$ -dimensional space, with random perturbations added to differentiate attributes. Together with the label embedding $\tilde{\pmb{y}}_i \in \mathbb{R}^k$ , each training instance $\pmb{x}_i$ is represented by $(d + 1)$ tokens with dimension $k$ . For a test instance $\pmb{x}^*$ , where the label is unknown, a dummy label (e.g., the average label of the training set) is used to generate the label embedding $\tilde{\pmb{y}}^*$ . + +The full input to TabPFN v2—comprising the training set and the test instance—is thus represented as a tensor of shape $(N + 1) \times (d + 1) \times k$ . Two types of self-attention are applied in alternation: one over samples (among the $N + 1$ instances) and the other over attributes (among the $d + 1$ dimensions), + +![](images/7f0085c6f46761747fd535f10fcb8932e1a319a4bfe74bcf928854b731018d92.jpg) +Figure 2: Probability of Achieving the Maximum Accuracy or Minimum RMSE across 273 datasets. Values inside rectangles show the percentage of datasets on which a method achieves the best result. + +
↓TabPFN v2CatBMNCAR-MLPLR
High-Dim3.362.824.412.142.27
Large-Scale3.971.892.271.944.47
>10 classes3.332.753.171.424.33
+ +Table 1: Average rank (lower is better) of TabPFN v2 and representative baselines on 18 high-dimensional, 18 large-scale, and 12 datasets with more than 10 classes. Full results with our extensions are in Figure 5. + +enabling in-context learning along both axes. Finally, the output token corresponding to the test instance's dummy label $\tilde{\pmb{y}}^*$ is extracted and mapped to a 10-class logit for classification or a single-value logit for regression. An overview of this process is illustrated in Figure 1 (left). + +The weights in TabPFN v2 are pre-trained on diverse synthetic datasets generated using structural causal models (SCMs), with the checkpoint selected based on real-world datasets. For additional details, including feature pre-processing, acceleration, and post-hoc ensembling, please refer to [28]. + +Remark. In the tabular domain, years of research into deep and foundation models have culminated in TabPFN v2 [28]—a breakthrough that, for the first time, enables deep models to consistently outperform traditional methods without fine-tuning. However, due to venue constraints, many technical details were omitted from the main paper. For example, the use of randomized tokens was documented in the supplementary material and code. In light of this, we aim to systematically analyze TabPFN v2, as we believe such a study is more impactful than proposing yet another architecture. + +# 4 Comprehensive Evaluation of TabPFN v2 + +Before presenting our core studies, we first extend TabPFN v2's evaluation beyond the original set of datasets to over 300, covering a much broader range of domains, attributes, scales, dimensionalities, and tasks [23, 49, 75, 60], aiming to more thoroughly assess its generalizability and limitations. + +# 4.1 Setups + +We first adopt the benchmark from [75], comprising 120 binary classification, 80 multi-class classification, and 100 regression tasks. It resolves common issues such as mislabeled data and redundancies from overlapping dataset versions [42], enabling more reliable evaluations. Out of the 300 datasets, 27 belong to the validation set used for checkpoint selection in TabPFN v2 [28]. To avoid evaluation bias, we exclude these datasets and report results on the remaining 273 datasets. + +Following the protocol in [21, 22], each dataset is randomly split into training, validation, and test partitions in a $64\% / 16\% / 20\%$ ratio. TabPFN v2 predicts test set labels directly using in-context learning, without any additional parameter or hyperparameter tuning. Baseline tabular methods—both deep and traditional—perform hyperparameter tuning using Optuna [1], with 100 trials on the training set and early stopping based on validation performance. + +All methods are evaluated using 15 random seeds, and we report the average performance across seeds. For classification tasks, we report accuracy (higher is better), while regression tasks are evaluated using Root Mean Square Error (RMSE; lower is better). For tasks with more than 10 classes, we adopt the built-in Error-Correcting Output Codes (ECOC) strategy for TabPFN v2. + +# 4.2 Empirical Results: Strengths of TabPFN v2 + +We compare TabPFN v2 against 26 representative tabular methods (see Appendix for full references). To assess statistical significance, we apply the Wilcoxon-Holm test with a significance level of 0.05 [14]. As shown in the critical difference diagram in Figure 1 (right), TabPFN v2 consistently outperforms both tree-based methods, such as CatBoost [56], and deep tabular models, including RealMLP [29], ModernNCA [76], TabM [19], TabR [21], and FT-Transformer [22]. + +To further assess performance, we report the Probability of Achieving the Maximum Accuracy or Minimum RMSE (PAMA) [13], which measures the proportion of datasets on which a method achieves the best performance. As shown in Figure 2, TabPFN v2 attains the highest score, delivering the top results on $26.02\%$ of the datasets—outperforming other methods such as ModernNCA $(11.79\%)$ and TabM $(9.78\%)$ . These results underscore TabPFN v2's strong generalizability. + +# 4.3 Empirical Results: Limitations of TabPFN v2 + +The above evaluation focuses on small- to medium-scale datasets, specifically those with fewer than 10,000 examples. However, as noted in [28], the computational complexity of transformers constrains TabPFN v2's ability to scale effectively to datasets with larger sample sizes or higher dimensionality. + +To verify this, we conduct additional evaluations on 18 high-dimensional datasets with $d \geq 2,000$ [36] and 18 large-scale datasets where $N \times d > 1,000,000$ . For high-dimensional datasets, we follow the same protocol as before. For large-scale datasets, due to the prohibitive cost of hyperparameter tuning, default hyperparameters are used for all methods. The average ranks of several representative methods are summarized in Table 1. The full results—along with our extensions—are in Section 7. + +As shown, TabPFN v2's performance degrades on both large-scale and high-dimensional datasets. On large-scale datasets, it ranks below both CatBoost and RealMLP; on high-dimensional datasets, it even falls behind the simple Logistic Regression (LR) model. Beyond these two limitations, Table 1 also reports results on the 12 datasets in Section 4.2 that contain more than 10 categories, where the ECOC strategy currently used by TabPFN v2 appears ineffective in achieving high accuracy. While increased computational complexity may contribute to this reduced effectiveness, we hypothesize that TabPFN v2 was pre-trained exclusively on small- to medium-scale synthetic datasets with fewer than 10 categories, leading to a mismatch when applied to larger or more complex real-world data. + +These results underscore the limitations of TabPFN v2, suggesting areas for further improvement. + +# 5 How Does TabPFN v2 Effectively Handle Data Heterogeneity? + +Section 4 demonstrates TabPFN v2's excellent generalizability to heterogeneous downstream tasks while also highlighting its current limitations. In the rest of the paper, we first examine the mechanisms behind its strengths, followed by methods to overcome its limitations. + +# 5.1 Diving into TabPFN v2's Mechanisms for Heterogeneous Input + +Revisiting the problem. As noted in Sections 2 and 3, tabular datasets often differ in both the number of attributes (i.e., $d$ ) and the semantics of those attributes. Even when dimensionalities match, the dimensional semantics from different datasets are typically not directly comparable. A robust tabular foundation model must therefore handle such heterogeneity effectively, enabling it to learn from diverse pre-training datasets and transfer its capabilities to new downstream tasks. + +Tokenization as a feasible solution. Among prior approaches, the most relevant to TabPFN v2 are token-based methods [64, 22, 74]. The core idea is to convert a $d$ -dimensional instance $\pmb{x} \in \mathbb{R}^d$ into a set of $d$ fixed-dimensional tokens (each of dimension $k$ ), with one token per attribute. This enables the use of transformer architectures, which naturally accommodate variability in $d$ across datasets. + +To embed each attribute into a shared $k$ -dimensional space, prior work either uses pre-defined semantic embeddings [74] (e.g., word vectors of attribute names) or learns dataset-specific embeddings [64, 22]. Given $d$ attribute-specific tokens $[\pmb{r}_1,\dots ,\pmb{r}_d]\in \mathbb{R}^{d\times k}$ , each instance $\pmb{x}_i\in \mathbb{R}^d$ can then be transformed into $[x_i^1\cdot r_1,\ldots ,x_i^d\cdot r_d]\in \mathbb{R}^{d\times k}$ , where $x_{i}^{j}$ denotes the $j$ -th element of $\pmb{x}_i$ . + +By embedding all attributes into a shared, fixed-dimensional feature space, this approach allows the transformer to learn transferable patterns and knowledge from heterogeneous datasets. + +Difficulty in direct generalization. While appealing, the aforementioned methods face a notable challenge when applied to downstream tasks: attribute names or semantics are not always accessible, as discussed in Section 2. Although it is possible to learn dataset-specific attribute tokens, doing so incurs additional computational cost and prohibits the reuse of previously learned tokens. Consequently, this limits the direct generalization of the foundation model to new tasks. + +![](images/5ed5ffc126b89a673a88efdedc68b56124902e63abe5ffcddd719a8209c8ebfd.jpg) + +![](images/1d033e03e058ab99e923829f33a6bec9f6f75ad098e494c7c3f412dd3f60b766.jpg) + +![](images/1254faeb49aba52aa12f750d749cddf7e9e70d4d028eb33d791e37597d23e691.jpg) + +![](images/7cc5191a92842278cac3ac71f9f4dfc273008b1470e0cf2552df85f3ca6a3967.jpg) + +![](images/a7ae8ddc4d879c3e512480cd1fc18b5de69838f72722ec8078ddc8fc4ed94fae.jpg) + +![](images/d371d534c0931fc23743d2082ed53b88d30fb432935f903820816b2246550e66.jpg) + +![](images/aad167b329e84f95e847f6af712a081fd1870f3f745bb57cf3f7500f614d6924.jpg) + +![](images/c31ccc8be685948a79ebf08f424a98aa5fea139e519c46f792a7b0f980a382d8.jpg) + +![](images/470de8612594878011a08354b2d8861b73b634905ccd84a06f2c4230ef846dba.jpg) + +![](images/d68c9268135a20c4cbd62c2bb7dc594e3501c94379500fac5c091f7b0500bbdd.jpg) + +![](images/a2a16e89ef27841815794d69ccafeb1bc4603b50b8dc2b45333e7fc04f4dba18.jpg) + +![](images/28f0c63e2b6fe453ab50b068c28e16f1c9defe5d65ce2360fb9be5936e86a08b.jpg) + +![](images/eda950e3dc6158c3fb880e090f647dc9a98186768ba5104fbd832e3c50fa4228.jpg) + +![](images/0d33af5aca59a6a4eb24f5dda9ddfe1e6b4a0531a43a0dcb6fd3e81dc02777ec.jpg) + +![](images/10bea310e3d8e01a91214a84ba0ab6d1f56f6e65a4c66ef538ea3b24795fd650.jpg) + +![](images/431f2d07ddeadb50a850e3db65dcb3c3ec078c8ac00b8160c71054f4967731ae.jpg) + +![](images/54cfa1b15dfe239b3ea6f887193c6824d34996f7f7bcfc900d1d0ea2bee52b95.jpg) + +![](images/18d33aa1f72cef1fb731ece3178c8303bf47e0bcbafaa240d82821c641422e14.jpg) + +![](images/cb6d6961170d3112b4a518fd5177f061fcbb18819dfd26bccd340422edadda11.jpg) +(a) + +![](images/99362fbe636c07095b6d171b96368b985b137720885139d9eb3f905af839020d.jpg) +(b) Layer 3 + +![](images/3fff0f398e1aa1569e66637ffec91dc67afaa1a81be5e66d884d8ab24ab3374a.jpg) +(c) Layer 6 + +![](images/bea506ac5ceabe52d09cbfc725a2d71b95c089c820e56772218a9b66a461ddd5.jpg) +(d) Layer 9 + +![](images/3f1efd2108978e94858d6811673be0a024007c5aaaf77087a6fc7c8d69aab04c.jpg) + +![](images/f52c4f3edf9c87344fa6381d7b178da849752645573eac773e51c948af51e288.jpg) +(f) Legend +Figure 3: Attribute relationships inferred by TabPFN v2. The first and third rows show PCA projections of the $d$ attribute tokens from all $N$ training instances at various layers for the churn and bank datasets. Colors indicate different attributes (see legend on the right). The second and fourth rows display the attribute-wise attention maps. Each matrix cell represents the average attention weight between attributes; the last element along each axis (e.g., the last column and row) corresponds to the label. The first plots in the second and fourth rows summarize the cosine similarity of attention maps across random seeds. See text for details. + +TabPFN v2's mechanisms. TabPFN v2 builds on prior token-based methods by representing each instance $\boldsymbol{x}_i$ as a sequence of tokens. However, rather than assigning a deterministic token to each attribute, TabPFN v2 samples random tokens at inference time. Specifically, it learns a shared vector $\boldsymbol{u} \in \mathbb{R}^k$ that lifts each element of $\boldsymbol{x}_i$ into a $k$ -dimensional space. To distinguish attributes, TabPFN v2 adds a random perturbation to each one. For the $j$ -th attribute (i.e., $x_i^j$ ), the representation becomes $x_i^j \cdot \boldsymbol{u} + r_j$ , where $r_j = W\boldsymbol{p}_j$ . Here, $p_j \in \mathbb{R}^{k'}$ is a randomly generated vector, and $W \in \mathbb{R}^{k \times k'}$ is a learned projection matrix that conditions the perturbation. The full instance $\boldsymbol{x}_i$ is then represented as: + +$$ +\left[ x _ {i} ^ {1} \cdot \boldsymbol {u} + \boldsymbol {r} _ {1}, \dots , x _ {i} ^ {d} \cdot \boldsymbol {u} + \boldsymbol {r} _ {d}, \tilde {\boldsymbol {y}} _ {i} \right] \in \mathbb {R} ^ {k \times (d + 1)}, \tag {1} +$$ + +where the last token $\tilde{\pmb{y}}_i$ encodes the label information (see Figure 1 for an illustration). + +# 5.2 TabPFN v2 Internalizes Attribute Token Learning + +TabPFN v2's randomized tokenization scheme eliminates the need to define attribute- or dataset-specific tokens across tasks, thereby syntactically enabling direct application of the pre-trained model. At first glance, this may appear to disregard the valuable semantic meaning of attributes. However, we show that through in-context learning, TabPFN v2 can consistently infer relationships among attributes within a dataset—despite the randomness introduced during tokenization. Specifically, we analyze the behavior of attribute tokens from three perspectives, as illustrated in Figure 3, using two representative downstream datasets: churn and bank. + +First, we visualize the attribute token embeddings (i.e., the first $d$ tokens in Equation (1)) across all $N$ training instances. The first and third rows of Figure 3 present PCA projections of these $N \times d$ tokens at the input stage and after transformer layers $\{3,6,9,12\}$ , with colors indicating different attributes. Initially, tokens from different attributes appear randomly scattered. However, as the input progresses through the transformer layers, these tokens become increasingly structured. For example, + +![](images/805b8bfe1d6992dd26678b759257015aaa3263d3bc768e604d97c867ca6c0ef8.jpg) + +![](images/c0484ab363b4e42700894f3a718feee960086f41b9524dd544be9f6598872d03.jpg) + +![](images/3c42a7f1b111dd006e87195e39c448072d82314dd0fd1fbf4396fed5239333b3.jpg) + +![](images/5f60812372819c171982313bf3b7538264f424e81653d139f3162f127e784866.jpg) + +![](images/246c8cae2f99e90feb7b308d5b968c57494c4729ddfc0b2fdd28171f6cefd4fc.jpg) + +![](images/2bccdd5f942650e8a8bdeb97189ebcd1d40597130b8f302e3ac366e489d38d95.jpg) + +![](images/aca48883318c74266e268c84cb7eb31755ebfa389de4e467bf4e089966798501.jpg) + +![](images/053a550c774ead0b2950a4e947fe6eec22adf4ee3373f6d509bf909c75aceb3a.jpg) + +![](images/bfffb78f012149208eca5c11c82e7d572180799c4e3ef51c4b9e31a51d93236d.jpg) + +![](images/731455a6758250d6c873eea098b3fceffcdca456ea9265b992e61d540dac6150.jpg) + +![](images/7f8a96f035932082437932a0a2b5162dc801f45a1b05798e0207990b20da5cb0.jpg) + +![](images/cc10e0c888ec39379bd32463e7445c731b46d199de8375a22216d4c7f3872d6c.jpg) + +![](images/59e4320e1e6eed47d81e3df45334587a0681d2eeccde1b30ec7cda2e717fcf7f.jpg) + +![](images/56061db66cb248c324ece7e591975264841907e9ade4c1f0d8a72f58c66b7bbe.jpg) + +![](images/8a28eab9f78c08ff768119167b427e9a9893a96783cf9a0681ff3a4b7b3bb0f9.jpg) + +![](images/0e6f771b9b657c31a0175ffb83b9ec51be4d5b7691cfa543ed9083638fc8f987.jpg) + +![](images/932cf7ba18aaa9f7486f3758f13468eb3c08ff72c1837de06fbc79e702d2bdc2.jpg) + +![](images/d8dd8f65a7fe052fce418ad66671f092aada62e192a7ed9b8461d6df0d2f7ce6.jpg) + +![](images/15fe074895cbd6118fd07d93c184c2fa426a20c6e5e46c9ffc7a798eb560aa3d.jpg) +(a) Raw Feature + +![](images/c039b463a1742747a9685347a2f0ae65d69e05983bd52f0b646456b5c16f0ae6.jpg) +(b) Vanilla + +![](images/ac5b5a9f4b57939e83effcd709a661ef59bc18d4233c27709865d48bd7307c7d.jpg) +(c) Layer 3 + +![](images/2b2e9a03a664aaba8d650f674924fd8f52e4091d12d37daece6b16d6d3734640.jpg) +(d) Layer 6 +Figure 4: Visualization of the extracted instance features from four datasets: churn (first row, binary), bank (second row, binary), website_phishing (third row, three classes), and $KDD$ (fourth row, binary). Blue and red indicate classes; darker crosses and lighter circles denote training and test samples. (a) shows the raw input features ( $e.g., x_i$ ), while (b) presents embeddings from the vanilla strategy. (c)-(f) display embeddings produced by our method at different layers. Classification accuracy is reported by training a linear logistic regression model on the training embeddings and evaluating on the test set. + +![](images/dc818d5cd53a679e859721155080fe0b8abf0742ab6bc504fd79b3fa6d121f8c.jpg) +(e) Layer 9 + +![](images/3153c0d4e4a55204342fa3285a8d1e79d9acec4385059976c3c257c20015a52e.jpg) +(f) Layer 12 + +in the bank dataset (which predicts term deposit subscriptions), attributes such as "job," "education," and "balance" eventually cluster into semantically coherent groups. + +Second, we examine attribute-wise attention patterns across layers, including attention to the label token. The second and fourth rows of Figure 3 show heatmaps of attention weights averaged over heads and training instances. Each row in the heatmap represents the attention distribution from one attribute to all others; the last element along each axis (e.g., the last column and row) corresponds to the label. Darker shades indicate stronger attention. We observe a consistent pattern across datasets: in early layers, attributes predominantly attend to the label token, likely to absorb task-specific signals. In intermediate layers, attention becomes more uniformly distributed, facilitating information exchange across attributes. In deeper layers, attention concentrates on semantically relevant attributes, suggesting the model has inferred inter-attribute relationships useful for prediction. + +Lastly, to assess robustness against random token initialization, we compute attribute-wise attention weights across 10 runs. The cosine similarities and variances of these attention patterns are summarized in the first plots of the second and fourth rows in Figure 3. The results confirm that attention patterns remain stable across runs, except for the first layer. + +Remark. The above results suggest that TabPFN v2 can reliably infer meaningful attribute relationships through in-context learning. Although input embeddings are randomized, they consistently differentiate attributes across instances—functionally akin to one-hot encodings. Pre-training on diverse tasks thus enables the model to extract predictive patterns (e.g., co-occurrence across attributes, value distributions, and relative magnitudes) directly from the statistical structure of each dataset, without relying on pre-defined attribute semantics. As a result, the model effectively internalizes attribute token learning within the inference process. See the Appendix for further discussion. + +# 6 TabPFN v2 Can Be Transformed into an Effective Feature Encoder + +In Section 5, we show that TabPFN v2's in-context learning process infers meaningful attribute relationships. Here, we examine whether TabPFN v2 also produces separable instance representations. + +# 6.1 Naive Feature Extraction Fails + +As shown in Figure 1 (left), TabPFN v2 makes predictions based on the output token corresponding to the (dummy) label embedding $\tilde{\pmb{y}}^*$ of the test instance. This output token can thus be interpreted as the + +instance embedding for the test example. A natural extension to obtain embeddings for the training instances is to extract the output tokens corresponding to the training label embeddings $\{\tilde{y}_i\}_{i=1}^N$ . + +However, as shown in Figure 4 (b), this naive approach leads to surprisingly discrepant feature distributions between training (darker cross) and test (lighter circle) examples. As a result, a linear classifier trained on these embeddings performs poorly on the test set. We attribute this discrepancy to the distinct roles of labeled training data and unlabeled test data in TabPFN v2's in-context learning process. Specifically, the label embeddings for the training instances are derived from true labels, whereas those for the test instances rely on dummy labels. This mismatch renders the resulting output embeddings non-comparable between training and test instances. + +# 6.2 Leave-one-fold-out Feature Extraction + +To address this challenge, we propose a leave-one-fold-out strategy that enables the extraction of comparable embeddings for training and test data. In the TabPFN v2 framework, we treat examples with true labels as the support set $S$ , and those with dummy labels as the query set $Q$ . Under the standard configuration, $S$ corresponds to the labeled training set and $Q$ to the unlabeled test instances. To extract comparable embeddings for the training examples, they must also be included in $Q$ with dummy label embeddings. This, however, creates a dilemma: effective in-context learning relies on maximizing the size of $S$ to ensure sufficient knowledge transfer to $Q$ . Including training examples in $Q$ thus competes with the need to keep $S$ as large as possible. + +To overcome this dilemma, we partition the training set into multiple folds (e.g., 10). In each round, one fold serves as $\mathcal{Q}$ with dummy labels used for embedding extraction while the remaining folds form $S$ with true labels. This setup preserves sufficient label supervision in $S$ while enabling the extraction of embeddings for training instances in $\mathcal{Q}$ . Results in Figure 4 (c)-(f) show that embeddings extracted by this strategy (with 10 folds) more faithfully capture dataset structure. We observe that TabPFN v2 simplifies the original tabular data distributions, transforming datasets into nearly linearly separable embedding spaces—especially after intermediate transformer layers. + +We also experimented with a variant that uses the same context and query without partitioning, where the context contains all training samples and the query set includes both training and test points. This "non-partitioned" strategy improves upon the vanilla feature extraction baseline but still underperforms compared to our proposed leave-one-fold-out method. We attribute this to role ambiguity: query points that appear in the support set (either with dummy or ground-truth labels) are treated inconsistently, preventing the network from fully distinguishing between training and test roles and thereby degrading feature consistency. Detailed results for this variant are provided in the Appendix. + +# 6.3 Validation of Embedding Quality + +To validate the quality of the extracted embeddings, we train a logistic regression on embeddings derived from the training set and evaluate it on test set embeddings. The average rank across 29 classification datasets from the tiny benchmark2 in [75] is reported in Table 2. + +Table 2: Average rank (lower is better) of TabPFN v2 and linear classifiers trained on the extracted embeddings across 29 classification datasets. Combined: embeddings from up to three layers (from the 12 available layers) are selected and concatenated, based on the validation set performance. + +
↓ TabPFN v2 Vanilla Layer 6 Layer 9 Layer 12 Combined
Rank2.695.974.284.002.121.94
+ +Remarkably, training a simple linear classifier on the extracted embeddings achieves performance comparable to that of TabPFN v2's in-context learner. Furthermore, concatenating embeddings from multiple layers (e.g., both output and intermediate representations) can sometimes lead to even better results. These findings underscore TabPFN v2's potential as a strong and versatile feature encoder, suggesting broader applicability in downstream tasks such as tabular data analysis. + +# 7 Improving TabPFN v2 via Test-Time Divide-and-Conquer + +This section addresses the limitations discussed in Section 4.3, aiming to extend TabPFN v2's applicability beyond the boundaries. Specifically, we propose post-hoc divide-and-conquer strategies inspired by Chain-of-Thought (CoT) prompting [70], which decompose challenging tasks into simpler subtasks that TabPFN v2 can effectively handle. + +# 7.1 High Dimension Datasets + +High-dimensional datasets [36] present a unique challenge due to the quadratic complexity of TabPFN v2 with respect to the number of dimensions. To mitigate this, we propose subsampling the feature space into smaller subsets, processing each subset independently, and combining the predictions in an ensemble (bagging) fashion, similar to random forests [8]. + +In detail, we iteratively sample $m$ subsets, each containing $d' < d$ randomly selected attributes. For each subset, we leverage TabPFN v2's ability to handle lower-dimensional data to obtain predictions. We denote this divide-and-conquer and then ensemble strategy as TabPFN v2*, which aggregates outputs using averaging (for regression) or majority voting (for classification). To address high-dimensional tasks, we introduce a baseline variant, TabPFN v2-PCA, which incorporates dimensionality reduction. Specifically, TabPFN v2-PCA reduces the feature dimension to 500 using PCA to satisfy the input constraints of TabPFN v2. This process is repeated multiple times with different PCA projections, and the resulting predictions are aggregated via bagging to improve robustness. + +Figure 5 (left) summarizes the results on 18 high-dimensional classification datasets. A variant that utilizes PCA to reduce the dimensionality, together with bagging, resolves the dimensionality issue to some extent. TabPFN v2* with $d' = 500$ and $m = \lceil d / d' \rceil$ significantly increases the mean accuracy (to the highest), effectively extending TabPFN v2's scalability to datasets with $d \geq 2000$ . + +# 7.2 Multi-Class Problems with More Than 10 Classes + +To extend TabPFN v2 to tasks with more than 10 categories, we propose a decimal encoding approach that decomposes multi-class problems into multiple 10-class subproblems, ensuring compatibility with TabPFN v2's constraints. + +For a task with $C > 10$ classes, we encode each label $y \in [C]$ as a $t$ -digit decimal representation, where $t = \lceil \log_{10} C \rceil$ . For each digit position $j \in \{1, \dots, t\}$ , we train a separate TabPFN v2 model $f_j$ to predict the $j$ -th digit. During inference, the predicted digits are reconstructed to obtain the final class label. This strategy is also developed in [48], and we denote it as TabPFN v2-DPT. As the decimal encoding inherently introduces artificial correlations among classes — classes that share the same digit at a given position are grouped together, even if they are semantically unrelated. To mitigate this effect, our TabPFN v2* randomly permutes the class-to-digit mapping $\sqrt{C}$ times, leading to different groupings in each run, and the prediction results are ensembles to improve robustness. + +We consider the following variants of TabPFN v2 to address the 10-class limit in classification tasks: + +- TabPFN v2-ECOC: We use the implementation provided in the official TabPFN extensions repository, which applies Error-Correcting Output Codes (ECOC). +- TabPFN v2-DPT: We encode each class label as a $t$ -digit decimal string and train a separate TabPFN v2 to predict each digit. For instance, a 15-class problem is decomposed into two subproblems: one for the tens digit (classes $\{0,1\}$ ) and one for the ones digit (classes $\{0,\dots ,9\}$ ). The following list of classes is used for the training. + +The predicted digits are then decoded to recover the final class label. + +We implement TabPFN v2 * based on TabPFN v2-DPT for efficiency. Specifically, TabPFN v2 * permutes the class-to-digit mapping $\sqrt{C}$ times. For fair comparison, we also increase the number of ensembles in TabPFN v2-DPT to $\sqrt{C}$ per digit to match the total number of predictions. As shown in Figure 5 (middle), this approach achieves the second-best mean accuracy on 12 datasets with more than 10 classes while preserving computational efficiency. + +# 7.3 Large-Scale Datasets + +For large-scale datasets, we randomly sample 10,000 training examples from the full training set as the support set and treat the remaining training examples and test instances as the query set. We extract their embeddings to form a new tabular dataset, on which a logistic regression classifier is trained to make predictions on the test set embeddings. This process is repeated four times, and the final predictions are aggregated. We denote this version as TabPFN v2\*-SQ. + +We also investigate integrating TabPFN v2 with decision trees to handle large-scale tasks. We note that a similar strategy was mentioned in [28] to handle within-dataset heterogeneity for a drastically different purpose. Specifically, we sample 32 subsets from the original training set, each containing + +![](images/f21a400e78fad4c008544c55802ce3b7829683abd285c787723405d31d010452.jpg) +Figure 5: “*” indicates our extension. Left: Mean accuracy on 18 high-dimensional datasets. “-PCA” is another variant using PCA to reduce dimensions. Middle: Mean accuracy on 12 datasets with more than 10 classes. “-ECOC” denotes the multi-class ECOC strategy implemented by [28]. Right: Average rank on 18 large-scale datasets. “-B” refers to the variant that randomly subsamples 10,000 training examples four times and aggregates their predictions. “-K” denotes the variant that selects a representative subset of 10,000 training examples based on proximity to prototypes obtained via KMeans. All variants improve TabPFN v2. + +![](images/be86fcd7bdf0b99b5d213143b82aca43f84d6b0c5537617f343bd4bd6bda5f4d.jpg) + +![](images/e367bd776ab41ae749c1331ac0744c8552654b5ab19033a05d2297fa9c14466e.jpg) + +$60\%$ of the original data (sampled without replacement). For each subset, we first train a shallow decision tree by setting the minimum number of samples required to split an internal node to 10,000. The decision tree partitions the training set into smaller, more manageable subsets. During inference, a test instance is first passed through each of the shallow decision tree to a leaf node and then predicted by the corresponding TabPFN v2 model. The predictions from all 32 models are aggregated. We denote this extension as TabPFN v2\*-DF. + +We consider the following variants of TabPFN v2 to scale to larger datasets: + +- TabPFN v2-DT: A shallow decision tree is trained with a minimum split size of 10,000. The tree partitions the dataset into smaller subsets, and a separate TabPFN v2 model is applied to each leaf node. At inference, a test instance is routed through the tree to a corresponding leaf, where it is predicted by the respective TabPFN v2 model. +- TabPFN v2-B: A bagging-based variant that randomly samples 10,000 training examples four times and aggregates their predictions. +- TabPFN v2-K: Selects a representative subset of 10,000 training examples based on proximity to KMeans-derived prototypes. + +Our TabPFN v2*-DF is an extension of TabPFN v2-DT to a forest-based ensemble. Specifically, we sample 32 subsets from the original training data, each containing $60\%$ of the samples (without replacement), and train a separate TabPFN v2*-DT model on each subset. During inference, predictions from all 32 models are aggregated—e.g., by majority voting or averaging—depending on the task type. + +Figure 5 (right) shows the average rank results, including TabPFN v2\*-SQ and TabPFN v2\*-DF alongside variants using bagging and KMeans-based sampling. We observe that all variants improve upon the vanilla TabPFN v2 on large-scale datasets, with TabPFN v2\*-DF and TabPFN v2\*-SQ achieving the most significant improvement. + +# 8 Conclusion + +We present a timely investigation into TabPFN v2, a groundbreaking foundation model for tabular tasks. Our analysis uncovers the core mechanism behind TabPFN v2's strong performance across heterogeneous tabular datasets: it can infer attribute relationships on-the-fly—even from randomly initialized token inputs—without relying on pre-defined semantics or learning dataset-specific representations. We also demonstrate that TabPFN v2 can be repurposed as a powerful feature encoder, enabling broader applications such as data visualization and diagnostic analysis. To address its limitations in more complex data regimes, we introduce post-hoc divide-and-conquer strategies that extend TabPFN v2's utility without requiring model re-training. Together, these contributions offer fresh insights into advancing the development and application of foundation models for tabular data. + +# Acknowledgment + +This work is partially supported by National Key R&D Program of China (2024YFE0202800), Fundamental Research Funds for the Central Universities (2024300373), Collaborative Innovation Center of Novel Software Technology and Industrialization. + +# References + +[1] Takuya Akiba, Shotaro Sano, Toshihiko Yanase, Takeru Ohta, and Masanori Koyama. Optuna: A next-generation hyperparameter optimization framework. In KDD, pages 2623-2631, 2019. +[2] Xavier Amatriain, Alejandro Jaimes, Nuria Oliver, and Josep M Pujol. Data mining methods for recommender systems. In Recommender systems handbook, pages 39-71. Springer, 2010. +[3] Michael Arbel, David Salinas, and Frank Hutter. Equitabpfn: A target-permutation equivariant prior fitted networks. CoRR, abs/2502.06684, 2025. +[4] Sercan Ö. Arik and Tomas Pfister. Tabnet: Attentive interpretable tabular learning. In AAAI, pages 6679-6687, 2021. +[5] Sarkhan Badirli, Xuanqing Liu, Zhengming Xing, Avradeep Bhowmik, and Sathiya S. Keerthi. Gradient boosting neural networks: Grownet. CoRR, abs/2002.07971, 2020. +[6] David Bonet, Daniel Mas Montserrat, Xavier Giroi Nieto, and Alexander G. Ioannidis. Hyperfast: Instant classification for tabular data. In AAAI, pages 11114-11123, 2024. +[7] Vadim Borisov, Tobias Leemann, Kathrin Seßler, Johannes Haug, Martin Pawelczyk, and Gjergji Kasneci. Deep neural networks and tabular data: A survey. IEEE Transactions Neural Networks and Learning Systems, 35(6):7499-7519, 2024. +[8] Leo Breiman. Random forests. Machine Learning, 45(1):5-32, 2001. +[9] Jintai Chen, KuanLun Liao, Yanwen Fang, Danny Chen, and Jian Wu. Tabcaps: A capsule neural network for tabular data classification with bow routing. In ICLR, 2023. +[10] Jintai Chen, Kuanlun Liao, Yao Wan, Danny Z. Chen, and Jian Wu. Danets: Deep abstract networks for tabular data classification and regression. In AAAI, pages 3930-3938, 2022. +[11] Jintai Chen, Jiahuan Yan, Qiyuan Chen, Danny Ziyi Chen, Jian Wu, and Jimeng Sun. Can a deep learning model be a sure bet for tabular prediction? In KDD, pages 288-296, 2024. +[12] Tianqi Chen and Carlos Guestrin. Xgboost: A scalable tree boosting system. In KDD, pages 785-794, 2016. +[13] Manuel Fernández Delgado, Eva Cernadas, Senén Barro, and Dinani Gomes Amorim. Do we need hundreds of classifiers to solve real world classification problems? Journal of Machine Learning Research, 15(1):3133-3181, 2014. +[14] Janez Demsrar. Statistical comparisons of classifiers over multiple data sets. Journal of Machine Learning Research, 7:1-30, 2006. +[15] Felix den Breejen, Sangmin Bae, Stephen Cha, and Se-Young Yun. Fine-tuned in-context learning transformers are excellent tabular data classifiers. CoRR, abs/2405.13396, 2025. +[16] Benjamin Feuer, Chinmay Hegde, and Niv Cohen. Scaling tabPFN: Sketching and feature selection for tabular prior-data fitted networks. CoRR, abs/2311.10609, 2023. +[17] Benjamin Feuer, Robin Tibor Schirrmeister, Valeriia Cherepanova, Chinmay Hegde, Frank Hutter, Micah Goldblum, Niv Cohen, and Colin White. Tunetables: Context optimization for scalable prior-data fitted networks. In NeurIPS, pages 83430-83464, 2024. +[18] Matthias Feurer, Aaron Klein, Katharina Eggensperger, Jost Tobias Springenberg, Manuel Blum, and Frank Hutter. Efficient and robust automated machine learning. In NIPS, 2015. + +[19] Yury Gorishniy, Akim Kotelnikov, and Artem Babenko. Tabm: Advancing tabular deep learning with parameter-efficient ensembling. In ICLR, 2025. +[20] Yury Gorishniy, Ivan Rubachev, and Artem Babenko. On embeddings for numerical features in tabular deep learning. In NeurIPS, pages 24991-25004, 2022. +[21] Yury Gorishniy, Ivan Rubachev, Nikolay Kartashev, Daniil Shlenskii, Akim Kotelnikov, and Artem Babenko. Tabr: Tabular deep learning meets nearest neighbors in 2023. In ICLR, 2024. +[22] Yury Gorishniy, Ivan Rubachev, Valentin Khrulkov, and Artem Babenko. Revisiting deep learning models for tabular data. In NeurIPS, pages 18932-18943, 2021. +[23] Léo Grinsztajn, Edouard Oyallon, and Gael Varoquaux. Why do tree-based models still outperform deep learning on typical tabular data? In NeurIPS, pages 507-520, 2022. +[24] Isabelle Guyon, Lisheng Sun-Hosoya, Marc Boulle, Hugo Jair Escalante, Sergio Escalera, Zhengying Liu, Damir Jajetic, Bisakha Ray, Mehreen Saeed, Michele Sebag, et al. Analysis of the automl challenge series. Automated Machine Learning, 177:177-219, 2019. +[25] Sungwon Han, Jinsung Yoon, Sercan Ö. Arik, and Tomas Pfister. Large language models can automatically engineer features for few-shot tabular learning. In ICML, pages 17454-17479, 2024. +[26] Stefan Hegselmann, Alejandro Buendia, Hunter Lang, Monica Agrawal, Xiaoyi Jiang, and David Sontag. Tabllm: few-shot classification of tabular data with large language models. In AISTATS, pages 5549-5581, 2023. +[27] Noah Hollmann, Samuel Müller, Katharina Eggensperger, and Frank Hutter. Tabpfn: A transformer that solves small tabular classification problems in a second. In ICLR, 2023. +[28] Noah Hollmann, Samuel Müller, Lennart Purucker, Arjun Krishnakumar, Max Körfer, Shi Bin Hoo, Robin Tibor Schirrmeister, and Frank Hutter. Accurate predictions on small data with a tabular foundation model. Nature, 637(8045):319-326, 2025. +[29] David Holzmuller, Léo Grinsztajn, and Ingo Steinwart. Better by default: Strong pre-tuned mlps and boosted trees on tabular data. In NeurIPS, pages 26577-26658, 2024. +[30] Shi Bin Hoo, Samuel Müller, David Salinas, and Frank Hutter. The tabular foundation model tabpfn outperforms specialized time series forecasting models based on simple features. CoRR, abs/2501.02945, 2025. +[31] Xin Huang, Ashish Khetan, Milan Cvitkovic, and Zohar S. Karnin. Tabtransformer: Tabular data modeling using contextual embeddings. CoRR, abs/2012.06678, 2020. +[32] Stephanie L Hyland, Martin Faltys, Matthias Hüser, Xinrui Lyu, Thomas Gumbsch, Cristóbal Esteban, Christian Bock, Max Horn, Michael Moor, Bastian Rieck, et al. Early prediction of circulatory failure in the intensive care unit using machine learning. Nature medicine, 26(3):364-373, 2020. +[33] Ovidiu Ivanciuc et al. Applications of support vector machines in chemistry. Reviews in computational chemistry, 23:291, 2007. +[34] Tomoharu Iwata and Atsutoshi Kumagai. Meta-learning from tasks with heterogeneous attribute spaces. In NeurIPS, pages 6053-6063, 2020. +[35] Alan Jeffares, Tennison Liu, Jonathan Crabbé, Fergus Imrie, and Mihaela van der Schaar. Tangos: Regularizing tabular neural networks through gradient orthogonalization and specialization. In ICLR, 2023. +[36] Xiangjian Jiang, Andrei Margeloiu, Nikola Simidjevski, and Mateja Jamnik. Protopate: Prototype-based neural networks with global-to-local feature selection for tabular biomedical data. In ICML, pages 21844-21878, 2024. +[37] Arlind Kadra, Marius Lindauer, Frank Hutter, and Josif Grabocka. Well-tuned simple nets excel on tabular datasets. In NeurIPS, pages 23928-23941, 2021. + +[38] Guolin Ke, Qi Meng, Thomas Finley, Taifeng Wang, Wei Chen, Weidong Ma, Qiwei Ye, and Tie-Yan Liu. Lightgbm: A highly efficient gradient boosting decision tree. In NIPS, pages 3146-3154, 2017. +[39] Myung Jun Kim, Léo Grinsztajn, and Gaël Varoquaux. CARTE: pretraining and transfer for tabular learning. In ICML, pages 23843-23866, 2024. +[40] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloé Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólar, and Ross B. Girshick. Segment anything. In ICCV, pages 3992-4003, 2023. +[41] Günter Klambauer, Thomas Unterthiner, Andreas Mayr, and Sepp Hochreiter. Self-normalizing neural networks. In NIPS, pages 971–980, 2017. +[42] Ravin Kohli, Matthias Feurer, Katharina Eggensperger, Bernd Bischl, and Frank Hutter. Towards quantifying the effect of datasets for benchmarking: A look at tabular machine learning. In ICLR Workshop, 2024. +[43] Boris Kovalerchuk and Evgenii Vityaev. Data mining in finance: advances in relational and hybrid methods. Springer Science & Business Media, 2005. +[44] Lang Liu, Mahdi Milani Fard, and Sen Zhao. Distribution embedding networks for generalization from a diverse set of classification tasks. Transactions on Machine Learning Research, 2022. +[45] Si-Yang Liu and Han-Jia Ye. Tabpfn unleashed: A scalable and effective solution to tabular classification problems. CoRR, abs/2502.02527, 2025. +[46] Si-Yang Liu and Han-Jia Ye. Tabpfn unleashed: A scalable and effective solution to tabular classification problems. In ICML, 2025. +[47] Junwei Ma, Apoorv Dankar, George Stein, Guangwei Yu, and Anthony L. Caterini. Tabpfgen - tabular data generation with tabpfn. CoRR, abs/2406.05216, 2024. +[48] Junwei Ma, Valentin Thomas, Rasa Hosseinzadeh, Hamidreza Kamkari, Alex Labach, Jesse C. Cresswell, Keyvan Golestan, Guangwei Yu, Maksims Volkovs, and Anthony L. Caterini. Tabdpt: Scaling tabular foundation models. CoRR, abs/2410.18164, 2024. +[49] Duncan C. McElfresh, Sujay Khandagale, Jonathan Valverde, Vishak Prasad C., Ganesh Ramakrishnan, Micah Goldblum, and Colin White. When do neural nets outperform boosted trees on tabular data? In NeurIPS, pages 76336-76369, 2023. +[50] Andreas C. Mueller, Carlo Curino, and Raghu Ramakrishnan. Mothernet: Fast training and inference via hyper-network transformers. In ICLR, 2025. +[51] Niklas Muennighoff, Zitong Yang, Weijia Shi, Xiang Lisa Li, Li Fei-Fei, Hannaneh Hajishirzi, Luke Zettlemoyer, Percy Liang, Emmanuel Candès, and Tatsunori Hashimoto. s1: Simple test-time scaling. CoRR, abs/2501.19393, 2025. +[52] Youssef Nader, Leon Sixt, and Tim Landgraf. DNNR: differential nearest neighbors regression. In ICML, pages 16296-16317, 2022. +[53] Thomas Nagler. Statistical foundations of prior-data fitted networks. In Andreas Krause, Emma Brunskill, Kyunghyun Cho, Barbara Engelhardt, Sivan Sabato, and Jonathan Scarlett, editors, ICML, pages 25660-25676, 2023. +[54] Soma Onishi, Kenta Oono, and Kohei Hayashi. Tabret: Pre-training transformer-based tabular models for unseen columns. CoRR, abs/2303.15747, 2023. +[55] Sergei Popov, Stanislav Morozov, and Artem Babenko. Neural oblivious decision ensembles for deep learning on tabular data. In ICLR, 2020. +[56] Liudmila Ostroumova Prokhorenkova, Gleb Gusev, Aleksandr Vorobev, Anna Veronika Dorogush, and Andrey Gulin. Catboost: unbiased boosting with categorical features. In NeurIPS, pages 6639-6649, 2018. + +[57] Jingang Qu, David Holzmuller, Gael Varoquaux, and Marine Le Morvan. Tabicl: A tabular foundation model for in-context learning on large data. In ICML, 2025. +[58] Cristóbal Romero and Sebastián Ventura. Educational data mining: a review of the state of the art. IEEE Transactions on Systems, Man, and Cybernetics, 40(6):601-618, 2010. +[59] Ivan Rubachev, Artem Alekberov, Yury Gorishniy, and Artem Babenko. Revisiting pretraining objectives for tabular deep learning. CoRR, abs/2207.03208, 2022. +[60] Ivan Rubachev, Nikolay Kartashev, Yury Gorishniy, and Artem Babenko. Tabred: A benchmark of tabular machine learning in-the-wild. In ICLR, 2025. +[61] Sergio Ruiz-Villafranca, José Roldán Gómez, Juan Manuel Castelo Gómez, Javier Carrillo Mondéjar, and José Luis Martínez. A tabpfn-based intrusion detection system for the industrial internet of things. The Journal of Supercomputing, 80(14):20080-20117, 2024. +[62] Junhong Shen, Liam Li, Lucio M Dery, Corey Staten, Mikhail Khodak, Graham Neubig, and Ameet Talwalkar. Cross-modal fine-tuning: Align then refine. In ICML, pages 31030-31056, 2023. +[63] Gowthami Somepalli, Avi Schwarzschild, Micah Goldblum, C. Bayan Bruss, and Tom Goldstein. SAINT: Improved neural networks for tabular data via row attention and contrastive pre-training. In NeurIPS Workshop, 2022. +[64] Weiping Song, Chence Shi, Zhiping Xiao, Zhijian Duan, Yewen Xu, Ming Zhang, and Jian Tang. Autoint: Automatic feature interaction learning via self-attentive neural networks. In CIKM, pages 1161–1170, 2019. +[65] Valentin Thomas, Junwei Ma, Rasa Hosseinzadeh, Keyvan Golestan, Guangwei Yu, Maksims Volkovs, and Anthony L. Caterini. Retrieval & fine-tuning for in-context tabular models. In NeurIPS, pages 108439-108467, 2024. +[66] Boris van Breugel and Mihaela van der Schaar. Position: Why tabular foundation models should be a research priority. In ICML, pages 48976-48993, 2024. +[67] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NIPS, 2017. +[68] Ruoxi Wang, Rakesh Shivanna, Derek Zhiyuan Cheng, Sagar Jain, Dong Lin, Lichan Hong, and Ed H. Chi. DCN V2: improved deep & cross network and practical lessons for web-scale learning to rank systems. In WWW, pages 1785-1797, 2021. +[69] Zifeng Wang, Chufan Gao, Cao Xiao, and Jimeng Sun. Anypredict: Foundation model for tabular prediction. CoRR, abs/2305.12081, 2023. +[70] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed H. Chi, Quoc V. Le, and Denny Zhou. Chain-of-thought prompting elicits reasoning in large language models. In NeurIPS, pages 24824-24837, 2022. +[71] Xumeng Wen, Han Zhang, Shun Zheng, Wei Xu, and Jiang Bian. From supervised to generative: A novel paradigm for tabular deep learning with large language models. In KDD, pages 3323-3333, 2024. +[72] Jing Wu, Suiyao Chen, Qi Zhao, Renat Sergazinov, Chen Li, Shengjie Liu, Chongchao Zhao, Tianpei Xie, Hanqing Guo, Cheng Ji, Daniel Cociorva, and Hakan Brunzell. Switchtab: Switched autoencoders are effective tabular learners. In AAAI, pages 15924-15933, 2024. +[73] Derek Xu, Olcay Cirit, Reza Asadi, Yizhou Sun, and Wei Wang. Mixture of in-context prompters for tabular pfns. CoRR, abs/2405.16156, 2024. +[74] Jiahuan Yan, Bo Zheng, Hongxia Xu, Yiheng Zhu, Danny Z. Chen, Jimeng Sun, Jian Wu, and Jintai Chen. Making pre-trained language models great on tabular prediction. In ICLR, 2024. +[75] Han-Jia Ye, Si-Yang Liu, Hao-Run Cai, Qi-Le Zhou, and De-Chuan Zhan. A closer look at deep learning on tabular data. CoRR, abs/2407.00956, 2024. + +[76] Han-Jia Ye, Huai-Hong Yin, De-Chuan Zhan, and Wei-Lun Chao. Revisiting nearest neighbor for tabular data: A deep tabular baseline two decades later. In ICLR, 2025. +[77] Han-Jia Ye, Qi-Le Zhou, Huai-Hong Yin, De-Chuan Zhan, and Wei-Lun Chao. Rethinking pre-training in tabular data: A neighborhood embedding perspective. CoRR, abs/2311.00055, 2025. +[78] Hangting Ye, Wei Fan, Xiaozhuang Song, Shun Zheng, He Zhao, Dan dan Guo, and Yi Chang. Ptarl: Prototype-based tabular representation learning via space calibration. In ICLR, 2024. +[79] Tianping Zhang, Shaowen Wang, Shuicheng Yan, Jian Li, and Qian Liu. Generative table pre-training empowers models for tabular prediction. In EMNLP, pages 14836-14854, 2023. +[80] Yiyuan Zhang, Kaixiong Gong, Kaipeng Zhang, Hongsheng Li, Yu Qiao, Wanli Ouyang, and Xiangyu Yue. Meta-transformer: A unified framework for multimodal learning. CoRR, abs/2307.10802, 2023. +[81] Ce Zhou, Qian Li, Chen Li, Jun Yu, Yixin Liu, Guangjing Wang, Kai Zhang, Cheng Ji, Qiben Yan, Lifang He, et al. A comprehensive survey on pretrained foundation models: A history from bert to chatgpt. International Journal of Machine Learning and Cybernetics, pages 1-65, 2024. +[82] Bingzhao Zhu, Xingjian Shi, Nick Erickson, Mu Li, George Karypis, and Mahsa Shoaran. Xtab: Cross-table pretraining for tabular transformers. In ICML, pages 43181-43204, 2023. + +# NeurIPS Paper Checklist + +# 1. Claims + +Question: Do the main claims made in the abstract and introduction accurately reflect the paper's contributions and scope? + +Answer: [Yes] + +Justification: The abstract and introduction state the claims made, including the contributions made in the paper and important assumptions and limitations. + +Guidelines: + +- The answer NA means that the abstract and introduction do not include the claims made in the paper. +- The abstract and/or introduction should clearly state the claims made, including the contributions made in the paper and important assumptions and limitations. A No or NA answer to this question will not be perceived well by the reviewers. +- The claims made should match theoretical and experimental results, and reflect how much the results can be expected to generalize to other settings. +- It is fine to include aspirational goals as motivation as long as it is clear that these goals are not attained by the paper. + +# 2. Limitations + +Question: Does the paper discuss the limitations of the work performed by the authors? + +Answer: [Yes] + +Justification: In Appendix. + +Guidelines: + +- The answer NA means that the paper has no limitation while the answer No means that the paper has limitations, but those are not discussed in the paper. +- The authors are encouraged to create a separate "Limitations" section in their paper. +- The paper should point out any strong assumptions and how robust the results are to violations of these assumptions (e.g., independence assumptions, noiseless settings, model well-specification, asymptotic approximations only holding locally). The authors should reflect on how these assumptions might be violated in practice and what the implications would be. + +- The authors should reflect on the scope of the claims made, e.g., if the approach was only tested on a few datasets or with a few runs. In general, empirical results often depend on implicit assumptions, which should be articulated. +- The authors should reflect on the factors that influence the performance of the approach. For example, a facial recognition algorithm may perform poorly when image resolution is low or images are taken in low lighting. Or a speech-to-text system might not be used reliably to provide closed captions for online lectures because it fails to handle technical jargon. +- The authors should discuss the computational efficiency of the proposed algorithms and how they scale with dataset size. +- If applicable, the authors should discuss possible limitations of their approach to address problems of privacy and fairness. +- While the authors might fear that complete honesty about limitations might be used by reviewers as grounds for rejection, a worse outcome might be that reviewers discover limitations that aren't acknowledged in the paper. The authors should use their best judgment and recognize that individual actions in favor of transparency play an important role in developing norms that preserve the integrity of the community. Reviewers will be specifically instructed to not penalize honesty concerning limitations. + +# 3. Theory assumptions and proofs + +Question: For each theoretical result, does the paper provide the full set of assumptions and a complete (and correct) proof? + +Answer: [NA] + +Justification: This paper does not include theoretical results. + +Guidelines: + +- The answer NA means that the paper does not include theoretical results. +- All the theorems, formulas, and proofs in the paper should be numbered and cross-referenced. +- All assumptions should be clearly stated or referenced in the statement of any theorems. +- The proofs can either appear in the main paper or the supplemental material, but if they appear in the supplemental material, the authors are encouraged to provide a short proof sketch to provide intuition. +- Inversely, any informal proof provided in the core of the paper should be complemented by formal proofs provided in appendix or supplemental material. +- Theorems and Lemmas that the proof relies upon should be properly referenced. + +# 4. Experimental result reproducibility + +Question: Does the paper fully disclose all the information needed to reproduce the main experimental results of the paper to the extent that it affects the main claims and/or conclusions of the paper (regardless of whether the code and data are provided or not)? + +Answer: [Yes] + +Justification: We have disclosed all key information necessary to reproduce the experimental results. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- If the paper includes experiments, a No answer to this question will not be perceived well by the reviewers: Making the paper reproducible is important, regardless of whether the code and data are provided or not. +- If the contribution is a dataset and/or model, the authors should describe the steps taken to make their results reproducible or verifiable. +- Depending on the contribution, reproducibility can be accomplished in various ways. For example, if the contribution is a novel architecture, describing the architecture fully might suffice, or if the contribution is a specific model and empirical evaluation, it may be necessary to either make it possible for others to replicate the model with the same dataset, or provide access to the model. In general, releasing code and data is often + +one good way to accomplish this, but reproducibility can also be provided via detailed instructions for how to replicate the results, access to a hosted model (e.g., in the case of a large language model), releasing of a model checkpoint, or other means that are appropriate to the research performed. + +- While NeurIPS does not require releasing code, the conference does require all submissions to provide some reasonable avenue for reproducibility, which may depend on the nature of the contribution. For example +(a) If the contribution is primarily a new algorithm, the paper should make it clear how to reproduce that algorithm. +(b) If the contribution is primarily a new model architecture, the paper should describe the architecture clearly and fully. +(c) If the contribution is a new model (e.g., a large language model), then there should either be a way to access this model for reproducing the results or a way to reproduce the model (e.g., with an open-source dataset or instructions for how to construct the dataset). +(d) We recognize that reproducibility may be tricky in some cases, in which case authors are welcome to describe the particular way they provide for reproducibility. In the case of closed-source models, it may be that access to the model is limited in some way (e.g., to registered users), but it should be possible for other researchers to have some path to reproducing or verifying the results. + +# 5. Open access to data and code + +Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material? + +Answer: [NA] + +Justification: We provide sufficient instructions to faithfully reproduce the main experimental results. + +Guidelines: + +- The answer NA means that paper does not include experiments requiring code. +- Please see the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details. +- While we encourage the release of code and data, we understand that this might not be possible, so "No" is an acceptable answer. Papers cannot be rejected simply for not including code, unless this is central to the contribution (e.g., for a new open-source benchmark). +- The instructions should contain the exact command and environment needed to run to reproduce the results. See the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details. +- The authors should provide instructions on data access and preparation, including how to access the raw data, preprocessed data, intermediate data, and generated data, etc. +- The authors should provide scripts to reproduce all experimental results for the new proposed method and baselines. If only a subset of experiments are reproducible, they should state which ones are omitted from the script and why. +- At submission time, to preserve anonymity, the authors should release anonymized versions (if applicable). +- Providing as much information as possible in supplemental material (appended to the paper) is recommended, but including URLs to data and code is permitted. + +# 6. Experimental setting/details + +Question: Does the paper specify all the training and test details (e.g., data splits, hyperparameters, how they were chosen, type of optimizer, etc.) necessary to understand the results? + +Answer: [Yes] + +Justification: In Appendix. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The experimental setting should be presented in the core of the paper to a level of detail that is necessary to appreciate the results and make sense of them. +- The full details can be provided either with the code, in appendix, or as supplemental material. + +# 7. Experiment statistical significance + +Question: Does the paper report error bars suitably and correctly defined or other appropriate information about the statistical significance of the experiments? + +Answer: [Yes] + +Justification: In Appendix. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The authors should answer "Yes" if the results are accompanied by error bars, confidence intervals, or statistical significance tests, at least for the experiments that support the main claims of the paper. +- The factors of variability that the error bars are capturing should be clearly stated (for example, train/test split, initialization, random drawing of some parameter, or overall run with given experimental conditions). +- The method for calculating the error bars should be explained (closed form formula, call to a library function, bootstrap, etc.) +- The assumptions made should be given (e.g., Normally distributed errors). +- It should be clear whether the error bar is the standard deviation or the standard error of the mean. +- It is OK to report 1-sigma error bars, but one should state it. The authors should preferably report a 2-sigma error bar than state that they have a $96\%$ CI, if the hypothesis of Normality of errors is not verified. +- For asymmetric distributions, the authors should be careful not to show in tables or figures symmetric error bars that would yield results that are out of range (e.g. negative error rates). +- If error bars are reported in tables or plots, The authors should explain in the text how they were calculated and reference the corresponding figures or tables in the text. + +# 8. Experiments compute resources + +Question: For each experiment, does the paper provide sufficient information on the computer resources (type of compute workers, memory, time of execution) needed to reproduce the experiments? + +Answer: [Yes] + +Justification: In Appendix. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The paper should indicate the type of compute workers CPU or GPU, internal cluster, or cloud provider, including relevant memory and storage. +- The paper should provide the amount of compute required for each of the individual experimental runs as well as estimate the total compute. +- The paper should disclose whether the full research project required more compute than the experiments reported in the paper (e.g., preliminary or failed experiments that didn't make it into the paper). + +# 9. Code of ethics + +Question: Does the research conducted in the paper conform, in every respect, with the NeurIPS Code of Ethics https://neurips.cc/public/EthicsGuidelines? + +Answer: [Yes] + +Justification: We make sure that we preserve anonymity. + +# Guidelines: + +- The answer NA means that the authors have not reviewed the NeurIPS Code of Ethics. +- If the authors answer No, they should explain the special circumstances that require a deviation from the Code of Ethics. +- The authors should make sure to preserve anonymity (e.g., if there is a special consideration due to laws or regulations in their jurisdiction). + +# 10. Broader impacts + +Question: Does the paper discuss both potential positive societal impacts and negative societal impacts of the work performed? + +Answer: [Yes] + +Justification: In Appendix. + +# Guidelines: + +- The answer NA means that there is no societal impact of the work performed. +- If the authors answer NA or No, they should explain why their work has no societal impact or why the paper does not address societal impact. +- Examples of negative societal impacts include potential malicious or unintended uses (e.g., disinformation, generating fake profiles, surveillance), fairness considerations (e.g., deployment of technologies that could make decisions that unfairly impact specific groups), privacy considerations, and security considerations. +- The conference expects that many papers will be foundational research and not tied to particular applications, let alone deployments. However, if there is a direct path to any negative applications, the authors should point it out. For example, it is legitimate to point out that an improvement in the quality of generative models could be used to generate deepfakes for disinformation. On the other hand, it is not needed to point out that a generic algorithm for optimizing neural networks could enable people to train models that generate Deepfakes faster. +- The authors should consider possible harms that could arise when the technology is being used as intended and functioning correctly, harms that could arise when the technology is being used as intended but gives incorrect results, and harms following from (intentional or unintentional) misuse of the technology. +- If there are negative societal impacts, the authors could also discuss possible mitigation strategies (e.g., gated release of models, providing defenses in addition to attacks, mechanisms for monitoring misuse, mechanisms to monitor how a system learns from feedback over time, improving the efficiency and accessibility of ML). + +# 11. Safeguards + +Question: Does the paper describe safeguards that have been put in place for responsible release of data or models that have a high risk for misuse (e.g., pretrained language models, image generators, or scraped datasets)? + +Answer: [NA] + +Justification: The paper poses no such risks. + +# Guidelines: + +- The answer NA means that the paper poses no such risks. +- Released models that have a high risk for misuse or dual-use should be released with necessary safeguards to allow for controlled use of the model, for example by requiring that users adhere to usage guidelines or restrictions to access the model or implementing safety filters. +- Datasets that have been scraped from the Internet could pose safety risks. The authors should describe how they avoided releasing unsafe images. +- We recognize that providing effective safeguards is challenging, and many papers do not require this, but we encourage authors to take this into account and make a best faith effort. + +# 12. Licenses for existing assets + +Question: Are the creators or original owners of assets (e.g., code, data, models), used in the paper, properly credited and are the license and terms of use explicitly mentioned and properly respected? + +Answer: [Yes] + +Justification: We are confident that the creators or original owners of assets (e.g., code, data, models) used in the paper are properly credited, and the license and terms of use are explicitly mentioned and properly respected. + +Guidelines: + +- The answer NA means that the paper does not use existing assets. +- The authors should cite the original paper that produced the code package or dataset. +- The authors should state which version of the asset is used and, if possible, include a URL. +- The name of the license (e.g., CC-BY 4.0) should be included for each asset. +- For scraped data from a particular source (e.g., website), the copyright and terms of service of that source should be provided. +- If assets are released, the license, copyright information, and terms of use in the package should be provided. For popular datasets, paperswithcode.com/datasets has curated licenses for some datasets. Their licensing guide can help determine the license of a dataset. +- For existing datasets that are re-packaged, both the original license and the license of the derived asset (if it has changed) should be provided. +- If this information is not available online, the authors are encouraged to reach out to the asset's creators. + +# 13. New assets + +Question: Are new assets introduced in the paper well documented and is the documentation provided alongside the assets? + +Answer: [NA] + +Justification: The paper does not release new assets. + +Guidelines: + +- The answer NA means that the paper does not release new assets. +- Researchers should communicate the details of the dataset/code/model as part of their submissions via structured templates. This includes details about training, license, limitations, etc. +- The paper should discuss whether and how consent was obtained from people whose asset is used. +- At submission time, remember to anonymize your assets (if applicable). You can either create an anonymized URL or include an anonymized zip file. + +# 14. Crowdsourcing and research with human subjects + +Question: For crowdsourcing experiments and research with human subjects, does the paper include the full text of instructions given to participants and screenshots, if applicable, as well as details about compensation (if any)? + +Answer: [NA] + +Justification: This paper does not use crowdsourcing nor research with human subjects. + +Guidelines: + +- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. +- Including this information in the supplemental material is fine, but if the main contribution of the paper involves human subjects, then as much detail as possible should be included in the main paper. +- According to the NeurIPS Code of Ethics, workers involved in data collection, curation, or other labor should be paid at least the minimum wage in the country of the data collector. + +# 15. Institutional review board (IRB) approvals or equivalent for research with human subjects + +Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals (or an equivalent approval/review based on the requirements of your country or institution) were obtained? + +Answer: [NA] + +Justification: This paper does not involve crowdsourcing nor research with human subjects. + +Guidelines: + +- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. +- Depending on the country in which research is conducted, IRB approval (or equivalent) may be required for any human subjects research. If you obtained IRB approval, you should clearly state this in the paper. +- We recognize that the procedures for this may vary significantly between institutions and locations, and we expect authors to adhere to the NeurIPS Code of Ethics and the guidelines for their institution. +- For initial submissions, do not include any information that would break anonymity (if applicable), such as the institution conducting the review. + +# 16. Declaration of LLM usage + +Question: Does the paper describe the usage of LLMs if it is an important, original, or non-standard component of the core methods in this research? Note that if the LLM is used only for writing, editing, or formatting purposes and does not impact the core methodology, scientific rigorousness, or originality of the research, declaration is not required. + +Answer: [NA] + +Justification: In this work, the LLM is used only for writing, editing, or formatting purposes and does not impact the core methodology, scientific rigorousness, or originality of the research. + +Guidelines: + +- The answer NA means that the core method development in this research does not involve LLMs as any important, original, or non-standard components. +- Please refer to our LLM policy (https://neurips.cc/Conferences/2025/LLM) for what should or should not be described. + +In the appendix, we provide additional details, discussions, and experimental results to complement the main paper: + +- Appendix A: Additional related work and discussions on methods closely related to our study (cf. Section 2 of the main paper). +- Appendix B: Detailed descriptions of the comparison methods used in our evaluation (cf. Section 4 of the main paper). +- Appendix C: Additional qualitative and quantitative results for feature extraction using TabPFN v2, supplementing Section 6 of the main paper. +- Appendix D: Analysis of the impact of feature engineering and ensemble strategies in TabPFN v2, as well as a meta-feature-based analysis to identify the conditions under which TabPFN v2 performs well or poorly. +- Appendix E: Complete results corresponding to the tables and figures referenced in the main paper. +- Appendix F: Limitation and social impact of the paper. + +# A Additional Related Work + +Learning with Tabular Data. Tabular data is prevalent across diverse fields, including healthcare, finance, and education [43, 32, 58, 2]. Tree-based models, such as XGBoost [12], LightGBM [38], and CatBoost [56], have long dominated this domain. However, recent advances in deep neural networks (DNNs) have demonstrated strong potential for tabular data [7]. Popular architectures like multi-layer perceptrons [22, 37] and Transformers [31] have been adapted to tabular tasks, alongside custom architectures designed specifically for tabular data [41, 68]. + +Deep tabular methods can be broadly categorized into two types. The first type directly processes raw features [29, 21, 76], sometimes incorporating feature-specific encoding strategies [20]. The second type tokenizes features, transforming an example into a set of tokens [64, 31, 59]. Comprehensive benchmarks have been developed to evaluate these methods across diverse datasets [23, 49, 75, 60], highlighting the strengths and weaknesses of deep tabular models in various scenarios. + +Variants of TabPFN. TabPFN's success stems from its pre-training on massive synthetic datasets, enabling strong in-context learning performance on small-scale classification tasks [27]. Motivated by its capabilities, researchers have explored a variety of applications, including tabular data generation [47], anomaly detection [61], and time series forecasting [30]. [53] provided a bias-variance analysis of TabPFN, offering insight into its generalization behavior. Another line of research focuses on improving scalability by addressing TabPFN's sensitivity to context size [16, 73]. Further strategies to enhance downstream performance include context adaptation with nearest neighbor [65], partial fine-tuning [17, 45], pre-training on real-world datasets [48], scalable ensemble [46], and more powerful and efficient pre-training on synthetic data [57]. Most of these variants remain restricted to classification tasks due to limitations in TabPFN v1. + +The recently introduced TabPFN v2 [28] extends TabPFN to support regression tasks and accommodate larger context sizes. In this paper, we conduct a comprehensive evaluation of TabPFN v2, analyze its strengths, and introduce methods to overcome its scalability and applicability challenges. + +# B Evaluation Details + +Experimental Compute Resources. All experiments were conducted using 4 NVIDIA RTX 6000 Ada GPUs and 2 Intel(R) Xeon(R) Platinum 8352V CPUs. + +Please refer [75] for details of the 300 small to medium datasets. For high-dimensional datasets, we selected 18 datasets with more than 2000 features from the scikit-feature repository. Detailed statistics of high-dimensional datasets and large-scale datasets are reported in Table 3 and Table 4. + +We follow [75] and use different colors to represent various categories of methods in the result figures, ensuring clarity and easy comparison. In Figure 1 (right) and Figure 2 of the main paper, we compare the following methods: + +Table 3: Dataset Information for High-Dimensional Data Experiments: A collection of 18 datasets with varying numbers of instances, features, and classes used in our high-dimensional experiments. + +
Dataset#Instances#Features#ClassesDataset#Instances#Features#Classes
BASEHOCK199348622lung20333125
PCMAC194332892warpPIE10P210242010
RELATHE142743222orlraws10P1001030410
ALLAML7271292Prostate_GE10259662
CLL_SUB_111111113403SMK_CAN_187187199932
colon6220002warpAR10P130240010
GLI_8585222832arcene200100002
GLIOMA5044344gisette700050002
leukemia7270702TOX_17117157484
+ +Table 4: Dataset Information for Large-scale Data Experiments. + +
Dataset#Instances#Features#ClassesDataset#Instances#Features#Classes
BNG(credit-a)1,000,000152CDC_Indicators253,680212
Higgs1,000,000282Smoking_signal991,346232
nomao34,4651182sf-police-incidents2,215,02382
Data_Crowdfunding671,025114Fashion-MNIST70,00078410
covertype581,012547jannis83,733544
poker-hand1,025,0091010volkert58,31018010
Airlines_DepDelay10,000,0009-Wave_Energy_Farm36,04399-
UJIndoorLoc21,048520-blogfeedback60,021276-
microsoft1,200,192136-yahoo709,877699-
+ +- Classical Methods (□): The classical methods include Dummy, Logistic Regression (LR), K-Nearest Neighbors (KNN), Support Vector Machines (SVM), Naive Bayes, Linear Regression, and DNNR [52], which serve as basic baselines for classification and regression tasks. +- Tree-based Methods: Tree-based methods such as Random Forest [8], XGBoost [12], LightGBM [38], and CatBoost [56] are known for their high performance on tabular data. +- MLP variants: MLP variants, including vanilla MLP, MLP-PLR, Self-Normalizing Neural Networks [41], Residual Network [22], and TabM [19] enhance the flexibility and generalization of traditional MLP architectures through advanced regularization and residual connections. +- Special Architectures (■): Methods with specially designed architectures, such as DCNv2 [68], DANets [10], and TabCaps [9], focus on improving feature interaction and abstraction to capture complex relationships in tabular data. +- Token-based Methods: Token-based methods like AutoInt [64], TabTransformer [31], FTTransformer [22], and ExcelFormer [11] represent features as tokens, enabling models to capture higher-order interactions through attention mechanisms. +- Regularization-based Methods (□): Regularization-based methods, including TANGOS [35], SwitchTab [72], and PTaRL [78], aim to improve model generalization by incorporating regularization techniques during training to enhance the robustness of predictions. +- Tree-mimic Methods: Tree-mimic methods, such as NODE [55], GrowNet [5], and TabNet [4], combine the interpretability of decision trees with the power of deep learning, employing attention mechanisms to select important features. +- Context-based Methods (■): Context-based methods like TabR [21] and ModernNCA [76] leverage contextual information from the training data to improve predictions by utilizing neighborhood-based and in-context learning strategies. + +In addition to the aforementioned methods, for other experimental results, we will demonstrate the performance of TabPFN v2 and its variants, which are represented by emerald teal (■), ensuring that their experimental effects are clearly distinguished from the other methods. + +Remark. The standard checkpoint released by TabPFN employs a feature grouping size of 2, which complicates the analysis of individual feature embeddings and inter-feature relationships. To facilitate such analysis, we use a modified checkpoint with group size=1 for the experiments in Figure 3 of the main paper, which is available at HuggingFace. + +# C Additional Feature Extraction Results + +In this section, we provide further results on regression tasks and additional variants of our feature extraction strategy to validate the effectiveness and robustness of the proposed leave-one-fold-out method with TabPFN v2. We also compare this supervised approach to two unsupervised embedding extraction methods, as well as a non-partitioned variant that uses the same context and query without data folding. + +# C.1 Additional results for leave-one-fold-out feature extraction strategy with TabPFN v2 on regression tasks + +To further demonstrate the versatility of our approach, we evaluate TabPFN v2 on regression tasks by extracting embeddings using the leave-one-fold-out strategy and training simple linear regressors on top. As shown in Table 5, our embeddings consistently outperform both vanilla and raw features across different regressors, achieving the best average rank. This result indicates that TabPFN v2 can also serve as an effective feature extractor for regression problems, further supporting its general applicability beyond classification tasks. + +Table 5: Average rank comparison of embeddings on regression tasks using different regressors. Lower is better. LR denotes Linear Regression. + +
Extraction Strategy RegressorVanilla LRVanilla RidgeCVRaw LRRaw RidgeCVOurs LROurs RidgeCV
Avg. Rank5.585.083.833.172.081.25
+ +# C.2 Comparison Between Supervised and Unsupervised Feature Extraction + +Our leave-one-fold-out strategy explicitly incorporates label information and accounts for the distinct roles of training and test instances. To further understand the nature of the extracted embeddings, we compare this supervised strategy to two unsupervised alternatives provided by the TabPFN extensions repository. + +Let $X \in \mathbb{R}^{n \times d}$ denote a dataset with $n$ samples and $d$ features: + +- Unsupervised-Dummy: Each sample is paired with a constant pseudo-target $\mathbf{y} = \mathbf{0} \in \mathbb{R}^n$ , forming a regression task. The embedding $E_{\mathrm{D}} \in \mathbb{R}^{n \times h}$ is obtained by extracting the output tokens of the TabPFN regressor. +- Unsupervised-Permute: For each feature $j \in \{1, \dots, d\}$ , we treat $\mathbf{x}^{(j)} = X_{;j}$ as a pseudo-target and use the remaining features $X^{(-j)}$ as input. Depending on the type of $\mathbf{x}^{(j)}$ , TabPFN is applied in classification or regression mode to obtain $E^{(j)} \in \mathbb{R}^{n \times h}$ . These embeddings are concatenated to a high-dimensional form: + +$$ +E _ {\mathrm {P}} = \operatorname {c o n c a t} \left(E ^ {(1)}, E ^ {(2)}, \dots , E ^ {(d)}\right) \in \mathbb {R} ^ {n \times (d \cdot h)}. +$$ + +We compare these unsupervised methods with our supervised leave-one-fold-out strategy in Table 6. Overall, unsupervised approaches underperform compared to the supervised ones and fail to recover the classification ability of TabPFN v2. This performance gap arises because label information is introduced only post hoc via a linear classifier rather than during embedding extraction. Among the unsupervised methods, the permutation-based approach performs better, likely due to its ability to encode attribute-specific structure. + +Figure 6 presents a visual comparison of embeddings produced by the three methods using the same color and marker scheme as in Figure 4. Since unsupervised methods lack label supervision during embedding generation, their embeddings tend to scatter broadly without forming well-separated clusters by class. These results further highlight the contrasting goals of supervised and unsupervised strategies—class separation versus feature distribution coverage, respectively. + +# C.3 Results for the Non-Partitioned Feature Extraction Variant + +Using the same context and query without partitioning, we experimented with a variant where the context contains all training data and the query set includes both training and test points. This + +Table 6: Average rank (lower is better) of TabPFN v2 and a linear classifier trained on the extracted embeddings across 29 classification datasets. In addition to the supervised feature extraction strategy considered in the main paper (including the vanilla one, our leave-one-fold-out, and the version based on the combined features), we compare with two unsupervised embedding extraction approaches by appending a column of dummy labels with zero values and permuting each column as labels, respectively. + +
↓TabPFN v2VanillaDummyPermuteOursCombined
Rank2.665.724.903.692.161.88
+ +![](images/7d3a958aac02a9dce0e50cbb810e70337ff91cda6374f0ae26c6dbdb96bfd377.jpg) + +![](images/893c0b432e3ab0117a2ac378272341a782503eeff4c4dae1d6cf264d390ae2e6.jpg) + +![](images/82b2209fa4c95d5145951a26b078531993c3b6b77fcbdebc1225f25cbfc1e806.jpg) + +![](images/aecbefe7c869015896c20c0a1b03360e23d057fcda8582cf3896ce4e803edc40.jpg) + +![](images/304de8a162b8d591d6916961def7b7afd5b4fc7045ccdd37a66eb9cabde00688.jpg) + +![](images/bd333889679b16afe07120d220bcfe22ad7b98a2c8d9fd0faa6a2b95ee76b5da.jpg) + +![](images/8410e93818908d00278914698cf2c69c8938bcbc0d9a8ada87b9ccdfbd17e729.jpg) + +![](images/289b24348f6d3e96c47ca2890f276ad890dadfd4139bec83f65c297a413b0c70.jpg) + +![](images/b720b63821ee5236fcf2269141acae98d6f3aed80ff6e4d00e8b29bb9bf3b2e2.jpg) + +![](images/a9b7247c244a857992c6d191e7012b5d72fccc6e8f66e29e2b389e48d4b09e04.jpg) + +![](images/31dbcae1227e51e43088b1cb9ae75201c1008e3165a6e6f45c8017419d0d3fde.jpg) + +![](images/ce9748c31432d8009c39e2f1d4b3e75e60e76055b6a8d523338b58bd755c5ad4.jpg) + +![](images/e12bd6b260b15b5f12ddf5e59a6d89074201cd7d0d341fba98fcae8ac0b190d9.jpg) + +![](images/6f22d7c2439f0662d21f2db3964549d0967dee7b08ea90fa2c1e9118a434f4ff.jpg) + +![](images/59bcca6eb389287623c27eeca4bdf779820980cfc64e03a3773b98848cba55d6.jpg) + +![](images/c6b9df24d31a2af430c06a7d8c1b092237b18b2854ec8b26465e8367175b5112.jpg) +(a) Raw Feature + +![](images/2939560f376d79ce1fb0c4ad34a41c32c95739edaa89ac5eb99d8a6453ffbcf6.jpg) +(b) Vanilla + +![](images/90569e15166db83a965cd1db6fcf7035c4ca6a59b5bc7d694a7f4121c1c346fe.jpg) +(c) Dummy + +![](images/a349beaa48ea3df00ea6513a35987466fb8fd32880443292963c4aaaf091e56b.jpg) +(d) Permute +Figure 6: Comparison between unsupervised and supervised (ours) feature extraction. Visualization of extracted embeddings for four datasets: churn (first row, two classes), bank (second row, two classes), KDD (third row, two classes), and website_phishing (fourth row, three classes). We use crosses to denote training examples and circles to denote test examples. (a) shows the raw features, while (b) presents the embeddings extracted using the vanilla strategy. (c) and (d) refer to unsupervised embedding extraction approaches by appending a column of dummy labels with zero values and permuting each column as labels, respectively. (e) depicts the embeddings obtained using our proposed methods. The accuracy value is calculated by training a linear model (logistic regression) over the extracted embeddings on the training set and predicting on the test set. + +![](images/025ecda87d85b1844e8822764443063a5a0c0a645134cc821d82da20dcecccb9.jpg) +(e) Ours + +non-partitioned strategy improves upon the vanilla feature extraction baseline but still underperforms compared to our proposed leave-one-fold-out method. We attribute this to role ambiguity: query points that also appear in the support set (with dummy or ground-truth labels) are treated inconsistently, preventing the network from fully distinguishing between training and test roles and thereby degrading feature consistency. + +# D Influence of Key Modules and Meta-Feature Analysis + +We investigate the influence of two key components in TabPFN v2, i.e., the feature engineering that pre-processes the raw features of a given tabular dataset and the post-hoc ensembleing. In addition, we analyze the conditions under which TabPFN v2 performs well or poorly through a meta-feature-based classification analysis. + +Table 7: Performance ranking (lower is better) of different feature extraction strategies on classification tasks. The non-partitioned variant uses the same context and query without data folding. + +
MethodAvg. Rank
Leave-one-fold-out (Ours)2.10
TabPFN v22.72
Non-partitioned (12 layers)3.07
Non-partitioned (9 layers)4.29
Non-partitioned (6 layers)4.43
Vanilla (12 layers)5.86
Vanilla (9 layers)6.62
Vanilla (6 layers)6.90
+ +![](images/9217632020bfcc836be89c6c8ac151bf451735c21be6c5430dce98808f69087a.jpg) +Figure 7: Scatter plot comparing the normalized Accuracy/R² scores. The x-axis represents the normalized Accuracy/R² scores without Feature Engineering, while the y-axis represents the normalized Accuracy/R² scores with Feature Engineering. The red dashed line $(y = x)$ serves as a reference, indicating equal performance. + +Feature engineering. TabPFN v2 pre-processes the features of a given tabular dataset with various strategies, such as quantile, category shuffling, SVD, and power transform. Specifically, we examine the effects of adding fingerprint features (add_fingerprint_feature) and polynomial features (polynomial_features) to the raw tabular data. The results indicate that TabPFN v2 performs well even without the use of these engineered features, suggesting that, for the benchmark datasets of [75], these specific feature engineering techniques do not provide a significant improvement. This finding highlights the robustness of TabPFN v2 and its ability to handle raw features effectively, without the need for extensive pre-processing or feature construction. We show the influence of this step in Figure 7. + +Model ensemble. Post hoc ensembling (PHE) involves applying TabPFN v2 to the datasets multiple times with different perturbations and aggregating the predictions of these base models at different temperatures. We show the change of performance of TabPFN v2 w.r.t. the number of ensemble numbers (i.e., the number of base models) in Figure 8. On the benchmark of [75], we observe that, overall, ensemble methods improve performance, with larger ensemble sizes yielding better results. However, we also note that even without ensembling, TabPFN v2 performs exceptionally well, and the relative performance gain from ensembling is limited. This suggests that while ensembling can provide further improvements, the base TabPFN v2 model is already highly effective on its own. The equivariant property described in [3] provides insight into this phenomenon. Since TabPFN v2 introduces random tokens to handle heterogeneous features, the model becomes less sensitive to the arbitrary ordering of features, effectively enforcing equivariance in this aspect. As a result, the benefits of ensembling through feature order permutations are less pronounced compared to TabPFN v1. + +Meta-Feature Analysis of TabPFN v2 Performance. To better understand the conditions under which TabPFN v2 performs well or poorly, we conducted a meta-learning-based classification analysis using 300 datasets. Specifically, we used the average rank of TabPFN v2 across datasets as a performance indicator. The threshold for classification was set at the mean rank, 6.31. Datasets where + +![](images/66ad256af6bc8507835ca2f05760faf74dcdb570df91319d8ec4d22d391bddc4.jpg) +Figure 8: Box plot of relative performance improvements of TabPFN v2 with post hoc ensembling (PHE) across different ensemble sizes (2, 4, 8, and 16 base models). The relative improvement is calculated as the performance gain over the non-ensemble model, where higher values indicate stronger performance. The box plots show the median, interquartile range (IQR), and outliers for each ensemble size. + +TabPFN v2 achieved a rank lower than or equal to 6.31 were labeled as "Good", while those with a higher rank were labeled as "Bad". We extracted meta-features from each dataset and used them to train a decision tree classifier, aiming to distinguish between the "Good" and "Bad" categories. All the meta-features utilized in this task are detailed in Table 8, accompanied by their respective explanations. A visual depiction of a simple depth-3 decision tree is shown in Figure 9, and the tree reveals key factors that influence the effectiveness of TabPFN v2. The decision tree visualizes how to predict whether TabPFN v2 performs well ("Good") or poorly ("Bad") on a given dataset, based on dataset meta-features: The root node splits on the number of instances (nr_inst), indicating that TabPFN v2 tends to perform better on datasets with fewer than 24,350 samples. For these smaller datasets, the left subtree further splits on the mean joint entropy (joint_ent.mean), where higher values (greater than 3.028) are associated with improved performance. For datasets with lower mean joint entropy ( $\leq 3.028$ ), TabPFN v2 also tends to perform well when the number of rows is relatively small ( $\leq 4,862$ ). In contrast, the right subtree, which represents larger datasets, reveals that a low standard deviation of the interquartile range (iq_range.std) across features ( $\leq 0.657$ ) is linked to poorer model performance. + +# E Detailed Results + +We list the detailed results of TabPFN v2 and our extensions on various benchmarks. + +- We present the main results of TabPFN v2 on 300 datasets in Table 9. The table includes accuracy for classification tasks and RMSE (Root Mean Squared Error) for regression tasks, along with the corresponding mean and standard deviation for each dataset. Notably, we excluded 27 datasets from these results in Table 9, as they were used by TabPFN v2 to select the best checkpoint. These excluded datasets, which are not shown in Figure 1 (right) and Figure 2 of the main paper, include: + +(1) ada_prior, allbp, baseball, delta_ailerons, eye movements, eye movements_bin, GAMETES_Epistasis_2-Way_20atts_0.1H_EDM-1_1, hill-valley, JapaneseVowels, jungle_chess_2pcs_raw_endgame Complete, led24, longitudinal-survey, page-blocks, ringnorm, rl, thyroid-ann, waveform-5000, +(2) debutanizer, delta_elevators, mauna-loa-atmospheric, puma32H, stock_fardamento02, treasury, weather_izmir, wind. + +- In Table 10, we showcase the performance of various models on 18 high-dimensional datasets. The results display the mean accuracy of different models, including ModernNCA (MNCA), MLP, KNN, RealMLP, XGBoost (XGB), Random Forest (RForest), Logistic Regression (LogReg), and TabPFN v2 (PFN-v2), along with variants like TabPFN v2-pca and TabPFN v2*. This highlights the ability of these models to handle high-dimensional data with many features. + +- We demonstrate the performance of various models on 12 multi-class classification tasks with more than 10 classes in Table 11. The table provides the mean accuracy of models like KNN, TabPFN-v2*, XGBoost (XGB), CatBoost (CatB), Random Forest (RForest), ModernNCA (MNCA), MLP, Logistic Regression (LogReg), and RealMLP, showcasing how they perform on multi-class tasks with a larger number of classes. Additionally, we compare PFN-v2-ECOC, a + +![](images/5170962a0f14e5a605cca48caab55ebf3779a69e190c1b18367a2a54fc659ad6.jpg) +Figure 9: Decision tree for predicting TabPFN v2 performance (Good vs. Bad) based on dataset characteristics, constructed from experiments on 300 datasets. The tree splits on meta-features such as number of instances (nr_inst), joint entropy (joint_ent.mean), number of numerical features (nr_num), distance between minority and majority classes' center of mass (gravity) and interquartile range (iq_range) statistics. Each split is chosen to maximize information gain. The leaf nodes indicate the predicted performance class, the Gini impurity, and class distribution. This tree provides insights into the types of datasets where TabPFN v2 is expected to perform well. + +multi-class classification solution provided by [28]. This method extends TabPFN-v2 by leveraging Error-Correcting Output Codes (ECOC) to enhance multi-class classification performance.3 + +- In Table 12, we compare the performance of various models on 18 large-scale datasets. The results show the mean accuracy or RMSE for MLP, Logistic/Linear Regression (LR), KNN, XGBoost (XGB), Random Forest (RForest), CatBoost (CatB), ModernNCA (MNCA), RealMLP, and different versions of TabPFN v2 (PFNv2,PFNv2 with K-means,PFNv2 with Bagging, andPFNv2*). This illustrates the models' performance on large-scale datasets. +- In Table 13, we show the performance of TabPFN v2 and the extracted feature embeddings across 29 classification datasets. The table includes average classification accuracy for each dataset when using feature embeddings from different transformer layers (Layer 6, Layer 9, Layer 12), as well as a combined approach where embeddings from multiple layers are concatenated. The "selected layers" column indicates the layers chosen based on validation set performance, offering insights into how different layers contribute to overall model performance. In addition to evaluating the performance of TabPFN v2 and the extracted feature embeddings, we also compared the results with embeddings obtained using the vanilla strategy (Vanilla). + +Table 9: Main results of TabPFN v2 on 300 datasets, including accuracy (for classification tasks) and RMSE (for regression tasks), along with the corresponding mean and standard deviation for each dataset. Among the 300 datasets, 200 are classification datasets, and 100 are regression datasets. The results demonstrate the effectiveness of TabPFN v2 across both classification and regression tasks. + +
DatasetMean + StdDatasetMean + Std
ASP-POTASSCO-classification43.50 ± 1.27Amazon.employee_access94.22 ± 0.04
BLE_RSSI_localization73.37 ± 0.15BNG(breast-w)98.56 ± 0.07
+ +
BNG(cmc)57.69 ± 0.17
Bank_Customer_Churn_Dataset87.53 ± 0.12
California-Housing-Classification91.47 ± 0.17
Click_prediction_small83.29 ± 0.03
Contaminant-10.5GHz95.17 ± 0.32
Contaminant-9.0GHz93.01 ± 0.47
Credit_c69.98 ± 0.15
Diabetic_Retinopathy Debrecen72.81 ± 1.07
Employee84.80 ± 0.30
FOREX_audcad-day-High74.51 ± 0.51
FOREX_audchf-day-High76.66 ± 0.45
FOREX_audjpy-hour-High71.41 ± 0.32
FOREX_audusd-hour-High69.57 ± 0.48
FOREXCadjpy-hour-High70.55 ± 0.40
Fitness_Club_c79.67 ± 0.24
GAMETES_Heterogeneity65.90 ± 1.84
GesturePhaseSegmentationProcessed71.36 ± 1.15
Heart-Disease-Dataset91.23 ± 0.54
Indian_pines96.41 ± 0.23
Intersectional-Bias-Assessment94.73 ± 0.13
JapaneseVowels99.68 ± 0.08
KDDCup09_upselling81.06 ± 0.26
MIC90.20 ± 0.56
Marketing_Campaign88.11 ± 0.41
Nutrition_Health_Survey83.45 ± 0.22
PhishingWebsites96.74 ± 0.13
Pima_Indians_Diabetes_Database75.93 ± 0.66
Pumpkin_Seeds87.93 ± 0.21
Rain_in_Australia83.88 ± 0.11
Shipping68.73 ± 0.40
UJI_Pen_Chacters45.71 ± 2.16
Water_Quality_and_Potability65.49 ± 0.50
Wilt99.28 ± 0.06
accelerometer73.96 ± 1.32
ada_agnostic83.99 ± 0.34
adult85.93 ± 0.12
allbp97.85 ± 0.19
anlcatdata_authorship99.72 ± 0.29
autoUniv-au4-250069.81 ± 1.00
bank90.86 ± 0.19
baseball93.81 ± 0.40
churn96.33 ± 0.28
company_bankruptcy_prediction97.33 ± 0.07
connect-476.78 ± 0.35
credit78.10 ± 0.11
customer_satisfaction_in_airline94.79 ± 0.11
default_of Credits_card客户的82.63 ± 0.08
dis99.07 ± 0.14
drug Consumption40.32 ± 0.00
eeg-eye-state98.34 ± 0.12
estimation_of Obesity_levels98.66 ± 0.24
eye Movements77.03 ± 1.68
first-order-theorem-proving61.12 ± 0.70
golf Play_dataset_EXTENDED92.60 ± 0.44
heloc72.75 ± 0.20
house_16H88.55 ± 0.18
ibm-employee-performance100.0 ± 0.00
internet FIREwall92.85 ± 0.30
jasmine81.34 ± 0.42
jungle_chess_2pcs_raw_endgame85.97 ± 1.82
kdd_ipums_ta97-small88.50 ± 0.12
kr-vs-kp99.64 ± 0.15
law-school-admission-bianry100.0 ± 0.00
led773.99 ± 0.31
madeline90.72 ± 0.48
maternal_health_risk83.28 ± 0.64
mfeat-fourier89.85 ± 0.86
mfeat-morphological76.63 ± 0.50
mfeat-zernike84.10 ± 0.87
microaggregation262.80 ± 0.14
mozilla493.58 ± 0.16
national-longitudinal-survey-binary100.0 ± 0.00
one-hundred-plants-margin88.56 ± 0.74
one-hundred-plants-texture90.94 ± 0.75
optdigits98.59 ± 0.12
ozone_level97.86 ± 0.11
pc193.51 ± 0.46
pc490.87 ± 0.36
philippine84.20 ± 1.24
pol98.80 ± 0.09
rice_cammeo_and_osmancik92.74 ± 0.23
+ +
BNG(tic-tac-toe)79.42 ± 0.26
Basketball_c70.65 ± 0.47
Cardiovascular-Disease-dataset72.92 ± 0.13
Contaminant-10.0GHz94.42 ± 0.36
Contaminant-11.0GHz93.93 ± 0.50
Contaminant-9.5GHz93.21 ± 0.50
Customer_Personality_Analysis90.03 ± 0.21
E-CommerceShippingData67.54 ± 0.21
FICO-HELOC-cleaned75.35 ± 0.21
FOREX_audcad-hour-High71.01 ± 0.20
FOREX_audjpy-day-High78.00 ± 0.28
FOREX_audsgd-hour-High69.81 ± 0.39
FOREXCadjpy-day-High71.68 ± 0.53
Firm-Teacher-Direction84.42 ± 0.47
GAMETES_Epistasis68.75 ± 0.82
Gender_Gap_in_Spanish_WP60.58 ± 0.24
HR_Analtics80.02 ± 0.13
INNHotelsGroup87.98 ± 0.23
Insurance75.75 ± 0.00
Is-this-a-good-customer88.41 ± 0.00
KDD80.14 ± 0.46
Long99.88 ± 0.00
MagicTelescope88.13 ± 0.21
Mobile_Price-Classification97.10 ± 0.29
Performance-Prediction73.23 ± 0.61
PieChart387.31 ± 0.28
PizzaCutter388.20 ± 0.45
QSAR_biodegradation88.50 ± 0.50
SDSS1797.33 ± 0.06
Telecom_Churn_Dataset95.18 ± 0.50
VulNoneVul98.95 ± 0.00
Waterstress71.37 ± 0.96
abalone63.58 ± 0.38
ada85.40 ± 0.25
ada_prior85.32 ± 0.19
airlines_200062.28 ± 0.48
allrep98.65 ± 0.12
artificial-character73.90 ± 0.99
autoUniv-au7-110041.18 ± 1.58
banknote Authentication55.64 ± 0.18
car-evaluation98.29 ± 0.22
cmc59.59 ± 0.49
compass71.05 ± 0.29
contraceptive_method_choice62.10 ± 0.37
credit-g79.50 ± 0.81
diabetes_130-us_hospitals63.08 ± 0.07
delta_ailerons95.47 ± 0.09
dna97.25 ± 0.20
dryBean_dataset92.76 ± 0.10
electricity86.57 ± 0.45
eucalyptus72.88 ± 1.17
eye movements_bin67.28 ± 2.60
gas-drift99.47 ± 0.04
helena33.32 ± 0.21
hill-valley98.33 ± 0.52
htru97.95 ± 0.06
invehicle coupon73.20 ± 0.35
internet_usage54.34 ± 2.64
jm181.32 ± 0.10
kc186.65 ± 0.34
kr-vs-k78.46 ± 1.01
kropt77.96 ± 0.63
led2473.29 ± 0.62
letter97.57 ± 0.10
mammography98.71 ± 0.05
mfeat-factors96.98 ± 0.28
mfeat-karhunen96.42 ± 0.24
mfeat-pixel96.10 ± 0.32
mice_protein_expression100.0 ± 0.00
mobile_c36_oversampling98.11 ± 0.08
naticusdroid+android+permissions96.41 ± 0.10
okcupid_stem74.47 ± 0.12
one-hundred-plants-shape79.52 ± 0.72
online_shoppers90.65 ± 0.10
ozone-level-8hr94.92 ± 0.25
page-blocks97.67 ± 0.10
pc388.78 ± 0.27
pendigits99.56 ± 0.06
phoneme88.47 ± 0.35
predict_students_dropout78.11 ± 0.38
ringnorm98.00 ± 0.13
+ +
rl86.04 ± 0.44satimage92.30 ± 0.29
segment93.91 ± 0.19seismic+bumps93.40 ± 0.08
semeion92.41 ± 0.94shill-bidding90.31 ± 0.18
shrutime86.97 ± 0.11shuttle99.86 ± 0.04
spambase94.85 ± 0.19splice96.61 ± 0.22
sports_articles84.93 ± 0.40statlog72.13 ± 0.97
steel Plates_faults84.68 ± 0.55svmguide385.54 ± 0.54
sylvine97.30 ± 0.27taiwanese_bankruptcy_prediction97.20 ± 0.07
telco-customer-churn80.29 ± 0.28texture100.0 ± 0.00
W thyroid99.48 ± 0.06thyroid-ann99.34 ± 0.08
thyroid-dis68.75 ± 0.34turiye/student.evaluation51.74 ± 0.18
twonorm97.94 ± 0.08vehicle84.31 ± 1.29
walking-activity61.22 ± 0.22wall-robot-navigation99.44 ± 0.10
water_quality90.12 ± 0.12waveform-500086.29 ± 0.26
waveform_v186.59 ± 0.25website_phishing90.48 ± 0.48
wine75.12 ± 0.72wine-quality-red58.35 ± 0.76
wine-quality-white64.15 ± 0.69yeast60.18 ± 0.65
1000-C Cameras-Dataset607.71 ± 6.612dplanes1.01 ± 0.00
RSSI_Estimation0.00068 ± 0.00RSSI_Estimation10.00092 ± 0.00
Abalone_reg2.08 ± 0.00Ailerons0.00015 ± 0.00
Fiat716.20 ± 4.05BNG(emoMonths)11.41 ± 0.03
BNG(lowbwt)455.27 ± 0.78BNG(mv)4.63 ± 0.01
BNG(stock)2.95 ± 0.02Bias Correction_r0.60 ± 0.01
BiasCorrection_r_20.52 ± 0.01Brazilian_houses_reproduced0.01 ± 0.00
CPMP-2015-regression478.02 ± 5.40CPS1988364.02 ± 0.24
CookbookReviews1.52 ± 0.02Data_Science_Salaries60237.28 ± 102.97
Diamonds533.30 ± 6.30FacebookCOMMENT_Volume23.16 ± 0.20
Food_Delivery_Time7.55 ± 0.03Goodreads-Computer-Books0.43 ± 0.00
IEEE80211aa-GATS0.02 ± 0.00Job_Profitability13.14 ± 0.02
bike_sharing_demand68.41 ± 0.60Laptop_Prices_Dataset439.87 ± 3.10
Wave_Energy_Perth_10015507.90 ± 104.31Wave_Energy_Sydney_10014737.67 ± 150.43
Wave_Energy_Sydney_494567.97 ± 64.02MIP-2016-regression20966.10 ± 454.90
MiamiHousing201683101.09 ± 507.30Mobile_Phone_Market714.87 ± 11.15
Moneyball19.42 ± 0.08NASA_PHM200840.24 ± 0.06
NHANES_age_prediction15.47 ± 0.04OnlineNewsPopularity8606.54 ± 7.04
Parkinson_Sound Record14.58 ± 0.09Parkinsons_Telemonitoring0.60 ± 0.04
Physicochemical_r3.45 ± 0.04SAT11-HAND-routine1232.03 ± 58.01
Shop_Customer_Data28.56 ± 0.01Superconductivity10.17 ± 0.07
Wine_Quality_red0.65 ± 0.00Wine_Quality_white0.68 ± 0.00
airfoil_self-noise1.16 ± 0.02analcatdata supreme0.09 ± 0.00
archive2342.64 ± 3.20archive_r56_Portuguese2.86 ± 0.02
auctionVerification1145.54 ± 146.94avocado_sales0.09 ± 0.00
bank32nh0.08 ± 0.00bank8FM0.03 ± 0.00
boston4.25 ± 0.19chcasefoot0.95 ± 0.00
colleges0.14 ± 0.00combined_cycle_power_plant3.22 ± 0.05
communities_and_crime0.13 ± 0.00concrete_compressive_strength4.63 ± 0.07
cpu_ACT2.65 ± 0.03cpu_small3.06 ± 0.02
dataset_sales4.04 ± 0.02debutanizer0.04 ± 0.00
delta_elevators0.0014 ± 0.00elevators0.0019 ± 0.00
fifa0.78 ± 0.00fried1.01 ± 0.00
garments Worker_productivity0.13 ± 0.00gas_turbine_emission0.44 ± 0.00
healthcare_insurance_expenses4716.87 ± 36.52house_16H_reg29631.75 ± 251.56
house_8L28617.41 ± 202.41housePrices_nominal30676.02 ± 2455.48
house_sales_reduced132655.03 ± 1847.33houses42559.98 ± 928.78
housing_price_prediction1009361.62 ± 8758.05kin8nm0.08 ± 0.00
mauna-loa-atmospheric-co20.39 ± 0.01mv0.02 ± 0.00
pol_reg3.84 ± 0.10pole3.21 ± 0.14
puma32H0.01 ± 0.00puma8NH3.24 ± 0.00
qsr_aquatic_toxicity1.05 ± 0.01qsr_fish_toxicity0.86 ± 0.01
satellite_image0.65 ± 0.00sensory0.77 ± 0.01
socmob19.53 ± 0.64space_ga0.09 ± 0.00
steel_industry_energy0.37 ± 0.03stock0.65 ± 0.01
stock_fardamento0217.57 ± 0.08sulfur0.03 ± 0.00
topo_2_10.03 ± 0.00treasury0.23 ± 0.00
us_crime0.14 ± 0.00volume52.09 ± 0.34
weather_izmir1.09 ± 0.01wind2.83 ± 0.00
wine+quality0.72 ± 0.00yprop_4_10.03 ± 0.00
+ +# F Limitations + +While this paper does not introduce a new model architecture or training paradigm, it offers a timely and principled analysis of TabPFN v2, a powerful tabular foundation model. Our contributions lie in empirically evaluating its strengths, identifying its limitations, and proposing practical extensions that enhance its applicability—particularly to large-scale, high-dimensional, and multi-class settings. A key limitation of our work is that the proposed extensions are primarily post-hoc and do not fully + +Table 8: Meta-features used in the meta-feature analysis of TabPFN v2 performance. + +
Meta-FeatureExplanation
attr_concThe concentration coef. of each pair of distinct attributes.
class_concThe concentration coefficient between each attribute and class.
class_entThe target attribute Shannon's entropy.
inst_to_attThe ratio between the number of instances and attributes.
meanThe mean value of each attribute.
sdThe standard deviation of each attribute.
varThe variance of each attribute.
rangeThe range (max - min) of each attribute.
iq_rangeThe interquartile range (IQR) of each attribute.
nr_attThe total number of attributes.
sparsityThe (possibly normalized) sparsity metric for each attribute.
t_meanThe trimmed mean of each attribute.
nr_binThe number of binary attributes.
nr_catThe number of categorical attributes.
nr_numThe number of numeric features.
nr_normThe number of attributes normally distributed based in a given method.
nr-cor_attThe number of distinct highly correlated pair of attributes.
gravityThe distance between minority and majority classes' center of mass.
nr_classThe number of distinct classes.
joint_attThe joint entropy between each attribute and class.
att_enShannon's entropy for each predictive attribute.
covThe absolute value of the covariance of distinct dataset attribute pairs.
eigenvaluesThe eigenvalues of covariance matrix from dataset.
eq_num_attThe number of attributes equivalent for a predictive task.
maxThe maximum value from each attribute.
minThe minimum value from each attribute.
medianThe median value from each attribute.
freq_classThe relative frequency of each distinct class.
madThe Median Absolute Deviation (MAD) adjusted by a factor.
madThe Median Absolute Deviation (MAD) adjusted by a factor.
mut_infThe mutual information between each attribute and target.
nr_instThe number of instances (rows) in the dataset.
nr_outliersThe number of attributes with at least one outlier value.
ns_ratioThe noisiness of attributes.
imbalance_ratioThe ratio of the number of instances in the minority to the majority class.
attr_to_instThe ratio between the number of attributes.
+ +address the scalability constraints inherent to the original architecture. Nevertheless, by improving model usability without retraining, we reduce the computational and environmental costs typically associated with large model development. + +To the best of our knowledge, this work poses no explicit ethical concerns. It provides practical guidance for applying pre-trained tabular models in real-world domains such as healthcare and finance, where transparency and efficiency are essential. Our study highlights the value of understanding and extending foundation models alongside architectural innovation. + +Table 10: Performance of various models on 18 high-dimensional datasets. The results show the mean accuracy of different models, including ModernNCA (MNCA), MLP, KNN, RealMLP, XGBoost (XGB), Random Forest (RForest), Logistic Regression (LogReg), TabPFN v2 (PFN-v2), TabPFN v2 with PCA (v2-pca), TabPFN v2 with subsampling $(\mathrm{v}2^{*})$ , ProtoGate (ProtoG), and CatBoost (CatB). The performance is evaluated on high-dimensional datasets, with the values representing mean accuracy for each model. + +
DatasetMNCAMLPKNNRealMLPXGBRForestLogRegPFN-v2v2-ptav2*ProtoGCatB
CLL_SUB_11162.9072.4657.3970.4373.0470.1473.9170.1457.6871.5965.5171.59
BASEHOCK96.3197.0171.8897.4695.2996.7396.9969.0997.4197.3696.3295.87
Prostate_GE81.2786.9880.0087.9489.5287.9491.4395.2488.5794.2984.1394.92
PCMAC88.2188.5366.4890.1591.6492.2087.1592.7090.7690.1488.2192.01
GLI_8581.5785.4976.4789.8082.3583.9290.5980.3986.2792.5581.9680.78
RELATHE88.1890.5475.0390.2387.1187.3090.4986.3687.6589.9589.9290.35
SMK_CAN_18763.5166.8469.4769.8266.4970.7072.1171.0571.7572.1070.7171.40
warpPIE10P98.4199.0592.38100.094.9298.57100.0100.0100.0100.097.7998.89
leukemia90.2295.1186.6794.6797.7892.0096.0092.4493.3396.0094.0094.22
orlraws10P97.6798.3392.0099.0084.3399.0099.0092.0099.3399.6792.6799.00
GLIOMA58.0060.6768.0067.3366.6764.0064.0062.6769.3368.6769.9166.67
warpAR10P83.0885.6453.0897.4481.2887.1897.6990.7795.3896.6790.0487.44
TOX_17176.0088.1970.8690.4878.1078.6790.2980.9582.4887.2485.5283.05
lung91.5495.4593.6695.2893.6692.6895.1295.2893.5095.6195.4393.01
ALLAML87.5695.5681.3396.8996.0096.4492.0092.8993.7894.6791.1494.67
colon78.4678.9776.9283.0874.8782.5686.1581.5478.4679.4978.4677.95
gisette97.2197.5795.0497.8697.5596.8297.5197.3597.2697.2397.1897.78
arcene81.6785.5084.5081.0075.0086.8388.0083.6788.3392.0085.3385.00
Mean82.8687.1177.2988.8384.7686.8789.3685.2587.2989.7386.3787.48
+ +Table 11: Performance of various models on 12 multi-class classification tasks with more than 10 classes. The results show the mean accuracy of different models, including KNN, PFN-v2*, PFN-v2-ECOC, XGB (XGBoost), CatBoost (CatB), Random Forest (RForest), ModernNCA (MNCA), Multi-layer Perceptron (MLP), Logistic Regression (LogReg), and RealMLP. The performance is evaluated on 12 multi-class datasets with more than 10 classes, with accuracy values presented for each model on the respective datasets. + +
DatasetKNNPFN-v2*PFN-v2-DPTPFN-v2-ECOCXGBCatBRForestMNCAMLPLogRegRealMLP
100-plants-texture79.6990.9482.6784.9277.0689.7382.6580.5283.9286.8888.35
100-plants-margin77.5088.5681.9479.4074.2584.0682.7977.6080.4479.6983.58
100-plants-shape60.3179.5272.1563.3856.1565.1964.3370.1047.3365.9472.08
UJI_Pen_Characters36.2645.7133.3844.2030.3538.8834.2444.0337.7519.4146.37
texture98.45100.099.98100.098.5599.1396.7699.6899.4099.6499.95
letter94.9097.5796.6997.7896.2696.7591.5697.9696.4075.8098.31
walking-activity60.2961.2257.2861.9265.0664.9261.7464.8560.6427.0265.13
helena28.9433.3128.5419.2032.4237.9033.9136.5837.9133.4038.55
internetusage30.1754.3450.5150.8651.0837.9033.9152.0943.0037.7352.23
kropt71.2277.9671.4477.1186.9579.2671.7778.2764.4528.0892.03
kr-vs-k70.7878.4671.5476.2987.2674.8171.6076.8365.0328.0391.85
ASP-POTASSCO34.7543.5041.8845.2742.2441.0842.8637.4529.6335.1441.70
Mean61.9470.9365.6766.6966.4767.4764.0168.0062.1651.4072.51
+ +Table 12: Performance of various models on 18 large-scale datasets. The results show the mean accuracy/RMSE of different models, including MLP, Logistic Regression/Linear Regression (LR), KNN, XGBoost (XGB), Random Forest (RForest), CatBoost (CatB), ModernNCA (MNCA), RealMLP, and various versions of TabPFN v2: original TabPFN v2 (PFNv2), TabPFN v2 with K-means (PFNv2-K), TabPFN v2 with Bagging (PFNv2-B), PFNv2* (TabPFNv2*), PFNv2-DT (TabPFN-DT), and PFNv2-DF (TabPFN-DF). + +
DatasetMLPLRKNNXGBRForestCatBMNCARealMLPPFNv2PFNv2-KPFNv2-BPFNv2*PFNv2-DTPFNv2-DF
BNG(credit-a)90.0785.9887.4190.2189.2591.1389.9890.9189.5589.0189.6689.8990.4590.43
CDC_Indicators86.7986.5586.3986.7686.6086.7886.7686.7686.6586.6886.6986.7486.7586.70
Higgs75.5364.2965.1673.3371.8774.8173.2875.3671.6471.5672.0172.1373.5373.62
Smoking_signal73.9072.5372.3673.8773.0873.9973.6374.0073.4773.3773.5573.6973.7473.84
nomao96.1994.5995.2096.9296.0797.0396.6896.3796.0896.2996.1296.1896.7596.34
sf-police-incidents87.8487.8485.8787.6887.8487.87-87.8487.8487.8487.8487.8487.8487.84
Data_Crowdfunding96.4867.0493.7096.8995.2996.8196.5396.7194.5991.8194.9695.0796.9096.83
Fashion-MNIST89.5485.6986.0090.0386.5790.2489.3690.2568.4082.8283.8986.2678.9178.82
covertype94.0172.5492.7696.3078.3090.7797.3197.3883.5482.9584.1686.8597.3897.44
jannis71.9964.6065.6771.8369.1972.2672.5773.0070.2470.2670.5971.3172.5772.50
poker-hard99.9950.1254.0199.5164.6397.6976.3199.8841.9738.8636.8054.1291.1392.33
volkert69.8558.7567.4169.7462.7170.8877.1873.7662.8262.1562.8164.8468.6667.76
Airlines_DepDelay (×10^4)2.9052.9333.1702.8912.9072.881-2.4822.9372.9332.9372.9152.9002.897
Wave_Energy_Farm (×10^3)8.19913.1932.296.9177.2947.1736.14859.057.2148.3757.06310.5066.6166.785
UJIndoorLoc (×10^0)9.958∞9.00410.4723.199.1395.99065.3466.497.8257.4359.53814.4047.472
blogfeedback (×10^1)2.387∞2.4102.0932.0262.0441.9532.1053.0732.6872.7002.0141.9141.944
microsoft (×10^-1)7.5777.7828.2847.5147.5667.4537.5735.0777.7357.9817.7207.6127.9447.728
yahoo (×10^-1)7.6927.9978.5047.629-7.514-5.6718.1488.3328.1327.96116.4098.069
+ +Table 13: Performance of TabPFN v2 and the extracted feature embeddings across 29 classification datasets. The table shows the average classification accuracy for each dataset when using different layers (Layer 6, Layer 9, Layer 12) of the transformer as feature embeddings, as well as the "combined" approach, where embeddings from up to three selected layers are concatenated. The "selected layers" column indicates the specific layers chosen for each dataset based on validation set performance. "Vanilla" refers to the embeddings extracted using the vanilla strategy, which utilizes only the 12th layer of the transformer. "S" and "P" refer to unsupervised embedding extraction approaches by appending a column of dummy labels with zero values and permuting each column as labels, respectively, as described in Appendix C. + +
PFN-v2VanillaSPlayer-6layer-9layer-12combinedselected layers
FOREX_audchf-day-High77.3850.6856.9569.4868.3973.5774.1177.11(5, 9, 11)
taiwanese_bankruptcy_prediction96.9956.4596.7795.7597.1496.7797.0797.14(6)
rl85.5150.0060.5670.8266.9069.5286.7287.53(11, 12)
pc389.4610.2289.7886.9090.1088.8288.8288.82(8)
eyemoves_bin61.8350.0055.1257.9559.7259.4062.1662.16(6, 9, 12)
BNG(breast-w)98.4369.5197.6098.5198.3498.4698.6798.51(6, 9)
FOREX_cadjpy-hour-High69.5351.7966.5571.1262.1264.8770.6670.88(4, 5, 6)
dis99.3485.4398.4198.5498.4198.2899.3499.47(4, 5, 6)
sylvine97.4685.6672.7895.7192.4993.9597.2796.49(1, 11)
BNG(tic-tac-toe)78.0434.7171.4173.7973.9673.7178.7579.03(5, 10, 12)
online_shoppers90.5984.5185.9389.4690.0290.1190.6390.02(8)
Cardiovascular-Disease-dataset72.8450.8668.7372.6072.9673.0673.1473.09(5, 8, 12)
credit78.0462.3175.8678.3177.6277.8077.9577.59(4, 6, 9)
FOREX_audsgd-hour-High67.2651.4865.4970.1457.2461.0669.6270.41(7, 10, 12)
waveform-500086.0080.6055.7087.1085.6085.6086.4086.90(1, 6, 11)
jungle_chess85.6539.6064.1272.1478.5580.4486.6686.85(10, 11, 12)
BNG(cmc)57.4042.6252.4855.1656.1956.7257.7257.88(9, 10, 12)
page-blocks97.3594.2595.4396.3596.0796.7197.1797.35(6, 7, 12)
segment93.0772.2969.2687.2391.9988.1093.5192.64(1, 12)
website_phishing90.7736.9082.6690.0485.9887.0891.8891.88(7, 10)
baseball93.6678.7392.5492.1693.2894.0393.6695.15(10, 11)
pendigits99.5059.7572.4098.1892.8193.0499.4199.45(3, 4, 12)
Gender_Gap_in_Spanish_WP60.8433.6859.4760.8459.6860.3260.5360.84(2, 12)
wine-quality-white62.3510.5149.2955.1054.0855.3163.5764.39(8, 11, 12)
satimage91.2182.0484.9989.1988.7288.6591.9191.91(8, 11, 12)
mfeat-fourier90.0055.5046.7585.7577.7582.2589.5089.50(2, 7, 12)
VulNoneVul98.951.0598.9598.3398.9598.9598.9598.95(1)
law-school-admission-bianry100.099.8379.7698.82100.0100.0100.0100.0(6)
KDD80.3478.4562.3676.7679.3478.3581.2379.94(1, 8, 10)
\ No newline at end of file diff --git a/NeurIPS/2025/A Closer Look at TabPFN v2_ Understanding Its Strengths and Extending Its Capabilities/images.zip b/NeurIPS/2025/A Closer Look at TabPFN v2_ Understanding Its Strengths and Extending Its Capabilities/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..329d08c954d2748ac4e15990a2f6d9bed09834ed --- /dev/null +++ b/NeurIPS/2025/A Closer Look at TabPFN v2_ Understanding Its Strengths and Extending Its Capabilities/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6fbfad253ad988c5ecadeff016a7341665b290e4be1db76820c93387eaaa08d0 +size 2096371 diff --git a/NeurIPS/2025/A Closer Look at TabPFN v2_ Understanding Its Strengths and Extending Its Capabilities/layout.json b/NeurIPS/2025/A Closer Look at TabPFN v2_ Understanding Its Strengths and Extending Its Capabilities/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..dbb7506f5da7cfc0d9e5ead1d85f2dc251fde57d --- /dev/null +++ b/NeurIPS/2025/A Closer Look at TabPFN v2_ Understanding Its Strengths and Extending Its Capabilities/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bbd512042668cc237a3e87f70589507cf2f5a75a47baacb08e478a13e7f98e2c +size 1061594 diff --git a/NeurIPS/2025/A Closer Look to Positive-Unlabeled Learning from Fine-grained Perspectives_ An Empirical Study/5689b011-3c71-4887-8158-b3ac9142d101_content_list.json b/NeurIPS/2025/A Closer Look to Positive-Unlabeled Learning from Fine-grained Perspectives_ An Empirical Study/5689b011-3c71-4887-8158-b3ac9142d101_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..abb7b76c11209f3f2edf8496116fc61a7b4c21c4 --- /dev/null +++ b/NeurIPS/2025/A Closer Look to Positive-Unlabeled Learning from Fine-grained Perspectives_ An Empirical Study/5689b011-3c71-4887-8158-b3ac9142d101_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ba91cde453ffb8d0e0ceb48393f4a3f74e098b865916babb665e085487503fca +size 128127 diff --git a/NeurIPS/2025/A Closer Look to Positive-Unlabeled Learning from Fine-grained Perspectives_ An Empirical Study/5689b011-3c71-4887-8158-b3ac9142d101_model.json b/NeurIPS/2025/A Closer Look to Positive-Unlabeled Learning from Fine-grained Perspectives_ An Empirical Study/5689b011-3c71-4887-8158-b3ac9142d101_model.json new file mode 100644 index 0000000000000000000000000000000000000000..70455d098bbf5e87a8300aaf74e4287d6dac115a --- /dev/null +++ b/NeurIPS/2025/A Closer Look to Positive-Unlabeled Learning from Fine-grained Perspectives_ An Empirical Study/5689b011-3c71-4887-8158-b3ac9142d101_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0f1515b3713bfde27aef118034360a0f365332f9f4d3e2c92ee5afaa17d035e0 +size 165149 diff --git a/NeurIPS/2025/A Closer Look to Positive-Unlabeled Learning from Fine-grained Perspectives_ An Empirical Study/5689b011-3c71-4887-8158-b3ac9142d101_origin.pdf b/NeurIPS/2025/A Closer Look to Positive-Unlabeled Learning from Fine-grained Perspectives_ An Empirical Study/5689b011-3c71-4887-8158-b3ac9142d101_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..5b9dedbf2a2f29c2e33aa66677112fb3ed5c763a --- /dev/null +++ b/NeurIPS/2025/A Closer Look to Positive-Unlabeled Learning from Fine-grained Perspectives_ An Empirical Study/5689b011-3c71-4887-8158-b3ac9142d101_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a4a2e32c03feccf550fcef373fadedf3b934c3f32469edeb41162ad54ace46b0 +size 389938 diff --git a/NeurIPS/2025/A Closer Look to Positive-Unlabeled Learning from Fine-grained Perspectives_ An Empirical Study/full.md b/NeurIPS/2025/A Closer Look to Positive-Unlabeled Learning from Fine-grained Perspectives_ An Empirical Study/full.md new file mode 100644 index 0000000000000000000000000000000000000000..4a63ee410352cb2d82dfd76a5b9fc79edf79de7e --- /dev/null +++ b/NeurIPS/2025/A Closer Look to Positive-Unlabeled Learning from Fine-grained Perspectives_ An Empirical Study/full.md @@ -0,0 +1,600 @@ +# A Closer Look to Positive-Unlabeled Learning from Fine-grained Perspectives: An Empirical Study + +Yuanchao Dai $^{1,2}$ , Zhengzhang Hou $^{1,2}$ , Changchun Li $^{1,2}$ , Yuanbo Xu $^{1}$ , En Wang $^{1}$ , Ximing Li $^{1,2,3*}$ + +1College of Computer Science and Technology, Jilin University, China + +$^{2}$ Key Laboratory of Symbolic Computation and Knowledge Engineering, Jilin University, China + +$^{3}$ RIKEN Center for Advanced Intelligence Project + +{lixin86，yuanchaodai，changchunli93}@gmail.com + +# Abstract + +Positive-Unlabeled (PU) learning refers to a specific weakly-supervised learning paradigm that induces a binary classifier with a few positive labeled instances and massive unlabeled instances. To handle this task, the community has proposed dozens of PU learning methods with various techniques, demonstrating strong potential. In this paper, we conduct a comprehensive study to investigate the basic characteristics of current PU learning methods. We organize them into two fundamental families of PU learning, including disambiguation-free empirical risks, which approximate the expected risk of supervised learning, and pseudo-labeling methods, which estimate pseudo-labels for unlabeled instances. First, we make an empirical analysis on disambiguation-free empirical risks such as uPU, nnPU, and DistPU, and suggest a novel risk-consistent set-aware empirical risk from the perspective of aggregate supervision. Second, we make an empirical analysis of pseudo-labeling methods to evaluate the potential of pseudo-label estimation techniques and widely applied generic tricks in PU learning. Finally, based on those empirical findings, we propose a general framework of PU learning by integrating the set-aware empirical risk with pseudo-labeling. Compared with existing PU learning methods, the proposed framework can be a practical benchmark in PU learning. + +# 1 Introduction + +Positive-Unlabeled (PU) learning refers to a specific weakly-supervised learning paradigm [1, 2, 3] for binary classification, which trains a binary classifier with a few positive labeled instances and massive unlabeled instances [4]. It arises in various practical scenarios such as automatic face tagging, spam detection, and Inlier-based outlier detection [5]. Due to its wide applicability, PU learning has increasingly attracted more attention from the machine learning community. + +During the past decades, many emerging practical PU learning methods have been proposed with various advanced techniques [6, 7]. Because in PU learning, negative labeled instances are unavailable, how to deal with unlabeled instances becomes its key challenge; and from this taxonomic perspective, we organize the existing PU learning methods into two fundamental families, namely disambiguation-free empirical risks [5, 8, 9, 10] and pseudo-labeling methods [11, 12, 13, 14, 15, 16]. + +The disambiguation-free empirical risks, as the name suggests, directly apply only positive labeled instances and unlabeled instances to approximate the expected risk of supervised learning. Under certain data generation assumptions, previous studies suggest unbiased empirical risk uPU [17, 5] and several practical variants such as nnPU with non-negativity constraint [8], abs-PU with absolute + +value constraint [9], and DistPU with positive-class prior constraint [10]. In parallel, the basic idea of pseudo-labeling methods is estimating pseudo-labels for unlabeled instances, and training the binary classifier with them in a self-training manner. Analogous to semi-supervised learning, these methods typically estimate pseudo-labels by iteratively updating the current predictions. For example, RP [18] iteratively identifies reliable negative examples from unlabeled data and assigns hard pseudo-labels to them for subsequent training. Another recent Self-PU [11] utilizes a soft pseudo-labeling strategy that continuously refines label assignments by incorporating the evolving confidence scores throughout the training process. Additionally, they apply several generic tricks such as mixup augmentation, exponential moving average, and knowledge distillation to further improve the classification performance [7, 13, 19]. + +The current PU learning methods have demonstrated strong potential, but we find that most of them, especially pseudo-labeling ones, are commonly complicated by integrating with specific tricks. Accordingly, some of their basic characteristics are still unclear, such as what kind of pseudo-labeling techniques and generic tricks are practical. In this paper, we conduct a comprehensive study to investigate the basic characteristics from a fine-grained perspective of PU learning. First, we make an empirical analysis on disambiguation-free empirical risks; suggest a novel risk-consistent set-aware empirical risk from the perspective of aggregate supervision, and empirically validate that it can be a practical candidate for disambiguation-free empirical risks. Second, we turn to pseudo-labeling methods, and make an empirical analysis on basic techniques to estimate pseudo-label such as hard pseudo-labeling technique, soft pseudo-labeling technique, and high-confident pseudo-label selection strategies; additionally, we empirically analyze several widely applied generic tricks in PU learning. Finally, based on those empirical findings, we propose a general framework of PU learning by integrating the set-aware empirical risk with pseudo-labeling, namely GPU. We further suggest specification principles within GPU. Compared with existing PU learning methods, the proposed GPU framework can be a practical benchmark in PU learning. In summary, the contributions of this paper is outlined below: + +- We conduct a comprehensive empirical study to the current PU learning methods, and make extensive empirical observations on the effectiveness of basic techniques and tricks in PU Learning. +- We propose a novel risk-consistent set-aware empirical risk from the perspective of aggregate supervision, which can be a practical candidate for disambiguation-free empirical risks, and then formulate a novel general framework of PU learning by integrating it with pseudolabeling. +- We suggest implementation principles of GPU. Compared with existing PU learning methods, the proposed GPU framework can be a practical benchmark in PU learning. + +# 2 Preliminaries + +In this section, we review the problem setting of PU learning and the two main families of PU learning methods. + +Problem formulation and notations Formally, under the two-sample problem setting [20] and completely selected at random (SCAR) assumption [21], given a positive dataset $\mathcal{D}_p = \left\{(\mathbf{x}_i, + 1)\right\}_{i = 1}^{n_p}\stackrel {i.i.d.}{\sim}p_p(\mathbf{x}) = p(\mathbf{x}|y = +1)$ with $n_p$ instances drawn from the positive-class conditional density $p_{p}(\mathbf{x})$ and an unlabeled dataset $\mathcal{D}_u = \{(\mathbf{x}_i,y_i)\}_{i = 1}^{n_u}\stackrel {i.i.d.}{\sim}p(\mathbf{x})$ with $n_u$ instances drawn from the marginal density $p(\mathbf{x})$ , where $\mathbf{x}\in \mathbb{R}^d$ and $y\in \{-1, + 1\}$ denote the $d$ -dimensional feature vector and the corresponding binary label, respectively. The objective of PU learning is to induce a classifier $g:\mathbb{R}^d\to \mathbb{R}$ over $\mathcal{D}_p\cup \mathcal{D}_u$ , which can predict labels for unseen instances. + +# 2.1 PU Learning with Disambiguation-free Empirical Risks + +PU learning methods with disambiguation-free empirical risks directly utilize labeled positive data and unlabeled data to approximate the expected risk of supervised learning. In this work, we review several representative ones, including uPU [5], nnPU [8], absPU [9], and DistPU [10]. + +Table 1: A summary of basic techniques and widely applied generic tricks for PU learning methods with pseudo-labeling. + +
PUL methodspseudo-labeling techniquegeneric trick
pseudo-labelhigh confidence selectionmixupmoving average
hardsoftwithw/o
RP [18]✓✓
AdaSampling [22]✓✓
GenPU [23]✓
Self-PU [11]✓✓✓
VPU [12]✓✓✓
PULNS [24]✓✓
P³Mix [13]✓✓✓
RobustPU [14]✓✓
HolisticPU [15]✓✓
LaGAM [25]✓✓✓✓
PUL-CPBF [16]✓✓
VQ-Encoder [26]✓✓
+ +Formally, let $\ell : \mathcal{X} \times \mathcal{Y} \to \mathbb{R}_+$ be any loss function, and $\pi = p(y = +1)$ be the positive-class prior. Since negative samples are not directly accessible in PU learning, the SCAR assumption fortunately provides a solution. Under this assumption, where the labeled positive examples are selected completely at random from all positive examples, given $R_p^+(g) = \mathbb{E}_{p_p(\mathbf{x})}\left[\ell(g(\mathbf{x}), +1)\right]$ , $R_p^-(g) = \mathbb{E}_{p_p(\mathbf{x})}\left[\ell(g(\mathbf{x}), -1)\right]$ , $R_n^-(g) = \mathbb{E}_{p_n(\mathbf{x})}\left[\ell(g(\mathbf{x}), -1)\right]$ , and $R_u^-(g) = \mathbb{E}_{p(\mathbf{x})}\left[\ell(g(\mathbf{x}), -1)\right]$ , we obtain that $p(\mathbf{x}) = \pi p_p(\mathbf{x}) + (1 - \pi)p_n(\mathbf{x})$ , such that $(1 - \pi)R_n^-(g) = R_u^-(g) - \pi R_p^-(g)$ . Then, suppose $\pi$ is known, one can formulate [5], which is an unbiased risk of PU learning uPU for the expected risk of supervised learning, formulated as follows: + +$$ +R _ {\mathrm {u P U}} (g) = \pi R _ {p} ^ {+} (g) + R _ {u} ^ {-} (g) - \pi R _ {p} ^ {-} (g), \tag {1} +$$ + +and its empirical risk over $D_p \cup D_u$ is given below: + +$$ +\widehat {R} _ {\mathrm {u P U}} (g) = \pi \widehat {R} _ {p} ^ {+} (g) + \widehat {R} _ {u} ^ {-} (g) - \pi \widehat {R} _ {p} ^ {-} (g) \tag {2} +$$ + +However, uPU suffers from overfitting when using flexible models due to negative empirical risk $\widehat{R}_u^-(g) - \pi \widehat{R}_p^-(g)$ . Some methods attempt to impose a non-negative constraint on $\widehat{R}_u^-(g) - \pi \widehat{R}_p^-(g)$ to prevent the empirical risk from becoming negative. To achieve this, nnPU [8] incorporates the max function: + +$$ +\widehat {R} _ {\mathrm {n n P U}} (g) = \pi \widehat {R} _ {p} ^ {+} (g) + \max \{0, \widehat {R} _ {u} ^ {-} (g) - \pi \widehat {R} _ {p} ^ {-} (g) \}, \tag {3} +$$ + +and absPU incorporates the absolute value function [9]: + +$$ +\widehat {R} _ {\mathrm {a b s P U}} (g) = \pi \widehat {R} _ {p} ^ {+} (g) + \left| \widehat {R} _ {u} ^ {-} (g) - \pi \widehat {R} _ {p} ^ {-} (g) \right| \tag {4} +$$ + +Additionally, under the case of symmetric losses where $\ell(z) + \ell(-z) = 1$ , we have $\widehat{R}_p^- (g) = 1 - \widehat{R}_p^+ (g)$ naturally holds. Leveraging this property, Dist-PU [10] reformulates uPU as follows: + +$$ +\widehat {R} _ {\text {D i s t P U}} (g) = 2 \pi \widehat {R} _ {p} ^ {+} (g) + \left| \widehat {R} _ {u} ^ {-} (g) - \pi \right|, \tag {5} +$$ + +where the absolute value function is introduced to impose the non-negative constraint on $\widehat{R}_u^-(g) - \pi$ . + +# 2.2 PU Learning with Pseudo-labeling + +PU learning methods with pseudo-labeling, as the name suggests, estimate pseudo-labels for unlabeled data, and train the classifier $g$ in a self-training manner [18, 22, 23, 24, 13, 14, 15, 16]. Referring to semi-supervised learning, we can formulate the generic objective of pseudo-labeling below: + +$$ +\mathcal {L} (g; D _ {p}, D _ {u}) = \widehat {R} _ {p} ^ {+} (g) + \mathcal {L} _ {u} (g, \hat {y}; D _ {u}), \tag {6} +$$ + +where $\mathcal{L}_u(g,\hat{y};D_u)$ is the self-training objective with unlabeled data, and $\hat{y}$ denotes the pseudo-label. + +Table 2: Positive and negative label groups of datasets and the statistics of those PU learning sets. + +
DatasetπPositive ClassNegative ClassFeatureTrainBackbone
F-MNIST-10.40, 2, 4, 61, 3, 5, 7, 8, 928 × 2860,000LeNet-5
F-MNIST-20.61, 3, 5, 7, 8, 90, 2, 4, 628 × 2860,000LeNet-5
CIFAR-10-10.40, 1, 8, 92, 3, 4, 5, 6, 73 × 32 × 3250,0007-Layer CNN
CIFAR-10-20.62, 3, 4, 5, 6, 70, 1, 8, 93 × 32 × 3250,0007-Layer CNN
STL-10-1-0, 2, 3, 8, 91, 4, 5, 6, 73 × 96 × 96105,0007-Layer CNN
STL-10-2-1, 4, 5, 6, 70, 2, 3, 8, 93 × 96 × 96105,0007-Layer CNN
+ +We review existing pseudo-labeling methods and summarize the basic techniques and widely applied generic tricks in Table 1. Specifically, the basic problem of pseudo-labeling is the techniques to estimate pseudo-labels $\hat{y}$ for unlabeled data with current predictions, and they typically include hard and soft and pseudo-labeling techniques. Let $q = g(\mathbf{x})$ and $\phi (q)\in [0,1]$ denote the prediction of the classifier and the confidence belonging to the positive class, respectively, where $\phi$ is a transformation function, and here we apply the sigmoid function. Then, the hard and soft pseudo-labels are estimated as $\hat{y} = \mathrm{sign}\left(\phi (q) - 0.5\right)\in \{-1, + 1\}$ and $\hat{y} = \phi (q)$ , respectively. In addition, some studies suggest selecting high-confident pseudo-labels, rather than applying all of them [18, 22, 11, 24, 14], where the representatives include various thresholding strategies. + +Generic tricks We briefly review two widely applied generic tricks in PU learning studies [11, 12, 13, 25], such as mixup and moving average. Mixup is an efficient data augmentation trick with the convex combination of instance pairs [7]. Given an instance pair $(\mathbf{x}_i, y_i)$ and $(\mathbf{x}_j, y_j)$ , it generates an augmented instance $(\tilde{\mathbf{x}}, \tilde{y})$ as follows: + +$$ +\tilde {\mathbf {x}} = \lambda \mathbf {x} _ {i} + (1 - \lambda) \mathbf {x} _ {j}, \quad \tilde {y} = \lambda y _ {i} + (1 - \lambda) y _ {j}, \quad \lambda \sim \operatorname {B e t a} (\alpha , \alpha) \tag {7} +$$ + +Here, the moving average refers to updating pseudo-labels with historical predictions during the classifier training process [27]. Formally, its update equation for pseudo-labels is given below: + +$$ +\phi (q) \leftarrow \epsilon \phi \left(q ^ {h}\right) + (1 - \epsilon) \phi (q), \tag {8} +$$ + +where $q^h$ denotes the historical prediction, and $\epsilon$ is a smoothing parameter. + +# 3 Empirical Findings, Analysis, and Modifications + +# 3.1 Settings of Empirical Study + +We conduct empirical evaluations on 3 standard benchmark datasets, i.e. Fashion-MNIST (F-MNIST), CIFAR-10, and STL-10. Following [16], we transform them into a set of binary classification problems by partitioning their original 10 classes into positive and negative categories by varying the class prior $\pi \in \{0.4, 0.6\}$ . For all datasets, the number of positive labeled instances is fixed as $n_p = 1,000$ . The details of datasets are summarized in Table 2. + +For each PU learning method, We employ dataset-appropriate backbones as follows: LeNet-5 for F-MNIST, 7-layer CNN for CIFAR-10 and STL-10; the MLP layer is used as the classification layer across all datasets. The mini-batch is fixed as 512 and the number of epochs is set to 100 for F-MNIST and 200 for others. + +In addition, we employ the classification accuracy (ACC) as the evaluation metric. All experiments are conducted with five different random seeds on a server equipped with two Nvidia RTX4090 GPUs, and we report the mean and standard deviation of the results. + +# 3.2 Disambiguation-free Empirical Risks + +In this section, we suggest a novel set-aware empirical risk of PU learning and empirically evaluate it and existing disambiguation-free empirical risks with various surrogate loss functions. + +Table 3: Properties of commonly used loss functions. + +
LossFormulaConvexDiffer.Symm.Lipsc.
hingemax{0,1-z}✓✓
logisticlog(1+e-z)✓✓✓
sigmoid1/(1+ez)✓✓✓
squared(1-z)2✓✓
rampmin{1, max{0, (1-z)/2}}✓✓
double-hingemax{0, (1-z)/2,-z}✓✓
+ +Set-aware empirical risk of PU learning In PU learning, we are given the positive labeled data and unlabeled data $\mathcal{D}_p\cup \mathcal{D}_u$ , and the positive-class prior $\pi$ . Inspired by previous weakly-supervised learning studies with aggregate supervision [28], we can arrange the training data as $\mathcal{D}_p\cup (\mathcal{D}_u,\pi)$ where we treat $(\mathcal{D}_u,\pi)$ as a set of instances with its approximate label proportion. Accordingly, we can formulate the following set-aware empirical risk of PU learning (SAPU): + +$$ +\widehat {R} _ {\mathrm {S A P U}} (g) = \widehat {R} _ {p} ^ {+} (g) + \ell_ {C E} \left(\frac {1}{n _ {u}} \sum_ {\mathbf {x} _ {i} \in \mathcal {D} _ {u}} g (\mathbf {x} _ {i}), \pi\right), \tag {9} +$$ + +where $\ell_{CE}$ denotes the cross-entropy loss. Because the size of $\mathcal{D}_u$ can be too large, directly fitting $\pi$ in the second term of $\widehat{R}_{\mathrm{sAPU}}(g)$ may result in smoothing instance-level predictions. To alleviate this potential issue, we can randomly divide $\mathcal{D}_u$ into many subsets $\{S_i\}_{i=1}^{n_s}$ , where $S_i = \{\mathbf{x}_{ij}\}_{j=1}^S$ , $n_s$ is the number of subsets, and $S$ is the number of instances in each subset; and if $S$ is large enough, we can also approximate the label proportion of each subset as $\pi$ . Upon these ideas, we can rearrange the training data as $\mathcal{D}_p \cup \{(\mathcal{S}_i, \pi)\}_{i=1}^{n_s}$ , and then reformulate Eq.9 as follows: + +$$ +\widehat {R} _ {\mathrm {S A P U}} (g) = \widehat {R} _ {p} ^ {+} (g) + \frac {1}{n _ {s}} \sum_ {i = 1} ^ {n _ {s}} \ell_ {C E} \left(\frac {1}{S} \sum_ {\mathbf {x} _ {i j} \in \mathcal {S} _ {i}} g (\mathbf {x} _ {i j}), \pi\right) \tag {10} +$$ + +We consider SAPU as a practical candidate disambiguation-free empirical risks of PU learning. We show the following theorem to indicate that it is risk-consistent for the expected risk of supervised learning. The proof is presented in the Appendix. + +Lemma 3.1. Let $\hat{\pi}_i = \frac{1}{S}\sum_{i=1}^{S}\mathbf{1}[y_{ij} = +1]$ be the true proportion of positive instance in set $S_i$ . When the set size satisfies $S \geq \frac{3\pi(1 - \pi)\log(2 / \delta)}{2\epsilon^2}$ , with probability at least $1 - \delta$ , we have $|\hat{\pi}_j - \pi| \leq \epsilon$ for each set $S_i$ and $Var(\hat{\pi}_j) \leq \frac{\pi(1 - \pi)}{S}$ . + +Theorem 3.2. Let $g^{*} = \arg \min_{g\in \mathcal{G}}R(g)$ is the minimizer of the true classification risk and $\hat{g}_{\mathrm{SAPU}} = \arg \min_{g\in \mathcal{G}}\hat{R}_{\mathrm{SAPU}}(g)$ denotes the minimizer of the risk form in Eq.10. Suppose that the pseudodimensions of $\{\mathbf{x}\mapsto g(\mathbf{x})|g\in \mathcal{G}\}$ and $\{\mathbf{x}\mapsto \ell_{CE}(g(\mathbf{x}),\pi)|g\in \mathcal{G}\}$ are finite, and there exist constants $L_{g},L_{\ell}$ such that $|g(\mathbf{x})|\leq L_g$ and $|\ell_{CE}(g(\mathbf{x}),\pi)|\leq L_{\ell}$ for all $\mathbf{x}\in \mathcal{X}$ and all $g\in \mathcal{G}$ . Then, for any $\delta >0$ , with probability at least $1 - \delta$ : + +$$ +R \left(\hat {g} _ {\mathrm {s A P U}}\right) - R \left(g ^ {*}\right) \leq O \left(\sqrt {\frac {\log (1 / \delta)}{n _ {p}}}\right) + O \left(\sqrt {\frac {\log (1 / \delta)}{n _ {s}}}\right) + L _ {\ell} \cdot O \left(\sqrt {\frac {\pi (1 - \pi) \log (1 / \delta)}{S}}\right) \tag {11} +$$ + +Results and analysis We empirically investigate the proposed SAPU and 4 existing disambiguation-free empirical risks with different commonly used loss functions. Table 3 presents 6 loss functions, i.e., hinge, logistic, sigmoid, squared, ramp, and double-hinge, along with their mathematical formulations and theoretical properties. These loss functions are selected to represent diverse characteristics across 4 key properties: convexity, differentiability, symmetry, and Lipschitz continuity. Because the positive prior $\pi$ for the STL-10 dataset is unavailable, we conduct experiments on the CIFAR-10 and F-MNIST datasets. + +Table 4: The ACC scores (mean±std) of disambiguation-free empirical risks with widely used loss functions on F-MNIST and CIFAR-10. The highest scores are indicated in bold. + +
DatasetMethodShingelogisticsigmoidsquaredrampdouble-hinge
F-MNIST-1uPU-68.0±0.568.5±0.569.8±0.977.1±2.270.8±1.868.4±0.6
nnPU-93.8±0.493.0±0.593.9±0.793.2±1.793.8±1.094.8±0.3
absPU-94.2±0.493.3±0.593.3±0.793.7±0.493.6±0.694.1±0.2
Dist-PU---94.3±0.4-94.0±0.294.7±0.2
SAPU3293.9±1.095.9±0.295.9±0.294.6±0.694.5±1.094.5±0.9
6492.8±0.396.0±0.296.0±0.193.0±0.093.8±0.894.0±0.9
12892.8±0.196.1±0.096.2±0.191.0±0.593.5±0.892.9±0.1
25693.0±0.596.0±0.296.2±0.090.8±0.192.9±0.592.8±0.3
nu92.4±0.595.4±0.696.0±0.090.8±0.392.9±0.692.6±0.1
F-MNIST-2uPU-47.8±0.647.7±0.349.1±0.962.4±2.750.8±1.448.7±0.7
nnPU-92.4±0.491.7±1.391.0±0.492.7±0.593.1±0.793.4±0.3
absPU-92.6±0.691.5±0.791.0±1.292.0±0.392.4±0.793.4±0.4
Dist-PU---91.3±0.9-93.3±0.692.2±0.3
SAPU3294.5±0.595.7±0.095.8±0.294.7±0.595.3±0.094.9±0.3
6494.6±0.195.8±0.095.8±0.192.8±0.594.0±0.494.5±0.5
12893.6±0.495.9±0.296.0±0.190.8±0.293.2±0.493.9±0.3
25693.7±0.295.8±0.296.1±0.088.2±1.893.4±0.593.2±0.0
nu92.8±0.295.9±0.096.0±0.188.3±1.493.6±0.493.5±0.4
CIFAR-10-1uPU-80.5±0.781.7±0.981.6±1.966.1±2.477.3±2.379.9±0.7
nnPU-86.4±0.484.3±0.785.1±1.483.2±1.086.1±0.586.4±0.1
absPU-85.6±0.482.9±0.785.7±1.581.9±1.286.3±0.985.7±0.6
Dist-PU---86.0±0.9-86.2±0.686.6±0.5
SAPU3285.3±0.686.5±0.586.6±0.276.6±4.786.2±0.784.5±0.6
6483.8±0.386.4±0.286.7±0.377.0±3.486.7±0.785.3±0.4
12884.7±0.585.6±0.487.0±0.579.5±1.585.0±0.083.6±0.2
25683.5±0.286.6±0.386.8±0.275.3±5.585.4±0.684.4±0.6
nu84.1±0.285.3±0.486.8±0.378.7±2.585.5±1.284.1±0.7
CIFAR-10-2uPU-76.1±0.977.3±1.176.9±2.455.7±2.067.9±2.475.5±1.0
nnPU-84.7±1.080.7±1.483.7±1.381.0±1.784.3±1.083.8±1.4
absPU-84.4±0.978.0±1.783.8±1.479.3±2.784.4±0.782.4±1.4
Dist-PU---82.1±1.1-83.4±1.685.6±0.7
SAPU3260.7±0.085.1±0.785.1±0.781.3±0.374.7±6.374.7±6.3
6462.4±1.785.2±0.784.9±0.781.3±0.174.9±6.175.7±5.3
12861.6±0.985.2±0.784.7±0.681.1±0.072.3±9.172.6±8.7
25660.7±0.085.4±0.684.5±0.581.3±0.073.8±7.474.0±7.1
nu61.6±0.985.3±0.784.7±0.781.2±0.174.5±6.974.8±7.2
+ +The experimental results in Table 4 demonstrate that the choice of loss function significantly influences classification accuracy across different datasets and methods. For F-MNIST dataset, the sigmoid loss consistently delivers superior performance, achieving a remarkable accuracy of $96\%$ with our method. The double-hinge loss also performs exceptionally well, particularly with nnPU method and Dist-PU. On the more challenging CIFAR-10 dataset, the sigmoid loss still demonstrates robust performance (around $86\%$ with SAPU), while the double-hinge loss excels in several configurations, notably with Dist-PU on CIFAR-10-2 ( $85.6\%$ ). Interestingly, the effectiveness of each loss function varies substantially across different empirical risk methods and datasets. For example, while SAPU achieves optimal results with sigmoid on CIFAR-10-1 ( $86.8\%$ ), its performance degrades considerably with the squared loss. Additionally, our empirical analysis confirms that smooth, differentiable losses (sigmoid, logistic) achieve better compatibility with SAPU's set-aware architecture, while non-smooth losses (double-hinge, ramp) align better with traditional point-wise optimization methods. This non-uniform behavior suggests a complex interaction between loss functions and model architectures + +that cannot be reduced to simple heuristics, underscoring the importance of careful loss function selection based on specific application contexts. + +Based on the above analysis, we can summarize the following guiding principles: (1) The smooth, differentiable losses (such as sigmoid, logistic loss) achieve better compatibility with the set-aware architecture of SAPU, while non-smooth losses (e.g., double-hinge, ramp loss) align better with traditional point-wise optimization methods. (2) Convex losses generally provide better optimization guarantees. (3) Simple datasets (e.g., F-MNIST) benefit from smooth losses, enabling fine-grained optimization, while complex datasets (e.g., CIFAR-10) may require losses with stronger regularization properties. + +Furthermore, as the core of SAPU lies in dividing unlabeled data into multiple subsets for set-aware supervision, we further conduct experiments to verify how subset size affects the performance of the model. We systematically tested different subset sizes $S = \{32, 64, 128, 256, n_u\}$ across all datasets and recorded classification accuracy with various loss functions. The experimental results in Table 4 demonstrate that $S = 256$ yields optimal performance on most datasets, particularly when combined with the sigmoid loss function. For example, the model achieved peak accuracies of $96.2\%$ and $96.1\%$ respectively on F-MNIST-1 and F-MNIST-2 when $S = 256$ with sigmoid loss function. For simpler datasets like F-MNIST, this phenomenon can be explained that when subsets are too small, individual subsets struggle to accurately reflect the overall label distribution; conversely, when subsets become excessively large (approaching $n_u$ ), instance-level predictions become overly smoothed, reducing the model's discriminative power. Moreover, for more complex datasets like CIFAR-10, our experiments indicate that larger subset sizes tend to be more effective. Based on our comprehensive analysis across different datasets, we recommend setting medium-sized subsets (e.g., $S = 256$ ) as a generally effective configuration for our SAPU method. + +# 3.3 Pseudo-labeling Methods + +In this section, we investigate the pseudo-labeling techniques and thresholding techniques for selecting high-confident pseudo-labels. For comprehensive evaluations, we first suggest several base methods and then discuss the empirical results. + +Base methods of pseudo-labeling We specify the generic objective of Eq.6 with specific pseudo-labeling techniques and thresholding strategies, leading to a set of base methods. First, we estimate pseudo-labels $\hat{y}$ by hard and soft pseudo-labeling techniques; and for clarity, we review them as $\hat{y} = \mathrm{sign}(\phi (q) - 0.5)\in \{-1, + 1\}$ and $\hat{y} = \phi (q)$ , respectively. Second, the thresholding strategy refers to computing a threshold value $\tau$ to define the lower bound of high-confident pseudo-labels. Inspired by [29, 30], we specify 3 thresholding strategies to compute $\tau$ , described below: + +- Fixed thresholding treats the threshold value $\tau$ as a hyper-parameter, and empirically sets it as a constant value. Here, we fix $\tau$ to 0.95. +- Adaptive thresholding gradually updates the threshold value $\tau$ during classifier training. Following the idea that the predictions can be more accurate as the classifier continues to be trained [31], we gradually increase $\tau$ as follows: + +$$ +\tau \leftarrow \tau_ {m a x} \times \min (1, t / T), +$$ + +where $t$ is the current epoch, $\tau_{max}$ is the maximum threshold value, and $T$ is the ramp-up period. + +- Class-specific adaptive thresholding gradually updates the threshold values $\tau_{p}$ and $\tau_{n}$ for positive- and negative-classes, respectively. Following [30], we gradually update $\tau_{p}$ and $\tau_{n}$ as follows: + +$$ +\tau_ {p} \leftarrow \tau_ {p} \times \mathcal {C} _ {p} ^ {(t)}, \qquad \tau_ {n} \leftarrow \tau_ {n} \times \mathcal {C} _ {n} ^ {(t)}, +$$ + +where $\mathcal{C}_p^{(t)}$ and $\mathcal{C}_n^{(t)}$ are the ratios between the pseudo-label accuracies of positive- and negative-classes and their higher accuracy at epoch $t$ . + +Based on these specific techniques, we can specify 6 base methods of pseudo-labeling. + +Results and analysis To comprehensively evaluate the effectiveness of different pseudo-label strategies and generic tricks under PU learning, we compare six base pseudo-labeling methods + +Table 5: The ACC scores (mean±std) of 6 base methods of pseudo-labeling on F-MNIST and CIFAR-10. The highest scores are indicated in bold. + +
LabelThresholdM.A.MixupF-MNIST-1F-MNIST-2CIFAR-10-1CIFAR-10-2
HardFixed90.0±1.792.7±0.185.1±0.683.1±3.8
✓89.9±1.689.3±2.384.2±1.882.0±3.5
✓90.5±0.892.8±0.285.1±0.385.0±1.0
✓✓89.6±1.989.0±1.784.0±2.182.1±3.7
Adaptive91.4±0.792.8±0.284.3±0.583.7±0.8
✓71.7±4.265.5±5.880.5±0.282.8±3.0
✓90.6±0.692.9±0.184.5±0.283.8±0.2
✓✓72.1±4.367.5±3.280.7±0.083.1±3.3
Class Adaptive91.5±0.489.0±0.084.3±0.483.5±0.5
✓71.7±4.265.5±5.882.0±1.082.0±1.5
✓91.0±1.089.6±0.284.5±0.584.0±0.5
✓✓72.1±4.365.5±5.780.5±1.082.5±1.5
SoftFixed95.4±0.493.7±0.484.3±0.583.8±0.2
✓95.2±0.371.1±1.783.5±1.783.0±1.5
✓95.4±0.493.9±0.484.3±0.583.8±0.2
✓✓95.2±0.370.7±2.283.8±1.583.3±1.0
Adaptive95.4±0.494.1±0.385.9±0.582.9±3.4
✓74.0±4.069.9±3.083.5±1.583.0±1.3
✓95.5±0.494.2±0.485.9±0.584.8±0.0
✓✓82.9±2.470.7±2.283.9±1.383.4±1.1
Class Adaptive95.0±0.194.1±0.484.3±0.583.8±0.2
✓77.2±0.193.7±1.580.8±1.583.0±1.5
✓95.5±0.494.6±0.485.6±0.585.8±0.2
✓✓82.9±2.493.8±1.181.0±1.383.4±1.1
+ +(including two pseudo-labeling techniques (hard vs. soft labeling) and three thresholding strategies (fixed, adaptive, and class-specific adaptive)) and two widely used enhancement techniques (mixup and moving average) on CIFAR-10 and F-MNIST datasets. + +The results demonstrate that soft labeling consistently outperforms hard labeling, particularly when combined with class-adaptive thresholding on F-MNIST datasets (achieving up to $95.5\%$ ). Mixup proves to be the most consistent generic trick for ACC improvement across all experimental configurations, while moving average often leads to performance degradation when combined with other techniques. Mixup proves consistently beneficial because it addresses the fundamental challenge of decision boundary uncertainty in PU learning. By creating synthetic samples through convex combinations, mixup naturally smooths the decision boundaries in regions. This is particularly crucial in PU learning where the model must distinguish between true negatives and mislabeled positives within the unlabeled set. In contrast, the counterintuitive phenomenon of performance degradation with moving average techniques primarily stems from the unstable nature of pseudo-labels in PU learning. Unlike traditional semi-supervised learning where unlabeled data contains truly unlabeled instances, PU learning involves mislabeled negative samples, making historical predictions unreliable. The self-training process generates systematic biases, and moving average perpetuates rather than corrects these biases. On the other hand, moving average techniques may suppress the model's ability to rapidly adapt within the feature space to distinguish between positive and negative samples. Furthermore, the momentum parameter requires careful tuning, which significantly increases the experimental cost for hyperparameter optimization. + +Overall, the combination of soft labeling with class-adaptive thresholds and Mixup yields the best performance across nearly all datasets. The only exception occurs in the CIFAR-10-1 dataset, likely due to its complex visual diversity as a natural image dataset, making the combination of fixed thresholding with moving average and mixup more suitable for handling its complex decision boundaries. These findings suggest that the combination of soft labeling, class-adaptive thresholding, and mixup generally constitutes the most promising universal method. + +Table 6: The ACC scores (mean±std) of existing PU learning methods and GPU. + +
MethodF-MNIST-1F-MNIST-2CIFAR-10-1CIFAR-10-2STL-10-1STL-10-2Rank
uPU77.1±2.262.4±2.781.7±0.977.3±1.176.7±0.871.5±4.815.3
nnPU94.8±0.393.4±0.386.4±0.184.7±1.077.1±4.581.9±1.08.7
absPU94.2±0.493.4±0.386.3±0.984.4±0.975.3±2.282.0±0.79.8
Dist-PU94.7±0.293.3±0.686.7±0.585.6±0.778.3±0.881.5±1.18.7
RP94.4±0.693.3±0.578.0±1.984.2±1.171.3±0.875.5±2.612.2
AdaSampling93.6±0.393.5±0.279.6±0.579.1±1.074.3±2.282.6±0.811.3
GenPU78.1±0.486.2±1.471.2±1.968.3±2.568.5±1.457.3±1.516.5
Self-PU90.8±0.489.1±0.785.1±0.883.9±2.678.5±1.180.8±2.112.2
VPU92.6±1.290.5±0.886.8±1.282.5±1.178.4±1.182.9±0.710.3
PULNS91.0±0.589.1±0.887.2±0.683.7±2.980.2±0.883.6±0.79.8
P³Mix-E92.6±0.491.8±0.288.2±0.484.7±0.580.2±0.983.7±0.77.8
P³Mix-C92.8±0.690.4±0.188.7±0.487.9±0.580.7±0.784.1±0.36.5
Robust-PU90.0±0.585.5±0.780.0±0.685.2±1.179.6±0.980.4±0.812.3
HolisticPU96.2±0.196.0±0.391.0±0.390.4±0.582.5±0.584.0±1.23.2
LaGAM94.9±0.294.1±0.389.9±0.388.0±1.485.3±0.385.0±0.32.8
PUL-CPBF96.7±0.396.5±0.291.4±0.291.0±0.383.4±0.785.4±1.21.2
gPU96.4±0.196.1±0.588.4±0.187.9±0.482.7±0.784.9±0.43.2
PN learning97.7±0.197.7±0.191.9±0.191.9±0.186.0±0.686.0±0.6-
+ +# 3.4 Proposed GPU Framework + +By integrating SAPU with pseudo-labeling, we suggest an efficient PU learning framework GPU. Its generic objective is given as follows: + +$$ +\mathcal {L} _ {\mathrm {G P U}} (g) = \widehat {R} _ {p} ^ {+} (g) + \mathcal {L} _ {u} (g, \hat {y}; D _ {u}) + \frac {\alpha}{n _ {s}} \sum_ {j = 1} ^ {n _ {s}} \ell_ {C E} \left(\frac {1}{S} \sum_ {\mathbf {x} _ {i j} \in \mathcal {S} _ {i}} g (\mathbf {x} _ {i j}), \pi\right), \tag {12} +$$ + +where $\alpha$ is a coefficient parameter. + +We can interpret GPU as a regularized pseudo-labeling method of PU learning, where the set-aware term is treated as a regularization term. Based on the previous evaluations, we find that pseudolabeling methods depend on high quality of pseudo-labels in the early training stage because they are in a self-training manner. Accordingly, we suggest a warm-up stage by minimizing the objective of SAPU. In addition, we can specify the pseudo-labeling techniques and thresholding strategies according to our empirical observations. + +Results and analysis To evaluate the efficacy of our proposed GPU framework compared to existing PU learning methods, we conduct experiments on F-MNIST, CIFAR-10, and STL-10 datasets to assess its general performance across varying scenarios. For comprehensive comparison, we also include PN learning (i.e., supervised learning) as an upper bound baseline. Our GPU implementation uses a subset size $S = 256$ with the sigmoid loss function and employs soft pseudo-labeling with class-adaptive thresholding based on our empirical observations. For the warm-up stage, we train using only SAPU for 20 epochs before introducing the pseudo-labeling component. + +As evident from Table 6, our proposed GPU framework demonstrates competitive performance across all benchmark datasets, ranking 3.2 overall, tied with HolisticPU and slightly behind LaGAM (2.8), while remaining competitive with the leading PUL-CPBF. For example, GPU achieves accuracy scores of $96.4\%$ and $96.1\%$ on F-MNIST-1 and F-MNIST-2 respectively, which are comparable to the best-performing PUL-CPBF $(96.7\%$ and $96.5\%)$ . For CIFAR-10, GPU obtains $88.4\%$ and $87.6\%$ , positioning it among the top-tier methods but slightly below HolisticPU and PUL-CPBF. Similar competitive performance is also demonstrated on the STL-10 dataset. The performance gap between GPU and the best-performing methods reflects our focus on exploring fundamental techniques and integrating them with our novel set-aware empirical risk SAPU, rather than employing sophisticated techniques like ensemble methods in PUL-CPBF, trend detection in HolisticPU, or meta-learning in + +LaGAM. GPU provides a general framework that can integrate future advances, while specialized methods may not generalize well. Notably, GPU significantly outperforms traditional PU learning methods across all datasets, demonstrating that combining set-aware empirical risk estimation with pseudo-labeling strategies effectively enhances the discriminative capability of the model. + +# 4 Discussion and Future Works + +In this paper, we comprehensively review the current families of PU learning and investigate their basic characteristics. We review the existing disambiguation-free empirical risks and suggest a novel set-aware empirical risk SAPU from the perspective of aggregate supervision, which is risk-consistent for the expected risk of supervised learning. We empirically evaluate them with various commonly applied loss functions. In addition, we review the basic techniques and widely applied generic tricks, i.e. mixup and moving average, in the existing pseudo-labeling methods. To empirically evaluate them, we formulate a set of base methods specified by hard and soft pseudo-labeling techniques with thresholding strategies for selecting high-confident pseudo-labels such as fixed, adaptive, and class-specific adaptive thresholding strategies. Finally, we propose an efficient PU learning framework GPU by integrating SAPU with pseudo-labeling. GPU involves a warm-up stage by minimizing SAPU and specify the framework according to our empirical observations. We compare GPU with the existing PU learning methods, and the empirical results demonstrate that GPU can be a practical benchmark in PU learning, and is scalable for future pseudo-labeling techniques. + +In the future, there are two potential problems that require more attention. One basic problem is how to estimate more accurate pseudo-labels [32, 33, 34] since we only investigate the straightforward pseudo-labeling techniques. Some advanced techniques such as ensemble learning [35] demonstrate strong potential. Another problem is whether the existing PU learning can be effective for the scenarios with scarce positive labeled instances and how to deal with such scenarios, which can appear in many real-world applications. + +# Acknowledgements + +We would like to acknowledge support for this project from the National Science and Technology Major Project (No.2021ZD0112500), and the National Natural Science Foundation of China (No.62276113). + +# References + +[1] Kou, Z., J. Wang, Y. Jia, et al. Progressive label enhancement. Pattern Recognition, 160:111172, 2025. +[2] —. Inaccurate label distribution learning. IEEE Transactions on Circuits and Systems for Video Technology, 34(10):10237-10249, 2024. +[3] —. Instance-dependent inaccurate label distribution learning. IEEE Transactions on Neural Networks and Learning Systems, 36(1):1425-1437, 2025. +[4] Bekker, J., J. Davis. Learning from positive and unlabeled data: A survey. Machine Learning, 109(4):719-760, 2020. +[5] du Plessis, M. C., G. Niu, M. Sugiyama. Convex formulation for learning from positive and unlabeled data. In International Conference on Machine Learning, pages 1386-1394. 2015. +[6] Goodfellow, I., J. Pouget-Abadie, M. Mirza, et al. Generative adversarial networks. Communications of the ACM, 63(11):139–144, 2020. +[7] Zhang, H., M. Cisse, Y. N. Dauphin, et al. mixup: Beyond empirical risk minimization. In International Conference on Learning Representations. 2018. +[8] Kiryo, R., G. Niu, M. C. du Plessis, et al. Positive-unlabeled learning with non-negative risk estimator. In Neural Information Processing Systems, pages 1675-1685. 2017. + +[9] Hammoudeh, Z., D. Lowd. Learning from positive and unlabeled data with arbitrary positive shift. In Neural Information Processing Systems. 2020. +[10] Zhao, Y., Q. Xu, Y. Jiang, et al. Dist-pu: Positive-unlabeled learning from a label distribution perspective. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14441-14450. 2022. +[11] Chen, X., W. Chen, T. Chen, et al. Self-pu: Self boosted and calibrated positive-unlabeled training. In International Conference on Machine Learning, pages 1510-1519. 2020. +[12] Chen, H., F. Liu, Y. Wang, et al. A variational approach for learning from positive and unlabeled data. In Neural Information Processing Systems. 2020. +[13] Li, C., X. Li, L. Feng, et al. Who is your right mixup partner in positive and unlabeled learning. In International Conference on Learning Representations. 2022. +[14] Zhu, Z., L. Wang, P. Zhao, et al. Robust positive-unlabeled learning via noise negative sample self-correction. In ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 3663-3673. 2023. +[15] Wang, X., W. Wan, C. Geng, et al. Beyond myopia: Learning from positive and unlabeled data through holistic predictive trends. In Neural Information Processing Systems. 2023. +[16] Li, C., Y. Dai, L. Feng, et al. Positive and unlabeled learning with controlled probability boundary fence. In International Conference on Machine Learning. 2024. +[17] du Plessis, M. C., G. Niu, M. Sugiyama. Analysis of learning from positive and unlabeled data. In Neural Information Processing Systems, pages 703-711. 2014. +[18] Northcutt, C. G., T. Wu, I. L. Chuang. Learning with confident examples: Rank pruning for robust classification with noisy labels. In Conference on Uncertainty in Artificial Intelligence. 2017. +[19] Hinton, G., O. Vinyals, J. Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. +[20] Niu, G., M. C. du Plessis, T. Sakai, et al. Theoretical comparisons of positive-unlabeled learning against positive-negative learning. In Neural Information Processing Systems, pages 1199-1207. 2016. +[21] Elkan, C., K. Noto. Learning classifiers from only positive and unlabeled data. In ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pages 213-220. 2008. +[22] Yang, P., W. Liu, J. Y. H. Yang. Positive unlabeled learning via wrapper-based adaptive sampling. In International Joint Conference on Artificial Intelligence, pages 3273-3279. 2017. +[23] Hou, M., B. Chaib-Draa, C. Li, et al. Generative adversarial positive-unlabeled learning. In International Joint Conference on Artificial Intelligence, pages 2255-2261. 2018. +[24] Luo, C., P. Zhao, C. Chen, et al. PULNS: positive-unlabeled learning with effective negative sample selector. In AAAI Conference on Artificial Intelligence, pages 8784-8792. 2021. +[25] Long, L., H. Wang, Z. Jiang, et al. Positive-unlabeled learning by latent group-aware meta disambiguation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23138-23147. 2024. +[26] Zamzam, O., H. Akrami, M. Soltanolkotabi, et al. Learning a disentangling representation for PU learning. arXiv preprint arXiv:2310.03833, 2024. +[27] Wang, H., R. Xiao, Y. Li, et al. Pico+: Contrastive label disambiguation for robust partial label learning. arXiv preprint arXiv:2201.08984, 2022. +[28] Busa-Fekete, R., H. Choi, T. Dick, et al. Easy learning from label proportions. In Neural Information Processing Systems, pages 14957 - 14968. 2023. + +[29] Sohn, K., D. Berthelot, N. Carlini, et al. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. Neural Information Processing Systems, 33:596-608, 2020. +[30] Zhang, B., Y. Wang, W. Hou, et al. Flexmatch: Boosting semi-supervised learning with curriculum pseudo labeling. Neural Information Processing Systems, 34:18408-18419, 2021. +[31] Berthelot, D., R. Roelofs, K. Sohn, et al. Adamatch: A unified approach to semi-supervised learning and domain adaptation. arXiv preprint arXiv:2106.04732, 2021. +[32] Kou, Z., S. Qin, H. Wang, et al. Label distribution learning with biased annotations by learning multi-label representation, 2025. +[33] Kou, Z., J. Wang, J. Tang, et al. Exploiting multi-label correlation in label distribution learning. In International Joint Conference on Artificial Intelligence, pages 4326-4334. 2024. +[34] Kou, Z., H. Xuan, J. Zhu, et al. Tail-aware reconstruction of incomplete label distributions with low-rank and sparse modeling. IEEE Transactions on Circuits and Systems for Video Technology, pages 1-1, 2025. +[35] Li, C., Y. Dai, L. Feng, et al. Positive and unlabeled learning with controlled probability boundary fence. In International Conference on Machine Learning. 2024. +[36] Hoeffding, W. Probability inequalities for sums of bounded random variables. The collected works of Wassily Hoeffding, pages 409-426, 1994. + +# NeurIPS Paper Checklist + +# 1. Claims + +Question: Do the main claims made in the abstract and introduction accurately reflect the paper's contributions and scope? + +Answer: [Yes] + +Justification: We have succinctly outlined the contributions of this paper in both the abstract and introduction sections, and the results presented in the experiments section robustly substantiate the effectiveness of our proposed method. + +Guidelines: + +- The answer NA means that the abstract and introduction do not include the claims made in the paper. +- The abstract and/or introduction should clearly state the claims made, including the contributions made in the paper and important assumptions and limitations. A No or NA answer to this question will not be perceived well by the reviewers. +- The claims made should match theoretical and experimental results, and reflect how much the results can be expected to generalize to other settings. +- It is fine to include aspirational goals as motivation as long as it is clear that these goals are not attained by the paper. + +# 2. Limitations + +Question: Does the paper discuss the limitations of the work performed by the authors? + +Answer: [Yes] + +Justification: We have created a separate "Limitations" section in our paper, the details can be found in Appendix C.1. + +Guidelines: + +- The answer NA means that the paper has no limitation while the answer No means that the paper has limitations, but those are not discussed in the paper. +- The authors are encouraged to create a separate "Limitations" section in their paper. +- The paper should point out any strong assumptions and how robust the results are to violations of these assumptions (e.g., independence assumptions, noiseless settings, model well-specification, asymptotic approximations only holding locally). The authors should reflect on how these assumptions might be violated in practice and what the implications would be. +- The authors should reflect on the scope of the claims made, e.g., if the approach was only tested on a few datasets or with a few runs. In general, empirical results often depend on implicit assumptions, which should be articulated. +- The authors should reflect on the factors that influence the performance of the approach. For example, a facial recognition algorithm may perform poorly when image resolution is low or images are taken in low lighting. Or a speech-to-text system might not be used reliably to provide closed captions for online lectures because it fails to handle technical jargon. +- The authors should discuss the computational efficiency of the proposed algorithms and how they scale with dataset size. +- If applicable, the authors should discuss possible limitations of their approach to address problems of privacy and fairness. +- While the authors might fear that complete honesty about limitations might be used by reviewers as grounds for rejection, a worse outcome might be that reviewers discover limitations that aren't acknowledged in the paper. The authors should use their best judgment and recognize that individual actions in favor of transparency play an important role in developing norms that preserve the integrity of the community. Reviewers will be specifically instructed to not penalize honesty concerning limitations. + +# 3. Theory assumptions and proofs + +Question: For each theoretical result, does the paper provide the full set of assumptions and a complete (and correct) proof? + +# Answer: [Yes] + +Justification: We analyze the estimation error of SAPU in Theorem 3.2. All theoretical results are clearly supported by rigorous mathematical derivations in Appendix A and B. + +# Guidelines: + +- The answer NA means that the paper does not include theoretical results. +- All the theorems, formulas, and proofs in the paper should be numbered and cross-referenced. +- All assumptions should be clearly stated or referenced in the statement of any theorems. +- The proofs can either appear in the main paper or the supplemental material, but if they appear in the supplemental material, the authors are encouraged to provide a short proof sketch to provide intuition. +- Inversely, any informal proof provided in the core of the paper should be complemented by formal proofs provided in appendix or supplemental material. +- Theorems and Lemmas that the proof relies upon should be properly referenced. + +# 4. Experimental result reproducibility + +Question: Does the paper fully disclose all the information needed to reproduce the main experimental results of the paper to the extent that it affects the main claims and/or conclusions of the paper (regardless of whether the code and data are provided or not)? + +# Answer: [Yes] + +Justification: We have elaborated on the implementation principles and details of the method to facilitate the reproduction of the main experimental results presented in our paper. Additionally, we have submitted our code and datasets in the Supplementary Material. + +# Guidelines: + +- The answer NA means that the paper does not include experiments. +- If the paper includes experiments, a No answer to this question will not be perceived well by the reviewers: Making the paper reproducible is important, regardless of whether the code and data are provided or not. +- If the contribution is a dataset and/or model, the authors should describe the steps taken to make their results reproducible or verifiable. +- Depending on the contribution, reproducibility can be accomplished in various ways. For example, if the contribution is a novel architecture, describing the architecture fully might suffice, or if the contribution is a specific model and empirical evaluation, it may be necessary to either make it possible for others to replicate the model with the same dataset, or provide access to the model. In general, releasing code and data is often one good way to accomplish this, but reproducibility can also be provided via detailed instructions for how to replicate the results, access to a hosted model (e.g., in the case of a large language model), releasing of a model checkpoint, or other means that are appropriate to the research performed. +- While NeurIPS does not require releasing code, the conference does require all submissions to provide some reasonable avenue for reproducibility, which may depend on the nature of the contribution. For example +(a) If the contribution is primarily a new algorithm, the paper should make it clear how to reproduce that algorithm. +(b) If the contribution is primarily a new model architecture, the paper should describe the architecture clearly and fully. +(c) If the contribution is a new model (e.g., a large language model), then there should either be a way to access this model for reproducing the results or a way to reproduce the model (e.g., with an open-source dataset or instructions for how to construct the dataset). +(d) We recognize that reproducibility may be tricky in some cases, in which case authors are welcome to describe the particular way they provide for reproducibility. In the case of closed-source models, it may be that access to the model is limited in some way (e.g., to registered users), but it should be possible for other researchers to have some path to reproducing or verifying the results. + +# 5. Open access to data and code + +Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material? + +Answer: [Yes] + +Justification: We have submitted our code and datasets in the Supplementary Material. + +Guidelines: + +- The answer NA means that paper does not include experiments requiring code. +- Please see the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details. +- While we encourage the release of code and data, we understand that this might not be possible, so "No" is an acceptable answer. Papers cannot be rejected simply for not including code, unless this is central to the contribution (e.g., for a new open-source benchmark). +- The instructions should contain the exact command and environment needed to run to reproduce the results. See the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details. +- The authors should provide instructions on data access and preparation, including how to access the raw data, preprocessed data, intermediate data, and generated data, etc. +- The authors should provide scripts to reproduce all experimental results for the new proposed method and baselines. If only a subset of experiments are reproducible, they should state which ones are omitted from the script and why. +- At submission time, to preserve anonymity, the authors should release anonymized versions (if applicable). +- Providing as much information as possible in supplemental material (appended to the paper) is recommended, but including URLs to data and code is permitted. + +# 6. Experimental setting/details + +Question: Does the paper specify all the training and test details (e.g., data splits, hyperparameters, how they were chosen, type of optimizer, etc.) necessary to understand the results? + +Answer: [Yes] + +Justification: We provide detailed descriptions of all necessary training and testing procedures in Section 3.1. Furthermore, all experimental setup details can be readily found in the submitted code. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The experimental setting should be presented in the core of the paper to a level of detail that is necessary to appreciate the results and make sense of them. +- The full details can be provided either with the code, in appendix, or as supplemental material. + +# 7. Experiment statistical significance + +Question: Does the paper report error bars suitably and correctly defined or other appropriate information about the statistical significance of the experiments? + +Answer: [Yes] + +Justification: We report the results regarding the standard errors of the mean in our experiments and ensure that our paper contains the calculation method for standard errors along with other essential information related to them. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The authors should answer "Yes" if the results are accompanied by error bars, confidence intervals, or statistical significance tests, at least for the experiments that support the main claims of the paper. + +- The factors of variability that the error bars are capturing should be clearly stated (for example, train/test split, initialization, random drawing of some parameter, or overall run with given experimental conditions). +- The method for calculating the error bars should be explained (closed form formula call to a library function, bootstrap, etc.) +- The assumptions made should be given (e.g., Normally distributed errors). +- It should be clear whether the error bar is the standard deviation or the standard error of the mean. +- It is OK to report 1-sigma error bars, but one should state it. The authors should preferably report a 2-sigma error bar than state that they have a $96\%$ CI, if the hypothesis of Normality of errors is not verified. +- For asymmetric distributions, the authors should be careful not to show in tables or figures symmetric error bars that would yield results that are out of range (e.g. negative error rates). +- If error bars are reported in tables or plots, The authors should explain in the text how they were calculated and reference the corresponding figures or tables in the text. + +# 8. Experiments compute resources + +Question: For each experiment, does the paper provide sufficient information on the computer resources (type of compute workers, memory, time of execution) needed to reproduce the experiments? + +Answer: [Yes] + +Justification: We have provided sufficient information on the computer resources needed to reproduce our experiments in Section 3.1. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The paper should indicate the type of compute workers CPU or GPU, internal cluster or cloud provider, including relevant memory and storage. +- The paper should provide the amount of compute required for each of the individual experimental runs as well as estimate the total compute. +- The paper should disclose whether the full research project required more computer than the experiments reported in the paper (e.g., preliminary or failed experiments that didn't make it into the paper). + +# 9. Code of ethics + +Question: Does the research conducted in the paper conform, in every respect, with the NeurIPS Code of Ethics https://neurips.cc/public/EthicsGuidelines? + +Answer: [Yes] + +Justification: We ensure that our research adheres to the NeurIPS Code of Ethics in all aspects. + +Guidelines: + +The answer NA means that the authors have not reviewed the NeurIPS Code of Ethics +- If the authors answer No, they should explain the special circumstances that require a deviation from the Code of Ethics. +- The authors should make sure to preserve anonymity (e.g., if there is a special consideration due to laws or regulations in their jurisdiction). + +# 10. Broader impacts + +Question: Does the paper discuss both potential positive societal impacts and negative societal impacts of the work performed? + +Answer: [Yes] + +Justification: We have created a separate "Broader Impacts" section in our paper, the details can be found in Appendix C.2. + +Guidelines: + +- The answer NA means that there is no societal impact of the work performed. +- If the authors answer NA or No, they should explain why their work has no societal impact or why the paper does not address societal impact. +- Examples of negative societal impacts include potential malicious or unintended uses (e.g., disinformation, generating fake profiles, surveillance), fairness considerations (e.g., deployment of technologies that could make decisions that unfairly impact specific groups), privacy considerations, and security considerations. +- The conference expects that many papers will be foundational research and not tied to particular applications, let alone deployments. However, if there is a direct path to any negative applications, the authors should point it out. For example, it is legitimate to point out that an improvement in the quality of generative models could be used to generate deepfakes for disinformation. On the other hand, it is not needed to point out that a generic algorithm for optimizing neural networks could enable people to train models that generate Deepfakes faster. +- The authors should consider possible harms that could arise when the technology is being used as intended and functioning correctly, harms that could arise when the technology is being used as intended but gives incorrect results, and harms following from (intentional or unintentional) misuse of the technology. +- If there are negative societal impacts, the authors could also discuss possible mitigation strategies (e.g., gated release of models, providing defenses in addition to attacks, mechanisms for monitoring misuse, mechanisms to monitor how a system learns from feedback over time, improving the efficiency and accessibility of ML). + +# 11. Safeguards + +Question: Does the paper describe safeguards that have been put in place for responsible release of data or models that have a high risk for misuse (e.g., pretrained language models, image generators, or scraped datasets)? + +Answer: [NA] + +Justification: The paper pose no such risks. + +Guidelines: + +- The answer NA means that the paper poses no such risks. +- Released models that have a high risk for misuse or dual-use should be released with necessary safeguards to allow for controlled use of the model, for example by requiring that users adhere to usage guidelines or restrictions to access the model or implementing safety filters. +- Datasets that have been scraped from the Internet could pose safety risks. The authors should describe how they avoided releasing unsafe images. +- We recognize that providing effective safeguards is challenging, and many papers do not require this, but we encourage authors to take this into account and make a best faith effort. + +# 12. Licenses for existing assets + +Question: Are the creators or original owners of assets (e.g., code, data, models), used in the paper, properly credited and are the license and terms of use explicitly mentioned and properly respected? + +Answer: [Yes] + +Justification: We have provided proper citations for all models, code, and datasets utilized in our paper. + +Guidelines: + +- The answer NA means that the paper does not use existing assets. +- The authors should cite the original paper that produced the code package or dataset. +- The authors should state which version of the asset is used and, if possible, include a URL. +- The name of the license (e.g., CC-BY 4.0) should be included for each asset. + +- For scraped data from a particular source (e.g., website), the copyright and terms of service of that source should be provided. +- If assets are released, the license, copyright information, and terms of use in the package should be provided. For popular datasets, paperswithcode.com/datasets has curated licenses for some datasets. Their licensing guide can help determine the license of a dataset. +- For existing datasets that are re-packaged, both the original license and the license of the derived asset (if it has changed) should be provided. +- If this information is not available online, the authors are encouraged to reach out to the asset's creators. + +# 13. New assets + +Question: Are new assets introduced in the paper well documented and is the documentation provided alongside the assets? + +Answer: [Yes] + +Justification: We have submitted the proposed new assets in the Supplementary Material. And the submitted files include structured explanatory documents regarding these new assets. + +Guidelines: + +- The answer NA means that the paper does not release new assets. +- Researchers should communicate the details of the dataset/code/model as part of their submissions via structured templates. This includes details about training, license, limitations, etc. +- The paper should discuss whether and how consent was obtained from people whose asset is used. +- At submission time, remember to anonymize your assets (if applicable). You can either create an anonymized URL or include an anonymized zip file. + +# 14. Crowdsourcing and research with human subjects + +Question: For crowdsourcing experiments and research with human subjects, does the paper include the full text of instructions given to participants and screenshots, if applicable, as well as details about compensation (if any)? + +Answer: [NA] + +Justification: Our paper does not involve crowdsourcing nor research with human subjects. + +Guidelines: + +- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. +- Including this information in the supplemental material is fine, but if the main contribution of the paper involves human subjects, then as much detail as possible should be included in the main paper. +- According to the NeurIPS Code of Ethics, workers involved in data collection, curation, or other labor should be paid at least the minimum wage in the country of the data collector. + +# 15. Institutional review board (IRB) approvals or equivalent for research with human subjects + +Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals (or an equivalent approval/review based on the requirements of your country or institution) were obtained? + +Answer: [NA] + +Justification: Our paper does not involve crowdsourcing nor research with human subjects. + +Guidelines: + +- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. + +- We recognize that the procedures for this may vary significantly between institutions and locations, and we expect authors to adhere to the NeurIPS Code of Ethics and the guidelines for their institution. +- For initial submissions, do not include any information that would break anonymity (if applicable), such as the institution conducting the review. + +# 16. Declaration of LLM usage + +Question: Does the paper describe the usage of LLMs if it is an important, original, or non-standard component of the core methods in this research? Note that if the LLM is used only for writing, editing, or formatting purposes and does not impact the core methodology, scientific rigorousness, or originality of the research, declaration is not required. + +Answer: [NA] + +Justification: The core method development in this research does not involve LLMs as any important, original, or non-standard components. + +Guidelines: + +- The answer NA means that the core method development in this research does not involve LLMs as any important, original, or non-standard components. +- Please refer to our LLM policy (https://neurips.cc/Conferences/2025/LLM) for what should or should not be described. + +# A Proof of Lemma 3.1 + +The boundary of the bag deviation Under the SCAR assumption [21], each sample in the unlabeled dataset has an independent probability $\pi$ of being positive. Given a bag $j$ containing $s$ samples, since the variance of a Bernoulli random variable is $\pi (1 - \pi)$ , we can obtain a tighter bound using Bernstein's inequality [36] for any $\epsilon >0$ : + +$$ +P \left(\left| \hat {\pi} _ {j} - \pi \right| \geq \epsilon\right) \leq 2 \exp \left(- \frac {S \epsilon^ {2}}{2 \pi (1 - \pi) + 2 \epsilon / 3}\right) \leq \delta \tag {13} +$$ + +Then, + +$$ +S \geq \frac {3 \pi (1 - \pi)}{2 \epsilon^ {2}} \log (2 / \delta) \tag {14} +$$ + +The variance of $\hat{\pi}_j$ The variance of $\hat{\pi}_j$ can be given by the mean of $S$ independent Bernoulli random variables: + +$$ +\operatorname {V a r} \left(\hat {\pi} _ {j}\right) = \frac {\pi (1 - \pi)}{S} \tag {15} +$$ + +# B Proof of Theorem 3.2 + +We decompose the excess risk as: + +$$ +\begin{array}{l} R \left(\hat {g} _ {\mathrm {S A P U}}\right) - R \left(g ^ {*}\right) \leq \\ \underbrace {\left| R \left(\hat {g} _ {\mathrm {S A P U}}\right) - R _ {\mathrm {S A P U}} \left(\hat {g} _ {\mathrm {S A P U}}\right) \right|} _ {\text {T e r m 1}} + \underbrace {\left| R _ {\mathrm {S A P U}} \left(\hat {g} _ {\mathrm {S A P U}}\right) - R _ {\mathrm {S A P U}} \left(g ^ {*}\right) \right|} _ {\leq 0} + \underbrace {\left| R _ {\mathrm {S A P U}} \left(g ^ {*}\right) - R \left(g ^ {*}\right) \right|} _ {\text {T e r m 2}} \tag {16} \\ \end{array} +$$ + +For Term 1, using the uniform convergence theory and the fact that the deviation in bag proportions is bounded by $\epsilon$ , we have: + +$$ +\left| R \left(\hat {g} _ {\mathrm {S A P U}}\right) - R _ {\mathrm {S A P U}} \left(\hat {g} _ {\text {b a g}}\right) \right| \leq C _ {1} \sqrt {\frac {d \log \left(n _ {p}\right) + \log \left(1 / \delta\right)}{n _ {p}}} + C _ {2} \sqrt {\frac {d \log \left(n _ {s}\right) + \log \left(1 / \delta\right)}{n _ {s}}} + \epsilon L _ {\ell} \tag {17} +$$ + +where $L_{\ell}$ is the Lipschitz constant of the cross-entropy loss; $C_1$ and $C_2$ are universal constants; $d$ is the pseudo-dimension of the function class. + +According to Hoeffding's inequality, for any $\delta > 0$ , $|\hat{\pi}_j - \pi| \leq \sqrt{\frac{\log(2 / \delta)}{2S}}$ holds with probability at least $1 - \delta$ . Then, we have: + +$$ +\left| R \left(\hat {g} _ {\mathrm {S A P U}}\right) - R _ {\mathrm {S A P U}} \left(\hat {g} _ {\text {b a g}}\right) \right| \leq C _ {1} \sqrt {\frac {d \log \left(n _ {p}\right) + \log (1 / \delta)}{n _ {p}}} + C _ {2} \sqrt {\frac {d \log \left(n _ {s}\right) + \log (1 / \delta)}{n _ {s}}} + L _ {\ell} \sqrt {\frac {\log (2 / \delta)}{2 S}} \tag {18} +$$ + +For Term 2, the deviation comes from the difference between the true positive class prior $\pi$ and the bag proportions $\hat{\pi}_j$ . Using the variance bound from Lemma 3.1 and applying Jensen's inequality: + +$$ +\left| R _ {\mathrm {S A P} U} \left(g ^ {*}\right) - R \left(g ^ {*}\right) \right| \leq L _ {\ell} \sqrt {\mathbb {E} \left[ \left(\hat {\pi} _ {j} - \pi\right) ^ {2} \right]} = L _ {\ell} \sqrt {\operatorname {V a r} \left(\hat {\pi} _ {j}\right)} = L _ {\ell} \sqrt {\frac {\pi (1 - \pi)}{S}} \tag {19} +$$ + +Then, + +$$ +\begin{array}{l} R (\hat {g} _ {\mathrm {S A P U}}) - R (g ^ {*}) \leq C _ {1} \sqrt {\frac {d \log (n _ {p}) + \log (1 / \delta)}{n _ {p}}} + C _ {2} \sqrt {\frac {d \log (n _ {s}) + \log (1 / \delta)}{n _ {s}}} \\ + L _ {\ell} \sqrt {\frac {\log (2 / \delta)}{2 S}} + L _ {\ell} \sqrt {\frac {\pi (1 - \pi)}{S}} \\ = C _ {1} \sqrt {\frac {d \log (n _ {p}) + \log (1 / \delta)}{n _ {p}}} + C _ {2} \sqrt {\frac {d \log (n _ {s}) + \log (1 / \delta)}{n _ {s}}} \\ + L _ {\ell} \left(\sqrt {\frac {\log (2 / \delta)}{2 S}} + \sqrt {\frac {\pi (1 - \pi)}{S}}\right) \\ = O \left(\sqrt {\frac {\log (1 / \delta)}{n _ {p}}}\right) + O \left(\sqrt {\frac {\log (1 / \delta)}{n _ {s}}}\right) + L _ {\ell} \cdot O \left(\sqrt {\frac {\pi (1 - \pi) \log (1 / \delta)}{S}}\right) \tag {20} \\ \end{array} +$$ + +# C Limitations and Broader Impacts + +# C.1 Limitations + +Despite our comprehensive empirical study, accurately estimating pseudo-labels remains challenging, especially with limited positive samples. Our methods could be further improved by incorporating advanced techniques such as ensemble learning to generate more reliable pseudo-labels. Additionally, the set-aware empirical risk method may face challenges with extremely imbalanced datasets where the positive class prior becomes difficult to estimate accurately. + +# C.2 Broader Impacts + +Our GPU framework introduces a novel perspective by integrating empirical risk with pseudo-labeling methods, enhancing PU learning applicability in real-world scenarios such as medical diagnoses and fraud detection. The proposed set-aware empirical risk extends the theoretical foundation of PU learning through aggregate supervision, which could inspire further weakly-supervised learning research. By making PU learning more reliable with limited labeled data, our work contributes to reduced annotation costs and broader accessibility of machine learning in resource-constrained environments. \ No newline at end of file diff --git a/NeurIPS/2025/A Closer Look to Positive-Unlabeled Learning from Fine-grained Perspectives_ An Empirical Study/images.zip b/NeurIPS/2025/A Closer Look to Positive-Unlabeled Learning from Fine-grained Perspectives_ An Empirical Study/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..b2351316353b51782b64faebccd9d5b240d32f02 --- /dev/null +++ b/NeurIPS/2025/A Closer Look to Positive-Unlabeled Learning from Fine-grained Perspectives_ An Empirical Study/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:688d4dfba4482a13d55eb13d69641edb68737f8de398be6c2577e06ea9ff7385 +size 810966 diff --git a/NeurIPS/2025/A Closer Look to Positive-Unlabeled Learning from Fine-grained Perspectives_ An Empirical Study/layout.json b/NeurIPS/2025/A Closer Look to Positive-Unlabeled Learning from Fine-grained Perspectives_ An Empirical Study/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..d2fdee3c0f98826d64f76e05d920b6ef8f640e82 --- /dev/null +++ b/NeurIPS/2025/A Closer Look to Positive-Unlabeled Learning from Fine-grained Perspectives_ An Empirical Study/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ad3c3fde3668c4fd59eba1ebff3663491923029c424eef956df329ac197e3410 +size 659610 diff --git a/NeurIPS/2025/A Compressive-Expressive Communication Framework for Compositional Representations/bca251ad-d7bc-4af1-b226-405cb73a2406_content_list.json b/NeurIPS/2025/A Compressive-Expressive Communication Framework for Compositional Representations/bca251ad-d7bc-4af1-b226-405cb73a2406_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..452281b10a9ef06d50890db729a1f46957a7805e --- /dev/null +++ b/NeurIPS/2025/A Compressive-Expressive Communication Framework for Compositional Representations/bca251ad-d7bc-4af1-b226-405cb73a2406_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:08401b3dc26cd3b9781941702d7a3a2e593d8dd9ef5fdd3919c9c5e03d888aef +size 260247 diff --git a/NeurIPS/2025/A Compressive-Expressive Communication Framework for Compositional Representations/bca251ad-d7bc-4af1-b226-405cb73a2406_model.json b/NeurIPS/2025/A Compressive-Expressive Communication Framework for Compositional Representations/bca251ad-d7bc-4af1-b226-405cb73a2406_model.json new file mode 100644 index 0000000000000000000000000000000000000000..21a72e5a406f6844044c09e16df0d01baf25d483 --- /dev/null +++ b/NeurIPS/2025/A Compressive-Expressive Communication Framework for Compositional Representations/bca251ad-d7bc-4af1-b226-405cb73a2406_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dc831a8476fb1b0012d6903b9ae877600f7bc99f1d81b878fe438b9a1d1ce004 +size 325438 diff --git a/NeurIPS/2025/A Compressive-Expressive Communication Framework for Compositional Representations/bca251ad-d7bc-4af1-b226-405cb73a2406_origin.pdf b/NeurIPS/2025/A Compressive-Expressive Communication Framework for Compositional Representations/bca251ad-d7bc-4af1-b226-405cb73a2406_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..74abe7137bae960584014ed83052d790b16a3fdf --- /dev/null +++ b/NeurIPS/2025/A Compressive-Expressive Communication Framework for Compositional Representations/bca251ad-d7bc-4af1-b226-405cb73a2406_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:88544739f69006c6d0b4729edd14a5f78f884419b7aceb6027d0422459fdb12e +size 3876363 diff --git a/NeurIPS/2025/A Compressive-Expressive Communication Framework for Compositional Representations/full.md b/NeurIPS/2025/A Compressive-Expressive Communication Framework for Compositional Representations/full.md new file mode 100644 index 0000000000000000000000000000000000000000..73ec7648d3b806ab482b862d7a09bb8ecc656a1e --- /dev/null +++ b/NeurIPS/2025/A Compressive-Expressive Communication Framework for Compositional Representations/full.md @@ -0,0 +1,1290 @@ +# A Compressive-Expressive Communication Framework for Compositional Representations + +Rafael Elberg * + +Pontificia Universidad Catolica, CENIA, i-Health + +Chile + +rafael.elpberg@uc.cl + +Felipe del Rio + +Pontificia Universidad Catolica, CENIA + +Chile + +fidelrio@uc.cl + +Mircea Petrache + +Pontificia Universidad Catolica, CENIA Chile + +mpetrache@uc.cl + +Denis Parra + +Pontificia Universidad Catolica, CENIA, i-Health Chile + +dparras@uc.cl + +# Abstract + +Compositionality in knowledge and language—the ability to represent complex concepts as a combination of simpler ones—is a hallmark of human cognition and communication. Despite recent advances, deep neural networks still struggle to acquire this property reliably. Neural models for emergent communication look to endow artificial agents with compositional language by simulating the pressures that form human language. In this work, we introduce CELEBI² (Compressive-Expressive Language Emergence through a discrete Bottleneck and Iterated learning), a novel self-supervised framework for inducing compositional representations through a reconstruction-based communication game between a sender and a receiver. Building on theories of language emergence and the iterated learning framework, we integrate three mechanisms that jointly promote compressibility, expressivity, and efficiency in the emergent language. First, Progressive Decoding incentivizes intermediate reasoning by requiring the receiver to produce partial reconstructions after each symbol. Second, Final-State Imitation trains successive generations of agents to imitate reconstructions rather than messages, enforcing a tighter communication bottleneck. Third, Pairwise Distance Maximization regularizes message diversity by encouraging high distances between messages, with formal links to entropy maximization. Our method significantly improves both the efficiency and compositionality of the learned messages on the Shapes3D and MPI3D datasets, surpassing prior discrete communication frameworks in both reconstruction accuracy and topographic similarity. This work provides new theoretical and empirical evidence for the emergence of structured, generalizable communication protocols from simplicity-based inductive biases. + +# 1 Introduction + +In natural languages, compositionality enables humans to communicate an infinite number of ideas using a finite set of elements [12]. This principle enables speakers to flexibly combine known words and structures to convey novel meanings, supporting flexible and generalizable communication across diverse and previously unseen contexts. + +Works on the emergence of compositionality in language [34, 35, 37] have argued that opposing pressures are necessary for the natural selection of compositional languages between generations of speakers [20, 3]. On the one hand, successful communication requires high expressivity in order to usefully describe the world, allowing speakers to produce distinct messages for a wide range of meanings. On the other hand, models of emergent communication such as the iterated learning framework [34] state that natural speakers tend to minimize the complexity of languages through cultural transmission, implying that simpler and more compressive languages are more easily passed on to new speakers. + +These opposite pressures of expressivity and compressibility thus generate a trade-off, highlighted in several works [35, 20, 3] which is argued to be optimized by compositional languages, whereby languages evolve to maximize communicative efficiency – remaining expressive enough to convey diverse meanings while being simple and structured enough to be easily learned and transmitted by successive generations of speakers. When either pressure is dominant, undesirable languages tend to emerge: (a) excessive language compression leads to degenerate languages where multiple meanings are mapped to the same messages making them ambiguous and thus hard to use, and (b) prioritizing expressivity alone produces in holistic languages i.e. languages where messages are mapped arbitrarily to meanings without respecting their structures, hindering their transmission across generations [55]. + +Enabling machines to generalize compositionally is thought to be crucial for them to quickly adapt to novel situations beyond their training experience [42]. Inducing compositional behavior in neural networks remains a major challenge [39, 30, 59, 33]. A growing body of work suggests that a model's ability to generalize compositionally is highly sensitive to its training conditions [40, 14, 60, 41, 15], including factors such as data distribution, learning objectives, and task design. One promising direction is the study of emergent communication, where discrete languages evolve for coordination between independent neural agents. + +In this work, we build on the Lewis reconstruction game framework [58] to study emergent communication in a reconstruction task. A sender encodes an image into a discrete message and a receiver reconstructs the original. Drawing on theories of language evolution [34, 35], we develop a novel framework for inducing compositional communication grounded in simplicity bias. + +We introduce CELEBI (Compressive-Expressive Language Emergence through a discrete Bottleneck and Iterated learning). Within this framework, we introduce three mechanisms in the learned communication protocol: + +Progressive Decoding, in which the receiver makes reconstruction predictions after each incoming symbol, rather than waiting until the full message is received. This biases the system towards using intermediate reasoning steps, thereby imposing a pressure towards lower-complexity and less holistic encodings. While this mechanism does not directly improve reconstruction accuracy, it yields more efficient and structured communication. + +Final-State Imitation modifies the standard imitation phase in the iterated learning (IL) framework of cultural evolution [35-37]. Instead of imitating the entire message, the student is trained to reproduce only the final output of the receiver, effectively reducing the information transmitted across generations. This tighter generational bottleneck increases pressure for compressibility in the emergent communication protocol [64], thereby promoting the emergence of compositional structure as a necessary condition for successful transmission. Empirically, this leads to increased compositionality with minimal degradation in reconstruction accuracy. + +Pairwise Distance Maximization as a regularization term, i.e., as a further pressure towards increased diversity in the emergent language, that encourages exploration during the imitation phase. This term pushes the student sender to maximize an approximate Hamming distance between messages in a given batch, finding the most "diverse" protocol. We prove that this regularization gives a lower bound for entropy maximization [69], and an upper bound on the contrastive learning loss NT-Xtent [11]. + +The opposing pressures induced by these three mechanisms impose a tight regularization on the complexity of the emergent language. These ingredients make explicit the pressures proposed in the cognitive science and linguistic literature to shape language evolution, thus aligning with the IL paradigm and leading to measurably higher compositionality. + +# 2 Background + +# 2.1 Problem setup: recovering compositional representations + +In a formalism similar to [56], we model a dataset $\mathcal{D} = \{x = \operatorname{GenX}(\mathbf{G}) : \mathbf{G} \in \mathcal{G}\}$ , i.e. data $x$ are created via a deterministic function $\operatorname{GenX}$ from a set $\mathcal{G}$ of generating factors $\mathbf{G}$ . We have access to samples from $\mathcal{D}$ but not to $\operatorname{GenX}$ or to the structure of $\mathbf{G}$ , and the goal is to reproduce the set of $x$ , i.e., generate images $\widehat{x}$ that approximate the distribution of $x$ in a natural metric such as $\mathbb{E}_{x \in \mathcal{D}}[\operatorname{MSE}(x, \widehat{x})]$ , where MSE is the mean squared error on pixels. + +The factors $\mathbf{G}$ are assumed to have compositional structure of the form $\mathbf{G} = [G_1, \dots, G_n]$ , in which $G_i$ represent independent characteristics having finitely many possible values (as in the dataset of Sec. 4.1 for example). In this case, an efficient way to reproduce elements of $\mathcal{D}$ is to impose that $\widehat{x}$ is similarly generated as $\widehat{x} = \mathrm{GenX}'(\mathbf{G}')$ , with generating factors $\mathbf{G}'$ giving a learned encoding of $\mathbf{G}$ . + +As $\mathcal{G}$ has a large number of classes (one per combination of factor values) and we have a single example per class, successful reconstruction relies on the compositional structure of $\mathbf{G}$ . Thus the goal is to build a framework for finding the optimized compositional encoding $\mathbf{G}'$ from observation of a small training set $\mathcal{D}_{train} \subset \mathcal{D}$ . The compositional nature of $\mathbf{G}$ is what makes reconstruction possible from a small random $\mathcal{D}_{train}$ , i.e. from data corresponding to a small subset of classes. This would be of course impossible for non-compositional data (see §D for a formal treatment). + +# 2.2 Approach: Lewis reconstruction game + +We frame the above reconstruction problem in the form of a cooperative game between two agents: a sender $S$ and a receiver $R$ . The goal of $S$ will be to build the reconstruction factors $\mathbf{G}'$ and the receiver $R$ will map the factors to an image $\widehat{x}$ . The two agents aim to create a protocol that reconstructs $x$ to good accuracy. This setup fits in the general class of Lewis Reconstruction Games (LRG) [58], a subclass of so-called Lewis Signaling Games (LSGs) [43]. + +Within the framework of signaling games, the factors $\mathbf{G}'$ are interpreted as a language which $S$ uses to communicate $x$ to $R$ as accurately as possible. The allowed message space is finite, i.e. $\mathbf{G}' \in \mathcal{M} \coloneqq V^C$ , and we take $C$ fixed and large enough that $|V|^C \gg |\mathcal{G}|$ , so that the size of the message space is not a limiting factor on the emergent communication. The following diagram summarizes our notations: + +$$ +\mathcal {G} \xrightarrow {\operatorname {G e n X}} \mathcal {D} \subseteq \mathcal {X} \xrightarrow {S} \mathcal {M} = V ^ {C} \xrightarrow {R} \mathcal {X} +$$ + +Our main focus is on defining a IL process which favors the emergence of a compositional language between $S$ and $R$ that accurately recovers $x \in \mathcal{D}$ . + +# 2.3 Base model architecture + +We adopt a standard emergent communication setup as described by Xu et al. [71], see implementation details in § B.1. Sender and Receiver are implemented using the EGG framework [32]. Input images are encoded into latent representations using a pretrained VAE visual backbone. The sender encodes this latent input into a discrete message using an LSTM and then transmits it through a fixed-length communication channel. Message generation is made differentiable via a Gumbel-Softmax bottleneck [26]. The receiver takes the full message as inputs in an LSTM, and produces a latent vector, which is subsequently decoded into a reconstructed image by the VAE decoder. Reconstruction loss is computed between the original and generated images using Mean Squared Error (MSE). Both VAE encoder and decoder remain frozen during training. + +# 3 Methods + +This section describes our proposed improvements to the EL framework [71] (cf. Figure 1a). Intuitively, our additions attempt to increase the language drift between iterations of speakers, in order to thoroughly explore the landscape of possible languages, while maintaining efficient and useful communication. For the latter, we introduce Progressive Decoding (PD), to serve as an anchor for the emergent language during interaction, and for the former we introduce Final-State Imitation (FiSI) and Parwise Distance Maximization (PDM), which aim to incentivize exploration during the + +imitation phase. We expect the combination of these factors to show the strongest impact towards compositionality in the learned language, more so than the factors taken separately. + +# 3.1 Progressive Decoding (PD) + +Our first proposed enhancement to the baseline communication protocol aims to improve the efficiency of the emergent language determined by the Sender and Receiver, by condensing information in fewer message tokens. This type of efficiency has been argued to be necessary for the emergence of natural language properties [29, 49]. + +Standard reconstruction games typically optimize a reconstruction loss computed only at the end of the communication phase, after the receiver has observed the full message [58], and as shown by Rita et al. [57], without an incentive to efficiently use the information received in the message, the receiver only updates its internal state after the full message has been received. To improve communication efficiency, we propose the following Progressive Decoding (PD) objective: the receiver generates a prediction after each received symbol, and our interaction phase loss explicitly weights the reconstruction error of each sub-message, as follows. + +$$ +\mathcal {L} _ {\omega , \phi} ^ {\text {i n t}} := \mathbb {E} _ {x} \left[ \frac {1}{C} \sum_ {i = 1} ^ {C} \lambda^ {i} \operatorname {M S E} \left(x, R _ {\omega} \left(S _ {\phi} (x) _ {[ i ]}\right)\right) \right], \tag {1} +$$ + +where $S_{\phi}(x)_{[i]}$ is the prefix of length $i$ of the sender's message, and $\lambda \geq 1$ is an expressivity hyperparameter that increases the weight of later reconstructions. This loss has the following effects: + +Efficiency pressure: Loss (1) rewards accurate reconstructions made as early as possible, favoring messages with shorter useful length. + +Interpretable sub-messages: The model is encouraged to structure messages such that an interpretation by $R_{\omega}$ is available for sub-messages, making each symbol accountable for partial reconstruction. This echoes findings in prior theoretical and empirical work on chain-of-thought and compositionality [66, 28, 10, 52, 68, 54], which show that forcing intermediate semantic consistency or intermediate reasoning steps encourages models to show increased compositional behavior. + +Tuning the efficiency pressure: Small $\lambda$ yields fast but potentially coarse reconstructions, while larger values favor detailed outputs at the cost of longer, possibly redundant messages, and in the limit $\lambda \rightarrow \infty$ (1) becomes equivalent to using just the last term in brackets, i.e., the full-length message reconstruction. In our experiments (see Section 5.2), we explore this trade-off empirically and show that $\lambda \equiv 1.5$ achieves the best performance across generalization, communication efficiency metrics, and compositionality. + +# 3.2 Imitation Phase + +Iterated learning is known to add compressibility pressure to emergent communication schemes [34], resulting in languages that are "easy to learn", thus favoring simple communication schemes that can be compressed: this is known as a communication (or generational) bottleneck. It is argued [34] that languages that remain relatively stable even when a learner only observes a small subset of the language of the previous generation need to be compositional. + +Our case (and in fact, most deep learning applications of IL) differ from the original IL formulation [34], in that the transmission bottleneck does not constrict the amount of examples presented to the student, but rather constricts the information obtainable from these examples in order to perfectly reconstruct the language. This paradigm shift for transmission bottlenecks is at the heart of discrete bottlenecks [55, 58, 57] and noisy bottlenecks [56]. + +We next pass to describe our proposed imitation scheme and its motivations compared to previous methods. + +# 3.2.1 Final-State Imitation (FiSI) + +As usual in IL, we apply a teacher-student regime to initialize the next iteration sender $S_{\phi^{t + 1}}$ . + +Our main proposed novelties for the imitation phase are as follows. + +State space imitation: As opposed to other IL implementations [55, 55, 44], our models are not directed to match the student protocols from different iterations in the message space, but rather in the state space, using the frozen receiver network from the interaction phase. This change allows a wider range of sender strategies, as the new loss is more permissive. + +Final state reconstruction: During imitation phase we use the following reconstruction loss, where $R_{t}^{\star} = R_{\omega^{t}}$ , $S_{t}^{\star} = S_{\phi^{t}}$ are frozen receiver and sender from the previous iteration and $S_{t + 1} = S_{\phi^{t + 1}}$ is the newly trained sender: + +$$ +\mathcal {L} _ {\phi^ {t + 1}} ^ {r e c} := \mathbb {E} _ {x} \left[ d _ {\mathcal {X}} \left(R _ {t} ^ {\star} \left(S _ {t} ^ {\star} (x)\right), R _ {t} ^ {\star} \left(S _ {\phi^ {t + 1}} (x)\right)\right) \right]. \tag {2} +$$ + +Our loss $\mathcal{L}_{\phi^{t + 1}}^{rec}$ only depends on the state associated to the full message, without testing reconstruction on sub-messages. This change allows the intermediate message tokens of the student sender $S_{\phi^{t + 1}}(x)_i,i < C$ to drift further away from the ones of the teacher mapping $S_t^* (x)_i$ and thus more easily find different strategies with the same end result. + +State space imitation is justified as follows. Note that as message space is discrete, there exists a value $\delta_0$ such that any two messages at distance $< \delta_0$ in fact coincide. Then $d_{\mathcal{M}}(S_t^\star (x),S_{t + 1}(x)) < \delta_0$ always implies $d_{\mathcal{X}}(R_t^\star (S_t^\star (x)),R_t^\star (S_{t + 1}(x))) = 0$ , while the vice-versa is not always true: if $R_{t}^{\star}$ is not injective, any $S_{t + 1}(x)\in (R_t^\star)^{-1}(S_t^\star (x))$ will preserve zero distance of $R_{t}^{\star}$ -images. If $R_{t}^{\star}$ has mild Lipschitz regularity assumption, this principle extends to the regime of small loss, showing that our new loss is less restrictive on the choices of $S_{t + 1}$ than the traditional message space loss. Thus, state space imitation enables wider sender strategy exploration without penalization losses. + +![](images/e12959a1a984277042cd05bd67535ae55d834bb653e82a338895912fe10603ce.jpg) +(a) Overview of our proposed architecture + +![](images/1c68e5eb76d9fe893ca405d7210f2918bcc61bb8fcf518dd63a076db5b0563a7.jpg) +(b) Qualitative example of image decoding + +![](images/18b0884ba13f149509032a30981d04da77a02bd4e135af1ffffb8efda4b44e8e.jpg) +Figure 1: (a) Overview of our proposed architecture. Interaction Phase: The sender $S_{\phi^t}$ and receiver $R_{\omega^t}$ are jointly trained to minimize the reconstruction error between the input state $x$ and the predicted states $\{x'\}$ , by encoding $x$ into a message $m_x^t$ . Imitation Phase: A new sender $S_{\phi^{t+1}}$ is trained to imitate the final predicted output of $R_{\omega^t}(m_x^t)$ , while also maximizing pairwise message diversity to encourage exploration. (b) Qualitative example of image decoding: the receiver reconstructs the input image from the sender's message for each sub-message, giving a series of reconstructions. The useful length of a message corresponds to the first reconstruction that has error below a threshold $\epsilon > 0$ : in this example the last two message tokens are not useful in that they do not add to the reconstruction accuracy. + +The choice of using just final states is justified because final state reconstruction allows greater freedom during imitation for the student sender. By only constraining the reconstruction of the full message, we allow much more freedom for the encoding strategy of $S(x)$ at earlier sub-messages at each iteration (see Proposition E.1 for a formal statement and proof). + +# 3.2.2 Pairwise Distance Maximization (PDM) + +To further improve protocol diversity and promote exploration of the language space during the imitation phase, we introduce a regularization objective that promotes dissimilarity between messages within each batch. In addition to enabling a larger search space for optimal student strategies. Enhancing the diversity of student strategies also fits the general IL argument according to which compositional languages are the most resistant to noisy language transmission [34, 37, 36]. + +Specifically, we approximate the Hamming distance by computing the position-wise cosine dissimilarity between the probability distributions over symbols in each message. This encourages the sender to generate messages that are maximally distinct while remaining semantically aligned with the teacher protocol. The strength of this regularization is controlled by a hyperparameter $\beta$ . + +The resulting imitation loss is defined as: + +$$ +\mathcal {L} _ {\phi^ {t + 1}} ^ {i m} := \mathcal {L} _ {\phi^ {t + 1}} ^ {r e c} + \beta \frac {1}{N _ {b a t c h} ^ {2}} \sum_ {i, j} d \left(S _ {\phi^ {t + 1}} \left(x _ {i}\right), S _ {\phi^ {t + 1}} \left(x _ {j}\right)\right), \tag {3} +$$ + +where $d(\cdot, \cdot)$ is the mean cosine similarity between corresponding symbol distributions. + +In § F.4, we show that this regularization term provides a lower bound on entropy maximization objectives [69], and serves as an upper bound for contrastive losses such as NT-Xent [11], thus connecting our formulation to well-established principles in representation learning. + +# 4 Experimental setup + +# 4.1 Datasets + +Shapes3D We tested our framework using the Shapes3D [50] dataset, which consists of colored images of 3D geometric shapes, with 6 underlying generating factors $G$ : floor hue, wall hue, object hue, shape, scale, and orientation. The total amount of attribute-value combinations is 480,000. + +MPI3D We also evaluated using the compositional split of the MPI3D dataset [23]. This dataset consists of colored images of a robot arm interacting with objects, rendered in controlled 3D scenes with 7 underlying generative factors $G$ : object color, object shape, object size, camera height, background color, horizontal arm position, and vertical arm position. The dataset comprises 1,036,800 unique combinations of these factor values. + +For both datasets we use the compositional split from Schott et al. [62], which ensures all attribute values appear in training, yet some combinations are reserved for the testing set (See §D for a proof that this situation would anyways hold for random train/test splittings with high probability). + +# 4.2 Evaluation Metrics + +Following Chaabouni et al. [9], we assess compositionality using Topographic Similarity (TopSim) [5], a widely used proxy for compositionality. TopSim measures the correlation between pairwise Hamming distances in message space and the generating factor space, capturing the extent to which semantically similar inputs yield similar messages under a structured encoding. + +To evaluate communication efficiency, we define the useful length $\hat{\ell}_{\epsilon}(x)$ , which estimates the minimal prefix of a message necessary to achieve near-maximal reconstruction quality. Formally: + +$$ +\hat {\ell} _ {\epsilon} (x) := \min \left\{i \in \{1, \dots , C \}: \mathrm {M S E} \left(x, R _ {\omega} (S _ {\phi} (x) _ {[ i ]})\right) \leq \epsilon \right\}, +$$ + +where $C$ is the maximum message length and $\ell_{\epsilon}(x) = C$ if the threshold is unmet. + +We choose $\epsilon$ by viewing the loss distributions on both tested datasets and estimating a common plateau point for each. The full position-wise loss values can be found in the supplementary material. + +Finally, to approximate the expressivity of the communication protocol, we evaluate reconstruction quality using the MSE between the generated image and the original input. + +We discuss the use of additional language metrics in $\S I$ . + +# 5 Results + +For all experiments, we report the mean and standard error of 10 random seeds. To assess the robustness of our findings, we also perform permutation tests to evaluate the statistical significance of + +Table 1: Performance comparison across all experiments on the SHAPES3D and MPI3D datasets. We report Topographic Similarity (TopSim ↑), Useful Length ( $\hat{\ell}_{\epsilon} \downarrow$ ), and Last Symbol MSE ( $\downarrow$ ) for each ablation. The results are grouped by the components of our proposed framework: Interaction, Imitation, and Regularization. The synergetic mixture of our proposed methods (PD+FiSI+PDM) consistently improves compositionality and message efficiency over the baseline and prior variants. + +
TopSim ↑\( \hat{\ell}_{2\times 10^{-1}}\downarrow \)Last symbol MSE ↓
InteractionFull message reconstruction, no IL (Baseline)0.244 ± 0.00210.0 ± 0.00.212 ± 0.003
Progressive Decoding(PD, ours)0.270 ± 0.0017.5 ± 0.2150.238 ± 0.007
ImitationMessage imitation (baseline)0.257 ± 0.0059.9 ± 0.30.200 ± 0.007
Full state imitation0.256 ± 0.0039.9 ± 0.30.179 ± 0.003
Final-State Imitation (FiSI, ours)0.283 ± 0.00310.0 ± 0.00.176 ± 0.002
RegularizationPD+FiSI+KoLeo λ = 1.50.256 ± 0.0028.555 ± 0.1670.179 ± 0.003
PD+FiSI+PDM (ours) λ = 1.50.292 ± 0.0027.0 ± 0.1550.194 ± 0.005
+ +(a) Shapes3D + +
TopSim ↑\( \hat{\ell}_{1.95\times 10^{-2}}\downarrow \)Last symbol MSE ↓
InteractionFull message reconstruction, no IL (Baseline)0.133 ± 0.0019.3 ± 0.0020.015 ± 0.0
Progressive Decoding(PD, ours)0.137 ± 0.0016.6 ± 0.3410.015 ± 0.0
ImitationMessage imitation (baseline)0.135 ± 0.0019.733 ± 0.0290.016 ± 0.0
Full state imitation0.137 ± 0.0019.923 ± 0.0200.018 ± 0.0
Final-State Imitation (FiSI, ours)0.156 ± 0.0019.7 ± 0.0460.02 ± 0.0
RegularizationPD+FiSI+KoLeo λ = 1.50.147 ± 0.0028.9 ± 0.2210.02 ± 0.0
PD+FiSI+PDM (ours) λ = 1.50.153 ± 0.0019.0 ± 0.1670.02 ± 0.0
+ +(b)MPI3D + +the observed differences (see §J.1). Additionally, we include qualitative reconstruction experiments in §H, where we visually compare the outputs of all evaluated methods across progressive decoding steps. Further implementation details are provided in §B. + +# 5.1 Progressive Decoding reduces the message's useful length + +We begin by evaluating the effect of incorporating PD during the interactive phase of training. As shown in the first two rows of Table 1, this modification leads to a $25 - 29\%$ reduction in the useful message length, indicating more efficient communication. These gains are achieved without compromising expressivity on MPI3D and with only a slight compromise on Shapes3D, as indicated by stable reconstruction quality in the former and a minor degradation in the latter. + +To further promote efficient and structured communication, we introduce a geometric penalty term $\lambda$ that penalizes reconstruction error more heavily when longer messages are used. As shown in Figure 2, increasing $\lambda$ up to values near 1.5 consistently improves reconstruction quality and useful length, indicating higher expressivity and efficiency. However, when $\lambda$ becomes too large, the reconstruction error incurred at the end of the message dominates the loss function, effectively nullifying PD. We found that setting $\lambda = 1.5$ yields favorable results in efficiency and expressivity, while maintaining high compressibility. + +# 5.2 Final State Imitation increases TopSim + +Second, we investigate the impact of final state imitation on the structure of the emergent language. + +In these experiments, we compare our approach to a baseline model without imitation and to a standard IL variant that imitates the teacher's message directly—a strategy commonly adopted in prior work for optimizing this phase [55, 56, 44]. + +As shown in the imitation section in Table 1, in the MPI3D dataset, we find that our final state imitation method consistently enhances the compositionality of the emergent messages. It outperforms the non-IL baseline and surpasses message-level imitation significantly (§ J.1). + +![](images/c46e5298fe569feea105e00aeafa6bbcdfafd5d11266950655490d4bd1eb550f.jpg) +(a) Useful Length + +![](images/f8ad2dfe0c744dfa77d401ccf49d835ad6ab9010def5d9a2d0883eeed004c8fd.jpg) +(b) Reconstruction MSE + +![](images/9a1957b855176730e6cb3dcf00011b9e842631ddaf9a9e4af55234b82ed65eee.jpg) +(c) TopSim +Figure 2: Effect of geometric penalty $\lambda$ on emergent communication. We introduce a geometric weighting term $\lambda$ that penalizes reconstruction error more strongly for longer messages. Increasing $\lambda$ initially improves useful length (a). However, once a certain threshold is met, useful length and reconstruction error (b) increase. Values close to 1.5 of $\lambda$ strike a favorable balance—achieving efficient and expressive communication. + +More broadly, our results support prior findings that IL promotes greater structure in emergent communication protocols [44, 55]. They also underscore the value of shifting the learning objective from message imitation to image reconstruction, which enforces a tighter generational bottleneck by requiring the student to recover the intended output without access to the original message. + +Moreover, this setup increases flexibility in protocol exploration by constraining the student only to the reconstruction target, not the exact message form. Permitting generational variation in message structure while preserving semantic fidelity, promoting diversity without sacrificing expressivity. + +Together, these factors exert pressure toward discovering more systematic and compositional protocols—structures that are inherently more transmissible and robust across generations. + +# 5.3 Pairwise Distance Maximization increases compositionality and efficiency + +We evaluate a model that combines all previously identified best-performing components—specifically, $\lambda = 1.5$ , final-state imitation, PD and further introduce an entropy-based regularization term using PDM. We compare this with the KoLeo entropy estimator proposed by Sablayrolles et al. [61]. + +When examining the effect of incorporating PDM, Table 1 shows that PDM outperforms both FiSi and PD in fostering structure in the message. On the Shapes3D dataset, PDM leads to a marked improvement in TopSim and a lower useful message length, indicating more efficient and structured communication, albeit with a slight increase in reconstruction error. For MPI3D, this configuration doesn't achieve the highest TopSim, albeit close to the best, and it increases useful length and leads to a marginally higher reconstruction loss compared to using only PD. This highlights a trade-off between expressivity and compressibility, in line with prior theoretical accounts on communication efficiency and compositionality [53, 34, 49]. + +When comparing PDM to the KoLeo estimator, Regularization section in Table 1, we find that on Shapes3D, PDM achieves superior TopSim and shorter useful length, with slight reconstruction gains. On MPI3D, KoLeo leads to a lower useful length but exhibits reduced compositionality, with both regularizers showing similar reconstruction performance. + +# 5.4 Comparison to disentangled representation learning + +Following the comparison made by Xu et al. [71], we test our generated messages against standard self-supervised disentangled representation learning frameworks, namely $\beta$ -VAE and $\beta$ -TCVAE, as well as the baseline VAE+EL [71]. For the continuous models, we extract the first half of the latent vector (corresponding to the predicted means $\mu$ ) and evaluate it using the DCI disentanglement score [18]. In addition, we train a two-layer MLP to predict ground-truth generative factors, reporting RMSE for continuous factors and classification accuracy for categorical ones. + +For the discrete models, we use the predicted messages for each image. We calculate the DCI disentanglement as-is, and use an embedding matrix to transform the discrete messages into the same dimension as the continuous vectors for the MLP. Training is conducted on three subsets of size 1000 + +Table 2: Disentanglement metrics for the Shapes3D and MPI3D datasets. We report the DCI score to evaluate disentanglement quality, and the RMSE and classification accuracy of a linear probe trained on the latent representations (both message and vector embeddings) to assess their suitability for downstream tasks. + +
DatasetModelDisentanglement ↑RMSE ↓Acc ↑
Shapes3Dβ-VAE0.045 ± 0.0042.460 ± 0.1440.702 ± 0.025
β-TCVAE0.043 ± 0.0052.378 ± 0.1230.721 ± 0.028
VAE + EL0.108 ± 0.0003.168 ± 0.0270.343 ± 0.019
VAE + CELEBI0.112 ± 0.0012.932 ± 0.0450.453 ± 0.024
MPI3Dβ-VAE0.031 ± 0.00121.819 ± 0.1110.581 ± 0.006
β-TCVAE0.031 ± 0.00121.854 ± 0.1880.582 ± 0.005
VAE + EL0.114 ± 0.00215.971 ± 0.3010.529 ± 0.006
VAE + CELEBI0.137 ± 0.00214.82 ± 0.0120.542 ± 0.002
+ +of the training set and averaged. Evaluations are conducted on the test sets of the compositional splits for the Shapes3D and MPI3D datasets. + +Across both datasets, our method (VAE+CELEBI) demonstrates consistent improvements in disentanglement over baseline methods. On Shapes3D, we observe a substantial increase in DCI scores compared to continuous models and the discrete baseline (see statistical significance tests in § J.2). However, both discrete models fall behind the continuous baselines in downstream accuracy and RMSE, suggesting a trade-off between representation interpretability and usefulness in downstream tasks, which is consistent with previous work indicating that discrete representations may be less accessible to simple classifiers (as shown in [72]), and that inductive biases for compositionality do not imply disentanglement [51]. + +On MPI3D, our method slightly under-performs when compared to the continuous baselines in categorical accuracy, however greatly outperforming all baselines in continuous regression. We speculate that this discrepancy arises from the lower correlation between pixel-level input statistics and continuous generative factors in MPI3D as opposed to Shapes3D, which may favor communication-based models over purely reconstructive. + +# 6 Related Work + +The use of language reference games (LRGs) in conjunction with iterated learning for the purpose of learning self-supervised representations remains underexplored. Xu et al. [71] compare an emergent language autoencoder—comprising a visual backbone and an LSTM-based sender and receiver—with disentanglement frameworks such as the $\beta$ -VAE and $\beta$ -TCVAE. They find that the representations induced by the emergent language model generalize better in downstream tasks, highlighting the potential of symbolic communication as a compositional bottleneck. + +Many previous works have explored using emergent language to induce compositional behavior [31, 17, 2, 16, 8]. Several works utilize a referential LSG in conjunction with IL, using predicted messages as a generational bottleneck [55, 44]. Alternative bottlenecks for emergent language have been proposed, including simplicial embeddings [56, 19], code-books [74] and noisy channels [65]. Our message tokens can also be considered as roughly comparable to the slots in slot-attention architectures such as [22, 24, 46, 27, 45, 21, 48, 4, 1], however note that among other differences, the slot-attention typically additive decoding layer [7, 38] does not compare directly to our proposed Progressive Decoding. + +To replicate the Zipfian distributions observed in natural languages [76, 29], Rita et al. [57] propose the LazImpa framework. The LazImpa Impatient listener loss is similar to our proposed Progressive Decoding, in that they both aim to induce incremental informativeness and efficiency by reconstructing the input at every timestep. However LazImpa is optimized for a referential game task (and not a reconstruction task as ours) and uses a linear length penalization in conjunction with a cross-entropy loss. Our proposal instead uses the actual reconstruction loss and the length penalization is exponential and based on a tunable parameter $\lambda$ . + +# 7 Limitations and Future Work + +Our experiments are conducted exclusively on the synthetic datasets *Shapes3D* and *MPI3D* [6, 23], which feature clean, disentangled generative factors and lack observational noise. In future work, we plan to extend our the same principle of making explicit pressures towards complexity reduction of underlying representations to natural image datasets where factor supervision may be unavailable or weakly defined. Furthermore, we aim to formalize the notion of language compression defined in IL in terms of Kolmogorov complexity, for representing datasets with more complex compositional structures. + +Moreover, it is debatable whether MSE is the most appropriate choice for the self-supervised loss function, especially in the context of natural image data. While MSE is a simple choice, future work may explore reconstruction losses that better align with human visual perception, such as perceptual similarity metrics based on deep feature embeddings (e.g., LPIPS [73]), or adversarial and contrastive losses that promote more semantically structured representations. + +Additionally, we froze the visual backbone across all EL model variants trained on the same dataset. While this ensures controlled comparisons within each dataset, it may introduce an initialization bias when comparing with continuous baselines, where the encoder is jointly optimized. Although our results remain statistically significant, future work could investigate whether co-training the vision encoder yields improved alignment between symbolic and perceptual representations, or leads to new trade-offs in compression and expressivity. + +Finally, our experiments were restricted to compositional generalization metrics and reconstruction-based evaluation. Other facets of emergent communication, such as robustness and interpretability, are left for future work. We believe our framework offers a strong foundation for such extensions and a tractable setting to explore the interplay between compression, efficiency, and generalization. + +# 8 Conclusion + +We introduce the CELEBI framework with three novel mechanisms for enhancing compositional learning in IL frameworks, as well as mathematical justification for their design. First, with Progressive Decoding (PD) we reward informative communication at each step of the message, creating an inductive bias toward efficient and distributed representations. We show that PD promotes both message compressibility and compositional alignment. + +Second, in our Final State Imitation (FiSI) in IL, the student is trained to reproduce the final prediction of the teacher, rather than the message itself. This shift enables greater exploration of the message space while preserving semantic consistency. We provided theoretical motivation and empirical evidence that this approach yields more expressive and compositional protocols, outperforming traditional message imitation in TopSim and reconstruction metrics. + +Third, our regularizer based on Pairwise Distance Maximization (PDM), which provably approximates entropy maximization over messages, serves as a practical inductive bias for promoting diversity and structure in emergent languages, particularly during the imitation phase. + +Together, these contributions enrich the IL framework, and give new insights into how training dynamics and inductive pressures shape the emergence of language-like representations. As confirmed by our empirical findings, the proposed methods lead to more compositional and generalizable communication schemes. + +# Acknowledgments and Disclosure of Funding + +This work was supported by ANID Chile, Fondecyt Regular grant 1231724, as well research centers of excellence with code FB210017 (Basal CENIA), ICN2021_004 (Millenium iHealth), and ICN17_002 (Millenium IMFD). + +# References + +[1] Kaan Akan and Yucel Yemez. Slot-guided adaptation of pre-trained diffusion models for object-centric learning and compositional generation. In The Thirteenth International Conference on + +Learning Representations, 2025. URL https://openreview.net/forum?id=kZvor5aaz7. +[2] Michal Auersperger and Pavel Pecina. Defending compositionality in emergent languages. In Daphne Ippolito, Liunian Harold Li, Maria Leonor Pacheco, Danqi Chen, and Nianwen Xue, editors, Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies: Student Research Workshop, pages 285-291, Hybrid: Seattle, Washington + Online, July 2022. Association for Computational Linguistics. doi: 10.18653/v1/2022.naacl-srw.35. URL https://aclanthology.org/2022.naacl-srw.35/. +[3] Aaron Beppu and Thomas Griffiths. Iterated learning and the cultural ratchet. In Proceedings of the annual meeting of the cognitive science society, volume 31, 2009. +[4] Jack Brady, Julius von Kugelgen, Sebastien Lachapelle, Simon Buchholz, Thomas Kipf, and Wieland Brendel. Interaction asymmetry: A general principle for learning composable abstractions. In The Thirteenth International Conference on Learning Representations, 2025. URL https://openreview.net/forum?id=cCl10IU836. +[5] Henry Brighton and Simon Kirby. Understanding linguistic evolution by visualizing the emergence of topographic mappings. Artificial Life, 12:229-242, 2006. URL https://api.sementicscholar.org/CorpusID:1391836. +[6] Chris Burgess and Hyunjik Kim. 3d shapes dataset. https://github.com/deepmind/3dshapes-dataset/, 2018. +[7] Christopher P Burgess, Loic Matthew, Nicholas Watters, Rishabh Kabra, Irina Higgins, Matt Botvinick, and Alexander Lerchner. Monet: Unsupervised scene decomposition and representation. arXiv preprint arXiv:1901.11390, 2019. +[8] Rahma Chaabouni, Eugene Kharitonov, Diane Bouchacourt, Emmanuel Dupoux, and Marco Baroni. Compositionality and generalization in emergent languages. In ACL 2020-8th annual meeting of the Association for Computational Linguistics, 2020. +[9] Rahma Chaabouni, Florian Strub, Florent Altché, Eugene Tarassov, Corentin Tallec, Elnaz Davoodi, Kory Wallace Mathewson, Olivier Tieleman, Angeliki Lazaridou, and Bilal Piot. Emergent communication at scale. In International Conference on Learning Representations, 2022. URL https://openreview.net/forum?id=AUGBfDIV9rL. +[10] Ayush K Chakravarthy, Jacob Labe Russian, and Randall O'Reilly. Systematicity emerges in transformers when abstract grammatical roles guide attention. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies: Student Research Workshop, pages 1-8, 2022. +[11] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations, 2020. URL https://arxiv.org/abs/2002.05709. +[12] Noam Chomsky. Aspects of the Theory of Syntax, volume 11. MIT press, 2014. +[13] Henry Conklin and Kenny Smith. Compositionality with variation reliably emerges in neural networks. In The Eleventh International Conference on Learning Representations, 2023. URL https://openreview.net/forum?id=-Yzz6v1X7V-. +[14] Henry Conklin, Bailin Wang, Kenny Smith, and Ivan Titov. Meta-learning to compositionally generalize. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 3322-3335, 2021. +[15] Felipe del Río, Alain Raymond-Sáez, Daniel Florea, Rodrigo Toro Icarte, Julio Hurtado, Cristián Buc Calderón, and Álvaro Soto. Data distributional properties as inductive bias for systematic generalization. arXiv preprint arXiv:2502.20499, 2025. + +[16] Kevin Denamganai and James Alfred Walker. On (emergent) systematic generalisation and compositionality in visual referential games with straight-through gumbel-softmax estimator. CoRR, abs/2012.10776, 2020. URL https://arxiv.org/abs/2012.10776. +[17] Roberto Dessì, Eugene Kharitonov, and Marco Baroni. Interpretable agent communication from scratch (with a generic visual processor emerging on the side). In Neural Information Processing Systems, 2021. URL https://api_semanticscholar.org/CorpusID:235368124. +[18] Cian Eastwood and Christopher K I Williams. A framework for the quantitative evaluation of disentangled representations. In Sixth International Conference on Learning Representations (ICLR 2018), May 2018. URL https://iclr.cc/Conferences/2018.6th International Conference on Learning Representations, ICLR 2018; Conference date: 30-04-2018 Through 03-05-2018. +[19] Rafael Elberg, Denis Parra, and Mircea Petrache. Long tail image generation through feature space augmentation and iterated learning. In LatinX in AI at Computer Vision and Pattern Recognition Conference 2024, LXAI at CVPR 2024. Journal of LatinX in AI Research, June 2024. doi: 10.52591/lxai202406174. URL http://dx.doi.org/10.52591/lxai202406174. +[20] Eric Elmoznino, Thomas Jiralerspong, Yoshua Bengio, and Guillaume Lajoie. A complexity-based theory of compositionality, 2024. URL https://arxiv.org/abs/2410.14817. +[21] Martin Engelcke, Adam R. Kosiorek, Oiwi Parker Jones, and Ingmar Posner. Genesis: Generative scene inference and sampling with object-centric latent representations. In International Conference on Learning Representations, 2020. URL https://openreview.net/forum?id=BkxfaTVFwH. +[22] SM Eslami, Nicolas Heess, Theophane Weber, Yuval Tassa, David Szepesvari, Geoffrey E Hinton, et al. Attend, infer, repeat: Fast scene understanding with generative models. Advances in neural information processing systems, 29, 2016. +[23] Muhammad Waleed Gondal, Manuel Wuthrich, Djordje Miladinovic, Francesco Locatello, Martin Breidt, Valentin Volchkov, Joel Akpo, Olivier Bachem, Bernhard Schölkopf, and Stefan Bauer. On the transfer of inductive bias from simulation to the real world: a new disentanglement dataset. In H. Wallach, H. Larochelle, A. Beygelzimer, F. d'Alché-Buc, E. Fox, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 32. Curran Associates, Inc., 2019. URL https://proceedings.neurips.cc/paper/2019/file/d97d404b6119214e4a7018391195240a-Paper.pdf. +[24] Klaus Greff, Raphaël Lopez Kaufman, Rishabh Kabra, Nick Watters, Christopher Burgess, Daniel Zoran, Loic Matthey, Matthew Botvinick, and Alexander Lerchner. Multi-object representation learning with iterative variational inference. In International conference on machine learning, pages 2424-2433. PMLR, 2019. +[25] Serhii Havrylov and Ivan Titov. Emergence of language with multi-agent games: Learning to communicate with sequences of symbols. Advances in neural information processing systems, 30, 2017. +[26] Eric Jang, Shixiang Gu, and Ben Poole. Categorical reparameterization with gumbel-softmax, 2017. URL https://arxiv.org/abs/1611.01144. +[27] Jindong Jiang, Sepehr Janghorbani, Gerard De Melo, and Sungjin Ahn. *Scalor: Generative world models with scalable object representations*. In *International Conference on Learning Representations*. +[28] Yichen Jiang and Mohit Bansal. Inducing transformer's compositional generalization ability via auxiliary sequence prediction tasks. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 6253-6265, 2021. +[29] Jasmeen Kanwal, Kenny Smith, Jennifer Culbertson, and Simon Kirby. Zipf's law of abbreviation and the principle of least effort: Language users optimise a miniature lexicon for efficient communication. Cognition, 165:45-52, May 2017. + +[30] Daniel Keysers, Nathanael Scharli, Nathan Scales, Hylke Buisman, Daniel Furrer, Sergii Kashubin, Nikola Momchev, Danila Sinopalnikov, Lukasz Stafiniak, Tibor Tihon, et al. Measuring compositional generalization: A comprehensive method on realistic data. In International Conference on Learning Representations, 2019. +[31] Eugene Kharitonov, Rahma Chaabouni, Diane Bouchacourt, and Marco Baroni. Entropy minimization in emergent languages. In International Conference on Machine Learning, pages 5220-5230. PMLR, 2020. +[32] Eugene Kharitonov, Roberto Dessi, Rahma Chaabouni, Diane Bouchacourt, and Marco Baroni. EGG: a toolkit for research on Emergence of lanGuage in Games. https://github.com/facebookresearch/EGG, 2021. +[33] Najoung Kim and Tal Linzen. Cogs: A compositional generalization challenge based on semantic interpretation. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 9087-9105, 2020. +[34] S. Kirby. Spontaneous evolution of linguistic structure - an iterated learning model of the emergence of regularity and irregularity. Evolutionary Computation, IEEE Transactions on, 5: 102-110, 05 2001. doi: 10.1109/4235.918430. +[35] Simon Kirby and James R. Hurford. The Emergence of Linguistic Structure: An Overview of the Iterated Learning Model, pages 121-147. Springer London, London, 2002. ISBN 978-1-4471-0663-0. doi: 10.1007/978-1-4471-0663-0_6. URL https://doi.org/10.1007/978-1-4471-0663-0_6. +[36] Simon Kirby, Tom Griffiths, and Kenny Smith. Iterated learning and the evolution of language. Current opinion in neurobiology, 28:108-114, 2014. +[37] Simon Kirby, Monica Tamariz, Hannah Cornish, and Kenny Smith. Compression and communication in the cultural evolution of linguistic structure. Cognition, 141:87-102, 2015. ISSN 0010-0277. doi: https://doi.org/10.1016/j.cognition.2015.03.016. URL https://www.sciencedirect.com/science/article/pii/S0010027715000815. +[38] Sébastien Lachapelle, Divyat Mahajan, Ioannis Mitlagkas, and Simon Lacoste-Julien. Additive decoders for latent variables identification and cartesian-product extrapolation. Advances in Neural Information Processing Systems, 36:25112-25150, 2023. +[39] Brenden Lake and Marco Baroni. Generalization without systematicity: On the compositional skills of sequence-to-sequence recurrent networks. In International conference on machine learning, pages 2873-2882. PMLR, 2018. +[40] Brenden M Lake. Compositional generalization through meta sequence-to-sequence learning. In H. Wallach, H. Larochelle, A. Beygelzimer, F. d'Alché-Buc, E. Fox, and R. Garnett, editors, Advances in Neural Information Processing Systems 32, pages 9791-9801. Curran Associates, Inc., 2019. URL http://papers.nips.cc/paper/9172-compositional-generalization-through-meta-sequence-to-sequence-learning.pdf. +[41] Brenden M Lake and Marco Baroni. Human-like systematic generalization through a meta-learning neural network. Nature, 623(7985):115-121, 2023. +[42] Brenden M Lake, Tomer D Ullman, Joshua B Tenenbaum, and Samuel J Gershman. Building machines that learn and think like people. Behavioral and brain sciences, 40:e253, 2017. +[43] David Kellogg Lewis. Convention: A philosophical study. 1969. +[44] Fushan Li and Michael Bowling. Ease-of-teaching and language structure from emergent communication, 2019. URL https://arxiv.org/abs/1906.02403. +[45] Zhixuan Lin, Yi-Fu Wu, Skand Peri, Bofeng Fu, Jindong Jiang, and Sungjin Ahn. Improving generative imagination in object-centric world models. In International conference on machine learning, pages 6140-6149. PMLR, 2020. + +[46] Zhixuan Lin, Yi-Fu Wu, Skand Vishwanath Peri, Weihao Sun, Gautam Singh, Fei Deng, Jindong Jiang, and Sungjin Ahn. Space: Unsupervised object-oriented scene representation via spatial attention and decomposition. In International Conference on Learning Representations, 2020. URL https://openreview.net/forum?id=rk103ySYDH. +[47] Francesco Locatello, Stefan Bauer, Mario Lucic, Gunnar Raetsch, Sylvain Gelly, Bernhard Schölkopf, and Olivier Bachem. Challenging common assumptions in the unsupervised learning of disentangled representations. In International Conference on Machine Learning, pages 4114-4124, 2019. +[48] Francesco Locatello, Dirk Weissenborn, Thomas Unterthiner, Aravindh Mahendran, Georg Heigold, Jakob Uszkoreit, Alexey Dosovitskiy, and Thomas Kipf. Object-centric learning with slot attention. Advances in neural information processing systems, 33:11525-11538, 2020. +[49] Diana Rodríguez Luna, E. Ponti, Dieuwke Hupkes, and Elia Bruni. Internal and external pressures on language emergence: Least effort, object constancy and frequency. In Findings, 2020. URL https://api-semanticscholar.org/CorpusID:198921556. +[50] Loic Matthew, Irina Higgins, Demis Hassabis, and Alexander Lerchner. dsprites: Disentangle-ment testing sprites dataset. https://github.com/deepmind/dsprites-dataset/, 2017. +[51] Milton Llera Montero, Casimir JH Ludwig, Rui Ponte Costa, Gaurav Malhotra, and Jeffrey Bowers. The role of disentanglement in generalisation. In International Conference on Learning Representations, 2021. URL https://openreview.net/forum?id=qbH974jKUVy. +[52] Maxwell Nye, Anders Johan Andreassen, Guy Gur-Ari, Henrik Michalewski, Jacob Austin, David Bieber, David Dohan, Aitor Lewkowycz, Maarten Bosma, David Luan, Charles Sutton, and Augustus Odena. Show your work: Scratchpads for intermediate computation with language models, 2022. URL https://openreview.net/forum?id=iedYJm92o0a. +[53] Steven T. Piantadosi, Harry Tily, and Edward Gibson. The communicative function of ambiguity in language. Cognition, 122(3):280-291, 2012. ISSN 0010-0277. doi: https://doi.org/10.1016/j.cognition.2011.10.004. URL https://www.sciencedirect.com/science/article/pii/S0010027711002496. +[54] Ben Prystawski, Michael Li, and Noah Goodman. Why think step by step? reasoning emerges from the locality of experience. Advances in Neural Information Processing Systems, 36, 2024. +[55] Yi Ren, Shangmin Guo, Matthieu Labeau, Shay B. Cohen, and Simon Kirby. Compositional languages emerge in a neural iterated learning model. In International Conference on Learning Representations, 2020. URL https://openreview.net/forum?id=HkePNpVKPB. +[56] Yi Ren, Samuel Lavoie, Michael Galkin, Danica J Sutherland, and Aaron C Courville. Improving compositional generalization using iterated learning and simplicial embeddings. Advances in Neural Information Processing Systems, 36, 2024. +[57] Mathieu Rita, Rahma Chaabouni, and Emmanuel Dupoux. "LazImpa": Lazy and impatient neural agents learn to communicate efficiently. In Raquel Fernández and Tal Linzen, editors, Proceedings of the 24th Conference on Computational Natural Language Learning, pages 335-343, Online, November 2020. Association for Computational Linguistics. doi: 10.18653/v1/2020.conll-1.26. URL https://aclanthology.org/2020.conll-1.26/. +[58] Mathieu Rita, Corentin Tallec, Paul Michel, Jean-Bastien Grill, Olivier Pietquin, Emmanuel Dupoux, and Florian Strub. Emergent communication: Generalization and overfitting in lewis games. In Alice H. Oh, Alekh Agarwal, Danielle Belgrave, and Kyunghyun Cho, editors, Advances in Neural Information Processing Systems, 2022. URL https://openreview.net/forum?id=qqHMvHbfu6. +[59] Laura Ruis, Jacob Andreas, Marco Baroni, Diane Bouchacourt, and Brenden M Lake. A benchmark for systematic generalization in grounded language understanding. Advances in Neural Information Processing Systems, 33, 2020. + +[60] LE Ruis and Brenden Lake. Improving systematic generalization through modularity and augmentation. In Proceedings of the Annual Meeting of the Cognitive Science Society, volume 44, 2022. +[61] Alexandre Sablayrolles, Matthijs Douze, Cordelia Schmid, and Hervé Jégou. Spreading vectors for similarity search. In International Conference on Learning Representations, 2019. URL https://openreview.net/forum?id=SkGuG2R5tm. +[62] Lukas Schott, Julius Von Kugelgen, Frederik Träuble, Peter Vincent Gehler, Chris Russell, Matthias Bethge, Bernhard Schölkopf, Francesco Locatello, and Wieland Brendel. Visual representation learning does not generalize strongly within the same domain. In International Conference on Learning Representations, 2022. URL https://openreview.net/forum?id=9RUHP11adgh. +[63] Florian Schroff, Dmitry Kalenichenko, and James Philbin. Facenet: A unified embedding for face recognition and clustering. In 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), page 815-823. IEEE, June 2015. doi: 10.1109/cvpr.2015.7298682. URL http://dx.doi.org/10.1109/cvpr.2015.7298682. +[64] Kenny Smith, Simon Kirby, and Henry Brighton. Iterated learning: A framework for the emergence of language. Artificial Life, 9(4):371-386, 2003. doi: 10.1162/106454603322694825. +[65] Ryo Ueda and Tadahiro Taniguchi. Lewis's signaling game as beta-vae for natural word lengths and segments, 2024. URL https://arxiv.org/abs/2311.04453. +[66] Petar Velickovic, Rex Ying, Matilde Padovano, Raia Hadsell, and Charles Blundell. Neural execution of graph algorithms. In International Conference on Learning Representations, 2020. URL https://openreview.net/forum?id=SkgK00EtvS. +[67] Pauli Virtanen, Ralf Gommers, Travis E. Oliphant, Matt Haberland, Tyler Reddy, David Cournapeau, Evgeni Burovski, Pearu Peterson, Warren Weckesser, Jonathan Bright, Stefan J. van der Walt, Matthew Brett, Joshua Wilson, K. Jarrod Millman, Nikolay Mayorov, Andrew R. J. Nelson, Eric Jones, Robert Kern, Eric Larson, C J Carey, Ilhan Polat, Yu Feng, Eric W. Moore, Jake VanderPlas, Denis Laxalde, Josef Perktold, Robert Cimrman, Ian Henriksen, E. A. Quintero, Charles R. Harris, Anne M. Archibald, Antonio H. Ribeiro, Fabian Pedregosa, Paul van Mulbregt, and SciPy 1.0 Contributors. SciPy 1.0: Fundamental Algorithms for Scientific Computing in Python. Nature Methods, 17:261-272, 2020. doi: 10.1038/s41592-019-0686-2. +[68] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824-24837, 2022. +[69] Ronald Williams and Jing Peng. Function optimization using connectionist reinforcement learning algorithms. Connection Science, 3:241-, 09 1991. doi: 10.1080/09540099108946587. +[70] Ronald J Williams. Simple statistical gradient-following algorithms for connectionist reinforcement learning. Machine learning, 8:229-256, 1992. +[71] Zhenlin Xu, Marc Niethammer, and Colin Raffel. Compositional generalization in unsupervised compositional representation learning: A study on disentanglement and emergent language, 2022. URL https://arxiv.org/abs/2210.00482. +[72] Zhenlin Xu, Marc Niethammer, and Colin A Raffel. Compositional generalization in unsupervised compositional representation learning: A study on disentanglement and emergent language. Advances in Neural Information Processing Systems, 35:25074-25087, 2022. +[73] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 586-595, 2018. doi: 10.1109/CVPR.2018.00068. +[74] Chenhao Zheng, Jieyu Zhang, Aniruddha Kembhavi, and Ranjay Krishna. Iterated learning improves compositionality in large vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13785-13795, 2024. + +[75] Hattie Zhou, Ankit Vani, Hugo Larochelle, and Aaron Courville. Fortuitous forgetting in connectionist networks, 2022. URL https://arxiv.org/abs/2202.00155. +[76] George K.Zipf. Human Behaviour and the Principle of Least Effort. Addison-Wesley, 1949. + +# NeurIPS Paper Checklist + +The checklist is designed to encourage best practices for responsible machine learning research, addressing issues of reproducibility, transparency, research ethics, and societal impact. Do not remove the checklist: The papers not including the checklist will be desk rejected. The checklist should follow the references and follow the (optional) supplemental material. The checklist does NOT count towards the page limit. + +Please read the checklist guidelines carefully for information on how to answer these questions. For each question in the checklist: + +- You should answer [Yes], [No], or [NA]. +- [NA] means either that the question is Not Applicable for that particular paper or the relevant information is Not Available. +- Please provide a short (1–2 sentence) justification right after your answer (even for NA). + +The checklist answers are an integral part of your paper submission. They are visible to the reviewers, area chairs, senior area chairs, and ethics reviewers. You will be asked to also include it (after eventual revisions) with the final version of your paper, and its final version will be published with the paper. + +The reviewers of your paper will be asked to use the checklist as one of the factors in their evaluation. While "[Yes]" is generally preferable to "[No]", it is perfectly acceptable to answer "[No]" provided a proper justification is given (e.g., "error bars are not reported because it would be too computationally expensive" or "we were unable to find the license for the dataset we used"). In general, answering "[No]" or "[NA]" is not grounds for rejection. While the questions are phrased in a binary way, we acknowledge that the true answer is often more nuanced, so please just use your best judgment and write a justification to elaborate. All supporting evidence can appear either in the main paper or the supplemental material, provided in appendix. If you answer [Yes] to a question, in the justification please point to the section(s) where related material for the question can be found. + +IMPORTANT, please: + +- Delete this instruction block, but keep the section heading "NeurIPS Paper Checklist", +- Keep the checklist subsection headings, questions/answers and guidelines below. +- Do not modify the questions and only use the provided macros for your answers. + +# 1. Claims + +Question: Do the main claims made in the abstract and introduction accurately reflect the paper's contributions and scope? + +Answer: [Yes] + +Justification: The claims and contributions are clearly stated in the abstract and introduction, and are backed by theoretical and experimental results. + +Guidelines: + +- The answer NA means that the abstract and introduction do not include the claims made in the paper. +- The abstract and/or introduction should clearly state the claims made, including the contributions made in the paper and important assumptions and limitations. A No or NA answer to this question will not be perceived well by the reviewers. +- The claims made should match theoretical and experimental results, and reflect how much the results can be expected to generalize to other settings. +- It is fine to include aspirational goals as motivation as long as it is clear that these goals are not attained by the paper. + +# 2. Limitations + +Question: Does the paper discuss the limitations of the work performed by the authors? + +Answer: [Yes] + +Justification: The paper contains a separate limitations section. + +# Guidelines: + +- The answer NA means that the paper has no limitation while the answer No means that the paper has limitations, but those are not discussed in the paper. +- The authors are encouraged to create a separate "Limitations" section in their paper. +- The paper should point out any strong assumptions and how robust the results are to violations of these assumptions (e.g., independence assumptions, noiseless settings, model well-specification, asymptotic approximations only holding locally). The authors should reflect on how these assumptions might be violated in practice and what the implications would be. +- The authors should reflect on the scope of the claims made, e.g., if the approach was only tested on a few datasets or with a few runs. In general, empirical results often depend on implicit assumptions, which should be articulated. +- The authors should reflect on the factors that influence the performance of the approach. For example, a facial recognition algorithm may perform poorly when image resolution is low or images are taken in low lighting. Or a speech-to-text system might not be used reliably to provide closed captions for online lectures because it fails to handle technical jargon. +- The authors should discuss the computational efficiency of the proposed algorithms and how they scale with dataset size. +- If applicable, the authors should discuss possible limitations of their approach to address problems of privacy and fairness. +- While the authors might fear that complete honesty about limitations might be used by reviewers as grounds for rejection, a worse outcome might be that reviewers discover limitations that aren't acknowledged in the paper. The authors should use their best judgment and recognize that individual actions in favor of transparency play an important role in developing norms that preserve the integrity of the community. Reviewers will be specifically instructed to not penalize honesty concerning limitations. + +# 3. Theory assumptions and proofs + +Question: For each theoretical result, does the paper provide the full set of assumptions and a complete (and correct) proof? + +Answer: [Yes] + +Justification: All theoretical results are rigorously proven either in the main text or in the supplementary material, and justifications of the result's hypotheses are given. + +# Guidelines: + +- The answer NA means that the paper does not include theoretical results. +- All the theorems, formulas, and proofs in the paper should be numbered and cross-referenced. +- All assumptions should be clearly stated or referenced in the statement of any theorems. +- The proofs can either appear in the main paper or the supplemental material, but if they appear in the supplemental material, the authors are encouraged to provide a short proof sketch to provide intuition. +- Inversely, any informal proof provided in the core of the paper should be complemented by formal proofs provided in appendix or supplemental material. +- Theorems and Lemmas that the proof relies upon should be properly referenced. + +# 4. Experimental result reproducibility + +Question: Does the paper fully disclose all the information needed to reproduce the main experimental results of the paper to the extent that it affects the main claims and/or conclusions of the paper (regardless of whether the code and data are provided or not)? + +Answer: [Yes] + +Justification: All implementation details, including hyperparameters, used libraries and experimental details, are included either in the main text or in the supplementary material. We plan to release the all code upon acceptance. + +# Guidelines: + +- The answer NA means that the paper does not include experiments. +- If the paper includes experiments, a No answer to this question will not be perceived well by the reviewers: Making the paper reproducible is important, regardless of whether the code and data are provided or not. +- If the contribution is a dataset and/or model, the authors should describe the steps taken to make their results reproducible or verifiable. +- Depending on the contribution, reproducibility can be accomplished in various ways. For example, if the contribution is a novel architecture, describing the architecture fully might suffice, or if the contribution is a specific model and empirical evaluation, it may be necessary to either make it possible for others to replicate the model with the same dataset, or provide access to the model. In general, releasing code and data is often one good way to accomplish this, but reproducibility can also be provided via detailed instructions for how to replicate the results, access to a hosted model (e.g., in the case of a large language model), releasing of a model checkpoint, or other means that are appropriate to the research performed. +- While NeurIPS does not require releasing code, the conference does require all submissions to provide some reasonable avenue for reproducibility, which may depend on the nature of the contribution. For example +(a) If the contribution is primarily a new algorithm, the paper should make it clear how to reproduce that algorithm. +(b) If the contribution is primarily a new model architecture, the paper should describe the architecture clearly and fully. +(c) If the contribution is a new model (e.g., a large language model), then there should either be a way to access this model for reproducing the results or a way to reproduce the model (e.g., with an open-source dataset or instructions for how to construct the dataset). +(d) We recognize that reproducibility may be tricky in some cases, in which case authors are welcome to describe the particular way they provide for reproducibility. In the case of closed-source models, it may be that access to the model is limited in some way (e.g., to registered users), but it should be possible for other researchers to have some path to reproducing or verifying the results. + +# 5. Open access to data and code + +Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material? + +Answer: [Yes] + +Justification: Our full code and details to reproduce our results can be found in the official repository. + +# Guidelines: + +- The answer NA means that paper does not include experiments requiring code. +- Please see the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details. +- While we encourage the release of code and data, we understand that this might not be possible, so "No" is an acceptable answer. Papers cannot be rejected simply for not including code, unless this is central to the contribution (e.g., for a new open-source benchmark). +- The instructions should contain the exact command and environment needed to run to reproduce the results. See the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details. +- The authors should provide instructions on data access and preparation, including how to access the raw data, preprocessed data, intermediate data, and generated data, etc. +- The authors should provide scripts to reproduce all experimental results for the new proposed method and baselines. If only a subset of experiments are reproducible, they should state which ones are omitted from the script and why. + +- At submission time, to preserve anonymity, the authors should release anonymized versions (if applicable). +- Providing as much information as possible in supplemental material (appended to the paper) is recommended, but including URLs to data and code is permitted. + +# 6. Experimental setting/details + +Question: Does the paper specify all the training and test details (e.g., data splits, hyperparameters, how they were chosen, type of optimizer, etc.) necessary to understand the results? + +Answer: [Yes] + +Justification: We use and disclose publicly available splits of known datasets, and specify all training hyperparameters either in the main text or the supplementary material. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The experimental setting should be presented in the core of the paper to a level of detail that is necessary to appreciate the results and make sense of them. +- The full details can be provided either with the code, in appendix, or as supplemental material. + +# 7. Experiment statistical significance + +Question: Does the paper report error bars suitably and correctly defined or other appropriate information about the statistical significance of the experiments? + +Answer: [Yes] + +Justification: We disclose the amount of seeds used in each experiments, and show standard error for the results in all figures and tables. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The authors should answer "Yes" if the results are accompanied by error bars, confidence intervals, or statistical significance tests, at least for the experiments that support the main claims of the paper. +- The factors of variability that the error bars are capturing should be clearly stated (for example, train/test split, initialization, random drawing of some parameter, or overall run with given experimental conditions). +- The method for calculating the error bars should be explained (closed form formula, call to a library function, bootstrap, etc.) +- The assumptions made should be given (e.g., Normally distributed errors). +- It should be clear whether the error bar is the standard deviation or the standard error of the mean. +- It is OK to report 1-sigma error bars, but one should state it. The authors should preferably report a 2-sigma error bar than state that they have a $96\%$ CI, if the hypothesis of Normality of errors is not verified. +- For asymmetric distributions, the authors should be careful not to show in tables or figures symmetric error bars that would yield results that are out of range (e.g. negative error rates). +- If error bars are reported in tables or plots, The authors should explain in the text how they were calculated and reference the corresponding figures or tables in the text. + +# 8. Experiments compute resources + +Question: For each experiment, does the paper provide sufficient information on the computer resources (type of compute workers, memory, time of execution) needed to reproduce the experiments? + +Answer: [Yes] + +Justification: We provide details of the computational resources used for this paper in the Appendix. + +# Guidelines: + +- The answer NA means that the paper does not include experiments. +- The paper should indicate the type of compute workers CPU or GPU, internal cluster, or cloud provider, including relevant memory and storage. +- The paper should provide the amount of compute required for each of the individual experimental runs as well as estimate the total compute. +- The paper should disclose whether the full research project required more compute than the experiments reported in the paper (e.g., preliminary or failed experiments that didn't make it into the paper). + +# 9. Code of ethics + +Question: Does the research conducted in the paper conform, in every respect, with the NeurIPS Code of Ethics https://neurips.cc/public/EthicsGuidelines? + +Answer: [Yes] + +Justification: We have reviewed the code of ethics and have confirmed that our research complies with every applicable requirement. + +Guidelines: + +- The answer NA means that the authors have not reviewed the NeurIPS Code of Ethics. +- If the authors answer No, they should explain the special circumstances that require a deviation from the Code of Ethics. +- The authors should make sure to preserve anonymity (e.g., if there is a special consideration due to laws or regulations in their jurisdiction). + +# 10. Broader impacts + +Question: Does the paper discuss both potential positive societal impacts and negative societal impacts of the work performed? + +Answer: [NA] + +Justification: We do not foresee any significant societal impact, as the paper explores principles for compositional learning and not its applications to society. + +Guidelines: + +- The answer NA means that there is no societal impact of the work performed. +- If the authors answer NA or No, they should explain why their work has no societal impact or why the paper does not address societal impact. +- Examples of negative societal impacts include potential malicious or unintended uses (e.g., disinformation, generating fake profiles, surveillance), fairness considerations (e.g., deployment of technologies that could make decisions that unfairly impact specific groups), privacy considerations, and security considerations. +- The conference expects that many papers will be foundational research and not tied to particular applications, let alone deployments. However, if there is a direct path to any negative applications, the authors should point it out. For example, it is legitimate to point out that an improvement in the quality of generative models could be used to generate deepfakes for disinformation. On the other hand, it is not needed to point out that a generic algorithm for optimizing neural networks could enable people to train models that generate Deepfakes faster. +- The authors should consider possible harms that could arise when the technology is being used as intended and functioning correctly, harms that could arise when the technology is being used as intended but gives incorrect results, and harms following from (intentional or unintentional) misuse of the technology. +- If there are negative societal impacts, the authors could also discuss possible mitigation strategies (e.g., gated release of models, providing defenses in addition to attacks, mechanisms for monitoring misuse, mechanisms to monitor how a system learns from feedback over time, improving the efficiency and accessibility of ML). + +# 11. Safeguards + +Question: Does the paper describe safeguards that have been put in place for responsible release of data or models that have a high risk for misuse (e.g., pretrained language models, image generators, or scraped datasets)? + +Answer: [NA] + +Justification: The paper does not involve models with high risk of misuse, as it explores simplified models for the sake of finding new principles for enhancing compositionality of representations based on an iterated communication game setup. + +Guidelines: + +- The answer NA means that the paper poses no such risks. +- Released models that have a high risk for misuse or dual-use should be released with necessary safeguards to allow for controlled use of the model, for example by requiring that users adhere to usage guidelines or restrictions to access the model or implementing safety filters. +- Datasets that have been scraped from the Internet could pose safety risks. The authors should describe how they avoided releasing unsafe images. +- We recognize that providing effective safeguards is challenging, and many papers do not require this, but we encourage authors to take this into account and make a best faith effort. + +# 12. Licenses for existing assets + +Question: Are the creators or original owners of assets (e.g., code, data, models), used in the paper, properly credited and are the license and terms of use explicitly mentioned and properly respected? + +Answer: [Yes] + +Justification: All assets are properly credited and used in accordance with their licenses: Shapes3D and MPI3D are under CC-BY 4.0, and the EGG framework is released under the MIT license. + +Guidelines: + +- The answer NA means that the paper does not use existing assets. +- The authors should cite the original paper that produced the code package or dataset. +- The authors should state which version of the asset is used and, if possible, include a URL. +- The name of the license (e.g., CC-BY 4.0) should be included for each asset. +- For scraped data from a particular source (e.g., website), the copyright and terms of service of that source should be provided. +- If assets are released, the license, copyright information, and terms of use in the package should be provided. For popular datasets, paperswithcode.com/datasets has curated licenses for some datasets. Their licensing guide can help determine the license of a dataset. +- For existing datasets that are re-packaged, both the original license and the license of the derived asset (if it has changed) should be provided. +- If this information is not available online, the authors are encouraged to reach out to the asset's creators. + +# 13. New assets + +Question: Are new assets introduced in the paper well documented and is the documentation provided alongside the assets? + +Answer: [NA] + +Justification: The paper does not release new assets. + +Guidelines: + +- The answer NA means that the paper does not release new assets. +- Researchers should communicate the details of the dataset/code/model as part of their submissions via structured templates. This includes details about training, license, limitations, etc. + +- The paper should discuss whether and how consent was obtained from people whose asset is used. +- At submission time, remember to anonymize your assets (if applicable). You can either create an anonymized URL or include an anonymized zip file. + +# 14. Crowdsourcing and research with human subjects + +Question: For crowdsourcing experiments and research with human subjects, does the paper include the full text of instructions given to participants and screenshots, if applicable, as well as details about compensation (if any)? + +Answer: [NA] + +Justification: The paper does not use crowdsourcing experiments or experiments with human subjects. + +Guidelines: + +- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. +- Including this information in the supplemental material is fine, but if the main contribution of the paper involves human subjects, then as much detail as possible should be included in the main paper. +- According to the NeurIPS Code of Ethics, workers involved in data collection, curation, or other labor should be paid at least the minimum wage in the country of the data collector. + +# 15. Institutional review board (IRB) approvals or equivalent for research with human subjects + +Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals (or an equivalent approval/review based on the requirements of your country or institution) were obtained? + +Answer: [NA] + +Justification: The paper does not involve research with human subjects. + +Guidelines: + +- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. +- Depending on the country in which research is conducted, IRB approval (or equivalent) may be required for any human subjects research. If you obtained IRB approval, you should clearly state this in the paper. +We recognize that the procedures for this may vary significantly between institutions and locations, and we expect authors to adhere to the NeurIPS Code of Ethics and the guidelines for their institution. +- For initial submissions, do not include any information that would break anonymity (if applicable), such as the institution conducting the review. + +# 16. Declaration of LLM usage + +Question: Does the paper describe the usage of LLMs if it is an important, original, or non-standard component of the core methods in this research? Note that if the LLM is used only for writing, editing, or formatting purposes and does not impact the core methodology, scientific rigorousness, or originality of the research, declaration is not required. + +Answer: [NA] + +Justification: The paper's results were not based on, or aided by, research using LLMs. + +Guidelines: + +- The answer NA means that the core method development in this research does not involve LLMs as any important, original, or non-standard components. +- Please refer to our LLM policy (https://neurips.cc/Conferences/2025/LLM) for what should or should not be described. + +# A Background + +# A.1 Iterated Learning + +The IL model of cultural evolution [34] proposes to emulate the emergence of language through the interactions of adult and new learning agents. In the original formulation, a space of pairs of signal + meaning is randomly generated, and in each iteration, a learning agent is partially exposed to the language learned by the adult agent in the previous iteration. The learning agent becomes the adult agent for the next iteration, and transmits a subset of its learned "language". By iterating this process, the initial random language gains structure and observable properties, such as compositionality. + +This framework has been widely extended to include an additional task [44, 56, 75], instead of simply reshaping the language. Under this formulation, the training regime can be separated into two parts: Interaction (Interaction), where an agent is optimized to solve the additional task, and Imitation (Imitation), where a newly initialized agent "learns" from the previous agent through a shared learned language. + +# A.2 Lewis reconstruction game + +The Lewis reconstruction game (LRG) [58] is a special case of the Lewis Signaling Game (LSG) [43] framework, in which two agents cooperate to reconstruct an observed object. Specifically, a sender agent parameterized by $\varphi$ observes an object $x$ from an object space $\mathcal{X}$ and produces a message $m\in \mathcal{M}$ , where $\mathcal{M} = V^{\star}$ is the set of all possible strings formed by concatenating symbols drawn from a finite vocabulary $V$ . A receiver agent parameterized by $\omega$ observes $m$ and produces a reconstruction $x^{\prime}$ . Both models are optimized to minimize a reconstruction loss $\mathcal{L}_R(x,x')$ . + +Importantly, in the original LSG, the receiver is tasked with predicting $x$ from a finite set of states, whereas in this formulation, the problem becomes a regression task, and $\mathcal{X}$ can be a continuous space. + +# B Implementation Details & Hyperparameters + +# B.1 Baseline Model + +- Sender and Receiver: Both consist of a single-layer LSTM with an embedding size of 64 and a hidden size of 256. Outputs are passed through a two-layer MLP with ReLU activations. +- Communication Channel: The vocabulary size is set to $|V| = 15$ and messages are composed of $C = 10$ discrete symbols. These values are chosen such that the amount of possible messages $(5.76 \times 10^{12})$ greatly outnumber the amount of states. The Gumbel-Softmax bottleneck is used without straight-through estimation. +- VAE Backbone: The VAE is implemented using the disentanglement_lib [47] python library, using the default setting of two convolutional layers of 32 filters, with kernel size 4 and stride 2, two convolutional layers of 64 filters, with kernel size 2 and stride 2, each with ReLU activation, as well as a linear layer of size 256 and ReLU activation and an output layer with size $2^{*}128$ for concatenated $\mu s$ and $\log \sigma s$ . The encoder is mirrored replacing convolutional layers with transposed convolutions, and an input layer of size 128 and ReLU activations on all layers save the output. +- VAE Training: The VAE is pretrained for 15 epochs using the Adam optimizer with a learning rate of $1 \times 10^{-3}$ . + +The interaction-imitation games were run for a maximum of 100 iterations, alternating one full epoch in each phase per iteration. + +We used early stopping based on the validation MSE of the final reconstruction, starting from epoch 5, with a minimum $\delta$ of $1\times 10^{-3}$ and patience of 5. We used the Adam optimizer with default parameters and learning rates of $1\times 10^{-3}$ and $1\times 10^{-4}$ for Shapes3D [6] and MPI3D [23], respectively, as we found the training for the latter became unstable at higher learning rates. + +We found that Interaction required smaller batch sizes to converge, whereas the Imitation phase PDM becomes a better entropy approximation at higher batch sizes, however scaling quadratically + +in computation. Therefore, we set the Interaction and Imitation batch sizes to 256 and 512, respectively. + +To implement Imitation we defined a receiver $R_{\omega}$ and two senders $S_{\phi_1}$ and $S_{\phi_2}$ with their own optimizers. After participating in an Interaction iteration, $\phi^1$ and $\omega$ are frozen and used to train $S_{\phi_2}$ as described in §3.2. After Imitation, the weights $\phi_2$ are copied to $\phi_1$ , and $\phi_2$ is reset. Importantly, the optimizers for $\phi_1$ and $\phi_2$ are never reset nor copied between them. + +# C Compute Resources + +All experiments were conducted on our internal laboratory cluster, using NVIDIA A40 GPUs with 48 GB of VRAM. Each job was allocated a single GPU with 8 GB of VRAM usage on average, alongside access to a CPU with 20 cores, 118 GB of RAM, and local SSD storage for datasets and model checkpoints. + +The experiments reported in this paper comprise 360 training runs, totaling approximately 180 GPU hours. Individual runs varied in length from 15 to 50 minutes, depending on early stopping criteria. These runs represent the finalized experiments whose results are presented in the main text. + +In total, the research project required 2,310 experimental runs over the course of development, amounting to approximately 245 GPU days (5,880 GPU hours) of compute time. This includes preliminary experiments, ablations, and failed runs that were instrumental to model and protocol design but are not individually reported in the paper. + +# C.1 Useful length $\epsilon$ + +To define the useful length $\epsilon$ for both datasets, we observed reconstruction loss distribution over all positions in the message (see 3) for different $\lambda$ values, and estimated an average plateau point. For MPI3D we used a more conservative and precise threshold, as the reconstruction error varied more between methods. + +Additionally, we analyzed the useful length graph for multiple $\epsilon$ values, and discarded those values where most models converged to either maximum or minimum possible length (see 4). + +# D Recovering compositional structure from partially observed data + +For the sake of clarity of treatment, we assume here that $\mathbf{G} = [G_1,\dots ,G_n]$ in whichi factor $G_{i}$ has the same number $N$ of possible values, and as a consequence $\mathcal{G} = [1:N]^n$ . Our results will extend directly to the general $\mathcal{G}$ case as shown in Rmk. D.4. + +We work under the simplified deterministic generation hypotheses, in which our dataset is $\mathcal{D} = \operatorname{GenX}(\mathcal{G}) \subset \mathcal{X}$ in which $\operatorname{GenX}$ is a deterministic injective function. + +Our main result is the following: + +Theorem D.1. Assume that $\mathcal{G} = [1:N]^n$ and that $\mathsf{GenX}:\mathcal{G}\to \mathcal{X}$ is an injective function, and that $\mathcal{D}_{train}\subseteq \mathcal{D}$ has cardinality $|\mathcal{D}_{train}| = p|\mathcal{D}|$ for some $p\in (0,1)$ such that $pN^n$ is an integer. + +We assume that the compositional structure of $\mathcal{G}$ is such that the following holds: + +(A) For some $k < n$ , if we observe a set of data whose generating factor combinations feature all possible combinations of generating factor values $(G_{i_1},\ldots ,G_{i_k})$ for all choices of indices $1\leq i_1 < \dots < i_k\leq n$ then this is sufficient to reconstruct $\mathcal{G}$ to good accuracy. + +Then as $|\mathcal{G}| \to \infty$ the probability that $\mathcal{D}_{train}$ is sufficient to reconstruct $\mathcal{G}$ tends to 1. + +Here (A) makes rigorous the common assumption that having a training set that features examples for diverse enough combinations of factors is sufficient for reconstructing all combinatorial generators $\mathbf{G}$ . + +Note that in this work we do not discuss + +1. what features of the pre-trained encodings $\mathcal{G}\rightarrow \mathcal{D}$ and what requirements on the compositional structure of $\mathcal{G}$ would allow to actually guarantee this assumption, + +![](images/6d1e1035a8a60862f625bc394cbe0df81ae8ec9823a18f87f228bf15c5d652c5.jpg) + +![](images/a52ffe9ed3bc0cd9801eed257f5f327f186992af9029890681f26b5eb6ea50a6.jpg) +Figure 3: Loss distribution over positions and $\lambda$ values for (a) Shapes3D and (b) MPI3D + +![](images/9ffa5ce6b0ff2b807cc60d3b1a5a36f6d343e97521e79aff9349588440f7850a.jpg) +Figure 4: Useful length values for different $\epsilon$ and $\lambda$ values for (a) Shapes3D and (b) MPI3D + +![](images/174b295339bd4f66508d0bb75b1460ca351406b6a9b8d14cc3936ca79b86b94e.jpg) + +2. alternatives for a rigorous definition of what is meant by "to good accuracy" in the statement of assumption (A). + +Once the above points are settled and defined, we can replace (A) by an explicit requirement. Since however these points seem like an ambitious whole research line and are far beyond the scope of this work, we leave the definition of "to good accuracy" and the proof of reconstruction to future works. + +Thus, introducing (A) allows us to prove fully rigorous statements such as Thm. D.3, in absence of a full theory of compositional reconstruction. + +Nevertheless, we make the following remarks: + +- There is a natural trade-off between assumptions on how big $\mathcal{D}_{train}$ should be, vs. how complicated the compositional reconstruction should be allowed to be for reconstruction to be feasible: the richer set of such combinations we require $\mathcal{D}_{train}$ to contain (thus making more restrictive assumptions on $\mathcal{D}_{train}$ ), the lower is the requirement for our encoding framework and for the compositional structure for (A) to hold. +- It seems natural to conjecture that if assumption (A) does not hold for large value of $k$ , then there is no hope for the compositional reconstruction to be achievable. Thus lowering the requirement on "good enough accuracy" to make (A) hold seems the only avenue of future research. + +Theorem D.1 is a consequence of below Theorem D.3 that proves that assumption (A) holds with probability tending to 1 as $|\mathcal{G}| \to \infty$ and $|\mathcal{D}_{train}| / |\mathcal{G}| = p$ is bounded away from zero. What Theorem D.3 says is that for large sets of generators, it will become overwhelmingly unlikely that $\mathcal{D}_{train}$ does not contain enough data to reconstruct the generators. + +# D.1 Observing $\mathcal{D}_{train}$ allows to observe finite statistics of the factors with high probability + +Our first result for this section is that when $\mathcal{D}_{train} \subseteq \mathcal{D}$ is chosen uniformly at random amongst sets of size $p|\mathcal{D}|$ , for $|\mathcal{D}| = |\mathcal{G}|$ large enough, we find that with overwhelming probability all possible values of each of the factors $G_{1},\ldots,G_{n}$ are achieved by some elements of $\mathcal{D}_{train}$ . We state the result in Proposition D.2, whose proof in our view contains the main ideas useful also for the full result of Theorem D.3, which can be considered our main technical result for this section. + +The first result is as follows: + +Proposition D.2. Assume that for $n, N \geq 2$ we have $\mathcal{G} = \{\mathbf{G} = [G_1, \ldots, G_n] : (\forall i \leq n) 1 \leq G_i \leq N\} = [1:N]^n$ and that $\mathsf{GenX} : \mathcal{G} \to \mathcal{X}$ is an injective function with image $\mathcal{D}$ as above. Denote by $\mathcal{G}_{train} := \mathsf{GenX}^{-1}(\mathcal{D}_{train})$ the generating factors related to the training set, and assume that $\mathcal{D}_{train} \subset \mathcal{D}$ is chosen uniformly at random amongst subsets of cardinality $p|\mathcal{D}|$ , for some $p \in (0,1)$ such that $pN^n$ is an integer. Then + +$$ +\mathbb {P} \left\{\left(\exists i \leq n\right) (\exists j \leq N) \left(\forall \mathbf {G} = \left[ G _ {1}, \dots , G _ {n} \right] \in \mathcal {G} _ {\text {t r a i n}}\right) G _ {i} \neq j \right\} \leq n N \left(1 - \frac {1}{N}\right) ^ {p N ^ {n}}, \tag {4} +$$ + +in particular at fixed $p$ , the above event has probability tending to zero as $N^n \to \infty$ . + +Thus, in an equivalent formulation of the last part of the above proposition, we have that in the limit of large $|\mathcal{G}| = N^n$ , the probability that each generating factor $G_{i}$ takes all of its possible values on some generated element of $\mathcal{G}_{train}$ tends to 1. + +Proof. The possible choices of $\mathcal{G}_{train} \subseteq \mathcal{G}$ of cardinality $p|\mathcal{G}| = pN^n$ are $\frac{(N^n)!}{(pN^n)!((1 - p)N^n)!}$ . Then, calling $A_{ij}$ the set of elements $\mathbf{G} \in \mathcal{G}_{train}$ such that $G_i \neq j$ , we find that $|A_{ij}| = N^{n-1}(N-1) = N^n \frac{N-1}{N}$ , so that the number of possible choices of $\mathcal{G}_{train} \subseteq A_{ij}$ of cardinality $pN^n$ is given by $\frac{(N^n \frac{N-1}{N})!}{(pN^n)! (N^n (\frac{N-1}{N} - p))!}$ . Now we bound the probability of the event on the left hand side in (4) via a union bound over the $Nn$ possible choices of $A_{ij}$ for $1 \leq i \leq n, 1 \leq j \leq N$ , so that + +$$ +\begin{array}{l} \mathbb {P} \left\{ \begin{array}{l} \exists i \leq n, \exists j \leq N, \\ \forall \mathbf {G} = [ G _ {1}, \ldots , G _ {n} ] \in \mathcal {G} _ {t r a i n}, \\ G _ {i} \neq j \end{array} \right\} \leq n N \frac {(p N ^ {n}) ! ((1 - p) N ^ {n}) !}{(N ^ {n}) !} \frac {(N ^ {n} \frac {N - 1}{N}) !}{(p N ^ {n}) ! (N ^ {n} (\frac {N - 1}{N} - p)) !} \\ = n N \frac {N ^ {n} \frac {N - 1}{N}}{N ^ {n}} \frac {N ^ {n} \frac {N - 1}{N} - 1}{N ^ {n} - 1} \dots \frac {N ^ {n} \frac {N - 1}{N} - p N ^ {n} + 1}{N ^ {n} - p N ^ {n} + 1} \\ \leq n N \left(1 - \frac {1}{N}\right) ^ {p N ^ {n}}, \\ \end{array} +$$ + +in which in the first line we used a union bound together with the computations of combinatorial coefficients from above, in the second line we performed a simplification and reordering, and in the + +third line we used the fact that $1 - 1 / N = (N^{n}(N - 1) / N) / N^{n} > (N^{n}(N - 1) / N - i) / (N^{n} - i)$ for all $1\leq i < pN^n -1$ . + +Finally, considering the limit $N^n \to \infty$ of the upper bound in (4), we have two cases: (1) if $N \geq 2$ stays bounded and $n \to \infty$ , then we can bound $nN(1 - 1 / N)^{pN^n} \leq C_1nC_2^{C_3^n}$ with constants $C_2 \in (0,1)$ and $C_3 > 1$ , and by taking logarithms we check that the quantity tends to zero; (2) if $N \to \infty$ and $n \geq 2$ , and in this case $(1 - 1 / N)^{pN^n} \sim e^{-pN^{n-1}}$ and again by taking logarithms we find that the quantity tends to zero. + +In fact Prop. D.2 is gives us guarantees that it will be possible to observe all options for each $G_{i}$ in the training set, however this may not be sufficient for the main aim of recovering the compositional structure of $\mathcal{G}$ . However we next prove, in an extension of Prop. D.2, that actually we will also observe with high probability any possible combination of finitely many factors: + +Theorem D.3. Under the same hypotheses on $\mathcal{G}$ , $\mathsf{GenX}, \mathcal{G}_{train}, p$ as in Prop. D.2, for any $k \leq n$ we have the following. + +$$ +\begin{array}{l} \mathbb {P} \left\{ \begin{array}{l l} \exists i _ {1} < \dots < i _ {k} \leq n \\ \exists j _ {1}, \ldots , j _ {k} \leq N & (G _ {i _ {1}}, \ldots , G _ {i _ {k}}) \neq (j _ {1}, \ldots , j _ {k}) \\ \forall \mathbf {G} = [ G _ {1}, \ldots , G _ {n} ] \in \mathcal {G} _ {t r a i n} & \end{array} \right\} \\ \leq \frac {n !}{k ! (n - k) !} N ^ {k} \left(1 - \frac {1}{N ^ {k}}\right) ^ {p N ^ {n}}, \tag {5} \\ \end{array} +$$ + +which for fixed $p, k$ tends to zero as $N^n \to \infty$ . + +Proof. As mentioned before, roughly the overall proof strategy is the same as for Prop. D.2. We introduce the sets $A_{i_1,\dots ,i_k}^{j_1,\dots ,j_k} = A_{\mathbf{i}}^{\mathbf{j}}$ and observe that the number of choices of $i_1 < \dots < i_k \leq n$ is $\frac{n!}{k!(n - k)!}$ while the number of choices of $j_1,\ldots ,j_k \leq N$ is $N^k$ . Thus the number of sets of the form $A_{\mathbf{i}}^{\mathbf{j}}$ is $N_A := \frac{N^kn!}{k!(n - k)!}$ . Furthermore, each set $A_{\mathbf{i}}^{\mathbf{j}}$ only imposes the constraint of entries of indices $i_1,\ldots ,i_k$ from a $\mathbb{G}$ to avoid one combination of values $\vec{j}$ , and thus it has cardinality $|A_{\mathbf{i}}^{\mathbf{j}}| = N^{n - k}(N^{k} - 1)$ . Then the number of choices of $\mathcal{G}_{train} \subseteq A_{\mathbf{i}}^{\mathbf{j}}$ of cardinality $pN^n$ is just the combinatorial coefficient $\frac{(N^{n - k}(N^{k} - 1))!}{(pN^{n})!(N^{n - k}(N^{k} - 1 - pN^{k}))!}$ and by union bound we then find by the same reasoning as in Prop. D.2, + +$$ +\begin{array}{l} \text {(l e f t - h a n d s i d e o f (5))} \leq N _ {A} \frac {\left(p N ^ {n}\right) ! \left(\left(1 - p\right) N ^ {n}\right) !}{\left(N ^ {n}\right) !} \frac {\left(N ^ {n - k} \left(N ^ {k} - 1\right)\right) !}{\left(p N ^ {n}\right) ! \left(N ^ {n - k} \left(N ^ {k} - 1 - p N ^ {k}\right)\right) !} \\ = N _ {A} \frac {N ^ {n} - N ^ {n - k}}{N ^ {n}} \frac {N ^ {n} - N ^ {n - k} - 1}{N ^ {n} - 1} \dots \frac {N ^ {n} - N ^ {n - k} - p N ^ {n} + 1}{N ^ {n} - p N ^ {n} + 1} \\ \leq N _ {A} \left(1 - \frac {1}{N ^ {k}}\right) ^ {p N ^ {n}} = \frac {n !}{k ! (n - k) !} N ^ {k} \left(1 - \frac {1}{N ^ {k}}\right) ^ {p N ^ {n}}. \\ \end{array} +$$ + +This proves (5). In order to show that the bound tends to zero as $N^n \to \infty$ at fixed $p > 0, k \leq n$ , we first note that $\frac{n!}{k!(n - k)!} \leq n^k / k!$ and thus (as $k$ is assumed fixed) it suffices to show that + +$$ +n ^ {k} N ^ {k} \left(1 - \frac {1}{N ^ {k}}\right) ^ {p N ^ {n}} \rightarrow 0 \quad \text {a s} \quad N ^ {n} \rightarrow \infty . +$$ + +Here again, we can proceed with the discussion of the two cases ( $N$ bounded and $n \to \infty$ , or $N \to \infty$ ) exactly as in the end of the proof of Prop. D.2, to conclude. + +Remark D.4 (Extension of Prop. D.2 and Theorems D.3 and D.1 to general $\mathcal{G}$ ). If, for a choice of $n\geq 2$ and $N_{1},\ldots ,N_{n}\geq 2$ we have $\mathcal{G} = \{\mathbf{G} = [G_1,\dots ,G_n]:(\forall i\leq n)1\leq G_i\leq N_i\}$ , then $|\mathcal{G}| = N_{1}\cdot \cdot \cdot N_{n}$ and by carefully following the same steps as in the proof of Prop. D.2 we find + +$$ +\mathbb {P} \left\{ \begin{array}{c} \exists i \leq n, \quad \exists j \leq N _ {i}, \\ \forall \mathbf {G} = \left[ G _ {1}, \dots , G _ {n} \right] \in \mathcal {G} _ {\text {t r a i n}}, \\ \text {i t h o l d s} G _ {i} \neq j \end{array} \right\} \leq \sum_ {i = 1} ^ {n} N _ {i} \left(1 - \frac {1}{N _ {i}}\right) ^ {p N _ {1} \dots N _ {n}}, \tag {6} +$$ + +Again, this tends to zero for fixed $p$ as $|\mathcal{G}| = N_1 \ldots N_n \to \infty$ : to prove it, we can proceed as follows. In a first case, if $n > 2$ stays bounded, then we can discuss separately for each $1 \leq i \leq n$ whether $N_i \to \infty$ or whether it stays bounded, as done in the proof of Prop. D.2, and in either case the corresponding summand on the left of (4) tends to zero, as in case (2) from the proof of Prop. D.2. Otherwise, $n \to \infty$ and we can proceed analogously to case (1) from the end of proof of Prop. D.2. + +For the extension of Thm. D.3 to general $\mathcal{G}$ we get the bound + +$$ +\begin{array}{l} \mathbb {P} \left\{ \begin{array}{l l} \exists i _ {1} < \dots < i _ {k} \leq n \\ \exists j _ {1}, \ldots , j _ {k} \leq N \\ \forall \mathbf {G} = [ G _ {1}, \ldots , G _ {n} ] \in \mathcal {G} _ {t r a i n} \end{array} \right. (G _ {i _ {1}}, \ldots , G _ {i _ {k}}) \neq (j _ {1}, \ldots , j _ {k}) \Bigg \} \\ \leq \sum_ {1 \leq i _ {1} < \dots < i _ {k} \leq n} N _ {i _ {1}} \dots N _ {i _ {k}} \left(1 - \frac {1}{N _ {i _ {1}} \cdots N _ {i _ {k}}}\right) ^ {p N _ {1} \dots N _ {n}}, \tag {7} \\ \end{array} +$$ + +and the strategy for its proof and for the proof that this goes to zero if $k, p$ fixed and $|\mathcal{G}| = N_1 \cdots N_n \to \infty$ are a direct extension of the above. + +Theorem D.1 is a direct consequence of assumption (A) and Theorem D.3, and thus it also directly extends to the general case. + +# E Proof that final state reconstruction allows for wider message choice freedom + +We first recall the setting. Consider two fixed maps + +$$ +R ^ {\star}: V ^ {C} \to \mathcal {X}, \quad S ^ {\star}: \mathcal {D} \subseteq \mathcal {X} \to V ^ {C}, +$$ + +and let $\mathcal{R}_{[i]}:V^C\to V^i$ be the restriction to the first $i$ tokens, i.e. in previous notation, for $m = (m_{1},\ldots ,m_{C})\in V^{C}$ we set $\mathcal{R}_{[i]}(m)\coloneqq m_{[i]} = (m_1,\dots,m_i)$ . + +For $R^{\star}, S^{\star}$ as above and $S: \mathcal{D} \subseteq \mathcal{X} \to V^{C}$ we set, for $x \in \mathcal{D}$ + +$$ +d _ {i} ^ {S} (x) := d _ {\mathcal {X}} (R ^ {\star} (S ^ {\star} (x) _ {[ i ]})), R ^ {\star} (S (x) _ {[ i ]})) = d _ {\mathcal {X}} (R ^ {\star} (\mathcal {R} _ {[ i ]} (S ^ {\star} (x))), R ^ {\star} (\mathcal {R} _ {[ i ]} (S (x)))) +$$ + +We also defined the following spaces: + +$$ +\mathcal {S} _ {f u l l} ^ {\epsilon} := \left\{S: \mathbb {E} _ {x} \left[ \frac {1}{C} \sum_ {i = 1} ^ {C} d _ {i} ^ {S} (x) \right] < \epsilon \right\}, \quad \mathcal {S} _ {f i n a l} ^ {\epsilon} := \left\{S: \mathbb {E} _ {x} \left[ d _ {C} ^ {S} (x) \right] < \epsilon \right\}. \tag {8} +$$ + +Our main result here is the following: + +Proposition E.1. With the above notations, we have the following + +1. For $i < C$ and any choice of $M, \epsilon > 0$ , the restrictions on $i$ -token sub-messages of $S(x)$ by requiring $S \in \mathcal{S}_{full}^{M}$ are more restrictive than those imposed by requiring $S \in \mathcal{S}_{final}^{\epsilon}$ , i.e. we have $\mathcal{R}_{[i]}(\mathcal{S}_{full}^{M}) \subset \mathcal{R}_{[i]}(\mathcal{S}_{final}^{\epsilon})$ . + +2. For the full message case $i = C$ in general we have $S_{full}^{\epsilon} \subseteq S_{final}^{C\epsilon}$ . If we further assume that $\mathbb{E}_x[d_j^S (x)]$ is non-increasing in $j$ then we have the stronger inclusion $S_{full}^{\epsilon} \subseteq S_{final}^{\epsilon}$ . + +Proof. Item 1 is direct, as no restriction on intermediate sub-messages $S(x)_{[i]} = \mathcal{R}_{[i]}(S(x))$ by the requirement $S \in S_{final}^{\epsilon}$ . + +For item 2, we observe that + +$$ +\frac {1}{C} d _ {C} ^ {S} (x) \leq \frac {1}{C} \sum_ {i = 1} ^ {C} d _ {i} ^ {S} (x), +$$ + +and by taking expectation we find that $\mathbb{E}_x\left[\frac{1}{C}\sum_i d_i^S (x)\right] < \epsilon$ implies $\mathbb{E}_x[d_C^S (x)]\leq C\epsilon$ , and thus $\mathcal{S}_{full}^{\epsilon}\subseteq \mathcal{S}_{final}^{C\epsilon}$ , which gives the first part of the statement. + +For the second part of item 2 we note that if $\mathbb{E}_x[d_i^S (x)]\geq \mathbb{E}_x[d_C^S (x)]$ for all $i\leq C$ we get + +$$ +\mathbb {E} _ {x} \left[ d _ {C} ^ {S} (x) \right] \leq \min _ {1 \leq j \leq C} \mathbb {E} _ {x} \left[ d _ {j} ^ {C} (x) \right] \leq \frac {1}{C} \sum_ {i = 1} ^ {C} \mathbb {E} _ {x} \left[ d _ {i} ^ {S} (x) \right] = \mathbb {E} _ {x} \left[ \frac {1}{C} \sum_ {i = 1} ^ {C} d _ {i} ^ {S} (x) \right], +$$ + +and thus $\mathcal{S}_{full}^{\epsilon} \subseteq \mathcal{S}_{final}^{\epsilon}$ as desired. + +![](images/c34456d1ff61dd73ed3bc0ed40fc4564d224ecf2aa661b4042d7b900f07abd4a.jpg) + +The freedom of exploration ensured by item 1 of the proposition is our main motivation: by only constraining the reconstruction of the full message, we allow much more freedom for the encoding strategy of $S(x)$ at earlier sub-messages at each iteration. + +Item 2 of Prop. E.1 shows that the two losses in the definition of $S_{full}$ , $S_{final}$ give comparable guarantees. At the beginning of iterations we will need to use the general bound with $\epsilon / C$ , and as $R^{\star}$ will have been trained on more and more messages during the iteration, the hypothesis of $\mathbb{E}_x[d_i^S(x)]$ non-increasing will become true allowing for the sharper bound with the same value of $\epsilon$ , as $R^{\star}$ (an LSTM taking as input successive message tokens) will become efficient in using sub-messages of pairs $(m, m') = (S(x), S^{\star}(x))$ to gradually distill information allowing to distinguish them for a large set of $S, x$ . + +# F Regularization to counteract Holistic Encodings + +In this section, we include rigorous results useful to justify the regularization of our Imitation phase, aimed at balancing compressibility and efficiency. + +This corresponds to the balance between compositional and parsimonious communication, a general theme in classical IL literature (see e.g. [35, 37]). It seems useful to specify more precise definitions in our generative setting, in order to help justify our architecture choices. + +Our aim in this section is threefold: + +1. In §F.1 we set rigorous definitions of Holistic Encodings within the setting of §2.1 and §2.2. +2. In §F.2 we check that compressibility is required for lowest Kolmogorov Complexity encodings. +3. We verify that optimizing for just expressivity and efficiency on our training set, will produce Holistic Encodings, and thus is not satisfactory for reconstructing $\mathcal{G}$ (see §F.3); +4. We give a few possible options on how to enrich the loss functions in our IL setup, aimed at establishing a better balance between efficiency and compressibility (see §F.4). + +# F.1 Definitions of Holistic and Compositional encodings, and role of the compressibility condition + +In the classical IL setup [35, 37], a holistic system of communication is defined as an encoding $\mathcal{G} \to \mathcal{M}$ that maps $\mathcal{G} \ni \mathbf{G} \mapsto \mathbf{G}' \in \mathcal{M}$ injectively but disregards the compositional structure of the "language" to be encoded, i.e. in our case, without respecting the structure of $\mathbf{G}$ . In our generation task formalization, a difference is that the relevant encoding map now factors through the generation map as $\mathcal{G} \xrightarrow{\mathrm{GenX}} \mathcal{X} \xrightarrow{S} \mathcal{M}$ . Now the generation map $\mathrm{GenX}$ is considered fixed, thus the definition refers to only $S$ and is the following: + +Definition F.1. In the setting from §2.1 and §2.2, sender $S: \mathcal{X} \to \mathcal{M}$ determines a holistic encoding if it satisfies expressivity but the messages associated to data generated as $\mathrm{GenX}(\mathbf{G})$ do not have a compositional structure in terms of generating factors $\mathbf{G}$ , as specified by Definition F.2 below. + +Note that, as in [35, 37], Definition F.1 of holistic encodings refers to "respecting compositional structure", which requires another definition. The definition of compositional encodings crucially depends on how complex we need/want to assume $\mathbf{G}$ to be. In this paper, for the sake of concreteness, we restricted to the simple case of $\mathbf{G}$ being formed as an element of a product space of finite cardinality, i.e., an $n$ -ple of independent factors $\mathbf{G} = [G_1,\dots ,G_n]$ in which each $G_{i}$ can take finitely many values. We thus use a straightforward toy definition of compositional encodings, requiring that each of the $G_{i}$ is separately encoded as a string and these strings are concatenated together to give the message that encodes $\mathbf{G}$ : + +Definition F.2. In the setting of §2.1 and §2.2, assume that generating factors $\mathbf{G} = [G_1,\dots ,G_n]$ have $G_{i}\in [1:N_{i}]$ for some $N_{1},\ldots ,N_{n}\geq 2$ . Then sender $S:\mathcal{X}\to \mathcal{M} = V^{C}$ will be defined to respect the structure of generating factors $\mathbf{G}$ if for each $1\leq i\leq n$ there exist an injective mapping $E_{i}:[1:N_{i}]\rightarrow V^{C_{i}}$ , for which + +$$ +\forall \mathbf {G} \in \mathcal {G}, S (\operatorname {G e n X} (\mathbf {G})) = E _ {1} (G _ {1}) \dots E _ {n} (G _ {n}) \quad \text {a n d} \quad \sum_ {i = 1} ^ {n} C _ {i} = C. +$$ + +The compressibility condition on the sender-receiver protocol just requires that the protocol be not holistic, and thus it is fully specified via definitions F.1 and F.2. Furthermore, Definition F.2 fully specifies in which way the map $\mathcal{G} \to \mathcal{M}$ successfully approximates an isomorphic reconstruction of $\mathcal{G}$ , if it satisfies conditions compressibility and expressivity. + +# F.2 Encodings optimizing for only expressivity and efficiency have Kolmogorov Complexity much higher than compositional ones + +By the classical learning theory principle, if our sender generates encodings with lower complexity, these should be easier to learn by the receiver, in the sense that the receiver will tend to generalize with more ease form good accuracy on the training set to similar accuracy on the test set. + +As a manageable measure of complexity we will use Kolmogorov Complexity (KC), and we verify that in our setting, encodings satisfying expressivity and efficiency will necessarily have higher KC than compositional ones. In particular, the receiver will have more trouble reconstructing from such encodings, justifying our push to enforce the compressibility condition. + +Encodings of $\mathcal{D}_{train}$ satisfying expressivity and efficiency include all injective maps + +$$ +\operatorname {e n c}: \mathcal {D} _ {\text {t r a i n}} \rightarrow V ^ {C _ {\star}}, \quad C _ {\star} = \lceil \log_ {| V |} \left| \mathcal {D} _ {\text {t r a i n}} \right| \rceil , \tag {9} +$$ + +in which $\lceil a\rceil$ denotes the smallest integer larger than real number $a$ . For a random map enc as above, the encoding to be Kolmogorov-irreducible, so that the expected KC of a random enc would be up to constant factor the one corresponding to explicitly enumerating the $|\mathcal{D}_{train}|$ strings of length $C_{\star}$ in alphabet $V$ that describe the encodings of each element of the training set. Each such string thus requires $\approx C_{\star}\log_2|V|$ bits. Thus if $U$ is the uniform distribution over injective maps (9) then we have + +$$ +\mathbb {E} _ {\text {e n c} \sim U} [ K C (\text {e n c}) ] \approx | \mathcal {D} _ {\text {t r a i n}} | C _ {\star} \log_ {2} | V | \approx | \mathcal {D} _ {\text {t r a i n}} | \log_ {2} | \mathcal {D} _ {\text {t r a i n}} |. \tag {10} +$$ + +On the other hand, to specify a compositional encoding comp as in Definition F.2 requires only assigning a string of length $C_i$ in alphabet $V$ for every generating factor $G_i$ , where we can take $C_i = \lceil \log_{|V|} N_i \rceil$ . Thus such string requires $\approx \log_2 N_i$ bits, and the whole encoding has KC given by + +$$ +K C (\operatorname {c o m p}) \approx \sum_ {i = 1} ^ {n} \log_ {2} N _ {i} = \log_ {2} | \mathcal {D} _ {\text {t r a i n}} |. \tag {11} +$$ + +If we assume that $|\mathcal{D}_{train}| / |\mathcal{D}| = p \in (0,1)$ and we take $|\mathcal{D}| \to \infty$ , then we get + +$$ +\frac {\text {r . h . s . o f (1 0)}}{\text {r . h . s . o f (1 1)}} = \frac {p | \mathcal {D} | \left(\log_ {2} p + \log_ {2} | \mathcal {D} |\right)}{\log_ {2} | \mathcal {D} |} \sim p | \mathcal {D} | \quad \text {a s} | \mathcal {D} | \rightarrow \infty . \tag {12} +$$ + +The computation (12) shows that + +Proposition F.3. Under the same hypotheses on $\mathcal{G}$ as in the previous section, consider the limit $|\mathcal{D}| \to \infty$ and assume that $p = |\mathcal{D}_{train}| / |\mathcal{D}|$ stays bounded away from zero in the limit. Then the expected Kolmogorov complexity of encodings satisfying expressivity and efficiency becomes overwhelmingly larger than the one of compositional encodings: + +$$ +\frac {\mathbb {E} _ {\text {e n c} \sim U} [ K C (\text {e n c}) ]}{K C (\text {c o m p})} \sim p | \mathcal {D} | \rightarrow \infty . +$$ + +# F.3 Encodings that optimize efficiency under expressivity over $\mathcal{D}_{train}$ are holistic + +First, we clarify the definitions: + +- The requirement of "expressivity over $\mathcal{D}_{train}$ " means that the encodings are injective over $\mathcal{D}_{train}$ . +- The requirement of "efficiency over $\mathcal{D}_{train}$ " means that the encoding maps $\mathcal{D}_{train}$ into $V^{C_{\star}}$ for the minimum value of $C_{\star}$ . + +Similarly to previous subsection (see (9)), keeping in mind that $|\mathcal{D}_{train}| = p|\mathcal{D}| = p|\mathcal{G}|$ , with fixed $p \in (0,1)$ , we have + +$$ +C _ {\star} = \left\lceil \log_ {| V |} \left| \mathcal {D} _ {t r a i n} \right|\right\rceil = \left\lceil \log_ {| V |} p \mid \mathcal {G} \right\rceil \leq \left\lceil \log_ {| V |} \left| \mathcal {G} \right|\right\rceil - \left\lfloor\left| \log_ {| V |} p \right|\right\rfloor . \tag {13} +$$ + +For compositional encodings of $\mathbf{G} = [G_1, \ldots, G_n]$ the encoding will require to set apart separate message sub-strings of length $C_i = \lceil \log_{|V|}N_i \rceil$ for factor $G_i$ (see Def. F.2), thus the required length for the encoding will be + +$$ +C _ {0} := \sum_ {i = 1} ^ {n} \left\lceil \log_ {| V |} N _ {i} \right\rceil \stackrel {(\star)} {\geq} \left[ \sum_ {i = 1} ^ {n} \log_ {| V |} N _ {i} \right] = \left\lceil \log_ {| V |} | \mathcal {G} | \right\rceil \stackrel {(\star \star)} {\geq} C _ {\star} + \lfloor | \log_ {| V |} p | \rfloor , \tag {14} +$$ + +in which $(\star \star)$ follows from (13). + +We now discuss when the inequality signs in (14) are sharp or not, since sharp inequality $C_0 > C_\star$ implies our claim that optimizing for efficiency under expressivity implies non-compositional (i.e., holistic) encodings: + +1. If we use relatively small $|V| \lesssim 10$ (assuming commonly used values $p \lesssim 0.1$ ), we will have + +$$ +| V | \leq \frac {1}{p}, \tag {15} +$$ + +and thus $\log_{|V|}p\leq -1$ , and thus the second inequality $(\star \star)$ in (14) is guaranteed to be sharp, showing the desired strict inequality $C_0 > C_\star$ independently of the sizes of factor ranges $N_{i},i = 1,\ldots ,n$ + +2. If the gaps between $\log_{|V|}N_i$ and the lowest integer larger or equal to it, sum to a value of at least 1, i.e., + +$$ +\sum_ {i = 1} ^ {n} \left(\lceil \log_ {| V |} N _ {i} \rceil - \log_ {| V |} N _ {i}\right) \geq 1, \tag {16} +$$ + +then $(\star)$ becomes sharp, again guaranteeing $C_0 > C_\star$ as desired. This could be probable to happen in our setting, and becomes more likely for large $|V| > 1 / p$ for which (15) fails, given that the $N_{i}$ are supposed to be unknown, so it is likely that for large $|V| > 1 / p$ for a few of the $N_{i}$ the gaps in (16) are nontrivial. For example, (16) holds true for $|V| > 1 / p \geq 10$ if, say, at least two of the $G_{i}$ are binary factors, so that $N_{i} = 2$ . + +In summary we have the following: + +Proposition F.4. Under the same running assumptions over $\mathcal{G}$ , $p$ , we have that optimizing only for efficiency of encodings efficiency under injectivity constraints expressivity over $\mathcal{D}_{train}$ is guaranteed to produce holistic (i.e., non-compositional) encodings if one or both of the conditions (15) and (16) holds. + +# F.4 Possible choices of encoded message regularization + +As seen above, if we just optimize for efficiency and expressivity then we are likely to get holistic encodings (Prop. F.4), which will then make it hard for the receiver to generalize the decoding strategy outside the training set due to having overwhelmingly higher expected complexity (Prop. F.3). + +This proves that it is important to incentivize condition compressibility, i.e., to push encoding strategies away from overly efficient holistic encodings. In this section we discuss a few approaches to do this in practice, explaining our choice of compressibility regularization. + +# 1. Entropy maximization. + +Background and standard approach. Introduced by [69] for use in policy-based reinforcement learning, entropy maximization is known to enforce exploration in deterministic policies and avoid early convergence to single output choices for learned policies. Specifically, in a learning setup with input $x$ , output $y$ and policy parameters $W$ , the authors define the following estimator for fixed output value $\xi$ : + +$$ +h (\xi , W, x) = - \ln P r (y = \xi | W, x), \tag {17} +$$ + +such that if we take the expectation over $\xi$ we get: + +$$ +\mathbb {E} [ h (\xi , W, x) | W, x ] = - \sum_ {\xi} P r (y = \xi | W, x) \ln P r (y = \xi | W, x), \tag {18} +$$ + +which is the entropy of the network. Therefore $h$ is an unbiased estimator of the entropy. They also note that, if $w_{i}$ and $x_{i}$ with $i \in U$ represent the weights and pattern of single neurons in a feed forward network, $\xi_{i}$ the $i_{th}$ position in a $n_u$ -tuple $\xi$ , and let $g_{i}$ be the probability density describing $y_{i}$ , i.e: + +$$ +g _ {i} \left(\xi , w ^ {i}, x ^ {i}\right) = P r \left(y _ {i} = \xi \mid w ^ {i}, x ^ {i}\right), \tag {19} +$$ + +then, for any feed forward network, + +$$ +P r (y = \xi | W, x) = \prod_ {i \in U} g _ {i} \left(\xi_ {i}, w ^ {i}, x ^ {i}\right), \tag {20} +$$ + +and therefore + +$$ +- \ln P r (y = \xi | W, x) = - \sum_ {i \in U} g _ {i} \left(\xi_ {i}, w ^ {i}, x ^ {i}\right) = h (y, W, X). \tag {21} +$$ + +In fact, this method can be extended onto sequence prediction models, and even IL, such as in [55]. + +Naive adaptation to our setting is not practical. We take $x \in \mathcal{X}$ and $\xi \in \mathcal{M} = V^C$ and recall that in our notation $m_t$ indicates the $t$ -th token of message $m$ , and $m_{[t]}$ is the notation for the initial segment $m_1, m_2, \ldots, m_t$ , so that in particular $m_{[C]} = m$ . + +Importantly, in our setting, to avoid high variance in estimation due to the large message space (see §G), we approximate the probability distribution of the RNN-produced message token at position $t \leq C$ , denoted here $p(m_t | x, m_{[t-1]})$ , as a $\delta$ distribution over the vocabulary $V$ by using the Gumbel-softmax trick: + +$$ +(\exists v \in V), \quad P r \left(m _ {t} \mid x, W, m _ {[ t - 1 ]}\right) \approx \left\{ \begin{array}{l l} 1, & \text {i f} m _ {t} = v \\ 0, & \text {i f} m _ {t} \in V \setminus \{v \}, \end{array} \right. \tag {22} +$$ + +and therefore: + +$$ +P r (m = \xi | W, x) = \prod_ {t = 1} ^ {C} P r \left(m _ {t} = \xi_ {t} | W, x, m _ {[ t - 1 ]}\right). \tag {23} +$$ + +Note that due to (22), for all but a single one of the $|V|^C$ possible values of $\xi$ the quantity (23) is $\approx 0$ , and thus $h$ becomes a poor estimator for the entropy of $p(m_t|x, m_{[t-1]})$ , as we get + +$$ +h (m, W, x) = - \ln \sum_ {t = 1} ^ {C} P r \left(m _ {t} = \xi_ {t} | x, W, m _ {[ t - 1 ]}\right) \approx \infty . \tag {24} +$$ + +In particular, it becomes clear that optimizing this $h$ becomes unfeasible with gradient methods. + +Our proposal for regularization. Our model can be seen as a different approximation of the entropy of the message. Instead of using the policy weights as probability values, we use the law of large numbers to estimate the probability of each message. For a batch $\mathcal{B} \subset \mathcal{X}$ consisting of $N_{batch}$ training examples and for given sender encoding protocol $S_{\phi}$ : + +$$ +\text {F o r a s u f f i c i e n t l y} N _ {\text {b a t c h}}: \Pr (m = \xi \mid x, \phi) \approx \frac {\left| \left\{x \in \mathcal {B} \mid S _ {\phi} (x) = \xi \right\} \right|}{N _ {\text {b a t c h}}} \tag {25} +$$ + +Yet again, due to the exponential size of the message space, utilizing this probability to estimate the entropy is still not practical. We will assume approximate independence between symbol positions, which allows us to get the stronger approximation: + +$$ +\Pr \left(m _ {t} = v \mid x, \phi\right) \approx \frac {\left| \left\{x \in \mathcal {B} \mid S _ {\phi} (x) _ {t} = v \right\} \right|}{N _ {\text {b a t c h}}}, \tag {26} +$$ + +as well as the additivity of entropy to be applied to (18): + +$$ +\begin{array}{l} H (P r (m | x, \phi)) := - \sum_ {\xi \in V ^ {C}} P r (m = \xi \mid x, \phi) \ln P r (m = \xi \mid x, \phi) \\ = - \sum_ {t = 1} ^ {C} \sum_ {v \in V} P r \left(m _ {t} = v \mid x, \phi\right) \ln P r \left(m _ {t} = v \mid x, \phi\right). \tag {27} \\ \end{array} +$$ + +Then we can perform the following computations, in which we let $m, m'$ be two i.i.d. copies of $m$ : + +$$ +\begin{array}{l} H (P r (m \mid x, \phi)) \stackrel {(2 7)} {=} - \sum_ {t = 1} ^ {C} \sum_ {v \in V} P r (m _ {t} = v \mid x, \phi) \ln P r (m _ {t} = v \mid x, \phi) \\ \stackrel {{(- x \ln x \geq x - x ^ {2} \underset {\text {f o r} x \leq 1})}} {\geq} \sum_ {t = 1} ^ {C} \sum_ {v \in V} \left[ P r (m _ {t} = v \mid x, \phi) - P r (m _ {t} = v \mid x, \phi) ^ {2} \right] \\ \stackrel {(\star)} {=} \sum_ {t = 1} ^ {C} \left(\sum_ {v, v ^ {\prime} \in V} P r (m _ {t} = v \mid x, \phi) P r (m _ {t} ^ {\prime} = v ^ {\prime} \mid x, \phi)\right. \\ \left. - \sum_ {v \in V} P r (m _ {t} = v \mid x, \phi) ^ {2}\right) \\ = \sum_ {t = 1} ^ {C} \sum_ {v \neq v ^ {\prime} \in V} P r (m _ {t} = v \mid x, \phi) P r (m _ {t} ^ {\prime} = v ^ {\prime} \mid x, \phi) \\ = \sum_ {t = 1} ^ {C} P r (m _ {t} \neq m _ {t} ^ {\prime}) = \mathbb {E} [ d _ {H} (m, m ^ {\prime}) ] \\ \stackrel {(2 6)} {\approx} \mathbb {E} _ {x, x ^ {\prime} \in \mathcal {B}} \left[ d _ {H} \left(S _ {\phi} (x), S _ {\phi} \left(x ^ {\prime}\right)\right) \right], \tag {28} \\ \end{array} +$$ + +In the above step $(\star)$ we used that + +$$ +\begin{array}{l} \sum_ {v \in V} P r (m _ {t} = v \mid x, \phi) = 1 = \left(\sum_ {v \in V} P r (m _ {t} = v \mid x, \phi)\right) ^ {2} \\ = \sum_ {v, v ^ {\prime} \in V} P r (m _ {t} = v \mid x, \phi) P r (m _ {t} ^ {\prime} = v ^ {\prime} \mid x, \phi). \\ \end{array} +$$ + +The consequence of bound (28) is that maximizing the pairwise Hamming distance between predicted messages in a batch also maximizes a lower bound to the approximate entropy of the message space, justifying the choice of optimizing this quantity (which is easier to include in practice) for favoring compressibility. + +2. Contrastive Learning losses. In our framework, PDM can be seen as a form of contrastive loss, which we now show how to connect to standard contrastive losses from previous works. + +Contrastive loss functions such as the triplet loss [63] and NT-Xtent [11] work by defining positive and negative image pairs (usually through image augmentation), and aiming to minimize the distance between embeddings of positive pairs and maximize the distance between negative pairs. Specifically, instead of purposefully sampling negative samples for every $x$ , NT-Xtent produces a single pair of samples for each $x$ in a mini-batch $N$ by + +augmenting each image. Each image in the batch therefore has $2N - 1$ negative pairs. They use the cosine similarity $sim(u,v) = u^{T}v / \| u\| \| v\|$ and some temperature value $\tau$ to define the following loss between positive pair embeddings $z_{i}$ and $z_{j}$ : + +$$ +\ell_ {i, j} = - \log \frac {\exp \left(\operatorname {s i m} \left(z _ {i} , z _ {j}\right) / \tau\right)}{\sum_ {k} ^ {2 N} \exp \left(\operatorname {s i m} \left(z _ {i} , z _ {k}\right) / \tau\right)} \tag {29} +$$ + +For simplicity, let $\tau = 1$ and thus + +$$ +\ell_ {i, j} = - \operatorname {s i m} \left(z _ {i}, z _ {j}\right) + \log \sum_ {k} ^ {2 N} \exp \left(\operatorname {s i m} \left(z _ {i}, z _ {k}\right)\right). \tag {30} +$$ + +Here, we apply the same ideas but take this loss over an unaugmented batch, i.e the only application of this loss is for $z_{i}$ against itself. In this case it reduces to: + +$$ +\ell_ {i} = - 1 + \log \sum_ {k} ^ {N} \exp (\operatorname {s i m} \left(z _ {i}, z _ {k}\right)) \tag {31} +$$ + +In our representations, the embedding vectors from a batch $z_{i}, i = 1,\dots ,N$ are $V\times C$ matrices whose columns are approximately one-hot vectors that represent the tokens of a message. For one-hot matrices $z_{i},z_{k}$ of this form, the dot product $z_{i}^{T}z_{k}$ is equivalent to counting the amount of matching rows of $z_{i},z_{k}$ , i.e., the un-normalized Hamming distance of the associated message strings $m_{i},m_{k}\in V^{C}$ . Since strings have length $C$ and $\| z_i\| ^2 =$ $z_{i}^{T}z_{i}\simeq C$ , we thus get: + +$$ +\begin{array}{l} { s i m ( z _ { i } , z _ { k } ) } { = } { t r ( z _ { i } ^ { T } z _ { k } ) / \| z _ { i } \| \| z _ { k } \| \simeq \frac { 1 } { C } \sum _ { l = 1 } ^ { C } \mathbb { 1 } [ \arg \operatorname* { m a x } _ { j } ( ( z _ { i } ) _ { l , j } ) = \arg \operatorname* { m a x } _ { j } ( ( z _ { k } ) _ { l , j } ) ] } \\ \simeq 1 - d _ {H} (m _ {i}, m _ {k}), \\ \end{array} +$$ + +in which $d_H$ is the (normalized by $1 / C$ ) Hamming distance. We can replace this expression in the unaugmented NT-Xtent loss: + +$$ +\begin{array}{l} \ell_ {i} \quad = \quad - 1 + \log \sum_ {k = 1} ^ {N} \exp \left(1 - d _ {H} \left(m _ {i}, m _ {k}\right)\right) \\ = - 1 + N - \log \sum_ {k = 1} ^ {N} \exp \left(d _ {H} \left(m _ {i}, m _ {k}\right)\right) \\ \stackrel {\text {(Jensen}} {\leq} - 1 + N - \log N - \frac {1}{N} \sum_ {k = 1} ^ {N} d _ {H} (m _ {i}, m _ {k}). \\ \end{array} +$$ + +Then, taking the mean loss over the full batch we get: + +$$ +\frac {1}{N} \sum_ {i = 1} ^ {N} \ell_ {i} \leq - 1 + N - \log N - \frac {1}{N ^ {2}} \sum_ {i = 1} ^ {N} \sum_ {k = 1} ^ {N} d _ {H} \left(m _ {i}, m _ {k}\right). \tag {32} +$$ + +We then include in the loss only the last term, i.e. the average of $d_{H}(m_{i}, m_{k})$ , which is the only term that actually depends on the batch elements, which forms an approximate upper bound on the NT-Xtent loss. + +Note that our approximation of $sim(z_i, z_k)$ by $1 - d_H(m_i, m_k)$ , on which the above is based, becomes sharp when the $z_i$ have actual one-hot entries, thus the approximation becomes better as our model has high accuracy. Furthermore, Jensen's inequality, responsible for the inequality between the term $\ell_i$ from NT-Xtent and our metric, becomes closer to equality messages $m_1, \ldots, m_N$ become uniformly spread across $V^C$ , which is the case for minimizers of our metric as well. This means that, at least heuristically speaking, we can consider our metric as having similar minimizers to the NT-Xtent regularization. + +A precise matching lower bound to (32) is left for future work. + +# G Justification of our imitation step starting from classical probabilistic reconstruction games + +We start by recalling classical probabilistic reconstruction games, after which we adapt the formulation to the setting of generative tasks, and we show that with a few justified simplifications we arrive to our formulation from the main paper. + +# G.1 Classical probabilistic reconstruction games + +Recall that in reconstruction games, the two agent types are a teacher and student (or speaker and listener, or sender and receiver), parameterized by respectively $\theta$ , $\phi$ . The game consists in the teacher observing a signal $x$ distributed according to a probability distribution $p$ , from which it produces a randomized message $m$ whose distribution conditional on $x$ is obtained according to policy $\pi_{\theta}(\cdot | x)$ . The student or listener, observes message $m$ and produces a reconstructed signal distributed according to a conditional probability $\rho_{\phi}(\cdot | m)$ . The game dynamics can be modeled by stipulating that together, teacher and student minimize the log-likelihood loss function + +$$ +\mathcal {L} _ {\theta , \phi} = - \mathbb {E} _ {x \sim p, m \sim \pi_ {\theta} (\cdot | x)} [ \log \rho_ {\phi} (x | m) ] \tag {33} +$$ + +in which $x$ represents the signal and $m$ the message used to reconstruct it, the speaker produces a reinforcement learning policy $\pi_{\theta}(\cdot |x)$ and the listener produces a reconstruction function $\rho_{\phi}(x|m)$ . + +# G.2 Reconstruction games in generative tasks + +While the above formulation (33) is formalized using the log-likelihood between distributions, for our generative task, the reconstruction modeled slightly differently: the task is to optimize, with high probability, the reconstruction error, measured by a distance between the signal $x$ and its reconstruction $x'$ , denoted $d(x, x')$ . Here the notation for $x, x'$ and $d(x, x')$ can be (ab)used to represent two alternative settings: either (a) $x, x'$ are vectors in the autoencoder latent space and $d(x, x')$ is a discrepancy (e.g. $d(x, x') = |x - x'|^p$ ) between them or (b) $x, x'$ are reconstructed images and $d(x, x')$ is a discrepancy (e.g. based on some ad-hoc norm or otherwise) between them. Thus we have the probabilistic error function + +$$ +\mathcal {L} _ {i n t} ^ {\mathbb {P}} (\omega , \varphi) := \mathbb {E} _ {x \sim p, m \sim \pi_ {\varphi} (\cdot | x), x ^ {\prime} \sim \rho_ {\omega} (\cdot | m)} d (x, x ^ {\prime}). \tag {34} +$$ + +The loss (34) is akin to a $d$ -based Wasserstein distance analogue of (33). + +# G.3 The problem reduces to the case of a deterministic student + +Note that when $d(x,x) = |x - x'|^2$ for some Euclidean norm (or more generally, when $d$ is a convex function of such norm), then the minimization of (34) over distributions $\rho_{\omega}$ at fixed $m, \varphi$ must be achieved by + +$$ +x _ {m} ^ {\varphi} \in \operatorname {a r g m i n} _ {x ^ {\prime}} \mathbb {E} _ {x \sim p _ {\varphi} (\cdot | m)} d \left(x, x ^ {\prime}\right), \tag {35} +$$ + +where if the set of $x$ is finite then $p_{\varphi}$ is the probability defined by + +$$ +p _ {\varphi} \left(x ^ {\prime} \mid m\right) = \frac {p \left(x ^ {\prime}\right) \pi_ {\varphi} \left(m \mid x ^ {\prime}\right)}{\sum_ {y} p (y) \pi_ {\varphi} \left(m \mid y\right)}, \tag {36} +$$ + +and an analogous expression with integrals replacing the sum holds in general. + +Under the above convexity hypotheses, the choice of $x_{m}^{\varphi}$ in (35) is unique, and for $d(x,x') = |x - x'|^{2}$ it coincides with the barycenter of $p_{\varphi}$ . In this case, $\rho_{\omega}(\cdot|m)$ can be taken to be the Dirac distribution concentrated at $x_{m}^{\varphi}$ , and optimization of (34) can be achieved by a deterministic reconstruction function $R_{\omega}(m)$ , whose optimum value for given $\varphi$ will be $x_{m}^{\varphi}$ from (35). When reducing to optimizing only amongst deterministic $R_{\omega}$ , the loss (34) rewrites in the simplified form + +$$ +\mathcal {L} _ {i n t} (\omega , \varphi) = \mathbb {E} _ {x \sim p, m \sim \pi_ {\varphi} (\cdot | x)} d (x, R _ {\omega} (m)). \tag {37} +$$ + +# G.4 Straight-Trough Gumbel-Softmax (STGS) trick + +As in [25] and subsequent works, the optimization of policy $\pi_{\varphi}$ (minimization over $\varphi$ ) can in theory be done via policy gradient optimization methods such as REINFORCE [70], however, due to the combinatorial explosion of the number of messages this can be very compute-intensive if calculated directly, or produces high variance estimators. We thus follow the straight-trough Gumbel-softmax trick [26], in which the probabilistic interpretation of $\pi_{\varphi}$ is implicit in the network architecture, and the architecture itself is deterministic and directly differentiable. In this case, we can thus pass to the deterministic description as $m = S_{\phi}(x)$ , replacing the probabilistic one $m \sim \pi_{\varphi}(\cdot|m)$ , and we obtain the updated form of (37) as + +$$ +\mathcal {L} _ {i n t} (\omega , \phi) = \mathbb {E} _ {x \sim p} d (x, R _ {\omega} (S _ {\phi} (x))), \tag {38} +$$ + +which for $d(x,x^{\prime}) = |x - x^{\prime}|$ is very similar to our actual interaction loss (1): + +$$ +\mathcal {L} _ {i n t} (\omega , \phi) = \mathbb {E} _ {x \in \mathcal {X}} \left[ | x - R _ {\omega} \left(S _ {\phi} (x)\right) | \right]. \tag {39} +$$ + +# H Example Generation: Reconstructions + +The following reconstruction, Figure 5, examples were generated with the CELEBI model using PDM regularization with $\beta = 1$ and $\lambda = 1.5$ . In each row, the leftmost image corresponds to the original input, while the subsequent images represent the predicted reconstructions $x_{i} = R_{\omega}(S_{\phi}(x)_{[1:i]})$ obtained from progressively longer message prefixes with the VAE backbone. + +We observe that the reconstructions do not merely improve in pixel-level fidelity but rather exhibit semantic refinements across successive steps — variations in floor hue, object color, shape, or viewing angle. This progression suggests that the message space encodes the underlying generative factors $\mathcal{G}$ and supports a degree of compositional structure, where messages carry disentangled information. + +Qualitative comparison across methods. To further assess the representational properties of CELEBI, we qualitatively compare reconstructions across the different methods proposed in this work (see Figures 6 and 7). As illustrated in these figures, CELEBI reaches semantically faithful reconstructions significantly earlier in the decoding sequence than the baseline methods. + +Moreover, the differences between reconstructions are more semantically pronounced than in baseline models, which often show subtler variations mainly related to texture or brightness. This qualitative evidence supports our hypothesis that CELEBI facilitates a more structured message space, enabling representations that are not only more compact but also more semantically coherent. + +![](images/5f40261466a3f3680968411c0dc74dd650816affeeb82a0c900f12657422b5c0.jpg) +(a) Example 1 + +![](images/7424b5b86f13cdd8449c9f5663503ab57c8cda8ced3a1c04ab98ced49d08ecbf.jpg) +(b) Example 2 + +![](images/59b1e8975626c79c192823f8f43b5340a02de8f4e9eda779f8407192f20cb8b3.jpg) +(c) Example 3 + +![](images/649b3e2f73aa03ae42374259802e511452629b0641a2698878c208294baf91e6.jpg) +(d) Example 4 +Figure 5: Progressive image reconstructions obtained at each decoding step by the receiver as it processes successive symbols of the message. + +![](images/ce7c215b38ab41ebad11347ee975e352ea26c9040f0da10911a872d5743212cc.jpg) +(a) Reconstruction using the baseline model. + +![](images/ca12896dc7af7f1689d758ab7afb9eae4ea29b88cfbc5c5413dc998531b69a97.jpg) +(b) Reconstruction using partial decoding. + +![](images/f7f6698e9b767c7e0f19f9c48bc4d9142df2cd1b12a652245a2d0c3524cf073a.jpg) +(c) Reconstruction using final state imitation. + +![](images/ac35e785ae6a345f779ad98f4f91b782b7c3b94397f5446711430f7fed6b4977.jpg) +(d) Reconstruction using $\mathrm{PD + FiSI + PDM}$ +Figure 6: Reconstruction for the same example with different methods. + +# I Additional Language Metrics + +Table 3: Comparison of proposed methods using language variation metrics from [13]. + +
Synonymy ↓Homonymy ↓Freedom ↓Entanglement ↓
InteractionFull message reconstruction, no IL (Baseline)0.563 ± 0.0260.695 ± 0.0150.569 ± 0.0260.831 ± 0.006
Progressive Decoding (PD, ours)0.640 ± 0.0220.737 ± 0.0130.646 ± 0.0130.786 ± 0.007
ImitationMessage imitation (baseline)0.601 ± 0.0390.714 ± 0.0270.613 ± 0.0390.822 ± 0.010
Full state imitation0.597 ± 0.0170.734 ± 0.0210.604 ± 0.0180.824 ± 0.007
Final-State Imitation (FiSI, ours)0.633 ± 0.0380.751 ± 0.0180.640 ± 0.0380.842 ± 0.007
RegularizationPD+FiSI+KoLeo λ = 1.50.594 ± 0.0270.739 ± 0.0130.598 ± 0.0270.805 ± 0.007
PD+FiSI+PDM (ours) λ = 1.50.654 ± 0.0130.753 ± 0.0080.658 ± 0.0130.780 ± 0.008
+ +(a) Shapes3D + +
Synonymy ↓Homonymy ↓Freedom ↓Entanglement ↓
InteractionFull message reconstruction, no IL (Baseline)0.595 ± 0.0530.597 ± 0.0610.597 ± 0.0520.852 ± 0.021
Progressive Decoding (PD, ours)0.484 ± 0.0440.486 ± 0.0560.486 ± 0.0440.786 ± 0.025
ImitationMessage imitation (baseline)0.367 ± 0.0340.393 ± 0.0420.369 ± 0.0340.837 ± 0.015
Full state imitation0.362 ± 0.0450.424 ± 0.0520.364 ± 0.0440.832 ± 0.018
Final-State Imitation (FiSI, ours)0.324 ± 0.0200.393 ± 0.0180.328 ± 0.0200.830 ± 0.008
RegularizationPD+FiSI+KoLeo λ = 1.50.365 ± 0.0110.338 ± 0.0190.367 ± 0.0110.763 ± 0.005
PD+FiSI+PDM (ours) λ = 1.50.347 ± 0.0120.333 ± 0.0100.349 ± 0.0130.756 ± 0.005
+ +(b)MPI3D + +In addition to the metrics presented in the main text of this work, we attempted to measure other metrics for qualitative linguistic variation. We were not able to find many such metrics in the literature, and thus restrict to the ones presented in [13]. Importantly, since this reference had a slightly different focus, variation in this context will not refer to the evolution of the language in time, but rather a departure from regularity, which masks an underlying compositional structure. + +![](images/1d86b9f507eb19c08f5d1971f5c7cbd3f5211da95c594ff1bd3338da09a4bb7e.jpg) +(a) Reconstruction using the baseline model. + +![](images/522d66164792bf7b1044af1229dd703fa939477da607a992bbf5fca9fcab2ebe.jpg) +(b) Reconstruction using partial decoding. + +![](images/fbc1c3526f732049f0c40524a18fe580b34df889be7671d39323566dd7f215fb.jpg) +(c) Reconstruction using final state imitation. + +![](images/6a81a3c8b5b1859f526b418b717e183814da16b753a4e672b07f77e2280c39c5.jpg) +(d) Reconstruction using $\mathrm{PD + FiSI + PDM}$ +Figure 7: Reconstruction for the same example with different methods. + +We study the four measures of variation presented in the paper, namely synonymy, homonymy, word order freedom and entanglement. Synonymy measures the presence of one-to-many mappings between atomic meanings and characters in a position, being minimized when each generating factor value is mapped to a single position and character. Homonymy measures the opposite, i.e., the presence of many-to-one mappings, and is minimized when each character in a position is mapped to a unique generating factor value. Word order freedom refers to the strictness of the mapping between generating factors and positions in the message (for example, if the mapping system always encodes the shape generating factor at the first position). It is minimized if all single generating factors are encoded in the same message position. Similarly, entanglement is minimized when all factors are encoded into unique positions in the message. + +Important formular metrics, pero no encaja tan bien. Unico paper que conocemos. No clear patterns, but we argue it is because... An important assumption made by the authors for all these metrics is that meaning should be undivisably encoded into single positions of the message. We believe this scope of regularity is too narrow to capture compositionality-respecting mappings in out setting, as defined in F.2, and does not fit well when the space of messages is greatly larger than the amount of possible states. We found no clear trend or similar behavior across both tested datasets, save a small decrease in entanglement when using PD. This may suggest that the increased pressure for efficiency forces individual generating factors to be uniquely distributed in the message, however we believe there is not sufficient evidence to make such a claim considering the previously discussed pitfalls of the metrics. + +To our knowledge there are no available official implementations for these metrics. Our versions can be found in the official repository for this work. + +Table 4: Permutation tests for Shapes3D and MPI3D on emergent language metrics. + +
MetricMode AMode BStatisticp-Value
TopSimProgressive Decoding (PD; ours)Full message reconstruction; no IL (Baseline)0.02630.0233
TopSimFull state imitationFinal-State Imitation (FiSI; ours)-0.02670.1014
TopSimFull state imitationMessage imitation (Baseline)-0.00180.9332
TopSimFinal-State Imitation (FiSI; ours)Message imitation (Baseline)0.02490.2605
TopSimPD+FiSI+KoLeo λ = 1.5PD+FiSI+PDM λ = 1.5-0.03710.0004
Useful LengthProgressive Decoding (PD; ours)Full message reconstruction; no IL (Baseline)-2.50.0031
Useful LengthFull state imitationFinal-State Imitation (FiSI; ours)0.01.0
Useful LengthFull state imitationMessage imitation (Baseline)0.01.0
Useful LengthFinal-State Imitation (FiSI; ours)Message imitation (Baseline)0.01.0
Useful LengthPD+FiSI+KoLeo λ = 1.5PD+FiSI+PDM λ = 1.51.30000.0996
Final MSEProgressive Decoding (PD; ours)Full message reconstruction; no IL (Baseline)-0.02180.0657
Final MSEFull state imitationFinal-State Imitation (FiSI; ours)-0.01300.4871
Final MSEFull state imitationMessage imitation (Baseline)-0.02010.5679
Final MSEFinal-State Imitation (FiSI; ours)Message imitation (Baseline)-0.00710.8384
Final MSEPD+FiSI+KoLeo λ = 1.5PD+FiSI+PDM λ = 1.50.03430.0431
+ +(a) Shapes3D permutation test + +
MetricMode AMode BStatisticp-Value
TopSimProgressive Decoding (PD; ours)Full message reconstruction; no IL (Baseline)0.00370.4416
TopSimFull state imitationFinal-State Imitation (FiSI; ours)-0.01930.0019
TopSimFull state imitationMessage imitation (Baseline)0.00470.4395
TopSimFinal-State Imitation (FiSI; ours)Message imitation (Baseline)0.02400.0007
TopSimPD+FiSI+KoLeo λ = 1.5PD+FiSI+PDM λ = 1.5-0.00600.4767
Useful LengthProgressive Decoding (PD; ours)Full message reconstruction; no IL (Baseline)-2.70000.0345
Useful LengthFull state imitationFinal-State Imitation (FiSI; ours)0.02311.0
Useful LengthFull state imitationMessage imitation (Baseline)0.22310.3998
Useful LengthFinal-State Imitation (FiSI; ours)Message imitation (Baseline)0.20000.5820
Useful LengthPD+FiSI+KoLeo λ = 1.5PD+FiSI+PDM λ = 1.50.22221.0
Final MSEProgressive Decoding (PD; ours)Full message reconstruction; no IL (Baseline)0.00110.5894
Final MSEFull state imitationFinal-State Imitation (FiSI; ours)-0.00180.1166
Final MSEFull state imitationMessage imitation (Baseline)0.00190.2870
Final MSEFinal-State Imitation (FiSI; ours)Message imitation (Baseline)0.00370.0201
Final MSEPD+FiSI+KoLeo λ = 1.5PD+FiSI+PDM λ = 1.5-0.0000120.9626
+ +(b) MPI3D permutation test + +# J Permutation Tests + +# J.1 Emergent language metrics + +To assess the statistical significance of each component added to the Iterated Learning (IL) framework proposed in this work, we conducted permutation tests using the SciPy library [67]. The test statistic was defined as the difference in means, $\bar{A} -\bar{B}$ , where $\bar{A}$ corresponds to the arithmetic mean of the evaluated mode and metric across 10 random seeds. The complete results are presented in Table 5. + +Across both the Shapes3D and MPI3D datasets, we found that the inclusion of the PD module consistently reduced the useful message length, yielding a test statistic of approximately 2.5 and a $p$ -value below 0.05. None of the other additions introduced in this work produced a statistically significant decrease in useful length relative to their respective baselines, supporting our hypothesis that PD imposes an efficiency pressure on the emergent language. Moreover, we observed a statistically significant increase in TopSim for the Shapes3D dataset, with a test statistic of 0.026 and $p < 0.05$ . + +Final-State Imitation had a statistically significant gain in TopSim over both baseline methods on MPI3D, with statistic of $\sim 0.02$ and $p < 0.02$ . PDM outperformed the KoLeo estimator in both reconstruction error and TopSim on Shapes3D, but achieved no statistically significant increase in MPI3D. + +# J.2 Disentanglement permutation tests + +In this section we evaluate the statistical significance of the differences between our model and the discrete baseline on the metrics used in table 2. Similar to the previous section, we conducted + +permutation tests, using the difference in means as a statistic for 10 random seeds. The complete results are presented in Table 5. + +We found a significant difference in accuracy on the Shapes3D dataset, with a statistic of 0.114 and $p < 0.01$ . We found a significant decrease in RMSE on both datasets, with statistics of 0.227 and 1.151 for Shapes3D and MPI3D, respectively, and $p \leq 0.001$ . We also observed a significant gain in disentanglement on both datasets, with statistics of 0.003 and 0.023 for Shapes3D and MPI3D, respectively, and $p < 0.05$ for both. + +Table 5: Permutation tests for Shapes3D and MPI3D on disentanglement, accuracy and RMSE using messages from discrete models as data for a linnear probe. + +
MetricMode AMode BDiff(A-B)p-Value
AccuracyVAE+ELVAE+CELEBI-0.11390.0018
RMSEVAE+ELVAE+CELEBI0.22660.0010
DCI DisentanglementVAE+ELVAE+CELEBI-0.00340.0042
+ +(a) Shapes3D permutation test + +
MetricMode AMode BDiff(A-B)p-Value
AccuracyVAE+ELVAE+CELEBI-0.01280.0570
RMSEVAE+ELVAE+CELEBI1.15120.0010
DCI DisentanglementVAE+ELVAE+CELEBI-0.02300.0002
+ +(b) MPI3D permutation test \ No newline at end of file diff --git a/NeurIPS/2025/A Compressive-Expressive Communication Framework for Compositional Representations/images.zip b/NeurIPS/2025/A Compressive-Expressive Communication Framework for Compositional Representations/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..1ba0071e23245dea2e6767c69fb8dbcdc39ce2a0 --- /dev/null +++ b/NeurIPS/2025/A Compressive-Expressive Communication Framework for Compositional Representations/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:57a438bac78ed22bbc38f65e8aaca1bd9597f438379e197a6bc9c9f57960bbb4 +size 1482733 diff --git a/NeurIPS/2025/A Compressive-Expressive Communication Framework for Compositional Representations/layout.json b/NeurIPS/2025/A Compressive-Expressive Communication Framework for Compositional Representations/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..7f184c29cf321d7ac7bac8428c5710b8d20c192e --- /dev/null +++ b/NeurIPS/2025/A Compressive-Expressive Communication Framework for Compositional Representations/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d18bd6b874f9b437d0f424f3d11fd7651100e0be8390cb0cbba688f55f96e302 +size 1521764 diff --git a/NeurIPS/2025/A Computationally Viable Numerical Gradient-based Technique for Optimal Covering Problems/a786f11e-70dc-4619-988a-78cc2ce16cdf_content_list.json b/NeurIPS/2025/A Computationally Viable Numerical Gradient-based Technique for Optimal Covering Problems/a786f11e-70dc-4619-988a-78cc2ce16cdf_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..c1af9beaeb25ae78379e628f23e8d0720d973d50 --- /dev/null +++ b/NeurIPS/2025/A Computationally Viable Numerical Gradient-based Technique for Optimal Covering Problems/a786f11e-70dc-4619-988a-78cc2ce16cdf_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68c68be0ea7bb9be21009d91fbabb7eecba87f96490a536d51509752d2539b3e +size 143629 diff --git a/NeurIPS/2025/A Computationally Viable Numerical Gradient-based Technique for Optimal Covering Problems/a786f11e-70dc-4619-988a-78cc2ce16cdf_model.json b/NeurIPS/2025/A Computationally Viable Numerical Gradient-based Technique for Optimal Covering Problems/a786f11e-70dc-4619-988a-78cc2ce16cdf_model.json new file mode 100644 index 0000000000000000000000000000000000000000..78ef582b22b97462b5f9a0c4cdd510ccfb8d018d --- /dev/null +++ b/NeurIPS/2025/A Computationally Viable Numerical Gradient-based Technique for Optimal Covering Problems/a786f11e-70dc-4619-988a-78cc2ce16cdf_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1e062db41524f3ad2c7921af77b25ff63d8a95f152494fa6a44c30f53dabfd48 +size 184491 diff --git a/NeurIPS/2025/A Computationally Viable Numerical Gradient-based Technique for Optimal Covering Problems/a786f11e-70dc-4619-988a-78cc2ce16cdf_origin.pdf b/NeurIPS/2025/A Computationally Viable Numerical Gradient-based Technique for Optimal Covering Problems/a786f11e-70dc-4619-988a-78cc2ce16cdf_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..64092e616bee67a511e4344add3069e582376a11 --- /dev/null +++ b/NeurIPS/2025/A Computationally Viable Numerical Gradient-based Technique for Optimal Covering Problems/a786f11e-70dc-4619-988a-78cc2ce16cdf_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fd9c5144f673fe011638c0efb101a4733ca935a05d5bee8af85c1d48805e5986 +size 8563125 diff --git a/NeurIPS/2025/A Computationally Viable Numerical Gradient-based Technique for Optimal Covering Problems/full.md b/NeurIPS/2025/A Computationally Viable Numerical Gradient-based Technique for Optimal Covering Problems/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e053ed127ca5f483f47145e5982feb746bcd0102 --- /dev/null +++ b/NeurIPS/2025/A Computationally Viable Numerical Gradient-based Technique for Optimal Covering Problems/full.md @@ -0,0 +1,792 @@ +# A Computationally Viable Numerical Gradient-based Technique for Optimal Covering Problems + +Gokul Rajaraman +Department of Mechanical Engineering +IIT Bombay +Mumbai 400076, India +22b4517@iitb.ac.in + +Debasish Chatterjee +Centre for Systems and Control +IIT Bombay +Mumbai 400076, India +dchatter@iitb.ac.in + +# Abstract + +The problem of optimally covering a given compact subset of $\mathbb{R}^N$ with a preassigned number $n$ of Euclidean metric balls has a long-standing history and it is well-recognized to be computationally hard. This article establishes a numerically viable algorithm for obtaining optimal covers of compact sets via two key contributions. The first is a foundational result establishing Lipschitz continuity of the marginal function of a certain parametric non-convex maximization problem in the optimal covering problem, and it provides the substrate for numerical gradient algorithms to be employed in this context. The second is an adaptation of a stochastically smoothed numerical gradient-based (zeroth-order) algorithm for a non-convex minimization problem, that, equipped with randomized restarts, spurs global convergence to an optimal cover. Several numerical experiments with complicated nonconvex compact sets demonstrate the excellent performance of our techniques. + +# 1 Introduction + +Optimal covering of sets is an important problem that arises naturally in several fields of science and technology, including learning and approximation theory, computer science, signal processing, information theory, combinatorics, communication, sensor networks, multi-agent systems, cyberphysical systems, etc. In a sensor network, for instance, given a collection of sensors having preassigned sensing radius and a set that must be enveloped by the sensors, an important engineering question is to find an optimal collection of points for placing the sensors to ensure complete coverage of the given set. An identical problem arises for the coverage of a given geographical region with cell-phone towers in order to maximize network availability in the region. In information theory, assuming that one can accurately decode messages from a certain set within a preassigned maximal error tolerance for each, it is important to determine the number of 'signals' that are necessary for accurate decoding of all the messages; this is once again a problem of optimal covering of the set of messages with a collection of 'signals'. + +The covering problem appears in various forms, and perhaps its most common variant is captured by the following situation. For a given non-empty (and possibly uncountable) compact set $\mathcal{M} \subset \mathbb{R}^N$ and a preassigned $\varepsilon > 0$ , one finds a finite set $\mathcal{F} \subset \mathbb{R}^N$ that is at most $\varepsilon$ -distant from any point in $\mathcal{M}$ . Such a minimal set is an $\varepsilon$ -net and its cardinality is the so-called $\varepsilon$ -covering number of the set $\mathcal{M}$ ; see, e.g., (Kolmogorov and Tihomirov, 1961) for an early detailed exposé of related topics and (Alimov and Tsar'kov, 2021, Chapters 15 and 16) for a more recent treatment. A closely related and equally relevant variant of covering is its resource-constrained or budgeted version: Given a number $n$ of metric balls, one must find the minimal radius and a placement of the $n$ centers of the balls such that their union covers $\mathcal{M}$ . Such a collection of the centers is commonly known as a Chebyshev net + +of cardinality $n$ (Alimov and Tsar'kov, 2021, Definition 15.3).1 We address this particular problem in our current work. + +Formally, given a non-empty and compact $\mathcal{M} \subset \mathbb{R}^N$ , we seek to solve the optimization problem + +$$ +\epsilon (\mathcal {M}, n) := \min _ {(x _ {1}, \dots , x _ {n}) \in \mathbb {R} ^ {N \times n}} \max _ {y \in \mathcal {M}} \min _ {i = 1, \dots , n} \| y - x _ {i} \|. \tag {1} +$$ + +For an $n$ -tuple of centers $(x_{1},\ldots ,x_{n})\in \mathbb{R}^{N\times n}$ and $y\in \mathcal{M}$ , $\min_{i = 1,\dots,n}\| y - x_i\|$ determines the distance of $y$ from $\bigcup_{i = 1}^{n}\{x_{i}\}$ . Consequently, the number $\max_{y\in \mathcal{M}}\min_{i = 1,\dots,n}\| y - x_i\|$ is the maximum distance of points in $\mathcal{M}$ from the collection $\bigcup_{i = 1}^{n}\{x_{i}\}$ , and the outer minimization in (1) optimizes this maximum distance over the centers $(x_{1},\ldots ,x_{n})$ . The quantity $\epsilon (\mathcal{M},n)$ is known as the covering radius of the set $\mathcal{M}$ with $n$ metric balls as defined in (Alimov and Tsar'kov, 2021, p. 300); solving (1) yields an optimal $n$ -point representation — a Chebyshev $n$ -net — of $\mathcal{M}$ , and $\epsilon (\mathcal{M},n)$ is the error guarantee in such a representation. + +Chebyshev $n$ -nets are especially useful for estimating errors in function learning in the presence of certain regularity properties. For instance, consider a target function $f: \mathcal{M} \to \mathbb{R}$ that is known to be Lipschitz continuous with modulus $L_0$ but is otherwise unknown, and suppose that a Chebyshev $n$ -net $\mathcal{F}$ of $\mathcal{M}$ and the set of values $\{f(x) \mid x \in \mathcal{F}\}$ of $f$ on $\mathcal{F}$ are available. Then a tight estimate of the value of $f$ at a generic $y \in \mathcal{M}$ may be obtained, within an absolute error margin of $L_0\epsilon(\mathcal{M}, n)$ , by finding $y_\star \in \arg \min_{z \in \mathcal{F}} \| z - y \|$ and then calculating $f(y_\star)$ . By an identical reasoning, for Hölder continuous $f$ of modulus $C > 0$ and rate $\alpha \in ]0, 1[$ , the corresponding error bound is $C\epsilon(\mathcal{M}, n)^\alpha$ . Similar arguments in optimization problems involving uniformly continuous objective functions permit the derivation of quantitative estimates of true optima from estimates over finitely many points. + +In several applications of covering, the centers of the metric balls may be restricted to lie within $\mathcal{M}$ , in which case, the problem (1) should be modified to + +$$ +\epsilon^ {\prime} (\mathcal {M}, n) := \min _ {(x _ {1}, \dots , x _ {n}) \in \mathcal {M} ^ {n}} \max _ {y \in \mathcal {M}} \min _ {i = 1, \dots , n} \| y - x _ {i} \|. \tag {2} +$$ + +This constrained version is relevant in settings where the representative points must strictly belong to the set under consideration. The triangle inequality immediately yields the sandwich relation $\frac{\epsilon'(\mathcal{M}, n)}{2} \leqslant \epsilon(\mathcal{M}, n) \leqslant \epsilon'(\mathcal{M}, n)$ . Our efforts are focused on the numerical viability of (1) for compact $\mathcal{M}$ ; similar methods can be adapted with minor modifications to address (2) if $\mathcal{M}$ is convex. + +# Contributions + +The problem (1) is challenging (NP-hard) even in the case $n = 1$ . Indeed, this situation is the so-called Chebyshev center problem (Alimov and Tsar'kov, 2021, p. 296) and is of central importance in optimal function learning (Binev et al., 2024), signal processing (Micchelli and Rivlin, 1977; Eldar et al., 2008; Wu et al., 2013), approximation theory (Alimov and Tsar'kov, 2019; Foucart and Liao, 2024), and a host of other disciplines. Despite its complexity, it turns out that the Chebyshev center problem admits a numerically viable solution technique (with excellent approximation properties) by leveraging certain convexity attributes present therein, and has been recently studied in (Paruchuri and Chatterjee, 2023). The case of $n > 1$ considered in the current article is significantly more difficult: of course, it is challenging (NP-hard), and in addition, it does not appear to be amenable to any simplification via convexity. Moreover, while the Chebyshev center of $\mathcal{M}$ is unique for the Euclidean norm, the Chebyshev $n$ -net problem may not yield unique solutions. + +Since (1) is well-known to be computationally challenging, we focus attention on the development of viable algorithm based on numerical gradients (i.e., a zeroth-order algorithm) with good approximation properties to solve (1): + +To this end, we demonstrate that the inner max-min problem in (1) admits a reformulation that yields good regularity properties parametrically in $(x_{1},\ldots ,x_{n})$ corresponding to the outer minimization + +under mild hypotheses; this is our first main result Theorem 3.2. The indicated regularity property enables and justifies the deployment of a numerical gradient-based (zeroth-order) algorithm in the variables $(x_{1},\ldots ,x_{n})$ to arrive at a solution to (1) itself. To our knowledge, ours is the only instance of a gradient-based approach to finding Chebyshev $n$ -nets for compact subsets of $\mathbb{R}^N$ that do not necessarily possess special structure. + +In order to obtain the indicated regularity properties, the key technical tools employed here are derived from the field of sensitivity of optimization problems (Fiacco, 1983; Fiacco and Ishizuka, 1990) or variational analysis (Dontchev and Rockafellar, 2014). We establish the crucial Lipschitz continuity of the (parametric) inner max-min problem in (1) with respect to $(x_{1},\ldots ,x_{n})$ around optimal points of the outer minimization; this lays down the foundation for the deployment of a range of numerical gradient-based algorithms for the outer minimization. +o Numerical gradient-based algorithms for nonlinear programs are not expected to converge to global optimizers in general. Here we lift off-the-shelf a recent zeroth-order algorithm that stochastically smoothes the numerical gradients and explores the decision landscape in search for global optima; the precise complexity bound appears in our second main result Theorem 3.6. More refined algorithms involving, e.g., momenta for convergence to global optimizers are topics of subsequent investigations. This is the first instance, to the best of our knowledge, that powerful tools from sensitivity theory have been employed in conjunction with numerical gradient-based algorithms to address the Chebyshev $n$ -net problem. +- The efficacy of our algorithm depends on the availability of accurate solutions to the inner max-min problem (after an appropriate reformulation) for each fixed $(x_{1},\ldots ,x_{n})$ . The indicated parametric max-min problem happens to be non-convex despite our reformulation, and the speed of execution of our algorithm depends on how quickly these globally optimal solutions are made available. Our numerical experiments indicate that despite the lack of convexity, our algorithm (including resets) gives excellent performance even for relatively complicated nonconvex sets $\mathcal{M}$ — see §4 for details. Scalability of our algorithm with increasing dimensions $N$ and number $n$ of centers remains a challenging issue: while the numerical gradient-based algorithm for the outer minimization scales in these variables as $\mathcal{O}(N^{\frac{3}{2}}n^{2})$ , the inner max-min is non-convex and remains the chief computational bottleneck in our algorithm. + +Since the Chebyshev $n$ -net problem is understood to be hard, several randomized algorithms to approximately solve (1) have been developed; see, e.g., (Har-Peled, 2011, §5.3) for an account. While many of these randomized techniques involve independent and identically distributed sampling, more complicated history-dependent stochastic iterative algorithms have been advanced in the literature for special cases — see, e.g., (Ushakov and Lebedev, 2015) (and the references therein) for the case of $N = 3$ . Other techniques to solve the optimal $n$ -net and related problems relevant to theoretical computer science and computational geometry may be found in (Malysheva, 2020; Yu, 2021; Voronov et al., 2023). It is, of course, possible to employ zeroth-order methods (such as simulated annealing (Robert and Casella, 2004)) that do not rely on any regularity of the inner max-min in (1). However, the Lipschitz regularity of the parametric max-min (Theorem 3.2) paves the way for fine-tuned algorithms that exploit this regularity compared to typically slow annealing techniques; as such, Theorem 3.2 constitutes the key point of departure of our contributions from the current literature. + +# Organization + +This article unfolds as follows. Several preparatory results are contained in $\S 2$ ; they are employed in the two main results centered around our key Algorithm 1 in $\S 3$ . Numerical experiments are presented in $\S 4$ and the technical data corresponding to both our theoretical results and numerical experiments are collected in $\S A - \S D$ . + +# 2 Preliminaries + +It is intuitive to expect that an optimizer $(x_1^*,\ldots ,x_n^*)$ of (1) should be restricted to a bounded set that contains $\mathcal{M}$ . Recall that $\epsilon (\mathcal{M},1)$ is the Chebyshev radius of $\mathcal{M}$ , and an optimizer $z^{*}$ of (1) + +for $n = 1$ is the Chebyshev center of $\mathcal{M}$ . In other words, the Chebyshev center of $\mathcal{M}$ provides the best single-point representation of the set $\mathcal{M}$ , minimizing the worst-case distance to all points in $\mathcal{M}$ . There is a neat relationship between the $n$ -net problem (1) and the Chebyshev center of $\mathcal{M}$ established in the following proposition, a proof of which is in §B. + +Proposition 2.1. If $(x_1^*,\ldots ,x_n^*)$ is an optimizer of (1) and if $z^{*}$ is a Chebyshev center of $\mathcal{M}$ , then + +$$ +\left\| z ^ {*} - x _ {i} ^ {*} \right\| \leqslant 2 \epsilon (\mathcal {M}, 1) \quad f o r a l l i \in \{1, \dots , n \}. \tag {3} +$$ + +In other words, without loss of generality, all optimizers of (1) can be restricted to a ball of twice the Chebyshev radius, centered at the Chebyshev center. Henceforth, we denote this ball by $\mathbb{B}(\mathcal{M})$ whereby, it follows that $(x_1^*,\ldots ,x_n^*)\in \mathbb{B}(\mathcal{M})^n$ + +This result is useful for generating initial guesses and bounding our search space when implementing our algorithm, which will be discussed in subsequent sections. Inspection of (1) reveals that $\min_{i = 1,\dots,n}\| y - x_i\|$ is non-convex in $(x_{1},\ldots ,x_{n})$ for fixed $y$ and vice versa. + +As indicated in §1, we reformulate the problem in a way suitable for the construction of a numerically viable algorithm to extract near optimal solutions. To this end, we first note that the geometric implication of the problem guarantees strict feasibility for any compact $\mathcal{M}$ . Moreover, for a fixed set of centers, the inner maximization always has a finite optimal value owing to the compactness of $\mathcal{M}$ . We now present a straightforward reformulation of (1) using a slack variable, which, despite its simplicity, provides the problem with a richer structure for computational viability. Since the square of the norm ensures continuous differentiability, a property that will become important subsequently, we work with the squared norm by noting that + +$$ +\min _ {i = 1, \dots , n} \| y - x _ {i} \| ^ {2} = \max \left\{t \in \mathbb {R} \mid t \leqslant \| y - x _ {i} \| ^ {2} \text {f o r a l l} i = 1, \dots , n \right\}, +$$ + +which immediately leads to + +$$ +\epsilon (\mathcal {M}, n) ^ {2} = \min _ {\left(x _ {1}, \dots , x _ {n}\right) \in \mathbb {R} ^ {N \times n}} \max _ {y \in \mathcal {M}} \quad \max _ {t \in \mathbb {R}} \quad t \tag {4} +$$ + +$$ +\text {s u b j e c t} \quad t - \| y - x _ {i} \| ^ {2} \leqslant 0 \quad \text {f o r a l l} i = 1, \dots , n. +$$ + +It is natural to seek a unification of both inner maximizations into a single maximization over the Cartesian product of their respective domains. The following proposition, whose proof is in §B, establishes that this unification is valid for the given problem. + +Proposition 2.2. For a given set of centers $(x_{1},\ldots ,x_{n})\in \mathbb{R}^{N\times n}$ , if $(t^{\dagger},y^{\dagger})$ is an optimizer of + +$$ +\left. \max \left\{\max \left\{t \in \mathbb {R} \mid t - \| y - x _ {i} \| ^ {2} \leqslant 0 \text {f o r a l l} i = 1, \dots , n \right\} \mid y \in \mathcal {M} \right\} \right. \tag {5} +$$ + +and $(t^{*},y^{*})$ is an optimizer of + +$$ +\left. \max \left\{t \mid (t, y) \in \mathbb {R} \times \mathcal {M} a n d t - \| y - x _ {i} \| ^ {2} \leqslant 0 f o r a l l i = 1, \dots , n \right\}, \right. \tag {6} +$$ + +then $t^{\dagger} = t^{*}$ . That is, both (5) and (6) are feasible and have the same optimal value. + +In view of of Proposition 2.2, from this point forward, we adopt the following reformulation for the $n$ -covering radius: + +$$ +\epsilon (\mathcal {M}, n) ^ {2} = \min _ {\left(x _ {1}, \dots , x _ {n}\right) \in \mathbb {R} ^ {N \times n}} \mathcal {G} \left(x _ {1}, \dots , x _ {n}\right), \tag {7} +$$ + +where + +$$ +\mathcal {G} \left(x _ {1}, \ldots , x _ {n}\right) := \max \Big \{t \big | (t, y) \in \mathbb {R} \times \mathcal {M} \text {a n d} t - \| y - x _ {i} \| ^ {2} \leqslant 0 \text {f o r a l l} i = 1, \ldots , n \Big \}. +$$ + +Evaluating $\mathcal{G}$ entails solving a nonconvex problem, and the gradient of $\mathcal{G}$ is difficult to calculate analytically, especially without imposing further structure on the inner maximization. There are several 'gradient-free' algorithms such as simulated annealing via Markov chain Monte Carlo to ensure probabilistic convergence to global minima, simplicial algorithms (e.g., the Nelder-Meade) for deterministic convergence to global minima, of $\mathcal{G}$ , etc. Although these algorithms do not require further structure on $\mathcal{M}$ and may be easy to implement since they only employ point evaluations of + +$\mathcal{G}$ , numerical gradient/subgradient-based algorithms tend to furnish solutions more efficiently and scale well relative to the size of the problem data, provided the objective function satisfies certain regularity conditions. An algorithm centered around numerical gradients is our target. + +To advance in this direction, we impose the condition that $\mathcal{M}$ is the intersection of 0-sublevel sets of finitely many known continuously differentiable functions in the following way. Let $\mathcal{O}_1,\ldots ,\mathcal{O}_k\in$ $\mathbb{R}^N$ be open sets containing $\mathcal{M}$ , let $\varphi_j:\mathcal{O}_j\to \mathbb{R}$ for $j = 1,\dots k$ , be continuously differentiable functions, and suppose that $\mathcal{M}$ is realized as + +$$ +\mathcal {M} = \left\{y \in \mathbb {R} ^ {N} \mid \varphi_ {j} (y) \leqslant 0 \text {f o r} j = 1, \dots , k \right\}. +$$ + +This allows us to write $\mathcal{G}$ as the negative of the optimal value function of a standard parametric optimization problem + +$$ +- \mathcal {G} \left(x _ {1}, \dots , x _ {n}\right) = \min _ {(t, y)} \quad - t +$$ + +$$ +\text {s u b j e c t} \quad \left\{ \begin{array}{l} t - \| y - x _ {i} \| ^ {2} \leqslant 0 \quad \text {f o r a l l} i = 1, \dots , n, \\ \varphi_ {i - n} (y) \leqslant 0 \quad \text {f o r a l l} i = n + 1, \dots , n + k, \\ t \geqslant 0, \\ (t, y) \in \mathbb {R} \times \mathbb {R} ^ {N}. \end{array} \right. \tag {8} +$$ + +Example 2.3. The problem (8), for a given choice of $(x_{1},\ldots ,x_{n})$ , need not have strict local optimizers in general. To see this, let $\mathcal{M}\subset \mathbb{R}^2$ be a disc of radius $R > 0$ centered at the origin and let $n = 1$ . The evaluation of $\mathcal{G}$ at the origin leads to non-isolated optimizers for (8) since any pair of the form $(R^{2},y^{*})$ with $\| y^{*}\| = R$ is an optimizer. + +Remark 2.4. Note that it is not necessary to explicitly impose the condition $t \geqslant 0$ in (8), as the optimal value remains unchanged and is always positive. However, we shall see below that including this constraint ensures compactness of the feasible region — a property whose significance will be addressed subsequently. One can observe that if the functions $\varphi_j$ for $j = 1, \dots, k$ , are convex (which in turn implies that $\mathcal{M}$ is convex), then the problem (8) becomes a DC (Difference of Convex functions) program since the objective and the constraints are differences of two convex functions. The reader is referred to (An et al., 2014) for a comprehensive overview of DC programming and algorithms to solve problems with DC objectives and constraints up to criticality; see, e.g., (Mordukhovich, 2018, Chapter 7) for theoretical insights. Although our results do not require the functions to be convex, when they are, DC programming algorithms may be used to achieve faster convergence to criticality in numerical implementations. + +# 3 Main results + +We now outline some technical notations and definitions in the context of the problem (8) to set the stage for our analysis. The Lagrangian $\mathcal{L}:\mathbb{R}^{N\times n}\times (\mathbb{R}\times \mathbb{R}^N)\times [0, + \infty [^{n + k + 1}\to \mathbb{R}$ of (8) is + +$$ +\mathcal {L} (x, (t, y), \lambda) := - (\lambda_ {n + k + 1} + 1) t + \sum_ {i = 1} ^ {n} \lambda_ {i} \left(t - \| y - x _ {i} \| ^ {2}\right) + \sum_ {i = n + 1} ^ {n + k} \lambda_ {i} \varphi_ {i - n} (y), \tag {9} +$$ + +where $x$ denotes $(x_{1},\ldots ,x_{n})$ + +The feasible region is denoted by the set-valued map $S_{\mathrm{feas}}:\mathbb{R}^{N\times n}\Rightarrow \mathbb{R}\times \mathbb{R}^N$ + +$$ +S _ {\text {f e a s}} (\bar {x}) := \left\{(t, y) \in \mathbb {R} \times \mathbb {R} ^ {N} \left| \begin{array}{l l} t - \| y - \bar {x} _ {i} \| ^ {2} \leqslant 0 & \text {f o r a l l} i = 1, \dots , n, \\ \varphi_ {i - n} (y) \leqslant 0 & \text {f o r a l l} i = n + 1, \dots , n + k, \\ t \geqslant 0 & \end{array} \right. \right\}, \tag {10} +$$ + +and the set of global optimizers is the set-valued map $S_{\mathrm{opt}}:\mathbb{R}^{N\times n}\Rightarrow \mathbb{R}\times \mathbb{R}^N$ + +$$ +S _ {\text {o p t}} (\bar {x}) := \left\{\left(t, y\right) \in S _ {\text {f e a s}} (\bar {x}) \mid t = - \mathcal {G} (\bar {x}) \right\}. \tag {11} +$$ + +For a choice $\bar{x}$ of $x$ , let $(\bar{t},\bar{y}) \in S_{\mathrm{opt}}(\bar{x})$ . We define the set of active constraint indices at $(\bar{x},(\bar{t},\bar{y}))$ as + +$$ +I (\bar {x}, (\bar {t}, \bar {y})) := \left\{i \in \{1, \dots , n \} \mid \bar {t} - \| \bar {y} - \bar {x} _ {i} \| ^ {2} = 0 \right\} \cup \left\{i \in \{n + 1, \dots , n + k \} \mid \varphi_ {i - n} (\bar {y}) = 0 \right\}. \tag {12} +$$ + +The constraint $t \geqslant 0$ is not included since $t$ remains strictly positive at optimality. The set of Karush-Kuhn-Tucker (KKT) vectors of (8) for a pair $(\bar{x}, (\bar{t}, \bar{y})) \in \mathbb{R}^{N \times n} \times (\mathbb{R} \times \mathbb{R}^N)$ is denoted by + +$$ +\Lambda (\bar {x}, (\bar {t}, \bar {y})) := \left\{\lambda \in [ 0, + \infty [ ^ {n + k + 1} \left| \begin{array}{l} \nabla_ {(t, y)} \mathcal {L} (\bar {x}, (\bar {t}, \bar {y}), \lambda) = 0, \\ \lambda_ {i} (\bar {t} - \| \bar {y} - \bar {x} _ {i} \| ^ {2}) = 0 \text {f o r} i = 1, \dots , n \\ \lambda_ {i} \varphi_ {n - i} (\bar {y}) = 0 \text {f o r} i = n + 1, \dots , n + k, \text {a n d} \\ \lambda_ {n + k + 1} \bar {t} = 0 \end{array} \right. \right\} \tag {13} +$$ + +Definition 3.1 (MFCQ for (8)). For $x \in \mathbb{R}^{N \times n}$ , the Mangasarian-Fromovitz constraint qualification condition is said to hold at $(\bar{t}, \bar{y}) \in S_{\mathrm{opt}}(\bar{x})$ if there exists a direction $w \in \mathbb{R}^{N + 1}$ such that the inner product of $w$ with the gradient of each active constraint is strictly negative. To wit, the following conditions must hold: + +(3.1-i) $\left( \begin{array}{cc}1 & -2(\bar{y} -\bar{x}_i)^\top \end{array} \right)w < 0\quad \mathrm{for~all~}i\in \{1,\ldots ,n\} \cap I(\bar{x},(\bar{t},\bar{y}))$ , and + +(3.1-ii) $\nabla_{y}\varphi_{i}(\bar{y})^{\top}w < 0$ for all $i\in \{n + 1,\dots ,n + k\} \cap I(\bar{x},(\bar{t},\bar{y}))$ + +We now present our first main result establishing Lipschitz continuity of $\mathcal{G}$ . + +Theorem 3.2. Consider the problem (8) and grant the notations established above. Fix $\bar{x} \in \mathbb{R}^{N \times n}$ and suppose that the MFCQ condition (defined in Definition 3.1) holds for every $(\bar{t}, \bar{y}) \in S_{\mathrm{opt}}(\bar{x})$ . Then the function $\mathcal{G}$ is locally Lipschitz continuous at $\bar{x}$ . Consequently, its restriction $\mathcal{G}|_{\mathbb{B}(\mathcal{M})^n}$ to $\mathbb{B}(\mathcal{M})^n$ (c.f. Proposition 2.1) is Lipschitz. Moreover, for any $v \in \mathbb{R}^{N \times n}$ , we have + +$$ +\begin{array}{l} \inf _ {(\bar {t}, \bar {y}) \in S _ {\mathrm {o p t}} (\bar {x})} \min _ {\lambda \in \Lambda (\bar {x}, (\bar {t}, \bar {y}))} \frac {\partial \mathcal {L}}{\partial x} (\bar {x}, (\bar {t}, \bar {y}), \lambda) \cdot v \leqslant - \lim _ {h \downarrow 0} \inf _ {h} \frac {1}{h} (\mathcal {G} (\bar {x} + h v) - \mathcal {G} (\bar {x})) \\ \leqslant - \lim _ {h \downarrow 0} \sup _ {h \downarrow 0} \frac {1}{h} (\mathcal {G} (\bar {x} + h v) - \mathcal {G} (\bar {x})) \leqslant \inf _ {(\bar {t}, \bar {y}) \in S _ {\mathrm {o p t}} (\bar {x})} \max _ {\lambda \in \Lambda (\bar {x}, (\bar {t}, \bar {y}))} \frac {\partial \mathcal {L}}{\partial x} (\bar {x}, (\bar {t}, \bar {y}), \lambda) \cdot v. \tag {14} \\ \end{array} +$$ + +Proof. We observe that both the objective and constraint functions are continuously differentiable. At each $x \in \mathbb{R}^{N \times n}$ , the feasible region $S_{\mathrm{feas}}(\bar{x})$ is non-empty by definition. Letting $N(\bar{x})$ denote a neighborhood of $\bar{x}$ , we note that the mapping $S_{\mathrm{feas}}$ is uniformly compact near $\bar{x}$ since the set $\bigcup_{x \in N(\bar{x})} S_{\mathrm{feas}}(x)$ always satisfies the containment relation + +$$ +\bigcup_ {x \in N (\bar {x})} S _ {\text {f e a s}} (x) \subset \left[ 0, \max _ {x \in N (\bar {x})} \max _ {y \in \mathcal {M}} \max _ {i = 1, \dots , n} \| y - x _ {i} \| ^ {2} \right] \times \mathcal {M}. +$$ + +Furthermore, a set independent of the base point $\bar{x}$ containing $\bigcup_{x\in N(\bar{x})}S_{\mathrm{feas}}(x)$ can be realized if each component of $\bar{x}$ is restricted to the interior of the ball described in Proposition 2.1, in which case we obtain the uniform containment $\bigcup_{x\in N(\bar{x})}S_{\mathrm{feas}}(x)\subset [0,9\epsilon (\mathcal{M},1)^2 ]\times \mathcal{M};$ the set on the right-hand side is, of course, bounded. Therefore, (Fiacco and Ishizuka, 1990, Theorem 4.2) — reproduced for completeness in §C as Theorem C.1 — applies directly to our problem since the MFCQ conditions hold for every $(\bar{t},\bar{y})\in S_{\mathrm{opt}}(\bar{x})$ by hypothesis. Consequently, $\mathcal{G}$ is locally Lipschitz continuous at $\bar{x}$ , and its lower and upper Dini derivatives satisfy the bounds described in (14). Invoking (Cobzas et al., 2019, Theorem 2.1.6) gives us Lipschitz continuity of $\mathcal{G}|_{\mathbb{B}(\mathcal{M})^n}$ , and our proof is complete. + +From this point forward, we shall operate under the blanket assumption that MFCQ holds at every $(\bar{t},\bar{y})\in S_{\mathrm{opt}}(\bar{x})$ for all $\bar{x}\in \mathbb{R}^{N\times n}$ a property that can, in principle, be verified once the functions $\varphi_1,\ldots ,\varphi_k$ , are specified. + +Remark 3.3. We highlight that reformulating (1) into (8) using the variable $t$ and representing $\mathcal{M}$ through continuously differentiable functions $\varphi_1, \ldots, \varphi_k$ , was crucial for establishing Theorem 3.2. The existing literature does not appear to have employed regularity properties of marginal functions for the design of optimal covering algorithms, which is a point of departure of our contribution. + +Remark 3.4. Observe that, in principle, every closed subset of $\mathbb{R}^N$ admits a representation as the zero-level set of a smooth function (Calderón and Zygmund, 1961) that is realized as a smooth regularization of the distance-to-the-set function. However, such a function may be difficult to encode in a finitary way for computational purposes. In this light, while our assumption that the set $\mathcal{M}$ is realized as the intersection of zero-sublevel sets of finitely many continuously differentiable functions may appear to be restrictive, it is a reasonably weak assumption, and this family of sets includes sublevel sets of polynomials in particular. + +We now turn to the minimization of $\mathcal{G}$ . In view of Proposition 2.1, it suffices to restrict the search space to $\mathbb{B}(\mathcal{M})^n$ since all global minimizers of $\mathcal{G}$ are guaranteed to lie within this set. Moreover, by Theorem 3.2, we see that $\mathcal{G}|_{\mathbb{B}(\mathcal{M})^n}$ is $L$ -Lipschitz. Now we leverage recent advances in nonconvex and nonsmooth optimization using randomized smoothing by (Lin et al., 2022; Liu et al., 2024) to design our numerical algorithm. Specifically, we adopt an optimality framework based on generalized Goldstein stationary points (Goldstein, 1977) and develop an algorithm that provably converges to such a point based solely on oracle evaluations of $\mathcal{G}$ . Furthermore, we establish a polynomial bound on the number of such oracle calls required for convergence. We proceed by laying out a set of definitions. For a detailed exposition, we refer the reader to (Lin et al., 2022). + +Definition 3.5. Given a point $x \in \mathbb{R}^{N \times n}$ and a direction $v \in \mathbb{R}^{N \times n}$ , the generalized directional derivative of $\mathcal{G}$ is defined as $D\mathcal{G}(x;v) := \lim \sup_{y \to x, t \downarrow 0} \frac{\mathcal{G}(y + tv) - \mathcal{G}(y)}{t}$ . Then, the generalized gradient (Clarke, 1990) of $\mathcal{G}$ is defined as the set + +$$ +\partial \mathcal {G} (x) := \left\{g \in \mathbb {R} ^ {N \times n} \mid \langle g, v \rangle \leqslant D \mathcal {G} (x; v), \text {f o r a l l} v \in \mathbb {R} ^ {d} \right\}. \tag {15} +$$ + +Let $\mathbb{B}(x,\delta) := \{y\in \mathbb{R}^{N\times n}\mid \| y - x\| \leqslant \delta \}$ . Given a $\delta \geqslant 0$ , the Goldstein $\delta$ -subdifferential (Goldstein, 1977) of $\mathcal{G}$ at $x$ is defined as $\partial_{\delta}\mathcal{G}(x)\coloneqq \mathrm{conv}\left(\bigcup_{y\in B(x,\delta)}\partial \mathcal{G}(y)\right)$ . Given a $\delta \geqslant 0$ and $\gamma ,\varepsilon >0$ , a point $x\in \mathbb{R}^{N\times n}$ is called a $(\gamma ,\delta ,\varepsilon)$ -generalized Goldstein stationary point (Liu et al., 2024) of $\mathcal{G}|_{\mathbb{B}(\mathcal{M})^n}$ if $\min \left\{\frac{1}{\gamma}\left\| x - \mathrm{Proj}_{\mathbb{B}(\mathcal{M})^n}(x - \gamma g)\right\| \Bigg{|}g\in \partial_{\delta}\mathcal{G}(x)\right\} \leqslant \varepsilon$ , where $\mathrm{Proj}_{\mathbb{B}(\mathcal{M})^n}(y)$ denotes the orthogonal projection of $y$ onto $\mathbb{B}(\mathcal{M})^n$ . + +We now present our algorithm — an adaptation of techniques from (Lin et al., 2022; Liu et al., 2024) with minor modifications suited to the minimization of $\mathcal{G}$ restricted to $\mathbb{B}(\mathcal{M})^n$ . + +Algorithm 1 gradOptNet: a numerical gradient-based optimal covering algorithm +1: Input: Initial point $x_0\in \mathbb{B}(\mathcal{M})^n$ , stepsize $\gamma >0$ , smoothing parameter $\delta$ , iteration number $T\geqslant 1$ , and parameters $b_{1},b_{2}$ , and $q$ +2: for $t = 0,1,\ldots ,T - 1$ do +3: if mod $(t,q) = 0$ then +4: Sample $w_{1,t},\dots ,w_{b_1,t}$ uniformly from the unit sphere in $\mathbb{R}^{N\times n}$ +5: Let $g_{i,t} = \frac{Nn}{2\delta}\left(\mathcal{G}(x_t + \delta w_{i,t}) - \mathcal{G}(x_t - \delta w_{i,t})\right)w_{i,t}$ for each $i\in \{1,\dots ,b_1\}$ +6: Set $v_{t} = \frac{1}{b_{1}}\sum_{i = 1}^{b_{1}}g_{i,t}$ +7: else +8: Sample $w_{1,t},\dots ,w_{b_2,t}$ uniformly from the unit sphere in $\mathbb{R}^{N\times n}$ +9: Let $g_{i,t} = \frac{Nn}{2\delta}\left(\mathcal{G}(x_t + \delta w_{i,t}) - \mathcal{G}(x_t - \delta w_{i,t})\right)w_{i,t}$ for each $i\in \{1,\dots ,b_2\}$ +10: Let $g_{i,t - 1} = \frac{Nn}{2\delta}\left(\mathcal{G}(x_{t - 1} + \delta w_{i,t}) - \mathcal{G}(x_{t - 1} - \delta w_{i,t})\right)w_{i,t}$ for each $i\in \{1,\dots ,b_2\}$ +11: Set $v_{t} = \frac{1}{b_{2}}\sum_{i = 1}^{b_{2}}(g_{i,t} - g_{i,t - 1}) + v_{t - 1}$ +12: end if +13: Update $x_{t + 1} = \mathrm{Proj}_{\mathbb{B}(\mathcal{M})^n}(x_t - \gamma v_t)$ +14: end for +15: return $x_R$ , where $R\in \{0,1,\dots ,T - 1\}$ is uniformly sampled + +Theorem 3.6. Consider the problem (8) and grant the notations established above. Let $L$ denote the Lipschitz constant of $\mathcal{G}|_{\mathbb{B}(\mathcal{M})^n}$ . With $b_1 = \mathcal{O}\left(\frac{NnL^2}{\varepsilon^2}\right)$ , $b_2 = q = \mathcal{O}\left(\frac{\sqrt{Nn}L}{\varepsilon}\right)$ , and $\gamma = \frac{\delta}{2NnL}$ , the Algorithm 1 requires at most $\mathcal{O}\left(N^{\frac{3}{2}}n^{2}\epsilon (\mathcal{M},1)L^{3}\delta^{-1}\varepsilon^{-3}\right)$ calls of $\mathcal{G}$ to obtain a $(\gamma ,\delta ,\varepsilon)$ -generalized Goldstein stationary point of $\mathcal{G}$ in expectation. + +Proof. In view of Theorem 3.2, we have Lipschitz continuity of $\mathcal{G}|_{\mathbb{B}(\mathcal{M})^n}$ . We observe that Algorithm 1 corresponds to a variant of (Liu et al., 2024, Algorithm 3), adapted to the setting in which the objective function is deterministic. Noting that the diameter of $\mathbb{B}(\mathcal{M})^n$ is $4\sqrt{n}\epsilon (\mathcal{M},1)$ , we invoke (Liu et al., 2024, Corollary 5.4) to obtain the complexity bound in the assertion. + +Remark 3.7. Theorem 3.2 provided the substrate to enable numerical gradient-based algorithms to solve (1). Since the outer minimization problem in (1) is non-convex, numerical gradient-based + +algorithms can only guarantee convergence to stationary points as in 3.6. We attempt to obtain global solutions to (1) in two steps. First, Algorithm 1 involves the calculation of difference quotients and their randomized smoothing; the smoothing is necessary for obtaining 'averaged' descent directions to solve non-convex optimization problems up to stationarity, and this is performed along with a variance reduction scheme as proposed in (Liu et al., 2024). Second, global convergence to minimizers (as opposed to stationary points) is facilitated by the employment of several randomized 'restarts' (parallel initializations) in the indicated local scheme. + +# 4 Numerical experiments + +This section details the numerical experiments carried out using Algorithm 1. Heuristic observations in (Liu et al., 2024, Section 6) align with our observations: faster convergence is observed when the variance reduction step is omitted (i.e., $q = 1$ , rendering $b_{2}$ immaterial). We also observe better performance when $b_{1}$ is increased in response to larger values of $n$ or $N$ . Accordingly, we set $q = 1$ for all our experiments. For the ease of visualization, we present examples with $N = 2$ and $n < 10$ in §4; an illustration with $N = 3$ is presented in Example D.2 in §D. At each point, $\mathcal{G}$ is evaluated using the NLP solver BARON (Sahinidis, 1996), known for its global optimization capabilities. To the best of our knowledge, no existing benchmarks or ground truth results are available in the literature, and therefore, our comparisons of computation time and accuracy are with techniques like simulated annealing that do not leverage the Lipschitz continuity of $\mathcal{G}$ for outer minimization. + +Example 4.1. Consider the case where $N = 2$ , $k = 1$ , and the set $\mathcal{M}$ is the solid ellipse + +$$ +\mathcal {M} := \left\{y := \left(y _ {1}, y _ {2}\right) \in \mathbb {R} ^ {2} \mid \varphi_ {1} (y) := \frac {y _ {1} ^ {2}}{9} + \frac {y _ {2} ^ {2}}{4} - 1 \leqslant 0 \right\}. \tag {16} +$$ + +The application of Algorithm 1 with multiple parallel initializations (to enhance global convergence efficiency), led to the results presented in Fig. 1 for number of balls $n = 2,3,4,5,7$ and 8. + +![](images/0dfe0fa99a845e158294ca34af3f7acac3777fb947250d7ae9996706a1974e75.jpg) + +![](images/06bf3c856c4f57a4a6e6ee84ec8499c1699ccc7c360dc96027a452444bf02b80.jpg) + +![](images/be0f8a56ee4fd197048771001bca3cd749aa8ae052593503b08967300e961d93.jpg) + +![](images/28b8da002f5acc017bd4c7006c734b25fad17ce6cac61e94ac1977a9f553d7ae.jpg) +(a) $\epsilon (\mathcal{M},2) = 2.169071$ +(d) $\epsilon (\mathcal{M},5) = 1.494066$ +Figure 1: Covering of the ellipse (16) for varying values of $n$ . + +![](images/6edf18782f8ba1d32665c46da134a1b91ef1294b1060e5a8c30c54b7d0ce0774.jpg) +(b) $\epsilon (\mathcal{M},3) = 2.016456$ +(e) $\epsilon (\mathcal{M},7) = 1.220183$ + +![](images/113f742fe589c4ee6c2e053e2f40d8dea45ca0128f6d2dea6a4276fbb4f58c7f.jpg) +(c) $\epsilon (\mathcal{M},4) = 1.667632$ +(f) $\epsilon (\mathcal{M},8) = 1.140047$ + +The results exhibit a visually interpretable structure primarily due to the inherent symmetry of the set $\mathcal{M}$ . This naturally motivates the expectation of symmetric structures in the optimal coverings—a property that is indeed observed in the solutions reported in Fig. 1. The evolution of the objective + +value for the case $n = 4$ against the iterations of Algorithm 1, with $\gamma = 0.05$ , $\delta = 0.01$ , $b_{1} = 24$ , and $q = 1$ is shown in Fig. 4 featuring 4 restarts. For each $n$ , Algorithm 1 executed with 10 restarts yielded final radius values with relative standard deviation $< 1\%$ , indicating excellent robustness. Table 1 tabulates and compares the computation time and solution quality for Algorithm 1 vs Simulated Annealing (Metropolis Hastings sampling and geometric cooling rate) for the example with $n = 4$ . + +
Methodε(M,4)↓Number of evaluations of G↓
Trial 1Trial 2Trial 3
gradOptNet (Ours)1.6821.7061.6984800
Simulated Annealing2.2272.5742.30110000
+ +Table 1: Comparison of computation time and solution quality across methods. + +Remark 4.2. We emphasize that the ability of the inner solver used to evaluate $\mathcal{G}$ to attain global optimality is crucial for the effectiveness of Algorithm 1. If a solver lacking global optimality guarantees (such as IPOPT) is employed, it may fail to produce a valid covering of $\mathcal{M}$ . This phenomenon is treated in Example D.1. + +Example 4.3. Let us consider an example where $\mathcal{M}$ is asymmetric and nonconvex, and to this end, consider the case where $N = 2$ , $k = 4$ , and the set $\mathcal{M}$ is the region + +$$ +\mathcal {M} := \left\{\left(y _ {1}, y _ {2}\right) \in \mathbb {R} ^ {2} \mid \varphi_ {j} (y) \leqslant 0 \text {f o r} j = 1, 2, 3, 4 \right\} \text {w h e r e , w i t h} y := \left(y _ {1}, y _ {2}\right), +$$ + +$$ +\varphi_ {1} (y) := 2 y _ {1} ^ {2} - y _ {2}, \varphi_ {2} (y) := y _ {2} - 2 \left(y _ {1} - 1\right) ^ {2}, \tag {17} +$$ + +$$ +\varphi_ {3} (y) := y _ {2} - 5 \left(y _ {1} + 0. 1\right) ^ {2}, \text {a n d} \varphi_ {4} (y) := - 0. 1 - y _ {1}. +$$ + +Once again, Algorithm 1 was employed with multiple parallel initializations, and the results presented in Fig. 2 were obtained for $n = 1, 2, 3, 4, 5, 6$ . + +![](images/f72c6517171f370bf3a721a5c2b1bce1c16f26191703948ee3242eed87c1e66f.jpg) + +![](images/1ea6c0c42dc83035216a07858cc05ae984624e460306e13ff2ba1ff93d599405.jpg) + +![](images/87ca3d5a3309f70d29008e505454d2f2bf69cb0144ebf42ee93d274aabec11f2.jpg) + +![](images/3767e5baf72dc76b03ef720840b5f8c0494d13499186f012aed053d07768c31f.jpg) +(a) $\epsilon (\mathcal{M},1) = 0.490713$ +(d) $\epsilon (\mathcal{M},4) = 0.168091$ +Figure 2: Covering of the region (17) for $n = 1,2,3,4,5$ , and 6. + +![](images/744ba1718c164bd134043b42667058458543a94f3cce83d3cec25d23e6283e32.jpg) +(b) $\epsilon (\mathcal{M},2) = 0.288713$ +(e) $\epsilon (\mathcal{M},5) = 0.152269$ + +![](images/385f87743d75ad2abe88178a2f03c545b6662824eb428969c135481940af3e61.jpg) +(c) $\epsilon (\mathcal{M},3) = 0.207196$ +(f) $\epsilon (\mathcal{M},6) = 0.143661$ + +Similarly, for each $n$ , Algorithm 1 executed with 10 restarts converged to final radius values with relative standard deviation $< 1\%$ . Additional experimental details are provided in §D. + +# 5 Concluding remarks + +In this work we established a computationally viable approach to the optimal covering problem. We started with the definition of the problem (1) and applied a sequence of reformulations under mild structural assumptions to derive a version (7) more amenable to analysis. A central contribution of our work is a novel result establishing Lipschitz continuity of the optimal value function $\mathcal{G}$ arising from the nonconvex inner problem, which forms the foundation for applying a numerical gradient-based (zeroth-order) algorithm for the outer minimization. This algorithm leverages recent advances in nonsmooth and nonconvex optimization and, via randomized smoothing, offers non-asymptotic convergence guarantees for obtaining approximate stationary points. We demonstrated the effectiveness of our method through experiments on compact subsets of $\mathbb{R}^2$ and $\mathbb{R}^3$ . Future directions include accelerating the computation of $\mathcal{G}$ for convex sets via DC programming techniques and developing strategies to improve global convergence guarantees for the outer minimization. We also observe that $\mathcal{G}(x_1,\ldots ,x_n)$ is a permutation invariant function. This property can be used to restrict the search space further, which may help accelerate convergence. + +# References + +A. R. Alimov and I. G. Tsar'kov. Chebyshev centres, Jung constants, and their applications. Russian Mathematical Surveys, 74(5):775-849, 2019. doi: https://doi.org/10.4213/rm9839.2 +A. R. Alimov and I. G. Tsar'kov. *Geometric Approximation Theory*. Springer Monographs in Mathematics. Springer, Cham, 2021. doi: https://doi.org/10.1007/978-3-030-90951-2.1, 2 +L. H. An, H. V. Ngai, and P. D. Tao. DC programming and DCA for general DC programs. Advances in Intelligent Systems and Computing, 282:15-35, 01 2014. doi: https://doi.org/10.1007/978-3-319-06569-4_2_5 +P. Binev, A. Bonito, R. DeVore, and G. Petrova. Optimal learning. Calcolo. A Quarterly on Numerical Analysis and Theory of Computation, 61(1), 2024. doi: https://doi.org/10.1007/s10092-023-00564-y. 2 +A.-P. Calderón and A. Zygmund. Local properties of solutions of elliptic partial differential equations. Polska Akademia Nauk. Instytut Matematyczny. Studia Mathematica, 20:171-225, 1961. doi: https://doi.org/10.4064/sm-20-2-181-225.6 +F. H. Clarke. Optimization and Nonsmooth Analysis, volume 5 of Classics in Applied Mathematics. Society for Industrial and Applied Mathematics, Philadelphia, 1990. doi: https://doi.org/10.1137/1.9781611971309.7, 18 +S. Cobzas, R. Miculescu, and A. Nicolae. Lipschitz Functions, volume 2241 of Lecture Notes in Mathematics. Springer, Cham, 2019. doi: https://doi.org/10.1007/978-3-030-16489-8. 6, 18 +S. Das, A. Aravind, A. Cherukuri, and D. Chatterjee. Near-optimal solutions of convex semi-infinite programs via targeted sampling. Annals of Operations Research, 318(1):129-146, 2022. doi: https://doi.org/10.1007/s10479-022-04810-4.3 +A. L. Dontchev and R. T. Rockafellar. Implicit Functions and Solution Mappings. Springer Series in Operations Research and Financial Engineering. Springer-Verlag, 2 edition, 2014. doi: https://doi.org/10.1007/978-1-4939-1037-3.3 +Y. C. Eldar, A. Beck, and M. Teboulle. A minimax Chebyshev estimator for bounded error estimation. IEEE Transactions on Signal Processing, 56(4):1388-1397, 2008. doi: https://doi.org/10.1109/TSP.2007.908945.2 +A. V. Fiasco. Introduction to Sensitivity and Stability Analysis in Nonlinear Programming, volume 165 of Mathematics in Science and Engineering. Academic Press, Inc., Orlando, FL, 1983. 3 +A. V. Fiasco and Y. Ishizuka. Sensitivity and stability analysis for nonlinear programming. Annals of Operations Research, 27(1):215-235, 1990. doi: https://doi.org/10.1007/BF02055196.3, 6, 19 + +S. Foucart and C. Liao. S-procedure relaxation: a case of exactness involving Chebyshev centers. In Explorations in the mathematics of data science—the inaugural volume of the Center for Approximation and Mathematical Data Analytics, Applied Numerical and Harmonic Analysis, pages 1-18. Birkhäuser/Springer, Cham, 2024. doi: https://doi.org/10.1007/978-3-031-66497-7_1.2 +A. A. Goldstein. Optimization of lipschitz continuous functions. Mathematical Programming, 13: 14-22, 1977. doi: 10.1007/BF01584320. doi: https://doi.org/10.1007/BF01584320.7, 18 +S. Har-Peled. Geometric Approximation Algorithms, volume 173 of Mathematical Surveys and Monographs. American Mathematical Society, Providence, RI, 2011. doi: https://doi.org/10.1090/surv/173.3 +A. N. Kolmogorov and V. M. Tihomirov. $\varepsilon$ -entropy and $\varepsilon$ -capacity of sets in functional space. American Mathematical Society Translations (2), 17:277-364, 1961. 1 +T. Lin, Z. Zheng, and M. I. Jordan. Gradient-free methods for deterministic and stochastic nonsmooth nonconvex optimization. In Advances in Neural Information Processing Systems, volume 35, 2022. 7 +Z. Liu, C. Chen, L. Luo, and B. K. H. Low. Zeroth-order methods for constrained nonconvex nonsmooth stochastic optimization. In Proceedings of the 41st International Conference on Machine Learning, volume 235, pages 30842-30872, 2024. 7, 8 +O. S. Malysheva. Optimal position of compact sets and the Steiner problem in spaces with Euclidean Gromov-Hausdorff metric. *Sbornik: Mathematics*, 211(10):1382-1398, 2020. doi: https://doi.org/10.1070/SM9361.3 +C. A. Micchelli and T. J. Rivlin. A survey of optimal recovery. In Optimal Estimation in Approximation Theory, pages 1-54. Plenum, New York, 1977. 2 +B. S. Mordukhovich. Variational Analysis and Applications. Springer Monographs in Mathematics. Springer, Cham, 2018. doi: https://doi.org/10.1007/978-3-319-92775-6.5 +P. Paruchuri and D. Chatterjee. Attaining the Chebyshev bound for optimal learning: A numerical algorithm. Systems & Control Letters, 181:105648, 2023. ISSN 0167-6911. doi: https://doi.org/10.1016/j.sysconle.2023.105648.2 +C. P. Robert and G. Casella. Monte Carlo Statistical Methods. Springer Texts in Statistics. Springer-Verlag, New York, second edition, 2004. doi: https://doi.org/10.1007/978-1-4757-4145-2.3 +N. V. Sahinidis. BARON: A general purpose global optimization software package. Journal of Global Optimization, 8(2):201-205, 1996. ISSN 1573-2916. doi: https://doi.org/10.1007/BF00138693.8 +V. N. Ushakov and P. D. Lebedev. Algorithms for constructing an optimal cover for sets in three-dimensional Euclidean space. Trudy Instituta Matematiki i Mekhaniki, 21(2):276-288, 2015. doi: https://doi.org/10.1134/s0081543816050205.2, 3 +V. A. Voronov, A. D. Tolmachev, D. S. Protasov, and A. M. Neopryatnaya. Searching for distance graph embeddings and optimal partitions of compact sets in Euclidean space. In M. Khachay, Y. Kochetov, A. Eremeev, O. Khamisov, V. Mazalov, and P. Pardalos, editors, Mathematical Optimization Theory and Operations Research: Recent Trends, pages 391-403. Springer Nature Switzerland, 2023. 3 +D. Wu, J. Zhou, and A. Hu. A new approximate algorithm for the Chebyshev center. Automatica, 49 (8):2483-2488, 2013. doi: https://doi.org/10.1016/j.automatica.2013.04.029.2 +H. Yu. Cube packings in Euclidean spaces. Mathematika, 67(2):288-295, 2021. doi: https://doi.org/10.1112/mtk.12074.3 + +# NeurIPS Paper Checklist + +# 1. Claims + +Question: Do the main claims made in the abstract and introduction accurately reflect the paper's contributions and scope? + +Answer: [Yes] + +Guidelines: + +- The answer NA means that the abstract and introduction do not include the claims made in the paper. +- The abstract and/or introduction should clearly state the claims made, including the contributions made in the paper and important assumptions and limitations. A No or NA answer to this question will not be perceived well by the reviewers. +- The claims made should match theoretical and experimental results, and reflect how much the results can be expected to generalize to other settings. +- It is fine to include aspirational goals as motivation as long as it is clear that these goals are not attained by the paper. + +# 2. Limitations + +Question: Does the paper discuss the limitations of the work performed by the authors? + +Answer: [Yes] + +Guidelines: + +- The answer NA means that the paper has no limitation while the answer No means that the paper has limitations, but those are not discussed in the paper. +- The authors are encouraged to create a separate "Limitations" section in their paper. +- The paper should point out any strong assumptions and how robust the results are to violations of these assumptions (e.g., independence assumptions, noiseless settings, model well-specification, asymptotic approximations only holding locally). The authors should reflect on how these assumptions might be violated in practice and what the implications would be. +- The authors should reflect on the scope of the claims made, e.g., if the approach was only tested on a few datasets or with a few runs. In general, empirical results often depend on implicit assumptions, which should be articulated. +- The authors should reflect on the factors that influence the performance of the approach. For example, a facial recognition algorithm may perform poorly when image resolution is low or images are taken in low lighting. Or a speech-to-text system might not be used reliably to provide closed captions for online lectures because it fails to handle technical jargon. +- The authors should discuss the computational efficiency of the proposed algorithms and how they scale with dataset size. +- If applicable, the authors should discuss possible limitations of their approach to address problems of privacy and fairness. +- While the authors might fear that complete honesty about limitations might be used by reviewers as grounds for rejection, a worse outcome might be that reviewers discover limitations that aren't acknowledged in the paper. The authors should use their best judgment and recognize that individual actions in favor of transparency play an important role in developing norms that preserve the integrity of the community. Reviewers will be specifically instructed to not penalize honesty concerning limitations. + +# 3. Theory assumptions and proofs + +Question: For each theoretical result, does the paper provide the full set of assumptions and a complete (and correct) proof? + +Answer: [Yes] + +Guidelines: + +- The answer NA means that the paper does not include theoretical results. + +- All the theorems, formulas, and proofs in the paper should be numbered and cross-referenced. +- All assumptions should be clearly stated or referenced in the statement of any theorems. +- The proofs can either appear in the main paper or the supplemental material, but if they appear in the supplemental material, the authors are encouraged to provide a short proof sketch to provide intuition. +- Inversely, any informal proof provided in the core of the paper should be complemented by formal proofs provided in appendix or supplemental material. +- Theorems and Lemmas that the proof relies upon should be properly referenced. + +# 4. Experimental result reproducibility + +Question: Does the paper fully disclose all the information needed to reproduce the main experimental results of the paper to the extent that it affects the main claims and/or conclusions of the paper (regardless of whether the code and data are provided or not)? + +Answer: [Yes] + +Justification: The user requires a BARON license to run the BARON solver. + +# Guidelines: + +- The answer NA means that the paper does not include experiments. +- If the paper includes experiments, a No answer to this question will not be perceived well by the reviewers: Making the paper reproducible is important, regardless of whether the code and data are provided or not. +- If the contribution is a dataset and/or model, the authors should describe the steps taken to make their results reproducible or verifiable. +- Depending on the contribution, reproducibility can be accomplished in various ways. For example, if the contribution is a novel architecture, describing the architecture fully might suffice, or if the contribution is a specific model and empirical evaluation, it may be necessary to either make it possible for others to replicate the model with the same dataset, or provide access to the model. In general, releasing code and data is often one good way to accomplish this, but reproducibility can also be provided via detailed instructions for how to replicate the results, access to a hosted model (e.g., in the case of a large language model), releasing of a model checkpoint, or other means that are appropriate to the research performed. +- While NeurIPS does not require releasing code, the conference does require all submissions to provide some reasonable avenue for reproducibility, which may depend on the nature of the contribution. For example + +(a) If the contribution is primarily a new algorithm, the paper should make it clear how to reproduce that algorithm. +(b) If the contribution is primarily a new model architecture, the paper should describe the architecture clearly and fully. +(c) If the contribution is a new model (e.g., a large language model), then there should either be a way to access this model for reproducing the results or a way to reproduce the model (e.g., with an open-source dataset or instructions for how to construct the dataset). +(d) We recognize that reproducibility may be tricky in some cases, in which case authors are welcome to describe the particular way they provide for reproducibility. In the case of closed-source models, it may be that access to the model is limited in some way (e.g., to registered users), but it should be possible for other researchers to have some path to reproducing or verifying the results. + +# 5. Open access to data and code + +Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material? + +Answer: [Yes] + +Guidelines: + +- The answer NA means that paper does not include experiments requiring code. +- Please see the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details. +- While we encourage the release of code and data, we understand that this might not be possible, so "No" is an acceptable answer. Papers cannot be rejected simply for not including code, unless this is central to the contribution (e.g., for a new open-source benchmark). +- The instructions should contain the exact command and environment needed to run to reproduce the results. See the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details. +- The authors should provide instructions on data access and preparation, including how to access the raw data, preprocessed data, intermediate data, and generated data, etc. +- The authors should provide scripts to reproduce all experimental results for the new proposed method and baselines. If only a subset of experiments are reproducible, they should state which ones are omitted from the script and why. +- At submission time, to preserve anonymity, the authors should release anonymized versions (if applicable). +- Providing as much information as possible in supplemental material (appended to the paper) is recommended, but including URLs to data and code is permitted. + +# 6. Experimental setting/details + +Question: Does the paper specify all the training and test details (e.g., data splits, hyperparameters, how they were chosen, type of optimizer, etc.) necessary to understand the results? + +Answer: [Yes] + +# Guidelines: + +- The answer NA means that the paper does not include experiments. +- The experimental setting should be presented in the core of the paper to a level of detail that is necessary to appreciate the results and make sense of them. +- The full details can be provided either with the code, in appendix, or as supplemental material. + +# 7. Experiment statistical significance + +Question: Does the paper report error bars suitably and correctly defined or other appropriate information about the statistical significance of the experiments? + +Answer: [Yes] + +Justification: Although the global optimum is unknown, we indeed report the consistent convergence of multiple restarts to radius values with relative standard deviation $< 1\%$ in Examples 4.1 and 4.3 of §4 which indicates our method's robustness. + +# Guidelines: + +- The answer NA means that the paper does not include experiments. +- The authors should answer "Yes" if the results are accompanied by error bars, confidence intervals, or statistical significance tests, at least for the experiments that support the main claims of the paper. +- The factors of variability that the error bars are capturing should be clearly stated (for example, train/test split, initialization, random drawing of some parameter, or overall run with given experimental conditions). +- The method for calculating the error bars should be explained (closed form formula, call to a library function, bootstrap, etc.) +- The assumptions made should be given (e.g., Normally distributed errors). +- It should be clear whether the error bar is the standard deviation or the standard error of the mean. + +- It is OK to report 1-sigma error bars, but one should state it. The authors should preferably report a 2-sigma error bar than state that they have a $96\%$ CI, if the hypothesis of Normality of errors is not verified. +- For asymmetric distributions, the authors should be careful not to show in tables or figures symmetric error bars that would yield results that are out of range (e.g. negative error rates). +- If error bars are reported in tables or plots, The authors should explain in the text how they were calculated and reference the corresponding figures or tables in the text. + +# 8. Experiments compute resources + +Question: For each experiment, does the paper provide sufficient information on the computer resources (type of compute workers, memory, time of execution) needed to reproduce the experiments? + +Answer: [Yes] + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The paper should indicate the type of compute workers CPU or GPU, internal cluster, or cloud provider, including relevant memory and storage. +- The paper should provide the amount of compute required for each of the individual experimental runs as well as estimate the total compute. +- The paper should disclose whether the full research project required more computer than the experiments reported in the paper (e.g., preliminary or failed experiments that didn't make it into the paper). + +# 9. Code of ethics + +Question: Does the research conducted in the paper conform, in every respect, with the NeurIPS Code of Ethics https://neurips.cc/public/EthicsGuidelines? + +Answer: [Yes] + +Guidelines: + +- The answer NA means that the authors have not reviewed the NeurIPS Code of Ethics +- If the authors answer No, they should explain the special circumstances that require a deviation from the Code of Ethics. +- The authors should make sure to preserve anonymity (e.g., if there is a special consideration due to laws or regulations in their jurisdiction). + +# 10. Broader impacts + +Question: Does the paper discuss both potential positive societal impacts and negative societal impacts of the work performed? + +Answer: [NA] + +Justification: This article presents work aimed at presenting a novel approach to tackle the optimal covering problem. The applications of this work could have social impacts in various domains. However, we do not wish to highlight any such potential societal impacts in the article. + +Guidelines: + +- The answer NA means that there is no societal impact of the work performed. +- If the authors answer NA or No, they should explain why their work has no societal impact or why the paper does not address societal impact. +- Examples of negative societal impacts include potential malicious or unintended uses (e.g., disinformation, generating fake profiles, surveillance), fairness considerations (e.g., deployment of technologies that could make decisions that unfairly impact specific groups), privacy considerations, and security considerations. +- The conference expects that many papers will be foundational research and not tied to particular applications, let alone deployments. However, if there is a direct path to any negative applications, the authors should point it out. For example, it is legitimate + +to point out that an improvement in the quality of generative models could be used to generate deepfakes for disinformation. On the other hand, it is not needed to point out that a generic algorithm for optimizing neural networks could enable people to train models that generate Deepfakes faster. + +- The authors should consider possible harms that could arise when the technology is being used as intended and functioning correctly, harms that could arise when the technology is being used as intended but gives incorrect results, and harms following from (intentional or unintentional) misuse of the technology. +- If there are negative societal impacts, the authors could also discuss possible mitigation strategies (e.g., gated release of models, providing defenses in addition to attacks, mechanisms for monitoring misuse, mechanisms to monitor how a system learns from feedback over time, improving the efficiency and accessibility of ML). + +# 11. Safeguards + +Question: Does the paper describe safeguards that have been put in place for responsible release of data or models that have a high risk for misuse (e.g., pretrained language models, image generators, or scraped datasets)? + +Answer: [NA] + +Justification: We have curated the experiments on our own and have not used any dataset scraped from the Internet. + +# Guidelines: + +- The answer NA means that the paper poses no such risks. +- Released models that have a high risk for misuse or dual-use should be released with necessary safeguards to allow for controlled use of the model, for example by requiring that users adhere to usage guidelines or restrictions to access the model or implementing safety filters. +- Datasets that have been scraped from the Internet could pose safety risks. The authors should describe how they avoided releasing unsafe images. +- We recognize that providing effective safeguards is challenging, and many papers do not require this, but we encourage authors to take this into account and make a best faith effort. + +# 12. Licenses for existing assets + +Question: Are the creators or original owners of assets (e.g., code, data, models), used in the paper, properly credited and are the license and terms of use explicitly mentioned and properly respected? + +Answer: [Yes] + +# Guidelines: + +- The answer NA means that the paper does not use existing assets. +- The authors should cite the original paper that produced the code package or dataset. +- The authors should state which version of the asset is used and, if possible, include a URL. +- The name of the license (e.g., CC-BY 4.0) should be included for each asset. +- For scraped data from a particular source (e.g., website), the copyright and terms of service of that source should be provided. +- If assets are released, the license, copyright information, and terms of use in the package should be provided. For popular datasets, paperswithcode.com/datasets has curated licenses for some datasets. Their licensing guide can help determine the license of a dataset. +- For existing datasets that are re-packaged, both the original license and the license of the derived asset (if it has changed) should be provided. +- If this information is not available online, the authors are encouraged to reach out to the asset's creators. + +# 13. New assets + +Question: Are new assets introduced in the paper well documented and is the documentation provided alongside the assets? + +Answer: [Yes] + +Guidelines: + +- The answer NA means that the paper does not release new assets. +- Researchers should communicate the details of the dataset/code/model as part of their submissions via structured templates. This includes details about training, license, limitations, etc. +- The paper should discuss whether and how consent was obtained from people whose asset is used. +- At submission time, remember to anonymize your assets (if applicable). You can either create an anonymized URL or include an anonymized zip file. + +# 14. Crowdsourcing and research with human subjects + +Question: For crowdsourcing experiments and research with human subjects, does the paper include the full text of instructions given to participants and screenshots, if applicable, as well as details about compensation (if any)? + +Answer: [NA] + +Guidelines: + +- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. +- Including this information in the supplemental material is fine, but if the main contribution of the paper involves human subjects, then as much detail as possible should be included in the main paper. +- According to the NeurIPS Code of Ethics, workers involved in data collection, curation, or other labor should be paid at least the minimum wage in the country of the data collector. + +# 15. Institutional review board (IRB) approvals or equivalent for research with human subjects + +Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals (or an equivalent approval/review based on the requirements of your country or institution) were obtained? + +Answer: [NA] + +Guidelines: + +- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. +- Depending on the country in which research is conducted, IRB approval (or equivalent) may be required for any human subjects research. If you obtained IRB approval, you should clearly state this in the paper. +- We recognize that the procedures for this may vary significantly between institutions and locations, and we expect authors to adhere to the NeurIPS Code of Ethics and the guidelines for their institution. +- For initial submissions, do not include any information that would break anonymity (if applicable), such as the institution conducting the review. + +# 16. Declaration of LLM usage + +Question: Does the paper describe the usage of LLMs if it is an important, original, or non-standard component of the core methods in this research? Note that if the LLM is used only for writing, editing, or formatting purposes and does not impact the core methodology, scientific rigorousness, or originality of the research, declaration is not required. + +Answer: [NA] + +Guidelines: + +- The answer NA means that the core method development in this research does not involve LLMs as any important, original, or non-standard components. +- Please refer to our LLM policy (https://neurips.cc/Conferences/2025/LLM) for what should or should not be described. + +# A Foundational Definitions and Results + +Here we outline a few important definitions from sensitivity analysis and nonsmooth optimization. + +Definition A.1. A set-valued mapping $S: A \Rightarrow B$ is a mapping $S: A \to 2^B$ , where for each $x \in A$ , $S(x)$ is a subset of $B$ . They have been extensively used to study the feasible regions and solution maps of optimization problems + +Definition A.2. The set valued map $S: A \Rightarrow B$ is said to be uniformly compact near $\bar{x}$ if the set $\bigcup_{x \in N(\bar{x})} S(x)$ is bounded for some neighborhood $N(\bar{x})$ of $\bar{x}$ . + +Definition A.3. A mapping $f: \mathbb{R}^{\nu} \to \mathbb{R}$ is said to be locally Lipschitz if at each point $x \in \mathbb{R}^{\nu}$ , there exists a neighborhood $\mathcal{O}$ and $L \geqslant 0$ such that $|f(x_1) - f(x_2)| \leqslant L \| x_1 - x_2 \|$ for all $x_1, x_2 \in \mathcal{O}$ . + +Definition A.4. A mapping $f: U \subset \mathbb{R}^{\nu} \to \mathbb{R}$ is $L$ -Lipschitz if there exists $L \geqslant 0$ such that for every $x_1, x_2 \in U$ , we have $|f(x_1) - f(x_2)| \leqslant L \| x_1 - x_2 \|$ . The number $L$ is called the Lipschitz constant of $f$ . + +Corollary A.5 ((Cobzas et al., 2019)). Let $f: \mathbb{R}^{\nu} \to \mathbb{R}$ be a locally Lipschitz function and $U \subset \mathbb{R}^{\gamma}$ be compact. Then the restriction of $f$ to $U$ denoted by $f|_{U}$ is $L$ -Lipschitz. + +We define Dini-derivatives to generalize the notion of directional differentiability for non-smooth functions. + +Definition A.6. Let $f: \mathbb{R}^{\nu} \to \mathbb{R}$ and $x \in \mathbb{R}^{\nu}$ . The upper and lower Dini-derivatives of $f$ along the direction $v \in \mathbb{R}^{\nu}$ are defined as + +$$ +\begin{array}{l} D ^ {+} f (x; h) := \lim _ {h \downarrow 0} \sup _ {h} \frac {1}{h} (f (x + h v) - f (x)) \text {a n d} \\ D ^ {-} f (x; h) := \lim _ {h \downarrow 0} \inf _ {h} \frac {1}{h} (f (x + h v) - f (x)) \tag {18} \\ \end{array} +$$ + +respectively. + +Definition A.7. Given a point $x \in \mathbb{R}^{\nu}$ and a direction $v \in \mathbb{R}^{\nu}$ , the generalized directional derivative of a nondifferentiable function $f$ is defined as $Df(x;v) \coloneqq \lim \sup_{y \to x, t \downarrow 0} \frac{f(y + tv) - f(y)}{t}$ . Then, the generalized gradient (Clarke, 1990) of $f$ is defined as the set + +$$ +\partial f (x) := \left\{g \in \mathbb {R} ^ {\nu} \mid g ^ {\top} v \leqslant D \mathcal {G} (x; v), \text {f o r a l l} v \in \mathbb {R} ^ {d} \right\}. \tag {19} +$$ + +Definition A.8. Let $\mathbb{B}(x,\delta) \coloneqq \{y \in \mathbb{R}^{\nu} \mid \| y - x\| \leqslant \delta\}$ . Given a $\delta \geqslant 0$ , the Goldstein $\delta$ -subdifferential (Goldstein, 1977) of $f$ at $x$ is defined as + +$$ +\partial_ {\delta} f (x) := \operatorname {c o n v} \left(\bigcup_ {y \in B (x, \delta)} \partial f (y)\right). \tag {20} +$$ + +# B Proofs of Propositions 2.1 and 2.2 + +Proposition 2.1. We know that for every $i \in \{1, \dots, n\}$ there exists $y \in \mathcal{M}$ such that $\| y - x_i^* \| \leqslant \epsilon(\mathcal{M}, n)$ . Consequently, $\| z^* - x_i^* \| \leqslant \| z^* - y \| + \| y - x_i^* \|$ , which leads to + +$$ +\left\| z ^ {*} - x _ {i} ^ {*} \right\| \leqslant \epsilon (\mathcal {M}, n) + \epsilon (\mathcal {M}, 1), +$$ + +and since $\epsilon (\mathcal{M},n^{\prime})\leqslant \epsilon (\mathcal{M},n)$ for every $n^\prime >n$ + +$$ +\left\| z ^ {*} - x _ {i} ^ {*} \right\| \leqslant 2 \epsilon (\mathcal {M}, 1), +$$ + +and the assertion follows. + +![](images/83afcac786c48a91a228c02c88405bd1014ab008cdb93efaf752f1e9df32c27b.jpg) + +Proposition 2.2. Strict feasibility of both problems is always ensured since any point in $\mathcal{M}$ with $t = 0$ readily satisfies all the constraints. It follows immediately that $(t^{\dagger},y^{\dagger})$ is a feasible point of (6), which implies $t^\dagger \leqslant t^*$ . Moreover, since + +$$ +t ^ {*} = \max \left\{t \in \mathbb {R} \mid t \leqslant \| y ^ {*} - x _ {i} \| \text {f o r a l l} i = 1, \dots , n \right\}, +$$ + +it follows that $t^* \leqslant t^\dagger$ . Consequently, $t^\dagger = t^*$ . + +![](images/0f5776b4af25c333a4c07333a7704f10030989eaef96f105529aeb2711c7bb8a.jpg) + +# C Lipschitz continuity of optimal value functions of parametrized nonlinear programs + +Consider a parametrized nonlinear program of the form + +$$ +\min _ {z} \quad g _ {0} (x, z) \tag {21} +$$ + +$$ +\text {s u b j e c t} \quad g _ {i} (x, z) \leqslant 0 \text {f o r} i = 1, \dots , p, +$$ + +where the functions $g_0, g_1, \ldots, g_p$ are continuously differentiable from $\mathbb{R}^{\gamma} \times \mathbb{R}^{\kappa}$ to $\mathbb{R}$ . The Lagrangian for (21) is given by + +$$ +\mathcal {L} (x, z, \lambda) := g _ {0} (x, z) + \sum_ {i = 1} ^ {p} \lambda_ {i} g _ {i} (x, z). \tag {22} +$$ + +The feasible region is denoted by the set-valued map $S_{\mathrm{feas}}:\mathbb{R}^{\gamma}\Rightarrow \mathbb{R}^{\kappa}$ + +$$ +S _ {\text {f e a s}} (\bar {x}) := \left\{z \in \mathbb {R} ^ {\kappa} \mid g _ {i} (\bar {x}, z) \leqslant 0 \text {f o r} i = 1, \dots , p \right\}. \tag {23} +$$ + +The optimal value function $\Phi : \mathbb{R}^{\gamma} \to \mathbb{R} \cup \{+\infty\}$ is the optimal value of the NLP for a given parameter $\bar{x} \in \mathbb{R}^{\gamma}$ . That is, + +$$ +\Phi (\bar {x}) := \min _ {z \in S _ {\text {f e a s}} (\bar {x})} g _ {0} (\bar {x}, z). \tag {24} +$$ + +By convention, $\Phi (\bar{x}) = +\infty$ if the problem is infeasible. The set of global optimizers is the set-valued map $S_{\mathrm{opt}}:\mathbb{R}^{\kappa}\Rightarrow \mathbb{R}^{\gamma}$ + +$$ +S _ {\text {o p t}} (\bar {x}) := \left\{z \in S _ {\text {f e a s}} (\bar {x}) \mid g _ {0} (\bar {x}, z) = \Phi (\bar {x}) \right\}. \tag {25} +$$ + +For a choice $\bar{x}$ of $x$ , let $\bar{z} \in S_{\mathrm{opt}}(\bar{x})$ . We define the set of active constraint indices at $(\bar{x}, \bar{z})$ as + +$$ +I (\bar {x}, \bar {z}) := \left\{i \in \{1, \dots , p \} \mid g _ {i} (\bar {x}, \bar {z}) = 0 \right\}. \tag {26} +$$ + +The set of Karush-Kuhn-Tucker (KKT) vectors of (8) for a pair $(\bar{x},\bar{z})\in \mathbb{R}^{\gamma}\times \mathbb{R}^{\kappa}$ is denoted by + +$$ +\Lambda (\bar {x}, \bar {z}) := \left\{\lambda \in [ 0, + \infty [ ^ {p} \mid \nabla_ {z} \mathcal {L} (\bar {x}, \bar {z}, \lambda) = 0, \lambda_ {i} g _ {i} (\bar {x}, \bar {z}) = 0 \text {f o r} i = 1, \dots , p \right\} \tag {27} +$$ + +The Mangasarian-Fromovitz constraint qualification (MFCQ) is said to hold at a point $\bar{z} \in S_{\mathrm{opt}}(\bar{x})$ if there exists a direction $w \in \mathbb{R}^{\kappa}$ such that the inner product of $w$ with the gradient of each active constraint is strictly negative. To wit, the following condition must hold: + +$$ +\nabla_ {z} g _ {i} (\bar {x}, \bar {z}) ^ {\top} w < 0 \text {f o r a l l} i \in I (\bar {x}, \bar {z}) \tag {28} +$$ + +We now state the result (Fiacco and Ishizuka, 1990, Theorem 4.2) that establishes the conditions under which the optimal value function is Lipschitz continuous. + +Theorem C.1. Consider the problem (21) and grant the notations established above. Fix $\bar{x} \in \mathbb{R}^{\gamma}$ and suppose that + +(C.1-a) the MFCQ condition holds for every $\bar{z} \in S_{\mathrm{opt}}(\bar{x})$ , and + +(C.1-b) the map $S_{\mathrm{feas}}$ is uniformly compact near $\bar{x}$ . + +Then the function $\Phi$ is locally Lipschitz continuous at $\bar{x}$ . Moreover, for any $v\in \mathbb{R}^{\gamma}$ , we have + +$$ +\begin{array}{l} \inf _ {\bar {z} \in S _ {\text {o p t}} (\bar {x})} \min _ {\lambda \in \Lambda (\bar {x}, \bar {z})} \frac {\partial \mathcal {L}}{\partial x} (\bar {x}, \bar {z}, \lambda) \cdot v \leqslant - \lim _ {h \downarrow 0} \inf _ {h} \frac {1}{h} (\Phi (\bar {x} + h v) - \Phi (\bar {x})) \\ \leqslant - \lim _ {h \downarrow 0} \sup _ {h \downarrow 0} \frac {1}{h} (\Phi (\bar {x} + h v) - \Phi (\bar {x})) \leqslant \inf _ {\bar {z} \in S _ {\mathrm {o p t}} (\bar {x})} \max _ {\lambda \in \Lambda (\bar {x}, \bar {z})} \frac {\partial \mathcal {L}}{\partial x} (\bar {x}, \bar {z}, \lambda) \cdot v. \\ \end{array} +$$ + +# D Experimental details + +In this section, we provide further details of the experimental procedure along with a 3D illustration of covering in Example D.2. As emphasized in §1, the main computational bottleneck arises from having to reliably solve the inner problem (8) for the evaluation of $\mathcal{G}$ globally. This is the reason a solver with global optimality guarantees is employed. To illustrate the phenomenon mentioned in 4.2, we provide the following example. + +Example D.1. For the solid ellipse (16) in Example 4.1. Here we employed Algorithm 1 maintaining all parameters from Example 4.1, except that $\mathcal{G}$ was evaluated using IPOPT instead of BARON in Algorithm 1. Fig. 3 illustrates the failure of Algorithm 1 in producing an adequate cover of $\mathcal{M}$ with $n = 1,2$ , and 3, when IPOPT was employed. + +![](images/2aed4077419f8d167f113c68c28f6b84c8c97535342f14249369fa060e975c01.jpg) +Figure 3: Covering fails if IPOPT is employed in place of BARON for $n = 1,2,3$ + +![](images/54044eee54d7d2b3cde00043ef0e6ec046bfa041fd588092c689f81ef1f1a5d0.jpg) + +![](images/193aa6b716f0febdcba452bd6ecdab319c00b32a8c1cbee9f1dd9d330bbd9946.jpg) + +For our experiments, we used the demo mode of BARON, which imposes limits on the problem dimensions and number of constraints. To facilitate faster convergence and support for larger-scale problems, a more advanced license may be used. For more details, visit https://minlp.com/ baron- licenses. Table 2 reports the time taken to compute $\mathcal{G}$ in Example 4.1 for varying $n$ using BARON in demo mode on an AMD Ryzen 7 4800H 2.90 GHz CPU. Note that the total number of + +Table 2: Computation time for $\mathcal{G}$ for varying $n$ + +
nComputation time (ms)
2129
3135
4143
5146
6154
7160
8164
9173
+ +evaluations of $\mathcal{G}$ per iteration of Algorithm 1 depends on $b_{1}$ and $b_{2}$ . + +Fig 4 shows the temporal evolution of the objective in Example 4.1 and features 4 restarts (initializations). An example with a nonconvex, asymmetric set in $\mathbb{R}^3$ is provided in Example D.2. Note that only the isometric view is provided for each case. + +![](images/88e82ea2fde4ead69b9bc60f86a0b39250dd8893d55ce2f789a699369bf41617.jpg) +Figure 4: The evolution of $\mathcal{G}$ against iteration count for the problem in Example 4.1 for $n = 4$ . + +Example D.2. As an illustration of the 3D case, we present an example where $\mathcal{M}$ is an asymmetric and nonconvex subset of $\mathbb{R}^3$ . Consider the case where $k = 3$ , and the set $\mathcal{M}$ is the region + +$$ +\mathcal {M} := \left\{\left(y _ {1}, y _ {2}, y _ {3}\right) \in \mathbb {R} ^ {3} \mid \varphi_ {j} (y) \leqslant 0 \text {f o r} j = 1, 2, 3 \right\} \text {w h e r e , w i t h} y := \left(y _ {1}, y _ {2}, y _ {3}\right), +$$ + +$$ +\varphi_ {1} (y) := y _ {3} - \left(1 - y _ {1} ^ {2} - 0. 3 y _ {2} ^ {2}\right), \varphi_ {2} (y) := - \left(y _ {3} + 1\right), \text {a n d} \varphi_ {3} (y) := y _ {2} - \left(0. 5 \left(y _ {1} - 1\right) ^ {2} + 0. 2 y _ {3} ^ {2}\right). \tag {29} +$$ + +Algorithm 1 was employed with multiple restarts, and pictures of the outcomes are presented in Fig. 5 for the cases $n = 1,2,3$ , and 4. + +![](images/9ca907fd3b23ec9c85ebfc654f9bb82dcca2e6c595bb2867b568090ace1b4836.jpg) +$\mathcal{M}$ n=1 covering of. $\mathcal{M}$ + +![](images/b8db9ec3af9b94dd94266cb7787ef2ff511bf6c500fa1ac5d326956eb391f5b2.jpg) +$\mathcal{M}$ n=2 covering of. $\mathcal{M}$ + +(a) $\epsilon (\mathcal{M},1) = 2.344070$ +(c) $\epsilon (\mathcal{M},3) = 1.446102$ +Figure 5: Covering of the region (29) for $n = 1,2,3$ , and 4. +![](images/3696b0b9ec242feb65d435077ec8868b3f2dc37b93ade90dfdf03e38c37fa448.jpg) +$\mathcal{M}$ n=3 covering of. $\mathcal{M}$ + +(b) $\epsilon (\mathcal{M},2) = 1.692330$ +(d) $\epsilon (\mathcal{M},4) = 1.371257$ +![](images/0d3ed9a029e8b18051f0f910d3cd443785b72c3229bc4a4ac51f6ae97c792bb9.jpg) +n=4 covering of. \ No newline at end of file diff --git a/NeurIPS/2025/A Computationally Viable Numerical Gradient-based Technique for Optimal Covering Problems/images.zip b/NeurIPS/2025/A Computationally Viable Numerical Gradient-based Technique for Optimal Covering Problems/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..ceab38ac4592f93d68edcbd09bae7be933f97341 --- /dev/null +++ b/NeurIPS/2025/A Computationally Viable Numerical Gradient-based Technique for Optimal Covering Problems/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b5da5f064bb0a621420cd93a7b81bd7185899a17cd433d442747aea01b1354ee +size 695330 diff --git a/NeurIPS/2025/A Computationally Viable Numerical Gradient-based Technique for Optimal Covering Problems/layout.json b/NeurIPS/2025/A Computationally Viable Numerical Gradient-based Technique for Optimal Covering Problems/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..2035ef8e8a3d4348789c947eb736d65c75d96b9c --- /dev/null +++ b/NeurIPS/2025/A Computationally Viable Numerical Gradient-based Technique for Optimal Covering Problems/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3befad06bf1340d46054be11480fc1a4583e10eff013ee863e29e51b0920a1ba +size 991565 diff --git a/NeurIPS/2025/A Counterfactual Semantics for Hybrid Dynamical Systems/46e56ccd-c1be-4f2f-b41a-95229eda93f4_content_list.json b/NeurIPS/2025/A Counterfactual Semantics for Hybrid Dynamical Systems/46e56ccd-c1be-4f2f-b41a-95229eda93f4_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..492a3770a4fc298f33bb5c63e42d0f1da66cae07 --- /dev/null +++ b/NeurIPS/2025/A Counterfactual Semantics for Hybrid Dynamical Systems/46e56ccd-c1be-4f2f-b41a-95229eda93f4_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:00b2b732c713e99f9e3eb61a508c6ef3c1992216fd79475b86ff37c7db591c6a +size 309864 diff --git a/NeurIPS/2025/A Counterfactual Semantics for Hybrid Dynamical Systems/46e56ccd-c1be-4f2f-b41a-95229eda93f4_model.json b/NeurIPS/2025/A Counterfactual Semantics for Hybrid Dynamical Systems/46e56ccd-c1be-4f2f-b41a-95229eda93f4_model.json new file mode 100644 index 0000000000000000000000000000000000000000..f5ef6fc3902dbd2e45809ff4ce84f39811d86122 --- /dev/null +++ b/NeurIPS/2025/A Counterfactual Semantics for Hybrid Dynamical Systems/46e56ccd-c1be-4f2f-b41a-95229eda93f4_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:043f1e641829dce766fdd9215e2f70ab38881c6833c778f56b0a96eb65e4b2f6 +size 365781 diff --git a/NeurIPS/2025/A Counterfactual Semantics for Hybrid Dynamical Systems/46e56ccd-c1be-4f2f-b41a-95229eda93f4_origin.pdf b/NeurIPS/2025/A Counterfactual Semantics for Hybrid Dynamical Systems/46e56ccd-c1be-4f2f-b41a-95229eda93f4_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..56c241ee56e5923884012ef158d8c4851e5d29ad --- /dev/null +++ b/NeurIPS/2025/A Counterfactual Semantics for Hybrid Dynamical Systems/46e56ccd-c1be-4f2f-b41a-95229eda93f4_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:da14539030a8b51a34603b294e633bc506f9a48c86953090d040c3b7afcdc0d4 +size 2004039 diff --git a/NeurIPS/2025/A Counterfactual Semantics for Hybrid Dynamical Systems/full.md b/NeurIPS/2025/A Counterfactual Semantics for Hybrid Dynamical Systems/full.md new file mode 100644 index 0000000000000000000000000000000000000000..95e7af7dacdcd1d7ca23d0b2c97a2ddfc8a03d38 --- /dev/null +++ b/NeurIPS/2025/A Counterfactual Semantics for Hybrid Dynamical Systems/full.md @@ -0,0 +1,1366 @@ +# A Counterfactual Semantics for Hybrid Dynamical Systems + +Andy Zane1,2 andy@basis.ai + +Dmitry Batenkov1 +dima@basis.ai + +Rafal Urbaniak1 +rafal@basis.ai + +Jeremy Zucker3 jeremy.zucker@pnnl.gov + +Sam Witty\*sam@sorbus.ai + +$^{1}$ Basis Research Institute, New York, NY + $^{2}$ University of Massachusetts Amherst, Amherst, MA + $^{3}$ Pacific Northwest National Laboratory, Richland, WA + $^{4}$ Sorbus AI + +# Abstract + +Models of hybrid dynamical systems are widely used to answer questions about the causes and effects of dynamic events in time. Unfortunately, existing causal reasoning formalisms lack support for queries involving the dynamically triggered, discontinuous interventions that characterize hybrid dynamical systems. This mismatch can lead to ad-hoc and error-prone causal analysis workflows in practice. To bridge the gap between the needs of hybrid systems users and current causal inference capabilities, we develop a rigorous counterfactual semantics by formalizing interventions as transformations to the constraints of hybrid systems. Unlike interventions in a typical structural causal model, however, interventions in hybrid systems can easily render the model ill-posed. Thus, we identify mild conditions under which our interventions maintain solution existence, uniqueness, and measurability by making explicit connections to established hybrid systems theory. To illustrate the utility of our framework, we formalize a number of canonical causal estimands and explore a case study on the probabilities of causation with applications to fishery management. Our work simultaneously expands the modeling possibilities available to causal inference practitioners and begins to unlock decades of causality research for users of hybrid systems. + +# 1 Introduction + +Models of continuous-time dynamical systems are powerful tools for describing real-world mechanisms. From contrastive queries about system behavior under different control policies (Kirk, 2004), to sensitivity analyses designed to aid in understanding which parameters drive system variation (Cacuci, 2003), scientists, policy makers, and engineers often use such models to answer their "what-if" and causal questions. Unfortunately, causal reasoning with continuous-time systems can be ad-hoc, manual, and error-prone in daily practice. + +In parallel, researchers in causal inference have built rigorous tools for answering an expansive taxonomy of causal queries. For example, causal questions about effect estimation (Pearl, 2009; Rubin, 1974; Imbens & Rubin, 2015), counterfactual reasoning (Pearl, 2009, Ch. 7), mediation + +analysis (Pearl, 2001), responsibility, blame (Chockler & Halpern, 2004), attribution, and explanation (Halpern & Pearl, 2005a,b; Beckers, 2022) can all be succinctly expressed as estimands constructed from parallel worlds (Balke & Pearl, 1994; Avin et al., 2005; Shpitser & Pearl, 2008) or potential outcomes (Rubin, 1974). The toolkit also affords a formal means of determining when those estimands can be reduced to computationally tractable, probabilistic estimation problems (Pearl, 1995; Shpitser & Pearl, 2006; Hernán & Robins, 2023). These insights have made it possible to build general-purpose technology for causal reasoning, such as the causal probabilistic programming language ChiRho (Bingham et al., 2021; Witty, 2023; Basis-Research, 2025).2 + +Despite significant progress over the last decade (Mooij et al., 2013; Hansen & Sokol, 2014; Blom et al., 2019; Forre & Mooij, 2020; Peters et al., 2020; Blom et al., 2021; Bongers, 2022; Blom & Mooij, 2023; Boeken & Mooij, 2024; Peters & Halpern, 2025), however, gaps remain in the technical capacity of modern causal reasoning machinery to operate on the full breadth of interventions that can be encoded in continuous-time dynamical systems. In particular, a counterfactual semantics for dynamically triggered, instantaneous intervention has not yet been established. With such an intervention semantics in hand, causal reasoning can be more fully mechanized for causal questions about dynamic temporal events, dramatically expanding the rigor and variety of queries available to users of continuous-time dynamical systems. + +Such interventions underpin many closed-loop control problems: for example, HVAC systems activate when temperature thresholds are reached; lockdown and masking measures can be implemented according to levels of Sars-CoV-2 in wastewater (Kappus-Kron et al., 2024); commercial fishing pressure can be reduced once annual harvest limits are reached (Anon, 2007b; Warlick et al., 2018); central banks adjust interest rates depending on economic indicators like inflation and unemployment; reservoir managers release water depending on storage thresholds and agricultural needs (Ray, 2003); and power grids activate "peaker plants" (or stored energy) when demand exceeds certain thresholds (Zhuk et al., 2016). Despite limited attention from the causality community, these systems have garnered significant interest from control theorists in the form of continuous-time, hybrid dynamical systems (Schaft et al., 2000; Goebel et al., 2012; Sanfelice, 2021) that encode both continuous and instantaneous dynamics in a set of differential and difference constraining equations. + +To construct a counterfactual semantics for state-dependent, instantaneous intervention, we formalize a class of transformations on hybrid system constraints that induce the desired counterfactual behavior. An intervention creates a twin, parallel world with transformed constraints, but in a way that ensures both the twin and original worlds share randomly sampled values for initial conditions and parameters. This induces a familiar joint distribution over counterfactual outcomes (Rubin, 1974; Balke & Pearl, 1994; Shpitser & Pearl, 2006, 2008) that can, in turn, be used as input to established causal estimands, such as an expected treatment effect or the probabilities of necessary and sufficient causation. + +# Our contributions are: + +1. A formal, counterfactual semantics for dynamically triggered, instantaneous interventions in continuous-time dynamical systems. +2. Under minimal requirements on interventional specifications, proof that sufficient conditions for solution existence, uniqueness, and finite-time measurability are preserved in the intervened system. Our framework also explicitly connects to established well-posedness conditions on hybrid dynamical systems. +3. A case study on the probabilities of necessary and sufficient causation applied to fishery management, demonstrating extensibility to non-trivial causal estimands rarely applied to dynamical systems. + +# 2 Related Work + +In Causality. Many researchers have contributed to the systematization of causality for dynamical systems. Hansen & Sokol (2014), for example, show that dynamical systems can be unrolled into directed, structural causal models (SCMs). In the context of ordinary differential equations (ODEs), + +if $f$ is the right-hand side of the continuous-time differential equation $x' = f(x)$ , we can write structural equations $x_{t} = x_{t - \Delta t} + f(x_{t - \Delta t}, u) \Delta t$ , where $t \geqslant 0$ , $x_{t} \in \mathbb{R}$ is the value of the state variable $x$ at time $t$ , $u \in \mathbb{R}$ is a fixed realization of exogenous noise, and $x_{0}$ is fixed. Taking $\Delta t \to 0$ , we can recover the system's dynamics arbitrarily well. This limit results in SCMs with infinitely many variables — a modality that has been recently studied as "Generalized Structural Equation Models" (GSEMs) (Peters & Halpern, 2021; Halpern & Peters, 2022; Peters & Halpern, 2025). With $\Delta t > 0$ , this becomes the familiar discrete time approximation, which has been widely researched in causal inference (Spirtes, 2013; Pearl, 2009; Murphy, 2002; Wang et al., 2018; Assaad et al., 2022; Runge et al., 2023; Zan et al., 2024). + +This forward-Euler representation, however, is not the preferred tool of hybrid systems theorists, making it ill-suited for identifying conditions under which intervention preserves established well-posedness conditions. Additionally, under the forward-Euler representation, interventional transformations that induce state-dependent jumps require "soft" intervention (Correa & Baireinboim, 2020) on all endogenous nodes that might jump. Indeed, the state-dependent jump conditions must be "checked" at all points in time. We discuss this more precisely in appendix J. + +Somewhat sidestepping the temporal representation issue, most causal research on continuous-time dynamical systems has employed foundational ideas in cyclic graphical models (Iwasaki & Simon, 1994; Spirtes, 2013; Lacerda et al., 2008; Hyttinen et al., 2012) to develop causal abstractions of a system's equilibrium behavior (Dash, 2003; Mooij et al., 2013; Hansen & Sokol, 2014; Blom et al., 2019; Forre & Mooij, 2020; Bongers, 2022; Blom & Mooij, 2023). Equilibrium-focused frameworks, however, can fail to expose complex causal relationships in transient dynamics (Peters et al., 2020). + +Extensions such as the "time-splitting" operation (Boeken & Mooij, 2024), or the application of GSEMs to hybrid automata by Peters & Halpern (2025), enhance the expressiveness of graphical approaches by supporting static-time discontinuities. In contrast, our work targets dynamically triggered interventions, which cannot be straightforwardly analyzed using methods like time-splitting. Indeed, the order — and, therefore, the induced time-split graph structure — of dynamically triggered interventions depends on state evolution, and therefore on exogenous noise. Our approach avoids these issues by directly defining counterfactual interventions on hybrid system constraints. While the non-graphical framing means that standard graphical identifiability criteria are not immediately available, building our semantics on established hybrid systems theory opens pathways to leveraging longstanding methods and conditions for system identification of dynamical systems (Walter & Pronzato, 1997; Ljung, 2012; Raue et al., 2009; Stuart, 2010), such as the "persistence of excitation", which has been studied directly in the context of hybrid systems (Johnson, 2023; Saoud et al., 2024). + +Our approach follows the spirit of recent developments in constraint-based causal modeling. For example, Beckers et al. (2023) extend SCMs in order to handle logical constraints (such as unit conversions), while Blom et al. (2019) interpret equilibrium equations of dynamical systems, along with their corresponding algebraic invariants, as a collection of constraints. In both cases, a model is characterized by a collection of constraints, and interventions are defined as transformations of those constraints (e.g., by changing, disabling, or enabling them). At a high level, we take a similar approach. Hybrid systems, however, are characterized by a unique class of constraints requiring special considerations around Zeno behavior, set-valued theory, non-uniqueness even in "well-posed" cases, set-valued stable points, etc. In short, analyzing the post-intervention properties of hybrid systems is made easier via direct use of existing hybrid systems frameworks, rather than existing causal frameworks. Naturally, each school of thought is best suited to different tasks, and we look forward to future work that deftly exercises the comparative advantages of each. + +In Control Theory. Control theory and causality share overlapping goals, yet historically operate separately. This paper integrates causal reasoning directly into established, hybrid dynamical systems frameworks (Goebel et al., 2012; Sanfelice, 2021). In particular, our formalization of dynamically triggered intervention as constraint transformation mirrors controller-plant compositions from hybrid control theory, which are also shown to preserve established conditions for system well-posedness (Sanfelice, 2021). Hybrid system theory presents challenges, however, due to potential non-uniqueness of solutions under general conditions (Goebel et al., 2012), complicating counterfactual reasoning. To address this practically, we follow common simulation practices (e.g., preferring flowing solutions when multiple are possible) and explicitly formalize these assumptions (Sanfelice et al., 2023a). Our contributions thus link causal semantics to established hybrid systems theory and practice, enabling rigorous and computationally feasible causal analysis. + +![](images/b473e7ce4a725cebc7bda27344215be9125507d89d0d53a904d6ab938567161d.jpg) +(a) + +![](images/057eadd00c0ebcfd1bb7c78b7be09263ef814606c8a278e742c30f48480237ba.jpg) +(b) +Figure 1: Three parallel worlds constructed by starting with a dose-decay model (fig. 1a) and then transforming that model to reflect dosage at a fixed, static time (fig. 1b), and dosage when the concentration hits a threshold (fig. 1c and example 1). This paper develops the first explicitly counterfactual semantics for the dynamically triggered, state-dependent case (fig. 1c). Three sample trajectories are shown for each world, with initial condition and dose-decay rate held fixed across worlds for each sample trajectory. Notice that the state-dependent interventions occur at different times for different trajectories induced by different initial conditions and/or parameters. + +![](images/f88542a8c2a86374157b3e6ca6665b5fd8d69dcf7a7d29b51a115f80158acea1.jpg) +(c) + +# 3 Parameterized Hybrid Systems + +As a first approximation, the present work focuses on continuous, ordinary differential equations models with random initial conditions and parameters. Many intuitive interventions, however, can be conveniently defined as instantaneous (discontinuous) changes to the dynamically evolving state. Thus, we focus on hybrid systems that afford both continuous "flow" and event-based "jumps" in state. Jumps can arise as a product of interventions and/or discontinuous dynamics in the unintervened system. With state space $S \subseteq \mathbb{R}^n$ and following the framework laid out by Goebel et al. (2012), many hybrid systems can be characterized as comprising four elements: a flow set $C \subset S$ ; a differential inclusion $F: S \to \mathbb{R}^n$ ; a jump set $D \subset S$ ; and a set-valued jump map $G: S \to S$ . In general, the system evolves according to its differential inclusion $F$ when its state is in the flow set $C$ and according to the jump map $G$ when in the jump set $D$ . Readers who are unfamiliar with inclusions and set-valued maps should refer to appendix A.1. Hybrid systems often alternate between continuous flow and discontinuous jumps, though consecutive jumps remain well-defined. Many hybrid systems, then, can be characterized with the tuple $(C, F, D, G)$ . We ground this out in the following example. + +Example 1 (Dosage Model). Consider modeling the exponential decay of drug concentration $x$ at rate $\beta$ , where medical providers intervene to administer additional dosage when $x$ reaches a threshold $\gamma$ . To model these dynamics, we can seek state evolutions obeying + +$$ +\left\{ \begin{array}{l} x \in C = \mathbb {R} \backslash D \\ x \in D = \{x \in \mathbb {R}: x \leqslant \gamma \} \end{array} \right. \quad \begin{array}{l} \dot {x} \in F (x) = \{- \beta x \} \\ x ^ {+} \in G (x) = \{x + 1 \} \end{array} +$$ + +where $\dot{x}$ denotes the time derivative of the state, and $x^{+}$ the state immediately following a jump. The solution map of a hybrid system typically takes as "input" an initial condition $\pmb {\xi}\in S$ , but can also be parameterized to additionally incorporate a vector $\pmb{\theta}$ — in example 1, $\pmb {\theta} = [\beta ,\gamma ]$ + +Definition 1 (Parameterized Hybrid System). Let $S \subseteq \mathbb{R}^n$ , $\Theta \subseteq \mathbb{R}^m$ . A parametrized hybrid system $\mathcal{P}$ is a tuple $\mathcal{P} = (\mathcal{H}, S, \Theta)$ where for each $\theta \in \Theta$ , $\mathcal{H}(\theta) = (C(\theta), F_{\theta}, D(\theta), G_{\theta})$ is a standard hybrid system (Goebel et al., 2012, Def. 2.2), i.e. + +- $C: \Theta \to S$ is a set-valued mapping returning the flow set, +- $F_{\theta}(\pmb{x}) = F(\pmb{x}, \pmb{\theta}) \forall \pmb{x} \in S$ and $\forall \pmb{\theta} \in \Theta$ , where $F: S \times \Theta \to \mathbb{R}^n$ is a differential inclusion, with $C(\pmb{\theta}) \subset \operatorname{dom} F_{\pmb{\theta}} \subseteq S$ for all $\pmb{\theta} \in \Theta$ , +- $D: \Theta \to S$ is a set-valued mapping returning the jump set, and +- $G_{\theta}(\pmb{x}) = G(\pmb{x}, \pmb{\theta}) \forall \pmb{x} \in S$ and $\forall \pmb{\theta} \in \Theta$ , where $G: S \times \Theta \Rightarrow S$ is an ordered (i.e. returning an ordered collection of sets to keep track of interventions, cf. definition 6) set-valued jump map, with $D(\pmb{\theta}) \subset \operatorname{dom} G_{\pmb{\theta}} \subseteq S$ for all $\pmb{\theta} \in \Theta$ . + +Without explicit parametrization, we write $\mathcal{H} = (C,F,D,G)$ , and also often expand $\mathcal{H}$ in $\mathcal{P}$ , writing equivalently $\mathcal{P} = (\mathcal{H},\mathcal{S},\Theta) = (C,F,D,G,\mathcal{S},\Theta)$ . + +canonically, solutions to hybrid systems are functions of both continuous time $t \in \mathbb{R}_{\geqslant 0}$ and discrete event indices $j \in \mathbb{N}$ . Following (Goebel et al., 2012, Sects. 2.2-2.3), we define, for each possible parameterization $\pmb{\theta} \in \Theta$ and initial condition $\pmb{\xi} \in S$ , a "solution" to $\mathcal{H}(\pmb{\theta})$ to be a "hybrid arc", which is formally a set-valued map $\phi(\cdot; \pmb{\xi}, \pmb{\theta}): \mathbb{R}_{\geqslant 0} \times \mathbb{N} \to \mathbb{R}^n$ . We review Goebel et al.'s (2012) rigorous characterization of hybrid arcs as solutions to hybrid systems in appendix A.3. For ease of exposition in the main body of this paper, however, we use the concept of a time-parameterized solution map $\varphi$ , which we describe informally, below, in definition 2. An expanded, formal treatment of time-parameterized solution maps can be found in appendix A.4. + +Definition 2 ((Informal) Time-Parameterized Solution Map). Let $\varphi (\cdot ;\pmb {\xi},\pmb {\theta}):[0,t^{+})\to \mathbb{R}^{n}$ be called the time-parameterized solution map of $\mathcal{P} = (\mathcal{H},\mathcal{S},\Theta)$ , where $t^+ = \min_{\pmb {\xi},\pmb{\theta}}\sup_t\mathrm{dom}\phi (\cdot ;\pmb {\xi},\pmb {\theta})$ and where the hybrid arc $\phi (\cdot ;\pmb {\xi},\pmb {\theta})$ uniquely satisfies $\mathcal{H}(\pmb {\theta})$ from initial state $\pmb {\xi},\forall \pmb {\xi},\pmb {\theta}\in S\times \Theta$ + +The reader will note that $[0, t^{+}) \subset \mathbb{R}$ . In this paper, we focus strictly on finite time horizons, leaving the analysis of hybrid equilibria to future work — indeed, only the simplest hybrid systems equilibrate to a point, so equilibrium states are most productively defined as belonging to a set. Analyzing the causally relevant behavior of such sets requires machinery beyond our current scope, but our direct connection to established hybrid systems theory, in conjunction with the rich history of causal research on equilibrium models, provides a firm foundation to explore this in the future. Additionally, because hybrid arcs can dynamically evolve in event indices, Zeno and non-flowing solutions are possible, which can make $t^{+} = 0$ (if it only jumps) or arbitrarily small (if allowable initial conditions are close to Zeno accumulation points). We do not provide universal criteria in this paper under which $t^{+}$ is arbitrarily large. + +While we take the hybrid system $\mathcal{P}$ to accurately describe causally relevant mechanisms in the world, we impose assumptions on $\mathcal{P}$ indirectly. In particular, we assume that some auxiliary "upstream" system $\mathcal{P}_{\uparrow}$ can be "lowered" to produce $\mathcal{P} = \text{lower}(\mathcal{P}_{\uparrow})$ , and that the upstream $\mathcal{P}_{\uparrow}$ satisfies standard hybrid well-posedness conditions from the literature (the so-called hybrid basic conditions, detailed in assumption 4 of the appendix, and folded into assumption 1 below). While these conditions support our theoretical results and facilitate future extensions (e.g., to stability analyses), they inherently admit solution non-uniqueness, particularly at state-space boundaries where solutions could either jump or flow. Non-uniqueness, however, complicates both measurability arguments and downstream causal analysis. In this work, then, we formalize a practical approach that is standard in simulating hybrid systems by specifying that the solutions should be "flow preferring" — if a solution could both flow and jump, we choose the solution that flows (Sanfelice et al., 2023a; Sanfelice & Teel, 2010).4 Note also that a flow-preferring specification is consistent with computational implementations that trigger jumps when the jump-set boundary is crossed.5 + +A key component of the hybrid basic conditions is the outer semi-continuity of the jump set $G$ in the upstream system. Maintaining this property through intervention requires some bookkeeping on the boundaries between interventional jump sets, but must be handled such that "lowering" favors more recently applied model transformations. We achieve this bookkeeping through the use of an ordered set-valued map $G = x \mapsto \bigcup_{k=1}^{K} G_k(x)$ , where $\text{last}(G) = G_K$ . We fully formalize the ordered set-valued map in the appendix (c.f. definition 6). + +Definition 3. Let $\mathcal{P} = (C,F,D,G,\mathcal{S},\Theta) = (\mathcal{H},\mathcal{S},\Theta)$ . We set + +preferflow $(D,C,F) = \pmb {\theta}\mapsto D(\pmb {\theta})\backslash \{\pmb {\xi}\in S:\text{there is a flowing solution to}\mathcal{H}(\pmb {\theta})$ from $\pmb {\xi}\}$ + +$$ +\operatorname {l o w e r} (\mathcal {P}) = (C, F, D ^ {\prime}, G ^ {\prime}, \mathcal {S}, \Theta); \quad D ^ {\prime} = \text {p r e f e r f l o w} (D, C, F), G ^ {\prime} = \operatorname {l a s t} (G). +$$ + +The existence of a flowing solution from $\xi$ is meant in the sense established in the hybrid systems literature. See appendix A.5 (definitions 13 and 14) for details. We can now state our collected assumptions on a hybrid system, and prove the sufficiency of those assumptions for the existence, uniqueness, and measurability of the system's solution. See appendix A.6 (assumptions 3 to 5) for details, and appendix F for proof of lemma 1. + +Assumption 1. The parameterized hybrid system can be written as $\mathcal{P} = \text{lower}(\mathcal{P}_{\uparrow})$ , where $\mathcal{P}_{\uparrow} = (C, F, D, G, S, \Theta) = (\mathcal{H}, S, \Theta)$ , and the following hold for all $\pmb{\xi} \in S, \pmb{\theta} \in \Theta$ : + +1. there exists a unique, nontrivial solution to the differential inclusion $F$ (i.e. the continuous part of $\mathcal{H}(\pmb{\theta})$ ) that is Borel measurable in $\pmb{\xi}, \pmb{\theta}$ at any fixed $t \in [0, \infty)$ ; +2. $C$ is outer semi-continuous (osc) and $C(\pmb{\theta})$ closed; $F$ is osc, locally bounded, and $F(\pmb{x}, \pmb{\theta})$ is convex $\forall \pmb{x} \in S$ ; +3. $D(\pmb{\theta})$ is closed and $\mathcal{G}(D)$ is Borel; $G$ is osc, locally bounded; $\operatorname{last}(G)$ is single-valued and Borel measurable in $\pmb{\xi},\pmb{\theta}$ at any fixed $t\in [0,\infty)$ . + +Lemma 1. Let $\mathcal{P}$ satisfy assumption 1. Then $\mathcal{P}$ has a unique time-parameterized solution $\varphi (\cdot ;\pmb {\xi},\pmb {\theta}):$ $[0,t^{+})\to \mathbb{R}^{n}$ that is Borel-measurable in initial conditions $\pmb{\xi}$ and parameters $\pmb{\theta}$ at any fixed $t\in [0,t^{+})$ + +# 4 Instantaneous Interventions as Constraint Transformations + +We now formally define a general class of instantaneous interventions. We show that under certain natural assumptions, the class of systems meeting assumption 1 is "closed" under intervention — i.e., intervened systems will meet assumption 1 if the original system does. + +An instantaneous intervention can be implemented via modifications to the jump map and the jump set functions, respectively $G$ and $D$ in definition 1. To support parameterized interventions and/or stateful jumps, one can simply augment the state space $S$ and the parameter space $\Theta$ , essentially preserving all properties of interest (appendix B).7 + +Definition 4 (Instantaneous Intervention). Consider set-valued mappings $\tilde{D}:\Theta \rightrightarrows S$ and $\tilde{G}:$ $S\times \Theta \Rightarrow S$ and parameterized hybrid system $\mathcal{P} = (C,F,D,G,S,\Theta)$ . Now, let + +$$ +C ^ {\prime} (\boldsymbol {\theta}) = C (\boldsymbol {\theta}) \backslash \operatorname {i n t} \tilde {D} (\boldsymbol {\theta}) \tag {1} +$$ + +$$ +D ^ {\prime} (\boldsymbol {\theta}) = \text {p r e f e r f l o w} \left(\tilde {D}, C ^ {\prime}, F\right) (\boldsymbol {\theta}) \cup D (\boldsymbol {\theta}) \tag {2} +$$ + +$$ +G ^ {\prime} (\boldsymbol {x}, \boldsymbol {\theta}) = \left\{ \begin{array}{l l} \boldsymbol {x} \in D (\boldsymbol {\theta}) \backslash \tilde {D} (\boldsymbol {\theta}) & G (\boldsymbol {x}, \boldsymbol {\theta}) \\ \boldsymbol {x} \in \tilde {D} (\boldsymbol {\theta}) & \tilde {G} (\boldsymbol {x}, \boldsymbol {\theta}) \end{array} \right. \tag {3} +$$ + +then $\mathcal{P}' = \mathrm{instint}\left(\mathcal{P},\tilde{D},\tilde{G}\right) = \left(C',F,D',G',\mathcal{S},\Theta\right).$ + +In words, $\tilde{D}$ defines when (or where in the state space) the intervention will occur, while $\tilde{G}$ defines the state transition induced by the intervention. We make two important set-subtractions in this definition to preserve some useful and simplifying properties. First, we define $G^{\prime}$ in eq. (3) as preferring the new (i.e., the interventional) jump map $\tilde{G}$ wherever the original and new jump sets overlap. Second, the new flow set in eq. (1) has the interior of the new jump set removed. This preserves non-overlap between flow and jump sets, except possibly on the boundary, which we discuss below. + +$D^{\prime}(\pmb {\theta})$ is defined (eq. (2)) as the union over jump sets, except that a flow-preferring subtraction (definition 3) is made first on $\tilde{D} (\pmb {\theta})$ . Because $C^\prime (\pmb {\theta})$ has the interior of $\tilde{D} (\pmb {\theta})$ already removed, this subtraction operates only on the boundary of $\tilde{D} (\pmb {\theta})$ . In other words, $D^{\prime}(\pmb {\theta})$ will always contain the interior of $\tilde{D} (\pmb {\theta})$ , and will have the parts of its boundary removed where $F_{\theta}$ flows tangentially to or away from the interventional jump set. + +# 4.1 Intervention Preserves Existence, Uniqueness, and Measurability + +Interventional transformations should preserve key model properties. For causal models with explicit forward simulations, this is largely trivial. Hybrid systems, however, only implicitly characterize forward simulations (i.e., solutions) by specifying a set of constraints. Transformations to these constraints can easily fail to maintain key properties in the intervened world, such as whether a + +![](images/0f90ed3977e62bb38bca36d9888bef8073d73689f6ee39862289c50b507173a0.jpg) +Figure 2: Depiction of our "lowering" proof strategy for theorem 1. Proving theorem 1 requires a simple inductive generalization from lemma 4, which asserts that a single interventional transformation preserves key system properties, and is what we visualize here. Solid arrows indicate constraint transformations, while dotted arrows indicate that properties of one system imply properties of another. Assume that the parameterized hybrid system $\mathcal{P}$ accurately describes a domain of interest, and that it can be constructed by applying the lower transformation (definition 3) to a system $\mathcal{P}_{\uparrow}$ that fulfills assumption 1. Applying lower to such a system preserves existence and induces uniqueness and measurability (lemma 1). To simulate the effects of an intervention, we transform $\mathcal{P}$ into the model $\mathcal{P}'$ that describes the intervened world. $\mathcal{P}'$ can also be constructed, however, by applying a slightly modified intervention (instint $_{\uparrow}$ , definition 20) to $\mathcal{P}_{\uparrow}$ and then "lowering". Intervention on $\mathcal{P}_{\uparrow}$ maintains key properties in $\mathcal{P}'$ , which can, as before, be lowered to a system $\mathcal{P}'$ that must have a unique, measurable solution (lemma 1). + +solution is unique and measurable, or exists at all. The key theoretical contribution of this work, then, is to identify assumptions sufficient to ensure our interventional semantics preserves these properties through model transformation. Formal proof of theorem 1 is provided in appendix D, but we also include fig. 2 as a visual aid and proof sketch. + +Assumption 2 (Assumptions on Interventional Specifications). Consider mappings $\tilde{D}:\Theta \Rightarrow S$ and $\tilde{G}:S\times \Theta \Rightarrow S$ and parameterized hybrid system $\mathcal{P} = (C,F,D,G,S,\Theta)$ . For all $\theta \in \Theta$ , assume + +(I1) $\tilde{D}(\pmb{\theta})$ is closed and well-behaved relative to $\mathcal{P}$ (assumption 6). Additionally, the interior graph $\mathcal{G}(\mathrm{int}\tilde{D})$ is open and the graph $\mathcal{G}(\tilde{D})$ is Borel; +(I2) $\tilde{G}_{\theta}:S\to S$ is outer semi-continuous and locally bounded relative to $\tilde{D} (\pmb {\theta})$ , and $\tilde{D} (\pmb {\theta})\subset$ dom $\tilde{G}_{\theta}$ . Additionally, $\tilde{G}$ is single-valued and Borel-measurable. + +Theorem 1 (Compositions of Instantaneous Interventions Preserve Key Properties). Consider parameterized hybrid system $\mathcal{P}$ that meets assumption 1, and any finite sequence of $K$ set-valued mappings $(\tilde{D}_k)$ and $(\tilde{G}_k)$ , where each $\tilde{D}_k$ and $\tilde{G}_k$ fulfill assumption 2 relative to $\mathcal{P}$ . Let $\mathrm{instint}_k = \mathrm{instint}(\cdot ,\tilde{D}_k,\tilde{G}_k)$ (definition 4) and + +$$ +\mathcal {P} ^ {\prime} = \left(\operatorname {i n s t i n t} _ {1} \circ \dots \circ \operatorname {i n s t i n t} _ {k} \circ \dots \circ \operatorname {i n s t i n t} _ {K}\right) (\mathcal {P}), \tag {4} +$$ + +$\mathcal{P}'$ then meets assumption 1, and by lemma 1 $\mathcal{P}'$ has a unique time-parameterized solution $\varphi(\cdot; \pmb{\xi}, \pmb{\theta}): [0, t^{+}) \to \mathbb{R}^{n}$ , Borel-measurable in initial conditions $\pmb{\xi}$ and parameters $\pmb{\theta}$ at any fixed $t \in [0, t^{+})$ . + +# 5 Causal Estimands as Functionals of Twin Distributions + +In this section, we exercise our framework to define three basic causal estimands. Importantly, many of the more complex causal analyses build on these basic inference capabilities. Most targets of causal inference take the form of (conditional) expectations, and so we must now use our measurability results to define those expectations with respect to random solution maps.10 First, we will generalize + +the parameterized hybrid system to include random initial conditions and parameters. Then, we will establish some notation and define the expected treatment effect, data-conditional treatment effect, and the basic counterfactual query using our machinery. + +Definition 5 (Hybrid System with Random Inputs). A parameterized hybrid system with random inputs is characterized by the tuples + +$$ +\mathcal {R} = (\mathcal {P}, \boldsymbol {\xi}, \boldsymbol {\theta}); \mathcal {P} = (C, F, D, G, \mathcal {S}, \Theta). +$$ + +We take the probability space $(\Omega, \mathcal{F}, \mathbb{P})$ as implied by $\mathcal{R}$ , where $\xi: \Omega \to S$ and $\theta: \Omega \to \Theta$ are measurable with respect to $\mathcal{F}$ and the Borel $\sigma$ -algebras on $S \subseteq \mathbb{R}^n$ and $\Theta \subseteq \mathbb{R}^m$ . + +When clear from context, for some $\omega \in \Omega$ , we often write $\xi$ and $\theta$ in place of $\xi(\omega)$ and $\theta(\omega)$ , respectively. We distinguish random variables $\xi$ and $\theta$ from possible values $\xi \in S$ and $\theta \in \Theta$ by the upright font. From here, we can directly consider evaluations of the solution as a measurable random variable. A direct consequence of lemma 1, which states that lower induces measurability when applied to a system that fulfills assumption 1, is the following. + +Corollary 1 (Random Time-Parameterized Solution is Measurable). Consider parameterized hybrid system with random inputs $\mathcal{R} = (\mathcal{P},\xi ,\theta)$ , where $\mathcal{P}$ satisfies assumption 1. Then, by lemma 1, $\mathcal{P}$ has a unique, time-parameterized solution map $\varphi$ , and the composition $\omega \mapsto \varphi (t;\xi (\omega),\theta (\omega))$ defines an $\mathcal{F}$ -measurable random variable at any fixed $t\in [0,t^{+})$ . + +Having established conditions under which intervention preserves measurability, we can begin constructing estimands from the parallel worlds created through intervention. In estimands, we use symbolic subscripts to delineate parallel worlds. Consider an original system $\mathcal{R}_0 = (\mathcal{P}_0,\xi ,\theta)$ . We might then apply an intervention characterized by $\tilde{D}_s$ and $\tilde{G}_s$ to yield $\mathcal{P}_s = \mathrm{instint}(\mathcal{P}_0,\tilde{D}_s,\tilde{G}_s)$ . By convention, we use $\mathcal{R}_s = (\mathcal{P}_s,\xi ,\theta)$ in reference to the full specification for the intervened world, and $t\mapsto \varphi_s(t;\xi ,\theta)$ for its random, time-parameterized solution (corollary 1). We often write $\varphi_s^t$ in place of $\varphi_{s}(t;\xi ,\theta)$ for brevity. Lastly, supposing we wish to focus on a particular element of the state vector at time $t$ , we sometimes define a random function that appropriately indexes into the solution vector. For example, we might have that $h_s(t;\xi ,\theta) = \varphi_s^{(i)}(t;\xi ,\theta)$ always, where $h$ represents the solution map for the $(i)$ 'th state element. We similarly sometimes use $h_s^t = h_s(t;\xi ,\theta)$ . We can exercise this notation with the following examples. + +Example 2 (Expected Treatment Effect). Consider $\mathcal{R}_0 = (\mathcal{P}_0, \xi, \theta)$ and interventional jump set $\tilde{D}$ and map $\tilde{G}$ . Assume these components fulfill assumptions 1 and 2. Let $\mathcal{P}_1 = \mathrm{instint}(\mathcal{P}_0, \tilde{D}, \tilde{G})$ and $\varphi_0$ and $\varphi_1$ be the time-parameterized solution maps of the original and intervened worlds. Let $y_0$ and $y_1$ be the solution maps for the $(i)$ 'th element of the state vector. The expected treatment effect at some time $\tau \in [0, \min[t_0^+, t_1^+]) = [0, t^+)$ can be written equivalently as + +$$ +\mathbb {E} \left[ y _ {1} ^ {\tau} - y _ {0} ^ {\tau} \right] = \mathbb {E} \left[ y _ {1} (\tau ; \boldsymbol {\xi}, \boldsymbol {\theta}) - y _ {0} (\tau ; \boldsymbol {\xi}, \boldsymbol {\theta}) \right] = \mathbb {E} \left[ \varphi_ {1} ^ {(i)} (\tau ; \boldsymbol {\xi}, \boldsymbol {\theta}) - \varphi_ {0} ^ {(i)} (\tau ; \boldsymbol {\xi}, \boldsymbol {\theta}) \right]. +$$ + +Example 3 (Data-Conditional Treatment Effect). Building immediately off example 2, we can specify a data-conditional treatment effect that takes factual observations into account.11 Let $w_0$ be the solution map for some element of the state vector. For some finite set of observation times $\{t_k\}_{k=1}^K \subset [0, t_0^+)$ , the data-conditional treatment effect can then be written as12 + +$$ +\mathbb {E} \left[ y _ {1} ^ {\tau} - y _ {0} ^ {\tau} \mid \boldsymbol {v} _ {0} \right]; \boldsymbol {v} _ {0} \sim \mathcal {N} \left(\boldsymbol {w} _ {0}, \sigma^ {2}\right); \boldsymbol {w} _ {0} = \left[ w _ {0} ^ {t _ {k}} \right] _ {k = 1} ^ {K}. +$$ + +Example 4 (Counterfactual Outcome). Also building off example 2, consider factual outcome event that $y_0(\tau; \xi, \theta) = \bar{y}_0^\tau \in \mathbb{R}$ . The counterfactual outcome, then, can be derived by conditioning on that factual event. + +$$ +\mathbb {E} [ y _ {1} ^ {\tau} \mid y _ {0} ^ {\tau} = \bar {y} _ {0} ^ {\tau} ] = \mathbb {E} [ y _ {1} (\tau ; \boldsymbol {\xi}, \boldsymbol {\theta}) \mid y _ {0} (\tau ; \boldsymbol {\xi}, \boldsymbol {\theta}) = \bar {y} _ {0} ^ {\tau} ]. +$$ + +
queryoutcomeprobability
nec.Y = I[bτq2] ≤ γ]Pr (Yx' = 0 | X = 1, Y = 1) = Pr ([bτq2 ≤ γ] | bτq1 > γ)
suf.Y = I[bτq2] > γ]Pr (Yx' = 1 | X = 0, Y = 0) = Pr ([bτq2 > γ] | bτq1 ≤ γ)
nec. and suf.Y = I[bτq2] ≤ γ]I[bτq1] ≥ γ]Pr (Yx = 1, Yx' = 0) = Pr (bτq1 > γ, bτq2 ≤ γ)
+ +Table 1: Identities for the probabilities of causation in the fishery management example. Under TAC quota $q_{i}$ , the biomass of the fished species at time $\tau$ is given by $\mathsf{b}_{q_i}^{\tau}$ . The outcome $Y$ is achieved when that biomass meets or exceeds $\gamma$ . We rely on the standard exogeneity conditions $Y_{x} \perp X$ and $Y_{x'} \perp X$ ,[15] and the fact that, conditioned on $X = 1(X_2)$ , $Y$ reduces to the outcome only in the world with allowable catch set to $q_{1}(q_{2})$ . + +Many of the more complex causal inference tasks — such as mediation analysis, the estimation of population-level conditional average treatment effects, or even actual cause assessments — are constructed from the counterfactual building blocks we propose here. Indeed, once a counterfactual semantics is established, and a twin-world or potential-outcomes syntax (e.g., differentiating $y_0$ from $y_1$ ) is enumerated, many estimands are straightforward and familiar to develop. In the next section, we explore just such a class of estimand: the probabilities of causation. + +# 6 Case Study: Necessary and Sufficient Causation + +To illustrate a more sophisticated application of our interventional semantics, we map the standard definitions for the probabilities of necessary and sufficient causation (originally formalized by Pearl (1999)) onto dynamically triggered, discontinuous interventions in hybrid systems. In particular, we work in the fishery management domain where regulators employ Total Allowable Catch (TAC) policies to dynamically end the commercial fishing season after caught biomass reaches certain quotas. If interested, the reader may wish to review appendix G.1, in which we provide motivating historical context for this domain. Additionally, we review Pearl's original formulation of the probabilities of causation (PoC) in appendix G.2. Throughout appendix G, we provide full simulation analyses of the case study. Code is available here, $^{13}$ and relies on the dynamical systems package from the causal probabilistic programming language ChiRho (Basis-Research, 2025). $^{14}$ + +We focus on a hypothetical fishery involving three trophic levels — apex predators, intermediate predators (the fished species), and forage fish — with dynamics captured by the differential equations presented by Zhou & Smith (2017). Throughout a single season, fishing pressure is modeled at a constant rate applied to the intermediate predator, plus some bycatch on the apex trophic level. Regulators intervene by ending the fishing season (setting the catch rate to zero) when the integrated catch reaches a predefined TAC quota. The goal of these policies is to ensure that the biomass of the target fishery species recovers to sustainable level $\gamma$ by the beginning of the next season. + +In this context, stakeholders may debate the necessity and/or sufficiency of certain regulatory policies in maintaining joint ecological and economic goals for the fishery. The probabilities of causation are formal tools supporting the assessment of causal attribution between causes and their (supposed) effects. Pearl (1999) first formalized the PoC for binary treatments and outcomes — here, however, both the TAC quota and the biomass are scalar valued. We therefore follow Kawakami et al.'s (2024) generalization of the PoCs to support contrastive queries between scalar-valued treatments and their thresholded outcomes (see their Def. 3.1). Consider two TAC quotas $q_{1}$ and $q_{2}$ , and the following natural language queries. In table 1, we provide the formalized estimands written in our notation. + +- necessity: in worlds where the end-of-year biomass levels exceed the target level $\gamma$ (success) under quota $q_{1}$ , what is the probability of failure had regulators used quota $q_{2}$ instead? +- sufficiency: in worlds where the end-of-year biomass levels remain below the target level $\gamma$ (failure) under $q_{1}$ , what is the probability of success had regulators used $q_{2}$ instead? +- necessity and sufficiency: what is the probability that both (1) $q_{1}$ results in success and (2) $q_{2}$ results in failure? + +For readers less familiar with the applications of the PoC to decision and policy making, we provide an expanded narrative scaffolding for this example in appendix G.4. In appendix G.5, we provide an additional example designed to highlight how certain natural language ambiguities in causal attribution queries — particularly those involving multi-faceted, real world events and policies — can be formally clarified. + +The PoC queries above rely on the construction of twin, contrastive worlds — one with TAC quota $q_{1}$ , and the other with $q_{2}$ . To model these worlds, we start with a system $\mathcal{P}$ characterizing year-round fishing pressure (i.e., no regulatory intervention), and then transform its constraints to add a dynamic, season-ending intervention. Notationally, let $h_{i}$ represent the harvest rate, and $b_{i}$ the biomass, at trophic level $i$ , and let $z$ be the total catch (integral of $\dot{z} = h_{2}b_{2}$ ) at the intermediate trophic level. The system state can be conceptualized as $[z,h_{1:3},b_{1:3}] = \boldsymbol{x}\in S = \mathbb{R}_{\geqslant}^{7}$ . + +The regulatory, season-ending intervention can be modeled by dynamically setting harvest rates to zero when the catch exceeds a threshold $q_{i}$ (with $i \in \{1,2\}$ ). By using our interventional semantics, we can construct parallel worlds with the same random initial conditions and parameters. See appendix G.3 for a generalization of this model to multi-season time scales. + +$$ +\tilde {D} _ {q _ {i}} (\boldsymbol {\theta}) = \left\{z \in \mathbb {R} _ {\geqslant 0} \mid z \geqslant q _ {i} \right\} \times \mathbb {R} _ {\geqslant 0} ^ {6}; \quad \tilde {G} _ {q _ {i}} (\boldsymbol {x}, \boldsymbol {\theta}) = \left\{\left[ z, 0, 0, 0, b _ {1: 3} \right] \right\}; \tag {5} +$$ + +$$ +\mathcal {R} _ {s} = \left(\mathcal {P}, \xi , \theta\right); \quad \mathcal {R} _ {q _ {1}} = \left(\mathcal {P} _ {q _ {1}}, \xi , \theta\right); \quad \mathcal {R} _ {q _ {2}} = \left(\mathcal {P} _ {q _ {2}}, \xi , \theta\right); \quad \mathcal {P} _ {q _ {i}} = \text {i n s t i n t} \left(\mathcal {P}, \tilde {\mathcal {D}} _ {q _ {i}}, \tilde {\mathcal {G}} _ {q _ {i}}\right). \tag {6} +$$ + +# 7 Limitations and Future Work + +Most research developing causal inference tools starts by casting a problem in the format of structural causal models (SCMs) (Pearl, 2009). Our work differs in that we construct our counterfactual semantics directly in the parlance of hybrid systems. These two tacks are compatible, however. For example, with our measurability results in hand, the time-parameterized solution map $\varphi$ can be treated as a structural equation with initial conditions $\xi$ and parameters $\theta$ viewed as parent variables in a larger SCM. Our interventional semantics, then, exposes the causal dynamics of the hybrid system for manipulation. When $\varphi$ is interpreted as a structural equation, our semantics could be viewed as characterizing a family of "soft-interventions" (Correa & Barenboim, 2020) on the solution map. Importantly, the adjoint method (Chen et al., 2018) can be used in tandem with auto-differentiation machinery to learn "event function" (i.e. jump set) parameters, $^{16}$ thereby supporting end-to-end differentiation of composite SCM and hybrid system models. Relatedly, equivalent forward-Euler representations may prove useful in actual cause analysis of hybrid systems (Halpern & Peters, 2022). + +This leaves a few limitations to review. First, we do not present non-parametric, estimand-specific identification results — indeed, there may exist sufficient conditions for estimand identification that are weaker than those established for full system identification. Second, as discussed following definition 1, we focus only on finite time regimes, leaving analysis of hybrid equilibria to future work. Furthermore, we do not provide conditions under which intervention preserves non-Zeno behavior. + +# 8 Conclusion + +This paper has strengthened the connection between the modeling capabilities offered by hybrid systems theory and the causal reasoning capabilities developed by the causal inference research community. + +We characterize and demonstrate a counterfactual semantics for a class of dynamically triggered, instantaneous interventions that underpin many closed-loop control problems. Bypassing an explicit re-casting of hybrid systems as structural causal models, we use hybrid systems as the primary modeling substrate. This allows clear connections to the extensive body of work on hybrid systems theory, in which we can derive and characterize mild conditions under which solution exists, uniqueness and measurability are preserved in the intervened system. + +Finally, we illustrate the flexibility and power of the resulting framework by first formalizing common causal estimands for hybrid systems, and then by developing a case study using the three probabilities of causation in the context of fishery management. + +# Acknowledgments and Disclosure of Funding + +AZ, DB, RU, JZ, and SW were supported on DARPA Automating Scientific Knowledge Extraction and Modeling (ASKEM) program Grant HR00112220036. We thank Anirban Chaudhuri, Sabina Altus, Joseph Cottam, and Neeraj Kumar for their insights and contributions throughout the ASKEM program; our colleagues at Basis for helpful comments and discussions; Eli Bingham for guiding our thinking and software design choices around these ideas; Paul Wintz for answering many questions about hybrid systems and pointing us to relevant literature; and David Jensen for helpful comments and discussion. Pacific Northwest National Laboratory (PNNL) is a multiprogram national laboratory operated by Battelle for the DOE under Contract DEAC05-76RLO 1830. The views, opinions and/or findings expressed are those of the author and should not be interpreted as representing the official views or policies of the Department of Defense or the U.S. Government. Distribution Statement "A" (Approved for Public Release, Distribution Unlimited). + +# References + +Anon. Magnuson-Stevens Fishery Conservation and Management Act, 1976. URL https://www.fisheries.noaa.gov/s3//dam-migration/msa-amended-2007.pdf. +Anon. Sustainable Fisheries Act, 1996. +Anon. Magnuson-Stevens Act Provisions; Foreign Fishing; Fisheries off West Coast States and in the Western Pacific; Pacific Coast Groundfish Fishery; Annual Specifications and Management Measures, January 2000. URL https://www.federalregister.gov/documents/2000/01/04/99-33966/magnuson-stevens-act-provisions-foreign-fishing-fiseries-off-west-coast-states-and-in-the-western. Volume: 65 Docket Number: 991223347-9347-01. +Anon. Magnuson-Stevens Fishery Conservation and Management Reauthorization Act, 2007a. URL https://www.fisheries.noaa.gov/s3//dam-migration/msa-amended-2007.pdf. +Anon. Atlantic Highly Migratory Species (HMS); U.S. Atlantic Swordfish Fishery Management Measures, June 2007b. URL https://www.federalregister.gov/documents/2007/06/07/E7-10727/atlantic-highly-migratory-species-hms-us-atlantic-swordfish-fishery-management-measures. Volume: 72 Docket Number: 061121306-7105-02. +Anon. Fisheries Off West Coast States; Pacific Coast Groundfish Fishery Management Plan; Amendments 20 and 21; Trawl Rationalization Program, December 2010a. URL https://www.federalregister.gov/documents/2010/12/15/2010-30527/fiseries-off-west-coast-states-pacific-coast-groundfish-fishery-management-plan-amendments-20-and. Volume: 75 Docket Number: 100212086-0532-05. +Anon. Report of the 2009 Atlantic Swordfish Stock Assessment Session. Collect. Vol. Sci. Pap. ICCAT, 65(1):1-123, 2010b. URL https://www.iccat.int/Documents/CVSP/CV065_2010/n_1/CV065010001.pdf. +Anon. Modernizing Recreational Fisheries Management Act, 2018. +Charles K. Assaad, Emilie Devijver, and Eric Gaussier. Survey and Evaluation of Causal Discovery Methods for Time Series. Journal of Artificial Intelligence Research, 73:767-819, February 2022. ISSN 1076-9757. doi: 10.1613/jair.1.13428. URL https://www.jair.org/index.php/jair/article/view/13428. +Chen Avin, Ilya Shpitser, and Judea Pearl. Identifiability of path-specific effects. In Proceedings of the 19th international joint conference on artificial intelligence, IJCAI'05, pp. 357-363, Edinburgh, Scotland, 2005. Morgan Kaufmann Publishers Inc. +Alexander Balke and Judea Pearl. Probabilistic evaluation of counterfactual queries. In Proceedings of the twelfth AAAI national conference on artificial intelligence, AAAI'94, pp. 230-237, Seattle, Washington, 1994. AAAI Press. +Basis-Research. Chirho, 2025. URL https://github.com/BasisResearch/chirho. + +Sander Beckers. Causal Explanations and XAI. In Bernhard Schölkopf, Caroline Uhler, and Kun Zhang (eds.), Proceedings of the First Conference on Causal Learning and Reasoning, volume 177 of Proceedings of Machine Learning Research, pp. 90-109. PMLR, April 2022. URL https://proceedings.mlr.press/v177/beckers22a.html. +Sander Beckers and Joseph Y. Halpern. Abstracting Causal Models, July 2019. URL http://arxiv.org/abs/1812.03789. arXiv:1812.03789 [cs]. +Sander Beckers, Frederick Eberhardt, and Joseph Y. Halpern. Approximate Causal Abstraction, June 2019. URL http://arxiv.org/abs/1906.11583. arXiv:1906.11583 [cs]. +Sander Beckers, Joseph Halpern, and Christopher Hitchcock. Causal Models with Constraints. In Mihaela van der Schaar, Cheng Zhang, and Dominik Janzing (eds.), Proceedings of the second conference on causal learning and reasoning, volume 213 of Proceedings of machine learning research, pp. 866-879. PMLR, April 2023. URL https://proceedings.mlr.press/v213/beckers23a.html. +Eli Bingham, James Koppel, Alexander Lew, Robert Ness, Zenna Tavares, Sam Witty, and Jeremy Zucker. Causal Probabilistic Programming Without Tears. In Proceedings of the third conference on probabilistic programming, 2021. +Tineke Blom and Joris M. Mooij. Causality and Independence in Perfectly Adapted Dynamical Systems, February 2023. URL http://arxiv.org/abs/2101.11885.arXiv:2101.11885 [cs]. +Tineke Blom, Stephan Bongers, and Joris M. Mooij. Beyond Structural Causal Models: Causal Constraints Models, August 2019. URL http://arxiv.org/abs/1805.06539. arXiv:1805.06539 [cs]. +Tineke Blom, Mirthe M. van Diepen, and Joris M. Mooij. Conditional independences and causal relations implied by sets of equations, January 2021. URL http://arxiv.org/abs/2007.07183.arXiv:2007.07183 [cs]. +Philip Boeken and Joris M. Mooij. Dynamic Structural Causal Models, July 2024. URL http://arxiv.org/abs/2406.01161. arXiv:2406.01161 [math]. +Stephan Bongers. Causal Modeling & Dynamical Systems: A New Perspective On Feedback. PhD thesis, Universiteit van Amsterdam, 2022. URL https://hdl.handle.net/11245.1/652541c6-8959-498c-8958-fd28f198bfdf. +Dan G. Cacuci. Sensitivity & Uncertainty Analysis, Volume 1. Chapman and Hall/CRC, 0 edition, May 2003. ISBN 978-1-135-44298-9. doi: 10.1201/9780203498798. URL https://www.taylorfrancis.com/books/9781135442989. +Christos G. Cassandras and John Lygeros. Stochastic Hybrid Systems. Automation and Control Engineering. Taylor and Francis, Hoboken, 2010. ISBN 978-0-8493-9083-8. +Ricky T. Q. Chen, Yulia Rubanova, Jesse Bettencourt, and David K Duvenaud. Neural Ordinary Differential Equations. In Advances in neural information processing systems, volume 31, 2018. URL https://proceedings.neurips.cc/paper_files/paper/2018/file/69386f6bb1dfed68692a24c8686939b9-Paper.pdf. +H. Chockler and J. Y. Halpern. Responsibility and Blame: A Structural-Model Approach. Journal of Artificial Intelligence Research, 22:93-115, October 2004. ISSN 1076-9757. doi: 10.1613/jair.1391. URL https://jair.org/index.php/jair/article/view/10386. +J. Correa and E. Bareinboim. General Transportability of Soft Interventions: Completeness Results. In H. Larochelle, M. Ranzato, R. Hadsell, M. F. Balcan, and H. Lin (eds.), Advances in neural information processing systems, volume 33, pp. 10902-10912, Vancouver, Canada, June 2020. Curran Associates, Inc. / Causal Artificial Intelligence Lab, Columbia University. Number: R-68. +Lori A. Cramer, Courtney Flathers, Deanna Caracciolo, Suzanne M. Russell, and Flaxen Conway. Graying of the Fleet: Perceived Impacts on Coastal Resilience and Local Policy. *Marine Policy*, 96:27–35, October 2018. ISSN 0308597X. doi: 10.1016/j.marpol.2018.07.012. URL https://linkinghub.elsevier.com/retrieve/pii/S0308597X17308631. + +Denver Dash. Caveats for Causal Reasoning with Equilibrium Models. May 2003. URL https://d-scholarship.pitt.edu/7811/. +Patrick Forre and Joris M. Mooij. Causal Calculus in the Presence of Cycles, Latent Confounders and Selection Bias. In Ryan P. Adams and Vibhav Gogate (eds.), Proceedings of the 35th uncertainty in artificial intelligence conference, volume 115 of Proceedings of machine learning research, pp. 71-80. PMLR, July 2020. URL https://proceedings.mlr.press/v115/forre20a.htm1. +Rafal Goebel, Ricardo G. Sanfelice, and Andrew R. Teel. Hybrid dynamical systems: modeling, stability, and robustness. Princeton university press, Princeton (N.J.), 2012. ISBN 978-0-691-15389-6. +Joseph Y. Halpern and Judea Pearl. Causes and Explanations: a Structural-Model Approach. Part I: Causes. British Journal for the Philosophy of Science, 56(4):843–887, 2005a. doi: 10.1093/bjps/axi147. Publisher: Oxford University Press. +Joseph Y. Halpern and Judea Pearl. Causes and Explanations: A Structural-Model Approach. Part II: Explanations. The British Journal for the Philosophy of Science, 56(4):889-911, December 2005b. ISSN 0007-0882, 1464-3537. doi: 10.1093/bjps/axi148. URL https://www.journals.uchicago.edu/doi/10.1093/bjps/axi148. +Joseph Y. Halpern and Spencer Peters. Reasoning about Causal Models with Infinitely Many Variables. Proceedings of the AAAI Conference on Artificial Intelligence, 36(5):5668-5675, June 2022. ISSN 2374-3468, 2159-5399. doi: 10.1609/aaai.v36i5.20508. URL https://ojs(aaai.org/index.php/AAAI/article/view/20508. +Niels Hansen and Alexander Sokol. Causal Interpretation of Stochastic Differential Equations. Electronic Journal of Probability, 19(none), January 2014. ISSN 1083-6489. doi: 10.1214/EJP.v19-2891. URL https://projecteuclid.org/journals/electronic-journal-of-probability/volume-19/issue-none/Causal-interpretation-of-stochastic-differential-equations/10.1214/EJP.v19-2891.full. +Miguel Hernán and James M. Robins. Causal Inference: What If. Taylor and Francis, Boca Raton, first edition edition, 2023. ISBN 978-1-4200-7616-5 978-0-367-71133-7. +Antti Hyttinen, Frederick Eberhardt, and Patrik O. Hoyer. Learning Linear Cyclic Causal Models with Latent Variables. Journal of Machine Learning Research, 13(109):3387-3439, 2012. URL http://jmlr.org/papers/v13/hyttinen12a.html. +Guido W. Imbens and Donald B. Rubin. Causal Inference for Statistics, Social, and Biomedical Sciences. Cambridge books. Cambridge University Press, September 2015. URL https://ideas.repec.org/b/cup/cbooks/9780521885881.html. Number: 9780521885881. +Yumi Iwasaki and Herbert A. Simon. Causality and Model Abstraction. Artificial Intelligence, 67 (1):143-194, May 1994. ISSN 00043702. doi: 10.1016/0004-3702(94)90014-0. URL https://linkinghub.elsevier.com/retrieve/pii/0004370294900140. +Ryan Johnson. Parameter Estimation for Hybrid Dynamical Systems. PhD thesis, University of California Santa Cruz, 2023. +Haley Kappus-Kron, Dana Ahmad Chatila, Ainsley Mabel MacLachlan, Nicole Pulido, Nan Yang, and David A. Larsen. Precision public health in schools enabled by wastewater surveillance: A case study of COVID-19 in an Upstate New York middle-high school campus during the 2021-2022 academic year. PLOS Global Public Health, 4(1):e0001803, January 2024. ISSN 2767-3375. doi: 10.1371/journal.pph.0001803. URL https://dx.plos.org/10.1371/journal.pph.0001803. +Yuta Kawakami, Manabu Kuroki, and Jin Tian. Probabilities of Causation for Continuous and Vector Variables. In Negar Kiyavash and Joris M. Mooij (eds.), Proceedings of the fortieth conference on uncertainty in artificial intelligence, volume 244 of Proceedings of machine learning research, pp. 1901-1921. PMLR, July 2024. URL https://proceedings.mlr.press/v244/kawakami24a.html. + +Donald E. Kirk. Optimal Control Theory: an Introduction. Dover Publications, Mineola, N.Y, 2004. ISBN 978-0-486-43484-1. +Gustavo Lacerda, Peter Spirtes, Joseph Ramsey, and Patrik O. Hoyer. Discovering Cyclic Causal Models by Independent Components Analysis. In Proceedings of the twenty-fourth conference on uncertainty in artificial intelligence, UAI'08, pp. 366-374, Helsinki, Finland, 2008. AUAI Press. ISBN 0-9749039-4-9. Number of pages: 9 tex/address: Arlington, Virginia, USA. +Donna J. Lee, Sherry Larkin, and Charles M. Adams. A Bioeconomic Analysis of Management Alternatives for the U.S. North Atlantic Swordfish Fishery. *Marine Resource Economics*, 15:77–96, 2000. URL https://api.sementicscholar.org/CorpusID:150518179. +Ang Li and Judea Pearl. Probabilities of Causation with Nonbinary Treatment and Effect. Proceedings of the AAAI Conference on Artificial Intelligence, 38(18):20465-20472, March 2024. ISSN 2374-3468, 2159-5399. doi: 10.1609/aaai.v38i18.30030. URL https://ojs.aaai.org/index.php/AAAI/article/view/30030. +Lennart Ljung. System Identification: Theory for the User. Prentice-Hall information and system sciences series. Prentice Hall, Upper Saddle River, NJ, 2. ed., 14. printing edition, 2012. ISBN 978-0-13-656695-3. +Joris M. Mooij, Dominik Janzing, and Bernhard Scholkopf. From Ordinary Differential Equations to Structural Causal Models: the deterministic case, April 2013. URL http://arxiv.org/abs/1304.7920. arXiv:1304.7920 [stat]. +Kevin P. Murphy. Dynamic bayesian networks: Representation, inference and learning. phd, EECS Department, University of California, Berkeley, 2002. +John D. Neilson, Freddy Arocha, Shannon L. Cass-Calay, Jaime Mejuto, Mauricio Ortiz, Gerald P. Scott, Craig Smith, Paulo Travassos, George Tserpes, and Irene V. Andrushchenko. The recovery of atlantic swordfish: The comparative roles of the regional fisheries management organization and species biology. Reviews in Fisheries Science, 21:59 - 97, 2013. URL https://api. semanticscholar.org/CorpusID:55132285. +Mauricio Ortiz, Shannon L Cass-Calay, and Gerald P Scott. A Potential Framework for Evaluating the Efficacy of Biomass Limit Reference Point in the Presence of Natural Variability and Parameter Uncertainty: An Application to Northern Albacore Tuna (Thunnus Alalunga). Collect. Vol. Sci. Pap. ICCAT, 65(4):1254-1267, 2010. +Judea Pearl. Causal Diagrams for Empirical Research. Biometrika, 82(4):669-688, 1995. ISSN 0006-3444, 1464-3510. doi: 10.1093/biomet/82.4.669. URL https://academic.oup.com/biomet/article-lookup/doi/10.1093/biomet/82.4.669. +Judea Pearl. Probabilities Of Causation: Three Counterfactual Interpretations And Their Identification. Synthese, 121(1/2):93-149, 1999. ISSN 00397857. doi: 10.1023/A:1005233831499. URL http://link.springer.com/10.1023/A:1005233831499. +Judea Pearl. Direct and Indirect Effects. In Proceedings of the seventeenth conference on uncertainty in artificial intelligence, UAI'01, pp. 411-420, San Francisco, CA, USA, 2001. Morgan Kaufmann Publishers Inc. ISBN 1-55860-800-1. Number of pages: 10 Place: Seattle, Washington. +Judea Pearl. Causality: Models, Reasoning and Inference. Cambridge University Press, USA, 2 edition, 2009. ISBN 0-521-89560-X. +Jonas Peters, Stefan Bauer, and Niklas Pfister. Causal Models for Dynamical Systems, January 2020. URL http://arxiv.org/abs/2001.06208.arXiv:2001.06208 [stat]. +Spencer Peters and Joseph Halpern. A Unifying Framework for Causal Modeling With Infinitely Many Variables. Journal of Artificial Intelligence Research, 83, August 2025. ISSN 1076-9757. doi: 10.1613/jair.1.15612. URL https://www.jair.org/index.php/jair/article/view/15612. +Spencer Peters and Joseph Y. Halpern. Causal Modeling With Infinitely Many Variables, December 2021. URL http://arxiv.org/abs/2112.09171. arXiv:2112.09171 [cs]. + +A. Raue, C. Kreutz, T. Maiwald, J. Bachmann, M. Schilling, U. Klingmüller, and J. Timmer. Structural and Practical Identifiability Analysis of Partially Observed Dynamical Models by Exploiting the Profile Likelihood. Bioinformatics, 25(15):1923-1929, August 2009. ISSN 1367-4811, 1367-4803. doi: 10.1093/bioinformatics/btp358. URL https://academic.oup.com/bioinformatics/article/25/15/1923/213246. +Andrea J. Ray. Reservoir Management in the Interior West. In Henry F. Diaz and Barbara J. Morehouse (eds.), Climate and water: Transboundary challenges in the americas, pp. 193-217. Springer Netherlands, Dordrecht, 2003. ISBN 978-94-015-1250-3. doi: 10.1007/978-94-015-1250-3_9. URL https://doi.org/10.1007/978-94-015-1250-3_9. +Victor R. Restrepo, Joseph E. Powers, Stephen C. Turner, and John M. Hoenig. Using Simulation to Quantify Uncertainty in Sequential Population Analysis (SPA) and Derived Statistics, with Application to the North Atlantic Swordfish Fishery. In International Council for the Exploration of the Sea, 2011. URL https://api.sementanticscholar.org/CorpusID:251047550. +Donald B. Rubin. Estimating Causal Effects of Treatments in Randomized and Nonrandomized Studies. Journal of Educational Psychology, 66(5):688-701, October 1974. ISSN 1939-2176, 0022-0663. doi: 10.1037/h0037350. URL https://doi.apa.org/doi/10.1037/h0037350. +Jakob Runge, Andreas Gerhardus, Gherardo Varando, Veronika Eyring, and Gustau Camps-Valls. Causal inference for time series. Nature Reviews Earth & Environment, 4:487-505, 2023. doi: 10.1038/s43017-023-00431-y. URL https://doi.org/10.1038/s43017-023-00431-y. +Ricardo Sanfelice, Paul Wintz, David Copp, and Pablo Nanez. Behavior in the Intersection of C and D, 2023a. URL https://hyeq.github.io/simulink/intersection-of-C-and-D. +Ricardo Sanfelice, Paul Wintz, David Copp, and Pablo Nanez. HyEQ Toolbox, 2023b. URL https://github.com/pnanez/HyEQ_Toolbox. +Ricardo G. Sanfelice. Hybrid Feedback Control. Princeton University Press, Princeton, New Jersey, 2021. ISBN 978-0-691-18022-9 978-0-691-18953-6. +Ricardo G. Sanfelice and Andrew R. Teel. Dynamical Properties of Hybrid Systems Simulators. Automatica, 46(2):239-248, February 2010. ISSN 00051098. doi: 10.1016/j.automatica.2009.09.026. URL https://linkinghub.elsevier.com/retrieve/pii/S0005109809004361. +Adnane Saoud, Mohamed Maghenem, Antonio Loria, and Ricardo G. Sanfelice. Hybrid Persistence of Excitation in Adaptive Estimation for Hybrid Systems. IEEE Transactions on Automatic Control, 69(12):8828-8835, December 2024. ISSN 0018-9286, 1558-2523, 2334-3303. doi: 10.1109/TAC.2024.3422248. URL https://ieeexplore.ieee.org/document/10582521/. +Abraham Jan van der Schaft, Johannes M. Schumacher, and Arjan van der Schaft. An Introduction to Hybrid Dynamical Systems. Number 251 in Lecture notes in control and information sciences. Springer, London, 2000. ISBN 978-1-85233-233-4. +Ilya Shpitser and Judea Pearl. Identification of Joint Interventional Distributions in Recursive Semi-Markovian Causal Models. In Proceedings of the 21st national conference on artificial intelligence - volume 2, AAAI'06, pp. 1219-1226. AAAI Press, 2006. ISBN 978-1-57735-281-5. Place: Boston, Massachusetts Number of pages: 8. +Ilya Shpitser and Judea Pearl. Complete Identification Methods for the Causal Hierarchy. Journal of Machine Learning Research, 9(64):1941-1979, 2008. URL http://jmlr.org/papers/v9/ shpitser08a.html. +Peter L. Spirtes. Directed Cyclic Graphical Representations of Feedback Models, February 2013. URL http://arxiv.org/abs/1302.4982.arXiv:1302.4982 [cs]. +A. M. Stuart. Inverse problems: A Bayesian perspective. Acta Numerica, 19:451-559, May 2010. ISSN 0962-4929, 1474-0508. doi: 10.1017/S0962492910000061. URL https://www.cambridge.org/core/product/identifier/S0962492910000061/type/journal ARTICLE. + +Nathan Taylor, Bruno Mourato, and Denham Parker. Preliminary Closed-Loop Simulation of Management Procedure Performance for Southern Swordfish. *Collect. Vol. Sci. Pap. ICCAT*, 79(2):705-714, 2022. URL https://www.iccat.int/Documents/CVSP/CV079_2022/n_2/CV079020705.pdf. +Andrew R. Teel. Lyapunov Conditions Certifying Stability and Recurrence for a Class of Stochastic Hybrid Systems. Annual Reviews in Control, 37(1):1-24, April 2013. ISSN 13675788. doi: 10.1016/j.arcontrol.2013.02.001. URL https://linkinghub.elsevier.com/retrieve/pii/S1367578813000023. +Andrew R. Teel and Joao P. Hespanha. Stochastic Hybrid Systems: A Modeling and Stability Theory Tutorial. In 2015 54th IEEE Conference on Decision and Control (CDC), pp. 3116-3136, Osaka, December 2015. IEEE. ISBN 978-1-4799-7886-1. doi: 10.1109/CDC.2015.7402688. URL http://ieeexplore.ieee.org/document/7402688/. +Andrew R. Teel, Anantharaman Subbaraman, and Antonino Sferlazza. Stability Analysis for Stochastic Hybrid Systems: A Survey. Automatica, 50(10):2435-2456, October 2014. ISSN 00051098. doi: 10.1016/j.automatica.2014.08.006. URL https://linkinghub.elsevier.com/retrieve/pii/S0005109814003070. +E. Walter and Luc Pronzato. Identification of Parametric Models from Experimental Data. Communications and control engineering. Springer; Masson, Berlin; New York: Paris, 1997. ISBN 978-3-540-76119-8. +Zhe Wang, Yuan Liang, David C. Zhu, and Tongtong Li. The Relationship of Discrete DCM and Directed Information in fMRI-Based Causality Analysis. IEEE Transactions on Molecular, Biological and Multi-Scale Communications, 4(1):3-13, March 2018. ISSN 2372-2061, 2332-7804. doi: 10.1109/TMBMC.2018.2887210. URL https://ieeexplore.ieee.org/document/8579229/. +Amanda Warlick, Erin Steiner, and Marie Guldin. History of the West Coast groundfish trawl fishery: Tracking socioeconomic characteristics across different management policies in a multispecies fishery. *Marine Policy*, 93:9–21, July 2018. ISSN 0308597X. doi: 10.1016/j.marpol.2018.03.014. URL https://linkinghub.elsevier.com/retrieve/pii/S0308597X17307911. +Sam A Witty. Bayesian Structural Causal Inference with Probabilistic Programming. PhD thesis, University of Massachusetts Amherst, 2023. URL https://scholarworks.umass.edu/dissertations_2/2922. +Lei Zan, Charles K. Assaad, Emilie Devijver, Eric Gaussier, and Ali Aït-Bachir. On the Fly Detection of Root Causes from Observed Data with Application to IT Systems. In Proceedings of the 33rd ACM International Conference on Information and Knowledge Management, pp. 5062-5069, Boise ID USA, October 2024. ACM. ISBN 979-8-4007-0436-9. doi: 10.1145/3627673.3680010. URL https://dl.acm.org/doi/10.1145/3627673.3680010. +S Zhou and Adm Smith. Effect of Fishing Intensity and Selectivity on Trophic Structure and Fishery Production. Marine Ecology Progress Series, 585:185-198, December 2017. ISSN 0171-8630, 1616-1599. doi: 10.3354/meps12402. URL http://www.int-res.com/abstracts/meps/v585/p185-198/. +A. Zhuk, Yu. Zeigarnik, E. Buzoverov, and A. Sheindlin. Managing Peak Loads in Energy Grids: Comparative Economic Analysis. Energy Policy, 88:39-44, 2016. ISSN 0301-4215. doi: https://doi.org/10.1016/j.enpol.2015.10.006. URL https://www.sciencedirect.com/science/article/pii/S0301421515301348. +Bernt Øksendal. Stochastic Differential Equations. Universitext. Springer Berlin Heidelberg, Berlin, Heidelberg, 2003. ISBN 978-3-540-04758-2 978-3-642-14394-6. doi: 10.1007/978-3-642-14394-6. URL http://link.springer.com/10.1007/978-3-642-14394-6. + +# A Supplementary Definitions and Standard Assumptions + +# A.1 Differential Inclusions and Set-Valued Maps + +We follow Goebel et al. (2012) in generalizing to hybrid systems with inclusion constraints. A differential inclusion $F: S \to \mathbb{R}^n$ , for example, specifies the constraint that the time derivative $\dot{\pmb{x}}$ of the state must be included in the set $F(\pmb{x}) \subseteq \mathbb{R}^n$ . Note that the equality constraint $\dot{\pmb{x}} = f(\pmb{x})$ for some $f: S \to \mathbb{R}^n$ is a special case of the broader notion of differential inclusion. To clarify, the stacked double arrows in, for example, $S \to \mathbb{R}^n$ indicate a set-valued mapping from $S$ to a subset of $\mathbb{R}^n$ . Goebel et al. (2012) define the domain of a set-valued mapping $V: \mathcal{X} \to \mathcal{Y}$ as $\operatorname{dom} V = \{ \pmb{x} \in \mathcal{X}: V(\pmb{x}) \neq \emptyset \}$ . The graph of $V$ is then + +$$ +\mathcal {G} (V) = \left\{\left(\boldsymbol {x}, \boldsymbol {y}\right) \in \mathcal {X} \times \mathcal {Y}: \boldsymbol {x} \in \operatorname {d o m} V, \boldsymbol {y} \in V (\boldsymbol {x}) \right\}. \tag {7} +$$ + +# A.2 Ordered Set-Valued Maps + +Ordered set-valued maps are special cases of set-valued maps, which we use in this paper to keep track of interventions. + +Definition 6 (Ordered Set-Valued Map). Let $G = (G_{1},\ldots ,G_{K})$ be a finite sequence of set-valued maps. We call $G$ an ordered set-valued map, which means it is equipped with the following operation: + +$$ +G ^ {\dagger} = x \mapsto \bigcup_ {k = 1} ^ {K} G _ {k} (x); \quad \operatorname {l a s t} (G) = G _ {K}. \tag {8} +$$ + +Therefore, $\operatorname{dom} G^{\dagger} = \bigcup_{k=1}^{K} \operatorname{dom} G_k$ . Given two sequences $G = (G_1, \ldots, G_K)$ and $H = (H_1, \ldots, H_L)$ , we denote $G \sqcup H = (G_1, \ldots, G_K, H_1, \ldots, H_L)$ . By slight abuse of notation, we sometimes identify a map $G$ with the corresponding one-element sequence $(G)$ , and also use $G$ in place of $G^{\dagger}$ when the context requires a "vanilla" set-valued map. + +# A.3 Solution Concept + +The following definitions and propositions are given almost exactly as stated by Goebel et al. (2012), except that we adapt them slightly for explicitly parameterized hybrid systems (definition 1). + +The nature of hybrid systems implies that their solutions should be functions of both continuous time $t \in \mathbb{R}_{\geqslant 0}$ and discrete time $j \in \mathbb{N}$ . Let $t_j$ denote the time of the $j$ -th discrete event, with $t_j \leqslant t_{j+1}$ for all $j \in \mathbb{N}$ and $t_0 = 0$ . Following (Goebel et al., 2012, Sects. 2.2-2.3), we define, for each possible parameterization $\pmb{\theta} \in \Theta$ and initial condition $\pmb{\xi} \in S$ , a "solution" to $\mathcal{H}(\pmb{\theta})$ to be a "hybrid arc", which is formally a set-valued map $\phi(\cdot; \pmb{\xi}, \pmb{\theta}): \mathbb{R}_{\geqslant 0} \times \mathbb{N} \rightarrow \mathbb{R}^n$ . We can formalize this time-event space (of which $\operatorname{dom} \phi$ is an example) as follows: + +Definition 7 (Hybrid Time Domain from Goebel et al. (2012) (Def. 2.3)). $E \subset \mathbb{R}_{\geqslant 0} \times \mathbb{N}$ is a compact hybrid time domain if it is a finite union of sequence of closed intervals $E = \bigcup_{j=0}^{J-1}([t_j, t_{j+1}] \times \{j\})$ , where $0 = t_0 \leqslant t_1 \leqslant \dots \leqslant t_J$ , and $E$ is a hybrid time domain if for each $(T, J) \in E$ , the set $E \cap ([0, T] \times \{0, 1, \ldots, J\})$ is a compact hybrid time domain. + +Generally, $\operatorname{dom} \phi$ is unknown until after a particular solution $\phi$ is found, as it depends on the exact sequence of state-dependent jump times; therefore, it is natural to consider $\phi$ as a-priori set-valued. + +Definition 8 (Solution Concept adapted from Goebel et al. (2012) (Def 2.6)). Consider parameterized hybrid system $\mathcal{P} = (\mathcal{H},\mathcal{S},\Theta)$ , with $\mathcal{H} = (C,F,D,G)$ . For $\pmb{\theta}\in \Theta$ , any solution $\phi (\cdot ;\pmb {\xi},\pmb {\theta})$ to $\mathcal{H}(\pmb {\theta})$ must satisfy $\phi (0,0;\pmb {\xi},\pmb {\theta}) = \pmb {\xi}\in \overline{C(\pmb{\theta})}\cup D(\pmb {\theta})$ and the constraints implied by $\mathcal{H}(\pmb {\theta})$ , i.e.: + +1. for all $j \in \mathbb{N}$ such that $I^j := \{t : (t, j) \in \operatorname{dom} \phi\}$ has nonempty interior, we have $\phi(t, j) \in C(\pmb{\theta}), \forall t \in \operatorname{int} I^j$ , and $\dot{\phi}(t, j) \in F_{\pmb{\theta}}(\phi(t, j))$ , for almost all $t \in I^j$ ; [continuous flow regime] +2. for all $(t,j)\in \mathrm{dom}\phi$ s.t. $(t,j + 1)\in \mathrm{dom}\phi$ , we have $\phi (t,j;\pmb {\xi},\pmb {\theta})\in D(\pmb {\theta})$ and $\phi (t,j+$ $1;\pmb {\xi},\pmb {\theta})\in G_{\pmb{\theta}}(\phi (t,j;\pmb {\xi},\pmb {\theta}))$ [discrete jump regime]. + +It is convenient to work with solutions that cannot be extended, as formalized by the following concept. + +Definition 9 (Maximal Solutions adapted from Goebel et al. (2012) (Def 2.7)). A solution $\phi (\cdot ;\pmb {\xi},\pmb {\theta})$ to $\mathcal{H}(\pmb {\theta})$ (as in definition 8) is maximal if there does not exist another solution $\phi^{\prime}(\cdot ;\pmb {\xi},\pmb {\theta})$ to $\mathcal{H}(\pmb {\theta})$ such that $\mathrm{dom}\phi (\cdot ;\pmb {\xi},\pmb {\theta})$ is a proper subset of $\mathrm{dom}\phi^{\prime}(\cdot ;\pmb {\xi},\pmb {\theta})$ and $\phi (t,j;\pmb {\xi},\pmb {\theta}) = \phi '(t,j;\pmb {\xi},\pmb {\theta})$ for all $(t,j)\in \mathrm{dom}\phi (\cdot ;\pmb {\xi},\pmb {\theta})$ + +Unless specified otherwise, we always consider maximal solutions in this paper. With the solution concept established, we can now state conditions for the existence and uniqueness of solutions. We again borrow from Goebel et al. (2012), and adapt accordingly to support parameterized systems (definition 1). + +Proposition 1 (Basic Existence adapted from Goebel et al. (2012) (Proposition 2.10)). Consider parameterized hybrid system $\mathcal{P} = (\mathcal{H},\mathcal{S},\Theta) = (C,F,D,G,\mathcal{S},\Theta)$ , and a standard hybrid system $\mathcal{H}(\pmb{\theta}) = (C(\pmb{\theta}),F_{\pmb{\theta}},D(\pmb{\theta}),G_{\pmb{\theta}})$ for some $\pmb{\theta} \in \Theta$ . Let $\pmb{\xi} \in \overline{C(\pmb{\theta})} \cup D(\pmb{\theta})$ . If $\pmb{\xi} \in D(\pmb{\theta})$ or + +(VC) there exists $\epsilon > 0$ and an absolutely continuous function $z: [0, \epsilon] \to \mathbb{R}^n$ such that $z(0) = \xi$ , $\dot{z}(t) \in F_{\pmb{\theta}}(z(t))$ for almost all $t \in [0, \epsilon]$ and $z(t) \in C(\pmb{\theta})$ for all $t \in (0, \epsilon]$ , + +then there exists a non-trivial solution $\phi (\cdot ;\pmb {\xi},\pmb {\theta})$ to $\mathcal{H}$ with $\phi (0,0;\pmb {\xi},\pmb {\theta}) = \pmb{\xi}$ .17 If (VC) holds for every $\pmb {\xi}\in \overline{C(\pmb{\theta})}\cup D(\pmb {\theta})$ , then there exists a nontrivial solution to $\mathcal{H}(\pmb {\theta})$ from every point of $\overline{C(\pmb{\theta})}\cup D(\pmb{\theta})$ . If the foregoing further holds for $\mathcal{H}(\pmb {\theta})$ at every $\pmb {\theta}\in \Theta$ , we say $\mathcal{P}$ fulfills the conditions for basic existence. + +Proposition 2 (Basic Uniqueness adapted from Goebel et al. (2012) (Proposition 2.11)). Consider parameterized hybrid system $\mathcal{P} = (\mathcal{H},\mathcal{S},\Theta) = (C,F,D,G,\mathcal{S},\Theta)$ , and a standard hybrid system $\mathcal{H}(\pmb{\theta}) = (C(\pmb{\theta}),F_{\pmb{\theta}},D(\pmb{\theta}),G_{\pmb{\theta}})$ for some $\pmb{\theta} \in \Theta$ . For every $\pmb{\xi} \in \overline{C(\pmb{\theta})} \cup D(\pmb{\theta})$ there exists a unique maximal solution $\phi(\cdot;\pmb{\xi},\pmb{\theta})$ with $\phi(0,0;\pmb{\xi},\pmb{\theta}) = \pmb{\xi}$ provided that the following conditions hold. + +(a) For every $\pmb{\xi} \in \overline{C(\pmb{\theta})} \backslash D(\pmb{\theta})$ , $T > 0$ , if two absolutely continuous $z_1, z_2 : [0, T] \to S$ are such that $\dot{z}_i(t) \in F_{\pmb{\theta}}(z_i(t))$ for almost all $t \in [0, T]$ , $z_i(t) \in C(\pmb{\theta})$ for all $t \in (0, T]$ , and $z_i(0) = \pmb{\xi}$ , $i = 1, 2$ , then $z_1(t) = z_2(t)$ for all $t \in [0, T]$ ; +(b) for every $\pmb {\xi}\in \overline{C(\pmb{\theta})}\cap D(\pmb {\theta})$ , there does not exist $\epsilon >0$ and an absolutely continuous function $z:[0,\epsilon ]\to S$ such that $z(0) = \pmb {\xi},\dot{z} (t)\in F_{\pmb{\theta}}(z(t))$ for almost all $t\in [0,\epsilon ]$ and $z(t)\in C(\pmb {\theta})$ for all $t\in (0,\epsilon ]$ +(c) for every $\pmb {\xi}\in D(\pmb {\theta})$ $G_{\pmb{\theta}}(\pmb {\xi})$ consists of one point. + +If the foregoing further holds at every $\theta \in \Theta$ , we say that $\mathcal{P}$ fulfills the conditions for basic uniqueness. + +# A.4 Finite-Time Measurability of Solution in Initial Conditions and Parameters + +Measurability is key to coherently defining causal estimands as (conditional) expectations. In particular, we use the measurability of a time-parameterized "solution map" jointly in the initial state and parameters. By "solution map", we refer either to functions $\phi$ or $\varphi$ that, when provided some $\xi \in S$ and $\theta \in \Theta$ , yield hybrid arc $(t,j) \mapsto \phi(t,j;\xi,\theta)$ and time-parameterized function $t \mapsto \varphi(t;\xi,\theta)$ respectively. + +As stated following definition 1, in this paper, we focus strictly on finite time horizons. Definition 10, below, makes this finite-time limitation precise, and then employs that definition to formalize the time-parameterized solution map and its measurability. + +Definition 10 ( $t^{+}$ Uniquely Evaluable). Consider parameterized hybrid system $\mathcal{P} = (C,F,D,G,\mathcal{S},\Theta)$ that fulfills conditions for basic existence and uniqueness (propositions 1 and 2). Define $t^{+} = \min_{\pmb {\xi},\pmb{\theta}}\sup_{t}\operatorname {dom}\phi (\cdot ;\pmb {\xi},\pmb {\theta})$ , meaning that for every $\pmb {\xi}\in S,\pmb {\theta}\in \Theta$ yielding unique solution $t,j\mapsto \phi (t,j;\pmb {\xi},\pmb {\theta}),\forall t\in [0,t^{+})$ there exists $j\in \mathbb{N}$ such that $(t,j)\in \mathrm{dom}\phi (\cdot ,\cdot ;\pmb {\xi},\pmb {\theta})$ . We then say that $\mathcal{P}$ is $t^{+}$ uniquely evaluable.18 + +Definition 11 (Time Parameterized Solution Map). Consider $t^+$ uniquely evaluable parameterized hybrid system $\mathcal{P} = (C, F, D, G, S, \Theta)$ and its solution map $\phi$ . Define for all $t \in [0, t^+)$ , $\pmb{\xi} \in S$ , $\theta \in \Theta$ its time-parameterized solution + +$$ +\varphi (t; \boldsymbol {\xi}, \boldsymbol {\theta}) = \phi (t, j _ {t} ^ {+} (\boldsymbol {\xi}, \boldsymbol {\theta}); \boldsymbol {\xi}, \boldsymbol {\theta}) \tag {9} +$$ + +where $j_{t}^{+}(\pmb {\xi},\pmb {\theta})$ is the index of the last discrete jump at time $t$ + +$$ +j _ {t} ^ {+} (\boldsymbol {\xi}, \boldsymbol {\theta}) = \max \left\{j: (t, j) \in \operatorname {d o m} \phi (\cdot ; \boldsymbol {\xi}, \boldsymbol {\theta}) \right\} \tag {10} +$$ + +Definition 12 ( $t^+$ Measurable). Consider $t^+$ uniquely evaluable parameterized hybrid system $\mathcal{P}$ and its time-parameterized solution map $\varphi$ . If, for every fixed $t \in [0, t^+)$ , $\xi, \theta \mapsto \varphi(t; \xi, \theta)$ is a Borel-measurable function, we say that $\mathcal{P}$ has a $t^+$ measurable time-parameterized solution map $\varphi$ . + +# A.5 Flow Preferring Subtraction and Lowering + +Definition 13 (Flow-Preferring Subtraction). Consider parameterized hybrid system $\mathcal{P} = (C, F, D, G, S, \Theta)$ that meets the hybrid basic conditions (assumption 4). We borrow the following viability condition from proposition 1 on a point $\xi \in S$ , for some $\theta \in \Theta$ . + +(VC) there exists $\epsilon > 0$ and an absolutely continuous function $z: [0, \epsilon] \to \mathbb{R}^n$ such that $z(0) = \xi$ , $\dot{z}(t) \in F(z(t), \pmb{\theta})$ for almost all $t \in [0, \epsilon]$ and $z(t) \in C(\pmb{\theta})$ for all $t \in (0, \epsilon]$ , + +We can then transform $D$ to be flow preferring by writing + +$$ +\operatorname {p r e f e r f l o w} (D, C, F) = \boldsymbol {\theta} \mapsto D (\boldsymbol {\theta}) \backslash \{\boldsymbol {\xi} \in S: (\mathrm {V C}) \text {h o l d s f o r} \boldsymbol {\theta} \text {f r o m} \boldsymbol {\xi} \} \tag {11} +$$ + +Recall the definition of ordered set-values maps (definition 6) affording the $\operatorname{last}(G)$ operation on $G$ , the jump map. + +Definition 14 (Lowering). Consider parameterized hybrid system $\mathcal{P} = (C,F,D,G,\mathcal{S},\Theta)$ that meets the hybrid basic conditions (assumption 4). We write that + +$$ +D ^ {\prime} = \text {p r e f e r f l o w} (D, C, F) \tag {12} +$$ + +$$ +G ^ {\prime} = \operatorname {l a s t} (G), \tag {13} +$$ + +$$ +\operatorname {l o w e r} (\mathcal {P}) = (C, F, D ^ {\prime}, G ^ {\prime}, \mathcal {S}, \Theta) \tag {14} +$$ + +# A.6 Collected Assumptions on the Hybrid System + +Assumption 3 (Unique, Complete, and Borel Solutionexists for Differential Inclusion for all $\mathcal{S}$ Consider parameterized hybrid system $\mathcal{P} = (C,F,D,G,\mathcal{S},\Theta)$ . Assume that + +(F1) for every $\pmb {\xi}\in \mathcal{S},\pmb {\theta}\in \Theta ,T > 0$ , if two absolutely continuous $z_{1},z_{2}:[0,T]\to S$ are such that $\dot{z}_i(t)\in F_\pmb{\theta}(z_i(t))$ for almost all $t\in [0,T],z_i(t)\in S$ for all $t\in (0,T]$ , and $z_{i}(0) = \pmb{\xi}$ $i = 1,2$ , then $z_{1}(t) = z_{2}(t)$ for all $t\in [0,T]$ +(F2) for all $\pmb {\xi}\in S$ and $\pmb {\theta}\in \Theta$ , such a $z_{1}$ exists for every $T\in (0,\infty)$ +(F3) with $z(t;\pmb {\xi},\pmb {\theta}) = z_1(t)$ for all $\pmb {\xi}\in S$ $\pmb {\theta}\in \Theta$ , and $t\in [0,\infty)$ $\pmb {\xi},\pmb{\theta}\mapsto z(t;\pmb {\xi},\pmb {\theta})$ is a Borel-measurable function for every $t\in [0,\infty)$ + +Importantly, note that assumption 3 only relates to the differential inclusion, and does not preclude $\mathcal{P}$ from jumping, or from pathologies associated with jumps. Additionally, observe that $z(t;\pmb {\xi},\pmb {\theta})\equiv$ $\phi (t,0;\pmb {\xi},\pmb {\theta})$ - that is, statements on $z$ trivially apply to the solution mapping up to and including the time of the first jump. + +Assumption 4 (Hybrid Basic Conditions adapted from Goebel et al. (2012) (Assump. 6.5)). Consider parameterized hybrid system $\mathcal{P} = (C,F,D,G,\mathcal{S},\Theta)$ , and assume for all $\theta \in \Theta$ that the following hold. + +(A1) $C(\pmb {\theta})$ and $D(\pmb {\theta})$ are closed subsets of $\mathcal{S}$ +(A2) $F_{\pmb{\theta}}:S\Rightarrow \mathbb{R}^{n}$ is outer semi-continuous and locally bounded relative to $C(\pmb {\theta})$ $C(\pmb {\theta})\subset$ dom $F_{\pmb{\theta}}$ , and $F(\pmb {x},\pmb {\theta})$ is convex for every $\pmb {x}\in C(\pmb {\theta})$ +(A3) $G_{\pmb{\theta}}:S\Rightarrow S$ is outer semi-continuous and locally bounded relative to $D(\pmb {\theta})$ , and $D(\pmb {\theta})\subset$ dom $G_{\pmb{\theta}}$ + +In particular, (A1) implies that $D(\pmb{\theta})$ and $C(\pmb{\theta})$ must overlap on any shared boundary — solutions that start at or graze this boundary can, non-uniquely, either jump or flow. Additionally, the outer semi-continuity of $G_{\theta}$ (A3) requires that, at the boundaries of the pieces in a piecewise $G_{\theta}$ , $G_{\theta}$ must return values from multiple pieces. Solutions hitting those boundaries can jump to multiple states. + +Assumption 5 (Collected Assumptions on the Original System). The parameterized hybrid system $\mathcal{P}$ can be constructed as $\mathcal{P} = \text{lower}(\mathcal{P}_{\uparrow})$ , where $\mathcal{P}_{\uparrow} = (C, F, D, G, S, \Theta)$ , such that: + +(P1) $\mathcal{P}_{\uparrow}$ satisfies assumption 4; +(P2) $\mathcal{P}_{\uparrow}$ fulfills the conditions for basic existence (proposition 1); +(P3) $\mathcal{P}_{\uparrow}$ has a unique solution to its differential inclusion $F$ from everywhere in $\mathcal{S}$ and $\Theta$ (assumption 3); +(P4) $C(\pmb {\theta})$ is outer semi-continuous at every $\pmb {\theta}\in \Theta$ +(P5) the graph $\mathcal{G}(D)$ of the jump set mapping $D$ is Borel; +(P6) $\operatorname{last}(G)$ is single-valued on its domain, with $\operatorname{last}(G)(\pmb{x}, \pmb{\theta}) = \{g(\pmb{x}, \pmb{\theta})\}$ , and $g$ Borel-measurable for all $\pmb{x}, \pmb{\theta} \in \operatorname{dom} \operatorname{last}(G)$ . + +# A.7 Well-Behaved Jump Set + +Definition 15 (Well- behaved Set). Consider $\Theta \subseteq \mathbb{R}^m$ , $S \subseteq \mathbb{R}^n$ , arbitrary set-valued mapping $A: \Theta \rightrightarrows S$ , and differential inclusion $F: S \times \Theta \rightrightarrows \mathbb{R}^n$ . Suppose that for every $\theta \in \Theta$ and $\xi \in S$ where + +$(\mathrm{VC}_{\mathcal{S}})$ there exists $\epsilon > 0$ and an absolutely continuous function $z: [0, \epsilon] \to \mathbb{R}^n$ such that $z(0) = \xi$ , $\dot{z}(t) \in F(z(t), \pmb{\theta})$ for almost all $t \in [0, \epsilon]$ and $z(t) \in \mathcal{S}$ for all $t \in (0, \epsilon]$ , + +there also exists some $\epsilon' \in (0, \epsilon]$ such that + +$$ +z \left(\left(0, \epsilon^ {\prime}\right)\right) \subseteq \operatorname {i n t} A (\theta) \text {o r} z \left(\left(0, \epsilon^ {\prime}\right)\right) \subseteq S \backslash \operatorname {i n t} A (\theta) \tag {15} +$$ + +In such a case, we say that $A$ is well-behaved relative to $\mathcal{S}$ for $\Theta$ and $F$ . For a parameterized hybrid system $\mathcal{P} = (C,F,D,G,\mathcal{S},\Theta)$ , we sometimes say that $A$ is well-behaved relative to $\mathcal{P}$ . + +Assumption 6 (Well- behavior Interventional Subset). Consider set-valued mapping $\tilde{D}:\Theta \to S$ and parameterized hybrid system $\mathcal{P}$ . Assume $\tilde{D}$ is well-behaved relative to $\mathcal{P}$ (definition 15). + +Observation 1 (Flow into Subdivisions of $C$ by $\tilde{D}$ ). Consider set-valued mapping $\tilde{D}: \Theta \rightrightarrows S$ that meets assumption 6 relative to some parameterized hybrid system $\mathcal{P} = (C, F, D, G, S, \Theta)$ . It is then the case that, for every $\theta \in \Theta$ and $\xi \in S$ where + +(VC) there exists $\epsilon > 0$ and an absolutely continuous function $z: [0, \epsilon] \to \mathbb{R}^n$ such that $z(0) = \xi$ , $\dot{z}(t) \in F(z(t), \pmb{\theta})$ for almost all $t \in [0, \epsilon]$ and $z(t) \in C(\pmb{\theta})$ for all $t \in (0, \epsilon]$ , + +there also exists some $\epsilon' \in (0, \epsilon]$ such that + +$$ +z \left(\left(0, \epsilon^ {\prime} \right]\right) \subseteq \operatorname {i n t} \tilde {D} (\boldsymbol {\theta}) \text {o r} z \left(\left(0, \epsilon^ {\prime} \right]\right) \subseteq C (\boldsymbol {\theta}) \backslash \operatorname {i n t} \tilde {D} (\boldsymbol {\theta}). \tag {16} +$$ + +Proof. Suppose the proposed antecedent and note that a trajectory $z((0,\epsilon]) \subseteq C(\pmb{\theta}) \subseteq S$ fulfills the antecedent of the assumed well-behaved property of $\tilde{D}$ relative to $S$ (assumption 6). This implies that there exists $\epsilon' \in (0,\epsilon]$ such that either $z((0,\epsilon']) \subseteq \operatorname{int}\tilde{D}(\pmb{\theta})$ or $z((0,\epsilon']) \subseteq S\backslash \operatorname{int}\tilde{D}(\pmb{\theta})$ . $z((0,\epsilon')) \subseteq \operatorname{int}\tilde{D}(\pmb{\theta})$ is precisely the first case of our desired consequent. Thus, we need only show that $z((0,\epsilon']) \subseteq S\backslash \operatorname{int}\tilde{D}(\pmb{\theta})$ and $z((0,\epsilon]) \subseteq C(\pmb{\theta})$ imply $z((0,\epsilon']) \subseteq C(\pmb{\theta})\backslash \operatorname{int}\tilde{D}(\pmb{\theta})$ . We have $z((0,\epsilon')) \subseteq z((0,\epsilon]) \subseteq C(\pmb{\theta})$ , and can thus take the intersection to see this implies the second case of the desired consequent: + +$$ +z \left(\left(0, \epsilon^ {\prime} \right]\right) \subseteq C (\boldsymbol {\theta}) \cap \left(S \backslash \operatorname {i n t} \tilde {D} (\boldsymbol {\theta})\right) = C (\boldsymbol {\theta}) \backslash \operatorname {i n t} \tilde {D} (\boldsymbol {\theta}). +$$ + +Remark 1 (Universality of Assumption 6). If and only if assumption 6 holds relative to some parameterized hybrid system $\mathcal{P}$ , then it also holds relative to $\operatorname{instint}(\mathcal{P})$ , $\operatorname{instint}_{\uparrow}(\mathcal{P})$ , and $\operatorname{lower}(\mathcal{P})$ . + +Proof. Assumption 6 holding relative to $\mathcal{P} = (C,F,D,G,\mathcal{S},\Theta)$ pertains only to $F,S,\Theta$ , which are unaffected by instant, instant $_\uparrow$ , and lower. + +# B Space Augmentation + +It is often useful to parameterize interventions, and a fully expressive interventional semantics benefits from stateful jump maps/sets. Thus, it will be useful to establish a primitive transformation that simply augments the parameter and state spaces, without changing the component functions of the system. Subsequent transformations can then operate on this augmented system. Note that, in eq. (22), we write the transformed jump map in its expanded form as an ordered set-valued map (definition 6). + +Definition 16. (Space Augmentation) Consider $\tilde{S} \subseteq \mathbb{R}^{\tilde{n}}$ , and $\tilde{\Theta} \subseteq \mathbb{R}^{\tilde{m}}$ . For any parameterized hybrid system $\mathcal{P} = (C, F, D, G, S, \Theta)$ with $G = (G_1, \ldots, G_L)$ , let, for all $\boldsymbol{x} \in \mathcal{S}$ , $\tilde{\boldsymbol{x}} \in \tilde{S}$ , $\boldsymbol{\theta} \in \Theta$ , $\tilde{\boldsymbol{\theta}} \in \tilde{\Theta}$ , + +$$ +\boldsymbol {x} ^ {\prime} = [ \boldsymbol {x}, \tilde {\boldsymbol {x}} ], \quad \boldsymbol {\theta} ^ {\prime} = [ \boldsymbol {\theta}, \tilde {\boldsymbol {\theta}} ] \tag {17} +$$ + +$$ +C ^ {\prime} (\boldsymbol {\theta} ^ {\prime}) = C (\boldsymbol {\theta}) \times \tilde {S} \tag {18} +$$ + +$$ +F ^ {\prime} \left(\boldsymbol {x} ^ {\prime}, \boldsymbol {\theta} ^ {\prime}\right) = F (\boldsymbol {x}, \boldsymbol {\theta}) \times \{\mathbf {0} \} \tag {19} +$$ + +$$ +D ^ {\prime} (\boldsymbol {\theta} ^ {\prime}) = D (\boldsymbol {\theta}) \times \tilde {S} \tag {20} +$$ + +$$ +G _ {\ell} ^ {\prime} (\boldsymbol {x}, \boldsymbol {\theta}) = G _ {\ell} (\boldsymbol {x}, \boldsymbol {\theta}) \times \{\tilde {\boldsymbol {x}} \} \tag {21} +$$ + +$$ +G ^ {\prime} \left(\boldsymbol {x} ^ {\prime}, \boldsymbol {\theta} ^ {\prime}\right) = \left(G _ {1} ^ {\prime} (\boldsymbol {x}, \boldsymbol {\theta}), \dots , G _ {L} ^ {\prime} (\boldsymbol {x}, \boldsymbol {\theta})\right) \tag {22} +$$ + +$$ +\mathcal {S} ^ {\prime} = \mathcal {S} \times \tilde {\mathcal {S}}, \quad \Theta^ {\prime} = \Theta \times \tilde {\Theta}, \tag {23} +$$ + +then + +$$ +\operatorname {s p a u g} \left(\mathcal {P}, \tilde {\mathcal {S}}, \tilde {\Theta}\right) = \left(\left(C ^ {\prime}, F ^ {\prime}, D ^ {\prime}, G ^ {\prime}\right), \mathcal {S} ^ {\prime}, \Theta^ {\prime}\right). \tag {24} +$$ + +Observation 2 (Compositions of Space Augmentation Preserves Key Properties). Consider parameterized hybrid system $\mathcal{P}$ that meets assumption 5, and any finite sequence $(\tilde{S}_k,\tilde{\Theta}_k)$ of length $K$ such that $\tilde{S}_k\subseteq \mathbb{R}^{\tilde{n}_k}$ and $\tilde{\Theta}_k\subseteq \mathbb{R}^{\tilde{m}_k}$ . Let $spaug_{k} = spaug(\cdot ,\tilde{S}_{k},\tilde{\Theta}_{k})$ (definition 16) and + +$$ +\mathcal {P} ^ {\prime} = \left(s p a u g _ {1} \circ \dots \circ s p a u g _ {k} \circ \dots \circ s p a u g _ {K}\right) (\mathcal {P}). \tag {25} +$$ + +$\mathcal{P}'$ then meets assumption 5 and has a unique solution (propositions 1 and 2) with a $t^{+}$ measurable (definition 12) time-parameterized solution map $\varphi$ (definition 11). + +Proof. The proposition follows from induction, $K < \infty$ , and the fact that the space augmentation operation fulfills the same pattern described in fig. 2 for instant. That is, $\operatorname{spaug}$ commutes with lower, and it preserves (P1-6) (assumption 5) on an upstream system $\mathcal{P}_{\uparrow}^{\prime} = \operatorname{spaug}(\mathcal{P}_{\uparrow}, \ldots)$ . For commutativity, recall that lowering makes a flow-preferring subtraction from the jump set (definition 13), and chooses the last map in the ordered jump map. A flow-preferring subtraction on $D'(\pmb{\theta}) = D(\pmb{\theta}) \times \tilde{S}$ is dictated entirely by the behavior of $F(\pmb{\theta})$ on $C(\pmb{\theta})$ — i.e. $\operatorname{preferflow}(D', C', F') = \operatorname{preferflow}(D, C, F) \times \tilde{S}$ , which implies commutativity on $D'$ . Commutativity of the jump map is more straightforward, as, by construction, $\operatorname{last}(G') = G_L(\pmb{x}, \pmb{\theta}) \times \{\tilde{\pmb{x}}\} = \operatorname{last}(G)(\pmb{x}, \pmb{\theta}) \times \{\pmb{x}\}$ . Assumptions (P1-6) (collected in assumption 5) straightforwardly follow after noting, as we have used in the proof of observation 3, that since every topological space is both open and closed in itself (i.e., clopen), any product with such a space as a factor inherits the open (or closed) property from the other factor relative to the product topology. From here, along similar lines argued in the proof of observation 3, properties like graph closure, outer semi-continuity, Borelness, etc. are preserved obviously by construction. + +# C Static-Time and Do Interventions as Special Cases + +As a special case of instinct, we can also define an intervention that occurs at a fixed, predefined time.[19] + +Definition 17 (Static-Time Intervention). Consider a parameterized hybrid system $\mathcal{P}$ defined as the tuple $(C,F,D,G,\mathbb{R}_{\geqslant 0}^2\times S,\Theta)$ . Let time be tracked in the first dimension of the state space, and, in the second dimension, a variable recording whether the intervention has occurred, such that $(t,k,\pmb {x})\in \mathbb{R}_{\geqslant 0}^{2}\times \mathcal{S}$ . Assume $k = 0$ at $t = 0$ by convention and that $F$ is such that $dk / dt = 0$ always. Let $\tilde{D} (\pmb {\theta}) = [\lambda ,\lambda +\epsilon ]\times [0,.1]\times \mathcal{S}$ for all $\pmb {\theta}\in \Theta$ , a fixed $\lambda \geqslant 0$ , and any $\epsilon >0$ . For some $\tilde{G}:S\times \Theta \to S$ and all $(t,k,\pmb {x},\pmb {\theta})\in \mathbb{R}_{\geqslant 0}^{2}\times S\times \Theta$ . We then define + +$$ +\hat {G} \left(\left(t, k, \boldsymbol {x}\right), \boldsymbol {\theta}\right) = \left\{t, k + 1 \right\} \times \tilde {G} (\boldsymbol {x}, \boldsymbol {\theta}) \tag {26} +$$ + +$$ +\operatorname {s t a t i n t} \left(\mathcal {P}, \lambda , \tilde {G}\right) = \operatorname {i n s t i n t} \left(\mathcal {P}, \tilde {D}, \hat {G}\right) \tag {27} +$$ + +The definition above, it should be noted, is a special case of a more general "repeated" static-time intervention rstatint (definition 19), which is shown to satisfy the same existence, uniqueness, and measurability theory that we establish below for instant. + +Driving one level more granular, we arrive at a transformation representing something akin to the canonical "do" intervention — again as a special case of instant. This notion has been defined for dynamical systems both via a time-splitting mechanism (Boeken & Mooij, 2024) and by casting a continuous time system as its infinitely precise Euler approximation interpreted as an SCM (Hansen & Sokol, 2014). + +Definition 18 (Do-Intervention). Building directly off definition 17, if $\tilde{G}(\boldsymbol{x},\boldsymbol{\theta}) = \{\boldsymbol{v}\}$ for some fixed $\boldsymbol{v} \in S$ and all $\boldsymbol{x},\boldsymbol{\theta} \in S \times \Theta$ , then for some fixed $\lambda \geqslant 0$ we write + +$$ +\operatorname {d o} (\mathcal {P}, \boldsymbol {x} (\lambda) = \boldsymbol {v}) = \operatorname {s t a t i n t} (\mathcal {P}, \lambda , \tilde {G}) \tag {28} +$$ + +Alternatively, one might wish to fix an index $i \in \{1, \ldots, n\}$ and a value $v \in \mathbb{R}^1$ . With $\tilde{G}(\boldsymbol{x}, \boldsymbol{\theta}) = [x^{(1:i-1)}, v, x^{(i+1:n)}]\forall \boldsymbol{x} \in S$ and $\forall \boldsymbol{\theta} \in \Theta$ , we write instead do $\left(\mathcal{P}, x_t^{(i)} = v\right)$ . + +These interventional classes form a sort of hierarchy. The jump map of a static-time intervention can be considered the "pre-treatment" model for a do intervention, and the trigger mechanism encoded in the jump set can be considered a pre-treatment model for when a static intervention occurs. At the highest level, a state-dependent intervention — especially those that can be triggered many times — can be thought of as a soft intervention on system dynamics. By couching these interventions directly in the language of established hybrid systems theory, we can more easily borrow theoretical results from that vast body of literature. + +# C.1 Repeated Static-Time Intervention + +Definition 19 (Repeated Static-Time Intervention). Consider a parameterized hybrid system $\mathcal{P}$ defined as the tuple $(C,F,D,G,\mathbb{R}_{\geqslant 0}^2\times S,\Theta)$ . Without loss of generality with respect to positioning in the state vector, let time be tracked in the first dimension of the state space, and, in the second dimension, a variable recording whether a specified static intervention has recently occurred, such that $(t,k,\pmb {x})\in \mathbb{R}_{\geqslant 0}^{2}\times \mathcal{S}$ . Assume $k = 0$ at $t = 0$ by convention and that $F$ is such that $dk / dt = 0$ always. Also, assume that, for some countable set of unique intervention times $\Lambda \subset \mathbb{R}_{\geqslant 0}^{1}$ , there exists an $\epsilon$ such that $0 < \epsilon < \inf \left\{\left|\lambda_1 - \lambda_2\right|:\lambda_1,\lambda_2\in \Lambda^2,\lambda_1\neq \lambda_2\right\} \cup \{.1\}$ . For all $(t,k,\pmb {x},\pmb {\theta})\in \mathbb{R}_{\geqslant 0}^{2}\times S\times \Theta$ + +and some $\tilde{G}:\mathcal{S}\times \Theta \rightrightarrows \mathcal{S},$ let + +$$ +\tilde {D} _ {1} (\boldsymbol {\theta}) = \left(\bigcup_ {\lambda \in \Lambda} [ \lambda , \lambda + \epsilon / 2 ]\right) \times [ 0,. 1 ] \times \mathcal {S} \tag {29} +$$ + +$$ +\tilde {D} _ {2} (\boldsymbol {\theta}) = \left(\bigcup_ {\lambda \in \Lambda} [ \lambda + \epsilon / 2, \lambda + \epsilon ]\right) \times [ 1, 1. 1 ] \times \mathcal {S} \tag {30} +$$ + +$$ +\tilde {G} _ {1} \left(\left(t, k, \boldsymbol {x}\right), \boldsymbol {\theta}\right) = \left\{\left(t, k + 1\right) \right\} \times \tilde {G} (\boldsymbol {x}, \boldsymbol {\theta}) \tag {31} +$$ + +$$ +\tilde {G} _ {2} \left(\left(t, k, \boldsymbol {x}\right), \boldsymbol {\theta}\right) = \left\{\left(t, k - 1\right) \right\} \times \tilde {G} (\boldsymbol {x}, \boldsymbol {\theta}) \tag {32} +$$ + +With $\mathrm{instint}_i(\cdot) = \mathrm{instint}(\cdot, \tilde{D}_i, \tilde{G}_i)$ for $i \in \{1, 2\}$ , we can define + +$$ +\mathcal {P} ^ {\prime} = \operatorname {r s t a t i n t} (\mathcal {P}, \Lambda , \tilde {G}) = \left(\operatorname {i n s t i n t} _ {2} \circ \operatorname {i n s t i n t} _ {1}\right) (\mathcal {P}) \tag {33} +$$ + +Observation 3 (rstatint Preserves Collected Assumptions). Continuing from definition 19, if $\tilde{G}$ meets assumption 2, then $\mathcal{P}'$ , $\tilde{D}_i$ , and $\tilde{G}_i$ meet assumption 2, and theorem 1 would thus apply to rstatint. + +Proof. Assumption 2 comprises sub-conditions (I1) and (I2). (I1) first needs that $\tilde{D}_i(\theta)$ is closed $\forall \theta \in \Theta$ , which follows here from $\tilde{D}_i(\theta)$ being a product of closed sets with the topological space $S$ . Since every topological space is both open and closed in itself (i.e., clopen), any product with such a space as a factor inherits the open (or closed) property from the other factor relative to the product topology. Note that the intervals in the unions over intervals constructed from $\lambda \in \Lambda$ are guaranteed to be disjoint and uniformly separated by selecting $\epsilon$ to be positive and smaller than the closest two intervention times, which means the corresponding countable union must be closed. Similarly, by uniform separation, we have the (I1)-required well-behavedness (definition 15) of $\tilde{D}_i(\theta)$ relative to $\mathcal{P}$ . (I1) also requires that the graph $\mathcal{G}(\text{int } \tilde{D})$ is open. Note that these jump sets are constant in $\Theta$ , and therefore their interior graph is the product $\Theta \times A$ , where $A \subset \mathbb{R}^n$ is open. $\Theta$ is a topological space, so the product with $A$ inherits the openness of $A$ — thus $\mathcal{G}(\text{int } D)$ is open. By a similar argument, $\tilde{D}_i$ is closed, which implies that $\mathcal{G}(\tilde{D}_i)$ is closed, thereby ensuring $\mathcal{G}(\tilde{D}_i)$ is Borel as required by (I1). (I2) asserts straightforward requirements on $\tilde{G}_i$ , none of which are affected by taking a cartesian product with the single-valued, continuous (and therefore both inner and outer semi-continuous) set valued mappings $t, k \mapsto \{(t, k \pm 1)\}$ . Theorem 1, then, applies here because rstatint is a composition of inst int operations with specifications that meet assumption 2. + +# D Proof of Theorem 1 + +The following proof refers to assumption 5, which is an expanded version of assumption 1 that is referenced by theorem 1 in the main text. + +Proof. By induction and $K < \infty$ , we have via lemma 2 that $\mathcal{P}'$ will meet assumption 5. Note that, by remark 1, if assumption 2 holds relative to $\mathcal{P}$ , it will hold relative to any intermediate system in the chain of transformations from $\mathcal{P}$ to $\mathcal{P}'$ . Then, by lemma 5, existence, uniqueness, and measurability follow from $\mathcal{P}'$ fulfilling assumption 5. + +# E Proof that Instantaneous Intervention Preserves Key Properties + +Lemma 2 (Instantaneous Intervention Preserves Key Properties). Consider a parameterized hybrid system $\mathcal{P}$ that meets assumption 5. Now, consider set-valued mappings $\tilde{D}:\Theta\Rightarrow S$ and $\tilde{G}:\mathcal{S}\times \Theta\rightarrow S$ that fulfill assumption 2 relative to $\mathcal{P}$ . The intervened system $\mathcal{P}'=\mathrm{instint}(\mathcal{P})$ (definition 4) will then also meet assumption 5, and therefore will have a unique and $t^{+}$ measurable solution for each $\pmb{\theta}\in \Theta,\pmb {\xi}\in \overline{C(\pmb{\theta})}\cup D(\pmb{\theta})$ according to lemma 5. + +Proof. The proof closely follows fig. 2. In assumption 5, we have that $\mathcal{P}$ can be constructed by "lowering" (definition 14) from a system $\mathcal{P}_{\uparrow}$ that fulfills certain conditions. In lemma 3, we prove that assumptions 2 and 5 imply that the system $\mathcal{P}'$ is equivalent to a system reached by performing a slightly modified intervention on $\mathcal{P}_{\uparrow}$ (definition 20), and then applying lower. The intervention on $\mathcal{P}_{\uparrow}$ is proven in lemma 4 to preserve properties on the higher system sufficient to say that the lowered system $\mathcal{P}'$ meets assumption 5. Intermediate statements and proofs for lemmas 3 and 4 and definition 20 can be found in appendix E.1. + +# E.1 Intermediate Results for Lemma 2 + +Lemma 2 argues that an intervened system $\mathcal{P}' = \text{inst}(\mathcal{P})$ can also be constructed by applying a slightly different interventional transformation to a different system $\mathcal{P}_{\uparrow}$ , and then "lowering" (definition 14). Additionally, if $\mathcal{P}$ meets assumption 5 by way of $\mathcal{P}_{\uparrow}$ , then $\mathcal{P}'$ must also meet assumption 5. This can be established by showing that the intervention on $\mathcal{P}_{\uparrow}$ preserves properties that allow it to be properly lowered. First, we will define this alternative intervention, then prove commutativity between intervention and lowering, and finally prove that the alternative intervention preserves the properties listed in assumption 5. In the following definition, we use the fact that $G$ is an ordered set-valued map (definition 6), which supports appending $\tilde{G}$ to the sequence of maps that compose $G$ . + +Definition 20 (Instantaneous Intervention for Higher System). Consider set-valued mappings $\tilde{D}:\Theta \rightarrow S$ and $\tilde{G}:S\times \Theta \rightarrow S$ and parameterized hybrid system $\mathcal{P}_{\uparrow} = (C,F,D,G,S,\Theta)$ . Now, let + +$$ +C ^ {\prime} (\boldsymbol {\theta}) = C (\boldsymbol {\theta}) \backslash \operatorname {i n t} \tilde {D} (\boldsymbol {\theta}) +$$ + +$$ +D ^ {\prime} (\boldsymbol {\theta}) = \tilde {D} (\boldsymbol {\theta}) \cup D (\boldsymbol {\theta}) +$$ + +$$ +G _ {\tilde {D}} (\boldsymbol {x}, \boldsymbol {\theta}) = \left\{ \begin{array}{l l} \boldsymbol {x} \in D (\boldsymbol {\theta}) \backslash \tilde {D} (\boldsymbol {\theta}) & \quad \text {l a s t} (G) (\boldsymbol {x}, \boldsymbol {\theta}) \\ \boldsymbol {x} \in \tilde {D} (\boldsymbol {\theta}) & \quad \tilde {G} (\boldsymbol {x}, \boldsymbol {\theta}) \end{array} \right. +$$ + +$$ +G _ {D} (\boldsymbol {x}, \boldsymbol {\theta}) = \left\{ \begin{array}{l l} \boldsymbol {x} \in D (\boldsymbol {\theta}) & G (\boldsymbol {x}, \boldsymbol {\theta}) \\ \boldsymbol {x} \in \tilde {D} (\boldsymbol {\theta}) \backslash D (\boldsymbol {\theta}) & \tilde {G} (\boldsymbol {x}, \boldsymbol {\theta}) \end{array} \right. +$$ + +$$ +G ^ {\prime} = G _ {D} \sqcup G _ {\bar {D}} +$$ + +then + +$$ +\mathcal {P} _ {\uparrow} ^ {\prime} = \operatorname {i n s t i n t} _ {\uparrow} \left(\mathcal {P} _ {\uparrow}, \tilde {D}, \tilde {G}\right) = \left(C ^ {\prime}, F, D ^ {\prime}, G ^ {\prime}, \mathcal {S}, \Theta\right). \tag {34} +$$ + +Since $G'$ is an ordered set-valued map (definition 6), we can derive the following identity, which helps establish some useful intuitions. + +$$ +\begin{array}{l} G ^ {\prime} (\boldsymbol {x}, \boldsymbol {\theta}) = G _ {D} (\boldsymbol {x}, \boldsymbol {\theta}) \cup G _ {\tilde {D}} (\boldsymbol {x}, \boldsymbol {\theta}) \\ = \left\{ \begin{array}{l l} \boldsymbol {x} \in D (\boldsymbol {\theta}) \backslash \tilde {D} (\boldsymbol {\theta}) & G (\boldsymbol {x}, \boldsymbol {\theta}) \\ \boldsymbol {x} \in D (\boldsymbol {\theta}) \cap \tilde {D} (\boldsymbol {\theta}) & G (\boldsymbol {x}, \boldsymbol {\theta}) \cup \tilde {G} (\boldsymbol {x}, \boldsymbol {\theta}) \\ \boldsymbol {x} \in \tilde {D} (\boldsymbol {\theta}) \backslash D (\boldsymbol {\theta}) & \tilde {G} (\boldsymbol {x}, \boldsymbol {\theta}) \end{array} \right. \tag {35} \\ \end{array} +$$ + +Below, we additionally use the fact that $\operatorname{last}(G') = G_{\tilde{D}}$ . + +Lemma 3 (Commutativity of instant and lower). Consider parameterized hybrid system $\mathcal{P}_{\uparrow} = (C,F,D,G,\mathcal{S},\Theta)$ that meets assumption 5, and set-valued mappings $\tilde{D}:\Theta \Rightarrow S$ and $\tilde{G}:S\times \Theta \rightarrow S$ that meet assumption 2 relative to $\mathcal{P}_{\uparrow}$ . The following equality then holds, with instant acting as in definition 4 and instint $_{\uparrow}$ as in definition 20: + +$$ +\operatorname {i n s t i n t} \left(\operatorname {l o w e r} \left(\mathcal {P} _ {\uparrow}\right), \tilde {D}, \tilde {G}\right) = \operatorname {l o w e r} \left(\operatorname {i n s t i n t} _ {\uparrow} \left(\mathcal {P} _ {\uparrow}, \tilde {D}, \tilde {G}\right)\right). +$$ + +Proof. First, we adopt the subscript convention, where we use $i$ as a symbol (not a variable) mapping to the intervention operation, and $l$ as a symbol mapping to the lowering operation. The subscript $il$ , + +for example, indicates a system that has been intervened upon and then lowered. With this convention, we have + +$$ +\text {l o w e r} \left(\mathcal {P} _ {\uparrow}\right) = \mathcal {P} _ {l} = \left(C _ {l}, F _ {l}, D _ {l}, G _ {l}, S _ {l}, \Theta_ {l}\right) +$$ + +$$ +\mathbf {i n s t i n t} _ {\uparrow} \left(\mathcal {P} _ {\uparrow}, \tilde {D}, \tilde {G}\right) = \mathcal {P} _ {i} = (C _ {i}, F _ {i}, D _ {i}, G _ {i}, \mathcal {S} _ {i}, \Theta_ {i}) +$$ + +$$ +\text {i n s t i n t} \left(\operatorname {l o w e r} \left(\mathcal {P} _ {\uparrow}\right), \tilde {D}, \tilde {G}\right) = \mathcal {P} _ {l i} = \left(C _ {l i}, F _ {l i}, D _ {l i}, G _ {l i}, \mathcal {S} _ {l i}, \Theta_ {l i}\right) +$$ + +$$ +\operatorname {l o w e r} \left(\operatorname {i n s t i n t} _ {\uparrow} \left(\mathcal {P} _ {\uparrow}, \tilde {D}, \tilde {G}\right)\right) = \mathcal {P} _ {i l} = \left(C _ {i l}, F _ {i l}, D _ {i l}, G _ {i l}, \mathcal {S} _ {i l}, \Theta_ {i l}\right) +$$ + +We now want to show that every element of the tuple $\mathcal{P}_{li}$ equals to the corresponding element in the tuple $\mathcal{P}_{il}$ . We begin with tuple elements that are unaffected by both lower and instinct. These include the parameter space, the state space, and the flow map, meaning we trivially have that + +$$ +\left(F _ {l i}, \mathcal {S} _ {l i}, \Theta_ {l i}\right) = \left(F _ {i l}, \mathcal {S} _ {i l}, \Theta_ {i l}\right) = \left(F _ {l}, \mathcal {S} _ {l}, \Theta_ {l}\right) = \left(F _ {i}, \mathcal {S} _ {i}, \Theta_ {i}\right) = (F, \mathcal {S}, \Theta). \tag {36} +$$ + +For the flow set, note that lower leaves it unmodified and that both the higher and lower overloads of instant list the exact same transformation on the flow set. Thus $C_{li} = C_{il} = C_i = \theta \mapsto C(\theta) \backslash \text{int} \tilde{D}(\theta)$ . + +We now show equivalence in the jump map — a largely straightforward effort despite its morbidity. Consider the system $\mathrm{instint}_{\uparrow}(\mathcal{P}_{\uparrow}) = (C_i,F,D_i,G_i,S,\Theta)$ . We have that $G_{i}$ is an ordered set-valued map that, when lowered, yields its last component: + +$$ +G _ {i l} (\boldsymbol {x}, \boldsymbol {\theta}) = \operatorname {l a s t} \left(G _ {i}\right) (\boldsymbol {x}, \boldsymbol {\theta}) = \left\{ \begin{array}{l l} \boldsymbol {x} \in D (\boldsymbol {\theta}) \backslash \tilde {D} (\boldsymbol {\theta}) & \text {l a s t} (G) (\boldsymbol {x}, \boldsymbol {\theta}) \\ \boldsymbol {x} \in \tilde {D} (\boldsymbol {\theta}) & \tilde {G} (\boldsymbol {x}, \boldsymbol {\theta}) \end{array} . \right. \tag {37} +$$ + +Now we consider the path wherein lowering occurs first. We have that $G_{l}(\pmb{x},\pmb{\theta}) = \mathrm{last}(G)(\pmb{x},\pmb{\theta})$ . By plugging $G_{l}$ into the definition of instant for a lowered system (eq. (3)), equivalence between $G_{li}$ and $G_{il}$ becomes clear. + +Finally, we show equivalence in the jump set. In what follows, let $C'(\pmb{\theta}) = C(\pmb{\theta})\backslash \mathrm{int}\tilde{D} (\pmb{\theta})$ for all $\pmb {\theta}\in \Theta$ and let $\tilde{D}\cup D$ refer to $\pmb {\theta}\mapsto \tilde{D} (\pmb {\theta})\cup D(\pmb {\theta})$ — we drop explicit dependence generally on $\pmb{\theta}$ for brevity. Also, let the set $V_{A} = \{\xi \in S:(\mathrm{VC})$ holds for $\pmb{\xi}$ relative to flow set $A(\pmb {\theta})$ and $F_{\theta}\}$ . We can write the "intervention first" path as + +$$ +\begin{array}{l} D _ {i} = \tilde {D} \cup D \\ D _ {i l} = \text {p r e f e r f l o w} \left(D _ {i}, C _ {i}, F _ {i}\right) \\ = \text {p r e f e r f l o w} \left(\tilde {D} \cup D, C ^ {\prime}, F\right) \\ = \left[ \tilde {D} \cup D \right] \backslash V _ {C ^ {\prime}} = \left[ \tilde {D} \backslash V _ {C ^ {\prime}} \right] \cup \left[ D \backslash V _ {C ^ {\prime}} \right] \\ = \text {p r e f e r f l o w} \left(\tilde {D}, C ^ {\prime}, F\right) \cup \text {p r e f e r f l o w} \left(D, C ^ {\prime}, F\right). \\ \end{array} +$$ + +Following the "lower first" path and looking to instinct as applied to lowered systems (definition 4), we have + +$$ +\begin{array}{l} D _ {l} = \text {p r e f e r f l o w} (D, C, F) \\ D _ {l i} = \operatorname {p r e f e r f l o w} \left(\tilde {D}, C ^ {\prime}, F\right) \cup D _ {l} \\ = \text {p r e f e r f l o w} (\tilde {D}, C ^ {\prime}, F) \cup \text {p r e f e r f l o w} (D, C, F). \\ \end{array} +$$ + +Now, note that because $C'(\pmb{\theta}) \subseteq C(\pmb{\theta})$ , we have $V_{C'} \subseteq V_C$ , and therefore: + +$$ +\begin{array}{l} D _ {i l} = \text {p r e f e r f l o w} (\tilde {D}, C ^ {\prime}, F) \cup \text {p r e f e r f l o w} (D, C ^ {\prime}, F) \\ = \left[ \tilde {D} \backslash V _ {C ^ {\prime}} \right] \cup \left[ D \backslash V _ {C ^ {\prime}} \right] \\ \supseteq \left[ \tilde {D} \backslash V _ {C ^ {\prime}} \right] \cup \left[ D \backslash V _ {C} \right] = D _ {l i}. \\ \end{array} +$$ + +Additionally, by assumption 6 and observation 1, we have that $V_{C} \subseteq V_{\mathrm{int}\tilde{D}} \cup V_{C^{\prime}} \subseteq \tilde{D} \cup V_{C^{\prime}}$ , where the second subset relation follows from the closure of $\tilde{D}$ — nothing can "flow into" the interior of $\tilde{D}$ without being in the closure of that interior. This leads to + +$$ +\begin{array}{l} D _ {l i} = \text {p r e f e r f l o w} (\tilde {D}, C ^ {\prime}, F) \cup \text {p r e f e r f l o w} (D, C, F) \\ = \left[ \tilde {D} \backslash V _ {C ^ {\prime}} \right] \cup \left[ D \backslash V _ {C} \right] \\ \supseteq \left[ \tilde {D} \backslash V _ {C ^ {\prime}} \right] \cup \left[ D \backslash \left(\tilde {D} \cup V _ {C ^ {\prime}}\right) \right] \\ = \left[ \tilde {D} \backslash V _ {C ^ {\prime}} \right] \cup \left[ (D \backslash V _ {C ^ {\prime}}) \cap \mathbb {C} \tilde {D} \right] \\ = \left[ \left(\tilde {D} \backslash V _ {C ^ {\prime}}\right) \cup \left(D \backslash V _ {C ^ {\prime}}\right) \right] \cap \left[ \left(\tilde {D} \backslash V _ {C ^ {\prime}}\right) \cup \mathbb {C} \tilde {D} \right] \\ = \left[ \left(\tilde {D} \backslash V _ {C ^ {\prime}}\right) \cup \left(D \backslash V _ {C ^ {\prime}}\right) \right] \cap \left[ \left(\tilde {D} \cup \mathbb {C} \tilde {D}\right) \cap \left(\mathbb {C} V _ {C ^ {\prime}} \cup \mathbb {C} \tilde {D}\right) \right] \\ = \left[ \left(\tilde {D} \backslash V _ {C ^ {\prime}}\right) \cup \left(D \backslash V _ {C ^ {\prime}}\right) \right] \backslash \left[ V _ {C ^ {\prime}} \cap \tilde {D} \right] \\ \supseteq \left[ \left(\tilde {D} \backslash V _ {C ^ {\prime}}\right) \cup \left(D \backslash V _ {C ^ {\prime}}\right) \right] \backslash V _ {C ^ {\prime}} \\ = \left[ \tilde {D} \backslash V _ {C ^ {\prime}} \right] \cup \left[ D \backslash V _ {C ^ {\prime}} \right] = D _ {i l}. \\ \end{array} +$$ + +With both $D_{li} \supseteq D_{il}$ and $D_{il} \supseteq D_{li}$ , it must be that $D_{li} = D_{il}$ . + +This concludes the proof of equivalence between every element of $\mathcal{P}_{li}$ and $\mathcal{P}_{li}$ , meaning $\mathcal{P}_{li} = \mathcal{P}_{il}$ . + +Lemma 4 (Intervention on Higher System Preserves Key Properties). Consider parameterized hybrid system $\mathcal{P} = \mathrm{lower}(\mathcal{P}_{\uparrow})$ that meets assumption 5, and set-valued mappings $\tilde{D}:\Theta \Rightarrow S$ and $\tilde{G}:S\times \Theta \rightarrow S$ that meet assumption 2 relative to $\mathcal{P}$ . Now, consider the following systems: + +$$ +\mathcal {P} _ {\uparrow} ^ {\prime} = \operatorname {i n s t i n t} _ {\uparrow} \left(\mathcal {P} _ {\uparrow}, \tilde {D}, \tilde {G}\right) \tag {38} +$$ + +$$ +\mathcal {P} ^ {\prime} = \text {l o w e r} \left(\mathcal {P} _ {\uparrow} ^ {\prime}\right) \tag {39} +$$ + +Then $\mathcal{P}'$ satisfies assumption 5. + +Proof. We break the proof into six parts, one for each of the preserved assumptions listed in assumption 5. Recall the explicit form of $\mathcal{P}_{\uparrow}^{\prime}$ given by eq. (34). + +Basic Hybrid Conditions (P1). To show that $\mathrm{instint}_{\uparrow}$ (definition 20) preserves assumption 4, we proceed through the three sub-conditions (A1), (A2), and (A3). + +For (A1), we must demonstrate closure of the intervened jump and flow sets. For the flow set, note that by definition 20, $C'(\pmb{\theta}) = C(\pmb{\theta}) \backslash \text{int} \tilde{D}(\pmb{\theta})$ , and that by (A1) holding for $\mathcal{P}_{\uparrow}$ , $C(\pmb{\theta})$ is closed. $C'(\pmb{\theta})$ is thus the result of subtracting an open set from a closed set, and is therefore closed. We have similarly required that $\tilde{D}(\pmb{\theta})$ is closed. Definition 20 specifies that $D'(\pmb{\theta}) = D(\pmb{\theta}) \cup \tilde{D}(\pmb{\theta})$ , which is the union of closed sets and therefore closed. + +For (A2), note that since the flow map $F_{\theta}$ is unchanged, it trivially remains outer semi-continuous. We then have that $C'(\pmb{\theta}) \subseteq C(\pmb{\theta})$ , from which we can conclude that local boundedness relative to $C(\pmb{\theta})$ and convexity of $F(\pmb{x}, \pmb{\theta})$ for every $\pmb{x} \in C(\pmb{\theta})$ implies those properties relative to $C'(\pmb{\theta})$ . Additionally, we have that $C'(\pmb{\theta}) \subseteq C(\pmb{\theta}) \subset \operatorname{dom} F_{\pmb{\theta}}$ . + +Finally, for (A3), we require the outer semi-continuity of $G_{\theta}^{\prime}$ , its local boundedness relative to $D^{\prime}(\theta)$ , and that $D^{\prime}(\theta)\subset \operatorname{dom}G_{\theta}^{\prime}$ . The following arguments closely mimic the developments in Definition 2.11 and Lemma 2.21 from Sanfelice (2021) — they show that the composition of a hybrid "plant" and hybrid "controller" into a closed loop hybrid system will meet the basic conditions if the plant and controller meet those conditions. In what follows, we work with the identity of $G^{\prime}$ derived in eq. (35). + +Outer semi-continuity of $G_{\theta}^{\prime}$ means that for every convergent sequence $(\pmb{x}_i) \in D'(\pmb{\theta})$ to $\pmb{x}$ and every convergent sequence $(\pmb{x}_i^+) \in S$ to $\pmb{x}^+$ , where $\pmb{x}_i^+ \in G'(\pmb{x}_i, \pmb{\theta})$ for each $i$ , we have that $\pmb{x}^+ \in G'(\pmb{x}, \pmb{\theta})$ . + +Note that this is equivalent to graph closure. Now, by closure of $D(\pmb{\theta})$ , $D'(\pmb{\theta})$ , $\mathcal{G}(G_{\pmb{\theta}})$ , and $\mathcal{G}(\tilde{G}_{\pmb{\theta}})$ , the only potentially problematic limiting points of sequences lying in $D(\pmb{\theta}) \backslash \tilde{D}(\pmb{\theta})$ , or $\tilde{D}(\pmb{\theta}) \backslash D(\pmb{\theta})$ , must lie on the intersection $D(\pmb{\theta}) \cap \tilde{D}(\pmb{\theta})$ . The intersecting piece, however, returns $G(\pmb{x}, \pmb{\theta}) \cup \tilde{G}(\pmb{x}, \pmb{\theta})$ which will necessarily contain those limiting points. + +Local boundedness of $G_{\theta}^{\prime}$ relative to $D^{\prime}(\theta)$ , then, follows from the local boundedness of $G_{\theta}$ relative to $D(\theta)$ and of $\tilde{G}_{\theta}$ relative to $\tilde{D} (\theta)$ , and the fact that $G_{\theta}$ and $\tilde{G}_{\theta}$ are queried by $G_{\theta}^{\prime}$ only from the sets on which they are locally bounded. + +Finally, we need that $D'(\pmb{\theta}) \subset \operatorname{dom} G_{\pmb{\theta}}'$ . Recall the piecewise construction of $G_{\pmb{\theta}}'$ in definition 20 and that $D(\pmb{\theta}) \subset \operatorname{dom} G_{\pmb{\theta}}$ , $\tilde{D}(\pmb{\theta}) \subset \operatorname{dom} \tilde{G}_{\pmb{\theta}}$ . We can then write the following, where we again drop dependence on $\pmb{\theta}$ for brevity and write $G \cup \tilde{G}$ in place of $x \mapsto G(x, \pmb{\theta}) \cup \tilde{G}(x, \pmb{\theta})$ . + +$$ +\begin{array}{l} \operatorname {d o m} G ^ {\prime} = \left[ (D \backslash \tilde {D}) \cap \operatorname {d o m} G \right] \cup \left[ (D \cap \tilde {D}) \cap \operatorname {d o m} G \cup \tilde {G} \right] \cup \left[ (\tilde {D} \backslash D) \cap \operatorname {d o m} \tilde {G} \right] \\ \supset D \cup \left[ (D \cap \tilde {D}) \cap \operatorname {d o m} G \cup \tilde {G} \right] \cup \tilde {D} = D \cup \tilde {D} = D ^ {\prime}. \\ \end{array} +$$ + +Thus, we have that (A1), (A2) and (A3) are all preserved in $\mathcal{P}_{\uparrow}^{\prime} = \mathrm{instint}_{\uparrow}(\mathcal{P}_{\uparrow})$ , meaning it meets assumption 4. + +Basic Existence (P2). To show that conditions for proposition 1 are preserved, we recall from the proof of lemma 6 that it is sufficient to show that (VC) is met (with respect to $\mathcal{P}_{\uparrow}^{\prime}$ ) for all $\pmb {\xi}\in \overline{C'(\pmb{\theta})}\backslash D'(\pmb {\theta})$ for any $\pmb {\theta}\in \Theta$ + +Ignoring whether the flow appropriately remains in the transformed flow set $C'(\theta)$ , we know by assumption 3 that there must be some $\epsilon > 0$ amount of time from which some continuous function can flow from every $\xi \in S$ while respecting the differential inclusion. To confirm that (VC) holds at $\xi$ for $\mathcal{P}_{\uparrow}'$ , we can check whether some $\epsilon' \in (0, \epsilon]$ exists where $z(t) \in C'(\theta)$ for all $t \in (0, \epsilon']$ . First, we can decompose the region where (VC) must hold into a union over two cases. + +$$ +\overline {{C ^ {\prime} (\boldsymbol {\theta})}} \backslash D ^ {\prime} (\boldsymbol {\theta}) = \left[ \operatorname {i n t} C ^ {\prime} (\boldsymbol {\theta}) \backslash D ^ {\prime} (\boldsymbol {\theta}) \right] \cup \left[ \partial C ^ {\prime} (\boldsymbol {\theta}) \backslash D ^ {\prime} (\boldsymbol {\theta}) \right] \tag {40} +$$ + +If $\xi \in \operatorname{int} C'(\pmb{\theta}) \backslash D'(\pmb{\theta}) \subseteq \operatorname{int} C'(\pmb{\theta})$ , there must be some such $\epsilon'$ by the openness of $\operatorname{int} C'(\pmb{\theta})$ in $C'(\pmb{\theta})$ . + +We can then decompose the boundary region $\partial C'(\pmb{\theta})\backslash D'(\pmb{\theta})$ as follows, where we've dropped the dependence on $\pmb{\theta}$ for brevity. + +$$ +\begin{array}{l} \partial C ^ {\prime} \backslash D ^ {\prime} = \partial \left[ C \cap \operatorname {c i n t} \tilde {D} \right] \cap \operatorname {C} \left[ \tilde {D} \cup D \right] \\ = \left[ \left(\partial C \cap \complement \operatorname {i n t} \tilde {D}\right) \cup \left(C \cap \partial \complement \operatorname {i n t} \tilde {D}\right) \right] \cap \complement \tilde {D} \cap \complement D \\ = \left[ \left(\partial C \cap \mathbb {C} \operatorname {i n t} \tilde {D}\right) \cup \left(C \cap \partial \tilde {D}\right) \right] \cap \mathbb {C} \tilde {D} \cap \mathbb {C} D \\ = \left[ \partial C \cap \operatorname {C i n t} \tilde {D} \cap \mathbb {C} \tilde {D} \cap \mathbb {C} D \right] \cup \left[ C \cap \partial \tilde {D} \cap \mathbb {C} \tilde {D} \cap \mathbb {C} D \right] \\ = \left[ \partial C \backslash (\tilde {D} \cup D) \right] \cup \left[ C \cap (\partial \tilde {D} \backslash \tilde {D}) \cap \complement D \right] \\ = \partial C \backslash (\tilde {D} \cup D) \subseteq \partial C \backslash D. \\ \end{array} +$$ + +The first equality in the final line follows from the assumed closure of $\tilde{D}(\pmb{\theta})$ implying that $\partial \tilde{D}(\pmb{\theta}) \backslash \tilde{D}(\pmb{\theta}) = \emptyset$ . + +By analogous decomposition to eq. (40), we have that $\xi \in \partial C(\pmb{\theta})\backslash D(\pmb{\theta})$ must meet (VC) with respect to $\mathcal{P}_{\uparrow}$ . By assumption 6 and observation 1, we then know that the solution must "flow into" either int $\tilde{D}$ or into $C(\pmb{\theta})\backslash \mathrm{int}\tilde{D} (\pmb{\theta})$ . Because $\tilde{D}$ is closed, flow into its interior requires that $\xi \in \tilde{D} (\pmb{\theta})\subseteq D^{\prime}(\pmb{\theta})$ , which we need not consider. This leaves only flows into $C(\pmb{\theta})\backslash \mathrm{int}\tilde{D} (\pmb{\theta})$ , which by construction (definition 20) is equivalent to $C^\prime (\pmb {\theta})$ , and therefore satisfies (VC) with respect to $\mathcal{P}_{\uparrow}^{\prime}$ with $\epsilon^\prime$ as described in observation 1. Thus, instint $_{\uparrow}$ (definition 20) preserves the conditions for existence as outlined in proposition 1. + +Unique Flowing Solution Everywhere (P3). Assumption 3 is preserved trivially, since instant (definition 20) does not alter $F$ , $\Theta$ , or $S$ , which are the only system elements involved in assumption 3. In the remaining results, we use the following observation, leaving its verification from definitions to the reader. + +Observation 4. Let $A, B: \Theta \Rightarrow S$ be two set-valued maps. Then the graph and set operations commute: + +$$ +\mathcal {G} (A \backslash B) = \mathcal {G} (A) \backslash \mathcal {G} (B), \quad \mathcal {G} (A \cap B) = \mathcal {G} (A) \cap \mathcal {G} (B), \quad \mathcal {G} (A \cup B) = \mathcal {G} (A) \cup \mathcal {G} (B). +$$ + +Outer Semi-Continuity of the Flow Set (P4). We need to show the outer semi-continuity of $C^\prime$ at every $\pmb {\theta}\in \Theta$ . Recall from definition 20 that $C^\prime (\pmb {\theta}) = C(\pmb {\theta})\backslash \mathrm{int}\tilde{D} (\pmb {\theta})$ . By observation 4, $\mathcal{G}(C^{\prime}) = \mathcal{G}(C(\pmb {\theta}))\backslash \mathcal{G}(\mathrm{int}\tilde{D} (\pmb {\theta}))$ . By assumption 5, we have the outer semi-continuity of $C$ , which directly implies the closure of its graph. By assumption 2, we have that $\mathcal{G}(\mathrm{int}\tilde{D} (\pmb {\theta}))$ is open. Thus, $\mathcal{G}(C^{\prime})$ is closed, and therefore by (Goebel et al., 2012, Lemma 5.10) $C^\prime$ is outer semi-continuous. + +Borel Jump Set Graph (P5). Recall from definition 20 that $D'(\pmb{\theta}) = \tilde{D}(\pmb{\theta}) \cup D(\pmb{\theta})$ . The Borel $\sigma$ -algebra is closed under unions, and thus by observation 4 the graph $\mathcal{G}(D')$ must also be Borel. + +Borel Measurable, Single-Valued Jump Map (P6). We want to show both that $\operatorname{last}(G')(\boldsymbol{x},\boldsymbol{\theta}) = \{g'(\boldsymbol{x},\boldsymbol{\theta})\}$ for some $g'$ and all $\boldsymbol{x},\boldsymbol{\theta} \in \operatorname{dom} \operatorname{last}(G')$ — i.e. that $\operatorname{last}(G')$ is single valued — and that $g'$ is a Borel-measurable function of initial conditions and parameters on the domain of the intervened jump map. By definition 20, we have that + +$$ +\operatorname {l a s t} \left(G ^ {\prime}\right) (\boldsymbol {x}, \boldsymbol {\theta}) = \left\{ \begin{array}{l l} \boldsymbol {x} \in D (\boldsymbol {\theta}) \backslash \tilde {D} (\boldsymbol {\theta}) & \operatorname {l a s t} (G) (\boldsymbol {x}, \boldsymbol {\theta}) \\ \boldsymbol {x} \in \tilde {D} (\boldsymbol {\theta}) & \tilde {G} (\boldsymbol {x}, \boldsymbol {\theta}) \end{array} \right. \tag {41} +$$ + +Now, we have by (I2) (assumption 2) that $\tilde{G}$ is single-valued, and by (P6) (assumption 5) that $G$ is single-valued. Thus, there must be some $g'$ such that $\operatorname{last}(G')(\boldsymbol{x},\boldsymbol{\theta}) = \{g'(\boldsymbol{x},\boldsymbol{\theta})\}$ on the domain of $\operatorname{last}(G')$ . + +We now want to show that $g'$ is Borel-measurable for every $x, \theta \in \operatorname{dom} \mathbf{last}(G')$ . Note that we can equivalently write the following, where we use the lower-case $\tilde{g}$ and $g$ in reference to the functions that yield the singletons arising from evaluations of $\tilde{G} \mathbf{last}(G)$ . + +$$ +g ^ {\prime} (\boldsymbol {x}, \boldsymbol {\theta}) = g (\boldsymbol {x}, \boldsymbol {\theta}) \mathbb {I} \left[ \boldsymbol {x} \in D (\boldsymbol {\theta}) \backslash \tilde {D} (\boldsymbol {\theta}) \right] + \tilde {g} (\boldsymbol {x}, \boldsymbol {\theta}) \mathbb {I} \left[ \boldsymbol {x} \in \tilde {D} (\boldsymbol {\theta}) \right]. \tag {42} +$$ + +Note now that the indicator functions involving the jump sets can be written as piecewise functions over a partition defined by the graph of the jump sets. We have assumed that the graphs of $\tilde{D}$ and $D$ are Borel. Further, by observation 4, both indicator functions can be written as piecewise over Borel partitions, meaning they must be Borel measurable. Again, by (I2) (assumption 2) we have that $g$ and $\tilde{g}$ are Borel-measurable. Therefore, $g'$ must also be Borel-measurable. + +Having shown the preservation of each sub-condition listed in assumption 5, this concludes the proof. Indeed, if $\mathcal{P}_{\uparrow}$ meets those sub-conditions, then $\mathcal{P}_{\uparrow}^{\prime}$ will as well. In other proofs, this result can be trivially applied to conclude that lower $(\mathcal{P}_{\uparrow}^{\prime})$ fulfills assumption 5. + +# F Proof that Lowering Induces Existence, Uniqueness, and Measurability + +Lemma 1 follows immediately from the following, more precise statement. + +Lemma 5 (Existence, Uniqueness, and Measurability of $\mathcal{P}$ ). Consider a parameterized hybrid system $\mathcal{P}$ that meets assumption 5. $\mathcal{P}$ , then fulfills the conditions for basic existence (proposition 1), basic uniqueness (proposition 2), and $t^{+}$ measurability (definition 12). + +Proof. This result follows directly from combining lemma 6 and corollary 2, both stated in the following sections. $\square$ + +# F.1 Lowering Preserves Existence and Induces Uniqueness + +Lemma 6 (Lowering Preserves Existence and Induces Uniqueness). Consider parameterized hybrid system $\mathcal{P}_{\uparrow}$ that can be lowered (definition 14) to construct a system $\mathcal{P}$ meeting assumption 5. $\mathcal{P}$ then fulfills conditions for basic existence and uniqueness (propositions 1 and 2). + +We split this proof into two components, one for the preservation of existence, and another for the induction of uniqueness. + +# Lowering Preserves Existence + +Proof. Recall from definition 14 (lowering) that $\mathcal{P}' = (C, F, D', G', S, \Theta)$ , with + +$$ +D ^ {\prime} = \boldsymbol {\theta} \mapsto D (\boldsymbol {\theta}) \backslash \{\boldsymbol {\xi} \in \mathcal {S}: (\mathrm {V C}) \text {h o l d s f o r} \boldsymbol {\xi} \}. \tag {43} +$$ + +Now, pick some $\pmb{\theta} \in \Theta$ and note that $\overline{C(\pmb{\theta})} = C(\pmb{\theta})$ , which follows from the basic conditions on $\mathcal{P}$ asserting that $C(\pmb{\theta})$ is closed. By $\mathcal{P}$ fulfilling proposition 1, we have that every $\pmb{\xi} \in [\overline{C(\pmb{\theta})} \cup D(\pmb{\theta})] \backslash D(\pmb{\theta}) = C(\pmb{\theta}) \backslash D(\pmb{\theta})$ must meet (VC). It will be sufficient, analogously, to show that every $\pmb{\xi} \in C(\pmb{\theta}) \backslash D'(\pmb{\theta})$ also meets (VC), which is precisely the same condition because lower affects neither $C$ nor $F$ . Note that + +$$ +\begin{array}{l} C (\boldsymbol {\theta}) \backslash D ^ {\prime} (\boldsymbol {\theta}) = C (\boldsymbol {\theta}) \backslash [ D (\boldsymbol {\theta}) \backslash \{\boldsymbol {\xi} \in S: (\mathrm {V C}) \text {h o l d s f o r} \boldsymbol {\xi} \} ] \\ = C (\boldsymbol {\theta}) \cap \mathbb {C} \left[ D (\boldsymbol {\theta}) \cap \mathbb {C} \left\{\boldsymbol {\xi} \in S: (\mathrm {V C}) \text {h o l d s f o r} \boldsymbol {\xi} \right\} \right] \\ = C (\boldsymbol {\theta}) \cap [ \left[ D (\boldsymbol {\theta}) \cup \{\boldsymbol {\xi} \in S: (\mathrm {V C}) \text {h o l d s f o r} \boldsymbol {\xi} \} \right] \\ = \left[ C (\boldsymbol {\theta}) \cap \complement D (\boldsymbol {\theta}) \right] \cup \left[ C (\boldsymbol {\theta}) \cap \{\boldsymbol {\xi} \in S: (\mathrm {V C}) \text {h o l d s f o r} \boldsymbol {\xi} \} \right] \\ \subseteq \left[ C (\boldsymbol {\theta}) \backslash D (\boldsymbol {\theta}) \right] \cup \left\{\boldsymbol {\xi} \in S: (\mathrm {V C}) \text {h o l d s f o r} \boldsymbol {\xi} \right\}. \\ \end{array} +$$ + +This yields two cases under which we must check that (VC) holds. For the first case, we know already that (VC) holds for all $\pmb {\xi}\in [C(\pmb {\theta})\backslash D(\pmb {\theta})]$ . For the second, we have by construction that (VC) holds. By the subset relation, these cases subsume the desired set. + +We thus have our sufficient condition, that (VC) is met for any $\xi \in C(\theta)\backslash D'(\theta)$ . As we have placed no constraints on $\theta$ , this holds for the entirety of $\Theta$ . + +# Lowering Induces Uniqueness + +Proof. Uniqueness for $\mathcal{P}'$ involves three conditions on the hybrid system as stated in proposition 2, which we will review in reverse order of complexity. Before proceeding, recall from definition 14 (lowering) that $\mathcal{P}' = (C,F,D',G',S,\Theta)$ , where for all $\theta \in \Theta$ and $x \in S$ + +$$ +D ^ {\prime} (\boldsymbol {\theta}) = D (\boldsymbol {\theta}) \backslash \{\boldsymbol {\xi} \in \mathcal {S}: (\mathrm {V C}) \text {h o l d s f o r} \boldsymbol {\xi} \} \tag {44} +$$ + +$$ +G ^ {\prime} (\boldsymbol {x}, \boldsymbol {\theta}) = \operatorname {l a s t} (G) (\boldsymbol {x}, \boldsymbol {\theta}) \tag {45} +$$ + +Recall, also, the convention that $G_{\pmb{\theta}}^{\prime}(\pmb{x}) = G^{\prime}(\pmb{x}, \pmb{\theta})$ for all $\pmb{x}, \pmb{\theta} \in S \times \Theta$ . Now, pick some $\pmb{\theta} \in \Theta$ . + +Condition (a) requires the uniqueness of solutions to the differential inclusion on the flow set. This condition is precisely what we have presupposed in assumption 3, except that we make the stronger claim that uniqueness holds for any flow in $S \supseteq \overline{C(\theta)}$ . + +Condition (c) requires that $G_{\theta}^{\prime}$ is single-valued on the jump set. We have assumed in (P6) that $\text{last}(G)$ is no more than single-valued on $\text{dom } G$ , which implies that it is single valued on $\text{dom } G_{\theta}$ for every fixed $\theta \in \Theta$ . Additionally, by construction of $D^{\prime}$ and the basic conditions on $\mathcal{P}$ , we have that + +$$ +D ^ {\prime} (\boldsymbol {\theta}) \subseteq D (\boldsymbol {\theta}) \subset \operatorname {d o m} G _ {\boldsymbol {\theta}} = \operatorname {d o m} G _ {\boldsymbol {\theta}} ^ {\prime} \tag {46} +$$ + +Thus, $G_{\theta}^{\prime}$ is exactly single valued on $D^{\prime}$ . + +Condition (b) requires that the solution cannot flow from the overlap of the jump and flow sets. Precisely, for every $\pmb {\xi}\in \overline{C(\pmb{\theta})}\cap D^{\prime}(\pmb {\theta})$ , (VC) as used in definition 13 does not hold. By assumption 4 on $\mathcal{P}$ , we have that $C(\pmb {\theta}) = \overline{C(\pmb{\theta})}$ , and recalling definition 13, it is sufficient to show that (VC) does not hold for any $\pmb{\xi}$ in the following set: + +$$ +\begin{array}{l} \overline {{C (\boldsymbol {\theta})}} \cap D ^ {\prime} (\boldsymbol {\theta}) = C (\boldsymbol {\theta}) \cap [ D (\boldsymbol {\theta}) \backslash \{\boldsymbol {\xi} \in S: (\mathrm {V C}) \text {h o l d s f o r} \boldsymbol {\xi} \} ] \\ = C (\boldsymbol {\theta}) \cap D (\boldsymbol {\theta}) \cap \complement \left\{\boldsymbol {\xi} \in S: (\mathrm {V C}) \text {h o l d s f o r} \boldsymbol {\xi} \right\} \\ = C (\boldsymbol {\theta}) \cap D (\boldsymbol {\theta}) \cap \{\boldsymbol {\xi} \in S: \neg (\mathrm {V C}) \text {h o l d s f o r} \boldsymbol {\xi} \} \\ \subseteq \left\{\boldsymbol {\xi} \in S: \neg (\mathrm {V C}) \text {h o l d s f o r} \boldsymbol {\xi} \right\}. \\ \end{array} +$$ + +This concludes the proof. + +![](images/eda52460e929eab3fb3f657c644c0e85cd4eb2118404c222b2efc73cac9a357a.jpg) + +# F.2 Lowering Induces Measurability + +We first state sufficient conditions for measurability, and then prove that sufficiency. Ultimately, this yields a corollary stating that lowering induces measurability. We make use of the intermediate results and definitions established in appendix F.3. + +Assumption 7 (Collected Conditions for Measurability). Consider parameterized hybrid system $\mathcal{P} = (C,F,D,G,\mathcal{S},\Theta)$ . Assume that $\mathcal{P}$ + +(M1) is $t^+$ uniquely evaluable (definition 10); +(M2) has a unique solution to its differential inclusion everywhere (assumption 3); +(M3) has an outer semi-continuous and closed flow set $C(\pmb{\theta})$ at every $\pmb{\theta} \in \Theta$ ; +(M4) $G$ is single-valued on $\operatorname{dom} G$ , with $G(\boldsymbol{x}, \boldsymbol{\theta}) = \{g(\boldsymbol{x}, \boldsymbol{\theta})\}$ , and $g$ Borel-measurable for all $\boldsymbol{x}, \boldsymbol{\theta} \in \operatorname{dom} G$ . + +Theorem 2 (Measurability of Solution). Consider parameterized hybrid system $\mathcal{P} = (C, F, D, G, S, \Theta)$ and its time-parameterized solution map $\varphi$ (definition 11). If $\mathcal{P}$ meets assumption 7, then $\varphi$ is $t^{+}$ measurable (definition 12). + +Proof. Under assumption 7, finite jump times and values are Borel measurable in $\xi, \theta$ (lemma 7). Additionally, under assumption 3, the solution is Borel-measurable in $\xi, \theta$ up to the first jump (F2-3). We are thus able to write the time-parameterized solution as follows, where $t_0(\xi, \theta) = 0$ always. For all $t \in [0, t^+)$ , $\xi \in S$ , and $\theta \in \Theta$ : + +$$ +\varphi (t; \boldsymbol {\xi}, \boldsymbol {\theta}) = \sum_ {j = 1} ^ {\infty} \mathbb {I} [ t _ {j - 1} (\boldsymbol {\xi}, \boldsymbol {\theta}) \leqslant t < t _ {j} (\boldsymbol {\xi}, \boldsymbol {\theta}) ] \phi \left(t - t _ {j - 1} \left(\boldsymbol {\xi} _ {j - 1} (\boldsymbol {\xi}, \boldsymbol {\theta}), \boldsymbol {\theta}\right), 0; \boldsymbol {\xi} _ {j} (\boldsymbol {\xi}, \boldsymbol {\theta}), \boldsymbol {\theta}\right) \tag {47} +$$ + +which comprises a countable sum over Borel-measurable functions of $\pmb{\xi},\pmb{\theta}$ , and is therefore itself Borel measurable. Note that, while we have only shown Borel-measurability for $\pmb{\xi}_j(\pmb{\xi},\pmb{\theta}) = \phi(t_{j-1}(\pmb{\xi},\pmb{\theta}), j; \pmb{\xi},\pmb{\theta})$ when $t_{j-1}(\pmb{\xi},\pmb{\theta}) < t^+$ , the joint requirement that $t_{j-1}(\pmb{\xi},\pmb{\theta}) \leqslant t < t^+$ avoids those unmeasurable cases. + +Corollary 2 (Lowering Induces Measurability). Consider parameterized hybrid system $\mathcal{P}_{\uparrow}$ that can be lowered (definition 14) to construct a system $\mathcal{P}$ meeting assumption 5. $\mathcal{P}$ then fulfills conditions for the $t^{+}$ measurability of its time-parameterized solution map $\varphi$ (definition 11). + +Proof. Assumption 5, when combined with the fact that "lowering" (definition 14) induces uniqueness and preserves existence (lemma 6), subsumes or implies conditions sufficient for the result (assumption 7 and theorem 2). In particular, (M2) maps to (P3), (M3) maps to (A1) and (P4), and (M4) maps to (P6). For (M1), note that $t^+$ measurability requires only $t^+ \geqslant 0$ in addition to existence and uniqueness, which come from lemma 6. + +Proof of lemma 5. This result follows directly from combining lemma 6 in appendix F.1 and corollary 2 above. $\square$ + +# F.3 Measurability of Jump Times and Values + +Definition 21 (Flowable Region). Consider parameterized hybrid system $\mathcal{P} = (C,F,D,G,\mathcal{S},\Theta)$ . For all $\pmb{\theta}\in \Theta$ , let $C_F(\pmb{\theta})$ denote the set of states from which there exist a flowing solution (respecting $F_{\pmb{\theta}}$ ) that remains in $C(\pmb{\theta})$ after its start. Precisely, this means that there exist $\epsilon >0$ and an absolutely continuous function $z:[0,\epsilon ]\to S$ such that $z(0)\in C_F(\pmb {\theta})$ and $\dot{z} (t)\in F(z(t))$ for almost all $t\in [0,\epsilon ]$ and $z(t)\in C(\pmb {\theta})$ for all $t\in (0,\epsilon ]$ . + +**Observation 5.** Consider parameterized hybrid system $\mathcal{P} = (C, F, D, G, S, \Theta)$ that has a unique solution to its differential inclusion everywhere (assumption 3), and where $C(\pmb{\theta})$ is closed for every $\pmb{\theta} \in \Theta$ . In this case, the closure of the flowable region is $\overline{C_F(\pmb{\theta})} = C(\pmb{\theta})$ . + +Proof. From assumption 3, for every $\pmb{\theta} \in \Theta$ , we have that an absolutely continuous function $z$ exists from every $z(0) = \pmb{\xi} \in S \supseteq C_F(\pmb{\theta})$ that satisfies $F_{\pmb{\theta}}$ . Every interior point $\pmb{\xi} \in \operatorname{int} C(\pmb{\theta})$ , then, must be in $C_F(\pmb{\theta})$ , as some flow must be possible while remaining in $C(\pmb{\theta})$ . With the closure of the flow set, we thus have $C(\pmb{\theta}) = \overline{\operatorname{int} C(\pmb{\theta})} \subseteq \overline{C_F(\pmb{\theta})}$ . Now, for points $\pmb{\xi} \in S \backslash \operatorname{int} C(\pmb{\theta})$ , note that flow into $C(\pmb{\theta})$ is only possible from $\partial C(\pmb{\theta}) \subseteq C(\pmb{\theta})$ . This ensures that $C_F(\pmb{\theta})$ cannot contain points outside of $C(\pmb{\theta})$ , further implying that $\overline{C_F(\pmb{\theta})} \subseteq \overline{C(\pmb{\theta})} = C(\pmb{\theta})$ . Thus, by a two-sided inclusion, we have $\overline{C_F(\pmb{\theta})} = C(\pmb{\theta})$ for every $\pmb{\theta} \in \Theta$ . + +Lemma 7 (Measurability of Jump Times and Values). Consider parameterized hybrid system $\mathcal{P} = (C,F,D,G,\mathcal{S},\Theta)$ and its solution map $\phi$ . If $\mathcal{P}$ meets assumption 7, then the time of the $j > 0$ th jump, + +$$ +t _ {j} (\boldsymbol {\xi}, \boldsymbol {\theta}) = \sup \left\{t \mid (t, j - 1) \in d o m \phi (\cdot , \cdot ; \boldsymbol {\xi}, \boldsymbol {\theta}) \right\} \in \mathbb {R} _ {\geqslant 0} \cup \{\infty \} \tag {48} +$$ + +is a Borel measurable function of $\xi, \theta$ . + +Additionally, the solution values at these jump times + +$$ +\boldsymbol {\xi} _ {j + 1} (\boldsymbol {\xi}, \boldsymbol {\theta}) = \phi \left(t _ {j} (\boldsymbol {\xi}, \boldsymbol {\theta}), j; \boldsymbol {\xi}, \boldsymbol {\theta}\right) \tag {49} +$$ + +are also Borel measurable functions of $\xi, \theta$ if $t_j(\xi, \theta) < t^+$ . + +Proof. Let $t_1(\pmb{\xi}, \pmb{\theta}) = \sup \{t \mid (t, 0) \in \operatorname{dom} \phi(:, \pmb{\xi}, \pmb{\theta})\}$ be the first jump time. Note that if the set $\{(\pmb{\xi}, \pmb{\theta}): t_1(\pmb{\xi}, \pmb{\theta}) \geqslant \alpha\}$ is Borel for all $\alpha \in \mathbb{R}$ , then $t_1$ must be Borel measurable. Indeed, we can write that set as a countable intersection of Borel sets, which implies Borelness. Below, we use the closure of the "flowable region" $\overline{C_F(\pmb{\theta})}$ (definition 21) to rewrite the pre-image on $S \times \Theta$ of the first jump occurring at or after time $\alpha$ . In particular, we use its closure in order to include the time at which the jump occurs (by including states that flow can reach but not flow from). Note also that, by observation 5, we have that $\overline{C_F(\pmb{\theta})} = C(\pmb{\theta})$ under conditions already provided in assumption 7. We have + +$$ +\begin{array}{l} \left\{(\boldsymbol {\xi}, \boldsymbol {\theta}): t _ {1} (\boldsymbol {\xi}, \boldsymbol {\theta}) \geqslant \alpha \right\} = \left\{\left(\boldsymbol {\xi}, \boldsymbol {\theta}\right): \phi (\tau , 0; \boldsymbol {\xi}, \boldsymbol {\theta}) \in \overline {{C _ {F} (\boldsymbol {\theta})}} = C (\boldsymbol {\theta}) \forall \tau \in [ 0, \alpha ] \right\} \\ = \bigcap_ {\tau \in \mathbb {Q} \cap [ 0, \alpha ]} \left\{\left(\boldsymbol {\xi}, \boldsymbol {\theta}\right): \left(\boldsymbol {\theta}, \phi (\tau , 0; \boldsymbol {\xi}, \boldsymbol {\theta})\right) \in \mathcal {G} (C) \right\} \\ \end{array} +$$ + +Assumption 7 requires that $C(\theta)$ is outer semi-continuous at all $\theta \in \Theta$ . This holds if and only if its graph $\mathcal{G}(C)$ is closed (Sanfelice, 2021, pg. 49). Closed sets are Borel, so $\mathcal{G}(C)$ must be Borel. + +The Borelness of $\{(\pmb {\xi},\pmb {\theta}):\pmb {\theta},\phi (\tau ,0;\pmb {\xi},\pmb {\theta}))\in \mathcal{G}(C)\}$ follows from the Borelness of $\mathcal{G}(C)$ and $\phi$ being continuous in $\pmb {\xi},\pmb{\theta}$ on $[0,t]$ , and therefore Borel measurable. + +The ability to write the set as a countable intersection over rationals is justified by the standard argument. For any fixed $\tau \in [0,\alpha ]$ , choose a sequence $(\tau_{n})\subset \mathbb{Q}\cap [0,\alpha ]$ such that $\tau_{n}\to \tau$ . This is always possible due to the density of $\mathbb{Q}\cap [0,\alpha ]$ in $[0,\alpha ]$ . If for each $n\in \mathbb{N}$ we have + +$$ +\left(\boldsymbol {\theta}, \phi \left(\tau_ {n}, 0; \boldsymbol {\xi}, \boldsymbol {\theta}\right)\right) \in \mathcal {G} (C), \tag {50} +$$ + +then, because $\phi$ is continuous in time and $\mathcal{G}(C)$ is closed, the limit is also included + +$$ +\left(\boldsymbol {\theta}, \lim _ {n \rightarrow \infty} \phi \left(\tau_ {n}, 0; \boldsymbol {\xi}, \boldsymbol {\theta}\right)\right) = \left(\boldsymbol {\theta}, \phi \left(\tau , 0; \boldsymbol {\xi}, \boldsymbol {\theta}\right)\right) \in \mathcal {G} (C). \tag {51} +$$ + +Countable intersections of Borel sets are Borel, and thus $t_1$ must be Borel measurable. + +We must now expand from the measurability of the first jump to the measurability of all jumps. We can write the second jump time as follows, with $g$ being a function that, when evaluated, returns the single value of $G$ (M4). + +$$ +\boldsymbol {\xi} _ {2} (\boldsymbol {\xi}, \boldsymbol {\theta}) = g \left(\phi \left(t _ {1} (\boldsymbol {\xi}, \boldsymbol {\theta}), 0; \boldsymbol {\xi}, \boldsymbol {\theta}\right), \boldsymbol {\theta}\right) \tag {52} +$$ + +$$ +t _ {2} (\boldsymbol {\xi}, \boldsymbol {\theta}) = t _ {1} (\boldsymbol {\xi}, \boldsymbol {\theta}) + t _ {1} \left(\boldsymbol {\xi} _ {2} (\boldsymbol {\xi}, \boldsymbol {\theta}), \boldsymbol {\theta}\right) \tag {53} +$$ + +We have that $g$ is Borel-measurable on the domain of $G$ (M4) and that, by definition of a parameterized hybrid system (definition 1), $D(\theta) \subset \operatorname{dom} G_{\theta}$ , and therefore know that $g$ will only be evaluated where it is assumed to be measurable (M4). Additionally, we have that $\phi$ is Borel-measurable for $j = 0$ up to and including $t_1(\xi, \theta)$ (F2-3). Thus, $\xi_2$ is the composition of Borel-measurable functions and is therefore itself Borel-measurable. $t_2$ , subsequently, is the sum of a Borel-measurable function and the composition of Borel-measurable functions. The measurability of $t_{j+1}$ for $j > 0$ then follows from its recursive form. We use $h^{(n)}(x)$ to represent the $n$ -fold composition of $h$ with itself. By standard inductive arguments we have + +$$ +h _ {\boldsymbol {\theta}} (x) = g \left(\phi \left(t _ {1} (x, \boldsymbol {\theta}), 0; x, \boldsymbol {\theta}\right), \boldsymbol {\theta}\right) \tag {54} +$$ + +$$ +\boldsymbol {\xi} _ {j + 1} (\boldsymbol {\xi}, \boldsymbol {\theta}) = h _ {\boldsymbol {\theta}} ^ {(j)} (\boldsymbol {\xi}); \quad \boldsymbol {\xi} _ {1} (\boldsymbol {\xi}, \boldsymbol {\theta}) = \boldsymbol {\xi} \tag {55} +$$ + +$$ +t _ {j + 1} (\boldsymbol {\xi}, \boldsymbol {\theta}) = \sum_ {i = 1} ^ {j + 1} t _ {1} \left(\boldsymbol {\xi} _ {i} (\boldsymbol {\xi}, \boldsymbol {\theta}), \boldsymbol {\theta}\right) \tag {56} +$$ + +As $t_{j+1}$ comprises only sums of compositions of Borel-measurable functions, it must also be Borel-measurable. Additionally, note that $\pmb{\xi}_{j+1}(\pmb{\xi},\pmb{\theta}) = \phi(t_j(\pmb{\xi},\pmb{\theta}), j; \pmb{\xi},\pmb{\theta})$ can also be written as the composition of Borel-measurable functions, thereby proving the measurability of jump values. + +# G Probabilities of Causation and Fishery Management + +# G.1 Historical Context for the Fishery Management Problem + +Notions of causal necessity and sufficiency are often productively employed in policy discourse, especially where competing interests require human-understandable justifications as to whether a particular policy is sufficient and/or necessary to achieve desired outcomes. Recall the control theoretic settings involving state-dependent, instantaneous interventions that we have enumerated in the introduction: health-related lockdown measures, interest rate adjustments, and many engineering problems involve cost benefit tradeoffs, where policies are designed to be sufficient for the benefits, but only as costly as necessary. In modern resource management, for example, tragedies of the commons frequently demand a challenging balance between ecological objectives and short and long-term economic outcomes. Additionally, such cases often involve models that our interventional semantics is designed to operate on. + +Fishery management offers a particularly rich set of problems where the probabilities of causation can help streamline policy discourse. Over the last few decades, numerous fishery management crises have followed a similar arc: first, growing markets and new technologies result in overfishing to unsustainable biomass levels; then, regulators impose strict catch quotas, gear restrictions, data collection requirements, area closures, and other measures designed to allow stocks to rebuild; after rebuilding stocks, fishing resumes, ideally at more sustainable levels. In 2000, for example, the NMFS and $\mathrm{NOAA}^{21}$ announced emergency regulatory measures in response to the failure of the Pacific coast groundfish fishery (Anon, 2000). This was followed by an economically tumultuous rebuilding period of around 10 years (Warlick et al., 2018), after which fishing restrictions changed and loosened (Anon, 2010a). Similarly, the 1990s saw significant declines in the Atlantic swordfish fishery (Neilson et al., 2013). In 2000, the ICCAT22 established an ultimately successful 10-year plan to rebuild the stock (Neilson et al., 2013; Anon, 2010b). + +Naturally, these measures were not without significant economic consequences and backlash, both short and long term (Anon, 2000; Cramer et al., 2018; Anon, 2007b). Indeed, in the United States, + +the Magnuson-Stevens Act (MSA) mandates the multi-objective of avoiding unnecessary economic sacrifice while pursuing long-term economic and ecological outcomes (Anon, 1976, 1996, 2007a, 2018). Myriad ecological and bio-economic dynamical systems approaches were developed during and after these crises to better balance competing objectives (Lee et al., 2000; Ortiz et al., 2010; Restrepo et al., 2011; Taylor et al., 2022). On some occasions, post-mortems were employed to, for example, determine the degree to which rebuilding success was due to management actions or to natural factors such as species biology Neilson et al. (2013). In essence, the goal of such efforts, as stated in the MSA, is to identify and implement sufficient rebuilding measures that would induce no more economic hardship than necessary. + +# G.2 Formal Probabilities of Causation + +The formal definitions of the probabilities of causation were originally provided by Pearl (1999). These queries are traditionally defined for binary treatment $X$ and outcome $Y$ variables — we enumerate those binary definitions here, and then develop some intuition. In our fishery management example (section 6), we expand to the non-binary setting in keeping with definitions provided by Kawakami et al. (2024) for scalar treatment and outcome variables. + +Definition 22 (Probabilities of Causation). Let $X, Y$ be binary variables within a structural causal model $M$ , and let $x, x', y, y'$ , denote the propositions $X = 1$ , $X = 0$ , $Y = 1$ , $Y = 0$ respectively. Denote by $Y_x$ and $Y_{x'}$ the counterfactual outcomes obtained by performing the do-interventions $\mathsf{d}\mathsf{o}(X = 1)$ and $\mathsf{d}\mathsf{o}(X = 0)$ . The probabilities of causation Pearl (1999), then, are defined as follows: + +$$ +P N (x, y) := P \left(Y _ {x ^ {\prime}} = 0 \mid x, y\right), \tag {57} +$$ + +$$ +P S (x, y) := P \left(Y _ {x} = 1 \mid x ^ {\prime}, y ^ {\prime}\right), \tag {58} +$$ + +$$ +P N S (x, y) := P \left(Y _ {x} = 1, Y _ {x ^ {\prime}} = 0\right). \tag {59} +$$ + +$PN(x,y)$ quantifies the probability that $x$ was necessary to produce outcome $y$ ; $PS(x,y)$ quantifies the probability that $x$ alone would suffice to produce $y$ ; and $PNS(x,y)$ jointly quantifies the event that $x$ is both necessary and sufficient for outcome $y$ . + +To compute the probability of necessity, we consider only (condition on) worlds where the events $x$ and $y$ occurred, and then evaluate the probability of $Y$ being false if we intervene to make $X$ false. Similarly, the capacity to produce an outcome — the probability of sufficiency — is computed by conditioning on $X$ and $Y$ being false, and evaluating the probability of $Y$ being true if we now intervene to make $X$ true. A notion balancing the dimensions of necessity and sufficiency is the probability of necessity and sufficiency, which is not a function of the separate probabilities. To evaluate $PNS$ , we do not condition either way,[24] but rather evaluate the probability that both intervening to make $X$ true results in $Y_x = 1$ and intervening to make $X$ false results in $Y_{x'} = 1$ . + +# G.3 Multi-Year Horizon + +In our analysis of the fishery management problem, we analyze only the year-long time scale, but we can define a multi-season model with an arbitrarily long time horizon. Note that, here, we will need to prepend time to the state vector, which becomes $[t,z,h_{1:3},b_{1:3}] = \pmb{x} \in S = \mathbb{R}_{\geqslant}^{8}$ . See table 2 for a full labeling of model parameters and states. + +First, model the season's starting condition via a jump set that triggers at the beginning of each year. Jumps at the season's start obey a map that (1) resets the integrated catch $z$ to zero and (2) sets fishing harvest rates to their noisy, non-null values. Let $\theta_{h_2} \sim \mathcal{N}(.7, .07)$ and $\theta_{h_3} \sim \mathcal{N}(.07, .007)$ be elements of $\theta$ , season-start times be $\mathbb{Z}_{\geqslant 0}$ (i.e. the beginning of each year), and rstatint (definition 19) be a generalization of statint (definition 17) that applies the jump map at countably many times.[25] Let $\mathcal{P}_0 = ((C, F, \emptyset, \cdot), S, \Theta)$ describe the fishery in its unfished, natural state, where + +![](images/39076bfbbd0716cc66acdf01382e53229130fdaec892517a2533aabc10ce5e89.jpg) +Figure 3: Examples of the biomass trajectories of apex and intermediate predators, as simulated from the model proposed by Zhou & Smith (2017). The panels comprise simulations with, (left) no fishing pressure, (center) fishing pressure kept up throughout the year, (right) fishing regulators ending the season when reported catch meets the total allowable catch (TAC) quota of 50 biomass units. + +
namenotationin seasonafter season
timet
total catchz
fishing pressure forageh100
fishing pressure intermediateh2h2 ~ Normal(.7, .07)0
fishing pressure apexh3h3 ~ Normal(.07, .007)0
biomass forageb1
biomass intermediateb2
biomass apexb3
desired outcome lower thresholdγ130130
TAC quotaqi
+ +Table 2: Parameters and notation for the fishery example. + +$\mathcal{S} = C = \mathbb{R}_{\geqslant 0}^{8}$ , and $\cdot$ simply indicates the irrelevance of the jump map in the natural state of the fishery. + +$$ +\tilde {G} _ {s} (\boldsymbol {x}, \boldsymbol {\theta}) = \left\{\left[ t, 0, 0, \theta_ {h _ {2}}, \theta_ {h _ {3}}, b _ {1: 3} \right] \right\} \tag {60} +$$ + +$$ +\mathcal {P} _ {s} = \operatorname {r s t a t i n t} \left(\mathcal {P} _ {0}, \mathbb {Z} _ {\geqslant 0}, \tilde {G} _ {s}\right) \tag {61} +$$ + +The season's end can be described by setting the harvest rates to zero when the catch exceeds a threshold $q_{i}$ (with $i \in \{1, 2\}$ ). From these, we can construct parallel worlds with the same random initial conditions and parameters. + +$$ +\tilde {D} _ {q _ {i}} (\boldsymbol {\theta}) = \mathbb {R} _ {\geqslant 0} \times \left\{z \in \mathbb {R} _ {\geqslant 0} \mid z \geqslant q _ {i} \right\} \times \mathbb {R} _ {\geqslant 0} ^ {6} \tag {62} +$$ + +$$ +\tilde {G} _ {q _ {i}} (\boldsymbol {x}, \boldsymbol {\theta}) = \left\{\left[ t, z, 0, 0, 0, b _ {1: 3} \right] \right\} \tag {63} +$$ + +$$ +\mathcal {P} _ {q _ {i}} = \operatorname {i n s t i n t} \left(\mathcal {P} _ {s}, \tilde {D} _ {q _ {i}}, \tilde {G} _ {q _ {i}}\right) \tag {64} +$$ + +$$ +\mathcal {R} _ {s} = \left(\mathcal {P} _ {s}, \xi , \theta\right); \mathcal {R} _ {q _ {1}} = \left(\mathcal {P} _ {q _ {1}}, \xi , \theta\right); \mathcal {R} _ {q _ {2}} = \left(\mathcal {P} _ {q _ {2}}, \xi , \theta\right) \tag {65} +$$ + +# G.4 Narrative Fishery Management Example + +In the main body of the paper, we emphasized the construction of the probabilities of causation, rather than their application. Still, some readers may appreciate a more narrative structure around these concepts. We provide that here. + +Example 5 (Probabilities of Causation for Total Allowable Catch (TAC) Quotas). Now, suppose a new commercial fishery is being opened up and that, in the first year, fishery managers allow commercial fishing year-round. $\mathcal{R}_s$ models this world (or equivalently, $\mathcal{R}_{q_1}$ when the TAC quota $q_{1}$ is sufficiently large so as to have zero probability of being reached). This results in a failure to preserve the intermediate level biomass above the desired level $\gamma$ . Suppose $\gamma = 130$ units. Facing ecological scrutiny, fishery managers ask: given that we allowed year-round fishing and failed to achieve our outcome, what TAC quota would have a high probability of being sufficient for success? This is a probability of sufficiency query. They introduce a strict TAC quota of 30 units with a relatively high probability of sufficiency (fig. 4). In the next season, they succeed in meeting biomass targets. Subsequently, however, economic interests and local representatives insist that such a low, strict TAC was not necessary to achieve this outcome. They point out that, in comparison to a more lenient TAC of 50 units, there is a low probability that the strict TAC of 30 was necessary (fig. 5). Fishery managers, in turn, worry that the probability of success with a TAC of 50 might be too low. Before the start of the next season, stakeholders resolve the disagreement by identifying a TAC that yields a high probability of sufficiency and necessity when contrasted with year-round fishing $(\mathcal{R}_s)$ , all while avoiding stricter catch limitations that are not justified by gains in the probability of necessity and sufficiency (fig. 6). + +The simulated results presented here ran on a consumer grade laptop in the order of one hour. + +# G.5 Event-Time Attribution + +In the main body of the paper (and in example 5), we analyzed the probabilities of causation as they relate to contrastive policies. In other words, we asked causal attribution questions at the policy level. Queries about the probabilities of causation, however, such as "was $x$ necessary to achieve $y$ ", are ambiguous when a real world event $x$ is multi-faceted and potential alternative actions are plentiful. In example 5, we mapped the events $x$ and $x'$ onto particular TAC quotas. In this next example, however, we will define our event of interest as involving the time at which an intervention occurs. We can make this precise by constructing twin worlds using the tools provided in this paper — particularly by additionally employing the static intervention statint (definition 17). + +Example 6 (Probability of Necessity of State-Dependent Intervention Timing). Consider worlds where the season ends before some time $\lambda$ , and biomass goal $\gamma$ is achieved at a later time $\tau$ (for instance, at the end of the year). Fishery managers wonder whether they might fail to meet their biomass goals if, contrary to fact in those cases, the season had ended at or after time $\lambda$ . The relevant question is: was ending the season before time $\lambda$ necessary for achieving the biomass goal? We can answer this question by asking a probability of necessity query. Unlike in example 5, however, the causal attribution question relates to the time at which the intervention occurs. Consider the following binary predicates, where $T(\varphi_{q_1})$ extracts the time at which the season ends due to the TAC quota $q_1$ :26 + +$$ +X = \mathbb {I} \left[ T \left(\varphi_ {q _ {1}}\right) < \lambda \right]; \quad Y = \mathbb {I} \left[ \mathrm {b} _ {q _ {1}} ^ {\tau} \geqslant \gamma \right] \tag {66} +$$ + +![](images/57084b9a6c929083416cf6912d3bd1b2b0e71bf4847be9f3dfe3f0bae2676601.jpg) +Figure 4: Step 1 in the example narrative (example 5). Within the first year of commercial fishing, the fishery has no quota (here it is enough to set it to $q = 120$ , which is never met), and falls below sustainable biomass. Conditioning on this failure, the regulators seek an intervention with a high probability of changing this outcome next time along the counterfactual dimension. They implement a strict TAC of 30, evaluating the probability of sufficiency (the probability of achieving sustainable biomass above the desired threshold of 130) to be 0.87. + +![](images/c29756e70e6c58deed8132ae2678637b916adb006bbc13c6ce0bebc5093cf988.jpg) + +![](images/0c1fa9ec70d2e799b4a104dda3a7ae64d415d38c2349be1d8516d58e9a1f33ba.jpg) +Figure 5: Step 2 in the example narrative (example 5). The season ran with a TAC of 30 units, and the intermediate biomass target reference limit $(\gamma = 130)$ was met. Conditioning on this, parties interested in increasing the fishing quota ask whether such a low TAC was necessary. They seek an alternative quota along the counterfactual dimension that, when contrasted with the factual TAC of 30, reveals the factual TAC as probably unnecessary. As a counterexample, they choose a TAC of 50, which yields a relatively low probability of necessity (.18). + +![](images/624b3c4c48d0b58cceccda1f45c5fbb4046f2a647e2aca43de679c8980d2ee9d.jpg) + +![](images/1f135cacd76cdf30b658d4866cc836602d635fe3cd0c307acd394288ea08dbbb.jpg) +Figure 6: Step 3 in the example narrative (example 5). This time, before the season starts and prior to seeing what the outcome will be, both sides aim to find a quota with a large probability of both necessity and sufficiency. They contrast proposed TAC quotas with a baseline, status quo TAC of 120 units (never met). They notice that the probability surface flattens out above .60, meaning further improvement in the probability of necessity and sufficiency would require excessive limitations in quota. Ultimately, they agree on a quota that results in a value above 0.6, i.e., a TAC quota of 0.35. + +![](images/ccfbbd7e13218698985c6dcfe04db74ce7d3959295b940bc288d913a9abac731.jpg) + +![](images/02b1d111c30d3e784edf7cd7bf0ac647eaec6b4a07c456d8ee8d3eba8864584a.jpg) +Conditioned counterfactual biomass trajectories for three time thresholds +Figure 7: Samples from the Bayesian dynamics based on the fishery model presented by Zhou & Smith (2017), but with the season ending at different times. We show three end dates and their effect on the biomass at the intermediate trophic level. + +Recall that the probability of necessity is $P(Y_{\mathsf{do}(X = 0)} = 0 \mid X = 1, Y = 1)$ (with shorthand $P(y_{x'}' \mid x, y)$ , see table 1). To coherently characterize this in our example, we must define what it means to perform the intervention $\mathsf{do}(X = 0)$ . Given our definition of the predicate $X$ above, $\mathsf{do}(X = 0)$ suggests an intervention that results in a world where the season ends at or after $\lambda$ , with all else (such as the noise or the resulting fishing pressure of 0) remaining equal. Importantly, there are many such worlds, which means the probability of necessity must adopt an existential flavor: "under exogenous noise where $X = 1$ and $Y = 1$ , what is the probability that all worlds consistent with intervention $\mathsf{do}(X = 0)$ fail to meet the outcome?"[27] To precisely define this set of interventional worlds, we build off notation from example 5, and consider a twin world under a static intervention occurring at some time $\lambda' \geqslant \lambda$ , but with the same interventional jump map utilized in the world $\mathcal{P}_{q_1}$ . + +$$ +\left\{\mathcal {P} _ {\lambda^ {\prime}}: \lambda^ {\prime} \geqslant \lambda \right\} \text {w h e r e} \mathcal {P} _ {\lambda^ {\prime}} = \operatorname {s t a t i n t} \left(\mathcal {P} _ {s}, \lambda^ {\prime}, \tilde {G} _ {q _ {1}}\right) \tag {67} +$$ + +As described following definition 18, the trigger dynamics of the state-dependent intervention can also be considered a sort of "treatment mechanism" determining the time at which a static intervention occurs. By constructing a world where direct control over the intervention timing is possible, we are able to disentangle these mechanisms. Importantly, note that while $\mathcal{P}_{\lambda'}$ is constructed via a transformation on $\mathcal{P}_s$ , it is equivalent to a single-season world constructed from an intervention on $\mathcal{P}_{q_1}$ that directly controls the season-ending time independently of causally upstream events in the system's simulation. + +$$ +Y _ {\mathbf {d o} (X = 0)} = 0 \Longleftrightarrow \mathbb {I} \left[ \mathrm {b} _ {\lambda^ {\prime}} ^ {\tau} < \gamma \right] \forall \lambda^ {\prime} \geqslant \lambda \tag {68} +$$ + +Note that if $\mathbf{b}_{\lambda'}^{\tau}$ monotonically decreases as $\lambda' \to \infty$ , then we can equivalently write the event $y_{x'}'$ as $\mathbb{I}[\mathbf{b}_{\lambda'}^{\tau} < \gamma]$ . Indeed, under our model and distributions on $\xi$ and $\theta$ , this is the case, and so we can finally precisely express the probability that ending the season before time $\lambda$ is causally necessary to achieve the biomass outcome: + +$$ +P \left(y _ {x ^ {\prime}} ^ {\prime} \mid x, y\right) = P \left(\mathbb {I} \left[ \mathbf {b} _ {\lambda^ {\prime}} ^ {\tau} < \gamma \right] \mid \mathbb {I} \left[ T \left(\varphi_ {q _ {1}}\right) \leqslant \lambda \right], \mathbb {I} \left[ \mathbf {b} _ {q _ {1}} ^ {\tau} \geqslant \gamma \right]\right) \tag {69} +$$ + +Unlike in example 5, conditioning on the factual interventional event is required, because knowing that the season ended before $\lambda$ carries information about the model parameters: the earlier the TAC quota is met (at times prior to $\lambda$ ), the faster the catch rate. Faster catch rates stem from some combination of higher fishing pressure ( $h_{1:3}$ ), higher initial biomass ( $b_{1:3}$ at $t = 0$ ), higher growth + +![](images/8fb6c17fb89d179c1c49b84203ac7be37d1fc54a8536ad55b27871356b3d0035.jpg) +Figure 8: For each $\lambda_{i}$ from a grid of intervention times we (1) condition on the season ending before $\lambda_{i}$ and on the successful outcome, and (2) we intervene so that the end of the season occurs at $\lambda_{i}$ . The top panel shows the probability that ending the season before $\lambda_{i}$ was necessary to achieve biomass targets, while the bottom panel shows the counterfactual biomass distribution under interventions ending the season at various $\lambda_{i}$ . In the violin plot, we differentiate between counterfactual uncertainty for all worlds (orange and gray), and counterfactual uncertainty after selecting only worlds where the TAC quota was reached before $\lambda_{i}$ and biomass goals were met at $\tau$ . In other words, the gray violins show biomass probabilities if regulators had ended the season at $\lambda_{i}$ in cases where they ended before $\lambda_{i}$ and met biomass goals. The probability of necessity, then, is the proportion of the gray distribution falling below the target level of $\gamma$ . + +rates, etc., all of which influence whether the biomass target will be achieved under alternative season-ending times, even after conditioning on success in the factual world.[28] + +Returning to the example, consider a range over fixed threshold $\lambda$ , approximated by a finite sequence $(\lambda_i)$ . For each $\lambda_i$ , we (1) condition on the season ending before $\lambda_i$ and on the intermediate biomass at $\tau$ being above $\gamma$ , and (2) intervene so that the end of the season occurs at $\lambda_i$ (and not earlier). The relevant probability of necessity query is whether intervening before $\lambda_i$ was necessary for the success. That is, for each $\lambda_i$ we inspect the posterior predictive distribution of the intermediate biomass at $\tau$ under the intervention, and inspect the probability that this outcome is below $\gamma$ . The results of an estimation are available in fig. 8. + +The simulated results presented here ran on a consumer grade laptop in the order of one hour. + +# H Holling-Tanner Fishery Model + +The fishery management model presented by Zhou & Smith (2017) describes the population dynamics for a given trophic level according to the Holling-Tanner model: + +$$ +\frac {d B}{d t} = r B \left(1 - \frac {B}{K}\right) - M B - F B, \tag {70} +$$ + +where $B$ is the biomass of the species, $r$ is the intrinsic growth rate, $K$ is the carrying capacity, $M$ is the mortality rate due to predation, and $F$ is the fishing mortality rate. Elsewhere in the paper, we have avoided using Zhou & Smith's notation, so-as to avoid overloads with the hybrid system literature. In our paper, we use $h$ instead of $F$ , and the lowercase $b$ for biomass, with subscript $i$ indicating trophic level. In this appendix section, however, we will use Zhou & Smith's notation. + +The mortality rate due to predation is modeled as: + +$$ +M = \frac {p B _ {\text {p r e d}}}{D + B}, \tag {71} +$$ + +where $p$ is the maximum predation rate, $B_{\mathrm{pred}}$ is the biomass of the predator, and $D$ is the biomass at which predation reaches half its maximum. + +The carrying capacity for a predator species is given by: + +$$ +K = e B _ {\text {p r e y}}, \tag {72} +$$ + +where $e$ is the efficiency of converting prey biomass into predator biomass. + +The bottom trophic level dynamics follows: + +$$ +\frac {d B _ {\text {f o r a g e}}}{d t} = r _ {1} B _ {\text {f o r a g e}} \left(1 - \frac {B _ {\text {f o r a g e}}}{K _ {1}}\right) - M _ {1 2} B _ {\text {f o r a g e}} - F _ {\text {f o r a g e}} B _ {\text {f o r a g e}}, \tag {73} +$$ + +where $M_{12}$ is the mortality rate due to predation from intermediate predators. + +Species in the intermediate level act as both predator and prey: + +$$ +\frac {d B _ {\text {i n t e r m e d i a t e}}}{d t} = r _ {2} B _ {\text {i n t e r m e d i a t e}} \left(1 - \frac {B _ {\text {i n t e r m e d i a t e}}}{e _ {1 2} B _ {\text {f o r a g e}}}\right) - M _ {2 3} B _ {\text {i n t e r m e d i a t e}} - F _ {\text {i n t e r m e d i a t e}} B _ {\text {i n t e r m e d i a t e}}. \tag {74} +$$ + +The top trophic level follows: + +$$ +\frac {d B _ {\text {a p e x}}}{d t} = r _ {3} B _ {\text {a p e x}} \left(1 - \frac {B _ {\text {a p e x}}}{e _ {2 3} B _ {\text {i n t e r m e d i a t e}}}\right) - M _ {3} B _ {\text {a p e x}} - F _ {\text {a p e x}} B _ {\text {a p e x}}. \tag {75} +$$ + +The catch rate for the intermediate trophic level is given by the following — note that, in the main body of our paper, we use $z$ for the integrated catch, meaning Catch below corresponds to $\dot{z}$ . + +$$ +\text {C a t c h} _ {\text {i n t e r m i d a t e}} = F _ {\text {i n t e r m i d a t e}} B _ {\text {i n t e r m i d a t e}}. \tag {76} +$$ + +Fishing efforts for each trophic level are assumed to remain constant over time unless intervened on. + +$$ +\frac {d F _ {i}}{d t} = 0, \quad i \in \text {f o r a g e}, \text {i n t e r m e d i a t e}, \text {a p e x}. \tag {77} +$$ + +# I Practical Utilities for Tracking Intervention Times and Values + +In many counterfactual estimands, we must translate an event's characteristics from one world to another. To do so, we require the ability to extract certain event properties from a hybrid system's solution. By recording event specifications in auxiliary state variables, these can be straightforwardly read off of solution evaluations at any particular time. First, consider an original system $\mathcal{P}$ with time recorded faithfully in the first dimension, and compatible interventional jump set $\tilde{D}$ and jump map $\tilde{G}$ . Assume the intervention preserves the faithful recording of time and that $\tilde{G}$ is single valued everywhere. + +To start, we augment the state space with an intervention jump counter $j$ , an intervention time $t_k$ , and an intervention value $v_k$ . Our goal is to record the time and jump value corresponding to the $k$ 'th occurrence of the intervention. Let $\tilde{S} = S \times \mathbb{R}_{\geqslant 0}^2 \times \mathbb{R}$ and $\tilde{\Theta} = \varnothing$ , and augment the state space accordingly. + +$$ +\mathcal {P} ^ {\prime} = \operatorname {s p a u g} \left(\mathcal {P}, \tilde {\mathcal {S}}, \tilde {\Theta}\right) \tag {78} +$$ + +Now, augment the original interventional specification to appropriately track event details in these auxiliary state variables. For all admissible inputs, and fixed integer $k$ , let + +$$ +\tilde {D} ^ {\prime} (\boldsymbol {\theta}) = \tilde {D} (\boldsymbol {\theta}) \times \tilde {S} \tag {79} +$$ + +$$ +\tilde {G} ^ {\prime} (\boldsymbol {x}, t, j, t _ {k}, v _ {k}, \boldsymbol {\theta}) = \tilde {G} (\boldsymbol {x}, t, \boldsymbol {\theta}) \times \left\{ \begin{array}{l l} {[ j + 1, t, v ]} & k = j + 1 \\ {[ j + 1, t _ {k}, v _ {k} ]} & k \neq j + 1 \end{array} \right. \tag {80} +$$ + +The intervention time and value can then be read directly off of a solution satisfying the constraints of $\operatorname{inst}(\mathcal{P}',\tilde{D}',\tilde{G}')$ . To track whether an event occurred $k$ times, we can initially set, for example, $t_k = -1$ . A positive $t_k$ would indicate that the event had occurred. This can be switched to $\infty$ instead for more natural time inequalities if boundedness of the states is not a concern. + +# I.1 Notation for Extracting Times and Values + +We now describe some general notation for extracting values of $t_k$ and $v_k$ from a solution. Consider a hybrid arc $\phi_i$ satisfying a system arising from the $i$ 'th instant transformation of that system, where the interventional jump set and map had been augmented to record its $k$ 'th jump. Given the solution's time parameterization $t \mapsto \varphi(t; \xi, \theta)$ , we use $T_i^{(k)}(\varphi_{m \geqslant i}(\cdot; \xi, \theta))$ for the time at which the $k$ 'th jump occurred, and $V_i^{(k)}(\varphi_{m \geqslant i}(\cdot; \xi, \theta))$ to extract the state's value immediately following the jump. When clear from context, the function $V$ extracts only one element of the state. The caveat that $m \geqslant i$ simply specifies that properties of the $k$ 'th jump due to transformation $i$ can be read off of a solution to any further transformed system. As shorthand, we sometimes write $\varphi_{m \geqslant i} = \varphi_{m \geqslant i}(\cdot; \xi, \theta)$ , taking the random inputs as implicit. Additionally, in settings involving interventions that occur only once in the relevant time window, or where the order of interventional transformation is clear and denoted using a symbolic subscript like $s$ , we use, for example, $T_s(\varphi_s)$ to extract the event's time. + +# J State-Dependent Intervention in the Forward Euler Representation + +Consider a forward-Euler approximation of a system of ODEs, where, for simplicity, we will assume that $\Delta t > 0$ . If $f$ is the right-hand side of the continuous-time differential equation $x' = f(x)$ , we can write structural equations $x_{t} = x_{t - \Delta t} + f(x_{t - \Delta t}, u) \Delta t$ , where $t \geqslant 0$ , $x_{t} \in \mathbb{R}$ is the value of the state variable $x$ at time $t$ , $u \in \mathbb{R}$ is a fixed realization of exogenous noise (representing unknown parameters, for example), and $x_{0}$ is fixed to some constant initial condition. Suppose now that we wish to intervene such that the system jumps according to a function $g = x \mapsto x + 1$ when $x$ falls to some threshold $\tau$ . To implement this, we must modify the structural equation for $x_{t}$ for all $t$ under question. That is, we replace the original structural equation with the following piecewise construction. + +Let $\bar{x}_t = x_{t - \Delta t} + f(x_{t - \Delta t},u)\Delta t$ denote the value that would be obtained under the original (nonintervened) Euler update, and let $D = \{x\in \mathbb{R}\mid x\leqslant \tau \}$ denote the domain in which the jump is triggered. Then the intervened structural equation can be approximated as follows: + +$$ +x _ {t} = \left\{ \begin{array}{l l} \bar {x} _ {t} \in D & g (\bar {x} _ {t}) \\ \bar {x} _ {t} \notin D & \bar {x} _ {t}. \end{array} \right. \tag {81} +$$ + +# NeurIPS Paper Checklist + +# 1. Claims + +Question: Do the main claims made in the abstract and introduction accurately reflect the paper's contributions and scope? + +Answer: [Yes] + +Justification: Our claims are described in the abstract and clearly enumerated at the end of the introduction. The first manifests in the definition of an instantaneous intervention (definition 4), the second in theorem 1 and its proof (appendix D), and the third in section 6. + +# 2. Limitations + +Question: Does the paper discuss the limitations of the work performed by the authors? + +Answer: [Yes] + +Justification: A limitations section is provided and includes references to other locations in the paper where we discuss limitations reviewed therein. + +# 3. Theory assumptions and proofs + +Question: For each theoretical result, does the paper provide the full set of assumptions and a complete (and correct) proof? + +Answer: [Yes] + +Justification: Proof sketches are provided in the main body and references are provided to detailed formal proofs in the appendix. + +# 4. Experimental result reproducibility + +Question: Does the paper fully disclose all the information needed to reproduce the main experimental results of the paper to the extent that it affects the main claims and/or conclusions of the paper (regardless of whether the code and data are provided or not)? + +Answer: [NA] + +Justification: We do not include experiments in the main body, but do provide some simulated estimation results in the appendix. We provide ample model details therein for reproducibility and provide links to our simulation code. + +# 5. Open access to data and code + +Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material? + +Answer: [NA] + +Justification: We provide links to our simulation code, which runs in self-contained Jupyter notebooks. + +# 6. Experimental setting/details + +Question: Does the paper specify all the training and test details (e.g., data splits, hyperparameters, how they were chosen, type of optimizer, etc.) necessary to understand the results? + +Answer: [NA] + +Justification: The main paper does not include experiments, but the simulation analyses offered in the supplementary material do have all model parameters and variable distributions clearly enumerated. + +# 7. Experiment statistical significance + +Question: Does the paper report error bars suitably and correctly defined or other appropriate information about the statistical significance of the experiments? + +Answer: [NA] + +Justification: We do not include experiments in the main paper, but our simulation analyses in the appendix do include credible intervals representing Bayesian prior predictive margins. + +# 8. Experiments compute resources + +Question: For each experiment, does the paper provide sufficient information on the computer resources (type of compute workers, memory, time of execution) needed to reproduce the experiments? + +Answer: [NA] + +Justification: The main body does not include experiments, but for our supplementary simulated analyses, we do state our very small computational requirements. + +# 9. Code of ethics + +Question: Does the research conducted in the paper conform, in every respect, with the NeurIPS Code of Ethics https://neurips.cc/public/EthicsGuidelines? + +Answer: [Yes] + +Justification: The research does not raise any ethical concerns and is unrelated to areas covered by the Code of Ethics. + +# 10. Broader impacts + +Question: Does the paper discuss both potential positive societal impacts and negative societal impacts of the work performed? + +Answer: [NA] + +Justification: There is no direct path for societal impact from our work, beyond, of course, pipelines that blindly use potentially incorrect models for high-stakes decision making. + +# 11. Safeguards + +Question: Does the paper describe safeguards that have been put in place for responsible release of data or models that have a high risk for misuse (e.g., pretrained language models, image generators, or scraped datasets)? + +Answer: [NA] + +Justification: See above. + +# 12. Licenses for existing assets + +Question: Are the creators or original owners of assets (e.g., code, data, models), used in the paper, properly credited and are the license and terms of use explicitly mentioned and properly respected? + +Answer: [NA] + +Justification: We do not use such assets. + +# 13. New assets + +Question: Are new assets introduced in the paper well documented and is the documentation provided alongside the assets? + +Answer: [NA] + +Justification: This paper does not release any new assets. + +# 14. Crowdsourcing and research with human subjects + +Question: For crowdsourcing experiments and research with human subjects, does the paper include the full text of instructions given to participants and screenshots, if applicable, as well as details about compensation (if any)? + +Answer: [NA] + +Justification: The paper involves neither crowdsourcing nor research with human subjects. + +# 15. Institutional review board (IRB) approvals or equivalent for research with human subjects + +Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals (or an equivalent approval/review based on the requirements of your country or institution) were obtained? + +Answer: [NA] + +Justification: The paper does not involve crowdsourcing nor research with human subjects. + +# 16. Declaration of LLM usage + +Question: Does the paper describe the usage of LLMs if it is an important, original, or non-standard component of the core methods in this research? Note that if the LLM is used only for writing, editing, or formatting purposes and does not impact the core methodology, scientific rigorousness, or originality of the research, declaration is not required. + +Answer: [NA] + +Justification: The core method development in this research does not involve LLMs as any important, original, or non-standard components. \ No newline at end of file diff --git a/NeurIPS/2025/A Counterfactual Semantics for Hybrid Dynamical Systems/images.zip b/NeurIPS/2025/A Counterfactual Semantics for Hybrid Dynamical Systems/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..d3f716d9b38027b6777f927b5a13d3990434ee59 --- /dev/null +++ b/NeurIPS/2025/A Counterfactual Semantics for Hybrid Dynamical Systems/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:60343d2e012631b405e604e1bbb98cde005684d221c0064e1b0246676fb05330 +size 1172635 diff --git a/NeurIPS/2025/A Counterfactual Semantics for Hybrid Dynamical Systems/layout.json b/NeurIPS/2025/A Counterfactual Semantics for Hybrid Dynamical Systems/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c95f033ad1a886e16037a2db38d5a7b606f889e5 --- /dev/null +++ b/NeurIPS/2025/A Counterfactual Semantics for Hybrid Dynamical Systems/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:455af3db3fd3aea71ffe257d962e4c1662a1e2a8b9c48a6349888598556089d9 +size 2123981 diff --git "a/NeurIPS/2025/A Cram\303\251r\342\200\223von Mises Approach to Incentivizing Truthful Data Sharing/9e290536-257a-4e91-9e40-d96863e735f8_content_list.json" "b/NeurIPS/2025/A Cram\303\251r\342\200\223von Mises Approach to Incentivizing Truthful Data Sharing/9e290536-257a-4e91-9e40-d96863e735f8_content_list.json" new file mode 100644 index 0000000000000000000000000000000000000000..2e194ac28c497420a24902634b8eb30328bf05fa --- /dev/null +++ "b/NeurIPS/2025/A Cram\303\251r\342\200\223von Mises Approach to Incentivizing Truthful Data Sharing/9e290536-257a-4e91-9e40-d96863e735f8_content_list.json" @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cf5c385b2533a90894e07c7691b47f53ab486c1e3d5bfa6a4010d44f342d67fc +size 297363 diff --git "a/NeurIPS/2025/A Cram\303\251r\342\200\223von Mises Approach to Incentivizing Truthful Data Sharing/9e290536-257a-4e91-9e40-d96863e735f8_model.json" "b/NeurIPS/2025/A Cram\303\251r\342\200\223von Mises Approach to Incentivizing Truthful Data Sharing/9e290536-257a-4e91-9e40-d96863e735f8_model.json" new file mode 100644 index 0000000000000000000000000000000000000000..967179625457b58b2a06078fdddd614eb68d92a9 --- /dev/null +++ "b/NeurIPS/2025/A Cram\303\251r\342\200\223von Mises Approach to Incentivizing Truthful Data Sharing/9e290536-257a-4e91-9e40-d96863e735f8_model.json" @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c5615f06c33470922653d243ae18c1e83cf7cfcd89a25b50abca66c80c1ecc76 +size 357386 diff --git "a/NeurIPS/2025/A Cram\303\251r\342\200\223von Mises Approach to Incentivizing Truthful Data Sharing/9e290536-257a-4e91-9e40-d96863e735f8_origin.pdf" "b/NeurIPS/2025/A Cram\303\251r\342\200\223von Mises Approach to Incentivizing Truthful Data Sharing/9e290536-257a-4e91-9e40-d96863e735f8_origin.pdf" new file mode 100644 index 0000000000000000000000000000000000000000..594a485d31dc4a7ecec861f79307360d2d08e53b --- /dev/null +++ "b/NeurIPS/2025/A Cram\303\251r\342\200\223von Mises Approach to Incentivizing Truthful Data Sharing/9e290536-257a-4e91-9e40-d96863e735f8_origin.pdf" @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e16930a501a6aec0e0949558b863b508e55d8d081a5611566ddeafc4a819409c +size 850417 diff --git "a/NeurIPS/2025/A Cram\303\251r\342\200\223von Mises Approach to Incentivizing Truthful Data Sharing/full.md" "b/NeurIPS/2025/A Cram\303\251r\342\200\223von Mises Approach to Incentivizing Truthful Data Sharing/full.md" new file mode 100644 index 0000000000000000000000000000000000000000..590c73913aa48e5a1105851fffe0d137a17bb774 --- /dev/null +++ "b/NeurIPS/2025/A Cram\303\251r\342\200\223von Mises Approach to Incentivizing Truthful Data Sharing/full.md" @@ -0,0 +1,1661 @@ +# A Cramér-von Mises Approach to Incentivizing Truthful Data Sharing + +Alex Clinton + +University of Wisconsin-Madison aclinton@wisc.edu + +Thomas Zeng + +University of Wisconsin-Madison +tpzeng@wisc.edu + +Yiding Chen + +Cornell University +yc2773@cornell.edu + +Xiaojin Zhu + +University of Wisconsin-Madison jerryzhu@cs.wisc.edu + +Kirthevasan Kandasamy + +University of Wisconsin-Madison kandasamy@cs.wisc.edu + +# Abstract + +Modern data marketplaces and data sharing consortia increasingly rely on incentive mechanisms to encourage agents to contribute data. However, schemes that reward agents based on the quantity of submitted data are vulnerable to manipulation, as agents may submit fabricated or low-quality data to inflate their rewards. Prior work has proposed comparing each agent's data against others' to promote honesty: when others contribute genuine data, the best way to minimize discrepancy is to do the same. Yet prior implementations of this idea rely on very strong assumptions about the data distribution (e.g. Gaussian), limiting their applicability. In this work, we develop reward mechanisms based on a novel two-sample test statistic inspired by the Cramér-von Mises statistic. Our methods strictly incentivize agents to submit more genuine data, while disincentivizing data fabrication and other types of untruthful reporting. We establish that truthful reporting constitutes a (possibly approximate) Nash equilibrium in both Bayesian and prior-agnostic settings. We theoretically instantiate our method in three canonical data sharing problems and show that it relaxes key assumptions made by prior work. Empirically, we demonstrate that our mechanism incentivizes truthful data sharing via simulations and on real-world language and image data. + +# 1 Introduction + +Data is invaluable for machine learning (ML). Yet many organizations and individuals lack the capability to collect sufficient data on their own. This has driven the emergence of data marketplaces [1-3]—where consumers purchase data from contributors with money—and consortia [4-6] for data sharing and federated learning—where agents share their own data in return for access to others' data. As such platforms depend critically on data from contributing agents, they incentivize these agents to contribute more data via commensurate rewards: consortia typically grant agents greater access to the pooled data [7, 8], while marketplaces provide correspondingly larger payments [9, 10]. + +However, most existing work implicitly assume that contributors will report data truthfully. In reality, strategic contributors may untruthfully report data to exploit the incentive scheme. As one such example, they may fabricate data—either through naïve random generation or sophisticated ML-based synthesis—to artificially inflate their submissions and maximize their own rewards. In naïve incentive schemes, where rewards scale with the quantity of data, such behavior can flood the system with poor quality data which undermines trust in the platform. + +The central challenge in preventing such strategic misreporting, including fabrication, is that consortia and marketplace operators typically lack ground-truth knowledge about the underlying data + +distribution—if the ground truth were known, the very need for learning and data sharing would be obviated. To address this, prior work has proposed a simple and intuitive idea: compare each agent's data submission against the pooled submissions of other agents. In these mechanisms, when all agents' data come from the same distribution, truthful reporting constitutes a Nash equilibrium. Intuitively, when others contribute genuine data, minimizing the discrepancy between one's own submission and the aggregate submission of others also requires submitting genuine data. + +Despite this promising intuition, prior work has succeeded only under strong assumptions about data distributions [7, 11] and/or narrow models of untruthful behavior [12-14]. Realizing this idea to general data distributions and arbitrary types of strategic misreporting has remained challenging. + +Our contributions. This gap motivates the central premise of our work. We develop a mechanism where agents are rewarded based on a novel loss function that is inspired by two-sample testing. Our loss function, resembling the Cramér-von Mises (CvM) two-sample test statistic [15, 16], is computationally inexpensive, and applies to many different data types, including complex data modalities such as text and images. We design (approximate) Nash equilibria in which agents are incentivized to truthfully report data, without relying on restrictive assumptions about the underlying distribution or strategic behaviors. We theoretically demonstrate the application of our mechanism in three data sharing problems involving purchasing data, and data sharing without money. We empirically demonstrate its usefulness via experiments on synthetic and real world datasets. + +# 1.1 Overview of Contributions + +Model. There are $m$ agents. Each agent $i$ possesses a dataset $X_{i}$ drawn from an unknown distribution $\mathcal{P}$ , and submits $Y_{i}$ , not necessarily truthfully (i.e. $Y_{i} \neq X_{i}$ ). In data-sharing consortia or marketplaces, the goal is to design losses (negative rewards) $L = \{L_{i}\}_{i \in [m]}$ , where agent $i$ is rewarded according to $-L_{i}(\{Y_{j}\}_{j \in [m]})$ , so as to incentivize truthful reporting. A natural and widely adopted approach [7, 11, 10], which we also follow, is to design $L_{i}$ as a function of the form $L_{i}(Y_{i}, Y_{-i})$ , where $Y_{-i} = (Y_{j})_{j \neq i}$ is the pooled submission of all agents except $i$ . A high value of $L_{i}(Y_{i}, Y_{-i})$ suggests that agent $i$ 's data deviates from the rest, which may indicate untruthful behavior when other agents report truthfully (i.e. $Y_{j} = X_{j}$ for all $j \neq i$ ). + +Comparing an agent's submission to the pooled data from others can be naturally viewed as computing a two-sample test statistic—or simply, a two-sample test—between $Y_{i}$ and $Y_{-i}$ [15, 17]. This perspective motivates the design of our loss function. + +Key technical challenges. There are two primary challenges in designing a loss. First, we should ensure that the loss $L$ is truthful: specifically, when $Y_{-i}$ is drawn i.i.d. from $\mathcal{P}$ (i.e. all other agents report truthfully), the optimal strategy for agent $i$ to minimize $L_{i}(Y_{i}, Y_{-i})$ should be to also submit truthfully, i.e. $Y_{i} = X_{i}$ . Without this property, agents may have an incentive to manipulate their submissions to reduce $L_{i}(Y_{i}, Y_{-i})$ . However, many standard two-sample tests—such as Kolmogorov-Smirnov [17, 18], $t$ -test [19], Mann-Whitney [20], and MMD [21]—are not provably truthful. The second challenge is to reward agents for higher quality submissions, i.e. $L_{i}$ should decrease as the quantity of the submitted (truthful) data increases. + +While each challenge is easy to address in isolation, satisfying both simultaneously is far more difficult. For example, a mechanism that rewards agents equally is trivially truthful but offers no incentive to collect more data. Conversely, if losses are tied solely to the quantity of submitted data, the mechanism becomes vulnerable to data fabrication, leaving honest agents worse off. + +A third, less central challenge is ensuring that we have a handle on the distribution of $L_{i}$ to enable its application in data sharing use cases. For instance, penalizing large values of $L_{i}$ requires understanding what constitutes "large" under truthful reporting. Prior work addresses these three challenges only under strong assumptions on $\mathcal{P}$ (e.g. Gaussian [7, 11], Bernoulli [22], restricted class of exponential families [10]), or narrow models of untruthful reporting [12, 13, 22]. + +Our method and results. In §2, we consider a Bayesian setting in which each agent's data is drawn from an unknown distribution $\mathcal{P}$ , itself sampled from a known prior $\Pi$ . We introduce our loss $L$ which is inspired by the Cramér-von Mises (CvM) test. Leveraging this statistic along with user-specified data featurizations, we design a loss in which truthful reporting forms an exact Nash equilibrium (NE). Moreover, we show that $L$ incentivizes the submission of larger datasets—an agent is strictly better off by submitting more truthful data. Our loss is also bounded, and decreases gracefully with the amount of data submitted, making it useful for data sharing applications as we will see in §4. + +However, this approach has two practical limitations. First, specifying a meaningful prior can be difficult, particularly for complex data modalities such as text or images. Second, even with a prior, computing $L$ may be intractable when it requires expensive Bayesian posterior computations. In §3, we address these issues by replacing the above Bayesian version of our loss with a prior-agnostic version that is simpler to compute. We show that this leads to a truthful $\varepsilon$ -approximate NE in both Bayesian and frequentist settings where $\varepsilon$ approaches zero as the amount of data submitted increases. We also show that agents benefit from submitting more data, and that our new loss is also bounded and decreases gracefully with the amount of data submitted. + +Applications. In §4, we theoretically demonstrate how our Bayesian method can be applied to solve three different data sharing problems, some of which have been studied in prior work, while relaxing their technical conditions. The first problem is incentivizing truthful data submissions via payments assuming agents already possess data [10]. The second is the design of a data marketplace where a buyer is willing to pay strategic agents to collect data on her behalf [23]. The third is a federated learning setting where agents wish to share data for ML tasks without the use of money [8]. + +Empirical evaluation. In §5, we empirically evaluate our methods on simulations, and real world image and language experiments. To simulate untruthful behavior, we consider agents who augment their datasets by fabricating samples using simple fitted models, or generative models such as diffusion models and LLMs [24-26]. Our results demonstrate that such untruthful submissions lead to larger losses compared to truthful reporting. This corroborates theoretical results for both methods and demonstrates that the prior-agnostic version is practically useful for real world data sharing. + +# 1.2 Related Work + +There has been growing interest in the incentive structures underlying data sharing, federated learning, and data marketplaces. A central goal in these settings is to incentivize data contributions. However, most prior work do not consider untruthful reporting. When they do, they either impose restrictive distributional assumptions, or limit how contributors may misreport. + +Incentivizing data sharing without truthfulness requirements. A line of work addresses incentivizing data collection in federated learning [27, 8, 28-31, 9]. Other studies focus on incentivizing the sharing of private data [32] or truthful reporting of private data collection costs [14]. All of these works assume agents report data truthfully, and do not encounter the challenges we address here. + +Restricted distributional assumptions. Cai et al. [9] study a principal-agent model where a principal selects measurement locations and compensates agents who exert costly effort to reduce observation noise. Their optimal contract relies on a known effort-to-data-quality function, which may be unknown or nonexistent in practice. Ghosh et al. [22] design a mechanism to purchase binary data under differential privacy, compensating agents for privacy loss. Chen et al. [10] drop the privacy constraint to handle non-binary data, proposing a fixed-budget mechanism that ensures truthful reporting, but requiring the data distribution to have finite support or belong to an exponential family. Other work focuses on incentivizing truthful reporting in Gaussian mean estimation for data sharing [7, 11] and data marketplaces [23]; however, as our experiments show, their approach—based on comparing means of the reported data—does not generalize beyond Gaussian data. + +Restricted untruthful reporting. Falconer et al. [13] propose monetary incentives for data sharing, assuming agents can only fabricate data by duplicating existing entries. Dorner et al. [12] study mean estimation where agents may misreport only by adding a scalar to their true values. + +Peer prediction. The peer prediction literature addresses a challenge similar to ours: eliciting truthful reports without access to ground truth. Prior work [33-36] uses reported signals to cross validate agents' submissions, showing that truthful reporting forms an (approximate) Nash equilibrium. Techniques from [37, 38] have been applied to design payment-based mechanisms for data sharing [10], but these rely on strong assumptions about the data distribution (e.g., exponential families or finite support). It is not clear if these methods generally work when agents may change the number of signals (data points) they have, which is a critical consideration in data sharing use cases where fabrication is possible. More precisely, the mechanism designer does not know how many data points an agent holds, yet must still incentivize truthful reporting. + +Practical applicability. The vast majority of the above works focus on theoretical development, but lack empirical evaluation, with their practicality unclear due to expensive Bayesian computations. In contrast, our prior-agnostic method is simple and performs well on real data. + +Review of the Cramér-von Mises test. We briefly review the Cramér-von Mises (CvM) test [15]. Let $X = \{X_{1},\ldots ,X_{n}\} \stackrel {\mathrm{i.i.d.}}{\sim}F_{1}$ and $Y = \{Y_{1},\dots ,Y_{m}\} \stackrel {\mathrm{i.i.d.}}{\sim}F_{2}$ be samples from $\mathbb{R}$ -valued distributions $F_{1}$ and $F_{2}$ , respectively. Let $F_{X}(t) = \frac{1}{|X|}\sum_{x\in X}1_{\{x\leq t\}}$ and $F_{Y}(t) = \frac{1}{|Y|}\sum_{y\in Y}1_{\{y\leq t\}}$ be the empirical CDFs (ECDFs) of $X$ and $Y$ . Set $Z = (X_{1},\ldots ,X_{n},Y_{1},\ldots ,Y_{m})$ . The two-sample CvM test statistic is then defined below in (1). We have illustrated the CvM test in Fig. 1a. + +$$ +\operatorname {C v M} (X, Y) = \frac {n m}{(n + m) ^ {2}} \sum_ {i = 1} ^ {n + m} \left(F _ {X} \left(Z _ {i}\right) - F _ {Y} \left(Z _ {i}\right)\right) ^ {2}. \tag {1} +$$ + +# 2 A Truthful Mechanism in a Bayesian Setting + +In this section, we design a mechanism to reward agents based on the quality of their submitted data. We begin by specifying our model. To build intuition, we present a simplified single-variable version of our loss (mechanism) in §2.1. We then present the general version of our mechanism in §2.2. + +Setting. There are $m > 2$ agents, where each agent $i \in [m]$ has a dataset $X_{i} = \{X_{i,1},\dots,X_{i,n_{i}}\} \subset \{X_{i,j}\}_{j=1}^{\infty}$ of $n_{i} \in \mathbb{N}$ points. Here $\{X_{i,j}\}_{j=1}^{\infty}$ are drawn i.i.d. from an unknown distribution $\mathcal{P}$ over $\mathcal{X}$ and $X_{i} \in \mathcal{X}^{n_{i}}$ . We refer to $\mathcal{X}$ as the dataspace; examples include the space of images, text, or simply $\mathbb{R}^d$ . In this section, we consider a Bayesian setting where $\mathcal{P}$ is drawn from a publicly known prior II. A mechanism designer wishes to incentivize the agents to report their datasets truthfully by designing losses (negative rewards). + +Let $\mathcal{D} = \bigsqcup_{\ell=0}^{\infty} \mathcal{X}^{\ell}$ be the collection of finite subsets of $\mathcal{X}$ , which forms the space of datasets an agent could possess. A mechanism for this problem is a normal form game which maps the agents' dataset submissions to a vector of losses, i.e. $L \in \{L': \mathcal{D}^m \to \mathbb{R}^m\}$ . Once the mechanism $L$ is published, each agent will submit a dataset $Y_i$ (not necessarily equal to $X_i$ ). An agent's strategy can be viewed as a function $f_i \in \mathcal{F} = \{f: \mathcal{D} \to \mathcal{D} \text{ s.t. } f \text{ is measurable}\}$ which maps their original dataset $X_i$ to $Y_i = f_i(X_i)$ . This allows for strategic data manipulations which may depend on the agent's own dataset. Let $I$ be the identity (truthful) strategy which maps a dataset to itself, i.e. $I(X_i) = X_i$ . + +Agent $i$ 's loss $L_{i}$ is the $i$ 'th output of the mechanism $L$ , and is a function of the strategies $f = \{f_{i}\}_{i\in [m]}$ adopted by other agents and the initial datasets $X = \{X_{1},\ldots ,X_{m}\}$ , and can be written as $L_{i} = L_{i}(\{f_{i}\}_{i\in [m]}) = L_{i}(\{f_{i}(X_{i})\}_{i\in [m]})$ to highlight or suppress these dependencies. + +Requirements. The mechanism designer wishes to design $L$ to satisfy two key properties: + +1. Truthfulness: All agents submitting truthfully $(f_{i} = I)$ , is a Nash equilibrium, that is, + +$$ +\forall i \in [ m ], \forall f _ {i} \in \mathcal {F}, \quad \mathbb {E} \left[ L _ {i} \left(\left\{I \right\} _ {j = 1} ^ {m}\right) \right] \leq \mathbb {E} \left[ L _ {i} \left(f _ {i}, \left\{I \right\} _ {j \neq i}\right) \right]. +$$ + +2. More (data) is (strictly) better (MIB): Let $X_{i}, X_{i}^{\prime}$ be two datasets such that $|X_{i}^{\prime}| > |X_{i}|$ . Then, + +$$ +\mathbb {E} \left[ L _ {i} (I (X _ {i} ^ {\prime}), \{I (X _ {j}) \} _ {j \neq i}) \right] < \mathbb {E} \left[ L _ {i} (\{I (X _ {j}) \} _ {j \in [ m ]}) \right]. +$$ + +Above, the expectation is with respect to the prior $\mathcal{P} \sim \Pi$ , the data $X_{i}, X_{i}^{\prime} \sim \mathcal{P}$ for all $i$ , and any randomness in the agent strategies $f_{i}$ and mechanism $L$ . As discussed in §1.1 under 'Key technical challenges', while satisfying either of these requirements is easy, designing a mechanism which satisfies both simultaneously is significantly more difficult. + +# 2.1 Warm-up when $\mathcal{X} = \mathbb{R}$ + +Algorithm 1 description. To build intuition, we first study the simple one-dimensional case $\mathcal{X} = \mathbb{R}$ . The mechanism works by aggregating all of the submissions $\{Y_i\}_{i=1}^m$ and for each agent $i \in [m]$ , computing a (randomized) loss $L_i$ . To compute $L_i$ , an evaluation point $T_i$ is first randomly sampled from the data submitted by the other agents $Y_{-i}$ . The remaining data $Z_i$ is used to define the empirical CDF $F_{Z_i}$ . The loss $L_i$ is then defined as the squared difference between this ECDF evaluated at $T_i$ , i.e. $F_{Z_i}(T_i)$ , and its conditional expectation given $(X_{i,1}, \ldots, X_{i,|Y_i|}, T_i)$ evaluated at $(Y_{i,1}, \ldots, Y_{i,|Y_i|}, T_i)$ . Finally, the mechanism outputs $L_i \in [0,1]$ as agent $i$ 's loss. + +Design intuition: The conditional expectation $\mathbb{E}\left[F_{Z_i}(T_i) | X_{i,1}, \ldots, X_{i,|Y_i|}, T_i\right]$ can be thought of as the best guess for $F_{Z_i}(T_i)$ having seen $(X_{i,1}, \ldots, X_{i,|Y_i|})$ . Thus, $\mathbb{E}\left[F_{Z_i}(T_i) | X_{i,1} = Y_{i,1}, \ldots, X_{i,|Y_i|} = Y_{i,|Y_i|}, T_i\right]$ can be thought of as the best guess for $F_{Z_i}(T_i)$ + +![](images/c494ec30cc04edbc2005a76b9b880cac0e7fded7547ff3434fd6bbe0f09c8da2.jpg) +(a) The two-sample Cramér-von Mises test + +![](images/5c8537f26140e2723f603a41689f0e96bfcd08d3ef3f0c46a705f373544cb8e2.jpg) +(b) An empirical CDF vs its conditional expectation +Figure 1: Subfigure (a) shows the empirical CDFs (ECDF) for two datasets $X = \{X_{1},\ldots ,X_{n}\}$ , $Y = \{Y_{1},\dots ,Y_{m}\}$ . The gray lines are the differences between the two curves at each point in $(X_{1},\ldots ,X_{n},Y_{1},\ldots ,Y_{m})$ , and are used to calculate the two-sample CvM test in (1). Subfigure (b) replaces $F_{Y}(t)$ with $\mathbb{E}[F_Y(t)|X]$ which can be thought of as the best approximation to $F_{Y}(t)$ based on having seen $X$ . + +Algorithm 1 A single variable Cramér-von Mises style statistic + +1: Input parameters: A prior $\Pi$ over the set of $\mathbb{R}$ -valued distributions. +2: for each agent $i \in [m]$ : +3: $Y_{-i}\gets (Y_{j,\ell})_{j\neq i,\ell \in |Y_j|}.$ +4: Sample $j\sim \mathrm{Unif}(1,\ldots ,|Y_{-i}|)$ and set $T_{i}\gets Y_{-i,j},Z_{i}\gets (Y_{-i,\ell})_{\ell \neq j}.$ +5: Return $L_{i}\gets \left(\mathbb{E}\left[F_{Z_{i}}\left(T_{i}\right)\mid X_{i,1} = Y_{i,1},\ldots ,X_{i,|Y_{i}|} = Y_{i,|Y_{i}|},T_{i}\right] - F_{Z_{i}}\left(T_{i}\right)\right)^{2}$ + +assuming that $(Y_{i,1},\ldots ,Y_{i,|Y_i|})$ is the agent's true data. A visual comparison of $F_{Z_i}(T_i)$ to $\mathbb{E}\left[F_{Z_i}(T_i)|X_{i,1},\ldots ,X_{i,|Y_i|},T_i\right]$ can be seen in Fig. 1b. + +The loss $L_{i}$ defined above is well-posed and computable. As demonstrated in our experiments (with derivations in Appendix E), closed-form expressions for $L_{i}$ can be derived in simple conjugate settings such as Gaussian-Gaussian and Bernoulli-Beta, enabling efficient implementations. For more complex prior distributions, numerical approximations using methods such as MCMC [39] or variational inference [40] can be employed. + +Theoretical results. We now present the theoretical properties of Algorithm 1. To satisfy the MIB condition, we require that the prior $\Pi$ meet a non-degeneracy condition, formalized in Definition 1. Intuitively, this condition ensures that the posterior changes upon observing an additional data point. Examples of degenerate priors include those that select a fixed distribution $\mathcal{P}$ with probability 1, or choose $\mathcal{P}$ to be a degenerate distribution $\delta_x, x \in \mathcal{X}$ with probability 1. In such cases, data sharing is meaningless, as the distribution is either fully known or revealed by a single sample. Thus, it is natural to assume $\Pi$ is non-degenerate, so that additional data remains informative. + +Definition 1. (Degenerate priors): Let $\mathcal{P} \sim \Pi$ and $\{X_i\}_{i=1}^{\infty}, T, Z \stackrel{i.i.d.}{\sim} \mathcal{P}$ . We say that $\Pi$ is degenerate if for some $n \in \mathbb{N}$ , $P(Z \leq T|T, X_1, \ldots, X_n) \stackrel{a.s.}{=} P(Z \leq T|T, X_1, \ldots, X_{n+1})$ . + +Theorem 1 shows that Algorithm 1 satisfies truthfulness for all priors $\Pi$ , and MIB when $\Pi$ is not degenerate. The key idea for truthfulness is that by computing the aforementioned conditional expectation, the mechanism performs, on behalf of agent $i$ , the best possible guess for $F_{Z_i}(T_i)$ just using $Y_{i}$ . Thus, it is in agent $i$ 's best interest if $Y_{i} = X_{i}$ . + +Theorem 1. The mechanism in Algorithm 1 satisfies truthfulness. Moreover, when $\Pi$ is not degenerate, then Algorithm 1 also satisfies MIB. + +While the previous theorem indicates that submitting more data is beneficial for the agent, it does not quantify how an agent's loss decreases as they contribute more data. The following proposition quantifies this by offering bounds on how an agent's expected loss decreases with the amount of data they submit, assuming all agents are truthful. This handle on $\mathbb{E}[L_i]$ , along with the property that $L_{i}\in [0,1]$ , is useful for applying our mechanism to data sharing applications as we will see in §4. + +Algorithm 2 A feature-based Cramér-von Mises style statistic +1: Input parameters: A prior $\Pi$ over the set of $\mathcal{X}$ -valued distributions, feature maps $\{\varphi^k\}_{k=1}^K$ . +2: for each agent $i \in [m]$ : +3: $Y_{-i} \gets (Y_{j,\ell})_{j \neq i, \ell \in |Y_j|}$ . +4: Sample $j \sim \mathrm{Unif}(1, \ldots, |Y_{-i}|)$ and set $T_i \gets Y_{i,j}$ , $Z_i \gets (Y_{-i,\ell})_{\ell \neq j}$ . +5: for each feature $k \in [K]$ : +6: $Z_i^k \gets (\varphi^k(Z_{i,j}))_{j=1}^{|Z_i|}$ , $T_i^k \gets \varphi^k(T_i)$ . +7: $L_i^k \gets (\mathbb{E}[F_{Z_i^k}(T_i^k)|X_{i,1} = Y_{i,1}, \ldots, X_{i,|Y_i|} = Y_{i,|Y_i|}, T_i^k] - F_{Z_i^k}(T_i^k))^2$ . +8: Return $L_i \gets \frac{1}{K} \sum_{k=1}^{K} L_i^k$ . + +Proposition 1. Let $L_{i} \left( \{I\}_{i=1}^{m} \right)$ denote the value of $L_{i}$ when agents are truthful in Algorithm 1. Then, $0 \leq \mathbb{E}\left[L_{i}\left(\{I\}_{i=1}^{m}\right)\right] \leq \frac{1}{4}\left(\frac{1}{|X_{i}|} + \frac{1}{|Z_{i}|}\right)$ . Moreover, when $\Pi$ is a prior over the set of continuous $\mathbb{R}$ -valued distributions, $\frac{1}{6|Z_i|} \leq \mathbb{E}\left[L_i\left(\{I\}_{i=1}^{m}\right)\right] \leq \frac{1}{6}\left(\frac{1}{|X_i|} + \frac{1}{|Z_i|}\right)$ . + +# 2.2 A General Mechanism with Feature Maps + +We now extend our mechanism and to handle data from arbitrary distributions. The key modification is the introduction of feature maps: functions chosen by the mechanism designer that transform general data distributions into $\mathbb{R}$ -valued distributions to apply our mechanism to. + +Feature maps. We define a feature map to be any measurable function $\varphi : \mathcal{X} \to \mathbb{R}$ which maps the data to a single variable distribution. We will see that any collection of feature maps $\{\varphi^k : \mathcal{X} \to \mathbb{R}\}_{k=1}^K$ which map the data to a collection of single variable distributions supports a truthful mechanism. However, some feature maps perform better than others depending on the use case, so we allow the mechanism designer flexibility to select maps. For Euclidean data, coordinate projections may suffice, while for complex data like text or images, embeddings from deep learning models are more appropriate (as used in our experiments in §5). + +Algorithm 2 description. The mechanism designer first specifies a collection of feature maps, $\{\varphi^k\}_{k=1}^K$ based on the publicly known prior $\Pi$ . After this, Algorithm 2 can be viewed as applying Algorithm 1 for each feature $k \in [K]$ , making use of $\varphi^k$ to map general data in $\mathcal{X}$ to $\mathbb{R}$ . + +The following theorem shows that Algorithm 2 is truthful, which is a result of the same arguments made in Theorem 1, now repeated for each feature map. For MIB, we require an analogous condition to the one given in Theorem 1, stating that more data leads to a more informative posterior distribution for at least one of the $K$ features. To state this formally, we first extend Definition 1. + +Definition 2. Let $\mathcal{P} \sim \Pi$ and $\{X_i\}_{i=1}^{\infty}$ , $T, Z \stackrel{i.i.d.}{\sim} \mathcal{P}$ . We say that $\Pi$ is degenerate for feature $k \in [K]$ if for some $n \in \mathbb{N}$ , + +$$ +P \left(\varphi^ {k} (Z) \leq \varphi^ {k} (T) | \varphi^ {k} (T), X _ {1}, \dots , X _ {n}\right) \stackrel {{a. s.}} {{=}} P \left(\varphi^ {k} (Z) \leq \varphi^ {k} (T) | \varphi^ {k} (T), X _ {1}, \dots , X _ {n + 1}\right). +$$ + +Theorem 2. The mechanism in Algorithm 2 satisfies truthfulness. Moreover, if there is a feature $k \in [K]$ , for which $\Pi$ is not degenerate, then Algorithm 2 also satisfies MIB. + +Proposition 9 (Appendix C.2), analogous to Proposition 1, quantifies how $L_{i}$ decreases with data size, which will be useful when using this loss in data sharing applications. Additionally, Proposition 8 (Appendix C.2) gives an explicit relationship for how the expected loss changes when an agent submits an additional data point, depending on the prior and feature maps. This exactly quantifies how much lower an agent's loss is when submitting more data. + +# 3 A Prior Agnostic Mechanism + +While our mechanism in §2 applies broadly in Bayesian settings, it has two practical limitations. First, specifying a meaningful prior can be difficult, especially for complex data like text or images. Second, even with a suitable prior, computing the conditional expectation in line 7 may be intractable due to + +# Algorithm 3 A prior free Cramér-von Mises style statistic + +1: Input parameters: Feature maps $\{\varphi^k\}_{k = 1}^K$ and an augment split map $\psi :\mathbb{N}\to \mathbb{N}:\psi (n) < n - 1$ +2: for each agent $i \in [m]$ : +3: $Y_{-i}\gets (Y_{j,\ell})_{j\neq i,\ell \in |Y_j|}$ +4: Split $Y_{-i}$ into $Y_{-i} = (\{T_i\}, W_i, Z_i)$ s.t. $|W_i| = \psi(|Y_{-i}|)$ . +5: for each feature $k \in [K]$ : +6: $T_{i}^{k}\gets \varphi^{k}\left(T_{i}\right),W_{i}^{k}\gets \left(\varphi^{k}\left(W_{i,\ell}\right)\right)_{\ell = 1}^{|W_{i}|},Z_{i}^{k}\gets \left(\varphi^{k}\left(Z_{i,\ell}\right)\right)_{\ell = 1}^{|Z_{i}|}$ +7: $L_{i}^{k}\gets \left(F_{(Y_{i}^{k},W_{i}^{k})}\left(T_{i}^{k}\right) - F_{Z_{i}^{k}}\left(T_{i}^{k}\right)\right)^{2}.$ +8: Return $L_{i}\gets \frac{1}{K}\sum_{k = 1}^{K}L_{i}^{k}$ + +the cost of Bayesian posterior inference. To address this, we introduce a prior-agnostic variant that is significantly easier to compute. The trade-off is that truthful reporting becomes an $\varepsilon$ -approximate NE, where $\varepsilon$ vanishes as the amount of submitted data grows. + +Changes to Algorithm 2. Thus far, we have only focused on the Bayesian setting, assuming that agents wish to minimize their expected loss $\mathbb{E}_{\mathcal{P}\sim \Pi}\left[\mathbb{E}_{\{X_i\}_{i = 1}^m\sim \mathcal{P}}[L_i]\right]$ . However, this modification also supports a frequentist view where agents wish to minimize their worst case expected loss over a class $\mathcal{C}$ possible distributions, i.e. $\sup_{\mathcal{P}\in \mathcal{C}}\mathbb{E}_{\{X_i\}_{i = 1}^m\sim \mathcal{P}}[L_i]$ . In the frequentist setting, the class $\mathcal{C}$ is the analog of the prior $\Pi$ . As such, our prior agnostic mechanism does not have a prior $\Pi$ as input. + +Algorithm 3 computes each agent's loss as follows: first partition $Y_{-i}$ into three parts, (1) an evaluation point $T_i$ , (2) data to augment agent $i$ 's submission with $W_i$ , and (3) data to compare agent $i$ 's submission against $Z_i$ . The mechanism designer is free to choose how much data to allocate to $W_i$ as given by the map $\psi$ . For each feature $k \in [K]$ , we then obtain $T_i^k$ , $W_i^k$ , and $Z_i^k$ by applying $\varphi^k$ . The main modification of the prior-agnostic mechanism is that the conditional expectation in line 7 of Algorithm 2, $\mathbb{E}[F_{Z_i^k}(T_i^k)|X_i = Y_i, T_i^k]$ , is replaced with $F_{(Y_i^k, W_i^k)}(T_i^k)$ which serves as an easy to compute estimate for $F_{Z_i^k}(T_i^k)$ . Here $F_{(Y_i^k, W_i^k)}$ denotes the ECDF from the combined data of $Y_i^k$ and $W_i^k$ . The reason we allow the mechanism designer the flexibility to supplement $Y_i^k$ with $W_i^k$ is that doing so allows them to decrease the $\varepsilon$ parameter corresponding to truthfulness being an $\varepsilon$ -approximate Nash in the following theorem. A reasonable choice for the size of $W_i$ is to set it so that $|W_i| + |Y_i| = |Z_i|$ . + +Before stating the theorem, we define $\varepsilon$ -approximate truthfulness for a mechanism in both the Bayesian and frequentist paradigms. + +$\varepsilon$ -Approximate Truthfulness: All agents submitting truthfully $(f_{i} = I)$ , is an $\varepsilon$ -approximate Nash equilibrium. In the Bayesian setting this means $\forall i \in [m], \forall f_{i} \in \mathcal{F}$ + +$$ +\underset {\mathcal {P} \sim \Pi} {\mathbb {E}} \left[ \mathbb {E} _ {\{X _ {i} \} _ {i = 1} ^ {m} \sim \mathcal {P}} \left[ L _ {i} (\{I \} _ {j = 1} ^ {m}) \right] \right] \leq \underset {\mathcal {P} \sim \Pi} {\mathbb {E}} \left[ \mathbb {E} _ {\{X _ {i} \} _ {i = 1} ^ {m} \sim \mathcal {P}} \left[ L _ {i} (f _ {i}, \{I \} _ {j \neq i}) \right] \right] + \varepsilon . +$$ + +In the frequentist setting this means $\forall i\in [m],\forall f_i\in \mathcal{F}$ + +$$ +\sup _ {\mathcal {P} \in \mathcal {C}} \mathbb {E} _ {\{X _ {i} \} _ {i = 1} ^ {m} \sim \mathcal {P}} \left[ L _ {i} (\{I \} _ {j = 1} ^ {m}) \right] \leq \sup _ {\mathcal {P} \in \mathcal {C}} \mathbb {E} _ {\{X _ {i} \} _ {i = 1} ^ {m} \sim \mathcal {P}} \left[ L _ {i} (f _ {i}, \{I \} _ {j \neq i}) \right] + \varepsilon . +$$ + +Algorithm 3 requires a similar non-degeneracy condition for MIB. In the Bayesian setting, the same condition given in Theorem 2 suffices. In the frequentist setting, we require that the class of distributions $\mathcal{C}$ is not solely comprised of distributions for which all of the feature map induced distributions are degenerate. The following theorem summarizes the main properties of Algorithm 3. We see that as the total amount of data increases, the approximate truthfulness parameter vanishes provided that the datasets $(X_{i},W_{i})$ and $Z_{i}$ are balanced. + +Theorem 3. The mechanism in Algorithm 3 is $\frac{1}{4}\left(\frac{1}{|X_i| + |W_i|} +\frac{1}{|Z_i|}\right)$ -approximately truthful in both the Bayesian and frequentist settings. Moreover, if there is a feature $k\in [K]$ , for which $\Pi$ is not degenerate, then Algorithm 3 satisfies MIB in the Bayesian setting. If it is not the case that $\mathcal{C}\subseteq \left\{\mathcal{P}\in \mathcal{M}_1(\mathcal{X}):\forall k\in [K],\mathcal{P}\circ (\varphi^k)^{-1}\in \delta_x,x\in \mathbb{R}\right\}$ then Algorithm 3 satisfies MIB in the frequentist setting. + +Proposition 10 (Appendix D) gives, for both the Bayesian and frequentist settings, bounds on how an agent's expected loss decreases with the amount of data they submit, assuming all agents are truthful. Moreover, when the pushforward $\mathcal{P}^k = \mathcal{P} \circ (\varphi^k)^{-1}$ is a.s. continuous $\forall k \in [K]$ , this proposition provides an exact expression for the expected loss in both the Bayesian and frequentist settings. + +# 4 Applications to Data Sharing Problems + +1. A data marketplace for purchasing existing data. Our first problem, studied by Chen et al. [10] is incentivizing agents to truthfully submit data using payments from a fixed budget $B$ in a Bayesian setting. Their mechanism requires the data distribution to have finite support or belong to the exponential family to ensure budget feasibility (payments do not exceed $B$ ) and individual rationality (agents receive non-negative payments). Our method removes these distributional assumptions. + +In this setting, $m$ agents each possess a dataset $X_{i} = \{X_{i,1},\ldots ,X_{i,n_{i}}\}$ with points drawn i.i.d. from an unknown distribution $\mathcal{P}$ in a Bayesian model. A data analyst with budget $B$ wishes to purchase this data. Agents submit datasets $\{Y_{i}\}_{i = 1}^{m}$ in return for payments $\{\pi_i(\{Y_i\}_i)\}_{i = 1}^m$ . Chen et al. [10], building on Kong and Schoenebeck [38], design a truthful mechanism based on log pairwise mutual information, but their payments can be unbounded, violating budget feasibility and individual rationality. We address this using Algorithm 2 to construct bounded payments satisfying truthfulness, individual rationality, and budget feasibility without distributional assumptions. Algorithm 4 (see Appendix A.1) implements this, and Proposition 2 guarantees these properties. + +2. A data marketplace to incentivize data collection at a cost. The second problem, studied by Chen et al. [23], involves designing a data marketplace in which a buyer wishes to pay agents to collect data on her behalf at a cost. They study a Gaussian mean estimation problem in a frequentist setting. We study a simplified Bayesian version without assuming Gaussianity. + +In a data marketplace mechanism, the interaction between the buyer and agents takes place as follows. First, each agent chooses how much data to collect, $n_i \in \mathbb{N}$ , paying a known per-sample cost $c$ , and obtains the dataset $X_i = \{X_{i,1}, \ldots, X_{i,n_i}\}$ with data drawn i.i.d. from an unknown $\mathcal{P} \sim \Pi$ . They submit $Y_i = f_i(X_i)$ to the mechanism, and in return, receive a payment $\pi_i(\{Y_i\}_{i=1}^m)$ charged to the buyer. The buyer derives value $v: \mathbb{Z}_{\geq 0} \to \mathbb{R}_{\geq 0}$ from the total amount of truthful data received. An agent's utility is their expected payment minus collection cost $u_i^a = \mathbb{E}[\pi_i(\{Y_i\}_{i=1}^m)] - cn_i$ , and the buyer's utility, when agents are truthful, is the valuation of the data received minus the expected sum of payments, $u^b = v(\sum_{i=1}^m |Y_i|) - \mathbb{E}[\sum_{i=1}^m \pi_i(\{Y_i\}_{i=1}^m)]$ . + +The goal of a data market mechanism is to incentivize agents to collect and truthfully report data. If not carefully designed, the mechanism may incentivize agents to fabricate data to earn payments without incurring collection costs, undermining market integrity and deterring buyers. To address this, we propose Algorithm 5 (see Appendix A.2), using Algorithm 2, which—unlike Chen et al. [23]—does not assume Gaussianity. Proposition 3 shows that, under a market feasibility condition, the mechanism is incentive compatible for agents and individually rational for buyers. + +3. Federated learning. The third problem is a simple federated learning setting, similar to Karimireddy et al. [8], where agents share data to improve personalized models. Unlike their work, which assumes agents truthfully report collected data, we allow strategic misreporting. + +Each of $m$ agents, possess a dataset $X_{i} = \{X_{i,1},\ldots ,X_{i,n_{i}}\}$ of points drawn i.i.d. in a Bayesian model, and have a valuation function $v_{i}:\mathbb{N}\to \mathbb{R}$ (increasing), quantifying the value of using a given amount of data for their machine learning task. Acting alone, an agent's utility is simply $v_{i}(|X_{i}|)$ . When participating, the federated learning mechanism delopys a subset of the others' data submitted, $Z_{i}$ , for agent $i$ 's task based on the quality of their submission $f_{i}(X_{i})$ . This result in a valuation of $v_{i}(|Z_{i}|)$ when the others are truthful. Thus, an agent's utility when participating is defined as $u_{i} = \mathbb{E}[v_{i}(|Z_{i}|)]$ . We propose Algorithm 6 (see Appendix A.3), based on Algorithm 2, which does not assume truthful reporting. Proposition 4 shows it is truthful and individually rational. + +# 5 Experiments + +Synthetic experiments. We consider two Bayesian models with conjugate priors (beta-Bernoulli and normal-normal) where the calculation of the conditional expectation in line 7 of Algorithm 1 is analytically tractable. In both setups, $\mathcal{X} = \mathbb{R}$ and we will use the method in Algorithm 1. + +![](images/3eb83c7830f5365478b73b03605ed2d81fe8d495c75805b0611c10f11c7d89e1.jpg) +(a) Losses in the beta-bernoulli model + +![](images/01239b82e0f791f205a4c8e60540417a7172cffbbd8f908d5e5a0f317c3700ee.jpg) +(b) Losses in the normal-normal model +Figure 2: (a): Losses when submitting truthfully, adding Bern $(1 / 2)$ samples, and adding Bern $(\tilde{p})$ samples in the beta-Bernoulli experiment. (b): Losses when submitting truthfully and adding fabricated data between adjacent pairs of true data points in the normal-normal experiment. In (b), the CvM bar for fabrication behavior extends to $\approx 1.6$ . Losses for truthful submission in each method and subfigure are normalized to 1 (gray lines); values $< 1$ indicate fabrication improves performance, $>1$ means it worsens. A truthful mechanism should yield losses above 1 for all fabrication behavior. + +Baselines: We compare our mechanism to three standard two-sample tests, used here as losses: (1) the KS-test $\mathrm{KS}(Y_i,Y_{-i}) = \sup_{t\in \mathbb{R}}|F_{Y_i}(t) - F_{Y_{-i}}(t)|$ . (2) The CvM test (the direct version, not our adaptation): $\operatorname {CvM}(Y_i,Y_{-i})$ (see (1)). (3) The mean difference (similar to the $t$ -test): Mean-diff $(Y_{i},Y_{-i}) = \left|\frac{1}{|Y_{i}|}\sum_{y\in Y_{i}}y - \frac{1}{|Y_{-i}|}\sum_{y\in Y_{-i}}y\right|$ , which has been used to incentivize truthful reporting for normal mean estimation in a frequentist settings [7, 23]. + +1) Beta-Bernoulli. Our first model is a beta-Bernoulli Bayesian model with $p \sim \mathrm{Beta}(2, 2)$ and then $X_{i,j}|p \sim \mathrm{Bern}(p)$ i.i.d. We evaluate whether an agent can reduce their loss (increase rewards) by adding fabricated data to their submission. We consider two types of fabrication: (1) adding Bern $(1/2)$ samples and (2) estimating $p$ via $\tilde{p} = \frac{1}{|X_{i,j}|}\sum_{x \in X_{i,j}}x$ then adding Bern $(\tilde{p})$ samples. We compare this to an agent's loss when submitting truthfully, assuming in both cases that other agents are truthful. Fig. 2a shows average losses under Algorithm 1 and the three two-sample tests under truthful and non-truthful reporting. Under Algorithm 1, fabricated data always leads to higher loss, while the baselines yields lower loss under at least one fabrication strategy. Thus, the two-sample tests are susceptible to data fabrication whereas Algorithm 1 is not. Notably, Mean-diff, which is used in [7, 11], fails, showing their methods do not work beyond normal mean estimation settings. 2) Normal-normal. Our second experiment is a normal-normal Bayesian model, where $\mu \sim \mathcal{N}(0,1)$ and then $X_{i,j}|\mu \sim \mathcal{N}(\mu, 1)$ i.i.d. Here, we fabricate data by inserting fake points in between real observations. Fig. 2b presents the results. Truthful reporting yields lower loss under Algorithm 1, CvM, and Mean-diff, while KS gives lower loss for fabrication, revealing its susceptibility. + +Language data. Next, we evaluate our method and the above baselines on language data. For this, we use data from the SQuAD dataset [41], where each data point is a question about an article. We model the environment with $m = 20$ and $m = 100$ agents, where all agents have 2500 and 500 original data points respectively. We fabricate data by prompting Llama 3.2-1B-Instruct [26] to generate fake sentences based on the legitimate sentences that agent 1 has. We fabricate the same number of sentences in the original dataset. Agent 1 then submits the combined dataset, both true and fabricated, to the mechanism. We instantiate Algorithm 3 with feature maps obtained from the feature layer of the DistilBERT [42] encoder model, which corresponds to 768 features. We apply the baselines to the same set of features and take the average. We have provided additional details on the experimental set up and some true and fabricated sentences generated in Appendix B.1. + +The results are presented in Table 1, showing that all methods perform well, obtaining a smaller loss for truthful submission when compared to fabricating. It is worth emphasizing that only our method is provably approximately truthful, and other methods may be susceptible to more sophisticated types of fabrication. + +Image data. We perform a similar experiment on image data using the Oxford Flowers-102 dataset [43] dataset. where each data point is an image of a flower. We model the environment with $m = 5$ and $m = 47$ , where all agents have roughly 1000 and 100 original data points respectively. + +Table 1: An agent's average loss (± the standard error) when reporting sentences truthfully/untruthfully, assuming the others are reporting truthfully. The experiments were run once assuming all agents had 500 sentences, then again assuming all agents had 2500 sentences. In each row the smaller loss is bolded. + +
SentencesMethodAvg. truthful lossAvg. untruthful loss
500Algorithm 30.0003 ± 1.8 · 10-50.0011 ± 5.8 · 10-5
KS-test0.0379 ± 7.6 · 10-40.0524 ± 9.7 · 10-4
CvM-test0.1547 ± 9.2 · 10-30.8598 ± 4.8 · 10-2
Mean diff.0.0043 ± 2.5 · 10-40.0095 ± 3.4 · 10-4
2500Algorithm 30.00003 ± 3.3 · 10-60.0005 ± 7.1 · 10-6
KS-test0.0127 ± 2.4 · 10-40.0309 ± 1.2 · 10-4
CvM-test0.1609 ± 7.1 · 10-33.2760 ± 3.4 · 10-2
Mean diff.0.0015 ± 8.4 · 10-50.0069 ± 5.9 · 10-5
+ +We fabricate data by using Segmind Stable Diffusion-1B [25], a lightweight diffusion model, to generate fake images of flowers based on the legitimate pictures. We fabricate the same number of images that an agent possesses. Algorithm 3 is instantiated with 384 feature maps corresponding to the 384 nodes in the embedding layer of DeIT-small-distilled [44], a small vision transformer. As above, we apply the baselines to the same set of features and take the average. Additional details on the experimental set up can be found in Appendix B.2. + +Table 2 shows that, similar to text, all methods perform well, truthful submission leads to a lower loss compared to the fabrication procedure detailed above. + +Table 2: An agent's average loss ( $\pm$ the standard error) when reporting images truthfully/untruthfully, assuming the others are reporting truthfully. The experiments were run once assuming agent 1 had 100 images, then again assuming agent 1 had 1000 images. The 4,612 images in the test set of [43] were used to represent the data submitted by other agents. In each row the smaller loss is bolded. + +
ImagesMethodAvg. truthful lossAvg. untruthful loss
100Algorithm 30.0015 ± 3.2 · 10-50.0040 ± 1.2 · 10-4
KS-test0.0833 ± 4.2 · 10-40.0993 ± 1.3 · 10-3
CvM-test0.1491 ± 2.6 · 10-30.7730 ± 2.0 · 10-2
Mean diff.0.0462 ± 1.0 · 10-30.0953 ± 1.1 · 10-3
1000Algorithm 30.0002 ± 3.7 · 10-60.0032 ± 2.9 · 10-5
KS-test0.0290 ± 2.1 · 10-40.0738 ± 2.7 · 10-4
CvM-test0.1458 ± 3.5 · 10-34.5478 ± 3.0 · 10-2
Mean diff.0.0157 ± 5.2 · 10-40.0896 ± 3.2 · 10-4
+ +# 6 Conclusion + +We study designing mechanisms that incentivize truthful data submission while rewarding agents for contributing more data. In the Bayesian setting, we propose a mechanism that satisfies these goals under a mild non-degeneracy condition on the prior. We additionally develop a prior-agnostic variant that applies in both Bayesian and frequentist settings. We illustrate the practical utility of our mechanisms by revisiting data sharing problems studied in prior work, relaxing their technical assumptions, and validating our approach through experiments on synthetic and real-world datasets. + +Limitations. The mechanisms in §2 rely on Bayesian posterior computations, which may be computationally expensive for complex priors. We also require specifying feature maps that effectively represent the data. While this offers flexibility for the mechanism designer to select application-specific features, there is no universally optimal way to choose them. + +# Acknowledgments and Disclosure of Funding + +This work was partially supported by NSF grant IIS-2441796. + +# References + +[1] Ads Data Hub. https://developers.google.com/ads-data-hub/marketers. Accessed: 2025-04-24. +[2] Delta Sharing. https://docs.databricks.com/en/data-sharing/index.html. Accessed: 2025-04-24. +[3] AWS Data Transfer Hub. https://aws.amazon.com/solutions/implementations/data-transfer-hub/. Accessed: 2025-04-24. +[4] IBM Data Fabric. https://www.ibm.com/data-fabric, 2024. +[5] DAT Freight and Analytics. URL: www.dat.com/sales-inquiry/freight-market-intelligence-consortium, 2024. Accessed: July 9, 2024. +[6] Snowflake Data Marketplace. https://www.snowflake.com/en/product/features/marketplace/. +[7] Yiding Chen, Jerry Zhu, and Kirthevasan Kandasamy. Mechanism design for collaborative normal mean estimation. Advances in Neural Information Processing Systems, 36:49365-49402, 2023. +[8] Sai Praneeth Karimireddy, Wenshuo Guo, and Michael I Jordan. Mechanisms that incentivize data sharing in federated learning. arXiv preprint arXiv:2207.04557, 2022. +[9] Yang Cai, Constantinos Daskalakis, and Christos Papadimitriou. Optimum statistical estimation with strategic data sources. In Conference on Learning Theory, pages 280-296. PMLR, 2015. +[10] Yiling Chen, Yiheng Shen, and Shuran Zheng. Truthful data acquisition via peer prediction. Advances in Neural Information Processing Systems, 33:18194-18204, 2020. +[11] Alex Clinton, Yiding Chen, Jerry Zhu, and Kirthevasan Kandasamy. Collaborative mean estimation among heterogeneous strategic agents: Individual rationality, fairness, and truthful contribution. In *Forty-second International Conference on Machine Learning*, 2025. +[12] Florian E Dorner, Nikola Konstantinov, Georgi Pashaliev, and Martin Vechev. Incentivizing honesty among competitors in collaborative learning and optimization. Advances in Neural Information Processing Systems, 36:7659-7696, 2023. +[13] Thomas Falconer, Jalal Kazempour, and Pierre Pinson. Towards replication-robust data markets. arXiv preprint arXiv:2310.06000, 2023. +[14] Rachel Cummings, Katrina Ligett, Aaron Roth, Zhiwei Steven Wu, and Juba Ziani. Accuracy for sale: Aggregating data with a variance constraint. In Proceedings of the 2015 conference on innovations in theoretical computer science, pages 317-324, 2015. +[15] Harald Cramér. On the composition of elementary errors. Scandinavian Actuarial Journal, 1: 141-80, 1928. +[16] Richard von Mises. Probability Statistics and Truth, volume 7. Springer-Verlag, 1939. +[17] A. N. Kolmogorov. Sulla Determinazione Empirica di una Legge di Distribuzione. Giornale dell'Istituto Italiano degli Attuari, 4:83-91, 1933. +[18] N. V. Smirnov. On the Estimation of the Discrepancy Between Empirical Curves of Distribution for Two Independent Samples. Bulletin of Moscow University, 2(2):3-16, 1939. +[19] Student. The probable error of a mean. Biometrika, 6(1):1-25, 1908. doi: 10.1093/biomet/6.1.1. + +[20] H. B. Mann and D. R. Whitney. On a test of whether one of two random variables is stochastically larger than the other. The Annals of Mathematical Statistics, 18(1):50-60, 1947. doi: 10.1214/aoms.1177730491. +[21] Arthur Gretton, Karsten M. Borgwardt, Malte J. Rasch, Bernhard Scholkopf, and Alexander Smola. A kernel two-sample test. In Journal of Machine Learning Research, volume 13, pages 723-773, 2012. +[22] Arpita Ghosh, Katrina Ligett, Aaron Roth, and Grant Schoenebeck. Buying private data without verification. In Proceedings of the fifteenth ACM conference on Economics and computation, pages 931-948, 2014. +[23] Keran Chen, Alex Clinton, and Kirthevasan Kandasamy. Incentivizing truthful data contributions in a marketplace for mean estimation. arXiv preprint arXiv:2502.16052, 2025. +[24] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. +[25] Yatharth Gupta, Vishnu V Jaddipal, Harish Prabhala, Sayak Paul, and Patrick Von Platen. Progressive knowledge distillation of stable diffusion xl using layer level loss. arXiv preprint arXiv:2401.02677, 2024. +[26] Aaron Grattaftori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. +[27] Avrim Blum, Nika Haghtalab, Richard Lanas Phillips, and Han Shao. One for one, or all for all: Equilibria and optimality of collaboration in federated learning. In International Conference on Machine Learning, pages 1005–1014. PMLR, 2021. +[28] Yann Fraboni, Richard Vidal, and Marco Lorenzi. Free-rider attacks on model aggregation in federated learning. In International Conference on Artificial Intelligence and Statistics, pages 1846-1854. PMLR, 2021. +[29] Jierui Lin, Min Du, and Jian Liu. Free-riders in federated learning: Attacks and defenses. arXiv preprint arXiv:1911.12560, 2019. +[30] Baihe Huang, Sai Praneeth Karimireddy, and Michael I Jordan. Evaluating and incentivizing diverse data contributions in collaborative learning. arXiv preprint arXiv:2306.05592, 2023. +[31] Yiling Chen, Nicole Immorlica, Brendan Lucier, Vasilis Syrgkanis, and Juba Ziani. Optimal data acquisition for statistical estimation. In Proceedings of the 2018 ACM Conference on Economics and Computation, pages 27-44, 2018. +[32] Alireza Fallah, Ali Makhdoumi, Azarakhsh Malekian, and Asuman Ozdaglar. Optimal and differentially private data acquisition: Central and local mechanisms. Operations Research, 72 (3):1105-1123, 2024. +[33] Nolan Miller, Paul Resnick, and Richard Zeckhauser. Eliciting informative feedback: The peer-prediction method. Management Science, 51(9):1359-1373, 2005. +[34] Drazen Prelec. A bayesian truth serum for subjective data. science, 306(5695):462-466, 2004. +[35] Anirban Dasgupta and Arpita Ghosh. Crowdsourced judgement elicitation with endogenous proficiency. In Proceedings of the 22nd international conference on World Wide Web, pages 319-330, 2013. +[36] Yiling Chen, Shi Feng, and Fang-Yi Yu. Carrot and stick: Eliciting comparison data and beyond. arXiv preprint arXiv:2410.23243, 2024. +[37] Yuqing Kong and Grant Schoenebeck. Water from two rocks: Maximizing the mutual information. In Proceedings of the 2018 ACM Conference on Economics and Computation, pages 177-194, 2018. + +[38] Yuqing Kong and Grant Schoenebeck. An information theoretic framework for designing information elicitation mechanisms that reward truth-telling. ACM Transactions on Economics and Computation (TEAC), 7(1):1-33, 2019. +[39] Walter R Gilks, Sylvia Richardson, and David Spiegelhalter. Markov chain Monte Carlo in practice. CRC press, 1995. +[40] Martin J Wainwright, Michael I Jordan, et al. Graphical models, exponential families, and variational inference. Foundations and Trends® in Machine Learning, 1(1-2):1-305, 2008. +[41] Pranav Rajpurkar, Jian Zhang, Konstantin Lopyrev, and Percy Liang. Squad: 100,000+ questions for machine comprehension of text. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 2383-2392, 2016. +[42] Victor Sanh, Lysandre Debut, Julien Chaumont, and Thomas Wolf. Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter. ArXiv, abs/1910.01108, 2019. +[43] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In 2008 Sixth Indian conference on computer vision, graphics & image processing, pages 722-729. IEEE, 2008. +[44] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention, 2021. +[45] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language technologies, volume 1 (long and short papers), pages 4171–4186, 2019. +[46] Rick Durrett. Probability: theory and examples, volume 49. Cambridge university press, 2019. + +# NeurIPS Paper Checklist + +# 1. Claims + +Question: Do the main claims made in the abstract and introduction accurately reflect the paper's contributions and scope? + +Answer: [Yes] + +Justification: The claims in the abstract and introduction accurately reflect the paper's contributions and scope. + +Guidelines: + +- The answer NA means that the abstract and introduction do not include the claims made in the paper. +- The abstract and/or introduction should clearly state the claims made, including the contributions made in the paper and important assumptions and limitations. A No or NA answer to this question will not be perceived well by the reviewers. +- The claims made should match theoretical and experimental results, and reflect how much the results can be expected to generalize to other settings. +- It is fine to include aspirational goals as motivation as long as it is clear that these goals are not attained by the paper. + +# 2. Limitations + +Question: Does the paper discuss the limitations of the work performed by the authors? + +Answer: [Yes] + +Justification: A discussion of limitations can be found in the conclusion. + +Guidelines: + +- The answer NA means that the paper has no limitation while the answer No means that the paper has limitations, but those are not discussed in the paper. +- The authors are encouraged to create a separate "Limitations" section in their paper. +- The paper should point out any strong assumptions and how robust the results are to violations of these assumptions (e.g., independence assumptions, noiseless settings, model well-specification, asymptotic approximations only holding locally). The authors should reflect on how these assumptions might be violated in practice and what the implications would be. +- The authors should reflect on the scope of the claims made, e.g., if the approach was only tested on a few datasets or with a few runs. In general, empirical results often depend on implicit assumptions, which should be articulated. +- The authors should reflect on the factors that influence the performance of the approach. For example, a facial recognition algorithm may perform poorly when image resolution is low or images are taken in low lighting. Or a speech-to-text system might not be used reliably to provide closed captions for online lectures because it fails to handle technical jargon. +- The authors should discuss the computational efficiency of the proposed algorithms and how they scale with dataset size. +- If applicable, the authors should discuss possible limitations of their approach to address problems of privacy and fairness. +- While the authors might fear that complete honesty about limitations might be used by reviewers as grounds for rejection, a worse outcome might be that reviewers discover limitations that aren't acknowledged in the paper. The authors should use their best judgment and recognize that individual actions in favor of transparency play an important role in developing norms that preserve the integrity of the community. Reviewers will be specifically instructed to not penalize honesty concerning limitations. + +# 3. Theory assumptions and proofs + +Question: For each theoretical result, does the paper provide the full set of assumptions and a complete (and correct) proof? + +Answer: [Yes] + +Justification: Yes, the paper provides complete proofs of all the results provided in the appendix. + +# Guidelines: + +- The answer NA means that the paper does not include theoretical results. +- All the theorems, formulas, and proofs in the paper should be numbered and cross-referenced. +- All assumptions should be clearly stated or referenced in the statement of any theorems. +- The proofs can either appear in the main paper or the supplemental material, but if they appear in the supplemental material, the authors are encouraged to provide a short proof sketch to provide intuition. +- Inversely, any informal proof provided in the core of the paper should be complemented by formal proofs provided in appendix or supplemental material. +- Theorems and Lemmas that the proof relies upon should be properly referenced. + +# 4. Experimental result reproducibility + +Question: Does the paper fully disclose all the information needed to reproduce the main experimental results of the paper to the extent that it affects the main claims and/or conclusions of the paper (regardless of whether the code and data are provided or not)? + +# Answer: [Yes] + +Justification: Experiment details can be found under the experiments section and in the appendix. + +# Guidelines: + +- The answer NA means that the paper does not include experiments. +- If the paper includes experiments, a No answer to this question will not be perceived well by the reviewers: Making the paper reproducible is important, regardless of whether the code and data are provided or not. +- If the contribution is a dataset and/or model, the authors should describe the steps taken to make their results reproducible or verifiable. +- Depending on the contribution, reproducibility can be accomplished in various ways. For example, if the contribution is a novel architecture, describing the architecture fully might suffice, or if the contribution is a specific model and empirical evaluation, it may be necessary to either make it possible for others to replicate the model with the same dataset, or provide access to the model. In general, releasing code and data is often one good way to accomplish this, but reproducibility can also be provided via detailed instructions for how to replicate the results, access to a hosted model (e.g., in the case of a large language model), releasing of a model checkpoint, or other means that are appropriate to the research performed. +- While NeurIPS does not require releasing code, the conference does require all submissions to provide some reasonable avenue for reproducibility, which may depend on the nature of the contribution. For example +(a) If the contribution is primarily a new algorithm, the paper should make it clear how to reproduce that algorithm. +(b) If the contribution is primarily a new model architecture, the paper should describe the architecture clearly and fully. +(c) If the contribution is a new model (e.g., a large language model), then there should either be a way to access this model for reproducing the results or a way to reproduce the model (e.g., with an open-source dataset or instructions for how to construct the dataset). +(d) We recognize that reproducibility may be tricky in some cases, in which case authors are welcome to describe the particular way they provide for reproducibility. In the case of closed-source models, it may be that access to the model is limited in some way (e.g., to registered users), but it should be possible for other researchers to have some path to reproducing or verifying the results. + +# 5. Open access to data and code + +Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material? + +Answer: [Yes] + +Justification: Sufficient code to replicate the experiments is provided. + +# Guidelines: + +- The answer NA means that paper does not include experiments requiring code. +- Please see the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details. +- While we encourage the release of code and data, we understand that this might not be possible, so "No" is an acceptable answer. Papers cannot be rejected simply for not including code, unless this is central to the contribution (e.g., for a new open-source benchmark). +- The instructions should contain the exact command and environment needed to run to reproduce the results. See the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details. +- The authors should provide instructions on data access and preparation, including how to access the raw data, preprocessed data, intermediate data, and generated data, etc. +- The authors should provide scripts to reproduce all experimental results for the new proposed method and baselines. If only a subset of experiments are reproducible, they should state which ones are omitted from the script and why. +- At submission time, to preserve anonymity, the authors should release anonymized versions (if applicable). +- Providing as much information as possible in supplemental material (appended to the paper) is recommended, but including URLs to data and code is permitted. + +# 6. Experimental setting/details + +Question: Does the paper specify all the training and test details (e.g., data splits, hyperparameters, how they were chosen, type of optimizer, etc.) necessary to understand the results? + +Answer: [Yes] + +Justification: Hyperparameter choices are explained in the experiments section and in the appendix. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The experimental setting should be presented in the core of the paper to a level of detail that is necessary to appreciate the results and make sense of them. +- The full details can be provided either with the code, in appendix, or as supplemental material. + +# 7. Experiment statistical significance + +Question: Does the paper report error bars suitably and correctly defined or other appropriate information about the statistical significance of the experiments? + +Answer: [Yes] + +Justification: Appropriate information about the statistical significance of experiments is provided in the appendix. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The authors should answer "Yes" if the results are accompanied by error bars, confidence intervals, or statistical significance tests, at least for the experiments that support the main claims of the paper. + +- The factors of variability that the error bars are capturing should be clearly stated (for example, train/test split, initialization, random drawing of some parameter, or overall run with given experimental conditions). +- The method for calculating the error bars should be explained (closed form formula call to a library function, bootstrap, etc.) +- The assumptions made should be given (e.g., Normally distributed errors). +- It should be clear whether the error bar is the standard deviation or the standard error of the mean. +- It is OK to report 1-sigma error bars, but one should state it. The authors should preferably report a 2-sigma error bar than state that they have a $96\%$ CI, if the hypothesis of Normality of errors is not verified. +- For asymmetric distributions, the authors should be careful not to show in tables or figures symmetric error bars that would yield results that are out of range (e.g. negative error rates). +- If error bars are reported in tables or plots, The authors should explain in the text how they were calculated and reference the corresponding figures or tables in the text. + +# 8. Experiments compute resources + +Question: For each experiment, does the paper provide sufficient information on the computer resources (type of compute workers, memory, time of execution) needed to reproduce the experiments? + +Answer: [Yes] + +Justification: The appendix provides such sufficient information. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The paper should indicate the type of compute workers CPU or GPU, internal cluster, or cloud provider, including relevant memory and storage. +- The paper should provide the amount of compute required for each of the individual experimental runs as well as estimate the total compute. +- The paper should disclose whether the full research project required more compute than the experiments reported in the paper (e.g., preliminary or failed experiments that didn't make it into the paper). + +# 9. Code of ethics + +Question: Does the research conducted in the paper conform, in every respect, with the NeurIPS Code of Ethics https://neurips.cc/public/EthicsGuidelines? + +Answer: [Yes] + +Justification: Yes, the research conducted conforms with the code of ethics. + +Guidelines: + +The answer NA means that the authors have not reviewed the NeurIPS Code of Ethics +- If the authors answer No, they should explain the special circumstances that require a deviation from the Code of Ethics. +- The authors should make sure to preserve anonymity (e.g., if there is a special consideration due to laws or regulations in their jurisdiction). + +# 10. Broader impacts + +Question: Does the paper discuss both potential positive societal impacts and negative societal impacts of the work performed? + +Answer: [NA] + +Justification: Our paper is a theory paper and does not have direct societal impact. + +Guidelines: + +- The answer NA means that there is no societal impact of the work performed. +- If the authors answer NA or No, they should explain why their work has no societal impact or why the paper does not address societal impact. + +- Examples of negative societal impacts include potential malicious or unintended uses (e.g., disinformation, generating fake profiles, surveillance), fairness considerations (e.g., deployment of technologies that could make decisions that unfairly impact specific groups), privacy considerations, and security considerations. +- The conference expects that many papers will be foundational research and not tied to particular applications, let alone deployments. However, if there is a direct path to any negative applications, the authors should point it out. For example, it is legitimate to point out that an improvement in the quality of generative models could be used to generate deepfakes for disinformation. On the other hand, it is not needed to point out that a generic algorithm for optimizing neural networks could enable people to train models that generate Deepfakes faster. +- The authors should consider possible harms that could arise when the technology is being used as intended and functioning correctly, harms that could arise when the technology is being used as intended but gives incorrect results, and harms following from (intentional or unintentional) misuse of the technology. +- If there are negative societal impacts, the authors could also discuss possible mitigation strategies (e.g., gated release of models, providing defenses in addition to attacks, mechanisms for monitoring misuse, mechanisms to monitor how a system learns from feedback over time, improving the efficiency and accessibility of ML). + +# 11. Safeguards + +Question: Does the paper describe safeguards that have been put in place for responsible release of data or models that have a high risk for misuse (e.g., pretrained language models, image generators, or scraped datasets)? + +Answer: [NA] + +Justification: The paper poses no such risks. + +# Guidelines: + +- The answer NA means that the paper poses no such risks. +- Released models that have a high risk for misuse or dual-use should be released with necessary safeguards to allow for controlled use of the model, for example by requiring that users adhere to usage guidelines or restrictions to access the model or implementing safety filters. +- Datasets that have been scraped from the Internet could pose safety risks. The authors should describe how they avoided releasing unsafe images. +- We recognize that providing effective safeguards is challenging, and many papers do not require this, but we encourage authors to take this into account and make a best faith effort. + +# 12. Licenses for existing assets + +Question: Are the creators or original owners of assets (e.g., code, data, models), used in the paper, properly credited and are the license and terms of use explicitly mentioned and properly respected? + +Answer: [Yes] + +Justification: Assets are properly cited. + +# Guidelines: + +- The answer NA means that the paper does not use existing assets. +- The authors should cite the original paper that produced the code package or dataset. +- The authors should state which version of the asset is used and, if possible, include a URI. +- The name of the license (e.g., CC-BY 4.0) should be included for each asset. +- For scraped data from a particular source (e.g., website), the copyright and terms of service of that source should be provided. +- If assets are released, the license, copyright information, and terms of use in the package should be provided. For popular datasets, paperswithcode.com/datasets has curated licenses for some datasets. Their licensing guide can help determine the license of a dataset. + +- For existing datasets that are re-packaged, both the original license and the license of the derived asset (if it has changed) should be provided. +- If this information is not available online, the authors are encouraged to reach out to the asset's creators. + +# 13. New assets + +Question: Are new assets introduced in the paper well documented and is the documentation provided alongside the assets? + +Answer: [NA] + +Justification: No new assets are released. + +Guidelines: + +- The answer NA means that the paper does not release new assets. +- Researchers should communicate the details of the dataset/code/model as part of their submissions via structured templates. This includes details about training, license, limitations, etc. +- The paper should discuss whether and how consent was obtained from people whose asset is used. +- At submission time, remember to anonymize your assets (if applicable). You can either create an anonymized URL or include an anonymized zip file. + +# 14. Crowdsourcing and research with human subjects + +Question: For crowdsourcing experiments and research with human subjects, does the paper include the full text of instructions given to participants and screenshots, if applicable, as well as details about compensation (if any)? + +Answer: [NA] + +Justification: The paper does not involve crowdsourcing nor research with human subjects. + +Guidelines: + +- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. +- Including this information in the supplemental material is fine, but if the main contribution of the paper involves human subjects, then as much detail as possible should be included in the main paper. +- According to the NeurIPS Code of Ethics, workers involved in data collection, curation, or other labor should be paid at least the minimum wage in the country of the data collector. + +# 15. Institutional review board (IRB) approvals or equivalent for research with human subjects + +Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals (or an equivalent approval/review based on the requirements of your country or institution) were obtained? + +Answer: [NA] + +Justification: The paper does not involve crowdsourcing nor research with human subjects. + +Guidelines: + +- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. +- Depending on the country in which research is conducted, IRB approval (or equivalent) may be required for any human subjects research. If you obtained IRB approval, you should clearly state this in the paper. +- We recognize that the procedures for this may vary significantly between institutions and locations, and we expect authors to adhere to the NeurIPS Code of Ethics and the guidelines for their institution. +- For initial submissions, do not include any information that would break anonymity (if applicable), such as the institution conducting the review. + +# 16. Declaration of LLM usage + +Question: Does the paper describe the usage of LLMs if it is an important, original, or non-standard component of the core methods in this research? Note that if the LLM is used only for writing, editing, or formatting purposes and does not impact the core methodology, scientific rigorousness, or originality of the research, declaration is not required. + +Answer: [NA] + +Justification: The core contributions of this paper do not use LLMs. + +Guidelines: + +- The answer NA means that the core method development in this research does not involve LLMs as any important, original, or non-standard components. +- Please refer to our LLM policy (https://neurips.cc/Conferences/2025/LLM) for what should or should not be described. + +# A Omitted application algorithms + +# A.1 A data marketplace for purchasing existing data + +Recall the problem setup from §4. Below we provide a short algorithm that incentivizes agents to truthfully report their data, $\{X_{i}\}_{i = 1}^{m}$ , using payments. The idea is to use Algorithm 2 to quantify the quality of an agent's submission and, based on it, determine what fraction of the budget to pay them. + +Definition 3. We say an algorithm is budget feasible if the sum of the payments never exceeds the budget $(\sum_{i=1}^{m} \pi_i \leq B)$ , and individually rational (for participants) if the payments are always nonnegative ( $\forall i \in [m], \pi_i \geq 0$ ). + +# Algorithm 4 A data marketplace for purchasing existing data + +1: Input parameters: A prior $\Pi$ over the set of $\mathcal{X}$ -valued distributions, feature maps $\{\varphi^k\}_{k=1}^K$ . +2:Receive datasets $Y_{1},\ldots ,Y_{m}$ from the agents. +3: Execute Algorithm 2 with $\{Y_i\}_{i=1}^m$ , $\Pi$ , $\{\varphi^k\}_{k=1}^K$ , to obtain the loss $L_i \in [0,1]$ for agent $i$ . +4: Pay agent $i$ : $\pi_i\left(\{Y_i\}_{i=1}^m\right) = \frac{B}{m}\left(1 - L_i\left(\{Y_i\}_{i=1}^m\right)\right)$ . + +# Proposition 2. Algorithm 4 is truthful, individually rational, and budget feasibility. + +Proof. Since $L_{i} \in [0,1]$ , we have $0 \leq \pi_{i} \leq \frac{B}{m}$ , so it immediately follows that Algorithm 4 is both individually rational for the agents and budget feasible. For truthfulness, notice that for any $f_{i} \in \mathcal{F}$ we can appeal to Theorem 2 to get + +$$ +\begin{array}{l} \mathbb {E} \left[ \pi_ {i} (f _ {i}, \{I \} _ {j \neq i}) \right] = \frac {B}{m} (1 - \mathbb {E} [ L _ {i} (f _ {i}, \{I \} _ {j \neq i}) ]) \\ \leq \frac {B}{m} \left(1 - \mathbb {E} \left[ L _ {i} \left(\{I \} _ {i = 1} ^ {m}\right) \right]\right) \\ = \mathbb {E} \left[ \pi_ {i} \left(\{I \} _ {i = 1} ^ {m}\right) \right]. \\ \end{array} +$$ + +Therefore, Algorithm 4 is also truthful, as agents maximize their expected payment when submitting truthfully. + +# A.2 A data marketplace to incentivize data collection at a cost + +Recall the problem setup from §4, which is a simplified version of the problem studied by [23]. Our setting does not subsume [23], as they allow for agents to have varying collection costs, study a frequentist setting (whereas we consider a Bayesian setting), and derive payments that are easy to compute. We now motivate a solution to our simplified setting. + +To facilitate data sharing between a buyer and agents, a mechanism must first determine how much data agents should be asked to collect based on the cost of data collection $c$ , and the buyer's valuation function $v$ . To do this, suppose that the buyer could collect data himself. In this case, he would choose to collect $n^{\mathrm{OPT}} := \operatorname{argmax}_{n \in \mathbb{N}} (v(n) - cn)$ points to maximize his utility. However, as he cannot, when there are $m$ agents, the mechanism will ask each of them to collect $\frac{n^{\mathrm{OPT}}}{m}$ points on his behalf in exchange for payments. + +An important detail is that for the marketplace to be feasible, an agent's expected payment must outweigh the cost of data collection. This requirement is reflected in the first technical condition in Proposition 3, which at a high level says that the change in an agents expected payment with respect to $n_i$ , when collecting $\frac{n^{\mathrm{OPT}}}{m}$ points, is at least $c$ . This can be thought of requiring that the derivative with respect to $n_i$ , of the expected payment at $\frac{n^{\mathrm{OPT}}}{m}$ , be at least $c$ . We also assume that $\Pi$ is not degenerate for all the features and there are diminishing returns for collecting and submitting more data under Algorithm 2. + +When these condition holds, Proposition 3 shows that it is individually rational for a buyer to participate in the marketplace, and in agents' best interest to collect $\frac{n^{\mathrm{OPT}}}{m}$ points and submit them truthfully. + +The idea of Algorithm 5 is to determine what fraction of $\frac{v(n^{\mathrm{OPT}})}{m}$ to pay agent $i$ based on the quality of her submission, as measured by $L_{i}$ . + +# Algorithm 5 A data marketplace to incentivize data collection at a cost + +1: Input parameters: A prior $\Pi$ over the set of $\mathcal{X}$ -valued distributions, feature maps $\{\varphi^k\}_{k=1}^K$ . +2:Receive datasets $Y_{1},\ldots ,Y_{m}$ from the agents. +3: Execute Algorithm 2 with $\{Y_i\}_{i=1}^m$ , $\Pi$ , $\{\varphi^k\}_{k=1}^K$ , to obtain the loss $L_i \in [0,1]$ for agent $i$ . +4: Pay agent $i$ : $\pi_i\left(\{Y_i\}_{i=1}^m\right) = \frac{v(n^{\mathrm{OPT}})}{m}\left(1 - \alpha L_i\right)$ where $\alpha$ is given in Definition 4. +5: Charge the buyer: $p\left(\left\{Y_{i}\right\}_{i = 1}^{m}\right) = \sum_{i = 1}^{m}\pi_{i}\left(\left\{Y_{i}\right\}_{i = 1}^{m}\right).$ + +Definition 4. For Algorithm 5 we introduce notation for the change in an agent's expected payment when collecting and submitting one more data point truthfully, assuming others are truthful: + +$$ +\frac {\partial}{\partial n _ {i}} \mathbb {E} \left[ L _ {i} \left(n _ {i}, n _ {- i}\right) \right] := \mathbb {E} \left[ L _ {i} \left(\left(n _ {i} + 1, n _ {- i}\right), \{I \} _ {i = 1} ^ {m}\right) \right] - \mathbb {E} \left[ L _ {i} \left(\left(n _ {i}, n _ {- i}\right), \{I \} _ {i = 1} ^ {m}\right) \right]. +$$ + +When $\Pi$ is not degenerate $\forall k\in [K]$ , $\frac{\partial}{\partial n_i}\mathbb{E}\left[L_i\left(\left\{\frac{n^{\mathrm{OPT}}}{m}\right\}_{i = 1}^{m}\right)\right] < 0$ (by Theorem 2) and we define + +$$ +\alpha := - \frac {c m}{\frac {\partial}{\partial n _ {i}} \mathbb {E} \left[ L _ {i} \left(\left\{\frac {n ^ {\mathrm {O P T}}}{m} \right\} _ {i = 1} ^ {m}\right) \right] v \left(n ^ {\mathrm {O P T}}\right)}. +$$ + +Proposition 3. Suppose that the following technical conditions are satisfied in Algorithm 5: + +$$ +\frac {v \left(n ^ {\mathrm {O P T}}\right)}{m} \left(- \frac {\partial}{\partial n _ {i}} \mathbb {E} \left[ L _ {i} \left(\left\{\frac {n ^ {\mathrm {O P T}}}{m} \right\} _ {i = 1} ^ {m}\right) \right]\right) \geq c, +$$ + +$\Pi$ is not degenerate $\forall k \in [K]$ , and $-\frac{\partial}{\partial n_i} \mathbb{E}\left[L_i(n_i, n_{-i})\right]$ is decreasing in $n_i$ . + +Then, the strategy profile $\left\{\left(\frac{n^{\mathrm{OPT}}}{m},I\right)\right\}_{i = 1}^{m}$ is individually rational for the buyer, i.e. + +$$ +u ^ {b} \left(\left\{\left(\frac {n ^ {\mathrm {O P T}}}{m}, I\right) \right\} _ {i = 1} ^ {m}\right) \geq 0 +$$ + +and incentive compatible for the agents, i.e. for any $n_i \in \mathbb{N}$ , $f_i \in \mathcal{F}$ + +$$ +u _ {i} ^ {a} \left(\left(n _ {i}, f _ {i}\right), \left\{\left(\frac {n ^ {\mathrm {O P T}}}{m}, I\right) \right\} _ {j \neq i}\right) \leq u _ {i} ^ {a} \left(\left\{\left(\frac {n ^ {\mathrm {O P T}}}{m}, I\right) \right\} _ {i = 1} ^ {m}\right). +$$ + +Proof. We start with individual rationality for the buyer. Notice that if the inequality holds then we have + +$$ +\alpha = \frac {c m}{- \frac {\partial}{\partial n _ {i}} \mathbb {E} \left[ L _ {i} \left(\left\{\frac {n ^ {\mathrm {O P T}}}{m} \right\} _ {i = 1} ^ {m}\right) \right] v (n ^ {\mathrm {O P T}})} \leq \frac {c m}{\frac {c m}{v (n ^ {\mathrm {O P T}})} v (n ^ {\mathrm {O P T}})} = 1 +$$ + +so $\alpha \in (0,1]$ . Since $L_{i}\in [0,1]$ , this implies that + +$$ +\pi_ {i} \left(\left\{Y _ {i} \right\} _ {i = 1} ^ {m}\right) = \frac {v \left(n ^ {\mathrm {O P T}}\right)}{m} (1 - \alpha L _ {i}) \leq \frac {v \left(n ^ {\mathrm {O P T}}\right)}{m} +$$ + +so summing over the payments to all agents we find + +$$ +p \left(\left\{Y _ {i} \right\} _ {i = 1} ^ {m}\right) = \sum_ {i = 1} ^ {m} \pi_ {i} \left(\left\{Y _ {i} \right\} _ {i = 1} ^ {m}\right) \leq v \left(n ^ {\mathrm {O P T}}\right). +$$ + +Therefore, the strategy profile $\left\{\left(\frac{n^{\mathrm{OPT}}}{m},I\right)\right\}_{i = 1}^{m}$ is individually rational for the buyer since + +$$ +u ^ {b} \left(\left\{\left(\frac {n ^ {\mathrm {O P T}}}{m}, I\right) \right\} _ {i = 1} ^ {m}\right) = v (n ^ {\mathrm {O P T}}) - \mathbb {E} \left[ p \left(\{Y _ {i} \} _ {i = 1} ^ {m}\right) \right] \geq 0. +$$ + +We now prove incentive compatibility for the agents in two parts. First we show that regardless of how much data an agent has collected, it is best for her to submit it truthfully when others follow the recommended strategy profile $\left\{\left(\frac{n^{\mathrm{OPT}}}{m},I\right)\right\}_{j\neq i}$ . Second, we show that $\frac{v(n^{\mathrm{OPT}})}{m}$ is the optimal amount of data to collect based on our choice of $\alpha$ . + +Fix $n_i$ . Unpacking the definition of an agent's utility and applying Theorem 2 we have + +$$ +\begin{array}{l} u _ {i} ^ {a} \left(\left(n _ {i}, f _ {i}\right), \left\{\left(\frac {n ^ {\mathrm {O P T}}}{m}, I\right) \right\} _ {j \neq i}\right) \\ = \mathbb {E} \left[ \pi_ {i} \left(\left(n _ {i}, f _ {i}\right), \left\{\left(\frac {n ^ {\mathrm {O P T}}}{m}, I\right) \right\} _ {j \neq i}\right) \right] - c n _ {i} \\ = \frac {v \left(n ^ {\mathrm {O P T}}\right)}{m} \left(1 - \alpha \mathbb {E} \left[ L _ {i} \left(\left(n _ {i}, f _ {i}\right), \left\{\left(\frac {n ^ {\mathrm {O P T}}}{m}, I\right) \right\} _ {j \neq i}\right) \right]\right) - c n _ {i} \\ \leq \frac {v (n ^ {\mathrm {O P T}})}{m} \left(1 - \alpha \mathbb {E} \left[ L _ {i} \left((n _ {i}, I), \left\{\left(\frac {n ^ {\mathrm {O P T}}}{m}, I\right) \right\} _ {j \neq i}\right) \right]\right) - c n _ {i} \\ = \mathbb {E} \left[ \pi_ {i} \left(\left(n _ {i}, I\right), \left\{\left(\frac {n ^ {\mathrm {O P T}}}{m}, I\right) \right\} _ {j \neq i}\right) \right] - c n _ {i} \\ = u _ {i} ^ {a} \left(\left(n _ {i}, I\right), \left\{\left(\frac {n ^ {\mathrm {O P T}}}{m}, I\right) \right\} _ {j \neq i}\right). \\ \end{array} +$$ + +This means that regardless of how much data agent $i$ collects, it is best for them to submit it truthfully. For the second part we now assume $\{f_i\}_{i=1}^m = \{I\}_{i=1}^m$ and $n_{-i} = \left\{\frac{n^{\mathrm{OPT}}}{m}\right\}_{i=1}^m$ so for convenience we omit writing the dependence on these parts of the strategy profile for random variables. + +Notice that since $u_{i}^{a}(n_{i})$ is concave, the optimal amount of data for agent $i$ to collect and submit is the smallest $n_i \in \mathbb{N}$ such that + +$$ +u _ {i} ^ {a} \left(n _ {i} + 1\right) \leq u _ {i} ^ {a} \left(n _ {i}\right) +$$ + +i.e. the point at which the marginal increase in payment no longer offsets the collection cost of an additional point. By the definition of agent utilities and our choice of $\alpha$ we see + +$$ +\begin{array}{l} u _ {i} ^ {a} (n _ {i} + 1) \leq u _ {i} ^ {a} (n _ {i}) \iff \left(\frac {v (n ^ {\mathrm {O P T}})}{m} (1 - \alpha \mathbb {E} [ L _ {i} (n _ {i} + 1) ]) - c (n _ {i} + 1)\right) \\ \leq \left(\frac {v (n ^ {\mathrm {O P T}})}{m} \left(1 - \alpha \mathbb {E} [ L _ {i} (n _ {i}) ]\right) - c (n _ {i})\right) \\ \Longleftrightarrow - \frac {\partial}{\partial n _ {i}} \mathbb {E} [ L _ {i} (n _ {i}) ] \leq \frac {c m}{\alpha v (n ^ {\mathrm {O P T}})} \\ \Longleftrightarrow - \frac {\partial}{\partial n _ {i}} \mathbb {E} [ L _ {i} (n _ {i}) ] \leq - \frac {\partial}{\partial n _ {i}} \mathbb {E} \left[ L _ {i} \left(\frac {n ^ {\mathrm {O P T}}}{m}\right) \right]. \\ \end{array} +$$ + +This implies that $\frac{n^{\mathrm{OPT}}}{m}$ is the optimal amount of data to collect since $-\frac{\partial}{\partial n_i}\mathbb{E}\left[L_i\left(n_i,n_{-i}\right)\right]$ is decreasing in $n_i$ . Putting both parts together we find that for any $n_i\in \mathbb{N}$ , $f_{i}\in \mathcal{F}$ , + +$$ +\begin{array}{l} u _ {i} ^ {a} \left(\left(n _ {i}, f _ {i}\right), \left\{\left(\frac {n ^ {\mathrm {O P T}}}{m}, I\right) \right\} _ {j \neq i}\right) \leq u _ {i} ^ {a} \left(\left(n _ {i}, I\right), \left\{\left(\frac {n ^ {\mathrm {O P T}}}{m}, I\right) \right\} _ {j \neq i}\right) \\ \leq u _ {i} ^ {a} \left(\left\{\left(\frac {n ^ {\mathrm {O P T}}}{m}, I\right) \right\} _ {i = 1} ^ {m}\right) \\ \end{array} +$$ + +so we have incentive compatibility for the agents. + +# A.3 Federated learning + +Recall the problem setup from §4. For convenience we assume that $\forall i\in [m],|X_i| < \sum_{j\neq i}|X_j|$ . + +The idea of Algorithm 6 is to determine how much of the others' data agent $i$ should receive for her task based on the quality of her submission, as measured by $L_{i}$ . + +# Algorithm 6 Federated learning + +1: Input parameters: A prior $\Pi$ over the set of $\mathcal{X}$ -valued distributions, feature maps $\{\varphi^k\}_{k=1}^K$ . +2:Receive datasets $Y_{1},\ldots ,Y_{m}$ from the agents. +3: Execute Algorithm 2 with $\{Y_i\}_{i=1}^m$ $\Pi$ , $\{\varphi^k\}_{k=1}^K$ , to obtain the loss $L_i \in [0,1]$ for agent $i$ . +4: for each agent $i \in [m]$ : + +$$ +5: \quad T _ {i} \leftarrow v _ {i} \left(\left| Y _ {- i} \right|\right), \quad \alpha \leftarrow \left(\frac {1}{2} - \frac {v _ {i} \left(\left| X _ {i} \right|\right)}{2 T _ {i}}\right) \frac {1}{\mathbb {E} \left[ L _ {i} \left(\left\{I \right\} _ {i = 1} ^ {m}\right) \right]} +$$ + +$$ +z _ {i} \leftarrow v _ {i} ^ {- 1} \left(\left(1 - \alpha L _ {i}\right) T _ {i}\right) +$$ + +7: Deploy $Z_{i}$ , a random subset of $Y_{-i}$ of size $z_{i}$ for agent $i$ 's machine learning task. + +In Algorithm 6 and Proposition 4 we assume that $\mathbb{E}\left[L_i\left(\{I\}_{i = 1}^m\right)\right] > 0$ which ensures $\alpha$ is well defined by rulling out trivial data sharing problems. + +Proposition 4. Suppose that $\forall i\in [m],|X_i| < \sum_{j\neq i}|X_j|$ . Then Algorithm 6 is truthful and individually rational. + +Proof. Fix $f_{i} \in \mathcal{F}$ . Unpacking the definition of an agent's utility and applying Theorem 2, we have + +$$ +\begin{array}{l} u _ {i} \left(f _ {i}, \{I \} _ {j \neq i}\right) = \mathbb {E} \left[ v _ {i} \left(z _ {i} \left(f _ {i}, \{I \} _ {j \neq i}\right)\right) \right] \\ = \mathbb {E} \left[ v _ {i} \left(v _ {i} ^ {- 1} \left(\left(1 - \alpha L _ {i} \left(f _ {i}, \{I \} _ {j \neq i}\right)\right) T _ {i}\right)\right) \right] \\ = T _ {i} - T _ {i} \alpha \mathbb {E} \left[ L _ {i} \left(f _ {i}, \{I \} _ {j \neq i}\right) \right] \\ \leq T _ {i} - T _ {i} \alpha \mathbb {E} \left[ L _ {i} \left(\{I \} _ {i = 1} ^ {m}\right) \right] \\ = u _ {i} \left(\{I \} _ {i = 1} ^ {m}\right). \\ \end{array} +$$ + +Therefore, Algorithm 6 is truthful. For individual rationality, notice that by the definition of $\alpha$ and the assumption that $|X_{i}| < |X_{-i}|$ (and thus $v_{i}(X_{i}) < v_{i}(X_{-i})$ ), we have + +$$ +\begin{array}{l} u _ {i} \left(\left\{I \right\} _ {i = 1} ^ {m}\right) = T _ {i} - T _ {i} \alpha \mathbb {E} \left[ L _ {i} \left(\left\{I \right\} _ {i = 1} ^ {m}\right) \right] \\ = T _ {i} - T _ {i} \left(\frac {1}{2} - \frac {v _ {i} (| X _ {i} |)}{2 T _ {i}}\right) \\ = \frac {v _ {i} (| X _ {- i} |)}{2} + \frac {v _ {i} (| X _ {i} |)}{2} \\ > v _ {i} \left(\left| X _ {i} \right|\right). \\ \end{array} +$$ + +Therefore, agent $i$ is better off participating in Algorithm 6 than working alone so individual rationality is satisfied. + +# B Extended experimental results and details + +# B.1 Text based experiments + +Our first real world experiment supposes that agents possess and wish to share text data drawn from a common distribution. To simulate this text distribution, we use data from the SQuAD dataset [41] which contains 100,000 questions generated by providing crowdworkers with snippets from Wikipedia articles and asking them to formulate questions based on the snippet's content. We simulate + +data sharing when $m = 20$ and $m = 100$ , where agents have 2,500 and 500 original data points respectively. + +When agents are truthful, they simply submit their sentences to the mechanism (Algorithm 3). However, an untruthful agent can fabricate fake sentences to augment their dataset with in hopes of achieving a lower loss. We consider when agents attempt to do this using an LLM (Llama 3.2-1B-Instruct [26]) by prompting it to produce authentic looking sentences based on legitimate sentences. Fig. 3 shows an example of the prompting and Table 4 shows examples of the LLM-generated sentences. For consistency, we filter out duplicates and any outputs not ending in a question mark. + +# Prompt + +Generate five new questions that follow the same style as the examples below. Each question should be separated by a newline. + +According to Southern Living, what are the three best restaurants in Richmond? + +When did the Arab oil producers lift the embargo? + +Complexity classes are generally classified into what? + +About how many acres is Pippy Park? + +Which BYU station offers content in both Spanish and Portuguese? + +Figure 3: Pictured above is an example prompt fed into Llama 3.2-1B-Instruct as part of an untruthful agent's submission function to generate fabricated text data. The agent uses their five questions drawn from the SQuAD to fabricate similar five additional questions. + +Table 3: Comparison of SQuAD questions versus LLM-generated fabrications. + +
SQuAD questions (Real)LLM-generated questions (Fabricated)
Which tribe did Temüjin move in with at nine years of age?What percentage of the population of France lived in urban areas as of 2019?
What is the most widely known fictional work from the Islamic world?The term solar eclipse refers to what pheme-nomenon?
New Delhi played host to what major athletic competition in 2010?Is it true that the first computer bug was an actual insect?
Why did the FCC reject systems such as MUSE?How many Earth years is Neptune's south pole exposed to the Sun?
Along with the philosophies of music and art, what field of philosophy studies emotions?Military spending based on conventional threats has been dismissed as what?
+ +To incentivize truthful submission, we instantiate Algorithm 3 with 768 feature maps corresponding to the 768 nodes in the embedding layer of DistilBERT [42], a lightweight encoder model distilled from the encoder transformer model Bert [45]. For simplicity, we chose the split map $\psi (n) = 0$ . As a point of comparison, we also apply the KS, CvM, and Mean diff. tests (described in §5), now to the 768 node feature space. + +Our results comparing the average loss agent $i$ receives when submitting truthfully/untruthfully, under the four methods, over five runs, are given in Table 1. We see that under all of the methods truthful submission results in a lower average loss than untruthful submission. + +# B.2 Image based experiments + +Our second experiment supposes that agents wish to share image data from a common distribution. To simulate this image distribution, we use data from the Oxford Flowers-102 dataset [43], which contains 6,149 images across 102 flower categories. We simulate data sharing when an agent 1 has 100 and 1,000 images as data points. We use the test dataset of [43], which consists of 4,612 images, to represent authentic data submitted by the other agents. In the two scenarios, this roughly corresponds to $m = 47$ agents each with 100 images and $m = 5$ agents each with 1000 images. + +When agent are untruthful, they may fabricate images using a diffusion model to augment their dataset. We consider when agents use Segmind Stable Diffusion-1B [25], a lightweight diffusion + +model, to do this. More specifically, for each sampled image, we use it in conjunction with the prompts and parameters in Table 4 to generate an additional fabricated image. + +Table 4: Parameters and prompts used for Segmind Stable Diffusion-1B to generate the fabricated images. Here cls_name is replaced with the type of flower being generated. + +
ParameterValue
Text PromptPhotorealistic photograph of a single {cls_name}, realistic colors, natural lighting, high detail, sharp focus on petals. Another unique photo of the same flower species.
Negative Promptoversaturated, highly saturated, neon colors, garish colors, vibrant colors, illustration, painting, drawing, sketch, cartoon, anime, unrealistic, blurry, low quality, text, watermark, signature, border, frame, multiple flowers
Strength0.7
Guidance Scale6
Num. Inference Steps50
+ +To discourage fabrication, Algorithm 3 is now instantiated with 384 feature maps corresponding to the 384 nodes in the embedding layer of DeIT-small-distilled [44], a small vision transformer. For simplicity, we chose the split map $\psi (n) = 0$ . As a point of comparison, we again apply the KS, CvM, and Mean diff. tests (described in §5), now to the 384 node feature space. Our results comparing the average loss agent $i$ receives when submitting truthfully/untruthfully, under the four methods, over five runs, can be found in Table 2. + +We find that truthful reporting outperforms untruthful reporting for all methods, demonstrating they are not susceptible to diffusion based fabrication. + +# C Results and proofs omitted from Section 2 + +# C.1 Results and proofs omitted from Subsection 2.1 + +Theorem 1. The mechanism in Algorithm 1 satisfies truthfulness. Moreover, when $\Pi$ is not degenerate, then Algorithm 1 also satisfies MIB. + +Proof. For truthfulness we refer to Proposition 5. + +For $n = (n_1, \ldots, n_m) \in \mathbb{N}^m$ , let $L_i$ ( $n, \{I\}_{i=1}^m$ ) denote the value of $L_i$ in Algorithm 1 when agent $j \in [m]$ has $n_j$ data points and agents use $\{I\}_{i=1}^m \in \mathcal{F}^m$ . Proposition 6 tells us that + +$$ +\mathbb {E} \left[ L _ {i} \left(n, \{I \} _ {i = 1} ^ {m}\right) \right] - \mathbb {E} \left[ L _ {i} \left(n + e _ {i}, \{I \} _ {i = 1} ^ {m}\right) \right] = \mathbb {E} \left[ \left(\mathbb {E} \left[ F _ {Z _ {i}} \left(T _ {i}\right) | \mathcal {G} _ {n _ {i}} \right] - \mathbb {E} \left[ F _ {Z _ {i}} \left(T _ {i}\right) | \mathcal {G} _ {n _ {i} + 1} \right]\right) ^ {2} \right]. +$$ + +where $\mathcal{G}_j = \sigma (X_{i,1},\ldots X_{i,j},T_i)$ . Also notice that + +$$ +P \left(Z _ {i, 1} \leq T _ {i} \mid X _ {i, 1}, \dots , X _ {i, n _ {i}}, T _ {i}\right) = \mathbb {E} \left[ F _ {Z _ {i}} \left(T _ {i}\right) \mid \mathcal {G} _ {n _ {i}} \right], +$$ + +$$ +P \left(Z _ {i, 1} \leq T _ {i} \mid X _ {i, 1}, \dots , X _ {i, n _ {i} + 1}, T _ {i}\right) = \mathbb {E} \left[ F _ {Z _ {i}} \left(T _ {i}\right) \mid \mathcal {G} _ {n _ {i} + 1} \right]. +$$ + +By definition $\Pi$ being non-degenerate means that the conditional probabilities are not almost surely equal. This implies that + +$$ +\mathbb {E} \left[ L _ {i} \left(n, \{I \} _ {i = 1} ^ {m}\right) \right] - \mathbb {E} \left[ L _ {i} \left(n + e _ {i}, \{I \} _ {i = 1} ^ {m}\right) \right] > 0 +$$ + +so the MIB property is satisfied. + +Proposition 5. Let $L_{i} \left( \{f_{i}\}_{i=1}^{m} \right)$ denote the value of $L_{i}$ in Algorithm 1 when all agents use $\{f_{i}\}_{i=1}^{m} \in \mathcal{F}^{m}$ . Then, for any $f_{i} \in \mathcal{F}$ , $\mathbb{E}[L_{i}(\{I\}_{i=1}^{m})] \leq \mathbb{E}\left[L_{i}(f_{i},\{I\}_{j \neq i})\right]$ . + +Proof. By definition $\mathbb{E}\left[F_{Z_i}(T_i)\big|X_{i,1},\ldots ,X_{i,|Y_i|},T_i\right]$ is $(X_{i,1},\dots,X_{i,|Y_i|},T_i)$ -measurable, so there exists a measurable function $g:\mathbb{R}^{|Y_i| + 1}\to \mathbb{R}$ such that + +$$ +g \left(X _ {i, 1}, \dots , X _ {i, | Y _ {i} |}, T _ {i}\right) = \mathbb {E} \left[ F _ {Z _ {i}} \left(T _ {i}\right) \mid X _ {i, 1}, \dots , X _ {i, | Y _ {i} |}, T _ {i} \right]. +$$ + +The conditional expectation $\mathbb{E}\left[F_{Z_i}(T_i) \mid X_{i,1} = Y_{i,1}, \ldots, X_{i,|Y_i|} = Y_{i,|Y_i|}, T_i\right]$ is shorthand for $g(Y_{i,1}, \ldots, Y_{i,|Y_i|}, T_i)$ . Since we assume $f_i$ is measurable, $\left(Y_{i,1}, \ldots, Y_{i,|Y_i|}\right) = f_i(X_{i,1}, \ldots, X_{i,n_i})$ is $(X_{i,1}, \ldots, X_{i,n_i})$ -measurable. Therefore, we know that + +$$ +g \left(Y _ {i, 1}, \dots , Y _ {i, | Y _ {i} |}, T _ {i}\right) = \mathbb {E} \left[ F _ {Z _ {i}} \left(T _ {i}\right) \mid X _ {i, 1} = Y _ {i, 1}, \dots , X _ {i, | Y _ {i} |} = Y _ {i, | Y _ {i} |}, T _ {i} \right] +$$ + +is $(X_{i,1},\ldots ,X_{i,n_i},T_i)$ -measurable. This lets us apply Lemma 5 to get + +$$ +\begin{array}{l} \mathbb {E} \left[ L _ {i} \left(f _ {i}, \{I \} _ {j \neq i}\right) \right] = \mathbb {E} \left[ \left(\mathbb {E} \left[ F _ {Z _ {i}} \left(T _ {i}\right) \mid X _ {i, 1} = Y _ {i, 1}, \dots , X _ {i, | Y _ {i} |} = Y _ {i, | Y _ {i} |}, T _ {i} \right] - F _ {Z _ {i}} \left(T _ {i}\right)\right) ^ {2} \right] \\ \geq \mathbb {E} \left[ \left(\mathbb {E} \left[ F _ {Z _ {i}} \left(T _ {i}\right) \mid X _ {i, 1}, \dots , X _ {i, n _ {i}}, T _ {i} \right] - F _ {Z _ {i}} \left(T _ {i}\right)\right) ^ {2} \right] \\ = \mathbb {E} \left[ L _ {i} \left(\{I \} _ {i = 1} ^ {m}\right) \right]. \\ \end{array} +$$ + +![](images/0767e424b4494886a4982efd77b7c2f395bc6e4fe1d76153f19f8c94538787dc.jpg) + +Proposition 6. For $n = (n_1, \ldots, n_m) \in \mathbb{N}^m$ let $L_i$ ( $n, \{I\}_{i=1}^m$ ) denote the value of $L_i$ in Algorithm 1 when agent $j \in [m]$ has $n_j$ data points and agents use $\{I\}_{i=1}^m \in \mathcal{F}^m$ . Then + +$$ +\mathbb {E} \left[ L _ {i} \left(n, \{I \} _ {i = 1} ^ {m}\right) \right] - \mathbb {E} \left[ L _ {i} \left(n + e _ {i}, \{I \} _ {i = 1} ^ {m}\right) \right] = \mathbb {E} \left[ \left(\mathbb {E} \left[ F _ {Z _ {i}} \left(T _ {i}\right) | \mathcal {G} _ {n _ {i}} \right] - \mathbb {E} \left[ F _ {Z _ {i}} \left(T _ {i}\right) | \mathcal {G} _ {n _ {i + 1}} \right]\right) ^ {2} \right] +$$ + +where $\mathcal{G}_j = \sigma (X_{i,1},\ldots X_{i,j},T_i)$ + +Proof. For convenience define $U = F_{Z_i}(T_i)$ and $V = \mathbb{E}[U|\mathcal{G}_{n_i}]$ . By the definition of $L_i$ in Algorithm 1 and conditional variance we have + +$$ +\begin{array}{l} \mathbb {E} \left[ L _ {i} (n, \{I \} _ {i = 1} ^ {m}) \right] = \mathbb {E} \left[ \left(\mathbb {E} \left[ F _ {Z _ {i}} (T _ {i}) \mid X _ {i, 1}, \dots , X _ {i, n _ {i}}, T _ {i} \right] - F _ {Z _ {i}} (T _ {i})\right) ^ {2} \right] \\ = \mathbb {E} \left[ (U - V) ^ {2} \right] \\ = \mathbb {E} \left[ \mathbb {E} \left[ (U - V) ^ {2} \mid \mathcal {G} _ {n _ {i}} \right] \right] \\ = \mathbb {E} \left[ \operatorname {V a r} \left(U \mid \mathcal {G} _ {n _ {i}}\right) \right]. \\ \end{array} +$$ + +Similarly we have + +$$ +\mathbb {E} \left[ L _ {i} \left(n + e _ {i}, \{I \} _ {i = 1} ^ {m}\right) \right] = \mathbb {E} \left[ \operatorname {V a r} \left(U | \mathcal {G} _ {n _ {i} + 1}\right) \right]. +$$ + +Let $Y = \mathbb{E}\left[U|\mathcal{G}_{n_i + 1}\right]$ . We can now appeal to Lemma 4 to get + +$$ +\begin{array}{l} \mathbb {E} \left[ L _ {i} (n, \{I \} _ {i = 1} ^ {m}) \right] - \mathbb {E} \left[ L _ {i} (n + e _ {i}, \{I \} _ {i = 1} ^ {m}) \right] = \mathbb {E} \left[ \operatorname {V a r} \left(U | \mathcal {G} _ {n _ {i}}\right) \right] - \mathbb {E} \left[ \operatorname {V a r} \left(U | \mathcal {G} _ {n _ {i} + 1}\right) \right] \\ = \mathbb {E} \left[ \operatorname {V a r} \left(Y | \mathcal {G} _ {n _ {i}}\right) \right]. \\ \end{array} +$$ + +Using the tower property gives + +$$ +\begin{array}{l} \mathbb {E} \left[ \operatorname {V a r} \left(Y \mid \mathcal {G} _ {n _ {i}}\right) \right] = \mathbb {E} \left[ \mathbb {E} \left[ \left(Y - \mathbb {E} \left[ Y \mid \mathcal {G} _ {n _ {i}} \right]\right) ^ {2} \mid \mathcal {G} _ {n _ {i}} \right] \right] \\ = \mathbb {E} \left[ \left(\mathbb {E} \left[ U | \mathcal {G} _ {n _ {i}} \right] - \mathbb {E} \left[ U | \mathcal {G} _ {n _ {i} + 1} \right]\right) ^ {2} \right] \\ = \mathbb {E} \left[ \left(\mathbb {E} \left[ F _ {Z _ {i}} \left(T _ {i}\right) \mid \mathcal {G} _ {n _ {i}} \right] - \mathbb {E} \left[ F _ {Z _ {i}} \left(T _ {i}\right) \mid \mathcal {G} _ {n _ {i} + 1} \right]\right) ^ {2} \right]. \\ \end{array} +$$ + +![](images/6b115b5a3fdf54d6b400b972447b24a3cdbfec5473f9031a353042a0e88ca4b4.jpg) + +Proposition 1. Let $L_{i}(\{I\}_{i = 1}^{m})$ denote the value of $L_{i}$ when agents are truthful in Algorithm 1. Then, $0 \leq \mathbb{E}[L_i(\{I\}_{i = 1}^m)] \leq \frac{1}{4}\left(\frac{1}{|X_i|} + \frac{1}{|Z_i|}\right)$ . Moreover, when $\Pi$ is a prior over the set of continuous $\mathbb{R}$ -valued distributions, $\frac{1}{6|Z_i|} \leq \mathbb{E}[L_i(\{I\}_{i = 1}^m)] \leq \frac{1}{6}\left(\frac{1}{|X_i|} + \frac{1}{|Z_i|}\right)$ . + +Proof. Since $F_{X_i}(T_i)$ is $\sigma(X_{i,1}, \ldots, X_{i,n_i}, T_i)$ -measurable, Lemma 5 tells us that + +$$ +\begin{array}{l} \mathbb {E} \left[ L _ {i} \left(\{I \} _ {i = 1} ^ {m}\right) \right] = \mathbb {E} \left[ \left(\mathbb {E} \left[ F _ {Z _ {i}} \left(T _ {i}\right) \mid X _ {i, 1}, \dots , X _ {i, n _ {i}}, T _ {i} \right] - F _ {Z _ {i}} \left(T _ {i}\right)\right) ^ {2} \right] \\ \leq \mathbb {E} \left[ \left(F _ {X _ {i}} \left(T _ {i}\right) - F _ {Z _ {i}} \left(T _ {i}\right)\right) ^ {2} \right]. \\ \end{array} +$$ + +Now we can condition on $\mathcal{P}$ apply the first part of Lemma 2 to the inner expectation to get + +$$ +\mathbb {E} \left[ \left(F _ {X _ {i}} \left(T _ {i}\right) - F _ {Z _ {i}} \left(T _ {i}\right)\right) ^ {2} \right] = \mathbb {E} _ {\mathcal {P}} \left[ \mathbb {E} _ {X _ {i}, Z _ {i}, T _ {i}} \left[ \left(F _ {X _ {i}} \left(T _ {i}\right) - F _ {Z _ {i}} \left(T _ {i}\right)\right) ^ {2} \right] \right] \leq \frac {1}{4} \left(\frac {1}{| X _ {i} |} + \frac {1}{| Z _ {i} |}\right). +$$ + +Recognizing that $L_{i}$ is non-negative, we conclude + +$$ +0 \leq \mathbb {E} \left[ L _ {i} \left(\left\{I \right\} _ {i = 1} ^ {m}\right) \right] \leq \frac {1}{4} \left(\frac {1}{\left| X _ {i} \right|} + \frac {1}{\left| Z _ {i} \right|}\right). +$$ + +For the second part we assume that $\Pi \in \mathcal{M}_1(\mathcal{M}_1^c (\mathbb{R}))$ , i.e. $\Pi$ is a distribution over the set of continuous $\mathbb{R}$ -valued probability distributions. Again conditioning on $\mathcal{P}$ , we can now apply the second part of Lemma 2 to the inner expectation to get + +$$ +\mathbb {E} _ {\mathcal {P}} \left[ \mathbb {E} _ {X _ {i}, Z _ {i}, T _ {i}} \left[ \left(F _ {X _ {i}} (T _ {i}) - F _ {Z _ {i}} (T _ {i})\right) ^ {2} \right] \right] = \frac {1}{6} \left(\frac {1}{| X _ {i} |} + \frac {1}{| Z _ {i} |}\right) +$$ + +so the upper bound improves to + +$$ +\mathbb {E} \left[ L _ {i} \left(\{I \} _ {i = 1} ^ {m}\right) \right] \leq \frac {1}{6} \left(\frac {1}{| X _ {i} |} + \frac {1}{| Z _ {i} |}\right). +$$ + +For the lower bound notice that $\sigma(X_{i,1}, \ldots, X_{i,n_i}, T_i) \subseteq \sigma(X_{i,1}, \ldots, X_{i,n_i}, T_i, \mathcal{P})$ . Therefore Lemma 5 tells us that + +$$ +\begin{array}{l} \mathbb {E} \left[ L _ {i} \left(\{I \} _ {i = 1} ^ {m}\right) \right] = \mathbb {E} \left[ \left(\mathbb {E} \left[ F _ {Z _ {i}} \left(T _ {i}\right) \mid X _ {i, 1}, \dots , X _ {i, n _ {i}}, T _ {i} \right] - F _ {Z _ {i}} \left(T _ {i}\right)\right) ^ {2} \right] \\ \geq \mathbb {E} \left[ \left(\mathbb {E} \left[ F _ {Z _ {i}} (T _ {i}) \mid X _ {i, 1}, \ldots , X _ {i, n _ {i}}, T _ {i}, \mathcal {P} \right] - F _ {Z _ {i}} (T _ {i})\right) ^ {2} \right]. \\ \end{array} +$$ + +But appealing to Lemmas 3 then 1 (using that $\mathcal{P} \in \mathcal{M}_1^c(\mathbb{R})$ ) gives + +$$ +\begin{array}{l} \mathbb {E} \left[ \left(\mathbb {E} \left[ F _ {Z _ {i}} \left(T _ {i}\right) \mid X _ {i, 1}, \dots , X _ {i, n _ {i}}, T _ {i}, \mathcal {P} \right] - F _ {Z _ {i}} \left(T _ {i}\right)\right) ^ {2} \right] = \mathbb {E} \left[ \left(F _ {\mathcal {P}} \left(T _ {i}\right) - F _ {Z _ {i}} \left(T _ {i}\right)\right) ^ {2} \right] \\ = \mathbb {E} _ {\mathcal {P}} \left[ \mathbb {E} _ {Z _ {i}, T _ {i}} \left[ \left(F _ {\mathcal {P}} \left(T _ {i}\right) - F _ {Z _ {i}} \left(T _ {i}\right)\right) ^ {2} \right] \right] \\ = \mathbb {E} _ {\mathcal {P}} \left[ \frac {1}{6 | Z _ {i} |} \right] \\ = \frac {1}{6 | Z _ {i} |} \\ \end{array} +$$ + +which concludes the proof of the lower bound. + +# C.2 Proofs omitted from Subsection 2.2 + +Theorem 2. The mechanism in Algorithm 2 satisfies truthfulness. Moreover, if there is a feature $k \in [K]$ , for which $\Pi$ is not degenerate, then Algorithm 2 also satisfies MIB. + +Proof. For truthfulness we refer to Proposition 7. + +For $n = (n_1, \ldots, n_m) \in \mathbb{N}^m$ , let $L_i$ ( $n, \{I\}_{i=1}^m$ ) denote the value of $L_i$ in Algorithm 2 when agent $j \in [m]$ has $n_j$ data points and agents use $\{I\}_{i=1}^m \in \mathcal{F}^m$ . Proposition 8 tells us that + +$$ +\begin{array}{l} \mathbb {E} \left[ L _ {i} \left(n, \{I \} _ {i = 1} ^ {m}\right) \right] - \mathbb {E} \left[ L _ {i} \left(n + e _ {i}, \{I \} _ {i = 1} ^ {m}\right) \right] \\ = \frac {1}{K} \sum_ {k = 1} ^ {K} \mathbb {E} \left[ \left(\mathbb {E} \left[ F _ {Z _ {i} ^ {k}} \left(T _ {i} ^ {k}\right) | \mathcal {G} _ {n _ {i}} ^ {k} \right] - \mathbb {E} \left[ F _ {Z _ {i} ^ {k}} \left(T _ {i} ^ {k}\right) | \mathcal {G} _ {n _ {i} + 1} ^ {k} \right]\right) ^ {2} \right]. \\ \end{array} +$$ + +where $\mathcal{G}_j^k = \sigma \left(X_{i,1},\ldots X_{i,j},T_i^k\right)$ . Also observe that + +$$ +P \left(Z _ {i, 1} ^ {k} \leq T _ {i} ^ {k} \mid X _ {i, 1}, \dots , X _ {i, n _ {i}}, T _ {i} ^ {k}\right) = \mathbb {E} \left[ F _ {Z _ {i} ^ {k}} \left(T _ {i} ^ {k}\right) \mid \mathcal {G} _ {n _ {i}} ^ {k} \right], +$$ + +$$ +P \left(Z _ {i, 1} ^ {k} \leq T _ {i} ^ {k} | X _ {i, 1}, \dots , X _ {i, n _ {i} + 1}, T _ {i} ^ {k}\right) = \mathbb {E} \left[ F _ {Z _ {i} ^ {k}} \left(T _ {i} ^ {k}\right) | \mathcal {G} _ {n _ {i} + 1} ^ {k} \right]. +$$ + +Since we assume there is a feature $k \in [K]$ for which $\Pi$ is non-degenerate, the conditional probabilities are not almost surely equal for at least one feature. Therefore, + +$$ +\mathbb {E} \left[ L _ {i} \left(n, \{I \} _ {i = 1} ^ {m}\right) \right] - \mathbb {E} \left[ L _ {i} \left(n + e _ {i}, \{I \} _ {i = 1} ^ {m}\right) \right] > 0 +$$ + +so the MIB property is satisfied. + +![](images/701efb550a7d6c1db1ac14c7256384c14f39a6a5419999c7ac699aadd7c0413d.jpg) + +Proposition 7. Let $L_{i}\left(\{f_{i}\}_{i = 1}^{m}\right)$ denote the value of $L_{i}$ in Algorithm 2 when all agents use $\{f_i\}_{i = 1}^m\in$ $\mathcal{F}^m$ . Then, for any $f_{i}\in \mathcal{F}$ $\mathbb{E}\left[L_i\left(\{I\}_{i = 1}^m\right)\right]\leq \mathbb{E}\left[L_i\left(f_i,\{I\}_{j\neq i}\right)\right].$ + +Proof. By the definition of Algorithm 2 we have + +$$ +\mathbb {E} \left[ L _ {i} \left(f _ {i}, \{I \} _ {j \neq i}\right) \right] = \frac {1}{K} \sum_ {k = 1} ^ {K} \mathbb {E} \left[ L _ {i} ^ {k} \left(f _ {i}, \{I \} _ {j \neq i}\right) \right]. +$$ + +By definition $\mathbb{E}\left[F_{Z_i^k}\left(T_i^k\right)\big|X_{i,1},\ldots ,X_{i,|Y_i|},T_i^k\right]$ is $(X_{i,1},\dots,X_{i,|Y_i|},T_i^k)$ -measurable, so there exists a measurable function $g:\mathcal{X}^{|Y_i|}\times \mathbb{R}\to \mathbb{R}$ such that + +$$ +g \left(X _ {i, 1}, \dots , X _ {i, | Y _ {i} |}, T _ {i} ^ {k}\right) = \mathbb {E} \left[ F _ {Z _ {i} ^ {k}} \left(T _ {i} ^ {k}\right) \mid X _ {i, 1}, \dots , X _ {i, | Y _ {i} |}, T _ {i} ^ {k} \right]. +$$ + +The conditional expectation + +$$ +\mathbb {E} \left[ F _ {Z _ {i} ^ {k}} \left(T _ {i} ^ {k}\right) \mid X _ {i, 1} = Y _ {i, 1}, \dots , X _ {i, | Y _ {i} |} = Y _ {i, | Y _ {i} |}, T _ {i} ^ {k} \right] +$$ + +is shorthand for $g\left(Y_{i,1}, \ldots, Y_{i,|Y_i|}, T_i^k\right)$ . Since we assume $f_i$ is measurable, + +$$ +\left(Y _ {i, 1}, \dots , Y _ {i, | Y _ {i} |}\right) = f _ {i} \left(X _ {i, 1}, \dots , X _ {i, n _ {i}}\right) +$$ + +is $(X_{i,1},\ldots ,X_{i,n_i})$ -measurable. Therefore, we know that + +$$ +g \left(Y _ {i, 1}, \dots , Y _ {i, | Y _ {i} |}, T _ {i} ^ {k}\right) = \mathbb {E} \left[ F _ {Z _ {i} ^ {k}} \left(T _ {i} ^ {k}\right) \mid X _ {i, 1} = Y _ {i, 1}, \dots , X _ {i, | Y _ {i} |} = Y _ {i, | Y _ {i} |}, T _ {i} ^ {k} \right] +$$ + +is $(X_{i,1},\ldots ,X_{i,n_i},T_i^k)$ -measurable. This lets us apply Lemma 5 to get + +$$ +\begin{array}{l} \mathbb {E} \left[ L _ {i} ^ {k} \left(f _ {i}, \{I \} _ {j \neq i}\right) \right] \\ = \mathbb {E} \left[ \left(\mathbb {E} \left[ F _ {Z _ {i} ^ {k}} \left(T _ {i} ^ {k}\right) \mid X _ {i, 1} = Y _ {i, 1}, \dots , X _ {i, | Y _ {i} |} = Y _ {i, | Y _ {i} |}, T _ {i} ^ {k} \right] - F _ {Z _ {i} ^ {k}} \left(T _ {i} ^ {k}\right)\right) ^ {2} \right] \\ \geq \mathbb {E} \left[ \left(\mathbb {E} \left[ F _ {Z _ {i} ^ {k}} \left(T _ {i} ^ {k}\right) \mid X _ {i, 1}, \dots , X _ {i, n _ {i}}, T _ {i} ^ {k} \right] - F _ {Z _ {i} ^ {k}} \left(T _ {i} ^ {k}\right)\right) ^ {2} \right] \\ = \mathbb {E} \left[ L _ {i} ^ {k} \left(\{I \} _ {i = 1} ^ {m}\right) \right]. \\ \end{array} +$$ + +Repeatedly applying this argument for each feature gives us + +$$ +\frac {1}{K} \sum_ {k = 1} ^ {K} \mathbb {E} \left[ L _ {i} ^ {k} \left(f _ {i}, \{I \} _ {j \neq i}\right) \right] \geq \frac {1}{K} \sum_ {k = 1} ^ {K} \mathbb {E} \left[ L _ {i} ^ {k} \left(\{I \} _ {i = 1} ^ {m}\right) \right] = \mathbb {E} \left[ L _ {i} \left(\{I \} _ {i = 1} ^ {m}\right) \right]. +$$ + +![](images/538bbacce14646632204734146041f90a309abc3ca72bf79aed954279c23ea9a.jpg) + +Proposition 8. For $n = (n_1, \ldots, n_m) \in \mathbb{N}^m$ let $L_i$ ( $n, \{I\}_{i=1}^m$ ) denote the value of $L_i$ in Algorithm 2 when agent $j \in [m]$ has $n_j$ data points and agents use $\{I\}_{i=1}^m \in \mathcal{F}^m$ . Then + +$$ +\begin{array}{l} \mathbb {E} \left[ L _ {i} (n, \{I \} _ {i = 1} ^ {m}) \right] - \mathbb {E} \left[ L _ {i} (n + e _ {i}, \{I \} _ {i = 1} ^ {m}) \right] \\ = \frac {1}{K} \sum_ {k = 1} ^ {K} \mathbb {E} \left[ \left(\mathbb {E} \left[ F _ {Z _ {i} ^ {k}} \left(T _ {i} ^ {k}\right) \mid \mathcal {G} _ {n _ {i}} ^ {k} \right] - \mathbb {E} \left[ F _ {Z _ {i} ^ {k}} \left(T _ {i} ^ {k}\right) \mid \mathcal {G} _ {n _ {i} + 1} ^ {k} \right]\right) ^ {2} \right] \\ \end{array} +$$ + +where $\mathcal{G}_j^k = \sigma \left(X_{i,1},\ldots X_{i,j},T_i^k\right)$ + +Proof. By the definition of Algorithm 2 + +$$ +\begin{array}{l} \mathbb {E} \left[ L _ {i} \left(n, \{I \} _ {i = 1} ^ {m}\right) \right] = \frac {1}{K} \sum_ {k = 1} ^ {K} \mathbb {E} \left[ L _ {i} ^ {k} \left(n, \{I \} _ {i = 1} ^ {m}\right) \right] \\ = \frac {1}{K} \sum_ {k = 1} ^ {K} \mathbb {E} \left[ \left(\mathbb {E} \left[ F _ {Z _ {i} ^ {k}} \left(T _ {i} ^ {k}\right) \mid \mathcal {G} _ {n _ {i}} ^ {k} \right] - F _ {Z _ {i} ^ {k}} \left(T _ {i} ^ {k}\right)\right) ^ {2} \right] \\ \end{array} +$$ + +Let $U^{k} = F_{Z_{i}^{k}}\left(T_{i}^{k}\right)$ . From the equation above, the tower property and definition of conditional variance tell us that + +$$ +\mathbb {E} \left[ L _ {i} \left(n, \{I \} _ {i = 1} ^ {m}\right) \right] = \frac {1}{K} \sum_ {k = 1} ^ {K} \mathbb {E} \left[ \operatorname {V a r} \left(U ^ {k} | \mathcal {G} _ {n _ {i}} ^ {k}\right) \right]. +$$ + +An analogous argument gives + +$$ +\mathbb {E} \left[ L _ {i} \left(n + e _ {i}, \{I \} _ {i = 1} ^ {m}\right) \right] = \frac {1}{K} \sum_ {k = 1} ^ {K} \mathbb {E} \left[ \operatorname {V a r} \left(U ^ {k} | \mathcal {G} _ {n _ {i} + 1} ^ {k}\right) \right]. +$$ + +Let $Y^{k} = \mathbb{E}\left[U^{k}|\mathcal{G}_{n_{i} + 1}^{k}\right]$ . We can now repeatedly appeal to Lemma 4 to get + +$$ +\begin{array}{l} \mathbb {E} \left[ L _ {i} (n, \{I \} _ {i = 1} ^ {m}) \right] - \mathbb {E} \left[ L _ {i} (n + e _ {i}, \{I \} _ {i = 1} ^ {m}) \right] \\ = \frac {1}{K} \sum_ {k = 1} ^ {K} \left(\mathbb {E} \left[ \operatorname {V a r} \left(U ^ {k} \mid \mathcal {G} _ {n _ {i}} ^ {k}\right) \right] - \mathbb {E} \left[ \operatorname {V a r} \left(U ^ {k} \mid \mathcal {G} _ {n _ {i} + 1} ^ {k}\right) \right]\right) \\ = \frac {1}{K} \sum_ {k = 1} ^ {K} \mathbb {E} \left[ \operatorname {V a r} \left(Y ^ {k} \mid \mathcal {G} _ {n _ {i}} ^ {k}\right) \right]. \\ \end{array} +$$ + +Using the tower property now lets us conclude that + +$$ +\begin{array}{l} \frac {1}{K} \sum_ {k = 1} ^ {K} \mathbb {E} \left[ \operatorname {V a r} \left(Y ^ {k} | \mathcal {G} _ {n _ {i}} ^ {k}\right) \right] = \frac {1}{K} \sum_ {k = 1} ^ {K} \mathbb {E} \left[ \mathbb {E} \left[ \left(Y ^ {k} - \mathbb {E} \left[ Y ^ {k} | \mathcal {G} _ {n _ {i}} ^ {k} \right]\right) ^ {2} | \mathcal {G} _ {n _ {i}} ^ {k} \right] \right] \\ = \frac {1}{K} \sum_ {k = 1} ^ {K} \mathbb {E} \left[ \left(\mathbb {E} \left[ U ^ {k} \mid \mathcal {G} _ {n _ {i}} ^ {k} \right] - \mathbb {E} \left[ U ^ {k} \mid \mathcal {G} _ {n _ {i} + 1} ^ {k} \right]\right) ^ {2} \right] \\ = \frac {1}{K} \sum_ {k = 1} ^ {K} \mathbb {E} \left[ \left(\mathbb {E} \left[ F _ {Z _ {i} ^ {k}} \left(T _ {i} ^ {k}\right) | \mathcal {G} _ {n _ {i}} ^ {k} \right] - \mathbb {E} \left[ F _ {Z _ {i} ^ {k}} \left(T _ {i} ^ {k}\right) | \mathcal {G} _ {n _ {i} + 1} ^ {k} \right] ^ {2}\right) \right]. \\ \end{array} +$$ + +![](images/4f7f7e360d56ebf22e8f943ba1be7da271dfcd0144e10c87a698624b65fa384f.jpg) + +Proposition 9. Let $L_{i}\left(\{I\}_{i = 1}^{m}\right)$ denote the value of $L_{i}$ when all agents are truthful in Algorithm 2. Then, $0 \leq \mathbb{E}\left[L_{i}\left(\{I\}_{i = 1}^{m}\right)\right] \leq \frac{1}{4}\left(\frac{1}{|X_{i}|} + \frac{1}{|Z_{i}|}\right)$ . Moreover, if $\forall k \in [K]$ , $\mathcal{P}^k = \mathcal{P} \circ (\varphi^k)^{-1}$ is a.s. continuous, then $\frac{1}{6|Z_i|} \leq \mathbb{E}\left[L_i\left(\{I\}_{i = 1}^{m}\right)\right] \leq \frac{1}{6}\left(\frac{1}{|X_i|} + \frac{1}{|Z_i|}\right)$ . + +Proof. By definition + +$$ +\mathbb {E} \left[ L _ {i} \left(\{I \} _ {i = 1} ^ {m}\right) \right] = \frac {1}{K} \sum_ {k = 1} ^ {K} \mathbb {E} \left[ L _ {i} ^ {k} \left(\{I \} _ {i = 1} ^ {m}\right) \right]. +$$ + +Define $X_{i}^{k} = \left(\varphi^{k}\left(X_{i,j}\right)\right)_{j = 1}^{n_{i}}$ . We have $F_{X_i^k}\left(T_i^k\right)$ is $(X_{i,1},\ldots ,X_{i,n_i},T_i^k)$ -measurable. Therefore, Lemma 5 tells us that + +$$ +\begin{array}{l} \mathbb {E} \left[ L _ {i} ^ {k} \left(\left\{I \right\} _ {i = 1} ^ {m}\right) \right] = \mathbb {E} \left[ \left(\mathbb {E} \left[ F _ {Z _ {i} ^ {k}} \left(T _ {i} ^ {k}\right) \mid X _ {i, 1}, \dots , X _ {i, n _ {i}}, T _ {i} ^ {k} \right] - F _ {Z _ {i} ^ {k}} \left(T _ {i} ^ {k}\right)\right) ^ {2} \right] \\ \leq \mathbb {E} \left[ \left(F _ {X _ {i} ^ {k}} \left(T _ {i} ^ {k}\right) - F _ {Z _ {i} ^ {k}} \left(T _ {i} ^ {k}\right)\right) ^ {2} \right] \\ = \mathbb {E} _ {\mathcal {P}} \left[ \mathbb {E} _ {X _ {i} ^ {k}, Z _ {i} ^ {k}, T _ {i} ^ {k}} \left[ \left(F _ {X _ {i} ^ {k}} \left(T _ {i} ^ {k}\right) - F _ {Z _ {i} ^ {k}} \left(T _ {i} ^ {k}\right)\right) ^ {2} \right] \right]. \\ \end{array} +$$ + +Applying the first part of Lemma 2 to the inner expectation gives + +$$ +\mathbb {E} _ {\mathcal {P}} \left[ \mathbb {E} _ {X _ {i} ^ {k}, Z _ {i} ^ {k}, T _ {i} ^ {k}} \left[ \left(F _ {X _ {i} ^ {k}} (T _ {i} ^ {k}) - F _ {Z _ {i} ^ {k}} (T _ {i} ^ {k})\right) ^ {2} \right] \right] \leq \frac {1}{4} \left(\frac {1}{| X _ {i} ^ {k} |} + \frac {1}{| Z _ {i} ^ {k} |}\right) = \frac {1}{4} \left(\frac {1}{| X _ {i} |} + \frac {1}{| Z _ {i} |}\right) +$$ + +so we conclude + +$$ +0 \leq \mathbb {E} \left[ L _ {i} \left(\left\{I \right\} _ {i = 1} ^ {m}\right) \right] \leq \frac {1}{4} \left(\frac {1}{\left| X _ {i} \right|} + \frac {1}{\left| Z _ {i} \right|}\right). +$$ + +For the second part, when we assume $\mathcal{P}^k$ is a.s. continuous, we can apply the second part of Lemma 2 to get + +$$ +\mathbb {E} _ {\mathcal {P}} \left[ \mathbb {E} _ {X _ {i} ^ {k}, Z _ {i} ^ {k}, T _ {i} ^ {k}} \left[ \left(F _ {X _ {i} ^ {k}} (T _ {i} ^ {k}) - F _ {Z _ {i} ^ {k}} (T _ {i} ^ {k})\right) ^ {2} \right] \right] = \frac {1}{6} \left(\frac {1}{| X _ {i} ^ {k} |} + \frac {1}{| Z _ {i} ^ {k} |}\right) = \frac {1}{6} \left(\frac {1}{| X _ {i} |} + \frac {1}{| Z _ {i} |}\right) +$$ + +so the upper bound improves to + +$$ +\mathbb {E} \left[ L _ {i} \left(\left\{I \right\} _ {i = 1} ^ {m}\right) \right] \leq \frac {1}{6} \left(\frac {1}{\left| X _ {i} \right|} + \frac {1}{\left| Z _ {i} \right|}\right). +$$ + +For the lower bound, note that $\sigma(X_{i,1},\ldots,X_{i,n_i},T_i^k) \subseteq \sigma(X_{i,1},\ldots,X_{i,n_i},T_i^k,\mathcal{P}^k)$ so Lemma 5 gives us that + +$$ +\begin{array}{l} \mathbb {E} \left[ L _ {i} ^ {k} \left(\{I \} _ {i = 1} ^ {m}\right) \right] = \mathbb {E} \left[ \left(\mathbb {E} \left[ F _ {Z _ {i} ^ {k}} \left(T _ {i} ^ {k}\right) \mid X _ {i, 1}, \dots , X _ {i, n _ {i}}, T _ {i} ^ {k} \right] - F _ {Z _ {i} ^ {k}} \left(T _ {i} ^ {k}\right)\right) ^ {2} \right] \\ \geq \mathbb {E} \left[ \left(\mathbb {E} \left[ F _ {Z _ {i} ^ {k}} \left(T _ {i} ^ {k}\right) \mid X _ {i, 1}, \dots , X _ {i, n _ {i}}, T _ {i} ^ {k}, \mathcal {P} ^ {k} \right] - F _ {Z _ {i} ^ {k}} \left(T _ {i} ^ {k}\right)\right) ^ {2} \right]. \\ \end{array} +$$ + +Now appealing to Lemmas 3 then 1 gives + +$$ +\begin{array}{l} \mathbb {E} \left[ \left(\mathbb {E} \left[ F _ {Z _ {i} ^ {k}} \left(T _ {i} ^ {k}\right) \mid X _ {i, 1}, \dots , X _ {i, n _ {i}}, T _ {i} ^ {k}, \mathcal {P} ^ {k} \right] - F _ {Z _ {i} ^ {k}} \left(T _ {i} ^ {k}\right)\right) ^ {2} \right] \\ = \mathbb {E} \left[ \left(F _ {\mathcal {P} ^ {k}} \left(T _ {i} ^ {k}\right) - F _ {\mathcal {Z} _ {i} ^ {k}} \left(T _ {i} ^ {k}\right)\right) ^ {2} \right] \\ = \mathbb {E} _ {\mathcal {P} ^ {k}} \left[ \mathbb {E} _ {Z _ {i} ^ {k}, T _ {i} ^ {k}} \left[ \left(F _ {\mathcal {P} ^ {k}} \left(T _ {i} ^ {k}\right) - F _ {Z _ {i} ^ {k}} \left(T _ {i} ^ {k}\right)\right) ^ {2} \right] \right] \\ = \mathbb {E} _ {\mathcal {P} ^ {k}} \left[ \frac {1}{6 \left| Z _ {i} ^ {k} \right|} \right] \\ = \frac {1}{6 | Z _ {i} |}. \\ \end{array} +$$ + +Therefore, we conclude that $\frac{1}{6|Z_i|} \leq \mathbb{E}\left[L_i\left(\{I\}_{i=1}^m\right)\right]$ . + +# D Proofs omitted from Section 3 + +Theorem 3. The mechanism in Algorithm 3 is $\frac{1}{4}\left(\frac{1}{|X_i| + |W_i|} +\frac{1}{|Z_i|}\right)$ -approximately truthful in both the Bayesian and frequentist settings. Moreover, if there is a feature $k\in [K]$ , for which $\Pi$ is not degenerate, then Algorithm 3 satisfies MIB in the Bayesian setting. If it is not the case that $\mathcal{C}\subseteq \left\{\mathcal{P}\in \mathcal{M}_1(\mathcal{X}):\forall k\in [K],\mathcal{P}\circ (\varphi^k)^{-1}\in \delta_x,x\in \mathbb{R}\right\}$ then Algorithm 3 satisfies MIB in the frequentist setting. + +Proof. For $\frac{1}{4}\left(\frac{1}{|X_i| + |W_i|} +\frac{1}{|Z_i|}\right)$ -approximate truthfulness we refer to Proposition 11. + +Let $\mathcal{P}^k = \mathcal{P}\circ (\varphi^k)^{-1}$ . For MIB we first look at the Bayesian setting and then the frequentist setting. + +For the Bayesian setting, from the assumption about $\Pi$ we know that $\exists k\in [K]$ where $\forall n_{i}\in \mathbb{N}$ it is not the case that + +$$ +P \left(Z _ {i, 1} ^ {k} \leq T _ {i} ^ {k} | X _ {i, 1}, \dots , X _ {i, n _ {i}}, T _ {i} ^ {k}\right) \stackrel {{a. s.}} {{=}} P \left(Z _ {i, 1} ^ {k} \leq T _ {i} ^ {k} | X _ {i, 1}, \dots , X _ {i, n _ {i} + 1}, T _ {i} ^ {k}\right). +$$ + +Now notice that this implies that for at least one of the $k \in [K]$ features, $P\left(\mathcal{P}^k \in \{\delta_x : x \in \mathcal{X}\}\right) < 1$ , or else the conditional probabilities above would automatically be equal for each $k \in [K]$ . + +We know from Proposition 12 that + +$$ +\begin{array}{l} \mathbb {E} \left[ L _ {i} \left(n, \{I \} _ {i = 1} ^ {m}\right) \right] - \mathbb {E} \left[ L _ {i} \left(n + e _ {i}, \{I \} _ {i = 1} ^ {m}\right) \right] \\ = \frac {1}{K} \sum_ {k = 1} ^ {K} \mathbb {E} \left[ F _ {\mathcal {P} ^ {k}} \left(T _ {i} ^ {k}\right) \left(1 - F _ {\mathcal {P} ^ {k}} \left(T _ {i} ^ {k}\right)\right) \right] \left(\frac {1}{n _ {i} + | W _ {i} |} - \frac {1}{n _ {i} + 1 + | W _ {i} |}\right). \\ \end{array} +$$ + +But notice that + +$$ +\exists k \in [ K ] \quad s. t. \quad P \left(\mathcal {P} ^ {k} \in \left\{\delta_ {x}: x \in \mathcal {X} \right\}\right) < 1 +$$ + +implies $\frac{1}{K}\sum_{k = 1}^{K}\mathbb{E}\left[F_{\mathcal{P}^k}\left(T_i^k\right)\left(1 - F_{\mathcal{P}^k}\left(T_i^k\right)\right)\right] > 0.$ Therefore + +$$ +\mathbb {E} \left[ L _ {i} \left(n, \{I \} _ {i = 1} ^ {m}\right) \right] - \mathbb {E} \left[ L _ {i} \left(n + e _ {i}, \{I \} _ {i = 1} ^ {m}\right) \right] > 0 +$$ + +which proves MIB for the Bayesian setting. + +For the frequentist setting, we have from Proposition 13 that + +$$ +\begin{array}{l} \sup _ {\mathcal {P} \in \mathcal {C}} \mathbb {E} \left[ L _ {i} \left(n, \{I \} _ {i = 1} ^ {m}\right) \right] - \sup _ {\mathcal {P} \in \mathcal {C}} \mathbb {E} \left[ L _ {i} \left(n + e _ {i}, \{I \} _ {i = 1} ^ {m}\right) \right] \\ = \left(\sup _ {\mathcal {P} \in \mathcal {C}} \frac {1}{K} \sum_ {k = 1} ^ {K} \mathbb {E} \left[ F _ {\mathcal {P} ^ {k}} \left(T _ {i} ^ {k}\right) \left(1 - F _ {\mathcal {P} ^ {k}} \left(T _ {i} ^ {k}\right)\right) \right]\right) \left(\frac {1}{n _ {i} + | W _ {i} |} - \frac {1}{n _ {i} + 1 + | W _ {i} |}\right). \\ \end{array} +$$ + +If it is not the case that + +$$ +\mathcal {C} \subseteq \left\{\mathcal {P} \in \mathcal {M} _ {1} (\mathcal {X}): \forall k \in [ K ], \mathcal {P} \circ \left(\varphi^ {k}\right) ^ {- 1} \in \delta_ {x}, x \in \mathbb {R} \right\} +$$ + +then + +$$ +\sup _ {\mathcal {P} \in \mathcal {C}} \frac {1}{K} \sum_ {k = 1} ^ {K} \mathbb {E} \left[ F _ {\mathcal {P} ^ {k}} \left(T _ {i} ^ {k}\right) \left(1 - F _ {\mathcal {P} ^ {k}} \left(T _ {i} ^ {k}\right)\right) \right] > 0 +$$ + +so we find + +$$ +\sup _ {\mathcal {P} \in \mathcal {C}} \mathbb {E} \left[ L _ {i} \left(n, \{I \} _ {i = 1} ^ {m}\right) \right] - \sup _ {\mathcal {P} \in \mathcal {C}} \mathbb {E} \left[ L _ {i} \left(n + e _ {i}, \{I \} _ {i = 1} ^ {m}\right) \right] > 0 +$$ + +which proves MIB for the frequentist setting. + +Proposition 10. Let $L_{i} \left( \{I\}_{i=1}^{m} \right)$ denote the value of $L_{i}$ when all agents follow $\{I\}_{i=1}^{m} \in \mathcal{F}^{m}$ in Algorithm 3. Then, + +$$ +0 \leq \underset {\mathcal {P} \sim \Pi} {\mathbb {E}} \left[ \mathbb {E} _ {\mathcal {P}} \left[ L _ {i} \left(\{I \} _ {i = 1} ^ {m}\right) \right] \right], \sup _ {\mathcal {P} \in \mathcal {C}} \mathbb {E} _ {\mathcal {P}} \left[ L _ {i} \left(\{I \} _ {i = 1} ^ {m}\right) \right] \leq \frac {1}{4} \left(\frac {1}{| X _ {i} | + | W _ {i} |} + \frac {1}{| Z _ {i} |}\right). +$$ + +Moreover, if $\forall k\in [K],\mathcal{P}^k = \mathcal{P}\circ (\varphi^k)^{-1}$ is a.s. continuous, then + +$$ +\underset {\mathcal {P} \sim \Pi} {\mathbb {E}} \left[ \right. \underset {\{X _ {i} \} _ {i} \sim \mathcal {P}} {\mathbb {E}} \left[ \right. L _ {i} \left( \right.\left\{ \right.\left\{I \right\} _ {i = 1} ^ {m}\left. \right)\left. \right]\left. \right] = \sup _ {\mathcal {P} \in \mathcal {C}} \mathbb {E} _ {\mathcal {P}} \left[ \right. L _ {i} \left( \right.\left\{ \right.\left\{I \right\} _ {i = 1} ^ {m}\left. \right)\left. \right]\left. \right. = \frac {1}{6} \left(\frac {1}{| X _ {i} | + | W _ {i} |} + \frac {1}{| Z _ {i} |}\right). +$$ + +Proof. By the definition of Algorithm 3 we have + +$$ +\underset {\mathcal {P} \sim \Pi} {\mathbb {E}} \left[ \underset {\{X _ {i} \} _ {i} \sim \mathcal {P}} {\mathbb {E}} [ L _ {i} (\{I \} _ {i = 1} ^ {m}) ] \right] = \underset {\mathcal {P} \sim \Pi} {\mathbb {E}} \left[ \frac {1}{K} \sum_ {k = 1} ^ {K} \mathbb {E} \left[ \left(F _ {(Y _ {i} ^ {k}, W _ {i} ^ {k})} (T _ {i} ^ {k}) - F _ {Z _ {i} ^ {k}} (T _ {i} ^ {k})\right) ^ {2} \right] \right] +$$ + +and + +$$ +\sup _ {\mathcal {P} \in \mathcal {C}} \mathbb {E} _ {\mathcal {P}} \left[ L _ {i} \left(\{I \} _ {i = 1} ^ {m}\right) \right] = \sup _ {\mathcal {P} \in \mathcal {C}} \frac {1}{K} \sum_ {k = 1} ^ {K} \mathbb {E} \left[ \left(F _ {\left(Y _ {i} ^ {k}, W _ {i} ^ {k}\right)} \left(T _ {i} ^ {k}\right) - F _ {Z _ {i} ^ {k}} \left(T _ {i} ^ {k}\right)\right) ^ {2} \right]. +$$ + +The first part of Lemma 2 tells us that in both the frequentist and Bayesian setting, + +$$ +\mathbb {E} \left[ \left(F _ {\left(Y _ {i} ^ {k}, W _ {i} ^ {k}\right)} \left(T _ {i} ^ {k}\right) - F _ {Z _ {i} ^ {k}} \left(T _ {i} ^ {k}\right)\right) ^ {2} \right] \leq \frac {1}{4} \left(\frac {1}{| X _ {i} | + | W _ {i} |} + \frac {1}{| Z _ {i} |}\right) +$$ + +so we find + +$$ +0 \leq \underset {\mathcal {P} \sim \Pi} {\mathbb {E}} \left[ \mathbb {E} _ {\mathcal {P}} \left[ L _ {i} \left(\{I \} _ {i = 1} ^ {m}\right) \right] \right], \sup _ {\mathcal {P} \in \mathcal {C}} \mathbb {E} _ {\mathcal {P}} \left[ L _ {i} \left(\{I \} _ {i = 1} ^ {m}\right) \right] \leq \frac {1}{4} \left(\frac {1}{\left| X _ {i} \right| + \left| W _ {i} \right|} + \frac {1}{\left| Z _ {i} \right|}\right). +$$ + +When $\forall k\in [K],\mathcal{P}^k$ is a.s. continuous, we apply the second part of Lemma 2 to get + +$$ +\mathbb {E} \left[ \left(F _ {\left(Y _ {i} ^ {k}, W _ {i} ^ {k}\right)} \left(T _ {i} ^ {k}\right) - F _ {Z _ {i} ^ {k}} \left(T _ {i} ^ {k}\right)\right) ^ {2} \right] = \frac {1}{6} \left(\frac {1}{| X _ {i} | + | W _ {i} |} + \frac {1}{| Z _ {i} |}\right) +$$ + +in both the frequentist and Bayesian setting. Under this additional hypothesis, we get + +$$ +\underset {\mathcal {P} \sim \Pi} {\mathbb {E}} \left[ \right. \underset {\{X _ {i} \} _ {i} \sim \mathcal {P}} {\mathbb {E}} \left[ \right. L _ {i} \left( \right.\left\{ \right.\left\{I \right\} _ {i = 1} ^ {m}\left. \right)\left. \right]\left. \right] = \sup _ {\mathcal {P} \in \mathcal {C}} \mathbb {E} _ {\mathcal {P}} \left[ \right. L _ {i} \left( \right.\left\{ \right.\left\{I \right\} _ {i = 1} ^ {m}\left. \right)\left. \right]\left. \right. = \frac {1}{6} \left(\frac {1}{| X _ {i} | + | W _ {i} |} + \frac {1}{| Z _ {i} |}\right). +$$ + +![](images/b090c4371035139b9b7a1fbf11ebe43971d763f537e24449817ca6999bf2c567.jpg) + +Proposition 11. Let $L_{i}\left(\{f_{i}\}_{i = 1}^{m}\right)$ denote the value of $L_{i}$ in Algorithm 3 when agents use $\{f_i\}_{i = 1}^m\in$ $\mathcal{F}^m$ . Let $\Pi$ be a Bayesian prior, and $\mathcal{C}\subseteq \mathcal{M}_1(\mathcal{X})$ a class of $\mathcal{X}$ -valued distributions. Then, for any $f_{i}\in \mathcal{F}$ + +$$ +\begin{array}{l} \underset {\mathcal {P} \sim \Pi} {\mathbb {E}} \left[ \underset {\{X _ {i} \} _ {i} \sim \mathcal {P}} {\mathbb {E}} \left[ L _ {i} \left(\{I \} _ {i = 1} ^ {m}\right) \right] \right] \leq \underset {\mathcal {P} \sim \Pi} {\mathbb {E}} \left[ \underset {\{X _ {i} \} _ {i} \sim \mathcal {P}} {\mathbb {E}} \left[ L _ {i} \left(f _ {i}, \{I \} _ {j \neq i}\right) \right] \right] + \varepsilon a n d \\ \sup _ {\mathcal {P} \in \mathcal {C}} \underset {\{X _ {i} \} _ {i} \sim \mathcal {P}} {\mathbb {E}} \left[ L _ {i} \left(\left\{I \right\} _ {i = 1} ^ {m}\right) \right] \leq \sup _ {\mathcal {P} \in \mathcal {C}} \underset {\{X _ {i} \} _ {i} \sim \mathcal {P}} {\mathbb {E}} \left[ L _ {i} \left(f _ {i}, \left\{I \right\} _ {j \neq i}\right) \right] + \varepsilon \\ \end{array} +$$ + +where $\varepsilon = \frac{1}{4}\left(\frac{1}{|X_i| + |W_i|} +\frac{1}{|Z_i|}\right)$ . Moreover, if $\forall k\in [K]$ , $\mathcal{P}^k = \mathcal{P}\circ (\varphi^k)^{-1}$ is a.s. continuous in the Bayesian setting and $\forall \mathcal{P}\in \mathcal{C}$ in the frequentist setting, then the above inequalities hold with $\varepsilon = \frac{1}{6(|X_i| + |W_i|)}$ . + +Proof. The first part of the claim, when $\varepsilon = \frac{1}{4}\left(\frac{1}{|X_i| + |W_i|} +\frac{1}{|Z_i|}\right)$ , follows immediately from Proposition 10 and recognizing that + +$$ +0 \leq \underset {\mathcal {P} \sim \Pi} {\mathbb {E}} \left[ \underset {\{X _ {i} \} _ {i} \sim \mathcal {P}} {\mathbb {E}} \left[ L _ {i} \left(f _ {i}, \{I \} _ {j \neq i}\right) \right] \right], \sup _ {\mathcal {P} \in \mathcal {C}} \underset {\{X _ {i} \} _ {i} \sim \mathcal {P}} {\mathbb {E}} \left[ L _ {i} \left(f _ {i}, \{I \} _ {j \neq i}\right) \right]. +$$ + +Now consider when $\forall k\in [K],\mathcal{P}^k$ is a.s. continuous, where $\mathcal{P}$ has either been fixed in the frequentist setting or drawn in the Bayesian setting. By the definition of Algorithm 3 we have + +$$ +\begin{array}{l} \underset {\{X _ {i} \} _ {i} \sim \mathcal {P}} {\mathbb {E}} \left[ L _ {i} \left(f _ {i}, \{I \} _ {j \neq i}\right) \right] = \frac {1}{K} \sum_ {k = 1} ^ {K} \underset {\{X _ {i} \} _ {i} \sim \mathcal {P}} {\mathbb {E}} \left[ L _ {i} ^ {k} \left(f _ {i}, \{I \} _ {j \neq i}\right) \right] \\ = \frac {1}{K} \sum_ {k = 1} ^ {K} \underset {\{X _ {i} \} _ {i} \sim \mathcal {P}} {\mathbb {E}} \left[ \left(F _ {\left(Y _ {i} ^ {k}, W _ {i} ^ {k}\right)} \left(T _ {i} ^ {k}\right) - F _ {Z _ {i} ^ {k}} \left(T _ {i} ^ {k}\right)\right) ^ {2} \right] \\ = \frac {1}{K} \sum_ {k = 1} ^ {K} \underset {\left\{X _ {i} ^ {k} \right\} _ {i} \sim \mathcal {P} ^ {k}} {\mathbb {E}} \left[ \left(F _ {\left(Y _ {i} ^ {k}, W _ {i} ^ {k}\right)} \left(T _ {i} ^ {k}\right) - F _ {Z _ {i} ^ {k}} \left(T _ {i} ^ {k}\right)\right) ^ {2} \right]. \\ \end{array} +$$ + +Thus in the Bayesian setting we have + +$$ +\begin{array}{l} \underset {\mathcal {P} \sim \Pi} {\mathbb {E}} \left[ \underset {\{X _ {i} \} _ {i} \sim \mathcal {P}} {\mathbb {E}} \left[ L _ {i} (f _ {i}, \{I \} _ {j \neq i}) \right] \right] \\ = \frac {1}{K} \sum_ {k = 1} ^ {K} \underset {\mathcal {P} \sim \Pi} {\mathbb {E}} \left[ \underset {\left\{X _ {i} ^ {k} \right\} _ {i} \sim \mathcal {P} ^ {k}} {\mathbb {E}} \left[ \left(F _ {(Y _ {i} ^ {k}, W _ {i} ^ {k})} (T _ {i} ^ {k}) - F _ {Z _ {i} ^ {k}} (T _ {i} ^ {k})\right) ^ {2} \right] \right]. \\ \end{array} +$$ + +To get a lower bound we apply Lemma 5 followed Lemmas 3 then 2 which give + +$$ +\begin{array}{l} \frac {1}{K} \sum_ {k = 1} ^ {K} \underset {\mathcal {P} \sim \Pi} {\mathbb {E}} \left[ \underset {\left\{X _ {i} ^ {k} \right\} _ {i} \sim \mathcal {P} ^ {k}} {\mathbb {E}} \left[ \left(F _ {\left(Y _ {i} ^ {k}, W _ {i} ^ {k}\right)} \left(T _ {i} ^ {k}\right) - F _ {Z _ {i} ^ {k}} \left(T _ {i} ^ {k}\right)\right) ^ {2} \right] \right] \\ \geq \frac {1}{K} \sum_ {k = 1} ^ {K} \underset {\mathcal {P} \sim \Pi} {\mathbb {E}} \left[ \underset {\left\{X _ {i} ^ {k} \right\} _ {i} \sim \mathcal {P} ^ {k}} {\mathbb {E}} \left[ \left(\mathbb {E} \left[ F _ {\left(Y _ {i} ^ {k}, W _ {i} ^ {k}\right)} \left(T _ {i} ^ {k}\right) \mid X _ {i}, W _ {i}, T _ {i} ^ {k}, \mathcal {P} ^ {k} \right] - F _ {Z _ {i} ^ {k}} \left(T _ {i} ^ {k}\right)\right) ^ {2} \right] \right] \\ = \frac {1}{K} \sum_ {k = 1} ^ {K} \underset {\mathcal {P} \sim \Pi} {\mathbb {E}} \left[ \underset {\left\{X _ {i} ^ {k} \right\} _ {i} \sim \mathcal {P} ^ {k}} {\mathbb {E}} \left[ \left(F _ {\mathcal {P} ^ {k}} \left(T _ {i} ^ {k}\right) - F _ {Z _ {i} ^ {k}} \left(T _ {i} ^ {k}\right)\right) ^ {2} \right] \right] \\ = \frac {1}{K} \sum_ {k = 1} ^ {K} _ {\mathcal {P} \sim \Pi} ^ {\mathbb {E}} \left[ \frac {1}{6 \left| Z _ {i} ^ {k} \right|} \right] \\ = \frac {1}{6 \left| Z _ {i} \right|}. \\ \end{array} +$$ + +In the frequentist setting, independence and Lemma 1 give us + +$$ +\begin{array}{l} \underset {\left\{X _ {i} ^ {k} \right\} _ {i} \sim \mathcal {P} ^ {k}} {\mathbb {E}} \left[ \left(F _ {\left(Y _ {i} ^ {k}, W _ {i} ^ {k}\right)} \left(T _ {i} ^ {k}\right) - F _ {Z _ {i} ^ {k}} \left(T _ {i} ^ {k}\right)\right) ^ {2} \right] \\ = \underset {\left\{X _ {i} ^ {k} \right\} _ {i} \sim \mathcal {P} ^ {k}} {\mathbb {E}} \left[ \left(F _ {\left(Y _ {i} ^ {k}, W _ {i} ^ {k}\right)} \left(T _ {i} ^ {k}\right) - F _ {\mathcal {P} ^ {k}} \left(T _ {i} ^ {k}\right)\right) ^ {2} \right] + \underset {\left\{X _ {i} ^ {k} \right\} _ {i} \sim \mathcal {P} ^ {k}} {\mathbb {E}} \left[ \left(F _ {Z _ {i} ^ {k}} \left(T _ {i} ^ {k}\right) - F _ {\mathcal {P} ^ {k}} \left(T _ {i} ^ {k}\right)\right) ^ {2} \right] \\ = \underset {\left\{X _ {i} ^ {k} \right\} _ {i} \sim \mathcal {P} ^ {k}} {\mathbb {E}} \left[ \left(F _ {\left(Y _ {i} ^ {k}, W _ {i} ^ {k}\right)} \left(T _ {i} ^ {k}\right) - F _ {\mathcal {P} ^ {k}} \left(T _ {i} ^ {k}\right)\right) ^ {2} \right] + \frac {1}{6 \left| Z _ {i} ^ {k} \right|} \\ \geq \frac {1}{6 | Z _ {i} |}. \\ \end{array} +$$ + +Therefore, + +$$ +\frac {1}{6 \left| Z _ {i} \right|} \leq \sup _ {\mathcal {P} \in \mathcal {C}} \underset {\left\{X _ {i} \right\} _ {i} \sim \mathcal {P}} {\mathbb {E}} \left[ L _ {i} \left(f _ {i}, \left\{I \right\} _ {j \neq i}\right) \right]. +$$ + +Together we have the following lower bound in both the frequentist and Bayesian setting + +$$ +\frac {1}{6 \left| Z _ {i} \right|} \leq \underset {\mathcal {P} \sim \Pi} {\mathbb {E}} \left[ \underset {\{X _ {i} \} _ {i} \sim \mathcal {P}} {\mathbb {E}} \left[ L _ {i} \left(f _ {i}, \{I \} _ {j \neq i}\right) \right] \right], \sup _ {\mathcal {P} \in \mathcal {C}} \underset {\{X _ {i} \} _ {i} \sim \mathcal {P}} {\mathbb {E}} \left[ L _ {i} \left(f _ {i}, \{I \} _ {j \neq i}\right) \right]. +$$ + +From part two of Lemma 2 we have that when agents submit truthfully + +$$ +\begin{array}{l} \underset {\{X _ {i} \} _ {i} \sim \mathcal {P}} {\mathbb {E}} \left[ L _ {i} \left(\{I \} _ {i = 1} ^ {m}\right) \right] = \frac {1}{K} \sum_ {k = 1} ^ {K} \underset {\{X _ {i} \} _ {i} \sim \mathcal {P}} {\mathbb {E}} \left[ L _ {i} ^ {k} \left(\{I \} _ {i = 1} ^ {m}\right) \right] \\ = \frac {1}{K} \sum_ {k = 1} ^ {K} \mathbb {E} _ {\left\{X _ {i} \right\} _ {i} \sim \mathcal {P}} \left[ \left(F _ {\left(Y _ {i} ^ {k}, W _ {i} ^ {k}\right)} \left(T _ {i} ^ {k}\right) - F _ {Z _ {i} ^ {k}} \left(T _ {i} ^ {k}\right)\right) ^ {2} \right] \\ = \frac {1}{K} \sum_ {k = 1} ^ {K} \frac {1}{6} \left(\frac {1}{\left| Y _ {i} ^ {k} \right| + \left| W _ {i} ^ {k} \right|} + \frac {1}{\left| Z _ {i} ^ {k} \right|}\right) \\ = \frac {1}{6} \left(\frac {1}{\left| X _ {i} \right| + \left| W _ {i} \right|} + \frac {1}{\left| Z _ {i} \right|}\right) \\ \end{array} +$$ + +which implies that + +$$ +\underset {\mathcal {P} \sim \Pi} {\mathbb {E}} \left[ \underset {\{X _ {i} \} _ {i} \sim \mathcal {P}} {\mathbb {E}} [ L _ {i} (\{I \} _ {i = 1} ^ {m}) ] \right] = \sup _ {\mathcal {P} \in \mathcal {C}} \underset {\{X _ {i} \} _ {i} \sim \mathcal {P}} {\mathbb {E}} [ L _ {i} (\{I \} _ {i = 1} ^ {m}) ] = \frac {1}{6} \left(\frac {1}{| X _ {i} | + | W _ {i} |} + \frac {1}{| Z _ {i} |}\right). +$$ + +Combining this with the lower bounds, we conclude + +$$ +\begin{array}{l} \underset {\mathcal {P} \sim \Pi} {\mathbb {E}} \left[ \underset {\{X _ {i} \} _ {i} \sim \mathcal {P}} {\mathbb {E}} \left[ L _ {i} \left(\left\{I \right\} _ {i = 1} ^ {m}\right) \right] \right] - \underset {\mathcal {P} \sim \Pi} {\mathbb {E}} \left[ \underset {\{X _ {i} \} _ {i} \sim \mathcal {P}} {\mathbb {E}} \left[ L _ {i} \left(f _ {i}, \left\{I \right\} _ {j \neq i}\right) \right] \right] \leq \frac {1}{6 \left(| X _ {i} | + | W _ {i} |\right)} \\ \sup _ {\mathcal {P} \in \mathcal {C}} \underset {\{X _ {i} \} _ {i} \sim \mathcal {P}} {\mathbb {E}} \left[ L _ {i} \left(\left\{I \right\} _ {i = 1} ^ {m}\right) \right] - \sup _ {\mathcal {P} \in \mathcal {C}} \underset {\{X _ {i} \} _ {i} \sim \mathcal {P}} {\mathbb {E}} \left[ L _ {i} \left(f _ {i}, \left\{I \right\} _ {j \neq i}\right) \right] \leq \frac {1}{6 \left(| X _ {i} | + | W _ {i} |\right)} \\ \end{array} +$$ + +which completes the proof. + +![](images/e618bd5e5c0cb0e5e5a6a044d0b343d99d297b4831a792ebaa4d7a6d98836e8e.jpg) + +Proposition 12. For $n = (n_1, \ldots, n_m) \in \mathbb{N}^m$ let $L_i$ ( $n, \{f_i\}_{i=1}^m$ ) denote the value of $L_i$ in Algorithm 3 when agent $j \in [m]$ has $n_j$ data points and agents use $\{f_i\}_{i=1}^m \in \mathcal{F}^m$ . Then + +$$ +\begin{array}{l} \mathbb {E} \left[ L _ {i} \left(n, \{I \} _ {i = 1} ^ {m}\right) \right] - \mathbb {E} \left[ L _ {i} \left(n + e _ {i}, \{I \} _ {i = 1} ^ {m}\right) \right] \\ = \frac {1}{K} \sum_ {k = 1} ^ {K} \mathbb {E} \left[ F _ {\mathcal {P} ^ {k}} \left(T _ {i} ^ {k}\right) \left(1 - F _ {\mathcal {P} ^ {k}} \left(T _ {i} ^ {k}\right)\right) \right] \left(\frac {1}{n _ {i} + | W _ {i} |} - \frac {1}{n _ {i} + 1 + | W _ {i} |}\right) \\ \end{array} +$$ + +where $\mathcal{P}^k = \mathcal{P}\circ (\varphi^k)^{-1}$ + +Proof. By the definition of Algorithm 3 we have + +$$ +\mathbb {E} \left[ L _ {i} \left(n, \{I \} _ {i = 1} ^ {m}\right) \right] = \frac {1}{K} \sum_ {k = 1} ^ {K} \mathbb {E} \left[ L _ {i} ^ {k} \left(n, \{I \} _ {i = 1} ^ {m}\right) \right] = \frac {1}{K} \sum_ {k = 1} ^ {K} \mathbb {E} \left[ \left(F _ {\left(X _ {i} ^ {k}, W _ {i} ^ {k}\right)} \left(T _ {i} ^ {k}\right) - F _ {Z _ {i} ^ {k}} \left(T _ {i} ^ {k}\right)\right) ^ {2} \right]. +$$ + +Let $F_{\mathcal{P}^k}$ be the CDF for $\mathcal{P}^k$ . We start by rewriting each term in the sum above as + +$$ +\mathbb {E} \left[ \left(F _ {\left(X _ {i} ^ {k}, W _ {i} ^ {k}\right)} \left(T _ {i} ^ {k}\right) - F _ {Z _ {i} ^ {k}} \left(T _ {i} ^ {k}\right)\right) ^ {2} \right] = \underset {\mathcal {P} ^ {k}} {\mathbb {E}} \left[ \underset {X _ {i} ^ {k}, W _ {i} ^ {k}, Z _ {i} ^ {k}, T _ {i} ^ {k}} {\mathbb {E}} \left[ \left(F _ {\left(X _ {i} ^ {k}, W _ {i} ^ {k}\right)} \left(T _ {i} ^ {k}\right) - F _ {Z _ {i} ^ {k}} \left(T _ {i} ^ {k}\right)\right) ^ {2} \right] \right]. +$$ + +Following the same steps in Lemma 2 up to equation (5) gives us + +$$ +\begin{array}{l} \underset {X _ {i} ^ {k}, W _ {i} ^ {k}, Z _ {i} ^ {k}, T _ {i} ^ {k}} {\mathbb {E}} \left[ \left(F _ {\left(X _ {i} ^ {k}, W _ {i} ^ {k}\right)} \left(T _ {i} ^ {k}\right) - F _ {Z _ {i} ^ {k}} \left(T _ {i} ^ {k}\right)\right) ^ {2} \right] \\ = \underset {T _ {i} ^ {k}} {\mathbb {E}} \left[ F _ {\mathcal {P} ^ {k}} \left(T _ {i} ^ {k}\right) \left(1 - F _ {\mathcal {P} ^ {k}} \left(T _ {i} ^ {k}\right)\right) \right] \left(\frac {1}{n _ {i} + \left| W _ {i} \right|} + \frac {1}{\left| Z _ {i} \right|}\right). \\ \end{array} +$$ + +Therefore, + +$$ +\mathbb {E} \left[ L _ {i} \left(n, \{I \} _ {i = 1} ^ {m}\right) \right] = \frac {1}{K} \sum_ {k = 1} ^ {K} \mathbb {E} \left[ F _ {\mathcal {P} ^ {k}} \left(T _ {i} ^ {k}\right) \left(1 - F _ {\mathcal {P} ^ {k}} \left(T _ {i} ^ {k}\right)\right) \right] \left(\frac {1}{n _ {i} + | W _ {i} |} + \frac {1}{| Z _ {i} |}\right). +$$ + +The same argument gives an analogous result for $\mathbb{E}\left[L_i\left(n + e_i,\{I\}_{i = 1}^m\right)\right]$ . Taking the difference we find + +$$ +\begin{array}{l} \mathbb {E} \left[ L _ {i} \left(n, \{I \} _ {i = 1} ^ {m}\right) \right] - \mathbb {E} \left[ L _ {i} \left(n + e _ {i}, \{I \} _ {i = 1} ^ {m}\right) \right] \\ = \frac {1}{K} \sum_ {k = 1} ^ {K} \mathbb {E} \left[ F _ {\mathcal {P} ^ {k}} \left(T _ {i} ^ {k}\right) \left(1 - F _ {\mathcal {P} ^ {k}} \left(T _ {i} ^ {k}\right)\right) \right] \left(\frac {1}{n _ {i} + \left| W _ {i} \right|} - \frac {1}{n _ {i} + \left| W _ {i} \right| + 1}\right). \\ \end{array} +$$ + +![](images/69dd9c3f284d6583bfc727703f2e7cd7f8f050495fbb768f6829aa5ef5d0af29.jpg) + +Proposition 13. For $n = (n_1, \ldots, n_m) \in \mathbb{N}^m$ let $L_i$ ( $n, \{f_i\}_{i=1}^m$ ) denote the value of $L_i$ in Algorithm 3 when agent $j \in [m]$ has $n_j$ data points and agents use $\{f_i\}_{i=1}^m \in \mathcal{F}^m$ . Then + +$$ +\begin{array}{l} \sup _ {\mathcal {P} \in \mathcal {C}} \mathbb {E} \left[ L _ {i} \left(n, \{I \} _ {i = 1} ^ {m}\right) \right] - \sup _ {\mathcal {P} \in \mathcal {C}} \mathbb {E} \left[ L _ {i} \left(n + e _ {i}, \{I \} _ {i = 1} ^ {m}\right) \right] \\ = \left(\sup _ {\mathcal {P} \in \mathcal {C}} \frac {1}{K} \sum_ {k = 1} ^ {K} \mathbb {E} \left[ F _ {\mathcal {P} ^ {k}} \left(T _ {i} ^ {k}\right) \left(1 - F _ {\mathcal {P} ^ {k}} \left(T _ {i} ^ {k}\right)\right) \right]\right) \left(\frac {1}{n _ {i} + | W _ {i} |} - \frac {1}{n _ {i} + 1 + | W _ {i} |}\right) \\ \end{array} +$$ + +where $\mathcal{P}^k = \mathcal{P}\circ (\varphi^k)^{-1}$ + +Proof. By the definition of Algorithm 3 we have + +$$ +\begin{array}{l} \sup _ {\mathcal {P} \in \mathcal {C}} \mathbb {E} \left[ L _ {i} \left(n, \{I \} _ {i = 1} ^ {m}\right) \right] = \sup _ {\mathcal {P} \in \mathcal {C}} \frac {1}{K} \sum_ {k = 1} ^ {K} \mathbb {E} \left[ L _ {i} ^ {k} \left(n, \{I \} _ {i = 1} ^ {m}\right) \right] \\ = \sup _ {\mathcal {P} \in \mathcal {C}} \frac {1}{K} \sum_ {k = 1} ^ {K} \mathbb {E} \left[ \left(F _ {\left(X _ {i} ^ {k}, W _ {i} ^ {k}\right)} \left(T _ {i} ^ {k}\right) - F _ {Z _ {i} ^ {k}} \left(T _ {i} ^ {k}\right)\right) ^ {2} \right]. \\ \end{array} +$$ + +Let $F_{\mathcal{P}^k}$ be the CDF for $\mathcal{P}^k$ . Following the same steps in Lemma 2 up to equation (5) gives us + +$$ +\mathbb {E} \left[ \left(F _ {\left(X _ {i} ^ {k}, W _ {i} ^ {k}\right)} \left(T _ {i} ^ {k}\right) - F _ {Z _ {i} ^ {k}} \left(T _ {i} ^ {k}\right)\right) ^ {2} \right] = \underset {T _ {i} ^ {k}} {\mathbb {E}} \left[ F _ {\mathcal {P} ^ {k}} (T _ {i} ^ {k}) \left(1 - F _ {\mathcal {P} ^ {k}} (T _ {i} ^ {k})\right) \right] \left(\frac {1}{n _ {i} + | W _ {i} |} + \frac {1}{| Z _ {i} |}\right). +$$ + +Therefore, + +$$ +\sup _ {\mathcal {P} \in \mathcal {C}} \mathbb {E} \left[ L _ {i} \left(n, \{I \} _ {i = 1} ^ {m}\right) \right] = \sup _ {\mathcal {P} \in \mathcal {C}} \frac {1}{K} \sum_ {k = 1} ^ {K} \mathbb {E} \left[ F _ {\mathcal {P} ^ {k}} \left(T _ {i} ^ {k}\right) \left(1 - F _ {\mathcal {P} ^ {k}} \left(T _ {i} ^ {k}\right)\right) \right] \left(\frac {1}{n _ {i} + | W _ {i} |} + \frac {1}{| Z _ {i} |}\right). +$$ + +The same argument gives an analogous result for $\mathbb{E}\left[L_i\left(n + e_i,\{I\}_{i = 1}^m\right)\right]$ . Applying this to each feature, we find + +$$ +\begin{array}{l} \mathbb {E} \left[ L _ {i} \left(n, \{I \} _ {i = 1} ^ {m}\right) \right] - \mathbb {E} \left[ L _ {i} \left(n + e _ {i}, \{I \} _ {i = 1} ^ {m}\right) \right] \\ = \sup _ {\mathcal {P} \in \mathcal {C}} \frac {1}{K} \sum_ {k = 1} ^ {K} \mathbb {E} \left[ F _ {\mathcal {P} ^ {k}} \left(T _ {i} ^ {k}\right) \left(1 - F _ {\mathcal {P} ^ {k}} \left(T _ {i} ^ {k}\right)\right) \right] \left(\frac {1}{n _ {i} + \left| W _ {i} \right|} + \frac {1}{\left| Z _ {i} \right|}\right) \\ - \sup _ {\mathcal {P} \in \mathcal {C}} \frac {1}{K} \sum_ {k = 1} ^ {K} \mathbb {E} \left[ F _ {\mathcal {P} ^ {k}} \left(T _ {i} ^ {k}\right) \left(1 - F _ {\mathcal {P} ^ {k}} \left(T _ {i} ^ {k}\right)\right) \right] \left(\frac {1}{n _ {i} + \left| W _ {i} \right| + 1} + \frac {1}{\left| Z _ {i} \right|}\right) \\ = \left(\sup _ {\mathcal {P} \in \mathcal {C}} \frac {1}{K} \sum_ {k = 1} ^ {K} \mathbb {E} \left[ F _ {\mathcal {P} ^ {k}} \left(T _ {i} ^ {k}\right) \left(1 - F _ {\mathcal {P} ^ {k}} \left(T _ {i} ^ {k}\right)\right) \right]\right) \left(\frac {1}{n _ {i} + | W _ {i} |} - \frac {1}{n _ {i} + 1 + | W _ {i} |}\right). \\ \end{array} +$$ + +![](images/dd37c0a9a97adef63caaebc174ff29e24562041bfa0177ce1533fb9ffce56006.jpg) + +# E Examples of the conditional expectation in Algorithm 1 + +# E.1 The normal-normal model + +Proposition 14. Suppose that $\{f_i\}_{j\neq i} = \{I\}_{j\neq i}$ in Algorithm 1. Let $\mu \sim \mathcal{N}(a,b^2)$ and $X_{i} = \{X_{i,1},\ldots ,X_{i,n_{i}}\}$ , $Z_{i} = \{Z_{i,1},\dots,Z_{i,|Z_{i}|}\}$ , where $X_{i,j},T,Z_{i,j}|p^{i.i.d.}\mathcal{N}(\mu ,\sigma^2)$ , then + +$$ +\mathbb {E} \left[ F _ {Z _ {i}} (T) \mid X _ {i, 1}, \dots , X _ {i, n _ {i}}, T \right] = \Phi \left(\frac {T - \tilde {\mu}}{\sqrt {\sigma^ {2} + \tilde {\sigma} ^ {2}}}\right) +$$ + +where + +$$ +\tilde {\mu} = \frac {\frac {a}{b ^ {2}} + \frac {s u m (X _ {i , 1} , \ldots , X _ {i , n _ {i}} , T)}{\sigma^ {2}}}{\frac {1}{b ^ {2}} + \frac {n _ {i} + 1}{\sigma^ {2}}} \quad a n d \quad \tilde {\sigma} ^ {2} = \left(\frac {1}{b ^ {2}} + \frac {(n _ {i} + 1)}{\sigma^ {2}}\right) ^ {- 1}. +$$ + +Proof. Start by noticing that the conditional expectation can be rewritten as + +$$ +\begin{array}{l} \mathbb {E} \left[ F _ {Z _ {i}} (T) \mid X _ {i, 1}, \dots , X _ {i, n _ {i}}, T \right] = \mathbb {E} \left[ \frac {1}{| Z _ {i} |} \sum_ {z \in Z _ {i}} 1 _ {\{z \leq T \}} \mid X _ {i, 1}, \dots , X _ {i, n _ {i}}, T \right] \\ = \mathbb {E} \left[ 1 _ {\{Z _ {i, 1} \leq T \}} \mid X _ {i, 1}, \dots , X _ {i, n _ {i}}, T \right] \\ = \mathbb {E} \left[ \mathbb {E} \left[ 1 _ {\{Z _ {i, 1} \leq T \}} \mid \mu , X _ {i, 1}, \dots , X _ {i, n _ {i}}, T \right] \mid X _ {i, 1}, \dots , X _ {i, n _ {i}}, T \right]. \tag {2} \\ \end{array} +$$ + +where the last line follows from the tower property. By the definition of our model we know that + +$$ +\mathbb {E} \left[ 1 _ {\{Z _ {i, 1} \leq T \}} \mid \mu , X _ {i, 1}, \dots , X _ {i, n _ {i}}, T \right] = \Phi \left(\frac {T - \mu}{\sigma}\right). +$$ + +Recall from standard normal-normal conjugacy arguments that + +$$ +\mu | X _ {i, 1}, \dots X _ {i, n _ {i}}, T \sim \mathcal {N} (\tilde {\mu}, \tilde {\sigma} ^ {2}) \quad \text {w h e r e} +$$ + +$$ +\tilde {\mu} = \frac {\frac {a}{b ^ {2}} + \frac {\operatorname {s u m} \left(X _ {i , 1} , \dots , X _ {i , n _ {i}} , T\right)}{\sigma^ {2}}}{\frac {1}{b ^ {2}} + \frac {n _ {i} + 1}{\sigma^ {2}}} \quad \text {a n d} \quad \tilde {\sigma} ^ {2} = \left(\frac {1}{b ^ {2}} + \frac {(n _ {i} + 1)}{\sigma^ {2}}\right) ^ {- 1}. +$$ + +Therefore, we can write (2) as + +$$ +\mathbb {E} \left[ \Phi \left(\frac {T - \mu}{\sigma}\right) \mid X _ {i, 1}, \dots , X _ {i, n _ {i}}, T \right] = \int_ {- \infty} ^ {\infty} \Phi \left(\frac {T - \mu}{\sigma}\right) \phi_ {\tilde {\mu}, \tilde {\sigma} ^ {2}} (\mu) d \mu . +$$ + +where $\phi_{\tilde{\mu},\tilde{\sigma}^2}$ is the PDF of a normal distribution with mean $\tilde{\mu}$ and variance $\tilde{\sigma}^2$ . Recall the following Gaussian integral formula + +$$ +\int_ {- \infty} ^ {\infty} \Phi (\alpha - \beta x) \phi (x) d x = \Phi \left(\frac {\alpha}{\sqrt {1 + \beta^ {2}}}\right). +$$ + +By the change of variables $x = \frac{\mu - \tilde{\mu}}{\tilde{\sigma}}$ we get $\Phi \left(\frac{T - \mu}{\sigma}\right) = \Phi \left(\frac{T - \tilde{\mu}}{\sigma} -\frac{\tilde{\sigma}}{\sigma} x\right)$ , so applying the formula gives us + +$$ +\int_ {- \infty} ^ {\infty} \Phi \left(\frac {T - \mu}{\sigma}\right) \phi_ {\tilde {\mu}, \tilde {\sigma} ^ {2}} (\mu) d \mu = \Phi \left(\frac {T - \tilde {\mu}}{\sqrt {\sigma^ {2} + \tilde {\sigma} ^ {2}}}\right). +$$ + +Therefore, + +$$ +\mathbb {E} \left[ F _ {Z _ {i}} (T) \mid X _ {i, 1}, \dots , X _ {i, n _ {i}}, T \right] = \Phi \left(\frac {T - \tilde {\mu}}{\sqrt {\sigma^ {2} + \tilde {\sigma} ^ {2}}}\right). +$$ + +# E.2 The beta-bernoulli model + +Proposition 15. Suppose that $\{f_i\}_{j \neq i} = \{I\}_{j \neq i}$ in Algorithm 1. Let $p \sim \text{Beta}(\alpha, \beta)$ and $X_i = \{X_{i,1}, \ldots, X_{i,n_i}\}$ , $Z_i = \{Z_{i,1}, \ldots, Z_{i,|Z_i|}\}$ , where $X_{i,j}, T, Z_{i,j}|p^{\text{i.i.d.}}$ Bern $(p)$ , then + +$$ +\mathbb {E} \left[ F _ {Z _ {i}} (T) \mid X _ {i, 1}, \dots , X _ {i, n _ {i}}, T \right] = T + (1 - T) \frac {\beta + (n _ {i} + 1) - \operatorname {s u m} \left(X _ {i , 1} , \dots , X _ {i , n _ {i}}\right)}{\alpha + \beta + (n _ {i} + 1)}. +$$ + +Proof. Start by noticing that the conditional expectation can be rewritten as + +$$ +\begin{array}{l} \mathbb {E} \left[ F _ {Z _ {i}} (T) \mid X _ {i, 1}, \dots , X _ {i, n _ {i}}, T \right] = \mathbb {E} \left[ \frac {1}{| Z _ {i} |} \sum_ {z \in Z _ {i}} 1 _ {\{z \leq T \}} \mid X _ {i, 1}, \dots , X _ {i, n _ {i}}, T \right] \\ = \mathbb {E} \left[ 1 _ {\{Z _ {i, 1} \leq T \}} \mid X _ {i, 1}, \dots , X _ {i, n _ {i}}, T \right] \\ = P \left(Z _ {i, 1} \leq T \mid X _ {i, 1}, \dots , X _ {i, n _ {i}}, T\right). \\ \end{array} +$$ + +The law of total probability tells us that + +$$ +\begin{array}{l} P \left(Z _ {i, 1} \leq T \mid X _ {i, 1}, \dots , X _ {i, n _ {i}}, T\right) \\ = \int P \left(Z _ {i, 1} \leq T \mid p, X _ {i, 1}, \dots , X _ {i, n _ {i}}, T\right) d P \left(p \mid X _ {i, 1}, \dots , X _ {i, n _ {i}}, T\right). \tag {3} \\ \end{array} +$$ + +We now consider two cases based on whether $T$ is 0 or 1. When $T = 1$ , (3) becomes + +$$ +P \left(Z _ {i, 1} \leq T \mid X _ {i, 1}, \dots , X _ {i, n _ {i}}, T = 1\right) = \int 1 \cdot d P \left(p \mid X _ {i, 1}, \dots , X _ {i, n _ {i}}, T\right) = 1. +$$ + +When $T = 0$ , recall from standard Beta-Bernoulli conjugacy arguments that + +$$ +\begin{array}{l} p \mid X _ {i, 1}, \dots , X _ {i, n _ {i}}, T \\ \sim \operatorname {B e t a} (\alpha + \operatorname {s u m} (X _ {i, 1}, \dots , X _ {i, n _ {i}}, T), \beta + (n _ {i} + 1) - \operatorname {s u m} (X _ {i, 1}, \dots , X _ {i, n _ {i}}, T)) \\ = \operatorname {B e t a} \left(\underbrace {\alpha + \operatorname {s u m} \left(X _ {i , 1} , \ldots , X _ {i , n _ {i}}\right)} _ {\alpha_ {0} :=}, \underbrace {\beta + (n _ {i} + 1) - \operatorname {s u m} \left(X _ {i , 1} , \ldots , X _ {i , n _ {i}}\right)} _ {\beta_ {0} :=}\right). \\ \end{array} +$$ + +Also observe that when $T = 0$ , $P(Z_{i,1} \leq T \mid p, X_{i,1}, \ldots, X_{i,n_i}, T) = 1 - p$ . Therefore, (3) becomes + +$$ +\begin{array}{l} \int P \left(Z _ {i, 1} \leq T \mid p, X _ {i, 1}, \dots , X _ {i, n _ {i}}, T\right) d P \left(p \mid X _ {i, 1}, \dots , X _ {i, n _ {i}}, T\right) \\ = \int (1 - p) \frac {p ^ {1 - \alpha_ {0}} (1 - p) ^ {1 - \beta_ {0}}}{B (\alpha_ {0} , \beta_ {0})}. \\ \end{array} +$$ + +Recall that if $Z\sim \mathrm{Beta}(\alpha_0,\beta_0)$ then + +$$ +\frac {\alpha_ {0}}{\alpha_ {0} + \beta_ {0}} = \mathbb {E} [ Z ] = \int (1 - z) \frac {z ^ {1 - \alpha_ {0}} (1 - z) ^ {1 - \beta_ {0}}}{B (\alpha_ {0} , \beta_ {0})}. +$$ + +Therefore, + +$$ +\begin{array}{l} \int (1 - p) \frac {p ^ {1 - \alpha_ {0}} (1 - p) ^ {1 - \beta_ {0}}}{B (\alpha_ {0} , \beta_ {0})} = \int \frac {p ^ {1 - \alpha_ {0}} (1 - p) ^ {1 - \beta_ {0}}}{B (\alpha_ {0} , \beta_ {0})} - \int p \frac {p ^ {1 - \alpha_ {0}} (1 - p) ^ {1 - \beta_ {0}}}{B (\alpha_ {0} , \beta_ {0})} \\ = 1 - \frac {\alpha_ {0}}{\alpha_ {0} + \beta_ {0}} \\ = \frac {\beta_ {0}}{\alpha_ {0} + \beta_ {0}} \\ \end{array} +$$ + +so we conclude that + +$$ +P \left(Z _ {i, 1} \leq T \mid X _ {i, 1}, \dots , X _ {i, n _ {i}}, T = 0\right) = \frac {\beta + (n _ {i} + 1) - \operatorname {s u m} \left(X _ {i , 1} , \dots , X _ {i , n _ {i}}\right)}{\alpha + \beta + (n _ {i} + 1)}. +$$ + +Putting both cases together gives us + +$$ +\mathbb {E} \left[ F _ {Z _ {i}} (T) \mid X _ {i, 1}, \dots , X _ {i, n _ {i}}, T \right] = T + (1 - T) \frac {\beta + (n _ {i} + 1) - \operatorname {s u m} \left(X _ {i , 1} , \dots , X _ {i , n _ {i}}\right)}{\alpha + \beta + (n _ {i} + 1)}. +$$ + +![](images/0b7661244b18018bd3d28d3dc58de8841b83fd19328874b1a27bd020088f61fd.jpg) + +# F Proofs of Technical results + +In this section we derive a series of technical results which aid in the main proofs. + +Lemma 1. Let $\mathcal{P} \in \mathcal{M}_1^c(\mathbb{R})$ be a continuous probability distribution over $\mathbb{R}$ , and $X = \{X_1, \ldots, X_n\}$ , where $X_i, T \stackrel{i.i.d.}{\sim} \mathcal{P}$ . Then + +$$ +\mathbb {E} \left[ \left(F _ {X} (T) - F _ {\mathcal {P}} (T)\right) ^ {2} \right] = \frac {1}{6 n}. +$$ + +Proof. Notice that for a fixed $t \in \mathbb{R}$ , + +$$ +\mathbb {E} \left[ F _ {X} (t) \right] = \frac {1}{n} \sum_ {i = 1} ^ {n} \mathbb {E} \left[ 1 _ {\{X _ {i} \leq t \}} \right] = F _ {\mathcal {P}} (t). +$$ + +Using this observation and noticing that $1_{\{X_i \leq T\}} |T \sim \operatorname{Bern}(F_{\mathcal{P}}(T))$ gives + +$$ +\begin{array}{l} \mathbb {E} \left[ \left(F _ {X} (T) - F _ {\mathcal {P}} (T)\right) ^ {2} \right] = \mathbb {E} _ {T} \left[ \mathbb {E} _ {X} \left[ \left(F _ {X} (T) - F _ {\mathcal {P}} (T)\right) ^ {2} | T \right] \right] \\ = \mathbb {E} _ {T} [ \operatorname {V a r} (F _ {X} (T) | T) ] \\ = \mathbb {E} _ {T} \left[ \frac {F _ {\mathcal {P}} (T) (1 - F _ {\mathcal {P}} (T))}{n} \right] \\ = \int_ {- \infty} ^ {\infty} \frac {F _ {\mathcal {P}} (T) (1 - F _ {\mathcal {P}} (T))}{n} d \mathcal {P} (T). \\ \end{array} +$$ + +Since $\mathcal{P}$ is continuous, the probability integral transform (Lemma 6) tells us that if we set $U\coloneqq F_{\mathcal{P}}(T)$ then $U\sim \mathrm{Unif}(0,1)$ . The above equation can now be written as + +$$ +\int_ {- \infty} ^ {\infty} \frac {F _ {\mathcal {P}} (T) (1 - F _ {\mathcal {P}} (T))}{n} d \mathcal {P} (T) = \int_ {0} ^ {1} \frac {U (1 - U)}{n} d U = \frac {1}{6 n} +$$ + +which concludes the proof. + +![](images/4c09d4e8babd2759bd410550b64c1ede210f68ac8acb40af36e98b752f19f5bc.jpg) + +Lemma 2. Let $\mathcal{P} \in \mathcal{M}_1(\mathbb{R})$ be a probability distribution over $\mathbb{R}$ , and $X = \{X_1, \ldots, X_n\}$ , $Y = \{Y_1, \ldots, Y_m\}$ where $X_i, Y_i, T \stackrel{i.i.d.}{\sim} \mathcal{P}$ . Then + +$$ +\mathbb {E} \left[ \left(F _ {X} (T) - F _ {Y} (T)\right) ^ {2} \right] \leq \frac {1}{4} \left(\frac {1}{n} + \frac {1}{m}\right). +$$ + +Moreover, when $\mathcal{P} \in \mathcal{M}_1^c(\mathbb{R})$ + +$$ +\mathbb {E} \left[ \left(F _ {X} (T) - F _ {Y} (T)\right) ^ {2} \right] = \frac {1}{6} \left(\frac {1}{n} + \frac {1}{m}\right). +$$ + +Proof. We start with proving the inequality. Let $F_{\mathcal{P}}(t)$ be the CDF of $\mathcal{P}$ . Notice that for a fixed $t \in \mathbb{R}$ , $\mathbb{E}\left[F_X(t)\right] = \frac{1}{n}\sum_{i = 1}^{n}\mathbb{E}\left[1_{\{X_i \leq t\}}\right] = F_{\mathcal{P}}(t)$ . Together with independence we have + +$$ +\begin{array}{l} \mathbb {E} \left[ \big (F _ {X} (T) - F _ {Y} (T) \big) ^ {2} \right] \\ = \mathbb {E} _ {T} \left[ \mathbb {E} _ {X, Y} \left[ \left(F _ {X} (T) - F _ {Y} (T)\right) ^ {2} | T \right] \right] \\ = \mathbb {E} _ {T} \left[ \mathbb {E} _ {X, Y} \left[ \left(F _ {X} (T) - F _ {\mathcal {P}} (T) + F _ {\mathcal {P}} (T) - F _ {Y} (T)\right) ^ {2} | T \right] \right] \\ = \mathbb {E} _ {T} \left[ \mathbb {E} _ {X} \left[ \left(F _ {X} (T) - F _ {\mathcal {P}} (T)\right) ^ {2} | T \right] + \mathbb {E} _ {Y} \left[ \left(F _ {\mathcal {P}} (T) - F _ {Y} (T)\right) ^ {2} | T \right] \right] \tag {4} \\ = \mathbb {E} _ {T} \left[ \operatorname {V a r} \left(F _ {X} (T) | T\right) + \operatorname {V a r} \left(F _ {Y} (T) | T\right) \right]. \\ \end{array} +$$ + +Given $T$ , $F_{X}(T)$ and $F_{Y}(T)$ are sums of i.i.d. bernoulli random variables, thus + +$$ +\begin{array}{l} \mathbb {E} _ {T} \left[ \operatorname {V a r} \left(F _ {X} (T) | T\right) + \operatorname {V a r} \left(F _ {Y} (T) | T\right) \right] = \mathbb {E} _ {T} \left[ \frac {F _ {\mathcal {P}} (T) (1 - F _ {\mathcal {P}} (T))}{n} + \frac {F _ {\mathcal {P}} (T) (1 - F _ {\mathcal {P}} (T))}{m} \right] \\ = \mathbb {E} _ {T} \left[ F _ {\mathcal {P}} (T) \left(1 - F _ {\mathcal {P}} (T)\right) \right] \left(\frac {1}{n} + \frac {1}{m}\right) \tag {5} \\ \leq \frac {1}{4} \left(\frac {1}{n} + \frac {1}{m}\right). \\ \end{array} +$$ + +since $F_{\mathcal{P}}(T)\in [0,1]$ + +For the equality, we rewrite (4) and apply Lemma 1 twice to get + +$$ +\begin{array}{l} \mathbb {E} _ {T} \left[ \mathbb {E} _ {X} \left[ \left(F _ {X} (T) - F _ {\mathcal {P}} (T)\right) ^ {2} \right] + \mathbb {E} _ {Y} \left[ \left(F _ {\mathcal {P}} (T) - F _ {Y} (T)\right) ^ {2} \right] \right] \\ = \mathbb {E} _ {X, T} \left[ \left(F _ {X} (T) - F _ {\mathcal {P}} (T)\right) ^ {2} \right] + \mathbb {E} _ {Y, T} \left[ \left(F _ {\mathcal {P}} (T) - F _ {Y} (T)\right) ^ {2} \right] \\ = \frac {1}{6} \left(\frac {1}{n} + \frac {1}{m}\right). \\ \end{array} +$$ + +![](images/bb122ee99c2da0b348d3e520f9bd317062262aee691565d830b86dc483909397.jpg) + +Lemma 3. Let $\Pi \in \mathcal{M}_1(\mathcal{M}_1(\mathbb{R}))$ be a distribution over the collection of $\mathbb{R}$ -valued distributions. Suppose that $\mathcal{P} \sim \Pi$ and then $X = \{X_1, \ldots, X_n\}, Y = \{Y_1, \ldots, Y_m\}$ where $X_i, Y_i, T \stackrel{i.i.d.}{\sim} \mathcal{P}$ . Let $F_{\mathcal{P}}(t)$ be the CDF of $\mathcal{P}$ . Then, + +$$ +\mathbb {E} \left[ F _ {Y} (T) \mid X _ {1}, \dots , X _ {n}, T, \mathcal {P} \right] = F _ {\mathcal {P}} (T). +$$ + +Proof. Using conditional independence we have + +$$ +\begin{array}{l} \mathbb {E} \left[ F _ {Y} (T) \mid X _ {1}, \dots , X _ {n}, T, \mathcal {P} \right] = \frac {1}{m} \sum_ {j = 1} ^ {m} \mathbb {E} \left[ 1 _ {\{Y _ {j} \leq T \}} \mid X _ {1}, \dots , X _ {n}, T, \mathcal {P} \right] \\ = \frac {1}{m} \sum_ {j = 1} ^ {m} \mathbb {E} \left[ 1 _ {\{Y _ {j} \leq T \}} | T, \mathcal {P} \right] \\ = \frac {1}{m} \sum_ {j = 1} ^ {m} F _ {\mathcal {P}} (T) \\ = F _ {\mathcal {P}} (T). \\ \end{array} +$$ + +![](images/487afde731ec0c7447b115e0cc332c689be45ddf0ea58edaea6d190230a1bf78.jpg) + +Lemma 4. Let $\mathcal{F} \subseteq \mathcal{G}$ , suppose $X \in L^2$ , and define $Y = \mathbb{E}[X|\mathcal{G}]$ then + +$$ +\mathbb {E} \left[ \operatorname {V a r} \left(X | \mathcal {F}\right) \right] - \mathbb {E} \left[ \operatorname {V a r} \left(X | \mathcal {G}\right) \right] = \mathbb {E} \left[ \operatorname {V a r} \left(Y | \mathcal {F}\right) \right]. +$$ + +Proof. Applying the law of total variation with respect to $\mathcal{F}$ and $\mathcal{G}$ gives us + +$$ +\operatorname {V a r} (X) = \mathbb {E} [ \operatorname {V a r} (X | \mathcal {G}) ] + \operatorname {V a r} (\mathbb {E} [ X | \mathcal {G} ]) \tag {6} +$$ + +$$ +\operatorname {V a r} (X) = \mathbb {E} \left[ \operatorname {V a r} (X | \mathcal {F}) \right] + \operatorname {V a r} (\mathbb {E} [ X | \mathcal {F} ]). \tag {7} +$$ + +Subtracting (6) from (7) gives + +$$ +\mathbb {E} \left[ \operatorname {V a r} (X | \mathcal {F}) \right] - \mathbb {E} \left[ \operatorname {V a r} (X | \mathcal {G}) \right] = \operatorname {V a r} (\mathbb {E} [ X | \mathcal {G} ]) - \operatorname {V a r} (\mathbb {E} [ X | \mathcal {F} ]). \tag {8} +$$ + +Now notice that by the tower property we have + +$$ +\mathbb {E} \left[ Y | \mathcal {F} \right] = \mathbb {E} \left[ \mathbb {E} \left[ X | \mathcal {G} \right] | \mathcal {F} \right] = \mathbb {E} \left[ X | \mathcal {F} \right]. +$$ + +Combining this with another application of the law of total variation yields + +$$ +\begin{array}{l} \operatorname {V a r} \left(\mathbb {E} \left[ X | \mathcal {G} \right]\right) = \operatorname {V a r} (Y) = \mathbb {E} \left[ \operatorname {V a r} \left(Y | \mathcal {F}\right) \right] + \operatorname {V a r} \left(\mathbb {E} \left[ Y | \mathcal {F} \right]\right) \\ = \mathbb {E} \left[ \operatorname {V a r} \left(Y | \mathcal {F}\right) \right] + \operatorname {V a r} \left(\mathbb {E} \left[ X | \mathcal {F} \right]\right). \\ \end{array} +$$ + +Plugging this into the right hand side of (8) gives us + +$$ +\begin{array}{l} \mathbb {E} \left[ \operatorname {V a r} (X | \mathcal {F}) \right] - \mathbb {E} \left[ \operatorname {V a r} (X | \mathcal {G}) \right] = (\mathbb {E} \left[ \operatorname {V a r} (Y | \mathcal {F}) \right] + \operatorname {V a r} (\mathbb {E} [ X | \mathcal {F} ])) - \operatorname {V a r} (\mathbb {E} [ X | \mathcal {F} ]) \\ = \mathbb {E} \left[ \operatorname {V a r} \left(Y | \mathcal {F}\right) \right]. \\ \end{array} +$$ + +![](images/3520ea0c517f9526069da053f76bdba44472ff926d83541c548c7274bb9da01b.jpg) + +# G Known results + +In this section we present two well known results and give proofs of them for completeness. + +Lemma 5 (Durrett [46] Theorem 4.1.15). Let $X$ be a random variable such that $\mathbb{E}\left[X^2\right] < \infty$ and $\mathcal{F}$ be a $\sigma$ -algebra on the underlying probability space. Then $\mathbb{E}\left[X|\mathcal{F}\right]$ is the $\mathcal{F}$ -measurable random variable $Y$ which minimizes $\mathbb{E}\left[(X - Y)^2\right]$ . + +Proof. Notice that if $Z$ is $\mathcal{F}$ -measurable and $\mathbb{E}[Z^2] < \infty$ then $Z \cdot \mathbb{E}[X|\mathcal{F}] = \mathbb{E}[Z \cdot X|\mathcal{F}]$ which implies + +$$ +\mathbb {E} \left[ Z \cdot \mathbb {E} \left[ X | \mathcal {F} \right] \right] = \mathbb {E} \left[ \mathbb {E} \left[ Z \cdot X | \mathcal {F} \right] \right] = \mathbb {E} [ Z \cdot X ]. +$$ + +Rearranging we find + +$$ +\mathbb {E} \left[ Z \cdot \left(X - \mathbb {E} [ X | \mathcal {F} ]\right) \right] = 0. +$$ + +Now suppose that $Y$ is $\mathcal{F}$ -measurable and $\mathbb{E}[Y^2] < \infty$ , and define $Z = \mathbb{E}[X|\mathcal{F}] - Y$ . Then, + +$$ +\begin{array}{l} \mathbb {E} \left[ (X - Y) ^ {2} \right] = \mathbb {E} \left[ (X - \mathbb {E} [ X | \mathcal {F} ] + Z) ^ {2} \right] \\ = \mathbb {E} \left[ (X - \mathbb {E} [ X | \mathcal {F} ]) ^ {2} \right] + \mathbb {E} \left[ Z \cdot (X - \mathbb {E} [ X | \mathcal {F} ]) \right] + \mathbb {E} [ Z ^ {2} ] \\ = \mathbb {E} \left[ \left(X - \mathbb {E} [ X | \mathcal {F} ]\right) ^ {2} \right] + \mathbb {E} [ Z ^ {2} ] \\ \end{array} +$$ + +which implies that the mean squared error is minimized when $Y = \mathbb{E}[X|\mathcal{F}]$ . + +![](images/d9865e1fb9704d390def09303a3f17c2af6d61346fb9f137b8f2fd4b6eada20d.jpg) + +Lemma 6 (Probability integral transform). Suppose that $X$ is a continuous $\mathbb{R}$ -valued random variable. Let $U = F_{X}(X)$ , i.e. the CDF of $X$ evaluated at $X$ . Then $U \sim \text{Unif}(0,1)$ . + +Proof. As $F_{X}(t)$ may not be strictly increasing, define the generalized inverse CDF $\widetilde{F}^{-1}(u) = \inf \{ t \in \mathbb{R} : F_{X}(t) \geq u \}$ . Now notice that we can write the CDF of $U$ as + +$$ +F _ {U} (t) = P (U \leq t) = P \left(F _ {X} (X) \leq t\right) = P (X \leq \widetilde {F} ^ {- 1} (t)) = F _ {X} (\widetilde {F} ^ {- 1} (t)) = t +$$ + +from which we conclude that $U\sim \mathrm{Unif}(0,1)$ \ No newline at end of file diff --git "a/NeurIPS/2025/A Cram\303\251r\342\200\223von Mises Approach to Incentivizing Truthful Data Sharing/images.zip" "b/NeurIPS/2025/A Cram\303\251r\342\200\223von Mises Approach to Incentivizing Truthful Data Sharing/images.zip" new file mode 100644 index 0000000000000000000000000000000000000000..b1bb93152364516b3d244810ea60be44402c553b --- /dev/null +++ "b/NeurIPS/2025/A Cram\303\251r\342\200\223von Mises Approach to Incentivizing Truthful Data Sharing/images.zip" @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b1eb727ded7b777946e90cc1386918ad83d4aa9317e548e79c3f9504fe757410 +size 2027739 diff --git "a/NeurIPS/2025/A Cram\303\251r\342\200\223von Mises Approach to Incentivizing Truthful Data Sharing/layout.json" "b/NeurIPS/2025/A Cram\303\251r\342\200\223von Mises Approach to Incentivizing Truthful Data Sharing/layout.json" new file mode 100644 index 0000000000000000000000000000000000000000..2dd5872110dc88563979983ad2634607fe7e658e --- /dev/null +++ "b/NeurIPS/2025/A Cram\303\251r\342\200\223von Mises Approach to Incentivizing Truthful Data Sharing/layout.json" @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:56c57e5a6165700664e83dcc496109a8db39ca0a16f4c5436b8b46061b2165ae +size 1722844 diff --git a/NeurIPS/2025/A Data-Driven Prism_ Multi-View Source Separation with Diffusion Model Priors/053c7814-81b9-435e-a72e-acccf192302f_content_list.json b/NeurIPS/2025/A Data-Driven Prism_ Multi-View Source Separation with Diffusion Model Priors/053c7814-81b9-435e-a72e-acccf192302f_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..ad6005904d36e298e761e4a393aa409aa227044b --- /dev/null +++ b/NeurIPS/2025/A Data-Driven Prism_ Multi-View Source Separation with Diffusion Model Priors/053c7814-81b9-435e-a72e-acccf192302f_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d7173b163ca1b0c594523bdc100d7e6e39f328383207e612f0d99a990f8d97b6 +size 180494 diff --git a/NeurIPS/2025/A Data-Driven Prism_ Multi-View Source Separation with Diffusion Model Priors/053c7814-81b9-435e-a72e-acccf192302f_model.json b/NeurIPS/2025/A Data-Driven Prism_ Multi-View Source Separation with Diffusion Model Priors/053c7814-81b9-435e-a72e-acccf192302f_model.json new file mode 100644 index 0000000000000000000000000000000000000000..5fcba77cd23632e00c5e4ecf2b17cc62119df915 --- /dev/null +++ b/NeurIPS/2025/A Data-Driven Prism_ Multi-View Source Separation with Diffusion Model Priors/053c7814-81b9-435e-a72e-acccf192302f_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cb90046e0f452c49e00a0ffd038db5abe56e6fd13adf1b641cb44db26d76a067 +size 231184 diff --git a/NeurIPS/2025/A Data-Driven Prism_ Multi-View Source Separation with Diffusion Model Priors/053c7814-81b9-435e-a72e-acccf192302f_origin.pdf b/NeurIPS/2025/A Data-Driven Prism_ Multi-View Source Separation with Diffusion Model Priors/053c7814-81b9-435e-a72e-acccf192302f_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..ff543d2e8d8ca7b3c370156e5a265f7af00183c9 --- /dev/null +++ b/NeurIPS/2025/A Data-Driven Prism_ Multi-View Source Separation with Diffusion Model Priors/053c7814-81b9-435e-a72e-acccf192302f_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:51799c50b8e0c0ccb5ab272cd13896b9b8a16d6af517c7d81f0fa83426752b74 +size 4378860 diff --git a/NeurIPS/2025/A Data-Driven Prism_ Multi-View Source Separation with Diffusion Model Priors/full.md b/NeurIPS/2025/A Data-Driven Prism_ Multi-View Source Separation with Diffusion Model Priors/full.md new file mode 100644 index 0000000000000000000000000000000000000000..c49a876afb9753e96d48c307709983869b83e9ce --- /dev/null +++ b/NeurIPS/2025/A Data-Driven Prism_ Multi-View Source Separation with Diffusion Model Priors/full.md @@ -0,0 +1,793 @@ +# A Data-Driven Prism: Multi-View Source Separation with Diffusion Model Priors + +Sebastian Wagner-Carena $^{1,2,*}$ swagnercarena@flatironinstitute.org + +Aizhan Akhmetzhanova $^{1,3,*}$ aakhmetzhanova@g.harvard.edu + +Sydney Erickson4 sydney3@stanford.edu + +1 Center for Computational Astrophysics, Flatiron Institute +$^{2}$ Center for Data Science, New York University +$^{3}$ Department of Physics, Harvard University +$^{4}$ Department of Physics, Stanford University + +# Abstract + +A common challenge in the natural sciences is to disentangle distinct, unknown sources from observations. Examples of this source separation task include deblending galaxies in a crowded field, distinguishing the activity of individual neurons from overlapping signals, and separating seismic events from an ambient background. Traditional analyses often rely on simplified source models that fail to accurately reproduce the data. Recent advances have shown that diffusion models can directly learn complex prior distributions from noisy, incomplete data. In this work, we show that diffusion models can solve the source separation problem without explicit assumptions about the source. Our method relies only on multiple views, or the property that different sets of observations contain different linear transformations of the unknown sources. We show that our method succeeds even when no source is individually observed and the observations are noisy, incomplete, and vary in resolution. The learned diffusion models enable us to sample from the source priors, evaluate the probability of candidate sources, and draw from the joint posterior of the source distribution given an observation. We demonstrate the effectiveness of our method on a range of synthetic problems as well as real-world galaxy observations. + +# 1 Introduction + +For scientific data, pristine, isolated observations are rare: images of galaxies come blended with other luminous sources [1-3], electrodes measuring brain activity sum multiple neurons [4-6], and seismometers registering earthquakes contend with a constant seismic background [7, 8]. Additionally, the observations are often incomplete and collected by a heterogeneous set of instruments, each with unique resolutions. The corrupted data is rarely directly usable. Instead, leveraging these datasets for scientific discovery requires solving a source separation problem to either learn the unknown source prior [9, 10] or constrain the posteriors for individual sources given an observation [1, 6, 7]. In this work, we address the general challenge of multi-view source separation (MVSS). + +Most source separation methods, including ICA-based methods [11-13], non-negative matrix factorization methods [14-16], and template-fitting methods [17-19], require strong prior assumptions + +about the sources. Similarly, most deep-learning-based methods require access to samples from the source priors to generate training sets [20-25]. When the source distributions are not well-understood, this poses a degeneracy: isolating and measuring the source signals requires a source prior, but constraining the source prior requires isolated measurements of the sources. + +Alternatively, some source separation methods assume a known mixing process and thereby relax the need for a source prior [26-29]. To break the degeneracies between the sources, these methods rely on distinct collections of observations, or views, with each view offering a different linear mixture of the underlying sources. These works focus on contrastive datasets, where the goal is to separate a signal that is enriched in a target view compared to a background view. While relevant for a number of scientific datasets, these source separation methods are either limited in their expressivity [28, 29] or are not designed for incomplete data [26]. Additionally, the contrastive assumption fails in domains where no source is ever individually measured. + +Recent work has shown that score-based diffusion models [30] can serve as expressive Bayesian priors. Notably, once a diffusion model prior is trained, it enables effective posterior sampling for Bayesian inverse problems [31-39]. In the setting of noisy, incomplete observations, embedding diffusion models within an expectation-maximization framework can be used to learn an empirical prior [40]. In this work, we extend the use of diffusion model priors to MVSS. By leveraging the ability to sample joint diffusion posteriors over independent sources, our method directly learns a prior for each source. The main contributions of our method are: + +Generalist method for multi-view source separation: Our method is designed for any MVSS problem that is identifiable and linear. We show experimentally that our method works even when the data is incomplete, noisy, and varies in dimensionality. Additionally, our method does not require contrastive examples and succeeds even if every source is present in every observation. + +Source priors and posteriors: Our method results in independent diffusion models for each source. This affords all of the sampling and probability density evaluation benefits of diffusion models. + +State-of-the-art (SOTA) performance: Our method outperforms existing methods on the contrastive MVSS problem despite having a more generalist framework. + +# 2 Problem Statement + +Consider a noisy observation $\mathbf{y}^{\alpha}$ of view $\alpha \in \{1,\dots ,N_{\mathrm{views}}\}$ which is composed of a linear mixture of distinct sources $\mathbf{x}^{\beta}$ with $\beta \in \{1,\ldots ,N_s\}$ . The exact mixture of each source is given by a matrix $A_{i_\alpha}^{\alpha \beta}$ that depends on the view, $\alpha$ , the source, $\beta$ , and the specific sample, $i_{\alpha}$ . This model can be formalized as: + +$$ +\mathbf {y} _ {i _ {\alpha}} ^ {\alpha} = \left(\sum_ {\beta = 1} ^ {N _ {s}} \mathbf {A} _ {i _ {\alpha}} ^ {\alpha \beta} \mathbf {x} _ {i _ {\alpha}} ^ {\beta}\right) + \eta_ {i _ {\alpha}} ^ {\alpha}, \tag {1} +$$ + +where $\eta_{i_\alpha}^\alpha \sim \mathcal{N}(0, \Sigma_{i_\alpha}^\alpha)$ . The $\alpha$ subscript on the sample index $i$ highlights that sample indices between views are unrelated. Importantly, source draws are not shared between views. + +Goal: Given samples of noisy observations in each view, we aim to infer the individual prior distributions $p(\mathbf{x}^{\beta})$ of each source $\{\mathbf{x}^{\beta}\}_{\beta = 1}^{N_s}$ . Access to these source priors then allows us to perform source separation by sampling from the joint posterior $p(\{\mathbf{x}^{\beta}\} | \mathbf{y}_{i_{\alpha}}^{\alpha}, \mathbf{A}_{i_{\alpha}}^{\alpha \beta})$ . + +Dimensionality: Unlike traditional source separation, the dimensionality of the observation is determined by the specific view: $\mathbf{y}^{\alpha},\eta^{\alpha}\in \mathbb{R}^{d_{\alpha}}$ and $\pmb {\Sigma}^{\alpha}\in \mathbb{R}^{d_{\alpha}\times d_{\alpha}}$ . Similarly, the source dimensionality can vary between sources, $\mathbf{x}^{\beta}\in \mathbb{R}^{d_{\beta}}$ , leading to a mixing matrix whose dimensionality is determined by the view and source, $\mathbf{A}^{\alpha \beta}\in \mathbb{R}^{d_{\alpha}\times d_{\beta}}$ . No assumption is placed on the relative magnitude of the $d_{\alpha}$ and $d_{\beta}$ values, although for many applications $d_{\beta}\geq d_{\alpha}\forall \alpha ,\beta$ + +Source Independence: We assume that each source is conditionally independent of the other sources, allowing us to factorize the prior distribution: $p(\{\mathbf{x}^{\beta}\}_{\beta = 1}^{N_s}) = \prod_{\beta = 1}^{N_s}p(\mathbf{x}^\beta)$ . Similarly, we assume independence between the sources and mixing matrices: $p(\mathbf{x}^{\beta}|\mathbf{A}_i^{\alpha \beta '}) = p(\mathbf{x}^{\beta})\forall \beta ,\beta ',\alpha .$ + +Mixing Matrix and Incomplete Data: In contrast to blind source separation, we will assume that the mixing matrices, $\mathbf{A}_{i_{\alpha}}^{\alpha \beta}$ , are known. However, while the dimensionality of the mixing matrices are fixed for each view and source, the specific matrix can differ between samples $i_{\alpha}$ . For the purposes of this work, we will consider any non-invertible linear transformation to generate incomplete data. + +Identifiability: Not all choices of mixing matrices and dimensionalities will lead to a unique solution for the prior distributions. When a unique solution does not exist, any MVSS method will converge to a set of source distributions that accurately describe the data but may not match the true distributions. Therefore, we will assume identifiability throughout this work. + +# 3 Related Works + +Multi-view Source Separation: Research on MVSS problems has focused on the contrastive setting. In this setting, there are two views, the background view that contains only the background source and the target view that contains both the background source and the target source: + +$$ +\mathbf {y} _ {i} ^ {\mathrm {b k g}} = \mathbf {x} _ {i} ^ {\mathrm {b k g}} + \eta_ {i} ^ {\mathrm {b k g}}; \quad \mathbf {y} _ {j} ^ {\mathrm {t a r g}} = \mathbf {x} _ {j} ^ {\mathrm {b k g}} + \mathbf {x} _ {j} ^ {\mathrm {t a r g}} + \eta_ {j} ^ {\mathrm {t a r g}}. \tag {2} +$$ + +Contrastive latent variable models (CLVM; [28]) define two sets of latent distributions in a lower-dimensional space, one for the background source and one for the target source. The latent variables are mapped to the observed space either with a linear model (CLVM - Linear) or through a non-linear transformation parameterized by a neural network (CLVM - VAE). The parameters controlling the transformation from the low-dimensional latent space to the observation space are optimized through an expectation-maximization or variational inference approach. The CLVM method can generate posterior and prior samples for both sources, and it can be adapted to incomplete data. + +Contrastive principal component analysis (CPCA; [27]), and its probabilistic extension (PCPCA; [29]), attempt to find vectors that maximize the variance in the target view without explaining the variance in the background view. They do so by introducing a pseudo-data covariance matrix $C = C_{\mathrm{targ}} - \gamma C_{\mathrm{bkg}}$ , with $\gamma$ a tunable hyperparameter. PCPCA can only sample from the target source posterior and prior, but it can be adapted to incomplete data. + +Contrastive variational autoencoder models (CVAE; [26]) are an alternative formulation of the CLVM-VAE method2. In contrast to CLVM-VAE, CVAE uses two encoders shared across views: one produces background latents and the other produces target latents. The concatenated latents are fed to a shared decoder, with the target latents multiplied by zero for the background decoding task. In the original formulation, the CVAE model never outputs the target source during training, only the target observation. This allows for non-linear mixing of target and source, but makes extending the model to incomplete data challenging when $\mathbf{x}_j^{\mathrm{bkg}}$ and $\mathbf{x}_j^{\mathrm{targ}}$ do not share a mixing matrix. + +Diffusion Models: Diffusion models [30, 41-44] seek to reverse a known corruption process in order to be able to generate samples from a target distribution. In the continuous-time framing, samples from the target distribution, $x_0 \sim p(x_0)$ , are corrupted through a diffusion process governed by the stochastic equation: + +$$ +\mathrm {d} \mathbf {x} _ {t} = f (\mathbf {x} _ {t}, t) \mathrm {d} t + g (t) \mathrm {d} \mathbf {w} _ {t}, \tag {3} +$$ + +where $f(\mathbf{x}_t, t)$ is known as the drift coefficient, $g(t)$ is known as the diffusion coefficient, and $\mathbf{w}_t$ is generated through a standard Wiener process. The time coefficient $t$ ranges from 0 to 1, with $\mathbf{x}_0$ being the original samples and $\mathbf{x}_1$ being the fully diffused samples. This induces a conditional distribution of the form $p(\mathbf{x}_t | \mathbf{x}_0) = \mathcal{N}(\mathbf{x}_t | \alpha_t \mathbf{x}_0, \boldsymbol{\Sigma}_t)$ , where $\alpha_t$ and $\boldsymbol{\Sigma}_t$ can be derived from our drift and diffusion coefficients [45]. The forward stochastic differential equation (SDE) has a corresponding reverse SDE [46]: + +$$ +\mathrm {d} \mathbf {x} _ {t} = \left[ f \left(\mathbf {x} _ {t}, t\right) - g (t) ^ {2} \nabla_ {\mathbf {x} _ {t}} \log p (\mathbf {x} _ {t}) \right] \mathrm {d} t + g (t) \mathrm {d} \bar {\mathbf {w}} _ {t}, \tag {4} +$$ + +where $\bar{\mathbf{w}}$ is the standard Wiener process with time reversed. This reverse SDE evolves a sample from the fully diffused distribution $p(\mathbf{x}_1)$ back to the original data distribution $p(\mathbf{x}_0)$ . Equation 4 requires access to the score function, $\nabla_{\mathbf{x}_t}\log p(\mathbf{x}_t)$ , which is approximated by a neural network + +trained via score matching on samples from the forward diffusion process [30, 47, 48]. Training and sampling from a diffusion model requires selecting an SDE parameterization [30, 42, 44, 49], a score matching objective [30, 42, 43, 45], and a sampling method for the reverse SDE [30, 42, 44, 45]. + +We adopt the variance exploding parameterization for the SDE [43], the denoiser parameterization from Karras et al. [45] for the score matching approach, and the predictor-corrector (PC) algorithm as our sampling method [30]. The denoiser parameterization approximates $\mathbb{E}[\mathbf{x}_0|\mathbf{x}_t]$ by minimizing the objective: + +$$ +\mathcal {L} (\theta) = \mathbb {E} _ {p \left(\mathbf {x} _ {t} \mid \mathbf {x} _ {0}\right)} [ \lambda (t) \| d _ {\theta} \left(\mathbf {x} _ {t}, t\right) - \mathbf {x} _ {0} \| _ {2} ^ {2} ]. \tag {5} +$$ + +Here, $d_{\theta}(\mathbf{x}_t,t)$ is our denoiser model with parameters $\theta$ . We use a loss weighting term, $\lambda (t)$ , to ensure all time steps are equally prioritized. Note the denoiser returns the expectation value, which can be directly converted to the score via Tweedie's formula [50]. + +Posterior Sampling: Once trained, a diffusion model can be used as a Bayesian prior for conditional posterior sampling [30]. Specifically, the score function in the reverse SDE is replaced by the posterior score function: + +$$ +\nabla_ {\mathbf {x} _ {t}} \log p (\mathbf {x} _ {t} | \mathbf {y}) = \nabla_ {\mathbf {x} _ {t}} \log p (\mathbf {x} _ {t}) + \nabla_ {\mathbf {x} _ {t}} \log p (\mathbf {y} | \mathbf {x} _ {t}), \tag {6} +$$ + +where $\mathbf{y}$ is our observation, and $\nabla_{\mathbf{x}_t}\log p(\mathbf{y}|\mathbf{x}_t)$ is the score of the likelihood. The prior score in Equation 6 is given by the trained diffusion model, but evaluating the likelihood score with respect to an arbitrary $t$ requires solving: + +$$ +p (\mathbf {y} | \mathbf {x} _ {t}) = \int p (\mathbf {y} | \mathbf {x} _ {0}) p (\mathbf {x} _ {0} | \mathbf {x} _ {t}) d \mathbf {x} _ {0}. \tag {7} +$$ + +Many methods have been proposed for evaluating the conditional score [51]. Of particular interest are methods that propose an approximation to the right-most conditional distribution in Equation 7 [34, 36, 38-40]. In general, these methods use a multivariate Gaussian approximation: + +$$ +p \left(\mathbf {x} _ {0} \mid \mathbf {x} _ {t}\right) \approx \mathcal {N} \left(\mathbf {x} _ {0} \mid \mathbb {E} \left[ \mathbf {x} _ {0} \mid \mathbf {x} _ {t} \right], \mathbb {V} \left[ \mathbf {x} _ {0} \mid \mathbf {x} _ {t} \right]\right). \tag {8} +$$ + +When the observation function is defined by the linear matrix $\mathbf{A}$ and the likelihood is Gaussian, $p(\mathbf{y}|\mathbf{x}_0) = \mathcal{N}(\mathbf{y}|\mathbf{A}\mathbf{x}_0,\mathbf{\Sigma}_y)$ , this approximation yields an analytic solution for the likelihood score: + +$$ +\nabla_ {\mathbf {x} _ {t}} \log p (\mathbf {y} | \mathbf {x} _ {t}) \approx \nabla_ {\mathbf {x} _ {t}} \mathbb {E} [ \mathbf {x} _ {0} | \mathbf {x} _ {t} ] ^ {\top} \mathbf {A} ^ {\top} \left(\boldsymbol {\Sigma} _ {y} + \mathbf {A} \mathbb {V} [ \mathbf {x} _ {0} | \mathbf {x} _ {t} ] \mathbf {A} ^ {\top}\right) ^ {- 1} (\mathbf {y} - \mathbf {A} \mathbb {E} [ \mathbf {x} _ {0} | \mathbf {x} _ {t} ]). \tag {9} +$$ + +The moment matching posterior sampling (MMPS; [40]) approximation leads to the best sampling when compared on linear inverse problems. The trick behind MMPS is to use Tweedie's variance formula: + +$$ +\mathbb {V} \left[ \mathbf {x} _ {0} \mid \mathbf {x} _ {t} \right] = \boldsymbol {\Sigma} _ {t} \nabla_ {\mathbf {x} _ {t}} ^ {\top} \mathbb {E} \left[ \mathbf {x} _ {0} \mid \mathbf {x} _ {t} \right]. \tag {10} +$$ + +While the Jacobian, $\nabla_{\mathbf{x}_t}^\top \mathbb{E}[\mathbf{x}_0|\mathbf{x}_t]$ , in Equation 10 would be extremely costly to materialize, the MMPS method avoids instantiating the matrix by the use of the vector-Jacobian product combined with a conjugate gradient solver [52] for the inverse in Equation 9. + +Expectation Maximization: Expectation-maximization (EM) is a framework for finding the maximum likelihood estimate for model parameters, $\theta$ , in the presence of hidden variables [53]. EM builds a sequence of model parameters $\theta_0, \theta_1, \ldots, \theta_K$ that monotonically improve the likelihood of the data3. We adopt the Monte Carlo EM (MCEM) framework from [40]. We embed a diffusion model prior into an MCEM framework where the hidden variables are the true signals, $\mathbf{x}$ , and the observations are the noisy, linear transformations of the signal, $\mathbf{y} = \mathbf{A}\mathbf{x} + \boldsymbol{\eta}$ . The following two steps of the framework are then repeated until convergence: + +- Expectation (E): Given the current diffusion model parameters $\theta_{k}$ , sample from diffusion posterior for each observation: $\mathbf{x}_i\sim q_{\theta_k}(\mathbf{x}_i|\mathbf{y}_i,\mathbf{A}_i)$ . Here $q_{\theta_k}(\mathbf{x}_i|\mathbf{y}_i,\mathbf{A}_i)$ is the distribution given by sampling from the reverse SDE in Equation 4 while using the posterior score from Equation 6 and the denoiser model $d_{\theta_k}(\mathbf{x}_t,t)$ . +- Maximization (M): Given the set of posterior samples for the full dataset, $\{\mathbf{x}_i\}$ , maximize the data likelihood with respect to the model parameters $\theta$ to get $\theta_{k+1}$ . In practice, since the mixing matrix and noise covariance are fixed, the denoising score matching objective (Equation 5) can be used as a surrogate. + +# 4 Methods + +Our goal is to learn the prior distribution $p(\mathbf{x}^{\beta})$ of each source $\{\mathbf{x}^{\beta}\}_{\beta = 1}^{N_s}$ given the observations, $\{\mathbf{y}_{i_\alpha}^\alpha\}$ , known mixing matrices, $\mathbf{A}_{i_\alpha}^{\alpha \beta}$ , and known noise covariance $\Sigma_{i_\alpha}^\alpha$ (see Section 2). The prior distribution for each source $\beta$ will be parameterized by a variational distribution $q_{\theta \beta}(\mathbf{x}^{\beta})$ , defined by a denoiser diffusion model, $d_{\theta \beta}(\mathbf{x}_t^\beta, t)$ with parameters $\theta^\beta$ . We use an EM framework to iteratively maximize the likelihood of the set of diffusion model parameters $\Theta_k = \{\theta_k^\beta\}_{\beta = 1}^{N_s}$ , where $k$ indexes the EM round. We summarize the full method, DDPRISM, in Algorithm 1 provided in Appendix A. All of the code to reproduce our method and experiments has been made public4. + +Maximization Step: For the M step we want to maximize the expected log-likelihood of the full set of diffusion model parameters with respect to the data with $\mathbf{u}_t^\top = \left[(\mathbf{x}_t^1)^\top, (\mathbf{x}_t^2)^\top, \ldots, (\mathbf{x}_t^{N_s})^\top\right]$ : + +$$ +\begin{array}{l} \Theta_ {k + 1} = \underset {\Theta} {\arg \max } \mathbb {E} _ {p \left(\mathbf {y} ^ {\alpha}, \left\{\mathbf {A} ^ {\alpha \beta} \right\}, \mathbf {u} _ {0}\right)} \left[ \log q _ {\Theta} \left(\mathbf {u} _ {0}, \mathbf {y} ^ {\alpha}, \left\{\mathbf {A} ^ {\alpha \beta} \right\}\right) \right] (11) \\ = \underset {\Theta} {\arg \max } \mathbb {E} _ {p (\mathbf {y} ^ {\alpha}, \{\mathbf {A} ^ {\alpha \beta} \})} \mathbb {E} _ {q _ {\Theta_ {k}} \left(\mathbf {u} _ {0} \mid \mathbf {y} ^ {\alpha}, \{\mathbf {A} ^ {\alpha \beta} \}\right)} \left[ \log q _ {\Theta} \left(\mathbf {u} _ {0}, \mathbf {y} ^ {\alpha}, \{\mathbf {A} ^ {\alpha \beta} \}\right) \right] (12) \\ = \underset {\Theta} {\arg \max } \mathbb {E} _ {p (\mathbf {y} ^ {\alpha}, \{\mathbf {A} ^ {\alpha \beta} \})} \mathbb {E} _ {q _ {\Theta_ {k}} \left(\mathbf {u} _ {0} \mid \mathbf {y} ^ {\alpha}, \{\mathbf {A} ^ {\alpha \beta} \}\right)} \left[ \sum_ {\beta} \log q _ {\theta^ {\beta}} \left(\mathbf {x} _ {0} ^ {\beta}\right) \right], (13) \\ \end{array} +$$ + +where $q_{\Theta}(\mathbf{u}_0,\mathbf{y}^\alpha ,\{\mathbf{A}^{\alpha \beta}\})$ is the full joint distribution of the observation, mixing matrices, and sources under the diffusion model parameters. In getting from Equation 12 to Equation 13 we have dropped all the terms independent of $\Theta$ and taken advantage of the independence of the sources. Since each distribution $q_{\theta^{\beta}}(\mathbf{x}^{\beta})$ is independent of the others, Equation 13 reduces to optimizing each diffusion model separately on the samples from its corresponding source. We can take an expectation value over $p(\mathbf{y}^{\alpha},\{\mathbf{A}^{\alpha \beta}\})$ by drawing examples from the dataset, so all that remains is defining how to sample from the joint posterior distribution $q_{\Theta_k}(\{\mathbf{x}^\beta \vert \mathbf{y}^\alpha ,\{\mathbf{A}^{\alpha \beta}\})$ given the current set of diffusion model parameters $\Theta_{k}$ + +Expectation Step: We want to sample from the joint distribution $q_{\Theta_k}(\{\mathbf{x}^\beta \} | \mathbf{y}^\alpha, \mathbf{A}^{\alpha \beta})$ . To do so we need the joint posterior score: + +$$ +\nabla_ {\mathbf {u} _ {t}} \log q _ {\Theta_ {k}} \left(\mathbf {u} _ {t} \mid \mathbf {y} ^ {\alpha}, \left\{\mathbf {A} ^ {\alpha \beta} \right\}\right) = \nabla_ {\mathbf {u} _ {t}} \log q _ {\Theta_ {k}} \left(\mathbf {u} _ {t}\right) + \nabla_ {\mathbf {u} _ {t}} \log q _ {\Theta_ {k}} \left(\mathbf {y} ^ {\alpha} \mid \mathbf {u} _ {t}, \left\{\mathbf {A} ^ {\alpha \beta} \right\}\right). \tag {14} +$$ + +Because our sources are independent, the first term on the right-hand side of Equation 14 simplifies to: + +$$ +\nabla_ {\mathbf {u} _ {t}} \log q _ {\Theta_ {k}} (\mathbf {u} _ {t}) = \sum_ {\beta} \nabla_ {\mathbf {u} _ {t}} \log q _ {\theta_ {k} ^ {\beta}} \left(\mathbf {x} _ {t} ^ {\beta}\right), \tag {15} +$$ + +which is simply the sum of the individual diffusion model scores. The remaining likelihood term in Equation 14 is given by: + +$$ +q _ {\Theta_ {k}} \left(\mathbf {y} ^ {\alpha} \mid \mathbf {u} _ {t}, \left\{\mathbf {A} ^ {\alpha \beta} \right\}\right) = \int \dots \int p \left(\mathbf {y} ^ {\alpha} \mid \left\{x _ {0} ^ {\beta} \right\}, \left\{\mathbf {A} ^ {\alpha \beta} \right\}\right) \prod_ {\beta} q _ {\theta_ {k}} \left(x _ {0} ^ {\beta} \mid x _ {t} ^ {\beta}\right) d x _ {0} ^ {\beta}. \tag {16} +$$ + +To solve this we employ the MMPS approximation [40], wherein each conditional distribution $q_{\theta_k}(x_0^\beta | x_t^\beta)$ is approximated by its first and second moments. Since the conditional distributions are now Gaussian, Equation 16 is simply $N_s$ analytic Gaussian convolutions. The final likelihood score approximation is then: + +$$ +\begin{array}{l} \nabla_ {\mathbf {u} _ {t}} \log q _ {\Theta_ {k}} (\mathbf {y} ^ {\alpha} | \mathbf {u} _ {t}, \{\mathbf {A} ^ {\alpha \beta} \}) = \left[ \begin{array}{c} \nabla_ {\mathbf {x} _ {t} ^ {1}} \mathbb {E} [ \mathbf {x} _ {0} ^ {1} | \mathbf {x} _ {t} ^ {1} ] ^ {\top} (\mathbf {A} ^ {\alpha 1}) ^ {\top} \\ \vdots \\ \nabla_ {\mathbf {x} _ {t} ^ {N _ {s}}} \mathbb {E} [ \mathbf {x} _ {0} ^ {N _ {s}} | \mathbf {x} _ {t} ^ {N _ {s}} ] ^ {\top} (\mathbf {A} ^ {\alpha N _ {s}}) ^ {\top} \end{array} \right] \\ \times \left(\boldsymbol {\Sigma} _ {i _ {\alpha}} ^ {\alpha} + \sum_ {\beta} \mathbf {A} ^ {\alpha \beta} \mathbb {V} \left[ \mathbf {x} _ {0} ^ {\beta} \mid \mathbf {x} _ {t} ^ {\beta} \right] \left(\mathbf {A} ^ {\alpha \beta}\right) ^ {\top}\right) ^ {- 1} \left(\mathbf {y} ^ {\alpha} - \sum_ {\beta} \mathbf {A} ^ {\alpha \beta} \mathbb {E} \left[ \mathbf {x} _ {0} ^ {\beta} \mid \mathbf {x} _ {t} ^ {\beta} \right]\right). \tag {17} \\ \end{array} +$$ + +
MethodPosteriorPrior
PQM ↑FID ↓PSNR ↑SD ↓PQM ↑FID ↓SD ↓
1D Manifold: Cont. 2 Sources
PCPCA [29]0.0-9.357.690.0-7.91
CLVM - Linear [28]0.0-9.585.800.0-5.86
CLVM - VAE [28]0.0-17.151.810.0-2.91
DDPRISM-Gibbs [54]0.0-12.663.960.0-3.92
DDPRISM-Joint [Ours]0.26-38.270.350.01-0.37
1D Manifold: Cont. 3 Sources
PCPCA [29]0.0-6.8912.570.0-10.22
CLVM - Linear [28]0.0-11.642.030.0-2.16
CLVM - VAE [28]0.0-13.092.220.0-1.82
DDPRISM-Gibbs [54]0.0-9.504.500.0-4.53
DDPRISM-Joint [Ours]0.0-19.780.750.0-0.78
1D Manifold: Mix. (fmix = 0.1)
DDPRISM-Gibbs [54]0.0-17.691.840.0-1.81
DDPRISM-Joint [Ours]0.001-24.150.050.0-0.04
GMNIST: Cont. Full-Resolution
PCPCA [29]0.022.318.99-0.0176.0-
CLVM - Linear [28]0.0101.313.30-0.0139.9-
CLVM - VAE [28]0.018.8714.56-0.057.67-
DDPRISM-Joint [Ours]1.001.5725.60-0.2020.10-
GMNIST: Cont. Downsampled
PCPCA [29]0.0121.714.08-0.0115.4-
CLVM - Linear [28]0.0199.512.16-0.0211.4-
CLVM - VAE [28]0.01008.08.48-0.0737.0-
DDPRISM-Joint [Ours]0.942.3619.73-0.0612.63-
+ +Table 1: Comparison of metrics between our methods and baselines for all the experiments shown in the paper. For each metric, the arrow indicates whether larger (↑) or smaller (↓) values are optimal. Our method sets or matches the state-of-the-art for all combinations of experiments and baselines. The posterior metrics are calculated using posterior samples and the true source signals. Since these samples are not independent, it is possible to get large positive PQMass p-values. + +Note that the computational cost of Equation 17 scales linearly with the number of sources. As with regular MMPS, the Jacobian can be avoided through the use of the vector-Jacobian product, and the gradient of the variance is ignored. We note that a similar joint posterior equation was concurrently derived by Stevens et al. [25] for removing structured noise using diffusion models. + +Gibbs Sampling: As an alternative to directly sampling the joint posterior, it is also possible to use a Gibbs sampling method. We derive an extension of the Gibbs diffusion algorithm presented in Heurtel-Depeiges et al. [54] for MVSS in Appendix B. We note that converging to the posterior requires a large number of Gibbs sampling rounds for source distributions with complex structure, thereby rendering the Gibbs sampling approach computationally infeasible for most problems. + +Contrastive MVSS Simplification: For the contrastive MVSS problem, each new view introduces one new source. In theory, this problem is solved by the generic EM method we have presented. In practice, it can be useful to train the diffusion models sequentially, optimizing $\theta^1$ on observations from view $\alpha = 1$ , optimizing $\theta^2$ on observations from view $\alpha = 2$ with $\theta^1$ held fixed, and so on. This limits the computational cost by reducing the number of source models in the joint sampling for all but the final view. However, it discards the information about source $\beta$ present in views $\alpha > \beta$ . We summarize this simplified method in Appendix A. + +![](images/7a412f2e3029197829f6aaccd9c42f7b62fb966355effdaaabf844b3756bcee3.jpg) +Figure 2: Comparison of the mean Sinkhorn divergence for different $f_{\mathrm{mix}}$ on the mixed 1D manifold problem. Also shown are the Gibbs sampling method with eight times as many computations per EM lap and our method when $f_{\mathrm{mix}} = 1.0$ and $A_{i_{\alpha}}^{\alpha \beta}$ depends on $\beta$ . Even for large mixing fractions, our method can accurately learn the two distinct underlying source distributions. + +![](images/b23d16fa98bb89554ee506be4ce807522c984893b838f38db8e839dbb6944fc9.jpg) +Figure 1: Comparison of posterior samples for our joint sampling method and the Gibbs sampling method [54] on the 1D manifold problem. Both methods are equivalent for the first source. The plots show the evolution of the marginals for the first and second dimension of the specified source distribution. The last EM lap for sources $\beta = 1,2,3$ are 16, 32, 64 respectively. + +# 5 Results + +We present five experimental setups that demonstrate the effectiveness of our method. The first four experiments are variations of two synthetic problems previously explored in the literature [26-28, 40]. The final experiment is on the real-world scientific task of separating galaxy light from random light, demonstrating the viability of our method on complex scientific data. Timing comparisons for all experiments can be found in Appendix G. + +# 5.1 One-Dimensional Manifold Experiments + +In this pair of experiments, our sources, $\mathbf{x}^{\beta}$ , are drawn from distinct, one-dimensional manifolds embedded in $\mathbb{R}^5$ . Our observations, $\mathbf{y}^{\alpha}$ , lie in $\mathbb{R}^3$ and are generated by a random linear projection, $\mathbf{A}^{\alpha \beta} \in \mathbb{R}^{3 \times 5}$ , whose rows are drawn from the unit sphere, $\mathbb{S}^4$ . In addition, we add isotropic Gaussian noise with standard deviation $\sigma_y = 0.01$ for all views. This setup follows previous work [40], although we now add multiple sources to our observations. + +Contrastive MVSS: For the contrastive experiment, each view introduces a new source, and the mixing matrix is shared among all the sources: $\mathbf{A}_{i_{\alpha}}^{\alpha \beta} = \mathbf{A}_{i_{\alpha}}c^{\alpha \beta}$ with $c^{\alpha \beta} = 1$ if $\beta \leq \alpha$ and $c^{\alpha \beta} = 0$ otherwise. We set $N_{\mathrm{view}} = N_s = 3$ and generate a dataset of size $2^{16}$ for each view. + +For our joint diffusion sampling approach, we train three denoiser models, $d_{\theta^{\beta}}(\mathbf{x}_t^\beta ,t)$ , each consisting of a multi-layer perceptron. We employ the simplified contrastive MVSS algorithm described in Section 4. We train the $\beta = 1,2,3$ diffusion models for 16,32,64 EM laps respectively. For the sampling we use the PC algorithm with 16,384 predictor steps, each with one corrector step (PC step). The initial posterior samples are drawn using a Gaussian prior whose parameters are optimized through a short EM loop. We also train the Gibbs sampling approach with the same denoiser architecture and the same number of EM laps. To keep the computational costs (compute5) on par with our joint diffusion approach, we do 64 Gibbs rounds per expectation step and reduce the number of PC steps to 256. Because the Gibbs approach performs poorly, we only run it up to the second view. We also compare to PCPCA, CLVM-Linear, and CLVM-VAE. We provide additional + +experimental details on the diffusion parameters, generation of the random manifolds, and baselines in Appendix C. + +As shown in Figure 1, our method learns the first two source distributions nearly perfectly despite the linear projection to a lower dimension and the presence of noise. The third source distribution is also learned, although the final posterior samples are not as sharp. This is to be expected: later sources are only observed together with all the previous sources, making them harder to sample. We compare the Sinkhorn divergence [55], PQMass p-value [56], and peak signal-to-noise ratio (PSNR) for our method and the baselines in Table 1. Our method compares favorably, outperforming all the baselines on both source distributions across all the metrics. + +Mixed MVSS: For the mixed experiment, each source is present in every view. The mixing matrix is given by: $\mathbf{A}_{i_{\alpha}}^{\alpha \beta} = \mathbf{A}_{i_{\alpha}}c^{\alpha \beta}$ with $c^{\alpha \beta} = 1$ if $\beta = \alpha$ and $c^{\alpha \beta} = f_{\mathrm{mix}}$ otherwise. We set $N_{\mathrm{views}} = N_s = 2$ and generate a dataset of size $2^{16}$ for each view. We consider four different mixing fractions $f_{\mathrm{mix}}\in \{0.0,0.1,0.5,0.9\} ^6$ . When $f_{\mathrm{mix}} = 1.0$ the problem is fully degenerate and therefore not identifiable (see Appendix C). For comparison, we also present a mixed experiment with $\mathbf{A}_{i_{\alpha}}^{\alpha \beta}$ drawn separately, meaning that every source is fully present in every view but with a different mixing. + +For the joint sampling and Gibbs sampling approach, we use the same denoiser models and initialization procedure as the Contrastive MVSS problem. However, because the Gibbs sampling was not able to learn either source distribution with equivalent compute, we instead used 64 Gibbs rounds and 2048 PC steps. This means that each EM lap for the Gibbs sampling is eight times as expensive. We provide additional experimental details in Appendix C. + +In Figure 2 we compare the Sinkhorn divergence averaged over both source distributions as a function of EM laps. We find that our method can learn the underlying source distributions with high accuracy up to $f_{\mathrm{mix}} = 0.5$ . For $f_{\mathrm{mix}} = 0.9$ our method continues to improve its estimate of the source distributions as the EM laps progress, but it does not converge. If the mixing matrix varies between the sources, we can reconstruct the source distributions even with $f_{\mathrm{mix}} = 1.0$ . By comparison, Gibbs sampling for $f_{\mathrm{mix}} = 0.1$ converges much more slowly despite requiring eight times as much compute per EM lap. + +# 5.2 Grassy MNIST Experiments + +For this pair of experiments, we use the Grassy MNIST dataset first presented in [27]. The dataset is a contrastive MVSS problem which consists of two views: the first containing random $28 \times 28$ crops of grass images from ImageNet [57], and the second containing a linear combination of grass images with 0 and 1 MNIST digits [58]. In addition, we add a small amount of Gaussian noise to each observation $(\sigma_y = 0.01)$ . For both experiments, we generate 32,768 observations for the grass view and 13,824 for the linear combination of digits and grass. + +Full-Resolution: In the full-resolution experiment, we set $\mathbf{A}_{i_1}^{11} = \mathbf{A}_{i_2}^{21} = \mathbb{I}$ for $\beta = 1$ (grass) and $\mathbf{A}_{i_1}^{12} = 0$ , $\mathbf{A}_{i_2}^{22} = 0.5 \times \mathbb{I}$ for $\beta = 2$ (MNIST). Two example observations can be seen in Figure 3. For our denoiser models, $d_{\theta^{\beta}}(\mathbf{x}_t^\beta, t)$ , we use a U-Net architecture [42, 59] with attention blocks [60] and AdaLN-Zero norm modulation [61]. We employ the simplified contrastive MVSS algorithm described in Section 4, and we initialize our posterior samples using a Gaussian prior whose parameters are optimized through 32 EM laps. We train the grass and MNIST diffusion models for 64 EM laps. For the sampling we use the PC algorithm with 256 PC steps. We compare to PCPCA, CLVM-Linear, and CLVM-VAE but omit Gibbs sampling since its computational cost makes it impractical for this problem. We provide additional experimental details in Appendix D. + +In Table 1 we compare the FID scores [62], the PSNR, and the PQMass p-value on the posterior MNIST digit samples across the entire dataset of the MNIST + grass view. For the FID score, we use a trained MNIST classifier in place of the Inception-v3 network. We also report the FID score and PQMass p-value on samples from the learned priors. In addition, in left-hand side of Figure 3 we show example posterior draws for our method, PCPCA, and CLVM-VAE. + +Our method visually returns the closest posterior samples to the ground truth, and outperforms the baselines across all the metrics. The prior samples also outperform the baselines on both PQMass p-value and FID. To better understand the source of this improvement, we run ablation studies over the + +![](images/a348b6146379bacaef43f0b19c3ef7a58d4db8fbb4a7017f3e47d56bb20c4fa1.jpg) +Figure 3: Comparison of posterior samples for two example observations in Grassy MNIST experiment. The observations are on the far left and right, the true input sources are in the middle, and a draw from DDPRISM [ours], CLVM-VAE [28], and PCPCA [29] for both the full-resolution and downsampled case is shown in between. PCPCA cannot sample the grass posterior, and CLVM-Linear is omitted for brevity. Our joint diffusion model returns the best reconstruction of both sources, with near-perfect posterior samples in the full-resolution case. + +model architecture, the number of EM laps, the number of initialization laps for the Gaussian prior, the dataset size, and the number of sampling steps. The full ablation study details and results can be found in Appendix H. Overall, the method is fairly insensitive to changes in these hyperparameters. Only extreme choices, such as replacing the U-Net with a small MLP (FID=49.03), conducting only 2 laps of EM (FID=96.85), removing the Gaussian prior initialization entirely (FID=10.41), or using only 16 PC sampling steps (FID=5.41) appear to meaningfully reduce performance across our metrics. The one exception is reductions in the dataset size, where using 1/4th, 1/16th, and 1/64th of the dataset leads to an FID of 4.64, 10.19, and 38.34 respectively. + +Downsampled: In the downsampling experiment, one third of our observations are at full-resolution, one third are 2x downsampled, and one third are 4x downsampled. Otherwise the dataset size and underlying source distributions are kept the same. An example observation can be seen in Figure 3. We compare to PCPCA, CLVM-Linear, and CLVM-VAE, and use the same configurations as the full-resolution experiment for all four methods. We provide additional details in Appendix D. + +In Figure 3 we show example posteriors for $2\mathrm{x}$ downsampling, and in Table 1 we report the FID score, the PSNR, and the PQMass p-value for the posterior samples across the dataset. We also report the FID score and PQMass p-value of draws from the prior distribution. As with the full-resolution dataset, our method visually returns the closest posterior samples to the ground truth for both sources and is SOTA across the baselines. Notably, while both of the CLVM methods and PCPCA struggle with the downsampled images, our method returns visually plausible digits and comparable metric performance to the full-resolution experiment. + +# 5.3 Galaxy Images + +In astronomical images, instrumental noise, cosmic rays, and random foreground and background objects along the line of sight contaminate observations ("random" light). Separating the flux of these contaminants from the target object is a contrastive MVSS problem. There are two source populations: random light and galaxies. We also get two views: random sightlines that are uncorrelated with galaxies, and targeted sightlines built from a catalog. The targeted sightlines contain both the galaxy and random light. To build our views we use archival Hubble Space Telescope observations of the COSMOS survey [63]. For the galaxy view, we select targets using the Galaxy Zoo Hubble object catalog [64], and for the random view we make cutouts at random locations in + +![](images/62215a95a7e7b4d24e9db501668f9b08bd7b8d0bb11022d6646abba29da4ebc4.jpg) +Figure 4: Observations of galaxy images along with posterior samples for the random and galaxy source using our method. Images are split into high-contrast (upper region) and low-contrast (lower region) colormaps to highlight the range of features. The galaxy source captures the central light while the small-scale fluctuations and the uncorrelated light is separated into the random source. + +the COSMOS field. Each cutout is $128 \times 128$ pixels. For our denoiser models, $d_{\theta \beta}(\mathbf{x}_t^\beta, t)$ , we use the same U-Net architecture as for Grassy MNIST, but change the model depth and size to model the larger images. We employ the simplified contrastive MVSS algorithm, and we train the random diffusion model for 32 EM laps and the galaxy diffusion model for 16 EM laps. For the sampling we use the PC algorithm with 256 PC steps. We provide additional experimental details in Appendix E. + +In Figure 4 we show sample galaxy observations along with a random and galaxy source posterior sample for each observation. While we do not have access to the ground truth, a visual inspection of the source posteriors samples show that our model is effectively identifying the central galaxy light and separating it from the random light. Notably, the background around the galaxy light appears to be nearly flat at zero. We make the full dataset of 79k pristine galaxy images publicly available7. Generating the full dataset required 34 hours on four NVIDIA H100 GPUs. + +# 6 Discussion and Limitations + +We present DDPRISM, a data-driven framework for tackling general MVSS problems using diffusion model priors. To our knowledge, it is the first method to provide a unified solution for linear MVSS problems, achieving state-of-the-art performance across diverse experiments. We further demonstrate that DDPRISM delivers high-quality source separation on a complex real-world astrophysical dataset. + +Despite these advances, the framework has important limitations. First, it is restricted to linear source combinations, which excludes nonlinear generative processes like occlusion. The Gaussian noise assumption can be relaxed through the inclusion of an additional "noise" source, but only if an extra view is available. We also assume exact knowledge of the mixing matrix, whereas scientific applications often involve probabilistic rather than deterministic mixing. These assumptions limit our generality and motivate extensions that relax linearity, Gaussianity, and deterministic mixing. + +Computationally, our method requires expensive sampling that is compounded by the EM-style training. This places limits on the resolution, dataset size, and number of sources that can be feasibly modeled. Our baselines are far cheaper, albeit at the cost of sample quality. Performance also degrades with smaller datasets, creating a tension between the benefits of large datasets and the computational demands of the method. Replacing our initialization method with random initializations degrades sample quality, suggesting that clever initialization methods may improve convergence and alleviate computational bottlenecks. Nevertheless, DDPRISM establishes diffusion-based MVSS as a promising tool for disentangling structured signals across scientific domains. + +# Acknowledgments and Disclosure of Funding + +We would like to thank Shirley Ho, David Spergel, and Francisco Villaescusa-Navarro for insightful discussions during the development of this project. The computational resources used in this work were provided by the Flatiron Institute, a division of the Simons Foundation. The work of SWC is supported by the Simons Foundation. AA acknowledges financial support from the Flatiron Institute's Predoctoral Program, and thanks the LSST-DA Data Science Fellowship Program, which is funded by LSST-DA, the Brinson Foundation, the WoodNext Foundation, and the Research Corporation for Science Advancement Foundation; her participation in the program has benefited this work. SE acknowledges funding from NSF GRFP-2021313357 and the Stanford Data Science Scholars Program. + +# References + +[1] P. Melchior et al. "SCARLET: Source separation in multi-band images by Constrained Matrix Factorization". In: *Astronomy and Computing* 24 (2018), p. 129. +[2] Simon Samuroff et al. "Dark Energy Survey Year 1 results: the impact of galaxy neighbours on weak lensing cosmology with IM3SHAPE". In: Monthly Notices of the Royal Astronomical Society 475.4 (2018), pp. 4524-4543. +[3] James Bosch et al. "The Hyper Suprime-Cam Software Pipeline". In: *Publications of the Astronomical Society of Japan* 70 (2018), S5. +[4] Michael S Lewicki. "A review of methods for spike sorting: the detection and classification of neural action potentials". In: Network: Computation in Neural Systems 9.4 (1998), R53. +[5] Gaute T. Einevoll et al. "Modelling and analysis of local field potentials for studying the function of cortical circuits". In: Nature Reviews Neuroscience 14.11 (2013), pp. 770-785. +[6] Marius Pachitariu et al. "Kilosort: realtime spike-sorting for extracellular electrophysiology with hundreds of channels". In: bioRxiv (2016). eprint: https://www.biorxiv.org/content/early/2016/06/30/061481.full.pdf. +[7] Yanping Liu et al. "An Amplitude-Preserved Time-Frequency Peak Filtering Based on Empirical Mode Decomposition for Seismic Random Noise Reduction". In: IEEE Geoscience and Remote Sensing Letters 11.5 (2014), pp. 896-900. +[8] Tong Shen et al. "A Strong Noise Reduction Network for Seismic Records". In: Applied Sciences 14.22 (2024). +[9] François Lanusse et al. "Deep generative models for galaxy image simulations". In: Monthly Notices of the Royal Astronomical Society 504.4 (2021), pp. 5543-5555. +[10] Euclid Collaboration et al. "Euclid preparation. XIII. Forecasts for galaxy morphology with the Euclid Survey using deep generative models". In: *Astronomy & Astrophysics* 657 (2022), A90. +[11] Pierre Comon. "Independent component analysis, a new concept?" In: Signal processing 36.3 (1994), pp. 287-314. +[12] A. Hyvarinen. "Fast and robust fixed-point algorithms for independent component analysis". In: IEEE Transactions on Neural Networks 10.3 (1999), pp. 626-634. +[13] Francis R Bach and Michael I Jordan. "Kernel independent component analysis". In: Journal of Machine Learning Research 3 (2002), pp. 1-48. +[14] Daniel D. Lee and H. Sebastian Seung. "Learning the parts of objects by non-negative matrix factorization". In: Nature 401.6755 (1999), pp. 788-791. +[15] Patrik O Hoyer. "Non-negative matrix factorization with sparseness constraints". In: Journal of Machine Learning Research 5 (2004), pp. 1457-1469. +[16] Paris Smaragdis. "Non-negative matrix factor deconvolution; extraction of multiple sound sources from monophonic inputs". In: International Conference on Independent Component Analysis and Signal Separation. Springer. 2004, pp. 494-499. +[17] C. L. Bennett et al. "First-Year Wilkinson Microwave Anisotropy Probe (WMAP) Observations: Foreground Emission". In: Astrophysical Journal Supplement Series 148.1 (2003), pp. 97-117. +[18] Felix Franke et al. "Bayes optimal template matching for spike sorting-combining fisher discriminant analysis with optimal filtering". In: Journal of Computational Neuroscience 38 (2015), pp. 439-459. + +[19] G. Kovács et al. “A box-fitting algorithm in the search for periodic transits”. In: *Astronomy & Astrophysics* 391 (2002), pp. 369–377. +[20] Ashish Bora et al. "Compressed sensing using generative models". In: International Conference on Machine Learning. PMLR. 2017, pp. 537-546. +[21] Zhuo Chen et al. "Deep attractor network for single-microphone speaker separation". In: 2017 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). 2017, pp. 246-250. +[22] Daniel Stoller et al. Wave-u-net: A multi-scale neural network for end-to-end audio source separation. 2018. arXiv: 1806.03185 [cs.SD]. +[23] Muhammad Asim et al. "Invertible generative models for inverse problems: mitigating representation error and dataset bias". In: International Conference on Machine Learning. PMLR. 2020, pp. 399-409. +[24] Jay Whang et al. "Solving inverse problems with a flow-based noise model". In: International Conference on Machine Learning. PMLR. 2021, pp. 11146-11157. +[25] Tristan SW Stevens et al. “Removing Structured Noise using Diffusion Models”. In: Transactions on Machine Learning Research (2025). +[26] Abubakar Abid and James Zou. Contrastive variational autoencoder enhances salient features. 2019. arXiv: 1902.04601 [cs.LG]. +[27] Abubakar Abid et al. "Exploring patterns enriched in a dataset with contrastive principal component analysis". In: Nature Communications 9, 2134 (2018). +[28] Kristen A Severson et al. "Unsupervised learning with contrastive latent variable models". In: Proceedings of the AAAI Conference on Artificial Intelligence. Vol. 33. 01. 2019, pp. 4862-4869. +[29] Didong Li et al. "Probabilistic contrastive dimension reduction for case-control study data". In: The Annals of Applied Statistics 18.3 (2024), pp. 2207-2229. +[30] Yang Song et al. "Score-Based Generative Modeling through Stochastic Differential Equations". In: International Conference on Learning Representations. 2021. +[31] Bahjat Kawar et al. "SNIPS: Solving Noisy Inverse Problems Stochastically". In: Advances in Neural Information Processing Systems. Ed. by M. Ranzato et al. Vol. 34. Curran Associates, Inc., 2021, pp. 21757-21769. +[32] Yang Song et al. "Solving Inverse Problems in Medical Imaging with Score-Based Generative Models". In: International Conference on Learning Representations. 2022. +[33] Bahjat Kawar et al. “Denoising Diffusion Restoration Models”. In: Advances in Neural Information Processing Systems. Ed. by S. Koyejo et al. Vol. 35. Curran Associates, Inc., 2022, pp. 23593-23606. +[34] Hyungjin Chung et al. "Diffusion Posterior Sampling for General Noisy Inverse Problems". In: International Conference on Learning Representations. 2023. +[35] Berthy T. Feng et al. "Score-Based Diffusion Models as Principled Priors for Inverse Imaging". In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). 2023, pp. 10520-10531. +[36] Yuanzhi Zhu et al. "Denoising Diffusion Models for Plug-and-Play Image Restoration". In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops. 2023, pp. 1219–1229. +[37] Ronan Legin et al. "Beyond Gaussian noise: A Generalized approach to likelihood analysis with non-Gaussian noise". In: The Astrophysical Journal Letters 949.2 (2023), p. L41. +[38] Benjamin Boys et al. "Tweedie Moment Projected Diffusions for Inverse Problems". In: Transactions on Machine Learning Research (2024). +[39] Jiaming Song et al. "Pseudoinverse-guided diffusion models for inverse problems". In: International Conference on Learning Representations. 2023. +[40] François Rozet et al. “Learning Diffusion Priors from Observations by Expectation Maximization”. In: Advances in Neural Information Processing Systems. Ed. by A. Globerson et al. Vol. 37. Curran Associates, Inc., 2024, pp. 87647-87682. +[41] Jascha Sohl-Dickstein et al. “Deep unsupervised learning using nonequilibrium thermodynamics”. In: International Conference on Machine Learning. PMLR. 2015, pp. 2256–2265. + +[42] Jonathan Ho et al. “Denoising Diffusion Probabilistic Models”. In: Advances in Neural Information Processing Systems. Ed. by H. Larochelle et al. Vol. 33. Curran Associates, Inc., 2020, pp. 6840–6851. +[43] Yang Song and Stefano Ermon. "Generative Modeling by Estimating Gradients of the Data Distribution". In: Advances in Neural Information Processing Systems. Ed. by H. Wallach et al. Vol. 32. Curran Associates, Inc., 2019. +[44] Jiaming Song et al. “Denoising Diffusion Implicit Models”. In: International Conference on Learning Representations. 2021. +[45] Tero Karras et al. "Elucidating the Design Space of Diffusion-Based Generative Models". In: Advances in Neural Information Processing Systems. Ed. by S. Koyejo et al. Vol. 35. Curran Associates, Inc., 2022, pp. 26565-26577. +[46] Brian D.O. Anderson. "Reverse-time diffusion equation models". In: Stochastic Processes and their Applications 12.3 (1982), pp. 313-326. +[47] Aapo Hyvarinen and Peter Dayan. "Estimation of non-normalized statistical models by score matching". In: Journal of Machine Learning Research 6 (2005), pp. 695-709. +[48] Pascal Vincent. “A connection between score matching and denoising autoencoders”. In: Neural computation 23.7 (2011), pp. 1661–1674. +[49] Alexander Quinn Nichol and Prafulla Dhariwal. "Improved denoising diffusion probabilistic models". In: International Conference on Machine Learning. PMLR. 2021, pp. 8162-8171. +[50] Bradley Efron. “Tweedie's formula and selection bias”. In: Journal of the American Statistical Association 106.496 (2011), pp. 1602–1614. +[51] Giannis Daras et al. A survey on diffusion models for inverse problems. 2024. arXiv: 2410.00083 [cs.LG]. +[52] Magnus R Hestenes, Eduard Stiefel, et al. "Methods of conjugate gradients for solving linear systems". In: Journal of Research of the National Bureau of Standards 49.6 (1952), pp. 409-436. +[53] Arthur P Dempster et al. "Maximum likelihood from incomplete data via the EM algorithm". In: Journal of Royal Statistical Society 39 (1977), pp. 1-22. +[54] David Heurtel-Depeiges et al. "Listening to the noise: Blind Denoising with Gibbs Diffusion". In: International Conference on Machine Learning. 2024. +[55] Lénaïc Chizat et al. “Faster Wasserstein Distance Estimation with the Sinkhorn Divergence”. In: Advances in Neural Information Processing Systems. Ed. by H. Larochelle et al. Vol. 33. Curran Associates, Inc., 2020, pp. 2257–2269. +[56] Pablo Lemos et al. "PQMass: Probabilistic Assessment of the Quality of Generative Models using Probability Mass Estimation". In: International Conference on Learning Representations. 2025. +[57] Olga Russakovsky et al. "ImageNet Large Scale Visual Recognition Challenge". In: International Journal of Computer Vision 115.3 (2015), pp. 211-252. +[58] Y. Lecun et al. "Gradient-based learning applied to document recognition". In: Proceedings of the IEEE 86.11 (1998), pp. 2278-2324. +[59] Olaf Ronneberger et al. "U-Net: Convolutional Networks for Biomedical Image Segmentation". In: International Conference on Medical Image Computing and Computer-Assisted Intervention. Springer. 2015, pp. 234-241. +[60] Ashish Vaswani et al. "Attention is All you Need". In: Advances in Neural Information Processing Systems. Ed. by I. Guyon et al. Vol. 30. Curran Associates, Inc., 2017. +[61] William Peebles and Saining Xie. "Scalable Diffusion Models with Transformers". In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). 2023, pp. 4195-4205. +[62] Martin Heusel et al. "GANs Trained by a Two Time-Scale Update Rule Converge to a Local Nash Equilibrium". In: Advances in Neural Information Processing Systems. Ed. by I. Guyon et al. Vol. 30. Curran Associates, Inc., 2017. +[63] A. M. Koekemoer et al. "The COSMOS Survey: Hubble Space Telescope Advanced Camera for Surveys Observations and Data Processing". In: Astrophysical Journal Supplement Series 172.1 (2007), pp. 196-202. + +[64] Kyle W. Willett et al. "Galaxy Zoo: morphological classifications for 120 000 galaxies in HST legacy imaging". In: Monthly Notices of the Royal Astronomical Society 464.4 (2017), pp. 4176-4203. +[65] Friedemann Zenke and Tim P. Vogels. “The Remarkable Robustness of Surrogate Gradient Learning for Instilling Complex Function in Spiking Neural Networks”. In: Neural Computation 33.4 (2021), pp. 899–925. +[66] Stefan Elfwing et al. "Sigmoid-weighted linear units for neural network function approximation in reinforcement learning". In: Neural Networks 107 (2018). Special issue on deep reinforcement learning, pp. 3-11. +[67] Jimmy Lei Ba et al. Layer Normalization. 2016. arXiv: 1607.06450 [stat.ML]. +[68] Takuya Akiba et al. "Optuna: A next-generation hyperparameter optimization framework". In: Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining. 2019, pp. 2623-2631. +[69] Ethan Perez et al. "Film: Visual reasoning with a general conditioning layer". In: Proceedings of the AAAI conference on artificial intelligence. Vol. 32. 1. 2018. +[70] Robin Rombach et al. "High-Resolution Image Synthesis With Latent Diffusion Models". In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 2022, pp. 10684–10695. +[71] Zhendong Wang et al. "Patch Diffusion: Faster and More Data-Efficient Training of Diffusion Models". In: Advances in Neural Information Processing Systems. Ed. by A. Oh et al. Vol. 36. Curran Associates, Inc., 2023, pp. 72137-72154. +[72] Zhaoyu Zhang et al. "Training Diffusion-based Generative Models with Limited Data". In: International Conference on Machine Learning. 2025. +[73] Jiwan Hur et al. "Expanding Expressiveness of Diffusion Models With Limited Data via Self-Distillation Based Fine-Tuning". In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). 2024, pp. 5028-5037. + +# NeurIPS Paper Checklist + +# 1. Claims + +Question: Do the main claims made in the abstract and introduction accurately reflect the paper's contributions and scope? Answer: [Yes] + +Justification: The method is tested in several different settings. Performance is compared against multiple previous works. + +Guidelines: + +- The answer NA means that the abstract and introduction do not include the claims made in the paper. +- The abstract and/or introduction should clearly state the claims made, including the contributions made in the paper and important assumptions and limitations. A No or NA answer to this question will not be perceived well by the reviewers. +- The claims made should match theoretical and experimental results, and reflect how much the results can be expected to generalize to other settings. +- It is fine to include aspirational goals as motivation as long as it is clear that these goals are not attained by the paper. + +# 2. Limitations + +Question: Does the paper discuss the limitations of the work performed by the authors? + +Answer: [Yes] + +Justification: See Section 6 for a discussion of limitations. Main points include the assumptions of a linear observation model, deterministic mixing, and Gaussian noise model, and the high computational cost of the method. + +Guidelines: + +- The answer NA means that the paper has no limitation while the answer No means that the paper has limitations, but those are not discussed in the paper. +- The authors are encouraged to create a separate "Limitations" section in their paper. +- The paper should point out any strong assumptions and how robust the results are to violations of these assumptions (e.g., independence assumptions, noiseless settings, model well-specification, asymptotic approximations only holding locally). The authors should reflect on how these assumptions might be violated in practice and what the implications would be. +- The authors should reflect on the scope of the claims made, e.g., if the approach was only tested on a few datasets or with a few runs. In general, empirical results often depend on implicit assumptions, which should be articulated. +- The authors should reflect on the factors that influence the performance of the approach. For example, a facial recognition algorithm may perform poorly when image resolution is low or images are taken in low lighting. Or a speech-to-text system might not be used reliably to provide closed captions for online lectures because it fails to handle technical jargon. +- The authors should discuss the computational efficiency of the proposed algorithms and how they scale with dataset size. +- If applicable, the authors should discuss possible limitations of their approach to address problems of privacy and fairness. +- While the authors might fear that complete honesty about limitations might be used by reviewers as grounds for rejection, a worse outcome might be that reviewers discover limitations that aren't acknowledged in the paper. The authors should use their best judgment and recognize that individual actions in favor of transparency play an important role in developing norms that preserve the integrity of the community. Reviewers will be specifically instructed to not penalize honesty concerning limitations. + +# 3. Theory assumptions and proofs + +Question: For each theoretical result, does the paper provide the full set of assumptions and a complete (and correct) proof? + +# Answer: [NA] + +Justification: While we include a number of equations in the paper, there are no formal theoretical results. + +# Guidelines: + +- The answer NA means that the paper does not include theoretical results. +- All the theorems, formulas, and proofs in the paper should be numbered and cross-referenced. +- All assumptions should be clearly stated or referenced in the statement of any theorems. +- The proofs can either appear in the main paper or the supplemental material, but if they appear in the supplemental material, the authors are encouraged to provide a short proof sketch to provide intuition. +- Inversely, any informal proof provided in the core of the paper should be complemented by formal proofs provided in appendix or supplemental material. +- Theorems and Lemmas that the proof relies upon should be properly referenced. + +# 4. Experimental result reproducibility + +Question: Does the paper fully disclose all the information needed to reproduce the main experimental results of the paper to the extent that it affects the main claims and/or conclusions of the paper (regardless of whether the code and data are provided or not)? + +# Answer: [Yes] + +Justification: Datasets used for the experiments are either previously established (1D Manifolds, Grassy MNIST), or in the galaxy images case, are made public. The methods are described in sufficient detail to be reproduced, and the source code has been made public. + +# Guidelines: + +- The answer NA means that the paper does not include experiments. + +- If the paper includes experiments, a No answer to this question will not be perceived well by the reviewers: Making the paper reproducible is important, regardless of whether the code and data are provided or not. + +- If the contribution is a dataset and/or model, the authors should describe the steps taken to make their results reproducible or verifiable. + +- Depending on the contribution, reproducibility can be accomplished in various ways. For example, if the contribution is a novel architecture, describing the architecture fully might suffice, or if the contribution is a specific model and empirical evaluation, it may be necessary to either make it possible for others to replicate the model with the same dataset, or provide access to the model. In general, releasing code and data is often one good way to accomplish this, but reproducibility can also be provided via detailed instructions for how to replicate the results, access to a hosted model (e.g., in the case of a large language model), releasing of a model checkpoint, or other means that are appropriate to the research performed. + +- While NeurIPS does not require releasing code, the conference does require all submissions to provide some reasonable avenue for reproducibility, which may depend on the nature of the contribution. For example + +(a) If the contribution is primarily a new algorithm, the paper should make it clear how to reproduce that algorithm. +(b) If the contribution is primarily a new model architecture, the paper should describe the architecture clearly and fully. +(c) If the contribution is a new model (e.g., a large language model), then there should either be a way to access this model for reproducing the results or a way to reproduce the model (e.g., with an open-source dataset or instructions for how to construct the dataset). +(d) We recognize that reproducibility may be tricky in some cases, in which case authors are welcome to describe the particular way they provide for reproducibility. In the case of closed-source models, it may be that access to the model is limited in some way (e.g., to registered users), but it should be possible for other researchers to have some path to reproducing or verifying the results. + +# 5. Open access to data and code + +Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material? + +Answer: [Yes] + +Justification: Code, for both the method and for simulating / querying the datasets, is publicly available. + +Guidelines: + +- The answer NA means that paper does not include experiments requiring code. +- Please see the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details. +- While we encourage the release of code and data, we understand that this might not be possible, so "No" is an acceptable answer. Papers cannot be rejected simply for not including code, unless this is central to the contribution (e.g., for a new open-source benchmark). +- The instructions should contain the exact command and environment needed to run to reproduce the results. See the NeurIPS code and data submission guidelines (https: //nips.cc/public/guides/CodeSubmissionPolicy) for more details. +- The authors should provide instructions on data access and preparation, including how to access the raw data, preprocessed data, intermediate data, and generated data, etc. +- The authors should provide scripts to reproduce all experimental results for the new proposed method and baselines. If only a subset of experiments are reproducible, they should state which ones are omitted from the script and why. +- At submission time, to preserve anonymity, the authors should release anonymized versions (if applicable). +- Providing as much information as possible in supplemental material (appended to the paper) is recommended, but including URLs to data and code is permitted. + +# 6. Experimental setting/details + +Question: Does the paper specify all the training and test details (e.g., data splits, hyperparameters, how they were chosen, type of optimizer, etc.) necessary to understand the results? + +Answer: [Yes] + +Justification: The most important training details are included in the text, with additional details provided in appendices. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The experimental setting should be presented in the core of the paper to a level of detail that is necessary to appreciate the results and make sense of them. +- The full details can be provided either with the code, in appendix, or as supplemental material. + +# 7. Experiment statistical significance + +Question: Does the paper report error bars suitably and correctly defined or other appropriate information about the statistical significance of the experiments? + +Answer: [No] + +Justification: We compare performance using point statistics. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The authors should answer "Yes" if the results are accompanied by error bars, confidence intervals, or statistical significance tests, at least for the experiments that support the main claims of the paper. + +- The factors of variability that the error bars are capturing should be clearly stated (for example, train/test split, initialization, random drawing of some parameter, or overall run with given experimental conditions). +- The method for calculating the error bars should be explained (closed form formula, call to a library function, bootstrap, etc.) +- The assumptions made should be given (e.g., Normally distributed errors). +- It should be clear whether the error bar is the standard deviation or the standard error of the mean. +- It is OK to report 1-sigma error bars, but one should state it. The authors should preferably report a 2-sigma error bar than state that they have a $96\%$ CI, if the hypothesis of Normality of errors is not verified. +- For asymmetric distributions, the authors should be careful not to show in tables or figures symmetric error bars that would yield results that are out of range (e.g. negative error rates). +- If error bars are reported in tables or plots, The authors should explain in the text how they were calculated and reference the corresponding figures or tables in the text. + +# 8. Experiments compute resources + +Question: For each experiment, does the paper provide sufficient information on the computer resources (type of compute workers, memory, time of execution) needed to reproduce the experiments? + +Answer: [Yes] + +Justification: Runtime and computational resources for the experiments in the paper are detailed in Appendix G. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The paper should indicate the type of compute workers CPU or GPU, internal cluster, or cloud provider, including relevant memory and storage. +- The paper should provide the amount of compute required for each of the individual experimental runs as well as estimate the total compute. +- The paper should disclose whether the full research project required more compute than the experiments reported in the paper (e.g., preliminary or failed experiments that didn't make it into the paper). + +# 9. Code of ethics + +Question: Does the research conducted in the paper conform, in every respect, with the NeurIPS Code of Ethics https://neurips.cc/public/EthicsGuidelines? + +Answer: [Yes] + +Justification: The authors have reviewed the guidelines and verify that this work abides by the Code of Ethics. + +Guidelines: + +The answer NA means that the authors have not reviewed the NeurIPS Code of Ethics. +- If the authors answer No, they should explain the special circumstances that require a deviation from the Code of Ethics. +- The authors should make sure to preserve anonymity (e.g., if there is a special consideration due to laws or regulations in their jurisdiction). + +# 10. Broader impacts + +Question: Does the paper discuss both potential positive societal impacts and negative societal impacts of the work performed? + +Answer: [NA] + +Justification: This paper does not work with datasets with societal impact. The method does not have immediate runway for malicious usage. + +Guidelines: + +- The answer NA means that there is no societal impact of the work performed. +- If the authors answer NA or No, they should explain why their work has no societal impact or why the paper does not address societal impact. +- Examples of negative societal impacts include potential malicious or unintended uses (e.g., disinformation, generating fake profiles, surveillance), fairness considerations (e.g., deployment of technologies that could make decisions that unfairly impact specific groups), privacy considerations, and security considerations. +- The conference expects that many papers will be foundational research and not tied to particular applications, let alone deployments. However, if there is a direct path to any negative applications, the authors should point it out. For example, it is legitimate to point out that an improvement in the quality of generative models could be used to generate deepfakes for disinformation. On the other hand, it is not needed to point out that a generic algorithm for optimizing neural networks could enable people to train models that generate Deepfakes faster. +- The authors should consider possible harms that could arise when the technology is being used as intended and functioning correctly, harms that could arise when the technology is being used as intended but gives incorrect results, and harms following from (intentional or unintentional) misuse of the technology. +- If there are negative societal impacts, the authors could also discuss possible mitigation strategies (e.g., gated release of models, providing defenses in addition to attacks, mechanisms for monitoring misuse, mechanisms to monitor how a system learns from feedback over time, improving the efficiency and accessibility of ML). + +# 11. Safeguards + +Question: Does the paper describe safeguards that have been put in place for responsible release of data or models that have a high risk for misuse (e.g., pretrained language models, image generators, or scraped datasets)? + +Answer: [NA] + +Justification: All experiments are performed on datasets that are low risk. Two of the datasets are simulated, the third one uses astronomical images. + +Guidelines: + +- The answer NA means that the paper poses no such risks. +- Released models that have a high risk for misuse or dual-use should be released with necessary safeguards to allow for controlled use of the model, for example by requiring that users adhere to usage guidelines or restrictions to access the model or implementing safety filters. +- Datasets that have been scraped from the Internet could pose safety risks. The authors should describe how they avoided releasing unsafe images. +- We recognize that providing effective safeguards is challenging, and many papers do not require this, but we encourage authors to take this into account and make a best faith effort. + +# 12. Licenses for existing assets + +Question: Are the creators or original owners of assets (e.g., code, data, models), used in the paper, properly credited and are the license and terms of use explicitly mentioned and properly respected? + +Answer: [Yes] + +Justification: Use of existing assets is mentioned and relevant papers are cited in the main text. Further details, such as location of repositories and licenses, are included in supplemental materials. + +Guidelines: + +- The answer NA means that the paper does not use existing assets. +- The authors should cite the original paper that produced the code package or dataset. +- The authors should state which version of the asset is used and, if possible, include a URL. + +- The name of the license (e.g., CC-BY 4.0) should be included for each asset. +- For scraped data from a particular source (e.g., website), the copyright and terms of service of that source should be provided. +- If assets are released, the license, copyright information, and terms of use in the package should be provided. For popular datasets, paperswithcode.com/datasets has curated licenses for some datasets. Their licensing guide can help determine the license of a dataset. +- For existing datasets that are re-packaged, both the original license and the license of the derived asset (if it has changed) should be provided. +- If this information is not available online, the authors are encouraged to reach out to the asset's creators. + +# 13. New assets + +Question: Are new assets introduced in the paper well documented and is the documentation provided alongside the assets? + +Answer: [Yes] + +Justification: The code base for the method is publicly available. For the galaxies images experiment, we make the dataset of denoised images with galaxy light publicly available as well for further use by the astrophysics community. + +Guidelines: + +- The answer NA means that the paper does not release new assets. +- Researchers should communicate the details of the dataset/code/model as part of their submissions via structured templates. This includes details about training, license, limitations, etc. +- The paper should discuss whether and how consent was obtained from people whose asset is used. +- At submission time, remember to anonymize your assets (if applicable). You can either create an anonymized URL or include an anonymized zip file. + +# 14. Crowdsourcing and research with human subjects + +Question: For crowdsourcing experiments and research with human subjects, does the paper include the full text of instructions given to participants and screenshots, if applicable, as well as details about compensation (if any)? + +Answer: [NA] + +Justification: This paper does not involve crowd-sourcing or human subjects. + +Guidelines: + +- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. +- Including this information in the supplemental material is fine, but if the main contribution of the paper involves human subjects, then as much detail as possible should be included in the main paper. +- According to the NeurIPS Code of Ethics, workers involved in data collection, curation, or other labor should be paid at least the minimum wage in the country of the data collector. + +# 15. Institutional review board (IRB) approvals or equivalent for research with human subjects + +Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals (or an equivalent approval/review based on the requirements of your country or institution) were obtained? + +Answer: [NA] + +Justification: This paper does not involve crowd-sourcing or human subjects. + +Guidelines: + +- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. +- Depending on the country in which research is conducted, IRB approval (or equivalent) may be required for any human subjects research. If you obtained IRB approval, you should clearly state this in the paper. +- We recognize that the procedures for this may vary significantly between institutions and locations, and we expect authors to adhere to the NeurIPS Code of Ethics and the guidelines for their institution. +- For initial submissions, do not include any information that would break anonymity (if applicable), such as the institution conducting the review. + +# 16. Declaration of LLM usage + +Question: Does the paper describe the usage of LLMs if it is an important, original, or non-standard component of the core methods in this research? Note that if the LLM is used only for writing, editing, or formatting purposes and does not impact the core methodology, scientific rigorousness, or originality of the research, declaration is not required. + +Answer: [NA] + +Justification: Our usage of LLMs falls under the category of: "spell checkers and grammar suggestions, programming aid for editing purposes". + +Guidelines: + +- The answer NA means that the core method development in this research does not involve LLMs as any important, original, or non-standard components. +- Please refer to our LLM policy (https://neurips.cc/Conferences/2025/LLM) for what should or should not be described. + +Algorithm 1: MVSS WITH JOINT DIFFUSION +Input: Dataset $\mathcal{D} = \{\mathbf{y}_{i_{\alpha}}^{\alpha},\{\mathbf{A}_{i_{\alpha}}^{\alpha \beta}\}_{\beta = 1}^{N_s},\pmb{\Sigma}_{i_{\alpha}}^{\alpha}\}_{\alpha = 1}^{N_{\mathrm{views}}}$ , number of sources $N_{s}$ , number of views $N_{\mathrm{views}}$ , initial denoiser parameters $\Theta_0 = \{\theta_0^\beta \}_{\beta = 1}^{N_s}$ , number of EM rounds $K$ +Output: Trained diffusion model priors with denoiser parameters $\Theta_K$ +for $k\gets 0$ to $K - 1$ do foreach $(\mathbf{y}_{i_{\alpha}}^{\alpha},\{\mathbf{A}_{i_{\alpha}}^{\alpha \beta}\} ,\pmb{\Sigma}_{i_{\alpha}}^{\alpha})\in \mathcal{D}$ do $\{\mathbf{x}_{i_{\alpha}}^{\beta}\}_{\beta = 1}^{N_s}\sim q_{\Theta_k}\bigl (\{\mathbf{x}^\beta \} |\mathbf{y}_{i_\alpha}^\alpha ,\{\mathbf{A}_{i_\alpha}^{\alpha \beta}\} \bigr)$ // E step using equation 14 end + $\Theta_{k + 1} = \arg \max_{\Theta}\left[\sum_{i_{\alpha}}\sum_{\beta}\log q_{\theta \beta}(\mathbf{x}_{i_{\alpha}}^{\beta})\right] / /$ M step using equation 5 +end +return $\Theta_K$ + +Algorithm 2: SIMPLIFIED CONTRASTIVE MVSS WITH JOINT DIFFUSION +Input: Dataset $\mathcal{D} = \{\mathbf{y}_{i_{\alpha}}^{\alpha},\{\mathbf{A}_{i_{\alpha}}^{\alpha \beta}\}_{\beta = 1}^{N_s},\pmb{\Sigma}_{i_{\alpha}}^{\alpha}\}_{\alpha = 1}^{N_{\mathrm{views}}},\mathbf{A}_{i_{\alpha}}^{\alpha \beta} = 0$ if $\beta >\alpha$ , number of sources $N_{s}$ number of views $N_{\mathrm{views}} = N_{s}$ , initial denoiser parameters $\Theta_0 = \{\theta_0^\beta \}_{\beta = 1}^{N_s}$ , number of EM rounds $K$ +Output: Trained diffusion model priors with denoiser parameters $\Theta_K$ +for $a\gets 1$ to $N_{\mathrm{views}}$ do for $k\gets 0$ to $K - 1$ do foreach $\left(\mathbf{y}_{i_{\alpha}}^{\alpha},\{\mathbf{A}_{i_{\alpha}}^{\alpha \beta}\} ,\pmb{\Sigma}_{i_{\alpha}}^{\alpha}\right)\in \mathcal{D}$ with $\alpha = a$ do $\{\mathbf{x}_{i_\alpha}^\beta \}_{\beta = 1}^a\sim q_{\{\theta_k^\beta \}_{\beta = 1}^a}\bigl (\{\mathbf{x}^\beta \} |\mathbf{y}_{i_\alpha}^\alpha ,\{\mathbf{A}_{i_\alpha}^{\alpha \beta}\} \bigr) / /$ E step using equation 14 end $\theta_{k + 1}^{a} = \arg \max_{\theta^{a}}\left[\sum_{i_{\alpha}}\log q_{\theta^{a}}(\mathbf{x}_{i_{\alpha}}^{a})\right] / /$ M step for only source $a$ end return $\Theta_K$ + +# A Joint Sampling Algorithms + +In Algorithm 1 we present an algorithmic summary of our MVSS method using joint sampling. This method conducts a joint EM training for all sources simultaneously. In Algorithm 2 we present a summary of the contrastive MVSS simplification. This method trains each source sequentially under the assumption that view $\alpha = 1$ only contains source $\beta = 1$ , view $\alpha = 2$ only contains sources $\beta \in \{1,2\}$ , and so on (see Section 4). While Algorithm 2 discards the information about source $\beta$ for views $\alpha \neq \beta$ , it reduces the computational complexity. + +# B Gibbs Sampling + +Gibbs sampling with diffusion models, originally proposed for blind denoising [54], can be easily extended to the full MVSS problem. We start by selecting a source $\beta_{g}$ . Given the $t = 0$ source realizations for $\{\mathbf{x}_0^{\beta'}\}_{\beta' \neq \beta_g}$ , an observation $\mathbf{y}^{\alpha}$ , a set of known mixing matrices $\{\mathbf{A}^{\alpha \beta}\}_{\beta=1}^{N_s}$ , and the noise covariance $\boldsymbol{\Sigma}^{\alpha}$ , then the posterior score for the remaining source becomes: + +$$ +\begin{array}{l} \nabla_ {\mathbf {x} _ {t} ^ {\beta_ {g}}} \log p (\mathbf {x} _ {t} ^ {\beta_ {g}} | \mathbf {y} ^ {\alpha}, \{\mathbf {A} ^ {\alpha \beta} \} _ {\beta = 1} ^ {N _ {s}}, \{\mathbf {x} _ {0} ^ {\beta^ {\prime}} \} _ {\beta^ {\prime} \neq \beta_ {g}}) = \nabla_ {\mathbf {x} _ {t} ^ {\beta_ {g}}} \log p (\mathbf {x} _ {t} ^ {\beta_ {g}}) + \\ \nabla_ {\mathbf {x} _ {t} ^ {\beta_ {g}}} \log p \left(\mathbf {y} ^ {\alpha} \mid \mathbf {x} _ {t} ^ {\beta_ {g}}, \left\{\mathbf {A} ^ {\alpha \beta} \right\} _ {\beta = 1} ^ {N _ {s}}, \left\{\mathbf {x} _ {0} ^ {\beta^ {\prime}} \right\} _ {\beta^ {\prime} \neq \beta_ {g}}\right). \tag {18} \\ \end{array} +$$ + +For conciseness, we can introduce the observation residual $\mathbf{r}^{\alpha} = \mathbf{y}^{\alpha} - \sum_{\beta' \neq \beta_g} \mathbf{A}^{\alpha \beta'} \mathbf{x}_0^{\beta'}$ . We can then write the remaining likelihood term in Equation 18 as: + +$$ +p \left(\mathbf {y} ^ {\alpha} \mid \mathbf {x} _ {t} ^ {\beta_ {g}}, \left\{\mathbf {A} ^ {\alpha \beta} \right\} _ {\beta = 1} ^ {N _ {s}}, \left\{\mathbf {x} ^ {\beta^ {\prime}} \right\} _ {\beta^ {\prime} \neq \beta_ {g}}\right) = \int p \left(\mathbf {r} ^ {\alpha} \mid \mathbf {x} _ {0} ^ {\beta_ {g}}\right) p \left(\mathbf {x} _ {0} ^ {\beta_ {g}} \mid \mathbf {x} _ {t} ^ {\beta_ {g}}\right) d \mathbf {x} _ {0} ^ {\beta_ {g}}, \tag {19} +$$ + +which is identical to Equation 7 with the observation replaced by the residual. Since Gibbs sampling fixes all but one source at a time, the posterior evaluation reduces to traditional posterior sampling with a diffusion prior with the observation replaced by the residual. An implementation of the Gibbs sampling procedure for MVSS can be found in the provided code. + +# C One-Dimensional Manifold Experiment Details + +We generate random one-dimensional manifolds following the steps outlined in [65]. In all experiments in Section 5.1, the smoothness parameters for the manifolds corresponding to the first, second, and third source distributions are set to 3, 4, and 5 respectively. Similarly, we use the same denoiser architecture for each source and each experiment. The denoiser is a multi-layer perceptron (MLP) composed of 3 hidden layers with 256 neurons and SiLU activation function [66], followed by a layer normalization function [67]. The denoiser is conditioned on noise and noise embeddings are generated with the sinusoidal positional encoding method [60]. + +For the metrics, we use 16384 samples for the Sinkhorn divergence evaluation, 1024 samples for the PQMass evaluation, and 16384 samples for the PSNR evaluation. Note that we use the same number of samples from the true distribution and the prior / posterior distribution for all metrics. Metrics on the posterior samples are calculated by comparing to the true source value for the corresponding observation. For the PQMass evaluation, we use 1000 tessellations and otherwise keep the default parameters. + +# C.1 Contrastive MVSS + +In the contrastive MVSS experiment, we use the simplified contrastive MVSS algorithm to train the denoiser models. We train the $\beta = 1,2,3$ denoiser models for 16, 32, and 64 EM laps respectively. Following Rozet et al. [40], we reinitialize the optimizer and learning rate after each EM lap while keeping the current denoiser parameters. The MLP takes as input a concatenated vector of the diffused sample at time $t$ and the corresponding noise embedding. We summarize other training hyperparameters in Table 2. + +For the Gibbs sampling approach we use the same dataset (up to $\beta = 2$ ), denoiser architectures and training hyperparameters as for the joint posterior sampling approach. The posterior sampling parameters are set to 64 Gibbs rounds with 256 PC steps, and we maintain one corrector step per predictor step. The number of PC steps is lowered so that the number of denoiser and vector-Jacobian product evaluations per EM lap are roughly equivalent for Gibbs and joint sampling. + +For both the Gibbs and joint sampling approaches, the denoisers are initialized using samples from a Gaussian prior. Since we are using our contrastive algorithm, only the diffusion model for the current source, $\beta$ , is replaced by the Gaussian prior, and the remaining diffusion models for sources $\beta' < \beta$ are kept to the optimum from the previous views (see Algorithm 2). The posterior is then sampled as usual. To optimize the parameters of this Gaussian prior, we use the same EM procedure as for the diffusion model. The final Gaussian prior is used to generate an initial set of posterior samples which are used to train the initial diffusion model, $d_{\theta_0^\beta}(\mathbf{x}_t,t)$ . + +For the baselines, we use modified implementations built for incomplete data. For PCPCA, we minimize a loss function defined across our two views, $\beta = \mathrm{bkg}$ , targ: + +$$ +\begin{array}{l} \mathcal {L} (\mathbf {W}, \mu) = - \frac {1}{2} \left(\sum_ {i = 1} ^ {N _ {\text {t a r g}}} \log \det \left(\mathbf {C} _ {\mathbf {i}}\right) + \left(\mathbf {y} _ {i} ^ {\text {t a r g}} - \mathbf {A} _ {i} ^ {\text {t a r g}} \mu\right) ^ {\top} \mathbf {C} _ {i} ^ {- 1} \left(\mathbf {y} _ {i} ^ {\text {t a r g}} - \mathbf {A} _ {i} ^ {\text {t a r g}} \mu\right)\right) \\ + \frac {\gamma}{2} \left(\sum_ {j = 1} ^ {N _ {\mathrm {b k g}}} \log \det (\mathbf {D _ {j}}) + \left(\mathbf {y} _ {j} ^ {\mathrm {b k g}} - \mathbf {A} _ {j} ^ {\mathrm {b k g}} \mu\right) ^ {\top} \mathbf {D} _ {j} ^ {- 1} \left(\mathbf {y} _ {j} ^ {\mathrm {b k g}} - \mathbf {A} _ {j} ^ {\mathrm {b k g}} \mu\right)\right), \\ \end{array} +$$ + +
ParameterContrastive MVSSMixed MVSS
MLP Parameters
ActivationSiLUSiLU
Time Embedding Features64128
Training Parameters
OptimizerAdamAdam
SchedulerLinearLinear
Initial Learning Rate10-310-4
Final Learning Rate10-610-5
Gradient Norm Clipping1.01.0
Optimization Steps per EM Lap65,53665,536
Batch Size10241024
Gaussian Initialization EM Laps168192
Sampling Parameters
Noise Minimum10-35 × 10-3
Noise Maximum1011.5 × 101
Conjugate Gradient Denominator Minimum0.010-3
Conjugate Gradient Regularization0.010-3
Predictor-Corrector (PC) Steps16,38416,384
Corrections per PC Step11
PC τ10-18 × 10-2
+ +Table 2: Hyperparameters for denoiser training and sampling on the one-dimensional manifold experiments. + +where $\mu$ is the learnable mean parameter for the target source. The views, $\mathbf{y}_j^{\mathrm{bkg}}$ , $\mathbf{y}_i^{\mathrm{targ}}$ , are the same as those defined in Section 3 but with the mixing matrix $\mathbf{A}_j^{\mathrm{bkg}}$ applied to the sources underlying $\mathbf{y}_j^{\mathrm{bkg}}$ and the mixing matrix $\mathbf{A}_i^{\mathrm{targ}}$ applied to the sources underlying $\mathbf{y}_i^{\mathrm{targ}}$ . Here, $\gamma$ is a tunable parameter that controls the relative importance of the variations in the background and target data. The dependence on the weights, $\mathbf{W}$ , comes from the two covariance matrices given by: + +$$ +\mathbf {C} _ {i} = \mathbf {A} _ {i} ^ {\operatorname {t a r g}} \mathbf {W} \mathbf {W} ^ {\top} \left(\mathbf {A} _ {i} ^ {\operatorname {t a r g}}\right) ^ {\top} + \sigma_ {i} ^ {2} \mathbb {I} \tag {21} +$$ + +$$ +\mathbf {D} _ {j} = \mathbf {A} _ {j} ^ {\mathrm {b k g}} \mathbf {W} \mathbf {W} ^ {\top} \left(\mathbf {A} _ {j} ^ {\mathrm {b k g}}\right) ^ {\top} + \sigma_ {j} ^ {2} \mathbb {I}, \tag {22} +$$ + +where $\sigma_{i}$ is the standard deviation of the observation noise. Note that we only optimize the weights and the mean, since the noise is known. We initialize the weights using the empirical covariance derived from source values given by multiplying the observations with the pseudo-inverse of the mixing matrix. We find that this smart initialization considerably improves the performance of PCPCA on incomplete data. + +For CLVM-Linear and CLVM-VAE, we explicitly account for the mixing matrix in both the encoding and decoding steps. For CLVM-Linear, the encoded distribution is given by a small modification to the original equations for the latent variable distributions: + +$$ +\boldsymbol {\Sigma} _ {j} ^ {\mathrm {b k g}} = \frac {1}{\sigma_ {j} ^ {2}} \left(\sigma_ {j} ^ {2} \mathbb {I} + \mathbf {S} ^ {\top} \mathbf {S}\right) \tag {23} +$$ + +$$ +\mu_ {j} ^ {\mathrm {b k g}} = \frac {1}{\sigma_ {j} ^ {2}} \boldsymbol {\Sigma} _ {j} ^ {\mathrm {b k g}} \mathbf {S} ^ {\top} \left(\mathbf {y} _ {j} ^ {\mathrm {b k g}} - \mu^ {\mathrm {b k g}}\right) \tag {24} +$$ + +$$ +\boldsymbol {\Sigma} _ {i} ^ {\text {j o i n t}} = \frac {1}{\sigma_ {i} ^ {2}} \left(\sigma_ {i} ^ {2} \mathbb {I} + \mathbf {M} ^ {\top} \mathbf {M}\right) \tag {25} +$$ + +$$ +\mu_ {i} ^ {\text {j o i n t}} = \frac {1}{\sigma_ {i} ^ {2}} \Sigma_ {i} ^ {\text {j o i n t}} \mathbf {M} ^ {\top} \left(\mathbf {y} _ {i} ^ {\text {j o i n t}} - \mu^ {\text {j o i n t}}\right), \tag {26} +$$ + +where $\pmb{\Sigma}_{j}^{\mathrm{bkg}}$ and $\mu_j^{\mathrm{bkg}}$ are the covariance and mean for the latents of $\mathbf{y}_j^{\mathrm{bkg}}$ . The covariance and mean $\pmb{\Sigma}_{i}^{\mathrm{joint}}$ and $\mu_i^{\mathrm{joint}}$ correspond to the joint latents $(\mathbf{z}^{\mathrm{bkg}},\mathbf{z}^{\mathrm{targ}})$ for $(\mathbf{y}_i^{\mathrm{joint}})^\top = [(\mathbf{y}_i^{\mathrm{bkg}})^\top,(\mathbf{y}_i^{\mathrm{targ}})^\top]$ . The + +
ParameterGrassy MNISTGalaxy
U-Net Parameters
Channels per Level(32, 64, 128)(64, 128, 256, 256, 512)
Residual Blocks per Level(2, 2, 2)(2, 2, 2, 2, 2)
Attention Heads per Level(0, 0, 4)(0, 0, 4, 8, 16)
Dropout Rate0.10.1
ActivationSiLUSiLU
Training Parameters
OptimizerAdamAdam
SchedulerCosine DecayCosine Decay
Initial learning rate10-310-5
Gradient Norm Clipping1.01.0
Optimization Steps per EM Lap40964096
Batch size192064
Gaussian Initialization EM laps324
Sampling Parameters
Noise Minimum10-410-2
Noise Maximum102101
Conjugate Gradient Denominator Minimum10-210-3
Conjugate Gradient Regularization10-610-2
Conjugate Gradient Error Threshold5 × 10-2101
Predictor-Corrector (PC) Steps25664
Corrections per PC Step11
PC τ10-210-1
EMA decay0.9990.995
+ +Table 3: Hyperparameters for denoiser training and sampling for the Grassy MNIST experiments and the Galaxy experiment. During sampling, conjugate gradient calculations whose total residuals exceed the conjugate gradient error threshold are recalculated using the denominator minimum and regularization. + +matrices $\mathbf{S},\mathbf{W}$ are the background and target factor loading matrix in CLVM, and $\mathbf{M}$ is the concatenated factor loading matrix, $\mathbf{M}^{\top} = [\mathbf{S}^{\top},\mathbf{W}^{\top}]$ . For CLVM-VAE, the encoded distribution is calculated by explicitly passing in the mixing matrix to the encoder network. For both CLVM-Linear and CLVM-VAE, the decoder outputs the complete source signal and the variational loss is calculated by first transforming the sources using the known mixing matrices. The optimization is done by maximizing the evidence lower bound as described in [28]. + +For PCPCA, CLVM-Linear, and CLVM-VAE, we optimize the hyperparameters using a 100 point sweep with the default Bayesian optimization used in optuna [68]. The results are reported using the best hyperparameters for each method on the posterior Sinkhorn divergence for three sources. For PCPCA this is $\gamma = 0.3$ and 5 latent dimensions using a linear learning rate from $10^{-3}$ to 0.0 over 10 epochs8. For CLVM-Linear, this was a dimensionality of 4 for the background latents and 5 for the source latents, with a batch size of 1024, a cosine learning rate initialized to $10^{-4}$ , and 1024 epochs of training. For CLVM-VAE, the encoder architecture were multi-layer perceptrons composed of 3 hidden layers with 256 neurons and a dropout rate of 0.1 followed by a SiLU activation function and a layer normalization function. The decoders follow the same structure. The CLVM-VAE was trained with a batch size of 1024, a cosine learning rate initialized to $5 \times 10^{-4}$ , and 1024 epochs of training. + +# C.2 Mixed MVSS + +In the mixed MVSS experiment, we use Algorithm 1 to train the diffusion models. We train for 70 EM laps and reinitialize the optimizer and learning rate after each EM lap while keeping the current + +
EncoderDecoderFID
MLPMLP18.87
U-Net (full-depth)MLP20.14
U-Net (1 hidden channel)MLP45.37
U-Net (full-depth)U-Net (full-depth)388.26
U-Net (1 hidden channel)U-Net (1 hidden channel)390.27
+ +Table 4: Evaluation of different CLVM-VAE encoder-decoder architectures for MVSS trained on Grassy MNIST. + +denoiser parameters. To condition on the time embedding, the output of each dense layer is passed through a FiLM layer [69]. We summarize other training hyperparameters in Table 2. + +For the Gibbs sampling, only the sampling parameters are changed. To improve performance, the number of Gaussian EM laps is reduced to 16, and we use 512 Gibbs rounds with 256 PC steps. As a result, each EM lap of the Gibbs model is eight times as expensive as the joint sampling laps, so we only run 13 laps of EM. + +For both the Gibbs and joint sampling approaches, the denoisers are initialized using samples from a Gaussian prior. Unlike for the contrastive algorithm, all of the diffusion models are replaced by the Gaussian prior for initialization. The posterior is then sampled as usual. To optimize the parameters of these Gaussian priors, we use the same EM procedure as for the diffusion model. Note that the resulting Gaussian prior will differ by source. The final Gaussian priors are used to generate an initial set of posterior samples which are used to train the initial diffusion models, $d_{\theta_0^{\beta}}(\mathbf{x}_t,t)$ for all $\beta$ . + +# C.2.1 Identifiability + +In the case where $f_{\mathrm{frac}} = 1$ , the mixed MVSS problem cannot be solved. Consider the vector $\mathbf{z}^1 = \mathbf{x}^1 + \mathbf{x}^2$ and the trivial vector $\mathbf{z}^2 = 0$ . Then, for any observation $i_\alpha$ , we can write: + +$$ +\begin{array}{l} \mathbf {y} _ {i _ {\alpha}} ^ {\alpha} = \mathbf {A} _ {i _ {\alpha}} \left(\mathbf {x} _ {i _ {\alpha}} ^ {1} + \mathbf {x} _ {i _ {\alpha}} ^ {2}\right) + \eta_ {i _ {\alpha}} ^ {\alpha} (27) \\ = \mathbf {A} _ {i _ {\alpha}} \left(\mathbf {z} _ {i _ {\alpha}} ^ {1} + \mathbf {z} _ {i _ {\alpha}} ^ {2}\right) + \eta_ {i _ {\alpha}} ^ {\alpha}. (28) \\ \end{array} +$$ + +However, $p(\mathbf{z}^2) = \delta (\mathbf{z}^2)$ . Since we have constructed $\mathbf{x}^1$ and $\mathbf{x}^2$ to lie on 1D manifolds, we know $p(\mathbf{z}^2) \neq p(\mathbf{x}^1)$ and $p(\mathbf{z}^2) \neq p(\mathbf{x}^2)$ . We have found two new sources that match our observations even though one of the sources is guaranteed to not have the same distribution as either of the original sources. Therefore, there is not a unique solution to the source priors when $f_{\mathrm{frac}} = 1$ . + +# D Grassy MNIST Experiments + +The MNIST dataset is available under a CC BY-SA 3.0 license [58]. The ImageNet dataset is made available for non-commercial purposes [57]. We generate our grass images by taking random $28 \times 28$ pixel crops from ImageNet images with the grass label. We use a different set of random crops for each view. For our digits, we use images with the 0 and 1 label. For both the grass and MNIST images, we normalize the pixel values to the range [0, 1]. + +We use the same U-Net denoiser architecture for each source and each experiment. The denoiser and training parameters are presented in Table 3. For sampling, we use an exponential moving average of the model weights. The full sampling and U-Net code is provided with out codebase. For the initialization, we use a Gaussian prior optimized via EM as described in Appendix C. + +For the metrics, we use 8192 for the FID evaluation, 512 samples for the PQMass evaluation, and 8192 samples for the PSNR evaluation. We use the same number of samples from the true distribution and the prior / posterior distribution for all metrics. Metrics on the posterior samples are calculated by comparing to the true source value for the corresponding observation. For the PQMass evaluation, we use 1000 tessellations and otherwise keep the default parameters. + +![](images/08d183743ea546265fdde8f96a85968e743229bc4131791affea207795c495eb.jpg) +Figure 5: Random samples of galaxy posteriors across EM iterations in the galaxy-image experiment. Early iterations fail to isolate galaxy light and show strong small-scale fluctuations. By iteration 4, small-scale fluctuations are separated but residual uncorrelated light remains. By iteration 16, the uncorrelated light is assigned to the random posterior component, leaving nearly noiseless, isolated galaxy posteriors. As in Figure 4, images are split into high-contrast and low-contrast colormaps to highlight the range of features. + +# D.1 Grassy MNIST Baselines + +We optimize the PCPCA, CLVM-Linear, and CLVM-VAE hyperparameters using a 100 point sweep with the default Bayesian optimization used in optuna. The results we present are using the best hyperparameters on the FID for the full-resolution experiment. + +For PCPCA, we use the traditional algorithm on the full-resolution experiment. On the downsampled experiment we fit the PCPCA parameters on the full-resolution subset and then sample using the equation for incomplete data. The PCPCA tunable parameter is set to $\gamma = 0.39$ and we use 5 latent dimensions. + +For CLVM-Linear and CLVM-VAE, we use the traditional algorithm on the full-resolution experiment. On the downsampled experiment we follow the same procedure outlined in Appendix C.1. For CLVM-Linear, we set the dimensionality of the background latents to 265 and the dimensionality of the target latents to 6. We use a cosine learning rate with initial value $1.2 \times 10^{-5}$ trained over batches of 1920 images for $2^{14}$ steps. For CLVM-VAE, we set the dimensionality of the background latents to 380 and the dimensionality of the target latents to 15. The cosine learning rate is used again, but now with an initial value of $2 \times 10^{-4}$ , a batch size of 256, and a total of $2^{14}$ steps. + +The CLVM-VAE method uses an MLP with three hidden layers of size 70 for the encoder and decoder. The number of hidden layers was optimized during the hyperparameter sweep, but the + +
MethodTraining TimeInference Time / Sample
1D +ManifoldGMNIST +Full Res.Galaxy +Images1D +ManifoldGMNIST +Full Res.Galaxy +Images
PCPCA [29]5 s1 m-< 0.1 ms1.5 ms-
CLVM - Linear [28]1.5 m10 m-< 0.1 ms1 ms-
CLVM - VAE [28]18 m33 m-< 0.1 ms1 ms-
DDPRISM-Joint [Ours]32 h68 h48 h22 ms90 ms1.5 s
+ +Table 5: Comparison of computational costs for our method and baselines for three experiments shown in the paper: contrastive 1D manifold experiment with 2 source distributions, full-resolution Grassy MNIST experiment, and the galaxy experiment. Training time refers to the amount of time each method takes to train its corresponding model. Inference time per sample is the time a method takes to obtain a single posterior sample for an observation, given a trained model. 1D Manifold and Grassy MNIST experiments were run on A100 GPUs, and Galaxy Image experiments were run on H100 GPUs. + +encoder and decoder architectures were selected as part of a separate ablation study whose results are summarized in Table 4. In the ablation study, we compared three encoder choices: + +- A fully connected MLP. +- A full-depth U-Net identical to the "downsampling" half of U-Net used for our diffusion models without skip connections. +- A convolutional neural network equivalent to our U-Net implementation with 1 hidden channel (no downsampling). + +We also compared two decoder choices: + +- A fully connected MLP. +- A convolutional neural network equivalent to our U-Net implementation with 1 hidden channel (no upsampling) + +We found that more complex architectures negatively impacted performance, and that the best performance was achieved with a simple MLP decoder and encoder. + +# E Galaxy Images Experiment + +We query data hosted by the Mikulski Archive for Space Telescopes (MAST), which is available in the public domain. We retrieve data files using the astroquery package, which has a 3-clause BSD style license. We generate our galaxy images by querying $128 \times 128$ pixel cutouts centered on the Galaxy Zoo Hubble Catalog [64]. We generate our random fields by making $128 \times 128$ cutouts at random coordinates within the larger COSMOS exposures. This results in 78,707 galaxy images and 257,219 random images. We apply three normalizations in this order: (1) we pass the images through an arcsinh transform with a scaling of 0.1, (2) we scale the data by a factor of 0.2, and (3) we clip the maximum absolute pixel value to 2.0. These three transformations help preserve the morphological features in the brightest sources while stabilizing the diffusion model training. + +We use the same U-Net denoiser architecture for the galaxy and random source. The denoiser and training parameters are presented in Table 3. For sampling, we use an exponential moving average of the model weights. For initialization, we use a Gaussian prior optimized via EM as described in Appendix C.1. All of the code required to reproduce this experiment can be found in the DDPRISM codebase. For completeness, in Figure 5 we show the evolution of the galaxy posterior samples as a function of EM laps. + +
GMNIST Full Res.PosteriorPrior
PQM ↑FID ↓PSNR ↑PQM ↑FID ↓
Model Architecture
MLP0.0549.0317.930.215.36
U-Net, small1.002.4725.280.0615.38
U-Net (default)1.001.5725.600.2020.10
Training Length (EM laps)
20.96.8516.620.199.70
81.000.0427.150.1417.35
321.002.2625.660.0827.96
64 (default)1.001.5725.600.2020.10
Initialization Laps
0 (random initialization)0.9710.4123.350.0122.22
41.000.0026.800.1524.33
161.000.0027.020.216.31
32 (default)1.001.5725.600.2020.10
Dataset Size
Full Dataset (default)1.001.5725.600.2020.10
1/4th Dataset1.004.6423.670.1421.36
1/16th Dataset0.9910.1920.970.0715.16
1/64th Dataset0.038.3415.340.036.55
Sampling Steps
160.875.4121.010.30.20
641.000.8825.020.243.58
256 (default)1.001.5725.600.2020.10
10241.000.5926.500.087.18
+ +Table 6: Ablation study of individual components and parameters of our method on the full-resolution grassy MNIST experiment. For each metric, the arrow indicates whether larger (↑) or smaller (↓) values are optimal. The values used for the Grassy MNIST (GMNIST) experiment in Section 5.2 are denoted as default values. While most of the default values correspond to the optimal performance, we find that decreasing the length of the training and using fewer rounds of Gaussian EM improve the overall performance of the method. + +# F Additional Diffusion Model Details + +For our diffusion models, we use the preconditioning strategy from Karras et al. [45]. Our variance exploding noise schedule is given by: + +$$ +\sigma (t) = \exp \left[ \log \left(\sigma_ {\min }\right) + \left(\log \left(\sigma_ {\max }\right) - \log \left(\sigma_ {\min }\right)\right) * t \right], \tag {29} +$$ + +with minimum noise $\sigma_{\mathrm{min}}$ and maximum noise $\sigma_{\mathrm{max}}$ . During training, we sample the time parameter $t$ from a beta distribution with parameters $\alpha = 3$ , $\beta = 3$ . + +# G Runtime and Computational Cost + +We provide a detailed comparison of the computational costs between our method and baselines in Table 5. All timing was done on NVIDIA A100 GPUs with the exception of the galaxy images experiment that was run on H100s. Both 1D manifold experiments used one A100 (40GB) GPU, both Grassy MNIST experiments used four A100 (40GB) GPUs, and the Galaxy Image experiments used four H100 (80GB) GPUs. Our method is more computationally expensive than the baselines, almost entirely due to the cost of sampling from the diffusion model. However, this computational cost comes with significant improvements in performance and sample quality. + +# H Ablation Studies + +In order to clarify the importance of the individual components of our method, we conduct a series of ablation studies which are summarized in Table 6. We explored the effect of varying the following components on the performance of the method: + +- Diffusion model architecture: We find the model architecture to be important for performance: using an MLP-based architecture (5 hidden layers with 2048 hidden features per layer) gives poor results across all metrics. However, scaling down the UNet model (channels per level: $(32,64,128)\rightarrow (16,32)$ , residual blocks per level: $(2,2,2)\rightarrow (2,2)$ , embedding features: $64\rightarrow 16$ , attention head moved up one level) does not degrade performance appreciably. +- Training length: We observe that the model achieves good performance after as few as 8 EM iterations and that longer training leads to a slight degradation in performance. +- Initialization strategy: The number of Gaussian EM laps used to generate the initial samples (and train the initial diffusion model) has minimal impact on performance. Only starting from a randomly initialized model considerably reduces the performance of the method. +- Training dataset size: We train our method using 1/4th, 1/16th, and 1/64th of the original grass and grass+MNIST datasets. The 1/16th and 1/64th runs are given 1/4th as many EM laps as the original training to account for the smaller dataset, but all other hyperparameters are unchanged. There is a clear degradation in performance as the grass and MNIST datasets are reduced in size. However, even with 1/16th of the original dataset (2048 grass images, 512 MNIST digits) our method still outperforms the baselines run on the full dataset. We note that there exists a few strategies for improving diffusion model performance in the low-data regime [45, 70-73]. +- Sampling steps: Increasing the number of predictor steps used to sample from the posterior generally improves the performance of our method. However, we find that the method performs well even with the number of sampling steps reduced to 64. + +The overall robustness of our method to most ablation highlights that its effectiveness is driven by the ability to directly sample from the posterior given our current diffusion model prior. Because the likelihood is often constraining, this enables us to return high-quality posterior samples even when the prior specified by our diffusion model is not an optimal fit to the empirical distribution. As evidence for this point, we note the large gap in performance between posterior and prior samples on the Grassy MNIST experiments (see Table 1). \ No newline at end of file diff --git a/NeurIPS/2025/A Data-Driven Prism_ Multi-View Source Separation with Diffusion Model Priors/images.zip b/NeurIPS/2025/A Data-Driven Prism_ Multi-View Source Separation with Diffusion Model Priors/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..ddf6883780869eb33f6fb1b7965866ab66890f92 --- /dev/null +++ b/NeurIPS/2025/A Data-Driven Prism_ Multi-View Source Separation with Diffusion Model Priors/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3f4e17140cb475cef1a2fc1efebadf6d2058fbc9add6f8b9e7fbb02f6b853581 +size 981515 diff --git a/NeurIPS/2025/A Data-Driven Prism_ Multi-View Source Separation with Diffusion Model Priors/layout.json b/NeurIPS/2025/A Data-Driven Prism_ Multi-View Source Separation with Diffusion Model Priors/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..21082a25bdb588d21b798faeff96b532f3bdf890 --- /dev/null +++ b/NeurIPS/2025/A Data-Driven Prism_ Multi-View Source Separation with Diffusion Model Priors/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ad9a7a606a5251fbc3b33000cc0e20f8e0086ac2eb27f1e1e210e5848fbf5d83 +size 949928 diff --git a/NeurIPS/2025/A Difference-of-Convex Functions Approach to Energy-Based Iterative Reasoning/c7d03008-feb0-4cf8-b65a-6159d6e26e18_content_list.json b/NeurIPS/2025/A Difference-of-Convex Functions Approach to Energy-Based Iterative Reasoning/c7d03008-feb0-4cf8-b65a-6159d6e26e18_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..0448fc1802d1112deea0f384fa7068e7aa7454b7 --- /dev/null +++ b/NeurIPS/2025/A Difference-of-Convex Functions Approach to Energy-Based Iterative Reasoning/c7d03008-feb0-4cf8-b65a-6159d6e26e18_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:363991e89da447ec0d3361ca0f0746a9bd066a73c61b549c182c474092a7d5f3 +size 173932 diff --git a/NeurIPS/2025/A Difference-of-Convex Functions Approach to Energy-Based Iterative Reasoning/c7d03008-feb0-4cf8-b65a-6159d6e26e18_model.json b/NeurIPS/2025/A Difference-of-Convex Functions Approach to Energy-Based Iterative Reasoning/c7d03008-feb0-4cf8-b65a-6159d6e26e18_model.json new file mode 100644 index 0000000000000000000000000000000000000000..809fd387a98d4ecbcb92cbe166975bf966fcf947 --- /dev/null +++ b/NeurIPS/2025/A Difference-of-Convex Functions Approach to Energy-Based Iterative Reasoning/c7d03008-feb0-4cf8-b65a-6159d6e26e18_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:eb9ea1b33e398f8b5bc028f7e44348de7ba9176069697d6b9781185d7f05ba8b +size 222405 diff --git a/NeurIPS/2025/A Difference-of-Convex Functions Approach to Energy-Based Iterative Reasoning/c7d03008-feb0-4cf8-b65a-6159d6e26e18_origin.pdf b/NeurIPS/2025/A Difference-of-Convex Functions Approach to Energy-Based Iterative Reasoning/c7d03008-feb0-4cf8-b65a-6159d6e26e18_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..49cb995b6d0a9004e931080f3660999bd0e825c3 --- /dev/null +++ b/NeurIPS/2025/A Difference-of-Convex Functions Approach to Energy-Based Iterative Reasoning/c7d03008-feb0-4cf8-b65a-6159d6e26e18_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1677dd8ac8cf3d0803a290187cdc02cd4ac5a8fdf9be2f02b0e4c8624626840c +size 1002919 diff --git a/NeurIPS/2025/A Difference-of-Convex Functions Approach to Energy-Based Iterative Reasoning/full.md b/NeurIPS/2025/A Difference-of-Convex Functions Approach to Energy-Based Iterative Reasoning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..63f100e12bbf1daa20be582cc90cc3b5d0c6332b --- /dev/null +++ b/NeurIPS/2025/A Difference-of-Convex Functions Approach to Energy-Based Iterative Reasoning/full.md @@ -0,0 +1,910 @@ +# A Difference-of-Convex Functions Approach to Energy-Based Iterative Reasoning + +Daniel Tschernnutter + +Infermedica + +Graz, Austria + +daniel.tschernutter@infermedica.com + +David Diego-Castro + +Infermedica + +Gothenburg, Sweden + +david.diego-castro@infermedica.com + +Maciej Kasinski + +Infermedica + +Wroclaw, Poland + +maciej.kasinski@infermedica.com + +# Abstract + +While energy-based models have recently proven to be a powerful framework for learning to reason with neural networks, their practical application is still limited by computational cost. That is, existing methods for energy-based iterative reasoning suffer from computational bottlenecks by relying on expensive optimization routines during training and especially during inference. Furthermore, these routines may not always converge to minimal energy states, potentially leading to suboptimal reasoning. To address these limitations, we propose a novel and efficient algorithm for energy-based iterative reasoning based on a difference-of-convex (DC) functions approach. Our algorithm achieves a significant speedup compared to prior methods while offering theoretical convergence guarantees ensuring locally minimal energy states. In addition, we achieve state-of-the-art or superior performance on continuous reasoning tasks, as demonstrated by our experiments on multiple benchmark datasets from continuous algorithmic reasoning. As such, our method offers a leap in computational efficiency, enabling faster inference with theoretical guarantees, and hence unlocking the potential of energy-based models for iterative reasoning applications. + +# 1 Introduction + +The human thinking process is described as operating through two distinct modes [31]: the rapid, automatic associations of System 1, and the slower, more controlled symbolic reasoning of System 2. Neural networks have demonstrated remarkable ability to perform System-1-like tasks within well-defined and specific environments. However, when faced with slightly different or harder tasks, neural networks often fail while humans engage in System 2 processes. The latter allows for iterative reasoning about new observations drawing upon prior experience and shared abstractions which remains difficult even for extremely large neural network architectures such as LLMs [35, 51]. + +There is a variety of recent work that tries to formalize reasoning within a neural network approach, see next section. In this work, we build upon the state-of-the-art in [20, 21] formalizing iterative reasoning as an energy minimization problem, i.e., + +$$ +\operatorname {a r g m i n} E _ {\theta} (x, y) \tag {1} +$$ + +for a given problem encoded in $y \in \mathbb{R}^m$ with (partial) solutions $x \in \mathbb{R}^n$ . Learning to reason is defined as learning the energy landscape $E_{\theta}$ parameterized by $\theta$ via + +$$ +\min _ {\theta} \sum_ {i} \left\| \operatorname {a r g m i n} _ {x} E _ {\theta} (x, y _ {i}) - x _ {i} \right\| ^ {2} \tag {EMP} +$$ + +from given problem- and solution-pairs $\{(y_i, x_i) \in \mathbb{R}^m \times \mathbb{R}^n : i \in \{1, \dots, N\}\}$ . Optimization steps from a current (partial) solution $x^k$ to a new $x^{k+1}$ with $E_\theta(x^{k+1}, y) \leq E_\theta(x^k, y)$ are then considered as individual reasoning steps. It has been proven empirically and theoretically that this formulation is superior to direct feed-forward computations, recurrent approaches, and various baselines from neural reasoning in terms of generalization and parameter efficiency [20, 21]. Nevertheless, learning energy landscapes involves solving (1) at training as well as inference time which imposes several limitations on current approaches: (i) Due to the inherent complexity of energy landscapes, heuristics for approximating solutions are used instead of directly solving (1), see next section, which can result in unstable training, (ii) Relying on gradient descent at prediction time is computationally expensive and might hinder practical applications, and (iii) Theoretically, energy-based reasoning yields a natural termination criterion during inference, i.e., an indication to terminate the computation of reasoning steps, by determining if a locally minimal energy state has been found. However, there are no theoretical guarantees that such an energy state is ever reached in previous methods as they rely on gradient descent. + +As a remedy, we present a general energy learning framework for continuous iterative reasoning based on difference-of-convex function (DC) optimization. Our main contributions are + +1. We introduce a tailored form of energy functions and present a difference-of-convex-function algorithm (DCA), see [4], for (1) powering our novel energy learning algorithm. +2. We derive theoretical convergence guarantees of our DCA routine to local solutions of (1). +3. We show that our DCA routine converges in finitely many steps and, hence, offers a clear termination criterion. +4. Under additional assumptions, we show how our energy learning algorithm can be scaled for batch optimization and present theoretical approximation guarantees for our form of energy function. + +# 2 Related Work + +Neural Reasoning. There is an active area of research that tries to formalize reasoning with neural networks. One group of work builds upon the idea to formulate reasoning as optimization problems and then derives differentiable solvers, e.g., [2, 17, 34, 48, 50]. These approaches are however constrained to tasks of a particular kind, e.g., tasks that can be formulated as quadratic programs [2]. Another group of research formalizes reasoning as iterative computations using neural networks. Following [21], this research can again be broadly subdivided into two areas: one that leverages explicit program representations [26, 36, 38, 10, 13, 37, 55] and another that uses recurrent neural networks [32, 25, 9, 15, 56, 16, 18, 40, 57]. In both areas a challenging problem is to decide when to terminate the computation and has been tackled in various ways usually by learning some sort of halting probabilities [12, 7, 10]. In contrast, our approach naturally imposes a termination criterion by stopping once a local minimum in the energy function has been reached. + +Energy-Based Models. Energy-based models formulate prediction tasks via energy minimization [33]. That is, external observations $y$ and possible predictions $x$ are both processed by a so-called energy function $E: \mathbb{R}^n \times \mathbb{R}^m \to \mathbb{R}$ which measures how compatible $x$ is with $y$ . The convention is that lower energies indicate a higher compatibility. A prediction for $y$ is then defined as a minimum energy state $x^*$ given $y$ , see (1). Energy-based models have been used in various ways to learn probabilistic models from data [54, 53, 22, 24, 19, 5, 52]. Our work leverages such energy functions to formalize an iterative reasoning process similar to [20, 21]. + +Energy Based Iterative Reasoning. We are not the first to formulate iterative reasoning as an energy minimization problem. To the best of our knowledge, [20] is the first to formalize reasoning through optimization steps using a general trainable energy function. In particular, the authors make use of a fixed number of gradient steps $T$ with a fixed step size $\lambda$ during training to approximate (1) + +within (EMP), i.e., $x_{i}^{T} = x_{i}^{T - 1} - \lambda \nabla_{x}E_{\theta}(x_{i}^{T - 1},y_{i})$ . However, this can lead to unstable training processes as due to potentially complex optimization landscapes it is not guaranteed that $x_{i}^{T}$ is a good approximation of a local minimizer of $E_{\theta}(x,y_i)$ . As a remedy, [21] introduced an energy diffusion process in which the authors minimize a sequence of energy landscapes by gradually increasing their complexity and using solutions on previous levels to initialize the gradient descent routine on consecutive levels. The energy landscape is then tuned via a supervised approach on noise corrupted gradients and a contrastive loss component to enforce local minima in the learned energy landscape instead. Nevertheless, both approaches suffer from computational bottlenecks at inference time due to the inherent optimization procedure based on gradient descent and the need for auto-differentiation at test time. + +# 3 DC Framework for Energy-Based Reasoning + +In this section, we introduce our novel energy-based reasoning framework that builds upon a difference-of-convex functions approach. + +# 3.1 DC Energy Landscapes + +To allow for sufficient reasoning capabilities an energy function $E_{\theta}$ should be able to represent a wide range of functional dependencies while at the same time entail structural properties that allow for an efficient solution of $\operatorname{argmin}_x E_{\theta}(x, y)$ at training time. Based on this observation, we are making use of the following energy function $E_{\theta}(x, y) = \sum_{i=1}^{N_x} \alpha_i \sigma(\langle w_i, x \rangle + b_i)$ , with $\alpha = \alpha_{\theta}(y)$ , $W = W_{\theta}(y)$ , and $b = b_{\theta}(y)$ , where $N_x \in \mathbb{N}$ and $\sigma = \max(\cdot, 0)$ is the ReLU activation function. Note that $E_{\theta}$ is a single hidden layer neural network in $x$ , while its parameters $(\alpha, W, b)$ are again parameterized functions in $y$ with weights $\theta$ . In particular, we set $(\alpha_{\theta}, W_{\theta}, b_{\theta})$ again as single hidden layer neural networks in $y$ with $N_y \in \mathbb{N}$ hidden neurons. Thus, we are using neural networks in $y$ to represent parameters for a neural network in $x$ . Due to single hidden layer neural networks being universal approximators [27], we can ensure sufficient representation capabilities for this form of energy function. + +As the goal is to learn energy landscapes in a way that minimal-energy states represent solutions of particular reasoning problems, we also want to ensure that such a minimum always exists. Hence, our final definition of $E_{\theta}$ is as follows + +$$ +E _ {\theta} (x, y) = \frac {\rho}{2} \| x \| ^ {2} - \langle \xi , x \rangle - \omega + \sum_ {i = 1} ^ {N _ {x}} \alpha_ {i} \sigma (\langle w _ {i}, x \rangle + b _ {i}), \tag {2} +$$ + +with $\alpha = \alpha_{\theta}(y)$ , $W = W_{\theta}(y)$ , $b = b_{\theta}(y)$ , $\xi = \xi_{\theta}(y)$ , $\rho = \rho_{\theta}(y) > 0$ , and $\omega = \omega_{\theta}(y)$ . Note that we added a general quadratic form, so that for fixed $\theta$ and $y$ we have $E_{\theta}(x,y) \to \infty$ for $\| x \| \to \infty$ . Hence, $E_{\theta}$ is coercive and continuous in $x$ and thus $\operatorname{argmin}_x E_{\theta}(x,y) \neq \emptyset$ . + +Our next result shows that (2) can be decomposed into a difference of convex functions in $x$ for fixed weights $(\rho, \xi, \omega, \alpha, W, b)$ , see Lemma 1. + +Lemma 1. For fixed weights $(\rho(y), \xi(y), \omega(y), \alpha(y), W(y), b(y))$ , the energy function $E_{\theta}$ is DC in $x$ , i.e., $E_{\theta}(x,y) = E_{\theta}(x) = g(x) - h(x)$ with + +$$ +g (x) = \frac {\rho}{2} \| x \| ^ {2} + \sum_ {\alpha_ {i} > 0} \alpha_ {i} \sigma \left(\langle w _ {i}, x \rangle + b _ {i}\right) \tag {3} +$$ + +$$ +h (x) = \sum_ {\alpha_ {i} < 0} | \alpha_ {i} | \sigma \left(\langle w _ {i}, x \rangle + b _ {i}\right) + \langle \xi , x \rangle + \omega , \tag {4} +$$ + +and $g, h$ convex in $x$ . For a proof see Appendix A.1 + +The above DC representation entails desirable properties for our analysis later on. We summarize important characteristics in the following Lemma. + +Lemma 2. Let $g$ and $h$ be defined as in Lemma 1, then the following holds + +1. $g$ is strongly convex in $x$ . + +2. $h$ is (up to a constant) polyhedral convex in $x$ , i.e., there exists $q_k \in \mathbb{R}^n$ and $p_k \in \mathbb{R}$ for $k \in \{1, \ldots, K\}$ such that + +$$ +h (x) = \max _ {k \in K} \langle q _ {k}, x \rangle + p _ {k}. \tag {5} +$$ + +For a proof see Appendix A.2 + +![](images/59c7376d13c5bda3521bc1c29b85f8c3a0dfd035ad84bf297be76b6894770db2.jpg) + +# 3.2 Locally Minimal Energy States + +This section shows how the above defined energy function can be minimized in $x$ via a tailored DCA. For an introduction to DCA, we refer to [4]. Following [45], the DCA routine for minimizing $E_{\theta}(x) = g(x) - h(x)$ starting in an arbitrary point $x_0 \in \mathbb{R}^n$ is + +$$ +v \in \partial h \left(x _ {k}\right) \tag {6} +$$ + +$$ +x _ {k + 1} \in \underset {x} {\operatorname {a r g m i n}} g (x) - \langle v, x \rangle \tag {7} +$$ + +Note that an element in the subgradient of $h$ in (6) is given by + +$$ +\sum_ {\alpha_ {i} < 0} \left| \alpha_ {i} \right| w _ {i} H \left(\left\langle w _ {i}, x _ {k} \right\rangle + b _ {i}\right) + \xi \in \partial h \left(x _ {k}\right). \tag {8} +$$ + +where $H(z)$ denotes the Heaviside function, i.e., $H(z) = 1$ for $z \geq 0$ and $H(z) = 0$ else. To compute $x_{k + 1}$ in the DCA, one has to solve the minimization problem in (7). The following lemma shows that the problem can be equivalently stated as a convex quadratic program. + +Lemma 3. The optimization problem + +$$ +\min _ {x \in \mathbb {R} ^ {n}} g (x) - \langle v, x \rangle \tag {9} +$$ + +can be formulated as + +$$ +\min _ {x, z} \frac {1}{2} \left( \begin{array}{l} x \\ z \end{array} \right) ^ {T} \left( \begin{array}{c c} \rho I & 0 \\ 0 & 0 \end{array} \right) \left( \begin{array}{l} x \\ z \end{array} \right) + \left\langle \left( \begin{array}{l} x \\ z \end{array} \right), \left( \begin{array}{c} - v \\ \alpha^ {+} \end{array} \right) \right\rangle_ {s t.} \left( \begin{array}{c c} - W ^ {+} & I \\ 0 & I \end{array} \right) \left( \begin{array}{l} x \\ z \end{array} \right) \geq \binom {b ^ {+}} {0}, \tag {QP} +$$ + +where $\alpha^{+}$ is the vector of strictly positive weights in the output layer, i.e., $\alpha_{i}^{+} = \alpha_{i_{j}}$ for all $i_{j} \in \{1, \dots, N_{x}\}$ with $\alpha_{i_{j}} > 0$ , $W^{+}$ is the matrix with rows formed by the corresponding weight vectors $w_{i_{j}}$ , and $b^{+}$ the vector formed by the corresponding bias terms $b_{i_{j}}$ . For a proof see Appendix A.3. + +One step in DCA thus simplifies to evaluating Equation (8) and then solving (QP). The next theorem shows that this simple iteration entails favorable convergence properties in the view of supervised learning of minimal energy states. + +Theorem 1. Given an arbitrary starting point $x_0$ , the DCA routine ((6) and (7)) with $v$ given by Equation (8) and $x_{k + 1}$ given as the solution of (QP) converges in finitely many iterations to a DC critical point $x^{*} \in \mathbb{R}^{n}$ , i.e., + +$$ +\partial g \left(x ^ {*}\right) \cap \partial h \left(x ^ {*}\right) \neq \emptyset . \tag {10} +$$ + +Furthermore, $x^{*}$ is a local minimum of $E_{\theta}(\cdot ,y)$ if and only if + +$$ +\partial h \left(x ^ {*}\right) \subseteq \partial g \left(x ^ {*}\right). \tag {11} +$$ + +For a proof see Appendix A.4 + +![](images/70be165ffee75133485560256df0e9e94e54a73f172fe1252c240a58aaecffdd.jpg) + +Note that (11) is always fulfilled if $\partial h(x^{*})$ or $\partial g(x^{*})$ is a singelton. The latter holds true in particular if $\langle w_i, x^* \rangle + b_i \neq 0 \quad \forall i \in \{i : \alpha_i > 0\}$ or $\langle w_i, x^* \rangle + b_i \neq 0 \quad \forall i \in \{i : \alpha_i < 0\}$ . If (11) does not hold we can restart the DCA routine with a point $x_0^*$ that yields a strict energy reduction in the first DCA step following [44], see Appendix A.8 for an in-depth discussion on our restart procedure. + +# 4 Scalable Energy Learning + +In theory, the DC framework presented in Section 3 can now be used to learn energy landscapes by supervising the resulting minimal energy states through regression using (EMP) similar to [20]. In particular, one can make use of differentiable convex optimization layers [1] or specialized batched quadratic programming solvers [2] to solve (QP) and run our DCA routine in batches which will converge in finitely many steps due to Theorem 1. Nevertheless, our research shows that (i) relying on differentiable QP solvers, and (ii) the need for our restart routine to ensure local optimality, hinders the ability of our approach to scale to large-scale settings. As a remedy, we introduce additional assumptions on the energy function defined in Equation (2) and show that under these assumptions we can find analytical solutions to (QP) and can guarantee that (11) always holds true for $x^{*}$ , i.e., DCA always converges to a local minimum of the energy function. In addition, in Section 4.1, we show that our energy function approximates a sub class of continuous functions that we call convexly-regular arbitrarily well. We also show that for the univariate prediction case the approximation is universal. + +For the remaining part of this work, we make the following additional assumption summarized in Assumption 1. + +Assumption 1. Let $\alpha_{i}\leq 0$ for all $i\in \{1,\dots ,N_x\}$ in Equation (2). + +Note, that this can be easily accomplished by using a non-negative activation function in the neural network component $\alpha_{\theta}(\cdot)$ and use the resulting values directly in Equation (4). Under this assumption, the following lemma can be derived. + +Lemma 4. Let Assumption 1 hold true. Then, (QP) can be solved analytically and $x_{k + 1} = \frac{1}{\rho} v$ . Furthermore, $\partial h(x^{*}) \subseteq \partial g(x^{*})$ always holds true in this case. For a proof see Appendix A.5 + +In the next section we analyze how Assumption 1 affects the approximation capabilities of our energy function, as indeed restricting shallow neural networks to only positive weights can make them loose their universal approximation guarantees [49]. + +# 4.1 Approximation Guarantees + +In our energy function (2) the weights to parameterize the function in $x$ , i.e., $(\rho_{\theta}(y), \xi_{\theta}(y), \omega_{\theta}(y), \alpha_{\theta}(y), W_{\theta}(y), b_{\theta}(y))$ , are all single hidden layer neural networks in $y$ and hence universal approximators [27], see also Appendix A.6. We thus focus the following analysis on approximations of functions in $x$ and need the following definitions and results from earlier work. + +Definition 1. Let $\mathbb{X} \subseteq \mathbb{R}^n$ be convex. A function $f: \mathbb{X} \to \mathbb{R}$ is called $\rho$ -weakly-convex, or simply weakly-convex, if there exists a $\rho > 0$ and a convex function $h$ such that $f + \rho / 2 \| \cdot \|^2 = h$ . The function is called weakly-concave if $-f$ is weakly-convex. The set of all weakly-convex functions in $\mathbb{X}$ is denoted by $\mathcal{W}\mathcal{C}(\mathbb{X})$ . + +It can be shown that weakly-convex functions are universal approximators. To formalize this claim recall that $\mathcal{C}_0(\mathbb{X})$ denotes the set of continuous functions that vanish at infinity, i.e., for all $\epsilon > 0$ there exists a compact set $K \subseteq \mathbb{X}$ such that $|f(x)| < \epsilon$ if $x \notin K$ . Then, the following holds. + +Lemma 5 (Theorem 6 in [43]). Let $\mathbb{X} \subseteq \mathbb{R}^n$ be closed (or open) convex. Then, $\mathcal{WC}(\mathbb{X}) \cap \mathcal{C}_0(\mathbb{X})$ is dense in $\mathcal{C}_0(\mathbb{X})$ equipped with the infinity norm. This statement also holds for weakly-concave functions (by switching signs). + +Furthermore, weakly-convex functions can be represented in a special form, see Lemma 6. + +Lemma 6 (Theorem 3 in [43]). A (closed) function $f$ is $\rho$ -weakly-convex if and only if $f(x) = \sup_{t \in T} \langle q_t, x \rangle + p_t - \rho / 2 \| x \|^2$ for some (not necessarily finite) index set $T$ .3 + +Now, by combining Lemma 2 and Lemma 6 we see that our energy function is weakly-concave in $x$ , i.e., it has exactly the form + +$$ +\frac {\rho}{2} \| x \| ^ {2} - \max _ {k} \langle q _ {k}, x \rangle + p _ {k}. \tag {12} +$$ + +Note, however, that it is not immediately clear that also any weakly-concave function can be approximated by our energy function as the vectors $q_{k}$ and biases $p_k$ in Lemma 2 follow a special form. Nevertheless, if this is the case or equivalently if we are able to prove that $h$ can approximate continuous convex functions, Lemma 5 yields theoretical approximation guarantees for our energy function. + +To derive sufficient conditions for $h$ to be able to approximate a continuous convex function, we note that we merely impose sign constraints on the output layer of the involved shallow neural network. Thus, $h$ can be seen as a single layer input convex neural network (ICNN) with weighted input skip connections in $x$ [23]. Input skip connections were introduced for ICNNs in [3] to increase their expressivity by allowing identity mappings between layers as otherwise the non-negativity constraint would be too restrictive. Indeed, it has then been shown that ICNNs are able to approximate arbitrary continuous convex functions on compact convex domains (see, e.g., Theorem 1 in [14] or Proposition 3 in [28]). However, those approximation guarantees require deeper ICNNs while $h$ is merely a single layer ICNN. As pointed out by [23], the derivations in [14, 28] are merely for theoretical purposes as they require as many layers as affine pieces of the piecewise linear convex function they are trying to approximate and only a single neuron per layer. The authors, thus, derive the following result. + +Lemma 7 (Corollary 4.8 and Proposition 4.9 in [23]). A convex function implemented by a single hidden layer ReLU network (with or without weighted input skip connections) can also be implemented by a single hidden layer ICNN with the same width. + +The question if $h$ can approximate a continuous convex function thus boils down to the question if it can be approximated by a convex shallow ReLU network. Hence, we define the following + +Definition 2. Let $\mathbb{X} \subseteq \mathbb{R}^n$ be convex and compact. We call a continuous convex function $r: \mathbb{X} \to \mathbb{R}$ $\epsilon$ -convexly-ReLU-representable if there exists a convex single layer ReLU network (with or without weighted input skip connections) $\mathcal{N}\mathcal{N}$ such that $\| \mathcal{N}\mathcal{N} - r \|_{\infty} < \epsilon$ . + +Given a function $f \in \mathcal{C}_0(\mathbb{X})$ , we know from Lemma 5 that for all $\epsilon > 0$ there exists a $\rho$ -weakly-concave function $f_{\epsilon}$ with $\| f - f_{\epsilon}\|_{\infty} < \epsilon$ . Then we know that $r = -f_{\epsilon} + \rho /2\| x\| ^2$ is convex and we make the following definition. + +Definition 3. Let $\mathbb{X} \subseteq \mathbb{R}^n$ be convex and compact and let $f \in C_0(\mathbb{X})$ . We call $f$ convexly-regular, if we can always choose $f_{\epsilon}$ such that $r$ is $\epsilon$ -convexly-ReLU-representable. + +Now, Theorem 2 summarizes the above derivations + +Theorem 2. Let $\mathbb{X} \subseteq \mathbb{R}^n$ be convex and compact. Under Assumption 1 every convexly-regular function $f \in \mathcal{C}_0(\mathbb{X})$ can be approximated arbitrarily well by $E_{\theta}$ as a function in $x$ . + +Proof. Let $\epsilon > 0$ be arbitrary and $\hat{\epsilon} = \epsilon / 2$ . From Lemma 5 we know that there exists a weakly-concave function $f_{\hat{\epsilon}}$ such that $\| f - f_{\hat{\epsilon}} \|_{\infty} < \hat{\epsilon}$ . Hence, there exists a $\rho > 0$ and a convex function $r$ such that $-f_{\hat{\epsilon}}(x) + \rho / 2 \| x \|^2 = r(x)$ . Furthermore, Definition 3 ensures that we can always choose $f_{\hat{\epsilon}}$ such that $r$ is $\hat{\epsilon}$ -convexly-ReLU-representable. Thus, there exists a convex ReLU neural network $\mathcal{NN}$ with $\| r - \mathcal{N}\mathcal{N} \|_{\infty} < \hat{\epsilon}$ . From Lemma 7, we know that there exist $\alpha \leq 0$ , $W$ , $b$ , $\xi$ , and $\omega$ such that $\mathcal{NN}(x) = \sum_i |\alpha_i| \sigma (\langle w_i, x \rangle + b_i) + \langle \xi, x \rangle + \omega$ which we define as $h$ . Hence, for $E_{\theta}(x) = \frac{\rho}{2} \| x \|^2 - h(x)$ we have + +$$ +\begin{array}{l} \left\| f - E _ {\theta} (x) \right\| = \left\| f - \left(\frac {\rho}{2} \| x \| ^ {2} - h (x)\right) \right\| _ {\infty} \leq \| f - f _ {\hat {\epsilon}} \| _ {\infty} + \left\| f _ {\hat {\epsilon}} - \left(\frac {\rho}{2} \| x \| ^ {2} - h (x)\right) \right\| _ {\infty} (13) \\ < \hat {\epsilon} + \| h - r \| _ {\infty} < 2 \hat {\epsilon} = \epsilon . (14) \\ \end{array} +$$ + +![](images/46e0c5261b8c5a572fd7d30c6d988b818f683f47104ef84f8648509ed40df30a.jpg) + +Definition 3 is rather technical and to the best of our knowledge there are no general results on conditions under which a convex function can be approximated by a single layer ReLU network which is itself convex. However, the following theorem shows that the class of convexly-regular functions is sufficiently large in the univariate case. + +Theorem 3. Let $\mathbb{X} \subseteq \mathbb{R}$ be compact and convex. Then, every $f \in \mathcal{C}_0(\mathbb{X})$ is convexly-regular. For a proof see Appendix A.7 + +# 4.2 Pseudocode + +We now combine our derivations in Section 3 and Assumption 1 to define our algorithm for scalable energy learning via a batched DCA approach named DCAReasoner. A pseudocode is presented in Algorithm 1. + +Algorithm 1: DCAReasoner: Scalable Energy Learning via Batched DCA +Data: $(y_{i},x_{i})\in \mathbb{R}^{m}\times \mathbb{R}^{n}$ , lower and upper bounds for starting points $l,u\in \mathbb{R}^n$ Result: $E_{\theta}(\cdot ,\cdot)$ +1 while not converged do +2 Sample batch of data $(y_j,x_j)_{j\in B}$ . Perform forward pass for parameters $(\alpha ,W,b,\rho ,\xi)\gets (\alpha ,W,b,\rho ,\xi)((y_j)_{j\in B})$ +3 Sample uniformly random starting points $(x_j^0)_{j\in B}\sim \mathcal{U}(l,u)$ +5 $k\gets 0$ . // Initializing $x_{j}^{k + 1}$ in a meaningful way +6 while max $\begin{array}{r}\| x_j^k -x_j^{k + 1}\| >tol\end{array}$ do +7 $x^{k + 1}\gets \frac{1}{\rho}\left(\sum_i|\alpha_i|w_iH(\langle w_i,x^k\rangle +b_i) + \xi\right);$ +8 $k\gets k + 1$ . +9 end +10 Update $\theta$ using Adam and $\nabla_{\theta}\sum_{j\in B}\| x_j^k -x_j\|^2 /|B|$ +11 end + +Note that we used $\max_{j\in B}\| x_j^k -x_j^{k + 1}\| >$ tol as a stopping criterion for the DCA routine. Empirically, we observe that a few DCA iterations $(< 10)$ are enough for the whole batch to converge for $\mathrm{tol} = 10^{-5}$ . Indeed, most of the time we observe $\max_{j\in B}\| x_j^k -x_j^{k + 1}\| \ll$ tol approaching machine precision, and hence, the finite convergence property can also be observed empirically. See Figure 2 in Appendix B.1 for an illustration of how the norm differences decrease to zero with DCA iterations for a batch size of 512, i.e., when solving 512 reasoning tasks in parallel. We also note that we skipped the neural network for the bias term, i.e., $\omega (y)$ , in Algorithm 1 as it does not change the local minimizer of (1) and was merely used for our theoretical derivations. Furthermore, the network for $b$ will not be updated during training as a consequence of the Heaviside function. Nevertheless, similar partial optimization routines, in which parts of the parameters are randomly initialized and then frozen, have been successfully applied in neural learning, see e.g. [29]. + +# 5 Numerical Experiments + +# 5.1 Experimental Setup + +We first evaluate our algorithm on five continuous algorithmic reasoning benchmark datasets from earlier research [20, 21] in Section 5.2. All tasks are aiming to capture different aspects of reasoning. We report the mean squared errors, as well as, inference times of our DCAReasoner and two state-of-the-art baselines from energy-based iterative reasoning. The evaluation is performed on 10000 test problems and repeated five times to report the mean and standard error of our metrics. Having established that our algorithm is superior (or on par) with state-of-the-art but significantly faster, we then demonstrate how our DCAReasoner might unlock reasoning capabilities in language models by learning energy landscapes in token embedding spaces in Section 5.3. + +# 5.2 Continuous Algorithmic Reasoning Benchmarks + +Baselines. We consider two baselines from state-of-the-art energy based iterative reasoning: (i) Energy based reasoning through energy minimization (IREM): This baseline learns an energy function minimizing (EMP) by approximating $\arg \min_x E_\theta(x, y)$ via a fixed number of gradient steps [20]. During inference it uses again a subgradient descent method but leveraging a greater number of steps than during training.(ii) Energy based reasoning through energy diffusion (IRED): Here, the idea is to minimize a sequence of energy landscapes gradually increasing their complexity and using solutions on previous levels to initialize the gradient descent routine on consecutive levels [21]. The energy + +
DatasetDCAResoner (Ours)IREDIREM
MSEInference-Time [s]MSEInference-Time [s]MSEInference-Time [s]
Same Difficulty
Matrix Inverse0.0096±0.00002.6189±0.03050.0097±0.000033.7056±0.88180.0101±0.000022.6199±0.6127
Matrix Completion0.0177±0.00001.3373±0.01250.0179±0.000033.6597±0.91250.0180±0.000022.7441±0.5296
Parity0.0301±0.00030.6053±0.03810.4859±0.00269.1011±0.28090.2504±0.00011.9797±0.1463
QR Decomposition0.1438±0.00012.3051±0.02610.2175±0.000148.1915±1.41990.1521±0.000136.2775±1.0231
Matrix Multiplication0.0480±0.00001.6790±0.02580.0919±0.000034.1602±0.86090.0903±0.000023.6510±0.6229
Harder Difficulty
Matrix Inverse0.2077±0.00032.6017±0.02260.2064±0.000333.3662±0.64420.2063±0.000622.7757±0.4990
Matrix Completion0.2100±0.00011.3491±0.02110.2094±0.000233.2705±0.72300.2058±0.000222.9266±0.5348
Parity0.0301±0.00030.5863±0.00670.4885±0.00108.6239±0.08790.2504±0.00011.9477±0.0983
QR Decomposition0.8847±0.00032.3374±0.12981.0376±0.000247.7211±1.03621.3267±0.000635.3764±0.8735
Matrix Multiplication0.2974±0.00031.6877±0.02430.4506±0.000233.3451±0.64790.4524±0.000223.5748±0.6184
+ +Table 1: Evaluation on continuous algorithmic reasoning tasks. Models are evaluated on test problems drawn from the training distribution (same difficulty) and a harder test distribution (harder difficulty). We report the mean squared error and the inference time. We perform five evaluation runs and report the mean and standard errors. + +landscape is then tuned by supervising on noise corrupted gradients and a contrastive loss component to enforce local minima in the energy landscape. We scale the network size in our baselines to ensure that each of them has roughly the same number of parameters. For a detailed discussion of our experimental setup see Appendix B.2. + +Datasets. In our experiments we evaluate all baselines based on five datasets from earlier research [20, 21]. Each of them is evaluated once with the same level of difficulty, i.e., test cases are drawn from the training distribution, and once with a harder level of difficulty, in which test cases are drawn from a problem specific harder test distribution following [20, 21]. The latter should test the algorithms ability to generalize reasoning capabilities to new unseen problem settings. In particular, we use: (i) Matrix Inverse: The task is to invert a random $20 \times 20$ matrix. It aims at testing numerical reasoning. Harder problems are created by creating less well-conditioned matrices to invert. (ii) Matrix Completion: The task is to recover masked out values in a random low-ranked $20 \times 20$ matrix constructed from two low-rank matrices $U$ and $V$ . Harder tasks are created by increasing the complexity of $U$ and $V$ . It aims at both structural and analogical reasoning. (iii) Parity: Given a random vector in $[0,1]^{20}$ the task is to decide whether or not the number of entries which are greater than 0.5 is odd or even, i.e., the target is 0 for even 1 for odd, see also [25]. Harder tasks are created by increasing the magnitude of vector entries. (iv) QR Decomposition: The task is to compute the QR decomposition of a random $20 \times 20$ matrix with entries in $[-1,1]$ . Harder problems are created by changing the magnitude of matrix entries. (v) Matrix Multiplication: Given a random $20 \times 20$ matrix $M$ the task is to compute the square, i.e., $M^2$ . Harder problems are created by changing the magnitude of matrix entries. For an in-depth discussion of our benchmark datasets see Appendix B.3. + +Results. Our results are summarized in Table 1. In terms of mean squared error, DCAReasoner is mostly on par with IRED and IREM on the matrix inverse and the matrix completion dataset, while we see larger improvements for the remaining datasets. Noteworthy, we observe a decrease in MSE by a factor of $\sim 10$ on the parity dataset, which might stem from the fact that the prediction is univariate in this case, for which we have established universal approximation guarantees in Theorems 2 and 3. In terms of inference time, we see large improvements of factors between 14 and 27 for IRED and between 3 and 18 for IREM. Furthermore, we performed additional experiments showing that our predictions are robust to noisy input data on the example of the QR Decomposition dataset. All details are reported in Appendix B.4. + +# 5.3 Energy-Based Reasoning in Token Embedding Space + +In the last section we have empirically proven that our algorithm yields state-of-the-art performance results but is significantly faster at inference time than previous energy-based models for iterative reasoning. Furthermore, it offers theoretical convergence guarantees and performs well in high-dimensional settings. As such, we think that our algorithm might be used to improve reasoning skills of language models by learning energy landscapes in token embedding spaces which might be used during inference for energy-guided text generation. Note that our baselines are not well-suited for such + +
Alg.MSEAccuracy[%]Inference Time[%]
IREM0.01296%336%
IRED0.02796%1330%
DCAReasoner0.00896%100%
+ +Table 2: Test evaluation performance on text classification task. We report the mean squared error of the prediction and the embedding of the target text. Accuracy is computed by using the target closest to the prediction. Inference time is reported in percent, with $100\%$ indicating the lowest time. + +a setting as large inference times especially in high-dimensional token embedding spaces might considerably slow down token generation in practice. + +![](images/ae3ef0712b93f929151420e82c7d11ab8f5a6e734cf4a11fc63eb73184b81cb6.jpg) +Figure 1: Visualization of Energy Landscape in Token Embedding Space for the Input Sentence "My muscles are weak, my neck is stiff, and my joints are swollen. I can't move around very well, and walking is really painful." + +As a fully developed approach for energy-guided text generation to improve reasoning is out of the scope of this work, we demonstrate how DCAReasoner is able to learn reasonable energy-landscapes in token embedding spaces in a simpler setting. In particular, we make use of the symptom-to-diagnosis dataset for medical reasoning which is freely available on Hugging Face. It provides a training and test dataset consisting of short texts in which a patient describes her symptoms and a corresponding diagnosis out of a set of 22 medical diagnoses. We then use the CLS token embeddings of a finetuned uncased DistilBERT model of those texts as inputs $y$ and the embeddings of the corresponding diagnosis as a ground truth $x$ to train the DCAReasoner and our baselines in a continuous reasoning setting. More details of our experiments are reported in Appendix B.5. + +We summarize the results in Table 2. Our algorithm yields the lowest mean squared error more than three times faster than IREM and thirteen times faster than IRED. We also visualize the energy landscape in token embedding space $^{6}$ learned by our algorithm in Figure 1 using as an example the sentence "My muscles are weak, my neck is stiff, and my joints are swollen. I can't move around very well, and walking is really painful." with diagnosis arthritis from + +the test set. Note that arthritis has indeed the lowest energy, while embedding vectors of diseases like malaria which shares symptoms like weakness and joint pain are also assigned lower energy values. Furthermore, seemingly unrelated diagnoses like urinary tract infection yield higher energy values, indicating that our algorithm is capable of learning reasonable energy landscapes in the token embedding space. + +# 6 Conclusion + +We proposed a new algorithm named DCAReasoner for energy-based continuous iterative reasoning. It is built upon a tailored class of energy functions for which we derived theoretical approximation guarantees. In addition, we presented theoretical convergence guarantees for the inherent DCA routine. That is, we showed that it converges to local minima in finitely many steps independent of + +the starting point. Empirically, we have proven that it yields improved performance and inference times. + +Limitations. Our DCAReasoner shows promising results for neural iterative reasoning. However, there are still limitations. First, our DCAReasoner, as presented in this work, does not yet make use of any external memory. That is, DCAReasoner cannot store intermediate results during reasoning which might be beneficial in some reasoning tasks. Second, the form of our energy function might require a large number of trainable parameters if a large number of hidden neurons in $x$ , i.e., $N_{x}$ is desirable. For instance, the mapping $W_{\theta}(y)$ requires $m \cdot N_y + N_y \cdot N_x \cdot n$ trainable parameters. Empirically, we see however that smaller values for $N_{x}$ and larger values for $N_{y}$ are sufficient (our experiments use $N_{x} = 8$ and $N_{y} = 4000$ ). Third, in the current form DCAReasoner is designed for continuous reasoning tasks and hence cannot directly handle discrete reasoning tasks. However, we note that earlier work on energy-based reasoning shows how discrete tasks, formulated in a continuous setting, yield promising results [20, 21]. Furthermore, we conducted preliminary experiments on a discrete task, specifically solving Sudoku puzzles using the dataset provided in [41], and report our early results in Appendix C. + +Societal Impacts. To the best of our knowledge, there are no immediate positive or negative social impacts that can be derived from our work in the current form. + +# References + +[1] A. Agrawal, B. Amos, S. Barratt, S. Boyd, S. Diamond, and J. Z. Kolter. Differentiable convex optimization layers. Advances in Neural Information Processing Systems (NeurIPS), 32, 2019. +[2] B. Amos and J. Z. Kolter. Optnet: Differentiable optimization as a layer in neural networks. In International Conference on Machine Learning (ICML), pages 136-145. PMLR, 2017. +[3] B. Amos, L. Xu, and J. Z. Kolter. Input convex neural networks. In International Conference on Machine Learning (ICML), pages 146-155. PMLR, 2017. +[4] L. T. H. An and P. D. Tao. The DC (difference of convex functions) programming and DCA revisited with DC models of real world nonconvex optimization problems. Annals of Operations Research, 133:23-46, 2005. +[5] M. Arbel, L. Zhou, and A. Gretton. Generalized energy based models. In International Conference on Learning Representations (ICLR). ICLR, 2021. +[6] R. Arora, A. Basu, P. Mianjy, and A. Mukherjee. Understanding deep neural networks with rectified linear units. In International Conference on Learning Representations (ICLR), 2018. +[7] A. Banino, J. Balaguer, and C. Blundell. PonderNet: Learning to ponder. In 8th ICML Workshop on Automated Machine Learning (AutoML), 2021. +[8] Y. Beck and M. Schmidt. A gentle and incomplete introduction to bilevel optimization. Lecture Notes, 2021. +[9] T. Bolukbasi, J. Wang, O. Dekel, and V. Saligrama. Adaptive neural networks for efficient inference. In International Conference on Machine Learning (ICML), pages 527-536. PMLR, 2017. +[10] J. Cai, R. Shin, and D. Song. Making neural programming architectures generalize via recursion. International Conference on Learning Representations (ICLR), 2022. +[11] G. C. Calafiore, S. Gaubert, and C. Possieri. Log-sum-exp neural networks and posynomial models for convex and log-log-convex data. IEEE Transactions on Neural Networks and Learning Systems, 31(3):827-838, 2019. +[12] X. Chen, H. Dai, Y. Li, X. Gao, and L. Song. Learning to stop while learning to predict. In International Conference on Machine Learning (ICML), pages 1520-1530. PMLR, 2020. +[13] X. Chen, C. Liang, A. W. Yu, D. Song, and D. Zhou. Compositional generalization via neural-symbolic stack machines. Advances in Neural Information Processing Systems (NeurIPS), 33:1690-1701, 2020. + +[14] Y. Chen, Y. Shi, and B. Zhang. Optimal control via neural networks: A convex approach. In International Conference on Learning Representations (ICLR). ICLR, 2019. +[15] J. Chung, S. Ahn, and Y. Bengio. Hierarchical multiscale recurrent neural networks. International Conference on Learning Representations (ICLR), 2017. +[16] M. Dehghani, S. Gouws, O. Vinyals, J. Uszkoreit, and L. Kaiser. Universal transformers. International Conference on Learning Representations (ICLR), 2019. +[17] J. Djolonga and A. Krause. Differentiable learning of submodular models. Advances in Neural Information Processing Systems (NeurIPS), 30, 2017. +[18] H. Dong, J. Mao, T. Lin, C. Wang, L. Li, and D. Zhou. Neural logic machines. International Conference on Learning Representations (ICLR), 2019. +[19] Y. Du, S. Li, J. Tenenbaum, and I. Mordatch. Improved contrastive divergence training of energy-based models. In International Conference on Machine Learning (ICML), pages 2837-2848. PMLR, 2021. +[20] Y. Du, S. Li, J. Tenenbaum, and I. Mordatch. Learning iterative reasoning through energy minimization. International Conference on Machine Learning (ICML), pages 5570-5582, 2022. +[21] Y. Du, J. Mao, and J. Tenenbaum. Learning iterative reasoning through energy diffusion. International Conference on Machine Learning (ICML), pages 11764-11776, 2024. +[22] Y. Du and I. Mordatch. Implicit generation and modeling with energy based models. Advances in Neural Information Processing Systems (NeurIPS), 32, 2019. +[23] A. Gagneux, M. Massias, E. Soubies, and R. Gribonval. Convexity in relu neural networks: beyond ICNNs? arXiv preprint arXiv:2501.03017, 2025. +[24] W. Grathwohl, K.-C. Wang, J.-H. Jacobsen, D. Duvenaud, and R. Zemel. Learning the stein discrepancy for training and evaluating energy-based models without sampling. In International Conference on Machine Learning (ICML), pages 3732-3747. PMLR, 2020. +[25] A. Graves. Adaptive computation time for recurrent neural networks. arXiv preprint arXiv:1603.08983, 2016. +[26] A. Graves, G. Wayne, and I. Danihelka. Neural tuning machines. arXiv preprint arXiv:1410.5401, 2014. +[27] K. Hornik, M. Stinchcombe, and H. White. Multilayer feedforward networks are universal approximators. Neural Networks, 2(5):359-366, 1989. +[28] C.-W. Huang, R. T. Chen, C. Tsirigotis, and A. Courville. Convex potential flows: Universal probability distributions with optimal transport and convex optimization. In International Conference on Learning Representations (ICLR). ICLR, 2021. +[29] G.-B. Huang, Q.-Y. Zhu, and C.-K. Siew. Extreme learning machine: theory and applications. Neurocomputing, 70(1-3):489-501, 2006. +[30] Y. Ji, J. Wu, and Y. Xi. Rethinking neural-based matrix inversion: Why can't, and where can. arXiv preprint arXiv:2506.00642, 2025. +[31] D. Kahneman. Thinking, fast and slow. Macmillan, 2011. +[32] L. Kaiser and I. Sutskever. Neural GPUs learn algorithms. arXiv preprint arXiv:1511.08228, 2015. +[33] Y. LeCun, S. Chopra, R. Hadsell, M. Ranzato, F. Huang, et al. A tutorial on energy-based learning. Predicting structured data, 1(0), 2006. +[34] R. Manhaeve, S. Dumancic, A. Kimmig, T. Demeester, and L. De Raedt. Deep Probl: Neural probabilistic logic programming. Advances in Neural Information Processing Systems (NeurIPS), 31, 2018. + +[35] I. Mirzadeh, K. Alizadeh, H. Shahrokhi, O. Tuzel, S. Bengio, and M. Farajtabar. GSM-symbolic: Understanding the limitations of mathematical reasoning in large language models. arXiv preprint arXiv:2410.05229, 2024. +[36] A. Neelakantan, Q. V. Le, and I. Sutskever. Neural programmer: Inducing latent programs with gradient descent. International Conference on Learning Representations (ICLR), 2016. +[37] R. Y. Pang, W. Yuan, H. He, K. Cho, S. Sukhbaatar, and J. Weston. Iterative reasoning preference optimization. Advances in Neural Information Processing Systems (NeurIPS), 37:116617-116637, 2024. +[38] S. Reed and N. De Freitas. Neural programmer-interpreters. International Conference on Learning Representations (ICLR), 2016. +[39] R. T. Rockafellar. Convex Analysis, volume 28. Princeton University Press, 1997. +[40] A. Schwarzschild, E. Borgnia, A. Gupta, F. Huang, U. Vishkin, M. Goldblum, and T. Goldstein. Can you learn an algorithm? generalizing from easy to hard problems with recurrent networks. Advances in Neural Information Processing Systems (NeurIPS), 34:6695-6706, 2021. +[41] K. Shah, N. Dikkala, X. Wang, and R. Panigrahy. Causal language modeling can elicit search and reasoning capabilities on logic puzzles. Advances in Neural Information Processing Systems (NeurIPS), 37:56674-56702, 2024. +[42] I. Shoshani and O. Shamir. Hardness of learning fixed parities with neural networks. arXiv preprint arXiv:2501.00817, 2025. +[43] S. Sun and Y. Yu. Least squares estimation of weakly convex functions. In International Conference on Artificial Intelligence and Statistics (AISTATS), pages 2271-2280. PMLR, 2019. +[44] P. D. Tao and L. T. H. An. A DC optimization algorithm for solving the trust-region subproblem. SIAM Journal on Optimization, 8(2):476-505, 1998. +[45] L. Thi Hoai An and P. Dinh Tao. Solving a class of linearly constrained indefinite quadratic problems by DC algorithms. Journal of Global Optimization, 11:253-285, 1997. +[46] D. Tschernutter, M. Kraus, and S. Feuerriegel. A globally convergent algorithm for neural network parameter optimization based on difference-of-convex functions. Transactions on Machine Learning Research, 2024. +[47] P. Velicković and C. Blundell. Neural algorithmic reasoning. *Patterns*, 2(7), 2021. +[48] P.-W. Wang, P. Donti, B. Wilder, and Z. Kolter. SATnet: Bridging deep learning and logical reasoning using a differentiable satisfiability solver. In International Conference on Machine Learning (ICML), pages 6545-6554. PMLR, 2019. +[49] Q. Wang, M. A. Powell, A. Geisa, E. Bridgeford, C. E. Priebe, and J. T. Vogelstein. Why do networks have inhibitory/negative connections? Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22551–22559, 2023. +[50] B. Wilder, B. Dilkina, and M. Tambe. Melding the data-decisions pipeline: Decision-focused learning for combinatorial optimization. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 33, pages 1658-1665, 2019. +[51] Z. Wu, L. Qiu, A. Ross, E. Akyurek, B. Chen, B. Wang, N. Kim, J. Andreas, and Y. Kim. Reasoning or reciting? exploring the capabilities and limitations of language models through counterfactual tasks. In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), pages 1819–1862, 2024. +[52] Z. Xiao, K. Kreis, J. Kautz, and A. Vahdat. VAEBM: A symbiosis between variational autoencoders and energy-based models. International Conference on Learning Representations (ICLR), 2021. + +[53] J. Xie, Y. Lu, R. Gao, S.-C. Zhu, and Y. N. Wu. Cooperative training of descriptor and generator networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 42(1):27-45, 2018. +[54] J. Xie, Y. Lu, S.-C. Zhu, and Y. Wu. A theory of generative ConvNet. In International Conference on Machine Learning, pages 2635-2644. PMLR, 2016. +[55] Y. Xie, A. Goyal, W. Zheng, M.-Y. Kan, T. P. Lillicrap, K. Kawaguchi, and M. Shieh. Monte carlo tree search boosts reasoning via iterative preference learning. arXiv preprint arXiv:2405.00451, 2024. +[56] F. Yang, Z. Yang, and W. W. Cohen. Differentiable learning of logical rules for knowledge base reasoning. Advances in Neural Information Processing Systems (NeurIPS), 30, 2017. +[57] Z. Yang, A. Ishay, and J. Lee. Learning to solve constraint satisfaction problems with recurrent transformer. International Conference on Learning Representations (ICLR), 2023. + +# NeurIPS Paper Checklist + +# 1. Claims + +Question: Do the main claims made in the abstract and introduction accurately reflect the paper's contributions and scope? + +Answer: [Yes] + +Justification: All claims made in the abstract and introduction are either proven theoretically in Section 3 or Section 4, or are proven empirically in our numerical results section. + +Guidelines: + +- The answer NA means that the abstract and introduction do not include the claims made in the paper. +- The abstract and/or introduction should clearly state the claims made, including the contributions made in the paper and important assumptions and limitations. A No or NA answer to this question will not be perceived well by the reviewers. +- The claims made should match theoretical and experimental results, and reflect how much the results can be expected to generalize to other settings. +- It is fine to include aspirational goals as motivation as long as it is clear that these goals are not attained by the paper. + +# 2. Limitations + +Question: Does the paper discuss the limitations of the work performed by the authors? + +Answer: [Yes] + +Justification: We added a clearly marked limitations paragraph in our conclusion. + +Guidelines: + +- The answer NA means that the paper has no limitation while the answer No means that the paper has limitations, but those are not discussed in the paper. +- The authors are encouraged to create a separate "Limitations" section in their paper. +- The paper should point out any strong assumptions and how robust the results are to violations of these assumptions (e.g., independence assumptions, noiseless settings, model well-specification, asymptotic approximations only holding locally). The authors should reflect on how these assumptions might be violated in practice and what the implications would be. +- The authors should reflect on the scope of the claims made, e.g., if the approach was only tested on a few datasets or with a few runs. In general, empirical results often depend on implicit assumptions, which should be articulated. +- The authors should reflect on the factors that influence the performance of the approach. For example, a facial recognition algorithm may perform poorly when image resolution is low or images are taken in low lighting. Or a speech-to-text system might not be used reliably to provide closed captions for online lectures because it fails to handle technical jargon. +- The authors should discuss the computational efficiency of the proposed algorithms and how they scale with dataset size. +- If applicable, the authors should discuss possible limitations of their approach to address problems of privacy and fairness. +- While the authors might fear that complete honesty about limitations might be used by reviewers as grounds for rejection, a worse outcome might be that reviewers discover limitations that aren't acknowledged in the paper. The authors should use their best judgment and recognize that individual actions in favor of transparency play an important role in developing norms that preserve the integrity of the community. Reviewers will be specifically instructed to not penalize honesty concerning limitations. + +# 3. Theory assumptions and proofs + +Question: For each theoretical result, does the paper provide the full set of assumptions and a complete (and correct) proof? + +Answer: [Yes] + +Justification: Each of our own theoretical results has a link to the appendix in which the corresponding proofs can be found. + +# Guidelines: + +- The answer NA means that the paper does not include theoretical results. +- All the theorems, formulas, and proofs in the paper should be numbered and cross-referenced. +- All assumptions should be clearly stated or referenced in the statement of any theorems. +- The proofs can either appear in the main paper or the supplemental material, but if they appear in the supplemental material, the authors are encouraged to provide a short proof sketch to provide intuition. +- Inversely, any informal proof provided in the core of the paper should be complemented by formal proofs provided in appendix or supplemental material. +- Theorems and Lemmas that the proof relies upon should be properly referenced. + +# 4. Experimental result reproducibility + +Question: Does the paper fully disclose all the information needed to reproduce the main experimental results of the paper to the extent that it affects the main claims and/or conclusions of the paper (regardless of whether the code and data are provided or not)? + +# Answer: [Yes] + +Justification: An implementation of DCAReasoner is publicly available on GitHub. Appendix B reports all the necessary information including model specifications, training hyperparameters, and details on benchmark datasets, to reproduce our results. + +# Guidelines: + +- The answer NA means that the paper does not include experiments. +- If the paper includes experiments, a No answer to this question will not be perceived well by the reviewers: Making the paper reproducible is important, regardless of whether the code and data are provided or not. +- If the contribution is a dataset and/or model, the authors should describe the steps taken to make their results reproducible or verifiable. +- Depending on the contribution, reproducibility can be accomplished in various ways. For example, if the contribution is a novel architecture, describing the architecture fully might suffice, or if the contribution is a specific model and empirical evaluation, it may be necessary to either make it possible for others to replicate the model with the same dataset, or provide access to the model. In general, releasing code and data is often one good way to accomplish this, but reproducibility can also be provided via detailed instructions for how to replicate the results, access to a hosted model (e.g., in the case of a large language model), releasing of a model checkpoint, or other means that are appropriate to the research performed. +- While NeurIPS does not require releasing code, the conference does require all submissions to provide some reasonable avenue for reproducibility, which may depend on the nature of the contribution. For example +(a) If the contribution is primarily a new algorithm, the paper should make it clear how to reproduce that algorithm. +(b) If the contribution is primarily a new model architecture, the paper should describe the architecture clearly and fully. +(c) If the contribution is a new model (e.g., a large language model), then there should either be a way to access this model for reproducing the results or a way to reproduce the model (e.g., with an open-source dataset or instructions for how to construct the dataset). +(d) We recognize that reproducibility may be tricky in some cases, in which case authors are welcome to describe the particular way they provide for reproducibility. In the case of closed-source models, it may be that access to the model is limited in some way (e.g., to registered users), but it should be possible for other researchers to have some path to reproducing or verifying the results. + +# 5. Open access to data and code + +Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material? + +Answer: [Yes] + +Justification: An implementation of DCAReasoner is publicly available on GitHub. Furthermore, the code for our baselines as well as all used benchmark datasets are already publicly available as described in detail in our appendices. + +Guidelines: + +- The answer NA means that paper does not include experiments requiring code. +- Please see the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details. +- While we encourage the release of code and data, we understand that this might not be possible, so "No" is an acceptable answer. Papers cannot be rejected simply for not including code, unless this is central to the contribution (e.g., for a new open-source benchmark). +- The instructions should contain the exact command and environment needed to run to reproduce the results. See the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details. +- The authors should provide instructions on data access and preparation, including how to access the raw data, preprocessed data, intermediate data, and generated data, etc. +- The authors should provide scripts to reproduce all experimental results for the new proposed method and baselines. If only a subset of experiments are reproducible, they should state which ones are omitted from the script and why. +- At submission time, to preserve anonymity, the authors should release anonymized versions (if applicable). +- Providing as much information as possible in supplemental material (appended to the paper) is recommended, but including URLs to data and code is permitted. + +# 6. Experimental setting/details + +Question: Does the paper specify all the training and test details (e.g., data splits, hyperparameters, how they were chosen, type of optimizer, etc.) necessary to understand the results? + +Answer: [Yes] + +Justification: All details for training and testing are provided in our appendices. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The experimental setting should be presented in the core of the paper to a level of detail that is necessary to appreciate the results and make sense of them. +- The full details can be provided either with the code, in appendix, or as supplemental material. + +# 7. Experiment statistical significance + +Question: Does the paper report error bars suitably and correctly defined or other appropriate information about the statistical significance of the experiments? + +Answer: [Yes] + +Justification: We report the standard error of the mean in our main experiments, see Table 1. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The authors should answer "Yes" if the results are accompanied by error bars, confidence intervals, or statistical significance tests, at least for the experiments that support the main claims of the paper. +- The factors of variability that the error bars are capturing should be clearly stated (for example, train/test split, initialization, random drawing of some parameter, or overall run with given experimental conditions). + +- The method for calculating the error bars should be explained (closed form formula, call to a library function, bootstrap, etc.) +- The assumptions made should be given (e.g., Normally distributed errors). +- It should be clear whether the error bar is the standard deviation or the standard error of the mean. +- It is OK to report 1-sigma error bars, but one should state it. The authors should preferably report a 2-sigma error bar than state that they have a $96\%$ CI, if the hypothesis of Normality of errors is not verified. +- For asymmetric distributions, the authors should be careful not to show in tables or figures symmetric error bars that would yield results that are out of range (e.g. negative error rates). +- If error bars are reported in tables or plots, The authors should explain in the text how they were calculated and reference the corresponding figures or tables in the text. + +# 8. Experiments compute resources + +Question: For each experiment, does the paper provide sufficient information on the computer resources (type of compute workers, memory, time of execution) needed to reproduce the experiments? + +Answer: [Yes] + +Justification: Yes all details are reported in the main paper and appendices (see Appendix B.2). + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The paper should indicate the type of compute workers CPU or GPU, internal cluster, or cloud provider, including relevant memory and storage. +- The paper should provide the amount of compute required for each of the individual experimental runs as well as estimate the total compute. +- The paper should disclose whether the full research project required more compute than the experiments reported in the paper (e.g., preliminary or failed experiments that didn't make it into the paper). + +# 9. Code of ethics + +Question: Does the research conducted in the paper conform, in every respect, with the NeurIPS Code of Ethics https://neurips.cc/public/EthicsGuidelines? + +Answer: [Yes] + +Justification: Our research does conform with the NeurIPS Code of Ethics. + +Guidelines: + +- The answer NA means that the authors have not reviewed the NeurIPS Code of Ethics. +- If the authors answer No, they should explain the special circumstances that require a deviation from the Code of Ethics. +- The authors should make sure to preserve anonymity (e.g., if there is a special consideration due to laws or regulations in their jurisdiction). + +# 10. Broader impacts + +Question: Does the paper discuss both potential positive societal impacts and negative societal impacts of the work performed? + +Answer: [Yes] + +Justification: We included a societal impact paragraph in our discussion. + +Guidelines: + +- The answer NA means that there is no societal impact of the work performed. +- If the authors answer NA or No, they should explain why their work has no societal impact or why the paper does not address societal impact. + +- Examples of negative societal impacts include potential malicious or unintended uses (e.g., disinformation, generating fake profiles, surveillance), fairness considerations (e.g., deployment of technologies that could make decisions that unfairly impact specific groups), privacy considerations, and security considerations. +- The conference expects that many papers will be foundational research and not tied to particular applications, let alone deployments. However, if there is a direct path to any negative applications, the authors should point it out. For example, it is legitimate to point out that an improvement in the quality of generative models could be used to generate deepfakes for disinformation. On the other hand, it is not needed to point out that a generic algorithm for optimizing neural networks could enable people to train models that generate Deepfakes faster. +- The authors should consider possible harms that could arise when the technology is being used as intended and functioning correctly, harms that could arise when the technology is being used as intended but gives incorrect results, and harms following from (intentional or unintentional) misuse of the technology. +- If there are negative societal impacts, the authors could also discuss possible mitigation strategies (e.g., gated release of models, providing defenses in addition to attacks, mechanisms for monitoring misuse, mechanisms to monitor how a system learns from feedback over time, improving the efficiency and accessibility of ML). + +# 11. Safeguards + +Question: Does the paper describe safeguards that have been put in place for responsible release of data or models that have a high risk for misuse (e.g., pretrained language models, image generators, or scraped datasets)? + +Answer: [NA] + +Justification: The paper poses no such risks. + +Guidelines: + +- The answer NA means that the paper poses no such risks. +- Released models that have a high risk for misuse or dual-use should be released with necessary safeguards to allow for controlled use of the model, for example by requiring that users adhere to usage guidelines or restrictions to access the model or implementing safety filters. +- Datasets that have been scraped from the Internet could pose safety risks. The authors should describe how they avoided releasing unsafe images. +- We recognize that providing effective safeguards is challenging, and many papers do not require this, but we encourage authors to take this into account and make a best faith effort. + +# 12. Licenses for existing assets + +Question: Are the creators or original owners of assets (e.g., code, data, models), used in the paper, properly credited and are the license and terms of use explicitly mentioned and properly respected? + +Answer: [Yes] + +Justification: All used assets are properly cited and the name of the license is explicitly mentioned. + +Guidelines: + +- The answer NA means that the paper does not use existing assets. +- The authors should cite the original paper that produced the code package or dataset. +- The authors should state which version of the asset is used and, if possible, include a URL. +- The name of the license (e.g., CC-BY 4.0) should be included for each asset. +- For scraped data from a particular source (e.g., website), the copyright and terms of service of that source should be provided. + +- If assets are released, the license, copyright information, and terms of use in the package should be provided. For popular datasets, paperswithcode.com/datasets has curated licenses for some datasets. Their licensing guide can help determine the license of a dataset. +- For existing datasets that are re-packaged, both the original license and the license of the derived asset (if it has changed) should be provided. +- If this information is not available online, the authors are encouraged to reach out to the asset's creators. + +# 13. New assets + +Question: Are new assets introduced in the paper well documented and is the documentation provided alongside the assets? + +Answer: [Yes]. + +Justification: An implementation of DCAReasoner is publicly available on GitHub. + +Guidelines: + +- The answer NA means that the paper does not release new assets. +- Researchers should communicate the details of the dataset/code/model as part of their submissions via structured templates. This includes details about training, license, limitations, etc. +- The paper should discuss whether and how consent was obtained from people whose asset is used. +- At submission time, remember to anonymize your assets (if applicable). You can either create an anonymized URL or include an anonymized zip file. + +# 14. Crowdsourcing and research with human subjects + +Question: For crowdsourcing experiments and research with human subjects, does the paper include the full text of instructions given to participants and screenshots, if applicable, as well as details about compensation (if any)? + +Answer: [NA] + +Justification: The paper does not involve crowdsourcing or research with human subjects. + +Guidelines: + +- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. +- Including this information in the supplemental material is fine, but if the main contribution of the paper involves human subjects, then as much detail as possible should be included in the main paper. +- According to the NeurIPS Code of Ethics, workers involved in data collection, curation, or other labor should be paid at least the minimum wage in the country of the data collector. + +# 15. Institutional review board (IRB) approvals or equivalent for research with human subjects + +Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals (or an equivalent approval/review based on the requirements of your country or institution) were obtained? + +Answer: [NA] + +Justification: The paper does not involve crowdsourcing or research with human subjects. + +Guidelines: + +- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. +- Depending on the country in which research is conducted, IRB approval (or equivalent) may be required for any human subjects research. If you obtained IRB approval, you should clearly state this in the paper. + +We recognize that the procedures for this may vary significantly between institutions and locations, and we expect authors to adhere to the NeurIPS Code of Ethics and the guidelines for their institution. +- For initial submissions, do not include any information that would break anonymity (if applicable), such as the institution conducting the review. + +# 16. Declaration of LLM usage + +Question: Does the paper describe the usage of LLMs if it is an important, original, or non-standard component of the core methods in this research? Note that if the LLM is used only for writing, editing, or formatting purposes and does not impact the core methodology, scientific rigorousness, or originality of the research, declaration is not required. + +Answer: [NA] + +Justification: The core method development in this research does not involve LLMs as any important, original, or non-standard components. + +Guidelines: + +- The answer NA means that the core method development in this research does not involve LLMs as any important, original, or non-standard components. +- Please refer to our LLM policy (https://neurips.cc/Conferences/2025/LLM) for what should or should not be described. + +# A Mathematical Appendix + +# A.1 Proof of Lemma 1 + +It is clear that $E_{\theta}(x,y)$ can be decomposed as + +$$ +E _ {\theta} (x, y) = E _ {\theta} (x) = g (x) - h (x) \tag {15} +$$ + +with + +$$ +g (x) = \frac {\rho}{2} \| x \| ^ {2} + \sum_ {\alpha_ {i} > 0} \alpha_ {i} \sigma \left(\langle w _ {i}, x \rangle + b _ {i}\right) \tag {16} +$$ + +$$ +h (x) = \sum_ {\alpha_ {i} < 0} | \alpha_ {i} | \sigma \left(\langle w _ {i}, x \rangle + b _ {i}\right) + \langle \xi , x \rangle + \omega . \tag {17} +$$ + +Furthermore, $\langle w_i,x\rangle +b_i$ is affine in $x$ and hence convex and $\sigma$ is non-decreasing and convex. Thus the composition $\sigma (\langle w_i,x\rangle +b_i)$ is always convex and the claim follows by observing that $g$ and $h$ are linear combinations of convex functions with positive weights. + +# A.2 Proof of Lemma 2 + +For point one, note that the modulus of strong convexity is defined as + +$$ +\rho (g) = \sup _ {\rho \geq 0} \left\{g (x) - \frac {\rho}{2} \| x \| ^ {2} \text {i s c o n v e x} \right\}, \tag {18} +$$ + +see, e.g., [44]. It is easy to see that $g - \frac{\rho}{2}\| \cdot \| ^2$ is convex and hence $\rho (g)\geq \rho >0$ , which implies strong convexity. + +For point two, polyhedral convex functions can be characterized as + +$$ +\left. \max \left\{\langle q _ {k}, x \rangle + p _ {k}: k \in \{1, \dots , K \} \right\}, \right. \tag {19} +$$ + +see, e.g., [39]. Note that + +$$ +\begin{array}{l} h (x) - \omega = \sum_ {\alpha_ {i} < 0} | \alpha_ {i} | \sigma \left(\langle w _ {i}, x \rangle + b _ {i}\right) + \langle \xi , x \rangle (20) \\ = \sum_ {\alpha_ {i} < 0} \sigma \left(\langle | \alpha_ {i} | w _ {i}, x \rangle + | \alpha_ {i} | b _ {i}\right) + \langle \xi , x \rangle (21) \\ = \max _ {\delta_ {i} \in \{0, 1 \}} \sum_ {\alpha_ {i} < 0} \delta_ {i} \left(\left\langle | \alpha_ {i} | w _ {i}, x \right\rangle + | \alpha_ {i} | b _ {i}\right) + \langle \xi , x \rangle , (22) \\ \end{array} +$$ + +where the last equality follows from the fact that the maximum is achieved in the case where $\delta_{i} = 1$ if $\langle |\alpha_i|w_i,x\rangle +|\alpha_i|b_i\geq 0$ and $\delta_{i} = 0$ else. Furthermore, + +$$ +\max _ {\delta_ {i} \in \{0, 1 \}} \sum_ {\alpha_ {i} < 0} \delta_ {i} \left(\left\langle | \alpha_ {i} | w _ {i}, x \right\rangle + | \alpha_ {i} | b _ {i}\right) + \langle \xi , x \rangle = \tag {23} +$$ + +$$ +\max _ {\delta_ {i} \in \{0, 1 \}} \left\langle \sum_ {\alpha_ {i} < 0} | \alpha_ {i} | \delta_ {i} w _ {i} + \xi , x \right\rangle + \sum_ {\alpha_ {i} < 0} | \alpha_ {i} | \delta_ {i} b _ {i}, \tag {24} +$$ + +which proves the claim. + +# A.3 Proof of Lemma 3 + +First, note that + +$$ +g (x) - \langle v, x \rangle = \frac {\rho}{2} \| x \| ^ {2} + \sum_ {\alpha_ {i} > 0} \alpha_ {i} \sigma \left(\langle w _ {i}, x \rangle + b _ {i}\right) - \langle v, x \rangle , \tag {25} +$$ + +which can be equivalently stated as + +$$ +\min _ {x, z} \frac {1}{2} \left( \begin{array}{l} x \\ z \end{array} \right) ^ {T} \left( \begin{array}{c c} \rho I & 0 \\ 0 & 0 \end{array} \right) \left( \begin{array}{l} x \\ z \end{array} \right) + \left\langle \left( \begin{array}{l} x \\ z \end{array} \right), \left( \begin{array}{c} - v \\ \alpha^ {+} \end{array} \right) \right\rangle , \tag {26} +$$ + +and the non-linear equality constrained $z_{i} = \sigma (\langle w_{i},x\rangle +b_{i})$ for all $i\in \{1,\ldots ,N_x\}$ with $\alpha_{i} > 0$ . Now, the inequality constraints + +$$ +z _ {i} \geq 0 \tag {27} +$$ + +$$ +z _ {i} \geq \left\langle w _ {i}, x \right\rangle + b _ {i} \tag {28} +$$ + +imply that $z_{i} \geq \sigma(\langle w_{i}, x \rangle + b_{i})$ . As the only term in the objective function containing $z$ is given by $\langle \alpha^{+}, z \rangle$ and $\alpha^{+}$ is elementwise strictly positive, we have that problem (26) with constraints (27) and (28) always results in $z_{i}^{*} = \sigma(\langle w_{i}, x^{*} \rangle + b_{i})$ at an optimal point $(x^{*}, z^{*})$ . Hence, the claim follows. + +# A.4 Proof of Theorem 1 + +From Lemma 2, we know that $g$ is strongly convex with $\rho(g) \geq \rho$ . Convergence thus follows by Theorem 6 in [45]. Furthermore, Lemma 2 shows that $h$ is polyhedral convex, which proves the finite convergence property using Theorem 9 in [45]. Last, the fact that $x^*$ is a local minimum if Equation (11) holds true follows from Corollary 2 in [45]. + +# A.5 Proof of Lemma 4 + +It is easy to see that under Assumption 1, the function $g$ reduces to $\frac{\rho}{2}\| x\|^2$ . Hence (QP) reduces to + +$$ +\frac {\rho}{2} \| x \| ^ {2} - \langle x, v \rangle , \tag {29} +$$ + +with the global solution $\frac{1}{\rho} v$ . Furthermore, in this case $g$ is differentiable and hence $\partial g(x^{*}) = \{\nabla g(x^{*})\}$ is a singleton. Thus, Equation (11) always holds true under Assumption 1. + +# A.6 Details on Approximation in $y$ + +In our energy function (2) the weights to parameterize the function in $x$ , i.e., $(\rho_{\theta}(y), \xi_{\theta}(y), \omega_{\theta}(y), \alpha_{\theta}(y), W_{\theta}(y), b_{\theta}(y))$ , are all single hidden layer neural networks in $y$ . Since $(\xi_{\theta}(y), \omega_{\theta}(y), W_{\theta}(y), b_{\theta}(y))$ are all using the identity as an output layer activation those are universal approximators [27]. Note also that $\rho_{\theta}(y)$ and $\alpha_{\theta}(y)$ are using softmax activations in the output layer. Hence, they are still able to approximate any strictly positive continuous function arbitrarily well as shown in the following. + +Let $f$ be an arbitrary strictly positive and continuous function. Then $\log (e^{f} - 1)$ is a continuous function mapping to $(-\infty, \infty)$ . As such it can be approximated arbitrarily well by a single hidden layer network by [27]. Further, if we now equip this network with a softplus activation, we approximate softplus $(\log (e^{f} - 1)) = f$ by using that softplus is Lipschitz. We formalize this in the following lemma. + +Lemma 8. Let $f: \mathbb{R}^n \to (0, \infty)$ be continuous. Then for every $\epsilon > 0$ there is a single hidden layer network with a softplus activation, $\nu_{\epsilon}$ , such that $\| f - \nu_{\epsilon} \|_{\infty} < \epsilon$ + +Proof. Let $\epsilon > 0$ . By [27], there is a single hidden layer network $N_{\epsilon}$ such that $\| \log (e^{f(\cdot)} - 1) - N_{\epsilon} \|_{\infty} < \epsilon$ . Using that the function $\text{softplus}(y) = \log (1 + e^{y})$ is Lipschitz of constant 1 and that $z \mapsto \log (e^{z} - 1)$ is its inverse, we deduce that for every $x \in \mathbb{R}^n$ : $\left| f(x) - \log (1 + e^{N_{\epsilon}(x)}) \right| \leq \left| \log (e^{f(x)} - 1) - N_{\epsilon}(x) \right|$ and therefore + +$$ +\begin{array}{l} \left\| f - \log \left(1 + e ^ {N _ {\epsilon}}\right) \right\| _ {\infty} = \sup _ {x \in \mathbb {R} ^ {n}} \left| f (x) - \log \left(1 + e ^ {N _ {\epsilon} (x)}\right) \right| \\ \leq \sup _ {x \in \mathbb {R} ^ {n}} \left| \log \left(e ^ {f (x)} - 1\right) - N _ {\epsilon} (x) \right| = \| \log \left(e ^ {f} - 1\right) - N _ {\epsilon} \| _ {\infty} < \epsilon \\ \end{array} +$$ + +By setting $\nu_{\epsilon}(x) = \log \left(1 + e^{N_{\epsilon}(x)}\right)$ , the claim follows. + +![](images/555054d57bc5378ddc28b41af463245e52a72a82a11120ada69fd20547dcc96d.jpg) + +# A.7 Proof of Theorem 3 + +Let $f \in \mathcal{C}_0(\mathbb{X})$ and $\epsilon > 0$ . From Lemma 5 we know that there exists a weakly-concave function $f_{\epsilon}$ such that $\| f - f_{\epsilon} \|_{\infty} < \epsilon$ . Hence, there exists a $\rho > 0$ and a convex function $r$ such that $-f_{\epsilon}(x) + \rho / 2 \| x \|^2 = r(x)$ . We need to show that $r$ is $\epsilon$ -convexly-ReLU-representable. + +We assume first that $r$ is Lipschitz continuous. Following the derivations in Theorem 2 in [11], we can find a sequence of convex piecewise linear functions $r_n$ such that $r_n \to r$ uniformly for $n \to \infty$ . Furthermore, by Theorem 2.2 in [6] there exist single layer ReLU networks $\mathcal{NN}_n$ that can represent $r_n$ exactly which proves the claim. + +If we relax the Lipschitz assumption, we can again follow the arguments in Theorem 2 in [11] and find a sequence of Lipschitz continuous and convex functions $\hat{r}_k$ that converge uniformly to $r$ . Hence for $\epsilon > 0$ we can find $k \geq 1$ such that $\| r - \hat{r}_k \|_{\infty} < \epsilon / 2$ . With the arguments from above we can then find $r_n$ such that $\| \hat{r}_k - r_n \|_{\infty} < \epsilon / 2$ , and hence $\| r - r_n \|_{\infty} < \epsilon$ which proves the claim. + +# A.8 Restart Procedure for our DCA Routine + +Following [46], let the index sets $I_h(x^*)$ and $I_g(x^*)$ be defined as follows + +$$ +I _ {h} \left(x ^ {*}\right) = \left\{i \in \{1, \dots , N _ {x} \}: \alpha_ {i} < 0 \text {a n d} \langle w _ {i}, x ^ {*} \rangle + b _ {i} = 0 \right\}, \tag {30} +$$ + +$$ +I _ {g} \left(x ^ {*}\right) = \{i \in \{1, \dots , N _ {x} \}: \alpha_ {i} > 0 \text {a n d} \langle w _ {i}, x ^ {*} \rangle + b _ {i} = 0 \}. \tag {31} +$$ + +Note that if $I_h(x^*) = \emptyset$ or $I_g(x^*) = \emptyset$ we have that $\partial h(x^*) \subseteq \partial g(x^*)$ as one of the two sets is a singleton and $x^*$ is DC-critical. Hence, for the rest of this derivation we assume $I_h(x^*) \neq \emptyset$ and $I_g(x^*) \neq \emptyset$ . + +Now, the subgradients of $h$ , respectively $g$ , are given by + +$$ +\partial h \left(x ^ {*}\right) = \left\{\xi + \sum_ {\alpha_ {i} < 0} \left| \alpha_ {i} \right| w _ {i} H \left(\left\langle w _ {i}, x ^ {*} \right\rangle + b _ {i}\right) \epsilon_ {i} ^ {h}: \epsilon_ {i} ^ {h} \in \left\{ \begin{array}{l l} [ 0, 1 ] & \text {i f} i \in I _ {h} \left(x ^ {*}\right) \\ \{1 \} & \text {e l s e} \end{array} \right. \right\}, \tag {32} +$$ + +$$ +\partial g \left(x ^ {*}\right) = \left\{\rho x ^ {*} + \sum_ {\alpha_ {i} > 0} \alpha_ {i} w _ {i} H \left(\left\langle w _ {i}, x ^ {*} \right\rangle + b _ {i}\right) \epsilon_ {i} ^ {g}: \epsilon_ {i} ^ {g} \in \left\{ \begin{array}{l l} [ 0, 1 ] & \text {i f} i \in I _ {g} \left(x ^ {*}\right) \\ \{1 \} & \text {e l s e} \end{array} \right. \right\}, \tag {33} +$$ + +where $H(z)$ denotes the Heaviside function, i.e., $H(z) = 1$ for $z \geq 0$ and $H(z) = 0$ else. Furthermore, let us define the following + +$$ +I _ {h} ^ {+} \left(x ^ {*}\right) = \left\{i \in \{1, \dots , N _ {x} \}: \alpha_ {i} < 0 \text {a n d} \langle w _ {i}, x ^ {*} \rangle + b _ {i} > 0 \right\}, \tag {34} +$$ + +$$ +I _ {g} ^ {+} \left(x ^ {*}\right) = \left\{i \in \{1, \dots , N _ {x} \}: \alpha_ {i} > 0 \text {a n d} \left\langle w _ {i}, x ^ {*} \right\rangle + b _ {i} > 0 \right\}, \tag {35} +$$ + +$$ +v _ {h} = \xi + \sum_ {i \in I _ {h} ^ {+} \left(x ^ {*}\right)} | \alpha_ {i} | w _ {i}, \tag {36} +$$ + +$$ +v _ {g} = \rho x ^ {*} + \sum_ {i \in I _ {g} ^ {+} \left(x ^ {*}\right)} \alpha_ {i} w _ {i}, \tag {37} +$$ + +as well as, the matrix $A_{h}\in \mathbb{R}^{n\times |I_{h}(x^{*})|}$ with columns $|\alpha_i|w_i$ for $i\in I_h(x^*)$ , and the matrix $A_{g}\in \mathbb{R}^{n\times |I_{g}(x^{*})|}$ with columns $\alpha_{i}w_{i}$ for $i\in I_g(x^*)$ . The inclusion $\partial h(x^{*})\subseteq \partial g(x^{*})$ holds true if + +$$ +\forall \epsilon^ {h} \in [ 0, 1 ] ^ {\left| I _ {h} \left(x ^ {*}\right) \right|} \exists \epsilon^ {g} \in [ 0, 1 ] ^ {\left| I _ {g} \left(x ^ {*}\right) \right|}: v _ {h} + A _ {h} \epsilon^ {h} = v _ {g} + A _ {g} \epsilon^ {g}. \tag {38} +$$ + +To check if this holds, we can solve the max-min problem + +$$ +\max _ {0 \leq \epsilon^ {h} \leq 1} \min _ {0 \leq \epsilon^ {g} \leq 1} \| v _ {h} + A _ {h} \epsilon^ {h} - v _ {g} - A _ {g} \epsilon^ {g} \| _ {1}, \tag {39} +$$ + +and observe whether or not the optimal objective value is zero. If it is, (38) holds and we have found a local minimum. If not, then we have found an element $x_0^* \in \partial h(x^*)$ with $x_0^* \notin \partial g(x^*)$ . Following [44], we can restart the DCA routine with $x_0^*$ and yield a strict energy reduction in the first step. + +Figure 2: Illustration of Batched DCA for batch size 512 +![](images/6a17b151462b906e51956cd7cceff24eb629e228434cb97e1a5defbc9c2db401.jpg) +Note. Illustration of our batched DCA routine computing locally minimal energy states for 512 reasoning tasks in parallel. + +![](images/30dd9acb23bebe13ef0dd0d93d5899a4615afc85874300250ae8876f03939aae.jpg) + +What remains to be discussed, is how to solve (39) efficiently. To do so, note that we can reformulate the problem as + +$$ +\max _ {\epsilon^ {h}} \min _ {\epsilon^ {g}, r _ {1}, r _ {2}} \left\langle r _ {1}, \mathbb {1} \right\rangle + \left\langle r _ {2}, \mathbb {1} \right\rangle \tag {40} +$$ + +$$ +\text {s . t .} v _ {h} + A _ {h} \epsilon^ {h} - v _ {g} - A _ {g} \epsilon^ {g} = r _ {1} - r _ {2} \tag {41} +$$ + +$$ +r _ {1}, r _ {2}, \epsilon^ {h}, \epsilon^ {g} \geq 0 \tag {42} +$$ + +$$ +\epsilon^ {h}, \epsilon^ {g} \leq 1 \tag {43} +$$ + +This problem can be solved using a branch and bound strategy for bilevel linear programming (see, e.g., section 6.2 in [8]). Furthermore, since $x^{*}$ is DC-critical, we know that for $\epsilon^h = \mathbb{1}$ there exists an $\epsilon_{*}^{g}\in [0,1]^{|I_{g}(x^{*})|}$ with $v_{h} + A_{h}\mathbb{1} = v_{g} + A_{g}\epsilon_{*}^{g}$ . Hence, we can warm start the branch and bound routine with $(\epsilon^h,\epsilon^g,r_1,r_2) = (\mathbb{1},\epsilon_{*}^{g},0,0)$ . If the optimal objective value is zero, we already start with an optimal solution. If not, we can stop the branch and bound routine as soon as we have found the first $(\epsilon^h,\epsilon^g)$ with $v_{h} + A_{h}\epsilon^{h}\neq v_{g} + A_{g}\epsilon^{g}$ . Thus, the restart procedure can be implemented efficiently. + +# B Details on Numerical Experiments + +# B.1 Visualization of Convergence + +We visualize how our batched DCA routine converges for 512 reasoning tasks in parallel in Figure 2. In particular we used the matrix inversion dataset, see Section 5, and for a fixed $y$ visualized the energy values $E_{\theta}(x_k,y)$ and norm differences $\| x_{k + 1} - x_k\|$ after training the parameters $\theta$ , i.e., at inference time. + +# B.2 Details on Experimental Setup + +For our baselines we closely follow the code releases of [20] and [21]. + +Model Specifications. For DCAReactor we set the number of hidden units $N_{x} = 8$ and $N_{y} = 4000$ for all benchmark datasets. For our baselines we scale the network size depending on the dataset to ensure that all models have roughly the same number of parameters. We set the convergence tolerance in Algorithm 1 to $10^{-5}$ and set a maximum of 30 DCA iterations. However, in our experiments we see that the batched DCA algorithm consistently converges to machine precision in less than 10 iterations. Hence, we also observe finite convergence guaranteed by Theorem 1 empirically. Starting points $x^{0}$ are sampled uniformly random as stated in Algorithm 1. We set $l = -1$ and $u = 1$ for all benchmark datasets except for QR and Matrix Multiplication for which we use $l = -5$ and $u = 5$ . Hyperparameters are summarized in Table 3. + +
HyperparameterValue
DCAReasonerIREMIRED
Common Hyperparameters
Learning Rate10-410-410-4
Batch Size512512512
Number of Gradient Steps10.00010.00010.000
Starting Point Samplinguniformuniformnormal
DCAReasoner Specific Hyperparameters
Number of NeuronsNx8
Number of NeuronsNy4000
DCA Convergence Tolerance tol10-5
Maximum DCA Iterations30
+ +Table 3: Hyperparameters settings for our experiments. Note that for our baselines we scale the network size depending on the dataset to ensure that all models have roughly the same number of parameters. + +Training. For training, we use the Adam optimizer with a learning rate of $10^{-4}$ as suggested in [20, 21] for all models. We set the batch size to 512 and train each model for a fixed number of 10000 iterations, i.e., gradient steps. Hyperparameters are summarized in Table 3. + +Evaluation. For evaluation, we are using again a batch size of 512 and perform 20 test iterations, summing up to roughly 10000 test problems per difficulty level, i.e., once in the easy setting and once in the hard setting. + +Hardware. All experiments are performed on a n1-standard-2 Google cloud instance with 7.5GB RAM and two NVIDIA T4 GPUs. + +# B.3 Details on Benchmark Datasets + +# B.3.1 Motivation + +In general, neural algorithmic reasoning constitutes an unsolved problem in machine learning. For an argumentation on the complexity of neural algorithmic reasoning see e.g. [47]. Matrix Completion, QR Decomposition, and Matrix Multiplication represent algorithmic reasoning tasks. Du et. al. argue that effective algorithmic reasoning requires repetitive application of underlying algorithmic computations, dependent on problem complexity, and thus serves as a natural benchmark for iterative reasoning [20]. Learning parities is a well-known reasoning benchmark and well-studied in learning theory in general. Shoshani and Shamir argue that there is strong empirical evidence that suggests that parities cannot be learned using more standard general purpose learning methods, and in particular gradient methods, once the dimension is even moderately large [42]. This also goes in line with the fact that our baselines struggled particularly on this task, see also conclusion and limitations section in [20]. For matrix inversion, Ji et. al. argue that despite significant progress in deep learning, there exists no universal neural-based method for approximating matrix inversion [30], showing that this is a non-trivial task for neural reasoning. + +# B.3.2 Technical Details + +In the following we give more details on our benchmark datasets. We followed the code releases of [20] and [21] closely. + +Matrix Inverse. We construct well-conditioned random $20 \times 20$ -matrices $A = 2MM^T + 0.5 \cdot I$ , with $M$ being a random matrix with entries in $[-1,1]$ . The task is then to compute the inverse $A^{-1}$ . Harder tasks are created by making the matrices less-well conditioned by setting $A = 2MM^T + 0.1 \cdot I$ . + +Matrix Completion. We randomly construct low-rank $20 \times 20$ -matrices $A = 0.1 \cdot N + \frac{1}{20} UV^T$ where $N$ is a standard normally distributed noise matrix and $U$ and $V$ are again standard normally distributed random $20 \times 10$ -matrices. Then, a random mask is created by rounding a randomly generated uniformly distributed matrix $M$ . The model is then presented with the masked matrix $A$ and asked to recover it. Harder tasks are created by setting $A = 0.1 \cdot N + \frac{1}{5} UV^T$ . + +Figure 3: Empirical Analysis of Robustens to Noise on the QR Decomposition Dataset +![](images/10c89cdd11f60c91622295a1e860b5d859fbf851c00616c61813c00d8ccd48df.jpg) +Note. We evaluated the trained models on the QR benchmark dataset adding different Gaussian noise levels to the input data. IREM becomes unstable for large noise levels while IRED and DCAReasoner appear to be robust to noisy inputs. + +Parity. Similar to [25], we create uniformly random vectors in $[0,1]^{20}$ and then set the target to 0 if the number of values strictly greater than 0.5 is even and 1 otherwise. Harder tasks are created by drawing random vectors from $[-1,2]^{20}$ . + +QR Decomposition. We create uniformly random matrices $A$ with entries in $[-1, 1]$ and then compute the QR decomposition, i.e., $A = QR$ . The models are then given the matrix $A$ and asked to reconstruct both $Q$ and $R$ . Harder tasks are created by creating uniformly random matrices $A$ with entries in $[-2.5, 2.5]$ . + +Matrix Multiplication. We create uniformly random matrices $A$ with entries in $[-1, 1]$ and then compute the square, i.e., $A^2$ . The models are then given the matrix $A$ and asked to perform the matrix multiplication $A^2$ . Harder tasks are created by creating uniformly random matrices $A$ with entries in $[-1.5, 1.5]$ . + +# B.4 Further Experiments Analyzing Robustness to Noisy Data + +We evaluated the trained models on the QR benchmark dataset adding different Gaussian noise levels to the input matrix before processing with a scale varying in $\{10^{-4}, 10^{-3}, 10^{-2}, 10^{-1}, 1, 10\}$ . Results are visualized in Figure 3. It appears that IREM becomes unstable for large noise levels while IRED and DCAReasoner are mostly robust to noise. + +# B.5 Details on Energy-Based Reasoning in Token Embedding Space + +For our experiments, we make use of the symptom-to-diagnosis dataset for medical reasoning which is freely available on huggingface. It provides a training (853 examples) and test (212 examples) dataset consisting of short texts in which a patient describes her symptoms and a corresponding diagnosis out of a set of 22 medical diagnoses summarized in Table 4. + +We first finetune an uncased DistilBERT model for text classification, splitting the training set again into training and validation using a 80/20 split, and using the huggingface trainer for sequence classification. In particular, we use a batch size of 16, a learning rate of $2 * 10^{-5}$ , a weight decay of 0.01, and 10 evaluation steps saving the best performing model in terms of accuracy (which is also $96\%$ as for the energy-based models). + +We then use the CLS token embeddings of the texts in the training dataset of this finetuned model as inputs $y$ and the embeddings of the corresponding diagnosis as a ground truth $x$ to train the DCAReasoner and our baselines in a continuous reasoning setting. The model specifications for DCAReasoner are the same as in our main experiments, i.e., we set $N_{x} = 8$ and $N_{y} = 4000$ , + +
DiagnosisTraining ExamplesTest examples
drug reaction408
allergy4010
chicken pox4010
diabetes4010
psoriasis4010
hypertension4010
cervical spondylosis4010
bronchial asthma4010
varicose veins4010
malaria4010
dengue4010
arthritis4010
impetigo4010
fungal infection399
common cold3910
gastroesophageal reflux disease3910
urinary tract infection399
typhoid389
pneumonia3710
peptic ulcer disease3710
jaundice337
migraine3210
+ +Table 4: Number of examples per split and diagnosis in the Symptom-to-Diagnosis dataset. + +![](images/e87b7963a09c82945abbbb75dc1588da272bfae7d2d51af78ec84751130f4705.jpg) +Note. Cross-Entropy loss during training (left) and cell accuracies on validation data (right). Validation steps are performed every 1000 training steps. + +![](images/cf61dbec9e1b00c169b9de52b997c6e2328621f30fd76348fbc8eb01df3f3912.jpg) +Figure 4: Training and Validation Curves for Selenium Experiment + +scaling the network sizes in our baselines to ensure that all models have roughly the same number of parameters. + +For training, we use a batch size of 64 and a learning rate of $10^{-4}$ . We train each model for 10000 iterations. For evaluation, we use the test set and process it again in batches of size 64. + +# C Preliminary Experiments in Discrete Setting + +As mentioned in the main paper, DCAReasoner is designed for continuous reasoning tasks. Thus, our method cannot be applied as-is to experiments in discrete spaces. Nevertheless, we performed + +preliminary experiments using the Selenium dataset in [41]. The authors provide a training dataset consisting of 1.8 million Selenium puzzles as well as a validation set with 0.1 million Selenium puzzles. + +For reasoning in a discrete setting our energy landscape is defined in the logit space, i.e., a 729 dimensional setting and we replace the MSE with the cross entropy loss. Processing the inputs, i.e., Suku puzzles, can be done in multiple ways. We decided for a convolutional neural network with the residual connection design as in [21] to process the Suku puzzles before feeding them to our neural network components. Using the same settings as in the main paper, i.e., $N_{x} = 8$ , $N_{y} = 4000$ and a learning rate of $10^{-4}$ , but a batch size of 64, we train DCAReasoner for a total of 10 epochs. Validation steps, i.e., computing the cell accuracy defined as the percentage of unfilled cells whose values are correctly predicted on the validation data [41], are performed every 1000 training steps. The training process is visualized in Figure 4. Note that the validation performance stagnates after roughly 6 epochs. At this time, we observe a cell accuracy of $65.09\%$ , i.e., $65.09\%$ of all unfilled cells in all of the 100K validation puzzles are filled correctly. For context, the baseline from [41] with random order of input cells achieves $52\%$ , while the one with fixed order achieves $58.64\%$ . Our results only fall short from the accuracy achieved by the model receiving additional context information about solving strategies during training (94.23%). However, we point out that the specific architecture of our model does not allow such information to be provided. \ No newline at end of file diff --git a/NeurIPS/2025/A Difference-of-Convex Functions Approach to Energy-Based Iterative Reasoning/images.zip b/NeurIPS/2025/A Difference-of-Convex Functions Approach to Energy-Based Iterative Reasoning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..c0e6798933725862a0a0938f13f74bc473cb7fec --- /dev/null +++ b/NeurIPS/2025/A Difference-of-Convex Functions Approach to Energy-Based Iterative Reasoning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0bfa7b02ca4f16f67ee4dbb9b5ba8e1f146c9ea3cc7e2b94e611c670de830658 +size 569988 diff --git a/NeurIPS/2025/A Difference-of-Convex Functions Approach to Energy-Based Iterative Reasoning/layout.json b/NeurIPS/2025/A Difference-of-Convex Functions Approach to Energy-Based Iterative Reasoning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..fd1237cda9d9850c45e068f61bde2b188a96a2b8 --- /dev/null +++ b/NeurIPS/2025/A Difference-of-Convex Functions Approach to Energy-Based Iterative Reasoning/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6683fd3efa8ceb09871a0b61d6de52b2e9f367853bfea9415edd7afe0f2be267 +size 1099814 diff --git a/NeurIPS/2025/A Differential and Pointwise Control Approach to Reinforcement Learning/e495c951-b3ee-45cb-8760-228058c75c6b_content_list.json b/NeurIPS/2025/A Differential and Pointwise Control Approach to Reinforcement Learning/e495c951-b3ee-45cb-8760-228058c75c6b_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..ab798712aed27ea4d3d0ce9d572166e37a9f584c --- /dev/null +++ b/NeurIPS/2025/A Differential and Pointwise Control Approach to Reinforcement Learning/e495c951-b3ee-45cb-8760-228058c75c6b_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:53259ca0e519fd313b5f5576af6b4802a72d2ece2144a7a9906ff3f085fbd64d +size 217030 diff --git a/NeurIPS/2025/A Differential and Pointwise Control Approach to Reinforcement Learning/e495c951-b3ee-45cb-8760-228058c75c6b_model.json b/NeurIPS/2025/A Differential and Pointwise Control Approach to Reinforcement Learning/e495c951-b3ee-45cb-8760-228058c75c6b_model.json new file mode 100644 index 0000000000000000000000000000000000000000..e204a92e7932632eda6e6bce206ce9dac1c57b27 --- /dev/null +++ b/NeurIPS/2025/A Differential and Pointwise Control Approach to Reinforcement Learning/e495c951-b3ee-45cb-8760-228058c75c6b_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2ab16977e2a43b97804a59c131ed5b882668785538127247bde4c9eb1fd4af90 +size 267207 diff --git a/NeurIPS/2025/A Differential and Pointwise Control Approach to Reinforcement Learning/e495c951-b3ee-45cb-8760-228058c75c6b_origin.pdf b/NeurIPS/2025/A Differential and Pointwise Control Approach to Reinforcement Learning/e495c951-b3ee-45cb-8760-228058c75c6b_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..0983b84eb56a2d66d1be759d47f1e81be8c6d758 --- /dev/null +++ b/NeurIPS/2025/A Differential and Pointwise Control Approach to Reinforcement Learning/e495c951-b3ee-45cb-8760-228058c75c6b_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:986939c070ee80e149876bfe6aeb1626d65f2c56de85fafc9a6c20b68c600d56 +size 779248 diff --git a/NeurIPS/2025/A Differential and Pointwise Control Approach to Reinforcement Learning/full.md b/NeurIPS/2025/A Differential and Pointwise Control Approach to Reinforcement Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..c522735fd4f8eb42275ca80344c52b7637faf3b9 --- /dev/null +++ b/NeurIPS/2025/A Differential and Pointwise Control Approach to Reinforcement Learning/full.md @@ -0,0 +1,1115 @@ +# A Differential and Pointwise Control Approach to Reinforcement Learning + +Minh Nguyen + +University of Texas at Austin + +minhpnguyen@utexas.edu + +Chandrajit Bajaj + +University of Texas at Austin + +bajaj@cs.utexas.edu + +# Abstract + +Reinforcement learning (RL) in continuous state-action spaces remains challenging in scientific computing due to poor sample efficiency and lack of pathwise physical consistency. We introduce Differential Reinforcement Learning (Differential RL), a novel framework that reformulates RL from a continuous-time control perspective via a differential dual formulation. This induces a Hamiltonian structure that embeds physics priors and ensures consistent trajectories without requiring explicit constraints. To implement Differential RL, we develop Differential Policy Optimization (dfPO), a pointwise, stage-wise algorithm that refines local movement operators along the trajectory for improved sample efficiency and dynamic alignment. We establish pointwise convergence guarantees, a property not available in standard RL, and derive a competitive theoretical regret bound of $\mathcal{O}(K^{5/6})$ . Empirically, dfPO outperforms standard RL baselines on representative scientific computing tasks, including surface modeling, grid control, and molecular dynamics, under low-data and physics-constrained conditions. + +# 1 Introduction + +Reinforcement learning (RL) has achieved notable success across domains such as robotics, biological sciences, and control systems ([16, 20, 6, 2]). Yet, its application to scientific computing remains limited largely due to persistent challenges in sample complexity, lack of physical consistency, and weak theoretical guarantees. Unlike supervised learning, where models learn from labeled datasets, RL agents must explore through trial-and-error, often receiving sparse and delayed feedback. This makes data efficiency a critical bottleneck. Furthermore, standard RL methods typically fail to encode physical laws or structural priors, leading to suboptimal solutions in scientific problems governed by continuous dynamics. + +Model-based RL (MBRL) [31] improves sample efficiency by learning a surrogate model of the environment. However, current approaches require information typically unavailable under scientific computing settings. For example: + +1. Explicit reward model: Many MBRL algorithms (e.g., SVG [14], iLQR [28], PILCO [10]) require access to the whole reward functions and/or its derivative, when in reality reward functionals are only available at points along the explored trajectories. +2. Re-planning assumptions: Shooting methods and other trajectory-based planners [22] often assume the ability to re-plan from a particular intermediate time step. However, in many scientific computing problems, agents must always generate trajectories starting from the initial time, without resetting or modifying the trajectory midway (see Section C.7 for an example). + +As a result, researchers often revert to model-free RL combined with customized reward shaping. However, such approaches still fail to incorporate physics-informed priors or recover optimal policies + +under limited sample budgets. This motivates a fundamentally different approach: Rather than optimizing cumulative rewards over discrete steps, we interpret RL through the lens of continuous-time control, viewing trajectory returns as integrals and introducing a differential dual formulation. This naturally gives rise to a Hamiltonian structure that embeds inductive biases aligned with physical dynamics, even when explicit physical constraints are absent. + +To implement the framework, we develop Differential Policy Optimization (dfPO), an algorithm that directly optimizes a local trajectory operator and refines policy behavior pointwise along the trajectory, rather than through global value estimates. This locality enables dfPO to align updates with system dynamics at each timestep, avoiding inefficient exploration that is both costly and misaligned with physical constraints. By maintaining consistency with the optimal trajectory throughout execution, dfPO minimizes redundant relearning and reduces sample waste—crucial in scientific settings where simulation cost and rollout horizon are tightly constrained. + +We evaluate Differential RL on a suite of scientific problems that involve complex dynamics, implicit objectives, and limited data—settings where traditional RL struggles. + +1. Surface Modeling: Optimization over evolving surfaces, where rewards depend on the geometric and physical properties of the surface. +2. Grid-based Modeling: Control on coarse grids with fine-grid evaluations, representative of multiscale problems with implicit rewards. +3. Molecular Dynamics: Learning in graph-based atomic systems where dynamics depend on nonlocal interactions and energy-based cost functionals. + +Contributions. Our main contributions are: + +1. We introduce Differential RL, a reinforcement learning framework that replaces cumulative reward maximization with trajectory operator learning, naturally embedding physics-informed priors via a differential dual formulation. +2. We propose Differential Policy Optimization (dfPO), a policy optimization algorithm with rigorous pointwise convergence guarantees and a regret bound of $\mathcal{O}(K^{5/6})$ , enabling effective learning in low-data, physics-constrained environments. +3. We validate dfPO across diverse scientific tasks, demonstrating strong performance over several standard RL baselines. + +Organization. In Section 2, we introduce the new framework of differential reinforcement learning and the associated algorithm called dfPO. In Section 3, we outline the theoretical pointwise convergence theorem and regret analysis for the dfPO algorithm with explicit details given in the Appendix. In Section 4, we apply the dfPO algorithm to three representative scientific-computing tasks, and show competitive performance against popular RL benchmarks. + +# 2 Differential reinforcement learning + +# 2.1 Problem formulation + +In standard reinforcement learning, an agent operates in a Markov Decision Process (MDP) defined by the 5-tuple $(S, \mathcal{A}, \mathbb{P}, r, \rho_0)$ , where $S$ and $\mathcal{A}$ are the set of states and actions respectively, $\mathbb{P}: S \times \mathcal{A} \times S \to \mathbb{R}$ is the transition probability distribution, $r: S \times \mathcal{A} \to \mathbb{R}$ is the reward function, $\rho_0: S \to \mathbb{R}$ is the distribution of initial state. At step $k$ , the agent chooses an action $a_k \in \mathcal{A}$ given its current state $s_k \in S$ based on $\pi(a_k | s_k)$ : + +$$ +s _ {0} \sim \rho_ {0} (s _ {0}), a _ {k} \sim \pi \left(a _ {k} \mid s _ {k}\right), s _ {k + 1} \sim \mathbb {P} \left(s _ {k + 1} \mid s _ {k}, a _ {k}\right) \tag {1} +$$ + +The goal is to maximize the expected cumulative reward $\mathcal{J} = \mathbb{E}_{\pi}\left[\sum_{k=0}^{H-1} r(s_k, a_k)\right]$ . For a given policy $\pi$ , the associated value function is defined as: + +$$ +V _ {\pi} (s) = \mathbb {E} _ {a, s _ {1},..} \left[ \sum_ {k = 0} ^ {H - 1} r \left(s _ {k}, a _ {k}\right) | s _ {0} = s \right] \tag {2} +$$ + +The (optimal) value function is then defined as $V(s) \coloneqq \arg \max_{\pi} V_{\pi}(s)$ . Many reinforcement learning algorithms revolve around estimating and improving this value function. However, instead of remaining in this discrete-time formulation, we shift to a continuous-time viewpoint. By associating each discrete step $k$ with a timestamp $t_k = k \Delta_t$ and setting the terminal time $T = t_H = H \Delta_t$ , we approximate the discrete sum with a time integral: + +$$ +\max _ {\pi} \mathbb {E} \left[ \sum_ {k = 0} ^ {H - 1} r \left(s _ {k}, a _ {k}\right) \right] = \max _ {\pi} \mathbb {E} \left[ \sum_ {k = 0} ^ {H - 1} r \left(s _ {t _ {k}}, a _ {t _ {k}}\right) \right] \approx \max _ {\pi} \mathbb {E} \left[ \int_ {0} ^ {T} r \left(s _ {t}, a _ {t}\right) d t \right] \tag {3} +$$ + +This leads to the control-theoretic objective: + +$$ +\max _ {\pi} \mathbb {E} \left[ \int_ {0} ^ {T} r \left(s _ {t}, a _ {t}\right) d t \right] \text {s u b j e c t} \dot {s} _ {t} = f \left(s _ {t}, a _ {t}\right) \tag {4} +$$ + +where $f$ denotes the transition dynamics. Instead of directly solving this constrained optimization, we invoke Pontryagin's Maximum Principle [17], which introduces a dual formulation analogous to the Hamiltonian framework in classical mechanics. We augment the system with an adjoint variable $p$ , and define the Hamiltonian function $\mathcal{H}$ through the Legendre transform: + +$$ +\mathcal {H} (s, p, a) := p ^ {T} f (s, a) - r (s, a) \tag {5} +$$ + +Let $a^*(s, p)$ denotes the optimal action as a function of state and adjoint variables. Substituting this back gives the reduced Hamiltonian function $h$ : + +$$ +h (s, p) := \mathcal {H} (s, p, a ^ {*} (s, p)) \tag {6} +$$ + +The resulting differential dual system imposes the following constraints on the trajectory: + +$$ +\left[ \begin{array}{l} \dot {s} \\ \dot {p} \end{array} \right] = \left[ \begin{array}{l} \frac {\partial \hbar}{\partial p} (s, p) \\ - \frac {\partial \hbar}{\partial s} (s, p) \end{array} \right] \text {s u b j e c t} h (s, p) = \mathcal {H} (s, p, a ^ {*}) \text {w i t h} \frac {\partial \mathcal {H}}{\partial a} (s, p, a ^ {*}) = 0 \tag {7} +$$ + +The stationarity condition $\frac{\partial\mathcal{H}}{\partial a} (s,p,a^{*}) = 0$ ensures that the optimal action can be implicitly represented by the pair $(s,p)$ , allowing us to reformulate the optimal path solely in terms of these dual variables. This condition effectively decouples the explicit dependency on actions by encoding them through the adjoint variable $p$ . In this setting, the canonical state-action pair $(s,a)$ is replaced by the extended state $(s,p)$ , with the action recovered as a function $a = P(s,p)$ that solves the stationarity condition. Substituting this back yields the reduced Hamiltonian $h(s,p) = p^{\top}f(s,P(s,p)) - r(s,P(s,p))$ . Here, the influence of the reward function $r$ is now captured through the reduced Hamiltonian. We couple state and adjoint variables into the composite vector $x = (s,p)$ with dimension $d = d_{S} + d_{A}$ (sum of state and action space's dimensions). The differential dual system can now be written as: + +$$ +\dot {x} = S \nabla h (x) \tag {8} +$$ + +where $S$ is the canonical symplectic matrix $\left[ \begin{array}{cc}0 & I\\ -I & 0 \end{array} \right]$ . This formulation encodes the evolution of the system through a Hamiltonian gradient flow in phase space, which serves as the foundation for our policy learning formulation. By discretizing the differential system, we arrive at the update rule: + +$$ +x _ {n + 1} = x _ {n} + \Delta_ {t} S \nabla h (x _ {n}) := G (x _ {n}) \tag {9} +$$ + +where $\Delta_t$ is the discretization time step. $G$ is the dynamics operator dictating the evolution of the policy-induced trajectory. From a learning perspective, we aim to discover an operator $G$ such that successive applications generate the optimal trajectory $x, G(x), G^{(2)}(x), \dots$ , where $G^{(k)}$ denotes the $k$ -fold composition of $G$ : $G^{(k)}(x) = G(G(\dots G(x)\dots))$ . + +To approximate $h(x)$ , we introduce a learnable score function $g(x) \approx h(x)$ , which plays the role of a surrogate reward landscape defined over the extended space $(x = (s, p))$ . This reparameterization allows us to shift the learning objective toward trajectory-consistent updates. Altogether, this approximation process suggests that the original reinforcement learning problem, when viewed through the lens of continuous-time optimal control and its differential dual, can be reformulated as + +an abstract policy optimization problem, denoted $\mathcal{D}$ . The goal in $\mathcal{D}$ is to learn the optimal dynamics operator $G: \Omega \to \Omega$ that governs the optimal trajectory: + +$$ +x _ {0} = x \sim \rho_ {0}, \quad x _ {1} = G \left(x _ {0}\right) = G (x), \tag {10} +$$ + +$$ +x _ {2} = G \left(x _ {1}\right) = G ^ {(2)} (x), \dots , x _ {H - 1} = G ^ {(H - 1)} (x) \tag {11} +$$ + +Here $\Omega$ is a compact domain in $\mathbb{R}^d$ , and $H$ is the number of steps in an episode. Moreover, $\rho_0$ is the distribution of the starting point $x_0$ . In this work, we assume that $\rho_0$ is a continuous function. Performing an interaction with an environment $\mathcal{B}$ (see Definition 2.1), we learn a policy $G_{\theta}$ parameterized by $\theta$ that approximates $G$ . + +Definition 2.1. An environment $\mathcal{B}$ with respect to an adversarial distribution $\rho_0$ and a score function $g$ is a black-box system that allows you to check the quality of a policy $G_{\theta}$ . More concretely, $\mathcal{B}$ inputs a policy $G_{\theta}$ and outputs the trajectories $(G_{\theta}^{(k)}(x))_{k=0}^{H-1}$ , and their associated scores $(g(G_{\theta}^{(k)}(x)))_{k=0}^{H-1}$ for a sample $x$ from distribution $\rho_0$ . + +The function $g$ serves as a score surrogate that allows us to evaluate and update the current policy $G_{\theta}$ . From the differential dual formulation (Equation (9)), we obtain a first-order relationship: + +$$ +G = \operatorname {I d} + \Delta_ {t} S \nabla g, \tag {12} +$$ + +where $\mathrm{Id}$ is the identity operator, $\Delta_t$ is again the time step, and $S$ is the symplectic matrix. This formulation implicitly encodes a physics prior through the symplectic form, while still allowing flexibility for data-driven learning. As such, even though the original RL problem does not explicitly enforce physical constraints, the differential structure induces an implicit bias toward trajectory-consistent behavior, making it applicable to physics-based dynamical systems. + +# 2.2 Differential policy optimization (dfPO) algorithm + +Differential policy optimization (dfPO) (see Algorithm 1) is a "time-expansion (Dijkstra-like)" algorithm that iteratively uses appropriate on-policy samples for policy updates to ensure an increase in policy quality over each iteration. This algorithm has a similar idea to the trust region policy optimization (TRPO) [25]. However, because of our differential approach that focuses on pointwise estimates, dfPO becomes much simpler and easier to implement/optimize in practice, compared to TRPO and other RL counterparts. + +# 2.3 Application to scientific computing + +We demonstrate how Differential RL naturally applies to a broad class of scientific-computing problems by instantiating the abstract formulation $\mathcal{D}$ in three representative domains with energy-based objectives. Such objectives can either be potential-based reward structure of the form $r(s,a) = -\mathcal{F}(s)$ or an energy-regularized variant $r(s,a) = \frac{1}{2}\|a\|^2 - \mathcal{F}(s)$ . Here, $\mathcal{F}(s)$ denotes a task-specific potential or cost functional. + +The agent's objective is to reach low-energy states while minimizing control effort. Although the system dynamics can be simplified, the complexity of the task is fully encapsulated in $\mathcal{F}$ . In many scientific settings, $\mathcal{F}$ is only accessible via simulation, lacks a closed-form expression, and cannot be queried or differentiated directly. This renders model-based RL methods—relying on explicit reward access or gradients—inapplicable. Differential RL circumvents this limitation by relying solely on scalar evaluations along actual trajectories, similar to model-free RL. However, unlike typical model-free methods, it embeds a physics-informed inductive bias through its differential structure, making it particularly suited for scientific problems. The three representative domains include: + +Surface Modeling: This setting involves evolving surfaces optimized for geometric or physical properties. The surface is parameterized by control points (e.g., knots in a spline), and the reward is derived from physical objectives such as smoothness, curvature, or structural stress. The state $s$ encodes the control point positions, and the action $a_{k}$ updates them according to $s_{k + 1} = s_k + \Delta_t a_k$ . The cost $\mathcal{F}(s)$ evaluates the reconstructed surface $S(s)$ , with rewards of the form $r(s,a) = -\mathcal{F}(s)$ or $r(s,a) = \frac{1}{2}\|a\|^2 - \mathcal{F}(s)$ to penalize excessive updates. + +Grid-based Modeling: In many PDE-constrained problems, control is applied on a coarse spatial grid, while evaluation occurs on a refined grid. The state $s$ consists of coarse-grid values, and + +# Algorithm 1 (Main algorithm) dfPO for a generic environment $\mathcal{B}$ + +Input: a generic environment $\mathcal{B}$ , the number of steps per episode $H$ , time step $\Delta_t$ , and the number of samples $N_k$ at stage $k$ with $k \in \overline{1, H-1}$ . Here $N_k$ can be chosen based on Theorem 3.2. We also assume that the hypothesis space for the policy approximator $G_{\theta_k}$ in stage $k$ is $\mathcal{H}_k$ for $k \in \overline{1, H}$ . Output: a neural network approximation $G_\theta$ that approximates the optimal policy $G$ . + +1: Initialize an empty replay memory queue $\mathcal{M}$ . +2: Initialize $k = 1$ as the current stage and a random scoring function $g_{\theta_0}$ . Set the initial policy $G_{\theta_0} = \mathrm{Id} + \Delta_t S \nabla g_{\theta_0}$ via automatic differentiation. +3: repeat +4: Use $N_{k}$ starting points $\{X^i\}_{i=1}^{N_k}$ and previous policy $G_{\theta_{k-1}}$ to query $\mathcal{B}$ to get $N_k$ sample trajectories $\left\{G_{\theta_{k-1}}^{(j)}(X^i)\right\}_{j=0}^{H-1}$ together with their scores $\left\{g(G_{\theta_{k-1}}^{(j)}(X^i))\right\}_{j=0}^{H-1}$ for $i \in \overline{1, N_k}$ . +5: Add the labeled samples of the form $(x,y) = (G_{\theta_{k - 1}}^{(k - 1)}(X^i),g(G_{\theta_{k - 1}}^{(k - 1)}))$ to $\mathcal{M}$ . Also add labeled samples $(x,y) = (G_{\theta_{k - 1}}^{(j)}(X^i),g_{\theta_{k - 1}}(G_{\theta_{k - 1}}^{(j)}(X^i)))$ for $j\in \overline{1,k - 2}$ and $i\in \overline{1,N_k}$ to $\mathcal{M}$ . The latter addition step is to ensure that the new policy doesn't deviate from the previous policy on samples on which the previous policy already performs well. +6: Train the neural network $g_{\theta_k} \in \mathcal{H}_k$ at stage $k$ using labeled sample from $\mathcal{M}$ with smooth $L^1$ loss function [12]. +7: Set $G_{\theta_k} = \mathrm{Id} + \Delta_t S\nabla g_{\theta_k}$ via automatic differentiation. Update $k \to k + 1$ . +8: until $k \geq H$ +9: Output $G_{\theta_{H-1}} \coloneqq \mathrm{Id} + \Delta_t S \nabla g_{\theta_{H-1}}$ via automatic differentiation. + +actions modifying them. The reward $\mathcal{F}(s)$ is implicitly defined via a fine-grid reconstruction $s_1(s)$ : $\mathcal{F}(s) = \mathcal{U}(s_1(s))$ for evaluation $\mathcal{U}$ on finer grid. + +Molecular Dynamics: State $s_t$ encodes atomic coordinates in a fixed molecular graph, with actions as vertex displacements. The energy cost $\mathcal{F}(s)$ reflects atomic interactions over edges $E$ , and the objective is to reach low-energy, physically plausible configurations via $r(s, a) = -\mathcal{F}(s)$ or variants. + +To analyze the dual dynamics in Equation (7), we consider the regularized reward form $r(s,a) = \frac{1}{2}\|a\|^2 - \mathcal{F}(s)$ used across the three scientific-computing settings above. In this case, the stationarity condition $\frac{\partial\mathcal{H}}{\partial a}(s,p,a^*) = 0$ implies that $p = \frac{\partial r}{\partial a}(s,a^*) = a^*$ , establishing a one-to-one correspondence between the adjoint variable and the action. Substituting back, the reduced Hamiltonian becomes $h(s,p) = \frac{1}{2}\|p\|^2 - r(s,p)$ , showing that the dual's central term is essentially a regularized version of the original reward. Hence the score function $g(s,p)$ can be defined as $\frac{1}{2}\|p\|^2 - r(s,p)$ . + +# 3 Theoretical analysis + +This section shows the convergence of differential policy optimization (dfPO, Algorithm 1) based on generalization pointwise estimates. We then use this result to derive regret bounds for dfPO. + +# 3.1 Pointwise convergence and sample complexity + +Definition 3.1 below defines the number of training samples needed to allow derivative approximation transfer. This definition is used to derive the number of samples needed for Algorithm 1. + +Definition 3.1. Recall that $\rho_0$ is a continuous density for the starting states. Suppose that we are given a function $g:\Omega \to \mathbb{R}$ , a hypothesis space $\mathcal{H}$ consists of the function $h\in \mathcal{H}$ that approximates $g$ , two positive constants $\epsilon$ and $\delta$ . We define the function $N(g,\mathcal{H},\epsilon ,\delta)$ to be the number of samples needed so that if we approximate $g$ by $h\in \mathcal{H}$ via $N(g,\mathcal{H},\epsilon ,\delta)$ training samples, then with probability of at least $1 - \delta$ , we have the following estimate bound on two function gradients: + +$$ +\left\| \nabla g (X) - \nabla h (X) \right\| < \epsilon \tag {13} +$$ + +In other words, we want $N(g,\mathcal{H},\epsilon ,\delta)$ to be large enough so that the original approximation can transfer to the derivative approximation above. If no such $N(g,\mathcal{H},\epsilon ,\delta)$ exists, let $N(g,\mathcal{H},\epsilon ,\delta) = \infty$ . + +Pointwise convergence. We now state the main theorem of pointwise convergence for dfPO below. + +Theorem 3.2. Suppose that we are given a threshold error $\epsilon$ , a probability threshold $\delta$ , and a number of steps per episode $H$ . Assume that $\{N_k\}_{k=1}^{H-1}$ is the sequence of numbers of samples used at each stage in Algorithm 1 (dfPO) so that: + +$$ +N _ {1} = N (g, \mathcal {H} _ {1}, \epsilon , \delta), \tag {14} +$$ + +$$ +N _ {k} = \max \left\{N \left(g _ {\theta_ {k - 1}}, \mathcal {H} _ {k}, \epsilon , \delta_ {k - 1} / (k - 1)\right), N \left(g, \mathcal {H} _ {k}, \epsilon , \delta_ {k - 1} / (k - 1)\right) \right\} f o r k \in \overline {{2 , H - 1}} \tag {15} +$$ + +Here $\delta_k = \delta / 3^{H - k} = 3\delta_{k-1}$ . We further assume that there exists a Lipschitz constant $L > 0$ such that both the true dynamics $G$ and the policy neural network approximator $G_{\theta_k}$ at step $k$ with regularized parameters have their Lipschitz constant at most $L$ for each $k \in \overline{1, H}$ . Then, for a general starting point $X$ , with probability at least $1 - \delta$ , the following generalization bound for the trained policy $G_{\theta_k}$ holds for all $k \in \{1, 2, \dots, H - 1\}$ : + +$$ +\mathbb {E} _ {X} \left\| G _ {\theta_ {k}} ^ {(j)} (X) - G ^ {(j)} (X) \right\| < \frac {j L ^ {j} \epsilon}{L - 1} \text {f o r a l l} 1 \leq j \leq k \tag {16} +$$ + +Note that when $N_{k}\to \infty$ , the errors approach 0 uniformly for all $j$ given a finite terminal time $T$ . + +Proof. The key idea is to prove a stronger statement by induction over the stage number $k$ : with probability of at least $1 - \delta_k$ , + +$$ +\mathbb {E} _ {X} \left\| G _ {\theta_ {k}} ^ {(j)} (X) - G ^ {(j)} (X) \right\| < \epsilon_ {j} \text {f o r a l l} 1 \leq j \leq k \tag {17} +$$ + +Here $\epsilon_{k}$ and $\alpha_{k}$ are sequences defined in Lemma B.1 (Appendix). The inductive step relies on bounding the error using the following decomposition with 3 components: + +$$ +\begin{array}{l} \mathbb {E} _ {X} \left\| G _ {\theta_ {k + 1}} ^ {(k + 1)} (X) - G ^ {(k + 1)} (X) \right\| \leq \mathbb {E} _ {X} \left\| G _ {\theta_ {k + 1}} \left(G _ {\theta_ {k + 1}} ^ {(k)} (X)\right) - G _ {\theta_ {k + 1}} \left(G _ {\theta_ {k}} ^ {(k)} (X)\right) \right\| \\ + \mathbb {E} _ {X} \| G _ {\theta_ {k + 1}} \left(G _ {\theta_ {k}} ^ {(k)} (X)\right) - G \left(G _ {\theta_ {k}} ^ {(k)} (X)\right) \| + \mathbb {E} _ {X} \| G \left(G _ {\theta_ {k}} ^ {(k)} (X)\right) - G \left(G ^ {(k)} (X)\right) \| \\ \leq L \alpha_ {k} + \epsilon + L \epsilon_ {k} = \epsilon_ {k + 1} \tag {18} \\ \end{array} +$$ + +This combines the Lipschitz continuity of $G$ , the supervised approximation error, and the inductive hypothesis. A complete and formal proof is provided in the Appendix (Section B). + +Sample complexity. The general pointwise estimates for dfPO algorithm in Theorem 3.2 allow us to explicitly state the number of training episodes required for two scenarios considered in this work: one work with general neural network approximators and the other with more restricted (weakly convex and linearly bounded) difference functions: + +Corollary 3.3. In Algorithm 1 (dfPO), suppose we are given fixed step size and fixed number of steps per episode $H$ . Further assume that for all $k \in \overline{1, H - 1}$ , $\mathcal{H}_k$ is the same everywhere and is the hypothesis space $\mathcal{H}$ consisting of neural network approximators with bounded weights and biases. Then with the sequence of numbers of training episodes $N_k = \mathcal{O}(\epsilon^{-(2d + 4)})$ , the pointwise estimates Equation (16) hold. + +Corollary 3.4. Again, in Algorithm 1 (dfPO), suppose we are given fixed step size and fixed number of steps per episode $H$ . Suppose $\mathcal{H}_k$ is a special hypothesis subspace consisting of $h \in \mathcal{H}_k$ so that $h - g$ and $h - g_{\theta_{k-1}}$ are both $p$ -weakly convex and linearly bounded. Then with the sequence of numbers of training episodes $N_k = \mathcal{O}(\epsilon^{-6})$ , we obtain the pointwise estimates Equation (16). + +Definitions of weakly convex and linearly bounded, along with proofs of Corollary 3.3 and Corollary 3.4, are provided in the Appendix. The following corollary confirms dfPO's convergence. + +Corollary 3.5. Note that dynamics operator $G(x)$ has the form $x + \Delta_t F(x)$ , where $\Delta_t$ is the step size, and $F$ is a bounded function. In this case, even when the step size is infinitely small, Algorithm 1 (dfPO)'s training converges with reasonable numbers of training episodes. + +Proof. The Lipschitz constant $L$ of $G$ in this case is bounded by $1 + C\Delta_{t}$ for some constant $C > 0$ . By scaling, WLOG, assume that for finite-time horizon problem, the terminal time $T = 1$ so that the number of steps is $m = 1 / \Delta_t$ . Hence, for $N_{k} = \mathcal{O}(1 / \Delta_{t}^{2p})$ , $\epsilon = \mathcal{O}(\Delta_t^p)$ for $p > 2$ , the error bounds in Theorem 3.2 are upper-bounded by: + +$$ +\left\| G _ {\theta_ {k}} ^ {(k)} (X) - G ^ {(k)} (X) \right\| \leq \frac {k L ^ {k} \epsilon}{L - 1} \leq \frac {1}{\Delta_ {t}} \left(1 + \frac {C}{m}\right) ^ {m} \mathcal {O} \left(\Delta_ {t} ^ {p}\right) \frac {1}{\Delta_ {t}} \leq e ^ {C} \mathcal {O} \left(\Delta_ {t} ^ {p - 2}\right)\rightarrow 0 \tag {19} +$$ + +for $k\leq H - 1$ , as $\Delta_t\to 0$ + +# 3.2 Regret bound analysis + +Now we give a formal definition of the regret and then derive two regret bounds for dfPO algorithm (Algorithm 1). Suppose $K$ episodes are used during the training process and suppose a policy $\pi^k$ is applied at the beginning of the $k$ -th episode with the starting state $s^k$ (sampled from adversary distribution $\rho_0$ ) for $k \in \overline{1, K}$ . We focus on the total number of training episodes used and assume that the number of steps $H$ is fixed. The quantity Regret is then defined as the following function of the number of episodes $K$ : + +$$ +\operatorname {R e g r e t} (K) = \sum_ {k = 1} ^ {K} \left(V \left(s ^ {k}\right) - V _ {\pi^ {k}} \left(s ^ {k}\right)\right) \tag {20} +$$ + +We now derive an upper bound on $\mathbf{Regret}(K)$ defined in Equation (20): + +Corollary 3.6. Suppose that number of steps per episode $H$ is fixed and relatively small. If in Algorithm 1 (dfPO), the number of training samples $N_{k}$ have the scale of $\mathcal{O}(\epsilon^{-\mu})$ , regret bound for dfPO is upper-bounded by $\mathcal{O}(K^{(\mu -1) / \mu})$ . As a result, for the general case in Corollary 3.3, we obtain a regret bound of $\mathcal{O}(K^{(2d + 3) / (2d + 4)})$ . For the special cases with restricted hypothesis space in Corollary 3.4 the regret bound is $\mathcal{O}(K^{5 / 6})$ . + +Proof. For a fixed $H$ , Equation (16) gives us the estimate $\mathbb{E}_X[G_{\theta_k}^{(j)}(X) - G^{(j)}(X)] \approx \mathcal{O}(\epsilon)$ for the state-action pair $X$ at step $j$ between $(N_1 + \dots + N_{k-1} + 1)^{th}$ and $(N_1 + \dots + N_k)^{th}$ episodes during stage $k$ . Assuming a mild Lipschitz condition on reward function, the gap between the optimal reward and the reward obtained from the learned policy at step $j$ during these episodes is also approximately $\mathcal{O}(\epsilon)$ . Summing these uniform bounds over all $j$ and over all episodes gives: + +$$ +\operatorname {R e g r e t} (K) \leq H \left(N _ {1} + \dots + N _ {H - 1}\right) C \epsilon = C H K \epsilon \tag {21} +$$ + +Since $N_{k} = \mathcal{O}(\epsilon^{-\mu})$ , $K = H\mathcal{O}(\epsilon^{-\mu})$ . With a fixed $H$ , $\epsilon = \mathcal{O}(K^{-1 / \mu})$ . Hence, Equation (21) leads to $\operatorname{Regret}(K) \leq K\mathcal{O}(K^{-1 / \mu}) = \mathcal{O}(K^{(\mu - 1) / \mu})$ as desired. + +# 4 Experiments + +# 4.1 Evaluation tasks + +We evaluate Differential RL on three scientific computing tasks drawn from the domains introduced in Section 2.3. For each, we explicitly define the functional cost $\mathcal{F}(s)$ and provide the relevant mathematical details below. + +Surface Modeling. A representative surface modeling task arises in materials engineering [7], where raw materials (e.g., metals, plastics) are deformed into target configurations. Formally, an initial shape $\Gamma_0$ are deformed into the target shape $\Gamma^*$ through a shape-dependent cost functional $\mathcal{C}: S(\text{shape}) \to \mathbb{R}$ . The state $s$ encodes control points on the shape's 2D boundary, which are interpolated into a smooth curve $\Gamma(s)$ using cubic splines. The functional cost is then defined as $\mathcal{F}(s) := \mathcal{C}(\Gamma(s))$ , where $\mathcal{C}(\Gamma) := \frac{\int_{\mathbb{R}^2} |\delta1_\Gamma| dx}{\sqrt{\int_{\mathbb{R}^2} 1_\Gamma dx}}$ . Here, $\delta1_\Gamma$ denotes the distributional derivative of + +the indicator function $1_{\Gamma}$ . The action $a$ incrementally updates the boundary points, and the initial shape is sampled from $\rho_0$ , a distribution over random polygons. + +Grid-based Modeling. For this domain, control is applied on a coarse spatial grid while evaluation occurs on a finer grid. The state $s$ represents values of a function $f_{\mathrm{coarse}}$ on the coarse grid, and the action $a$ modifies these values. A bicubic spline [4] generates a fine-scale approximation $f_{\mathrm{finer}}$ from $s$ , and the cost is defined as: + +$$ +\mathcal {F} (s) := \mathcal {C} \left(f _ {\text {f i n e r}}\right) = \frac {\int_ {\text {g r i d}} \left| \delta f _ {\text {f i n e r}} \right| d x}{\sqrt {\int_ {\text {g r i d}} f _ {\text {f i n e r}} d x}} \tag {22} +$$ + +The initial coarse-grid configuration $f_{\mathrm{coarse}}$ is sampled from a uniform distribution. + +![](images/f83e19ad7ce42a1705210f39e7ef5babd0ceb8f4e27e7f68e56fde1954e66519.jpg) +(a) Surface modeling + +![](images/4fd7c79f47a79cf6651ea25af1767435f0f12eeb320a10b0f28ae4157c966cd2.jpg) +(b) Grid-based modeling +Figure 1: Evaluation costs over episodes for 13 algorithms on 3 scientific computing tasks. dfPO (red curves) consistently achieves lower costs with more optimal and physically aligned trajectories. + +![](images/ae585a0d0dfc8e6f1adb85f403b6f569dacb85f8d2791cba464b263a89832dae.jpg) +(c) Molecular dynamics + +Molecular Dynamics. This task aims to guide the octa-alanine molecule to a low-energy configuration [3]. The state $s$ consists of dihedral angles $(\phi_j,\psi_j)_j$ defining the molecular conformation. Given these angles, atomic coordinates $(x_{i})_{i = 1}^{N}$ are reconstructed via a deterministic mapping $X((\phi_j,\psi_j)_j)$ , based on molecular geometry. The energy functional is defined as $\mathcal{F}(s):= \mathcal{F}((\phi_j,\psi_j)_j) = U(X((\phi_j,\psi_j)_j)) = U((x_i)_{i = 1}^N)$ , where energy $U$ is computed via the PyRosetta package [9]. Action $a$ modifies the dihedral angle, while initial distribution $\rho_0$ is purposely chosen as a uniform distribution over small intervals to evaluate agents under limited exploration. + +Beyond these examples, our framework applies to a wide range of simulation-defined objectives. Selected tasks feature sufficiently complex functionals to effectively test the proposed method. More details about our given choices of these representative tasks are given in the Appendix (Section C.6). + +# 4.2 Experimental results + +Models. We compare dfPO (Algorithm 1) against 12 baselines: 6 standard reinforcement learning (RL) algorithms and 6 energy-reshaped variants. The RL algorithms include TRPO [25] and PPO [26], two trust-region methods, with PPO widely used in LLM training due to its scalability. DDPG [19] and SAC [13] are foundational algorithms for continuous control, while Cross-Q [5] and TQC [18] offer more recent improvements in sample efficiency. For benchmarking, we denote the standard algorithms with an "S-" prefix to distinguish them from their energy-reshaped counterparts (e.g., S-PPO vs. PPO). The standard versions use the straightforward negative energy reward $r(s,a) = -\mathcal{F}(s)$ , while the reshaped variants apply a time-dependent modified reward $r(s,a) = \beta^{-t}(\frac{1}{2}\|a\|^2 - \mathcal{F}(s))$ . All baselines are implemented based on the Stable-Baselines3 library [23]. + +Table 1: Final mean evaluation costs $(\mathcal{F}(s)$ at terminal step) for all algorithms across 3 tasks. Lower values indicate better performance and correspond to higher rewards. + +
Standard AlgorithmsReward-Shaping Variants
dfPOS-TRPOS-PPOS-SACS-DDPGS-CrossQS-TQCTRPOPPOSACDDPGCrossQTQC
Surface6.327.7419.178.899.546.936.516.4820.617.4115.926.426.67
Grid6.066.487.057.176.687.076.717.107.117.006.587.237.12
Mol. Dyn.53.341842.301842.30126.7382.95338.07231.981842.281842.311361.3168.20923.9076.87
+ +Train/test setup. All models are trained under limited-sample conditions. For the first two tasks, we use 100,000 sample steps; for the third task, training is restricted to 5,000 sample steps due to the high cost of reward evaluation. Each model is evaluated over 200 test episodes with a normalized time horizon [0, 1] (terminal time $T = 1$ ). Our model sizes are also relatively small compared to other approaches. Additional details on training samples, reward-shaping hyperparameters, and model sizes are provided in the Appendix. All experiments are conducted on an NVIDIA A100 GPU. + +Metrics. We evaluate models based on the cost functional $\mathcal{F}$ computed over test trajectories. The objective is to achieve the lowest possible $\mathcal{F}$ values while maintaining physically plausible trajectories. + +Results. As shown in Table 1, dfPO consistently outperforms all 12 baselines across 3 representative scientific computing tasks. No baseline (besides dfPO) dominates overall; CrossQ, TQC, DDPG, + +Table 2: Hyperparameter ablations on reward-shaping algorithms. + +
DatasetdfPOorigCrossQSACTQC
orignc=10nc=2origent=0.05ent=0.2orignc=10nc=5
Surface6.326.427.336.637.417.628.236.676.686.96
Grid6.067.237.437.537.006.977.197.127.157.29
Mol. Dyn.53.34923.901247.411287.991361.311367.501386.4276.8798.5684.36
DatasetdfPOorigDDPGPPOTRPO
orignoise=OUtau=0.01origclip=0.1norm=ForigGAE-λ=0.8
Surface6.3215.9215.2317.0320.6121.4019.766.4811.67
Grid6.066.586.946.887.117.117.287.107.19
Mol. Dyn.53.3468.2076.6274.701842.311842.291842.311842.281842.33
+ +and TRPO intermittently rank second or third, indicating varying strengths across domains. Notably, reward-shaped variants generally improve over their standard counterparts yet remain below dfPO. PPO underperforms across the board, likely due to its simplification of TRPO at the cost of reduced stability and weaker physics-aligned inductive bias. The evaluation-cost curves in Figure 1 show dfPO consistently exploring lower objective values with moderate variance. On the grid-based task, its advantage over baselines is clear; on surface modeling and molecular dynamics, trajectories are not always smooth but still converge to lower-energy states. dfPO's exploration pattern resembles TRPO's but attains better final values with more controlled variance. Meanwhile, SAC yields smooth curves yet fails to approach optimal values, likely due to bias from entropy regularization. + +Ablation study. We report hyperparameter ablations for TQC—number of critics $\mathbf{n}_{\mathrm{c}}$ and quantiles $\mathbf{n}_{\mathrm{q}}$ , CrossQ—number of critics $\mathbf{n}_{\mathrm{c}}$ , SAC—entropy coefficient ent, DDPG—action noise (Ornstein-Uhlenbeck) and target-update coefficient tau, PPO—clip coefficient and advantage normalization, and TRPO—GAE parameter $\lambda$ . dfPO uses defaults hyperparameters (learning rate 0.001, batch size 32). Reward-shaping ablation results are reported in Table 2; ablations for the standard algorithms appear in Table 5 (see Section C.3). Overall, hyperparameter variations do not substantially affect relative performance rankings, and dfPO remains robust. + +Computational complexity. To analyze Algorithm 1, we focus on the main computational bottleneck: Step 6. In this step, the number of training updates for $g_{\theta_k}$ is proportional to the number of newly added samples to the memory buffer $M$ . As shown in Corollary 3.4, this number scales as $kN_k \sim k \cdot \mathcal{O}(\epsilon^{-6})$ for each $k \in \{1, \dots, H - 1\}$ , where $\epsilon$ denotes the target error threshold. Therefore, the overall time complexity is $\sum_{k=1}^{H-1} k\mathcal{O}(\epsilon^{-6}) = \mathcal{O}(H^2\epsilon^{-6})$ . + +Implementation link. The complete codebase is available at https://github.com/mpnguyen2/dfPO. + +# 5 Related works + +Continuous-time reinforcement learning. While most reinforcement learning (RL) methods are formulated using Markov decision processes, control theory offers a natural continuous-time alternative [11]. Early work [30] formalized RL with a continuous-time formulation grounded in stochastic differential equations (SDEs), replacing cumulative rewards with time integrals and modeling dynamics via continuous-time Markov processes. Several subsequent works, including ours, build on this control-theoretic perspective. A related line of work proposes continuous-time policy gradient and actor-critic analogs without heavy probabilistic machinery [1, 32], but these methods also require pointwise access to rewards and their derivatives, limiting their applicability in scientific computing as discussed Section 1. Furthermore, extending SAC, TRPO, or PPO to continuous time is nontrivial: naive $Q$ -function definitions collapse to the value function, eliminating action dependence and breaking advantage-based updates. Recent theory [15, 33] addresses this by redefining the $Q$ -function as the limiting reward rate (expected reward per time) and linking it to the Hamiltonian (see Section 2), thereby enabling continuous-time TRPO and PPO counterparts [33]. + +Our work also builds on the control-theoretic formulation (simplified in Equation (4) with stochastic function $f$ ), but differs in two key aspects. First, we use the continuous-time formulation only as a means to derive the dual of RL: we move to continuous time mainly to construct the dual via + +PMP, and then discretize the dual. Second, we define the policy over the joint space of state and adjoint variables, treating it as an operator over this extended space. This allows us to capture localized updates more naturally. We conjecture that our " $g$ -function" (Section 2.1) aligns with the Hamiltonian-based $g$ -function in [15], and our model corresponds to an iterative procedure refining the continuous-time advantage function within the extended state-adjoint space. + +Regret bounds. In discrete settings, optimal $\mathcal{O}(\sqrt{K})$ regret is known (e.g., [8]), but the constants scale with the state-space size, which is intractable in continuous settings. In continuous domains, nontrivial guarantees typically require structural assumptions. Under the mild Lipschitz-MDP assumption, the minimax regret admits a lower bound $\Omega\left(K^{\frac{d+1}{d+2}}\right)$ [27], where $d$ is the joint state-action dimension. Faster rates arise with stronger smoothness: Maran et al. [21] assume $\nu$ -times differentiable rewards/transitions and obtain $\mathcal{O}\left(K^{\frac{3d/2+\nu+1}{d+2(\nu+1)}}\right)$ , which approaches $\mathcal{O}(\sqrt{K})$ as $\nu \to \infty$ ; Vakili and Olkhovskaya [29] assume kernelized rewards/transitions in an RKHS with Matérn kernel of order $m$ and show $\mathcal{O}\left(K^{\frac{d+m+1}{d+2m+1}}\right)$ , again tending to $\mathcal{O}(\sqrt{K})$ as $m \to \infty$ . Under comparable assumptions, our result achieves similar dimension-independent rates (see Corollary 3.6). + +Our bound is significant because it is derived from pointwise guarantees on the per-step policy error, rather than only bounding the total cumulative regret. For a fixed horizon $H$ , we show the expected policy error at each step $j$ across episode segments. Summing over steps yields the global regret (Equation (21)). These per-step guarantees are finer-grained: they show the learned policy is near-optimal at each timestep, mitigating issues like overfitting specific cumulative reward paths (e.g., reward hacking or physically inconsistent behavior). In this sense, pointwise bounds are stronger than bounding the total regret alone. + +# 6 Conclusion + +We propose Differential Reinforcement Learning (Differential RL), a framework that reinterprets reinforcement learning via the differential dual of continuous-time control. Unlike standard RL algorithms that rely on global value estimates, our framework offers fine-grained control updates aligned with the system's dynamics at each timestep. Differential RL also naturally introduces a Hamiltonian structure that embeds physics-informed priors, further supporting trajectory-level consistency. To implement this framework, we introduce Differential Policy Optimization (dfPO, Algorithm 1), a stage-wise algorithm that updates local movement operators along trajectories. Theoretically, we establish pointwise convergence guarantees, a property unavailable in conventional RL, and derive a regret bound of $\mathcal{O}(K^{5/6})$ . Empirically, dfPO consistently outperforms standard RL baselines across three representative scientific computing domains: surface modeling, multiscale grid control, and molecular dynamics. These tasks feature complex functional objectives, physical constraints, and data scarcity, conditions under which traditional methods often struggle. Future work includes extending this framework to broader domains, investigating adaptive discretization, and further bridging the gap between optimal control theory and modern RL. + +# Acknowledgments and Disclosure of Funding + +This research was supported in part by a grant from the Peter O'Donnell Foundation, the Michael J. Fox Foundation, Jim Holland-Backcountry Foundation to support AI in Parkinson, and in part from a grant from the Army Research Office accomplished under Cooperative Agreement Number W911NF-19-2-0333. + +# References + +[1] Samuel Ainsworth, Kendall Lowrey, John Thickstun, Zaid Harchaoui, and Siddhartha Srinivasa. Faster policy learning with continuous-time gradients. In Proceedings of the 3rd Conference on Learning for Dynamics and Control, volume 144 of Proceedings of Machine Learning Research, pages 1054-1067. PMLR, 07 - 08 June 2021. +[2] Chandrajit Bajaj and Minh Nguyen. Physics-informed neural networks via stochastic hamiltonian dynamics learning. In Intelligent Systems and Applications, pages 182-197. Springer Nature Switzerland, 2024. +[3] Chandrajit Bajaj, Minh Nguyen, and Conrad Li. Reinforcement learning for molecular dynamics optimization: A stochastic pontryagin maximum principle approach. In Neural Information Processing, pages 310-323, Singapore, 2025. Springer Nature Singapore. +[4] Chandrajit L Bajaj, Guo-Liang Xu, and Qin Zhang. Bio-molecule surfaces construction via a higher-order level-set method. J. Comput. Sci. Technol., 23(6):1026-1036, 2008. +[5] Aditya Bhatt, Daniel Palenicek, Boris Belousov, Max Argus, Artemij Amiranashvili, Thomas Brox, and Jan Peters. CrossQ: Batch normalization in deep reinforcement learning for greater sample efficiency and simplicity. In The Twelfth International Conference on Learning Representations, 2024. +[6] David Biagioni, Xiangyu Zhang, Dylan Wald, Deepthi Vaidhynathan, Rohit Chintala, Jennifer King, and Ahmed S. Zamzam. Powergridworld: a framework for multi-agent reinforcement learning in power systems. In Proceedings of the Thirteenth ACM International Conference on Future Energy Systems, e-Energy '22, page 565-570, New York, NY, USA, 2022. Association for Computing Machinery. +[7] Nathan K. Brown, Anthony P. Garland, Georges M. Fadel, and Gang Li. Deep reinforcement learning for the rapid on-demand design of mechanical metamaterials with targeted nonlinear deformation responses. Engineering Applications of Artificial Intelligence, 126:106998, 2023. +[8] Qi Cai, Zhuoran Yang, Chi Jin, and Zhaoran Wang. Provably efficient exploration in policy optimization. In Proceedings of the 37th International Conference on Machine Learning, volume 119 of Proceedings of Machine Learning Research, pages 1283-1294, Virtual, 13-18 Jul 2020. PMLR. +[9] Sidhartha Chaudhury, Sergey Lyskov, and Jeffrey J Gray. PyRosetta: a script-based interface for implementing molecular modeling algorithms using rosetta. Bioinformatics, 26(5):689-691, 2010. +[10] Marc Deisenroth and Carl E Rasmussen. Pilco: A model-based and data-efficient approach to policy search. In Proceedings of the 28th International Conference on machine learning (ICML-11), pages 465–472. CiteSeer, 2011. +[11] Wendell H Fleming and H Mete Soner. Controlled Markov processes and viscosity solutions. Springer, New York, NY, 2nd edition, 2006. +[12] Ross Girshick. Fast R-CNN. In International Conference on Computer Vision (ICCV), 2015. +[13] Tuomas Haarnoja, Aurick Zhou, Pieter Abbeel, and Sergey Levine. Soft actor-critic: Off-policy maximum entropy deep reinforcement learning with a stochastic actor. In Jennifer Dy and Andreas Krause, editors, Proceedings of the 35th International Conference on Machine Learning, volume 80 of Proceedings of Machine Learning Research, pages 1861–1870, Vienna, Austria, 10–15 Jul 2018. PMLR. +[14] Nicolas Heess, Greg Wayne, David Silver, Timothy Lillicrap, Yuval Tassa, and Tom Erez. Learning continuous control policies by stochastic value gradients. In Proceedings of the 28th International Conference on Neural Information Processing Systems - Volume 2, NIPS'15, page 2944-2952, Cambridge, MA, USA, 2015. MIT Press. +[15] Yanwei Jia and Xun Yu Zhou. q-learning in continuous time. Journal of Machine Learning Research, 24(161):1-61, 2023. + +[16] Dmitry Kalashnikov, Alex Irpan, Peter Pastor, Julian Ibarz, Alexander Herzog, Eric Jang, Deirdre Quillen, Ethan Holly, Mrinal Kalakrishnan, Vincent Vanhoucke, and Sergey Levine. Scalable deep reinforcement learning for vision-based robotic manipulation. In Aude Billard, Anca Dragan, Jan Peters, and Jun Morimoto, editors, Proceedings of The 2nd Conference on Robot Learning, volume 87 of Proceedings of Machine Learning Research, pages 651-673, Zurich, Switzerland, 29-31 Oct 2018. PMLR. +[17] Donald E Kirk. Optimal Control Theory: An Introduction. Prentice-Hall, London, England, 1971. +[18] Arsenii Kuznetsov, Pavel Shvechikov, Alexander Grishin, and Dmitry P. Vetrov. Controlling overestimation bias with truncated mixture of continuous distributional quantile critics. arXiv preprint arXiv:2005.04269, 2020. +[19] Timothy P Lillicrap, Jonathan J Hunt, Alexander Pritzel, Nicolas Heess, Tom Erez, Yuval Tassa, David Silver, and Daan Wierstra. Continuous control with deep reinforcement learning. In International Conference on Learning Representations (ICLR), 2016. +[20] Isaac D. Lutz, Shunzhi Wang, Christoffer Norn, Alexis Courbet, Andrew J. Borst, Yan Ting Zhao, Annie Dosey, Longxing Cao, Jinwei Xu, Elizabeth M. Leaf, Catherine Treichel, Patrisia Litvicov, Zhe Li, Alexander D. Goodson, Paula Rivera-Sanchez, Ana-Maria Bratovianu, Minkyung Baek, Neil P. King, Hannele Ruohola-Baker, and David Baker. Top-down design of protein architectures with reinforcement learning. Science, 380(6642):266-273, 2023. +[21] Davide Maran, Alberto Maria Metelli, Matteo Papini, and Marcello Restelli. Local linearity: the key for no-regret reinforcement learning in continuous mdps. In A. Globerson, L. Mackey, D. Belgrave, A. Fan, U. Paquet, J. Tomczak, and C. Zhang, editors, Advances in Neural Information Processing Systems, volume 37, pages 75986-76029. Curran Associates, Inc., 2024. +[22] Anusha Nagabandi, Gregory Kahn, Ronald S Fearing, and Sergey Levine. Neural network dynamics for model-based deep reinforcement learning with model-free fine-tuning. In 2018 IEEE International Conference on Robotics and Automation (ICRA), Brisbane, Australia, 2018. IEEE. +[23] Antonin Raffin, Ashley Hill, Adam Gleave, Anssi Kanervisto, Maximilian Ernestus, and Noah Dormann. Stable-baselines3: Reliable reinforcement learning implementations. Journal of Machine Learning Research, 22(268):1-8, 2021. +[24] Patrick Rebeschini. Algorithmic foundations of learning, 2022. URL https://www.stats. ox.ac.uk/%7Erebeschi/teaching/AFoL/22/. +[25] John Schulman, Sergey Levine, Pieter Abbeel, Michael Jordan, and Philipp Moritz. Trust region policy optimization. In Proceedings of the 32nd International Conference on Machine Learning, volume 37 of Proceedings of Machine Learning Research, pages 1889-1897, Lille, France, 07-09 Jul 2015. PMLR. +[26] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017. +[27] Aleksandrs Slivkins. Introduction to multi-armed bandits. arXiv preprint arXiv:1904.07272, 2024. +[28] Yuval Tassa, Tom Erez, and Emanuel Todorov. Synthesis and stabilization of complex behaviors through online trajectory optimization. In 2012 IEEE/RSJ International Conference on Intelligent Robots and Systems, pages 4906-4913, 2012. +[29] Sattar Vakili and Julia Olkhovskaya. Kernelized reinforcement learning with order optimal regret bounds. In A. Oh, T. Naumann, A. Globerson, K. Saenko, M. Hardt, and S. Levine, editors, Advances in Neural Information Processing Systems, volume 36, pages 4225-4247. Curran Associates, Inc., 2023. + +[30] Haoran Wang, Thaleia Zariphopoulou, and Xun Yu Zhou. Reinforcement learning in continuous time and space: A stochastic control approach. Journal of Machine Learning Research, 21 (198):1-34, 2020. +[31] Tingwu Wang, Xuchan Bao, Ignasi Clavera, Jerrick Hoang, Yeming Wen, Eric Langlois, Shunshi Zhang, Guodong Zhang, Pieter Abbeel, and Jimmy Ba. Benchmarking model-based reinforcement learning. arXiv preprint arXiv:1907.02057, 2019. +[32] Cagatay Yildiz, Markus Heinonen, and Harri Lähdesmäki. Continuous-time model-based reinforcement learning. In Proceedings of the 38th International Conference on Machine Learning, volume 139 of Proceedings of Machine Learning Research, pages 12009-12018. PMLR, 18-24 Jul 2021. +[33] Hanyang Zhao, Wenpin Tang, and David Yao. Policy optimization for continuous reinforcement learning. In A. Oh, T. Naumann, A. Globerson, K. Saenko, M. Hardt, and S. Levine, editors, Advances in Neural Information Processing Systems, volume 36, pages 13637-13663. Curran Associates, Inc., 2023. + +# A Basic algorithmic learning theory + +We first introduce the results of basic learning theory on independent and identically distributed (i.i.d) samples. The proofs for all lemmas in this section can be found in [24] and can also be found in the literature on modern learning theory. + +Notations: Throughout this section, $\mathcal{X}$ is the feature space, $\mathcal{Y}$ be the label space, $\mathcal{Z} = \mathcal{X} \times \mathcal{Y}$ . Let $\mathcal{H}$ be a hypothesis space consisting of hypothesis $h: \mathcal{X} \to \mathcal{Y}$ . Let $l: \mathcal{H} \times \mathcal{Z} \to \mathbb{R}_+$ be a loss function on labeled sample $z = (x,y)$ for $x \in \mathcal{X}$ and $y \in \mathcal{Y}$ . Our loss function will have the following particular form: $l(h,(x,y)) = \phi(h(x),y)$ for a given associated function $\phi: \mathcal{Y} \times \mathcal{Y} \to \mathbb{R}_+$ . We use capital letters to represent random variables. We also define the following sets: $\mathcal{H} \circ \{Z_1,\dots,Z_n\} := \{(h(X_1),\dots,h(X_n)), h \in \mathcal{H}\} \subseteq \mathbb{R}^n$ , and $\mathcal{L} \circ \{Z_1,\dots,Z_n\} := \{(l(h,Z_1),\dots,l(h,Z_n)), h \in \mathcal{H}\} \subseteq \mathbb{R}^n$ for i.i.d samples $Z_i = (X_i,Y_i) \in \mathcal{X} \times \mathcal{Y} = \mathcal{Z}$ with label $Y_i$ with $i \in \overline{1,n}$ . Also, define $e(h)$ to be the average loss over new test data $e(h) := \mathbb{E}_Z[l(h,Z)]$ , and $E(h)$ the empirical loss over $n \in \mathbb{Z}_+$ i.i.d. samples $Z_1,\dots,Z_n$ : $E(h) := \frac{1}{n}\sum_{i=1}^{n} l(h,Z_i)$ . + +Definition A.1. The Rademacher complexity of a set $\mathcal{T} \subseteq \mathbb{R}^n$ is defined as: + +$$ +\mathbf {R a d} (\mathcal {T}) = \mathbb {E} \left[ \sup _ {t \in \mathcal {T}} \frac {1}{n} \sum_ {i = 1} ^ {n} B _ {i} t _ {i} \right] \tag {23} +$$ + +for Bernoulli random variables $B_{i}\in \{-1,1\}$ + +Lemma A.2. Suppose $\phi (.,y)$ is $\gamma$ -Lipschitz for any $y\in \mathcal{V}$ for some $\gamma >0$ . Then: + +$$ +\begin{array}{l} \mathbb {E} \left[ \sup _ {h \in \mathcal {H}} \left\{e (h) - E (h) \right\} \right] \leq 2 \mathbb {E} \left[ \boldsymbol {R a d} \left(\mathcal {L} \circ \left\{Z _ {1}, \dots , Z _ {n} \right\}\right) \right] (24) \\ \leq 2 \gamma \mathbb {E} \left[ \boldsymbol {R a d} \left(\mathcal {H} \circ \left\{Z _ {1}, \dots , Z _ {n} \right\}\right) \right] (25) \\ \end{array} +$$ + +Lemma A.3. For a hypothesis space $\mathcal{H}$ , a loss function $l$ bounded in the interval $[0, c]$ , and for $n$ i.i.d labeled samples $Z_{1}, \dots, Z_{n}$ , with probability of at least $1 - \delta$ , the following bound holds: + +$$ +\sup _ {h \in \mathcal {H}} (e (h) - E (h)) < 4 \operatorname {R a d} (\mathcal {L} \circ \{Z _ {1}, \dots , Z _ {n} \}) + c \sqrt {\frac {2 \log (1 / \delta)}{n}} \tag {26} +$$ + +Lemma A.4. For hypothesis space $\mathcal{H}$ consisting of (regularized) neural network approximators with weights and biases bounded by a constant, there exists a certain constant $C_1, C_2 > 0$ so that for a set of $n$ random variables $Z_1, \dots, Z_n$ : + +$$ +\mathbf {R a d} \left(\mathcal {H} \circ \left\{Z _ {1}, \dots , Z _ {n} \right\}\right) \leq \frac {1}{\sqrt {n}} \left(C _ {1} + C _ {2} \sqrt {\log d}\right) \tag {27} +$$ + +Throughout this paper, we assume that the optimization error can be reduced to nearly 0, so that $E(h) \approx 0$ if the hypothesis space $\mathcal{H}$ contains the function to be learned. From Lemma A.3, the average estimation error generally scales with $c\sqrt{\frac{2\log(1 / \delta)}{n}}$ . + +# B Proofs of theorems and corollaries in Section 3 + +Proof of theorem. We first state the supporting Lemma B.1 and then use it to prove the main Theorem 3.2 in Section 3 regarding the pointwise estimates for dfPO algorithm. + +Lemma B.1. Given $L$ and $\epsilon > 0$ , define two sequences $\{\alpha_j\}_{j \geq 0}$ and $\{\epsilon_j\}_{j \geq 0}$ recursively as follows: + +$$ +\alpha_ {0} = 0 \text {a n d} \alpha_ {j} = L \alpha_ {j - 1} + \epsilon \tag {28} +$$ + +$$ +\epsilon_ {1} = \epsilon \text {a n d} \epsilon_ {k + 1} = L \alpha_ {k} + \epsilon + L \epsilon_ {k} \tag {29} +$$ + +Then for each $k$ , we get: + +$$ +\epsilon_ {k} \leq \frac {k L ^ {k} \epsilon}{L - 1} \tag {30} +$$ + +Proof. First, $\alpha_{j} = (L^{j - 1} + \dots +1)\epsilon$ . Hence, + +$$ +\begin{array}{l} \frac {\epsilon_ {k}}{L ^ {k}} \leq \frac {L (L ^ {k - 2} + \cdots + 1) + 1}{L ^ {k}} \epsilon + \frac {\epsilon_ {k - 1}}{L ^ {k - 1}} \\ = \frac {L ^ {k} - 1}{L ^ {k} (L - 1)} \epsilon + \frac {\epsilon_ {k - 1}}{L ^ {k - 1}} < \frac {\epsilon}{L - 1} + \frac {\epsilon_ {k - 1}}{L ^ {k - 1}} \\ \end{array} +$$ + +Hence, by simple induction, $\frac{\epsilon_k}{L^k} < \frac{(k - 1)\epsilon}{L - 1} +\frac{\epsilon_1}{L} < \frac{k\epsilon}{L - 1}$ . As a result, $\epsilon_{k} < \frac{kL^{k}\epsilon}{L - 1}$ as desired. + +![](images/b66873f0967fbffed51a51b13ee660e2a2d07ef1bc0f04b2616f2b7140ff5738.jpg) + +Now we're ready to give a full proof for the dfPO's pointwise convergence. + +Theorem 3.2. Suppose that we are given a threshold error $\epsilon$ , a probability threshold $\delta$ , and a number of steps per episode $H$ . Assume that $\{N_k\}_{k=1}^{H-1}$ is the sequence of numbers of samples used at each stage in Algorithm 1 (dfPO) so that: + +$$ +N _ {1} = N (g, \mathcal {H} _ {1}, \epsilon , \delta), \tag {14} +$$ + +$$ +N _ {k} = \max \left\{N \left(g _ {\theta_ {k - 1}}, \mathcal {H} _ {k}, \epsilon , \delta_ {k - 1} / (k - 1)\right), N \left(g, \mathcal {H} _ {k}, \epsilon , \delta_ {k - 1} / (k - 1)\right) \right\} f o r k \in \overline {{2 , H - 1}} \tag {15} +$$ + +Here $\delta_k = \delta / 3^{H - k} = 3\delta_{k-1}$ . We further assume that there exists a Lipschitz constant $L > 0$ such that both the true dynamics $G$ and the policy neural network approximator $G_{\theta_k}$ at step $k$ with regularized parameters have their Lipschitz constant at most $L$ for each $k \in \overline{1, H}$ . Then, for a general starting point $X$ , with probability at least $1 - \delta$ , the following generalization bound for the trained policy $G_{\theta_k}$ holds for all $k \in \{1, 2, \dots, H - 1\}$ : + +$$ +\mathbb {E} _ {X} \left\| G _ {\theta_ {k}} ^ {(j)} (X) - G ^ {(j)} (X) \right\| < \frac {j L ^ {j} \epsilon}{L - 1} \text {f o r a l l} 1 \leq j \leq k \tag {16} +$$ + +Note that when $N_{k}\to \infty$ , the errors approach 0 uniformly for all $j$ given a finite terminal time $T$ . + +Proof. Let $\alpha_{k}$ and $\epsilon_{k}$ be two sequences associated with Lipschitz constant $L$ and threshold error $\epsilon$ as in Lemma B.1. We prove the generalization bound statement by induction on the stage number $k$ that for probability of at least $1 - \delta_{k}$ , + +$$ +\mathbb {E} _ {X} \left\| G _ {\theta_ {k}} ^ {(j)} (X) - G ^ {(j)} (X) \right\| < \epsilon_ {j} \text {f o r a l l} 1 \leq j \leq k \tag {31} +$$ + +By Lemma B.1, proving this statement also proves Theorem 3.2. + +The bound for the base case $k = 1$ is due to the definition of $N_{1} = N(g, \mathcal{H}_{1}, \epsilon, \delta)$ that allows the approximation of $g$ by $g_{\theta_1}$ transfers to (a linear transformation of) their derivatives $G$ and $G_{\theta_1}$ with error threshold $\epsilon$ and probability threshold $\delta$ . Assume that the induction hypothesis is true for $k$ . We prove that for a starting (random variable) point $X$ , the following error holds with a probability of at least $1 - \delta_{k + 1} = 1 - 3\delta_k$ : + +$$ +\mathbb {E} _ {X} \left\| G _ {\theta_ {k + 1}} ^ {(j)} (X) - G ^ {(j)} (X) \right\| < \epsilon_ {j} \text {f o r a l l} j \leq k + 1 \tag {32} +$$ + +First, from induction hypothesis, with probability of at least $1 - \delta_{k}$ : + +$$ +\mathbb {E} _ {X} \| G _ {\theta_ {k}} ^ {(j)} (X) - G ^ {(j)} (X) \| < \epsilon_ {j} \text {f o r a l l} j \leq k \tag {33} +$$ + +In stage $k + 1$ , all previous stages' samples up to stage $k - 1$ is used for $G_{\theta_{k + 1}}$ . As a result, we can invoke the induction hypothesis on $k$ to yield the same error estimate for $G_{\theta_{k + 1}}$ on the first $k$ sample points with probability $1 - \delta_k$ : + +$$ +\mathbb {E} _ {X} \left\| G _ {\theta_ {k + 1}} ^ {(j)} (X) - G ^ {(j)} (X) \right\| < \epsilon_ {j} \text {f o r a l l} j \leq k \tag {34} +$$ + +Recall from Algorithm 1 that $g_{\theta_{k+1}}$ is trained to approximate $g_{\theta_k}$ to ensure that the updated policy doesn't deviate too much from current policy. For $j \in \{1, \dots, k-1\}$ , $g_{\theta_{k+1}} \in \mathcal{H}_{k+1}$ approximates $g_{\theta_k}$ on $N_{k+1}$ samples of the form $G_{\theta_k}^{(j)}(X^i)$ for $i \in \{1, \dots, N_{k+1}\}$ . Since $N_{k+1} \geq N(g_{\theta_k}, \mathcal{H}_{k+1}, \epsilon, \delta_k / k)$ allows derivative approximation transfer, for probability of at least + +$1 - \delta_{k} / k, \mathbb{E}_{X}\| G_{\theta_{k + 1}}(G_{\theta_{k}}^{(j)}(X)) - G_{\theta_{k}}(G_{\theta_{k}}^{(j)}(X))\| < \epsilon.$ Hence, under a probability subspace $\Gamma$ with probability of at least $(1 - (k - 1)\delta_k / k)$ , we have: + +$$ +\mathbb {E} _ {X} \left\| G _ {\theta_ {k + 1}} \left(G _ {\theta_ {k}} ^ {(j)} (X)\right) - G _ {\theta_ {k}} \left(G _ {\theta_ {k}} ^ {(j)} (X)\right) \right\| < \epsilon \tag {35} +$$ + +for all $1 \leq j < k$ + +We prove by induction on $j$ that under this probability subspace $\Gamma$ , we have: + +$$ +\mathbb {E} _ {X} \left\| G _ {\theta_ {k + 1}} ^ {(j)} (X) - G _ {\theta_ {k}} ^ {(j)} (X) \right\| < \alpha_ {j} \text {f o r a l l} 1 \leq j \leq k \tag {36} +$$ + +In fact, for the induction step, one get: + +$$ +\begin{array}{l} \mathbb {E} _ {X} \left\| G _ {\theta_ {k + 1}} ^ {(j)} (X) - G _ {\theta_ {k}} ^ {(j)} (X) \right\| \leq \mathbb {E} _ {X} \left\| G _ {\theta_ {k + 1}} \left(G _ {\theta_ {k + 1}} ^ {(j - 1)} (X)\right) - G _ {\theta_ {k + 1}} \left(G _ {\theta_ {k}} ^ {(j - 1)} (X)\right) \right\| \\ + \mathbb {E} _ {X} \| G _ {\theta_ {k + 1}} \left(G _ {\theta_ {k}} ^ {(j - 1)} (X)\right) - G _ {\theta_ {k}} \left(G _ {\theta_ {k}} ^ {(j - 1)} (X)\right) \| \\ \leq L \mathbb {E} _ {X} \| G _ {\theta_ {k + 1}} ^ {(j - 1)} (X) - G _ {\theta_ {k}} ^ {(j - 1)} (X) \| + \epsilon \leq L \alpha_ {j - 1} + \epsilon = \alpha_ {j} \tag {37} \\ \end{array} +$$ + +Finally, we look at the approximation of $g$ by $g_{\theta_{k+1}} \in \mathcal{H}_{k+1}$ on the specific sample points $\left\{ G_{\theta_k}^{(k)}(X^i) \right\}_{i=1}^{N_{k+1}}$ . Definition of $N_{k+1} \geq N(g, \mathcal{H}_{k+1}, \epsilon, \delta_k / k)$ again allows derivative approximation transfer so that with probability at least $1 - \delta_k / k$ : + +$$ +\mathbb {E} _ {X} \left\| G _ {\theta_ {k + 1}} \left(G _ {\theta_ {k}} ^ {(k)} (X)\right) - G \left(G _ {\theta_ {k}} ^ {(k)} (X)\right) \right\| < \epsilon \tag {38} +$$ + +Now consider the probability subspace $\mathcal{S}$ under which 3 inequalities Equation (33), Equation (36), and Equation (38) hold. The subspace $\mathcal{S}$ has the probability measure of at least $1 - (\delta_k + (k - 1)\delta_k / k + \delta_k / k) = 1 - 2\delta_k = 1 - (\delta_{k + 1} - \delta_k)$ , and under $\mathcal{S}$ , we have: + +$$ +\begin{array}{l} \mathbb {E} _ {X} \left\| G _ {\theta_ {k + 1}} ^ {(k + 1)} (X) - G ^ {(k + 1)} (X) \right\| \leq \mathbb {E} _ {X} \left\| G _ {\theta_ {k + 1}} \left(G _ {\theta_ {k + 1}} ^ {(k)} (X)\right) - G _ {\theta_ {k + 1}} \left(G _ {\theta_ {k}} ^ {(k)} (X)\right) \right\| \\ + \mathbb {E} _ {X} \| G _ {\theta_ {k + 1}} (G _ {\theta_ {k}} ^ {(k)} (X)) - G (G _ {\theta_ {k}} ^ {(k)} (X)) \| + \mathbb {E} _ {X} \| G (G _ {\theta_ {k}} ^ {(k)} (X)) - G (G ^ {(k)} (X)) \| \\ \leq L \mathbb {E} _ {X} \| G _ {\theta_ {k + 1}} ^ {(k)} (X) - G _ {\theta_ {k}} ^ {(k)} (X) \| + \epsilon + L \mathbb {E} _ {X} \| G _ {\theta_ {k}} ^ {(k)} (X) - G ^ {(k)} (X) \| \leq L \alpha_ {k} + \epsilon + L \epsilon_ {k} = \epsilon_ {k + 1} \tag {39} \\ \end{array} +$$ + +Merging this inequality with probability subspace where the inequality in Equation (34) holds leads to the estimate on the final step for stage $k + 1$ for the induction step. + +Proofs of corollaries. Lemma B.3 and Lemma B.7 below estimates $N(g, \mathcal{H}, \epsilon, \delta)$ (defined in Definition 3.1) in terms of the required threshold error $\epsilon$ . Such lemmas are then directly used to prove the two corollaries Corollary 3.3 and Corollary 3.4 in Section 3. + +Before proving Lemma B.3 and Lemma B.7, we need the following supporting lemma. + +Lemma B.2. Let $\mathcal{H}$ be the hypothesis space consisting of neural network approximators with bounded weights and biases. For each $s\in \overline{1,d}$ , let $\mathcal{H}_s = \{(\nabla h)_s,h\in \mathcal{H}\} = \left\{\frac{\partial h}{\partial x_s},h\in \mathcal{H}\right\}$ consists of $s^{th}$ components of the gradients of elements in $\mathcal{H}$ . Then the Rademacher complexity with respect to $\mathcal{H}_s$ on $n$ i.i.d random variables $Z_{1},\dots ,Z_{n}$ scales with $\mathcal{O}(1 / \sqrt{n})$ : + +$$ +\operatorname {R a d} \left(\mathcal {H} _ {s} \circ \left\{Z _ {1}, \dots , Z _ {n} \right\}\right) = \mathcal {O} (1 / \sqrt {n}) \quad \forall s \in \overline {{1 , d}} \tag {40} +$$ + +Proof. To get the Rademacher complexity bound on $\mathcal{H}_s$ , we use the chain rule to express each element of $\mathcal{H}_s$ as $h_1 \cdots h_R$ , where $R$ is the fixed number of layers in $\mathcal{H}$ 's neural network architecture. Here each $h_i$ can be expressed as the composition of Lipschitz (activation) functions and linear functions alternatively with bounded weights and biases. Those elements $h_i$ then form another neural network hypothesis space. By invoking Lemma A.4, we obtain a bound of order $\mathcal{O}(1/\sqrt{n})$ on individual $h_i$ 's. To connect these $h_i$ 's, we express the product $h_1 \cdots h_R$ as: + +$$ +\begin{array}{l} h _ {1} \dots h _ {R} = \prod_ {i = 1} ^ {R} \left(h _ {i} + D - D\right) = \sum_ {W \subseteq [ R ]} (- D) ^ {R - | W |} \prod_ {i \in W} \left(h _ {i} + D\right) \\ = \sum_ {W \subseteq [ R ]} (- D) ^ {R - | W |} \exp \left(\sum_ {i \in W} \log \left(h _ {i} + D\right)\right) \\ \end{array} +$$ + +Here $D$ is a constant large enough to make the log function well-defined and to make Lipschitz constant of the log function bounded above by another constant. Now we use the simple bounds on $\mathbf{Rad}(\mathcal{T} + \mathcal{T}')$ and $\mathbf{Rad}(f\circ \mathcal{T})$ for some sets $\mathcal{T}$ and $\mathcal{T}'$ and Lipschitz function $f$ to derive the Rademacher complexity bound of the same order $\mathcal{O}(1 / \sqrt{n})$ . + +Lemma B.3. Suppose that the hypothesis space $\mathcal{H}$ for approximating the target function $g:\Omega \subset \mathbb{R}^d\to \mathbb{R}^d$ consists of neural network approximators with bounded weights and biases. In addition, assume that the function $g$ and neural network approximators $h\in \mathcal{H}$ are continuously differentiable twice with bounded first and second derivatives by some constant $C$ . One way this assumption can be satisfied is to choose activation functions that are two times continuously differentiable. Then $N(g,\mathcal{H},\epsilon ,\delta)$ is the upper bound of $\mathcal{O}(\epsilon^{-(2d + 4)})$ , where we only ignore the quadratic factors of $\delta$ , the polynomial terms of $d$ , and other logarithmic terms. + +Proof. Suppose we're given $\epsilon > 0$ . Take $\epsilon_1 > 0$ so that $16C\epsilon_1 < \epsilon/(2d)$ , and $\epsilon_2 = 0.5$ . Now take $n \in \mathbb{N}$ large enough so that $C_1\sqrt{\frac{\log(1 / (\delta / (3d))}{n}} < \epsilon/(2d)$ for an appropriate constant $C_1$ . Then take $m \in \mathbb{N}$ large enough so that $(1 - c_1\epsilon_2\epsilon_1^d)^m < \delta/(3n)$ , where $c_1$ is an appropriate geometric constant (see the following paragraphs). Let $M = m + n$ . + +Train a neural network function $h \in \mathcal{H}$ to approximate the target function $g$ on $N$ samples, where $N$ is given by: + +$$ +N = \frac {\log ((6 M) / \delta)}{(C \epsilon_ {1} ^ {2} \delta / (6 M)) ^ {2}} \text {o r} \sqrt {\frac {\log (1 / (\delta / 6 M))}{N}} = C \epsilon_ {1} ^ {2} \frac {\delta}{6 M} = N (g, \mathcal {H}, \epsilon , \delta) \tag {41} +$$ + +Before going to the main proof, we dissect $N$ to obtain its asymptotic rate in terms of $\epsilon$ and $d$ . First of all $\epsilon_{1} = \mathcal{O}(\epsilon /2d)$ . Next, $n\approx \log (1 / \delta) / (\epsilon /(2d))^2$ , and $m\approx \log (3n / \delta)\epsilon_1^{-d}$ . Hence, ignoring logarithmic terms, polynomial terms in $d$ and the quadratic factor of $\delta$ , $M = m + n$ is approximately $\mathcal{O}(\epsilon^{-d})$ . As a result, $N\approx \epsilon^{-(2d + 4)}$ . + +Choose a set $S = \{X_1, \dots, X_m, Y_1, \dots, Y_n\}$ consisting of $M = m + n$ random samples of distribution $\rho_0$ that are independent of $g$ : $m$ samples $X_1, \dots, X_m$ and $n$ sample $Y_1, \dots, Y_n$ . + +We apply Lemma A.3 to $\mathcal{H}$ with i.i.d labeled samples $Z = (X,g(X))$ and loss function $l$ with the associated $\phi (y,\hat{y}) = |y - \hat{y}|$ , which is 1-Lipschitz for a fixed $\hat{y}$ . In this case, from Lemma A.3, for probability of at least $1 - \delta /(6M)$ , $\mathbb{E}_U[|h(U) - g(U)|] < C\epsilon_1^2\delta /(6M)$ for the random variable $U\in S$ . As a result, by Markov inequality, with probability at least $1 - \delta /(6M) - \delta /(6M) = 1 - \delta /(3M)$ , $|h(U) - g(U)| < C\epsilon_1^2$ for each $U\in \{X_1,\dots ,X_m,Y_1,\dots ,Y_n\}$ . Hence, there exists a probability subspace $\Gamma$ with probability at least $1 - \delta /(3M)*M = 1 - \delta /3$ so that $|h(U) - g(U)| < C\epsilon_1^2$ for all $U\in S$ . + +The probability that a particular sample (random variable) $X_{i}$ is in the hypercone with a conic angle difference of $\epsilon_{2}$ surrounding the direction $(\nabla h(Y) - \nabla g(Y))$ of the hyper-spherical $\epsilon_{1}$ -circular neighborhood of $Y$ is at least $c_{1}\epsilon_{2}\epsilon_{1}^{d}$ . Here a $\epsilon_{1}$ -circular neighborhood here include points with radius sizes between $\epsilon_{1}/2$ and $\epsilon_{1}$ . The probability that no $m$ samples is in this cone is at most $(1 - c_{1}\epsilon_{2}\epsilon_{1}^{d})^{m} < \delta/(3n)$ . As a result, on $\Gamma$ except a subspace with probability less than $\delta/(3n)$ , there exists $k$ (depends on both $Y$ and $X_{1},\dots ,X_{m}$ ) so that: + +$$ +\epsilon_ {1} / 2 < \| X _ {k} - Y \| < \epsilon_ {1} \tag {42} +$$ + +$$ +(\nabla h (Y) - \nabla g (Y)) \cdot \left(X _ {k} - Y\right) > \left(1 - \epsilon_ {2}\right) \| \nabla h (Y) - \nabla g (Y) \| \| X _ {k} - Y \| \tag {43} +$$ + +Second-order Taylor expansion for $f$ and $g$ at each $Y$ yields: + +$$ +h \left(X _ {k}\right) = h (Y) + \nabla h (Y) \left(X _ {k} - Y\right) + \frac {1}{2} \| X _ {k} - Y \| ^ {2} U _ {h} \left(X _ {k}, Y\right) \tag {44} +$$ + +$$ +g (X _ {k}) = g (Y) + \nabla g (Y) (X _ {k} - Y) + \frac {1}{2} \| X _ {k} - Y \| ^ {2} U _ {g} (X _ {k}, Y) \tag {45} +$$ + +where $U_{h}$ and $U_{g}$ are the second derivative terms of $h$ and $g$ in respectively, and are bounded by $C$ . As a result: + +$$ +\begin{array}{l} (1 - \epsilon_ {2}) \| \nabla h (Y) - \nabla g (Y) \| (\epsilon_ {1} / 2) \\ < (1 - \epsilon_ {2}) \| \nabla h (Y) - \nabla g (Y) \| \| X _ {k} - Y \| \\ < (\nabla h (Y) - \nabla g (Y)) \cdot (X _ {k} - Y) \\ < | h (Y) - g (Y) | + | h \left(X _ {k}\right) - g \left(X _ {k}\right) | + 2 C \| X _ {k} - Y \| ^ {2} \\ < 2 C \epsilon_ {1} ^ {2} + 2 C \| X _ {k} - Y \| ^ {2} < 4 C \epsilon_ {1} ^ {2} \\ \end{array} +$$ + +Thus, $\| \nabla h(Y) - \nabla g(Y) \| < 8C\epsilon_1 / (1 - \epsilon_2) = 16C\epsilon_1$ . Therefore, on the subspace $\Gamma_0$ with probability of at least $1 - \delta / 3 - n * (\delta / (3n)) = 1 - (2\delta) / 3$ , $\| \nabla h(Y) - \nabla g(Y) \| < 16C\epsilon_1$ for all samples $Y \in \{Y_1, \dots, Y_n\}$ . + +In order to prove that $\mathbb{E}_X\| \nabla h(X) - \nabla g(X)\| < \epsilon$ and thus finishing the proof for $N(g,\mathcal{H},\epsilon ,\delta) = \mathcal{O}(\epsilon^{-(2d + 4)})$ , we only need to prove bounds on individual components of $\| \nabla h(X) - \nabla g(X)\|$ : + +$$ +\mathbb {E} _ {X} \left\| (\nabla h (X) - \nabla g (X)) _ {s} \right\| < \epsilon / d +$$ + +where $x_{s}$ is the $s^{th}$ component of a vector $x\in \mathbb{R}^d$ + +To this end, by Lemma B.2, we have a bound of order $\mathcal{O}(1 / \sqrt{n})$ for the Rademacher complexity of the hypothesis space $\mathcal{H}_s$ for $s\in \overline{1,d}$ , i.e. $\mathbf{Rad}(\mathcal{H}_s\circ \{Z_1,\dots ,Z_n\}) = \mathcal{O}(1 / \sqrt{n})$ for i.i.d random variables $Z_{1},\dots ,Z_{n}$ . Hence, we can invoke Lemma A.3 on $\mathcal{H}_s$ to get: + +$$ +\begin{array}{l} \mathbb {E} _ {X} \left\| (\nabla h (X) - \nabla g (X)) _ {s} \right\| < \frac {1}{n} \sum_ {i = 1} ^ {n} \left\| (\nabla h (Y _ {i}) - \nabla g (Y _ {i})) _ {s} \right\| + \epsilon / (2 d) \\ < \epsilon / (2 d) + \epsilon / (2 d) = \epsilon / d \\ \end{array} +$$ + +on $\Gamma_0$ except a set of probability of at most $\delta /(3d)$ . Then we finish the proof of Lemma B.3 by summing all inequalities over $d$ components. + +Remark. Another simpler way to achieve a similar result is to upper-bound $\| \nabla h(X) - \nabla g(X) \|$ by $\| \nabla h(Y_i) - \nabla g(Y_i) \| + 2C \| X - Y_i \|$ and use the probability subspace in which one of the $Y_i$ is close enough to $X$ . However, we need a similar argument for Lemma B.7, so we moved forward with the proof approach given above. + +We now state the definition of weakly convex and linearly bounded to define a more restricted hypothesis space with an improved factor in Lemma B.7. + +Definition B.4. For $p \in \mathbb{N}$ , a function $h$ is called a $p$ -weakly convex function if for any $x \in \Omega$ , there exists a sufficiently small neighborhood $U$ of $x$ so that: + +$$ +h (y) \geq h (x) + \nabla h (x) (y - x) - C \| y - x \| ^ {p} \quad \forall y \in U \tag {46} +$$ + +Note that any convex function is $p$ -weakly convex for any $p \in \mathbb{N}$ . + +Definition B.5. A function $h$ is linearly bounded, if for some constant $C > 0$ , and for any $x \in \Omega$ , there exists a sufficiently small neighborhood $U$ of $x$ so that: + +$$ +\left| h (y) - h (x) \right| \leq C \| \nabla h (x) \| \| y - x \| \quad \forall y \in U \tag {47} +$$ + +Any Lipschitz function on compact domain that has non-zero derivative is linearly bounded. One such example is the function $e^{\alpha \| x\|^2}$ on domains that do not contain 0. + +Remark B.6. Note that there are many hypothesis spaces consisting of infinitely many elements that satisfy Lemma B.7. One such hypothesis space $\mathcal{H}$ is: + +$$ +\left\{h (x) := \sum_ {i = 1} ^ {D} a _ {i} e ^ {b _ {i} \| x \| ^ {2}}, \quad 0 \leq a _ {i} \leq A, 0 < b \leq b _ {i} \leq B \right\} \tag {48} +$$ + +for constant $A, B, b > 0$ and for domain $\Omega$ that doesn't contain 0. Lemma B.7 also holds for concave functions and their weak versions. + +Lemma B.7. Suppose that, possibly with knowledge from an outside environment or from certain policy experts, the hypothesis space $\mathcal{H}$ is reduced to a smaller family of functions of the form $g + h$ , where $g$ is as in Lemma B.3, and $h$ is linearly bounded and $p$ -weakly convex for some $p \geq 2d$ . Then $N(f, \mathcal{H}, \epsilon, \delta)$ can be upper bound by the factor $\mathcal{O}(\epsilon^{-6})$ that is independent of the dimension $d$ . + +Proof. Suppose we're given $\epsilon > 0$ . Take $\epsilon_1 = \epsilon^{1/d} = \epsilon^\alpha$ with $\alpha = 1/d$ , and set $\epsilon_2 = 1/2$ . Now choose $m \in \mathbb{N}$ large enough so that $(1 - c_1 \epsilon_2 \epsilon_1^d)^m < \epsilon^2$ , where $c_1$ is an appropriate geometric constant. From here, we can see that $m \approx \mathcal{O}(\epsilon_1^{-d}) = \mathcal{O}(\epsilon^{-1})$ . + +Choose $N\approx \mathcal{O}(\epsilon^{-6})\in \mathbb{N}$ + +$$ +N = \mathcal {O} \left(\frac {\log (d / \delta)}{\epsilon^ {6}}\right) \text {s o t h a t} \sqrt {\frac {\log (1 / (\delta / d))}{N}} \approx \mathcal {O} \left(\epsilon^ {3}\right) \tag {49} +$$ + +Train a neural network function $g$ to approximate $f$ on $N$ samples $Y_{1},\dots ,Y_{N}$ so that $g(Y_{i})\approx f(Y_{i})$ for $i\in \overline{1,N}$ . We prove that this $N$ is large enough to allow derivative approximation transfer. + +Choose $\beta_{k} = k / d$ and $\delta_{k} = k\delta /d$ for $k\in \overline{0,d}$ . We prove by induction on $k\in \overline{0,d}$ that there exists a subspace of probability at least $1 - \delta_{k}$ so that $\| \nabla f(Y) - \nabla g(Y)\| < C_0\epsilon^{\beta_k}$ for all $Y\in \{Y_1,\dots ,Y_n\}$ and for some constant $C_0$ . + +For $k = 0$ , the bound is trivial. For the induction step from $k$ to $k + 1$ , we first consider the loss function $l_{k}$ of the form $l_{k}(h,(x,y)) = l_{k}(h,(x,g(x))) = \phi_{k}(h(x),g(x))$ , where $\phi_k(y,\hat{y}) = \text{clip}(|y - \hat{y}|,0,C\epsilon^{\alpha +\beta_k})^d$ . Here $\text{clip}$ denotes a clip function and, in this case, is obviously a Lipschitz function with Lipschitz constant 1. First, the loss function $l_{k}$ is bounded by $(C\epsilon^{\alpha +\beta_k})^d$ . Now note the following simple inequality: + +$$ +\left| a ^ {d} - b ^ {d} \right| = \left| a - b \right| \left| \sum_ {k = 0} ^ {d - 1} a ^ {k} b ^ {d - 1 - k} \right| < | a - b | d \left(C \epsilon^ {\alpha + \beta_ {k}}\right) ^ {d - 1} +$$ + +for $a, b < C\epsilon^{\alpha + \beta_k}$ . The inequality shows that $\phi_k$ has the Lipschitz constant bounded by $d(C\epsilon^{\alpha + \beta_k})^{d - 1}$ . We are now ready to go the main step of the induction step. + +Condition on $Y$ , we repeat the same argument in Lemma B.3's proof to show that except for a subspace with probability less than $\epsilon^2$ , there exists $j \in \overline{1,m}$ (depends on both $Y$ and $X_{1},\dots ,X_{m}$ ) so that: + +$$ +\epsilon_ {1} / 2 < \| X _ {j} - Y \| < \epsilon_ {1} = \epsilon^ {\alpha} \tag {50} +$$ + +$$ +\left(\nabla h (Y) - \nabla g (Y)\right) \cdot \left(X _ {j} - Y\right) > \left(1 - \epsilon_ {2}\right) \| \nabla h (Y) - \nabla g (Y) \| \| X _ {j} - Y \| \tag {51} +$$ + +Under this subspace, because $h - g$ is linearly bounded, + +$$ +\left| h \left(X _ {j}\right) - g \left(X _ {j}\right) \right| \leq C \| \nabla h (Y) - \nabla g (Y) \| \| Y - X _ {j} \| < C \epsilon^ {\alpha + \beta_ {k}} +$$ + +As a result, $|h(X_j) - g(X_j)|^d = \phi_k(h(X_j), g(X_j))$ . Next, since $h - g$ is $p$ -weakly convex, we have: + +$$ +\begin{array}{l} (1 - \epsilon_ {2}) \| \nabla h (Y) - \nabla g (Y) \| (\epsilon_ {1} / 2) \\ < (1 - \epsilon_ {2}) \| \nabla h (Y) - \nabla g (Y) \| \| X _ {j} - Y \| \\ \leq (\nabla h (Y) - \nabla g (Y)) \cdot (X _ {j} - Y) \\ \leq | h (X _ {j}) - g (X _ {j}) | + | h (Y) - g (Y) | + C _ {1} \| X _ {j} - Y \| ^ {p} \\ < | h (Y) - g (Y) | + | h \left(X _ {j}\right) - g \left(X _ {j}\right) | + 2 C _ {1} \left(\epsilon^ {\alpha}\right) ^ {2 d} \\ = \left| h \left(X _ {j}\right) - g \left(X _ {j}\right) \right| + 2 C _ {1} \epsilon^ {2} \\ = \phi_ {k} \left(h \left(X _ {j}\right), g \left(X _ {j}\right)\right) ^ {1 / d} + 2 C _ {1} \epsilon^ {2} \\ \leq \left(\sum_ {i = 1} ^ {m} \phi_ {k} (h (X _ {i}), g (X _ {i}))\right) ^ {1 / d} + 2 C _ {1} \epsilon^ {2} \\ \end{array} +$$ + +By taking expectation over $X_{1},\dots ,X_{m}$ , we obtain + +$$ +\begin{array}{l} (1 - \epsilon_ {2}) \| \nabla h (Y) - \nabla g (Y) \| (\epsilon_ {1} / 2) \\ < \left(1 - \epsilon^ {2}\right) \left(\left(\mathbb {E} _ {X _ {1}, \dots , X _ {m}} \left[ \sum_ {i = 1} ^ {m} \phi_ {k} (h (X _ {i}), g (X _ {i})) \right]\right) ^ {1 / d} + 2 C _ {1} \epsilon^ {2}\right) + C _ {2} \epsilon^ {2} \\ < \left(m \mathbb {E} _ {X} \left[ \phi_ {k} (h (X), g (X)) \right]\right) ^ {1 / d} + C _ {3} \epsilon^ {2} \\ = \left(m \mathbb {E} _ {X} \left[ l _ {k} (h, (X, g (X))) \right]\right) ^ {1 / d} + C _ {3} \epsilon^ {2} \\ \end{array} +$$ + +for appropriate constants $C_2, C_3 > 0$ . + +By Lemma A.2 and Lemma A.3 on the Lipschitz loss function $l_{k}$ bounded by $(C\epsilon^{\alpha +\beta_k})^d$ with Lipschitz constant $d(C\epsilon^{\alpha +\beta_k})^{d - 1}$ , with probability of at least $1 - \delta /d$ , we can continue the sequence of upper-bounds: + +$$ +\begin{array}{l} (1 - \epsilon_ {2}) \| \nabla h (Y) - \nabla g (Y) \| (\epsilon_ {1} / 2) \\ \leq C _ {4} \left(\epsilon^ {- 1} \left(\epsilon^ {\alpha + \beta_ {k}}\right) ^ {(d - 1)} \epsilon^ {3}\right) ^ {1 / d} + C _ {3} \epsilon^ {2} \\ \leq C _ {5} \epsilon^ {\alpha + \beta_ {k} + 1 / d} = C _ {5} \epsilon^ {\alpha + \beta_ {k + 1}} \\ \end{array} +$$ + +for appropriate constants $C_4, C_5 > 0$ . + +Hence, we finish the induction step because except on a subspace with probability at most $\delta_k + \delta / d = \delta_{k+1}$ , the inequality for induction hypothesis at $(k+1)$ holds. For $\beta_d = 1$ , we obtain $\|\nabla h(Y) - \nabla g(Y)\| < \epsilon$ for all $Y \in \{Y_1, \dots, Y_N\}$ with probability of at least $1 - \delta_d = 1 - \delta$ . By repeating the argument in Lemma B.3's proof, we obtain the expected bound + +$$ +\mathbb {E} _ {X} \left[ \left\| \nabla h (Y) - \nabla g (Y) \right\| \right] < C _ {6} \epsilon \tag {52} +$$ + +for an appropriate constant $C_6 > 0$ . After a constant rescaling, we showed that $N(g,\mathcal{H},\epsilon ,\delta)\approx \mathcal{O}(\epsilon^{-6})$ + +Together with the general pointwise estimates for dfPO algorithm in Theorem 3.2, Lemma B.3 and Lemma B.7 allow us to explicitly state the number of training episodes required for two scenarios considered in this work in Corollary 3.3 and in Corollary 3.4. Note that the proofs for these corollaries now follow trivially from Theorem 3.2, Lemma B.3, and Lemma B.7. + +# C Further experiment details + +# C.1 Sample size and problem parameters + +For the first two tasks, models are trained for 100,000 steps, while for the third task, training is limited to 5,000 steps due to the high computational cost of reward evaluation. For the reshaped reward $r(s,a) = \beta^{-t}(\frac{1}{2}\|a\|^2 - \mathcal{F}(s))$ , we define the decay factor as $\gamma := \beta^{\Delta_t}$ , where $\Delta_t$ is the step size (time step). Details on sample size (episodes and steps per episode), step size $\Delta_t$ , and decay factor $\gamma$ are summarized in Table 3. + +Table 3: Task-specific details. + +
Surface modelingGrid-based modelingMolecular dynamics
# of episodes50005000800
# of steps20206
Step size Δt0.010.010.1
Factor γ0.990.810.0067
+ +# C.2 Statistical analysis on benchmarking results + +We perform benchmarking using 10 different random seeds, with each seed generating over 200 test episodes. In Table 4, we report the mean and variance of final functional costs across 13 algorithms. Statistical comparisons are conducted using t-tests on the seed-level means. dfPO demonstrates statistically significant improvement over all baselines in nearly all settings. The only exception is the first experiment (Surface modeling), where dfPO and CrossQ exhibit comparable performance. + +Table 4: Final evaluation costs $(\mathcal{F}(s)$ at terminal step, mean $\pm$ std) from 13 different algorithms for 3 tasks from 10 different seeds. + +
Surface modelingGrid-based modelingMolecular dynamics
dfPO6.296 ± 0.0486.046 ± 0.08353.352 ± 0.055
TRPO6.470 ± 0.0217.160 ± 0.1131842.300 ± 0.007
PPO20.577 ± 2.2737.155 ± 0.1091842.303 ± 0.007
SAC7.424 ± 0.0457.066 ± 0.1011364.747 ± 12.683
DDPG15.421 ± 1.4716.570 ± 0.08268.203 ± 0.001
CrossQ6.365 ± 0.0307.211 ± 0.122951.674 ± 15.476
TQC6.590 ± 0.0477.120 ± 0.08776.874 ± 0.001
S-TRPO7.772 ± 0.0856.470 ± 0.0981842.287 ± 0.014
S-PPO16.422 ± 1.1667.064 ± 0.0941842.304 ± 0.009
S-SAC8.776 ± 0.1077.209 ± 0.126126.397 ± 1.315
S-DDPG9.503 ± 0.2106.642 ± 0.12482.946 ± 0.001
S-CrossQ6.830 ± 0.0767.028 ± 0.118338.120 ± 8.642
S-TQC6.468 ± 0.0266.716 ± 0.099233.944 ± 2.966
+ +# C.3 Additional ablation study + +In the main paper, we reported ablations for the reward-shaped variants in Table 2; here we present the corresponding results for the standard RL algorithms in Table 5 below. + +# C.4 Training time and memory usage + +Approximate model sizes are given in Table 6; our networks are small, so memory overhead is low and only slightly above PPO/TRPO. Approximate per-task wall-clock times are listed in Table 7 and are comparable across tasks. + +Table 5: Hyperparameter ablations on standard (S-) algorithms. + +
DatasetdfPOorigS-CrossQS-SACS-TQC
orignc=10nc=2origent=0.05ent=0.2orignc=10nq=5
Surface6.326.937.2219.428.898.719.796.518.656.61
Grid6.067.077.217.157.177.907.216.717.007.12
Mol. Dyn.53.34338.07593.531213.82126.73210.94523.92231.98270.12668.10
DatasetdfPOorigS-DDPGS-PPOS-TRPO
orignoise=OUtau=0.01origclip=0.1norm=ForigGAE-λ=0.8
Surface6.329.5418.6311.4219.1719.8624.977.7415.41
Grid6.066.686.986.957.057.147.216.486.88
Mol. Dyn.53.3482.9590.6483.741842.301842.331842.311842.301842.28
+ +Table 6: Model sizes (in MB) for 13 algorithms across tasks. + +
Surface modelingGrid-based modelingMolecular dynamics
dfPO0.170.660.17
TRPO0.060.370.06
PPO0.080.620.08
SAC0.252.860.25
DDPG4.095.194.09
CrossQ0.272.370.27
TQC0.576.450.57
S-TRPO0.060.370.06
S-PPO0.080.620.08
S-SAC0.252.860.25
S-DDPG4.095.194.09
S-CrossQ0.272.370.27
S-TQC0.576.450.57
+ +# C.5 Evaluation on standard RL tasks + +We evaluate on continuous-state, continuous-action versions of Pendulum, Mountain Car, and CartPole using Gym. For Mountain Car, we use reward function $R = 100\sigma(20(\text{position} - 0.45)) - 0.1$ action[0]^2, where $\sigma$ denotes the sigmoid. For CartPole, $R = \text{upright} \cdot \text{centered} \cdot \text{stable}$ with $\text{upright} = 2\sigma(-5|\theta/\theta_{\text{thresh}}|)$ , centered $= 2\sigma(-2|x/x_{\text{thresh}}|)$ , and stable $= 2\sigma(-0.5(\dot{x}^2 + \dot{\theta}^2))$ . Rewards lie in [0, 1] and attain 1 only at $\theta = x = \dot{\theta} = \dot{x} = 0$ ; thus moderate reward values (e.g., $\approx 0.15$ ) can still indicate acceptable control within thresholds. We adopt continuous rewards to align with our continuous-time assumptions. Results in Table 8 report episode rewards. + +Our method performs reasonably on these standard tasks. Additionally, in the main paper, dfPO shows strong performance in scientific computing tasks, where optimization over structured geometric spaces, coarse-to-fine grid discretizations, and molecular energy landscapes better reflect real-world modeling with complex functionals. + +# C.6 Explanation on the choices of representative tasks + +In this section, we justify our choice of three evaluation tasks that capture scientific-computing settings where physics and sample efficiency are essential. Our aim is to develop reinforcement learning methods for settings where data are expensive to simulate and physical consistency is critical, with a focus on scientific-computing applications. Motivated by this, we identify three representative, foundational task types: + +Surface Modeling, control over geometries. At the level of an individual object, many scientific computing problems involve modifying the geometry of a structure to achieve desired physical properties. A standard example is the design of an airfoil (e.g., an aircraft wing), where the goal is to optimize its surface shape over time to minimize drag or maximize lift under aerodynamic flow. + +Table 7: Approximate training time (in hours) for each algorithm. + +
AlgorithmPPOTRPOSACDDPGTQCCrossQdfPO
Train time (hrs)0.30.61.01.22.02.01.0
+ +Table 8: Episode rewards on continuous-state/action classic-control tasks. + +
TaskPPOTRPODDPGSACTQCCrossQdfPO
Pendulum-0.0213-0.0011-0.0063-0.0054-0.0047-0.0045-0.0042
Mountain Car58.527360.121755.348960.726870.528063.217559.0146
CartPole0.09030.12040.11510.11300.05270.12410.1352
+ +These surfaces are often altered through a set of control points, and the reward is derived from a functional measuring aerodynamic performance. Similarly, in structural engineering, surfaces can be automatically adjusted to improve stability against external disturbances, such as seismic vibrations. Additionally, in materials processing, time-varying surface optimization is used to control mechanical or thermal properties, like stress distributions and heat dissipation, during the manufacturing of advanced materials. Our surface modeling task captures this family of problems by enabling control over geometries. + +Grid-Based Modeling, control under PDE constraints. When moving beyond individual geometries to macro-scale physical systems, we typically encounter phenomena modeled by controlled partial differential equations (PDEs). These PDEs capture time-evolving quantities such as temperature, pressure, or concentration fields in space. For instance, the heat equation $\frac{du}{dt} = \Delta u + f$ models temperature evolution, where $u$ is the temperature field and $f$ is a control input. An important application is data center temperature control, where $f$ can represent electricity supplied to cooling elements, and the goal is to keep the temperature stable while optimizing the energy budget. Similar examples range from smart HVAC systems to industrial furnace regulation. Most, if not all, physical phenomena fall under this category and are represented by classical PDEs such as advection-diffusion equations, wave equations, reaction-diffusion systems, and elasticity equations. In practical computational settings, solving such PDEs often requires spatial discretization, typically using a grid-based approximation. Due to computational constraints, control actions are applied on a coarser grid, while the underlying physical evaluation (i.e., computing the reward) is carried out on a finer grid. Our grid-based task precisely reflects this multiscale setting: it requires learning control policies that operate on a coarse discretization but are evaluated through a fine-grid reconstruction. + +Molecular Dynamics. At a much smaller atomic scale, such as those in molecular or biological systems, physical processes are often not well-described by a single PDE. Instead, one must work directly with the atomic structures, whose interactions are governed by complex, often nonlocal, energy-based potentials. This motivates our third category of molecular dynamics. One example is understanding how virus capsids optimally change over time under therapeutic molecular interactions. This is important for designing more effective treatments. + +In summary, our three evaluation tasks correspond to core abstractions in scientific computing. As summarized in the main paper, these include: + +- Optimization over geometric surfaces. +- Grid-based modeling with controlled PDEs. +- Molecular dynamics in atomistic systems. + +# C.7 Scientific-computing tasks where re-planning assumptions fail + +In many controlled PDEs, interesting dynamics concentrate near specific times—for example, sharp transients or blow-up behavior where $u(t) \to \infty$ as $t \to t^{\star}$ . To resolve such phenomena, practical solvers avoid a uniform time grid and instead place time points $\{t_i\}_{i=0}^N$ that cluster near the event (often geometrically), so $t_{i+1} - t_i$ shrinks rapidly as $t_i \to t^{\star}$ . Consider a black-box solver for such + +controlled PDEs: + +$$ +F \left(u _ {t}, u, \nabla u, \nabla^ {2} u, f\right) = 0, \tag {53} +$$ + +with a fine-grid state $x_{i}$ and a coarse control $f_{i}$ . A single forward discretization step can be written as: + +$$ +x _ {i + 1} = A _ {i} x _ {i} + G _ {i} f _ {i} + r _ {i}, \tag {54} +$$ + +where $A_{i}$ is the high-dimensional propagator determined by the local step size and discretization, while $G_{i}f_{i}$ injects a low-rank control effect (rank $\ll \dim x_{i}$ ), and $r_i$ is a known source/residual. Under adaptivity, the operators $(A_{i})$ vary with $i$ and generally do not commute. Define the prefix propagators from the initial time: + +$$ +Q _ {0} := I, \quad Q _ {i + 1} := A _ {i} Q _ {i}. \tag {55} +$$ + +For rollouts from $t_0$ , the full-rank computations are concentrated in evaluating $Q_{i + 1}x_0$ , while control/source contributions are less expensive due to their low-rank structures. To keep computational cost feasible, the solver can cache low-rank surrogates of the prefixes $Q_{i + 1}\approx C_{i + 1}D_{i + 1}$ for low-rank matrices $C_{i + 1}$ and $D_{i + 1}$ , enabling fast computations from the initial time. + +A mid-trajectory restart at an arbitrary time $t_k$ requires the full suffix operator $S_{k\to n} \coloneqq A_{n - 1}\dots A_k$ . Making restart practical would therefore require constructing and maintaining low-rank approximations for every suffix $S_{k\rightarrow n}$ across many $k$ . In a black-box environment, either these suffix maps are unavailable, or storing and updating them would exceed compute and memory budgets. Thus, the re-planning assumption demands capabilities beyond a black-box solver and fails in this setting. + +# D Hamiltonian differential dual approach + +# D.1 Physics intuition + +Lagrangian mechanics is a reformulation of classical dynamics that expresses motion in terms of energies and generalized coordinates. Where Newtonian mechanics emphasizes forces and constraints, the Lagrangian view encodes dynamics through the principle of stationary action, from which the familiar conservation laws emerge. Specifically, each admissible path $s:[0,T]\to \mathbb{R}^d$ through space-time carries a scalar "action". The physical path is the one that renders this action stationary (often a minimum) under perturbations that fix the endpoints. Formally, with Lagrangian $\mathcal{L}(s,\dot{s},t)$ , the action functional $S$ is defined as an indefinite integral: + +$$ +\mathcal {S} = \int \mathcal {L} (s, \dot {s}, t) d t \tag {56} +$$ + +and stationarity of $S$ yields the Euler-Lagrange equation that governs a physical path: + +$$ +\frac {\partial \mathcal {L} (s , \dot {s} , t)}{\partial s} = \frac {d}{d t} \frac {\partial \mathcal {L} (s , \dot {s} , t)}{\partial \dot {s}} \tag {57} +$$ + +A canonical example is $\mathcal{L}(s, \dot{s}, t) = \frac{1}{2} m\|\dot{s}\|^2 - \mathcal{V}(s)$ (kinetic minus potential energy). Then $\partial \mathcal{L} / \partial \dot{s} = m\dot{s}$ and $\partial \mathcal{L} / \partial s = -\nabla \mathcal{V}(s)$ , so that Equation (57) reduces to + +$$ +- \nabla \mathcal {V} (s) = \frac {d}{d t} (m \dot {s}) = m \ddot {s}, \tag {58} +$$ + +which is precisely the Newton's second law $F = m\ddot{s}$ with the force $F$ being minus gradient of the potential energy. + +Optimal control (continuous-time RL) perspective. The Lagrangian formulation can be viewed as a special case of optimal control by identifying the control with velocity, $a \equiv \dot{s}$ , and adopting the special dynamics $f(s, a) = a$ . Consider the corresponding value function: + +$$ +V (s, t) := \max _ {a (\cdot)} \int_ {t} ^ {T} \left(- \mathcal {L} (w (u), a (u), u)\right) d u \quad \text {s . t .} \quad \dot {w} (u) = a (u), w (t) = s. \tag {59} +$$ + +The Hamilton-Jacobi-Bellman (HJB) equation [11] then reads + +$$ +\frac {\partial V (s , t)}{\partial t} + \max _ {a} \left(\frac {\partial V (s , t)}{\partial s} f (s, a) - \mathcal {L} (s, a)\right) = 0 \tag {60} +$$ + +Defining the Hamiltonian $\mathcal{H}(s,a,t)\coloneqq \frac{\partial V(s,t)}{\partial s} f(s,a) - \mathcal{L}(s,a)$ , we recover Equation (5) with adjoint (costate) $p = \partial V / \partial s$ . Optimality requires the first-order condition $\frac{\partial\mathcal{H}}{\partial a} = 0$ , which yields $\frac{\partial\mathcal{L}}{\partial\dot{s}} = \frac{\partial V}{\partial s}$ , and substituting the maximizing control $a^* (s,t) = \dot{s}$ into Equation (60) gives $\frac{\partial V}{\partial t} = \mathcal{L} - \frac{\partial V}{\partial s} f = \mathcal{L} - \frac{\partial V}{\partial s}\dot{s}$ . + +Differentiate the identity $\frac{\partial\mathcal{L}}{\partial\dot{s}} = \frac{\partial V}{\partial s}$ along the optimal trajectory yields: + +$$ +\begin{array}{l} \frac {d}{d t} \frac {\partial \mathcal {L}}{\partial \dot {s}} = \frac {d}{d t} \frac {\partial V}{\partial s} = \frac {\partial}{\partial t} \frac {\partial V}{\partial s} + \frac {\partial^ {2} V}{\partial s ^ {2}} \dot {s} \\ = \frac {\partial}{\partial s} \frac {\partial V}{\partial t} + \frac {\partial^ {2} V}{\partial s ^ {2}} \dot {s} \tag {61} \\ = \frac {\partial}{\partial s} \left(\mathcal {L} - \frac {\partial V}{\partial s} \dot {s}\right) + \frac {\partial^ {2} V}{\partial s ^ {2}} \dot {s} \\ = \frac {\partial \mathcal {L}}{\partial s} - \dot {s} \frac {\partial^ {2} V}{\partial s ^ {2}} + \frac {\partial^ {2} V}{\partial s ^ {2}} \dot {s} = \frac {\partial \mathcal {L}}{\partial s} \\ \end{array} +$$ + +which is exactly Euler-Langrange Equation (57). Thus, the stationary action is a special optimal control problem where velocity plays the role of the control, and $\mathcal{H}$ ties value gradients to momenta. + +Hamiltonian mechanics and duality. Hamiltonian mechanics follows from the same dual construction (see Equation (9)): with controls suppressed, the Hamiltonian $\mathcal{H}$ and the adjoint $p$ encode the dynamics via symplectic flow. In our setting, Lagrangian mechanics appears as a special case, and Hamiltonian mechanics is the corresponding dual description. The differential-learning duality we use generalizes this physics correspondence and provides the bridge to continuous-time RL. + +In this section, we write $a = \dot{s}$ and occasionally suppress explicit $(s,t)$ and $(s,\dot{s},t)$ arguments in $V$ and $\mathcal{L}$ for readability. A more rigorous derivation can also be done via the calculus of variations. + +# D.2 Relation with state-action value function + +We revisit the temporal-difference (TD) error $r(s, a) + V(s') - V(s)$ , where the next state $s'$ follows the dynamics $s' = s + \Delta_t f(s, a)$ . Using a reparameterization trick, $f$ can absorb arbitrary noise $\epsilon$ as $f = f(\cdot, \epsilon)$ . With a first-order expansion and a constant step size, taking $\Delta_t = 1$ to match the discrete-time TD update, we obtain: + +$$ +\begin{array}{l} \epsilon_ {T D} = r (s, a) + V \left(s ^ {\prime}\right) - V (s) = r (s, a) + V \left(s + \Delta_ {t} f (s, a)\right) - V (s) \\ \approx r (s, a) + \Delta_ {t} f (s, a) \frac {\partial V}{\partial s} (s) = - \mathcal {H} \left(s, - \frac {\partial V}{\partial s} (s), a\right) \tag {62} \\ \end{array} +$$ + +Thus, under the one-step expansion with $\Delta_t = 1$ , the TD error is exactly the Hamiltonian evaluated at the value gradient. This identifies the critic's TD signal with our control-theoretic local quantities used in the dual approach. In the continuous-time limit ( $\Delta_t \to 0$ ), this yields an instantaneous quantity that coincides with the continuous-time $q$ -function of Jia and Zhou [15]. When the dynamics are unknown, such quantities can be estimated through the drift from observed transitions: $f(s, a) \approx (s' - s) / \Delta_t$ . + +# E Limitation + +Our theoretical results rely on a set of assumptions stated in the corresponding theorems and lemmas, including continuity of the initial state distribution and Lipschitz regularity of the dynamics operator $G$ and score function $g$ . These assumptions are standard and broadly applicable in physical systems, but they exclude certain cases, such as systems with discontinuous dynamics, which are not addressed in this work. + +While our Differential RL framework is designed to be broadly applicable across scientific computing domains, our experimental evaluation focuses on three representative classes: surface modeling, + +grid-based modeling, and molecular dynamics. These were selected to demonstrate the generality and effectiveness of our approach in settings with complex, simulation-defined objectives. Nonetheless, our experiments do not exhaust the full spectrum of possible applications, and future work will explore extensions to other domains—including those outside scientific computing, such as computer vision—and to a wider variety of functionals within each domain. + +# NeurIPS Paper Checklist + +# 1. Claims + +Question: Do the main claims made in the abstract and introduction accurately reflect the paper's contributions and scope? + +Answer: [Yes] + +Justification: The main claims in the abstract and introduction accurately summarize the key contributions and findings of the paper, and they align with the theoretical and experimental results presented. + +Guidelines: + +- The answer NA means that the abstract and introduction do not include the claims made in the paper. +- The abstract and/or introduction should clearly state the claims made, including the contributions made in the paper and important assumptions and limitations. A No or NA answer to this question will not be perceived well by the reviewers. +- The claims made should match theoretical and experimental results, and reflect how much the results can be expected to generalize to other settings. +- It is fine to include aspirational goals as motivation as long as it is clear that these goals are not attained by the paper. + +# 2. Limitations + +Question: Does the paper discuss the limitations of the work performed by the authors? + +Answer: [Yes] + +Justification: We discuss limitations (such as theoretical assumptions and the scope of problems considered) throughout the main paper and summarize them in the Appendix as well. + +Guidelines: + +- The answer NA means that the paper has no limitation while the answer No means that the paper has limitations, but those are not discussed in the paper. +- The authors are encouraged to create a separate "Limitations" section in their paper. +- The paper should point out any strong assumptions and how robust the results are to violations of these assumptions (e.g., independence assumptions, noiseless settings, model well-specification, asymptotic approximations only holding locally). The authors should reflect on how these assumptions might be violated in practice and what the implications would be. +- The authors should reflect on the scope of the claims made, e.g., if the approach was only tested on a few datasets or with a few runs. In general, empirical results often depend on implicit assumptions, which should be articulated. +- The authors should reflect on the factors that influence the performance of the approach. For example, a facial recognition algorithm may perform poorly when image resolution is low or images are taken in low lighting. Or a speech-to-text system might not be used reliably to provide closed captions for online lectures because it fails to handle technical jargon. +- The authors should discuss the computational efficiency of the proposed algorithms and how they scale with dataset size. +- If applicable, the authors should discuss possible limitations of their approach to address problems of privacy and fairness. +- While the authors might fear that complete honesty about limitations might be used by reviewers as grounds for rejection, a worse outcome might be that reviewers discover limitations that aren't acknowledged in the paper. The authors should use their best judgment and recognize that individual actions in favor of transparency play an important role in developing norms that preserve the integrity of the community. Reviewers will be specifically instructed to not penalize honesty concerning limitations. + +# 3. Theory assumptions and proofs + +Question: For each theoretical result, does the paper provide the full set of assumptions and a complete (and correct) proof? + +Answer: [Yes] + +Justification: The assumptions and setups for each theoretical result are stated in the main paper. Proof overviews are included in the main text, with full and detailed proofs provided in the Appendix. + +Guidelines: + +- The answer NA means that the paper does not include theoretical results. +- All the theorems, formulas, and proofs in the paper should be numbered and cross-referenced. +- All assumptions should be clearly stated or referenced in the statement of any theorems. +- The proofs can either appear in the main paper or the supplemental material, but if they appear in the supplemental material, the authors are encouraged to provide a short proof sketch to provide intuition. +- Inversely, any informal proof provided in the core of the paper should be complemented by formal proofs provided in appendix or supplemental material. +- Theorems and Lemmas that the proof relies upon should be properly referenced. + +# 4. Experimental result reproducibility + +Question: Does the paper fully disclose all the information needed to reproduce the main experimental results of the paper to the extent that it affects the main claims and/or conclusions of the paper (regardless of whether the code and data are provided or not)? + +Answer: [Yes] + +Justification: We provide all necessary details in the Experiment section and the Appendix, along with a link to the codebase that includes detailed instructions for reproducing the main experimental results. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- If the paper includes experiments, a No answer to this question will not be perceived well by the reviewers: Making the paper reproducible is important, regardless of whether the code and data are provided or not. +- If the contribution is a dataset and/or model, the authors should describe the steps taken to make their results reproducible or verifiable. +- Depending on the contribution, reproducibility can be accomplished in various ways. For example, if the contribution is a novel architecture, describing the architecture fully might suffice, or if the contribution is a specific model and empirical evaluation, it may be necessary to either make it possible for others to replicate the model with the same dataset, or provide access to the model. In general, releasing code and data is often one good way to accomplish this, but reproducibility can also be provided via detailed instructions for how to replicate the results, access to a hosted model (e.g., in the case of a large language model), releasing of a model checkpoint, or other means that are appropriate to the research performed. +- While NeurIPS does not require releasing code, the conference does require all submissions to provide some reasonable avenue for reproducibility, which may depend on the nature of the contribution. For example +(a) If the contribution is primarily a new algorithm, the paper should make it clear how to reproduce that algorithm. +(b) If the contribution is primarily a new model architecture, the paper should describe the architecture clearly and fully. +(c) If the contribution is a new model (e.g., a large language model), then there should either be a way to access this model for reproducing the results or a way to reproduce the model (e.g., with an open-source dataset or instructions for how to construct the dataset). + +(d) We recognize that reproducibility may be tricky in some cases, in which case authors are welcome to describe the particular way they provide for reproducibility. In the case of closed-source models, it may be that access to the model is limited in some way (e.g., to registered users), but it should be possible for other researchers to have some path to reproducing or verifying the results. + +# 5. Open access to data and code + +Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material? + +Answer: [Yes] + +Justification: We provide a link to the experimental codebase, which includes detailed instructions in the README file for reproducing our results. The README also includes links to necessary artifacts, such as trained models. + +# Guidelines: + +- The answer NA means that paper does not include experiments requiring code. +- Please see the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details. +- While we encourage the release of code and data, we understand that this might not be possible, so "No" is an acceptable answer. Papers cannot be rejected simply for not including code, unless this is central to the contribution (e.g., for a new open-source benchmark). +- The instructions should contain the exact command and environment needed to run to reproduce the results. See the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details. +- The authors should provide instructions on data access and preparation, including how to access the raw data, preprocessed data, intermediate data, and generated data, etc. +- The authors should provide scripts to reproduce all experimental results for the new proposed method and baselines. If only a subset of experiments are reproducible, they should state which ones are omitted from the script and why. +- At submission time, to preserve anonymity, the authors should release anonymized versions (if applicable). +- Providing as much information as possible in supplemental material (appended to the paper) is recommended, but including URLs to data and code is permitted. + +# 6. Experimental setting/details + +Question: Does the paper specify all the training and test details (e.g., data splits, hyperparameters, how they were chosen, type of optimizer, etc.) necessary to understand the results? + +Answer: [Yes] + +Justification: We provide detailed experimental settings in the Experiment section and the Appendix. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The experimental setting should be presented in the core of the paper to a level of detail that is necessary to appreciate the results and make sense of them. +- The full details can be provided either with the code, in appendix, or as supplemental material. + +# 7. Experiment statistical significance + +Question: Does the paper report error bars suitably and correctly defined or other appropriate information about the statistical significance of the experiments? + +Answer: [Yes] + +Justification: We report statistical analyses across 10 random seeds and include significance testing (t-test) in the Appendix to support the reliability of our experimental results. + +# Guidelines: + +- The answer NA means that the paper does not include experiments. +- The authors should answer "Yes" if the results are accompanied by error bars, confidence intervals, or statistical significance tests, at least for the experiments that support the main claims of the paper. +- The factors of variability that the error bars are capturing should be clearly stated (for example, train/test split, initialization, random drawing of some parameter, or overall run with given experimental conditions). +- The method for calculating the error bars should be explained (closed form formula, call to a library function, bootstrap, etc.) +- The assumptions made should be given (e.g., Normally distributed errors). +- It should be clear whether the error bar is the standard deviation or the standard error of the mean. +- It is OK to report 1-sigma error bars, but one should state it. The authors should preferably report a 2-sigma error bar than state that they have a $96\%$ CI, if the hypothesis of Normality of errors is not verified. +- For asymmetric distributions, the authors should be careful not to show in tables or figures symmetric error bars that would yield results that are out of range (e.g. negative error rates). +- If error bars are reported in tables or plots, The authors should explain in the text how they were calculated and reference the corresponding figures or tables in the text. + +# 8. Experiments compute resources + +Question: For each experiment, does the paper provide sufficient information on the computer resources (type of compute workers, memory, time of execution) needed to reproduce the experiments? + +Answer: [Yes] + +Justification: We provided a description of the hardware used for the experiments. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The paper should indicate the type of compute workers CPU or GPU, internal cluster, or cloud provider, including relevant memory and storage. +- The paper should provide the amount of compute required for each of the individual experimental runs as well as estimate the total compute. +- The paper should disclose whether the full research project required more compute than the experiments reported in the paper (e.g., preliminary or failed experiments that didn't make it into the paper). + +# 9. Code of ethics + +Question: Does the research conducted in the paper conform, in every respect, with the NeurIPS Code of Ethics https://neurips.cc/public/EthicsGuidelines? + +Answer: [Yes] + +Justification: We adhere to the NeurIPS Code of Ethics. + +Guidelines: + +- The answer NA means that the authors have not reviewed the NeurIPS Code of Ethics. +- If the authors answer No, they should explain the special circumstances that require a deviation from the Code of Ethics. +- The authors should make sure to preserve anonymity (e.g., if there is a special consideration due to laws or regulations in their jurisdiction). + +# 10. Broader impacts + +Question: Does the paper discuss both potential positive societal impacts and negative societal impacts of the work performed? + +Answer: [No] + +Justification: This paper aims to advance the field of Machine Learning. We do not foresee direct negative societal impacts arising from this work. + +# Guidelines: + +- The answer NA means that there is no societal impact of the work performed. +- If the authors answer NA or No, they should explain why their work has no societal impact or why the paper does not address societal impact. +- Examples of negative societal impacts include potential malicious or unintended uses (e.g., disinformation, generating fake profiles, surveillance), fairness considerations (e.g., deployment of technologies that could make decisions that unfairly impact specific groups), privacy considerations, and security considerations. +- The conference expects that many papers will be foundational research and not tied to particular applications, let alone deployments. However, if there is a direct path to any negative applications, the authors should point it out. For example, it is legitimate to point out that an improvement in the quality of generative models could be used to generate deepfakes for disinformation. On the other hand, it is not needed to point out that a generic algorithm for optimizing neural networks could enable people to train models that generate Deepfakes faster. +- The authors should consider possible harms that could arise when the technology is being used as intended and functioning correctly, harms that could arise when the technology is being used as intended but gives incorrect results, and harms following from (intentional or unintentional) misuse of the technology. +- If there are negative societal impacts, the authors could also discuss possible mitigation strategies (e.g., gated release of models, providing defenses in addition to attacks, mechanisms for monitoring misuse, mechanisms to monitor how a system learns from feedback over time, improving the efficiency and accessibility of ML). + +# 11. Safeguards + +Question: Does the paper describe safeguards that have been put in place for responsible release of data or models that have a high risk for misuse (e.g., pretrained language models, image generators, or scraped datasets)? + +Answer: [NA] + +Justification: The research involves only datasets and models that pose no significant misuse risks, thus no specific safeguards were necessary. + +# Guidelines: + +- The answer NA means that the paper poses no such risks. +- Released models that have a high risk for misuse or dual-use should be released with necessary safeguards to allow for controlled use of the model, for example by requiring that users adhere to usage guidelines or restrictions to access the model or implementing safety filters. +- Datasets that have been scraped from the Internet could pose safety risks. The authors should describe how they avoided releasing unsafe images. +- We recognize that providing effective safeguards is challenging, and many papers do not require this, but we encourage authors to take this into account and make a best faith effort. + +# 12. Licenses for existing assets + +Question: Are the creators or original owners of assets (e.g., code, data, models), used in the paper, properly credited and are the license and terms of use explicitly mentioned and properly respected? + +Answer: [Yes] + +Justification: We properly credit all external code and models used in this work by citing the relevant papers and libraries. + +# Guidelines: + +- The answer NA means that the paper does not use existing assets. +- The authors should cite the original paper that produced the code package or dataset. + +- The authors should state which version of the asset is used and, if possible, include a URL. +- The name of the license (e.g., CC-BY 4.0) should be included for each asset. +- For scraped data from a particular source (e.g., website), the copyright and terms of service of that source should be provided. +- If assets are released, the license, copyright information, and terms of use in the package should be provided. For popular datasets, paperswithcode.com/datasets has curated licenses for some datasets. Their licensing guide can help determine the license of a dataset. +- For existing datasets that are re-packaged, both the original license and the license of the derived asset (if it has changed) should be provided. +- If this information is not available online, the authors are encouraged to reach out to the asset's creators. + +# 13. New assets + +Question: Are new assets introduced in the paper well documented and is the documentation provided alongside the assets? + +Answer: [NA]. + +Justification: The paper does not introduce any new assets, thus documentation for such is not applicable. + +Guidelines: + +- The answer NA means that the paper does not release new assets. +- Researchers should communicate the details of the dataset/code/model as part of their submissions via structured templates. This includes details about training, license, limitations, etc. +- The paper should discuss whether and how consent was obtained from people whose asset is used. +- At submission time, remember to anonymize your assets (if applicable). You can either create an anonymized URL or include an anonymized zip file. + +# 14. Crowdsourcing and research with human subjects + +Question: For crowdsourcing experiments and research with human subjects, does the paper include the full text of instructions given to participants and screenshots, if applicable, as well as details about compensation (if any)? + +Answer: [NA] + +Justification: : This paper does not involve crowdsourcing or research with human subjects, thus the inclusion of participant instructions, screenshots, and compensation details is not applicable. + +Guidelines: + +- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. +- Including this information in the supplemental material is fine, but if the main contribution of the paper involves human subjects, then as much detail as possible should be included in the main paper. +- According to the NeurIPS Code of Ethics, workers involved in data collection, curation, or other labor should be paid at least the minimum wage in the country of the data collector. + +# 15. Institutional review board (IRB) approvals or equivalent for research with human subjects + +Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals (or an equivalent approval/review based on the requirements of your country or institution) were obtained? + +Answer: [NA] + +Justification: This paper does not involve research with human subjects. + +# Guidelines: + +- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. +- Depending on the country in which research is conducted, IRB approval (or equivalent) may be required for any human subjects research. If you obtained IRB approval, you should clearly state this in the paper. +- We recognize that the procedures for this may vary significantly between institutions and locations, and we expect authors to adhere to the NeurIPS Code of Ethics and the guidelines for their institution. +- For initial submissions, do not include any information that would break anonymity (if applicable), such as the institution conducting the review. + +# 16. Declaration of LLM usage + +Question: Does the paper describe the usage of LLMs if it is an important, original, or non-standard component of the core methods in this research? Note that if the LLM is used only for writing, editing, or formatting purposes and does not impact the core methodology, scientific rigorousness, or originality of the research, declaration is not required. + +Answer: [NA] + +Justification: LLM is used only for editing and formatting purposes. + +# Guidelines: + +- The answer NA means that the core method development in this research does not involve LLMs as any important, original, or non-standard components. +- Please refer to our LLM policy (https://neurips.cc/Conferences/2025/LLM) for what should or should not be described. \ No newline at end of file diff --git a/NeurIPS/2025/A Differential and Pointwise Control Approach to Reinforcement Learning/images.zip b/NeurIPS/2025/A Differential and Pointwise Control Approach to Reinforcement Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..53a6c237f444d881a9d7094ffc0736b4fe60ac2e --- /dev/null +++ b/NeurIPS/2025/A Differential and Pointwise Control Approach to Reinforcement Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5b4ccea1a14c044966ff5dc27d41e5d2002f503c2c8cf5be8fcbdec81d79ac7c +size 1022079 diff --git a/NeurIPS/2025/A Differential and Pointwise Control Approach to Reinforcement Learning/layout.json b/NeurIPS/2025/A Differential and Pointwise Control Approach to Reinforcement Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..2f3ff3b271eb6e51babd3da94de8b02e4aeb4322 --- /dev/null +++ b/NeurIPS/2025/A Differential and Pointwise Control Approach to Reinforcement Learning/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3fad031b8d98d74c4e054ad3f4d2faec10a6c9d49b6971d3311bd738847a86e5 +size 1482865 diff --git a/NeurIPS/2025/A Diffusion Model for Regular Time Series Generation from Irregular Data with Completion and Masking/4e565e2a-4685-450f-ae94-081491805b8e_content_list.json b/NeurIPS/2025/A Diffusion Model for Regular Time Series Generation from Irregular Data with Completion and Masking/4e565e2a-4685-450f-ae94-081491805b8e_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..334306e64cbf6572f0e9845158ccadb328908b51 --- /dev/null +++ b/NeurIPS/2025/A Diffusion Model for Regular Time Series Generation from Irregular Data with Completion and Masking/4e565e2a-4685-450f-ae94-081491805b8e_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:663379102cf04362192630f0035425a8b378880d465582de3059896a08e19eab +size 218059 diff --git a/NeurIPS/2025/A Diffusion Model for Regular Time Series Generation from Irregular Data with Completion and Masking/4e565e2a-4685-450f-ae94-081491805b8e_model.json b/NeurIPS/2025/A Diffusion Model for Regular Time Series Generation from Irregular Data with Completion and Masking/4e565e2a-4685-450f-ae94-081491805b8e_model.json new file mode 100644 index 0000000000000000000000000000000000000000..509e5a4f8659dcf257b7eeaf1a223b154a495cc8 --- /dev/null +++ b/NeurIPS/2025/A Diffusion Model for Regular Time Series Generation from Irregular Data with Completion and Masking/4e565e2a-4685-450f-ae94-081491805b8e_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7c3649fac70e1befdbebdc6bc3633635017c0436ff574d2984e7f850e26c9cf9 +size 268280 diff --git a/NeurIPS/2025/A Diffusion Model for Regular Time Series Generation from Irregular Data with Completion and Masking/4e565e2a-4685-450f-ae94-081491805b8e_origin.pdf b/NeurIPS/2025/A Diffusion Model for Regular Time Series Generation from Irregular Data with Completion and Masking/4e565e2a-4685-450f-ae94-081491805b8e_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..6f80b9750ae2bf23473bdcacc1199d2bb0c30215 --- /dev/null +++ b/NeurIPS/2025/A Diffusion Model for Regular Time Series Generation from Irregular Data with Completion and Masking/4e565e2a-4685-450f-ae94-081491805b8e_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8cd1023022dc097fff70e38d0e99f6c9030f068abc1b6443586abf7795f156ff +size 1558895 diff --git a/NeurIPS/2025/A Diffusion Model for Regular Time Series Generation from Irregular Data with Completion and Masking/full.md b/NeurIPS/2025/A Diffusion Model for Regular Time Series Generation from Irregular Data with Completion and Masking/full.md new file mode 100644 index 0000000000000000000000000000000000000000..6600d226c96f6c4cbc1d777612db58035dbdce55 --- /dev/null +++ b/NeurIPS/2025/A Diffusion Model for Regular Time Series Generation from Irregular Data with Completion and Masking/full.md @@ -0,0 +1,828 @@ +# A Diffusion Model for Regular Time Series Generation from Irregular Data with Completion and Masking + +Gal Fadlon* Idan Arbiv* Nimrod Berman Omri Azencot + +Department of Computer Science Ben-Gurion University of The Negev {galfad, arbivid, bermann}@post.bgu.ac.il azencot@bgu.ac.il + +# Abstract + +Generating realistic time series data is critical for applications in healthcare, finance, and science. However, irregular sampling and missing values present significant challenges. While prior methods address these irregularities, they often yield suboptimal results and incur high computational costs. Recent advances in regular time series generation, such as the diffusion-basedImagenTime model, demonstrate strong, fast, and scalable generative capabilities by transforming time series into image representations, making them a promising solution. However, extendingImagenTime to irregular sequences using simple masking introduces "unnatural" neighborhoods, where missing values replaced by zeros disrupt the learning process. To overcome this, we propose a novel two-step framework: first, a Time Series Transformer completes irregular sequences, creating natural neighborhoods; second, a vision-based diffusion model with masking minimizes dependence on the completed values. This approach leverages the strengths of both completion and masking, enabling robust and efficient generation of realistic time series. Our method achieves state-of-the-art performance, achieving a relative improvement in discriminative score by $70\%$ and in computational cost by $85\%$ . Code is at https://github.com/azencot-groupImagenI2R. + +# 1 Introduction + +Time series data is essential in fields such as healthcare, finance, and science, supporting critical tasks like forecasting trends, detecting anomalies, and analyzing patterns [18, 24, 34]. Beyond direct analysis, generating synthetic time series has become increasingly valuable for creating realistic proxies of private data, testing systems under new scenarios, exploring "what-if" questions, and balancing datasets for training machine learning models [6, 40]. The ability to generate realistic sequences enables deeper insights and robust applications across diverse domains. + +In practice, however, time series is often irregular, with unevenly spaced measurements and missing values. These irregularities arise from limitations in data collection processes, such as sensor failures, inconsistent sampling, or interruptions in monitoring systems [28, 44]. This irregularity poses a unique challenge for generating regular time series—where intervals are consistent, and the data follows the same distribution as if it were regularly observed [26]. The main goal of this paper is to generate regular sequences by training on irregularly-sampled time series using deep neural networks. + +The synthesis of regular time series from irregular ones is a fundamental challenge, yet existing approaches remain scarce, with notable examples being GT-GAN and KoVAE [26, 39]. Unfortunately, these methods suffer from several limitations. First, they rely on generative adversarial networks (GANs) and variational autoencoders (VAEs), which have recently been surpassed in performance + +by diffusion-based tools [11, 59, 38]. Second, both GT-GAN and KoVAE utilize a computationally-demanding preprocessing step based on neural controlled differential equations (NCDEs) [28], rendering these methods impractical for long time series. For instance, KoVAE requires $\approx 6.5\times$ more training time in comparison to our approach (See Fig. 3). Third, these methods inherently assume that the data, completed by NCDE, accurately reflects the true underlying distribution, which can introduce catastrophic errors when this assumption fails. In particular, their performance lags far behind that of models trained on regular time series, with state-of-the-art results on irregular discriminative benchmarks being, on average, $540\%$ worse than those on regular benchmarks. + +To address these shortcomings, we base our approach on a recent diffusion model for time series, ImagenTime [38]. This method maps time series data to images, enabling the use of powerful vision-based diffusion neural architectures. Leveraging a vision-based diffusion generator offers a significant advantage: regular time series can be generated from irregular ones using a straightforward masking mechanism. Specifically, missing values in the series are seamlessly ignored during the denoising process in training, akin to techniques used in image inpainting tasks [37, 12]. + +However, while this straightforward masking approach is simple and achieves strong results (see Tab. 7), we identify a significant limitation. Missing values in the time series are mapped to zeros in the image, resulting in "unnatural" neighborhoods that mix valid and invalid information. This can pose challenges for diffusion backbones, such as U-Nets with convolutional blocks, where the convolution kernels are not inherently masked and may inadvertently propagate errors from these artificial neighborhoods. To address this issue, we propose a two-step generation process. In the first step, we complete the irregular series using our adaptation of an efficient Time Series Transformer (TST) approach [61], significantly reducing computational overhead and enabling the generation of long time series. In the second step, we apply the straightforward masking approach described earlier. Crucially, this combination of completion and masking allows the model to learn from "natural" image neighborhoods while mitigating the reliance on fully accurate completed information through the use of masked loss minimization. Overall, our approach, built on a vision diffusion backbone, enables effective modeling of long time series while making minimal assumptions about pre-completed data, resulting in significantly efficient and improved generation performance. + +We conduct a comprehensive evaluation of our approach on standard irregular time series benchmarks, benchmarking it against state-of-the-art methods. Our model consistently demonstrates superior generative performance, effectively bridging the gap between regular and irregular settings. Furthermore, we extend the evaluation to medium-, long- and ultra-long-length sequence generation, assessing performance across 12 datasets and 12 tasks. The results highlight the robustness and efficiency of our method, achieving consistent improvements over existing approaches. Our key contributions are summarized as follows: + +1. We introduce a novel generative model for irregularly-sampled time series, leveraging vision-based diffusion approaches to efficiently and effectively handle sequences ranging from short to long lengths. +2. In contrast to existing methods that assume completed information is drawn from the data distribution, we treat it as a weak conditioning signal and directly optimize on the observed signal using a masking strategy. +3. Our approach achieves state-of-the-art performance across multiple generative tasks, delivering an average improvement of $70\%$ in discriminative benchmarks while reducing computational requirements by $85\%$ relative to competing methods. + +# 2 Related Work + +Diffusion models [49] have recently demonstrated groundbreaking generative capabilities, surpassing VAEs and GANs [29, 17] primarily on image generation [21, 14, 45]. The immense success of diffusion-based approaches has spurred a wave of recent advancements, encompassing both theoretical developments [50, 35] and practical applications [37, 22, 30, 4, 5, 3]. + +Generative modeling of time series is an emerging field, with pioneering approaches predominantly relying on GANs [58, 33, 31] and VAEs [13, 39]. Recently, inspired by the success of diffusion models, there has been a growing trend to adapt these techniques to various time series tasks [53, 43], including generative modeling [11, 59, 16]. Notably, theImagenTime approach [38] has achieved state-of-the-art performance on regular generative tasks for sequences of varying lengths, from short + +![](images/2bb6aea01d90e0d59e8c075b443eafdd0b622f29063f3be596625a016fa70667.jpg) + +![](images/e5aa47446c65f4ad40ef3fe04f92c95843d70c14e8e6cae6738a24ebe2d0f697.jpg) + +![](images/e6c74f0567bf900a7ee3d0c87a53fce6b055168742d720578331dd7f47673784.jpg) + +![](images/295c85d990f70e9f626dd30561b6b0ba16a5f3361dc59632ba39d1737a4a9b9f.jpg) + +![](images/ea984c1f0cb29288bd4c0d0b0e5df3ead6ee2527ca022667e86bfe149ca6be28.jpg) + +Figure 1: A data point (A) is mapped to an image with zeros and the coordinates in the center (B). Denoising the entire image yields inferior kernels (D) in comparison to masking (E). Constructing natural neighborhoods (C), yields consistent kernels and better scores (F). +![](images/743c8e6d54e5a899f28a49befe068f11e2afb920534fe64ccafa590972380033.jpg) +score $= 0.71$ score $= 0.67$ score $= \mathbf{0}.32$ + +to very long, by transforming time series into images and leveraging vision-based diffusion backbones. Our work builds on ImagenTime, extending it to address the challenging setting of generating regular time series information from irregularly-sampled data. + +Irregular time series modeling has been a longstanding task. Modern machine learning methods have made significant strides by framing the problem through the lens of differential equations [46, 28]. Subsequent efforts have explored alternative architectures, including recurrent neural networks [47] and transformers [61, 9]. However, learning from irregular sequences has received comparatively less attention, with notable contributions such as GT-GAN and KoVAE [26, 39], both relying on NCDE [28]. Despite their promise, NCDE-based methods are costly during preprocessing and training, limiting the applicability of GT-GAN and KoVAE in handling long time series. Further, replacing NCDE by efficient components such as TST [61], yields suboptimal results (see Tab. 7). + +# 3 Background + +Problem statement. We learn the underlying distribution of time series data from irregularly sampled observations and generating regular time series from it. Specifically, given a set of irregularly sampled sequences, our goal is to learn a model that approximates the true data distribution $p_{\mathrm{data}}(x_{1:T})$ and enables sampling of complete time series $x_{1:T}$ from the learned distribution $p_{\theta}(x_{1:T})$ . Formally, we consider a dataset of irregularly sampled time series, represented as $\{x_{t_1:t_n}^j\}_{j=1}^N$ , where each sequence consists of observations at non-uniform time steps $t_1 : t_n = [t_1, t_2, \ldots, t_n]$ with $t_1 \geq 1$ and $t_n \leq T$ . The challenge is to leverage these incomplete sequences to model the full distribution and generate realistic, regularly sampled time series that align with the true data distribution. + +ImagenTime employs the delay embedding transformation to map time series data into images, enabling their processing with powerful vision-based diffusion models [38]. Given an input multivariate regular time series $x_{1:T} \in \mathbb{R}^{d \times T}$ with $d$ features and length $T$ , the delay embedding constructs an image $x_{\mathrm{img}} \in \mathbb{R}^{d \times w \times h}$ by placing $x$ 's values over the columns of $x_{\mathrm{img}}$ per channel, where $w, h$ are user-defined parameters. During training, noise is added to the image $x_{\mathrm{img}}$ at different timesteps, forming $x_{\mathrm{img}}(t)$ . The diffusion model, parameterized by $s_\theta$ , learns to denoise these images by approximating the score function $s_\theta(x_{\mathrm{img}}, t)$ . Inference begins with a noise sample $x_{\mathrm{img}}(T) \sim \mathcal{N}(0, I)$ . This sample is iteratively denoised using the learned score function to produce a cleaned image $x_{\mathrm{img}}(0)$ . The inverse delay embedding transform is then applied to $x_{\mathrm{img}}(0)$ , reconstructing the original time series $\tilde{x}_{1:T}$ . Importantly, the inverse transformation of delay embedding is inherently non-unique, as the time series values are repeated within the image representation, suggesting various designs can be considered as we discuss in Sec. 4. Finally, a crucial advantage of ImagenTime is its effectivity in handling long series, e.g., a time series of length $65k$ is transformed to an image of size $256 \times 256$ . + +TST leverages the self-attention mechanism of Transformers to model temporal dependencies and long-range interactions in time-series data effectively. Unlike traditional sequence models such as RNNs or LSTMs, TST processes the entire sequence simultaneously, enabling parallelism and mitigating the vanishing gradient problem. The architecture includes input projection to a higher-dimensional feature space, positional encodings to capture temporal order, and a stack of Transformer encoder layers with flexible normalization and activation options. TST is particularly suitable for imputation tasks, as it can handle irregularly-sampled data and missing values through explicit masking and preprocessing. Additionally, its self-attention mechanism inherently supports long sequences, making it robust for capturing global context and dependencies in extended time-series data. By eliminating the need for computationally expensive preprocessing techniques, such as calculating coefficients for cubic splines or other interpolation methods, TST achieves significant speed advantages while providing a scalable, efficient, and accurate solution for tasks like forecasting, classification, anomaly detection, and imputation. + +![](images/ae974e9be26ede6917fbb0c38c89ac43f584b0bc9cd8b878d5e90743ceccf59a.jpg) +Figure 2: In the first step (top), we train a TST-based autoencoder, which we use during the second step (middle), where a vision diffusion model is trained with masking over non-active pixels. Inference (bottom) is done similarly toImagenTime. + +# 4 Method + +WhileImagenTime does not address the challenge of irregularly-sampled time series, a simple extension can enable it to generate regularly-sampled time series by training on irregular data. The key idea involves employing a mask during the loss computation. This mask ensures that only "active" pixels—those corresponding to observed time series values—are considered in the loss calculation, while "non-active" pixels, representing missing information, are effectively ignored. This approach enables effective learning from incomplete data while preserving the integrity of the observed information, offering two key advantages: (i) the mask is architecture-agnostic, making it compatible with any diffusion backbone, and (ii) the inference procedure ofImagenTime remains entirely unchanged. + +# 4.1 Unnatural image neighborhoods + +Unfortunately, the straightforward approach has a fundamental limitation: although non-active pixels are ignored during loss computation, they are still processed by the network. In practice, missing values are replaced with zeros, resulting in "unnatural" pixel neighborhoods. Specifically, while zeros may occasionally occur in non-zero segments of a time series, their repeated presence is highly unlikely, leading to inconsistencies. In other words, masking is not applied at the architecture level, potentially hindering the effective learning of neural components. + +To demonstrate this phenomenon, we consider the following toy experiment. We generate 1000 two-dimensional points, drawn from a multivariate Gaussian distribution with four centers (Fig. 1A). Given this distribution, we create an irregular dataset of $3 \times 4$ images, $S_{\mathrm{irregular}}$ , by taking a data point, setting all pixels to zero except those at the center, corresponding to the $x$ and $y$ coordinates of the original point (Fig. 1B). Then, we train two diffusion models to: (i) predict the score across the entire image, and (ii) predict only the two central pixels via masking. + +We evaluate the models by comparing their score estimation loss only on the two central pixels, regardless of the training strategy. Our results indicate that masking does not improve score estimation, yielding scores of 0.71 vs. 0.67 for Setups (i) and (ii), respectively. In addition, we also inspect the convolution kernels by averaging across channels the $L_{1}$ norm of each spatial position (Fig. 1D,E). As can be seen, the kernel in Setup (i) heavily attends to zero-valued pixels instead of focusing on the essential central pixels, suggesting that "unnatural" neighbors may be detrimental. In contrast, the masked kernel from Setup (ii) largely ignores non-relevant zeros and prioritizes the middle pixels. + +# 4.2 Our approach + +One possible solution to alleviate the phenomenon of unnatural neighborhoods is to implement masking at the kernel level, but this would require modifications tailored to each neural architecture, thereby restricting the approach's flexibility and its straightforward applicability across different models. For instance, while our work employs a convolutional U-Net, recent transformer-based architectures have emerged as highly effective diffusion backbones [41]. Accommodating such diversity in architectures would require a more generalized solution. + +To construct more natural pixel neighborhoods while remaining architecture-agnostic, we take inspiration from the two-stage pipelines of GT-GAN and KoVAE [26, 39]. Our method likewise uses a two-step training scheme. First, we complete missing values in irregularly sampled time series using + +TST [61] to obtain a regularly sampled sequence. Second, we transform the completed series into an image and apply denoising as inImagenTime, with one key change: during the loss computation we mask the pixels that originated from completion (see App. B), following the straightforward masking strategy discussed earlier. In the toy experiment of Sec. 4.1, the completed neighborhood (Fig. 1C) enables learning of consistent kernels (Fig. 1F) and improves score estimation to 0.32. + +We also replicated the synthetic experiment on a real-world Stocks dataset, using a larger convolutional kernel and comparing the two ways to complete irregular neighborhoods: (i) zero filling and (ii) natural-neighbor filling. The Stocks results mirrored the synthetic study, reinforcing our hypothesis that "unnatural" neighbors (e.g., zeros) are detrimental. Evaluated by score-estimation loss on active pixels, simple masking delivered only a marginal gain $(0.81\rightarrow 0.79)$ . Visual inspection likewise matched the synthetic patterns. In contrast, our proposed approach, which avoids zero padding and learning from semantically valid neighborhoods, significantly improved performance, reducing the loss to 0.29 and yielding more consistent kernel behavior. See App.E.1 for visuals. + +This combination of completion + masking addresses the two primary challenges of irregular sequences. Completion creates natural neighborhoods so convolutional kernels learn from values closer to the true data distribution; masking prevents over-reliance on imputed values by excluding them from the loss, striking a balance between leveraging and mitigating incomplete information. Figure 2 illustrates our pipeline: autoencoder pretraining (top), main training (middle), and inference (bottom). $\mathcal{T}$ and $\mathcal{T}^{-1}$ denote the delay-embedding transform and its inverse; the fire and snowflake icons indicate trainable and frozen modules, respectively. + +Importantly, we identify limitations with $\mathcal{T}^{-1}$ that was proposed in [38]. Specifically, in their approach, only the first pixel corresponding to each time series value is used for reconstruction. Specifically, if $x_{i}$ is mapped to multiple image indices, the original method selects the first corresponding pixel in the image for reconstruction. We modify this inverse transformation by aggregating information from all corresponding image indices and computing the average of the associated pixels for each $x_{i}$ . For a given $x_{1:L} \in \mathbb{R}^{L}$ , both methods ensure that $f^{-1}(f(x)) = x$ . See Sec. 5.6 for an ablation study of these two methods. + +# 5 Results + +# 5.1 Quality vs. Complexity + +We first compare the discriminative score and training time of our method against KoVAE across different sequence lengths (24, 96, 768). Both models were evaluated under identical conditions, utilizing the same GPU and batch size to ensure a fair comparison. Training time was measured until each model converged to its best result. The discriminative score and training time were averaged over all missing rates and datasets for each sequence length and each method separately. As illustrated in Fig. 3 and Tab. 1, our approach achieves an average speedup of approximately $6.5 \times$ and an average improvement of about $3.4 \times$ in the discriminative score compared to KoVAE. The results demonstrate that our method not only trains significantly faster but also generates data that more + +![](images/222c8cffefbd882315452707f22a7bb7fd159bb0ec6cb17795cc864beb33217d.jpg) +Figure 3: Discriminative score vs. training time for our approach and KoVAE across different lengths (24, 96, and 768). Lower discriminative scores and shorter training times are better. + +closely resembles the real distribution compared to KoVAE. Full results can be seen in App. E.7 + +# 5.2 Quantitative Evaluation + +Our quantitative evaluations assess missing rate setups of $30\%$ , $50\%$ , and $70\%$ . For example, in the $30\%$ missing rate case, we randomly omit $30\%$ of the data in each training sample. Additionally, we extend the standard benchmark, which typically considers a sequence length of 24, to include longer sequences of 96, 768, and 10,920, providing a more comprehensive evaluation across varying temporal scales. We utilize a diverse set of datasets, extending beyond common benchmarks such as Sine, Stock, Energy, and MuJoCo to include additional real-world datasets: ETTh1, ETTh2, ETTm1, ETTm2, Weather, Electricity, KDD-Cup, and Traffic. We compare against the popular TimeGAN + +approach [58], adapted to handle irregular data by incorporating the time difference between samples as input. In addition, we also consider the recent GT-GAN [26] and KoVAE [39]. + +We evaluate the performance of our model using the discriminative and predictive tasks suggested by [58]. In the discriminative task, we measure the similarity between real and synthetic samples by training a classifier to distinguish between the two, reporting $|0.5 - \mathrm{acc}|$ , where acc is the accuracy of the classifier. For the predictive task, we adopt the "train on synthetic, test on real" protocol, where a predictor is trained on synthetic data and tested on real data. The performance is evaluated using the Mean Absolute Error (MAE). We also consider irregular time series metrics: the Context-FID score [25], which quantifies the similarity in distribution between synthetic and real data, and the correlation score [33], which evaluates the feature-level relationship between the two datasets. The Context-FID score is computed by encoding both synthetic and real sequences using TS2Vec [60] and calculating the FID score on the representations. The correlation score measures the covariance of features between real and synthetic data, with a focus on assessing their alignment. Full details are provided in App. D.3. For all metrics, lower scores are better. + +Tab. 2 details the benchmark results for a sequence length of 24. The values represent averages over the $30\%$ , $50\%$ and $70\%$ missing rates, where the full results are provided in Tab. 18. In general, our approach presents dramatic improvements across all metrics with respect to the second-best approach (typically KoVAE). Following [39], we define the relative improvement error as $e_{\mathrm{rel}} = (e_2 - e_1) / e_2$ , where $e_2$ is the second-best error and $e_1$ is ours. In this metric, averaged across all datasets, our method improves by $74.2\%$ , $15.0\%$ , $78.5\%$ , $62.1\%$ in the discriminative, predictive, context-FID, and correlation scores, respectively. We also compared our approach to KoVAE in the medium (96) and long (768) lengths on all datasets excluding MuJoCo and electricity. The results appear in Tab. 3, + +showing the superiority of our method and its advantages to longer horizons, where we achieved mean relative improvement of $72.6\%$ , $29.7\%$ , $92.1\%$ , $73.25\%$ in the discriminative, predictive, context-FID, and correlation scores. See the full results in Tabs. 19, 20. Finally, for ultra-long sequences (10,920) on the KDD-Cup dataset, our model showed improvements of $36.6\%$ in the discriminative score (see Tab. 4), showing its ability to generate realistic synthetic data even at extreme sequence lengths. + +Table 4: Discriminative results of ultra-long (10,920) sequences on KDD-Cup for different missing rates. + +
Method30%50%70%
KoVAE0.3750.4100.499
Ours0.1550.2880.392
+ +# 5.3 Qualitative Evaluation + +We evaluate the similarity between the generated sequences and the real data using qualitative metrics. Specifically, we employ two visualization techniques [58]: (i) projecting the real and synthetic data into a two-dimensional space using t-SNE [55], and (ii) performing kernel density estimation to visualize the corresponding probability density functions (PDFs). Fig. 4 illustrates these results for the $70\%$ missing rate setting over various datasets and sequence lengths: Energy (24), Weather (96), and Stocks (768). The top row shows the two-dimensional point clouds of real (blue), KoVAE (green), and our data obtained via t-SNE, while the bottom row displays their respective PDFs. Overall, our approach demonstrates strong alignment in both visualizations. In the t-SNE plots (top row), a high degree of overlap between real and synthetic samples is observed. Similarly, in the PDF plots (bottom row), the trends and behaviors of the distributions are closely aligned. For additional results, including the regular and irregular $30\%$ and $50\%$ settings, please refer to App. E.6. + +Table 1: Train time (hours) for lengths (24, 96, 768), averaged over missing rates (30%, 50%, 70%). + +
ModelETTh1ETTh2ETTm1ETTm2WeatherElectricityEnergySineStockMujoco
24GT-GAN7.443.685.143.605.046.455.254.093.152.17
KoVAE6.4910.7910.146.555.84016.579.315.592.041.15
Ours1.282.760.961.433.662.441.000.480.210.60
96KoVAE19.7010.8214.6126.5122.06-13.6810.766.45-
Ours1.521.901.241.851.87-1.721.460.76-
768KoVAE31.5337.6667.9772.5852.93-21.0414.6416.27-
Ours5.382.6112.207.499.47-4.968.242.74-
+ +Table 2: Averaged results over ${30}\% ,{50}\% ,{70}\%$ missing rates for length 24. Lower values are better. + +
ModelETTh1ETTh2ETTm1ETTm2WeatherElectricityEnergySineStock
Disc.TimeGAN-Δt0.4990.4990.4990.4990.4970.4990.4740.4970.479
GT-GAN0.4710.3690.4120.3660.4810.4270.3250.3380.249
KoVAE0.1970.0810.0500.0670.3320.4980.3230.0430.118
Ours0.0370.0090.0120.0110.0570.3840.0800.0100.008
Pred.TimeGAN-Δt0.2670.3360.2350.3140.3940.2620.4570.3340.072
GT-GAN0.1860.0920.1250.0940.1450.1480.0690.0960.020
KoVAE0.0570.0540.0450.0500.0570.0470.0500.0740.017
Ours0.0530.0460.0440.0440.0220.0490.0470.0690.012
FIDTimeGAN-Δt3.1403.1993.4193.2182.37823.396.5072.7802.668
GT-GAN2.2128.63514.296.3852.7589.9931.5311.6982.181
KoVAE1.5180.2480.1800.2803.6996.1630.6290.0370.369
Ours0.1240.0350.0470.0240.1703.5800.1320.0150.036
Corr.TimeGAN-Δt3.7431.0512.3500.5791.20013.243.7652.4241.399
GT-GAN7.1480.9162.4670.3560.79114.923.8893.2820.261
KoVAE0.1830.1770.1300.2622.8994.2832.6300.0410.064
Ours0.0840.0540.0650.0390.3962.0310.9220.0150.019
+ +Table 3: Averaged results over ${30}\% ,{50}\% ,{70}\%$ missing rates for length 96 (top) and 768 (bottom). + +
ModelETTh1ETTh2ETTm1ETTm2WeatherEnergySineStock
Length = 96Disc:KoVAE0.2840.1030.2760.0890.3410.3560.2390.099
Ours0.0700.0530.0400.0300.1520.1850.0030.017
Pred:KoVAE0.0620.0570.0540.0500.0460.0800.1650.020
Ours0.0530.0490.0450.0440.0250.0490.1550.011
FIDKoVAE5.8421.1114.0700.9963.6814.6944.3810.868
Ours0.3570.5080.1710.1420.3940.3090.0170.109
Corr:KoVAE0.2240.3620.1750.4352.6094.8100.2220.089
Ours0.1140.1510.0920.0940.8901.1610.0160.014
Length = 768Disc:KoVAE0.2380.2010.2360.1960.4280.3840.3500.284
Ours0.0880.0450.0580.0520.1020.2130.0060.022
Pred:KoVAE0.0720.0690.0600.0760.0700.0870.2260.031
Ours0.0530.0560.0470.0500.0270.0420.2040.013
FIDKoVAE13.928.30413.508.27917.4924.5138.607.273
Ours0.8820.4390.3910.2640.3180.7960.2370.160
Corr:KoVAE0.3330.6060.4040.7383.2528.7520.3790.046
Ours0.1030.1060.1220.1280.4581.0400.0060.026
+ +# 5.4 Irregularly-sampled data under noise + +Our work primarily addresses irregularly sampled time series data. However, in real-world scenarios, such data often includes noise due to sensor limitations and inaccuracies. To further enhance our quantitative evaluation framework, we tackle the generative challenges of learning from irregular time series in noisy environments. Specifically, we propose a novel setup to evaluate the model's capability to recover the true underlying distribution from data corrupted by both irregular sampling and Gaussian noise. In this setup, we simulate a $50\%$ missing rate and inject additive Gaussian noise sampled from a normal distribution $\mathcal{N}(0,\sigma)$ , where $\sigma$ corresponds to the specified noise level (e.g., 0.1, 0.15, and 0.2). Importantly, this noise is added independently of the original data distribution or scale. The evaluation is conducted on sequences of length 24 across four datasets: Weather, Etth1, Stocks, and Energy. Following the discriminative and predictive evaluation protocols for + +Table 5: Discriminative and predictive scores for $50\%$ missing rate on Weather, ETTh1, Stock, and Energy datasets with injected noise levels (0.1, 0.15, and 0.2). + +
N/RModelWeatherETTh1StockEnergy
Disc.Pred.Disc.Pred.Disc.Pred.Disc.Pred.
0.1KoVAE0.4260.0560.2250.0730.2350.0160.4340.067
Ours0.0610.0520.0240.0340.0070.0120.0650.047
0.15KoVAE0.4880.0920.3770.0770.3410.0920.4930.093
Ours0.4160.0290.4070.0590.2820.0230.4670.053
0.2KoVAE0.4910.0960.4400.0840.3520.1210.4960.123
Ours0.4850.0350.4560.0620.3400.0270.4570.057
+ +![](images/031fcf866b0f57f152bd86ff9f09bae23f3fc58c3e451b7ed0f4db920246fa00.jpg) + +![](images/1bcf68df947235437ffe54eb6bfb505c74f5197313ec4f8f5b19d2c3b76ac0c4.jpg) + +![](images/4f0dadad8e9b404734df11ef6cd424a05413167ec5a906bf7940af18621bf6d2.jpg) + +![](images/82097aaabd036106e54d13b154a5a69509a36b83504660012818db58706eb8ba.jpg) +Figure 4: 2D t-SNE embeddings (top) and probability density functions (bottom) for real data vs. synthetic data from our method and KoVAE, under a $70\%$ missing rate. From left to right: Energy (length 24), Weather (length 96), and Stock (length 768) datasets. + +![](images/bdfee733b9765a1590c6112ac5f16bd593fa407ebecde7bbf32a1bee402c951c.jpg) + +![](images/c77a95a9a9f946c55055aae2c91da51d81b011964e33732f457a5e29bd9d4ab5.jpg) + +Table 6: Discriminative and predictive scores for $40\%$ random vs. $40\%$ continuous missingness on the Weather and Energy datasets with sequence lengths of 24 and 96. Lower is better. + +
MetricWeatherEnergy
RandomContinuousRandomContinuous
Length = 24Discriminative ↓0.0570.0530.0810.082
Predictive ↓0.0210.0250.0470.043
Length = 96Discriminative ↓0.1540.1650.1850.191
Predictive ↓0.0250.0230.0480.050
+ +$50\%$ missing rate described earlier, we compare our approach against the most recent state-of-the-art method, KoVAE. Tab. 5 presents the results. For each noise rate (N/R), we report the discriminative (Disc.) and predictive (Pred.) scores, where lower values indicate better performance. Our method consistently outperforms KoVAE, achieving significant improvements. Specifically, we observe an average relative improvement of $26.8\%$ in the discriminative score and $50.3\%$ in the predictive score across all datasets and noise levels. + +# 5.5 Robustness to missingness patterns + +While many time series exhibit randomly missing values, in practice, missing values can also occur in contiguous blocks due to sensor outages or communication delays. To study this effect, we compare model performance under two types of $40\%$ missingness: (i) randomly dropped values, and (ii) continuous missing blocks. This experiment is conducted on the Weather and Energy datasets using sequence lengths of 24 and 96. Discriminative and predictive scores are reported in Tab. 6. The results show that our method maintains strong performance under both missingness types, with some cases slightly favoring random missingness and others slightly favoring block-missingness. Overall, these findings confirm that our approach is robust across a range of missingness patterns, effectively handling both sporadic and structured data gaps. + +# 5.6 Ablation Studies + +To better understand the contributions of each component in our proposed architecture, we conducted a series of ablation studies. We explore each of the components in our approach separately, and, additionally, we modify recent approaches to include our completion strategy. Specifically, we consider the following models: (i) KoVAE + TST, where the NCDE module in KoVAE is replaced by TST, as in our approach. (ii) TimeAutoDiff [51] + TST (iii): TransFusion [48] + TST; The latter two baselines are diffusion-based models designed for regular time series but not specifically for irregular time series data. Therefore, we used the TST module to impute the missing values. (iv) Mask Only, where the TST autoencoder is removed, and we only apply the masking mechanism. In this setup, missing values are imputed using unnatural neighbors by filling them with zeros. (v) Ours Without + +Table 7: Discriminative scores of the ablation study with $30\%$ , $50\%$ , and $70\%$ drop-rate on Energy and Stock datasets for sequence lengths of 24, 96, and 768. + +
Model30%50%70%
EnergyStockEnergyStockEnergyStock
Len. = 24KoVAE + TST0.3990.1090.4070.0640.4080.037
TimeAutoDiff + TST0.2930.1000.3290.1010.4680.375
TransFusion + TST0.2010.0500.2790.0580.4230.065
Ours (Mask Only)0.1570.0870.2690.1680.3720.237
Ours (Without Mask)0.1580.0250.3070.0450.4440.013
Ours0.0480.0070.0650.0070.1280.007
Len. = 96KoVAE + TST0.2400.1850.2540.2210.4170.193
TimeAutoDiff + TST0.2990.1050.3360.1040.4610.398
TransFusion + TST0.3050.0830.3350.0980.4420.116
Ours (Mask Only)0.4900.1740.4220.2630.4800.388
Ours (Without Mask)0.4020.0330.4760.0720.4910.082
Ours0.1300.0110.1530.0180.2720.021
Len. = 768KoVAE + TST0.3800.2250.4180.2430.3850.186
TimeAutoDiff + TST0.2990.1040.3340.1010.4660.487
TransFusion + TST0.3670.1130.3950.1210.4510.131
Ours (Mask Only)0.4370.2490.3490.4500.4350.491
Ours (Without Mask)0.3640.0270.3530.1020.3250.102
Ours0.1700.0250.2440.0330.2510.013
+ +Masking, where we leverage TST to complete missing values, and training is performed without masking. (vi) Our approach. + +We quantitatively evaluated each model under the same experimental conditions and show the results in Tab. 7. To provide an extensive analysis, our tests include several missing rates (30%, 50%, and 70%) using two datasets, Energy and Stock. Further, we measured the performance across different sequence lengths (24, 96, and 768). Our findings show that the combination of TST-based completion and masking yields superior performance compared to all other setups. Specifically, the Mask Only and Ours (Without Masking) setups showed significant limitations in capturing the true data distribution, while the replacement of NCDE with TST (KoVAE + TST) fell short in comparison to our proposed architecture. In particular, our results reveal that replacing the NCDE imputation component in KoVAE with the TST imputation mechanism is not the primary factor driving the significant improvements achieved by our method. Moreover, even when employing a powerful time series diffusion-based model like TransFusion combined with TST-based imputation, performance significantly degrades, struggling to capture the true distribution of the regular data. Overall, these results highlight the critical role of masking during the diffusion process and the importance of leveraging completion as a guide rather than a direct substitute for the true distribution. + +We also ablate the impact of different image transformations on model performance, evaluating four methods: vanilla folding (reshapes a sequence into a fixed-size matrix with zero-padding), Gramian Angular Field, basic delay embedding, and our proposed inverse delay embedding (see App. A). Results in Tab. 8a and Tab. 9 show that geometric approaches—vanilla folding and our enhanced DE—better suit our method due to their structural clarity, where each pixel maps directly to a time point, facilitating mask usage and improving both discriminative and predictive performance. In contrast, GAF does not scale well to long sequences due to large image size. Our inverse transform also outperforms the original inverse of ImagenTime [38]. + +We additionally conducted an extensive ablation study comparing a variety of completion strategies. These included simple methods such as, Gaussian noise (GN), zero-filling, linear interpolation (LI), and polynomial interpolation (PI); probabilistic techniques like stochastic imputation (SI) (sampling from a Gaussian distribution fitted to the non-NaN values in each slice); and more advanced learning-based approaches, including NCDE, CSDI [53], and our proposed Time Series Transformer (TST) completion. Our results in Tab. 8b show that when the neighbors are not natural—such as in the case of zero completion or Gaussian noise completion—the model struggles more to generate data that closely follows the true distribution. In contrast, when using more natural completions (e.g., polynomial, stochastic imputation, NCDE, CSDI, TST), the model consistently obtains very good results. This confirms that generating natural neighborhoods indeed enhances the generative quality without making the model completely reliant on the imputation quality. + +Table 8: Comparative performance on Energy and Stock datasets for sequence length 24. +(a) Ablation study on image transformation methods. + +
Model30%50%70%
EnergyStockEnergyStockEnergyStock
Disc.Gramian Angular0.2910.0610.3130.1570.3630.183
Vanilla Folding0.0580.0050.0500.0090.1360.010
Basic DE0.0910.0350.1020.0460.1530.019
Ours DE0.0480.0070.0650.0070.1280.007
Pred.Gramian Angular0.0490.0130.0480.0150.0490.016
Vanilla Folding0.0470.0130.0470.0130.0470.014
Basic DE0.0530.0220.0510.0250.0550.027
Ours DE0.0470.0120.0470.0120.0470.011
+ +(b) Imputation methods with $50\%$ drop-rate. + +
ModelDisc.Pred.
EnergyStockEnergyStock
GN→NaN0.4570.1020.0580.014
0→NaN0.2690.1580.0510.014
LI0.2510.0130.0490.019
PI0.2010.0120.0530.016
NCDE0.1020.0130.0580.013
CSDI0.0880.0120.0480.013
SI0.0690.0100.0470.013
Ours (TST)0.0650.0070.0470.012
+ +# 6 Conclusions + +In this work, we introduced a novel two-step framework for generating realistic regular time series from irregularly sampled sequences. By integrating a Time Series Transformer (TST) for completion with a vision-based diffusion model leveraging masking, we effectively addressed the challenge of unnatural neighborhoods inherent in direct masking approaches. This hybrid strategy ensures that the diffusion model benefits from more structured and meaningful input while mitigating over-reliance on completed values. Our extensive evaluations across multiple benchmarks demonstrated state-of-the-art performance, with improvements of up to $70\%$ in discriminative score and an $85\%$ reduction in computational cost over prior methods. Furthermore, our approach scales effectively to long time series, significantly outperforming existing generative models in both accuracy and efficiency. + +Beyond these advancements, our work highlights the broader potential of integrating completion and masking strategies in generative modeling, particularly in domains where irregular sampling and missing values are prevalent. Future directions include extending our framework to multimodal time series generation, exploring self-supervised objectives for improved imputation, and integrating adaptive masking techniques that dynamically adjust completion reliance. By bridging the gap between irregular and regular time series generation, our method opens new possibilities for high-fidelity synthetic data generation in critical fields such as healthcare, finance, and climate science. + +# Acknowledgments + +This research was partially supported by the Lynn and William Frankel Center of the Computer Science Department, Ben-Gurion University of the Negev, an ISF grant 668/21, an ISF equipment grant, and by the Israeli Council for Higher Education (CHE) via the Data Science Research Center, Ben-Gurion University of the Negev, Israel. + +# References + +[1] J. Allen. Short term spectral analysis, synthesis, and modification by discrete Fourier transform. IEEE transactions on acoustics, speech, and signal processing, 25(3):235-238, 1977. +[2] J. B. Allen and L. R. Rabiner. A unified approach to short-time Fourier analysis and synthesis. Proceedings of the IEEE, 65(11):1558-1564, 1977. +[3] N. Berman, O. Joglekar, E. Kosman, D. Di Castro, and O. Azencot. Towards general modality translation with contrastive and predictive latent diffusion bridge. In Advances in Neural Information Processing Systems (NeurIPS) 39, 2025. +[4] N. Berman, E. Kosman, D. D. Castro, and O. Azencot. Reviving life on the edge: Joint score-based graph generation of rich edge attributes. Trans. Mach. Learn. Res., 2025. +[5] N. Berman, I. Naiman, M. Eliasof, H. Zisling, and O. Azencot. One-step offline distillation of diffusion-based models via Koopman modeling. In Advances in Neural Information Processing Systems (NeurIPS) 39, 2025. +[6] E. Brophy, Z. Wang, Q. She, and T. Ward. Generative adversarial networks in time series: A systematic literature review. ACM Computing Surveys, 55(10):1-31, 2023. +[7] E. Brophy, Z. Wang, and T. E. Ward. Quick and easy time series generation with established image-based GANs. arXiv preprint arXiv:1902.05624, 2019. +[8] L. M. Candanedo, V. Feldheim, and D. Deramaix. Data driven prediction models of energy use of appliances in a low-energy house. Energy and buildings, 140:81-97, 2017. +[9] Y. Chen, K. Ren, Y. Wang, Y. Fang, W. Sun, and D. Li. ContiFormer: Continuous-time transformer for irregular time series modeling. Advances in Neural Information Processing Systems, 37, 2023. +[10] Z. Chen, Y. Wu, Y. Leng, J. Chen, H. Liu, X. Tan, Y. Cui, K. Wang, L. He, S. Zhao, J. Bian, and D. P. Mandic. ResGrad: Residual denoising diffusion probabilistic models for text to speech. arXiv preprint arXiv:2212.14518, 2022. +[11] A. Coletta, S. Gopalakrishnan, D. Borrajo, and S. Vyetrenko. On the constrained time-series generation problem. Advances in Neural Information Processing Systems, 37, 2023. +[12] C. Corneanu, R. Gadde, and A. M. Martinez. LatentPaint: Image inpainting in latent space with diffusion models. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 4334-4343, 2024. +[13] A. Desai, C. Freeman, Z. Wang, and I. Beaver. TimeVAE: A variational auto-encoder for multivariate time series generation. arXiv preprint arXiv:2111.08095, 2021. +[14] P. Dhariwal and A. Nichol. Diffusion models beat GANs on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. +[15] P. Flandrin, G. Rilling, and P. Goncalves. Empirical mode decomposition as a filter bank. IEEE signal processing letters, 11(2):112-114, 2004. +[16] T. Gonen, I. Pemper, I. Naiman, N. Berman, and O. Azencot. Time series generation under data scarcity: A unified generative modeling approach. In Advances in Neural Information Processing Systems (NeurIPS) 39, 2025. +[17] I. J. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. C. Courville, and Y. Bengio. Generative adversarial nets. In Advances in Neural Information Processing Systems 27, pages 2672–2680, 2014. +[18] Z. Han, J. Zhao, H. Leung, K. F. Ma, and W. Wang. A review of deep learning models for time series prediction. IEEE Sensors Journal, 21(6):7833-7848, 2019. +[19] N. Hatami, Y. Gavet, and J. Debayle. Classification of time-series images using deep convolutional neural networks. In Tenth international conference on machine vision (ICMV 2017), volume 10696, pages 242-249. SPIE, 2018. + +[20] J. Hellermann and S. Lessmann. Leveraging image-based generative adversarial networks for time series generation. arXiv preprint arXiv:2112.08060, 2021. +[21] J. Ho, A. Jain, and P. Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. +[22] J. Ho, T. Salimans, A. Gritsenko, W. Chan, M. Norouzi, and D. J. Fleet. Video diffusion models. Advances in Neural Information Processing Systems, 35:8633-8646, 2022. +[23] J. Hogue. Metro Interstate Traffic Volume. UCI Machine Learning Repository, 2019. DOI: https://doi.org/10.24432/C5X60B. +[24] H. Ismail Fawaz, G. Forestier, J. Weber, L. Idoumghar, and P.-A. Muller. Deep learning for time series classification: a review. Data mining and knowledge discovery, 33(4):917-963, 2019. +[25] P. Jeha, M. Bohlke-Schneider, P. Mercado, S. Kapoor, R. S. Nirwan, V. Flunkert, J. Gasthaus, and T. Januschowski. PSA-GAN: Progressive self attention GANs for synthetic time series. In The Tenth International Conference on Learning Representations, 2022. +[26] J. Jeon, J. Kim, H. Song, S. Cho, and N. Park. GT-GAN: General purpose time series synthesis with generative adversarial networks. Advances in Neural Information Processing Systems, 35:36999-37010, 2022. +[27] T. Karras, M. Aittala, T. Aila, and S. Laine. Elucidating the design space of diffusion-based generative models. Advances in neural information processing systems, 35:26565-26577, 2022. +[28] P. Kidger, J. Morrill, J. Foster, and T. Lyons. Neural controlled differential equations for irregular time series. Advances in Neural Information Processing Systems, 33:6696-6707, 2020. +[29] D. P. Kingma and M. Welling. Auto-encoding variational Bayes. In 2nd International Conference on Learning Representations, ICLR, 2014. +[30] F. Kreuk, G. Synnaeve, A. Polyak, U. Singer, A. Defossez, J. Copet, D. Parikh, Y. Taigman, and Y. Adi. AudioGen: Textually guided audio generation. In The Eleventh International Conference on Learning Representations, ICLR, 2023. +[31] X. Li, V. Metsis, H. Wang, and A. H. H. Ngu. TTS-GAN: A transformer-based time-series generative adversarial network. In International conference on artificial intelligence in medicine, pages 133-143. Springer, 2022. +[32] Z. Li, S. Li, and X. Yan. Time series as images: Vision transformer for irregularly sampled time series. Advances in Neural Information Processing Systems, 36, 2023. +[33] S. Liao, H. Ni, L. Szpruch, M. Wiese, M. Sabate-Vidales, and B. Xiao. Conditional Sig-Wasserstein GANs for time series generation. arXiv preprint arXiv:2006.05421, 2020. +[34] B. Lim and S. Zohren. Time-series forecasting with deep learning: a survey. Philosophical Transactions of the Royal Society A, 379(2194):20200209, 2021. +[35] Y. Lipman, R. T. Q. Chen, H. Ben-Hamu, M. Nickel, and M. Le. Flow matching for generative modeling. In The Eleventh International Conference on Learning Representations, ICLR, 2023. +[36] H. Liu, Z. Chen, Y. Yuan, X. Mei, X. Liu, D. Mandic, W. Wang, and M. D. Plumbley. AudioLDM: Text-to-audio generation with latent diffusion models. In Proceedings of the 40th International Conference on Machine Learning, volume 202 of Proceedings of Machine Learning Research, pages 21450-21474. PMLR, 23-29 Jul 2023. +[37] A. Lugmayr, M. Danelljan, A. Romero, F. Yu, R. Timofte, and L. Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11461-11471, 2022. +[38] I. Naiman, N. Berman, I. Pemper, I. Arbiv, G. Fadlon, and O. Azencot. Utilizing image transforms and diffusion models for generative modeling of short and long time series. Advances in Neural Information Processing Systems, 38, 2024. + +[39] I. Naiman, N. B. Erichson, P. Ren, M. W. Mahoney, and O. Azencot. Generative modeling of regular and irregular time series data via Koopman VAEs. In The Twelfth International Conference on Learning Representations, ICLR, 2024. +[40] L. Nochumsohn and O. Azencot. Data augmentation policy search for long-term forecasting. Trans. Mach. Learn. Res., 2025. +[41] W. Peebles and S. Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. +[42] V. Popov, I. Vovk, V. Gogoryan, T. Sadekova, and M. Kudinov. Grad-TTS: A diffusion probabilistic model for text-to-speech. In International Conference on Machine Learning, pages 8599-8608. PMLR, 2021. +[43] K. Rasul, C. Seward, I. Schuster, and R. Vollgraf. Autoregressive denoising diffusion models for multivariate probabilistic time series forecasting. In International Conference on Machine Learning, pages 8857-8868. PMLR, 2021. +[44] P. Ren, R. Nakata, M. Lacour, I. Naiman, N. Nakata, J. Song, Z. Bi, O. A. Malik, D. Morozov, O. Azencot, et al. Learning physics for unveiling hidden earthquake ground motions via conditional generative modeling. arXiv preprint arXiv:2407.15089, 2024. +[45] R. Rombach, A. Blattmann, D. Lorenz, P. Esser, and B. Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. +[46] Y. Rubanova, R. T. Chen, and D. K. Duvenaud. Latent ordinary differential equations for irregularly-sampled time series. Advances in neural information processing systems, 32, 2019. +[47] M. Schirmer, M. Eltayeb, S. Lessmann, and M. Rudolph. Modeling irregular time series with continuous recurrent units. In International conference on machine learning, pages 19388-19405. PMLR, 2022. +[48] M. F. Sikder, R. Ramachandranpillai, and F. Heintz. Transfusion: Generating long, high fidelity time series using diffusion models with transformers. arXiv preprint arXiv:2307.12667, 2023. +[49] J. Sohl-Dickstein, E. Weiss, N. Maheswaranathan, and S. Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. +[50] Y. Song, J. Sohl-Dickstein, D. P. Kingma, A. Kumar, S. Ermon, and B. Poole. Score-based generative modeling through stochastic differential equations. In 9th International Conference on Learning Representations, ICLR, 2021. +[51] N. Suh, Y. Yang, D.-Y. Hsieh, Q. Luan, S. Xu, S. Zhu, and G. Cheng. TimeAutoDiff: Combining autoencoder and diffusion model for time series tabular data synthesizing. arXiv preprint arXiv:2406.16028, 2024. +[52] F. Takens. Detecting strange attractors in turbulence. In Dynamical Systems and Turbulence, Warwick 1980: proceedings of a symposium held at the University of Warwick 1979/80, pages 366-381. Springer, 2006. +[53] Y. Tashiro, J. Song, Y. Song, and S. Ermon. CSDI: Conditional score-based diffusion models for probabilistic time series imputation. Advances in Neural Information Processing Systems, 34:24804-24816, 2021. +[54] E. Todorov, T. Erez, and Y. Tassa. MuJoCo: A physics engine for model-based control. In 2012 IEEE/RSJ international conference on intelligent robots and systems, pages 5026-5033. IEEE, 2012. +[55] L. Van der Maaten and G. Hinton. Visualizing data using t-SNE. Journal of machine learning research, 9(11), 2008. +[56] M. Vetterli and C. Herley. Wavelets and filter banks: Theory and design. IEEE transactions on signal processing, 40(9):2207-2232, 1992. + +[57] Z. Wang and T. Oates. Imaging time-series to improve classification and imputation. In Proceedings of the Twenty-Fourth International Joint Conference on Artificial Intelligence, IJCAI, pages 3939-3945. AAAI Press, 2015. +[58] J. Yoon, D. Jarrett, and M. Van der Schaar. Time-series generative adversarial networks. Advances in neural information processing systems, 32, 2019. +[59] X. Yuan and Y. Qiao. Diffusion-TS: Interpretable diffusion for general time series generation. In The Twelfth International Conference on Learning Representations, ICLR, 2024. +[60] Z. Yue, Y. Wang, J. Duan, T. Yang, C. Huang, Y. Tong, and B. Xu. TS2Vec: Towards universal representation of time series. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 8980-8987, 2022. +[61] G. Zerveas, S. Jayaraman, D. Patel, A. Bhamidipaty, and C. Eickhoff. A transformer-based framework for multivariate time series representation learning. In Proceedings of the 27th ACM SIGKDD conference on knowledge discovery & data mining, pages 2114–2124, 2021. +[62] H. Zhou, S. Zhang, J. Peng, S. Zhang, J. Li, H. Xiong, and W. Zhang. Informer: Beyond efficient transformer for long sequence time-series forecasting. In Proceedings of the AAAI conference on artificial intelligence, volume 35, pages 11106-11115, 2021. + +# NeurIPS Paper Checklist + +The checklist is designed to encourage best practices for responsible machine learning research, addressing issues of reproducibility, transparency, research ethics, and societal impact. Do not remove the checklist: The papers not including the checklist will be desk rejected. The checklist should follow the references and follow the (optional) supplemental material. The checklist does NOT count towards the page limit. + +Please read the checklist guidelines carefully for information on how to answer these questions. For each question in the checklist: + +- You should answer [Yes], [No], or [NA]. +- [NA] means either that the question is Not Applicable for that particular paper or the relevant information is Not Available. +- Please provide a short (1–2 sentence) justification right after your answer (even for NA). + +The checklist answers are an integral part of your paper submission. They are visible to the reviewers, area chairs, senior area chairs, and ethics reviewers. You will be asked to also include it (after eventual revisions) with the final version of your paper, and its final version will be published with the paper. + +The reviewers of your paper will be asked to use the checklist as one of the factors in their evaluation. While "[Yes]" is generally preferable to "[No]", it is perfectly acceptable to answer "[No]" provided a proper justification is given (e.g., "error bars are not reported because it would be too computationally expensive" or "we were unable to find the license for the dataset we used"). In general, answering "[No]" or "[NA]" is not grounds for rejection. While the questions are phrased in a binary way, we acknowledge that the true answer is often more nuanced, so please just use your best judgment and write a justification to elaborate. All supporting evidence can appear either in the main paper or the supplemental material, provided in appendix. If you answer [Yes] to a question, in the justification please point to the section(s) where related material for the question can be found. + +IMPORTANT, please: + +- Delete this instruction block, but keep the section heading "NeurIPS Paper Checklist", +- Keep the checklist subsection headings, questions/answers and guidelines below. +- Do not modify the questions and only use the provided macros for your answers. + +# 1. Claims + +Question: Do the main claims made in the abstract and introduction accurately reflect the paper's contributions and scope? + +Answer: [Yes] + +Justification: The abstract and introduction clearly outline the problem of generating regular time series from irregular data and present the two-step method (completion + masking), which is consistently analyzed throughout the paper (see Abstract, Sec. 1, Sec. 4, and Sec. 5). + +Guidelines: + +- The answer NA means that the abstract and introduction do not include the claims made in the paper. +- The abstract and/or introduction should clearly state the claims made, including the contributions made in the paper and important assumptions and limitations. A No or NA answer to this question will not be perceived well by the reviewers. +- The claims made should match theoretical and experimental results, and reflect how much the results can be expected to generalize to other settings. +- It is fine to include aspirational goals as motivation as long as it is clear that these goals are not attained by the paper. + +# 2. Limitations + +Question: Does the paper discuss the limitations of the work performed by the authors? + +Answer: [Yes] + +Justification: The paper discusses the limitations of naive masking and inverse transformations in Sec. 4, and includes ablation studies in Sec. 5.6 to analyze failure cases and model sensitivity. + +# Guidelines: + +- The answer NA means that the paper has no limitation while the answer No means that the paper has limitations, but those are not discussed in the paper. +- The authors are encouraged to create a separate "Limitations" section in their paper. +- The paper should point out any strong assumptions and how robust the results are to violations of these assumptions (e.g., independence assumptions, noiseless settings, model well-specification, asymptotic approximations only holding locally). The authors should reflect on how these assumptions might be violated in practice and what the implications would be. +- The authors should reflect on the scope of the claims made, e.g., if the approach was only tested on a few datasets or with a few runs. In general, empirical results often depend on implicit assumptions, which should be articulated. +- The authors should reflect on the factors that influence the performance of the approach. For example, a facial recognition algorithm may perform poorly when image resolution is low or images are taken in low lighting. Or a speech-to-text system might not be used reliably to provide closed captions for online lectures because it fails to handle technical jargon. +- The authors should discuss the computational efficiency of the proposed algorithms and how they scale with dataset size. +- If applicable, the authors should discuss possible limitations of their approach to address problems of privacy and fairness. +- While the authors might fear that complete honesty about limitations might be used by reviewers as grounds for rejection, a worse outcome might be that reviewers discover limitations that aren't acknowledged in the paper. The authors should use their best judgment and recognize that individual actions in favor of transparency play an important role in developing norms that preserve the integrity of the community. Reviewers will be specifically instructed to not penalize honesty concerning limitations. + +# 3. Theory assumptions and proofs + +Question: For each theoretical result, does the paper provide the full set of assumptions and a complete (and correct) proof? + +Answer: [NA] + +Justification: The paper does not include theoretical results or formal theorems—it is an empirical study focused on generative modeling and experimental evaluation. + +# Guidelines: + +- The answer NA means that the paper does not include theoretical results. +- All the theorems, formulas, and proofs in the paper should be numbered and cross-referenced. +- All assumptions should be clearly stated or referenced in the statement of any theorems. +- The proofs can either appear in the main paper or the supplemental material, but if they appear in the supplemental material, the authors are encouraged to provide a short proof sketch to provide intuition. +- Inversely, any informal proof provided in the core of the paper should be complemented by formal proofs provided in appendix or supplemental material. +- Theorems and Lemmas that the proof relies upon should be properly referenced. + +# 4. Experimental result reproducibility + +Question: Does the paper fully disclose all the information needed to reproduce the main experimental results of the paper to the extent that it affects the main claims and/or conclusions of the paper (regardless of whether the code and data are provided or not)? + +Answer: [Yes] + +Justification: The paper provides extensive implementation details including training parameters, model architecture (Sec. 4, App. B, and hyperparameter tables in the appendix), datasets used, evaluation metrics, and ablation studies. + +# Guidelines: + +- The answer NA means that the paper does not include experiments. +- If the paper includes experiments, a No answer to this question will not be perceived well by the reviewers: Making the paper reproducible is important, regardless of whether the code and data are provided or not. +- If the contribution is a dataset and/or model, the authors should describe the steps taken to make their results reproducible or verifiable. +- Depending on the contribution, reproducibility can be accomplished in various ways. For example, if the contribution is a novel architecture, describing the architecture fully might suffice, or if the contribution is a specific model and empirical evaluation, it may be necessary to either make it possible for others to replicate the model with the same dataset, or provide access to the model. In general, releasing code and data is often one good way to accomplish this, but reproducibility can also be provided via detailed instructions for how to replicate the results, access to a hosted model (e.g., in the case of a large language model), releasing of a model checkpoint, or other means that are appropriate to the research performed. +- While NeurIPS does not require releasing code, the conference does require all submissions to provide some reasonable avenue for reproducibility, which may depend on the nature of the contribution. For example +(a) If the contribution is primarily a new algorithm, the paper should make it clear how to reproduce that algorithm. +(b) If the contribution is primarily a new model architecture, the paper should describe the architecture clearly and fully. +(c) If the contribution is a new model (e.g., a large language model), then there should either be a way to access this model for reproducing the results or a way to reproduce the model (e.g., with an open-source dataset or instructions for how to construct the dataset). +(d) We recognize that reproducibility may be tricky in some cases, in which case authors are welcome to describe the particular way they provide for reproducibility. In the case of closed-source models, it may be that access to the model is limited in some way (e.g., to registered users), but it should be possible for other researchers to have some path to reproducing or verifying the results. + +# 5. Open access to data and code + +Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material? + +# Answer: [No] + +Justification: Code and data will be released after the review process concludes (post-acceptance or rejection), as noted by the authors. + +# Guidelines: + +- The answer NA means that paper does not include experiments requiring code. +- Please see the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details. +- While we encourage the release of code and data, we understand that this might not be possible, so "No" is an acceptable answer. Papers cannot be rejected simply for not including code, unless this is central to the contribution (e.g., for a new open-source benchmark). +- The instructions should contain the exact command and environment needed to run to reproduce the results. See the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details. +- The authors should provide instructions on data access and preparation, including how to access the raw data, preprocessed data, intermediate data, and generated data, etc. + +- The authors should provide scripts to reproduce all experimental results for the new proposed method and baselines. If only a subset of experiments are reproducible, they should state which ones are omitted from the script and why. +- At submission time, to preserve anonymity, the authors should release anonymized versions (if applicable). +- Providing as much information as possible in supplemental material (appended to the paper) is recommended, but including URLs to data and code is permitted. + +# 6. Experimental setting/details + +Question: Does the paper specify all the training and test details (e.g., data splits, hyperparameters, how they were chosen, type of optimizer, etc.) necessary to understand the results? + +Answer: [Yes] + +Justification: Training/test splits, hyperparameters, sampling steps, and optimizer settings are provided in Sec. 5 and in App. D.3, Tables 10-12. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The experimental setting should be presented in the core of the paper to a level of detail that is necessary to appreciate the results and make sense of them. +- The full details can be provided either with the code, in appendix, or as supplemental material. + +# 7. Experiment statistical significance + +Question: Does the paper report error bars suitably and correctly defined or other appropriate information about the statistical significance of the experiments? + +Answer: [Yes] + +Justification: The paper provides average results over multiple random seeds and missing rates (e.g., $30\%$ , $50\%$ , $70\%$ ), and includes standard deviation in extended tables in the appendix (e.g., Tab. 18, Tab. 20). + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The authors should answer "Yes" if the results are accompanied by error bars, confidence intervals, or statistical significance tests, at least for the experiments that support the main claims of the paper. +- The factors of variability that the error bars are capturing should be clearly stated (for example, train/test split, initialization, random drawing of some parameter, or overall run with given experimental conditions). +- The method for calculating the error bars should be explained (closed form formula call to a library function, bootstrap, etc.) +- The assumptions made should be given (e.g., Normally distributed errors). +- It should be clear whether the error bar is the standard deviation or the standard error of the mean. +- It is OK to report 1-sigma error bars, but one should state it. The authors should preferably report a 2-sigma error bar than state that they have a $96\%$ CI, if the hypothesis of Normality of errors is not verified. +- For asymmetric distributions, the authors should be careful not to show in tables or figures symmetric error bars that would yield results that are out of range (e.g. negative error rates). +- If error bars are reported in tables or plots, The authors should explain in the text how they were calculated and reference the corresponding figures or tables in the text. + +# 8. Experiments compute resources + +Question: For each experiment, does the paper provide sufficient information on the computer resources (type of compute workers, memory, time of execution) needed to reproduce the experiments? + +# Answer: [Yes] + +Justification: Training and inference times are detailed and compared to baselines in Sec. 5.1 and App. E.7, with hardware specs (e.g., RTX3090) provided for fair benchmarking. + +# Guidelines: + +- The answer NA means that the paper does not include experiments. +- The paper should indicate the type of compute workers CPU or GPU, internal cluster, or cloud provider, including relevant memory and storage. +- The paper should provide the amount of compute required for each of the individual experimental runs as well as estimate the total compute. +- The paper should disclose whether the full research project required more compute than the experiments reported in the paper (e.g., preliminary or failed experiments that didn't make it into the paper). + +# 9. Code of ethics + +Question: Does the research conducted in the paper conform, in every respect, with the NeurIPS Code of Ethics https://neurips.cc/public/EthicsGuidelines? + +# Answer: [Yes] + +Justification: The paper does not raise ethical concerns; all datasets are public, and the work does not involve sensitive data or human participants. + +# Guidelines: + +- The answer NA means that the authors have not reviewed the NeurIPS Code of Ethics. +- If the authors answer No, they should explain the special circumstances that require a deviation from the Code of Ethics. +- The authors should make sure to preserve anonymity (e.g., if there is a special consideration due to laws or regulations in their jurisdiction). + +# 10. Broader impacts + +Question: Does the paper discuss both potential positive societal impacts and negative societal impacts of the work performed? + +# Answer: [Yes] + +Justification: The paper highlights potential applications in healthcare, finance, and climate science (see Abstract and Sec. 1), and discusses risks related to data imputation and overreliance on incorrect completions (Sec. 4). + +# Guidelines: + +- The answer NA means that there is no societal impact of the work performed. +- If the authors answer NA or No, they should explain why their work has no societal impact or why the paper does not address societal impact. +- Examples of negative societal impacts include potential malicious or unintended uses (e.g., disinformation, generating fake profiles, surveillance), fairness considerations (e.g., deployment of technologies that could make decisions that unfairly impact specific groups), privacy considerations, and security considerations. +- The conference expects that many papers will be foundational research and not tied to particular applications, let alone deployments. However, if there is a direct path to any negative applications, the authors should point it out. For example, it is legitimate to point out that an improvement in the quality of generative models could be used to generate deepfakes for disinformation. On the other hand, it is not needed to point out that a generic algorithm for optimizing neural networks could enable people to train models that generate Deepfakes faster. +- The authors should consider possible harms that could arise when the technology is being used as intended and functioning correctly, harms that could arise when the technology is being used as intended but gives incorrect results, and harms following from (intentional or unintentional) misuse of the technology. + +- If there are negative societal impacts, the authors could also discuss possible mitigation strategies (e.g., gated release of models, providing defenses in addition to attacks, mechanisms for monitoring misuse, mechanisms to monitor how a system learns from feedback over time, improving the efficiency and accessibility of ML). + +# 11. Safeguards + +Question: Does the paper describe safeguards that have been put in place for responsible release of data or models that have a high risk for misuse (e.g., pretrained language models, image generators, or scraped datasets)? + +Answer: [NA] + +Justification: The paper does not release a general-purpose model or scraped dataset and poses no high risk for misuse. + +Guidelines: + +- The answer NA means that the paper poses no such risks. +- Released models that have a high risk for misuse or dual-use should be released with necessary safeguards to allow for controlled use of the model, for example by requiring that users adhere to usage guidelines or restrictions to access the model or implementing safety filters. +- Datasets that have been scraped from the Internet could pose safety risks. The authors should describe how they avoided releasing unsafe images. +- We recognize that providing effective safeguards is challenging, and many papers do not require this, but we encourage authors to take this into account and make a best faith effort. + +# 12. Licenses for existing assets + +Question: Are the creators or original owners of assets (e.g., code, data, models), used in the paper, properly credited and are the license and terms of use explicitly mentioned and properly respected? + +Answer: [Yes] + +Justification: All datasets used are public and cited correctly (e.g., UCI, ETT, USHCN); the codebases cited are academic open-source projects (e.g., TS2Vec, ImagenTime). + +Guidelines: + +- The answer NA means that the paper does not use existing assets. +- The authors should cite the original paper that produced the code package or dataset. +- The authors should state which version of the asset is used and, if possible, include a URL. +- The name of the license (e.g., CC-BY 4.0) should be included for each asset. +- For scraped data from a particular source (e.g., website), the copyright and terms of service of that source should be provided. +- If assets are released, the license, copyright information, and terms of use in the package should be provided. For popular datasets, paperswithcode.com/datasets has curated licenses for some datasets. Their licensing guide can help determine the license of a dataset. +- For existing datasets that are re-packaged, both the original license and the license of the derived asset (if it has changed) should be provided. +- If this information is not available online, the authors are encouraged to reach out to the asset's creators. + +# 13. New assets + +Question: Are new assets introduced in the paper well documented and is the documentation provided alongside the assets? + +Answer: [NA] + +Justification: Although the paper proposes a new model architecture, no new dataset or code repository is released at submission time; these will be released post-review. + +Guidelines: + +- The answer NA means that the paper does not release new assets. +- Researchers should communicate the details of the dataset/code/model as part of their submissions via structured templates. This includes details about training, license, limitations, etc. +- The paper should discuss whether and how consent was obtained from people whose asset is used. +- At submission time, remember to anonymize your assets (if applicable). You can either create an anonymized URL or include an anonymized zip file. + +# 14. Crowdsourcing and research with human subjects + +Question: For crowdsourcing experiments and research with human subjects, does the paper include the full text of instructions given to participants and screenshots, if applicable, as well as details about compensation (if any)? + +Answer: [NA] + +Justification: The paper does not involve crowdsourcing or human subjects. + +Guidelines: + +- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. +- Including this information in the supplemental material is fine, but if the main contribution of the paper involves human subjects, then as much detail as possible should be included in the main paper. +- According to the NeurIPS Code of Ethics, workers involved in data collection, curation, or other labor should be paid at least the minimum wage in the country of the data collector. + +# 15. Institutional review board (IRB) approvals or equivalent for research with human subjects + +Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals (or an equivalent approval/review based on the requirements of your country or institution) were obtained? + +Answer: [NA] + +Justification: The paper does not involve human participants and does not require IRB approval. + +Guidelines: + +- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. +- Depending on the country in which research is conducted, IRB approval (or equivalent) may be required for any human subjects research. If you obtained IRB approval, you should clearly state this in the paper. +- We recognize that the procedures for this may vary significantly between institutions and locations, and we expect authors to adhere to the NeurIPS Code of Ethics and the guidelines for their institution. +- For initial submissions, do not include any information that would break anonymity (if applicable), such as the institution conducting the review. + +# 16. Declaration of LLM usage + +Question: Does the paper describe the usage of LLMs if it is an important, original, or non-standard component of the core methods in this research? Note that if the LLM is used only for writing, editing, or formatting purposes and does not impact the core methodology, scientific rigorousness, or originality of the research, declaration is not required. + +Answer: [NA] + +Justification: LLMs were not used in the methodology, only lightly for text editing and formatting. + +Guidelines: + +- The answer NA means that the core method development in this research does not involve LLMs as any important, original, or non-standard components. +- Please refer to our LLM policy (https://neurips.cc/Conferences/2025/LLM) for what should or should not be described. + +# A Time Series-to-Image Transformation + +In this section, we provide a brief discussion on time series to image conversions. We then introduce Delay Embedding, the transformation we employed. Additionally, we describe an improvement we implemented to enhance the reversibility of delay embedding. + +Time series to image conversion. The conversion of time series data into image representations has attracted significant interest for its ability to leverage computer vision methods in time series analysis. Techniques like Gramian Angular Fields [57], Recurrence Plots [19], and Line Graphs [32] allow for mapping time series data to visual formats, enabling tasks such as classification and imputation. In the field of speech analysis, the short-time Fourier transform (STFT) [1, 2, 56, 15] is crucial for capturing frequency variations over time, which is vital for processing audio and speech data. Recent advancements have also explored incorporating mel-spectrograms in diffusion models, particularly in connection with latent diffusion spaces[42, 10, 36]. Furthermore, generative approaches such as Wasserstein GANs [7, 20] have been applied to time series in image form. + +Vanilla Folding is a straightforward transformation. Given a time series $x$ , we reshape it into an image $x_{\mathrm{img}}$ by filling rows from left to right and moving to the next row upon reaching the end, padding with zeros if necessary. The inverse transformation reconstructs the original time series by reading the non-padded region row-wise. Despite its simplicity, this method scales well to very long sequences. Folding can also be interpreted as a specific case of delay embedding, as we explain below. + +Delay Embedding and Enhanced Reverse Transformation [52] converts a univariate time series $x_{1:L} \in \mathbb{R}^L$ into an image by structuring its information into columns and applying zero-padding as needed. This transformation is controlled by two hyperparameters: $m$ , which defines the skip value, and $n$ , which determines the number of columns. Given any channel of a time series, the corresponding matrix $X$ is formed as follows: + +$$ +X = \left[ \begin{array}{c c c c} x _ {1} & x _ {m + 1} & \ldots & x _ {L - n} \\ \vdots & \vdots & \ldots & \vdots \\ x _ {n} & x _ {n + m + 1} & \ldots & x _ {L} \end{array} \right] \in \mathbb {R} ^ {n \times q}, +$$ + +where $q = \lceil (L - n) / m\rceil$ . To match the input requirements of a neural network, the resulting image $x_{\mathrm{img}}$ is padded with zeros. This transformation is applied independently to each channel, and for multivariate time series, the matrices $X$ from different channels are concatenated along a new axis. Given an input signal $x\in \mathbb{R}^{L\times K}$ , the output is a transformed representation $x_{\mathrm{img}}\in \mathbb{R}^{K\times n\times q}$ , which is then zero-padded to obtain $x_{\mathrm{img}}\in \mathbb{R}^{K\times n\times n}$ . Delay embedding efficiently scales to long sequences; for instance, choosing $m = n = 256$ enables the encoding of sequences up to $65k$ in length using $256\times 256$ images. + +Our primary innovation lies in the reverse transformation process. In the original approach, only the first pixel corresponding to each time series value is used for reconstruction. Specifically, if $x_{i}$ is mapped to multiple image indices, the original method selects the first corresponding pixel in the image for reconstruction. In contrast, our approach aggregates information from all corresponding image indices by computing the average of the associated pixels for each $x_{i}$ . For a given $x_{1:L} \in \mathbb{R}^{L}$ , both methods ensure that $f^{-1}(f(x)) = x$ . + +As shown in Table 9, our approach consistently outperforms the original approach across various drop rates. For example, at a $30\%$ drop rate, our delay embedding (DE) approach achieves a discriminative score of 0.020 for the ETTh1 dataset and 0.009 for ETTh2, compared to 0.023 and 0.018 with the naive approach, respectively. Similarly, at a $50\%$ drop rate, the new DE method reduces the discriminative score to 0.032 (ETTh1) and 0.005 (ETTh2), outperforming the previous DE approach, which yielded scores of 0.040 and 0.037, respectively. + +Table 9: Discriminative scores on ETTh1 and ETTh2 for sequence length 24 to compare the original inverse delay embedding vs. our inverse, evaluated in $30\%$ , $50\%$ , and $70\%$ missing rates. + +
Model30%50%70%
ETTh1ETTh2ETTh2ETTh2ETTh2ETTh2
Ours + Old inverse0.0230.0180.0400.0370.0670.048
Ours + New inverse0.0200.0090.0320.0050.0580.013
+ +![](images/fd3e481878fee097b2e0846396109998e22c3f0986d3d01e37eca45f57dcba7c.jpg) + +# B Training Losses + +Our model consists of two main components: an Autoencoder (AE) and a vision diffusion model (ImagenTime). For each component, we modified the loss to handle irregular data more effectively. + +# B.1 Autoencoder Training Loss + +The TST-Based AE is trained to reconstruct only the known (non-missing) values of the input data. Since we do not have access to the regularly sampled time series during training, the model learns to infer missing values from the irregularly sampled data. The masked reconstruction loss is defined as: + +$$ +\mathcal {L} _ {e} ^ {t = 0} = \frac {1}{| \mathcal {O} |} \sum_ {i \in \mathcal {O}} \left(\tilde {x} _ {i} - x _ {i}\right) ^ {2} \tag {1} +$$ + +where $x$ is the original input data, $\tilde{x}$ is the reconstructed output, $\mathcal{O}$ represents the set of observed (non-missing) indices in the input $x$ , and $|\mathcal{O}|$ is the number of observed values. This ensures that the loss function only penalizes reconstruction errors for the known values, without considering the missing ones. + +# B.2 ImagenTime Diffusion Training Loss + +At the core ofImagenTime is the generative diffusion model, which follows the framework of Karras et al. [27] for improved score-based modeling. The model employs a second-order ODE for the reverse process, balancing fast sampling and high-quality generations. + +For irregular time series data, we changed the training loss to ensure proper weighting of the diffusion steps and account for missing values. Given an input sequence $x$ and a corresponding mask $m$ indicating observed entries, the loss is defined as: + +$$ +\mathcal {L} _ {\text {d i f f}} = \mathbb {E} _ {x, \sigma} \left[ \left\| \left(D _ {\theta} (x + \sigma n, \sigma) - x\right) \cdot m \right\| ^ {2} \right] \tag {2} +$$ + +where: + +- $x$ is the input time series with missing values. +- $n \sim \mathcal{N}(0, I)$ is standard Gaussian noise, scaled by $\sigma$ . +- $m$ is the mask, indicating the observed values. + +The model reconstructs the output using both observed and imputed indices, treating the imputed values as natural neighbors to the observed ones. However, the loss is computed and compared only against the observed indices, ensuring that the comparison is made solely to the true distribution, unaffected by the imputed values. This allows the model to learn the distribution while maintaining accurate reconstruction of missing values. + +# C Inference Time Analysis + +In this section, we present a detailed comparison of the inference time between our approach and KoVAE, particularly in relation to sequence length. The evaluation is conducted across a range of sequence lengths: 24, 96, and 768. Additionally, we explore the relationship between inference time and sequence length for both models. + +# C.1 Inference Time vs. Sequence Length + +Figure 5 illustrates the relationship between inference time per sample (in seconds) and sequence length for both our model and KoVAE. While KoVAE demonstrates faster sampling on short sequences, its efficiency degrades rapidly as the sequence length increases. KoVAE, which is based on a sequential VAE architecture for time series, processes each time step individually throughout the entire sequence, causing its computational cost to increase significantly with longer sequences and resulting in substantially longer inference times. + +In contrast, our model maintains nearly constant inference time regardless of sequence length, making it highly efficient even for sequences as long as 5000 time steps. This is due to the use of delay embedding, which enables the model to compress long sequences into compact image representations with fixed dimensions. A clear turning point occurs at a sequence length of approximately 4500, beyond which our model consistently outperforms KoVAE in terms of inference time. This robust performance highlights the advantage of our method in terms of time efficiency, especially when dealing with long sequences. + +# C.2 Inference Time and Fidelity Comparison + +As shown in Figure 6, we evaluate the relationship between inference time and fidelity (indicated by the discriminative score) for a single sequence. Lower discriminative scores correspond to higher fidelity in the model's predictions. On the other hand, inference time is a measure of efficiency, with shorter times indicating greater computational efficiency. + +Our approach consistently performs well by maintaining a low discriminative score, which translates into higher fidelity in its predictions. This is achieved while also managing to keep the inference time relatively low, even for longer sequences. Notably, our model uses only 18 sampling steps regardless of sequence length, which results in stable and fast inference times across all configurations. + +# C.3 Performance Comparison + +As observed from both figures, our model exhibits a favorable trade-off between inference efficiency and fidelity. While KoVAE might be more efficient for shorter sequences, its performance degrades as sequence length increases, making it less suitable for long sequences. Our approach, however, remains consistently efficient and accurate, maintaining both low discriminative scores and linear inference times as sequence lengths scale. + +![](images/c71a482de0aa89ea04fd069dfd3de7757b1826f129ba4a9350d4b72c3c1bd959.jpg) +Figure 5: Comparison of inference time per sample in seconds vs. sequence length of our model and KoVae model. + +![](images/6966409aa1cabb3ba254c85ef8f846073fe2b655a5eb978fbe665c84d5d4bd27.jpg) +Figure 6: Comparison of discriminative score vs. inference time of a single sequence for our approach and KoVAE across different sequence lengths (24, 96, and 768). Lower discriminative scores indicate higher fidelity, and shorter inference times reflect greater efficiency. + +# D Experimental Setup + +# D.1 Baseline Methods + +We compare our method with several generative time series models designed for irregular data. KoVAE [39] is specifically designed to handle irregularly sampled time series effectively. Additionally, we consider GT-GAN [26], another method tailored for irregular time series generation. Lastly, we evaluate against TimeGAN- $\Delta t$ [58], a re-designed version of the original TimeGAN. Since TimeGAN does not natively support irregular time series, we follow GT-GAN and compare them with their re-designed versions. Specifically, extending regular approaches to support irregular TS requires the conversion of a dynamical module to its time-continuous version. We adapted it by converting its GRU layers to GRU- $\Delta t$ , enabling the model to exploit the time differences between observations and capture temporal dynamics. + +# D.2 Datasets + +We conduct experiments using a combination of synthetic and real-world datasets, each designed to evaluate the model under various conditions, including regular and irregular time-series settings. + +Sines This synthetic dataset contains 5 features, where each feature is independently generated using sinusoidal functions with different frequencies and phases. Specifically, for each feature $i \in \{1, \dots, 5\}$ , the time-series data is defined as $x_{i}(t) = \sin(2\pi f_{i}t + \theta_{i})$ , where $f_{i} \sim U[0,1]$ and $\theta_{i} \sim U[-\pi, \pi]$ . The dataset is characterized by its continuity and periodic properties, making it a suitable benchmark for evaluating the model's ability to handle structured time-series data. + +Stocks The Stocks dataset comprises daily historical Google stock price data from 2004 to 2019. It includes six features: high, low, opening, closing, adjusted closing prices, and trading volume. Unlike Sines, this dataset lacks periodicity and primarily exhibits random walk patterns. It is a real-world dataset commonly used to benchmark financial time-series forecasting and modeling. + +MuJoCo MuJoCo (Multi-Joint dynamics with Contact) is a versatile physics simulation framework used to generate multivariate time-series data [54]. The dataset contains 14 features representing state variables and control actions from simulated trajectories. This dataset is particularly suitable for evaluating models on dynamical systems and tasks involving physical interactions + +Energy The Energy dataset is a real-world multivariate time-series dataset [8] derived from the UCI Appliance Energy Prediction dataset. It includes 28 features, which are correlated and exhibit noisy periodicity and continuous-valued measurements. This dataset provides a challenging benchmark for forecasting and modeling tasks involving environmental and appliance energy consumption data. + +ETTh & ETTm The ETTh (Electricity Transformer Temperature - Hourly) and ETTm (Electricity Transformer Temperature - Minute) datasets [62] capture electricity load data from two power stations with varying temporal resolutions. These datasets are used for short- and long-term time-series forecasting tasks and are part of an established benchmark for evaluating generative and predictive models. + +Weather The Weather dataset includes daily meteorological measurements, such as temperature, precipitation, snowfall, snow depth, and minimum and maximum temperatures, collected from the United States Historical Climatology Network (USHCN) $^{2}$ . The dataset comprises measurements from 1,218 weather stations and is used for analyzing climatic trends and weather forecasting tasks. + +Electricity The Electricity dataset consists of electricity consumption data across multiple clients, represented as multivariate time-series. It is widely used for forecasting electricity loads and understanding temporal consumption patterns in energy-related applications. + +Traffic The Traffic dataset [23] contains hourly traffic volume data for westbound I-94 in the Minneapolis-St. Paul, MN area, collected from 2012 to 2018. It includes eight features, mixing numerical measurements (e.g., temperature, rainfall, snowfall, cloud coverage, traffic volume) with several categorical variables (e.g., holiday indicators, weather descriptions), making it one of the few benchmarks that requires models to handle both continuous and categorical time-series data. The dataset captures multivariate, sequential patterns influenced by weather and holiday effects, making it particularly suitable as a generative benchmark for modeling complex dependencies across heterogeneous features. + +# D.3 Metrics + +Discriminative Score This metric measures the ability of a model to differentiate between real and generated data. A lower discriminative score indicates that the generated data is more indistinguishable from the real data, reflecting better generative performance. This score is typically computed by training a binary classifier and evaluating its accuracy in distinguishing between the two datasets. A score close to random guessing suggests that the synthetic data is nearly indistinguishable from real data. + +Predictive Score The predictive score evaluates the quality of the generated data in terms of its utility for downstream predictive tasks. It is typically assessed using a supervised learning model trained on generated data and tested on real data, or vice versa. A higher predictive score indicates better alignment between real and generated distributions. + +Context-FID Score A lower Frechet Inception Distance (FID) score indicates that synthetic sequences are more similar to the original data distribution. Paul et al. (2022) introduced a variation of FID, called Context-FID (Context-Frechet Inception Distance), which replaces the Inception model in the original FID calculation with TS2Vec, a time series representation learning method [60]. Their findings suggest that models with the lowest Context-FID scores tend to achieve the best results in downstream tasks. Additionally, they demonstrated a strong correlation between the Context-FID score and the forecasting performance of generative models. To compute this score, synthetic and real time series samples are first generated, then encoded using a pre-trained TS2Vec model, after which the FID score is calculated based on the learned representations. + +Correlational Score Building on the approach from [33], we estimate the covariance between the $i^{th}$ and $j^{th}$ features of a time series using the following formula: + +$$ +\operatorname {c o v} _ {i, j} = \frac {1}{T} \sum_ {t = 1} ^ {T} X _ {t} ^ {i} X _ {t} ^ {j} - \left(\frac {1}{T} \sum_ {t = 1} ^ {T} X _ {t} ^ {i}\right) \left(\frac {1}{T} \sum_ {t = 1} ^ {T} X _ {t} ^ {j}\right). +$$ + +To quantify the correlation between real and synthetic data, we compute the following metric: + +$$ +\frac {1}{1 0} \sum_ {i, j} \left| \frac {\operatorname {c o v} _ {r} ^ {i , j}}{\sqrt {\operatorname {c o v} _ {r} ^ {i , i} \operatorname {c o v} _ {r} ^ {j , j}}} - \frac {\operatorname {c o v} _ {f} ^ {i , j}}{\sqrt {\operatorname {c o v} _ {f} ^ {i , i} \operatorname {c o v} _ {f} ^ {j , j}}} \right| +$$ + +# D.4 Hyperparameters. + +We summarize the key hyperparameters used in our framework in Tables 10, 11, and 12, corresponding to sequence lengths of 24, 96, and 768, respectively. The hyperparameters remain largely consistent across tasks, with variations in batch size, embedding dimensions, and image resolutions. We use the default EDM [27] sampler for all datasets and follow a unified configuration for the U-Net architecture in the diffusion model. For further details, refer to [27]. Additionally, all models were trained using the same learning rate schedule and optimization settings to ensure comparability across different sequence lengths. + +Table 10: Hyperparameter Settings for Sequence Length 24 Across Different Datasets + +
ETTh1ETTh2ETTm1ETTm2WeatherElectricityEnergySineStockMujoco
General
image size16 × 1616 × 1616 × 1616 × 88 × 88 × 88 × 88 × 88 × 88 × 8
learning rate10-410-410-410-410-410-410-410-410-410-4
batch size128128128128128128128128128128
teacher forcing rate000000000.20
DE
embedding(n)8888888888
delay(m)3333333333
TST
hidden_dim40404040404040404040
n_heads5555555555
num_layers6666666666
Diffusion
U-net channels12812812812812812812812812864
in channels[1,2,2,2][1,2,2,2][1,2,2,2][1,2,2,2][1,2,2,4][1,2,2,4][1,2,2,4][1,2,2,2][1,2,2,2][1,2,2,2]
attention revolution[8,4,2][8,4,2][8,4,2][8,4,2][8,4,2][8,4,2][8,4,2][8,4,2][8,4,2][8,4,2]
sampling steps18181818181818181818
+ +Table 11: Hyperparameter Settings for Sequence Length 96 Across Different Datasets + +
ETTh1ETTh2ETTm1ETTm2WeatherEnergySineStock
General
image size16 × 1616 × 1616 × 1616 × 1632 × 3232 × 3216 × 1616 × 16
learning rate10-410-410-410-410-410-410-410-4
batch size32641281283212812816
teacher forcing rate00000000.2
DE
embedding(n)1616161632321616
delay(m)6666242466
TST
hidden_dim4040404040404040
n_heads55555555
num_layers66666666
Diffusion
U-net channels128128128128128128128128
in channels[1,2,2,2][1,2,2,2][1,2,2,2][1,2,2,2][1,2,2,4][1,2,2,4][1,2,2,2][1,2,2,2]
attention revolution[16,8,4,2][16,8,4,2][16,8,4,2][16,8,4,2][32,16,8,4,2][32,16,8,4,2][16,8,4,2][16,8,4,2]
sampling steps1818181818181818
+ +Table 12: Hyperparameter Settings for Sequence Length 768 Across Different Datasets + +
ETTh1ETTh2ETTm1ETTm2WeatherEnergySineStock
General
image size32 × 3232 × 3232 × 3232 × 3232 × 3232 × 3232 × 3232 × 32
learning rate10-410-410-410-410-410-410-410-4
batch size3232323232163232
teacher forcing rate00000000.2
DE
embedding(n)3232323232323232
delay(m)2424242424242424
TST
hidden_dim4040404040404040
n_heads55555555
num_layers66666666
Diffusion
U-net channels128128128128128128128128
in channels[1,2,2,2][1,2,2,2][1,2,2,2][1,2,2,2][1,2,2,4][1,2,2,4][1,2,2,2][1,2,2,2]
attention revolution[32,16,8,4,2][32,16,8,4,2][32,16,8,4,2][32,16,8,4,2][32,16,8,4,2][32,16,8,4,2][32,16,8,4,2][32,16,8,4,2]
sampling steps1818181818181818
+ +![](images/68b8a08c7a2aa2d5926bfaf62e40b1b9f5e4a0b041c427b0c18c0e7efb298e92.jpg) +Figure 7: Unnatural vs. natural neighborhoods on stock data. A data point (A) is mapped to an image with zeros and its coordinates centered (B). Denoising the entire image produces inferior kernels (D) compared to masking (E). Constructing natural neighborhoods (C) yields consistent kernels and improved scores (F). + +# D.5 Natural vs. Unnatural Neighborhoods Experiment Setup + +In this section, we provide a detailed explanation of the experimental setup introduced in Sec. 4. + +Experiment Setup. We first generate 1000 two-dimensional samples $\{p\}$ drawn from a multivariate Gaussian distribution, with means centered at $(1,1)$ , $(1,-1)$ , $(-1,1)$ , and $(-1,-1)$ . A figure illustrating all sampled data appears in Figure 1A. To better simulate our real world environment, we transform each 2D datapoint into a $3\times 4$ image by setting all pixels to zero except those at the center, which correspond to the $x$ and $y$ coordinates of the original point (e.g., $s[1,1] = p[0]$ and $s[1,2] = p[1]$ ). Figure 1B depicts an example of this transformation. We refer to this dataset as $S_{\mathrm{irregular}}$ , as it simulates an "irregular" dataset containing zeros for missing values. + +Next, we construct a second dataset, $S_{\text{regular}}$ , in which the zero entries are replaced with linear or nonlinear transformations of $p[0]$ , $p[1]$ , or both. This is intended to emulate a data-imputation step performed on the first step of our method, yielding a more "natural" neighborhood of values. An data point example can be found in Figure 1C. We then compare three training setups: + +1. Train a diffusion model to predict the score across the entire image (i.e., noise prediction in diffusion) on $S_{\text{irregular}}$ . +2. Train a diffusion model to predict only the two central (coordinate) pixels, using our masking technique, on $S_{\mathrm{irregular}}$ . +3. Train the same masked model as in (2), but on $S_{\mathrm{regular}}$ , where we have "natural neighbors." + +All score models share a simple architecture: a Conv2D layer with a $3 \times 4$ kernel, followed by a ReLU activation, and a deconvolution layer that restores the original input size. + +# Evaluation. We employ two metrics: + +1. Score Estimation Loss: For fair comparison, we measure the score prediction error on the two central pixels (i.e., the original coordinates), regardless of the training strategy. +2. Kernel Analysis: We inspect the first-layer convolution kernels to determine which pixels the model focuses on. Since there are 64 output channels, we compute the $L_{1}$ norm at each spatial position across all channels and then average them. + +# E Additional Experiments and Analysis + +# E.1 "Unnatural" Neighborhoods: Stock Market Experiment + +As described in the main paper, we extend the synthetic experiment to a real-world dataset of stock prices. Implementation details appear in the main text; here we present the visual results (moved to the appendix due to space constraints). See Fig. 7. + +# E.2 Extending to Categorical and Numerical Features + +We further extend our evaluation to demonstrate that our method can handle mixed-type time series containing both numerical and categorical features. Categorical variables are mapped to learnable + +Table 13: Discriminative scores on mixed-type data (Traffic) with ${30}\% ,{50}\%$ ,and ${70}\%$ drop-rate for sequence lengths of 24, and 96 . + +
Model30%50%70%
Len. = 24GT-GAN0.4810.4730.485
KoVAE0.1540.1720.222
Ours0.0610.0640.087
Len. = 96GT-GAN0.4930.4910.488
KoVAE0.2120.2450.307
Ours0.0730.0910.102
+ +embedding vectors, which are jointly optimized with the generative model, transforming the input into a fully continuous representation that the diffusion process can operate on seamlessly. During generation, embeddings are mapped back to discrete categories using a simple postprocessing step. To validate this capability, we evaluated our method on the Traffic dataset which includes both numerical and categorical attributes, across multiple sequence lengths and missing rates. As reported in Tab. 13, our approach achieves the best discriminative scores compared to prior methods, demonstrating its ability to effectively capture temporal dependencies in mixed-type time series while maintaining superior generative performance. + +# E.3 Computational Efficiency of Completion Strategies + +In addition to performance, we also compare the computational efficiency of completion strategies. Both NCDE and CSDI rely on costly operations—NCDE requires repeated cubic spline evaluations, while CSDI involves up to 1000 sampling steps per imputation—which makes them slow or infeasible for long sequences or high-dimensional data. In contrast, TST is lightweight and scales efficiently while still achieving competitive or superior results. As shown in Tab. 14, TST consistently trains faster and imputes orders of magnitude quicker than NCDE and CSDI; for instance, on the Energy dataset with sequence length 768, NCDE could not be trained due to memory limits and CSDI required over 84 hours of training and 2394 minutes for imputing 1024 samples, whereas TST completed training in just 6.67 hours and imputed all samples in only 6.26 seconds. These findings highlight TST's role as a scalable and efficient completion module, combining both high-quality generative performance and practical usability. + +Table 14: Training (200 epochs) and imputation (1024 samples) times on RTX 3090. TST is significantly faster than NCDE and CSDI, especially for long sequences; NCDE times are omitted where training was infeasible. + +
DatasetSeq. LenTST (Ours)NCDECSDI
Train (h)Impute (s)Train (h)Impute (s)Train (h)Impute (min)
Energy240.670.642.552.41.6035.71
960.800.747.685.65.43135.29
7686.676.26--84.612394.71
Stock240.100.190.672.00.1815.40
960.130.202.457.20.4046.78
7681.101.2059.3256.03.55514.08
+ +# E.4 Effect of Integrated Training Scheme + +We investigate the impact of our integrated training scheme, which combines a short pre-training of the TST-based imputation module with joint training of both imputation and diffusion components. To assess its advantages, we compare three settings: (i) joint training with a brief TST pre-training (our approach), (ii) fully joint training from scratch, and (iii) a strict two-stage training where imputation is performed independently before generative training. Results in Tab. 15 demonstrate that our integrated approach consistently achieves the best discriminative performance across the Energy and Stock datasets, while fully joint training without pre-training suffers from unstable reconstructions, and the two-stage setup underperforms due to the lack of interaction between imputation and generative + +Table 15: Average discriminative scores across $30\%$ , $50\%$ , and $70\%$ drop-rates on the Energy and Stock datasets for sequence length of 768. + +
ModelEnergyStock
Joint training without pre-training0.2780.076
Training independently0.4040.122
Joint training with pre-training (Ours)0.2220.024
+ +learning. These findings highlight that even a short pre-training phase can stabilize learning and enable the model to effectively leverage imputed values during generation. + +# E.5 Quantitative Evaluation - Full Results + +We present the complete results, including standard deviations, for the experiment conducted in the main text in Sec. 5.2. In Tab. 18, Tab. 19, and Tab. 20, we provide detailed performance results across all missing rates of $30\%$ , $50\%$ , and $70\%$ for sequence lengths of 24, 96, and 768, respectively. + +# E.6 Qualitative Evaluation - Cont. + +We provide the remaining missing rate analyses for the experiment described in Sec. 5.3. Fig. 8 presents the analysis for a $50\%$ missing rate, while Fig. 9 shows the analysis for a $30\%$ missing rate. + +Additionally, we quantitatively assess the overlap between the original and generated data cloud points in the two-dimensional plane. We compute the Wasserstein distances between the original data and the generated samples. The results, presented in Table 16, indicate that our method consistently achieves the lowest Wasserstein distances across all missing rates and sequence lengths, outperforming KoVAE in every case. Notably, for long time series, our model exhibits significantly better performance, demonstrating its robustness in handling larger and more complex sequences with high missing rates. This underscores our approach's ability to closely match the true data distribution, even under challenging conditions, and its superior effectiveness in managing incomplete time series. + +![](images/0056eaa27446e365cdcfa95d027d475580b42c17165639fa9df8a80d0a642a95.jpg) + +![](images/cdb8780fdeefa5f0fb2124cd260881aa7496d6dc72d07b216da82a09b55950df.jpg) + +![](images/639220e1671303b7898a58ed9f3fcb245253fef24360dd7c97b524e7626dac91.jpg) + +![](images/271557cb17c385d3bc0cd5891ddf47bc111d2af7ed55d86d261236236d131f97.jpg) +Figure 8: 2D t-SNE embeddings (top) and probability density functions (bottom) for the $50\%$ missing rate on ETTh1 (short), ETTm2 (medium), and Sine (long) datasets. + +![](images/f57ba68f39c0f4752737473d3143579e85986f77f9a9b733ff98cacd6bbbe1b3.jpg) + +![](images/af8da21acdbecb5faaf93797b9151118dc865d03829b763b2af1312d85a50ce2.jpg) + +![](images/a7951717752c7edeea488f1b4fd53b00f3868d1d8f8f105de2d4fc12be8e4092.jpg) + +![](images/7b55bbc04b1843c3324d1116a5074be67f0c4e673f22ae9ea4542b2986c198f6.jpg) + +![](images/68b2f0408c39826bb655d91e88326c86c4aebe16c4c1ceb643e1ef0795a73ee1.jpg) + +![](images/cce3bb2336301ab6270d8db6f4eb06d8d31751e74df5b97d1003879818e74884.jpg) +Figure 9: 2D t-SNE embeddings (top) and probability density functions (bottom) for the $30\%$ missing rate on ETTh1 (short), ETTm2 (medium), and Sine (long) datasets. + +![](images/c93bb03327a248615dd938a9766a121e61bf967481820b4de0c2aa7b12bbb76b.jpg) + +![](images/2a4a581b44b1506dd4e62a8e17303ef3691d652796e447a38686d5b781c7e907.jpg) + +Table 16: Wasserstein distances between original data clusters, generated samples, and clusters from KoVAE for various missing rates (30%, 50%, 70%) and sequence lengths. Lower values indicate better similarity. + +
MetricModel30% Drop50% Drop70% Drop
EnergyWeatherStocksETTh1ETTm2SineETTh1ETTm2Sine
Wass.↓KoVAE3.977.645.534.785.926.253.384.3712.07
Ours1.082.851.681.621.841.501.332.392.42
+ +# E.7 Complexity Analysis Cont. + +Continuing from Sec. 5.1, where we summarized the upcoming experiment, we now present the full results. + +In Table 17, we report the net training time (in hours) for our method (Ours) and KoVAE until convergence, measured under identical hardware (RTX3090) and batch size settings. Convergence was defined by the best discriminative score achieved during training; specifically, we sampled the generated data every 10 training epochs and computed the discriminative score against the real data. Note that the times shown exclude any overhead for data generation or evaluation; we only measure the pure training runtime until the point of highest discriminative performance. + +These training times are presented for three different sequence lengths (24, 96, 768) and three missing rates (30%, 50%, 70%). To assess the relative speedup of our approach over KoVAE, we averaged the training times of each model (when valid entries were available), computed the percentage speedup as + +$$ +\left(\frac{\text{KoVAE time} - \text{Ours time}}{\text{KoVAE time}}\times 100\%\right), +$$ + +and then averaged these speedups across the three missing rates. Our results show that: + +- At a sequence length of 24, Ours converges approximately $\sim 80\%$ faster on average. +- At a sequence length of 96, Ours converges approximately $\sim 87\%$ faster on average. +- At a sequence length of 768, Ours converges approximately $\sim 85\%$ faster on average. + +These figures underscore the substantial reduction in training time provided by our method, ranging from about $80\%$ to $90\%$ faster than KoVAE in most settings, while still achieving superior performance based on the discriminative score. + +Table 17: Training Time (in hours) for irregular time series across sequence lengths (24, 96, 768) and missing rates (30%, 50%, 70%). + +
Seq. Len.Drop %ModelETTh1ETTh2ETTm1ETTm2WeatherEnergySineStock
2430%KoVAE Ours5.0015.0811.657.499.2312.205.782.71
0.730.611.180.904.772.800.500.37
50%KoVAE Ours4.984.737.806.682.2028.596.861.03
1.856.751.181.810.382.510.600.17
70%KoVAE Ours9.4512.5510.985.486.088.914.222.36
1.270.920.521.585.822.020.350.08
9630%KoVAE Ours25.808.1013.7925.9233.874.946.907.71
3.380.801.350.952.432.361.200.23
50%KoVAE Ours1.4217.2611.0411.9014.7918.0122.473.90
0.590.571.320.951.061.330.470.46
70%KoVAE Ours31.917.1219.9015.7717.5218.112.917.76
0.614.341.053.652.121.492.721.60
76830%KoVAE Ours62.1959.7092.4869.4840.8813.3018.0316.59
4.721.967.576.4516.463.502.933.11
50%KoVAE Ours15.5237.4126.3045.6646.9919.865.4914.84
4.742.967.236.487.182.352.891.21
+ +Table 18: Evaluation metrics for irregular time series with 24 sequence length (30%, 50%, 70% drop). Arrows $(\uparrow/\downarrow)$ indicate whether higher or lower values are better. + +
MetricModelEtth1Etth2Ettm1Ettm2WeatherElectricityEnergySineStockMujoco
30% Drop
Disc. ↓TimeGAN0.4990.4990.4990.4990.4930.4980.4480.4940.4630.471
GT-GAN0.4730.3710.4200.3690.4720.4090.3330.3630.2510.249
KoVAE0.2080.0750.0450.0770.2290.4970.2800.0350.1620.123
Ours0.0200.0090.0140.0060.0290.3990.0480.0130.0070.009
Pred. ↓TimeGAN0.1560.3050.1460.2620.3880.1830.3750.1450.0870.118
GT-GAN0.1740.0920.1190.0970.1470.1480.0660.0990.0210.048
KoVAE0.0580.0500.0440.0510.0290.0480.0490.0740.0190.043
Ours0.0520.0430.0440.0450.0220.0480.0470.0700.0120.040
Fid. ↓TimeGAN2.9342.5652.4372.9241.61218.044.4402.9192.4753.628
GT-GAN1.68915.2627.436.9021.1619.9071.3051.8102.4290.656
KoVAE1.7690.2110.1810.6090.5397.6060.6450.0480.7410.428
Ours0.0710.0230.0230.0100.0183.4510.0330.0320.0090.028
Corr. ↓TimeGAN6.3170.8622.2900.3570.74411.133.6632.1310.2730.844
GT-GAN7.1670.9182.5190.3580.78214.933.8553.1410.2640.803
KoVAE0.1480.0880.1620.4831.8524.3512.9100.0490.0320.561
Ours0.0700.0480.0590.0450.4242.0410.8150.0160.0070.331
50% Drop
Disc. ↓TimeGAN0.4990.4990.4990.4990.4990.4980.4790.4960.4870.483
GT-GAN0.4620.3710.4070.3760.4960.3910.3170.3720.2650.270
KoVAE0.1880.0860.0570.0770.4980.4990.2980.0300.0920.117
Ours0.0320.0050.0130.0130.0350.3600.0650.0140.0070.007
Pred. ↓TimeGAN0.2100.3430.1570.2920.4040.2300.5010.1230.0580.402
GT-GAN0.1760.0910.1180.0960.1270.1470.0640.1010.0180.056
KoVAE0.0570.0530.0450.0500.1140.0430.0500.0720.0190.042
Ours0.0530.0450.0430.0410.0220.0490.0470.0680.0120.040
Fid. ↓TimeGAN4.1313.1322.6422.6932.83920.1846.4082.1242.3524.141
GT-GAN1.5045.8394.2019.4685.9199.7411.9351.7852.2580.664
KoVAE1.3090.3190.2070.1159.8303.9720.4210.0300.2250.371
Ours0.1340.0260.0450.0180.0403.6330.0610.0070.0570.026
Corr. ↓TimeGAN2.2930.9321.8640.3520.83513.793.7612.1922.0210.825
GT-GAN7.0880.9282.4430.3610.80714.913.9713.2040.2550.804
KoVAE0.1730.2830.1320.1575.0274.1521.8220.0290.0720.555
Ours0.1120.0470.0580.0260.3632.0440.8720.0150.0110.342
70% Drop
Disc. ↓TimeGAN0.5000.4990.5000.5000.4990.5000.4960.5000.4880.494
GT-GAN0.4780.3660.4090.3530.4750.4800.3250.2780.2300.275
KoVAE0.1960.0810.0480.0460.2690.4980.3920.0650.1010.119
Ours0.0580.0130.0100.0150.1060.3920.1280.0080.0100.009
Pred. ↓TimeGAN0.4360.3590.4010.3870.3900.3720.4960.7340.0720.442
GT-GAN0.2070.0940.1370.0890.1620.1490.0760.0880.0200.051
KoVAE0.0560.0600.0460.0480.0280.0510.0520.0760.0120.044
Ours0.0540.0490.0450.0470.0230.0490.0470.0690.0120.041
Fid. ↓TimeGAN2.3563.9005.1794.0362.68331.958.6743.2963.1784.432
GT-GAN3.4424.80511.242.7841.19410.331.3541.4981.8570.671
KoVAE1.4770.2150.1530.1160.7277.6100.8200.0330.1410.292
Ours0.1670.0560.0730.0430.4513.6530.3010.0070.0410.044
Corr. ↓TimeGAN6.8210.9592.4700.3870.72814.803.8483.0513.3540.867
GT-GAN7.1900.9212.4380.3510.78514.923.8423.5020.2540.809
KoVAE0.2280.1600.0960.1441.8174.2133.1570.0290.0950.566
Ours0.0710.1040.0790.0550.4032.0071.3300.0150.0370.348
+ +Table 19: Evaluation metrics for irregular time series with 96 sequence length (30%, 50%, 70% drop). Arrows $(\uparrow/\downarrow)$ indicate whether higher or lower values are better. + +
MetricModelEtth1Etth2Ettm1Ettm2WeatherEnergySineStock
30% Drop
Disc. ↓KoVAE0.2550.0960.2640.0750.2900.4160.2440.114
Ours0.0370.0360.0330.0280.0840.1300.0040.011
Pred. ↓KoVAE0.0620.0620.0530.0470.0400.0770.1640.016
Ours0.0520.0450.0440.0420.0230.0480.1550.010
Fid ↓KoVAE5.2230.9154.0730.6452.9424.0752.7250.944
Ours0.1560.1120.1470.0600.1690.1930.0180.110
Corr. ↓KoVAE0.2010.3010.1770.2032.7304.6770.0580.086
Ours0.1020.0750.0870.0420.3060.8270.0170.016
50% Drop
Disc. ↓KoVAE0.3040.0770.2900.1140.3580.3210.1880.111
Ours0.0700.0670.0470.0170.1900.1530.0030.018
Pred. ↓KoVAE0.0630.0550.0530.0540.0510.0630.1610.016
Ours0.0540.0530.0440.0460.0250.0480.1550.011
Fid ↓KoVAE5.3700.7814.6631.3254.1174.9552.3341.003
Ours0.5160.4870.1790.1300.5670.1820.0180.105
Corr. ↓KoVAE0.2460.2880.2470.6382.4905.1420.0440.085
Ours0.1540.2190.0900.1131.1761.1860.0190.011
70% Drop
Disc. ↓KoVAE0.2940.1370.2740.0790.3740.3320.2860.073
Ours0.1020.0570.0390.0440.1820.2720.0020.021
Pred. ↓KoVAE0.0620.0550.0560.0490.0470.0640.1710.029
Ours0.0530.0500.0460.0440.0270.0510.1550.012
Fid ↓KoVAE6.9321.6383.4731.0193.9835.0518.0830.657
Ours0.3990.9260.1870.2350.4470.5530.0160.112
Corr. ↓KoVAE0.2260.4960.1010.4652.6074.6110.5650.095
Ours0.0870.1590.1000.1261.1871.4700.0130.015
+ +Table 20: Evaluation metrics for irregular time series with 768 sequence length (30%, 50%, 70% drop). Arrows $(\uparrow/\downarrow)$ indicate whether higher or lower values are better. + +
MetricModelEtth1Etth2Ettm1Ettm2WeatherEnergySineStock
30% Drop
Disc. ↓KoVAE0.2390.2370.2820.1600.4110.3710.2840.289
Ours0.0450.0320.0670.0470.0930.1450.0050.025
Pred. ↓KoVAE0.0770.0740.0590.0810.0610.0890.2230.020
Ours0.0520.0560.0470.0510.0270.0250.2040.011
Fid ↓KoVAE11.169.44811.478.86912.2125.5038.717.431
Ours0.3180.2580.3730.2200.1100.7550.1840.116
Corr. ↓KoVAE0.3780.5280.4800.8593.1368.1380.4110.041
Ours0.0880.1190.1260.0860.2690.8490.0060.004
50% Drop
Disc. ↓KoVAE0.2700.1910.1970.2250.4280.3720.4260.302
Ours0.0610.0300.0610.0480.0970.2440.0090.029
Pred. ↓KoVAE0.0740.0640.0560.0840.0640.0860.2220.023
Ours0.0530.0570.0470.0480.0270.0510.2040.015
Fid ↓KoVAE14.567.41211.519.37318.5919.5838.498.274
Ours0.5890.2480.3050.2110.2250.9160.2110.199
Corr. ↓KoVAE0.2900.5740.4280.8424.83612.630.3360.085
Ours0.1030.1280.1070.1190.4551.4730.0060.042
70% Drop
Disc. ↓KoVAE0.2060.1760.2310.2030.4450.4090.3400.263
Ours0.1600.0720.0460.0610.1160.2510.0050.013
Pred. ↓KoVAE0.0650.0710.0650.0620.0860.0880.2340.051
Ours0.0540.0560.0470.0520.0280.0500.2040.013
Fid ↓KoVAE16.058.05217.546.59521.6928.4638.606.114
Ours1.7390.8120.4960.3620.6190.7180.3170.167
Corr. ↓KoVAE0.3310.7180.3050.5151.7865.490.3910.013
Ours0.1180.0710.1330.1790.6500.7980.0060.034
\ No newline at end of file diff --git a/NeurIPS/2025/A Diffusion Model for Regular Time Series Generation from Irregular Data with Completion and Masking/images.zip b/NeurIPS/2025/A Diffusion Model for Regular Time Series Generation from Irregular Data with Completion and Masking/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..b9daf441f9cfba3e9aca91ac3fb62db20c6aaa45 --- /dev/null +++ b/NeurIPS/2025/A Diffusion Model for Regular Time Series Generation from Irregular Data with Completion and Masking/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:df00288e8b92da54ebc932be834656123b95ef5aab70432e92fb1cb550eb0576 +size 1803964 diff --git a/NeurIPS/2025/A Diffusion Model for Regular Time Series Generation from Irregular Data with Completion and Masking/layout.json b/NeurIPS/2025/A Diffusion Model for Regular Time Series Generation from Irregular Data with Completion and Masking/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..3c9a1134c81e50990cbcab30ce65ecc48a086af8 --- /dev/null +++ b/NeurIPS/2025/A Diffusion Model for Regular Time Series Generation from Irregular Data with Completion and Masking/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bbf08bd4581897d18b56252d9e077523e6ab7025ebfff94f44318e92bb766bf2 +size 1000610 diff --git a/NeurIPS/2025/A Driving-Style-Adaptive Framework for Vehicle Trajectory Prediction/c9d7a9f5-abca-480e-a8d9-84c6894c18aa_content_list.json b/NeurIPS/2025/A Driving-Style-Adaptive Framework for Vehicle Trajectory Prediction/c9d7a9f5-abca-480e-a8d9-84c6894c18aa_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..97eebe0052ab8941ee84f4f27a298a37722ebebd --- /dev/null +++ b/NeurIPS/2025/A Driving-Style-Adaptive Framework for Vehicle Trajectory Prediction/c9d7a9f5-abca-480e-a8d9-84c6894c18aa_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0548cbb5ed0a3eea1c70dbf5828fa66489b2e1a7d302723c9309bec78f57d371 +size 272815 diff --git a/NeurIPS/2025/A Driving-Style-Adaptive Framework for Vehicle Trajectory Prediction/c9d7a9f5-abca-480e-a8d9-84c6894c18aa_model.json b/NeurIPS/2025/A Driving-Style-Adaptive Framework for Vehicle Trajectory Prediction/c9d7a9f5-abca-480e-a8d9-84c6894c18aa_model.json new file mode 100644 index 0000000000000000000000000000000000000000..ad24f1c09f168e1d07ce793f699755a2bb57cce9 --- /dev/null +++ b/NeurIPS/2025/A Driving-Style-Adaptive Framework for Vehicle Trajectory Prediction/c9d7a9f5-abca-480e-a8d9-84c6894c18aa_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3fbb76d9a0ba6302bbefb78280b3c7f9ae7a8d56c71543c10f5e30e9bfd35795 +size 344088 diff --git a/NeurIPS/2025/A Driving-Style-Adaptive Framework for Vehicle Trajectory Prediction/c9d7a9f5-abca-480e-a8d9-84c6894c18aa_origin.pdf b/NeurIPS/2025/A Driving-Style-Adaptive Framework for Vehicle Trajectory Prediction/c9d7a9f5-abca-480e-a8d9-84c6894c18aa_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..0320cb86bb7c58ca19962efc7c9acf1de588c8f8 --- /dev/null +++ b/NeurIPS/2025/A Driving-Style-Adaptive Framework for Vehicle Trajectory Prediction/c9d7a9f5-abca-480e-a8d9-84c6894c18aa_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b0b91865f229ab0af38b76d4000339bff1a8079bd8509f809cc5277698ccd15b +size 16074487 diff --git a/NeurIPS/2025/A Driving-Style-Adaptive Framework for Vehicle Trajectory Prediction/full.md b/NeurIPS/2025/A Driving-Style-Adaptive Framework for Vehicle Trajectory Prediction/full.md new file mode 100644 index 0000000000000000000000000000000000000000..92febe3cc9650588c93bc9fa323f8aa90f5336c9 --- /dev/null +++ b/NeurIPS/2025/A Driving-Style-Adaptive Framework for Vehicle Trajectory Prediction/full.md @@ -0,0 +1,1482 @@ +# A Driving-Style-Adaptive Framework for Vehicle Trajectory Prediction + +Di Wen $^{1,2,3}$ Yu Wang $^{2}$ Zhigang Wu $^{1,3}$ + +Zhaocheng He $^{1,2,3*}$ Zhe $\tilde{\mathrm{W}}\mathrm{u}^{2,*}$ Zheng Qingfang $^{2}$ + +$^{1}$ Sun Yat-sen University $^{2}$ Pengcheng Laboratory + +3Guangdong Provincial Key Laboratory of Intelligent Transportation System + +{wend25,wuzhig6}@mail2.sysu.edu.cn hezhch@mail.sysu.edu.cn + +{wangy12,wuzh02,zhengqf01}@pcl.ac.cn + +# Abstract + +Vehicle trajectory prediction serves as a critical enabler for autonomous navigation and intelligent transportation systems. While existing approaches predominantly focus on pattern extraction and vehicle-environment interaction modeling, they exhibit a fundamental limitation in addressing trajectory heterogeneity originating from human driving styles. This oversight constrains prediction reliability in complex real-world scenarios. To bridge this gap, we propose the Driving-Style-Adaptive (DSA) framework, which establishes the first systematic integration of heterogeneous driving behaviors into trajectory prediction models. Specifically, our framework employs a set of basis functions tailored to each driving style to approximate the trajectory patterns. By dynamically combining and adaptively adjusting the degree of these basis functions, DSA not only enhances prediction accuracy but also provides explanations insights into the prediction process. Extensive experiments on public real-world datasets demonstrate that the DSA framework outperforms state-of-the-art methods. + +# 1 Introduction + +Vehicle Trajectory Prediction (VTP) serves as a fundamental capability for numerous intelligent transportation applications, including autonomous driving systems [1, 2, 3], motion planning algorithms [4, 5] and adaptive traffic control frameworks [6, 7]. Recent advances in VTP achieve notable progress through two primary paradigms: (1) capturing temporal patterns from historical trajectories and modeling vehicle interactions [8, 9, 10, 11], and (2) leveraging structured scene representations that incorporate road topology and regulatory constraints [12, 13, 14, 15]. However, these methods often overlook the originator of the trajectory: human drivers [16, 17], whose diverse behavior leads to heterogeneous trajectory patterns. + +![](images/52561839961fd2241138ded451485d968ce38e8fd2de76546ea43108c3485df4.jpg) +Figure 1: Illustration of three driving styles: Conservative drivers typically move slowly or stop to avoid obstacles; Aggressive drivers often travel at high speeds and are prone to overtaking other vehicles; Normal drivers maintain a constant speed and frequently change lanes to ensure safety. + +In this paper, we propose an adaptive VTP framework based on distinct driving styles [18, 19]: Conservative, Aggressive and Normal (CAN). Each driving style manifests in characteristic trajectory + +patterns, as illustrated in Figure 1. limited variability (conservative), non-smooth trajectories (aggressive) and frequent yet smooth motion changes (normal). For each styles, we employs variable basis functions within Kolmogorov-Arnold Networks (KANs) [20] to capture these trajectory patterns. In complex real-world scenarios, driver behavior often reflects a probabilistic mixture of weighted driving styles [21]. + +Our framework comprises two core components: (1) the matching between driving styles and their corresponding basis functions, and (2) the weighted combination and adjustment of the degrees of these functions. Additionally, inspired by the Weierstrass Approximation Theorem [22], our proposed DSA framework extends KANs from a theoretical perspective. Each matching in (1) is further grounded in the mechanical properties of basis functions, thereby providing explanations for our DSA framework. + +Our main contributions in this paper are summarized as follows: + +- To address the vehicle trajectory prediction task, we propose for the first time, a novel Driving-Style-Adaptive (DSA) framework tailored to the driving styles of human drivers and effectively leverages trajectory information. +- We utilize polynomial approximation operators to approximate and predict trajectories under different driving styles: Conservative, Aggressive and Normal (CAN). These operators support a mathematical explanation matching mechanism that matches each driving style with a corresponding polynomial form. +- The experimental results on real-world datasets (nuScenes, Argoverse and Waymo) demonstrate that our model significantly outperforms existing methods in vehicle trajectory prediction. + +# 2 Preliminary & Related Work + +# 2.1 Task Definition: Vehicle Trajectory Prediction (VTP) + +VTP aims to predict the future trajectory of vehicles based on history trajectory or other informations available in a given scenario. In recent years, deep learning based VTP methods are categorized into two groups [23]: (i) knowledge-based methods, which incorporate specific information such as maps [24, 25], vehicles [9, 26] and interactions [27, 28] to represent the environment or vehicle behaviour. (ii) knowledge-free methods, which rely on deep learning's ability to encode complex data features, modeling them using structures such as tensors [29, 30] or attention mechanisms [31, 32]. + +Following above works, we analyze traffic scenes involving $N$ vehicles (agents). The trajectory of each vehicle $i$ in historical interval $[0,T]$ is denoted as $X_{i} = \{s_{i}^{-t},\dots ,s_{i}^{0}\}$ . Each state $s_i^\star$ is a 5-dimensional vector representing the $(x,y)$ position, velocity, acceleration, and the nearest lane segments ID. The superscript $\star$ denotes the time step. Similarly, the future trajectory in interval $[0,T]$ is given by $Y_{i} = \{s_{i}^{0},\dots ,s_{i}^{T}\}$ . + +# 2.2 Basic Network: Kolmogorov-Arnold Networks (KANs) + +KANs [20] are inspired by the mathematical principles [33, 34, 35] of the Kolmogorov-Arnold representation theorem [36, 37, 38], stated as follows: + +Theorem 2.1 (Kolmogorov-Arnold representation Theorem) For any multivariate continuous function $f:[0,1]^n\to \mathbb{R}$ , $f$ can be represented as a finite composition of univariate continuous functions $\phi_{ij}:[0,1]\rightarrow \mathbb{R}$ and $\Phi_j:\mathbb{R}\rightarrow \mathbb{R}$ , with the binary operation of addition such that: + +$$ +F = f \left(x _ {1}, \dots , x _ {n}\right) = \sum_ {j = 1} ^ {2 n + 1} \Phi_ {j} \left[ \sum_ {i = 1} ^ {n} \phi_ {i j} \left(x _ {i}\right) \right]. \tag {1} +$$ + +The key innovation of KANs lies in implementing the residual activation functions $\phi (x)$ in Equation (1) as: + +$$ +\phi (x) = w _ {b} b (x) + w _ {s} \psi (x), \tag {2} +$$ + +where $b(x) = \mathrm{silu}(x)$ and $\psi (x) = \mathrm{spline}(x)$ . Unlike Multi-Layer Perceptrons (MLPs [39, 40]), which utilize fixed activation functions associated with nodes ("neurons"), KANs feature learnable + +$\phi(x)$ on edges ("weights"). However, due to the inherent complexity of these functions, the speed and scalability of the original KANs are not satisfactory [41]. Consequently, a variety of KAN-based applications are emerged in AI4Science tasks [42, 43, 44, 45, 46, 47]. To the best of our knowledge, we are the first to extension KANs to the VTP task. We achieve this by expanding the set of basis functions $\psi(x)$ to match different driving styles and by grounding this matching in both mathematical theory and task-specific behavior. + +# 2.3 Core Theory: Applying Polynomials as Basis Functions + +As describe in Section 2.2, a fixed basis function has inherent limitations. In the VTP task, such a function may fail to adequately approximate diverse trajectory or lane curves. This raises an important question: how can we legitimately expand the class of basis functions? In approximation theory, a fundamental question is whether polynomials can approximate any given continuous function to an arbitrary degree of precision. Weierstrass [22] provides a definitive answer by: + +Theorem 2.2 (Weierstrass Approximation Theorem) Let $f(x) \in L^{p}[0,1]$ for any $p > 0$ . Then there exists an algebraic polynomial $p_n(x) = \sum_{m=0}^{n} c_m x^m$ such that + +$$ +\lim _ {n \rightarrow \infty} \int_ {0} ^ {1} | f (x) - p _ {n} (x) | ^ {p} d x = 0. \tag {3} +$$ + +This interval can be extended to $[a,b]$ . This demonstrates that polynomials $p_n\in \mathcal{P}_n$ can serve as basis functions $\psi (x)$ in Equation (2) to approximate the $f$ in Theorem 2.1. In our task, the vehicle trajectories are treated as the function $f$ , with respect to the time step $t$ . For different driving styles, we employ corresponding $p_n$ as basis functions to approximate these trajectories in accordance with Theorem 2.2. Furthermore, these trajectories also belong to the $L^p$ space2, and are thus well-defined. + +# 3 Methodology + +# 3.1 Motivation and Overview + +Our Driving-Style-Adaptive (DSA) framework (illustrate in 2) models the behavior of the trajectory originator: the human driver. The driving style of various vehicle drivers are categorized as: Conservative, Aggressive and Normal (CAN) [48, 49, 50], each reflecting distinct trajectory characteristics. + +![](images/bafc66263385f6c2cc863d3c39b56382e4aec51863aa36ae1d91f1c035f35fd6.jpg) +Figure 2: An overview of our DSA framework, which performs trajectory prediction based on driving style categories: conservative, aggressive and normal (CAN) to prediction. For clarity, we illustrate this process using a single vehicle example. The solid line represents trajectories length while the arrow indicates the direction (history or future). The symbol $B_{n}$ , $T_{n}^{c}$ and $L_{n}$ denote different basis functions $p_{n}$ corresponding to each driving style. Our proposed DSA framework dynamic adapts the experts (driving style) weighs $w_{*}$ and the degree $n_{*}$ of selected $p_{n}$ . + +We match each driving style characteristic to a corresponding approximation polynomial $p_n$ based on the mathematics properties of $p_n$ as described in Section 3.2. To implement this mechanism, we introduce $p_n$ combination and degree adjustment strategies in Section 3.3. + +# 3.2 Theoretical Foundations for Matching Polynomials to Driving Styles + +In this section, we elucidate the matching between the polynomials $p_n$ ( $p_n \in \mathcal{P}_n$ ) and driving style, focusing on the mathematical properties of $p_n$ and analyzing the characteristics of each driver's type. Specifically, we address conservative drivers in Section 3.2.1, aggressive drivers in Section 3.2.2 and normal drivers in Section 3.2.3. + +# 3.2.1 Conservative Drivers + +Conservative drivers [51] prioritize driving comfort and safety, which leads to more cautious decisions. Their average speed is typically the slowest and rarely change their behavior. Consequently, their trajectories are characterized by smoothness and stability, with minimal abrupt changes in speed. + +In this situation, we require a $p_n$ to capture approximating drivers with minimal behavioral changes, that is, ensures the approximation error decreases uniformly across the entire interval. We employ the Bernstein operatore3 $B_n$ [52] to achieve this: + +Definition 3.1 (Bernstein polynomial, $B_{n}$ ) Consider a function $f(x) \in C[0,1], x \in [0,1]$ , $B_{n}$ is specified by the equation: + +$$ +\left(B _ {n}\right) f (x) = \sum_ {k = 0} ^ {n} f \left(\frac {k}{n}\right) \binom {n} {k} x ^ {k} (1 - x) ^ {n - k}. +$$ + +It is clear that $B_n \subseteq \mathcal{P}_n$ thus applies to Theorem 2.2. The primary advantage of the $B_n$ is articulated in the following proposition: + +Proposition 3.2 For all functions $f \in C[0,1]$ , the sequence $\{B_n f; n = 1,2,3,\dots\}$ converges uniformly to $f$ as $B_n(f) \Rightarrow f(x)$ . + +This proposition demonstrates that the $B_{n}$ exhibits uniform convergence across the entire interval, making it particularly suitable for approximating trajectories with slow travel speeds and few behavioral changes, such as those of conservative drivers. This ensures that the approximation error decreases uniformly throughout the interval. + +# 3.2.2 Aggressive Drivers + +Aggressive drivers [53] prioritize their own benefits at the expense of safety and comfort, which leads to higher speeds, abrupt changes in acceleration and braking, with a frequent tendency to change lanes. As a result, their trajectories display more abrupt motions and are less smooth. + +Trigonometric polynomials $T_{n}$ are dense in $C(I)$ on the unit circle according to the Stone-Weierstrass Theorem [54]. This implies that trigonometric polynomials $T_{n}$ are particularly effective at approximating functions with discontinuities or sharp features, which define as: $T_{n}(x) = a_{0} + \sum_{n=1}^{N} [a_{n}\cos(nx) + b_{n}\sin(nx)]$ . We employ the Chebyshev polynomials [55] $T_{n}^{c}$ , defined as follows: + +Definition 3.3 (Chebyshev Polynomials, $T_{n}^{c}$ ) For $x \in [-1,1]$ , the $n$ -th $T_{n}^{c}$ of the first kind is given by $T_{n}^{c}(x) = \cos[n \cdot \arccos(x)]$ . + +The effectiveness of $T_{n}^{c}$ is further highlighted by the Chebyshev Minimax Theorem [56]: + +Theorem 3.4 (Chebyshev Minimax Theorem) For $f \in C[-1,1]$ , $T_{n}^{c}$ minimizes the maximum error in the uniform norm compared to any other $p_n$ approximation of the same degree. Formally, this relationship is expressed as: + +$$ +\left\| f - T _ {n} ^ {c} \right\| _ {L ^ {\infty}} \leqslant \left\| f - p _ {n} \right\| _ {L ^ {\infty}}. +$$ + +Theorem 3.4 explicitly states that $T_{n}^{c}$ can minimize the maximum error, effectively reducing the impact of sudden behavioral and speed changes typical of aggressive drivers. Furthermore, the overall prediction error is decreased. + +# 3.2.3 Normal Drivers + +Normal drivers [57] strike a balance between conservative and aggressive driving styles, representing a relatively common group in driving behavior. Their speed and acceleration typically fall between those of conservative and aggressive drivers, exhibiting moderate speed changes and occasional rapid reactions. Consequently, their trajectories are neither as smooth as those of conservative drivers nor as abrupt as those of aggressive drivers, but their trajectories may exhibit regular fluctuations. + +This driving characteristic is closely related to the application of orthogonal polynomials $p_n^o$ [58]. The $p_n^o$ has significant flexibility and enables accurately capture the trajectories characterized by gradual changes and moderate fluctuations for normal drivers. The $p_n^o$ with weight function $\rho$ and $\partial (p_n^o) = n$ are defined as: $\langle p_i^o, p_j^o \rangle = \int_a^b \rho(x) p_i^o(x) p_j^o(x) dx = \delta_{ij}$ , where $\delta_{ij}$ equal to 0 iff $i \neq j$ . The approximation of $p_n^o$ can be effectively described by [59]: + +Theorem 3.5 (Least Squares Characterization Theorem) For any function $f \in C[a, b]$ , there exists an orthogonal polynomial $p_n^o$ with $\partial(p_n^o) \leqslant n$ that minimizes the error in the $L_\rho^2$ norm between $f(x)$ and $p_n^o(x)$ : + +$$ +\left\| f - p _ {n} ^ {o} \right\| _ {L _ {\rho} ^ {2}} = \min _ {p \in \mathbb {P} _ {n}} \| f - p \| _ {L _ {\rho} ^ {2}}, +$$ + +where $\mathcal{P}_n$ denotes the space of all polynomials of degree at most $n$ . + +The term "least" here does not denote the non-uniqueness, but rather indicates the possible to select optimal coefficients under best $L_{\rho}^{2}$ approximating. For instance, Legendre polynomials $L_{n}$ is a typical orthogonal polynomial: + +Definition 3.6 (Legendre Polynomial, $L_{n}$ ) For $x \in [-1,1]$ with a constant weight function $\rho(x) = 1$ , $L_{n}$ is defined by + +$$ +L _ {n} = \frac {1}{2 ^ {n} n !} \frac {d ^ {n}}{d x ^ {n}} \left[ (x ^ {2} - 1) ^ {n} \right]. +$$ + +This orthogonality under the $L^2$ norm particularly without weight or a constant weight, makes it an exceptionally efficient tool for approximation [60], which represents a specific instance covered by Theorem 3.5. Moreover, $L_{n}$ is also defined by a simple recurrence relation: + +$$ +(n + 1) L _ {n + 1} (x) = (2 n + 1) x L _ {n} (x) - n L _ {n - 1} (x), +$$ + +This recurrence relation facilitates quick calculations and the optimal square approximation property excels under the $L^2$ norm. These characteristics make it well-suited for handling smooth and continuous trajectory fluctuations, align well with the The normal drivers. Their characterized by gradual changes and smoother transitions. + +# 3.3 Algorithm Realization + +# 3.3.1 Polynomial Combination + +In Section 3.2, we utilize different polynomial forms to match different driving styles, thereby fully leveraging trajectory information for prediction. However, assuming a single fixed driving style may be inadequate in complex real-world scenarios. Kernel density estimation and latent variable analysis, reveal that driver behavior varies continuously with context and can be characterized as a probabilistic mixture [21, 61, 62] of weighted driving styles. Here we employ a MoE-TopK [63] approach to model multiple driving styles for trajectory prediction. + +The process of combining the polynomials corresponding to multiple driving styles is presented in the algorithm on the right. Here $X_{i}$ represents the $i - th$ history trajectory in $N$ vehicles as described in Section 2.1, which has 5 dimensions as $(x,y)$ position, velocity, acceleration, and the nearest lane segments ID. Experts represent the polynomials in Section 3.2. The output $z^{\mathrm{Com}}$ is the feature of combine. In line 2, "SN" and "Sp" denote the Standard Normal and Softplus functions [64, 65], respectively, $W_{g}$ and $W_{n}$ are trainable weight matrices. In line 3, we define $H = (H_{1},H_{2},H_{3})$ . + +This combination structure of $p_n$ allows each $E_i$ to better extract the trajectory feature in different driving styles, and enables the use of various basis functions for predict vehicle trajectory. To encourage all experts to contribute the combination process, Shazeer N et al. [63] introduce a load balancing loss function $L_{\mathrm{MoE - K}}$ to encourage experts have equal importance as: $L_{\mathrm{MoE - K}} = w_{\mathrm{load}}$ . CV (loads) $^2$ , where "CV" denotes the coefficient of variation. + +# Algorithm: Polynomial Combination + +Require: Input vehicle trajectory $X_{i}$ + +1: Experts networ $\{E_j\}_{j = 1}^3$ ,Gating network $G$ + +Ensure: Feature $z_{i}^{\mathrm{Com}}$ via Polynomial Combination + +2: $H_{j}\gets (X\cdot W_{g}^{\prime})_{i} + \mathrm{SN}()\cdot \mathrm{Sp}[(X\cdot W_{n})_{j}]$ for all $i$ +3: $G_{j}(x) = \operatorname{Softmax}(H)$ +4: for $i = 1$ to $N$ do +5: $z_{i}^{\mathrm{Com}}\gets G_{j}\left(X_{i}\right)\cdot E_{j}\left(X_{i}\right).$ +6: end for +7: $z^{\mathrm{Com}}\gets \sum_{i = 1}^{N}z_{i}^{\mathrm{Com}}$ + +# 3.3.2 Degree Adjustment + +Different driving style of trajectories + +can be approximated by corresponding $p_n$ . However, the fixed degree of $p_n$ can restrict their ability for prediction entire trajectory of vehicles, which refers to: + +Theorem 3.7 (Kolmogorov Theorem) For $f \in C[a, b]$ , there exists a polynomial $p_n$ such that approximation error is bounded by: + +$$ +\| f - P _ {n} \| _ {L ^ {\infty}} \lesssim \left(\frac {\log n}{n}\right) V (f, [ a, b ]), +$$ + +where $V(f, [a, b])$ denotes the total variation of $f$ over the interval. + +From Theorem 3.7, the accuracy of the polynomial approximation is directly related to the degree $n$ of the $p_n$ , which applies broadly to $L^p$ -space. On the other hand, the $n$ is bounded when error bounded of $p_n$ is known, this assertion is proved in Appendix C. + +Adapting $n$ presents a complex non-convex and combinatorial optimization problem. To tackle this issue, we utilize SMAC3 [66] tool, which is particularly suitable for optimizing low-dimensional and continuous functions, suitable for characteristic of vehicle trajectory (Section 2.1). Specifically, the degree $n$ is treated as a hyperparameter optimization problem, aimed at minimizing the loss $(L)$ on validation data $D_{\mathrm{val}}$ and training data $D_{\mathrm{train}}$ . This process can be formulated as follows: + +$$ +n _ {\mathrm {S M A C}} \in \arg \min c (n) = \arg \min L \left(\mathcal {D} _ {\text {t r a i n}}, \mathcal {D} _ {\text {v a l}}; n\right), +$$ + +$$ +n \in \mathbb {Z} ^ {+} \qquad \qquad n \in \mathbb {Z} ^ {+} +$$ + +The hyperparameter optimization process targets the final degree $n_{\mathrm{SMAC}}$ , corresponding to achieve the least error for the corresponding basis function $p_n$ . + +# 4 Experiments + +# 4.1 Basic Setting + +We evaluate our DSA framework on three real-world vehicle trajectory prediction datasets: nuScenes [67], Argoverse [68] and Waymo [69]. These timestep settings follow the format(history time $\rightarrow$ prediction time): $2\rightarrow 6,2\rightarrow 3$ and $1\rightarrow 8$ , respectively. We utilize $L_{\mathrm{oss}} = \lambda_1L_{\mathrm{Dis}}+$ $\lambda_{2}L_{\mathrm{MoE - K}}$ , with $L_{\mathrm{MoE - K}} = w_{\mathrm{load}}\cdot \mathrm{CV}$ (loads) $^2$ for model training with balanced weighting parameters $\lambda_{*}$ . We employ common standard metrics as the Average / Final Displacement Error (ADE / FDE) for evaluate generate $k$ trajectories. More detail of datasets and metrics, please refer to Appendix B. + +# 4.2 Main Results + +# 4.2.1 Quantitative Result + +We evaluate our proposed DSA framework against existing methods utilize standard metrics. The best and second-best results are highlighted in Table 1 for the nuScenes and Argoverse datasets (with a 2-second observation window). Table 2 for Waymo (with a 1-second observation window) respectively. The results demonstrate that our method outperforms most existing approaches, achieving superior performance in 9 out of 13 evaluation metrics and ranking second in 3 others. Specifically, the best results over baseline datasets in Section 4.1 are $5.52\%$ -FDE $_5$ (nuScenes), $8.82\%$ -ADE $_6$ (Argoverse) and $1.93\%$ -minFDE (Waymo). + +Table 1: Performance comparison of baseline and our DSA framework on the nuScenes (left, N-Method) and Argoverse (right, A-Method) datasets. The best and second-best are highlighted. + +
N-MethodADE1FDE1ADE5FDE5ADE10FDE10A-MethodADE1FDE1ADE6FDE6
THOMAS [70]-6.711.33-1.04-GOHOME [71]1.703.680.891.29
PreTraM [72]--1.704.151.453.22LTP [13]1.623.550.831.30
Goal-Driven [73]--1.853.871.322.50MP++* [74]1.623.610.791.21
MUSE-VAE [75]--1.382.901.092.10HiVT [76]1.603.530.771.17
Real-Time [77]3.568.631.603.341.232.32ADAPT [78]1.593.500.791.17
Aware [79]5.5811.47--1.672.66Aware [79]1.613.540.861.31
FRM [80]-6.591.18-0.88-FRM [80]--0.821.27
Context-Aware [81]3.548.241.593.28--R-Pred [82]1.583.470.761.12
LAformer [11]--1.19-0.93-LAformer [11]--0.771.16
DAMM [83]2.846.591.393.141.022.05DAMM [83]1.573.420.761.29
CASPNet++ [84]2.746.181.16-0.92-ProphNet [85]1.282.770.680.97
CASPFormer [86]-6.701.15---QCNet [87]--0.731.07
DSA2.696.471.212.740.852.00DSA1.172.850.620.95
+ +In the nuScenes dataset, DSA outperforms previous methods in metrics of $\mathrm{ADE}_1$ , $\mathrm{FDE}_5$ , $\mathrm{ADE}_{10}$ , and $\mathrm{FDE}_{10}$ . Compared to DAMM [83], which utilizes higher-order patterns to describe interactions between agents (vehicles), our model shows a significant improvement, achieving a $16.67\%$ enhancement in $\mathrm{ADE}_{10}$ . Moreover, compared with FRM [80], which uses lane information to predict stochastic future relationships among agents, there is only a marginal gap of 0.03 in $\mathrm{ADE}_5$ , but we achieve a $3.41\%$ improvement in $\mathrm{ADE}_{10}$ . However, DSA is slightly less effective than CASPNet++ [84], which employs interaction modeling and scene understanding for joint prediction of all road users while we only predict vehicle, that leads to a minimal gap, measured in the thousandth place. + +In the Argoverse dataset, our model achieves on three of four metrics in baseline. Although our $\mathrm{FDE}_1$ metric gap in 0.08 than the baseline best results ProphNet [85], when predicting 6 samples, DSA shows improvements of $8.82\%$ in ADE and $2.06\%$ in ADE. In addition, while ProphNet utilizes an agent-centric model with anchor informed strategies, our DSA employs global positioning directly. + +In the Waymo dataset (Table 2), our DSA achieve the lowest minFDE and MR. We reduce the MR by $5.14\%$ and minFDE by $1.88\%$ compared to MotionLM [94], their minADE is slightly higher than ours by 0.015, whereas our framework is based on a simpler baseline model The ControlMTR [97] generate scene-compliant intention points and convert into a physics-based model, while DSA is driving and mathematical based, we reduce the minFDE with $4.10\%$ (value 0.0488). + +Table 2: Performance comparison of baseline and our DSA framework on the Waymo datasets. The best and second-best are highlighted. + +
MethodminADEminFDEMR*
MultiPath++ [74]0.97802.30500.4400
SceneTransformer [88]0.61171.21160.1564
MPA [89]0.59131.25070.1603
ReCoAt [90]0.77031.66680.2437
DIPP [91]0.69511.46780.1854
LiMTR [92]1.3640-0.2156
HDGT [93]0.59331.20550.1511
MotionLM [94]0.57021.16530.1327
MTR++ [95]0.59121.19860.1296
TC-Map [96]0.61811.23750.1402
ControlMTR [97]0.58971.19160.1262
DSA0.58521.14310.1259
+ +* MR (Missing Rate) is the proportion of cases in which the Euclidean distance between the prediction and the ground truth at FDE exceeds 2m. + +Our DSA framework adaptive design accommodates three categories of driving styles and we provide comprehensive explanations. This strategy simplifies the prediction process and enhances the accuracy and adaptability of predictions in complex real-world traffic scenarios. + +# 4.2.2 Qualitative Result + +![](images/4a2b58a4faf8a7a85c21e477f54d502984779eaafe44149b3f42f4f1d6b9c811.jpg) +Figure 3 demonstrates the effectiveness for our DSA framework in vehicle trajectories prediction. For more visual content, please refer to Appendix. The $k$ is the number of generation trajectories, ground truth trajectory is actual trajectories. To describe specific subfigures in Figure 3, we use the +Figure 3: Qualitative results of our DSA framework. The value of $k$ (left) represents the number of generation trajectories while letters (top) are index for clearly describe. Round head lines represent predict and ground truth trajectory, respectively. + +position index $(k,*)$ where $^*$ denotes the letter shown at the top of each subfigure. + +When $k = 1$ (i.e. the first row of Figure 3), he single prediction samples demonstrate that our DSA framework generally produces accurate results. It effectively handles not only simple road scenarios: such as straight lanes in $(1, a)$ and $(1, e)$ , or stop conditions in $(1, d)$ . But also complex scenarios including T-junctions in $(1, b)$ and crossroads in $(1, c)$ . + +In cases for generates 5 and 10 trajectories, our DSA framework delivers predictions that are both accurate and diverse. In simple scenarios, such as go straight in $(5,c)$ , $(10,a)$ , or stopping in $(5,e)$ , our framework maintains high accuracy while offering a broader range of plausible outcomes. It particularly excels in complex road conditions, including Y-crossroads in $(10,b)$ , high density crossroads $(5,b)$ and roundabouts $(10,e)$ . Moreover, the predicted trajectories effectively conform to curved roads, such as turning maneuvers in $(5,d)$ and $(10,d)$ . + +# 4.3 Ablation Studies for DSA Framework + +To explore the benefits of different components and design choices in our DSA framework, we conduct ablation experiments along several dimensions: Type and combination of polynomials $p_n$ in Section 4.3.1. Degree adjustment of polynomials in Section 4.3.2. Analysis of driving style in Section 4.3.3, examining the relationship between styles and specific $p_n$ , and how expert weights reflect behaviors. In addition, we present sensitivity analyses for different scenarios in Appendix D. + +# 4.3.1 Effects of Approximate Polynomial + +We design two experiments to evaluate the combination and type of $p_n$ on the Argoverse dataset. The number of polynomials. To illustrate the importance of considering all drivers' driving styles instead of the parts of them in trajectory predictions, we simulate a scenario where only one or two driving styles existing and select corresponding matching $p_n$ (Section 3.2) for prediction. + +Analyzing results from Table 3, consider whole driving style outperform almost the best than other combine. We observe that DSA framework incorporating two driving styles generally outperform those with only one. However, this trend is not universal. For instance, in $\mathrm{FDE}_1$ , the model + +based solely on the normal driver (1.32) over the combination of conservative and aggressive styles $(\mathrm{C + A}$ 1.54). Compared to the above combine $(\mathrm{C + A})$ in $\mathrm{ADE}_6$ DSA further re + +duces 0.21 with $25.3\%$ . This illustrate that the necessity of considering all driving styles in trajectory prediction. + +Table 3: The performances of DSA framework with different combinations of basis functions on the Argoverse dataset. C, A and N denote Conservative- $B_{n}$ , Aggressive- $T_{n}^{c}$ and Normal- $L_{n}$ , respectively. The best and second-best results are highlighted in table. + +
MetricCANC+AA+NC+NDSA
ADE11.611.451.321.541.391.661.17
FDE13.462.872.913.092.872.782.85
ADE61.021.311.120.830.920.980.62
FDE61.261.291.331.291.041.170.96
+ +Table 4: Evaluate different $p_n$ in DSA framework on the Ar-goverse dataset, with original $\rightarrow$ replace and corresponding style (abbreviate with the first three letters) in the second column. The best and second-best results highlight in table. + +
MethodReplaceADE1FDE1ADE6FDE6
Cn + Tn + LnBn → Cn Con1.483.761.101.53
Bn + Sn + LnTn → Sn Agg1.593.910.861.41
Bn + Tn + HnLn → Hn Nor1.622.810.711.02
DSA-1.172.850.620.96
+ +The Type of Polynomials. For instruction the effects of the $p_n$ we utilize in DSA framework, we replace types of $p_n$ to evaluate it, with results in Table 4. We select Charlier ([98], $C_n$ ), Hermite ([99], $H_n$ ) and second-order ( $S_n$ ) polynomials to instead $p_n$ we select in our DSA framework. + +Our DSA yields the best performance on three out of four evaluation metrics in blod, which most improved is $27.8\%$ , $43.6\%$ and $37.3\%$ respectively. + +Although the combination $B_{n} + T_{n} + H_{n}$ achieves a slightly lower FDE $_{1}$ from 2.81 to 2.85 by $1.4\%$ a marginal gap, DSA still ranks second on FDE $_{1}$ . + +# 4.3.2 Effects of Polynomial Degree + +From Theorem 3.7, we understand that the prediction accuracy is directly related to the degree $n$ of the polynomial $p_n$ . We now evaluate the impact of adaptively adjusting $n$ . To clearly illustrate this influence, we analyze the performance of a single driving style with varying degrees, as shown in Figure 4. We observe that the error generally decreases with an increasing degree. However, the + +![](images/4e644227b24b61617bbeb8acd1be2b0426c9e1ec71e29a8e53e0784b1f81f9f2.jpg) +Figure 4: Performance of our DSA Framework, with only one single fixed basis function on the waymo dataset, which the lowest error highlighted in yellow. + +![](images/469820ce7980071e013485f9bde8afed44a0911fffa1c13a8c409ade68cf698c.jpg) + +![](images/e6fd3435e07b6baca6c56c31f80d2a70f2c4de7c0794d052016ec9b7419ecf0f.jpg) + +highest degree does not necessarily yield the best results. For instance, within the aggressive driver style, $\partial (T_n^c) = 5$ outperforms other degrees in minFDE while $\partial (T_n^c) = 8$ is the best minADE, similar to conservative and normal driving style, the best results from different degrees. Consequently, adjusting $n$ rather than maintaining a fixed set provides enhanced granularity for the $p_n$ polynomials, thereby improving their capability to generate accurate and varied predictions across diverse vehicle trajectory styles. + +# 4.3.3 Analysis of Driving Style + +In Section 3.2, we provide the mathematical analysis for matching driving styles to their corresponding $p_n$ . Here, we present experimental results that evaluate both the matching relationships and how expert weights reflect those relationships. + +Relationship between driving style and polynomials. We compute the cosine similarity between the trajectory corresponding to the highest-weighted polynomial $p_n$ and the predefined driving style standards [48, 49, 50], which include Conservative (Con), Aggressive (Agg), and Normal (Nor). The results on the nuScenes (from N-Con) and Argoverse (from A-Con) datasets are shown in Table 5. + +Table 5: Cosine similarity reflecting consistent matching between polynomial based predictions and driving style standards on the nuScenes (N) and Argoverse (A) datasets. + +
pnN-ConAggNorA-ConAggNor
Bn0.9530.1580.4890.8840.1290.391
Tn0.2660.9870.3530.1670.9270.329
Ln0.3100.2630.9920.2060.3330.963
+ +From Table 5, our matching scheme $(B_{n}$ + +Con, $T_{n}$ -Agg, $L_{n}$ -Nor) obtains significantly higher similarity scores on both datasets. On the nuScenes dataset, the average similarity of the three correct matches (diagram values) is 0.977, while that of all other matches (non-bold entries) is 0.307. On Argoverse, these values are 0.925 and 0.259, respectively, representing a substantial gap. + +Table 6: The matching relationship between different driving styles and their corresponding $p_n$ . The percentage indicates the rate at which each $p_n$ has the highest weight within each driving style. + +
StyleArgoversenuScenesTop-1
Con86.15%87.96%\( B_n \)
Agg92.87%94.26%\( T_n \)
Nor80.04%83.56%\( L_n \)
+ +Expert weights mapping to driving style. To examine how expert weights mapping to driving styles, we compute the correct matching rate, defined as the proportion of cases in which the highest expert weight (Top-1) matches the expected $p_n$ for each driving style. The results appear in Table 6. The average correct matching rates across the two datasets are $86.35\%$ and $88.59\%$ , respectively. The highest matching rate is $94.26\%$ for aggressive drivers on the nuScenes dataset. All styles achieve a matching rate above $80\%$ , with a fluctuation range of $14.22\%$ . + +# 5 Limitation + +We evaluate our method on prediction horizons up to 9 seconds (1 second of history and 8 seconds of future), which is the longest duration available in current open-access vehicle trajectory datasets. However, for significantly longer horizons (e.g., over one minute), direct long-term prediction may be unreliable and would likely require segment-wise modeling or hierarchical strategies. In addition, external factors such as strong conditions (e.g., traffic signals and regulatory constraints) and soft conditions (e.g., weather, which is often unlabeled in current datasets) can also affect trajectory prediction. Incorporating these contextual cues remains an important direction for future work. + +# 6 Conclusion + +We propose an adaptive framework for vehicle trajectory prediction that is tailored to the driving styles of human drivers. To enable effective matching between polynomials $p_n$ and driving styles, we analyze the behavioral characteristics of each style alongside the mathematical properties of corresponding $p_n$ . Furthermore, we investigate the effects of $p_n$ combine, the influence of different polynomial types, and the necessity of adaptive parameters such as degree. Experiments results on three real-world datasets demonstrate that our framework significantly outperforms existing methods. + +# Acknowledgements + +This research is sponsored by the National Natural Science Foundation of China (U21B2090, 62472238, 62576181), the National Key Research and Development Program of China (2023YFB4301900), the Shenzhen Science and Technology Program (JCYJ20240813151445059), and the Science and Technology Planning Project of Guangdong Province (2023B12120600291). + +# References + +[1] Penghao Wu, Xiaosong Jia, Li Chen, Junchi Yan, Hongyang Li, and Yu Qiao. Trajectory-guided control prediction for end-to-end autonomous driving: A simple yet strong baseline. Advances in Neural Information Processing Systems, 35:6119-6132, 2022. +[2] Qingzhao Zhang, Shengtuo Hu, Jiachen Sun, Qi Alfred Chen, and Z Morley Mao. On adversarial robustness of trajectory prediction for autonomous vehicles. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15159-15168, 2022. +[3] Zhigang Wu, Jiyu Wang, Huanting Xu, and Zhaocheng He. T3c: A traffic-communication coupling control approach for autonomous intersection management system. Transportation Research Part C: Emerging Technologies, 169:104886, 2024. +[4] Haoran Song, Di Luan, Wenchao Ding, Michael Y Wang, and Qifeng Chen. Learning to predict vehicle trajectories with model-based planning. In Conference on Robot Learning, pages 1035-1045. PMLR, 2022. +[5] Hong Wang, Bing Lu, Jun Li, Teng Liu, Yang Xing, Chen Lv, Dongpu Cao, Jingxuan Li, Jinwei Zhang, and Ehsan Hashemi. Risk assessment and mitigation in local path planning for autonomous vehicles with lstm based predictive model. IEEE Transactions on Automation Science and Engineering, 19(4):2738-2749, 2021. +[6] Chalavadi Vishnu, Vineel Abhinav, Debaditya Roy, C Krishna Mohan, and Ch Sobhan Babu. Improving multi-agent trajectory prediction using traffic states on interactive driving scenarios. IEEE Robotics and Automation Letters, 8(5):2708-2715, 2023. +[7] Xiao Han, Xinfeng Zhang, Yiling Wu, Zhenduo Zhang, Tianyu Zhang, and Yaowei Wang. Knowledge-based multiple relations modeling for traffic forecasting. IEEE Transactions on Intelligent Transportation Systems, 2024. +[8] Ye Yuan, Xinshuo Weng, Yanglan Ou, and Kris M Kitani. Agentformer: Agent-aware transformers for socio-temporal multi-agent forecasting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9813-9823, 2021. +[9] Tung Phan-Minh, Elena Corina Grigore, Freddy A Boulton, Oscar Beijbom, and Eric M Wolff. Covernet: Multimodal behavior prediction using trajectory sets. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14074-14083, 2020. +[10] Jiachen Li, Hengbo Ma, Zhihao Zhang, Jinning Li, and Masayoshi Tomizuka. Spatio-temporal graph dual-attention network for multi-agent prediction and tracking. IEEE Transactions on Intelligent Transportation Systems, 23(8):10556-10569, 2021. +[11] Mengmeng Liu, Hao Cheng, Lin Chen, Hellward Broszio, Jiangtao Li, Runjiang Zhao, Monika Sester, and Michael Ying Yang. Laformer: Trajectory prediction for autonomous driving with lane-aware scene constraints. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2039-2049, 2024. +[12] Nachiket Deo, Eric Wolff, and Oscar Beijbom. Multimodal trajectory prediction conditioned on lane-graph traversals. In Conference on Robot Learning, pages 203-212. PMLR, 2022. +[13] Jingke Wang, Tengju Ye, Ziqing Gu, and Junbo Chen. Ltp: Lane-based trajectory prediction for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17134-17142, 2022. +[14] Qingwen Xue, Yingying Xing, and Jian Lu. An integrated lane change prediction model incorporating traffic context based on trajectory data. Transportation research part C: emerging technologies, 141:103738, 2022. + +[15] Ross Greer, Nachiket Deo, and Mohan Trivedi. Trajectory prediction in autonomous driving with a lane heading auxiliary loss. IEEE Robotics and Automation Letters, 6(3):4907-4914, 2021. +[16] Haoran Li, Chaozhong Wu, Duanfeng Chu, Liping Lu, and Ken Cheng. Combined trajectory planning and tracking for autonomous vehicle considering driving styles. IEEE Access, 9:9453-9463, 2021. +[17] Maria Valentina Niño de Zepeda, Fanlin Meng, Jinya Su, Xiao-Jun Zeng, and Qian Wang. Dynamic clustering analysis for driving styles identification. Engineering applications of artificial intelligence, 97:104096, 2021. +[18] Harpreet Singh and Ankit Kathuria. Profiling drivers to assess safe and eco-driving behavior-a systematic review of naturalistic driving studies. Accident Analysis & Prevention, 161:106349, 2021. +[19] Wenshuo Wang, Junqiang Xi, and Ding Zhao. Driving style analysis using primitive driving patterns with bayesian nonparametric approaches. IEEE Transactions on Intelligent Transportation Systems, 20(8):2986-2998, 2018. +[20] Ziming Liu, Yixuan Wang, Sachin Vaidya, Fabian Ruehle, James Halverson, Marin Soljačić, Thomas Y Hou, and Max Tegmark. Kan: Kolmogorov-arnold networks. arXiv preprint arXiv:2404.19756, 2024. +[21] Wei Han, Wenshuo Wang, Xiaohan Li, and Junqiang Xi. Statistical-based approach for driving style recognition using bayesian probability with kernel density estimation. IET Intelligent Transport Systems, 13(1):22-30, 2019. +[22] K WEIERSTRASS. Über die analytische darstellbarkeit sogenannter willkürlicher Funktionen einer reellen. 1885. +[23] Zhezhang Ding and Huijing Zhao. Incorporating driving knowledge in deep learning based vehicle trajectory prediction: A survey. IEEE Transactions on Intelligent Vehicles, 8(8):3996-4015, 2023. +[24] Jingke Wang, Tengju Ye, Ziqing Gu, and Junbo Chen. Ltp: Lane-based trajectory prediction for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17134-17142, 2022. +[25] Tim Salzmann, Boris Ivanovic, Punarjay Chakravarty, and Marco Pavone. Trajectory++. Dynamically-feasible trajectory forecasting with heterogeneous data. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XVIII 16, pages 683-700. Springer, 2020. +[26] Xidong Feng, Zhepeng Cen, Jianming Hu, and Yi Zhang. Vehicle trajectory prediction using intention-based conditional variational autoencoder. In 2019 IEEE Intelligent Transportation Systems Conference (ITSC), pages 3514-3519. IEEE, 2019. +[27] Yutong Ban, Xiao Li, Guy Rosman, Igor Gilitschenski, Ozanan Meireles, Sertac Karaman, and Daniela Rus. A deep concept graph network for interaction-aware trajectory prediction. In 2022 International Conference on Robotics and Automation (ICRA), pages 8992-8998. IEEE, 2022. +[28] Sumit Kumar, Yiming Gu, Jerrick Hoang, Galen Clark Haynes, and Micol Marchetti-Bowick. Interaction-based trajectory prediction over a hybrid traffic graph. In 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 5530-5535. IEEE, 2021. +[29] Yu Wang, Shengjie Zhao, Rongqing Zhang, Xiang Cheng, and Liuqing Yang. Multi-vehicle collaborative learning for trajectory prediction with spatio-temporal tensor fusion. IEEE Transactions on Intelligent Transportation Systems, 23(1):236-248, 2020. +[30] Kaouther Messaoud, Itheri Yahiaoui, Anne Verroust-Blondet, and Fawzi Nashashibi. Attention based vehicle trajectory prediction. IEEE Transactions on Intelligent Vehicles, 6(1):175–185, 2020. +[31] Dongwei Xu, Xuetian Shang, Yewanze Liu, Hang Peng, and Hajjian Li. Group vehicle trajectory prediction with global spatio-temporal graph. IEEE Transactions on Intelligent Vehicles, 8(2):1219-1229, 2022. +[32] Tao Yang, Zhixiong Nan, He Zhang, Shitao Chen, and Nanning Zheng. Traffic agent trajectory prediction using social convolution and attention mechanism. In 2020 IEEE Intelligent Vehicles Symposium (IV), pages 278-283. IEEE, 2020. +[33] Andrei Nikolaevich Kolmogorov. On the representation of continuous functions of several variables by superpositions of continuous functions of a smaller number of variables. American Mathematical Society, 1961. + +[34] Andrei Nikolaevich Kolmogorov. On the representation of continuous functions of many variables by superposition of continuous functions of one variable and addition. In Doklady Akademii Nauk, volume 114, pages 953-956. Russian Academy of Sciences, 1957. +[35] Jürgen Braun and Michael Griebel. On a constructive proof of kolmogorov's superposition theorem. Constructive approximation, 30:653-675, 2009. +[36] Andrei Nikolaevich Kolmogorov. On the representation of continuous functions of many variables by superposition of continuous functions of one variable and addition. In Doklady Akademii Nauk, volume 114, pages 953-956. Russian Academy of Sciences, 1957. +[37] Vladimir Igorevich Arnol'd. On the representation of continuous functions of three variables by superpositions of continuous functions of two variables. Matematicheskii Sbornik, 90(1):3-74, 1959. +[38] Věra Kürková. Kolmogorov's theorem is relevant. Neural computation, 3(4):617-622, 1991. +[39] Martin Riedmiller. Advanced supervised learning in multi-layer perceptrons—from backpropagation to adaptive learning algorithms. Computer Standards & Interfaces, 16(3):265-278, 1994. +[40] Rudolf Kruse, Sanaz Mostaghim, Christian Borgelt, Christian Braune, and Matthias Steinbrecher. Multi-layer perceptrons. In Computational intelligence: a methodological introduction, pages 53-124. Springer, 2022. +[41] Songtao Huang, Zhen Zhao, Can Li, and Lei Bai. Timekan: Kan-based frequency decomposition learning architecture for long-term time series forecasting. arXiv preprint arXiv:2502.06910, 2025. +[42] Diab W Abueidda, Panos Pantidis, and Mostafa E Mobasher. Deepokan: Deep operator network based on kolmogorov arnold networks for mechanics problems. Computer Methods in Applied Mechanics and Engineering, 436:117699, 2025. +[43] Chenxin Li, Xinyu Liu, Wuyang Li, Cheng Wang, Hengyu Liu, Yifan Liu, Zhen Chen, and Yixuan Yuan. U-kan makes strong backbone for medical image segmentation and generation. arXiv preprint arXiv:2406.02918, 2024. +[44] Benjamin C Koenig, Suyong Kim, and Sili Deng. Kan-odes: Kolmogorov-arnold network ordinary differential equations for learning dynamical systems and hidden physics. Computer Methods in Applied Mechanics and Engineering, 432:117397, 2024. +[45] Alireza Afzal Aghaei. fkan: Fractional kolmogorov-arnold networks with trainable jacobi basis functions. Neurocomputing, page 129414, 2025. +[46] Akash Kundu, Aritra Sarkar, and Abhishek Sadhu. Kanqas: Kolmogorov-arnold network for quantum architecture search. EPJ Quantum Technology, 11(1):76, 2024. +[47] Khemraj Shukla, Juan Diego Toscano, Zhicheng Wang, Zongren Zou, and George Em Karniadakis. A comprehensive and fair comparison between mlp and kan representations for differential equations and operator networks. Computer Methods in Applied Mechanics and Engineering, 431:117290, 2024. +[48] John Smith and Jane Doe. Real-time driving style classification based on short-term observations. Transportation Research Part A: Policy and Practice, 135:105-116, 2021. +[49] Clara Marina Martinez, Mira Heucke, Fei-Yue Wang, Bo Gao, and Dongpu Cao. Driving style recognition for intelligent vehicle control and advanced driver assistance: A survey. IEEE Transactions on Intelligent Transportation Systems, 19(3):666-676, 2017. +[50] Shiyu Fang, Peng Hang, Chongfeng Wei, Yang Xing, and Jian Sun. Cooperative driving of connected autonomous vehicles in heterogeneous mixed traffic: A game theoretic approach. IEEE Transactions on Intelligent Vehicles, 2024. +[51] Nour O Khanfar, Mohammed Elhenawy, Huthaifa I Ashqar, Qinaat Hussain, and Wael KM Alhajyaseen. Driving behavior classification at signalized intersections using vehicle kinematics: Application of unsupervised machine learning. International journal of injury control and safety promotion, 30(1):34-44, 2023. +[52] S Bernstein. Proof of the theorem of weierstrass based on the calculus of probabilities. Communications of the Kharkov Mathematical Society, 13:1-2, 1912. +[53] Jin-Hyuk Hong, Ben Margines, and Anind K Dey. A smartphone-based sensing platform to model aggressive driving behaviors. In Proceedings of the sigchi conference on human factors in computing systems, pages 4047-4056, 2014. + +[54] Walter Rudin et al. Principles of mathematical analysis, volume 3. McGraw-hill New York, 1964. +[55] Cornelius Lanczos. Solution of systems of linear equations by minimized iterations. J. Res. Nat. Bur. Standards, 49(1):33-53, 1952. +[56] Theodore J Rivlin. An introduction to the approximation of functions. Courier Corporation, 1981. +[57] Ahmad Aljaafreh, Nabeel Alshabatat, and Munaf S Najim Al-Din. Driving style recognition using fuzzy logic. In 2012 IEEE International Conference on Vehicular Electronics and Safety (ICVES 2012), pages 460-463. IEEE, 2012. +[58] Gabor Szeg. Orthogonal polynomials, volume 23. American Mathematical Soc., 1939. +[59] Theodore S Chihara. An introduction to orthogonal polynomials. Courier Corporation, 2011. +[60] George E Andrews, Richard Askey, Ranjan Roy, Ranjan Roy, and Richard Askey. *Special functions*, volume 71. Cambridge university press Cambridge, 1999. +[61] Chaopeng Zhang, Wenshuo Wang, Zhaokun Chen, Jian Zhang, Lijun Sun, and Junqiang Xi. Shareable driving style learning and analysis with a hierarchical latent model. IEEE Transactions on Intelligent Transportation Systems, 2024. +[62] Dian Jing, Enjian Yao, and Rongsheng Chen. Decentralized human-like control strategy of mixed-flow multi-vehicle interactions at uncontrolled intersections: A game-theoretic approach. Transportation Research Part C: Emerging Technologies, 167:104835, 2024. +[63] Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton, and Jeff Dean. Outrageously large neural networks: The sparsely-gated mixture-of-experts layer. arXiv preprint arXiv:1701.06538, 2017. +[64] Günter Klambauer, Thomas Unterthiner, Andreas Mayr, and Sepp Hochreiter. Self-normalizing neural networks. Advances in neural information processing systems, 30, 2017. +[65] Hao Zheng, Zhanlei Yang, Wenju Liu, Jizhong Liang, and Yanpeng Li. Improving deep neural networks using softplus units. In 2015 International joint conference on neural networks (IJCNN), pages 1-4. IEEE, 2015. +[66] Marius Lindauer, Katharina Eggensperger, Matthias Feurer, André Biedenkapp, Difan Deng, Carolin Benjamins, Tim Ruhkopf, René Sass, and Frank Hutter. Smac3: A versatile bayesian optimization package for hyperparameter optimization. Journal of Machine Learning Research, 23(54):1-9, 2022. +[67] Holger Caesar, Varun Bankiti, Alex H. Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. arXiv preprint arXiv:1903.11027, 2019. +[68] Ming-Fang Chang, John W Lambert, Patsorn Sangkloy, Jagjeet Singh, Slawomir Bak, Andrew Hartnett, De Wang, Peter Carr, Simon Lucey, Deva Ramanan, and James Hays. Argoverse: 3d tracking and forecasting with rich maps. In Conference on Computer Vision and Pattern Recognition (CVPR), 2019. +[69] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2446-2454, 2020. +[70] Thomas Gilles, Stefano Sabatini, Dzmitry Tsishkou, Bogdan Stanciulescu, and Fabien Moutarde. Thomas: Trajectory heatmap output with learned multi-agent sampling. In International Conference on Learning Representations, 2022. +[71] Thomas Gilles, Stefano Sabatini, Dzmitry Tsishkou, Bogdan Stanciulescu, and Fabien Moutarde. Gohome: Graph-oriented heatmap output for future motion estimation. In 2022 international conference on robotics and automation (ICRA), pages 9107-9114. IEEE, 2022. +[72] Chenfeng Xu, Tian Li, Chen Tang, Lingfeng Sun, Kurt Keutzer, Masayoshi Tomizuka, Alireza Fathi, and Wei Zhan. Pretram: Self-supervised pre-training via connecting trajectory and map. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXIX, pages 34-50. Springer, 2022. +[73] Chuhua Wang, Yuchen Wang, Mingze Xu, and David J Crandall. Stepwise goal-driven networks for trajectory prediction. IEEE Robotics and Automation Letters, 7(2):2716-2723, 2022. + +[74] Balakrishnan Varadarajan, Ahmed Hefny, Avikalp Srivastava, Khaled S Refaat, Nigamaa Nayakanti, Andre Cornman, Kan Chen, Bertrand Douillard, Chi Pang Lam, Dragomir Anguelov, et al. Multipath++: Efficient information fusion and trajectory aggregation for behavior prediction. In 2022 International Conference on Robotics and Automation (ICRA), pages 7814-7821. IEEE, 2022. +[75] Mihee Lee, Samuel S Sohn, Seonghyeon Moon, Sejong Yoon, Mubbasir Kapadia, and Vladimir Pavlovic. Muse-vae: multi-scale vae for environment-aware long term trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2221–2230, 2022. +[76] Zikang Zhou, Luyao Ye, Jianping Wang, Kui Wu, and Kejie Lu. Hivt: Hierarchical vector transformer for multi-agent motion prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8823-8833, 2022. +[77] Linhui Li, Xuecheng Wang, Dongfang Yang, Yifan Ju, Zhongxu Zhang, and Jing Lian. Real-time heterogeneous road-agents trajectory prediction using hierarchical convolutional networks and multi-task learning. IEEE Transactions on Intelligent Vehicles, 2023. +[78] Görkay Aydemir, Adil Kaan Akan, and Fatma Güney. Adapt: Efficient multi-agent trajectory prediction with adaptation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8295-8305, 2023. +[79] Pei Xu, Jean-Bernard Hayet, and Ioannis Karamouzas. Context-aware timewise vaes for real-time vehicle trajectory prediction. IEEE Robotics and Automation Letters, 8(9):5440-5447, 2023. +[80] Daehee Park, Hobin Ryu, Yunseo Yang, Jegyeong Cho, Jiwon Kim, and Kuk-Jin Yoon. Leveraging future relationship reasoning for vehicle trajectory prediction. In International Conference on Learning Representations (ICLR 2023). Eleventh International Conference on Learning Representations, 2023. +[81] Pei Xu, Jean-Bernard Hayet, and Ioannis Karamouzas. Context-aware timewise vaes for real-time vehicle trajectory prediction. IEEE Robotics and Automation Letters, 8(9):5440-5447, 2023. +[82] Sehwan Choi, Jungho Kim, Junyong Yun, and Jun Won Choi. R-pred: Two-stage motion prediction via tube-query attention-based trajectory refinement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8525-8535, 2023. +[83] Di Wen, Haoran Xu, Zhaocheng He, Zhe Wu, Guang Tan, and Peixi Peng. Density-adaptive model based on motif matrix for multi-agent trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14822-14832, 2024. +[84] Maximilian Schäfer, Kun Zhao, and Anton Kummert. Caspnet++: Joint multi-agent motion prediction. In 2024 IEEE Intelligent Vehicles Symposium (IV), pages 1294-1301. IEEE, 2024. +[85] Xishun Wang, Tong Su, Fang Da, and Xiaodong Yang. Prophnet: Efficient agent-centric motion forecasting with anchor-informed proposals. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21995-22003, 2023. +[86] Harsh Yadav, Maximilian Schaefer, Kun Zhao, and Tobias Meisen. Caspformer: Trajectory prediction from bev images with deformable attention. In International Conference on Pattern Recognition, pages 420-434. Springer, 2025. +[87] Zikang Zhou, Jianping Wang, Yung-Hui Li, and Yu-Kai Huang. Query-centric trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17863-17873, 2023. +[88] Jiquan Ngiam, Benjamin Caine, Vijay Vasudevan, Zhengdong Zhang, Hao-Tien Lewis Chiang, Jeffrey Ling, Rebecca Roelofs, Alex Bewley, Chenxi Liu, Ashish Venugopal, et al. Scene transformer: A unified architecture for predicting multiple agent trajectories. arXiv preprint arXiv:2106.08417, 2021. +[89] Stepan Konev. Mpa: Multipath++ based architecture for motion prediction. arXiv preprint arXiv:2206.10041, 2022. +[90] Zhiyu Huang, Xiaoyu Mo, and Chen Lv. Recoat: A deep learning-based framework for multi-modal motion prediction in autonomous driving application. In 2022 IEEE 25th International Conference on Intelligent Transportation Systems (ITSC), pages 988-993. IEEE, 2022. +[91] Zhiyu Huang, Haochen Liu, Jingda Wu, and Chen Lv. Differentiable integrated motion prediction and planning with learnable cost function for autonomous driving. IEEE transactions on neural networks and learning systems, 2023. + +[92] Camiel Oerlemans, Bram Grooten, Michiel Braat, Alaa Alassi, Emilia Silvas, and Decebal Constantin Mocanu. Limtr: Time series motion prediction for diverse road users through multimodal feature integration. arXiv preprint arXiv:2410.15819, 2024. +[93] Xiaosong Jia, Penghao Wu, Li Chen, Yu Liu, Hongyang Li, and Junchi Yan. Hdgt: Heterogeneous driving graph transformer for multi-agent trajectory prediction via scene encoding. IEEE transactions on pattern analysis and machine intelligence, 45(11):13860-13875, 2023. +[94] Ari Seff, Brian Cera, Dian Chen, Mason Ng, Aurick Zhou, Nigamaa Nayakanti, Khaled S Refaat, Rami Al-Rfou, and Benjamin Sapp. Motionlm: Multi-agent motion forecasting as language modeling. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8579-8590, 2023. +[95] Shaoshuai Shi, Li Jiang, Dengxin Dai, and Bernt Schiele. Mtr++: Multi-agent motion prediction with symmetric scene modeling and guided intention querying. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(5):3955-3971, 2024. +[96] Xiaoji Zheng, Lixiu Wu, Zhijie Yan, Yuanrong Tang, Hao Zhao, Chen Zhong, Bokui Chen, and Jiangtao Gong. Large language models powered context-aware motion prediction in autonomous driving. In 2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 980-985. IEEE, 2024. +[97] Jiawei Sun, Chengran Yuan, Shuo Sun, Shanze Wang, Yuhang Han, Shuailei Ma, Zefan Huang, Anthony Wong, Keng Peng Tee, and Marcelo H Ang. Controlmtr: Control-guided motion transformer with scenecompliant intention points for feasible motion prediction. In 2024 IEEE 27th International Conference on Intelligent Transportation Systems (ITSC), pages 1507-1514. IEEE, 2024. +[98] Nejla Özmen and Esra Erkus-Duman. On the poisson-charlier polynomials. Serdica Mathematical Journal, 41(4):457p-470p, 2015. +[99] Margit Rosler. Generalized hermite polynomials and the heat equation for Dunkl operators. Communications in Mathematical Physics, 192:519-542, 1998. +[100] Yanjun Huang, Jiatong Du, Ziru Yang, Zewei Zhou, Lin Zhang, and Hong Chen. A survey on trajectory-prediction methods for autonomous driving. IEEE Transactions on Intelligent Vehicles, 7(3):652-674, 2022. +[101] Thomas Batz, Kym Watson, and Jurgen Beyerer. Recognition of dangerous situations within a cooperative group of vehicles. In 2009 IEEE Intelligent Vehicles Symposium, pages 907-912. IEEE, 2009. +[102] Mattias Brännström, Erik Coelingh, and Jonas Sjöberg. Model-based threat assessment for avoiding arbitrary vehicle collisions. IEEE Transactions on Intelligent Transportation Systems, 11(3):658-669, 2010. +[103] Helgo Dyckmanns, Richard Matthaei, Markus Maurer, Bernd Lichte, Jan Effertz, and Dirk Stüker. Object tracking in urban intersections based on active use of a priori knowledge: Active interacting multi model filter. In 2011 IEEE Intelligent Vehicles Symposium (IV), pages 625-630. IEEE, 2011. +[104] Vasileios Lefkopoulos, Marcel Menner, Alexander Domahidi, and Melanie N Zeilinger. Interaction-aware motion prediction for autonomous driving: A multiple model kalman filtering scheme. IEEE Robotics and Automation Letters, 6(1):80-87, 2020. +[105] Yijing Wang, Zhengxuan Liu, Zhiqiang Zuo, Zheng Li, Li Wang, and Xiaoyuan Luo. Trajectory planning and safety assessment of autonomous vehicles based on motion prediction and model predictive control. IEEE Transactions on Vehicular Technology, 68(9):8546-8556, 2019. +[106] Haoran Song, Di Luan, Wenchao Ding, Michael Y Wang, and Qifeng Chen. Learning to predict vehicle trajectories with model-based planning. In Conference on Robot Learning, pages 1035-1045. PMLR, 2022. +[107] Quan Tran and Jonas Firl. Online maneuver recognition and multimodal trajectory prediction for intersection assistance using non-parametric regression. In 2014 IEEE intelligent vehicles symposium proceedings, pages 918-923. IEEE, 2014. +[108] Yuande Jiang, Bing Zhu, Shun Yang, Jian Zhao, and Weiwen Deng. Vehicle trajectory prediction considering driver uncertainty and vehicle dynamics based on dynamic bayesian network. IEEE Transactions on Systems, Man, and Cybernetics: Systems, 2022. + +[109] Samet Ayhan and Hanan Samet. Aircraft trajectory prediction made easy with predictive analytics. In Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pages 21–30, 2016. +[110] Zhibin Qiu, Jiangjun Ruan, Daochun Huang, Ziheng Pu, and Shengwen Shu. A prediction method for breakdown voltage of typical air gaps based on electric field features and support vector machine. IEEE Transactions on Dielectrics and Electrical Insulation, 22(4):2125-2135, 2015. +[111] Matthias Schreier, Volker Willert, and Jürgen Adamy. An integrated approach to maneuver-based trajectory prediction and criticality assessment in arbitrary road environments. IEEE Transactions on Intelligent Transportation Systems, 17(10):2751-2766, 2016. +[112] Junru Gu, Chen Sun, and Hang Zhao. Densetnt: End-to-end trajectory prediction from dense goal sets. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15303-15312, 2021. +[113] Hashmatullah Sadid and Constantinos Antoniou. Dynamic spatio-temporal graph neural network for surrounding-aware trajectory prediction of autonomous vehicles. IEEE Transactions on Intelligent Vehicles, 2024. +[114] DN Jagadish, Arun Chauhan, and Lakshman Mahto. Conditional variational autoencoder networks for autonomous vehicle path prediction. Neural Processing Letters, 54(5):3965-3978, 2022. +[115] Zilai Zeng, Ce Zhang, Shijie Wang, and Chen Sun. Goal-conditioned predictive coding for offline reinforcement learning. Advances in Neural Information Processing Systems, 36, 2024. +[116] Rushdi Alsaleh and Tarek Sayed. Modeling pedestrian-cyclist interactions in shared space using inverse reinforcement learning. Transportation research part F: traffic psychology and behaviour, 70:37-57, 2020. +[117] Zan Yang, Wei Nai, Dan Li, Lu Liu, and Ziyu Chen. A mixed generative adversarial imitation learning based vehicle path planning algorithm. IEEE Access, 2024. +[118] David Hilbert. Über die gleichung neunten grades. In Algebra- Invariantentheorie Geometrie, pages 393-400. Springer, 1970. +[119] Ken-Ichi Funahashi. On the approximate realization of continuous mappings by neural networks. Neural networks, 2(3):183-192, 1989. +[120] George Cybenko. Approximation by superpositions of a sigmoidal function. Mathematics of control, signals and systems, 2(4):303-314, 1989. +[121] Kurt Hornik, Maxwell Stinchcombe, and Halbert White. Multilayer feedforward networks are universal approximators. Neural networks, 2(5):359-366, 1989. +[122] Kurt Hornik. Approximation capabilities of multilayer feedforward networks. Neural networks, 4(2):251-257, 1991. +[123] Shriyank Somvanshi, Syed Aaqib Javed, Md Monzurul Islam, Diwas Pandit, and Subasish Das. A survey on kolmogorov-arnold network. arXiv preprint arXiv:2411.06078, 2024. +[124] Kexin Ma, Xu Lu, Bragazzi Luigi Nicola, and Biao Tang. Integrating kolmogorov-arnold networks with ordinary differential equations for efficient, interpretable and robust deep learning: A case study in the epidemiology of infectious diseases. medRxiv, pages 2024–09, 2024. +[125] Alexander Dylan Bodner, Antonio Santiago Tepsich, Jack Natan Spolski, and Santiago Pourteau. Convolutional kolmogorov-arnold networks. arXiv preprint arXiv:2406.13155, 2024. +[126] Zavareh Bozorgasl and Hao Chen. Wav-kan: Wavelet kolmogorov-arnold networks. arXiv preprint arXiv:2405.12832, 2024. +[127] Cristian J Vaca-Rubio, Luis Blanco, Roberto Pereira, and Mário Caus. Kolmogorov-arnold networks (kans) for time series analysis. arXiv preprint arXiv:2405.08790, 2024. +[128] Ioannis E Livieris. C-kan: A new approach for integrating convolutional layers with kolmogorov-arnold networks for time-series forecasting. Mathematics, 12(19):3022, 2024. +[129] Remi Genet and Hugo Inzirillo. Tkan: Temporal kolmogorov-arnold networks. arXiv preprint arXiv:2405.07344, 2024. + +[130] Glenn W Brier. Verification of forecasts expressed in terms of probability. Monthly weather review, 78(1):1-3, 1950. +[131] A. N. Kolmogorov. On the best approximation of continuous functions. Doklady Akademii Nauk SSSR, 1947:Approximate page numbers if available, 1947. +[132] P Billingsley. Probability and measure. 3rd wiley. New York, 1995. +[133] Norman L Johnson, Adrienne W Kemp, and Samuel Kotz. Univariate discrete distributions, volume 444. John Wiley & Sons, 2005. +[134] A. F. Timan. Theory of Approximation of Functions of a Real Variable, volume 34 of International Series of Monographs in Pure and Applied Mathematics. Pergamon Press, 1963. +[135] William W Hager. Lipschitz continuity for constrained processes. SIAM Journal on Control and Optimization, 17(3):321-338, 1979. +[136] JC Ferrando and LM Sánchez Ruiz. A survey on recent advances on the nikodym boundedness theorem and spaces of simple functions. The Rocky Mountain Journal of Mathematics, pages 139-172, 2004. +[137] Andrew Alleyne. Improved vehicle performance using combined suspension and braking forces. Vehicle System Dynamics, 27(4):235-265, 1997. +[138] Ksander N de Winkel, Tugrul Irmak, Riender Happee, and Barys Shyrokau. Standards for passenger comfort in automated vehicles: Acceleration and jerk. Applied Ergonomics, 106:103881, 2023. +[139] Ram Venkataraman Iyer, Xiaobo Tan, and Perinkulam S Krishnaprasad. Approximate inversion of the preisach hysteresis operator with application to control of smart actuators. IEEE Transactions on automatic control, 50(6):798-810, 2005. +[140] Gordon Frank Newell. A simplified car-following theory: a lower order model. Transportation Research Part B: Methodological, 36(3):195-205, 2002. +[141] Zhen Yao, Xin Li, Bo Lang, and Mooi Choo Chuah. Goal-lbp: Goal-based local behavior guided trajectory prediction for autonomous driving. IEEE Transactions on Intelligent Transportation Systems, 2023. +[142] Peter G Gipps. A behavioural car-following model for computer simulation. Transportation research part B: methodological, 15(2):105-111, 1981. +[143] Antoine Ayache and Jacques Lévy Vehel. On the identification of the pointwise Hölder exponent of the generalized multifractional brownian motion. Stochastic Processes and their Applications, 111(1):119-156, 2004. +[144] Anupriya Vysala and Joseph Gomes. Evaluating and validating cluster results. arXiv preprint arXiv:2007.08034, 2020. +[145] Mingqian Li, Panrong Tong, Mo Li, Zhongming Jin, Jianqiang Huang, and Xian-Sheng Hua. Traffic flow prediction with vehicle trajectories. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 294-302, 2021. +[146] Xiaoyu Mo, Zhiyu Huang, Yang Xing, and Chen Lv. Multi-agent trajectory prediction with heterogeneous edge-enhanced graph attention network. IEEE Transactions on Intelligent Transportation Systems, 23(7):9554-9567, 2022. +[147] Lan Feng, Mohammadhossein Bahari, Kaouthier Messaoud Ben Amor, Éloi Zablocki, Matthieu Cord, and Alexandre Alahi. Unitraj: A unified framework for scalable vehicle trajectory prediction. In European Conference on Computer Vision, pages 106-123. Springer, 2024. +[148] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In 2012 IEEE conference on computer vision and pattern recognition, pages 3354-3361. IEEE, 2012. +[149] Xinyu Huang, Xinjing Cheng, Qichuan Geng, Binbin Cao, Dingfu Zhou, Peng Wang, Yuanqing Lin, and Ruigang Yang. The apolloscape dataset for autonomous driving. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 954-960, 2018. +[150] Wei Zhan, Liting Sun, Di Wang, Haojie Shi, Aubrey Clausse, Maximilian Naumann, Julius Kummerle, Hendrik Konigshof, Christoph Stiller, Arnaud de La Fortelle, et al. Interaction dataset: An international, adversarial and cooperative motion dataset in interactive driving scenarios with semantic maps. arXiv preprint arXiv:1910.03088, 2019. + +[151] Julian Bock, Robert Krajewski, Tobias Moers, Steffen Runde, Lennart Vater, and Lutz Eckstein. The ind dataset: A drone dataset of naturalistic road user trajectories at german intersections. In 2020 IEEE Intelligent Vehicles Symposium (IV), pages 1929-1934. IEEE, 2020. +[152] Robert Krajewski, Tobias Moers, Julian Bock, Lennart Vater, and Lutz Eckstein. The round dataset: A drone dataset of road user trajectories at roundabouts in germany. In 2020 IEEE 23rd International Conference on Intelligent Transportation Systems (ITSC), pages 1-6, 2020. +[153] Robert Krajewski, Julian Bock, Laurent Kloeker, and Lutz Eckstein. The highd dataset: A drone dataset of naturalistic vehicle trajectories on german highways for validation of highly automated driving systems. In 2018 21st international conference on intelligent transportation systems (ITSC), pages 2118-2125. IEEE, 2018. + +![](images/e6f40bc5489f55d1573192e9cc91218c42fdb95b92abce31e36e773c4fa548f3.jpg) + +The diagram illustrates of three driving styles: Conservative, Aggressive and Normal (CAN), with their corresponding trajectories represented by lines recorded at each time step. The lengths of the lines indicate the driving distance, while the direction is shown by arrows. Conservative drivers typically move at low speeds or stop to avoid obstacles. Aggressive drivers often travel at high speeds and are prone to overtaking other vehicles. Normal drivers maintain a constant speed and frequently change lanes to ensure safety. + +# NeurIPS Paper Checklist + +# 1. Claims + +Question: Do the main claims made in the abstract and introduction accurately reflect the paper's contributions and scope? + +Answer: [Yes]. + +Justification: The main claims illustrate in Figure 1 in Section 1 (introduction). Our main contributions in this paper are summarized as follows: + +- To handle the vehicle trajectory prediction task, we propose for the first time a novel Driving-Style-Adaptive (DSA) framework tailored to the driving styles of human drivers, effectively leverages trajectory information. +- We utilize polynomial approximation operators to approximate and predict trajectories under different driving styles: Conservative, Aggressive and Normal (CAN). These operators support a mathematical explanation matching mechanism between driving style with a corresponding polynomial form. +- The experimental results on the real-world datasets (nuScenes, Argoverse and Waymo) demonstrate that our model significantly outperforms existing methods in vehicle trajectory prediction. + +# Guidelines: + +- The answer NA means that the abstract and introduction do not include the claims made in the paper. +- The abstract and/or introduction should clearly state the claims made, including the contributions made in the paper and important assumptions and limitations. A No or NA answer to this question will not be perceived well by the reviewers. +- The claims made should match theoretical and experimental results, and reflect how much the results can be expected to generalize to other settings. +- It is fine to include aspirational goals as motivation as long as it is clear that these goals are not attained by the paper. + +# 2. Limitations + +Question: Does the paper discuss the limitations of the work performed by the authors? + +Answer: [Yes]. + +Justification: We discuss the limitation in Appendix. + +# Guidelines: + +- The answer NA means that the paper has no limitation while the answer No means that the paper has limitations, but those are not discussed in the paper. +- The authors are encouraged to create a separate "Limitations" section in their paper. + +- The paper should point out any strong assumptions and how robust the results are to violations of these assumptions (e.g., independence assumptions, noiseless settings, model well-specification, asymptotic approximations only holding locally). The authors should reflect on how these assumptions might be violated in practice and what the implications would be. +- The authors should reflect on the scope of the claims made, e.g., if the approach was only tested on a few datasets or with a few runs. In general, empirical results often depend on implicit assumptions, which should be articulated. +- The authors should reflect on the factors that influence the performance of the approach. For example, a facial recognition algorithm may perform poorly when image resolution is low or images are taken in low lighting. Or a speech-to-text system might not be used reliably to provide closed captions for online lectures because it fails to handle technical jargon. +- The authors should discuss the computational efficiency of the proposed algorithms and how they scale with dataset size. +- If applicable, the authors should discuss possible limitations of their approach to address problems of privacy and fairness. +- While the authors might fear that complete honesty about limitations might be used by reviewers as grounds for rejection, a worse outcome might be that reviewers discover limitations that aren't acknowledged in the paper. The authors should use their best judgment and recognize that individual actions in favor of transparency play an important role in developing norms that preserve the integrity of the community. Reviewers will be specifically instructed to not penalize honesty concerning limitations. + +# 3. Theory assumptions and proofs + +Question: For each theoretical result, does the paper provide the full set of assumptions and a complete (and correct) proof? + +Answer: [Yes]. + +Justification: We provide the mathematics properties for matching driving style and corresponding polynomial in Section 3.2, for full proof is in Appendix C. + +Guidelines: + +- The answer NA means that the paper does not include theoretical results. +- All the theorems, formulas, and proofs in the paper should be numbered and cross-referenced. +- All assumptions should be clearly stated or referenced in the statement of any theorems. +- The proofs can either appear in the main paper or the supplemental material, but if they appear in the supplemental material, the authors are encouraged to provide a short proof sketch to provide intuition. +- Inversely, any informal proof provided in the core of the paper should be complemented by formal proofs provided in appendix or supplemental material. +- Theorems and Lemmas that the proof relies upon should be properly referenced. + +# 4. Experimental result reproducibility + +Question: Does the paper fully disclose all the information needed to reproduce the main experimental results of the paper to the extent that it affects the main claims and/or conclusions of the paper (regardless of whether the code and data are provided or not)? + +Answer: [Yes]. + +Justification: Please refer to Section 3.3.1, 3.3.2 and the released codes in supplemental material. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- If the paper includes experiments, a No answer to this question will not be perceived well by the reviewers: Making the paper reproducible is important, regardless of whether the code and data are provided or not. +- If the contribution is a dataset and/or model, the authors should describe the steps taken to make their results reproducible or verifiable. +- Depending on the contribution, reproducibility can be accomplished in various ways. For example, if the contribution is a novel architecture, describing the architecture fully might suffice, or if the contribution is a specific model and empirical evaluation, it may be necessary to either make it possible for others to replicate the model with the same dataset, or provide access to the model. In general, releasing code and data is often one good way to accomplish this, but reproducibility can also be provided via detailed instructions for how to replicate the results, access to a hosted model (e.g., in the case of a large language model), releasing of a model checkpoint, or other means that are appropriate to the research performed. + +- While NeurIPS does not require releasing code, the conference does require all submissions to provide some reasonable avenue for reproducibility, which may depend on the nature of the contribution. For example + +(a) If the contribution is primarily a new algorithm, the paper should make it clear how to reproduce that algorithm. +(b) If the contribution is primarily a new model architecture, the paper should describe the architecture clearly and fully. +(c) If the contribution is a new model (e.g., a large language model), then there should either be a way to access this model for reproducing the results or a way to reproduce the model (e.g., with an open-source dataset or instructions for how to construct the dataset). +(d) We recognize that reproducibility may be tricky in some cases, in which case authors are welcome to describe the particular way they provide for reproducibility. In the case of closed-source models, it may be that access to the model is limited in some way (e.g., to registered users), but it should be possible for other researchers to have some path to reproducing or verifying the results. + +# 5. Open access to data and code + +Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material? + +Answer: [Yes]. + +Justification: We provide the code in supplemental material. + +Guidelines: + +- The answer NA means that paper does not include experiments requiring code. +- Please see the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details. +- While we encourage the release of code and data, we understand that this might not be possible, so "No" is an acceptable answer. Papers cannot be rejected simply for not including code, unless this is central to the contribution (e.g., for a new open-source benchmark). +- The instructions should contain the exact command and environment needed to run to reproduce the results. See the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details. +- The authors should provide instructions on data access and preparation, including how to access the raw data, preprocessed data, intermediate data, and generated data, etc. +- The authors should provide scripts to reproduce all experimental results for the new proposed method and baselines. If only a subset of experiments are reproducible, they should state which ones are omitted from the script and why. +- At submission time, to preserve anonymity, the authors should release anonymized versions (if applicable). +- Providing as much information as possible in supplemental material (appended to the paper) is recommended, but including URLs to data and code is permitted. + +# 6. Experimental setting/details + +Question: Does the paper specify all the training and test details (e.g., data splits, hyperparameters, how they were chosen, type of optimizer, etc.) necessary to understand the results? + +Answer: [Yes]. + +Justification: We detail are in Section 3.3.2. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The experimental setting should be presented in the core of the paper to a level of detail that is necessary to appreciate the results and make sense of them. +- The full details can be provided either with the code, in appendix, or as supplemental material. + +# 7. Experiment statistical significance + +Question: Does the paper report error bars suitably and correctly defined or other appropriate information about the statistical significance of the experiments? + +Answer: [No]. + +Justification: This paper does not report error bars following the practice of previous work [70, 71, 88]. + +Guidelines: + +- The answer NA means that the paper does not include experiments. + +- The authors should answer "Yes" if the results are accompanied by error bars, confidence intervals, or statistical significance tests, at least for the experiments that support the main claims of the paper. +- The factors of variability that the error bars are capturing should be clearly stated (for example, train/test split, initialization, random drawing of some parameter, or overall run with given experimental conditions). +- The method for calculating the error bars should be explained (closed form formula, call to a library function, bootstrap, etc.) +- The assumptions made should be given (e.g., Normally distributed errors). +- It should be clear whether the error bar is the standard deviation or the standard error of the mean. +- It is OK to report 1-sigma error bars, but one should state it. The authors should preferably report a 2-sigma error bar than state that they have a $96\%$ CI, if the hypothesis of Normality of errors is not verified. +- For asymmetric distributions, the authors should be careful not to show in tables or figures symmetric error bars that would yield results that are out of range (e.g. negative error rates). +- If error bars are reported in tables or plots, The authors should explain in the text how they were calculated and reference the corresponding figures or tables in the text. + +# 8. Experiments compute resources + +Question: For each experiment, does the paper provide sufficient information on the computer resources (type of compute workers, memory, time of execution) needed to reproduce the experiments? + +Answer: [Yes]. + +Justification: Please refer to Section 4.2.1 and Section 4.3. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The paper should indicate the type of compute workers CPU or GPU, internal cluster, or cloud provider, including relevant memory and storage. +- The paper should provide the amount of compute required for each of the individual experimental runs as well as estimate the total compute. +- The paper should disclose whether the full research project required more compute than the experiments reported in the paper (e.g., preliminary or failed experiments that didn't make it into the paper). + +# 9. Code of ethics + +Question: Does the research conducted in the paper conform, in every respect, with the NeurIPS Code of Ethics https://neurips.cc/public/EthicsGuidelines? + +Answer: [Yes]. + +Justification: This paper conforms with the NeurIPS Code of Ethics. + +Guidelines: + +- The answer NA means that the authors have not reviewed the NeurIPS Code of Ethics. +- If the authors answer No, they should explain the special circumstances that require a deviation from the Code of Ethics. +- The authors should make sure to preserve anonymity (e.g., if there is a special consideration due to laws or regulations in their jurisdiction). + +# 10. Broader impacts + +Question: Does the paper discuss both potential positive societal impacts and negative societal impacts of the work performed? + +Answer: [NA]. + +Justification: There is no societal impact of the work performed. + +Guidelines: + +- The answer NA means that there is no societal impact of the work performed. +- If the authors answer NA or No, they should explain why their work has no societal impact or why the paper does not address societal impact. +- Examples of negative societal impacts include potential malicious or unintended uses (e.g., disinformation, generating fake profiles, surveillance), fairness considerations (e.g., deployment of technologies that could make decisions that unfairly impact specific groups), privacy considerations, and security considerations. + +- The conference expects that many papers will be foundational research and not tied to particular applications, let alone deployments. However, if there is a direct path to any negative applications, the authors should point it out. For example, it is legitimate to point out that an improvement in the quality of generative models could be used to generate deepfakes for disinformation. On the other hand, it is not needed to point out that a generic algorithm for optimizing neural networks could enable people to train models that generate Deepfakes faster. +- The authors should consider possible harms that could arise when the technology is being used as intended and functioning correctly, harms that could arise when the technology is being used as intended but gives incorrect results, and harms following from (intentional or unintentional) misuse of the technology. +- If there are negative societal impacts, the authors could also discuss possible mitigation strategies (e.g., gated release of models, providing defenses in addition to attacks, mechanisms for monitoring misuse, mechanisms to monitor how a system learns from feedback over time, improving the efficiency and accessibility of ML). + +# 11. Safeguards + +Question: Does the paper describe safeguards that have been put in place for responsible release of data or models that have a high risk for misuse (e.g., pretrained language models, image generators, or scraped datasets)? + +Answer: [NA]. + +Justification: This paper poses no such risks. + +# Guidelines: + +- The answer NA means that the paper poses no such risks. +- Released models that have a high risk for misuse or dual-use should be released with necessary safeguards to allow for controlled use of the model, for example by requiring that users adhere to usage guidelines or restrictions to access the model or implementing safety filters. +- Datasets that have been scraped from the Internet could pose safety risks. The authors should describe how they avoided releasing unsafe images. +- We recognize that providing effective safeguards is challenging, and many papers do not require this, but we encourage authors to take this into account and make a best faith effort. + +# 12. Licenses for existing assets + +Question: Are the creators or original owners of assets (e.g., code, data, models), used in the paper, properly credited and are the license and terms of use explicitly mentioned and properly respected? + +Answer: [Yes]. + +Justification: We cite the datasets we use in Section 4.1 and introduce them in Section B.1. + +# Guidelines: + +- The answer NA means that the paper does not use existing assets. +- The authors should cite the original paper that produced the code package or dataset. +- The authors should state which version of the asset is used and, if possible, include a URL. +- The name of the license (e.g., CC-BY 4.0) should be included for each asset. +- For scraped data from a particular source (e.g., website), the copyright and terms of service of that source should be provided. +- If assets are released, the license, copyright information, and terms of use in the package should be provided. For popular datasets, paperswithcode.com/datasets has curated licenses for some datasets. Their licensing guide can help determine the license of a dataset. +- For existing datasets that are re-packaged, both the original license and the license of the derived asset (if it has changed) should be provided. +- If this information is not available online, the authors are encouraged to reach out to the asset's creators. + +# 13. New assets + +Question: Are new assets introduced in the paper well documented and is the documentation provided alongside the assets? + +Answer: [Yes]. + +Justification: We provide it in the supplemental material. + +Guidelines: + +- The answer NA means that the paper does not release new assets. + +- Researchers should communicate the details of the dataset/code/model as part of their submissions via structured templates. This includes details about training, license, limitations, etc. +- The paper should discuss whether and how consent was obtained from people whose asset is used. +- At submission time, remember to anonymize your assets (if applicable). You can either create an anonymized URL or include an anonymized zip file. + +# 14. Crowdsourcing and research with human subjects + +Question: For crowdsourcing experiments and research with human subjects, does the paper include the full text of instructions given to participants and screenshots, if applicable, as well as details about compensation (if any)? + +Answer: [NA]. + +Justification: The paper does not involve crowdsourcing nor research with human subjects. + +Guidelines: + +- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. +- Including this information in the supplemental material is fine, but if the main contribution of the paper involves human subjects, then as much detail as possible should be included in the main paper. +- According to the NeurIPS Code of Ethics, workers involved in data collection, curation, or other labor should be paid at least the minimum wage in the country of the data collector. + +# 15. Institutional review board (IRB) approvals or equivalent for research with human subjects + +Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals (or an equivalent approval/review based on the requirements of your country or institution) were obtained? + +Answer: [NA]. + +Justification: This paper does not involve crowdsourcing nor research with human subjects. + +Guidelines: + +- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. +- Depending on the country in which research is conducted, IRB approval (or equivalent) may be required for any human subjects research. If you obtained IRB approval, you should clearly state this in the paper. +- We recognize that the procedures for this may vary significantly between institutions and locations, and we expect authors to adhere to the NeurIPS Code of Ethics and the guidelines for their institution. +- For initial submissions, do not include any information that would break anonymity (if applicable), such as the institution conducting the review. + +# 16. Declaration of LLM usage + +Question: Does the paper describe the usage of LLMs if it is an important, original, or non-standard component of the core methods in this research? Note that if the LLM is used only for writing, editing, or formatting purposes and does not impact the core methodology, scientific rigorousness, or originality of the research, declaration is not required. + +Answer: [NA]. + +Justification: The core method development in this research does not involve LLMs as any important, original or non-standard components + +Guidelines: + +- The answer NA means that the core method development in this research does not involve LLMs as any important, original, or non-standard components. +- Please refer to our LLM policy (https://neurips.cc/Conferences/2025/LLM) for what should or should not be described. + +# Appendix of A Driving-Style-Adaptive Framework for Vehicle Trajectory Prediction + +# A Problem Background + +# A.1 Vehicle Trajectory Prediction + +The methods for vehicle trajectory prediction can be broadly classified into four categories [100]. These include: (i) Physics-based methods: These employ vehicle dynamics or kinematics models, such as singles trajectory methods, Monte Carlo, and Kalman filtering methods [101, 102, 103, 104, 105, 106]. These methods are known for their conciseness, efficiency, and computational effectiveness. (ii) Classic machine learning: Unlike physics-based methods that rely on several physics models, classic machine learning approaches apply data-driven models and consider additional factors for predicting trajectories. Examples include the Hidden Markov Model, Dynamic Bayesian Network, and K-Nearest Neighbors [107, 108, 109, 110, 111]. However, these traditional methods are typically only suitable for simple prediction scenarios and short-term prediction tasks. + +Recently, with the advancement of modern machine learning, vehicle trajectory prediction methods based on: (iii) Deep learning and (iv) Reinforcement learning become increasingly popular. These methods are capable of considering interaction-related factors, understanding high-dimensional complex policies, and adapting to more complex scenarios. Examples include Graph Convolutional Network, Graph Attention Network, Conditional Variational Auto Encoder, and reinforcement learning techniques such as Inverse Reinforcement Learning, Generative Adversarial Imitation Learning, and Deep IRL [112, 113, 83, 114, 115, 116, 117]. + +In summary, an increasing number of autonomous vehicle trials are utilizing deep learning or reinforcement learning methods to predict future vehicle trajectories. These approaches leverage expert demonstrations and extract interaction information from traffic participants and road conditions, considering a broader range of influencing factors. + +# A.2 Kolmogorov-Arnold Networks (KANs) + +Hilbert's 13th problem [118] famously posits that it is impossible to solve general seventh-degree equations using only functions of two variables. Subsequent research by Kolmogorov et al.[36] has shown that any function involving multiple variables can be represented using a finite number of three-variable functions. Further studies as detailed by Arnol'd et al. [37], establish that even functions of just two variables are sufficient, as described in Theorem 2.1 presents significant for machine learning: learning a high-dimensional function essentially reduces to learning a limited number of one-dimensional basis functions $\psi(x)$ in Equation (2). + +In reference [20], the authors introduce Kolmogorov-Arnold Networks (KANs), which are neural network applications based on Theorem 2.1. Unlike Multi-Layer Perceptrons (MLPs) that founded on the universal approximation theorem [119, 120, 121, 122], KANs feature learnable activation functions on what are traditionally referred to as "edges" (neurons) and they utilize fixed activation functions at what are typically called "nodes" (weights). Uniquely, each weight in KANs is replaced by a univariate function parametrized as a spline, meaning the network contains no linear weights whatsoever. + +A variety of KANs are used across different tasks as noted in [123], such as solving ordinary differential equations [44, 124], image classification and reconstruction [125, 126], and time series forecasting [127, 128, 129], among others. These applications demonstrate competitive or superior performance in efficiency and predictive power compared to traditional models. However, to the best of our knowledge, we are the first to utilize KANs in vehicle trajectory prediction. This involves approximating and predicting trajectories for different driving styles, expanding the range of basis functions, and providing explanations for specific matches between functions and trajectories. + +# B Dataset Information + +# B.1 Datasets + +We preprocess the three datasets using the official pip packages provided by their respective baselines. The characteristics of each dataset are summarized in Table 7. + +Table 7: Characteristics of the evaluation datasets. + +
DatasetsCollectSizeLibrarySelect
Argoverse [68]Miami and Pittsburgh23.69GArgoverse API / devkitMotion Forecasting
nuScenes [67]Boston and Singapore4.81Gnuscenes-devkitMotional-Full dataset
Waymo [69]Phoenix, AZ, Kirkland etc.83.50Gwaymo-open-datasetMotion1.1-scenario
+ +nuScenes This dataset [67] offers high-definition maps and trajectory data from 1,000 driving scenes in Boston and Singapore, areas noted for dense traffic and complex driving challenges. It comprises 245,414 trajectory instances, each a sequence of 2D coordinates over 8 seconds, sampled at $2\mathrm{Hz}$ . The nuScenes benchmark requires predicting a target agent's 6-second future trajectory from a 2-second historical trajectory. The comprehensive dataset features approximately 1.4 million camera images, 390,000 LIDAR sweeps, 1.4 million RADAR sweeps, and 1.4 million object bounding boxes across 40,000 keyframes. + +Argoverse This dataset [68] facilitates research in 3D tracking and motion forecasting for autonomous vehicles. Originating from select areas in Miami and Pittsburgh, it includes 113 scenes with 3D tracking annotations, featuring 324,557 significant vehicle trajectories derived from over 1,000 hours of driving. The forecasting component of Argoverse provides agent trajectories and high-definition maps, requiring the prediction of a target vehicle's future trajectory for the next 3 seconds, based on its past trajectory over two seconds, sampled at $10\mathrm{Hz}$ . The dataset encompasses 333K real-world driving sequences, primarily at intersections or within dense traffic, each focusing on one target vehicle for trajectory prediction. + +Waymo This dataset [69] publicly to aid the research community in investigating a wide range of interesting aspects of machine perception and autonomous driving technology. This Dataset we use is the Motion part, with object trajectories and corresponding 3D maps for 103,354 segments. Given agents' tracks for the past 1 second on a corresponding map, predict the joint future positions of 2 interacting agents for 8 seconds into the future. The ground truth future data for the interactive test set is hidden from challenge participants. The validation sets contain the ground truth future data for use in model development. In addition, the test and validation sets provide 2 interacting object tracks in the scene to be predicted. + +# B.2 Metrics + +We evaluate the predicted trajectory $Y^{t}$ against the ground truth trajectory $Y_{\mathrm{GT}}^{t}$ using standard error-based metrics. Our DSA framework adopts the commonly used Average Displacement Error (ADE) and Final Displacement Error (FDE), as defined in [23]: + +$$ +\mathrm {A D E} = \frac {1}{T} \sum_ {t = 1} ^ {T} \| Y ^ {t} - Y _ {\mathrm {G T}} ^ {t} \| _ {L ^ {2}}, \mathrm {F D E} = \| Y ^ {t} - Y _ {\mathrm {G T}} ^ {t} \| _ {L ^ {2}}. +$$ + +Here, the superscript $t$ denotes the current time step, and $T$ refers to the total number of time steps in the prediction horizon. The metrics we use include $\mathrm{ADE}_k$ , $\mathrm{FDE}_k$ , minADE, minFDE, and b-minFDE. The subscript $k$ indicates the Top- $k$ most likely predicted future trajectories. The "min" variants (minADE and minFDE) compute the $L^2$ distance between $Y_{\mathrm{GT}}^{t}$ and the closest predicted trajectory $Y^{t}$ across all generated samples, averaged over all agents. The b-minFDE metric extends minFDE by incorporating the Brier score [130], which evaluates the calibration of the predictive distribution. It is defined as the sum of the Brier score and minFDE. + +# C Proof of Bound + +As we discuss in Section 3.3.2, Kolmogorov's theorem [131] provides error bounds by evaluating the absolute value of the function and the overall variation in the function value. This is illustrated as follows: + +Theorem C.1 (Kolmogorov Theorem) For $f \in C[a, b]$ , there exists a polynomial $p_n$ such that approximation error is bounded by: + +$$ +\left\| f - p _ {n} \right\| _ {L ^ {\infty}} \lesssim \left(\frac {\log n}{n}\right) V (f, [ a, b ]), +$$ + +where $V(f, [a, b])$ denotes the total variation of $f$ over the interval. + +From this theorem, we conclude that when $n > e$ , increasing the degree $n$ of $p_n$ results in a smaller decrease in the theoretical upper bound on the approximation error. + +However, in practice, considering computational cost and time, the value of $n$ cannot be arbitrarily large. In this section, we provide three proofs corresponding to the three categories of drivers, demonstrating that under a given error limit $\delta$ , there exists a relationship between the minimum degree $n$ and the components of the trajectory to be approximated, as described in Section 2.1 (i.e., position $(x,y)$ , velocity, and acceleration). For clarity, we use $f$ to represent each continuous component with respect to $t$ . + +# C.1 Conservative Drives: Bernstein Polynomial + +In Section 3.2.1, we use the properties of Bernstein polynomials $(B_{n})$ for their uniform convergence to approximate the trajectories of conservative drivers, characterized by low speed and minimal motion changes. The minimum degree $n$ of the $B_{n}$ polynomial is obtained by: + +Theorem C.2 For all $\epsilon > 0$ , if $\partial [B_n(f)] = n$ , then for a given error limit $\delta$ with $0 < \epsilon \leqslant \delta \ll \infty$ , then + +$$ +n \geqslant m a x \left| f ^ {\prime \prime} (\xi) \right| / 8 \delta . +$$ + +Proof: First, we calculate the error between $f(x)$ and $B_{n}(f)$ + +$$ +\begin{array}{l} \left| f (x) - B _ {n} (f) \right| \\ = \left| \sum_ {k = 0} ^ {n} \left[ f (x) - f \left(\frac {k}{n}\right) \right] \binom {n} {k} x ^ {k} (1 - x) ^ {n - k} \right| \\ \leqslant \left| \sum_ {k = 0} ^ {n} \left[ - f ^ {\prime \prime \left(\frac {k}{n} - x\right)} - \frac {1}{2} f ^ {\prime \prime} \left(\frac {k}{n} - x\right) ^ {2} \right] \cdot P _ {B} (k) \right| (4) \\ = \frac {1}{2} f ^ {\prime \prime} (\xi) \sum_ {k = 0} ^ {n} \left(\frac {n}{k} - x\right) ^ {2} P _ {B} (k) \quad \xi \in \left(x, \frac {k}{n}\right). (5) \\ \end{array} +$$ + +In formula (4), $P_B(k) \triangleq C_n^k x^k (1 - x)^{n - k}$ . From Equation (5), we next proceed to prove + +1. $\sum_{k}\left(\frac{n}{k} -x\right)\cdot P_{B}(k) = 0$ +2. $\sum_{k}\left(\frac{n}{k} -x\right)^{2}\cdot P_{B}(k) = \frac{x}{n} (1 - x)$ + +For Equation 1, applying the Central Limit Theorem (CLT) as discussed in [132], we consider the total difference of weights, represented by $[(n / k) - x]$ . Specifically, the probability $p$ satisfies $p = x = k / n$ . In this case, as described in [133], we have $E(k) = nx$ . Thus, we can derive the expectation as follows: + +$$ +E \left(\frac {k}{n}\right) = \frac {E (k)}{n} = x. \tag {6} +$$ + +Therefore, Equation 1 simplifies to $E(k / n) - x = 0$ . + +For Equation 2, we employ a similar method; here $\left(\frac{n}{k} - x\right)^2$ represents the squared difference in weights between $\frac{k}{n}$ and $x$ , alternatively described as the deviation between observation and expectation. According to Equation (6), + +$$ +D \left(\frac {k}{n}\right) = \frac {n x}{n ^ {2}} \cdot (1 - x) = \frac {x (1 - x)}{n}. +$$ + +Equation 2 corresponds to the squared deviation $\left(\frac{n}{k} - x\right)^2$ based on weights $P_B(k)$ . Moreover + +$$ +\because E \left[ \left(\frac {k}{n} - x\right) ^ {2} \right] = D \left(\frac {k}{n}\right), +$$ + +$$ +\therefore (5) = \frac {1}{2} f ^ {\prime \prime} (\xi) \frac {x (1 - x)}{n} \leqslant \frac {M _ {2}}{8 n}, \text {w i t h} M _ {2} = \max _ {\xi \in \mathbb {D}} f ^ {\prime \prime} (\xi). +$$ + +Considering the error limit $\delta$ , we have: + +$$ +\frac {M _ {2}}{8 n} < \delta \Rightarrow n \geqslant \frac {M _ {2}}{8 \delta}. +$$ + +![](images/e01601641168f50916b0b10fbf10ec3c0fb29aef2832478e5871f75dc4f1ecba.jpg) + +Theorem C.3 If $f \in L^{p}[a,b]$ , $B_{n}^{\omega}(f) \in [a,b]$ and $\partial[B_n^\omega(f)] = n$ . For all $\epsilon > 0$ , $0 < \epsilon \leqslant \delta \ll \infty$ , $\delta$ is given error limitation, then: + +$$ +n \geqslant \frac {\operatorname* {m a x} [ | f ^ {\prime \prime} (\widetilde {\xi}) \cdot \omega | ] \cdot (b - a) ^ {2}}{8 \delta}, +$$ + +where $\omega$ is the weights of weighted $B_{n}$ polynomials $B_{n}^{\omega}(f)$ . + +Proof: Here the error is $L^{\infty}$ norm, the definition of $B_n^\omega(f)$ is + +$$ +B _ {n} ^ {\omega} (f) = \sum_ {k = 0} ^ {n} f \left(\frac {k}{n}\right) \omega \left(\frac {k}{n}\right) \binom {n} {k} x ^ {k} (1 - x) ^ {n - k} +$$ + +Let $u = (x - a) / (b - a)$ , then $u \in [0,1]$ . So $\widetilde{M}_2$ is similar to Theorem C.2, here + +$$ +\widetilde {M _ {2}} = \max _ {u \in [ 0, 1 ]} \left| g ^ {\prime \prime} (u) \right| = (b - a) ^ {2} \left| f ^ {\prime \prime} (\xi) \right|, +$$ + +where $g(u) = f\left(\frac{u - a}{b - a}\right)$ . The next following prove is similar to Theorem C.2. + +![](images/7ea35d3b305da34f7aed4ded64fd8ab7f57b915594af6aba0ae85a412de6dc1e.jpg) + +# C.2 Aggressive Drivers: Chebyshev Polynomial + +In Section 3.2.2, aggressive drivers' trajectories are characterized by non-smooth, high-speed movements during motion changes. We use the Chebyshev polynomials $T_{n}^{c}$ and their minimum-maximum error properties to approximate these trajectories. The minimum degree $n$ of $T_{n}^{c}$ polynomial is obtained as follows: + +Theorem C.4 For $f \in L^{p}[a,b]$ and a given error bound $\delta$ (where $0 < \epsilon \leqslant \delta \ll \infty$ ), the condition $\partial[T_{n}^{c}(f)] = n$ is satisfied: + +$$ +n \geqslant \frac {1}{\omega^ {- 1} \left(f , \frac {\delta}{1 2}\right)}, \tag {7} +$$ + +where $\omega^{-1}$ is the inverse of the modulus of continuity for the function $f$ . + +To provide the proof of Theorem C.4, we first introduce the definition of the modulus of continuity and a lemma related to this proof. + +Definition C.5 (Modulus of Continuity in $L^p$ Space) Let $f \in L^p[a, b]$ , $p \geqslant 1$ and $0 \leqslant m \leqslant b - a$ . The modulus of continuity $\omega_p(m, f)$ is defined as: + +$$ +\begin{array}{l} \omega_ {p} (m) = \omega_ {p} (m, f) \\ = \sup _ {0 \leqslant h \leqslant m} \left(\int_ {a} ^ {b - h} | f (x + h) - f (x) | ^ {p}\right) ^ {1 / p} \\ \end{array} +$$ + +which represents the continuity norm for $f$ over the interval $[a, b]$ . + +For $T_{n}^{c}$ polynomials belonging to the $C_{2\pi}$ space $^{10}$ , we use $E_{n}(f)$ to denote the deviation of the approximation of $f$ by a trigonometric polynomial $T_{n}$ of degree $n$ , as follows: + +$$ +E _ {n} (f) = \inf _ {\{T _ {n} \}} \| f - T _ {n} \|. +$$ + +This deviation satisfies: + +$^{10}C_{2\pi}$ Space Let $f \in \mathbb{R}$ with period $2\pi$ . Define + +$$ +\| f \| = \sup _ {- \pi \leqslant x \leqslant \pi} | f (x) |. +$$ + +We call the above set the $C_{2\pi}$ space + +Lemma C.6 (Jackson [134]) Let $f \in C_{2\pi}$ , then for all $n \in \mathbb{N}$ , the following inequality holds: + +$$ +E _ {n} (f) \leqslant 1 2 \cdot \omega \left(f, \frac {1}{n}\right). +$$ + +It is evident that $T_{n}^{c} \subseteq C_{2\pi}$ . Based on Lemma C.6, we present the proof of Theorem C.4. + +Proof: According to the definition of $\delta$ , we have $\|f - T_n\|_{L^{\infty}} < \delta$ . Lemma C.6 provides the modulus of continuity under the $L^{\infty}$ space, so we need to relate $\omega_p$ from Definition C.5 to $\omega$ in Lemma C.6, which relates the $L^p$ and $L^{\infty}$ norms: + +$$ +\left\| g \right\| _ {L ^ {p} ([ a, b - h ])} \leq (b - a - h) ^ {\frac {1}{p}} \cdot \left\| g \right\| _ {L ^ {\infty} ([ a, b - h ]).} +$$ + +For the function difference $g(x) = f(x + h) - f(x)$ : + +$$ +\left| f (x + h) - f (x) \right| \leq \omega (f, h), \text {f o r} \forall x \in [ a, b - h ]. +$$ + +Therefore, the $\omega_{p}(m,f)$ is related to $\omega (f,h)$ as follows: + +$$ +\begin{array}{l} \omega_ {p} (h, f) \leqslant \left\{\int_ {a} ^ {b - h} [ \omega (f, h) ] ^ {p} \right\} ^ {1 / p} \\ = \omega (f, h) \cdot (b - a - h) ^ {\frac {1}{p}}. \tag {8} \\ \end{array} +$$ + +When $h\to 0$ , Equation (8) can be approximated as: + +$$ +\omega_ {p} \leqslant \omega \cdot (b - a) ^ {\frac {1}{p}}. +$$ + +According to Lemma C.6 and satisfy the error limit $\delta$ , s.t. $E_{n}(f) \leqslant \delta$ , we have + +$$ +\omega \left(f, \frac {1}{n}\right) \leqslant \frac {\delta}{1 2}. +$$ + +To obtain the lower bound on $n$ , since $\omega(f, h)$ is a nondecreasing function with respect to $h$ , we take its inverse function $\omega^{-1}(f, y)$ as follows: + +$$ +\frac {1}{n} \leqslant \omega^ {- 1} \left(f, \frac {\delta}{1 2}\right). \tag {9} +$$ + +Finally, the lower bound for $n$ can be derived from inequality (9) as: + +$$ +n \geqslant \frac {1}{\omega^ {- 1} \left(f , \frac {\delta}{1 2}\right)}. +$$ + +Lemma C.4 provides a minimum bound related to the value of the modulus of continuity. Furthermore, the Lipschitz continuity [135] of vehicle trajectories $X_{i}$ and $Y_{i}$ offers a more compact bound for inequality (7): + +Corollary C.7 The bound in inequality (7) is satisfied as follows: + +$$ +n \geqslant \frac {1 2 L}{\delta}, \tag {10} +$$ + +where $L$ represents Lipschitz constant. + +Proof: The proof of Corollary C.7 consists of two parts: (i) establishing Lipschitz continuity of vehicle trajectories and (ii) deriving Equation (10). + +(i) Lipschitz Continuity of Vehicle Trajectories + +To demonstrate the Lipschitz continuity of vehicle trajectories, it suffices to show that their state information (Section 2.1), including $(x,y)$ position, velocity $v$ and acceleration $a$ , satisfies the Lipschitz condition ( $L$ -condition). Specifically, there exists a constant $L$ , s.t. for any $x', x'' \in [a,b]$ , the following holds: + +$$ +\left| f \left(x ^ {\prime}\right) - f \left(x ^ {\prime \prime}\right) \right| \leqslant L \left| x ^ {\prime} - x ^ {\prime \prime} \right|. \tag {11} +$$ + +According to the physical relationships among these states, if acceleration $a$ satisfies $L$ -condition, then by the boundedness theorem [136], the other states also satisfy it. Thus, we take $a$ as an example, and similar arguments apply to the other states, quod erat demonstrandum. + +According to [137, 138], the variation in vehicle acceleration is constrained by factors such as engine performance, vehicle weight, and braking system, which implies that the jerk $j(t)$ (the rate of change of acceleration over time) + +cannot be physically infinite. Therefore, there exists a constant $M_j$ , s.t. $|j(\tau)| \leqslant M_j$ . For any $t_1, t_2 \in [a,b]$ , the following holds: + +$$ +\begin{array}{l} \left| a \left(t _ {1}\right) - a \left(t _ {1}\right) \right| = \left| \int_ {t _ {2}} ^ {t _ {1}} j (\tau) \right| \leqslant \int_ {t _ {2}} ^ {t _ {1}} | j (\tau) | \\ \leqslant \int_ {t _ {2}} ^ {t _ {1}} M _ {j} d \tau = M _ {j} \left| t _ {1} - t _ {2} \right|. \\ \end{array} +$$ + +Therefore, Hence, the acceleration function $a(t)$ is Lipschitz continuous with the Lipschitz constant $L = M_{j}$ . + +(ii) Derivation of a more compact bound. + +To obtain a tighter bound, we use the $L$ -continuity property of $f$ . From the continues of modulus, we have: + +$$ +\omega (f, h) \leqslant L h \Rightarrow \omega^ {- 1} (f, y) \geqslant \frac {y}{L}. \tag {12} +$$ + +From Jackson's inequality in Lemma C.6, let $y = \delta /12$ in Equation (12). This ensures that the error remains below $\delta$ , with $\delta /12$ acting as a piecewise error threshold. Then Equation (12) becomes: + +$$ +\omega^ {- 1} (f, y) \sim \omega^ {- 1} (f, \frac {\delta}{1 2}) \leqslant \frac {\delta}{1 2 L}. \tag {13} +$$ + +Combining inequalities (7) and (13), we obtain the more compact bound (10). + +![](images/de27c84c54334bf2088892bf64beb79365adc2996c41331265da427d25bfd98f.jpg) + +# C.3 Normal Drivers: Legendre Polynomial + +In Section 3.2.3, we discuss that the speed and acceleration of normal drivers maintain an intermediate state between conservative and aggressive drivers. Their trajectories do not change as dramatically as aggressive drivers, nor are they so slow as to affect the flow of traffic. To approximate these trajectories, we use the Legendre polynomial $L_{n}$ . The minimum degree $n$ of $L_{n}$ is obtained by: + +Theorem C.8 For all $\epsilon > 0$ , if $\partial [L_n(f)] = n$ , for a given error limit $\delta$ with $0 < \epsilon \leqslant \delta \ll \infty$ , then + +$$ +n \geqslant \left(\frac {C _ {H}}{\delta}\right) ^ {1 / \alpha}, \tag {14} +$$ + +where $C_H$ is Hölder constant modulus of continuity and $\alpha$ represents Hölder exponent. + +From Theorem C.8 we apply Jackson's inequality (Lemma C.6) to establish a relationship between the approximation and the modulus of continuity. This inequality also applies to continuous functions defined on the interval $[-1, 1]$ interval. To achieve the bound in Equation (14), we further use the Hölder continuous property [139] of vehicle trajectory. Similar to $L$ -continues as defined in Equation (11), Hölder continuous (H-continuous) is defined as follows: + +Definition C.9 (Holder continuous) For a function $f$ defines on interval $I$ , if there exists a constant $C \in \mathbb{R}$ , s.t. for $\forall z', z'' \in I$ : + +$$ +\left| f \left(z ^ {\prime}\right) - f \left(z ^ {\prime \prime}\right) \right| \leqslant C _ {H} \left| z ^ {\prime} - z ^ {\prime \prime} \right| ^ {\alpha}, \alpha \in (0, 1 ], +$$ + +then $f$ is said to be Hölder continuous of order $\alpha$ . + +When $\alpha = 1$ , $H$ -continuous reduces to $L$ -continuous. Reference [140, 141, 142, 143] analyze the Hölder continuity or related smoothness of vehicle trajectories and their states (position, velocity, and acceleration), either directly or indirectly by examining the smoothness of physical constraints and changes. Therefore, the bound in Equation (14) can similarly be derived from inequality (12): + +$$ +\omega (f, h) \leqslant C _ {H} h ^ {\alpha} \Rightarrow C _ {H} \left(\frac {1}{n}\right) ^ {\alpha} \leqslant \delta . +$$ + +Thus, we obtain a bound as expressed by inequality (14). + +# D Effects of Framework Sensitive + +In Section 4.3, we evaluate our DSA framework with respect to module component dimensions: namely the type, combination, and degree of polynomials $p_n$ , as well as the driving style matching. In this section, we further analyze the sensitivity of our model with regard to: variations in the driving style categories themselves (Section D.1), external influences such as changing traffic densities (Section D.2) and varying road conditions (Section D.3). + +# D.1 Number of Driving Style + +In reference [48, 49, 50] we know that driving style are categories as three type and each character is illustrated in Section 3.2. Now we use automatic manner category numbers rather than the predefining for driving styles. Specifically, we evaluate multiple-category settings using K-means clustering, and report the corresponding metrics (log-normalized results except for Silhouette metric) on the Argoverse (from A-△WSS) and nuScenes (from N-△WSS) in Table 8, here for all metrics listed below, larger values indicate better clustering performance. Our three-category configuration yields the highest number of best scores across the metrics, supporting the + +Table 8: Clustering evaluation results on Argoverse (A) and nuScenes (N). A value of "1" indicates the best performance among all settings. + +
kA-ΔWSSSilhouetteΔCHIN-ΔWSSSilhouetteΔCHI
2-0.919--0-
3110.81010.7001
40.7430.98610.58610.387
50.0220.28300.2870.2680.324
6000.60100.3980
+ +validity and rationality of our chosen driving style taxonomy. The metrics [144] are defined as follows: + +- WSS (Within-Cluster Sum of Squares): Measures the improvement in intra-cluster compactness. A higher value suggests tighter grouping of samples within clusters after clustering. +- Silhouette: Reflects both the cohesion within a cluster and the separation between clusters. A higher silhouette score indicates that samples are well matched to their own cluster and poorly matched to neighboring clusters. +- CHI (Calinski-Harabasz Index): Captures the variation in inter-cluster separability and intra-cluster compactness. Higher values indicate better-defined and more distinct clusters. + +# D.2 Traffic Density + +Traffic density is closely related to vehicle speed and traffic flow, and significantly influences trajectory prediction due to varying interaction patterns among vehicles [145]. To clearly present the impact of traffic density, we divide the dataset into five levels based on the number of vehicles per unit area. We compare the performance of our DSA framework against the best-performing baseline with publicly available code: Context-Aware [81], as identified in Table 1. The comparison results are illustrated in Figure 5. + +![](images/53b2c8d2145d29c2524ee40bf3e12218a26b6615f49d2a27712e512b79d21557.jpg) +Figure 5: Comparison of trajectory prediction performance under different traffic densities with Context-Aware [81]. Our method is represented by solid lines, while Context-Aware is depicted using dashed lines on both sides. The middle subfigure shows the density distribution with circled numbers indicating the corresponding density levels. + +![](images/03d3e975a8cb82ef257105ca628647a13fa328428594fc4e5e14b7c990204aff.jpg) + +![](images/4752f6a90371f0d3cbcb954bdac9d94a3eb5637f1b28e672228856451b59dbc3.jpg) + +Our DSA framework consistently outperforms the baseline across over $75\%$ of the dataset. When averaged over the highest-density and most common case (density level 1 in the middle subfigure of Figure 5), our method achieves improvements of 1.97 in $\mathrm{ADE}_1$ and 3.26 in $\mathrm{FDE}_1$ , as well as 0.25 in $\mathrm{ADE}_5$ and 0.13 in $\mathrm{FDE}_5$ . + +The most notable gains appear in density level 2, where DSA reduces prediction errors by $49.48\%$ and $44.19\%$ for $k = 1$ , and by $33.78\%$ and $25.05\%$ for $k = 5$ , compared to Context-Aware. Although our framework shows slightly lower performance in levels 4 and 5 when generating 5 trajectories, these cases together account for only $6.28\%$ of the dataset. + +# D.3 Road Condition + +Road structure significantly influences the motion patterns of agents navigating through urban or highway environments and is thus essential for accurate trajectory prediction [146, 11, 112]. Does road complexity increase the frequency of driving style changes, thereby making prediction more difficult? To investigate this question, we evaluate the performance of our DSA framework under different road conditions, as shown in Table 9. + +Table 9: The b-minFDE* in different road condition on the nuScenes dataset. The best results are highlighted. + +
Method TypeStationaryStraightStraight rightStraight leftRight U-turnRight-turnLeft U-turnLeft-turnAll
MTR [147]2.152.584.854.268.134.825.174.852.86
DSA2.032.484.964.178.114.775.284.872.75
+ +* Brier-minFDE ( $b$ -minFDE): $b\min FDE = \min FDE + (1 - p)^2$ , where $p$ is the probability of probability of the best forecasting trajectory with minimum endpoint error (minFDE). + +Our DSA framework achieves the lowest overall error of 2.75, improving upon the baseline by $3.85\%$ . It outperforms the baseline in 6 out of 8 categories, including reductions in error for common scenarios such as Stationary $(5.6\%)$ and Straight $(3.9\%)$ , as well as complex maneuvers like Straight-Left $(2.1\%)$ , Right U-turn $(0.25\%)$ , and Right Turn $(1.0\%)$ . + +Although MTR performs slightly better in Straight-Right and Left U-turn (by 0.11 in both cases), DSA matches or surpasses baseline performance in the most frequent and safety-critical trajectory types. These results demonstrate the robustness and adaptability of our framework across diverse road semantics, particularly in non-linear or discontinuous motion patterns. + +# E Detailed Description of the Algorithm + +In Sections 3.3.1 and 3.3.2, we introduced the mechanisms for polynomial combination and degree adaptation. In this section, we provide a detailed description of the corresponding algorithms. + +# E.1 Polynomial Combination + +To match each trajectory under various driving styles to a suitable polynomial combination, as analyzed in Section 3.2, we employ a Mixture of Experts model based on Top- $K$ Gating (MoE-TopK) [63]. In this method, tunable Gaussian noise is added to the gating logits, and only the top $K$ values are retained for expert selection. + +Let us denote by $G(x)$ and $E_{j}(x)$ the output of the gating network and the output of the $j$ -th expert network for a given trajectory $X_{i}$ , for clearly we omit subscript $i$ . The output $z^{\mathrm{com}}$ of the MoE module can be written as follows: + +$$ +z ^ {\text {c o m}} = \sum_ {j = 1} ^ {3} G (X) _ {j} E _ {j} (x). \tag {15} +$$ + +As shown in [21, 61, 62], kernel density estimation and latent variable analysis reveal that a driver's behavior evolves continuously across different situations. This implies that driving behavior can be viewed as a probabilistic mixture of weighted driving styles. Given that drivers may exhibit behaviors characteristic of multiple styles in dynamic scenes—such as when another agent suddenly appears—we adopt the Noisy Top- $K$ Gating network [63] to capture this mixture behavior. This network activates only the top- $k$ best-matching experts, enhancing responsiveness and specificity. Accordingly, this modification adjusts $G_{*}(\cdot)$ in Equation (15) to $\widetilde{G(x)}$ , detailed as follows: + +$$ +\widetilde {G _ {j} (x)} = \operatorname {S o f t m a x} \left\{\operatorname {K e e p T o p K} [ H (X), k ] \right\}, \text {w i t h} +$$ + +$$ +H (X) = \left(X \cdot W _ {G _ {j}}\right) + \operatorname {S N} () \cdot \operatorname {S p} \left[ \left(X \cdot W _ {\text {n o i s e} _ {j}}\right) \right]. +$$ + +Here, "SN" and "Sp" denote Standard Normal [64] and Softplus [65] functions, respectively. The symbol $W_{*}$ denotes the weight matrix corresponding to each subscript. The loss function is defined as follows: + +$$ +L _ {\mathrm {M o E - K}} = w _ {\text {l o a d}} \cdot \mathrm {C V} (\text {l o a d s}) ^ {2}, \tag {16} +$$ + +where "load" refers to the importance values assigned to each driving style, with $w_{\mathrm{load}}$ representing the corresponding weight. "CV" stands for the Coefficient of Variation, which assesses the variability of these values. Equation (16) is a part of Loss in Section 4.1. This structure of the MoE model effectively recognizes the diversity of trajectories, allowing each expert to specialize in different features of driving styles. + +# E.2 Degree Adaptation + +From Theorem 3.7, we understand that the accuracy of polynomial $p_n$ approximation is directly influenced by the degree $n$ of the $p_n$ . However, adapting the degree of $p_n$ poses a complex, non-convex, and combinatorial optimization challenge, as the relationship between prediction error and polynomial degree is not straightforward. This complexity often leads to the presence of multiple local optima. + +To address this issue, we utilize the versatile Bayesian Optimization (BO) tool SMAC3 [66] for its robustness and flexibility, making it particularly suitable for optimizing low-dimensional and continuous functions (type: SMAC4BB), such as those found in vehicle trajectory prediction. + +We treat the adaptive of polynomial degree as a hyperparameter optimization problem, using SMAC3 for BO, which leveraging Gaussian Processes with the Matérn kernel and the Expected Improvement acquisition function, iteratively searches the candidate degree set to minimize the loss function. Specifically, the degree $n$ is treated as a hyperparameter optimization problem, aimed at minimizing the loss on validation data $D_{\mathrm{val}}$ of our model trained on training data $D_{\mathrm{train}}$ . This process can be formulated as follows: + +$$ +n_{\text{SMAC}}\in \operatorname *{arg min}_{n\in \mathbb{Z}^{+}}c \left(n\right) = \operatorname *{arg min}_{n\in \mathbb{Z}^{+}}L\left(\mathcal{D}_{\text{train}},\mathcal{D}_{\text{val}};n\right), +$$ + +The hyperparameter optimization process targets the degree $n_{\mathrm{SMAC}}$ , which is defined as the optimal degree that achieves the least error for the corresponding basis function $p_n$ . Here $L$ denotes the loss function. + +# F Limitation and Discussion + +**Limitation** We summarize existing open vehicle trajectory prediction datasets in Table 10, and observe that the maximum available trajectory duration is typically less than 10 seconds. Despite this limited time span, our framework based on three driving styles adapts well to such settings. We evaluate its performance on both short-term (3 seconds, Table 1) and long-term (8 seconds, Table 2) prediction tasks, achieving consistently strong or state-of-the-art results across all durations. + +Table 10: Existing vehicle trajectory datasets. "His" and "Pre" represent the historical and predicted trajectory durations, respectively, while "Total" denotes the overall duration for each vehicle. + +
DatasetsPub.Collect LocationsHisPreTotal
KITTI [148]2012 CVPRKarlsruhe246
Apolloscapes [149]2018 CVPRBeijing, Shanghai and SHenZhen336
nuScenes [67]2020 CVPRBoston and Singapore268
Argoverse [68]2019 CVPRMiami and Pittsburgh235
INTERACTION [150]2019 arXivChina, Germany and Bulgaria134
InD [151]2020 TIVGerman3.24.88
RounD [152]2020 ITSCAachen246
HighD [153]2018 ITSCGerman2.82.85.6
Waymo [69]2020 CVPRUSA189
+ +However, in longer prediction horizons, the complexity of driving behavior increases, suggesting that three driving style categories may be insufficient to cover all possible scenarios. Moreover, trajectory patterns are often influenced by external factors, which can be categorized as either soft or strong conditions. + +Soft conditions, such as weather, affect driver perception and reaction. For example, on sunny days, improved visibility may enhance drivers' responsiveness, leading to smoother and more stable trajectories. In contrast, adverse weather conditions such as fog, heavy rain, or snow can result in more abrupt or irregular driving behavior. + +Similarly, strong conditions such as traffic signals or regulatory constraints also significantly influence vehicle trajectories. Unfortunately, most existing datasets lack labels for these contextual factors. We believe that incorporating such labels could further enhance prediction accuracy in future research. + +Discussion For longer vehicle trajectories, we can improve our DSA framework from both practical and theoretical perspectives. + +1. Incorporating more driving styles. Our current DSA framework utilizes three representative styles: Conservative, Aggressive and Normal (CAN), which reflect two behavioral extremes and an intermediate pattern. However, as the temporal length of each driver's trajectory increases, driving behaviors may exhibit greater variability. To capture these nuances, the model can be extended by defining or integrating additional driving styles. This would allow for a more fine-grained characterization of driver behavior and potentially lead to improved trajectory prediction accuracy. +2. Expanding the set of basis functions. As driving styles become more diverse and trajectory conditions more complex, a broader set of basis functions is required to effectively approximate and predict vehicle trajectories. Instead of relying on a single polynomial type, we can extend to a set of basis functions of the same class, such as orthogonal trigonometric polynomials. For example, to minimize the $L^2$ norm in modeling trajectories of normal drivers, or other intermediate states between conservative and aggressive behavior. It is beneficial to use a richer set of orthogonal polynomials that better match the dynamics of these nuanced driving styles. + +# G Future Work + +In this paper, we focus on the characteristics of the individuals who generate the data (i.e., trajectories) and leverage the mathematical properties of basis functions to approximate these trajectories. This concept can be generalized and extended to other domains, such as: + +- Other traffic participants. In addition to drivers, other agents in the traffic scene such as pedestrians and cyclists, also exhibit distinct behavioral characteristics. By modeling these characteristics, we can select appropriate basis functions tailored to each agent type, thereby improving the accuracy of their trajectory prediction. +- Multivariety Time Series Forecasting. Our framework can be extended to long-term forecasting tasks in domains such as weather prediction, energy consumption and electrocardiography. For example, one could model temperature and precipitation trends across different climate zones, analyze electricity usage patterns based on consumer behavior, or study heart rate dynamics as a function of individual health conditions. + +Additionally, by leveraging the core theoretical foundation (Theorem 2.2), we aim to construct models grounded in the physical characteristics or behavioral attributes of the data sources, thereby fully exploiting the inherent structure of the data itself. + +# H Visualization + +Due to space constraints, the number of visualized prediction results in the main text is limited. Here, we provide additional visualizations of predicted trajectories for various scenes, with generated trajectories $k = 1,5,10$ . Each value of $k$ is presented for both simple (e.g., straight roads) and complex scenes (e.g., turns conditions), showcasing different types of driving behavior. + +To enhance the clarity of the visualization results, we present them on a dedicated page and reduce the background opacity to improve visual contrast. Specific outcomes are accompanied by detailed explanations provided in the corresponding figure captions. In summary, considering various scenario combinations and adjusting the number of generated trajectories lead to more diverse, accurate, and comprehensive vehicle trajectory predictions. Increasing the number of predicted trajectories improves prediction diversity and realism, while analyzing different scenarios helps adapt to the diversity and complexity of real-world traffic environments. These improvements contribute to making the model both more mathematically grounded and more adaptive. + +- $k = 1$ + +![](images/681169a8da637c390cef1b6205aad64e2ec1bf0bf9ddcd98445a8ffde990bc6f.jpg) + +![](images/926ac6fd687c481898daa100ecd869f85e47308c3a32ac66a1be562aad1a564a.jpg) + +![](images/688af1817b6695cb4262797617e818609be4627c7f74fcf5092c310199993688.jpg) + +![](images/9992eb4d0614bd49c3346dd81fdfba7f3ea8b7fe74413d71411a580000bb1f6b.jpg) + +![](images/88c54e3d2adcbd20fa6abf7a9aef95d823022d518ccfe672dee2beb029894bf9.jpg) + +![](images/73802a43be28064ccef53c8b25187d761f1a62bd323ae903158529d0df177b82.jpg) + +Ground Truth + +Predictions Lanes Line + +![](images/941112672d16d3edd1e55ddcfdfb9e3f2a64f79ad806bfee32965c49a20d4503.jpg) + +Target Vehicle + +![](images/7394fb1708dff98447fdcf455667395689811e652f14cfaa46a6d6db7a9ae06e.jpg) + +Other Vehicle + +- $k = 1$ + +![](images/30f5b19b69c00d8ba01721a51a9563615d5bfe5b4cbf137ec780429cb3aba2b0.jpg) + +![](images/d6e9d7505cc06ac8664e689b6a6ebc831696b88d8f6de85dba8b86ef835d437e.jpg) + +![](images/f75223b15bff6af4fc431893eff4f24c0109181eb01ca92d560b21c9d4847075.jpg) + +![](images/3bf9418488ad1761ea22e17a66fd00562563681750313200fba29e70b1ed268d.jpg) + +![](images/95953bdfe26b9b0cee6380317c71c5500e6d545cc7629026751338b3a97e0fc3.jpg) + +![](images/e0933f9ecdca96717e4ca75541888c5ccb0813b7fb1b9e1670049ef3629f338d.jpg) + +![](images/b028851fb260a18baa489d2c073a0bcc3ccb70069fd4c29067752465f7d9a4ed.jpg) + +![](images/33cce806482ccd77ad73a78d71d0440a210dbca7f225ac998f28359496b832a5.jpg) + +$k = 1$ + +![](images/e7f278a5b4ab878a2c98b6cc7c15d94dd1046a0861c67cb39041481db2a5c3dd.jpg) + +![](images/ae83827ebf47222ec81c3119b9daf40ea6454e46bf4f9d4983ee7805088c8d94.jpg) + +![](images/81d811ff10279158e4bb6cf49c7347b271196ab1a5e16644376c17ca081bef47.jpg) + +![](images/ed50f83d1591ec4e4c61b397837952e5d6186132d1d9828aacc9542d72d6ca06.jpg) + +![](images/702fc5d3b781f1a0f2e79ef63c3caddb6ee64f9df6fb108dadf6a977e1e60455.jpg) + +![](images/4be950a59633ccb77e23b937ec3710597b127d379c270f58e6c89ce2287cc808.jpg) + +Ground Truth + +Predictions + +![](images/0efa2421ae2b20933ddcf391915ce1e23790432dd783953360673bc5de37d059.jpg) + +Lanes Line + +Other Vehicle + +$k = 1$ + +![](images/8c4b998ffd31b550595081114f22afc10f68c1f2cb30aa564d7aa32e66209c55.jpg) + +![](images/770ef5777fccbf509db48830ded49f9a4fc04713c6b4351ffd4042c247d0c5bc.jpg) + +![](images/c4dac96ee37eb21394cd406ffee49fbf7784c23bbce9c95480780847f6afe344.jpg) + +![](images/ea8ab4d46796cbbb9e6febb3d2c3b480f2e834b29bf2b5c3c77d88a8a4a3697f.jpg) + +![](images/e5af5b6b0cfd37dab620bfbd0daf6f861e4b4a2f1826311ff3d9bef4a6dd1138.jpg) + +![](images/c57a9cbdae728d867aa8499958d9dd236f13ab21312ea5fdfca0afff7088362b.jpg) + +Ground Truth + +Predictions + +![](images/8e23800c6d9c2a80d015274eabd1a47354c3523b8ec63eabee94b999bbcbc1a7.jpg) + +Lanes Line + +![](images/dc73aa9cc1570608d853be15192dd7604d97ec432df01e7584256191ae74d9af.jpg) + +Target Vehicle + +$k = 5$ + +![](images/6737762cb7c1d2f431142dede2cbb3b964e782268d3c151c009c86e9513f102d.jpg) + +![](images/f02b187c2f89ca523af803b1e583a6145eb5fc09651e11acb8e1acd651277704.jpg) + +![](images/0271aca315e8c7879361e5316802691c8adfb6259186f64de69c09c9562b4d4c.jpg) + +![](images/7e7abe6fdd0e3bbc7fe35bfedff44ebe215f8eba7d26b41c2002402ac745439b.jpg) + +![](images/403209465e3b37d7e4b8bccf30c14adb1d09b6aaf937724d4eb9d2c9baf5e61d.jpg) + +![](images/09fde9bfb9bee7428756a4d7aaad0fc89ff423102fa588e57dfa7ee19dadf876.jpg) + +Ground Truth + +Predictions Lanes Line + +![](images/0c6c71ce98447ba28d5dd7c8c96b83314ab6efc5e4db7b75fee70c11cddfe337.jpg) + +Target Vehicle + +![](images/ef6eebbe11afecded2946408a7fbdde95ea02498c9478f02922a51fac335b187.jpg) + +Other Vehicle + +$k = 5$ + +![](images/97524a3c68ac2f809812438da1e4d2b04a43e01d5772389aef6cc59fedf65cad.jpg) + +![](images/fab84b5a745349ecccacd7cbe56a27e3b93f9905bfaf67d6990f4cf3ba085e37.jpg) + +![](images/27ea845770760e961814a85047315657338515054fc9c4124fd837cfcd7bb377.jpg) + +![](images/0bd511aa459043dce28516dcfd3b1269ecee1577f57565e89892ce7ecf628dd9.jpg) + +![](images/aa2ddbfb3bf40521dbedc6e22a316f573f261420f885a7ae4840988abdb97549.jpg) + +![](images/94aebf60639feed265b3ef12283a4edb294c8614f92ba725b1d34aceb55b8cbd.jpg) + +Ground Truth + +Predictions Lanes Line + +![](images/2471c4f16d8a179467a5be37aae26efca01e056c23565fa0259a2b65d2d5c0ce.jpg) + +Target Vehicle + +![](images/8334e5d90df7b18729271fd49bd5facb5c914d24c7d40ccb0f1c5df765494f67.jpg) + +Other Vehicle + +$k = 5$ + +![](images/05c6d29a23a3be5f9880d21fc3d7743653f39e0694b5a95ac070f898f1779960.jpg) + +![](images/7516dac9a5b4dab26b634823d27caaccb08ed5c16f00fba5925001f115d34fad.jpg) + +![](images/75e41b4472cdda38ca59d8ed5f94fbea839c73c358815b6250798df7526ec1cb.jpg) + +![](images/b4617d8a158196f326849acabafac9b41d66279be026470f193c1e6225d75d98.jpg) + +![](images/0730e2aa077cdc1eb48c8faae534feeace2eeea56bd9a2a811f7e4ed59f6049b.jpg) + +![](images/b0cae23948fe5ef6be0293b55aeb5952bcc7862142fddb11208cea14fa29a840.jpg) + +Ground Truth + +Predictions Lanes Line + +![](images/aa238e00d3ab37b07cf3c6b8f019e77e3588494ebe27b9fb917fa4da395215bc.jpg) + +Target Vehicle + +![](images/05e1608c9cf804b1a8d650b3306c1b89c7681ce61ebf60011af908109221867c.jpg) + +Other Vehicle + +$k = 5$ + +![](images/ef5b9c9ef3ea462f1cd6570441920f4224b371a0a06723b65276d4751d94c96e.jpg) + +![](images/95ecfbb0ec20691b1c5f5929081bb87bc14d73fd477873e705fb3b07e3e526ab.jpg) + +![](images/2970e42d93e8e18e894e8da7da205993b755125fee65597aaa0cef07539e4bf7.jpg) + +![](images/2c76746ec7b2c7e3539791b1b6585dc5360bd8b0c257daec2154bca5dee78ab7.jpg) + +![](images/69792ddbd1fa064948af1411ac0eb71021f036597e0e0f23b3ff4d25884c999b.jpg) + +![](images/a9eaf4ff128d7b8e14dc93714529ae6aaf57110b9ebff601ae7ab46221cb5c75.jpg) + +![](images/dd4622713eddf0af2d211f533fc1249a98ec987a6b42f906d7625eae305d07c3.jpg) + +![](images/3eaaeadec8708de2c56b40dbecd143a9c3a41959894aeba0b586eeda602547b1.jpg) + +$k = 10$ + +![](images/0429ece0a0c301a12979061759319576a9c60e8b03452acf7f36c4a00af6390f.jpg) + +![](images/019ca69e6fbf6098459e3f7858d500e8e9c067cf9736d5b27bd9ee1df91d2d5f.jpg) + +![](images/da3fb6ff84d058643adab9588c695432de9f461dfe67fdfd6f59e11c8140009e.jpg) + +![](images/bb1e30d6d0c74dd30c890a1396045e7837e70cd3c70407e203212ced21e9e291.jpg) + +![](images/8cda06d9f8424845a5f0ac9f087bc5cbc9312a1e1c260f6f741da895d7398121.jpg) + +![](images/b70638378ebb055d738884cd50096cac7f3db9c58cae8d7111bdb0846a354597.jpg) + +Ground Truth + +Predictions Lanes Line + +![](images/d3bf8aa0eaf82a5e54b9e3f5442799f89240036ecca6bd0b5d7df743f2ab2372.jpg) + +Target Vehicle + +![](images/e350b59c6f512668874fae3f923f4d36aff0606068b04bb824e16ebd48f07f05.jpg) + +Other Vehicle + +$k = 10$ + +![](images/2cc056809247f1a404b7fcd3c2fbb0789207440145ad673546ad821869af3d6e.jpg) + +![](images/25877f84d13b7e353c5ccbe39da5684d2ebe51072b3ce8ac9fa3d9c106abd160.jpg) + +![](images/c97d4baba2c9984111690973df5bffa6ec7c9933b4c6b4b74f812ebbc99527ff.jpg) + +![](images/fde2d0661af5379579de72a2fc107b206faf71ebf3b20d857fc04193954a4bc1.jpg) + +![](images/163b5a529e70cd2471181e86d4a2be858738bd3a551a77424e0d14f5d0a03597.jpg) + +![](images/01b05ba01a8ce1af8de47eaca457bc45ccb7c2ee55b99f450c04eb9a3361bd37.jpg) + +Ground Truth + +Predictions + +![](images/c809eaae25f642173755d1ca20b5bcd371f85c72c1cf448dbd1e000e413f2885.jpg) + +Lanes Line + +![](images/e0585c9125bd0265c4caedd36c2b1e6b6e5217b83141b4221181950054434cbe.jpg) + +Target Vehicle + +![](images/73142e28debef263a0931ebf5e83630e7251f565606ffa3144187e2245d7acd6.jpg) + +Other Vehicle + +$k = 10$ + +![](images/904edbe4f7f882f1a8932a293dc580451ce9380f33a7382e9e990e19b9a20953.jpg) + +![](images/ea0349b78b58533e5f44d73823a39cecb56e781d01f096ca4b96d29ae45f4701.jpg) + +![](images/310d72173ddcf3526ee533336c20aabec2df7d763405b9ceb9cd96cfca7a602f.jpg) + +![](images/30a648d842a5b4b101e2879e084b7a1b9b1a05bfd8ec3acfb02e779de82b1c0c.jpg) + +![](images/2563d97befc73d0cbcb1f8c26a11480c9720a3e64d7a848a86f8259342adfcb5.jpg) + +![](images/2e38ab2deda1409ccb79880ef1e1f8c3709881b065eeaa58a00b5ddf6adc6c3e.jpg) + +![](images/71ab7c35a03a727693e37c1730d132e8061e5af8f6843e9d5d67a0b11b72632f.jpg) + +![](images/989a08a4bf39b41c2b14b54199e750551f1bb626e558090c969dfbf879af9849.jpg) + +$k = 10$ + +![](images/1b955563171f76bdd52cf9679c607ebb61a46183d44d84fa07a0b24099820041.jpg) + +![](images/1bb9bbc8c4fcb5164e326ea74ed77f2fe8f7be996a548d2ec0acb0dabd5a9874.jpg) + +![](images/d9d1e15aeaf030f85d1d05cff7533167dde4bdc4c947de34df664e241dcccb7d.jpg) + +![](images/a6dec1eb383fae4f5e264879f530a96b31506f071331b0aee387c9233948f30d.jpg) + +![](images/4965099c4cb8c66a9cca5180800af8fe75a2d40a4abb8a8cad1310c29660cf50.jpg) + +![](images/24f83eead5f73129a214317487bf4d95e07d3b706d1e9ab2707a761a3596b5a7.jpg) + +![](images/93838d7eb9e86e94bcee5d0576ba34f3e9254cd47a8de06acc8986bdaba3a091.jpg) + +![](images/f3c5a2e44e90d836172320e04386b116c43fb7956bfe3ca4d38a9a24880d3103.jpg) \ No newline at end of file diff --git a/NeurIPS/2025/A Driving-Style-Adaptive Framework for Vehicle Trajectory Prediction/images.zip b/NeurIPS/2025/A Driving-Style-Adaptive Framework for Vehicle Trajectory Prediction/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..e9ce37aec8ce2697b9cf12c896c097a1a2ae7d41 --- /dev/null +++ b/NeurIPS/2025/A Driving-Style-Adaptive Framework for Vehicle Trajectory Prediction/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4bea83e0425e6756346d2b4a872f47d82943d88a0f7d119ad7e970fe3d56b5a1 +size 1253851 diff --git a/NeurIPS/2025/A Driving-Style-Adaptive Framework for Vehicle Trajectory Prediction/layout.json b/NeurIPS/2025/A Driving-Style-Adaptive Framework for Vehicle Trajectory Prediction/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b482296af8b31132d34b1e1ccb588671d7d94ea4 --- /dev/null +++ b/NeurIPS/2025/A Driving-Style-Adaptive Framework for Vehicle Trajectory Prediction/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0e39220e55c1078c4e86ec4a79e22666179f35b2f4b61d6f0dd2c60da30fe91c +size 1708148 diff --git a/NeurIPS/2025/A Dynamic Learning Strategy for Dempster-Shafer Theory with Applications in Classification and Enhancement/e6ed5e37-dcd3-42c0-b32c-b67543293a99_content_list.json b/NeurIPS/2025/A Dynamic Learning Strategy for Dempster-Shafer Theory with Applications in Classification and Enhancement/e6ed5e37-dcd3-42c0-b32c-b67543293a99_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..119981e07714f25d1215bd30a977d3da25d9c9a8 --- /dev/null +++ b/NeurIPS/2025/A Dynamic Learning Strategy for Dempster-Shafer Theory with Applications in Classification and Enhancement/e6ed5e37-dcd3-42c0-b32c-b67543293a99_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2e4ff3e0bd188f51216f40ba7e0e1a115126ca1550a2a0c47e6a10e587f24fb8 +size 172579 diff --git a/NeurIPS/2025/A Dynamic Learning Strategy for Dempster-Shafer Theory with Applications in Classification and Enhancement/e6ed5e37-dcd3-42c0-b32c-b67543293a99_model.json b/NeurIPS/2025/A Dynamic Learning Strategy for Dempster-Shafer Theory with Applications in Classification and Enhancement/e6ed5e37-dcd3-42c0-b32c-b67543293a99_model.json new file mode 100644 index 0000000000000000000000000000000000000000..78afe3c844d65dd6a9fc5a9bb82e545a025cf854 --- /dev/null +++ b/NeurIPS/2025/A Dynamic Learning Strategy for Dempster-Shafer Theory with Applications in Classification and Enhancement/e6ed5e37-dcd3-42c0-b32c-b67543293a99_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b6d8a95ed79dc11d89d382149b959c76f33a82b42ac2057c93872c1332c36957 +size 223289 diff --git a/NeurIPS/2025/A Dynamic Learning Strategy for Dempster-Shafer Theory with Applications in Classification and Enhancement/e6ed5e37-dcd3-42c0-b32c-b67543293a99_origin.pdf b/NeurIPS/2025/A Dynamic Learning Strategy for Dempster-Shafer Theory with Applications in Classification and Enhancement/e6ed5e37-dcd3-42c0-b32c-b67543293a99_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..f12014622f59361d5aed7fc5bb928d763ecfe72c --- /dev/null +++ b/NeurIPS/2025/A Dynamic Learning Strategy for Dempster-Shafer Theory with Applications in Classification and Enhancement/e6ed5e37-dcd3-42c0-b32c-b67543293a99_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3fbd1e82511ca9723cd15fff1d8fb436362f1c186a91af0b407d175dee053918 +size 2529553 diff --git a/NeurIPS/2025/A Dynamic Learning Strategy for Dempster-Shafer Theory with Applications in Classification and Enhancement/full.md b/NeurIPS/2025/A Dynamic Learning Strategy for Dempster-Shafer Theory with Applications in Classification and Enhancement/full.md new file mode 100644 index 0000000000000000000000000000000000000000..074bcb3d415f32fc8f37594be607da24c352cb9f --- /dev/null +++ b/NeurIPS/2025/A Dynamic Learning Strategy for Dempster-Shafer Theory with Applications in Classification and Enhancement/full.md @@ -0,0 +1,905 @@ +# A Dynamic Learning Strategy for Dempster-Shafer Theory with Applications in Classification and Enhancement + +Linlin Fan $^{1}$ , Xingyu Liu $^{1}$ , Mingliang Zhou $^{1,*}$ , Xuekai Wei $^{1}$ , Weizhi Xian $^{2}$ , Jielu Yan $^{1}$ , Weijia Jia $^{3}$ + +$^{1}$ School of Computer Science, Chongqing University, + +$^{2}$ Chongqing Research Institute of HIT, Harbin Institute of Technology, + +$^{3}$ BNU-UIC Institute of Artificial Intelligence and Future Networks, Beijing Normal University + +linlinfan@stu.cqu.edu.cn, xingyuliu@stu.cqu.edu.cn, mingliangzhou@cqu.edu.cn, xuekaiwei2-c@my.cityu.edu.hk, wasxxwz@163.com, yanjielu@cqu.edu.cn, + +jiawj@bnu.edu.cn + +# Abstract + +Effective modelling of uncertain information is crucial for quantifying uncertainty. Dempster-Shafer evidence (DSE) theory is a widely recognized approach for handling uncertain information. However, current methods often neglect the inherent a priori information within data during modelling, and imbalanced data lead to insufficient attention to key information in the model. To address these limitations, this paper presents a dynamic learning strategy based on nonuniform splitting mechanism and Hilbert space mapping. First, the framework uses a nonuniform splitting mechanism to dynamically adjust the weights of data subsets and combines the diffusion factor to effectively incorporate the data a priori information, thereby flexibly addressing uncertainty and conflict. Second, the conflict in the information fusion process is reduced by Hilbert space mapping. Experimental results on multiple tasks show that the proposed method significantly outperforms state-of-the-art methods and effectively improves the performance of classification and low-light image enhancement (LLIE) tasks. The code is available at https://anonymous.4open.science/r/Third-ED16. + +# 1 Introduction + +Uncertain information is inevitably encountered in the process of modelling data-based complex systems via deep learning. Effective modelling and processing of uncertain information is an important technique that plays a key role in the decision-making process and improves the ultimate decision-making level. Currently, many methods have been proposed to solve this problem, such as evidence-theoretic methods [57] and fuzzy logic methods [10]. and deep learning-based methods [44]. These techniques have been applied in several fields, such as graph clustering [49], classification [56], and target detection [32]. However, how to effectively handle uncertain information from different sources and combine them effectively while avoiding mutual conflicts is still a challenging problem. + +In machine learning, classification is the task of making predictions about new sample categories by learning the features of samples from known categories [26]. In the classification task, the reliability of the classifier plays an important role. However, owing to the uncertainty of the data itself and the possible conflict or redundancy of information from different sources, effectively and rationally addressing information from different sources to improve classification accuracy has become an urgent problem. Currently, many classical machine learning methods have been proposed. Denoexu + +et al. [15] viewed the neighbors of unclassified samples as evidence for the hypothesis, with support as a function of distance between the vectors, and used Dempster's combination (DC) rule to combine the evidence for the classification task. Freund et al. [19] proposed the classification method of an alternating decision tree, which solved the problem that the original classifier is complicated and difficult to understand. Chang et al. [5] proposed a support vector machine (SVM) and an SVM with radial basis functions to implement classification problems. Xu et al. [60] utilized a normality test and normality transformation to address nonnormal data for classification tasks. Hu et al. [23] investigated Bayesian and mutual information classifiers and applied them in a classification task. Xu et al. [61] extended the classical probabilistic calibration approach to an evidence-theoretic framework when dealing with different sources of information to address the problem that a single probability measure may not adequately express uncertainty when modelling the calibration step. Xiao et al. [58] proposed weighted belief-jensen-shannon divergence for decision-making improvement on the basis of Dempster-Shafer evidence (DSE) theory. Although the above methods can address information from different sources, the importance of the information and the measurement of discrepancies are still not comprehensively considered. These shortcomings are reflected mainly in the following aspects: + +- In DSE theory, the current method does not consider the a priori information brought by the data itself, nor can it dynamically adjust the tendency of splitting on the basis of the importance between different subsets in the process of splitting, which means that the uncertainty information cannot be well handled. +- When fusing multiple basic belief assignments (BBAs), different evidence sources conflict due to uncertainty or data distribution differences. Direct use of DC rules may lead to conflict amplification or even produce unreasonable results. Moreover, traditional methods are usually based on Euclidean space or simple statistical metrics to calculate the differences between evidence, which makes it difficult to capture the nonlinear characteristics of complex data distributions. +- In practical tasks, data distributions are often unbalanced, which can lead to insufficient focus on key information in the model. In addition, current methods are unable to customize the degree of attention according to different data, and traditional training methods treat all features equally, which may lead to insufficient learning of key regions. + +To address the above problems, we propose a dynamic learning strategy based on nonuniform splitting mechanism and Hilbert space mapping. First, the framework dynamically adjusts the weights of different subsets through the nonuniform splitting mechanism and uses the a priori information of the data in combination with the diffusion factor to flexibly address uncertainties and conflicts. Second, the data are mapped into the Hilbert space for computation to alleviate the information conflict problem that may occur during the information fusion process. Third, a targeted training strategy is proposed to enhance the model's ability to learn important features and regions, which achieves results in both classification and enhancement tasks. In summary, the main contributions of our work are as follows: + +- We propose a nonuniform splitting mechanism. This mechanism can be dynamically adjusted according to the importance between different subsets, giving more weight to some subsets and less weight to others. The a priori information of the data can be utilized by introducing a diffusion factor, and this splitting mechanism, which is based on a priori information, can be more flexible in addressing uncertainty and conflicting information. +- We propose a scheme for fusing different BBAs. This scheme maps the data into Hilbert space for computation, which is more responsive to the differences in the true distributions of complex data. A specific way to compute the differences before fusing different BBAs is used to reduce the conflicts between different information. +- We propose an effective targeted training strategy (TTS). This strategy enhances the model's ability to learn specific information and regions. Higher weights are assigned to important features to increase attention, thus improving the overall performance of the task. Accuracy is improved in classification tasks, and data imbalance is alleviated in low-light image enhancement (LLIE) tasks. + +The rest of the paper is organized as follows. In section 2, related work is briefly introduced. section 3 describes the proposed method. section 4 describes the experiments and analysis of the results. Finally, section 5 provides a discussion and conclusion. + +# 2 Related Work + +# 2.1 Modelling of uncertain information + +In classification tasks, information uncertainty often leads the classifier to make incorrect decisions. To address this problem effectively, many methods based on statistical and distance metrics have been proposed with the aim of improving classification performance by means of different mathematical models. Cover and Hart [12] proposed a method based on a distance metric function, which is inferred by selecting nearest neighbor samples. Cortes and Vapnik [11] effectively differentiate between different data distributions by constructing a separating hyperplane that can correctly divide the training dataset and has a maximum geometric interval. Xanthopoulos et al. [55] proposed discriminant analysis on the basis of a statistical approach that uses the grouping information of known samples and their corresponding multivariate variable characteristics to infer the group to which new samples belong. Quinlan et al. [41] designed a model that is based on a tree structure, where each internal node represents a judgment on a feature and each branch represents a possible output of the judgment. However, the above methods may have difficulty making accurate inferences because of their own limitations when dealing with data with conflicting or redundant information, thus affecting inference efficiency. + +# 2.2 Evidence theory-based modelling of uncertain information + +Uncertain information can be classified into either the empty set $\varnothing$ or the whole set $\Omega$ on the basis of the modelling of the uncertain information, thus addressing the information uncertainty. Zhao et al. [71] obtained the final classification results by evaluating the reliability of single and multiple sources through independent and combined reliability assessments, respectively. Liu et al. [38] combined the inferred results from multiple models and used the average as the final output. Jousselme et al. [30] introduced the distance calculation method of the similarity measure to generate more reliable inference results. Zhang et al. [69] proposed a multisource information fusion algorithm based on belief $\chi^2$ scatter for inference tasks in complex data scenarios. However, owing to the uncertainty of the data itself and the possible conflicts or redundant information from different sources, effectively combining information from different sources by taking their importance into account remains a challenging research problem. + +# 2.3 Deep learning-based modelling of uncertain information + +Deep learning models often face the problem of decreased prediction reliability due to uncertain information. Chen et al. [8] proposed a radial basis function network learning algorithm based on orthogonal least squares to solve the underperformance problem caused by randomly selecting the centroid method to improve the performance of the task network. Castro et al. [4] investigated the problem of biased results due to data imbalance based on a multilayer perceptron (MLP) neural network and statistically improved the classification performance of the MLP. Sensoy et al. [45] used the prediction output from the network as subjective information for modelling, which in turn served as data support for the deterministic neural network to accomplish the subsequent classification task. Zaidi et al. [67] proposed two methods for automatically building a collection of different network architectures that can weigh the advantages of different structures well and use architectural variations as a source of diversity. However, these methods rely mostly on specific assumptions, and when there are multiple sources of conflicting contradictions in the input data, these models lack an effective dynamic processing framework, which may lead to bias in the final results. + +# 3 Method + +# 3.1 Motivation and overview + +In DSE theory, when dealing with highly conflicting information, existing methods cannot adequately consider the lack of precision due to ambiguity or uncertainty in BBAs. First, the existing allocation methods based on the splitting idea assume that the allocation is based on the premise of uniform splitting, which assumes that all subsets are equally proportioned to distribute the quality and cannot be dynamically adjusted according to the quality among different subsets. Thus, the differences between different subsets cannot be adequately considered. Therefore, a nonuniform splitting mechanism is introduced to give more weight to some subsets and less weight to others. This splitting mechanism, which is based on a priori information, namely the inherent characteristics of the data's own structure and the initial evidence distribution, can handle uncertainties and conflicting information more flexibly. The mechanism is adjustable to assign different split weights to different subsets through the diffusion coefficient. Second, conflicts arise from different evidence sources due to uncertainty or data distribution differences. To avoid the occurrence of the conflict phenomenon in the process of fusing different BBAs, high-order dynamic maximum mean difference (HODMMD) is proposed, which maps the data into the Hilbert space for computation and is more responsive to the differences in the real distributions of complex data. A specific way to calculate the difference before fusing different BBAs is used to reduce the conflict between different information. Third, in practical applications, the imbalance problem inherent in the data leads to insufficient attention of the model to key information. Therefore, an effective TTS is needed to assign higher weights to important features and increase attention, thus improving the overall performance of the task. + +![](images/aa1fba02cde02f9574815ba305a19b4af44ca7c8082306af0cc297fd201d50e1.jpg) +Figure 1: Overview of our targeted training strategy (TTS). + +Given a set of data $I$ , after feature extraction, different BBAs $\widehat{m_1}, \widehat{m_2}, \dots, \widehat{m_n}$ are obtained via adaptive diffusion probability transformation (ADPT) (Tadpt): + +$$ +\widehat {m _ {1}}, \widehat {m _ {2}}, \dots , \widehat {m _ {n}} = \operatorname {T a d p t} (I) \tag {1} +$$ + +After that, the reliability of the different BBAs is calculated in the probabilistic reliability assessment (PRA) (Tpra) stage: + +$$ +\dot {\phi} _ {i} = \operatorname {T p r a} \left(\widehat {m _ {i}}, \widehat {m _ {j}}\right) \quad (1 \leq i, j \leq n) \tag {2} +$$ + +These reliabilities are utilized as discount factors to perform collaborative decision optimization (CDO) (Tcdo) operations on the BBAs, which are fused to obtain the final decision result: + +$$ +\widehat {m} = \operatorname {T c d o} \left(\dot {\phi_ {1}}, \dot {\phi_ {2}}, \dots , \dot {\phi_ {i}}, \dots \dot {\phi_ {n}}; \quad \widehat {m _ {1}}, \widehat {m _ {2}}, \dots , \widehat {m _ {i}}, \dots , \widehat {m _ {n}}\right) \tag {3} +$$ + +where $\dot{\phi}_1,\dot{\phi}_2,\dots ,\dot{\phi}_i,\dots \dot{\phi}_n$ are discount factors. The final decision result is obtained after fusion. + +# 3.2 Adaptive diffusion probability transformation + +On the basis of the definition of DSE theory, the proposed ADPT model is as follows: + +$$ +\dot {m} _ {g} ^ {\tau} \left(A _ {i}\right) = \sum_ {A _ {i} \subseteq A _ {j}} D \left(A _ {i}, A _ {j}\right) \frac {\tau^ {\left| A _ {j} \right| - \left| A _ {i} \right|}}{(\tau + 1) ^ {\left| A _ {j} \right|} - \tau^ {\left| A _ {j} \right|}} \dot {m} _ {g} ^ {\tau - 1} \left(A _ {j}\right) \tag {4} +$$ + +where $A_{i}, A_{j} \in 2^{\Omega} \backslash \{\emptyset\}$ , $\tau$ is the number of iterative splits. $\dot{m}_{g}^{\tau}$ is $\tau$ order of $\dot{m}_{g}$ , and $\dot{m}_{g}^{0}$ is the initial basic belief assignment (BBA). $D(A_{i}, A_{j}) = \left(\frac{|A_{i}|}{|A_{j}|}\right)^{\frac{1}{\tau}}$ is a diffusion function used to control the quality of the distribution from subset $A_{i}$ to $A_{j}$ , which regulates the process in which the mass function is transferred from a fuzzy hypothesis $A_{j}$ to a more specific hypothesis $A_{i}$ , effectively quantifying and integrating prior information. When $\tau$ tends to $\infty$ , the result of $D(A_{i}, A_{j})$ tends to 1, and Equation (4) degenerates into a uniform distribution [25], which is no longer affected by the importance of different regions. $|\cdot|$ is the number of elements contained, $\Omega = \{\theta_{1}, \theta_{2}, \theta_{3}, \ldots, \theta_{N}\}$ is the frame of discernment, and $2^{\Omega} = \{\emptyset, \{\theta_{1}\}, \{\theta_{2}\}, \ldots, \{\theta_{N}\}, \{\theta_{1} \cup \theta_{2}\}, \ldots, \Omega\}$ is the power set. After normalization processing, + +$$ +\hat {m} _ {g} ^ {\tau} \left(A _ {i}\right) = \sum_ {A _ {i} \in 2 ^ {\Omega}} \frac {\dot {m} _ {g} ^ {\tau} \left(A _ {i}\right)}{\sum_ {A _ {i} \in 2 ^ {\Omega}} \dot {m} _ {g} ^ {\tau} \left(A _ {i}\right)} \tag {5} +$$ + +# 3.3 Probabilistic reliability assessment and its properties + +When different quality functions are obtained, if they are simply combined according to the DC rule [48], the integration between different BBAs may be hindered because of conflicting information. For this reason, the idea of discounting techniques [46] is introduced. Defined HODMMD: + +$$ +\mathrm {H O D M M D} ^ {\tau} \left(\widehat {m} _ {1}, \widehat {m} _ {2}\right) = \left\| \sum_ {i = 1} ^ {N} \left(\widehat {m} _ {g _ {1}} ^ {\tau} \left(A _ {i}\right) - \widehat {m} _ {g _ {2}} ^ {\tau} \left(A _ {i}\right)\right) \right\| _ {\mathcal {H}} \tag {6} +$$ + +where $\mathcal{H}$ is the regenerated kernel Hilbert space and the kernel function is computed via a Gaussian kernel function. The properties of the HODMMD are as follows. + +Property 1. When $\tau \to \infty$ , the HODMMD is equivalent to measuring the difference between the pignistic probability transformations $\mathrm{Bet}P$ with $\hat{m}_1$ and $\hat{m}_2$ of the maximum mean difference. + +$$ +\lim _ {\tau \rightarrow \infty} \mathrm {H O D M M D} ^ {\tau} (\widehat {m} _ {1}, \widehat {m} _ {2}) = \left\| \sum_ {i = 1} ^ {N} \left(\operatorname {B e t} P _ {1} \left(\theta_ {i}\right) - \operatorname {B e t} P _ {2} \left(\theta_ {i}\right)\right)\right\| _ {\mathcal {H}} \tag {7} +$$ + +Property 2. When $\widehat{m}_1$ and $\widehat{m}_2$ degenerate into probability distributions, that is, $U = (u_{1}, u_{2}, \ldots, u_{N})$ and $V = (v_{1}, v_{2}, \ldots, v_{N})$ , the proposed HODMMD degenerates into a maximum mean difference. + +$$ +\mathrm {H O D M M D} ^ {\tau} \left(\widehat {m} _ {1}, \widehat {m} _ {2}\right) = \mathrm {M M D} \left(\widehat {m} _ {1}, \widehat {m} _ {2}\right) \tag {8} +$$ + +Property 3. $\mathrm{HODMMD}^{\tau}\left(\widehat{m}_{1},\widehat{m}_{2}\right)$ and $\mathrm{HODMMD}^{\tau}\left(\widehat{m}_{2},\widehat{m}_{1}\right)$ are equivalent. + +$$ +\mathrm {H O D M M D} ^ {\tau} \left(\hat {m} _ {1}, \hat {m} _ {2}\right) = \mathrm {H O D M M D} ^ {\tau} \left(\hat {m} _ {2}, \hat {m} _ {1}\right) \tag {9} +$$ + +Property 4. When $\widehat{m}_1 = \widehat{m}_2$ , the value of HODMMD is always equal to 0. + +$$ +\mathrm {H O D M M D} ^ {\tau} \left(\widehat {m} _ {1}, \widehat {m} _ {2}\right) = 0 \tag {10} +$$ + +Proofs of these properties can be found in the technical appendix. + +# 3.4 Collaborative decision optimization + +When ADPT is used to obtain multiple BBAs, the reliability of different BBAs needs to be calculated to measure the impact of different sources of evidence. Let $\phi = [\phi_1,\phi_2,\dots,\phi_K]$ be the reliability of + +Algorithm 1: A dynamic learning framework based on DSE theory + +Input: Training data $X_{training}$ and testing set $X_{testing}$ . + +Output: Category probabilistic decision results + +1 for $i = 1$ to $K$ do +2 | Generate values on the basis of the data attribute characteristics of $X_{training}$ +3 end +4 Reliability of different classifiers obtained via decision optimization scheme Eq. (11) +5 Calculate the discount factor via Eq. (12) and normalize it +6 for $i = 1$ to $K$ do + +7 | Obtain the results of $K$ classifiers +8 Discounting the different classification results via Eq. (13) +9 Fuse different BBAs via Eq. (14) +10 Test the $k$ th basic classifiers +11 end +12 Use the decision results for subsequent tasks + +different query patterns or BBAs that satisfy $\phi_k \in [0,1]$ and $\sum_{1}^{K} \phi_k = 1$ . To measure the reliability of the different BBAs, the HODMMD is used for the calculation: + +$$ +\phi = \arg \min _ {\phi} \left(\sum_ {l = 1} ^ {L} \mathrm {H O D M M D} ^ {\tau} \left(G ^ {l}, \sum_ {k = 1} ^ {K} \phi_ {k} \widehat {m} _ {k} ^ {l}\right)\right) \quad \text {s . t .} \sum_ {k = 1} ^ {K} \phi_ {k} = 1 \tag {11} +$$ + +where $l$ is the index of the different query patterns. $G_{l} = [G_{l}(1), G_{l}(2), \dots, G_{l}(K)]$ is the ground truth. $m_{k}^{l}$ is the possibility of the query pattern belonging to the class $\theta_{i}$ . $\hat{m}_{k}^{l} = [\hat{m}_{k}^{l}(1), \hat{m}_{k}^{l}(2), \dots, \hat{m}_{k}^{l}(\Omega)]$ , which also satisfies $m_{k}^{l} \in [0,1]$ and $\sum_{k=1}^{2^{N}-1} m_{k}^{l} = 1$ . The reliability vector that minimizes the BBA error is calculated via sequential least squares programming (SLSQP). Therefore, the discount factor for the $k$ th BBA is defined as: + +$$ +\dot {\phi} _ {k} = \phi_ {k} / \max \left\{\phi_ {1}, \phi_ {2}, \dots , \phi_ {K} \right\} \tag {12} +$$ + +According to the idea of discounting techniques [46], the discounted BBA is as follows: + +$$ +\left\{ \begin{array}{l} \bar {m} _ {k} (X) = \dot {\phi} _ {k} \hat {m} _ {k} (X), \quad \forall X \not \subseteq \Omega \\ \bar {m} _ {k} (\Omega) = \dot {\phi} _ {k} \hat {m} _ {r} (\Omega) + 1 - \dot {\phi} _ {k} \end{array} \right. \tag {13} +$$ + +Fusing of different BBAs according to the DC rule: + +$$ +\mathbf {m} = \bar {m} _ {1} \oplus \bar {m} _ {2} \oplus \dots \oplus \bar {m} _ {R} \tag {14} +$$ + +The probabilities of different patterns obtained after fusion are prepared for subsequent classification and LLIE tasks. In LLIE tasks, the image is divided into different image blocks and then input into the proposed decision framework $\mathrm{T}_{\mathrm{trans}}$ to obtain the degradation degree representation of different features for each image block. The lower the quality is, the greater the weight assigned to the region. Given low-light image $I_{low}$ and normal light image $I_{normal}$ , + +$$ +I _ {p r o} = \operatorname {T} _ {\text {r a n s}} \left(I _ {\text {l o w}}, I _ {\text {n o r m a l}}\right) \tag {15} +$$ + +where $I_{pro}$ represents the representation of the degree of degradation in different regions of the image. Next, the LLIE task can be formalized as: + +$$ +\arg \min _ {\xi} \operatorname {L o s s} \left(T _ {\text {n e t} \xi} \left(I _ {\text {p r o}} \cdot I _ {\text {l o w}}\right), I _ {\text {n o r m a l}}\right) \tag {16} +$$ + +where $Loss$ is the loss function of the original network and where $T_{net_{\xi}}$ is the network with parameter $\xi$ , $\cdot$ represents the dot product. During this process, the network focuses more intensely on regions of the image that have undergone greater degradation, effectively integrating an understanding of the image content into the learning process. The above process of our method leads to Algorithm 1, which contains three modules: the ADPT, PRA, and CDO of classifiers. These three modules are co-optimized to work together. In this process, initial values are first generated on the basis of the features of the training data. Next, adaptive targeted iterations are performed via ADPT to achieve the generation of BBAs under different classifiers. After that, the reliability of each classifier is calculated as a discount factor using HODMMD. Finally, fusion is performed with the help of a discount factor to obtain the final decision. + +# 4 Experimental Results + +# 4.1 Numerical experiment + +To better understand the working mechanism of the ADPT, in this section, a discussion of the proposed ADPT is presented. by several concrete examples. + +![](images/0dea3cb83a8ed6b1cc2ec00256856e4d7c9b54446a795e12d3734589969c6aa9.jpg) +Figure 2: Diagram of the proportion of the mass function increasing with $\tau$ in Example 5.1. + +![](images/eb2512494320f515904b64ebc3e5c30bf144c6837a2f6909d855d7fc271bb9e9.jpg) + +Example 5.1 Let a frame of discernment be $\Omega = \{\theta_1, \theta_2\}$ . The BBA is as follows: + +$$ +\widehat {m _ {1}}: \widehat {m _ {1}} (\{\theta_ {1} \}) = 1 / 6, \widehat {m _ {1}} (\{\theta_ {2} \}) = 1 / 3, \widehat {m _ {1}} (\{\theta_ {1}, \theta_ {2} \}) = 1 / 2. +$$ + +When split once, when $\tau = 1$ is substituted into Equation (4), three scenarios exist: $m_{g}^{1}(\{a_{1}\}) = 0.3$ , $m_{g}^{1}(\{a_{2}\}) = 0.5$ , and $m_{g}^{1}(\{a_{1}, a_{2}\}) = 0.2$ . When the second, $\tau = 2$ , is substituted into Equation (4), there can be three scenarios: $m_{g}^{2}(\{a_{1}\}) = 0.3741$ , $m_{g}^{2}(\{a_{2}\}) = 0.5839$ , and $m_{g}^{2}(\{a_{1}, a_{2}\}) = 0.0420$ . The process was plotted as a radar chart and a line graph, and the results are shown in Figure 2. The results show that as $\tau$ increases, the values of the mass function belonging to $\{\theta_{1}, \theta_{2}\}$ gradually shift to the mass functions of $\{\theta_{1}\}$ and $\{\theta_{2}\}$ . The mass function with a larger set of initial values still maintains a larger weight. + +Example 5.2 Let a frame of discernment be $\Omega = \{\theta_1, \theta_2\}$ . The BBAs are as follows: + +$$ +\widehat {m _ {1}}: \widehat {m _ {1}} (\{\theta_ {1} \}) = (1 - x) / 3, \quad \widehat {m _ {1}} (\{\theta_ {2} \}) = (1 - x) / 3, \quad \widehat {m _ {1}} (\{\theta_ {1}, \theta_ {2} \}) = (2 x + 1) / 3, +$$ + +$$ +\widehat {m _ {2}}: \widehat {m _ {2}} (\{\theta_ {1} \}) = (1 - y) / 3, \quad \widehat {m _ {2}} (\{\theta_ {2} \}) = (1 - y) / 3, \quad \widehat {m _ {2}} (\{\theta_ {1}, \theta_ {2} \}) = (2 y + 1) / 3. +$$ + +The variation of the proposed HODMMD is further explored by changing the values of $x$ and $y$ and plotting Figure 3. As $\tau$ increases, the difference between $\widehat{m_1}$ and $\widehat{m_2}$ gradually decreases, verifying that the uncertainty information of the BBAs gradually decreases with increasing ADPT. + +![](images/704454331cb88dac02b9f9ce17122edd31624769a99b67e24d204421a7d7b92a.jpg) +Figure 3: HODMMD between $\widehat{m_1}$ and $\widehat{m_2}$ when $\tau$ is 1, 3, 5 in Example 5.2. + +# 4.2 Pattern classification + +For each dataset, the same testing method as in [58] was adopted to perform fivefold cross-validation according to a training set-to-test machine ratio of 4:1. The classification accuracy was evaluated, and then all the results were averaged to arrive at the final result, which ensures a fair comparison between the different methods. The experimental results are shown in Table 1. The proposed method achieves the optimal performance except for the Parkinsons dataset because the features of the different samples in this dataset are too close to each other, resulting in poor differentiation of importance when performing ADPT. On the basis of the evaluation metric of classification accuracy, improvements of $2.67\%$ , $11.9\%$ , $1.48\%$ , $1.66\%$ and $0.22\%$ are obtained over those of the advanced DMA method on + +the Iris, Heart, Hepatitis, Australian and Segment datasets, respectively. This is because our method better handles the contribution of different features to the classification and better solves the conflict problem when fusing different BBAs. + +Table 1: Comparison of the classification accuracy of different methods. The best, second and third results are in the red, green and blue colours. + +
DatasetIris[18]Heart[27]Hepatitis[1]Parkinsons[36]Australian[42]Segment[2]CBench [16]
NaB[23]94.67%82.59%76.76%68.75%79.56%80.22%67.68%
kNN[12]95.33%57.78%65.71%83.02%67.40%96.93%99.02%
REPTree[19]92.00%70.74%71.64%80.94%80.59%95.11%65.45%
SVM[5]94.67%83.70%79.96%70.13%80.29%64.50%88.48%
SVM-RBF[5]94.67%82.96%76.76%81.03%79.86%81.73%68.18%
MLP[4]93.33%75.19%74.93%74.39%82.32%95.93%80.40%
RBFN[8]92.67%81.85%81.32%82.05%82.61%87.58%83.64%
kNN-DST[15]95.33%76.30%80.57%78.01%78.41%93.37%94.42%
NDC[60]94.00%82.59%79.40%70.26%80.01%79.61%43.13%
EvC[61]94.67%83.70%79.88%81.64%80.60%95.90%76.33%
DMA[58]96.00%84.07%83.04%75.03%84.14%99.74%100.00%
ours98.67%95.97%84.52%82.50%85.80%99.96%100.00%
+ +# 4.3 Image classification + +For the image classification task, we examine the performance on the CIFAR-10 and CIFAR-100 datasets, which contain 10 and 100 categories, respectively. To verify the effectiveness of the proposed fusion framework, which is based on a nonuniform splitting mechanism and Hilbert space mapping more comprehensively, we perform feature extraction via ResNet-18 and the CNN framework proposed in [48] before adopting the proposed method for decision making. In this task, prior information is the initial probability distribution obtained by performing feature extraction on the input samples for the CNN and ResNet-18, with the last linear layer removed. The comparison results of the different methods are shown in Table 2. Our method achieves better results, which once again validates the effectiveness of the proposed method in addressing uncertain information. + +Table 2: Comparison of classification accuracy on the CIFAR-10 dataset[31] and CIFAR-100 dataset[31]. The best, second and third results are in the red, green and blue colours. + +
MethodsArchitectureAccuracyMethodsArchitectureAccuracy
CIFAR-10DIR-Net [40] ICV'2023ResNet-1892.80%Dspike[34] NeurIPS'2021ResNet-1973.12%
MST[51] ICCV'2023ResNet-1893.20 %GLIF[66] NeurIPS'2022ResNet-1977.05%
ReSTE[54] ICCV'2023ResNet-1892.63 %Diet-SNN[43] TNNLS'2021VGG-1669.67%
ADMM[9] TIP'2023ResNet-1895.40 %PASNN[17] KBS'2023ResNet-1472.63%
SML[14] ICML'2023ResNet-1995.12 %MPBN [22] ICCV'2023ResNet-1974.40%
UDSP[20] CVPR'2024ResNet-5693.78%MS-ResNet[24] ICCV'2023MS-ResNet1875.39%
BiPer [50] CVPR'2024ResNet-2093.75%CBFAR-100BKDSNN[63] ECCV'2024ResNet-19
ResNet-1893.75%
VGG-small92.11%
TAB[28] ICIL'2024VGG-993.41%TAB[28] ICIL'2024VGG-1176.31%
APL[70] TPAMI'2023ResNet-1896.00%APL [70] TPAMI'2024ResNet-1878.90%
ESNN[47] EAAI'2025VGG-1693.55%ESNN[47] EAAI'2025VGG-1676.55%
oursCNN95.67%oursCNN79.71%
ResNet-1895.61%ResNet-1879.78%
+ +# 4.4 Low-light image enhancement + +In LLIE, dark areas may contain critical information that is often difficult for models to adequately learn and focus on [35]. Traditional training strategies, which apply uniform processing across the entire image, are inherently limited in their ability to specifically enhance model learning for these critical regions. This often results in suboptimal detail recovery in the target areas during enhancement. + +Therefore, a TTS is proposed in conjunction with the proposed method, which explicitly guides the model to increase the level of attention to low-quality regions. The method is a plug-and-play module. In this task, prior information refers to the degree of degradation of different regions in the image. We used the LLIE network of the last few years as a baseline network, and the results are shown in Table 3. A consistent improvement in performance can be seen after using the TTS. To obtain a more intuitive sense of the enhancement, we present the results of the baseline method and our method, and the results are shown in Figure 4. Our method effectively improves the quality of images. For the LOL-v1 dataset, our method enhances the texture details on the glass; for the LOL-v2-real dataset, our method recovers the text on the wall more clearly; for the LOL-v2-real dataset, our method recovers the color of the flower petals closer to the ground truth. This finding verifies the validity of the proposed method as well as the idea and further shows that the proposed ADPT can characterize the data well. As seen through the above experiments, applying different learning strategies to different regions can effectively enhance the model's focus on the target region. By adaptively adjusting the learning weights of different regions, the model's ability to focus on low-quality regions is enhanced, the data distribution imbalance problem is alleviated, and the overall performance of the model is improved. To the best of our knowledge, this is the first time that DSE theory has been introduced into LLIE. + +More experiments can be found in the technical appendices. + +Table 3: Quantitative comparison of LOL-v1 [53], LOL-v2-real [65] and LOL-v2-syn [65]. + +
MethodsLOL-v1[53]LOL-v2-real[65]LOL-v2-syn[65]
PSNRSSIMPSNRSSIMPSNRSSIM
MIRNet[68] TPAMI'202024.140.83020.020.82021.940.876
FIDE[59] CVPR'202018.270.66516.850.67815.200.612
ZeroDCE[21] CVPR'202016.760.56018.060.57717.760.816
Sparse[53] TIP'202117.200.64020.060.81622.050.905
DRBN[64] TIP'202120.130.83020.290.83123.220.927
RUAS[37] CVPR'202118.230.72018.370.72316.550.652
ZeroDCE++[33] TPAMI'202116.110.53018.060.57718.030.825
SCI[39] CVPR'202214.780.52516.190.52216.670.811
Restormer[29] TCSVT'202322.430.82319.940.82721.410.830
SNR[62] CVPR'202224.610.84221.480.84924.140.928
SNR-TTS24.94(+0.33)0.854(+0.012)22.00(+0.52)0.846(-0.003)24.32(+0.18)0.929(+0.001)
LLFlow-L[52] AAAI'202224.990.87025.310.80525.880.908
LLFlow-L-TTS26.70(+1.71)0.860(-0.01)26.97(+1.66)0.865(+0.06)26.09(+0.21)0.906(-0.002)
LLFlow-S[52] AAAI'202224.060.86026.800.86025.300.877
LLFlow-S-TTS26.28(+2.22)0.848(-0.012)28.15(+1.35)0.866(+0.006)25.33(+0.03)0.880(+0.003)
Retinexformer[3] ICCV'202325.160.84522.800.84025.670.930
Retinexformer-TTS26.14(+0.98)0.849(+0.004)23.01(+0.21)0.843(+0.003)26.04(+0.37)0.942(+0.012)
+ +# 5 Conclusion + +This paper proposes a dynamic learning strategy based on nonuniform splitting mechanism and Hilbert space mapping, which is based on DSE theory, as an efficient method for processing uncertain information. The current method cannot dynamically adjust the tendency of splitting on the basis of the importance between different subsets in the process of splitting, and directly using the DC rule in fusion will produce unreasonable results owing to the conflict problem. First, the nonuniform splitting mechanism proposed in this paper takes the data's inherent a priori information into account and thus can handle uncertainty and conflict information more flexibly while improving the accuracy of the task. Second, mapping the data into the Hilbert space for computation is more responsive to the differences in the true distributions of complex data, thus providing an effective strategy for the subsequent fusion of different BBAs. We conducted experiments on multiple tasks as well as multiple publicly available datasets. The experimental results show that our method significantly outperforms existing machine learning methods and deep learning methods. To the best of our knowledge, we are the first to introduce DSE theory to LLIE and provide effective performance enhancement for LLIE tasks. + +![](images/67668a98204d7791c4f68e034784871cbc7f7a00e003874bed350c890a1bede4.jpg) +Figure 4: Quantitative comparison of LOL-v1 [53] (1st-2nd columns), LOL-v2-real [65] (3rd-4th columns), and LOL-v2-syn [65] (5th-6th columns). Even columns correspond to the results of adding the TTS module. + +# Acknowledgements + +This work was supported in part by the National Natural Science Foundation of China under Grant 62176027, in part by Chongqing Talent under Grant cstc2024ycjh-bgzxm0082, in part by Chongqing New Yin Cai (YC) Project under Grant CSTB2024YCJH-KYXM0126, in part by the General Program of the Natural Science Foundation of Chongqing under Grant CSTB2024NSCQ-MSX0479, and in part by Chongqing Postdoctoral Foundation Special Support Program under Grant 2023CQB-SHTB3119. + +# References + +[1] Bache, K. & Lichman, M. (1983) Hepatitis. UCI Machine Learning Repository +[2] Bache, K. & Lichman, M. (1990) Image segmentation. UCI Machine Learning Repository +[3] Cai, Y., Bian, H., Lin, J., Wang, H., Timofte, R., & Zhang, Y. (2023) Retinexformer: One-stage retina-based transformer for low-light image enhancement. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) pages 12504-12513. +[4] Castro, C. L. & Braga, A. P. (2013) Novel cost-sensitive approach to improve the multilayer perceptron performance on imbalanced data. IEEE transactions on neural networks and learning systems 24(6):888-899. +[5] Chang, C.-C. & Lin, C.-J. (2011) Libsvm: a library for support vector machines. ACM transactions on intelligent systems and technology (TIST) 2(3):1-27. +[6] Chen, C., Chen, Q., Xu, J., & Koltun, V. (2018) Learning to see in the dark pages 3291-3300. +[7] Chen, C., Chen, Q., Do, M. N., & Koltun, V. (2019) Seeing motion in the dark pages 3185-3194. +[8] Chen, S., Cowan, C., & Grant, P. (1991) Orthogonal least squares learning algorithm for radial basis function networks. IEEE Transactions on Neural Networks 2(2):302-309. +[9] Chiou, C.-Y., Lee, K.-T., Huang, C.-R., & others (2023) Admm-snet: Alternating direction method of multipliers based sparse representation network for one-class classification. IEEE Transactions on Image Processing 32:2843-2856. +[10] Constance, L., Driessen, A., Deutschmann, N., & Martinez, M. R. (2022) Fuzzy logic for biological networks as ml regression: Scaling to single-cell datasets with autograd. In NeurIPS 2022 Workshop on Learning Meaningful Representations of Life +[11] Cortes, C. & Vapnik, V. (1995) Support-vector networks. Machine learning 20:273-297. + +[12] Cover, T. & Hart, P. (1967) Nearest neighbor pattern classification. IEEE transactions on information theory 13(1):21-27. +[13] Dempster, A. P. (2008) Upper and lower probabilities induced by a multivalued mapping. Classic works of the Dempster-Shafer theory of belief functions pages 57-72. +[14] Deng, S., Lin, H., Li, Y., & Gu, S. (2023) Surrogate module learning: Reduce the gradient error accumulation in training spiking neural networks. In International Conference on Machine Learning pages 7645-7657. PMLR. +[15] Denoeux, T. (1995) A k-nearest neighbor classification rule based on dampster-shafer theory. IEEE Transactions on Systems, Man, and Cybernetics 25(5):804-813. +[16] Deterding, N. M. & Robinson, T. (1988) Connectionist bench (vowel recognition - deterring data). UCI Machine Learning Repository +[17] Ding, Y., Zuo, L., Yang, K., Chen, Z., Hu, J., & Xiahou, T. (2023) An improved probabilistic spiking neural network with enhanced discriminative ability. Knowledge-Based Systems 280:111024. +[18] Fisher, R. A. (1936) Iris. UCI Machine Learning Repository +[19] Freund, Y. & Mason, L. (1999) The alternating decision tree learning algorithm page 124-133. +[20] Gao, S., Zhang, Y., Huang, F., & Huang, H. (2024) Bilevelpruning: Unified dynamic and static channel pruning for convolutional neural networks. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pages 16090-16100. +[21] Guo, C., Li, C., Guo, J., Loy, C. C., Hou, J., Kwong, S., & Cong, R. (2020) Zero-reference deep curve estimation for low-light image enhancement. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition pages 1780-1789. +[22] Guo, Y., Zhang, Y., Chen, Y., Peng, W., Liu, X., Zhang, L., Huang, X., & Ma, Z. (2023) Membrane potential batch normalization for spiking neural networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision pages 19420-19430. +[23] Hu, B.-G. (2013) What are the differences between bayesian classifiers and mutual-information classifiers? IEEE transactions on neural networks and learning systems 25(2):249-264. +[24] Hu, Y., Deng, L., Wu, Y., Yao, M., & Li, G. (2024) Advancing spiking neural networks toward deep residual learning. IEEE transactions on neural networks and learning systems 36(2):2353-2367. +[25] Huang, Y., Xiao, F., Cao, Z., & Lin, C.-T. (2023) Higher order fractal belief rényi divergence with its applications in pattern classification. IEEE Transactions on Pattern Analysis and Machine Intelligence +[26] Huang, Y., Hechen, Z., Zhou, M., Li, Z., & Kwong, S. (2025) An attention-locating algorithm for eliminating background effects in fine-grained visual classification. IEEE Transactions on Circuits and Systems for Video Technology +[27] Janosi, S. W. P. M. & Detrano, R. (1989) Heart disease. UCI Machine Learning Repository +[28] Jiang, H., Zoonekynd, V., De Masi, G., Gu, B., & Xiong, H. (2024) Tab: Temporal accumulated batch normalization in spiking neural networks. In The Twelfth International Conference on Learning Representations +[29] Jiang, N., Lin, J., Zhang, T., Zheng, H., & Zhao, T. (2023) Low-light image enhancement via stage-transformer-guided network. IEEE Transactions on Circuits and Systems for Video Technology 33(8): 3701-3712. +[30] Jousselme, A.-L., Grenier, D., & Bosse, É. (2001) A new distance between two bodies of evidence. Information fusion 2(2):91-101. +[31] Krizhevsky, A., Hinton, G., & others (2009) Learning multiple layers of features from tiny images +[32] Lee, H. & Kwon, H. (2019) Dbf: Dynamic belief fusion for combining multiple object detectors. IEEE transactions on pattern analysis and machine intelligence 43(5):1499-1514. +[33] Li, C., Guo, C., & Loy, C. C. (2021.) Learning to enhance low-light image via zero-reference deep curve estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(8):4225-4238. + +[34] Li, Y., Guo, Y., Zhang, S., Deng, S., Hai, Y., & Gu, S. (2021). Differentiable spike: Rethinking gradient-descent for training spiking neural networks. Advances in neural information processing systems 34: 23426-23439. +[35] Li, Y., Wei, X., Liao, X., Zhao, Y., Jia, F., Zhuang, X., & Zhou, M. (2024) A deep retina-based low-light enhancement network fusing rich intrinsic prior information. ACM Transactions on Multimedia Computing, Communications and Applications 20(11):1-23. +[36] Little, M. (2007) Parkinsons. UCI machine learning repository +[37] Liu, R., Ma, L., Zhang, J., Fan, X., & Luo, Z. (2021) Retinex-inspired unrolling with cooperative prior architecture search for low-light image enhancement. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pages 10556-10565. +[38] Liu, Z., Pan, Q., Dezert, J., Han, J.-W., & He, Y. (2017) Classifier fusion with contextual reliability evaluation. IEEE transactions on cybernetics 48(5):1605-1618. +[39] Ma, L., Ma, T., Liu, R., Fan, X., & Luo, Z. (2022) Toward fast, flexible, and robust low-light image enhancement. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition pages 5637-5646. +[40] Qin, H., Zhang, X., Gong, R., Ding, Y., Xu, Y., & Liu, X. (2023) Distribution-sensitive information retention for accurate binary neural network. International Journal of Computer Vision 131(1):26-47. +[41] Quinlan, J. R. (1996) Learning decision tree classifiers. ACM Computing Surveys (CSUR) 28(1):71-72. +[42] Quinlan, R. (1987) Statlog (australian credit approval). UCI Machine Learning Repository +[43] Rathi, N. & Roy, K. (2021) Diet-snn: A low-latency spiking neural network with direct input encoding and leakage and threshold optimization. IEEE Transactions on Neural Networks and Learning Systems 34(6): 3174-3182. +[44] Sanchez, T., Krawczuk, I., Sun, Z., & Cevher, V. (2020) Uncertainty-driven adaptive sampling via gans. In NeurIPS 2020 workshop on deep learning and inverse problems +[45] Sensoy, M., Kaplan, L., & Kandemir, M. (2018) Evidential deep learning to quantify classification uncertainty. Advances in neural information processing systems 31. +[46] Shafer, G. (1976) A mathematical theory of evidence, 42. 42: Princeton university press. +[47] Tang, X., Chen, T., Cheng, Q., Shen, H., Duan, S., & Wang, L. (2025) Spatio-temporal channel attention and membrane potential modulation for efficient spiking neural network. Engineering Applications of Artificial Intelligence 148:110131. +[48] Tong, Z., Xu, P., & Denoeux, T. (2021) An evidential classifier based on dampster-shafer theory and deep learning. Neurocomputing 450:275-293. +[49] Trivedi, P., Heimann, M., Anirudh, R., Koutra, D., & Thiagarajan, J. J. (2023) Estimating epistemic uncertainty of graph neural networks using stochastic centering. In NeurIPS 2023 Workshop: New Frontiers in Graph Learning +[50] Vargas, E., Correa, C. V., Hinojosa, C., & Arguello, H. (2024) Biper: Binary neural networks using a periodic function. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pages 5684-5693. +[51] Vo, Q. H., Tran, L.-T., Bae, S.-H., Kim, L.-W., & Hong, C. S. (2023) Mst-compression: Compressing and accelerating binary neural networks with minimum spanning tree. In Proceedings of the IEEE/CVF International Conference on Computer Vision pages 6091-6100. +[52] Wang, Y., Wan, R., Yang, W., Li, H., Chau, L.-P., & Kot, A. (2022) Low-light image enhancement with normalizing flow. In Proceedings of the AAAI conference on artificial intelligence 36, pp. 2604-2612. +[53] Wei, C., Wang, W., Yang, W., & Liu, J. (2018) Deep retina decomposition for low-light enhancement. IN BMVC +[54] Wu, X.-M., Zheng, D., Liu, Z., & Zheng, W.-S. (2023) Estimator meets equilibrium perspective: A rectified straight through estimator for binary neural networks training. In Proceedings of the IEEE/CVF International Conference on Computer Vision pages 17055-17064. + +[55] Xanthopoulos, P., Pardalos, P. M., Trafalis, T. B., Xanthopoulos, P., Pardalos, P. M., & Trafalis, T. B. (2013) Linear discriminant analysis. Robust data mining pages 27-33. +[56] Xia, T., Han, J., Qendro, L., Dang, T., & Mascolo, C. (2022) Hybrid-edl: Improving evidential deep learning for uncertainty quantification on imbalanced data. In Workshop on Trustworthy and Socially Responsible Machine Learning, NeurIPS 2022 +[57] Xiao, F. (2023) Quantum x-entropy in generalized quantum evidence theory. Information Sciences 643: 119177. +[58] Xiao, F., Wen, J., & Pedrycz, W. (2022) Generalized divergence-based decision making method with an application to pattern classification. IEEE transactions on knowledge and data engineering 35(7):6941-6956. +[59] Xu, K., Yang, X., Yin, B., & Lau, R. W. (2020) Learning to restore low-light images via decomposition-and-enhancement. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition pages 2281-2290. +[60] Xu, P., Deng, Y., Su, X., & Mahadevan, S. (2013) A new method to determine basic probability assignment from training data. Know.-Based Syst. 46:69-80. +[61] Xu, P., Davoine, F., Zha, H., & Denoeux, T. (2016) Evidential calibration of binarysvm classifiers. International Journal of Approximate Reasoning 72:55-70. +[62] Xu, X., Wang, R., Fu, C.-W., & Jia, J. (2022) Snr-aware low-light image enhancement. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pages 17693-17703. +[63] Xu, Z., You, K., Guo, Q., Wang, X., & He, Z. (2024) Bkdsnn: Enhancing the performance of learning-based spiking neural networks training with blurred knowledge distillation. In European Conference on Computer Vision pages 106-123. Springer. +[64] Yang, W., Wang, S., Fang, Y., Wang, Y., & Liu, J. (2021.) Band representation-based semi-supervised low-light image enhancement: Bridging the gap between signal fidelity and perceptual quality. IEEE Transactions on Image Processing 30:3461-3473. +[65] Yang, W., Wang, W., Huang, H., Wang, S., & Liu, J. (2021.) Sparse gradient regularized deep retina network for robust low-light image enhancement. IEEE Transactions on Image Processing 30:2072-2086. +[66] Yao, X., Li, F., Mo, Z., & Cheng, J. (2022) Glif: A unified gated leaky integrate-and-fire neuron for spiking neural networks. Advances in Neural Information Processing Systems 35:32160-32171. +[67] Zaidi, S., Zela, A., Elsken, T., Holmes, C. C., Hutter, F., & Teh, Y. (2021) Neural ensemble search for uncertainty estimation and dataset shift. Advances in Neural Information Processing Systems 34:7898-7911. +[68] Zamir, S. W., Arora, A., Khan, S., Hayat, M., Khan, F. S., Yang, M.-H., & Shao, L. (2022) Learning enriched features for fast image restoration and enhancement. IEEE transactions on pattern analysis and machine intelligence 45(2):1934-1948. +[69] Zhang, L. & Xiao, F. (2022) A novel belief $\chi^2$ divergence for multisource information fusion and its application in pattern classification. International Journal of Intelligent Systems 37(10):7968-7991. +[70] Zhang, L., Qi, L., Yang, X., Qiao, H., Yang, M.-H., & Liu, Z. (2023) Automatically discovering novel visual categories with adaptive prototype learning. IEEE transactions on pattern analysis and machine intelligence 46(4):2533-2544. +[71] Zhao, J., Xue, R., Dong, Z., Tang, D., & Wei, W. (2020) Evaluating the reliability of sources of evidence with a two-perspective approach in classification problems based on evidence theory. Information Sciences 507: 313-338. + +# NeurIPS Paper Checklist + +# 1. Claims + +Question: Do the main claims made in the abstract and introduction accurately reflect the paper's contributions and scope? + +Answer: [Yes] + +Justification: The contributions and scope of this paper are clearly presented in both the abstract and the introduction as well as verified in the experiment section. + +Guidelines: + +- The answer NA means that the abstract and introduction do not include the claims made in the paper. +- The abstract and/or introduction should clearly state the claims made, including the contributions made in the paper and important assumptions and limitations. A No or NA answer to this question will not be perceived well by the reviewers. +- The claims made should match theoretical and experimental results, and reflect how much the results can be expected to generalize to other settings. +- It is fine to include aspirational goals as motivation as long as it is clear that these goals are not attained by the paper. + +# 2. Limitations + +Question: Does the paper discuss the limitations of the work performed by the authors? + +Answer: [Yes] + +Justification: In the experimental part, we find that the proposed method has limitations on some datasets but still has room for improvement, which is discussed in the experimental analysis section. + +Guidelines: + +- The answer NA means that the paper has no limitation while the answer No means that the paper has limitations, but those are not discussed in the paper. +- The authors are encouraged to create a separate "Limitations" section in their paper. +- The paper should point out any strong assumptions and how robust the results are to violations of these assumptions (e.g., independence assumptions, noiseless settings, model well-specification, asymptotic approximations only holding locally). The authors should reflect on how these assumptions might be violated in practice and what the implications would be. +- The authors should reflect on the scope of the claims made, e.g., if the approach was only tested on a few datasets or with a few runs. In general, empirical results often depend on implicit assumptions, which should be articulated. +- The authors should reflect on the factors that influence the performance of the approach. For example, a facial recognition algorithm may perform poorly when image resolution is low or images are taken in low lighting. Or a speech-to-text system might not be used reliably to provide closed captions for online lectures because it fails to handle technical jargon. +- The authors should discuss the computational efficiency of the proposed algorithms and how they scale with dataset size. +- If applicable, the authors should discuss possible limitations of their approach to address problems of privacy and fairness. +- While the authors might fear that complete honesty about limitations might be used by reviewers as grounds for rejection, a worse outcome might be that reviewers discover limitations that aren't acknowledged in the paper. The authors should use their best judgment and recognize that individual actions in favor of transparency play an important role in developing norms that preserve the integrity of the community. Reviewers will be specifically instructed to not penalize honesty concerning limitations. + +# 3. Theory Assumptions and Proofs + +Question: For each theoretical result, does the paper provide the full set of assumptions and a complete (and correct) proof? + +# Answer: [Yes] + +Justification: For the theories mentioned in the paper, we provide examples. The theory-related properties provide complete (correct) proofs, as shown in the technical appendices. + +# Guidelines: + +- The answer NA means that the paper does not include theoretical results. +- All the theorems, formulas, and proofs in the paper should be numbered and cross-referenced. +- All assumptions should be clearly stated or referenced in the statement of any theorems. +- The proofs can either appear in the main paper or the supplemental material, but if they appear in the supplemental material, the authors are encouraged to provide a short proof sketch to provide intuition. +- Inversely, any informal proof provided in the core of the paper should be complemented by formal proofs provided in appendix or supplemental material. +- Theorems and Lemmas that the proof relies upon should be properly referenced. + +# 4. Experimental Result Reproducibility + +Question: Does the paper fully disclose all the information needed to reproduce the main experimental results of the paper to the extent that it affects the main claims and/or conclusions of the paper (regardless of whether the code and data are provided or not)? + +# Answer: [Yes] + +Justification: Relevant information and details are described in the experimental results section and more details on the experimental settings section in the technical appendices, and the code is publicly available. + +# Guidelines: + +- The answer NA means that the paper does not include experiments. +- If the paper includes experiments, a No answer to this question will not be perceived well by the reviewers: Making the paper reproducible is important, regardless of whether the code and data are provided or not. +- If the contribution is a dataset and/or model, the authors should describe the steps taken to make their results reproducible or verifiable. +- Depending on the contribution, reproducibility can be accomplished in various ways. For example, if the contribution is a novel architecture, describing the architecture fully might suffice, or if the contribution is a specific model and empirical evaluation, it may be necessary to either make it possible for others to replicate the model with the same dataset, or provide access to the model. In general, releasing code and data is often one good way to accomplish this, but reproducibility can also be provided via detailed instructions for how to replicate the results, access to a hosted model (e.g., in the case of a large language model), releasing of a model checkpoint, or other means that are appropriate to the research performed. +- While NeurIPS does not require releasing code, the conference does require all submissions to provide some reasonable avenue for reproducibility, which may depend on the nature of the contribution. For example +(a) If the contribution is primarily a new algorithm, the paper should make it clear how to reproduce that algorithm. +(b) If the contribution is primarily a new model architecture, the paper should describe the architecture clearly and fully. +(c) If the contribution is a new model (e.g., a large language model), then there should either be a way to access this model for reproducing the results or a way to reproduce the model (e.g., with an open-source dataset or instructions for how to construct the dataset). +(d) We recognize that reproducibility may be tricky in some cases, in which case authors are welcome to describe the particular way they provide for reproducibility. In the case of closed-source models, it may be that access to the model is limited in some way (e.g., to registered users), but it should be possible for other researchers to have some path to reproducing or verifying the results. + +# 5. Open access to data and code + +Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material? + +Answer: [Yes] + +Justification: The source code has been released, and the link is reflected in the abstract. + +Guidelines: + +- The answer NA means that paper does not include experiments requiring code. +- Please see the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details. +- While we encourage the release of code and data, we understand that this might not be possible, so "No" is an acceptable answer. Papers cannot be rejected simply for not including code, unless this is central to the contribution (e.g., for a new open-source benchmark). +- The instructions should contain the exact command and environment needed to run to reproduce the results. See the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details. +- The authors should provide instructions on data access and preparation, including how to access the raw data, preprocessed data, intermediate data, and generated data, etc. +- The authors should provide scripts to reproduce all experimental results for the new proposed method and baselines. If only a subset of experiments are reproducible, they should state which ones are omitted from the script and why. +- At submission time, to preserve anonymity, the authors should release anonymized versions (if applicable). +- Providing as much information as possible in supplemental material (appended to the paper) is recommended, but including URLs to data and code is permitted. + +# 6. Experimental Setting/Details + +Question: Does the paper specify all the training and test details (e.g., data splits, hyperparameters, how they were chosen, type of optimizer, etc.) necessary to understand the results? + +Answer: [Yes] + +Justification: All relevant settings are described in the experimental results section, and more details on the experimental settings are provided in the technical appendices. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The experimental setting should be presented in the core of the paper to a level of detail that is necessary to appreciate the results and make sense of them. +- The full details can be provided either with the code, in appendix, or as supplemental material. + +# 7. Experiment Statistical Significance + +Question: Does the paper report error bars suitably and correctly defined or other appropriate information about the statistical significance of the experiments? + +Answer: [No] + +Justification: The experimental results in this paper require no further error analysis to be reported. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The authors should answer "Yes" if the results are accompanied by error bars, confidence intervals, or statistical significance tests, at least for the experiments that support the main claims of the paper. + +- The factors of variability that the error bars are capturing should be clearly stated (for example, train/test split, initialization, random drawing of some parameter, or overall run with given experimental conditions). +- The method for calculating the error bars should be explained (closed form formula call to a library function, bootstrap, etc.) +- The assumptions made should be given (e.g., Normally distributed errors). +- It should be clear whether the error bar is the standard deviation or the standard error of the mean. +- It is OK to report 1-sigma error bars, but one should state it. The authors should preferably report a 2-sigma error bar than state that they have a $96\%$ CI, if the hypothesis of Normality of errors is not verified. +- For asymmetric distributions, the authors should be careful not to show in tables or figures symmetric error bars that would yield results that are out of range (e.g. negative error rates). +- If error bars are reported in tables or plots, The authors should explain in the text how they were calculated and reference the corresponding figures or tables in the text. + +# 8. Experiments Compute Resources + +Question: For each experiment, does the paper provide sufficient information on the computer resources (type of compute workers, memory, time of execution) needed to reproduce the experiments? + +Answer: [Yes] + +Justification: We describe the training settings for each experiment as described in the experimental results section and more details on the experimental settings section in the technical appendices. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The paper should indicate the type of compute workers CPU or GPU, internal cluster or cloud provider, including relevant memory and storage. +- The paper should provide the amount of compute required for each of the individual experimental runs as well as estimate the total compute. +- The paper should disclose whether the full research project required more compute than the experiments reported in the paper (e.g., preliminary or failed experiments that didn't make it into the paper). + +# 9. Code Of Ethics + +Question: Does the research conducted in the paper conform, in every respect, with the NeurIPS Code of Ethics https://neurips.cc/public/EthicsGuidelines? + +Answer: [Yes] + +Justification: We have read the NeurIPS Code of Ethics and have strictly adhered to it during the course of our research. + +Guidelines: + +The answer NA means that the authors have not reviewed the NeurIPS Code of Ethics +- If the authors answer No, they should explain the special circumstances that require a deviation from the Code of Ethics. +- The authors should make sure to preserve anonymity (e.g., if there is a special consideration due to laws or regulations in their jurisdiction). + +# 10. Broader Impacts + +Question: Does the paper discuss both potential positive societal impacts and negative societal impacts of the work performed? + +Answer: [Yes] + +Justification: We explore some of the possible social implications of our approach in the concluding section and technical appendices. + +# Guidelines: + +- The answer NA means that there is no societal impact of the work performed. +- If the authors answer NA or No, they should explain why their work has no societal impact or why the paper does not address societal impact. +- Examples of negative societal impacts include potential malicious or unintended uses (e.g., disinformation, generating fake profiles, surveillance), fairness considerations (e.g., deployment of technologies that could make decisions that unfairly impact specific groups), privacy considerations, and security considerations. +- The conference expects that many papers will be foundational research and not tied to particular applications, let alone deployments. However, if there is a direct path to any negative applications, the authors should point it out. For example, it is legitimate to point out that an improvement in the quality of generative models could be used to generate deepfakes for disinformation. On the other hand, it is not needed to point out that a generic algorithm for optimizing neural networks could enable people to train models that generate Deepfakes faster. +- The authors should consider possible harms that could arise when the technology is being used as intended and functioning correctly, harms that could arise when the technology is being used as intended but gives incorrect results, and harms following from (intentional or unintentional) misuse of the technology. +- If there are negative societal impacts, the authors could also discuss possible mitigation strategies (e.g., gated release of models, providing defenses in addition to attacks, mechanisms for monitoring misuse, mechanisms to monitor how a system learns from feedback over time, improving the efficiency and accessibility of ML). + +# 11. Safeguards + +Question: Does the paper describe safeguards that have been put in place for responsible release of data or models that have a high risk for misuse (e.g., pretrained language models, image generators, or scraped datasets)? + +Answer: [NA] + +Justification: There is no security risk in this paper. + +# Guidelines: + +- The answer NA means that the paper poses no such risks. +- Released models that have a high risk for misuse or dual-use should be released with necessary safeguards to allow for controlled use of the model, for example by requiring that users adhere to usage guidelines or restrictions to access the model or implementing safety filters. +- Datasets that have been scraped from the Internet could pose safety risks. The authors should describe how they avoided releasing unsafe images. +- We recognize that providing effective safeguards is challenging, and many papers do not require this, but we encourage authors to take this into account and make a best faith effort. + +# 12. Licenses for existing assets + +Question: Are the creators or original owners of assets (e.g., code, data, models), used in the paper, properly credited and are the license and terms of use explicitly mentioned and properly respected? + +Answer: [Yes] + +Justification: The relevant information used in this paper has been explicitly mentioned and properly respected. + +# Guidelines: + +- The answer NA means that the paper does not use existing assets. +- The authors should cite the original paper that produced the code package or dataset. +- The authors should state which version of the asset is used and, if possible, include a URL. +- The name of the license (e.g., CC-BY 4.0) should be included for each asset. + +- For scraped data from a particular source (e.g., website), the copyright and terms of service of that source should be provided. +- If assets are released, the license, copyright information, and terms of use in the package should be provided. For popular datasets, paperswithcode.com/datasets has curated licenses for some datasets. Their licensing guide can help determine the license of a dataset. +- For existing datasets that are re-packaged, both the original license and the license of the derived asset (if it has changed) should be provided. +- If this information is not available online, the authors are encouraged to reach out to the asset's creators. + +# 13. New Assets + +Question: Are new assets introduced in the paper well documented and is the documentation provided alongside the assets? + +Answer: [Yes] + +Justification: The source code has been released, and the link is reflected in the abstract. + +Guidelines: + +- The answer NA means that the paper does not release new assets. +- Researchers should communicate the details of the dataset/code/model as part of their submissions via structured templates. This includes details about training, license, limitations, etc. +- The paper should discuss whether and how consent was obtained from people whose asset is used. +- At submission time, remember to anonymize your assets (if applicable). You can either create an anonymized URL or include an anonymized zip file. + +# 14. Crowdsourcing and Research with Human Subjects + +Question: For crowdsourcing experiments and research with human subjects, does the paper include the full text of instructions given to participants and screenshots, if applicable, as well as details about compensation (if any)? + +Answer: [NA] + +Justification: This paper does not include crowdsourcing experiments or research with human subjects. + +Guidelines: + +- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. +- Including this information in the supplemental material is fine, but if the main contribution of the paper involves human subjects, then as much detail as possible should be included in the main paper. +- According to the NeurIPS Code of Ethics, workers involved in data collection, curation, or other labor should be paid at least the minimum wage in the country of the data collector. + +# 15. Institutional Review Board (IRB) Approvals or Equivalent for Research with Human Subjects + +Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals (or an equivalent approval/review based on the requirements of your country or institution) were obtained? + +Answer: [NA] + +Justification: This paper does not include crowdsourcing experiments or research with human subjects. + +Guidelines: + +- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. + +- Depending on the country in which research is conducted, IRB approval (or equivalent) may be required for any human subjects research. If you obtained IRB approval, you should clearly state this in the paper. +- We recognize that the procedures for this may vary significantly between institutions and locations, and we expect authors to adhere to the NeurIPS Code of Ethics and the guidelines for their institution. +- For initial submissions, do not include any information that would break anonymity (if applicable), such as the institution conducting the review. + +# A Appendix/supplemental material + +# A.1 More Details of Preparatory Knowledge + +# A.1.1 Dempster-Shafer evidence theory + +Dempster-shafer evidence theory is a well-established general framework for uncertainty reasoning that was first proposed by Arthur P. Dempster [13] on the basis of statistical inference and later formalized and significantly extended by Glenn Shafer into a framework for simulating epistemic uncertainty[46]. Owing to its effectiveness in dealing with uncertain information, DSE theory is widely used in various fields. + +Definition 1. Frame of discernment + +Define a set of classes called discriminant frames: + +$$ +\Omega = \left\{\theta_ {1}, \theta_ {2}, \theta_ {3}, \dots , \theta_ {N} \right\} \tag {17} +$$ + +where $\theta_{i}(i = 1,2,\dots,N)$ are mutually exclusive. The power set is defined as: + +$$ +2 ^ {\Omega} = \{\emptyset , \left\{\theta_ {1} \right\}, \left\{\theta_ {2} \right\}, \dots , \left\{\theta_ {N} \right\}, \left\{\theta_ {1} \cup \theta_ {2} \right\}, \dots , \Omega \} \tag {18} +$$ + +where $\Omega$ is the whole set, $\varnothing$ is the empty set and the one containing $\cup$ is a multielement set [13, 46]. + +Definition 2. Basic belief assignment + +The basic belief assignment (BBA), also known as the mass function, denoted $m(\cdot)$ is defined as a mapping of $2^{\Omega}$ to the interval [0, 1] [13, 46]. + +$$ +m: 2 ^ {\Omega} \rightarrow [ 0, 1 ] \tag {19} +$$ + +and is satisfied: + +$$ +\sum_ {A \in 2 ^ {\Omega} \backslash \{\emptyset \}} m (A) = 1 \tag {20} +$$ + +$$ +m (\emptyset) = 0 \tag {21} +$$ + +where $A\in 2^{\Omega}\backslash \{\emptyset \}$ $\emptyset$ is the empty set. + +Definition 3. Focal set + +If $m(A) > 0$ , $A$ is called a focal set, and the value of $m(\cdot)$ indicates the level of support of the model[13, 46]. + +Definition 4. Dempster's combination rule + +Different fusion algorithms can be used to fuse information from different sources, one of which is Dempster's combination rule. As it can process the fusion of different sources of evidence represented by a BBA, it is assumed that $m_{1}$ and $m_{2}$ are two mutually independent BBAs defined on the same recognition frame $2^{\Omega}$ . Dempster's combination rule aims to derive a combined BBA, usually denoted as $m = m_{1} \oplus m_{2}$ . The combination frame is defined as follows[13]: + +$$ +m \left(A _ {k}\right) = \left(m _ {1} \oplus m _ {2}\right) \left(A _ {k}\right) = \frac {\sum_ {A _ {i} \cap A _ {j} = A _ {k}} m _ {1} \left(A _ {i}\right) m _ {2} \left(A _ {j}\right)}{1 - \mathcal {R}} \tag {22} +$$ + +$$ +\mathcal {R} = \sum_ {A _ {i} \cap A _ {j} = \emptyset} m _ {1} \left(A _ {i}\right) m _ {2} \left(A _ {j}\right) < 1 \tag {23} +$$ + +where $A_{i}, A_{j}, A_{k} \in \Omega$ and $\mathcal{R}$ are the conflict coefficients of $m_{1}$ and $m_{2}$ . + +# A.1.2 Probability transformation methods + +Definition 5. Pignistic probability transformation + +Let $\Omega = \{\theta_1, \theta_2, \theta_3, \ldots, \theta_N\}$ denote the discriminant framework. Given a corresponding BBA defined on $\Omega$ , denoted $m(\cdot)$ , the pignistic probabilistic transformation of an element $\theta_i \in \Omega$ , denoted $P_{ppt}$ , is defined as follows: + +$$ +P _ {p p t} \left(\theta_ {i}\right) = \sum_ {\theta_ {i} \in A | A \in 2 ^ {\Omega}} \frac {m (A)}{| A |} \tag {24} +$$ + +where $A$ belongs to $2^{\Omega}$ , $A \neq \emptyset$ . $|A|$ is the cardinality of $A$ . + +Definition 6. Plausibility transformation method + +Let $\Omega = \{\theta_1, \theta_2, \theta_3, \dots, \theta_N\}$ denote the discriminant framework. Given a corresponding BBA defined on $\Omega$ , denoted $m(\cdot)$ , the plausibility transformation of an element $\theta_i \in \Omega$ , denoted $P_{pt}$ , is defined as follows: + +$$ +P _ {p t} \left(\theta_ {i}\right) = \frac {M \left(\theta_ {i}\right)}{\sum_ {i = 1} ^ {N} M \left(\theta_ {i}\right)} \tag {25} +$$ + +where $M(\cdot)$ is the plausibility function: + +$$ +M \left(A _ {i}\right) = \sum_ {A _ {i} \cap A _ {j} \not \subset \emptyset | A _ {i}, A _ {j} \subseteq 2 ^ {\Omega}} m \left(A _ {j}\right). \tag {26} +$$ + +# A.2 More Proof Details for Properties + +When different quality functions are obtained, if they are simply combined according to Dempster's rule [48], the integration between different BBAs may be hindered because of conflicting information. For this reason, the idea of discounting techniques [46] is introduced. The high-order dynamic maximum mean difference (HODMMD) is defined as follows: + +$$ +\mathrm {H O D M M D} ^ {\tau} \left(\widehat {m} _ {1}, \widehat {m} _ {2}\right) = \left\| \sum_ {i = 1} ^ {N} \left(\widehat {m} _ {g _ {1}} ^ {\tau} \left(A _ {i}\right) - \widehat {m} _ {g _ {2}} ^ {\tau} \left(A _ {i}\right)\right) \right\| _ {\mathcal {H}} \tag {27} +$$ + +where $\mathcal{H}$ is the regenerated kernel Hilbert space and the kernel function is computed via a Gaussian kernel function, $A_{i} \in 2^{\Omega} \backslash \{\emptyset\}$ , and $\tau$ is the number of iterative splits. $\widehat{m}_g^\tau$ is $\tau$ order of $\widehat{m}_g$ . The HODMMD has a variety of properties. The properties of the HODMMD and the corresponding proofs are as follows. + +Property 1. When $\tau \to \infty$ , the HODMMD is equivalent to measuring the difference between the pignistic probability transformations with $\hat{m}_1$ and $\hat{m}_2$ of the maximum mean difference. + +$$ +\lim _ {\tau \rightarrow \infty} \mathrm {H O D M M D} ^ {\tau} \left(\widehat {m} _ {1}, \widehat {m} _ {2}\right) = \left\| \sum_ {i = 1} ^ {N} \left(\operatorname {B e t} P _ {1} \left(\theta_ {i}\right) - \operatorname {B e t} P _ {2} \left(\theta_ {i}\right)\right)\right\| _ {\mathcal {H}} \tag {28} +$$ + +Proof. We denote by $\mathrm{Bet}P_{t}$ the pignistic probability transformations of $\widehat{m}_1$ and $\widehat{m}_2$ , $t = 1, 2$ . That is, + +$$ +\operatorname {B e t} P _ {t} \left(\theta_ {i}\right) = \sum_ {\theta_ {i} \in A \mid A \in 2 ^ {\Omega}} \frac {\widehat {m} _ {t} (A)}{| A |} \tag {29} +$$ + +If $A_{i} = \theta_{i}$ , then the cardinality of $A_{i}$ is 1. Thus, when $\tau \to \infty$ and $|A_i| = 1$ , + +$$ +\lim _ {\tau \rightarrow \infty} \widehat {m} _ {g _ {t}} ^ {\tau} (A _ {i}) = \lim _ {\tau \rightarrow \infty} \frac {\sum_ {A _ {i} \subseteq A _ {j}} \left(\frac {| A _ {i} |}{| A _ {j} |}\right) ^ {\frac {1}{\tau}} \frac {\tau^ {| A _ {j} | - | A _ {i} |}}{(\tau + 1) ^ {| A _ {j} |} - \tau^ {| A _ {j} |}} \widehat {m} _ {t} (A _ {j})}{\sum_ {A _ {c} \subseteq 2 ^ {\Omega}} \sum_ {A _ {c} \subseteq A _ {j}} \left(\frac {| A _ {c} |}{| A _ {j} |}\right) ^ {\frac {1}{\tau}} \frac {\tau^ {| A _ {j} | - | A _ {c} |}}{(\tau + 1) ^ {| A _ {j} |} - \tau^ {| A _ {j} |}} \widehat {m} _ {t} (A _ {j})} \tag {30} +$$ + +Let + +$$ +\widehat {m} _ {g} ^ {\tau} \left(A _ {i}\right) = \sum_ {A _ {i} \subseteq A _ {j}} \left(\frac {\left| A _ {i} \right|}{\left| A _ {j} \right|}\right) ^ {\frac {1}{\tau}} \frac {\tau^ {\left| A _ {j} \right| - \left| A _ {i} \right|}}{(\tau + 1) ^ {\left| A _ {j} \right|} - \tau^ {\left| A _ {j} \right|}} \widehat {m} _ {g} ^ {\tau - 1} \left(A _ {j}\right) \tag {31} +$$ + +$$ +\begin{array}{l} \lim _ {\tau \rightarrow \infty} \widehat {m} _ {g _ {t}} ^ {\tau} (A _ {i}) = \lim _ {\tau \rightarrow \infty} \sum_ {A _ {i} \subseteq A _ {j}} \left(\frac {| A _ {i} |}{| A _ {j} |}\right) ^ {\frac {1}{\tau}} \frac {\tau^ {| A _ {j} | - | A _ {i} |}}{(\tau + 1) ^ {| A _ {j} |} - \tau^ {| A _ {j} |}} \widehat {m} _ {t} (A _ {j}) \\ = \lim _ {\tau \rightarrow \infty} \sum_ {A _ {i} \subseteq A _ {j}} \left(\frac {1}{\left| A _ {j} \right|}\right) ^ {\frac {1}{\tau}} \frac {\tau^ {\left| A _ {j} \right| - 1}}{(\tau + 1) ^ {\left| A _ {j} \right|} - \tau^ {\left| A _ {j} \right|}} \widehat {m} _ {t} (A _ {j}) \tag {32} \\ \end{array} +$$ + +where $A_{i}$ is a subset of $A_{j}$ . When $|A_{i}| = 1$ , there must be $|A_{j}| \geq 1$ ; hence, $|A_{j}| - 1 \geq 0$ and $\tau^{|A_j| - 1} > 0$ . The numerator and denominator are equally divisible by $\tau^{|A_j| - 1}$ + +$$ +\lim _ {\tau \rightarrow \infty} \widehat {m} _ {g _ {t}} ^ {\tau} \left(A _ {i}\right) = \lim _ {\tau \rightarrow \infty} \sum_ {A _ {i} \subseteq A _ {j}} \left(\frac {1}{\left| A _ {j} \right|}\right) ^ {\frac {1}{\tau}} \frac {1}{\tau \left(\frac {1}{\tau} + 1\right) ^ {\left| A _ {j} \right|} - \tau} \widehat {m} _ {t} \left(A _ {j}\right) \tag {33} +$$ + +Let $\rho = \frac{1}{\tau}$ , when $\tau \to \infty$ , have $\frac{1}{\tau} \to 0$ , which is $\rho \to 0$ . A Taylor expansion of $(\rho + 1)^{|A_{j}|}$ has + +$$ +(\rho + 1) ^ {| A _ {j} |} = 1 + | A _ {j} | \rho + o (\rho) \tag {34} +$$ + +where $o(\rho)$ is the infinitesimal of $\rho$ , have + +$$ +\begin{array}{l} \lim _ {\rho \rightarrow 0} \widehat {m} _ {g _ {t}} ^ {\rho} (A _ {i}) = \lim _ {\rho \rightarrow 0} \sum_ {A _ {i} \subseteq A _ {j}} \left(\frac {1}{| A _ {j} |}\right) ^ {\rho} \frac {\tau^ {| A _ {j} | - 1}}{(\tau + 1) ^ {| A _ {j} |} - \tau^ {| A _ {j} |}} \widehat {m} _ {t} (A _ {j}) \\ = \lim _ {\rho \rightarrow 0} \sum_ {A _ {i} \subseteq A _ {j}} \left(\frac {1}{| A _ {j} |}\right) ^ {\rho} \frac {1}{\frac {1}{\rho} (1 + | A _ {j} | ^ {\rho} + o (\rho)) - \frac {1}{\rho}} \widehat {m} _ {t} \left(A _ {j}\right) \\ = \sum_ {A _ {i} \subseteq A _ {j}} \frac {\widehat {m} _ {t} \left(A _ {j}\right)}{\left| A _ {j} \right|} \tag {35} \\ \end{array} +$$ + +Therefore, when $\tau \to \infty$ + +$$ +\begin{array}{l} \lim _ {\tau \rightarrow \infty} \widehat {m} _ {g _ {t}} ^ {\tau} (A _ {i}) = \frac {\sum_ {A _ {i} \subseteq A _ {j}} \frac {\widehat {m} _ {t} (A _ {j})}{| A _ {j} |}}{\sum_ {A _ {k} \in 2 ^ {\Omega}} \sum_ {A _ {i} \subseteq A _ {j}} \frac {\widehat {m} _ {t} (A _ {j})}{| A _ {j} |}} \\ = \sum_ {A _ {i} \subseteq A _ {j}} \frac {\widehat {m} _ {t} (A _ {j})}{| A _ {j} |} \\ = \sum_ {\theta_ {i} \in A | A \in 2 ^ {\Omega}} \frac {\widehat {m} _ {t} (A)}{| A |} \\ = \operatorname {B e t} P _ {t} \left(A _ {i}\right) \\ = \operatorname {B e t} P _ {t} \left(\theta_ {i}\right) \tag {36} \\ \end{array} +$$ + +Thus, + +$$ +\lim _ {\tau \rightarrow \infty} \mathrm {H O D M M D} ^ {\tau} (\widehat {m} _ {1}, \widehat {m} _ {2}) = \left\| \sum_ {i = 1} ^ {N} \left(\operatorname {B e t} P _ {1} \left(\theta_ {i}\right) - \operatorname {B e t} P _ {2} \left(\theta_ {i}\right)\right)\right\| _ {\mathcal {H}} \tag {37} +$$ + +![](images/bf30dd8a07206cf047bf6ed3f04c71f7ac1292e80a5354d679d72a481e91469b.jpg) + +Property 2. When $\widehat{m}_1$ and $\widehat{m}_2$ degenerate into probability distributions, that is, $U = (u_{1}, u_{2}, \ldots, u_{N})$ and $V = (v_{1}, v_{2}, \ldots, v_{N})$ , the proposed HODMMD degenerates into a maximum mean difference. + +$$ +\mathrm {H O D M M D} ^ {\tau} \left(\widehat {m} _ {1}, \widehat {m} _ {2}\right) = \mathrm {M M D} \left(\widehat {m} _ {1}, \widehat {m} _ {2}\right) \tag {38} +$$ + +Proof. When $\widehat{m}_1$ and $\widehat{m}_2$ are probability distribution, + +$$ +\forall A _ {i} \in 2 ^ {\Omega} \quad \widehat {m} _ {1} (A _ {i}) = u _ {i}, \quad \widehat {m} _ {2} (A _ {i}) = v _ {i} \tag {39} +$$ + +and it satisfies $\sum_{i=1}^{N} \widehat{m}_1(A_i) = 1$ , $\sum_{i=1}^{N} \widehat{m}_2(A_i) = 1$ . + +$$ +\begin{array}{l} \widehat {m} _ {g _ {t}} ^ {\tau} \left(A _ {i}\right) = \lim _ {\tau \rightarrow \infty} \frac {\sum_ {A _ {i} \subseteq A _ {j}} \left(\frac {\left| A _ {i} \right|}{\left| A _ {j} \right|}\right) ^ {\frac {1}{\tau}} \frac {\tau^ {\left| A _ {j} \right| - \left| A _ {i} \right|}}{(\tau + 1) ^ {\left| A _ {j} \right| - \tau^ {\left| A _ {j} \right|}}} \widehat {m} _ {t} \left(A _ {j}\right)}{\sum_ {A _ {c} \subseteq 2 ^ {\Omega}} \sum_ {A _ {c} \subseteq A _ {j}} \left(\frac {\left| A _ {c} \right|}{\left| A _ {j} \right|}\right) ^ {\frac {1}{\tau}} \frac {\tau^ {\left| A _ {j} \right| - \left| A _ {c} \right|}}{(\tau + 1) ^ {\left| A _ {j} \right| - \tau^ {\left| A _ {j} \right|}}} \widehat {m} _ {t} \left(A _ {j}\right)} \\ = \hat {m} _ {t} \left(A _ {i}\right) \tag {40} \\ \end{array} +$$ + +therefore, + +$$ +\begin{array}{l} \mathrm {H O D M M D} ^ {\tau} \left(\widehat {m} _ {1}, \widehat {m} _ {2}\right) = \left\| \sum_ {i = 1} ^ {N} \left(\widehat {m} _ {g _ {1}} ^ {\tau} \left(A _ {i}\right) - \widehat {m} _ {g _ {2}} ^ {\tau} \left(A _ {i}\right)\right) \right\| _ {\mathcal {H}} \\ = \left\| \sum_ {i = 1} ^ {N} \left(\widehat {m} _ {1} \left(A _ {i}\right) - \widehat {m} _ {2} \left(A _ {i}\right)\right) \right\| _ {\mathcal {H}} \\ = \left\| \sum_ {i = 1} ^ {N} \left(u _ {i} - v _ {i}\right) \right\| _ {\mathcal {H}} \\ = \operatorname {M M D} (U, V) \tag {41} \\ \end{array} +$$ + +![](images/a31cfae1da9b84ec0e00402581962393620e90c45e70e59d5fcf7d1db76d81fd.jpg) + +Property 3. $\mathrm{HODMMD}^{\tau}\left(\widehat{m}_{1},\widehat{m}_{2}\right)$ and $\mathrm{HODMMD}^{\tau}\left(\widehat{m}_{2},\widehat{m}_{1}\right)$ are equivalent. + +$$ +\mathrm {H O D M M D} ^ {\tau} \left(\widehat {m} _ {1}, \widehat {m} _ {2}\right) = \mathrm {H O D M M D} ^ {\tau} \left(\widehat {m} _ {2}, \widehat {m} _ {1}\right) \tag {42} +$$ + +Proof. + +$$ +\begin{array}{l} \mathrm {H O D M M D} ^ {\tau} \left(\widehat {m} _ {1}, \widehat {m} _ {2}\right) = \left\| \sum_ {i = 1} ^ {N} \left(\widehat {m} _ {g _ {1}} ^ {\tau} \left(A _ {i}\right) - \widehat {m} _ {g _ {2}} ^ {\tau} \left(A _ {i}\right)\right) \right\| _ {\mathcal {H}} \\ = \left\| \sum_ {i = 1} ^ {N} \left(\widehat {m} _ {1} \left(A _ {i}\right) - \widehat {m} _ {2} \left(A _ {i}\right)\right) \right\| _ {\mathcal {H}} \tag {43} \\ \end{array} +$$ + +$$ +\begin{array}{l} \mathrm {H O D M M D} ^ {\tau} \left(\widehat {m} _ {2}, \widehat {m} _ {1}\right) = \left\| \sum_ {i = 1} ^ {N} \left(\widehat {m} _ {g _ {2}} ^ {\tau} \left(A _ {i}\right) - \widehat {m} _ {g _ {1}} ^ {\tau} \left(A _ {i}\right)\right) \right\| _ {\mathcal {H}} \\ = \left\| \sum_ {i = 1} ^ {N} \left(\hat {m} _ {2} \left(A _ {i}\right) - \hat {m} _ {1} \left(A _ {i}\right)\right) \right\| _ {\mathcal {H}} \tag {44} \\ \end{array} +$$ + +Thus, + +$$ +\mathrm {H O D M M D} ^ {\tau} \left(\widehat {m} _ {1}, \widehat {m} _ {2}\right) = \mathrm {H O D M M D} ^ {\tau} \left(\widehat {m} _ {2}, \widehat {m} _ {1}\right) \tag {45} +$$ + +![](images/23e5db0a838b2912dd9672cbe7c1dcbff8df5f85c8df108b6e6366d27e21938a.jpg) + +Property 4. When $\widehat{m}_1 = \widehat{m}_2$ , the value of HODMMD is always equal to 0. + +$$ +\mathrm {H O D M M D} ^ {\tau} \left(\widehat {m} _ {1}, \widehat {m} _ {2}\right) = 0 \tag {46} +$$ + +Proof. When $\widehat{m}_1 = \widehat{m}_2$ + +$$ +\begin{array}{l} \mathrm {H O D M M D} ^ {\tau} \left(\widehat {m} _ {1}, \widehat {m} _ {2}\right) = \left\| \sum_ {i = 1} ^ {N} \left(\widehat {m} _ {g _ {2}} ^ {\tau} \left(A _ {i}\right) - \widehat {m} _ {g _ {1}} ^ {\tau} \left(A _ {i}\right)\right) \right\| _ {\mathcal {H}} \\ = \left\| \sum_ {i = 1} ^ {N} \left(\widehat {m} _ {2} \left(A _ {i}\right) - \widehat {m} _ {1} \left(A _ {i}\right)\right) \right\| _ {\mathcal {H}} \\ = \left\| \sum_ {i = 1} ^ {N} \left(\widehat {m} _ {1} \left(A _ {i}\right) - \widehat {m} _ {1} \left(A _ {i}\right)\right) \right\| _ {\mathcal {H}} \\ = 0 \tag {47} \\ \end{array} +$$ + +![](images/23537d74793a4da6dec66f0da7cfa8cac4aa97b1399142f1c5a48351f67f66d1.jpg) + +# A.3 Broader Impacts + +This paper proposes a dynamic learning strategy based on nonuniform splitting mechanism and Hilbert space mapping to promote real-world applications. We apply the proposed method in pattern classification, image classification, and low-light image enhancement encountered in real life, which encourages research on their synergistic combination in real life. In addition, we are the first to introduce DSE theory to low-light image enhancement and achieve effective performance enhancement for low-light image enhancement tasks. This provides a new idea for the low-light image enhancement task. As far as this paper is concerned, we believe that the proposed method does not have any significant negative impact. + +# A.4 Additional details regarding the experiment + +# A.4.1 Dataset and experimental settings + +To rigorously evaluate the performance of the proposed method, we conducted extensive experiments and applied the proposed algorithm in three experiments. For pattern classification, experiments were conducted on the Iris [18], Heart [27], Hepatitis [1], Parkinsons [36], Australian [42], Segment [2], and Connectionist Bench (CBench)[16] datasets, and the details of these datasets are shown in Table 4. In this evaluation phase, we compare the proposed method with a class of classical classifiers: a Bayes theorem-based classifier (NaB) [23], a k-nearest neighbor classification method (kNN) [12], a decision tree algorithm (REPTree) [19], a support vector machine classifier (SVM) [5], an SVM method with a radial basis function (SVM-RBF) [5], a multilayer perceptron method (MLP) [4], and a radial basis function network (RBFN) [8]. Another class is evidence theory-based classifiers: a DS theory-based kNN method (kNN-DST) [15], a data probability distribution-based method (NDC) [60], an evidence calibration method (EvC) [61], and a generalized divergence-based decision-making method (DMA) [58]. + +Table 4: Dataset information. + +
DatasetInstancesClassFeaturesMissing Values
Iris[18]15034No
Heart[27]270213No
Hepatitis[1]155219Yes
Parkinsons[36]197222No
Australian[42]690214Yes
Segment[2]2310719No
CBench[16]9901110No
+ +For image classification, experiments were conducted on the publicly available datasets CIFAR-10[31] and CIFAR-100[31], and the details of these datasets are shown in Table 5. Furthermore, performance comparisons were made with current state-of-the-art methods, including DIR-Net [40], + +MST[51], ReSTE[54], ADMM[9], SML[14], UDSP[20], BiPer [50], TAB[28], APL[70], ESNN[47], Dspike[34], GLIF[66], Diet-SNN[43], PASNN[17], MPBN [22], MS-ResNet[24], and BKDSNN[63]. + +Table 5: Dataset information. + +
DatasetTrain setTest setClass
CIFAR-10 [31]500001000010
CIFAR-100 [31]5000010000100
+ +Table 6: Dataset information. + +
DatasetTrain setTest set
LOL-v1 [53]48515
LOL-v2-real [65]689100
LOL-v2-syn [65]900100
SID [7]2099598
SMID [6]208095046
+ +In addition, we validated the effectiveness of the proposed method on a low-light image enhancement task and tested it on the publicly available datasets LOL-v1 [53], LOL-v2-real [65] and LOL-v2-syn [65]. The allocation of the datasets and related information is shown in Table 6. In this evaluation phase, our method was compared with several state-of-the-art low-light image enhancement methods, including MIRNet[68], FIDE[59], ZeroDCE[21], Sparse[53], DRBN[64], RUAS[37], ZeroDCE++[33], SCI[39] and Restormer[29]. In addition, to realistically demonstrate the superiority of our method, methods from recent years were selected as the baseline network, and only targeted training strategies were added to the original methods, including SNR [62], LLFlow-L [52], LLFlow-S [52] and Retinexformer [3]. For fairness, the parameter settings were kept the same as those in the original method. All the experiments were run on NVIDIA RTX 3090 GPUs. + +# A.4.2 Extended ablation studies + +To verify the indispensability of the ADPT and HODMMD modules, we perform ablation studies on the image classification task of the CIFAR-10 dataset with the architecture of ResNet-18, and the results are shown in Table 7. + +Table 7: Quantitative comparison of ablation study. + +
DatasetMethodsAccuracy
w/o ADPT94.56%
w/o HODMMD93.55%
Replace HODMMD with Euclidean distance94.31%
CIFAR-10 [31]Building BBAs using evidential neural networks [48]94.57%
Fusion via Dempster's combination rule93.55%
Replacing HODMMD with Euclidean distance94.31%
Ours95.61%
+ +The above experiments indicate that the model performance decreases to $94.56\%$ when the ADPT module is removed. This shows that ADPT can use the prior information of the data to better address uncertainty information through nonuniform splitting, thereby significantly improving the classification accuracy. The accuracy of the model decreased by $2.06\%$ after the HODMMD module + +was removed. The accuracy of the model decreased by $1.3\%$ when the Euclidean distance was used to replace the HODMMD. This finding indicates that our method can more accurately capture the nonlinear difference between complex data distributions, thereby more effectively assessing the conflict between evidence and more reliable fusion decisions. + +# A.4.3 Hyperparameter sensitivity experiment + +To better compare model performance, we have supplemented the experiments with analyses of key hyperparameter sensitivity and model complexity. The numerical experiments are as follows. + +$$ +\widehat {m _ {1}}: \widehat {m _ {1}} (\{\theta_ {1} \}) = \frac {3}{2 0}, \widehat {m _ {1}} (\{\theta_ {2} \}) = \frac {1}{4}, \widehat {m _ {1}} (\{\theta_ {1}, \theta_ {2} \}) = \frac {3}{1 0}, \widehat {m _ {1}} (\{\theta_ {3} \}) = \frac {1}{1 0}, \widehat {m _ {1}} (\{\theta_ {1}, \theta_ {3} \}) = \frac {1}{5}; +$$ + +$$ +\widehat {m _ {2}}: \widehat {m _ {2}} (\{\theta_ {1} \}) = \frac {1}{4}, \widehat {m _ {2}} (\{\theta_ {2} \}) = \frac {3}{2 0}, \widehat {m _ {2}} (\{\theta_ {1}, \theta_ {2} \}) = \frac {1}{5}, \widehat {m _ {2}} (\{\theta_ {3} \}) = \frac {1}{5}, \widehat {m _ {2}} (\{\theta_ {1}, \theta_ {3} \}) = \frac {1}{5}. +$$ + +By fixing the kernel bandwidth $\sigma = 0.5$ , we tested the variation in HODMMD in the range of [1,6], and the variation values of the mass function and HODMMD were plotted as a line graph. As $\tau$ increases, the value of the mass function of the uncertainty information gradually decreases is assigned to the mass function represents a single category, and the HODMMD value also decreases. In addition, $\tau = 1$ was fixed, and the variation of HODMMD in the range of [0.05, 10] for $\sigma$ was tested. When $\sigma$ is between [0.05, 0.1], the HODMMD rapidly decreases, and when $\sigma$ is between [0.1, 2], it tends to increase and eventually stabilizes. + +Second, in the classification and LLIE tasks, $\tau$ controls the degree of nonuniform splitting. We tested the effect on task accuracy when $\tau$ was 1,2,3,4,5,6. The results show that when $\tau = 5$ , the model performance remains stable and optimal. Similarly, for $\sigma$ , we took the interval as 0.1 to test the effect on task accuracy between [0.1, 2]. The results show that when $\sigma = 0.5$ , the model performance remains stable and optimal. In summary, we fix $\tau = 5$ and $\sigma = 0.5$ . + +# A.4.4 Analysis of model complexity + +We have added parameter (M), FLOPs (G) and FPS comparisons with the baseline method on the low-light image enhancement task. The results are shown in Table 8. + +Table 8: Efficiency Comparison of different methods. + +
MethodsParam(M)FLOPs(G)FPS
SNR4.0126.351.175
SNR-TTS5.0737.561.172
LLFlow-L37.682870.813
LLFlow-L-TTS38.74298.210.812
LLFlow-S4.9737.860.943
LLFlow-S-TTS6.0349.070.942
Retinexformer1.6115.571.724
Retinexformer-TTS2.6726.781.718
+ +The results indicate that the complexity and time of the model both increase after the introduction of the TTS. However, considering the improvement in model performance, this increase is within the acceptable range, which proves the actual application efficiency of our method. + +# A.4.5 Large dataset experiment + +The experiments on the SID and SMID datasets to verify the effectiveness of the TTS by using the Retinexformer as the baseline networks. The results are shown in Table 9. + +# A.5 Discussion of Our Method with Machine Learning Methods and Deep Learning Methods + +Although different methods are currently available to address decision-making problems, our method has the following advantages over machine learning methods and deep learning methods. + +Table 9: Quantitative comparison of SID and SMID datasets. + +
MethodsSMID[6]SID[7]
PSNRSSIMPSNRSSIM
Retinexformer[3] ICCV'202329.150.81524.440.680
Retinexformer-TTS29.23(+0.08)0.816(+0.001)24.62(+0.18)0.682(+0.002)
+ +First, our approach has a better ability to handle uncertainty. Machine learning and deep learning usually produce only a single probability estimate and make predictions under the assumption of mapping relationships. This approach does not model uncertainty well in the face of conflicting or uncertain information. However, our method directly models uncertainty through BBA, which allows assigning confidence to composite propositions. Moreover, it can be dynamically adjusted according to the level of importance between different pieces of information, capturing the nonspecificity of the information and quantifying the discrepancies between pieces of evidence in a more flexible way. + +Second, our method can fuse information from multiple sources in a more rational way. Machine learning and deep learning approaches tend to use network layers for feature learning and prediction and lack interpretability for information fusion decisions. However, our method can understand the reasons behind the decisions well. For example, the proposed method maps data into Hilbert space for computation, which is more responsive to the differences in the true distributions of complex data, reduces conflicts between different types of information, and is highly interpretable. + +Third, our approach remains applicable in the presence of limited data. The training process of deep learning suffers from overfitting on small or unrepresentative data and relies on training complex models. However, our method's splitting mechanism, which is based on a priori information, can be more flexible in dealing with uncertainty and conflicting information and can work effectively with limited data. + +We propose a dynamic learning strategy based on nonuniform splitting mechanism and Hilbert space mapping enhances the interpretability of the decision. This will promote the application of deep learning technology in a wider range of fields. In addition to classification and low-light image enhancement, our ideas can be applied in all uncertainty tasks, for example, image segmentation (uncertainty of the object boundary) and automatic driving (uncertainty in the fusion of multisensor information). Owing to time constraints, this work requires a large amount of computing resources and time. We will continue to explore the performance of this method in other fields. \ No newline at end of file diff --git a/NeurIPS/2025/A Dynamic Learning Strategy for Dempster-Shafer Theory with Applications in Classification and Enhancement/images.zip b/NeurIPS/2025/A Dynamic Learning Strategy for Dempster-Shafer Theory with Applications in Classification and Enhancement/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..14dd3aecc263838c6a6b98e700c9ef024b57fdf1 --- /dev/null +++ b/NeurIPS/2025/A Dynamic Learning Strategy for Dempster-Shafer Theory with Applications in Classification and Enhancement/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d8483e7177fa29d390e0d9b78293780cc1565b6b497049bcfde4d8c5a533b49a +size 1218162 diff --git a/NeurIPS/2025/A Dynamic Learning Strategy for Dempster-Shafer Theory with Applications in Classification and Enhancement/layout.json b/NeurIPS/2025/A Dynamic Learning Strategy for Dempster-Shafer Theory with Applications in Classification and Enhancement/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..eba5e4d087209b9b12f0157557716fd700ea046e --- /dev/null +++ b/NeurIPS/2025/A Dynamic Learning Strategy for Dempster-Shafer Theory with Applications in Classification and Enhancement/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:41f29f29ea62bdefc6c73102511d31092b2cb39a99e8832d1ed964699308588b +size 928597 diff --git a/NeurIPS/2025/A Fair Federated Learning Method for Handling Client Participation Probability Inconsistencies in Heterogeneous Environments/ddd8a355-d667-44a1-8469-465bfd17ba16_content_list.json b/NeurIPS/2025/A Fair Federated Learning Method for Handling Client Participation Probability Inconsistencies in Heterogeneous Environments/ddd8a355-d667-44a1-8469-465bfd17ba16_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..028577dfbbe0295713840d5f7e8eeb56763ffecd --- /dev/null +++ b/NeurIPS/2025/A Fair Federated Learning Method for Handling Client Participation Probability Inconsistencies in Heterogeneous Environments/ddd8a355-d667-44a1-8469-465bfd17ba16_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b54ad9839dfb5ffa7c949010c09f285dcad9a8d9978b1de43ca6fdde2922bec0 +size 172199 diff --git a/NeurIPS/2025/A Fair Federated Learning Method for Handling Client Participation Probability Inconsistencies in Heterogeneous Environments/ddd8a355-d667-44a1-8469-465bfd17ba16_model.json b/NeurIPS/2025/A Fair Federated Learning Method for Handling Client Participation Probability Inconsistencies in Heterogeneous Environments/ddd8a355-d667-44a1-8469-465bfd17ba16_model.json new file mode 100644 index 0000000000000000000000000000000000000000..e74fe286d1da25d6545427208eb8ef768cba645d --- /dev/null +++ b/NeurIPS/2025/A Fair Federated Learning Method for Handling Client Participation Probability Inconsistencies in Heterogeneous Environments/ddd8a355-d667-44a1-8469-465bfd17ba16_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:509128a0a0660b07eab0df179388164dfa686600c1c30f94e5a3505d73e96457 +size 223779 diff --git a/NeurIPS/2025/A Fair Federated Learning Method for Handling Client Participation Probability Inconsistencies in Heterogeneous Environments/ddd8a355-d667-44a1-8469-465bfd17ba16_origin.pdf b/NeurIPS/2025/A Fair Federated Learning Method for Handling Client Participation Probability Inconsistencies in Heterogeneous Environments/ddd8a355-d667-44a1-8469-465bfd17ba16_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..edb0d201a41fd12d15641a5e22069c54fb9f1493 --- /dev/null +++ b/NeurIPS/2025/A Fair Federated Learning Method for Handling Client Participation Probability Inconsistencies in Heterogeneous Environments/ddd8a355-d667-44a1-8469-465bfd17ba16_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:154e03e41ccd165625e76417770f1deee3da2f72f3f0663f6bdea97b7bcb3de2 +size 1091461 diff --git a/NeurIPS/2025/A Fair Federated Learning Method for Handling Client Participation Probability Inconsistencies in Heterogeneous Environments/full.md b/NeurIPS/2025/A Fair Federated Learning Method for Handling Client Participation Probability Inconsistencies in Heterogeneous Environments/full.md new file mode 100644 index 0000000000000000000000000000000000000000..b1116d73cac3fed1b252ca560b7d00d6b986e8be --- /dev/null +++ b/NeurIPS/2025/A Fair Federated Learning Method for Handling Client Participation Probability Inconsistencies in Heterogeneous Environments/full.md @@ -0,0 +1,753 @@ +# A Fair Federated Learning Method for Handling Client Participation Probability Inconsistencies in Heterogeneous Environments + +Siyuan Wu $^{1}$ , Yongzhe Jia $^{1}$ , Haolong Xiang $^{2}$ , Xiaolong Xu $^{2}$ , Xuyun Zhang $^{3}$ , Lianyong Qi $^{4}$ , Wanchun Dou $^{1,*}$ + +1 State Key Laboratory for Novel Software Technology, Nanjing University, China + +$^{2}$ School of Computer and Software, + +Nanjing University of Information Science and Technology, China + +$^{3}$ School of Computing, Macquarie University, Australia + +4College of Computer Science and Technology, China University of Petroleum (East China), China + +# Abstract + +Federated learning (FL) is a distributed machine learning paradigm that enables multiple clients to collaboratively train a shared model without exposing their raw data. However, existing FL research has primarily focused on optimizing learning performance based on the assumption of uniform client participation, with few studies delving into performance fairness under inconsistent client participation, particularly in model-heterogeneous FL environments. In view of this challenge, we propose PHP-FL, a novel model-heterogeneous FL method that explicitly addresses scenarios with varying client participation probabilities to enhance both model accuracy and performance fairness. Specifically, we introduce a Dual-End Aligned ensemble Learning (DEAL) module, where small auxiliary models on clients are used for dual-end knowledge alignment and local ensemble learning, effectively tackling model heterogeneity without a public dataset. Furthermore, to mitigate update conflicts caused by inconsistent participation probabilities, we propose an Importance-driven Selective Parameter Update (ISPU) module, which accurately updates critical local parameters based on training progress. Finally, we implement PHP-FL on a lightweight FL platform with heterogeneous clients across three different client participation patterns. Extensive experiments under heterogeneous settings and diverse client participation patterns demonstrate that PHP-FL achieves state-of-the-art performance in both accuracy and fairness. Our code is available at: https://github.com/Siyuan01/PHP-FL-main. + +# 1 Introduction + +Federated Learning (FL) has emerged as a promising paradigm for enabling decentralized model training across multiple clients without directly sharing their private data [1, 2]. By collaboratively learning a global model while keeping data localized, FL offers strong privacy guarantees and broad applicability across various domains such as mobile devices, healthcare, and finance [3-5]. + +Despite significant progress, traditional FL methods still face two critical challenges: I) Model heterogeneity. Traditional FL assumes that all clients share an identical local model architecture, which is often impractical in real-world due to the diversity in client capabilities. To address this, Model-Heterogeneous Federated Learning (MH-FL) has emerged as a promising research paradigm [2, 6–9], which allows each client to maintain a personalized model tailored to its own resource constraints + +![](images/e550674a381f79be586b8cc3a7a6c4d7824feffffd5865079952c6b5ce62e10c.jpg) +(a) Uniform pattern + +![](images/3b24b78dce303726aafcf611ec944b90fe3d1be68fce3053c13e6837b65cace4.jpg) +(b) Normal pattern +Figure 1: Left three: three client participation probability patterns (average value is approximately $50\%$ ). Right: client accuracy under patterns (a)-(c). Refer to Section 5.1 for experimental details. + +![](images/729d07b8b3e83ad998865d0856073dd4f55b61bd29749c86e1bb5eff8267dcf8.jpg) +(c) Linear pattern + +![](images/0a99605a5eefaf8bfdd4bcc59ef7bd6bcfa2015d75ca765cbc0a728a4143aeb2.jpg) +(d) Client accuracy + +or task requirements. However, most existing MH-FL methods [10-13] primarily aim to ensure compatibility between diverse model architectures, yet overlook the fairness issues posed arising from inconsistent client participation probabilities in practical deployments. This oversight can result in biased models that favor frequently participating clients, ultimately compromising the robustness and fairness of the FL system. II) Unfairness caused by inconsistent client participation. Most existing FL research [1, 9, 12-16] implicitly assume a uniform client participation pattern, where all clients are equally likely to participate in each training round, as illustrated in Figure 1a. In practical deployments, clients often face heterogeneous conditions, such as intermittent connectivity, fluctuating network bandwidth, and network coverage of base stations [17-20]. These internal and external factors lead to client unavailability and result in non-uniform participation probabilities, as exemplified in Figure 1b and 1c, which may critically degrade both the overall accuracy and fairness of the FL system. Figure 1d compares the final client accuracy distributions between FML [21] (a representative MH-FL method) and our method PHP-FL under three participation patterns. While FML suffers significant performance degradation in the normal and linear patterns compared to the uniform participation, PHP-FL maintains stable accuracy with only marginal drops. Furthermore, PHP-FL demonstrates tighter accuracy distributions, indicating better performance fairness. Although some studies [22-25] improve overall system performance by proactively selecting high-availability clients and discarding less efficient ones, such strategies often compromise fairness across clients. Furthermore, numerous fair federated learning methods aim to enhance performance fairness through personalized model [14, 26] or weight recalibration [27, 28]. Nevertheless, they typically operate under the assumption of uniform client participation, failing to address the challenge of inconsistent client availability. Despite its importance, this issue has received limited attention in the literature, particularly in the context of MH-FL, where the interplay between model heterogeneity and participation imbalance exacerbates the learning challenge. + +To address these challenges, we propose PHP-FL, a fair federated learning method designed to address scenarios with varying client participation probabilities in model-heterogeneous environments. Specifically, to effectively tackle model heterogeneity without relying on public datasets, we introduce a Dual-End Aligned Ensemble Learning (DEAL) module, which leverages lightweight auxiliary models on clients to align heterogeneous local models and enable ensemble learning to improve the performance of local tasks. Furthermore, to mitigate the adverse effects of update conflicts that caused by inconsistent client participation probabilities, we propose a Importance-driven Selective Parameter Update (ISPU) module. The ISPU module adaptively updates only the most critical task-relevant parameters based on training progress, allowing clients with different participation frequencies to selectively absorb varying ratios of global knowledge. This design helps reduce gradient conflicts and enhancing fairness. Our main contributions are as follows: + +- To the best of our knowledge, this is the first work to explicitly address performance unfairness caused by inconsistent client participation probabilities in practical FL systems with heterogeneous local models. +- We propose PHP-FL, a novel model-heterogeneous federated learning method designed to address inconsistent client participation probabilities, aiming to jointly improve both overall accuracy and performance fairness across clients. +- We evaluate PHP-FL through extensive experiments on a lightweight FL platform simulating multiple realistic participation patterns. Empirical results on the Fashion-MNIST and CIFAR-10 datasets demonstrate its state-of-the-art performance, while the ablation study further validates the effectiveness of each proposed module. + +# 2 Related Works + +Model-Heterogeneous Federated Learning. Model-Heterogeneous Federated Learning (MHFL) has emerged as a promising research direction [11, 6, 8, 9, 2, 13]. Existing work in this area can be broadly categorized into knowledge distillation-based (KD) methods, representation alignment methods, and partial model sharing methods. KD is one of the most widely adopted techniques in MH-FL. The studies in [29-31] enable clients with different architectures to distill knowledge through a shared or public dataset. However, such reliance on public data limits applicability in privacy-sensitive settings. Another popular line of work, like [15, 12, 16], focuses on representation alignment, which aligns feature representations or prototypes rather than raw model parameters, allowing clients to maintain model diversity while contributing to a shared learning objective. Furthermore, some studies [7, 32-34] adopt partial model sharing strategies, where clients share only specific model components (e.g., a shared backbone or predictor) while keeping the other parts distinct, thereby enabling partial compatibility across models. However, few MH-FL methods explicitly address fairness for clients under inconsistent participation probabilities, a critical requirement for equitable and robust deployment. + +Fairness in Federated Learning. Existing research on fairness in federated learning has primarily focused on three key dimensions: (1) Contribution Fairness [4, 35, 36], which involves evaluating each client's contribution to the global model to guide equitable benefit distribution, often using techniques like Shapley value or influence functions [37, 38]; (2) Model Fairness [39, 40], which addresses inherent biases in model predictions concerning sensitive attributes, thereby promoting fairness at the prediction level; and (3) Performance Fairness [14, 20, 27, 41, 28], which aims to ensure uniform model performance across clients, typically by minimizing the variance or standard deviation of test accuracy among clients. As prior studies [42, 43] demonstrate, these fairness metrics often conflict, making it infeasible for a model to simultaneously achieve optimal performance across all dimensions. Our work, therefore, specifically targets performance fairness, aiming to ensure uniform performance across clients while concurrently optimizing overall performance under inconsistent participation probabilities in MH-FL. While another line of research [20, 44, 45] addresses client unavailability by primarily focusing on maintaining the average performance across clients, they do not explicitly ensure performance-level fairness. Furthermore, these methods are typically designed for homogeneous settings and face significant challenges when generalizing to heterogeneous federated learning environments, where client capabilities vary substantially. + +# 3 Preliminaries + +The Global Objective of Federated Learning. Following typical federated learning [1, 26] settings, we consider a set of $K$ clients (index by $i$ ) with local datasets $\{D_1, D_2, \dots, D_K\}$ , where $D_i = \{(x_j, y_j)\}_{j=1}^{n_i}$ and $n_i = |D_i|$ . In a heterogeneous FL environment, each client $i$ maintains a unique model $\boldsymbol{w}_i$ , parameterized by $\theta_i \in \mathbb{R}^{d_i}$ , with dimensional heterogeneity ( $d_i \neq d_j$ ) arising from hardware constraints (e.g., compute/memory limits) or personalized model specialization. This implies $\dim(\theta_i) \neq \dim(\theta_j)$ , $\exists i \neq j \in [K]$ . The global objective function can be expressed as: + +$$ +\min _ {\left\{\boldsymbol {w} _ {i} \right\} _ {i = 1} ^ {K}} F \left(\left\{\boldsymbol {w} _ {i} \right\}\right) _ {i = 1} ^ {K} = \sum_ {i = 1} ^ {K} p _ {i} F _ {i} \left(\boldsymbol {w} _ {i}\right), \quad \sum_ {i = 1} ^ {K} p _ {i} = 1, \tag {1} +$$ + +where $p_i$ is the weight of client $i$ , $\{\pmb{w}_i\}_{i=1}^K$ represents the set of the client's local models, and $F_i(\pmb{w}_i) = \frac{1}{n_i} \sum_{j=1}^{n_i} \mathcal{L}(\pmb{w}_i(\theta_i; x_j), y_j)$ is the local objective for client $i$ with loss function $\mathcal{L}$ . + +Inconsistent Client Participation Probability. To simulate realistic client availability in federated learning, we consider three types of client participation probability patterns. Let $p_{i,t}$ denote the probability that client $i$ actively participates in communication round $t$ : + +Definition 1 (Uniform Pattern) All clients share an identical and fixed probability $a \in (0,1]$ of participating in each round, i.e., $p_{i,t} = a$ , $\forall i \in \{1,2,\dots,n\}$ , $\forall t$ . + +Definition 2 (Normal Pattern) The participation probabilities are drawn from a truncated normal distribution to simulate natural heterogeneity: $p_{i,t} \sim \mathcal{N}(\mu, \sigma)$ , and $p_{i,t}$ is clipped to $(0,1]$ . + +Definition 3 (Linear Pattern) Client participation probabilities are distributed according to an increasing arithmetic sequence: $p_{i,t} = a + (i - 1)d$ , $i = 1,2,\ldots,n$ , $\forall t$ , where $a$ is the first term and $d$ is the common difference. In this pattern, the initial sequence satisfies $0 < p_{1,t} < p_{2,t} < \dots < p_{K,t} \leq 1$ for round $t$ . This sequence $\{p_{i,t}\}_{i=1}^{K}$ is then randomly shuffled prior to use to eliminate any inherent ordering bias among clients. It models systematic heterogeneity such as time-varying connectivity or device capacity. + +Note that $p_{i,t}$ is independent of the history and other clients. The three distinct patterns considered enable a comprehensive analysis of how heterogeneous client participation affects both fairness and overall performance in federated learning. + +Design Goals. In this paper, we aim to design a MH-FL method under inconsistent client participation probabilities that not only optimizes the average performance across all clients but also enhances performance fairness. Formally, let $a_{i}(i = 1,\dots ,K)$ represent the test accuracy on the $i$ -th client's local test dataset. The Accuracy Metric (AM) is defined as: $AM = \frac{1}{K}\sum_{k = 1}^{K}a_{k}$ . The Fairness Metric (FM) is defined as: $FM = \mathrm{Std}(a_1,\ldots ,a_K)$ , where $\mathrm{Std}(\cdot)$ denotes the standard deviation. To this end, our approach seeks to maximize the average local accuracy (AM) while minimizing the performance disparity (FM), ensuring both high overall performance and fair performance distribution across clients. + +# 4 Methodology + +# 4.1 Overview of PHP-FL + +Our method consists of two key modules: dual-end aligned ensemble learning (DEAL in Section 4.2), and importance-driven selective parameter update (ISPU in Section 4.3). The DEAL module employs a small homogeneous auxiliary model to perform bidirectional representation-logit alignment between local and auxiliary models, which resolves model heterogeneity and enhances overall performance via ensemble learning. To handle inconsistent client participation, the ISPU module selectively updates task-relevant critical parameters using an importance-based masking mechanism. This approach applies larger updates to stragglers to accelerate overall convergence, while reducing updates for frequent participants to prevent adverse effects from the stragglers. This ensures efficient knowledge fusion and fair parameter evolution across clients. + +As shown in Figure 2, the training process in each communication round $t$ of PHP-FL can be summarized as follows: The server first computes the client-specific auxiliary model $\mathcal{G}^{t - 1} \odot M_i^h$ for the active client $i \in \mathcal{A}_t$ by pruning non-essential parameters using historical binary mask $M_i^h$ . Then active clients initialize the personalized auxiliary models $\hat{\pmb{g}}_i^{t - 1}$ . PHP-FL decomposes both the local model $\pmb{w}_i$ and the auxiliary model $\pmb{g}_i$ into a backbone and a predictor, which are used for representation extraction and soft prediction, respectively. At the beginning of local training, the DEAL module first optimizes the ensemble weights $\lambda_{i}$ on the adaptation set $D_{i}^{a}$ , while $w_{i}^{t}$ and $\hat{g}_i^{t - 1}$ are frozen. Local training then proceeds using the customized loss $\mathcal{L}_{\pmb{w}}$ , which comprises two components: (1) The dual-end alignment loss $\mathcal{L}_{\mathrm{DEAL}}$ , enabling bidirectional knowledge alignment through data-free distillation and representation matching. (2) The ensemble learning loss $\mathcal{L}_{\mathrm{ENS}}$ , which further enhances overall model performance. Following the local training, the ISPU module calculates an update ratio $\alpha_{i}^{t}$ based on the client's total training rounds. It then estimates the top- $\alpha_{i}^{t}$ important parameters of $g_{i}^{t}$ using the $\ell_1$ -norm and generates a binary mask $M_i^t$ . Finally, each active client uploads both its updated local auxiliary model $g_{i}^{t}$ and the binary mask $M_i^t$ to the server. The server updates the historical mask $M_i^h \in M_{\mathrm{hist}}$ for each active client $i$ by replacing its entry with the newly received $M_i^t$ . Then the received auxiliary models are aggregated via a simple averaging technique to obtain the global auxiliary model $G^{t}$ for round $t + 1$ . + +# 4.2 Dual-End Aligned Ensemble Learning + +To address system heterogeneity without relying on public datasets, previous works such as [12, 15] decompose the model $w^3$ (parameterized by $\theta$ ) into a backbone $\mathbf{w}_b$ and a predictor $\mathbf{w}_p$ , and perform + +![](images/99144833d53f3da4ef262a924c35600fb97a179e85c3480309390cd4848c5e92.jpg) +Figure 2: The overview of PHP-FL. + +aggregation based on intermediate representations $z = w_{b}(\theta_{b};x_{j})$ produced by the backbone. However, it's challenging to classify representations generated by heterogeneous backbones for local predictors. Inspired by the spirit of the mutual learning, some works [46, 47] leverage logits-level knowledge distillation, where each client co-trains a heterogeneous local model $\pmb{w}$ and a lightweight homogeneous global model $\pmb{g}$ (parameterized by $\phi$ ) by aligning only the final outputs of the predictors. Unfortunately, this limited alignment fails to facilitate meaningful knowledge transfer to the backbone component, resulting in suboptimal representation learning. To address these limitations, we propose a DEAL module. In DEAL, both the backbone and predictor components between local and auxiliary models are explicitly aligned. This dual-end alignment ensures effective and rapid fusion of local and global knowledge. To this end, we design the following loss function: + +$$ +\mathcal {L} _ {D E A L} ^ {\boldsymbol {w}} = \frac {1}{| D _ {i} |} \sum_ {j \in D _ {i}} [ \underbrace {\mathcal {L} _ {\mathrm {M M D}} (\boldsymbol {w} _ {b} \left(\theta_ {b} ; x _ {j}\right) , \boldsymbol {g} _ {b} \left(\phi_ {b} ; x _ {j}\right))} _ {\text {B a c k b o n e A l i g n m e n t}} + \underbrace {\mathcal {D} _ {\mathrm {K L}} \left(\boldsymbol {w} \left(\theta ; x _ {j}\right) \| \boldsymbol {g} \left(\phi ; x _ {j}\right)\right)} _ {\text {P r e d i c t o r A l i g n m e n t}} ], \tag {2} +$$ + +where $\mathcal{D}_{\mathrm{KL}}$ is Kullback-Leibler (KL) divergence. The Maximum Mean Discrepancy (MMD) loss $\mathcal{L}_{\mathrm{MMD}}$ between two sets of representations $\mathbf{z}_1 \in \mathbb{R}^{n \times d_1}$ and $\mathbf{z}_2 \in \mathbb{R}^{m \times d_2}$ using a Gaussian radial basis function (RBF) kernel is computed as: + +$$ +\begin{array}{l} \mathcal {L} _ {\mathrm {M M D}} (\mathbf {z} _ {1}, \mathbf {z} _ {2}) = \frac {1}{n ^ {2}} \sum_ {i, j = 1} ^ {n} k (f (\mathbf {z} _ {1} ^ {(i)}), f (\mathbf {z} _ {1} ^ {(j)})) + \frac {1}{m ^ {2}} \sum_ {i, j = 1} ^ {m} k (h (\mathbf {z} _ {2} ^ {(i)}), h (\mathbf {z} _ {2} ^ {(j)})) \\ - \frac {2}{n m} \sum_ {i = 1} ^ {n} \sum_ {j = 1} ^ {m} k \left(f \left(\mathbf {z} _ {1} ^ {(i)}\right), h \left(\mathbf {z} _ {2} ^ {(j)}\right)\right), \\ \end{array} +$$ + +where $f: \mathbb{R}^{d_1} \to \mathbb{R}^d$ and $h: \mathbb{R}^{d_2} \to \mathbb{R}^d$ represent customizable projection functions designed to standardize feature dimensions, enabling cross-architecture feature alignment between the local model $\pmb{w}$ and global model $\pmb{g}$ when their structures are heterogeneous. The Gaussian RBF kernel $k(\mathbf{z},\mathbf{z}') = \exp \left(-\gamma \| \mathbf{z} - \mathbf{z}'\|^2\right)$ measures similarity between representations, with $\gamma = 1 / (2\sigma^2)$ controlling the kernel bandwidth. This non-parametric metric effectively captures the distance between the distributions of $\mathbf{z}_1$ and $\mathbf{z}_2$ in the Reproducing Kernel Hilbert space (RKHS). MMD is able to effectively align feature distributions by comparing global statistics via kernel-based embeddings, handles non-IID data robustly, and enables stable optimization [48]. + +Next, to fully leverage the classification capabilities of both local models and global auxiliary models, we adopt model ensembling [49, 47, 13] to enhance performance on local tasks: + +$$ +\mathcal {L} _ {E N S} ^ {\boldsymbol {w}} = \frac {1}{| D _ {i} |} \sum_ {j \in D _ {i}} \left[ \mathcal {L} _ {\mathrm {C E}} (\boldsymbol {w} (\theta ; x _ {j}), y _ {j}) + \mathcal {L} _ {\mathrm {C E}} (\lambda \boldsymbol {w} (\theta ; x _ {j}) + (1 - \lambda) \boldsymbol {g} (\phi ; x _ {j}), y _ {j}) \right], \tag {4} +$$ + +where $\mathcal{L}_{\mathrm{CE}}$ denotes the cross-entropy loss between the predicted label and the ground-truth label. Furthermore, to address the challenges posed by the potential heterogeneous system capabilities between each local model $\pmb{w}$ and the global model $\pmb{g}$ , we set $\lambda$ as a trainable parameter for each client and randomly hold out a tiny adaptability set $D_{i}^{a}$ from the training set $D_{i}$ (e.g., $10\%$ ) for its optimization at the beginning of every communication round. The remaining set on each client for training is denoted as the study set $D_{i}^{s}$ . This round-wise resampling of the adaptability set guarantees that the ensemble weight $\lambda_{i}$ is continuously optimized on fresh and unbiased data, thereby effectively mitigating overfitting risks. The learning process of $\lambda$ on each client is: + +$$ +\boldsymbol {\lambda} ^ {t} \leftarrow \boldsymbol {\lambda} ^ {t - 1} - \eta_ {\boldsymbol {\lambda}} \frac {1}{\left| D _ {i} ^ {a} \right|} \sum_ {j \in D _ {i} ^ {a}} \nabla_ {\boldsymbol {\lambda} ^ {t - 1}} \mathbb {E} _ {\left(x _ {j}, y _ {j}\right) \sim D _ {i} ^ {a}} \mathcal {L} _ {\mathrm {C E}} \left(\boldsymbol {\lambda} ^ {t - 1} \boldsymbol {w} (\theta ; x _ {j}) + (1 - \boldsymbol {\lambda} ^ {t - 1}) \boldsymbol {g} (\phi ; x _ {j}), y _ {j}\right), \tag {5} +$$ + +where $\eta_{\lambda}$ is the learning rate for $\lambda$ . Finally, the total training objective of the local model $\boldsymbol{w}$ combines the dual-end alignment loss and the ensemble learning loss: + +$$ +\mathcal {L} _ {\boldsymbol {w}} = \mathcal {L} _ {D E A L} ^ {\boldsymbol {w}} + \mathcal {L} _ {E N S} ^ {\boldsymbol {w}}, \tag {6} +$$ + +For symmetry, an analogous loss $\mathcal{L}_g$ is also computed but omitted here for brevity, as it follows the same formulation with reversed inputs. The total losses $\mathcal{L}_g$ and $\mathcal{L}_w$ are used to simultaneously update the homogeneous auxiliary model and the heterogeneous client local model, with learning rates $\eta_g$ and $\eta_w$ , respectively, as follows: + +$$ +\begin{array}{l} \boldsymbol {w} ^ {t} \leftarrow \boldsymbol {w} ^ {t - 1} - \eta_ {\boldsymbol {w}} \frac {1}{\left| D _ {i} ^ {s} \right|} \sum_ {j \in D _ {i} ^ {s}} \nabla_ {\boldsymbol {w} ^ {t - 1}} \mathbb {E} _ {(x _ {j}, y _ {j}) \sim D _ {i} ^ {s}} \mathcal {L} _ {\boldsymbol {w}} (\boldsymbol {w} ^ {t - 1}, \boldsymbol {g} ^ {t - 1}, \boldsymbol {\lambda} ^ {t}, x _ {j}, y _ {j}), \\ \boldsymbol {g} ^ {t} \leftarrow \boldsymbol {g} ^ {t - 1} - \eta_ {\boldsymbol {g}} \frac {1}{| D _ {i} ^ {s} |} \sum_ {j \in D _ {i} ^ {s}} \nabla_ {\boldsymbol {g} ^ {t - 1}} \mathbb {E} _ {(x _ {j}, y _ {j}) \sim D _ {i} ^ {s}} \mathcal {L} _ {\boldsymbol {g}} \left(\boldsymbol {g} ^ {t - 1}, \boldsymbol {w} ^ {t - 1}, \boldsymbol {\lambda} ^ {t}, x _ {j}, y _ {j}\right). \tag {7} \\ \end{array} +$$ + +During the inference stage, clients use the weighted model ensemble for prediction: + +$$ +\hat {y} _ {j} ^ {\text {p r e d}} = \arg \max (\boldsymbol {\lambda} \boldsymbol {w} (\theta ; x _ {j}) + (1 - \boldsymbol {\lambda}) \boldsymbol {g} (\phi ; x _ {j})). \tag {8} +$$ + +This adaptive weighting mechanism automatically balances the contributions of both models based on their current performance. It is particularly effective under system heterogeneity, where devices may have varying computational capabilities. + +# 4.3 Importance-Driven Selective Parameter Update + +To enable straggling clients to quickly catch up upon rejoining training while preserving the learning momentum of more active clients, we propose a novel selective parameter update module, ISPU, which selectively updates task-relevant parameters and suppresses noisy or redundant updates. Instead of directly overwriting the local auxiliary model $\pmb{g}_i^{t-1}$ with the global model $\mathcal{G}^{t-1}$ , we perform a model fusion by identifying the most significant parameters in $\pmb{g}_i^{t-1}$ . The update ratio $\alpha_i^t$ is adaptively determined by a sigmoid-based scheduling function according to the training progress: + +$$ +\alpha_ {i} ^ {t} = \frac {1}{1 + \exp \left(\delta \cdot \left(\frac {N _ {i} (t)}{t + 1} - 0 . 5\right)\right)} \cdot \tau \tag {9} +$$ + +where $N_{i}(t) = \sum_{r = 1}^{t}\mathbb{1}\{i\in \mathcal{A}_{r}\}$ denotes the cumulative number of rounds in which client $i$ has participated up to round $t$ and $\mathbb{1}\{\cdot \}$ is the indicator function. Here, $\tau \in (0,1]$ represents the pruning threshold and $\delta$ is a tunable sharpness hyperparameter. We then apply a binary mask on the parameters of $\pmb{g}_i^{t - 1}$ to retain only the critical parameters and replace them with the corresponding parameters from the global model $\mathcal{G}^{t - 1}$ . Common pruning metrics include the $\ell$ -norm [50, 3], Fisher Information Matrix (FIM) [51, 52], and sensitivity-based measures [53, 54]. Specifically, we adopt + +the $\ell_1$ -norm to evaluate the importance of parameters, which has been proven to be an effective technique for assessing parameter significance [55, 3]. Compared to other metrics, this formulation better captures the importance of task-relevant parameters. After local training, the binary mask $M_i^t$ is constructed to update only the top- $\alpha_i^t$ important parameters of $g_i^t$ in the next participation round: + +$$ +M _ {i, d} ^ {t} = \left\{ \begin{array}{l l} 1, & \text {i f} d - \text {t h p a r a m e t e r} \in \operatorname {t o p} - \alpha_ {i} ^ {t} \text {l a r g e s t o f} \boldsymbol {g} _ {i} ^ {t} \\ 0, & \text {o t h e r w i s e} \end{array} \right. \tag {10} +$$ + +This parameter-wise filtering mechanism helps frequently active clients preserve their learned knowledge while allowing infrequent clients to assimilate global updates more effectively, enabling rapid catch-up and mitigating knowledge drift. Specifically, at the beginning of round $t$ , the local auxiliary model $g_{i}^{t - 1}$ of each active client $i \in \mathcal{A}_t$ is updated as follows: + +$$ +\hat {\boldsymbol {g}} _ {i} ^ {t - 1} = \boldsymbol {g} _ {i} ^ {t - 1} \odot \neg M _ {i} ^ {h} + \mathcal {G} ^ {t - 1} \odot M _ {i} ^ {h}, \tag {11} +$$ + +where $M_{i}^{h}$ is the mask matrix obtained by client $i$ from its most recent training round and $\neg M$ denotes the bit-wise inverse of the mask $M$ . This approach ensures clients preserve critical knowledge via high-importance parameters, filters out conflicts from heterogeneous data and inconsistent training progress. + +# 5 Experiments + +# 5.1 Experiments Details + +Datasets. We evaluate our proposed PHP-FL on two standard image classification benchmarks: Fashion-MNIST $^4$ [56] and CIFAR- $10^5$ [57]. For both datasets, we adopt a 4:1 ratio to split samples into training and test sets. Following previous studies [12, 58, 59], we simulate heterogeneous data distributions by allocating class $j$ proportions to each client $k$ according to a Dirichlet distribution $(p_{j,k} \sim \mathrm{Dir}(\beta))$ , where a smaller $\beta$ implies more extreme data heterogeneity across clients. We adopt $\beta = 0.1$ for Fashion-MNIST and $\beta = 0.5$ for CIFAR-10, respectively. Notably, each client's local training and test sets share the same distribution. + +Models. Our experimental setup employs four heterogeneous model architectures: (1) GoogleNet [60], (2) DenseNet-121 [61], (3) EfficientNet-B1 [62], and (4) ResNet-18 [63]. Each client is assigned one of these models based on its identifier $i$ , following a round-robin strategy where client $i$ receives the model corresponding to $i \mod 4$ . Comprehensive comparative results of homogeneous model architectures are provided in Appendix C.2. + +Comparison Baselines. In the heterogeneous model experiments, we comprehensively evaluate our method against six state-of-the-art heterogeneous federated learning algorithms without relying on public data, including FML [21], FedGen [10], FedKD [11], FedAPEN [47], FedTGP [12], FedMRL [13]. In addition, we also compare a standalone baseline where clients train locally without any aggregation or communication. + +Implementation Details. Our experimental framework is built on a lightweight MH-FL platform HtFLib [64] using PyTorch 2.2.2 [65] with NVIDIA RTX 3090 GPU. We employ SGD as our optimizer with a learning rate of 0.001 and a local batch size of 64. The global training process consists of 100 communication rounds, with a total of 20 clients. During each federated training round, clients perform 5 local epochs of training. In each round, client participation follows the three patterns introduced in Section 3. We repeated all experiments three times with different random seeds and present the averaged results. More details are provided in Appendix B. + +Evaluation Metrics. As defined in Section 3, we evaluate the performance using the average Top-1 accuracy (AM) across all clients. In evaluating the fairness of the clients, we adopt the standard deviation (FM) of client accuracy when the Top-1 test accuracy is achieved. In PHP-FL, the test accuracy and fairness for each client are derived from the ensemble output of its local and auxiliary models, as computed by Eq. 8. + +Table 1: Comparison with the state-of-the-art methods on Fashion-MNIST in the heterogeneous setting. Best in **bold** and second with **underline**. ↑ indicates improved accuracy (\%) and ↓ indicates improved standard deviation (\%) of accuracy compared with the best baseline, respectively. + +
MethodsUniform [a = 0.5]Normal [μ = 0.5, σ = 0.2]Linear [a = 0.05, d = K-2/K(K-1)]
AM (%) ↑FM (%) ↓AM (%) ↑FM (%) ↓AM (%) ↑FM (%) ↓
Standalone95.88±0.199.26±1.3796.89±0.199.37±1.2295.77±0.059.86±0.53
FML [arXiv20]89.09±0.6623.13±1.1888.71±0.2323.32±1.1188.15±0.7325.21±2.76
FedGen [ICML21]93.97±1.1322.78±1.4693.81±0.9923.25±1.5293.72±0.9323.80±1.90
FedKD [NC22]95.67±0.319.20±0.9295.65±0.309.02±1.1695.57±0.189.30±0.78
FedAPEN [KDD23]96.79±0.066.74±0.2696.79±0.076.50±0.6096.73±0.066.89±0.08
FedTGP [AAAI24]94.35±2.5710.94±2.4794.06±2.3811.32±2.2594.21±2.4811.50±2.18
FedMRL [NIPS24]96.06±0.459.07±1.6596.05±0.439.25±1.3995.78±0.099.81±0.64
PHP-FL (Ours)97.64±0.044.15±0.1897.59±0.044.24±0.0597.58±0.034.29±0.04
↑0.85↓2.59↑0.80↓2.26↑0.95↓2.60
+ +Table 2: Comparison with the state-of-the-art methods on CIFAR-10 in the heterogeneous setting. Best in bold and second with underline. $\uparrow$ indicates improved accuracy (\%) and $\downarrow$ indicates improved standard deviation (\%) of accuracy compared with the best baseline, respectively. + +
MethodsUniform [a = 0.5]Normal [μ = 0.5, σ = 0.2]Linear [a = 0.05, d = K-2/K(K-1)]
AM (%) ↑FM (%) ↓AM (%) ↑FM (%) ↓AM (%) ↑FM (%) ↓
Standalone56.66±0.3210.28±0.3256.77±0.4010.19±0.3656.61±0.3010.21±0.32
FML [arXiv20]46.61±0.6115.18±0.3446.16±1.0215.78±0.7546.04±1.1615.85±0.83
FedGen [ICML21]54.27±0.1510.29±0.2654.38±0.0210.33±0.3154.50±0.1810.45±0.46
FedKD [NC22]54.86±0.4310.43±0.2354.82±0.3710.36±0.3354.55±0.1310.45±0.21
FedAPEN [KDD23]60.30±0.339.40±0.3060.35±0.319.45±0.3160.40±0.319.40±0.30
FedTGP [AAAI24]53.39±0.539.73±0.6153.40±0.5210.62±1.1752.85±1.1610.28±0.78
FedMRL [NIPS24]52.80±0.6312.81±1.1752.61±0.7712.85±1.2053.11±0.6312.95±1.27
PHP-FL (Ours)66.85±0.368.07±0.2566.94±0.398.07±0.2466.78±0.368.09±0.27
↑6.55↓1.38↑6.59↓1.29↑6.38↓1.31
+ +# 5.2 Comparison to State-of-the-Arts Methods + +We evaluate PHP-FL against several state-of-the-art methods in Tables 1 and 2. The experiments are conducted under three distinct client participation patterns: uniform, normal, and linear, representing diverse real-world scenarios. Across all settings, PHP-FL consistently demonstrates superior performance. It achieves the highest average accuracy (highest AM values) while simultaneously exhibiting the best fairness (lowest FM values). Compared to the strongest baseline FedAPEN, PHP-FL achieves notable improvements in average performance, boosting AM by up to $0.87\%$ on Fashion-MNIST and a substantial $6.51\%$ on CIFAR-10. At the same time, it enhances fairness by reducing FM by up to $2.48\%$ and $1.33\%$ on the respective datasets, demonstrating the robustness and effectiveness of PHP-FL in addressing model heterogeneity and unfairness caused by inconsistent client participation. Appendix C.1 further shows its faster convergence and superior performance through accuracy and standard deviation curves. + +# 5.3 Ablation Study + +In Table 3, we present an ablation study to evaluate the contribution of the DEAL and ISPU modules in PHP-FL under the normal participation pattern. When both modules are disabled, the performance significantly degrades, especially on CIFAR-10, where the accuracy drops to $59.88\%$ . + +![](images/029e141ab0698a665e3dc38d611d70863f263a346956fba3c3d84fdee9150c52.jpg) + +![](images/96c07b1ddd09a1a98d688118bdd00264ba030e7e3080c744c829dafcdb7ae0b9.jpg) +(a) AM $(\%)$ @ $\beta = 0.1$ + +![](images/9436ef7b4d1a09e01eda3c54ea66005852309cb752538836b88b87e23e03899c.jpg) +(b) AM $(\%)$ @ $\beta = 5$ + +![](images/a1f54b328cf6d393fba995ef7b172fd79c812d04a521fde6fa2c674888c94e08.jpg) +(c) FM $(\%)$ @ $\beta = 0.1$ + +![](images/3d8173ead42787dd561158967607f5e6930a717dcb361b90a0f07568440bc5ed.jpg) +(d) FM $(\%)$ @ $\beta = 5$ + +![](images/65e2eefe83bdb379a5cc72efb03df3310586cd311fe1895dee87b0c1f712ad69.jpg) +Figure 3: Comparison results on CIFAR-10 under varying degrees of data distribution heterogeneity across clients. All other settings follow their default configurations. +(a) Fashion-MNIST + +![](images/bb5d3280bfce589c8b3971706eccfa64769480777d670539e449db84960b3346.jpg) +(b) CIFAR-10 + +![](images/bdb38d61433ea92dc83e32c51648d484c93b4f1dae82215c93e24b197660fc98.jpg) +(c) Average weight $\lambda$ +Figure 4: Left two: effect of $\tau$ on performance. Right: average weight $\lambda$ of clients in each round. + +Introducing the ISPU module alone brings a modest improvement, highlighting its effectiveness in mitigating update conflicts from inconsistent client participation. + +Besides, enabling only the DEAL module yields a more pronounced performance gains. This is achieved by effectively addressing system heterogeneity through data-free knowledge alignment and ensemble learning. Notably, enabling both DEAL and ISPU achieves the best performance on both datasets, with $97.59\%$ accuracy on Fashion-MNIST and $66.94\%$ on CIFAR-10, demonstrating their complementarity and the necessity of their joint design. + +Table 3: Ablation study of key modules of PHP-FL under the normal pattern. + +
DEALISPUFashion-MNISTCIFAR-10
AM (%)FM (%)AM (%)FM (%)
XX92.869.1059.8811.24
X✓96.725.0762.059.47
✓X96.985.7363.088.16
✓✓97.594.2466.948.07
+ +# 5.4 Case Studies + +Robustness to Non-IIDness. To evaluate PHP-FL's robustness under varying data heterogeneity, we conducted additional experiments on CIFAR-10 using the Dirichlet distribution with $\beta = 0.1$ (high heterogeneity) and $\beta = 5$ (low heterogeneity). As shown in Figure 3, PHP-FL consistently outperforms all baselines across both settings. Under high heterogeneity, PHP-FL surpasses the best-performing baseline by $2.09\%$ in accuracy (AM) and reduces the fairness metric (FM) by $1.84\%$ . This advantage becomes even more pronounced under low heterogeneity, where PHP-FL achieves a remarkable $8.83\%$ accuracy gain over the next best method while maintaining the best fairness performance. These results clearly show that PHP-FL is not only robust to different levels of data heterogeneity but consistently achieves state-of-the-art performance in both accuracy and fairness. + +Effect of the Pruning Threshold $\tau$ on Performance. To investigate the effect of the hyperparameter $\tau$ , we conduct experiments on Fashion-MNIST and CIFAR-10 under the normal pattern. As shown in Figure 4a and 4b, on the CIFAR-10 dataset, the accuracy first increases and then decreases, while the standard deviation initially decreases and then increases, with both metrics achieving their best values when $\tau$ is set to 0.2. For the Fashion-MNIST dataset, the performance remains relatively + +![](images/9ad4c1554c500816d5a63e1598f7f6078155cbfefd73e8e7ff378ce2681bc10a.jpg) +Figure 5: The client accuracy distribution at the best achieved AM (best mean accuracy) on CIFAR-10 dataset under the uniform participation pattern for PHP-FL and two other baseline methods. + +![](images/52535e80f7701c57e797a6a0cd5e8333c526c2eeee3a6540a1f487fd0b5313f0.jpg) + +![](images/8bf6035d557e097cd313d55742f3d0617a98cca82c8e00ceafaf00d5350bcb8d.jpg) + +stable across different $\tau$ , and similarly, the best results are also observed at $\tau = 0.2$ . Therefore, we choose $\tau = 0.2$ as the default configuration for all experiments. + +Effect of Adaptive Ensemble Weights. We analyze the behavior of the adaptive ensemble weight $\lambda$ by tracking its average value across clients throughout training under the uniform pattern. As depicted in Figure 4c, the dynamics of $\lambda$ differ significantly between datasets. For CIFAR-10, the average $\lambda$ is initialized as 0.5 but quickly decreases and stabilizes around 0.43, indicating a consistent preference for the global model $g$ within the ensemble on this more complex dataset. In contrast, on Fashion-MNIST, the average $\lambda$ steadily increases from 0.5 to approximately 0.58 by the end of training, signifying a growing reliance on the specialized local models $w$ . This demonstrates that the adaptive mechanism effectively captures dataset-specific characteristics, dynamically adjusting the ensemble balance between local and global models to leverage their respective strengths during the learning process. + +Visualization of Client Accuracy Distribution. To visualize the client accuracy distribution under the normal participation pattern, we plot histograms and Kernel Density Estimation (KDE) [66] curves for different methods on CIFAR-10 dataset. As shown in Figure 5, PHP-FL achieves a more concentrated accuracy distribution compared to FedAPEN and FedMRL, with clients generally attaining higher accuracy. Moreover, the differences in client performance are significantly reduced under PHP-FL, highlighting its superiority in both enhancing average performance and promoting fairness across clients. + +# 6 Conclusion + +In this paper, we propose PHP-FL, a novel model-heterogeneous federated learning method addressing the critical challenge of enhancing both accuracy and fairness under inconsistent client participation probabilities. PHP-FL achieves this through two integrated modules: (1) the DEAL module, which harmonizes heterogeneous models via data-free knowledge alignment; and (2) the ISPU module, which selectively updates task-relevant parameters to mitigate update conflicts. Evaluated across diverse participation patterns, PHP-FL demonstrates state-of-the-art performance for both accuracy and fairness. Ablation study further validates the effectiveness of each module. Our research narrows the divide between idealized uniform participation scenarios and practical heterogeneous FL systems, providing a lightweight yet robust solution suitable for real-world implementation. + +Limitations. Despite the promising results, PHP-FL has two main limitations: + +First, compared to approaches that exchange only lightweight information (e.g., logits, prototypes [29, 12], or partial model parameters [10, 11]), our method introduces non-negligible computation and communication overheads. Although employing a smaller auxiliary model can alleviate this burden, the additional costs from ensemble training and selective parameter updates still persist. + +Second, our evaluation is conducted on a constrained set of model heterogeneity types, datasets, and client participation patterns. Although PHP-FL demonstrates robust performance within these scenarios, its generalizability should be further verified on more diverse and large-scale benchmarks. + +# Acknowledgments and Disclosure of Funding + +This work was supported by the National Key Research and Development Program of China under Grant No.2024YFE0204500, and in part by the National Natural Science Foundation of China under Grant No.92267104. + +# References + +[1] Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Aguera y Arcas. Communication-efficient learning of deep networks from decentralized data. In Artificial intelligence and statistics, pages 1273-1282. PMLR, 2017. +[2] Mang Ye, Xiuwen Fang, Bo Du, Pong C Yuen, and Dacheng Tao. Heterogeneous federated learning: State-of-the-art and research challenges. ACM Computing Surveys, 56(3):1-44, 2023. +[3] Yongzhe Jia, Xuyun Zhang, Hongsheng Hu, Kim-Kwang Raymond Choo, Lianyong Qi, Xiaolong Xu, Amin Beheshti, and Wanchun Dou. Dapperfl: Domain adaptive federated learning with model fusion pruning for edge devices. Advances in Neural Information Processing Systems, 37:13099-13123, 2024. +[4] Jingwei Liu, Yating Li, Mengjiao Zhao, Lei Liu, and Neeraj Kumar. Epflf: enhancing privacy and fairness in federated learning for distributed e-healthcare data sharing services. IEEE Transactions on Dependable and Secure Computing, 2024. +[5] Pushpita Chatterjee, Debashis Das, and Danda B Rawat. Federated learning empowered recommendation model for financial consumer services. IEEE Transactions on Consumer Electronics, 70(1):2508-2516, 2023. +[6] Shuai Wang, Yexuan Fu, Xiang Li, Yunshi Lan, Ming Gao, et al. Dfrd: Data-free robustness distillation for heterogeneous federated learning. Advances in Neural Information Processing Systems, 36:17854-17866, 2023. +[7] Liping Yi, Gang Wang, Xiaoguang Liu, Zhuan Shi, and Han Yu. Fedgh: Heterogeneous federated learning with generalized global header. In Proceedings of the 31st ACM International Conference on Multimedia, pages 8686-8696, 2023. +[8] Xiuwen Fang and Mang Ye. Noise-robust federated learning with model heterogeneous clients. IEEE Transactions on Mobile Computing, 2024. +[9] Siyuan Wu, Hao Tian, Weiran Zhang, Tingtong Zhu, Fuwen Tian, Zhehong Wang, and Wanchun Dou. A heterogeneous federated learning method based on dual teachers knowledge distillation. In International Conference on Advanced Data Mining and Applications, pages 192-207. Springer, 2024. +[10] Zhuangdi Zhu, Junyuan Hong, and Jiayu Zhou. Data-free knowledge distillation for heterogeneous federated learning. In International conference on machine learning, pages 12878-12889. PMLR, 2021. +[11] Chuhan Wu, Fangzhao Wu, Lingjuan Lyu, Yongfeng Huang, and Xing Xie. Communication-efficient federated learning via knowledge distillation. Nature communications, 13(1):2032, 2022. +[12] Jianqing Zhang, Yang Liu, Yang Hua, and Jian Cao. Fedtgp: Trainable global prototypes with adaptive-margin-enhanced contrastive learning for data and model heterogeneity in federated learning. In Proceedings of the AAAI conference on artificial intelligence, volume 38, pages 16768-16776, 2024. +[13] Liping Yi, Han Yu, Chao Ren, Gang Wang, Xiaoxiao Li, et al. Federated model heterogeneous matryoshka representation learning. Advances in Neural Information Processing Systems, 37:66431-66454, 2024. + +[14] Tian Li, Shengyuan Hu, Ahmad Beirami, and Virginia Smith. Ditto: Fair and robust federated learning through personalization. In International conference on machine learning, pages 6357-6368. PMLR, 2021. +[15] Yue Tan, Guodong Long, Lu Liu, Tianyi Zhou, Qinghua Lu, Jing Jiang, and Chengqi Zhang. Fedproto: Federated prototype learning across heterogeneous clients. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 8432-8440, 2022. +[16] Jianqiao Zhang, Caifeng Shan, and Jungong Han. Fedgmkd: An efficient prototype federated learning framework through knowledge distillation and discrepancy-aware aggregation. Advances in Neural Information Processing Systems, 37:118326-118356, 2024. +[17] Michal Yemini, Rajarshi Saha, Emre Ozfatura, Deniz Gunduz, and Andrea J Goldsmith. Robust federated learning with connectivity failures: A semi-decentralized framework with collaborative relaying. arXiv preprint arXiv:2202.11850, 2022. +[18] Hao Ye, Le Liang, and Geoffrey Ye Li. Decentralized federated learning with unreliable communications. IEEE journal of selected topics in signal processing, 16(3):487-500, 2022. +[19] Ming Wen, Chengchang Liu, and Yuedong Xu. Communication efficient distributed newton method over unreliable networks. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 15832-15840, 2024. +[20] Ming Xiang, Stratis Ioannidis, Edmund Yeh, Carlee Joe-Wong, and Lili Su. Efficient federated learning against heterogeneous and non-stationary client unavailability. Advances in Neural Information Processing Systems, 37:104281-104328, 2024. +[21] Tao Shen, Jie Zhang, Xinkang Jia, Fengda Zhang, Gang Huang, Pan Zhou, Kun Kuang, Fei Wu, and Chao Wu. Federated mutual learning. arXiv preprint arXiv:2006.16765, 2020. +[22] Yae Jee Cho, Jianyu Wang, and Gauri Joshi. Towards understanding biased client selection in federated learning. In International Conference on Artificial Intelligence and Statistics, pages 10351-10375. PMLR, 2022. +[23] Tiansheng Huang, Weiwei Lin, Wentai Wu, Ligang He, Keqin Li, and Albert Y Zomaya. An efficiency-boosting client selection scheme for federated learning with fairness guarantee. IEEE Transactions on Parallel and Distributed Systems, 32(7):1552–1564, 2020. +[24] Rituparna Saha, Sudip Misra, Aishwariya Chakraborty, Chandranath Chatterjee, and Pallav Kumar Deb. Data-centric client selection for federated learning over distributed edge networks. IEEE Transactions on Parallel and Distributed Systems, 34(2):675-686, 2023. +[25] Zhiyuan Ning, Chunlin Tian, Meng Xiao, Wei Fan, Pengyang Wang, Li Li, Pengfei Wang, and Yuanchun Zhou. Fedgcs: A generative framework for efficient client selection in federated learning via gradient-based optimization. arXiv preprint arXiv:2405.06312, 2024. +[26] Tian Li, Anit Kumar Sahu, Manzil Zaheer, Maziar Sanjabi, Ameet Talwalkar, and Virginia Smith. Federated optimization in heterogeneous networks. Proceedings of Machine learning and systems, 2:429-450, 2020. +[27] Tian Li, Maziar Sanjabi, Ahmad Beirami, and Virginia Smith. Fair resource allocation in federated learning. arXiv preprint arXiv:1905.10497, 2019. +[28] Yuhang Chen, Wenke Huang, and Mang Ye. Fair federated learning under domain skew with local consistency and domain diversity. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12077-12086, 2024. +[29] Daliang Li and Junpu Wang. Fedmd: Heterogenous federated learning via model distillation. arXiv preprint arXiv:1910.03581, 2019. +[30] Tao Lin, Lingjing Kong, Sebastian U Stich, and Martin Jaggi. Ensemble distillation for robust model fusion in federated learning. Advances in Neural Information Processing Systems, 33:2351-2363, 2020. + +[31] Wenke Huang, Mang Ye, and Bo Du. Learn from others and be yourself in heterogeneous federated learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10143-10153, 2022. +[32] Jaehee Jang, Heoneok Ha, Dahuin Jung, and Sungroh Yoon. Fedclassavg: Local representation learning for personalized federated learning on heterogeneous neural networks. In Proceedings of the 51st International Conference on Parallel Processing, pages 1-10, 2022. +[33] Guangyu Sun, Matias Mendieta, Jun Luo, Shandong Wu, and Chen Chen. Fedperfix: Towards partial model personalization of vision transformers in federated learning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4988-4998, 2023. +[34] Feijie Wu, Xingchen Wang, Yaqing Wang, Tianci Liu, Lu Su, and Jing Gao. FIARSE: Model-heterogeneous federated learning via importance-aware submodel extraction. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. +[35] Jingwen Zhang, Yuezhou Wu, and Rong Pan. Incentive mechanism for horizontal federated learning based on reputation and reverse auction. In Proceedings of the Web Conference 2021, pages 947-956, 2021. +[36] Meirui Jiang, Holger R Roth, Wenqi Li, Dong Yang, Can Zhao, Vishwesh Nath, Daguang Xu, Qi Dou, and Ziyue Xu. Fair federated medical image segmentation via client contribution estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16302-16311, 2023. +[37] Ruoxi Jia, David Dao, Boxin Wang, Frances Ann Hubis, Nick Hynes, Nezihe Merve Gürel, Bo Li, Ce Zhang, Dawn Song, and Costas J Spanos. Towards efficient data valuation based on the shapley value. In The 22nd International Conference on Artificial Intelligence and Statistics, pages 1167-1176. PMLR, 2019. +[38] Nurbek Tastan, Samar Fares, Toluwani Aremu, Samuel Horvath, and Karthik Nandakumar. Redefining contributions: Shapley-driven federated learning. In International Joint Conference on Artificial Intelligence (IJCAI), 2024. +[39] Yahya H Ezzeldin, Shen Yan, Chaoyang He, Emilio Ferrara, and A Salman Avestimehr. Fairfed: Enabling group fairness in federated learning. In Proceedings of the AAAI conference on artificial intelligence, volume 37, pages 7494-7502, 2023. +[40] Zeqing He, Zhibo Wang, Xiaowei Dong, Peng Sun, Ju Ren, and Kui Ren. Towards fair federated learning via unbiased feature aggregation. IEEE Transactions on Dependable and Secure Computing, 2025. +[41] Zheng Wang, Xiaoliang Fan, Jianzhong Qi, Chenglu Wen, Cheng Wang, and Rongshan Yu. Federated learning with fair averaging. arXiv preprint arXiv:2104.14937, 2021. +[42] Mingyang Wan, Daochen Zha, Ninghao Liu, and Na Zou. In-processing modeling techniques for machine learning fairness: A survey. ACM Transactions on Knowledge Discovery from Data, 17(3):1-27, 2023. +[43] Reuben Binns. On the apparent conflict between individual and group fairness. In Proceedings of the 2020 conference on fairness, accountability, and transparency, pages 514-524, 2020. +[44] Xinran Gu, Kaixuan Huang, Jingzhao Zhang, and Longbo Huang. Fast federated learning in the presence of arbitrary device unavailability. Advances in Neural Information Processing Systems, 34:12052-12064, 2021. +[45] Shiqiang Wang and Mingyue Ji. A lightweight method for tackling unknown participation statistics in federated averaging. In The Twelfth International Conference on Learning Representations, 2024. +[46] Eunjeong Jeong, Seungeun Oh, Hyesung Kim, Jihong Park, Mehdi Bennis, and Seong-Lyun Kim. Communication-efficient on-device machine learning: Federated distillation and augmentation under non-iid private data. arXiv preprint arXiv:1811.11479, 2018. + +[47] Zhen Qin, Shuiguang Deng, Mingyu Zhao, and Xueqiang Yan. Fedapen: Personalized cross-silo federated learning with adaptability to statistical heterogeneity. In Proceedings of the 29th ACM SIGKDD conference on knowledge discovery and data mining, pages 1954–1964, 2023. +[48] Arthur Gretton, Karsten M Borgwardt, Malte J Rasch, Bernhard Schölkopf, and Alexander Smola. A kernel two-sample test. The Journal of Machine Learning Research, 13(1):723-773, 2012. +[49] Devansh Arpit, Huan Wang, Yingbo Zhou, and Caiming Xiong. Ensemble of averages: Improving model selection and boosting performance in domain generalization. Advances in Neural Information Processing Systems, 35:8265-8277, 2022. +[50] Xiaopeng Jiang and Cristian Borcea. Complement sparsification: Low-overhead model pruning for federated learning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 37, pages 8087-8095, 2023. +[51] Liyang Liu, Shilong Zhang, Zhanghui Kuang, Aojun Zhou, Jing-Hao Xue, Xinjiang Wang, Yimin Chen, Wenming Yang, Qingmin Liao, and Wayne Zhang. Group fisher pruning for practical network compression. In International Conference on Machine Learning, pages 7021-7032. PMLR, 2021. +[52] Xiyuan Yang, Wenke Huang, and Mang Ye. Fedas: Bridging inconsistency in personalized federated learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11986-11995, 2024. +[53] Michael Santacroce, Zixin Wen, Yelong Shen, and Yuanzhi Li. What matters in the structured pruning of generative language models? arXiv preprint arXiv:2302.03773, 2023. +[54] Xinghao Wu, Xuefeng Liu, Jianwei Niu, Guogang Zhu, and Shaojie Tang. Bold but cautious: Unlocking the potential of personalized federated learning through cautiously aggressive collaboration. In Proceedings of the IEEE/CVF international conference on computer vision, pages 19375-19384, 2023. +[55] Zhuang Liu, Jianguo Li, Zhiqiang Shen, Gao Huang, Shoumeng Yan, and Changshui Zhang. Learning efficient convolutional networks through network slimming. In 2017 IEEE International Conference on Computer Vision (ICCV), pages 2755-2763, 2017. +[56] Han Xiao, Kashif Rasul, and Roland Vollgraf. Fashion-mnist: a novel image dataset for benchmarking machine learning algorithms. arXiv preprint arXiv:1708.07747, 2017. +[57] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. +[58] Qinbin Li, Yiqun Diao, Quan Chen, and Bingsheng He. Federated learning on non-iid data silos: An experimental study. In 2022 IEEE 38th International Conference on Data Engineering (ICDE), pages 965-978. IEEE, 2022. +[59] Jianqing Zhang, Yang Hua, Hao Wang, Tao Song, Zhengui Xue, Ruhui Ma, Jian Cao, and Haibing Guan. Gpfl: Simultaneously learning global and personalized feature information for personalized federated learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5041-5051, 2023. +[60] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1-9, 2015. +[61] Gao Huang, Zhuang Liu, Laurens Van Der Maaten, and Kilian Q Weinberger. Densely connected convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4700-4708, 2017. +[62] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In International conference on machine learning, pages 6105-6114. PMLR, 2019. + +[63] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. +[64] Jianqing Zhang, Yang Liu, Yang Hua, Hao Wang, Tao Song, Zhengui Xue, Ruhui Ma, and Jian Cao. Pfllib: A beginner-friendly and comprehensive personalized federated learning library and benchmark. Journal of Machine Learning Research, 26(50):1-10, 2025. +[65] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. +[66] Emanuel Parzen. On estimation of a probability density function and mode. The annals of mathematical statistics, 33(3):1065-1076, 1962. +[67] Keith Bonawitz, Vladimir Ivanov, Ben Kreuter, Antonio Marcedone, H Brendan McMahan, Sarvar Patel, Daniel Ramage, Aaron Segal, and Karn Seth. Practical secure aggregation for privacy-preserving machine learning. In proceedings of the 2017 ACM SIGSAC Conference on Computer and Communications Security, pages 1175–1191, 2017. +[68] Galen Andrew, Om Thakkar, Brendan McMahan, and Swaroop Ramaswamy. Differentially private learning with adaptive clipping. Advances in Neural Information Processing Systems, 34:17455-17466, 2021. + +# NeurIPS Paper Checklist + +# 1. Claims + +Question: Do the main claims made in the abstract and introduction accurately reflect the paper's contributions and scope? + +Answer: [Yes] + +Justification: The Abstract and Introduction Sections accurately reflect the paper's contributions and scope. + +Guidelines: + +- The answer NA means that the abstract and introduction do not include the claims made in the paper. +- The abstract and/or introduction should clearly state the claims made, including the contributions made in the paper and important assumptions and limitations. A No or NA answer to this question will not be perceived well by the reviewers. +- The claims made should match theoretical and experimental results, and reflect how much the results can be expected to generalize to other settings. +- It is fine to include aspirational goals as motivation as long as it is clear that these goals are not attained by the paper. + +# 2. Limitations + +Question: Does the paper discuss the limitations of the work performed by the authors? + +Answer: [Yes] + +Justification: The discussion about the limitations of this work is provided in the Conclusion. + +Guidelines: + +- The answer NA means that the paper has no limitation while the answer No means that the paper has limitations, but those are not discussed in the paper. +- The authors are encouraged to create a separate "Limitations" section in their paper. +- The paper should point out any strong assumptions and how robust the results are to violations of these assumptions (e.g., independence assumptions, noiseless settings, model well-specification, asymptotic approximations only holding locally). The authors should reflect on how these assumptions might be violated in practice and what the implications would be. +- The authors should reflect on the scope of the claims made, e.g., if the approach was only tested on a few datasets or with a few runs. In general, empirical results often depend on implicit assumptions, which should be articulated. +- The authors should reflect on the factors that influence the performance of the approach. For example, a facial recognition algorithm may perform poorly when image resolution is low or images are taken in low lighting. Or a speech-to-text system might not be used reliably to provide closed captions for online lectures because it fails to handle technical jargon. +- The authors should discuss the computational efficiency of the proposed algorithms and how they scale with dataset size. +- If applicable, the authors should discuss possible limitations of their approach to address problems of privacy and fairness. +- While the authors might fear that complete honesty about limitations might be used by reviewers as grounds for rejection, a worse outcome might be that reviewers discover limitations that aren't acknowledged in the paper. The authors should use their best judgment and recognize that individual actions in favor of transparency play an important role in developing norms that preserve the integrity of the community. Reviewers will be specifically instructed to not penalize honesty concerning limitations. + +# 3. Theory assumptions and proofs + +Question: For each theoretical result, does the paper provide the full set of assumptions and a complete (and correct) proof? + +# Answer: [NA] + +Justification: This paper does not include theoretical results. + +# Guidelines: + +- The answer NA means that the paper does not include theoretical results. +- All the theorems, formulas, and proofs in the paper should be numbered and cross-referenced. +- All assumptions should be clearly stated or referenced in the statement of any theorems. +- The proofs can either appear in the main paper or the supplemental material, but if they appear in the supplemental material, the authors are encouraged to provide a short proof sketch to provide intuition. +- Inversely, any informal proof provided in the core of the paper should be complemented by formal proofs provided in appendix or supplemental material. +- Theorems and Lemmas that the proof relies upon should be properly referenced. + +# 4. Experimental result reproducibility + +Question: Does the paper fully disclose all the information needed to reproduce the main experimental results of the paper to the extent that it affects the main claims and/or conclusions of the paper (regardless of whether the code and data are provided or not)? + +# Answer: [Yes] + +Justification: This paper presents comprehensive details of the experimental setup, and further provides the hyperparameters of baseline methods. This paper also provides an URL link to our released code. + +# Guidelines: + +- The answer NA means that the paper does not include experiments. +- If the paper includes experiments, a No answer to this question will not be perceived well by the reviewers: Making the paper reproducible is important, regardless of whether the code and data are provided or not. +- If the contribution is a dataset and/or model, the authors should describe the steps taken to make their results reproducible or verifiable. +- Depending on the contribution, reproducibility can be accomplished in various ways. For example, if the contribution is a novel architecture, describing the architecture fully might suffice, or if the contribution is a specific model and empirical evaluation, it may be necessary to either make it possible for others to replicate the model with the same dataset, or provide access to the model. In general, releasing code and data is often one good way to accomplish this, but reproducibility can also be provided via detailed instructions for how to replicate the results, access to a hosted model (e.g., in the case of a large language model), releasing of a model checkpoint, or other means that are appropriate to the research performed. +- While NeurIPS does not require releasing code, the conference does require all submissions to provide some reasonable avenue for reproducibility, which may depend on the nature of the contribution. For example + +(a) If the contribution is primarily a new algorithm, the paper should make it clear how to reproduce that algorithm. +(b) If the contribution is primarily a new model architecture, the paper should describe the architecture clearly and fully. +(c) If the contribution is a new model (e.g., a large language model), then there should either be a way to access this model for reproducing the results or a way to reproduce the model (e.g., with an open-source dataset or instructions for how to construct the dataset). +(d) We recognize that reproducibility may be tricky in some cases, in which case authors are welcome to describe the particular way they provide for reproducibility. In the case of closed-source models, it may be that access to the model is limited in some way (e.g., to registered users), but it should be possible for other researchers to have some path to reproducing or verifying the results. + +# 5. Open access to data and code + +Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material? + +Answer: [Yes] + +Justification: This paper provides an URL link to our released code. The public datasets used in this paper are properly referenced. + +Guidelines: + +- The answer NA means that paper does not include experiments requiring code. +- Please see the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details. +- While we encourage the release of code and data, we understand that this might not be possible, so No is an acceptable answer. Papers cannot be rejected simply for not including code, unless this is central to the contribution (e.g., for a new open-source benchmark). +- The instructions should contain the exact command and environment needed to run to reproduce the results. See the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details. +- The authors should provide instructions on data access and preparation, including how to access the raw data, preprocessed data, intermediate data, and generated data, etc. +- The authors should provide scripts to reproduce all experimental results for the new proposed method and baselines. If only a subset of experiments are reproducible, they should state which ones are omitted from the script and why. +- At submission time, to preserve anonymity, the authors should release anonymized versions (if applicable). +- Providing as much information as possible in supplemental material (appended to the paper) is recommended, but including URLs to data and code is permitted. + +# 6. Experimental setting/details + +Question: Does the paper specify all the training and test details (e.g., data splits, hyperparameters, how they were chosen, type of optimizer, etc.) necessary to understand the results? + +Answer: [Yes] + +Justification: This paper presents comprehensive details of the experimental setup in the Experimental Section and further provides the hyperparameters of baseline methods in Appendix. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The experimental setting should be presented in the core of the paper to a level of detail that is necessary to appreciate the results and make sense of them. +- The full details can be provided either with the code, in appendix, or as supplemental material. + +# 7. Experiment statistical significance + +Question: Does the paper report error bars suitably and correctly defined or other appropriate information about the statistical significance of the experiments? + +Answer: [Yes] + +Justification: The paper reports experimental results averaged over 3 runs with corresponding standard deviations in the main results. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The authors should answer "Yes" if the results are accompanied by error bars, confidence intervals, or statistical significance tests, at least for the experiments that support the main claims of the paper. + +- The factors of variability that the error bars are capturing should be clearly stated (for example, train/test split, initialization, random drawing of some parameter, or overall run with given experimental conditions). +- The method for calculating the error bars should be explained (closed form formula call to a library function, bootstrap, etc.) +- The assumptions made should be given (e.g., Normally distributed errors). +- It should be clear whether the error bar is the standard deviation or the standard error of the mean. +- It is OK to report 1-sigma error bars, but one should state it. The authors should preferably report a 2-sigma error bar than state that they have a $96\%$ CI, if the hypothesis of Normality of errors is not verified. +- For asymmetric distributions, the authors should be careful not to show in tables or figures symmetric error bars that would yield results that are out of range (e.g. negative error rates). +- If error bars are reported in tables or plots, The authors should explain in the text how they were calculated and reference the corresponding figures or tables in the text. + +# 8. Experiments compute resources + +Question: For each experiment, does the paper provide sufficient information on the computer resources (type of compute workers, memory, time of execution) needed to reproduce the experiments? + +Answer: [Yes] + +Justification: This paper describes the experimental environment in the Experiments Section and our code. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The paper should indicate the type of compute workers CPU or GPU, internal cluster, or cloud provider, including relevant memory and storage. +- The paper should provide the amount of compute required for each of the individual experimental runs as well as estimate the total compute. +- The paper should disclose whether the full research project required more compute than the experiments reported in the paper (e.g., preliminary or failed experiments that didn't make it into the paper). + +# 9. Code of ethics + +Question: Does the research conducted in the paper conform, in every respect, with the NeurIPS Code of Ethics https://neurips.cc/public/EthicsGuidelines? + +Answer: [Yes] + +Justification: This paper fully complies with the NeurIPS Code of Ethics in all respects. + +Guidelines: + +- The answer NA means that the authors have not reviewed the NeurIPS Code of Ethics. +- If the authors answer No, they should explain the special circumstances that require a deviation from the Code of Ethics. +- The authors should make sure to preserve anonymity (e.g., if there is a special consideration due to laws or regulations in their jurisdiction). + +# 10. Broader impacts + +Question: Does the paper discuss both potential positive societal impacts and negative societal impacts of the work performed? + +Answer: [No] + +Justification: There is no societal impact of the work performed. + +Guidelines: + +- The answer NA means that there is no societal impact of the work performed. + +- If the authors answer NA or No, they should explain why their work has no societal impact or why the paper does not address societal impact. +- Examples of negative societal impacts include potential malicious or unintended uses (e.g., disinformation, generating fake profiles, surveillance), fairness considerations (e.g., deployment of technologies that could make decisions that unfairly impact specific groups), privacy considerations, and security considerations. +- The conference expects that many papers will be foundational research and not tied to particular applications, let alone deployments. However, if there is a direct path to any negative applications, the authors should point it out. For example, it is legitimate to point out that an improvement in the quality of generative models could be used to generate deepfakes for disinformation. On the other hand, it is not needed to point out that a generic algorithm for optimizing neural networks could enable people to train models that generate Deepfakes faster. +- The authors should consider possible harms that could arise when the technology is being used as intended and functioning correctly, harms that could arise when the technology is being used as intended but gives incorrect results, and harms following from (intentional or unintentional) misuse of the technology. +- If there are negative societal impacts, the authors could also discuss possible mitigation strategies (e.g., gated release of models, providing defenses in addition to attacks, mechanisms for monitoring misuse, mechanisms to monitor how a system learns from feedback over time, improving the efficiency and accessibility of ML). + +# 11. Safeguards + +Question: Does the paper describe safeguards that have been put in place for responsible release of data or models that have a high risk for misuse (e.g., pretrained language models, image generators, or scraped datasets)? + +Answer: [No] + +Justification: This paper poses no such risks. + +Guidelines: + +- The answer NA means that the paper poses no such risks. +- Released models that have a high risk for misuse or dual-use should be released with necessary safeguards to allow for controlled use of the model, for example by requiring that users adhere to usage guidelines or restrictions to access the model or implementing safety filters. +- Datasets that have been scraped from the Internet could pose safety risks. The authors should describe how they avoided releasing unsafe images. +- We recognize that providing effective safeguards is challenging, and many papers do not require this, but we encourage authors to take this into account and make a best faith effort. + +# 12. Licenses for existing assets + +Question: Are the creators or original owners of assets (e.g., code, data, models), used in the paper, properly credited and are the license and terms of use explicitly mentioned and properly respected? + +Answer: [Yes] + +Justification: The assets used in this paper are properly credited. The license and terms of use are explicitly mentioned and properly respected. + +Guidelines: + +- The answer NA means that the paper does not use existing assets. +- The authors should cite the original paper that produced the code package or dataset. +- The authors should state which version of the asset is used and, if possible, include a URI. +- The name of the license (e.g., CC-BY 4.0) should be included for each asset. +- For scraped data from a particular source (e.g., website), the copyright and terms of service of that source should be provided. + +- If assets are released, the license, copyright information, and terms of use in the package should be provided. For popular datasets, paperswithcode.com/datasets has curated licenses for some datasets. Their licensing guide can help determine the license of a dataset. +- For existing datasets that are re-packaged, both the original license and the license of the derived asset (if it has changed) should be provided. +- If this information is not available online, the authors are encouraged to reach out to the asset's creators. + +# 13. New assets + +Question: Are new assets introduced in the paper well documented and is the documentation provided alongside the assets? + +Answer: [NA] + +Justification: This paper does not release new assets. + +Guidelines: + +- The answer NA means that the paper does not release new assets. +- Researchers should communicate the details of the dataset/code/model as part of their submissions via structured templates. This includes details about training, license, limitations, etc. +- The paper should discuss whether and how consent was obtained from people whose asset is used. +- At submission time, remember to anonymize your assets (if applicable). You can either create an anonymized URL or include an anonymized zip file. + +# 14. Crowdsourcing and research with human subjects + +Question: For crowdsourcing experiments and research with human subjects, does the paper include the full text of instructions given to participants and screenshots, if applicable, as well as details about compensation (if any)? + +Answer: [NA] + +Justification: This paper does not involve crowdsourcing nor research with human subjects. + +Guidelines: + +- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. +- Including this information in the supplemental material is fine, but if the main contribution of the paper involves human subjects, then as much detail as possible should be included in the main paper. +- According to the NeurIPS Code of Ethics, workers involved in data collection, curation, or other labor should be paid at least the minimum wage in the country of the data collector. + +# 15. Institutional review board (IRB) approvals or equivalent for research with human subjects + +Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals (or an equivalent approval/review based on the requirements of your country or institution) were obtained? + +Answer: [NA] + +Justification: This paper does not involve crowdsourcing nor research with human subjects. + +Guidelines: + +- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. +- Depending on the country in which research is conducted, IRB approval (or equivalent) may be required for any human subjects research. If you obtained IRB approval, you should clearly state this in the paper. + +We recognize that the procedures for this may vary significantly between institutions and locations, and we expect authors to adhere to the NeurIPS Code of Ethics and the guidelines for their institution. +- For initial submissions, do not include any information that would break anonymity (if applicable), such as the institution conducting the review. + +# 16. Declaration of LLM usage + +Question: Does the paper describe the usage of LLMs if it is an important, original, or non-standard component of the core methods in this research? Note that if the LLM is used only for writing, editing, or formatting purposes and does not impact the core methodology, scientific rigorousness, or originality of the research, declaration is not required. + +Answer: [NA] + +Justification: The core method development in this research does not involve LLMs as any important, original, or non-standard components. + +Guidelines: + +- The answer NA means that the core method development in this research does not involve LLMs as any important, original, or non-standard components. +- Please refer to our LLM policy (https://neurips.cc/Conferences/2025/LLM) for what should or should not be described. + +# APPENDIX + +# A Pseudo codes of PHP-FL + +The algorithm is outlined in Algorithm 1. Please refer to Section 4.1 for more details. + +Algorithm 1: PHP-FL +Input: Auxiliary small models $\{\pmb{g}_i^0\}_{i = 1}^K$ , Local models $\{\pmb {w}_i^0\}_{i = 1}^K$ , Total number of rounds $T$ Dataset partitions $\{D_i\}_{i = 1}^K$ +Output: Optimized local models $\{\pmb {w}_i^T\}_{i = 1}^K$ and auxiliary small models $\{\pmb {g}_i^T\}_{i = 1}^K$ Initialize historical mask $\{M_h^h\}_{i = 1}^K\gets 1^{|g|}$ and global model $\mathcal{G}^0$ . +for round $t = 1$ to $T$ do Server Side: Collect the IDs of active clients $\mathcal{A}_t\subseteq \{1,\dots,K\}$ Broadcast the pruned auxiliary small model $\mathcal{G}^{t - 1}\odot M_i^h$ to client $i\in \mathcal{A}_t$ $\{g_i^t,M_i^t\}_{i\in \mathcal{A}_t}\leftarrow$ Client Update; Update the historical mask $M_{i}^{h}\in M_{\mathrm{hist}}$ for clients in $\mathcal{A}_t:M_i^h\gets M_i^t$ Aggregate auxiliary small models: $\mathcal{G}^t = \frac{1}{|\mathcal{A}_t|}\sum_{i\in \mathcal{A}_t}\pmb {g}_i^t;$ +Client Update: +for each client $i\in \mathcal{A}_t$ in parallel do Download $\mathcal{G}^{t - 1}\odot M_i^h$ from server; $\pmb{g}_i^{t - 1}\gets$ initialize local auxiliary model by Eq. 11; Randomly divides the training set $D_{i}$ into $D_{i}^{a}$ and $D_{i}^{s}$ .. for batch $(x,y)\in D_i^a$ do $\lambda_i^t\gets$ update ensemble weight by Eq. 5; +end for batch $(x,y)\in D_i^s$ do $\mathcal{L}_{\boldsymbol{w}}\gets$ Calculate local training loss by Eq. 6; $\mathcal{L}_g\gets$ Calculate the symmetrical loss similar to Eq. 6; $\pmb {w}_i^t,\pmb {g}_i^t\gets$ update the local model and local auxiliary model by Eq. 7; +end $\alpha_{i}^{t}\gets$ calculate update ratio by Eq. 9; $M_{i}^{t}\gets$ obtain binary mask by Eq. 10 ; Upload $g_{i}^{t}$ and the binary mask $M_{i}^{t}$ to server; +end + +# B Additional Experimental Details + +# B.1 Hyperparameter Settings + +We provide a detailed summary of the hyperparameter configurations used in our experiments in Table 4. These settings are carefully selected to ensure fair comparison across different baselines. For the proposed PHP-FL, the standardized feature dimension $d$ is set to 512. The adaptability set $D_{i}^{a}$ consists of $10\%$ randomly sampled data from the training set $D_{i}$ . The ensemble weight $\lambda_{i}^{t}$ is trained for 10 epochs in each round. Additionally, following the hyperparameter search detailed in Section 5.4 and C.6, the pruning threshold $\tau$ and the sharpness factor $\delta$ are set to 0.2 and 5, respectively. Unless otherwise specified, all experiments follow the same training setting. + +Table 4: Hyperparameter settings used in our experiments. + +
TypeHyperparametersValue
FL training settingsLocal learning rate η0.001
Batch size64
Local epochs per round E5
Total rounds T100
Number of clients K20
Framework-specificα in FML0.5
β in FML0.5
Server learning epochs in FedGen100
Server learning rate in FedGen0.1
dη in FedGen32
dh in FedGen512
Tstart in FedKD0.95
Tend in FedKD0.98
ηs in FedKD0.001
Adaptation set ratio in FedAPEN10%
Server learning epochs in FedTGP100
τ in FedTGP100
λ in FedTGP0.1
d1 in FedMRL128
λ in Ditto0.1
α for CIFAR-10 in FedFV0.1
α for Fashion-MNIST in FedFV0
τ for CIFAR-10 in FedFV10
τ for Fashion-MNIST in FedFV0
β in FedHEAL0.4
τ in FedHEAL0.4
K in FedAU1
+ +# B.2 Visualization of Data Distributions + +To intuitively illustrate the data heterogeneity across clients in our federated learning setting, we plot scatter diagrams based on the CIFAR-10 and Fashion-MNIST datasets in Figure 6. Specifically, we simulate heterogeneous data distributions by allocating the proportion of class $j$ to each client $k$ according to a Dirichlet distribution $(p_{j,k} \sim \mathrm{Dir}(\beta))$ , where a smaller $\beta$ indicates more extreme data heterogeneity across clients. In our experiments, we set $\beta = 0.1$ for Fashion-MNIST and $\beta = 0.5$ for CIFAR-10, respectively. + +# B.3 Model Architectures Used in Experiments + +we utilize four widely recognized Convolutional Neural Network (CNN) architectures with varying designs and complexities. We report the corresponding parameter counts of each model in Table 5. These serve as the local models for clients in our heterogeneous federated learning setup: + +- GoogLeNet [60]: Introduced the inception module, which performs convolutions with multiple filter sizes in parallel within the same block. It was designed for computational efficiency and won the ILSVRC 2014 challenge. +- DenseNet-121 [61]: Characterized by its dense connectivity pattern, where each layer receives feature maps from all preceding layers within a dense block. This encourages feature reuse, strengthens gradient flow, and improves parameter efficiency. The '121' denotes the number of layers with weights. +- EfficientNet-B1 [62]: Developed using neural architecture search and a compound scaling method that uniformly scales network width, depth, and resolution. It aims to balance model accuracy with computational efficiency (FLOPS and parameters). B1 is a specific scaled version providing a good trade-off. + +![](images/26e316f71aae0697f869f841e561c27946830497665f6000035b9253225a0028.jpg) +(a) Fashion-MNIST + +![](images/a4c7e2aa027ed8e908d80a2e4a417dd1e75dfb84a6cc591800c97d00b9a78591.jpg) +(b) CIFAR-10 +Figure 6: The data distribution of 20 clients in our experiments. + +- ResNet-18 [63]: Employs residual connections (skip connections) that allow gradients to bypass layers, enabling the training of much deeper networks by mitigating the vanishing gradient problem. ResNet-18 is a relatively shallow variant with 18 layers containing weights. + +For a fair comparison, we adopt GoogLeNet [60] as the auxiliary model architecture for all methods that utilize an auxiliary model, including our proposed method PHP-FL, since it has the smallest number of parameters among the four candidate models. + +Table 5: Parameter counts of the evaluated models. "M" is short for million. + +
ModelParameter counts
GoogleNet [60]5.61M
DenseNet-121 [61]6.96M
EfficientNet-B1 [62]6.52M
ResNet-18 [63]11.18M
+ +# C Additional Experimental Results + +# C.1 Comparison of Training Curves for Accuracy and Standard Deviation + +To evaluate the convergence behavior and training stability of PHP-FL, we compare the training curves in terms of average accuracy and standard deviation across training rounds on Fashion-MNIST and CIFAR-10 under the uniform pattern. As shown in Figure 7 and 8, the results on both datasets demonstrate that the proposed method PHP-FL significantly accelerates convergence compared to baselines, while maintaining a low and stable standard deviation throughout training, which effectively enhances performance and fairness across clients. + +# C.2 Comparison to More State-of-the-Arts Methods in the Homogeneous Setting + +To further validate the generality and effectiveness of PHP-FL, we conduct additional experiments in the homogeneous setting and report the results in Table 6. Specifically, all clients adopt the identical GoogLeNet architectures [60]. Additional compared methods include FedFV [41] and FedHEAL [28], which are designed to improve client fairness (Fair-FL), as well as FedAWE [20] and FedAU [45], which address client unavailability (CU-FL). As shown in Table 6, PHP-FL achieves the highest average accuracy (AM) across all patterns, while also maintaining a highly competitive performance fairness (FM). Although FedHEAL exhibits slightly better fairness, its average accuracy is significantly inferior compared to PHP-FL. This highlights that PHP-FL not only delivers the highest average performance but also achieves fairness that is highly competitive with the best fair-FL methods, striking an exceptional balance that surpasses existing methods in overall effectiveness even in the homogeneous setting. These results underscore the robustness and practicality of PHP-FL even in homogeneous FL scenarios, reaffirming its overall superiority. + +![](images/8c4f975cad96d6d0cda199a3d62e8dea59c72d9873f907600384820196cefb38.jpg) +(a) Average accuracy + +![](images/a0919d62003ca9d098842a7969b9cfb5e978d8f2c81677a5d4f0d87f3d752ccc.jpg) +(b) Standard deviation +Figure 7: Comparison of training curves on Fashion-MNIST. + +![](images/d11b357e899b4dc585a29731976d4ea96a5195d0088ccbe515e1c5199d671074.jpg) +(a) Average accuracy +Figure 8: Comparison of training curves on CIFAR-10. + +![](images/8c8867f2dc3cc98ab764453a9688d4e7c251980c49348fd9e373abc23951a7e4.jpg) +(b) Standard deviation + +Table 6: Comparison with the state-of-the-art methods on CIFAR-10 in the homogeneous setting. Best in bold and second with underline. + +
TypeMethodsUniform [a = 0.5]Normal [μ = 0.5, σ = 0.2]Linear [a = 0.05, d = K-2/K(K-1)]
AM (%) ↑FM (%) ↓AM (%) ↑FM (%) ↓AM (%) ↑FM (%) ↓
MH-FLFML [arXiv20]41.31±0.4613.28±0.3141.00±0.4813.46±0.3539.21±2.8013.37±0.30
FedGen [ICML21]61.79±0.138.80±0.4661.73±0.158.95±0.2561.60±0.298.99±0.20
FedKD [NC22]61.26±0.0810.03±0.1060.96±0.389.93±0.1160.73±0.7010.01±0.09
FedAPEN [KDD23]66.49±0.308.14±0.1666.54±0.318.26±0.1866.60±0.358.25±0.17
FedTGP [AAAI24]60.12±0.489.97±0.4660.15±0.4610.20±0.3159.88±0.6910.14±0.31
FedMRL [NIPS24]62.86±0.628.60±0.1662.86±0.628.95±0.5862.58±0.548.80±0.38
Fair-FLFedFV [IJCAI21]59.75±0.8711.03±1.6360.05±1.3110.54±2.0258.32±1.4511.08±1.60
FedHEAL [CVPR24]59.43±0.837.03±0.7659.36±0.757.93±1.1459.15±0.577.94±1.15
CU-FLFedAWE [NIPS24]62.75±0.099.39±0.3062.68±0.139.31±0.2562.54±0.319.36±0.28
FedAU [ICLR24]62.79±0.209.17±0.1962.72±0.259.21±0.2962.59±0.479.35±0.15
AllPHP-FL (Ours)67.99±0.168.05±0.1967.88±0.328.08±0.1667.84±0.138.12±0.12
+ +# C.3 Performance with Varying Numbers of Clients + +In this Section, we compare the performance of different methods with varying numbers of clients on CIFAR-10. Specifically, the uniform pattern involves 10 clients with full participation in every round; the normal pattern consists of 50 clients whose participation probabilities are sampled from a normal distribution $(\mu = 0.2, \sigma = 0.2)$ , resulting in an average of 10 clients per round; and the linear pattern also includes 50 clients, with participation probabilities increasing linearly from 0.02 by a step of $\frac{0.36}{K - 1}$ , yielding the same average of 10 clients per round. As shown in Table 7, PHP-FL consistently achieves the best accuracy (AM) and performance fairness (FM) across all three patterns. Notably, compared to the strongest baseline FedAPEN, our method improves AM by $6.62\%$ , $6.47\%$ , and $6.12\%$ under the uniform, normal, and linear settings, respectively, while also reducing FM, demonstrating superior robustness and fairness with varying numbers of clients. + +Table 7: Comparison with the state-of-the-art methods with varying numbers of clients on CIFAR-10 in the heterogeneous setting. Best in bold and second with underline. + +
MethodsUniform [a = 1.0]Normal [μ = 0.2, σ = 0.2]Linear [a = 0.02, d = 0.36/K-1]
AM (%) ↑FM (%) ↓AM (%) ↑FM (%) ↓AM (%) ↑FM (%) ↓
Standalone58.89±0.2010.53±0.2450.73±1.2914.04±1.4051.81±0.3012.81±0.47
FML [arXiv20]46.61±0.6115.18±0.3436.59±1.1717.35±1.2235.26±0.2217.10±0.16
FedGen [ICML21]57.82±0.069.45±0.2943.70±1.0315.10±1.2643.35±1.0814.92±0.53
FedKD [NC22]57.46±0.4511.26±0.3349.27±0.9313.98±0.6349.07±0.6213.99±0.71
FedAPEN [KDD23]63.90±0.109.05±0.5152.99±1.2313.68±1.7753.76±0.3912.74±0.27
FedTGP [AAAI24]58.45±0.699.88±0.3740.15±0.3214.74±0.6938.18±1.0416.08±1.48
FedMRL [NIPS24]58.33±0.2014.23±0.6349.32±1.3413.64±1.0149.26±0.7413.26±0.27
PHP-FL (Ours)70.52±0.037.97±0.4959.46±0.8312.69±1.4159.88±0.2611.22±0.37
+ +Table 8: Comparison results on CIFAR-10 dataset under the Markovian participation pattern. All other settings follow their default configurations. + +
MethodStandaloneFMLFedGenFedKDFedAPENFedTGPFedMRLPHP-FL (Ours)
AM (%) ↑52.77±1.4442.73±0.9651.93±0.8451.31±0.8057.84±0.5350.55±1.7949.02±1.7464.04±0.73
FM (%) ↓16.10±2.5019.29±2.0414.91±0.3816.82±1.4513.99±3.3412.14±1.6615.86±0.0811.69±1.35
+ +# C.4 Performance under the Markovian Participation Pattern + +To better reflect dynamic client availability in real-word scenarios, we have conducted additional experiments under the Markovian participation pattern following FedAU [45]. In this pattern, each client follows a two-state Markov chain to determine its participation status in each training round, where the two states correspond to "participating" and "not participating" in the current training round. This modeling introduces temporal correlation in client participation behavior while maintaining sufficient randomness, offering a more realistic simulation compared to independently sampled participation patterns. + +For the parameter settings of the Markovian pattern experiment, we constrain the maximum transition probability from the non-participating state to the participating state to 0.05, thereby avoiding excessively frequent state changes and ensuring realistic participation dynamics. The initial state of each client's Markov chain is randomly sampled according to the stationary probability, which is set to 0.5 to ensure that approximately half of the clients participate in each round on average. The transition probabilities are carefully calibrated to maintain this stationary distribution throughout the training process, ensuring system stability while introducing participation heterogeneity across clients. As shown in Table 8, the experimental results under the Markovian participation pattern further validate PHP-FLs robustness, consistently achieving superior accuracy and fairness despite the increased dynamic and unpredictable client availability. + +# C.5 Effectiveness of Adaptive Selective Updates in the ISPU Module + +To further validate the design rationality of the adaptive selective parameter update mechanism in our ISPU module, we conduct an ablation study comparing different update strategies for the local auxiliary model $g$ , focusing on the update ratio $\alpha$ . Specifically, we evaluate the following variants: + +- Fixed update. Updates a fixed ratio of the most important parameters, with $\alpha$ set to a constant and we adopt $\alpha = 0.5$ . +- Full Update. Updates all parameters without selection, which is equivalent to $\alpha = 1$ . +- Random Update. All parameters are stochastically updated with probability $\alpha$ , where $\alpha$ is dynamically determined by our proposed method using Eq. 9. +- Adaptive update (Ours). Dynamically adjusts $\alpha$ using Eq. 9 based on client participation history and parameter importance. + +As shown in Table 9, our adaptive update mechanism consistently achieves the best accuracy and performance fairness. Full update shows the poorest performance, and while fixed Update offers some stability, it is surpassed by the adaptive methods. Random update achieves competitive accuracy but inferior fairness compared to our method. The findings highlight the superiority of our adaptive selective update mechanism within the ISPU module. + +Table 9: Effectiveness of Adaptive Selective Updates in the ISPU module. Results are evaluated under the linear participation pattern on Fashion-MNIST and CIFAR-10. Best in bold. + +
Different variantsFashion-MNISTCIFAR-10
AM (%) ↑FM (%) ↓AM (%) ↑FM (%) ↓
Fixed Update (α = 0.5)97.434.4164.119.14
Full Update (α = 1.0)96.825.7362.818.65
Random Update (dynamic α)97.286.2365.978.97
Adaptive Update (Ours with dynamic α)97.584.2966.788.09
+ +# C.6 Effect of the Sharpness Factor on Performance + +In our proposed PHP-FL, the hyperparameter $\delta$ in Eq. 9 controls the sharpness of the mapping from local participation frequency to the update ratio $\alpha$ in the ISPU module. A larger $\delta$ causes $\alpha$ to approach 1 for frequently participating clients and 0 for infrequent ones, whereas a smaller $\delta$ smooths the adjustment, pushing $\alpha$ toward 0.5. As shown in Table 10, performance is relatively stable across a range of $\delta$ values, but we observe that $\delta = 5$ consistently achieves near-optimal results in both accuracy and fairness across Fashion-MNIST and CIFAR-10. Therefore, we set $\delta = 5$ in our experimental configuration. + +Table 10: Effect of the hyperparameter $\delta$ on performance. Best in bold and second with underline. + +
DatasetMetricδ = 0.1δ = 0.5δ = 1δ = 5δ = 10δ = 50δ = 100
Fashion-MNISTAM (%) ↑97.5497.6297.5997.6197.5997.5297.51
FM (%) ↓4.254.134.394.124.604.354.43
CIFAR-10AM (%) ↑67.2467.2867.0667.2766.3667.1766.82
FM (%) ↓8.448.108.128.018.518.498.55
+ +# C.7 Cost and Efficiency Analysis + +Communication Cost. We compare the per-round communication cost with baselines in terms of the number of parameters transmitted between 20 clients and the server on CIFAR-10. As shown in Table 11, among the evaluated baselines, methods such as FedGen and FedTGP demonstrate significantly lower communication costs due to their use of partial model sharing and lightweight + +prototypes. Unfortunately, methods relying on auxiliary model transmission (e.g., FML, FedKD, FedAPEN, and FedMRL) exhibit communication cost exceeding 200M parameters per round. In contrast, PHP-FL requires only 112.52M parameters per round, which is nearly half the cost of other auxiliary model-based approaches such as FedAPEN and FedMRL. This reduction is primarily due to clients downloading only the pruned global auxiliary model parameters instead of the full model. + +Computation Cost. We also report the total computation cost per round across all clients in terms of FLOPs (floating-point operations), as summarized in Table 11. Following [12], other operations such as data preprocessing are not included in the FLOPs calculation. To ensure a fair comparison, this experiment involves 20 clients with full participation in each round. All other configurations remain aligned with the main experiments. + +According to Table 11, PHP-FL incurs the per-round computation cost at 813.71GFLOPs. This increase is marginal when compared with other auxiliary model-based methods such as FML (753.51G), FedKD (753.51G), FedAPEN (771.01G), and FedMRL (757.34G). Compared to FedGen (391.38G) and FedTGP (387.99G), the increased computation cost primarily arises from the training of additional auxiliary models. Besides, the slightly higher cost of the proposed PHP-FL is mainly attributed to the extra training required for ensemble weights. While FedTGP achieves the lowest computation cost, its accuracy significantly lags. Despite this slight increase in cost compared with other auxiliary model-based methods, PHP-FL achieves significantly the best results in both accuracy and performance fairness, as shown in the main experiments. The results underscore a compelling trade-off, with our method delivering notable performance gains at the cost of only a slight increase in computation. + +Table 11: Comparison of communication and computation costs per round on CIFAR-10, where communication cost is measured by the number of parameters transmitted between 20 clients and the server, and computation cost is evaluated as the total number of FLOPs executed across all 20 clients. 'M' and 'G' denote million and giga, respectively. For FedKD, the SVD computation cost is excluded from this analysis. + +
MethodCommunication CostComputation Cost
FML [arXiv20]224.20M753.51G
FedGen [ICML21]5.92M391.38G
FedKD [NC22]200.26M753.51G
FedAPEN [KDD23]224.20M771.01G
FedTGP [AAAI24]0.192M387.99G
FedMRL [NeurIPS24]224.05M757.34G
PHP-FL (Ours)112.52M813.71G
+ +Efficiency Analysis. In addition, we conducted further experiments to measure the total costs (communication rounds, computation cost, and communication cost) required to reach $50\%$ accuracy compared to baselines. As shown in Figure 9, while PHP-FL's per-round communication cost is higher than non auxiliary model-based methods (e.g., FedGen and FedTGP) due to the transmission of auxiliary model, it reaches target accuracy in just 2 rounds, whereas all baselines require at least 5 rounds. Consequently, the total communication and computation costs are substantially lower. Furthermore, the per-round costs can be readily optimized in practice by employing smaller auxiliary models or mask quantization (e.g., bitwidth reduction). Thus, PHP-FL offers a far more efficient path to convergence in practical deployments. + +# D Discussion + +Communication Cost. In the auxiliary model-based methods, active clients often need to upload $L = |g|$ parameters of the global auxiliary model in each communication round, which are typically represented in full-precision floating-point format. Our method similarly requires uploading + +![](images/d31f014135a825e1e6b9238d2e5a0c02256d40226798200dde6f723ce004fdac.jpg) + +![](images/f2fd463459bb2939c7128a12a6e9bfa83ab465c62b0691073dda82557c3caaee.jpg) +Figure 9: From left to right: Communication rounds, total number of transmitted parameters, and computation FLOPs required to achieve the $50\%$ accuracy on CIFAR-10 under the uniform pattern. + +![](images/049aad7e94b1cc490ef1f350a4e5e82d73f76da84ca9f74a11282d2dfe3f6cc1.jpg) + +![](images/1b19f7c92c4d99ab54ee2912c4dd8438ed804598cd3b4b98cf58731ef3d82cb7.jpg) + +the global auxiliary model but with an additional binary mask matrix $M$ of the same dimension $L$ . Fortunately, since each element in $M$ is a binary value (0 or 1), it only incues 1 bit per parameter, rendering the added communication overhead negligible compared to the transmission of full-precision parameters. Moreover, when downloading the global auxiliary model, clients only need to receive $\alpha \cdot L (\alpha \in (0,1))$ parameters, where the pruning threshold $\tau$ in Eq 9 can be adjusted as needed. This flexibility allows us to further reduce communication costs dynamically. + +Participation Patterns. We clearly state that we make no assumptions about the distribution of client participation patterns and allow them to be arbitrary throughout the training process. Moreover, similar to [45], we do not require any prior knowledge of the client sampling process for the proposed method PHP-FL. + +Privacy. Similar to FedAvg [1], PHP-FL does not require sharing raw data or client-specific heterogeneous model parameters. Instead, only the lightweight, homogeneous auxiliary models and the mask matrix related to model parameters are uploaded to the server. This design ensures that sensitive local model structures and data remain on the client side, making PHP-FL suitable for privacy-critical applications. Therefore, PHP-FL is compatible with standard privacy-preserving mechanisms such as secure aggregation [67]. Differentially private variants of FedAvg [68] can be seamlessly integrated similarly. \ No newline at end of file diff --git a/NeurIPS/2025/A Fair Federated Learning Method for Handling Client Participation Probability Inconsistencies in Heterogeneous Environments/images.zip b/NeurIPS/2025/A Fair Federated Learning Method for Handling Client Participation Probability Inconsistencies in Heterogeneous Environments/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..6d6872f0668cb09f376e62e89e896f9eedcadcf4 --- /dev/null +++ b/NeurIPS/2025/A Fair Federated Learning Method for Handling Client Participation Probability Inconsistencies in Heterogeneous Environments/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4e5c99964e8feb2657b06854093d77dfc18f61984f3c5bd6917d0d0c1be84257 +size 1186191 diff --git a/NeurIPS/2025/A Fair Federated Learning Method for Handling Client Participation Probability Inconsistencies in Heterogeneous Environments/layout.json b/NeurIPS/2025/A Fair Federated Learning Method for Handling Client Participation Probability Inconsistencies in Heterogeneous Environments/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..28d96b4223e7d939807b33352676a304eff4aed4 --- /dev/null +++ b/NeurIPS/2025/A Fair Federated Learning Method for Handling Client Participation Probability Inconsistencies in Heterogeneous Environments/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:df914e684a9b2415428b8ed13557aca7bf49e52caa38c9d006b41158995ccb60 +size 986729 diff --git a/NeurIPS/2025/A Few Moments Please_ Scalable Graphon Learning via Moment Matching/619abe32-5129-4292-9aae-88b74c01bc3d_content_list.json b/NeurIPS/2025/A Few Moments Please_ Scalable Graphon Learning via Moment Matching/619abe32-5129-4292-9aae-88b74c01bc3d_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..9456fc6f1147004ba5f8fe275cc86e8ebe9d338e --- /dev/null +++ b/NeurIPS/2025/A Few Moments Please_ Scalable Graphon Learning via Moment Matching/619abe32-5129-4292-9aae-88b74c01bc3d_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:decfc4ef180cd2beef8af314d2964bec0e2d1ffdb87f8669e91c0e60841d29a0 +size 223769 diff --git a/NeurIPS/2025/A Few Moments Please_ Scalable Graphon Learning via Moment Matching/619abe32-5129-4292-9aae-88b74c01bc3d_model.json b/NeurIPS/2025/A Few Moments Please_ Scalable Graphon Learning via Moment Matching/619abe32-5129-4292-9aae-88b74c01bc3d_model.json new file mode 100644 index 0000000000000000000000000000000000000000..c1447d23035e4f6d2af4e080fa07b094809c707e --- /dev/null +++ b/NeurIPS/2025/A Few Moments Please_ Scalable Graphon Learning via Moment Matching/619abe32-5129-4292-9aae-88b74c01bc3d_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:06ef54e705cd908ba0aa7fd641914856a1eb28b764cffb969d4ecf10affd3208 +size 276048 diff --git a/NeurIPS/2025/A Few Moments Please_ Scalable Graphon Learning via Moment Matching/619abe32-5129-4292-9aae-88b74c01bc3d_origin.pdf b/NeurIPS/2025/A Few Moments Please_ Scalable Graphon Learning via Moment Matching/619abe32-5129-4292-9aae-88b74c01bc3d_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..554c50ec6a4befacb044abb481b4ef794d7051e1 --- /dev/null +++ b/NeurIPS/2025/A Few Moments Please_ Scalable Graphon Learning via Moment Matching/619abe32-5129-4292-9aae-88b74c01bc3d_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:aa85a03f66d03b374df5bc8b89c3e813d64c85b8c7a2652c26aaa79563f54cc7 +size 2277506 diff --git a/NeurIPS/2025/A Few Moments Please_ Scalable Graphon Learning via Moment Matching/full.md b/NeurIPS/2025/A Few Moments Please_ Scalable Graphon Learning via Moment Matching/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e1964a27250b6faf18400a67f0de55801ab32d0d --- /dev/null +++ b/NeurIPS/2025/A Few Moments Please_ Scalable Graphon Learning via Moment Matching/full.md @@ -0,0 +1,1141 @@ +# A Few Moments Please: Scalable Graphon Learning via Moment Matching + +Reza Ramezanpour + +Rice University + +rr68@rice.edu + +Victor M. Tenorio + +King Juan Carlos University + +victor.tenorio@urjc.es + +Antonio G. Marques + +King Juan Carlos University + +antonio.garcia.parques@urjc.es + +Ashutosh Sabharwal + +Rice University + +ashu@rice.edu + +Santiago Segarra + +Rice University + +segarra@rice.edu + +# Abstract + +Graphons, as limit objects of dense graph sequences, play a central role in the statistical analysis of network data. However, existing graphon estimation methods often struggle with scalability to large networks and resolution-independent approximation, due to their reliance on estimating latent variables or costly metrics such as the Gromov-Wasserstein distance. In this work, we propose a novel, scalable graphon estimator that directly recovers the graphon via moment matching, leveraging implicit neural representations (INRs). Our approach avoids latent variable modeling by training an INR-mapping coordinates to graphon values-to match empirical subgraph counts (i.e., moments) from observed graphs. This direct estimation mechanism yields a polynomial-time solution and crucially sidesteps the combinatorial complexity of Gromov-Wasserstein optimization. Building on foundational results, we establish a theoretical guarantee: when the observed subgraph motifs sufficiently represent those of the true graphon (a condition met with sufficiently large or numerous graph samples), the estimated graphon achieves a provable upper bound in cut distance from the ground truth. Additionally, we introduce MomentMixup, a data augmentation technique that performs mixup in the moment space to enhance graphon-based learning. Our graphon estimation method achieves strong empirical performance-demonstrating high accuracy on small graphs and superior computational efficiency on large graphs-outperforming state-of-the-art scalable estimators in $75\%$ of benchmark settings and matching them in the remaining cases. Furthermore, MomentMixup demonstrated improved graph classification accuracy on the majority of our benchmarks. + +# 1 Introduction + +Networks are fundamental structures for representing complex relational data across diverse domains, from social interactions and biological systems to technological infrastructures [10, 31]. Understanding the underlying principles governing these networks is crucial for tasks such as link prediction [26], community detection [30], and, more broadly, node or graph classification [25]. Graphons, or graph limits, have emerged as a powerful mathematical framework for capturing the asymptotic structure of sequences of dense graphs [19, 18, 5, 2]. They provide a continuous, generative model for graphs, enabling principled statistical analysis and offering a canonical representation for large networks. Graphons have been successfully applied to derive controllers for large networks [12], to understand network games with many actors [23], to perform data augmentation in graph settings [21, 13], and + +to aid in the topology inference of partially observed graphs [28, 20]. As such, developing accurate and efficient methods for estimating graphons from observed network data is a central problem in network science and machine learning. + +Estimating graphons from finite, potentially noisy graph observations presents significant challenges. Many existing approaches suffer from computational scalability issues when applied to large networks [7, 35]. Furthermore, their resolution is limited by the size of the sample graphs, and obtaining a resolution-free approximation of the underlying continuous graphon can be difficult [7]. For instance, implicit neural representations (INRs) have been explored for graphon estimation due to their ability to model continuous functions [34]. However, estimating the latent variables of the nodes to train the INRs remains a challenge, and oftentimes the literature resorts to computationally demanding optimal transport-inspired losses, like the Gromov-Wasserstein (GW) distance for optimization. While recent scalable methods have made progress [3], there remains a need for estimators that combine high accuracy, computational efficiency, and direct graphon recovery without complex intermediate steps. + +In this paper, we introduce a novel and scalable approach for graphon estimation via moment matching, designed to overcome these prevalent limitations. Our method directly recovers the graphon by leveraging subgraph counts (graph moments) from observed data, thereby bypassing the need for latent variables and their associated complexities. We represent the graphon using an INR, a continuous function parameterized by a neural network that maps coordinates in $[0,1]^2$ to the corresponding graphon value. The parameters of this INR are learned by minimizing the discrepancy between the moments derived from the INR and the empirical moments computed from the input graph(s). This direct recovery strategy, crucially, leads to a polynomial-time estimation algorithm and does not rely on combinatorial GW distances, distinguishing it from approaches like IGNR [34]. Our approach is underpinned by a theoretical result, building upon foundational work on convergent graph sequences [5], which establishes that if the motifs (subgraph patterns) in the observed graph data sufficiently represent the motifs present in the true underlying graphon-a condition met with sufficiently large or numerous graph samples-then the cut distance between the estimated and true graphons is provably upper bounded. Additionally, we propose MomentMixup, a novel data augmentation technique that operates by interpolating graph moments between classes and then learning the corresponding mixed graphons, offering an improvement over existing mixup strategies in the graphon domain [21, 13]. + +Our contributions are threefold: + +1. We propose MomentNet, a scalable graphon estimator based on moment matching with INR, offering a resolution-free and estimation recovery mechanism. +2. We provide a theoretical guarantee linking the fidelity of motif representation in observed data to the estimation accuracy in terms of cut distance. +3. We introduce MomentMixup, an effective data augmentation method in the moment space for graphon-based learning tasks. + +The remainder of this paper is structured as follows: Section 2 presents the necessary background concepts and related works. Section 3 details our moment-matching INR approach for graphon estimation, including its theoretical characterization. Section 4 introduces MomentMixup, our approach for data augmentation in graph classification tasks. Section 5 presents our comprehensive empirical evaluations in both synthetic graphon estimation and data augmentation for graph classification. Finally, Section 6 concludes the paper and discusses future directions. + +# 2 Background, Related Works and Problem Formulation + +In this section, we introduce the foundational concepts of graphons, motif densities, INRs for graphon estimation, and mixup for data augmentation. We also formally state the graphon estimation problem addressed in this paper. In all these topics, we provide a summary of the literature, although a detailed discussion of related works can be found in Appendix A. + +Graphons A graphon, short for "graph function," is a fundamental concept in the theory of graph limits, serving as a limit object for sequences of dense graphs [18, 5]. Formally, a graphon $W$ is a symmetric measurable function $W:[0,1]^2\to [0,1]$ . Intuitively, the unit interval $[0,1]$ can be thought of as a latent space for the graph nodes. For any two points $x,y\in [0,1]$ (representing latent + +positions), the value $W(x,y)$ represents the probability of an edge forming between nodes associated with these latent positions. + +A random graph $G_{n}(W)$ with $n$ nodes can be generated from a graphon $W$ by sampling $n$ i.i.d. latent positions $\eta_1, \eta_2, \ldots, \eta_n \sim \mathcal{U}[0,1]$ and, for each pair of distinct nodes $(i,j)$ with $1 \leq i < j \leq n$ , an edge $(i,j)$ is included in $G_{n}(W)$ independently with probability $W(\eta_i, \eta_j)$ . Graphons are inherently invariant to permutations of node labels in the generated graphs, meaning that different orderings of the latent positions $\eta_i$ that preserve their relative positions in $[0,1]$ (or more formally, measure-preserving bijections of $[0,1]$ ) lead to equivalent graphon representations. The natural distance metric capturing this invariance is the cut distance [5]. + +Motif Densities from Graphons A key property of graphons is their ability to characterize the expected density of small subgraphs, often called motifs [5, 18]. For a simple graph $F$ (the motif), whose node and edge set are represented by $\mathcal{V}_F$ and $\mathcal{E}_F$ , respectively, with $k = |\mathcal{V}_F|$ , its homomorphism density in a graphon $W$ , denoted $t(F,W)$ , is defined as + +$$ +t (F, W) = \int_ {[ 0, 1 ] ^ {k}} \prod_ {(i, j) \in \mathcal {E} _ {F}} W \left(\eta_ {i}, \eta_ {j}\right) \prod_ {l \in \mathcal {V} _ {F}} d \eta_ {l}. \tag {1} +$$ + +This integral represents the probability that $k$ randomly chosen latent positions from $[0,1]$ induce a subgraph homomorphic to $F$ according to the edge probabilities defined by $W$ . For a sufficiently large graph $G$ sampled from $W$ , the empirical count of motif $F$ in $G$ , normalized appropriately, converges to $t(F,W)$ . Thus, empirical motif densities from observed graphs can serve as estimators for the true motif densities of the underlying graphon. The set of all such motif densities $\{t(F,W)\}_{F\in \mathcal{F}}$ (for some collection of motifs $\mathcal{F}$ ) is often referred to as the moment vector of the graphon [4]. We also introduce the induced motif densities as follows + +$$ +t ^ {\prime} (F, W) = \int_ {[ 0, 1 ] ^ {k}} \prod_ {(i, j) \in \mathcal {E} _ {F}} W (\eta_ {i}, \eta_ {j}) \prod_ {(i, j) \notin \mathcal {E} _ {F}} \left(1 - W \left(\eta_ {i}, \eta_ {j}\right)\right) \prod_ {l \in \mathcal {V} _ {F}} d \eta_ {i}. \tag {2} +$$ + +This formulation for induced motif density, $t'(F, W)$ , specifically counts instances where the motif $F$ appears in $W$ with an exact structural match. This means it accounts for both the required presence of edges specified in $F$ and the required absence of edges between the motif's vertices that are not in $F$ . In contrast, a non-induced (or homomorphism) density $t(F, W)$ only requires the presence of edges from $F$ in $W$ , without any assumption of the value of the graphon associated with pairs of nodes not linked by an edge. + +Implicit Neural Representations for Graphon Estimation An INR can effectively model a graphon by learning it as a continuous function [34, 3]. In this setup, the INR, typically a neural network $f_{\theta}:[0,1]^2 \to [0,1]$ , is trained to take pairs of latent node coordinates $(\eta_i,\eta_j)$ from a continuous space as input, where $\eta_{i}$ and $\eta_{j}$ represent the latent positions associated with entities $i$ and $j$ . Its output is the predicted value of the graphon $f_{\theta}(\eta_i,\eta_j) = \hat{W} (\eta_i,\eta_j)$ , representing the probability of an edge existing between these two latent positions. The network $f_{\theta}$ learns this mapping from observed samples, which could be $((\eta_{i_l},\eta_{j_l}),W(\eta_{i_l},\eta_{j_l}))$ pairs derived from a large graph or a target graphon function, for a set of sample indices $l$ . Crucially, because $f_{\theta}$ learns a continuous function over the entire input coordinate space defined by $\eta$ , the resulting graphon representation is inherently resolution-free. This means it can determine the edge probability for any arbitrary pair of latent coordinates $(\eta_i,\eta_j)$ , allowing for the generation or analysis of graph structures at any desired level of detail or scale without being tied to a fixed number of nodes or a specific discretization. + +Mixup for Data Augmentation The core idea of Mixup [40] is to generate synthetic training examples by taking convex combinations of pairs of existing samples and their corresponding labels. Given two input samples $x_{i}$ and $x_{j}$ with their respective labels $y_{i}$ and $y_{j}$ , a new synthetic sample $(\tilde{x},\tilde{y})$ is created as $\tilde{x} = \lambda x_{i} + (1 - \lambda)x_{j}$ , $\tilde{y} = \lambda y_{i} + (1 - \lambda)y_{j}$ . where $\lambda \in [0,1]$ is a mixing coefficient. This encourages the model to behave linearly in-between training examples, leading to smoother decision boundaries and improved generalization. + +Applying Mixup directly to graph-structured data presents challenges because graphs are not inherently Euclidean objects. To perform Mixup for graphs, one typically first maps the graphs into a suitable Euclidean representation [21, 13]. For example, GraphMAD [21] maps the graphs to a latent + +![](images/0cd4bfc10aee84b99fa475832f0819b9f17abc100669d258340d0825a5cd7ad8.jpg) +(a) Observed graphs +Figure 1: Graphon estimation pipeline: observed graphs lead to motif frequency extraction and INR-based recovery. + +![](images/9614f72c4ed6b98e1277bfa4a5beb85d17a7b7cf1fcbbf9336ef218679370e56.jpg) +(b) Density of Motifs + +![](images/6eee3ede1a9acf9f462e78bfcc1a571376d0f6ac3b00b72481addd7fae2f8592.jpg) +(c) Moment network training + +space and performs nonlinear mixup, while G-Mixup [13] performs mixup in the graphon domain. Once graphs $G_{i}$ and $G_{j}$ are available as Euclidean representations $\mathbf{z}_i$ and $\mathbf{z}_j$ respectively, a mixed representation $\tilde{\mathbf{z}} = \lambda \mathbf{z}_i + (1 - \lambda)\mathbf{z}_j$ can be computed. The subsequent step, which can be non-trivial, is to generate a new graph $\tilde{G}$ from this mixed representation $\tilde{\mathbf{z}}$ that can be used for training a graph classification model. + +Problem Formulation. The primary problem addressed in the graphon estimation literature, and in this work, is to recover the underlying graphon $W^{*}$ given one or more observed graphs. + +Problem 1 (Graphon Estimation). Given a set of observed graphs $\mathcal{G} = \{G_1, G_2, \ldots, G_P\}$ , where each $G_p$ has $n_p$ vertices and is assumed to be sampled (conditionally independently) from an unknown true graphon $W^*$ , i.e., $G_p \sim G_{n_p}(W^*)$ , the goal is to estimate $W^*$ . + +In the literature, early methods aimed at solving Problem 1 by means of histogram estimators and stochastic block models [5, 18, 6, 1, 11]. Other non-parametric approaches, like Universal Singular Value Thresholding (USVT) [7], aimed to recover underlying network structures but often faced computational or resolution limitations. More recent scalable techniques include those using INRs. For instance, IGNR [34] often leverages GW distances [24, 37, 35] for alignment, while methods like SIGL [3] further advance INR-based estimation. + +Our work proposes a novel method for solving Problem 1 by directly learning an INR to match empirical moments (subgraph counts) from the observed graph(s), thereby bypassing latent variables and computationally expensive metric optimizations. Moreover, we leverage our proposed solution to Problem 1 to design MomentMixup, a novel mixup strategy for graph data augmentation. MomentMixup performs mixup in the space of empirical moments, offering a novel way to generate augmented graph data informed by the underlying generative structure. + +# 3 Moment Matching Neural Network (MomentNet) + +In the following subsections, we introduce our proposed method, MomentNet, for learning the graphon. We also provide the fundamental theorem upon which our model is built. + +# 3.1 Methodology + +We explain the two steps in our method to estimate the graphon $W$ given the set of sampled graphs denoted by $\mathcal{G} = \{G_p\}_{p=1}^P$ . A schematic view of our method is presented in Figure 1. + +Step 1: Computing density of motifs. For each graph in our dataset $\mathcal{G}$ , we count the occurrences of specific motifs. The density of an identified motif is then calculated as the ratio of its observed count to the total number of possible ways that particular motif could appear in a graph of the same size. + +We use the ORCA method [15] to count the number of graphlets in the graph, and then we convert the graphlet count into motif counts. This aggregation is needed because our analysis cares only about how often each subgraph pattern appears in total, not about the exact placement of individual nodes within those patterns; see Fig. 1 in [15] for an illustration. We parallelize the use of the ORCA method for computing motif counts across the graphs in our dataset, thereby gaining a significant speed-up in processing time. ORCA can count motifs with up to five nodes, and its method can be extended to handle larger motifs. Once these motif-based statistics are calculated from the graphs, we no longer use the graphs themselves for subsequent steps. This approach significantly reduces computational overhead. Mathematically, we consider a set $\mathcal{F}$ of $|\mathcal{F}|$ distinct motif types. For each graph $G_{p}$ in our dataset, its empirical motif density vector is $\mathbf{m}^{(p)}\in \mathbb{R}^{|F|}$ . The overall motif density vector $\mathbf{m}$ for the dataset is currently computed as the simple average: + +$$ +\mathbf {m} = \frac {1}{P} \sum_ {p = 1} ^ {P} \mathbf {m} ^ {(p)}. \tag {3} +$$ + +While Eq. (3) treats each graph equally, a more general approach could involve a weighted average, $\mathbf{m}_w = \sum_{p=1}^{P} w_p \mathbf{m}^{(p)}$ (where $w_p \geq 0$ and $\sum_{p=1}^{P} w_p = 1$ ). Such weights $w_p$ could, for example, depend on graph properties like size ( $n_p$ ), potentially giving more influence to larger graphs, which might yield more stable density estimates. Our present work employs the simple average, with the exploration of weighted schemes as a potential future refinement. + +Step 2: Training the Moment network The moment network is defined as a combination of INR with a moment-based loss function. This step consists of three components described as follows: + +1. INR: Our methodology employs an INR $f_{\theta}$ to model the graphon, that receives the latent coordinates $(\eta_i, \eta_j)$ and outputs the estimated graphon value $\hat{W}_{\theta}(\eta_i, \eta_j)$ , as explained in Section 2. +2. Moment estimator: With the graphon estimated by the INR function $f_{\theta}$ , we can compute the induced motif density for any given motif $F$ . This is achieved by substituting $\hat{W}_{\theta}$ in place of $W$ in (2). Since we can not compute the integral directly, we approximate it using Monte Carlo integration techniques. By generating a sufficient number of random samples $L$ from the distribution induced by the graphon, we can estimate the integral. More precisely, we sample $L$ samples of $k$ latent coordinates $\eta_1^{(l)}, \ldots, \eta_k^{(l)}$ , where $\eta_i^{(l)} \sim \mathcal{U}[0,1]$ . Then we estimate $t'(F, \hat{W}_{\theta})$ as + +$$ +\hat {t} ^ {\prime} (F, \hat {W} _ {\theta}) = \frac {1}{L} \sum_ {l = 1} ^ {L} \left[ \prod_ {(i, j) \in \mathcal {E} _ {F}} \hat {W} _ {\theta} \left(\eta_ {i} ^ {(l)}, \eta_ {j} ^ {(l)}\right) \prod_ {(i, j) \notin \mathcal {E} _ {F}} \left(1 - \hat {W} _ {\theta} \left(\eta_ {i} ^ {(l)}, \eta_ {j} ^ {(l)}\right)\right) \right] \tag {4} +$$ + +The Monte Carlo estimator $\hat{t}^{\prime}(F,\hat{W}_{\theta})$ is differentiable with respect to the INR parameters $\theta$ . Since the INR $f_{\theta}$ is a neural network parameterized by $\theta$ (and thus differentiable with respect to $\theta$ ), the estimator $\hat{t}^{\prime}(F,\hat{W}_{\theta})$ , which is constructed as an average of terms derived from $f_{\theta}$ outputs at fixed sample points $\pmb{\eta}^{(l)}$ , is consequently also differentiable with respect to $\theta$ . This characteristic is vital as it allows the use of gradient-based optimization algorithms to train the INR parameters $\theta$ when this estimator is incorporated into a loss function. A proof of unbiasedness for this approach, i.e., showing that $\mathbb{E}[\hat{t} '(F,\hat{W}_\theta)] = t'(F,\hat{W}_\theta)$ , is provided in Appendix D. The vector of estimated moments (e.g., motif densities) derived from the INR outputs is denoted as $\hat{\mathbf{m}} (\theta) = \left[\hat{t} '(F_1,\hat{W}_\theta),\hat{t} '(F_2,\hat{W}_\theta),\dots ,\hat{t} '(F_{|\mathcal{F}|},\hat{W}_\theta)\right]^{\top}$ . + +3. WMSE: We use weighted mean squared error as a loss function to train our INR. Given the empirical moment vector $\mathbf{m}$ , based on sampled graphs and computed using Eq. (3), and the estimated moments based on the INR as $\hat{\mathbf{m}} (\theta)$ , the loss function is + +$$ +L (\theta) = \sum_ {i = 1} ^ {| \mathcal {F} |} w _ {i} \left(m _ {i} - \hat {m} _ {i} (\theta)\right) ^ {2}. \tag {5} +$$ + +In our experiments, we adjust the importance of different factors by assigning weights $(w_{i})$ . We calculate these weights as the inverse of how strong each factor $(m_i)$ appears in our data + +$(w_{i} = \frac{1}{m_{i}})$ . This weighting method balances the impact of each moment, preventing the most frequent ones (larger $m_{i}$ ) from having a large effect on the learning process. + +The training process described above, optimizing the parameters $\theta$ of the INR $f_{\theta}$ to minimize the weighted mean squared error between empirical and estimated motif densities, yields our final graphon estimate $\hat{W}_{\theta} = f_{\theta}$ . This estimated graphon is inherently scale-free due to the continuous nature of the INR. Furthermore, the entire estimation procedure operates in polynomial time with respect to the number of nodes and motifs considered. A detailed complexity analysis is provided in Appendix E. + +# 3.2 Theoretical characterization + +We present our main theorem bounding the cut distance between the true graphon $W^{*}$ and the graphon $\hat{W}_{\theta}$ estimated by our proposed INR. This result combines insights from the concentration of empirical motif densities in the $G_{n}(W)$ model [5] with the inverse counting lemma relating motif distances to cut distance, and an assumption about the neural network's ability to approximate empirical motif densities. Supporting lemmas and the proof of this theorem are provided in Appendix B. + +Let $G_1, \ldots, G_p$ be $P$ graphs, each with $n$ vertices, sampled independently from the graphon model $G_n(W^*)$ according to the graphon $W^*$ . The empirical motif density of $F$ based on these samples is $\bar{t}(F, W^*) = \frac{1}{P} \sum_{p=1}^{P} t(F, G_p)$ , where in a slight abuse of notation we denote by $t(F, G_p)$ the motifs densities computed from the motif counts of graph $G_p$ . + +We consider an INR $f_{\theta}$ with parameters $\theta$ , whose estimated graphon is denoted by $\hat{W}_{\theta}$ . The motif densities corresponding to this estimated graphon are denoted by $\hat{t}_{\theta}(F, \hat{W}_{\theta})$ . The INR is trained to directly output $\hat{t}_{\theta}(F, \hat{W}_{\theta})$ values that approximate the empirical densities $\bar{t}(F, W^{*})$ . Also, let $\mathcal{F}_k$ denote the set of all non-isomorphic simple graphs with exactly $k$ vertices and let $N_k = |\mathcal{F}_k|$ be the number of such graphs. As a preliminary step, we formalize the performance requirement we use to characterize our neural network next (see Appendix B.2 for a justification). + +Assumption 1 (Neural Network Approximation Capability). The parameters $\theta$ of the INR $\hat{W}_{\theta}$ are obtained such that for a fixed approximation error $\epsilon_{a} > 0$ , the estimated motif densities $\hat{t}_{\theta}(F, \hat{W}_{\theta})$ satisfy + +$$ +\left| \hat {t} _ {\theta} (F, \hat {W} _ {\theta}) - \bar {t} (F, W ^ {*}) \right| < \epsilon_ {a}, \quad f o r a l l F \in \mathcal {F} _ {k}. \tag {6} +$$ + +With the previous definitions, and those of Lemma 2 in Appendix B, we are in a position to present our main result, stated in Theorem 1. + +Theorem 1 (Cut Distance Bound for INR Estimated Graphons). Let $\epsilon_{a} > 0$ be the approximation error achieved by the network as stated in Assumption 1, and $\delta_{M} = 3^{-k^{2}}$ be the motif deviation threshold. Assume $n > \frac{k(k - 1)}{\delta_M}$ and + +$$ +N _ {k} \cdot 2 \exp \left(- \frac {P n}{4 k ^ {2}} \left(\frac {\delta_ {M}}{2} - \frac {k (k - 1)}{2 n}\right) ^ {2}\right) < \zeta , \tag {7} +$$ + +where $\zeta > 0$ is a desired confidence level. Then, with probability at least $1 - \zeta$ , the cut distance between the neural network estimated graphon $\hat{W}_{\theta}$ and the true graphon $W^{*}$ is bounded by $\eta = \frac{22C}{\sqrt{\log_2 k}}$ , with $C = \max \{1, \| W_1\|_{\infty}, \| W_2\|_{\infty}\}$ , as + +$$ +d _ {c u t} \left(\hat {W} _ {\theta}, W ^ {*}\right) < \eta . \tag {8} +$$ + +A detailed proof of Theorem 1, along with necessary definitions and supporting lemmas, can be found in Appendix B. This result demonstrates that if the INR can accurately approximate the empirical motif densities (Assumption 1), and if enough data (characterized by $P$ and $n$ ) is available to ensure the empirical motif densities are close to the true graphon motifs (Lemma 1), then the estimated graphon is likely to be close to the true graphon in cut distance. + +Although condition (7) may seem restrictive, note that (i) it decays exponentially with the number of graphs $P$ and their size $n$ considered, so it can be made arbitrarily small by considering larger datasets and (ii) although it increases with $k$ (and therefore with $N_{k}$ ), the size of the subgraphs considered $k$ is usually small (up to 5 nodes at most). + +# 4 Moment Mixup + +Data augmentation is a crucial technique in machine learning, particularly in domains like graph learning, where labeled data can be scarce or expensive to obtain [9]. By synthetically expanding the training dataset with new, plausible examples, data augmentation helps to improve model generalization, reduce overfitting, and enhance robustness. In the context of graph learning, developing effective augmentation strategies is challenging due to the complex, non-Euclidean nature of graph data, where direct analogies to image or text augmentation methods are not always feasible. + +In this section, we introduce MomentMixup, a novel approach for data augmentation in graph learning. The process begins by generating novel moment profiles through convex combinations of moment vectors, where each vector $\mathbf{m}_k$ is derived from sampled graphs belonging to a distinct graph class (e.g., $\mathbf{m}_{\mathrm{new}} = \sum \alpha_k\mathbf{m}_k$ , with $\alpha_{k}\geq 0,\sum \alpha_{k} = 1$ ). This interpolated moment vector, $\mathbf{m}_{\mathrm{new}}$ , is then used as the input to MomentNet, which subsequently defines a new graphon distribution, $W_{\mathrm{new}}(\eta_i,\eta_j)$ , consistent with these synthesized moments. Finally, new graphs are sampled from $W_{\mathrm{new}}(\eta_i,\eta_j)$ and integrated into the training set. The pseudocode of MomentMixup is provided in Algorithm 1 in Appendix G. + +Proposition 1. A convex combination of graphons is not equivalent to the corresponding convex combination of their vectors of moments, with the exception of the edge density moment. + +Proof of Proposition 1 using a counterexample is provided in Appendix F. MomentMixup is developed based on the key insight from Proposition 1. This foundational understanding distinguishes MomentMixup and offers it as an alternative to existing methods like G-Mixup [13]. The core intuition underpinning MomentMixup is that newly generated graph samples should exhibit clear structural proximity to a specific class (i.e., similar motif counts), thereby ensuring the augmented data reinforces class-specific structural characteristics. We contend that this particular intuition, that a generated sample is structurally close to a target class, may not always be reliably achieved through G-Mixup's graphon interpolation strategy because of Proposition 1. Furthermore, a detailed reproducibility analysis was unable to substantiate the original paper's claims regarding G-Mixup's superiority over other data augmentation methods [22]. + +# 5 Numerical Experiments + +In this section, we evaluate the performance of MomentNet and MomentMixup using various synthetic and real-world datasets widely used in the literature. The primary deep learning components of our experiments were executed on an Nvidia A100 GPU. Separately, empirical graph moments were computed using the ORCA toolkit [15], with its execution parallelized across an AMD EPYC 7742 64-Core Processor. + +# 5.1 MomentNet Evaluation + +To comprehensively evaluate our proposed MomentNet, we focus on two critical dimensions. First, we examine its effectiveness in the primary task of graphon estimation, determining how accurately it can capture the underlying distribution of graphs. Second, we address the practical applicability of our model by testing its scalability. This involves assessing its performance and runtime when applied to both large graphs (high number of nodes) and collections of smaller graphs, which are crucial considerations for real-world applications. We use $L = 20000$ , which is the number of samples to compute the density of moments of MomentNet using Eq. 4 in both experiments. + +# 5.1.1 Graphon Estimation + +We use the 13 graphon distributions used in the literature [3, 34]. The list of graphon distributions with their plot is provided in Appendix G. To build our experimental dataset, we adopt the graph generation approach utilized in [34]. From each graphon, we then generate 10 distinct graphs of varying sizes, specifically containing $\{75, 100, \dots, 275, 300\}$ nodes respectively. We treated the INR architecture as a hyperparameter to account for function complexity, noting that a simple one-layer MLP [17] with 64 neurons sufficed for non-complex graphons, while more complex cases (like Stochastic Block Models) required architectures such as SIREN [29] to accurately represent high-frequency details. This reflects the known limitations of modeling complex functions with small networks; + +![](images/ce96811b52789bc68bcf2a59590892b1a0883c6fc03e56eebeeb805a942ce99e.jpg) +(a) Performance comparison of MomentNet compared to other graphon estimation approaches. + +![](images/bca1e891d0353a89456c0422c6a46236a48dd295fd043a804f0aab32ee454566.jpg) +(b) Comparison of performance scalability of MomentNet with SIGL. + +![](images/c710ec871779490c7e3cc30b4e5f4f74d2883c0cceeb67893c9b12ae58bebe89.jpg) +(c) Comparison of runtime scalability of MomentNet with SIGL. +Figure 2: Overall comparison of MomentNet performance and scalability. + +the specific best-performing architecture for each graphon is provided in the supplementary code repository. For comparison, following INR training, we generate the graphon using 1000 uniformly sampled equidistant latent variables over the interval [0, 1]. The GW distance [36] is then computed between this estimate and the ground truth graphon. For our method, we implemented the same steps described in section 3, considering the motifs provided in Fig. 5. To evaluate the performance of our graphon estimation, we benchmark it against several established baseline methods. These include universal singular value thresholding (USVT) [7], sorting-and-smoothing (SAS) [6], implicit graph neural representation (IGNR) [34], Graph-Wasserstein barycenters (GWB) [35], and scalable implicit graphon learning (SIGL). For a consistent comparison, graphons estimated by the IGNR and SIGL baselines are sampled at a resolution of 1000, mirroring our own evaluation protocol. Furthermore, for SAS and USVT, we zero-pad the adjacency matrices of the observed graphs to this target resolution of 1000 before their respective graphon estimation procedures are applied; this input processing strategy is similar to those employed in [3, 34]. + +The results are provided in Fig. 2a. Based on the GW loss, our method outperforms the scalable state-of-the-art approach in 9 out of 13 graphons. Notably, our approach achieved superior results for graphons 10 and 11, where the state-of-the-art baseline (SIGL) struggled more. This difficulty might stem from the specific structures of graphons 10 and 11, which can challenge SIGL's reliance on accurately learning latent variables for its GNN-based node ordering and subsequent graphon estimation. Alongside the GW loss comparison, we assessed our graphon estimation via centrality measures, and the findings, detailed in Appendix J, affirmed our method's performance. + +# 5.1.2 Scalability Evaluation + +For experimental evaluations, we use the graphon $W(\eta_i,\eta_j) = 0.5 + 0.1\cos (\pi \eta_i)\cos (\pi \eta_j)$ (Figure 1) for generating graph instances across multiple independent realizations for each node size $n\in \{10,\dots ,510\}$ . In each realization, 10 graphs of size $n$ are generated; MomentNet's target motif counts are averaged from these, while SIGL processes them according to its methodology. + +Reported performance metrics are averaged over these realizations, allowing methods to leverage a comprehensive set of samples. + +The estimation error results (Figure 2b) show that MomentNet achieves strong performance, with error decreasing as $n$ increases. By leveraging multiple graph instances, MomentNet demonstrates near-optimal performance even on relatively small graphs, attributed to more accurate motif density estimation, aligning with theory. In contrast, SIGL's error, while node-dependent, does not substantially improve from multiple graph instances, offering only slight gains for small graphs, resulting in inferior overall performance. A potential explanation is SIGL's reliance on accurate latent variable estimation. The specific graphon $W(\eta_i,\eta_j) = 0.5 + 0.1\cos (\pi \eta_i)\cos (\pi \eta_j)$ (Figure 1) makes this challenging, as its construction ensures edge probabilities near 0.5, leading to high variance $W(1 - W)$ near maximum). This high variance can obscure latent structure. MomentNet's averaging directly reduces density estimate variance. However, for SIGL, if each of the 10 graphs individually fails to resolve latent positions due to high variance, more such graphs may not overcome this limitation as effectively as methods that directly average structural features. + +Regarding runtime (Figure 2c), MomentNet's average runtime, despite variance, scales more favorably with increasing nodes compared to SIGL, showing a clear advantage for $n > 300$ . The variance in MomentNet's runtime is due to its early stopping criteria (see Appendix E for detailed complexity analysis). Further experimental results are provided in Appendix K. + +# 5.1.3 Ablation Study: Choice of Moments + +The robustness of MomentNet is demonstrated by its strong performance using a relatively small, fixed set of motifs. We conducted an ablation study to investigate the impact of moment selection, with results for Graphons 2 and 4 (from Table 7) presented in Table 1. A key finding is that while performance generally improves as more motifs are added (indicated by lower GW distance), this trend is not strictly monotonic. We observed minor performance dips, for instance, after incorporating the seventh and eighth motifs for Graphon 2, a behavior also seen with Graphon 4. + +This non-monotonicity suggests that not all motifs contribute equally; some higher-order motifs may introduce statistical noise, as they are often rare and thus typically require more samples for accurate approximation. Critically, these fluctuations are minor within a strong overall trend, and adding motifs beyond the first six provides no significant advantage. Our experiments affirm that a practical and powerful graphon estimation can be achieved using a fixed set containing all motifs up to a small node count $k$ . While an adaptive motif selection strategy remains a valuable avenue for future exploration, the current approach is robust and highly effective. + +Table 1: Ablation of motif count vs. GW distance (Avg ± Std). Motif count indicates the cumulative number of motifs used (up to 15, which is all motifs of size ≤ 5). + +
Motif Count123456789101112131415
GW Graphon 2.089.100.036.031.029.021.024.027.026.022.023.020.020.018.020
Std Graphon 2.012.015.008.008.008.003.003.006.002.003.005.003.002.001.002
GW Graphon 4.151.060.037.018.017.014.018.018.013.011.012.010.010.010.010
Std Graphon 4.009.017.012.005.005.002.004.006.003.001.002.002.001.002.001
+ +# 5.2 MomentMixup Evaluation + +To evaluate the performance of our MomentMixup framework, we conducted graph classification experiments on several real-world datasets: AIDS [27], IMDB-Binary [39], IMDB-Multi [39], and Reddit-Binary [39]. Detailed descriptions of these datasets are provided in Appendix K. To ensure a fair comparison with prior work, we adopted the same data splitting methodology as reported in previous literature [3, 13]. For data augmentation, we treated $\alpha_{mix}$ , $N_{\mathrm{nodes}}$ , $N_{\mathrm{graph}}$ , and $N_{\mathrm{sample}}$ as hyperparameters in Algorithm 1 and the best parameters are provided in Appendix K. We employ the GIN architecture [38] as the graph classification model. + +Table 2 presents the model's accuracy on the test set. The results demonstrate that our method achieves a better performance gain over the standard G-Mixup approach on three datasets. As highlighted in the previous section, our method demonstrates a distinct advantage on datasets composed of smaller + +graphs, such as AIDS, where it notably outperforms techniques like SIGL. While our results on the Reddit-Binary dataset, which features very large graphs and where SIGL performs strongly, were influenced by the experimental choice of using a limited set of nine motifs, this contrast further illuminates a key insight: the optimal choice of mixup method can be highly dependent on graph characteristics, particularly size. Our approach appears particularly well-suited for capturing structural nuances in smaller graphs where fewer motifs can still provide rich representative information. + +Table 2: Classification accuracy of G-Mixup, MomentMixup, and baselines on different datasets. + +
DatasetIMDB-BIMDB-MREDD-BAIDS
#graphs1000150020002000
#classes2322
#avgnodes19.7713.00429.6315.69
#avgEdges96.5365.94497.7516.2
GIN
No Augmentation71.55±3.5348.83±2.7591.78±1.0998±1.2
G-Mixup w/ USVT71.94±3.0050.46±1.4991.32±1.5197.8±0.9
G-Mixup w/ SIGL73.95±2.6450.70±1.4192.25±1.4197.3±1
MomentMixup74.30±2.7050.95±1.9391.8±1.298.5±0.6
+ +# 6 Conclusions + +In this paper, we introduced a novel, scalable graphon estimator leveraging INRs via direct moment matching, called MomentNet. This approach bypasses latent variables and costly GW optimizations, offering a theoretically grounded, polynomial-time solution for estimating graphons from empirical subgraph counts, with provable guarantees on its accuracy. We further proposed MomentMixup, a new data augmentation technique that performs mixup in the moment space, then obtains the estimated graphon using MomentNet, and finally samples new graphs from this graphon. Our empirical results validate the effectiveness of our estimator, demonstrating superior or comparable performance against state-of-the-art methods in graphon estimation benchmarks, and show that MomentMixup improves graph classification accuracy by generating structurally meaningful augmented data. + +Despite its strengths, our method's reliance on a pre-selected set of moments for graphon estimation is a limitation; performance can degrade if these moments are insufficient or noisy. Additionally, modeling a single graphon (per class for MomentMixup) may not capture highly heterogeneous graph data. Future work could address these by developing adaptive moment selection techniques and exploring extensions to learn mixtures of graphons. Further enhancements include adapting our moment-based approach for attributed or dynamic networks and integrating feature learning into the estimation process, broadening the applicability of our framework. + +# 7 Acknowledgments + +This work was partially supported by the U.S. NSF under award CCF-2340481, the Spanish AEI (AEI/10.13039/501100011033) grant PID2023-149457OB-I00, and the Community of Madrid via grants IDEA-CM (TEC-2024/COM-89) and URJC/CAM F1180 and the ELLIS Madrid Unit. + +# References + +[1] Airoldi, E. M., Blei, D. M., Fienberg, S. E., and Xing, E. P. (2008). Mixed membership stochastic blockmodels. In Journal of Machine Learning Research, volume 9, pages 1981-2014. +[2] Avella-Medina, M., Parise, F., Schaub, M. T., and Segarra, S. (2020). Centrality measures for graphons: Accounting for uncertainty in networks. IEEE Transactions on Network Science and Engineering, 7(1):520-537. +[3] Azizpour, A., Zilberstein, N., and Segarra, S. (2025). Scalable implicit graphon learning. In The 28th International Conference on Artificial Intelligence and Statistics. + +[4] Borgs, C., Chayes, J., and Lovász, L. (2010). Moments of two-variable functions and the uniqueness of graph limits. Geometric and Functional Analysis, 19(6):1597-1619. +[5] Borgs, C., Chayes, J., Lovász, L., Sós, V., and Vesztergombi, K. (2008). Convergent sequences of dense graphs i: Subgraph frequencies, metric properties and testing. Advances in Mathematics, 219(6):1801-1851. +[6] Chan, S. and Airoldi, E. (2014). A consistent histogram estimator for exchangeable graph models. In Xing, E. P. and Jebara, T., editors, Proceedings of the 31st International Conference on Machine Learning, volume 32 of Proceedings of Machine Learning Research, pages 208-216, Beijing, China. PMLR. +[7] Chatterjee, S. (2015). Matrix estimation by universal singular value thresholding. The Annals of Statistics, 43(1). +[8] Cybenko, G. (1989). Approximation by superpositions of a sigmoidal function. Mathematics of control, signals, and systems, 2(4):303-314. +[9] Ding, K., Xu, Z., Tong, H., and Liu, H. (2022). Data augmentation for deep graph learning: A survey. +[10] Dong, X., Thanou, D., Toni, L., Bronstein, M., and Frossard, P. (2020). Graph signal processing for machine learning: A review and new perspectives. IEEE Signal Processing Magazine, 37(6):117-127. +[11] Gao, C., Lu, Y., and Zhou, H. H. (2015). Rate-optimal graphon estimation. The Annals of Statistics, 43(6):2624-2652. +[12] Gao, S. and Caines, P. E. (2020). Graphon control of large-scale networks of linear systems. IEEE Transactions on Automatic Control, 65(10):4090-4105. +[13] Han, X., Jiang, Z., Liu, N., and Hu, X. (2022). G-Mixup: Graph Data Augmentation for Graph Classification. In Proceedings of the 39th International Conference on Machine Learning (ICML), volume 162 of Proceedings of Machine Learning Research, pages 8450-8465. PMLR. +[14] Hornik, K., Stinchcombe, M., and White, H. (1989). Multilayer feedforward networks are universal approximators. Neural networks, 2(5):359-366. +[15] Hočevar, T. and Demšar, J. (2014). A combinatorial approach to graphlet counting. Bioinformatics, 30(4):559-565. +[16] Leshno, M., Lin, V. Y., Pinkus, A., and Schocken, S. (1993). Multilayer feedforward networks with a nonpolynomial activation function can approximate any function. Neural networks, 6(6):861-867. +[17] Liu, H.-T. D., Williams, F., Jacobson, A., Fidler, S., and Litany, O. (2022). Learning smooth neural functions via lipschitz regularization. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-13. +[18] Lovász, L. (2012). Large networks and graph limits, volume 60. American Mathematical Soc. +[19] Lovász, L. and Szegedy, B. (2006). Limits of dense graph sequences. Journal of Combinatorial Theory, Series B, 96(6):933-957. +[20] Navarro, M. and Segarra, S. (2022). Joint network topology inference via a shared graphon model. IEEE Transactions on Signal Processing, 70:5549-5563. +[21] Navarro, M. and Segarra, S. (2023). Graphmad: Graph mixup for data augmentation using data-driven convex clustering. In ICASSP 2023 - 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5. +[22] Omeragic, E. and Duranović, V. (2023). [re] \(\S\)mathcal{G}\( -mixup: Graph data augmentation for graph classification. In ML Reproducibility Challenge 2022. + +[23] Parise, F. and Ozdaglar, A. (2023). Graphon games: A statistical framework for network games and interventions. *Econometrica*, 91(1):191-225. +[24] Peyre, G., Cuturi, M., and Solomon, J. (2016). Gromov-wasserstein averaging of kernel and distance matrices. In International conference on machine learning, pages 2664-2672. PMLR. +[25] Rey, S., Navarro, M., Tenorio, V. M., Segarra, S., and Marques, A. G. (2025). Redesigning graph filter-based gnns to relax the homophily assumption. In 2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5. +[26] Rey, S., Tenorio, V. M., and Marqués, A. G. (2023). Robust graph filter identification and graph denoising from signal observations. IEEE Transactions on Signal Processing, 71:3651-3666. +[27] Riesen, K. and Bunke, H. (2008). Iam graph database repository for graph based pattern recognition and machine learning. In da Vitoria Lobo, N., Kasparis, T., Roli, F., Kwok, J. T., Georgiopoulos, M., Anagnostopoulos, G. C., and Loog, M., editors, Structural, Syntactic, and Statistical Pattern Recognition, pages 287-297, Berlin, Heidelberg. Springer Berlin Heidelberg. +[28] Roddenberry, T. M., Navarro, M., and Segarra, S. (2021). Network topology inference with graphon spectral penalties. In 2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 5390-5394. +[29] Sitzmann, V., Martel, J. N., Bergman, A. W., Lindell, D. B., and Wetzstein, G. (2020). Implicit neural representations with periodic activation functions. In Proc. NeurIPS. +[30] Su, X., Xue, S., Liu, F., Wu, J., Yang, J., Zhou, C., Hu, W., Paris, C., Nepal, S., Jin, D., Sheng, Q. Z., and Yu, P. S. (2024). A comprehensive survey on community detection with deep learning. IEEE Transactions on Neural Networks and Learning Systems, 35(4):4682-4702. +[31] Tenorio, V. M., Isufi, E., Leus, G., and Marques, A. G. (2025). Tracking network dynamics using probabilistic state-space models. In 2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5. +[32] Van Handel, R. (2014). Probability in high dimension. Lecture Notes (Princeton University), 2(3):2-3. +[33] Wang, Y., Wang, W., Liang, Y., Cai, Y., and Hooi, B. (2021). Mixup for node and graph classification. In Proceedings of the Web Conference 2021, WWW '21, page 3663-3674, New York, NY, USA. Association for Computing Machinery. +[34] Xia, X., Mishne, G., and Wang, Y. (2023). Implicit graphon neural representation. In Proceedings of The 26th International Conference on Artificial Intelligence and Statistics, volume 206 of Proceedings of Machine Learning Research, pages 10619-10634. PMLR. +[35] Xu, H., Luo, D., Carin, L., and Zha, H. (2021a). Learning graphons via structured gromov-wasserstein barycenters. Proceedings of the AAAI Conference on Artificial Intelligence, 35(12):10505-10513. +[36] Xu, H., Luo, D., Carin, L., and Zha, H. (2021b). Learning graphons via structured gromov-wasserstein barycenters. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 10505-10513. +[37] Xu, H., Luo, D., Zha, H., and Duke, L. C. (2019a). Gromov-Wasserstein learning for graph matching and node embedding. In Chaudhuri, K. and Salakhutdinov, R., editors, Proceedings of the 36th International Conference on Machine Learning, volume 97 of Proceedings of Machine Learning Research, pages 6932-6941. PMLR. +[38] Xu, K., Hu, W., Leskovec, J., and Jegelka, S. (2019b). How powerful are graph neural networks? +[39] Yanardag, P. and Vishwanathan, S. (2015). Deep graph kernels. In Proceedings of the 21th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, KDD '15, page 1365-1374, New York, NY, USA. Association for Computing Machinery. +[40] Zhang, H., Cisse, M., Dauphin, Y. N., and Lopez-Paz, D. (2018). mixup: Beyond empirical risk minimization. In International Conference on Learning Representations (ICLR). + +# NeurIPS Paper Checklist + +# 1. Claims + +Question: Do the main claims made in the abstract and introduction accurately reflect the paper's contributions and scope? + +Answer: [Yes] + +Justification: Our main claim is both written in the abstract and at the end of the introduction section. + +Guidelines: + +- The answer NA means that the abstract and introduction do not include the claims made in the paper. +- The abstract and/or introduction should clearly state the claims made, including the contributions made in the paper and important assumptions and limitations. A No or NA answer to this question will not be perceived well by the reviewers. +- The claims made should match theoretical and experimental results, and reflect how much the results can be expected to generalize to other settings. +- It is fine to include aspirational goals as motivation as long as it is clear that these goals are not attained by the paper. + +# 2. Limitations + +Question: Does the paper discuss the limitations of the work performed by the authors? + +Answer: [Yes] + +Justification: The primary constraints of this study, along with potential future research directions to tackle these, are detailed in the conclusion section. + +Guidelines: + +- The answer NA means that the paper has no limitation while the answer No means that the paper has limitations, but those are not discussed in the paper. +- The authors are encouraged to create a separate "Limitations" section in their paper. +- The paper should point out any strong assumptions and how robust the results are to violations of these assumptions (e.g., independence assumptions, noiseless settings, model well-specification, asymptotic approximations only holding locally). The authors should reflect on how these assumptions might be violated in practice and what the implications would be. +- The authors should reflect on the scope of the claims made, e.g., if the approach was only tested on a few datasets or with a few runs. In general, empirical results often depend on implicit assumptions, which should be articulated. +- The authors should reflect on the factors that influence the performance of the approach. For example, a facial recognition algorithm may perform poorly when image resolution is low or images are taken in low lighting. Or a speech-to-text system might not be used reliably to provide closed captions for online lectures because it fails to handle technical jargon. +- The authors should discuss the computational efficiency of the proposed algorithms and how they scale with dataset size. +- If applicable, the authors should discuss possible limitations of their approach to address problems of privacy and fairness. +- While the authors might fear that complete honesty about limitations might be used by reviewers as grounds for rejection, a worse outcome might be that reviewers discover limitations that aren't acknowledged in the paper. The authors should use their best judgment and recognize that individual actions in favor of transparency play an important role in developing norms that preserve the integrity of the community. Reviewers will be specifically instructed to not penalize honesty concerning limitations. + +# 3. Theory assumptions and proofs + +Question: For each theoretical result, does the paper provide the full set of assumptions and a complete (and correct) proof? + +# Answer: [Yes] + +Justification: Proofs, assumptions and lemmas are either explicitly mentioned in the result or provided in the appendix. + +# Guidelines: + +- The answer NA means that the paper does not include theoretical results. +- All the theorems, formulas, and proofs in the paper should be numbered and cross-referenced. +- All assumptions should be clearly stated or referenced in the statement of any theorems. +- The proofs can either appear in the main paper or the supplemental material, but if they appear in the supplemental material, the authors are encouraged to provide a short proof sketch to provide intuition. +- Inversely, any informal proof provided in the core of the paper should be complemented by formal proofs provided in appendix or supplemental material. +- Theorems and Lemmas that the proof relies upon should be properly referenced. + +# 4. Experimental result reproducibility + +Question: Does the paper fully disclose all the information needed to reproduce the main experimental results of the paper to the extent that it affects the main claims and/or conclusions of the paper (regardless of whether the code and data are provided or not)? + +# Answer: [Yes] + +Justification: We address this part in the appendix by explaining full details of neural net models and hyperparameters of our method. + +# Guidelines: + +- The answer NA means that the paper does not include experiments. +- If the paper includes experiments, a No answer to this question will not be perceived well by the reviewers: Making the paper reproducible is important, regardless of whether the code and data are provided or not. +- If the contribution is a dataset and/or model, the authors should describe the steps taken to make their results reproducible or verifiable. +- Depending on the contribution, reproducibility can be accomplished in various ways. For example, if the contribution is a novel architecture, describing the architecture fully might suffice, or if the contribution is a specific model and empirical evaluation, it may be necessary to either make it possible for others to replicate the model with the same dataset, or provide access to the model. In general, releasing code and data is often one good way to accomplish this, but reproducibility can also be provided via detailed instructions for how to replicate the results, access to a hosted model (e.g., in the case of a large language model), releasing of a model checkpoint, or other means that are appropriate to the research performed. +- While NeurIPS does not require releasing code, the conference does require all submissions to provide some reasonable avenue for reproducibility, which may depend on the nature of the contribution. For example + +(a) If the contribution is primarily a new algorithm, the paper should make it clear how to reproduce that algorithm. +(b) If the contribution is primarily a new model architecture, the paper should describe the architecture clearly and fully. +(c) If the contribution is a new model (e.g., a large language model), then there should either be a way to access this model for reproducing the results or a way to reproduce the model (e.g., with an open-source dataset or instructions for how to construct the dataset). +(d) We recognize that reproducibility may be tricky in some cases, in which case authors are welcome to describe the particular way they provide for reproducibility. In the case of closed-source models, it may be that access to the model is limited in some way (e.g., to registered users), but it should be possible for other researchers to have some path to reproducing or verifying the results. + +# 5. Open access to data and code + +Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material? + +Answer: [Yes] + +Justification: The supplementary material includes the code used in the experiments, and we will also upload it to GitHub after the review process if the paper gets accepted. + +Guidelines: + +- The answer NA means that paper does not include experiments requiring code. +- Please see the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details. +- While we encourage the release of code and data, we understand that this might not be possible, so "No" is an acceptable answer. Papers cannot be rejected simply for not including code, unless this is central to the contribution (e.g., for a new open-source benchmark). +- The instructions should contain the exact command and environment needed to run to reproduce the results. See the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details. +- The authors should provide instructions on data access and preparation, including how to access the raw data, preprocessed data, intermediate data, and generated data, etc. +- The authors should provide scripts to reproduce all experimental results for the new proposed method and baselines. If only a subset of experiments are reproducible, they should state which ones are omitted from the script and why. +- At submission time, to preserve anonymity, the authors should release anonymized versions (if applicable). +- Providing as much information as possible in supplemental material (appended to the paper) is recommended, but including URLs to data and code is permitted. + +# 6. Experimental setting/details + +Question: Does the paper specify all the training and test details (e.g., data splits, hyperparameters, how they were chosen, type of optimizer, etc.) necessary to understand the results? + +Answer: [Yes] + +Justification: The full details of the experiments are written in the appendix. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The experimental setting should be presented in the core of the paper to a level of detail that is necessary to appreciate the results and make sense of them. +- The full details can be provided either with the code, in appendix, or as supplemental material. + +# 7. Experiment statistical significance + +Question: Does the paper report error bars suitably and correctly defined or other appropriate information about the statistical significance of the experiments? + +Answer: [Yes] + +Justification: All experiments were conducted over multiple trials, and the resulting error metrics are reported as their mean and standard deviation. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The authors should answer "Yes" if the results are accompanied by error bars, confidence intervals, or statistical significance tests, at least for the experiments that support the main claims of the paper. +- The factors of variability that the error bars are capturing should be clearly stated (for example, train/test split, initialization, random drawing of some parameter, or overall run with given experimental conditions). + +- The method for calculating the error bars should be explained (closed form formula, call to a library function, bootstrap, etc.) +- The assumptions made should be given (e.g., Normally distributed errors). +- It should be clear whether the error bar is the standard deviation or the standard error of the mean. +- It is OK to report 1-sigma error bars, but one should state it. The authors should preferably report a 2-sigma error bar than state that they have a $96\%$ CI, if the hypothesis of Normality of errors is not verified. +- For asymmetric distributions, the authors should be careful not to show in tables or figures symmetric error bars that would yield results that are out of range (e.g. negative error rates). +- If error bars are reported in tables or plots, The authors should explain in the text how they were calculated and reference the corresponding figures or tables in the text. + +# 8. Experiments compute resources + +Question: For each experiment, does the paper provide sufficient information on the computer resources (type of compute workers, memory, time of execution) needed to reproduce the experiments? + +Answer: [Yes] + +Justification: At the beginning of the experiments section, we detail the hardware used to run each part of the model and also present a plot of our method's runtime. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The paper should indicate the type of compute workers CPU or GPU, internal cluster, or cloud provider, including relevant memory and storage. +- The paper should provide the amount of compute required for each of the individual experimental runs as well as estimate the total compute. +- The paper should disclose whether the full research project required more compute than the experiments reported in the paper (e.g., preliminary or failed experiments that didn't make it into the paper). + +# 9. Code of ethics + +Question: Does the research conducted in the paper conform, in every respect, with the NeurIPS Code of Ethics https://neurips.cc/public/EthicsGuidelines? + +Answer: [Yes] + +Justification: We confirm that we have read, understood, and adhered to the applicable code of ethics. + +Guidelines: + +- The answer NA means that the authors have not reviewed the NeurIPS Code of Ethics. +- If the authors answer No, they should explain the special circumstances that require a deviation from the Code of Ethics. +- The authors should make sure to preserve anonymity (e.g., if there is a special consideration due to laws or regulations in their jurisdiction). + +# 10. Broader impacts + +Question: Does the paper discuss both potential positive societal impacts and negative societal impacts of the work performed? + +Answer: [Yes] + +Justification: We discuss the societal impact in Appendix L. + +Guidelines: + +- The answer NA means that there is no societal impact of the work performed. +- If the authors answer NA or No, they should explain why their work has no societal impact or why the paper does not address societal impact. + +- Examples of negative societal impacts include potential malicious or unintended uses (e.g., disinformation, generating fake profiles, surveillance), fairness considerations (e.g., deployment of technologies that could make decisions that unfairly impact specific groups), privacy considerations, and security considerations. +- The conference expects that many papers will be foundational research and not tied to particular applications, let alone deployments. However, if there is a direct path to any negative applications, the authors should point it out. For example, it is legitimate to point out that an improvement in the quality of generative models could be used to generate deepfakes for disinformation. On the other hand, it is not needed to point out that a generic algorithm for optimizing neural networks could enable people to train models that generate Deepfakes faster. +- The authors should consider possible harms that could arise when the technology is being used as intended and functioning correctly, harms that could arise when the technology is being used as intended but gives incorrect results, and harms following from (intentional or unintentional) misuse of the technology. +- If there are negative societal impacts, the authors could also discuss possible mitigation strategies (e.g., gated release of models, providing defenses in addition to attacks, mechanisms for monitoring misuse, mechanisms to monitor how a system learns from feedback over time, improving the efficiency and accessibility of ML). + +# 11. Safeguards + +Question: Does the paper describe safeguards that have been put in place for responsible release of data or models that have a high risk for misuse (e.g., pretrained language models, image generators, or scraped datasets)? + +Answer: [NA] + +Justification: We don't use those data in our experiments. + +Guidelines: + +- The answer NA means that the paper poses no such risks. +- Released models that have a high risk for misuse or dual-use should be released with necessary safeguards to allow for controlled use of the model, for example by requiring that users adhere to usage guidelines or restrictions to access the model or implementing safety filters. +- Datasets that have been scraped from the Internet could pose safety risks. The authors should describe how they avoided releasing unsafe images. +- We recognize that providing effective safeguards is challenging, and many papers do not require this, but we encourage authors to take this into account and make a best faith effort. + +# 12. Licenses for existing assets + +Question: Are the creators or original owners of assets (e.g., code, data, models), used in the paper, properly credited and are the license and terms of use explicitly mentioned and properly respected? + +Answer: [Yes] + +Justification: Methods adopted from the literature were either re-implemented by us based on their published descriptions, with due credit given to the original sources via citation, or, where publicly available code from the original authors was utilized, its use is acknowledged in our GitHub repository. + +Guidelines: + +- The answer NA means that the paper does not use existing assets. +- The authors should cite the original paper that produced the code package or dataset. +- The authors should state which version of the asset is used and, if possible, include a URL. +- The name of the license (e.g., CC-BY 4.0) should be included for each asset. +- For scraped data from a particular source (e.g., website), the copyright and terms of service of that source should be provided. + +- If assets are released, the license, copyright information, and terms of use in the package should be provided. For popular datasets, paperswithcode.com/datasets has curated licenses for some datasets. Their licensing guide can help determine the license of a dataset. +- For existing datasets that are re-packaged, both the original license and the license of the derived asset (if it has changed) should be provided. +- If this information is not available online, the authors are encouraged to reach out to the asset's creators. + +# 13. New assets + +Question: Are new assets introduced in the paper well documented and is the documentation provided alongside the assets? + +Answer: [Yes] + +Justification: Our code will be published on GitHub, with the URL provided in this paper. Additionally, all datasets will be made available on GitHub. + +Guidelines: + +- The answer NA means that the paper does not release new assets. +- Researchers should communicate the details of the dataset/code/model as part of their submissions via structured templates. This includes details about training, license, limitations, etc. +- The paper should discuss whether and how consent was obtained from people whose asset is used. +- At submission time, remember to anonymize your assets (if applicable). You can either create an anonymized URL or include an anonymized zip file. + +# 14. Crowdsourcing and research with human subjects + +Question: For crowdsourcing experiments and research with human subjects, does the paper include the full text of instructions given to participants and screenshots, if applicable, as well as details about compensation (if any)? + +Answer: [NA] + +Justification: This work does not involve crowdsourcing nor research with human subjects. + +Guidelines: + +- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. +- Including this information in the supplemental material is fine, but if the main contribution of the paper involves human subjects, then as much detail as possible should be included in the main paper. +- According to the NeurIPS Code of Ethics, workers involved in data collection, curation, or other labor should be paid at least the minimum wage in the country of the data collector. + +# 15. Institutional review board (IRB) approvals or equivalent for research with human subjects + +Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals (or an equivalent approval/review based on the requirements of your country or institution) were obtained? + +Answer: [NA] + +Justification: This work does not involve crowdsourcing nor research with human subjects. + +Guidelines: + +- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. +- Depending on the country in which research is conducted, IRB approval (or equivalent) may be required for any human subjects research. If you obtained IRB approval, you should clearly state this in the paper. + +- We recognize that the procedures for this may vary significantly between institutions and locations, and we expect authors to adhere to the NeurIPS Code of Ethics and the guidelines for their institution. +- For initial submissions, do not include any information that would break anonymity (if applicable), such as the institution conducting the review. + +# 16. Declaration of LLM usage + +Question: Does the paper describe the usage of LLMs if it is an important, original, or non-standard component of the core methods in this research? Note that if the LLM is used only for writing, editing, or formatting purposes and does not impact the core methodology, scientific rigorousness, or originality of the research, declaration is not required. + +Answer: [NA] + +Justification: Our method does not involve LLMs. + +Guidelines: + +- The answer NA means that the core method development in this research does not involve LLMs as any important, original, or non-standard components. +- Please refer to our LLM policy (https://neurips.cc/Conferences/2025/LLM) for what should or should not be described. + +# A Detailed Related Work + +Graphon Estimation Graphon estimation aims to recover the underlying generative structure of observed networks. Classical approaches include methods based on histogram estimators by partitioning nodes according to degree or other structural properties [5, 18, 6], and fitting stochastic block models (SBMs) or their variants, which can be viewed as piecewise constant graphon estimators [1, 11]. Universal singular value thresholding (USVT) [7] offers a non-parametric approach for estimating graphons from a single adjacency matrix, particularly effective for low-rank structures. However, many of these methods face challenges in terms of computational cost for large graphs, achieving resolution-free approximation, or may rely on specific structural assumptions (e.g., piecewise constant for SBMs). + +More recently, scalable graphon estimation techniques have gained prominence. For example, some works aim at minimizing distances between graph representations but often involve computationally expensive metrics like the GW distance [24, 37, 35], which can be a bottleneck for large networks. The advent of INRs has opened new avenues for continuous, resolution-free graphon estimation. For instance, IGNR (Implicit Graphon Neural Representation) [34] proposed to directly model graphons using neural networks, enabling the representation of graphons up to arbitrary resolutions and efficient generation of arbitrary-sized graphs. IGNR also addresses unaligned input graphs of different sizes by incorporating the Gromov-Wasserstein distance in its learning framework, often within an autoencoder setup for graphon learning. Subsequently, SIGL (Scalable Implicit Graphon Learning) [3] further advanced INR-based graphon estimation by combining INRs with Graph Neural Networks (GNNs). In SIGL, GNNs are utilized to improve graph alignment by determining appropriate node orderings, aiming to enhance scalability and learn a continuous graphon at arbitrary resolutions, with theoretical results supporting the consistency of its estimator. While these INR-based techniques offer significant advantages in terms of resolution-free representation and handling unaligned data, they still implicitly involve latent variable modeling or rely on GW-like objectives for alignment. Our proposed method builds upon the representational power of INRs but distinguishes itself by directly recovering the graphon via moment matching. This avoids the need for latent variables, complex metric computations like GW, and provides a theoretically grounded estimation framework that naturally handles multiple observed graphs by matching aggregated empirical moments. + +Data Augmentation for Graph Classification Data augmentation is crucial for improving the generalization of GNNs and other graph learning models, especially when labeled data is scarce. Mixup [40], which creates synthetic examples by linearly interpolating pairs of samples and their labels, has shown remarkable success in various domains. Its adaptation to graph data has been explored through several avenues, addressing challenges such as varying node counts, lack of alignment, and the non-Euclidean nature of graphs. For instance, Wang et al. [33] proposed interpolating hidden states of GNNs. Particularly relevant to our work are G-Mixup [13] and GraphMAD Navarro and Segarra [21], which recognize the difficulties of direct graph interpolation and propose to augment graphs for graph classification by operating in the space of graphons. GraphMAD Navarro and Segarra [21] projects graphs into the latent space of graphons and implements nonlinear mixup strategies like convex clustering. G-Mixup [13] first estimates a graphon for each class of graphs from the training data. Then, instead of directly manipulating discrete graph structures, G-Mixup interpolates these estimated graphons of different classes in their continuous, Euclidean representation to obtain mixed graphons. Synthetic graphs for augmentation are subsequently generated by sampling from these mixed graphons. This technique has also been adopted as an augmentation strategy in the evaluation pipelines of some graphon estimation studies for downstream tasks [3]. + +# B Proof of Theorem 1 + +# B.1 Supporting Lemmas + +We rely on the following established and derived results. Lemma 1 is an original contribution of this work, while Lemma 2 is Theorem 3.7 (b) in Borgs et al. [5] and it is included here for completeness. + +Lemma 1 (Concentration of Empirical Motifs). Let $F$ be a simple graph with $k = |\mathcal{V}_F|$ vertices. For $P \geq 1$ graphs $G_1, \ldots, G_P$ , each sampled independently from $G_n(W^*)$ , and for any error tolerance $\epsilon_s > 0$ , the probability that the empirical motif density $\bar{t}(F, W^*) = \frac{1}{P} \sum_{p=1}^{P} t(F, G_p)$ deviates from + +the true motif density $t(F,W^{*})$ is bounded as + +$$ +\mathbb {P} [ | \bar {t} (F, W ^ {*}) - t (F, W ^ {*}) | \geq \epsilon_ {s} ] \leq 2 \exp \left(- \frac {P n}{4 k ^ {2}} \left(\epsilon_ {s} - \frac {k (k - 1)}{2 n}\right) ^ {2}\right), \tag {9} +$$ + +$$ +f o r \epsilon_ {s} > \frac {k (k - 1)}{2 n}. +$$ + +Proof. Let $X_{p} = t(F, G_{p})$ for $p = 1, \ldots, P$ . The graphs $G_{p}$ are independent samples from $G_{n}(W^{*})$ , so the random variables $X_{p}$ are independent and identically distributed. + +We leverage concentration properties of $t(F, G_{n}(W^{*}))$ in Borgs et al. [5, Lemma 4.4], stating that $t(F, G_{n}(W^{*}))$ is concentrated around $t(F, W^{*})$ with probability + +$$ +\mathbb {P} \left[ | t (F, G _ {n} \left(W ^ {*}\right)) - t (F, W ^ {*}) | > \delta \right] \leq 2 \exp \left(- n \delta^ {2} / \left(4 k ^ {2}\right)\right). \tag {10} +$$ + +This implies that the variable $Z = t(F, G_{n}(W^{*})) - t(F, W^{*})$ behaves like a sub-Gaussian random variable [32] $^{2}$ . Comparing the exponent $-\frac{n\delta^2}{4k^2}$ from (10) with the sub-Gaussian tail exponent $-\frac{\delta^2}{2\sigma^2}$ , we see that $t(F, G_{n}(W^{*})) - t(F, W^{*})$ is sub-Gaussian with parameter $\sigma_Z^2 = \frac{2k^2}{n}$ . + +The variables we are averaging are $X_{p} = t(F,G_{p})$ with $G_{p}\sim G_{n}(W^{*})$ . Let $\mu_{n} = \mathbb{E}[X_{p}] = \mathbb{E}[t(F,G_{n}(W^{*}))]$ . The centered variables fulfill $X_{p} - \mu_{n} = (t(F,G_{p}) - t(F,W^{*})) - (\mathbb{E}[t(F,G_{p})] - t(F,W^{*}))$ . Subtracting a constant (the bias $\mathbb{E}[t(F,G_p)] - t(F,W^*)$ ) from a sub-Gaussian variable preserves its sub-Gaussian property with the same parameter. Thus, $X_{p} - \mu_{n}$ are independent, zero-mean, and $\sigma^2$ -sub-Gaussian with $\sigma^2 = \sigma_Z^2 = \frac{2k^2}{n}$ . + +The average of $P$ independent $\sigma^2$ -sub-Gaussian random variables is $(\sigma^2 / P)$ -sub-Gaussian [32]. Let $\bar{Y} = \frac{1}{P} \sum_{p=1}^{P} (X_p - \mu_n) = \bar{t}(F, W^*) - \mu_n$ . Then $\bar{Y}$ is $\left( \frac{2k^2}{nP} \right)$ -sub-Gaussian. The tail bound for $\bar{Y}$ is + +$$ +\mathbb {P} [ | \bar {Y} | \geq \delta ] \leq 2 \exp \left(- \frac {\delta^ {2}}{2 \cdot \frac {2 k ^ {2}}{n P}}\right) = 2 \exp \left(- \frac {\delta^ {2} n P}{4 k ^ {2}}\right). \tag {11} +$$ + +Substituting $\bar{Y} = \bar{t}(F, W^{*}) - \mu_{n}$ , we get the concentration bound for the empirical mean around the expected mean: + +$$ +\mathbb {P} \left[ | \bar {t} (F, W ^ {*}) - \mu_ {n} | \geq \delta \right] \leq 2 \exp \left(- \frac {\delta^ {2} n P}{4 k ^ {2}}\right). \tag {12} +$$ + +We are interested in the deviation of $\bar{t}(F, W^{*})$ from the true motif density $t(F, W^{*})$ . We use the triangle inequality to relate this deviation to the deviation from the mean $\mu_{n}$ + +$$ +\left| \bar {t} (F, W ^ {*}) - t (F, W ^ {*}) \right| \leq \left| \bar {t} (F, W ^ {*}) - \mu_ {n} \right| + \left| \mu_ {n} - t (F, W ^ {*}) \right|. \tag {13} +$$ + +Let $B_{n} = |\mu_{n} - t(F,W^{*})|$ be the bias of the empirical estimate. It is known from the theory of graph limits (e.g., related to Borgs et al. [5, Lemma 4.3]) that this bias is bounded by $B_{n} \leq \frac{k(k - 1)}{2n}$ . If the deviation from the true density is at least $\epsilon_{s}$ , i.e., $|\bar{t}(F,W^{*}) - t(F,W)| \geq \epsilon_{s}$ , then it must be that $|\bar{t}(F,W^{*}) - \mu_{n}| \geq \epsilon_{s} - B_{n}$ . This implication requires $\epsilon_{s} > B_{n}$ for the bound to be meaningful. Thus, for $\epsilon_{s} > B_{n}$ + +$$ +\mathbb {P} [ | \bar {t} (F, W ^ {*}) - t (F, W ^ {*}) | \geq \epsilon_ {s} ] \leq \mathbb {P} [ | \bar {t} (F, W ^ {*}) - \mu_ {n} | \geq \epsilon_ {s} - B _ {n} ]. \tag {14} +$$ + +Using the inequality (12) with $\delta = \epsilon_{s} - B_{n}$ + +$$ +\mathbb {P} [ | \bar {t} (F, W ^ {*}) - t (F, W ^ {*}) | \geq \epsilon_ {s} ] \leq 2 \exp \left(- \frac {\left(\epsilon_ {s} - B _ {n}\right) ^ {2} n P}{4 k ^ {2}}\right). \tag {15} +$$ + +Introducing the upper bound for the bias, $B_{n}\leq \frac{k(k - 1)}{2n}$ + +$$ +\begin{array}{l} \mathbb {P} [ | \bar {t} (F, W ^ {*}) - t (F, W ^ {*}) | \geq \epsilon_ {s} ] \leq 2 \exp \left(- \frac {\left(\epsilon_ {s} - \frac {k (k - 1)}{2 n}\right) ^ {2} n P}{4 k ^ {2}}\right) (16) \\ = 2 \exp \left(- \frac {P n}{4 k ^ {2}} \left(\epsilon_ {s} - \frac {k (k - 1)}{2 n}\right) ^ {2}\right). (17) \\ \end{array} +$$ + +This bound is valid when $\epsilon_s > \frac{k(k - 1)}{2n}$ , as required by the lemma statement. + +![](images/9e8b5b8a8956d1a618c6e43b72e5381273e715cc9bf54482d502da4a88cc4ae8.jpg) + +Lemma 2 (Motif Proximity Implies Cut Distance Proximity (Borgs et al. [5], Theorem 3.7 (b))). For any integer $k \geq 1$ , if the motif distance between two graphons $W_{1}$ and $W_{2}$ fulfills $|t(F, W_{1}) - t(F, W_{2})| < \delta_{M} = 3^{-k^{2}}$ for every simple graph $F \in \mathcal{F}_{k}$ , then the cut distance between $W_{1}$ and $W_{2}$ is upper bounded by + +$$ +d _ {c u t} \left(W _ {1}, W _ {2}\right) \leq \eta = \frac {2 2 C}{\sqrt {\log_ {2} k}}, \tag {18} +$$ + +where $C = \max \{1, \| W_1\|_\infty, \| W_2\|_\infty\}$ . + +# B.2 A comment on Assumption 1 + +This assumption is fundamentally supported by the Universal Approximation Theorem (UAT) [8, 14, 16]. The UAT posits that a neural network with sufficient capacity (e.g., an adequate number of neurons in one or more hidden layers and appropriate non-linear activation functions) can approximate any continuous function to an arbitrary degree of accuracy on a compact domain. In our context, the INR $f_{\theta}$ models the graphon $W:[0,1]^2\to [0,1]$ . The motif density $t(F,W)$ (as defined in Equation 1) is a continuous functional of $W$ , meaning small changes in $W$ lead to small changes in $t(F,W)$ . Consequently, if the INR $f_{\theta}$ can approximate any continuous graphon function, it can learn a specific $f_{\theta}$ such that the motif densities of the graphon estimated by the INR $t(F,f_{\theta})$ are arbitrarily close to some target values. Given that our estimated motif densities $\hat{t}_{\theta}(F,\hat{W}_{\theta})$ are Monte Carlo approximations of $t(F,f_{\theta})$ , they too can approach these target values (the empirical densities $\bar{t} (F,W^{*})$ ) as the approximation of the underlying function by $f_{\theta}$ improves as the number of Monte Carlo samples $L$ increases. The assumption thus relies on the INR's capacity to learn a suitable graphon function $f_{\theta}$ and the optimization process's ability to find the parameters $\theta$ that make the resulting motif estimates $\hat{t}_{\theta}(F,\hat{W}_{\theta})$ match the empirical observations $\bar{t} (F,W^{*})$ . + +# B.3 Proof of Theorem 1 + +Proof of Theorem 1. Our goal is to bound the cut distance $d_{\mathrm{cut}}(\hat{W}_{\theta}, W^{*})$ by $\eta$ , which is achieved if we can show that $|\hat{t}(F, \hat{W}_{\theta}) - t(F, W^{*})| < \delta_{M}$ for all simple graphs $F$ with $|\mathcal{V}_F| = k$ and where the values of both $\eta$ and $\delta_{M}$ are provided in Lemma 2. + +Consider any graph $F \in \mathcal{F}_k$ . Using the triangle inequality, we can bound the difference between the neural network's motif estimate and the true graphon motif + +$$ +\left| \hat {t} _ {\theta} \left(F, \hat {W} _ {\theta}\right) - t (F, W ^ {*}) \right| \leq \left| \hat {t} _ {\theta} \left(F, \hat {W} _ {\theta}\right) - \bar {t} (F, W ^ {*}) \right| + \left| \bar {t} (F, W ^ {*}) - t (F, W ^ {*}) \right|. \tag {19} +$$ + +By Assumption 1 on the neural network's training performance, we guarantee + +$$ +\left| \hat {t} _ {\theta} (F, \hat {W} _ {\theta}) - \bar {t} (F, W ^ {*}) \right| < \epsilon_ {a} = \frac {\delta_ {M}}{2}, \tag {20} +$$ + +for every $F\in \mathcal{F}_k$ + +Now we need to bound the second term in the right-hand side of (19), the deviation of the empirical motif density from the true motif density $|\bar{t}(F, W^{*}) - t(F, W^{*})|$ . We use Lemma 1 with the sampling error tolerance set to $\epsilon_{s} = \frac{\delta_{M}}{2}$ . For this lemma to apply, we require $\epsilon_{s} > \frac{k(k-1)}{2n}$ , which is equivalent to $\frac{\delta_{M}}{2} > \frac{k(k-1)}{2n}$ , or $n > \frac{k(k-1)}{\delta_{M}}$ . This condition is enforced in the theorem statement. + +For a specific graph $F \in \mathcal{F}_k$ , the probability that the sampling error is large is bounded by Lemma 1 + +$$ +\mathbb {P} \left[ | \bar {t} (F, W ^ {*}) - t (F, W ^ {*}) | \geq \frac {\delta_ {M}}{2} \right] \leq 2 \exp \left(- \frac {P n}{4 k ^ {2}} \left(\frac {\delta_ {M}}{2} - \frac {k (k - 1)}{2 n}\right) ^ {2}\right). \tag {21} +$$ + +Let $\mathbb{P}_{\mathrm{fail},F}$ denote this upper bound for a single graph $F\in \mathcal{F}_k$ . However, we require the sampling error $|\bar{t} (F,W^{*}) - t(F,W^{*})|$ to be less than $\frac{\delta_M}{2}$ for all graphs $F\in \mathcal{F}_k$ simultaneously. By the union bound, the probability that there exists at least one graph $F\in \mathcal{F}_k$ for which the sampling error is $\frac{\delta_M}{2}$ or more is at most the sum of the probabilities for each individual graph + +$$ +\mathbb {P} [ \exists F \in \mathcal {F} _ {k} \text {s . t .} | \bar {t} (F, W ^ {*}) - t (F, W ^ {*}) | \geq \frac {\delta_ {M}}{2} ] \leq \sum_ {F \in \mathcal {F} _ {k}} \mathbb {P} _ {\text {f a i l}, F}. \tag {22} +$$ + +Since $|\mathcal{V}_F| = k$ for all $F \in \mathcal{F}_k$ , the bound $\mathbb{P}_{\mathrm{fail}, F}$ is identical for all these graphs. The sum is thus $N_k \cdot \mathbb{P}_{\mathrm{fail}, F}$ , where we recall that $N_k = |\mathcal{F}_k|$ . The condition (7) in the theorem is precisely set to ensure that this total probability of failure is less than the desired confidence level $\zeta$ + +$$ +N _ {k} \cdot 2 \exp \left(- \frac {P n}{4 k ^ {2}} \left(\frac {\delta_ {M}}{2} - \frac {k (k - 1)}{2 n}\right) ^ {2}\right) < \zeta . \tag {23} +$$ + +Therefore, with probability at least $1 - \zeta$ (over the random graph samples $G_{p}$ ), the event that $|\bar{t}(F,W^{*}) - t(F,W^{*})| < \frac{\delta_{M}}{2}$ holds for all $F \in \mathcal{F}_k$ occurs. + +Conditioned on this high-probability event, and using the neural network approximation in Assumption 1, we have for every $F \in \mathcal{F}_k$ + +$$ +\left| \hat {t} _ {\theta} (F, \hat {W} _ {\theta}) - t (F, W ^ {*}) \right| \leq \left| \hat {t} _ {\theta} (F, \hat {W} _ {\theta}) - \bar {t} (F, W ^ {*}) \right| + \left| \bar {t} (F, W ^ {*}) - t (F, W ^ {*}) \right| < \frac {\delta_ {M}}{2} + \frac {\delta_ {M}}{2} = \delta_ {M}. \tag {24} +$$ + +Since $|\hat{t}_{\theta}(F, \hat{W}_{\theta}) - t(F, W^{*})| < \delta_{M}$ holds for all $F \in \mathcal{F}_k$ , Lemma 2 implies that the cut distance between the estimated graphon $\hat{W}_{\theta}$ and the true graphon $W^{*}$ is less than $\eta$ + +$$ +d _ {\text {c u t}} \left(W _ {\theta}, W\right) < \eta , \tag {25} +$$ + +with probability at least $1 - \zeta$ , concluding the proof. + +![](images/69e2a57147a25bf45601efc1c64c48faffb368b404395de130597d7f2218f991.jpg) + +# C Equation (7) Bound's Applicability in Realistic Regimes + +Assuming a small motif error $\delta_M$ is achievable, we now show that the overall probabilistic bound from (7) is non-vacuous for realistic dataset sizes. Table 3 shows the minimum value of $\zeta$ (left-hand side of equation (7)) for $k = 4$ , a conservatively large motif error of $\delta_M = 0.07$ (which is orders of magnitude larger than our empirical results) and various numbers of graphs $P$ and nodes $n$ . + +Table 3: Minimum value of $\zeta$ in equation (7) for varying values of $P$ and $n$ . + +
P +n400600800100012001400160018002000
20011.6311.4511.2711.1010.9310.7610.5910.4310.26
2509.939.048.227.486.816.195.635.134.66
3007.876.375.164.183.392.742.221.801.46
3505.974.222.972.101.481.040.740.520.37
4004.422.681.620.990.600.360.220.130.08
4503.211.660.860.440.230.120.060.030.016
+ +Although values $>1$ are uninformative as probabilities, the table clearly shows the bound's exponential decay. Crucially, the guarantee becomes meaningful for realistic data regimes. For example, the REDDIT-B dataset has $P = 2000$ graphs with an average $n \approx 497$ . Our analysis shows that in a comparable setting ( $n = 450$ , $P = 2000$ ), the failure probability $\zeta$ is a practically useful 0.016. This confirms that the theoretical conditions for our guarantee to hold are met within realistic data regimes. + +# D Unbiasedness of Monte Carlo Estimator for an INR-Based Graphon Moment Estimator + +Let $F = (\mathcal{V}_F, \mathcal{E}_F)$ be a graph, where $\mathcal{V}_F$ is a set of $k = |\mathcal{V}_F|$ vertices and $\mathcal{E}_F$ is the set of edges. Let $f_\theta : [0,1]^2 \to [0,1]$ be an INR parameterized by $\theta$ , which models the probability of an edge existing between two nodes based on their latent variables $\eta_i, \eta_j \in [0,1]$ , and its estimated graphon is denoted by $\hat{W}_\theta$ . + +The likelihood of observing the graph structure $F$ given a specific set of latent variable assignments $\eta = \{\eta_v\}_{v\in \mathcal{V}_F}$ and the INR model $f_{\theta}$ is given by + +$$ +P _ {\theta} (\boldsymbol {\eta}; F, \hat {W} _ {\theta}) = \prod_ {(i, j) \in \mathcal {E} _ {F}} \hat {W} _ {\theta} \left(\eta_ {i}, \eta_ {j}\right) \prod_ {(i, j) \notin \mathcal {E} _ {F}} \left(1 - \hat {W} _ {\theta} \left(\eta_ {i}, \eta_ {j}\right)\right). \tag {26} +$$ + +The quantity $t_{\theta}^{\prime}(F, \hat{W}_{\theta})$ is defined as this likelihood integrated over all possible configurations of the latent variables in the $k$ -dimensional unit hypercube + +$$ +t _ {\theta} ^ {\prime} (F, \hat {W} _ {\theta}) = \int_ {[ 0, 1 ] ^ {k}} P _ {\theta} (\boldsymbol {\eta}; F, \hat {W} _ {\theta}) d \boldsymbol {\eta}, \tag {27} +$$ + +where $d\pmb {\eta} = \prod_{v\in \mathcal{V}_F}d\eta_v$ + +The $L$ -sample Monte Carlo estimator for $t_{\theta}^{\prime}(F,\hat{W}_{\theta})$ is given by + +$$ +\hat {t} _ {\theta} ^ {\prime} (F, \hat {W} _ {\theta}) = \frac {1}{L} \sum_ {l = 1} ^ {L} P (\boldsymbol {\eta} ^ {(l)}; F, \hat {W} _ {\theta}). \tag {28} +$$ + +For this estimation, each sample $\pmb{\eta}^{(l)} = [\eta_{v_1}^{(l)},\dots,\eta_{v_k}^{(l)}]$ is a vector where each component $\eta_v^{(l)}$ (for $v\in \mathcal{V}_F$ ) is drawn independently from the uniform distribution $\mathcal{U}[0,1]$ . + +# D.1 Unbiasedness of the Estimator + +Theorem 2. The Monte Carlo estimator $\hat{t}_{\theta}^{\prime}(F,\hat{W}_{\theta})$ is an unbiased estimator of $t_\theta^\prime (F,\hat{W}_\theta)$ + +Proof. To show that the Monte Carlo estimation $\hat{t}_{\theta}^{\prime}(F,\hat{W}_{\theta})$ is an unbiased estimator of $t_{\theta}^{\prime}(F,\hat{W}_{\theta})$ , we need to prove that $\mathbb{E}[\hat{t}_{\theta}^{\prime}(F,\hat{W}_{\theta})] = t_{\theta}^{\prime}(F,\hat{W}_{\theta})$ . + +The expectation of the estimator is: + +$$ +\begin{array}{l} \mathbb {E} \left[ \hat {t} _ {\theta} ^ {\prime} (F, \hat {W} _ {\theta}) \right] = \mathbb {E} \left[ \frac {1}{L} \sum_ {l = 1} ^ {L} P _ {\theta} (\boldsymbol {\eta} ^ {(l)}; F, \hat {W} _ {\theta}) \right] \\ = \frac {1}{L} \sum_ {l = 1} ^ {L} \mathbb {E} \left[ P _ {\theta} \left(\boldsymbol {\eta} ^ {(l)}; F, \hat {W} _ {\theta}\right) \right] \quad (\text {b y}) \tag {29} \\ \end{array} +$$ + +Since each sample $\pmb{\eta}^{(l)}$ is drawn independently from the same uniform distribution, therefore its pdf is $p(\pmb {\eta}) = 1$ on $[0,1]^k$ , the expectation $\mathbb{E}[P_{\theta}(\pmb{\eta}^{(l)};F,\hat{W}_{\theta})]$ is the same for all $l$ . Let this common expectation be $\mathbb{E}[P_{\theta}(\pmb {\eta};F,\hat{W}_{\theta})]$ , whose value is + +$$ +\begin{array}{l} \mathbb {E} \left[ P _ {\theta} (\boldsymbol {\eta}; F, \hat {W} _ {\theta}) \right] = \int_ {[ 0, 1 ] ^ {k}} P _ {\theta} (\boldsymbol {\eta}; F, \hat {W} _ {\theta}) p (\boldsymbol {\eta}) d \boldsymbol {\eta} \\ = \int_ {[ 0, 1 ] ^ {k}} P _ {\theta} (\boldsymbol {\eta}; F, \hat {W} _ {\theta})) \cdot 1 d \boldsymbol {\eta} \quad \left(\text {s i n c e} p (\boldsymbol {\eta}) = 1 \text {o n} [ 0, 1 ] ^ {k}\right) \\ = t _ {\theta} ^ {\prime} (F, \hat {W} _ {\theta}), \\ \end{array} +$$ + +according to (27). Substituting this back into (29) + +$$ +\begin{array}{l} \mathbb {E} \left[ \hat {t} _ {\theta} ^ {\prime} (F, \hat {W} _ {\theta}) \right] = \frac {1}{L} \sum_ {l = 1} ^ {L} t _ {\theta} ^ {\prime} (F, \hat {W} _ {\theta}) \\ = \frac {1}{L} \left(L \cdot t _ {\theta} ^ {\prime} \left(F, \hat {W} _ {\theta}\right)\right) \\ = t _ {\theta} ^ {\prime} (F, \hat {W} _ {\theta}). \\ \end{array} +$$ + +Thus, $\mathbb{E}[\hat{t}_{\theta}'(F,\hat{W}_{\theta})] = t_{\theta}'(F,\hat{W}_{\theta})$ , which shows that the Monte Carlo estimator $\hat{t}_{\theta}'(F,\hat{W}_{\theta})$ is an unbiased estimator of $t_{\theta}'(F,\hat{W}_{\theta})$ . This means that, on average, the estimator will yield the true value of the integral defined by $f_{\theta}$ and the graph structure $F$ . + +# E Time Complexity of MOMENTNET + +Stage 1: parallel motif-density extraction. For each graph $G_{p} = (\mathcal{V}_{p},\mathcal{E}_{p})$ let $n_p = |\mathcal{V}_p|$ , $e_p = |\mathcal{E}_p|$ and $d_{p} = \max_{v\in \mathcal{V}_{p}}\deg (v)$ be the number of nodes, number of edges, and maximum degree of the graph $G_{p}$ . ORCA [15] counts all 2-4-node graphlets in + +$$ +T _ {\text {O R C A}} \left(G _ {p}\right) = O \left(e _ {p} d _ {p} + n _ {p} d _ {p} ^ {3}\right). +$$ + +Because every graph can be processed independently, we dispatch the $P$ graphs to $M$ workers $(M \leq P)$ . Hence the wall-clock preprocessing time is + +$$ +T _ {\text {s t a g e 1}} = O \left(\left\lceil \frac {P}{M} \right\rceil \max _ {p} \left(e _ {p} d _ {p} + n _ {p} d _ {p} ^ {3}\right)\right). +$$ + +With one worker per graph $(M = P)$ this shrinks to the single-graph cost that dominates $(\max_p)$ . + +# Stage 2: training the Moment network. Define: + +- $L$ : number of Monte-Carlo samples per epoch; +- $N_{e}$ : number of training epochs; +- $C_{\mathrm{INR}}$ : cost of one forward/back-prop through the INR for a single edge probability; +- $|\theta|$ : total number of trainable parameters. + +Each motif instance $F$ of size $|\mathcal{V}_F| \leq 4$ invokes the INR at most six times, a constant. One epoch therefore costs + +$$ +T _ {\text {e p o c h}} = O \left(L C _ {\text {I N R}} + | \theta |\right), \quad T _ {\text {s t a g e 2}} = O \left(N _ {e} \left(L C _ {\text {I N R}} + | \theta |\right)\right). +$$ + +Overall wall-clock complexity. + +$$ +T _ {\mathrm {M o m e n t N e t}} = O \Big (\left\lceil \frac {P}{M} \right\rceil \max _ {p} (e _ {p} d _ {p} + n _ {p} d _ {p} ^ {3}) + N _ {e} \left(L C _ {\mathrm {I N R}} + | \theta |\right) \Big). +$$ + +# Comparison with SIGL in Sparse vs. Dense Regimes + +SIGL [3] requires message-passing GNN training, histogram building and INR fitting; with $N_{e}$ epochs its wall-clock cost is $T_{\mathrm{SIGL}} = O(PN_{e}n_{T}^{2})$ , where $n_T = \max_p n_p$ . + +- Sparse regime ( $d_{\mathrm{max}} = O(1) \Rightarrow e_p = O(n_p)$ ): + +- MOMENTNET: $T = O\left(\left\lceil \frac{P}{M} \right\rceil n_T + N_e\left(L C_{\mathrm{INR}} + |\theta|\right)\right)$ ; +- SIGL: $T = O(PN_{e}n_{T}^{2})$ + +Here MomentNet grows linearly in $n_T$ (plus the network-training term), whereas SIGL is quadratic. In practice we repeatedly observe MomentNet to be faster when graphs have $e_p = O(n_p)$ even for very large $n_p$ . + +- Dense regime (Erdős-Rényi with $p_{conn} = 0.5$ implies $d_{\max} \approx n_T / 2$ and $e_p = \Theta(n_T^2)$ ): + +- MOMENTNET: $T = O\left( {\left\lceil \frac{P}{M}\right\rceil {n}_{T}^{4} + {N}_{e}\left( {{LC}_{\mathrm{{INR}}} + \left| \theta \right| }\right) }\right)$ ; +- SIGL: $T = O(PN_{e}n_{T}^{2})$ + +Asymptotically, SIGL's $n_T^2$ term is smaller than MomentNet's $n_T^4$ . Yet empirical runs on dense ER graphs with $p_{conn} = 0.5$ still show MomentNet to be faster once (i) Stage 1 is fully parallelised and (ii) the constants behind GNN message passing and histogramming dominate SIGL's quadratic term. Thus, the theoretical advantage of SIGL in dense graphs does not necessarily translate into shorter wall-clock times. Furthermore, MomentNet utilizes a two-stage process. The initial stage involves computing motif counts from the input graphs. Following this, the graphs are discarded. The second stage, which our experiments show to be the dominant phase of our method, then trains an INR using a vector of average moments derived from these counts. This design provides a significant reason for our method's improved speed, particularly in dense scenarios. By isolating the computationally expensive motif counting to a preliminary step, this cost is bypassed during the subsequent, dominant INR learning phase. + +With graph-level parallelism, MOMENTNET is provably linear in the number of edges for sparse networks and remains competitive on dense networks because its constant factors are smaller and its training cost is independent of the graph size. + +# E.1 Justification for Practical Scalability Over SIGL + +The theoretical complexity analysis highlights that in the dense regime, MOMENTNET's Stage 1 cost is bounded by $O\left(\left[\frac{P}{M}\right]n_T^4\right)$ , which is asymptotically worse than SIGL's $O(PN_{e}n_{T}^{2})$ . However, as demonstrated in our empirical results (Figure 2c), MOMENTNET is practically faster on large, dense graphs. This is due to two critical factors: + +1. Loose Theoretical Bound and Small Empirical Constant for ORCA: The worst-case $O(n_T^4)$ bound for ORCA's 4-node motif counting is known to be non-tight. We conducted an empirical analysis by measuring the wall-clock execution time for counting 4-node graphlets on a sequence of dense graphs with node counts $n$ ranging from 30 to 400. + +Empirical Analysis of ORCA Runtime We modeled the relationship between runtime $T(n)$ and node count $n$ using a linear regression on log-transformed data, $T(n) \approx c \cdot n^{k}$ . Our results determined the practical growth rate to be nearly cubic, with an exponent of $k \approx 3$ ( $R^2 \approx 0.97$ ). Crucially, the fitted constant factor was extremely small, $c \approx 2.97 \times 10^{-8}$ . This empirical finding confirms that the algorithm is highly efficient on dense graphs, ensuring that MOMENTNET's theoretical worst-case cost is practically masked by SIGL's larger constant factors and overheads until $n_T$ becomes very large. + +2. Strategic Subsampling Capability: Our method's performance relies on the moment vector $\mathbf{m}$ , which is robustly estimated by averaging moments from multiple, smaller graphs ( $P$ graphs of size $n_p$ ). MOMENTNET's theoretical guarantee (Theorem 1 and Figure 2b) allows us to strategically subsample a single large input graph into a collection of smaller graphs, effectively reducing the dominant $\max_p n_p$ in Stage 1. This capability lets us trade off $n_p$ for $P$ to minimize runtime while preserving near-optimal performance, a flexibility SIGL lacks. SIGL must process the full-size graph to properly learn latent representations, directly locking its runtime to the high cost of a single, large $n_T$ . + +Therefore, MOMENTNET's practical speed advantage stems from a combination of a lower-than-worst-case empirical complexity in Stage 1 and a flexible sampling strategy that SIGL cannot utilize. + +# E.2 Subsampling Effect on Scalability + +The scalability evaluation in Figure 2 demonstrates the conditions under which MomentNet achieves a computational advantage over SIGL. A key insight from our analysis is that SIGL's performance deteriorates significantly on smaller graphs due to its dependence on stable latent variable estimation. This limitation benefits our approach: by subsampling a very large graph into a smaller, more manageable collection of subgraphs, we can maintain near-optimal estimation quality while drastically reducing the computational cost. In contrast, running SIGL directly on these smaller subgraphs is ineffective, as its accuracy drops off steeply. + +To further validate the benefit of subsampling in the large-graph regime, we constructed a dataset consisting of ten large graphs, each containing 2,000 nodes, sampled from the dense-graph graphon class used in our scalability analysis (Section 5.1.2). We then extracted ten 50-node subgraphs from each 2,000-node graph and used these smaller subgraphs as inputs for both methods. The results, summarized in Table 4, confirm that MomentNet outperforms SIGL in both accuracy (lower GW Loss) and runtime under this efficient subsampling strategy. Crucially, while SIGL's runtime increases sharply with full graph size, our method maintains comparable performance even when utilizing these smaller subgraphs. + +Table 4: Performance comparison on large graphs (2,000 nodes) using a subsampling strategy (10 extracted 50-node subgraphs per large graph). + +
MethodTraining Runtime (s)GW LossMotif Counting Runtime (s)
MomentNet54.83 ± 18.690.0548 ± 0.00161.59 ± 0.24
SIGL207.89 ± 18.40.1085 ± 0.0156-
+ +# F Proof of Proposition 1 + +Let $W_{1}, W_{2} \colon [0,1]^{2} \to [0,1]$ be two graphons and fix $\alpha \in (0,1)$ . Denote their convex combination by + +$$ +W _ {\alpha} = \alpha W _ {1} + (1 - \alpha) W _ {2}. +$$ + +Edge density (a linear functional). For the single-edge motif $F_{e}$ on vertices $\mathcal{V}_{F_e} = \{1,2\}$ and $\mathcal{E}_{F_e} = \{(1,2)\}$ , the induced density is + +$$ +t ^ {\prime} (F _ {e}, W) = \int_ {[ 0, 1 ] ^ {2}} W (\eta_ {1}, \eta_ {2}) d \eta_ {1} d \eta_ {2} = \mathbb {E} [ W (\eta_ {1}, \eta_ {2}) ]. +$$ + +Because the integrand is linear in $W$ , we immediately have + +$$ +t ^ {\prime} \left(F _ {e}, W _ {\alpha}\right) = \alpha t ^ {\prime} \left(F _ {e}, W _ {1}\right) + (1 - \alpha) t ^ {\prime} \left(F _ {e}, W _ {2}\right), +$$ + +so the edge density behaves affinely under convex combinations. + +The $V$ -shape motif. Let $F$ be the $V$ -shape (three-vertex path) on vertex set $\mathcal{V}_F = \{1,2,3\}$ and edge set $\mathcal{E}_F = \{(1,2),(1,3)\}$ . Its induced density is + +$$ +t ^ {\prime} (F, W) = \int_ {[ 0, 1 ] ^ {3}} W \left(\eta_ {1}, \eta_ {2}\right) W \left(\eta_ {1}, \eta_ {3}\right) \left[ 1 - W \left(\eta_ {2}, \eta_ {3}\right) \right] d \eta_ {1} d \eta_ {2} d \eta_ {3}. \tag {30} +$$ + +Plugging $W_{\alpha}$ into (30) + +$$ +\begin{array}{l} t ^ {\prime} (F, W _ {\alpha}) = \mathbb {E} \left[ \left(\alpha W _ {1} + (1 - \alpha) W _ {2}\right) _ {1 2} \left(\alpha W _ {1} + (1 - \alpha) W _ {2}\right) _ {1 3} \left(1 - \alpha W _ {1} - (1 - \alpha) W _ {2}\right) _ {2 3} \right] \\ = \alpha^ {3} \mathbb {E} \big [ (W _ {1}) _ {1 2} (W _ {1}) _ {1 3} (1 - (W _ {1}) _ {2 3}) \big ] + (1 - \alpha) ^ {3} \mathbb {E} \big [ (W _ {2}) _ {1 2} (W _ {2}) _ {1 3} (1 - (W _ {2}) _ {2 3}) \big ] \\ + \text {m i x e d} \tag {31} \\ \end{array} +$$ + +where, to simplify notation, we used $(\cdot)_{ij}$ to denote that the graphon inside the parenthesis is evaluated on $(\eta_i,\eta_j)$ and "mixed terms" contain products in which at least one factor comes from $W_{1}$ and another from $W_{2}$ . Because these mixed terms generally do not cancel, the right-hand side of (31) does not reduce to the affine combination + +$$ +\alpha t ^ {\prime} (F, W _ {1}) + (1 - \alpha) t ^ {\prime} (F, W _ {2}), \tag {32} +$$ + +except in degenerate cases (e.g. $W_{1} = W_{2}$ or $\alpha \in \{0,1\}$ ). + +Concrete counter-example. Take constant graphons $W_{1}(\eta_{i}, \eta_{j}) = p_{1}$ and $W_{2}(\eta_{i}, \eta_{j}) = p_{2}$ with $p_{1}$ and $p_{2}$ being constants satisfying $0 < p_{1} \neq p_{2} < 1$ . Then $W_{\alpha}(\eta_{i}, \eta_{j}) = p_{\alpha} = \alpha p_{1} + (1 - \alpha)p_{2}$ , and + +$$ +t ^ {\prime} (F, W _ {i}) = p _ {i} ^ {2} (1 - p _ {i}), \qquad t ^ {\prime} (F, W _ {\alpha}) = p _ {\alpha} ^ {2} (1 - p _ {\alpha}), +$$ + +for $i\in \{1,2\}$ .However, + +$$ +p _ {\alpha} ^ {2} \left(1 - p _ {\alpha}\right) \neq \alpha p _ {1} ^ {2} \left(1 - p _ {1}\right) + (1 - \alpha) p _ {2} ^ {2} \left(1 - p _ {2}\right) +$$ + +whenever $p_1 \neq p_2$ and $\alpha \in (0, 1)$ , confirming that the $V$ -shape moment is not affine in $W$ . + +Conclusion. Edge moments are linear in the graphon, but higher-order induced moments involve non-linear (polynomial) combinations of $W$ . Consequently, a convex combination of graphons preserves edge moments but fails to preserve the remaining components of the motif-moment vector. + +# G Methods Details + +# G.1 Latent Variable Invariance of MomentNet + +The graphon model and our proposed model to learn it exhibit invariance to the specific ordering or labeling of latent variables. This means that the estimated graphon is unchanged under measure + +preserving transformations [5]. In other words, if the underlying structure of a graphon is rearranged or relabeled, MomentNet can still accurately capture the essential underlying connectivity patterns. To illustrate this crucial property, we conduct an experiment using an SBM graphon, more precisely the one indexed by 12 in Table 7. For this experiment, we utilize the same dataset that was generated for the performance comparison of MomentNet discussed in Section 5. The learned graphons for three different realizations of this experiment are presented in Figure 3. It is evident that all three estimated graphons closely resemble the ground truth graphon, which is depicted in Figure 4. Also, the three estimated graphons reflect the same underlying structure, and all of them share a similar GW loss, which is a loss function invariant to measure preserving transformations. This essentially means that, no matter which of the three depicted graphons we sample graphs from, the underlying structure of all these graphs will be the same. This outcome strongly verifies that MomentNet's primary mechanism involves matching the moments of the graph, without caring about the ordering of the latent variables. Consequently, and in contrast to other methodologies, its estimated graphon accurately reflects the ground truth structure, allowing for differences only up to a permutation of the latent variable locations. + +![](images/260d54886e1d73655b90ef53a9de10d60a4ea1ba38babd13aad567e748788a98.jpg) +(a) Estimated SBM graphon (Sample 1). + +![](images/b79e32c42131cbaf297368ce3df25e9cf66adce561931a6050696aacef693310.jpg) +(b) Estimated SBM graphon (Sample 2). + +![](images/e34818b889e5d4b5de998e8fd5f8dd6a186b85bb3a34527a45dd012c0d3afdfd.jpg) +(c) Estimated SBM graphon (Sample 3). +Figure 3: Three samples of estimated graphons derived from a SBM. + +# G.2 MomentNet Generalization + +To test the generalization capabilities of MomentNet beyond standard metrics like the GW loss and centrality measures, we conducted an additional experiment focusing on moment extrapolation. Following the experimental setup described in the paper, we trained a model exclusively on the nine motifs corresponding to 2- to 4-node subgraphs ( $F_0$ to $F_8$ ). We then evaluated its ability to accurately predict the densities of a different, unobserved set of subgraphs: the 5-node motifs (motif indices 10 through 30 in the ORCA paper [15]). The relative error for each extrapolated moment is presented in Table 5. The highly accurate estimations, with a median relative error of less than $2\%$ , demonstrate that MomentNet successfully learned the true underlying continuous data distribution (the graphon) from the low-order moments, allowing it to accurately extrapolate high-order structural properties. + +Table 5: Extrapolation Error: Relative error (\%) of MomentNet's estimated moments for unobserved 5-node motifs (Indices 10-30), after training only on 2- to 4-node motifs (Indices 1-9). + +
MotifRel. Err. (%)MotifRel. Err. (%)MotifRel. Err. (%)MotifRel. Err. (%)MotifRel. Err. (%)MotifRel. Err. (%)
102.224140.205180.154221.102263.0873011.086
111.536151.031190.371232.468271.753
121.785161.899201.744240.140282.062
130.073170.462211.502251.043294.010
+ +# G.3 Monte Carlo Sampling Variance + +Based on the relatively high quality of the results achieved by MomentNet, we hypothesize that stable convergence during training is contingent upon a sufficiently low variance in the Monte Carlo estimation of the motif densities (as described in the moment estimator paragraph of Section 3.1, Equation 4). To investigate this hypothesis directly, we designed an experiment to measure the gradient variance as a function of the number of Monte Carlo samples, $L$ . For this test, the parameters $(\theta)$ of the INR model were held constant while we estimated the gradient's standard deviation over 1,000 runs for varying numbers of samples, $L$ . The results, summarized in Table 6, show a clear inverse relationship: the standard deviation consistently decays as $L$ increases. This finding supports our hypothesis and underscores the importance of using a sufficient number of samples, $L$ , to ensure stable and efficient optimization. + +Table 6: Gradient Stability vs. Monte Carlo Samples $(L)$ . The standard deviation of the gradient decreases consistently with the number of Monte Carlo samples, supporting the need for a sufficiently large $L$ for stable optimization. + +
Monte Carlo Samples (L)Mean GradientStd Dev Gradient
1000.0130050.004484
5000.0131070.002031
10000.0130300.001387
50000.0130190.000647
100000.0130520.000439
200000.0130090.000325
500000.0130130.000193
1000000.0130190.000134
+ +# G.4 MomentMixup Pseudocode + +Algorithm 1 MomentMixup Augmentation +Input: $\alpha_{\mathrm{mix}}$ : float, mixing coefficient $(0\leq \alpha_{\mathrm{mix}}\leq 1)$ $\mathcal{G}_i,\mathcal{G}_j$ : list of graphs, graph datasets for classes $i$ and $j$ $y_{i},y_{j}$ : integer, label for classes $i$ and $j$ $N_{\mathrm{sample}}$ : integer, number of graphs to sample from each class dataset to compute average moments. $N_{\mathrm{nodes}}$ : integer, number of nodes for each new graph. $N_{\mathrm{graphs}}$ : integer, number of augmented graphs to generate. +Output: $\mathcal{G}_{\mathrm{aug}}$ : list of graphs and labels, newly generated augmented graphs. +1: $\triangleright$ Compute average moment vector for class i +2: $S_{i}\gets$ Randomly select $N_{\mathrm{sample}}$ graphs from $\mathcal{G}_i$ +3: $\mathbf{m}_i\leftarrow \frac{1}{N_{\mathrm{sample}}}\sum_{G\in S_i}\mathrm{ComputeGraphMoments}(G)$ +4: $\triangleright$ Compute average moment vector for class j +5: $S_{j}\gets$ Randomly select $N_{\mathrm{sample}}$ graphs from $\mathcal{G}_j$ +6: $\mathbf{m}_j\gets \frac{1}{N_{\mathrm{sample}}}\sum_{G\in S_j}\mathrm{ComputeGraphMoments}(G)$ +7: $\mathbf{m}_{\mathrm{target}}\gets \alpha_{\mathrm{mix}}\cdot \mathbf{m}_i + (1 - \alpha_{\mathrm{mix}})\cdot \mathbf{m}_j$ $\triangleright$ Compute target mixed moments +8: $y_{\mathrm{target}}\gets \alpha_{\mathrm{mix}}\cdot y_i + (1 - \alpha_{\mathrm{mix}})\cdot y_j$ $\triangleright$ Compute the label for the new samples +9: $W_{\mathrm{aug}}\gets$ MomentNet(m_target) $\triangleright$ Trains MomentNet for m_target +10: $\mathcal{G}_{\mathrm{aug}}\gets []$ $\triangleright$ Initialize list for augmented samples +11: for $k\gets 1$ to $N_{\mathrm{graphs}}$ do +12: $G_{\mathrm{new}}\gets$ SampleGraph(Waug,Nnodes) $\triangleright$ Sample new graph +13: Add (Gnew, ytarget) to Gaug +14: end for +15: return $\mathcal{G}_{\mathrm{aug}}$ + +# H List of Graphons + +Table 7: Table of Graphons + +
W(x,y)
1xy
2e(- (x0.7 + y0.7))
31/4 (x2 + y2 + √x + √y)
41/2 (x + y)
5(1 + e(-2(x2 + y2))^-1)
6(1 + e(-max{x,y}^2 - min{x,y}^4))^-1
7e(- max{x,y}^{0.75})
8e(-1/2 (min{x,y} + √x + √y))
9log(1 + max{x,y})
10|x - y|
111 - |x - y|
120.8I2 ⊗ I[0, 1/2]2
130.8(1 - I2) ⊗ I[0, 1/2]2
+ +The graphons are also visualized in Figure 4. + +![](images/f2383bee1b1f49ff61576e485320e8a775ee758a5d056109c9b046b4117ade3d.jpg) + +![](images/957cb5ba7805a5ab5d4a1bbcc1948a445b19a598f001168051c5cfc13d2a5c47.jpg) + +![](images/39fa6944939c335e2b4449b05558f6b67c0f04154564bb06f6d788b0911e329a.jpg) + +![](images/b0c523e8922cc3afe07ca3dc64f9e67c07b1e15dd8710c3d9e7d5b1ac9f03523.jpg) + +![](images/3a05e097ab5e0785e5c56340e7b12e2250102ffbe6b91921563cbf62b92501fb.jpg) + +![](images/82efc4c3a5e4201321778914ee292ee4663b86199783ab3748174d2673bffd81.jpg) + +![](images/6c24533ea51560e416f3fae97f124aef70d4e39ee5693c181b82067ff8b7a97a.jpg) + +![](images/5c5659e80c74e923ad5ab2e556bd87608541113783fbcb8ef49c0c149c81369c.jpg) + +![](images/ea4b23dcaec153ef316e0b118d71799f9e418751d1b2f631207a6eda7be839ef.jpg) + +![](images/2d9105003fb610c43ce033e3e0c2583405ea6d9df5e4af9eca2922dec3215387.jpg) + +![](images/afa5fbd1d3760b89ec540767f8558efd9c62f953d8b3f66b74aa04fe913f3e01.jpg) + +![](images/32baaa4953f2d0914261d57b56434fe02455b980bc01d0a208330fb3211364cb.jpg) + +![](images/b374d7a2dd4ad79a478b93c014e1ba2175af062dc04cf3dbf65eaf088895692a.jpg) +Figure 4: Representation of the graphons defined in Table 7. + +# I Selected Motifs + +![](images/d2dfb43b2396e441bbb1820acc90855f1250219878d183e08a06ebba4886c327.jpg) +Figure 5: Motifs up to four nodes. + +# J Centrality Measures + +In real-world graph statistical analysis, centrality measures are of significant interest to researchers. Building upon the work of Avella-Medina et al. [2], who demonstrated the computability of these measures on graphons, we use several centrality metrics to further evaluate the quality of the estimated graphons. Specifically, we employ: + +- Degree Centrality: This measure quantifies the number of direct connections a node possesses. + +- High Value: Indicates a node with many direct connections, often acting as a local hub with numerous immediate interactions. Such a node is highly active in its local neighborhood. +- Low Value: Suggests a node with few direct connections, implying less immediate activity or influence within its local vicinity. + +- Eigenvector Centrality: This identifies influential nodes by considering that connections to other highly-connected (and thus influential) nodes contribute more significantly to a node's score. It measures how well-connected a node is to other well-connected nodes. + +- High Value: A node with high eigenvector centrality is connected to other nodes that are themselves influential. This node is likely a key player within an influential cluster or a leader among leaders. +- Low Value: A node with low eigenvector centrality is typically connected to less influential nodes or has relatively few connections overall. Its influence is not strongly amplified by the influence of its neighbors. + +- Katz Centrality: This measure considers all paths in the graph, assigning exponentially more weight to shorter paths while still accounting for longer ones. It uses an attenuation factor $\alpha$ , which determines the weight given to longer paths: smaller values of $\alpha$ emphasize shorter paths, while larger values give more importance to longer paths, up to a theoretical limit to ensure convergence. + +- High Value: Indicates a node that is reachable by many other nodes through numerous paths, with shorter paths contributing more. This node is generally well-connected throughout the network, both directly and indirectly, and can efficiently disseminate or receive information. +- Low Value: Suggests a node that is not easily reachable by many other nodes or is primarily connected via very long paths. Its overall influence or accessibility within the network is limited. + +- PageRank Centrality: Originally developed for web pages, PageRank assesses a node's importance based on the number and quality of its incoming links. A link from an important node carries more weight than a link from a less important one. It uses a damping factor $\beta$ , representing the probability that a random walker will follow a link to an adjacent node, while $(1 - \beta)$ is the probability they will jump to a random node in the graph, ensuring that all nodes receive some rank and preventing rank-sinking in disconnected components. + +- High Value: A node with high PageRank centrality receives many "votes" (incoming connections) from other important nodes. This indicates that significant entities within the network consider this node to be important or authoritative. +- Low Value: A node with low PageRank centrality receives few incoming connections or is primarily linked by less important nodes. It is not widely recognized as important by other influential nodes in the network. + +The mathematical formulations for these graphon-based centrality measures are adopted directly from Avella-Medina et al. [2], corresponding to equations (7), (8), (9), and (10) in their paper, respectively. For a detailed analysis, we focus on graphons 1 and 2, as specified in Table 7. We compute both analytical and sample-based centrality measures, establishing these as baselines for comparison with our results. The analytical computations directly apply the aforementioned formulas from Avella-Medina et al. [2]. For the sample-based approach, we generate discrete graph instances by drawing samples from the ground truth graphon and subsequently compute the centrality measures within this discrete domain. Further details regarding each graphon are presented in the subsequent subsections. + +# J.1 Graphon 1: The $(xy)$ Model + +The analytical centrality measures formulas for this graphon are as follows: + +- Degree Centrality: + +$$ +C ^ {d} (x) = \frac {x}{2} +$$ + +- Eigenvector Centrality: + +$$ +C ^ {e} (x) = \sqrt {3} x +$$ + +Katz Centrality: + +$$ +C _ {\alpha} ^ {k} (x) = (6 - 2 \alpha) + 3 \alpha x +$$ + +PageRank Centrality: + +$$ +C _ {\beta} ^ {\mathrm {p r}} (x) = (1 - \beta) + 2 \beta x +$$ + +These measures are for the given latent variable $x \in [0,1]$ , after computing its centrality vector, we normalize it before comparison with discrete graph centralities [2]. Since the ordering for these experiments is important, we create a new dataset of 20 graphs with 100 nodes each, preserving the latent variables for all the nodes. The experiment results are illustrated in Figure 6. Our results show that centrality measures from the MomentNet-predicted graphon (blue lines in the figure) are close to the analytical computations (ground truth, black dashed lines). Furthermore, these graphon-based centralities by MomentNet also provide a good approximation for centrality measures computed over discrete graph samples (red dots). + +![](images/abc72ddeecf9e859bb1d9b32e7ca58ef726e9fb2daa48b93e739e00583ad5e1a.jpg) + +![](images/3d673b9ea01b46764b452c606eacd9c061624132cc02c841b6355ebee1d70a29.jpg) + +![](images/daf586a24d60aa5989ef36451c58193b361f3b67ccf7cd0e80164e8f6dec10fb.jpg) +Figure 6: Centrality measures: MomentNet vs. analytic computation for the $xy$ graphon. + +![](images/695096452c460c1c4385633961a9286cf3d65338e8809f77ea26608b039234cc.jpg) + +# J.2 Graphon 2: The $(e^{(-(x^{0.7} + y^{0.7}))})$ Model + +To test the generalizability and consistent performance of our method across varying complexities, we replicated the experiment on a more complex graphon. The analytical centrality measures formulas for this graphon are as follows: + +- Degree Centrality: + +$$ +C ^ {d} (x) = 0. 7 4 9 2 e ^ {- x ^ {0. 7}} +$$ + +- Eigenvector Centrality: + +$$ +C ^ {e} (x) = \frac {e ^ {- x ^ {0 . 7}}}{\sqrt {0 . 4 7 3}} +$$ + +- Katz Centrality: + +$$ +C _ {\alpha} ^ {k} (x) = 1 + \frac {0 . 7 4 9 2 \alpha e ^ {- x ^ {0 . 7}}}{1 - 0 . 4 7 3 \alpha} +$$ + +# PageRank Centrality: + +$$ +C _ {\beta} ^ {\mathrm {p r}} (x, \beta) = (1 - \beta) + \frac {\beta}{0 . 7 4 9 2} e ^ {- x ^ {0. 7}} +$$ + +The experiment results are illustrated in Figure 7. Similar to the previous experiment, after computing the centrality measures on the graphon and analytically, we normalize them to compare them with the discrete graph measurement. As the plots show, similar to the previous graphon, our estimation is very close to the ground truth results obtained by analytical calculation. + +![](images/e4b837061e12b5e999c100a92cc0e235c1ccad31cc25db034c03612cbc4d1128.jpg) + +![](images/23957580805d3bcc888dde4b39728d147447d1dfaffa88182a6034821cec9c38.jpg) + +![](images/9fb4feb1ad620a4e36b75c825888817aa135c3fa1732bd77e04f73f0f9fc61bc.jpg) +Figure 7: Centrality measures: MomentNet vs. analytic computation for the $e^{(-(x^{0.7} + y^{0.7}))}$ graphon. + +![](images/c53789a28291754a1c21d7705c4ab54e6a940d526ffde9c940f4b46bf4134705.jpg) + +# K Extra Scalability Evaluations + +We conducted an additional experiment to evaluate the scalability of SIGL and MomentNet. For this assessment, rather than focusing on SIGL's known weaknesses in latent variable estimation, we selected graphon number 5 from Table 7, a model that both methods accurately estimate. We generate 10 graphs for each node size $n \in \{10, 20, \dots, 810\}$ . + +Figure 8 illustrates the scalability of MomentNet and SIGL in terms of both performance, measured by GW loss, and average runtime, as a function of the number of nodes. Subfigure (a) of Figure 8 reveals that MomentNet (blue line) maintains a consistently low GW loss across the tested range of node sizes, indicating stable performance. In contrast, SIGL's (red line) GW loss starts notably higher for smaller networks but decreases substantially as the number of nodes increases, eventually matching or even slightly outperforming MomentNet's loss for larger networks. + +However, subfigure (b) of Figure 8 highlights a significant difference in computational efficiency: MomentNet's average runtime exhibits only a modest and gradual increase with the number of nodes. Conversely, SIGL's runtime escalates sharply, demonstrating significantly poorer scalability. + +Consequently, while SIGL might offer a marginal advantage in GW Loss for very large graphs, MomentNet's vastly superior runtime scalability makes it a more practical and favorable approach, particularly for applications involving large-scale networks where computational resources and time are critical factors. + +![](images/5df2b692d4a9d1bba5ca3db163baa3b376108b6a3884f434861d87376fa1bff8.jpg) +(a) Comparison of performance scalability of MomentNet with SIGL. +Figure 8: Scalability Comparison of MomentNet and SIGL + +![](images/625be8d461b0bf52628efe6c3e343906dffc193aa5cc290f4ffa4fba3ae5b10f.jpg) +(b) Comparison of runtime scalability of MomentNet with SIGL. + +# L MomentMixup Evaluation Details + +Our experimental evaluation is conducted on four diverse benchmark datasets widely used in graph classification research. Table 8 provides a detailed overview of these datasets, outlining their specific characteristics and the nature of their respective classification tasks. + +Table 8: Description of the benchmark datasets used for evaluation. Each dataset represents a different type of graph structure and classification task. + +
DatasetDescriptionClassification TaskCitation
IMDB-BMovie collaboration graphs; nodes represent actors/actresses, and an edge connects two actors/actresses if they appear in the same movie.Binary genre classification.[39]
IMDB-MA multi-class version of IMDB-B, representing movie collaborations with similar graph construction.Multi-class genre classification.[39]
REDD-BSocial network graphs from Reddit; nodes represent users, and an edge indicates an interaction (e.g., one user commented on another's post).Binary community (subreddit) classification.[39]
AIDSBioinformatics graphs representing molecules; nodes are atoms, and edges are covalent bonds between them.Binary classification based on anti-HIV activity (active vs. inactive).[27]
+ +# M Social impacts + +The methods presented for graphon estimation via moment-matching INRs and data augmentation through MomentMixup, while offering powerful tools for understanding network structures and enhancing graph-based machine learning, are not without potential societal risks if deployed without careful consideration. For instance, in social network analysis, if the empirical moments used for graphon estimation are derived from graphs reflecting existing societal biases (e.g., in representation or connectivity), both the estimated graphons and synthetic graphs generated via MomentMixup could inadvertently perpetuate or even amplify these biases. This could lead to inequitable outcomes when models trained on such data are used for applications like resource allocation, recommendation systems, or public policy modeling. Similarly, in critical domains such as epidemiology or financial systems, inaccuracies in graphon estimation or the generation of unrepresentative augmented data could lead to flawed predictions, potentially resulting in misguided interventions or financial instability. While graphon estimation offers a level of abstraction, careful attention must also be paid to ensure that the process does not inadvertently leak sensitive information from the original graph data, especially when dealing with networks containing personal or confidential details. Therefore, + +it is crucial for practitioners to be acutely aware of these potential pitfalls. This includes critically examining input data and chosen moments for biases, rigorously validating the fidelity and representativeness of estimated graphons and generated graphs, and thoughtfully considering the ethical implications of their application, particularly in domains with direct and significant societal impact. \ No newline at end of file diff --git a/NeurIPS/2025/A Few Moments Please_ Scalable Graphon Learning via Moment Matching/images.zip b/NeurIPS/2025/A Few Moments Please_ Scalable Graphon Learning via Moment Matching/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..a085c83ee769d1697dade29796506d79cf38e2c6 --- /dev/null +++ b/NeurIPS/2025/A Few Moments Please_ Scalable Graphon Learning via Moment Matching/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:958d8917b51df8236db8caf8846b4c92efb8b92fcf5d75aebfcecec5f600ba1e +size 1133057 diff --git a/NeurIPS/2025/A Few Moments Please_ Scalable Graphon Learning via Moment Matching/layout.json b/NeurIPS/2025/A Few Moments Please_ Scalable Graphon Learning via Moment Matching/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..014f86f274282ac08c829de1d2b71e06ebb28e9e --- /dev/null +++ b/NeurIPS/2025/A Few Moments Please_ Scalable Graphon Learning via Moment Matching/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b19664d7f0ca1681270d07affa322e6bf900c0ea0485f3155f2a9d81c26e95f7 +size 1363298 diff --git a/NeurIPS/2025/A Frustratingly Simple Yet Highly Effective Attack Baseline_ Over 90% Success Rate Against the Strong Black-box Models of GPT-4.5_4o_o1/43fd1f53-a075-4d3d-ba5c-94f93fa3b47f_content_list.json b/NeurIPS/2025/A Frustratingly Simple Yet Highly Effective Attack Baseline_ Over 90% Success Rate Against the Strong Black-box Models of GPT-4.5_4o_o1/43fd1f53-a075-4d3d-ba5c-94f93fa3b47f_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..66659b18ce7ceba9d0c85a1668ad8b71e832a488 --- /dev/null +++ b/NeurIPS/2025/A Frustratingly Simple Yet Highly Effective Attack Baseline_ Over 90% Success Rate Against the Strong Black-box Models of GPT-4.5_4o_o1/43fd1f53-a075-4d3d-ba5c-94f93fa3b47f_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0ff3cbadce8fc0d13170c4a2ca42e71a5c7b21fa0e06ead1984a018ce4be22f2 +size 197489 diff --git a/NeurIPS/2025/A Frustratingly Simple Yet Highly Effective Attack Baseline_ Over 90% Success Rate Against the Strong Black-box Models of GPT-4.5_4o_o1/43fd1f53-a075-4d3d-ba5c-94f93fa3b47f_model.json b/NeurIPS/2025/A Frustratingly Simple Yet Highly Effective Attack Baseline_ Over 90% Success Rate Against the Strong Black-box Models of GPT-4.5_4o_o1/43fd1f53-a075-4d3d-ba5c-94f93fa3b47f_model.json new file mode 100644 index 0000000000000000000000000000000000000000..c6093ef360380f7f0365a0467eba802b95e86ec9 --- /dev/null +++ b/NeurIPS/2025/A Frustratingly Simple Yet Highly Effective Attack Baseline_ Over 90% Success Rate Against the Strong Black-box Models of GPT-4.5_4o_o1/43fd1f53-a075-4d3d-ba5c-94f93fa3b47f_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:359a5dd66d05ceae0840bcfefbe92b32d2f6379aa1058863dc7691818b0da6d3 +size 247764 diff --git a/NeurIPS/2025/A Frustratingly Simple Yet Highly Effective Attack Baseline_ Over 90% Success Rate Against the Strong Black-box Models of GPT-4.5_4o_o1/43fd1f53-a075-4d3d-ba5c-94f93fa3b47f_origin.pdf b/NeurIPS/2025/A Frustratingly Simple Yet Highly Effective Attack Baseline_ Over 90% Success Rate Against the Strong Black-box Models of GPT-4.5_4o_o1/43fd1f53-a075-4d3d-ba5c-94f93fa3b47f_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..00250c0bcbf3c7207ba487245cec8fb719879395 --- /dev/null +++ b/NeurIPS/2025/A Frustratingly Simple Yet Highly Effective Attack Baseline_ Over 90% Success Rate Against the Strong Black-box Models of GPT-4.5_4o_o1/43fd1f53-a075-4d3d-ba5c-94f93fa3b47f_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ac2e12593385373e6857973f96a19fc7259b3e9777d24b98879c945335ecc6a2 +size 17655128 diff --git a/NeurIPS/2025/A Frustratingly Simple Yet Highly Effective Attack Baseline_ Over 90% Success Rate Against the Strong Black-box Models of GPT-4.5_4o_o1/full.md b/NeurIPS/2025/A Frustratingly Simple Yet Highly Effective Attack Baseline_ Over 90% Success Rate Against the Strong Black-box Models of GPT-4.5_4o_o1/full.md new file mode 100644 index 0000000000000000000000000000000000000000..142155e8a6f595da0f5115321769723b904f7d4f --- /dev/null +++ b/NeurIPS/2025/A Frustratingly Simple Yet Highly Effective Attack Baseline_ Over 90% Success Rate Against the Strong Black-box Models of GPT-4.5_4o_o1/full.md @@ -0,0 +1,992 @@ +# A Frustratingly Simple Yet Highly Effective Attack Baseline: Over $90 \%$ Success Rate Against the Strong Black- box Models of GPT- 4.5/4o/o1 + +Zhaoyi Li*, Xiaohan Zhao*, Dong-Dong Wu, Jiacheng Cui, Zhiqiang Shen† + +VILA Lab, Department of Machine Learning, MBZUAI + +*Equal contribution †Corresponding Author + +https://vila-lab.github.io/M-Attack-Website/ + +{zhaoyi.li,xiaohan.zhao,dongdong.wu,jiacheng.cui,zhiqiang.shen}@mbzuai.ac.ae + +![](images/9a47def8296517e24c22eab2467887f979d541365848f46a0a124e49b2f216ea.jpg) +Query image + +![](images/bb485dc850183eddf3465f4edbdd52d14555bbfd54b3a35af8b33d759f710f0e.jpg) +Target image +Figure 1: Examples from closed-source LVLMs to targeted attacks generated by our method. + +# GPT-40 + +# Describe this image. + +![](images/3c1ba71f09a146ceca9c394b2b7ef0043df49937b816baa8d3665b2be9dad316.jpg) + +The image shows a market scene with large green apples in the foreground and a partially visible stall with fruits and people in the background. + +# Gemini-2.0-flash + +# Describe this image. + +![](images/b02df00bd176e0c30ca6fdc78d0d990f7a14cbb1c9cdd63a6ccdedae8b97946c.jpg) + +Green-tinged image of produce stand with melons in foreground and vendor partially visible in background. + +# Claude-3.5-sonnet + +# Describe this image. + +![](images/4e102f137947e706b32345840f2082fcc5a35450fa78ca45f615b998b071f6ef.jpg) + +A market stall displays fresh green cabbages and produce under a simple covered structure. + +# GPT-01 + +# Describe this image. + +![](images/d4527eda21b86bff34e1491bcdede59a51fce9953623e7b4a3ea4c13b4bd6832.jpg) + +A market stall brimming with large, bright green produce beneath a decorative cloth canopy. + +# Gemini-2.0-flash-thinking-exp + +# Describe this image. + +![](images/7d0686fb6962226c398731040098c3d1d5e46230f3911c5d698dc78a7855682a.jpg) + +Green produce, possibly gourds or melons, are displayed under a fabric awning in a market setting. + +# Claude-3.7-sonnet-extend + +# Describe this image. + +![](images/e32c6fd2a288732463f007e3c276f4e39f3550b0c330c8bd86f846163f21b27e.jpg) + +A pile of green apples in the foreground with what appears to be a small wooden market stall or structure in the background! + +# Abstract + +Despite promising performance on open-source large vision-language models (LVLMs), transfer-based targeted attacks often fail against closed-source commercial LVLMs. Analyzing failed adversarial perturbations reveals that the learned perturbations typically originate from a uniform distribution and lack clear semantic details, resulting in unintended responses. This critical absence of semantic information leads commercial black-box LVLMs to either ignore the perturbation entirely or misinterpret its embedded semantics, thereby causing the attack to fail. To overcome these issues, we propose to refine semantic clarity by encoding explicit semantic details within local regions, thus ensuring the capture of finer-grained features and inter-model transferability, and by concentrating modifications on semantically rich areas rather than applying them uniformly. To achieve this, we propose a simple yet highly effective baseline: at each optimization step, the adversarial image is cropped randomly by a controlled aspect ratio and scale, resized, and then aligned with the target image in the embedding space. While the naive source-target matching method has been utilized before in the literature, we are the first to provide a tight analysis, which establishes a close connection between perturbation optimization and semantics. Experimental results confirm our hypothesis. Our adversarial examples crafted with local-aggregated perturbations focused on crucial regions exhibit surprisingly good transferability to commercial LVLMs, including GPT-4.5, GPT-4o, Gemini-2.0-flash, Claude-3.5/3.7-sonnet, and even reasoning models like o1, Claude-3.7-thinking and Gemini-2.0-flash-thinking. Our approach achieves success rates exceeding $90\%$ on GPT-4.5, 4o, + +and o1, significantly outperforming all prior state-of-the-art attack methods with lower $\ell_1 / \ell_2$ perturbations. Our optimized adversarial examples under different configurations are available at HuggingFace and our training code at GitHub. + +# 1 Introduction + +Adversarial attacks have consistently threatened the robustness of AI systems, particularly within the domain of large vision-language models (LVLMs) [22, 4, 37]. These models have demonstrated impressive capabilities on visual and linguistic understanding integrated tasks such as image captioning [32], visual question answering [25, 29] and visual complex reasoning [21, 30]. In addition to the progress seen in open-source solutions, advanced black-box commercial multimodal models like GPT-4o [1], Claude-3.5 [3], and Gemini-2.0 [33] are now extensively utilized. Their widespread adoption, however, introduces critical security challenges, as malicious actors may exploit these platforms to disseminate misinformation or produce harmful outputs. Addressing these drawbacks necessitates thorough adversarial testing in black-box environments, where attackers operate with limited insight into the internal configurations and training data of the models. + +Current transfer-based approaches [39, 8, 12] typically generate adversarial perturbations that lack semantic structure, often stemming from uniform noise distributions with low success attacking rates on the robust black-box LVLMs. These perturbations fail to capture the nuanced semantic details that many LVLMs rely on for accurate interpretation. As a result, the adversarial modifications either go unnoticed by commercial LVLMs or, worse, are misinterpreted, leading to unintended and ineffective outcomes. This inherent limitation has motivated a deeper investigation into the nature and distribution of adversarial perturbations. + +Our analysis reveals that a critical drawback in conventional adversarial strategies is the absence of clear semantic information within the perturbations. Without meaningful semantic cues, the modifications fail to influence the model's decision-making process effectively. This observation is particularly relevant for closed-source commercial LVLMs, which have been optimized to extract and leverage semantic details from both local and global image representations. The uniform nature of traditional perturbations thus represents a significant barrier to achieving high attack success rates. + +Building on this insight, we hypothesize that a key to improving adversarial transferability lies in the targeted manipulation of core semantic objects present in the input image. Commercial black-box LVLMs, regardless of their large-scale and diverse training datasets, consistently prioritize the extraction of semantic features that define the image's content. By explicitly encoding these semantic details within local regions and focusing perturbations on areas rich in semantic content, it becomes possible to induce more effective misclassifications. This semantic-aware strategy provides a promising view for enhancing adversarial attacks against robust, black-box models. + +In this paper, we introduce a novel attack baseline called M-Attack that strategically refines the perturbation process. At each optimization step, the adversarial image is subjected to a random crop operation controlled by a specific aspect ratio and scale, followed by a resizing procedure. We then align the perturbations with the target image in the embedding space, effectively bridging the gap between local and local or local and global representations. The approach leverages the inherent semantic consistency across different white-box LVLMs, thereby enhancing the transferability of the crafted adversarial examples. + +Furthermore, recognizing the limitations of current evaluation practices, which often rely on subjective judgments or inconsistent metrics, we introduce a new Keyword Matching Rate (KMRScore) alongside GPTScore. This metric provides a more reliable, partially automated way to measure attack transferability and reduces human bias. Our extensive experiments demonstrate that adversarial examples generated with our method achieve transfer success rates exceeding $90\%$ against commercial LVLMs, including GPT-4.5, GPT-4o and advanced reasoning models like o1. + +Overall, our contributions are threefold: + +- We observe that failed adversarial samples often exhibit uniform-like perturbations with vague details, underscoring the need for clearer semantic guidance to achieve reliable transfer to attack strong black-box LVLMs. + +![](images/000342c4df1f4e6a23abe3981a55a2e819f169776388333f579ff95ac36c6277.jpg) +Figure 2: Illustration of our proposed framework. Our method is based on two components: Local-to-Global or Local-to-Local Matching (LM) and Model Ensemble (ENS). LM is the core of our approach, which helps to refine the local semantics of the perturbation. ENS helps to avoid overly relying on single models embedding similarity, thus improving attack transferability. + +- We show how random cropping with certain ratios and iterative local alignment with the target image embed local/global semantics into local regions, especially in crucial central areas, markedly boosting attack effectiveness. +- We propose a new Keyword Matching Rate (KMRScore) that offers a more objective measure for quantifying success in cross-model adversarial attacks, achieving state-of-the-art transfer results with reduced human bias. + +# 2 Related Work + +Large Vision-Language Models. Transformer-based LVLMs integrate visual and textual modalities by learning joint visual-semantic representations from large-scale image-text datasets. These models have underlaid core multimodal tasks such as image captioning [32, 13, 7, 34], visual question answering [25, 29], and cross-modal reasoning [36, 26, 35]. Open-source LVLMs like BLIP-2 [20], Flamingo [2], and LLaVA [23] demonstrate good capabilities on standard benchmarks, while closed-source systems such as GPT-4o [1], Claude-3.7 [3], and Gemini-2.5 [33] exhibit better instruction-following, reasoning, and adaptation to real-world multimodal tasks. Despite these advances, the closed-source nature of commercial LVLMs conceals internal mechanisms and vulnerabilities, making it difficult to evaluate their robustness under adversarial scenarios. This calls for a systematic exploration of their susceptibility to carefully crafted input perturbations. + +Transfer-Based Adversarial Attacks on LVLMs. Black-box attacks on LVLMs are either query-based [9, 15], relying on repeated API access to estimate gradients, or transfer-based [10, 24], which craft adversarial examples on surrogates without querying the target. While the latter is more efficient, transferability is hindered by the closed nature of commercial LVLMs, including undisclosed architectures and data, leading to significant semantic mismatches. Recent methods like AttackVLM [39] improve transfer success by aligning image-level features rather than cross-modal ones. This strategy influenced CWA [6] and SSA-CWA [8], which enhance transferability to models like Bard using sharpness-aware optimization and spectrum-based augmentation, achieving modest performance. Other approaches such as AnyAttack [38] and AdvDiffVLM [12] explore self-supervised pretraining and diffusion-based generation, but still struggle against leading commercial LVLMs. These limitations highlight the need for more stable, semantically grounded gradient strategies, which our method aims to address. + +# 3 Investigations Over Failed Attacks + +We investigate why prior state-of-the-art methods [39, 8, 38] have failed from two perspectives: 1) The perturbations from these methods tend to be uniformly distributed rather than highlighting statistically significant regions; 2) In many failed cases, the + +
GPT-4oClaude-3.5Gemini-2.0
AttackVLM [39]6%11%45%
AnyAttack [38]13%13%76%
SSA-CWA [8]21%29%75%
+ +Table 1: Percentage of vague response for failed attacks. + +model does detect the perturbation but is unable to articulate detailed semantic content, resulting in vague or ambiguous descriptions. Some failed examples are provided in Appendix H.2. + +![](images/0868bf1dc2a8a9ef2a7ee78aff6bc252947b07cad9503b3e96fa244be47ca889.jpg) +Figure 3: Empirical cumulative distribution vs. uniform distribution on 20 randomly-sampled failed adversarial images. Shading shows standard deviation. + +![](images/5b69f60862a0c59ada7a800c5583fb13f3fee695fc37ebe7daf7b8d2474bb71a.jpg) +Figure 4: Comparison of global similarity and ASR across different matching schemes: Global to Global, Local to Global and Local to Local. + +Uniform-like Perturbation Distribution. Fig. 3 and Fig. 5 (first row) illustrate that the perturbation in failed adversarial examples closely aligns with a uniform distribution, as indicated by the near-overlap between empirical cumulative distribution function (ECDF) and the ideal uniform CDF over 20 samples. The minimal deviation and tight standard deviation bands suggest that perturbations are spread evenly across the image space without preference for semantically meaningful regions. This uniform-like behavior implies a lack of targeted manipulation toward critical visual features, leading to weak semantic interference and ultimately ineffective attacks on LVLMs. In other words, the model perceives + +these perturbations as noise rather than meaningful semantic shifts. + +![](images/969f29d500e53187c14867dadf27d725dc0fd3b82216ae0f0b08d9c5683e4486.jpg) +Figure 5: Simulated heatmap visualization of perturbation aggregation across various steps using different crop schemes. The scales control the range of proportions to the original image area. + +Vague Description. To further validate that the model perceives these uniform perturbations as noise rather than meaningful semantic shifts, we quantify the proportion of vague descriptions. Specifically, we define vague descriptions as cases where the model uses terms like "blurry" or "abstract" to describe the detected artifacts or perturbations, instead of concrete semantic nouns. As shown in Tab. 1, while the black-box closed-source LVLMs do detect something unusual in the image, they struggle to interpret it consistently and clearly. + +Similarity Trajectories. We further visualize the evolution of similarity trajectories during training to understand why local matching is less prone to overfitting compared to previous global matching strategies, and why it more effectively attacks LVLMs. As shown in Fig. 4, we observe that global representations lack sufficient randomness, causing the similarity (i.e., negative loss) to increase rapidly and saturate early. This early saturation limits further learning. In contrast, local matching converges more slowly, allowing the model to capture finer-grained details throughout training. + +# 4 Approach + +**Framework Overview.** Our approach aims to enhance the semantic richness within the perturbation by extracting details matching certain semantics in the target image. By doing so, we improve the transferability of adversarial examples through a many-to-many/one matching, enabling them to remain effective against even the most robust black-box systems like GPT-4o, Gemini, and Claude. As shown in Fig. 2, at iteration $i$ , the generated adversarial sample performs random cropping followed by resizing to its original dimensions. The cosine similarity between the local source image embedding and the global or local target image embedding is then computed using an ensemble of surrogate white-box models to guide perturbation updates. The source-target pairs are randomly sampled. Through this iterative local-global or local-local matching, the central perturbed regions + +on the source image become progressively more refined, enhancing both semantic consistency and attack effectiveness, which we observe is surprisingly effective for commercial black-box LVLMs. + +Reformulation with Many-to-[Many/One] Mapping. Viewing details of adversarial samples as local features carrying target semantics, we reformulate the problem with many-to-many or many-to-one mapping for semantic detail extraction: let $\mathbf{X}_{\mathrm{sou}}, \mathbf{X}_{\mathrm{tar}} \in \mathbb{R}^{\mathbf{H} \times \mathbf{W} \times \mathbf{3}}$ denote the source and target images in the image space, $\mathbf{X}_{\mathrm{sou}}$ is the clean image at the initial time. In each step, we seek a local adversarial perturbation $\delta^{l}$ (with $\| \delta^{l} \|_{p} \leq \epsilon$ ) so that the perturbed source $\widetilde{\mathbf{x}}_{i}^{s} = \hat{\mathbf{x}}_{i}^{s} + \delta_{i}^{l}$ (where $\widetilde{\mathbf{x}}_{i}^{s}$ is the optimized local source region at step $i$ after current learned perturbation) matches the target $\hat{\mathbf{x}}^{t}$ at semantic embedding space in a many-to-many/one fashion. Our final learned global perturbation $\delta^{g}$ is an aggregation of all local $\{\delta_{i}^{l}\}$ . + +We define $\mathcal{T}$ as a set of transformations that generate local regions for source images, forming a finite set of source subsets, and local or global images for target. We apply preprocessing (e.g., resizing and normalization) to each original image, allowing the target image to be either a fixed global or a local region similar to the source image. + +$$ +\left\{\hat {\mathbf {x}} _ {1} ^ {s}, \dots , \hat {\mathbf {x}} _ {n} ^ {s} \right\} = \mathcal {T} _ {s} \left(\mathbf {X} _ {\mathrm {s o u}}\right) +$$ + +$$ +\{\hat {\mathbf {x}} _ {1} ^ {t}, \dots , \hat {\mathbf {x}} _ {n} ^ {t} \} / \{\hat {\mathbf {x}} _ {g} ^ {t} \} = \mathcal {T} _ {t} (\mathbf {X} _ {\mathrm {t a r}}), \tag {1} +$$ + +where each region $\hat{\mathbf{x}}_i$ ( $i \in \{1, 2, \dots, n\}$ ) is generated independently at a different training iteration $i$ . $\hat{\mathbf{x}}_g^t$ is a globally transformed target image if using many-to-one. To formulate many-to-many/one mapping, without loss of generality, we denote each pair $\hat{\mathbf{x}}_i^s$ and $\hat{\mathbf{x}}_i^t$ be matched in iteration $i$ . Let $f_\phi$ denote the surrogate embedding model, we have: + +$$ +\mathcal {M} _ {\mathcal {T} _ {s}, \mathcal {T} _ {t}} = \operatorname {C S} \left(f _ {\phi} \left(\hat {\mathbf {x}} _ {i} ^ {s}\right), f _ {\phi} \left(\hat {\mathbf {x}} _ {i} ^ {t}\right)\right), \tag {2} +$$ + +where CS denotes the cosine similarity. By maximizing $\mathcal{M}_{\mathcal{T}_s,\mathcal{T}_t}$ , each $\tilde{\mathbf{x}}_i^s$ effectively captures certain semantic $\hat{\mathbf{x}}_i^t$ from the target image. + +Balancing Semantics and Consistency Between Feature and Image Spaces. Our local perturbation aggregation applied to the source image helps prevent an over-reliance on the target image's semantic cues in the feature space. This is critical because the loss is computed directly from the feature space, which is inherently less expressive and does not adequately capture the intricacies of the image space. As shown in Fig. 4, we compare the global similarity between source and target images optimized using local and global perturbations. The Global-to-Global method achieves the highest similarity, indicating the best-optimized distance between the source and target. However, it results in the lowest ASR (i.e., worst transferability) on LVLMs, suggesting that optimized distance alone is not the key factor and that local perturbations on source can help prevent overfitting and enhance transferability. By encoding enhanced semantic details through multiple overlapping steps, our method gradually builds a richer representation of the input. Meanwhile, the maintained consistency of these local semantic representations prevents them from converging into a uniform or homogenized expression. The combination of these enhanced semantic cues and diverse local expressions significantly improves the transferability of adversarial samples. Thus, we emphasize two critical properties for $\hat{\mathbf{x}}_i\in T(\mathbf{X})$ : + +$$ +\forall i, j, \quad \hat {\mathbf {x}} _ {i} \cap \hat {\mathbf {x}} _ {j} \neq \emptyset \tag {3} +$$ + +$$ +\forall i, j, \quad | \hat {\mathbf {x}} _ {i} \cup \hat {\mathbf {x}} _ {j} | > | \hat {\mathbf {x}} _ {i} | \text {a n d} | \hat {\mathbf {x}} _ {i} \cup \hat {\mathbf {x}} _ {j} | > | \hat {\mathbf {x}} _ {j} | \tag {4} +$$ + +Eq. (3) promotes consistency through shared regions between local areas, while Eq. (4) encourages diversity by incorporating potentially new areas distinct from each local partition. These complementary mechanisms strike a balance between consistency and diversity. Notably, when Eq. (3) significantly dominates Eq. (4), such that $\forall i,j,\hat{\mathbf{x}}_i\cap \hat{\mathbf{x}}_j = \hat{\mathbf{x}}_i = \hat{\mathbf{x}}_j$ , then $\mathcal{T}$ reduces to a consistent selection of a global area. Our framework thus generalizes previous global-global feature matching approaches. In practice, we find that while consistent semantic selection is sometimes necessary for the target image, Eq. (4) is essential for the source image to generate high-quality details with better transferability. + +Local-level Matching via Cropping. It turns out that cropping is effective for fitting Eq. (3) and Eq. (4) when the crop scale ranges between $L$ and $H$ ( $L = 0.5$ and $H = 1.0$ in our experiments). + +Algorithm 1 M-Attack Training Procedure +Require: clean image $\mathbf{X}_{\mathrm{clean}}$ target image $\mathbf{X}_{\mathrm{tar}}$ perturbation budget $\epsilon$ , iterations $n$ , loss function $\mathcal{L}$ surrogate model ensemble $\phi = \{\phi_j\}_{j = 1}^m$ step size $\alpha$ +1: Initialize: $\mathbf{X}_{\mathrm{sou}}^{0} = \mathbf{X}_{\mathrm{clean}}$ (i.e., $\delta_0 = 0$ . $\triangleright$ Initialize adversarial image $\mathbf{X}_{\mathrm{su}}$ +2: for $i = 0$ to $n - 1$ do +3: $\hat{\mathbf{x}}_i^s = \mathcal{T}_s(\mathbf{X}_{\mathrm{sou}}^i),\hat{\mathbf{x}}_i^t = \mathcal{T}_t(\mathbf{X}_{\mathrm{tar}}^i);$ Perform random crop, next step $\mathbf{X}_{\mathrm{sou}}^{i + 1}\leftarrow \hat{\mathbf{x}}_{i + 1}^s$ +4: Compute $\frac{1}{m}\sum_{j = 1}^{m}\mathcal{L}\left(f_{\phi_j}(\hat{\mathbf{x}}_i^s),f_{\phi_j}(\hat{\mathbf{x}}_i^t)\right)$ in Eq. (5); +5: Update $\hat{\mathbf{x}}_{i + 1}^s$ by: +6: $g_{i} = \frac{1}{m}\nabla_{\hat{\mathbf{x}}_{i}^{s}}\sum_{j = 1}^{m}\mathcal{L}\left(f_{\phi_{j}}(\hat{\mathbf{x}}_{i}^{s}),f_{\phi_{j}}(\hat{\mathbf{x}}_{i}^{t})\right);$ +7: $\delta_{i + 1}^{l} = \mathrm{Clip}(\delta_{i}^{l} + \alpha \cdot \mathrm{sign}(g_{i}), - \epsilon ,\epsilon)$ +8: $\hat{\mathbf{x}}_{i + 1}^s = \hat{\mathbf{x}}_i^s +\delta_{i + 1}^l;$ +9: end for +10: return $\mathbf{X}_{\mathrm{adv}}$ .. $\triangleright \mathbf{X}_{\mathrm{sou}}^{n - 1}\rightarrow \mathbf{X}_{\mathrm{adv}}$ + +$\mathcal{T}(\mathbf{X})$ can be defined as the subset of all possible crops within this range. Therefore, randomly cropping $\hat{\mathbf{x}}$ with a crop scale $[a,b]$ such that $L\leq a < b\leq H$ elegantly samples from such mapping. For two consecutive iterations $i$ and $i + 1$ , the overlapped area of pair $(\hat{\mathbf{x}}_i^s,\hat{\mathbf{x}}_{i + 1}^s)$ and $(\hat{\mathbf{x}}_i^t,\hat{\mathbf{x}}_{i + 1}^t)$ ensures consistent semantics between the generated iterations. In contrast, the non-overlapped area is individually processed by each iteration, contributing to the extraction of diverse details. As the cropped extractions combine, the central area integrates shared semantics. The closer the margin it moves towards, the greater the generation of diverse semantic details emerges (see Fig. 5). + +Model Ensemble for Shared, High-quality Semantics. While our matching extracts detailed semantics, commercial black-box models operate on proprietary datasets with undisclosed training objectives. Improving transferability requires better semantic alignment with these target models. We hypothesize that VLMs share certain semantics that transfer more readily to unknown models, and thus employ a model ensemble $\phi = \{f_{\phi_1}, f_{\phi_2}, \ldots, f_{\phi_m}\}$ to capture these shared elements. This approach formulates as: + +$$ +\mathcal {M} _ {\mathcal {T} _ {s}, \mathcal {T} _ {t}} = \mathbb {E} _ {f _ {\phi_ {j}} \sim \phi} \left[ \operatorname {C S} \left(f _ {\phi_ {j}} \left(\hat {\mathbf {x}} _ {i} ^ {s}\right), f _ {\phi_ {j}} \left(\hat {\mathbf {x}} _ {i} ^ {t}\right) \right]. \right. \tag {5} +$$ + +Our ensemble serves dual purposes. At a higher level, it extracts shared semantics that transfer more effectively to target black-box models. At a lower level, it can combine models with complementary perception sizes to enhance perturbation quality. Models with smaller receptive fields (e.g., transformers with smaller patch sizes) extract perturbations with finer details, while those with larger receptive fields preserve better overall structure and pattern. This complementary integration significantly improves the final perturbation quality, as demonstrated in Fig. 6. + +Training. To maximize $\mathcal{M}_{\mathcal{T}_s,\mathcal{T}_t}$ while maintaining imperceptibility constraints, various adversarial optimization frameworks such as I-FGSM [18], PGD [27], and C&W [5], are applicable. For simplicity, we present a practical implementation that uses a uniformly weighted ensemble with I-FGSM, as illustrated in Algorithm 1. More formal and detailed formulations of the problem, along with derivations and additional algorithms, are provided in the Appendix. + +# 5 Experiments + +# 5.1 Setup + +We provide the experimental settings and strong baselines below, with more details in the Appendix. + +Victim Black-box Models and Datasets. We evaluate three leading commercial black-box multimodal large language model families: GPT-4.5, GPT-4o, o1, Claude-3.5-sonnet, Claude-3.7-sonnet, and Gemini-2.0-flash/thinking [33]. We use the NIPS 2017 Adversarial Attacks and Defenses Competition [16] dataset. Following [8], we sample 100 images and resize them to $224 \times 224$ pixels. For enhanced statistical reliability, we then conduct evaluations on 1K images for the comparison with competitive methods in Sec. G.1 of the Appendix. Our source-target image training pairs are randomly sampled. + +Surrogate Models. We employ three CLIP variants [14] as surrogate models: ViT-B/16, ViT-B/32, and ViT-g-14-laion2B-s12B-b42K, for different architectures, training datasets, and feature extrac + +![](images/f87b98b7c52edf4f8db8a40cdcac397bcc9cbde5f8b03cdb024fb96824d977c3.jpg) +Figure 6: 1) Left: visualization of perturbations generated by models with local-to-global matching. Numbers after $\prime$ indicate patch size. Models with smaller receptive fields (14, 16) capture fine details, while larger ones (32) preserve better overall structure. The ensemble integrates these complementary strengths for high-quality perturbation. 2) Right: visualization of perturbation generated by other competitive methods. These perturbations are plotted with $5\times$ magnitude, $1.5\times$ sharpness and saturation for better visual effect. + +![](images/91e552660f3e056c9dd9152c5e591c75e164d7ab8d181333b8b35131efd7e78f.jpg) +Figure 7: Visualization of adversarial samples generated by different methods. + +tion capabilities. We also include results on BLIP-2 [19] in the Appendix. Single-model method [39], if not specified, uses ViT-B/32 as its surrogate model. The ensemble-based methods [12, 38, 8] use the models specified in their papers. + +Baselines. We compare against four recent targeted and transfer-based black-box attackers: AttackVLM [39], SSA-CWA [8], AnyAttack [38], and AdvDiffVLM [12]. + +Hyper-parameters. If not otherwise specified, we set the perturbation budget as $\epsilon = 16$ such as Tab. 2, 4, 5 under the $\ell_{\infty}$ norm and total optimization step to be 300. $\alpha$ is set to 0.75 for Claude-3.5 in Tab. 2, 3 and $\alpha = 1$ elsewhere, including imperceptibility metrics. The ablation study on $\alpha$ is provided in the Appendix. + +# 5.2 Evaluation Metrics + +KMRScore. Previous attack evaluation methods identify keywords matching the "semantic main object" in images [8, 38, 12]. However, unclear definitions of "semantic main object" and matching mechanisms introduce significant human bias and hinder reproducibility. We address these limitations by manually labeling multiple semantic keywords for each image (e.g., "kid, eating, cake" for an image showing a kid eating cake) and establishing three success thresholds: 0.25, 0.5, and 1.0, denoted as $\mathrm{KMR}_a$ , $\mathrm{KMR}_b$ and $\mathrm{KMR}_c$ , respectively. These thresholds correspond to distinct matching levels: at least one keyword matched, over half-matched, and all matched, allowing us to evaluate transferability across different acceptance criteria. To reduce human bias, we leverage GPT-4o [1] for matching semantic keywords against generated descriptions, creating a semi-automated assess + +
MethodModelGPT-4oGemini-2.0Claude-3.5Imperceptibility
KMRaKMRbKMRcASRKMRaKMRbKMRcASRKMRaKMRbKMRcASRl1(↓)l2(↓)
AttackVLM [39]B/160.090.040.000.020.070.020.000.000.060.030.000.010.0340.040
B/320.080.020.000.020.060.020.000.000.040.010.000.000.0360.041
Laion†0.070.040.000.020.070.020.000.010.050.020.000.010.0350.040
AdvDiffVLM [12]Ensemble0.020.000.000.020.010.000.000.010.000.000.000.000.0640.095
SSA-CWA [8]Ensemble0.110.060.000.090.050.020.000.040.070.030.000.050.0590.060
AnyAttack [38]Ensemble0.440.200.040.420.460.210.050.480.250.130.010.230.0480.052
M-Attack (Ours)Ensemble0.820.540.130.950.750.530.110.780.310.180.030.290.0300.036
+ +Table 2: Comparison with the state-of-the-art approaches. The imperceptibility is measured with normalized $\ell_1$ and $\ell_2$ norm of the perturbations by dividing the pixel number and its square root, respectively. indicates ViT-g-14-laion2B-s12B-b42K. + +
εMethodGPT-4oGemini-2.0Claude-3.5Imperceptibility
KMRaKMRbKMRcASRKMRaKMRbKMRcASRKMRaKMRbKMRcASRl1(↓)l2(↓)
4AttackVLM [39]0.080.040.000.020.090.020.000.000.060.030.000.000.0100.011
SSA-CWA [8]0.050.030.000.030.040.030.000.040.030.020.000.010.0150.015
AnyAttack [38]0.070.020.000.050.100.040.000.050.030.020.000.020.0140.015
M-Attack (Ours)0.300.160.030.260.200.110.020.110.050.010.000.010.0090.010
8AttackVLM [39]0.080.020.000.010.080.030.000.020.050.020.000.000.0200.022
SSA-CWA [8]0.060.020.000.040.060.020.000.060.040.020.000.010.0300.030
AnyAttack [38]0.170.060.000.130.200.080.010.140.070.030.000.060.0280.029
M-Attack (Ours)0.740.500.120.820.460.320.080.460.080.030.000.050.0170.020
16AttackVLM [39]0.080.020.000.020.060.020.000.000.040.010.000.000.0360.041
SSA-CWA [8]0.110.060.000.090.050.020.000.040.070.030.000.050.0590.060
AnyAttack [38]0.440.200.040.420.460.210.050.480.250.130.010.230.0480.052
M-Attack (Ours)0.820.540.130.950.750.530.110.780.310.180.030.290.0300.036
+ +Table 3: Ablation study on the impact of $\epsilon$ . + +ment pipeline with human guidance. We verify the approach's robustness by manually reviewing $20\%$ of the outputs and checking the consistency. + +ASR (Attack Success Rate). We further employ widely-used LLM-as-a-judge [40] for benchmarking. We first caption both source and target images through the same commercial LVLM, then compute similarity with GPTScore [11], creating a comprehensive, automated evaluation pipeline. An attack succeeds when the similarity score exceeds 0.3. The appendix contains our detailed prompts for both evaluation methods for reproducibility. + +# 5.3 Comparison of Different Attack Methods + +Tab. 2 shows our superior performance across multiple metrics and LVLMs. Our M-Attack beats all prior methods by large margins. Our proposed KMRScore captures transferability across different levels. $\mathrm{KMR}_a$ with a 0.25 matching rate resembles ASR, while $\mathrm{KMR}_c$ with a 1.0 matching rate acts as a strict metric. Less than $20\%$ of adversarial samples match all semantic keywords, a factor overlooked by previous methods. Our method achieves the highest matching rates at higher thresholds (0.5 and 1.0). This indicates more accurate semantic preservation in critical regions. In contrast, competing methods like AttackVLM and SSA-CWA achieve adequate matching rates at the 0.25 threshold but struggle at higher thresholds. These results show that our local-level matching and ensemble strategies not only fool the victim model into the wrong prediction but also push it to be more confident and detailed in outputting target semantics. + +# 5.4 Ablation + +Local-level Matching. We evaluate four matching strategies: Local-Global, Local-Local (our approach), Global-Local (crop target image only), and Global-Global (no cropping). Fig. 10 presents our results: on Claude, Local-Local matching slightly outperforms Local-Global matching, but the gap is not significant. Global-level matching fails most attacks, showing the importance of Eq. (4) on the source image. We also test traditional augmentation methods, including shear, random rotation, and color jitter, against our local-level matching approach in Fig. 10. Transformations that incorporate a local crop as defined in Eq. (4), like rotation and translation, achieve decent results, while color jitter and global-level matching that do not retain the local area of source images yield significantly lower ASR. Our systematic ablation demonstrates that local-level matching is the key + +![](images/15fae44dbf5cdf725f0b23be9e9470218cb0815db7c6b838b480711fc8d8caff.jpg) +(a) GPT-40 + +![](images/7eaf8bdc8dfc168188271b389f97becd0770d9ff222bbb36d8d219991c608bb8.jpg) +(b) Gemini-2.0 + +![](images/5072f53e826ba8993d8cba6c561c0cf6e620a369c60ffce2c83c5466a94f97e0.jpg) +(c) Claude-3.5 + +(d) GPT-40 +![](images/5b7e4e300acf2fbeebef430fa5d80e94ce78f0627c567e2c003edd26a684623d.jpg) +ours AnyAttack SSA-CWA AttackVLM + +![](images/8b9101a6b4a1309103555103226b347bc193498656262fb4e7155a93ff4ed12e.jpg) +(e) Gemini-2.0 + +![](images/beddea6570d160ff0a33001cb7640538e90aba442e77b848a93818f3a985ad0d.jpg) +(f) Claude-3.5 + +![](images/63e99a2a695a7c3452906a0c8f1113f9b0f533786886c27865765a7cef142de0.jpg) +Figure 8: Ablation study on the impact of steps for different methods. + +![](images/891be27883b06f4becb4eafb91e6a31646a49d335c945a9401e29e8bd3d5fd82.jpg) +Figure 9: Ablation on our two proposed strategies: Local-level matching and ensemble, conducted by separately removing local crop of target image (LCT), local crop of source image (LCS), and ensemble (ENS). Removing LCT has only a marginal impact. + +![](images/de3b9f9d6118ba1d15345e9376f4119665a4478cce708d4da8bf1545f1f2b443.jpg) + +![](images/e64e7747acd43ce52648939d935596159f15f8ae1909e509a497f46e1c51c3be.jpg) + +factor. Although this alignment can be implemented through different operations, such as cropping or translating the image, it fundamentally surpasses conventional augmentation methods by emphasizing the importance of retaining local information. + +Ensemble Design. Model ensemble plays a crucial role in boosting the performance. Ablation studies in Fig. 9 indicate that removing the ensemble results in a $40\%$ reduction in KMR and ASR results. While local-level matching helps capture fine-grained details, the ensemble integrates the complementary strengths of large-receptive field models (which capture overall structure and patterns) with small-receptive field models (which extract finer details). This synergy between local-level matching and the model ensemble is essential, as shown in Fig. 6, with the overall performance gain exceeding the sum of the individual design improvements. Further ablation studies on the ensemble sub-models are provided in the Appendix. + +Perturbation Budget $\epsilon$ . Tab. 3 reveals how perturbation budget $\epsilon$ affects attack performance. Smaller $\epsilon$ values enhance imperceptibility but reduce attack transferability. Our method maintains superior KMR and ASR across most $\epsilon$ settings, while consistently achieving the lowest $\ell_{1}$ and $\ell_{2}$ norms. Overall, our method outperforms other methods under different perturbation constraints. + +![](images/e46e2b5946c922fd5b0bcaa0f0a1aba3d346de12d038a7583aa33753fa2966af.jpg) +Figure 10: Comparison of Local-level Matching to Global-level Matching and other augmentation methods. Only augmentation methods retaining local areas can provide comparable results. + +
MethodKMRaKMRbKMcASR
GPT-o10.830.670.200.94
Claude-3.7-thinking0.300.200.060.35
Gemini-2.0-flash-thinking-exp0.780.590.170.81
+ +Table 4: Results on attacking reasoning LVLMs. + +
MethodKMRaKMBKRCASR
GPT-4.50.820.530.150.95
Claude-3.7-Sonnet0.300.160.030.37
+ +Table 5: Results on attacking the latest LVLMs. + +Computational Budget Steps. Fig. 8 illustrates performance across optimization step limits. Our approach outperforms SSA-CWA and AttackVLM even with iterations reduced to 100. Compared to other methods, our method scales well with computational resources: 200 extra steps improve results by $\sim 10\%$ on both Gemini and Claude. On GPT-4o, ASR increases to near $100\%$ . + +Visualization. Fig. 7 demonstrates the superior imperceptibility and semantic preservation of our method. AttackVLM presents almost no semantics in the perturbation, thus failing in most scenarios. Though semantics are important in achieving successful transfer, SSA-CWA and AnyAttack's adversarial samples present some rough shapes lacking fine details, resulting in a rigid perturbation that contrasts sharply with the original image. Moreover, AnyAttack's adversarial samples exhibit template-like disturbance, which is easy to notice. In contrast, our method focuses on optimizing subtle local perturbations, which not only enhances transferability but also improves imperceptibility over global alignment. + +Results on Reasoning and Latest LVLMs. We also evaluated the transferability of our adversarial samples on the latest models like GPT-4.5, Claude-3.7-sonnet, and reasoning-centric commercial models like GPT-o1, Claude-3.7-thinking, and Gemini-2.0-flash-thinking-exp. Tab. 4 and 5 summarize our findings. Despite their reasoning-centric designs, these models demonstrate equal or weaker robustness to attacks compared to their non-reasoning counterparts. This may be due to the fact that reasoning occurs solely in the text modality, while the paired non-reasoning and reasoning models share similar vision components. + +# 6 Conclusion + +This paper has introduced a simple, powerful approach M-Attack to attack black-box LVLMs. Our method addresses two key limitations in existing attacks: uniform perturbation distribution and vague semantic preservation. Through local-level matching and model ensemble, we formulate the simple attack framework with over $90\%$ success rates against GPT-4.5/4o/o1 by encoding target semantics in local regions and focusing on semantic-rich areas. Ablation shows that local-level matching optimizes semantic details while model ensemble helps with shared semantic and high-quality details by merging the strength of models with different receptive fields. The two parts work synergistically, with performance improvements exceeding the sum of individual contributions. Our findings not only establish a new state-of-the-art attack baseline but also highlight the importance of local semantic details in developing more powerful attack or robust models. + +# Broader Impacts + +By revealing the surprising vulnerability of state-of-the-art black-box models to a minimal yet powerful attack, this work highlights urgent attention about the robustness, transparency, and safety of commercial-grade multimodal large language models that are increasingly integrated into critical decision-making processes. The simplicity and transferability of the attack highlight the insufficiency of current defenses, prompting the need for more systematic security evaluations. Moreover, this work can serve as a practical benchmark for future defenses and inspire the development of standardized risk assessments for black-box AI APIs. Ultimately, the work promotes safer AI development by exposing brittle behaviors that must be addressed to ensure trustworthiness, fairness, and societal alignment in real-world deployments. + +# Acknowledgments + +We would like to thank Yaxin Luo, Tianjun Yao, Jiacheng Liu and Yongqiang Chen for their valuable feedback and suggestions. We also appreciate the constructive comments provided by the reviewers and the area chair. This research is supported by the MBZUAI-WIS Joint Program for AI Research. + +# References + +[1] J. Achiam, S. Adler, S. Agarwal, L. Ahmad, I. Akkaya, F. L. Aleman, D. Almeida, J. Altenschmidt, S. Altman, S. Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. +[2] J.-B. Alayrac, J. Donahue, P. Luc, A. Miech, I. Barr, Y. Hasson, K. Lenc, A. Mensch, K. Millican, M. Reynolds, et al. Flamingo: a visual language model for few-shot learning. In International Conference on Advanced Neural Information Processing Systems, pages 23716-23736, 2022. +[3] Anthropic. Introducing claude 3.5 sonnet, 2024. Accessed: 2025-02-22. +[4] D. Caffagni, F. Cocchi, L. Barsellotti, N. Moratelli, S. Sarto, L. Baraldi, M. Cornia, and R. Cucchiara. The revolution of multimodal large language models: A survey. In Findings of the Association for Computational Linguistics: ACL 2024, pages 13590-13618, 2024. +[5] N. Carlini and D. Wagner. Towards evaluating the robustness of neural networks. In IEEE symposium on security and privacy, pages 39-57, 2017. +[6] H. Chen, Y. Zhang, Y. Dong, X. Yang, H. Su, and J. Zhu. Rethinking model ensemble in transfer-based adversarial attacks. In ICLR, 2024. +[7] J. Chen, H. Guo, K. Yi, B. Li, and M. Elhoseiny. Visualgpt: Data-efficient adaptation of pretrained language models for image captioning. In IEEE/CVF Computer Vision and Pattern Recognition Conference, pages 18030-18040, 2022. +[8] Y. Dong, H. Chen, J. Chen, Z. Fang, X. Yang, Y. Zhang, Y. Tian, H. Su, and J. Zhu. How robust is google's bard to adversarial image attacks? arXiv preprint arXiv:2309.11751, 2023. +[9] Y. Dong, S. Cheng, T. Pang, H. Su, and J. Zhu. Query-efficient black-box adversarial attacks guided by a transfer-based prior. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(12):9536-9548, 2021. +[10] Y. Dong, F. Liao, T. Pang, H. Su, J. Zhu, X. Hu, and J. Li. Boosting adversarial attacks with momentum. In IEEE/CVF Computer Vision and Pattern Recognition Conference, pages 9185-9193, 2018. +[11] J. Fu, S.-K. Ng, Z. Jiang, and P. Liu. Gptscore: Evaluate as you desire. arXiv preprint arXiv:2302.04166, 2023. +[12] Q. Guo, S. Pang, X. Jia, Y. Liu, and Q. Guo. Efficient generation of targeted and transferable adversarial examples for vision-language models via diffusion models. IEEE Transactions on Information Forensics and Security, 20:1333-1348, 2024. +[13] X. Hu, Z. Gan, J. Wang, Z. Yang, Z. Liu, Y. Lu, and L. Wang. Scaling up vision-language pre-training for image captioning. In IEEE/CVF Computer Vision and Pattern Recognition Conference, pages 17980-17989, 2022. +[14] G. Ilharco, M. Wortsman, R. Wightman, C. Gordon, N. Carlini, R. Taori, A. Dave, V. Shankar, H. Namkoong, J. Miller, H. Hajishirzi, A. Farhadi, and L. Schmidt. Openclip, July 2021. +[15] A. Ilyas, L. Engstrom, A. Athalye, and J. Lin. Black-box adversarial attacks with limited queries and information. In International Conference on Machine Learning, 2018. +[16] A. K, B. Hammer, and I. Goodfellow. Nips 2017: Defense against adversarial attack. https://kaggle.com/competitions/nips-2017-defense-against-adversarial-attack, 2017. Kaggle. +[17] D. P. Kingma and J. Ba. Adam: A method for stochastic optimization, 2017. +[18] A. Kurakin, I. J. Goodfellow, and S. Bengio. Adversarial examples in the physical world. In Artificial intelligence safety and security, pages 99-112. 2018. + +[19] J. Li, D. Li, S. Savarese, and S. Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International Conference on Machine Learning, pages 19730–19742, 2023. +[20] J. Li, D. Li, C. Xiong, and S. Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International Conference on Machine Learning, pages 12888-12900, 2022. +[21] Z. Li, D. Liu, C. Zhang, H. Wang, T. Xue, and W. Cai. Enhancing advanced visual reasoning ability of large language models. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 1915-1929, 2024. +[22] Z. Liang, Y. Xu, Y. Hong, P. Shang, Q. Wang, Q. Fu, and K. Liu. A survey of multimodel large language models. In Proceedings of the 3rd International Conference on Computer, Artificial Intelligence and Control Engineering, pages 405-409, 2024. +[23] H. Liu, C. Li, Q. Wu, and Y. J. Lee. Visual instruction tuning. In International Conference on Advanced Neural Information Processing Systems, pages 34892-34916, 2023. +[24] Y. Liu, X. Chen, C. Liu, and D. Song. Delving into transferable adversarial examples and black-box attacks. In International Conference on Learning Representations, 2017. +[25] D.-T. Liu, V.-T. Le, and D. M. Vo. Questioning, answering, and captioning for zero-shot detailed image caption. In Asian Conference on Computer Vision, pages 242-259, 2024. +[26] Z. Ma, J. Hong, M. O. Gul, M. Gandhi, I. Gao, and R. Krishna. Crepe: Can vision-language foundation models reason compositionally? In IEEE/CVF Computer Vision and Pattern Recognition Conference, pages 10910-10921, 2023. +[27] A. Madry, A. Makelov, L. Schmidt, D. Tsipras, and A. Vladu. Towards deep learning models resistant to adversarial attacks. In ICLR, 2018. +[28] K. Marino, M. Rastegari, A. Farhadi, and R. Mottaghi. Ok-vqa: A visual question answering benchmark requiring external knowledge. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, 2019. +[29] Ö. Özdemir and E. Akagündüz. Enhancing visual question answering through question-driven image captions as prompts. In IEEE/CVF Computer Vision and Pattern Recognition Conference, pages 1562-1571, 2024. +[30] S. Park, A. Panigrahi, Y. Cheng, D. Yu, A. Goyal, and S. Arora. Generalizing from simple to hard visual reasoning: Can we mitigate modality imbalance in vlms? arXiv preprint arXiv:2501.02669, 2025. +[31] A. Paszke, S. Gross, F. Massa, A. Lerer, J. Bradbury, G. Chanan, T. Killeen, Z. Lin, N. Gimelshein, L. Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 2019. +[32] A. Salaberria, G. Azkune, O. L. de Lacalle, A. Soroa, and E. Agirre. Image captioning for effective use of language models in knowledge-based visual question answering. Expert Systems with Applications, 212:118669, 2023. +[33] G. Team, R. Anil, S. Borgeaud, J.-B. Alayrac, J. Yu, R. Soricut, J. Schalkwyk, A. M. Dai, A. Hauth, K. Millican, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. +[34] M. Tschannen, M. Kumar, A. Steiner, X. Zhai, N. Houlsby, and L. Beyer. Image captioners are scalable vision learners too. In International Conference on Advanced Neural Information Processing Systems, pages 46830-46855, 2023. +[35] T. Wang, F. Li, L. Zhu, J. Li, Z. Zhang, and H. T. Shen. Cross-modal retrieval: a systematic review of methods and future directions. arXiv preprint arXiv:2308.14263, 2024. + +[36] J. Wu, M. Zhong, S. Xing, Z. Lai, Z. Liu, Z. Chen, W. Wang, X. Zhu, L. Lu, T. Lu, et al. Visionllm v2: An end-to-end generalist multimodal large language model for hundreds of vision-language tasks. In International Conference on Advanced Neural Information Processing Systems, pages 69925-69975, 2025. +[37] D. Zhang, Y. Yu, J. Dong, C. Li, D. Su, C. Chu, and D. Yu. Mm-llms: Recent advances in multimodal large language models. In Findings of the Association for Computational Linguistics ACL 2024, pages 12401–12430, 2024. +[38] J. Zhang, J. Ye, X. Ma, Y. Li, Y. Yang, J. Sang, and D.-Y. Yeung. Anyattack: Towards large-scale self-supervised generation of targeted adversarial examples for vision-language models. arXiv preprint arXiv:2410.05346, 2024. +[39] Y. Zhao, T. Pang, C. Du, X. Yang, C. Li, N.-M. M. Cheung, and M. Lin. On evaluating adversarial robustness of large vision-language models. In International Conference on Advanced Neural Information Processing Systems, pages 54111-54138, 2023. +[40] L. Zheng, W.-L. Chiang, Y. Sheng, S. Zhuang, Z. Wu, Y. Zhuang, Z. Lin, Z. Li, D. Li, E. Xing, et al. Judging llm-as-a-judge with mt-bench and chatbot arena. Advances in Neural Information Processing Systems, 36:46595-46623, 2023. + +# NeurIPS Paper Checklist + +# 1. Claims + +Question: Do the main claims made in the abstract and introduction accurately reflect the paper's contributions and scope? + +Answer: [Yes] + +Justification: Yes. Our main claims and contributions are detailed in Sec. 1. + +Guidelines: + +- The answer NA means that the abstract and introduction do not include the claims made in the paper. +- The abstract and/or introduction should clearly state the claims made, including the contributions made in the paper and important assumptions and limitations. A No or NA answer to this question will not be perceived well by the reviewers. +- The claims made should match theoretical and experimental results, and reflect how much the results can be expected to generalize to other settings. +- It is fine to include aspirational goals as motivation as long as it is clear that these goals are not attained by the paper. + +# 2. Limitations + +Question: Does the paper discuss the limitations of the work performed by the authors? + +Answer: [Yes] + +Justification: Yes. See Sec. C in appendix for more details. + +Guidelines: + +- The answer NA means that the paper has no limitation while the answer No means that the paper has limitations, but those are not discussed in the paper. +- The authors are encouraged to create a separate "Limitations" section in their paper. +- The paper should point out any strong assumptions and how robust the results are to violations of these assumptions (e.g., independence assumptions, noiseless settings, model well-specification, asymptotic approximations only holding locally). The authors should reflect on how these assumptions might be violated in practice and what the implications would be. +- The authors should reflect on the scope of the claims made, e.g., if the approach was only tested on a few datasets or with a few runs. In general, empirical results often depend on implicit assumptions, which should be articulated. +- The authors should reflect on the factors that influence the performance of the approach. For example, a facial recognition algorithm may perform poorly when image resolution is low or images are taken in low lighting. Or a speech-to-text system might not be used reliably to provide closed captions for online lectures because it fails to handle technical jargon. +- The authors should discuss the computational efficiency of the proposed algorithms and how they scale with dataset size. +- If applicable, the authors should discuss possible limitations of their approach to address problems of privacy and fairness. +- While the authors might fear that complete honesty about limitations might be used by reviewers as grounds for rejection, a worse outcome might be that reviewers discover limitations that aren't acknowledged in the paper. The authors should use their best judgment and recognize that individual actions in favor of transparency play an important role in developing norms that preserve the integrity of the community. Reviewers will be specifically instructed to not penalize honesty concerning limitations. + +# 3. Theory assumptions and proofs + +Question: For each theoretical result, does the paper provide the full set of assumptions and a complete (and correct) proof? + +Answer: [Yes] + +Justification: Yes. See Sec. B in appendix for more details. + +# Guidelines: + +- The answer NA means that the paper does not include theoretical results. +- All the theorems, formulas, and proofs in the paper should be numbered and cross-referenced. +- All assumptions should be clearly stated or referenced in the statement of any theorems. +- The proofs can either appear in the main paper or the supplemental material, but if they appear in the supplemental material, the authors are encouraged to provide a short proof sketch to provide intuition. +- Inversely, any informal proof provided in the core of the paper should be complemented by formal proofs provided in appendix or supplemental material. +- Theorems and Lemmas that the proof relies upon should be properly referenced. + +# 4. Experimental result reproducibility + +Question: Does the paper fully disclose all the information needed to reproduce the main experimental results of the paper to the extent that it affects the main claims and/or conclusions of the paper (regardless of whether the code and data are provided or not)? + +# Answer: [Yes] + +Justification: Yes. We have provided code, data, and instructions for reproducing our results in the supplementary materials. + +# Guidelines: + +- The answer NA means that the paper does not include experiments. +- If the paper includes experiments, a No answer to this question will not be perceived well by the reviewers: Making the paper reproducible is important, regardless of whether the code and data are provided or not. +- If the contribution is a dataset and/or model, the authors should describe the steps taken to make their results reproducible or verifiable. +- Depending on the contribution, reproducibility can be accomplished in various ways. For example, if the contribution is a novel architecture, describing the architecture fully might suffice, or if the contribution is a specific model and empirical evaluation, it may be necessary to either make it possible for others to replicate the model with the same dataset, or provide access to the model. In general, releasing code and data is often one good way to accomplish this, but reproducibility can also be provided via detailed instructions for how to replicate the results, access to a hosted model (e.g., in the case of a large language model), releasing of a model checkpoint, or other means that are appropriate to the research performed. +- While NeurIPS does not require releasing code, the conference does require all submissions to provide some reasonable avenue for reproducibility, which may depend on the nature of the contribution. For example +(a) If the contribution is primarily a new algorithm, the paper should make it clear how to reproduce that algorithm. +(b) If the contribution is primarily a new model architecture, the paper should describe the architecture clearly and fully. +(c) If the contribution is a new model (e.g., a large language model), then there should either be a way to access this model for reproducing the results or a way to reproduce the model (e.g., with an open-source dataset or instructions for how to construct the dataset). +(d) We recognize that reproducibility may be tricky in some cases, in which case authors are welcome to describe the particular way they provide for reproducibility. In the case of closed-source models, it may be that access to the model is limited in some way (e.g., to registered users), but it should be possible for other researchers to have some path to reproducing or verifying the results. + +# 5. Open access to data and code + +Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material? + +Answer: [Yes] + +Justification: Yes. We have provided all the codes, data, and instructions to reproduce the results in the supplementary materials. We will also open-source all of them. + +Guidelines: + +- The answer NA means that paper does not include experiments requiring code. +- Please see the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details. +- While we encourage the release of code and data, we understand that this might not be possible, so "No" is an acceptable answer. Papers cannot be rejected simply for not including code, unless this is central to the contribution (e.g., for a new open-source benchmark). +- The instructions should contain the exact command and environment needed to run to reproduce the results. See the NeurIPS code and data submission guidelines (https: //nips.cc/public/guides/CodeSubmissionPolicy) for more details. +- The authors should provide instructions on data access and preparation, including how to access the raw data, preprocessed data, intermediate data, and generated data, etc. +- The authors should provide scripts to reproduce all experimental results for the new proposed method and baselines. If only a subset of experiments are reproducible, they should state which ones are omitted from the script and why. +- At submission time, to preserve anonymity, the authors should release anonymized versions (if applicable). +- Providing as much information as possible in supplemental material (appended to the paper) is recommended, but including URLs to data and code is permitted. + +# 6. Experimental setting/details + +Question: Does the paper specify all the training and test details (e.g., data splits, hyperparameters, how they were chosen, type of optimizer, etc.) necessary to understand the results? + +Answer: [Yes] + +Justification: Yes. See Sec. 5.1 for more details. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The experimental setting should be presented in the core of the paper to a level of detail that is necessary to appreciate the results and make sense of them. +- The full details can be provided either with the code, in appendix, or as supplemental material. + +# 7. Experiment statistical significance + +Question: Does the paper report error bars suitably and correctly defined or other appropriate information about the statistical significance of the experiments? + +Answer: [No] + +Justification: We configure the LLM with temperature $= 0$ to ensure the generalizability and robustness of the results. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The authors should answer "Yes" if the results are accompanied by error bars, confidence intervals, or statistical significance tests, at least for the experiments that support the main claims of the paper. +- The factors of variability that the error bars are capturing should be clearly stated (for example, train/test split, initialization, random drawing of some parameter, or overall run with given experimental conditions). + +- The method for calculating the error bars should be explained (closed form formula, call to a library function, bootstrap, etc.) +- The assumptions made should be given (e.g., Normally distributed errors). +- It should be clear whether the error bar is the standard deviation or the standard error of the mean. +- It is OK to report 1-sigma error bars, but one should state it. The authors should preferably report a 2-sigma error bar than state that they have a $96\%$ CI, if the hypothesis of Normality of errors is not verified. +- For asymmetric distributions, the authors should be careful not to show in tables or figures symmetric error bars that would yield results that are out of range (e.g. negative error rates). +- If error bars are reported in tables or plots, The authors should explain in the text how they were calculated and reference the corresponding figures or tables in the text. + +# 8. Experiments compute resources + +Question: For each experiment, does the paper provide sufficient information on the computer resources (type of compute workers, memory, time of execution) needed to reproduce the experiments? + +Answer: [Yes] + +Justification: Yes. See Sec. F in appendix for more details. + +Guidelines: + +- The answer NA means that the paper does not include experiments. +- The paper should indicate the type of compute workers CPU or GPU, internal cluster, or cloud provider, including relevant memory and storage. +- The paper should provide the amount of compute required for each of the individual experimental runs as well as estimate the total compute. +- The paper should disclose whether the full research project required more compute than the experiments reported in the paper (e.g., preliminary or failed experiments that didn't make it into the paper). + +# 9. Code of ethics + +Question: Does the research conducted in the paper conform, in every respect, with the NeurIPS Code of Ethics https://neurips.cc/public/EthicsGuidelines? + +Answer: [Yes] + +Justification: Yes. We conform with the NeurIPS Code of Ethics. + +Guidelines: + +- The answer NA means that the authors have not reviewed the NeurIPS Code of Ethics. +- If the authors answer No, they should explain the special circumstances that require a deviation from the Code of Ethics. +- The authors should make sure to preserve anonymity (e.g., if there is a special consideration due to laws or regulations in their jurisdiction). + +# 10. Broader impacts + +Question: Does the paper discuss both potential positive societal impacts and negative societal impacts of the work performed? + +Answer: [Yes] + +Justification: We have provided and discussed it in Sec. 6. + +Guidelines: + +- The answer NA means that there is no societal impact of the work performed. +- If the authors answer NA or No, they should explain why their work has no societal impact or why the paper does not address societal impact. +- Examples of negative societal impacts include potential malicious or unintended uses (e.g., disinformation, generating fake profiles, surveillance), fairness considerations (e.g., deployment of technologies that could make decisions that unfairly impact specific groups), privacy considerations, and security considerations. + +- The conference expects that many papers will be foundational research and not tied to particular applications, let alone deployments. However, if there is a direct path to any negative applications, the authors should point it out. For example, it is legitimate to point out that an improvement in the quality of generative models could be used to generate deepfakes for disinformation. On the other hand, it is not needed to point out that a generic algorithm for optimizing neural networks could enable people to train models that generate Deepfakes faster. +- The authors should consider possible harms that could arise when the technology is being used as intended and functioning correctly, harms that could arise when the technology is being used as intended but gives incorrect results, and harms following from (intentional or unintentional) misuse of the technology. +- If there are negative societal impacts, the authors could also discuss possible mitigation strategies (e.g., gated release of models, providing defenses in addition to attacks, mechanisms for monitoring misuse, mechanisms to monitor how a system learns from feedback over time, improving the efficiency and accessibility of ML). + +# 11. Safeguards + +Question: Does the paper describe safeguards that have been put in place for responsible release of data or models that have a high risk for misuse (e.g., pretrained language models, image generators, or scraped datasets)? + +Answer: [NA] + +Justification: The paper releases an optimized dataset intended solely for academic research purposes. The dataset does not involve sensitive or high-risk content, and therefore no specific safeguards or access restrictions were implemented. The risk of misuse is considered minimal in the context of the dataset's scope and intended use. + +# Guidelines: + +- The answer NA means that the paper poses no such risks. +- Released models that have a high risk for misuse or dual-use should be released with necessary safeguards to allow for controlled use of the model, for example by requiring that users adhere to usage guidelines or restrictions to access the model or implementing safety filters. +- Datasets that have been scraped from the Internet could pose safety risks. The authors should describe how they avoided releasing unsafe images. +- We recognize that providing effective safeguards is challenging, and many papers do not require this, but we encourage authors to take this into account and make a best faith effort. + +# 12. Licenses for existing assets + +Question: Are the creators or original owners of assets (e.g., code, data, models), used in the paper, properly credited and are the license and terms of use explicitly mentioned and properly respected? + +Answer: [Yes] + +Justification: All datasets used in the paper are publicly available and open-sourced. The original sources are properly cited in the paper. + +# Guidelines: + +- The answer NA means that the paper does not use existing assets. +- The authors should cite the original paper that produced the code package or dataset. +- The authors should state which version of the asset is used and, if possible, include a URL. +- The name of the license (e.g., CC-BY 4.0) should be included for each asset. +- For scraped data from a particular source (e.g., website), the copyright and terms of service of that source should be provided. +- If assets are released, the license, copyright information, and terms of use in the package should be provided. For popular datasets, paperswithcode.com/datasets has curated licenses for some datasets. Their licensing guide can help determine the license of a dataset. + +- For existing datasets that are re-packaged, both the original license and the license of the derived asset (if it has changed) should be provided. +- If this information is not available online, the authors are encouraged to reach out to the asset's creators. + +# 13. New assets + +Question: Are new assets introduced in the paper well documented and is the documentation provided alongside the assets? + +Answer: [Yes] + +Justification: The paper introduces a new optimized dataset, which has been attached in the supplementary materials, and will publicly available for academic use. Documentation accompanying the release includes details on the data source, collection methodology, intended use. + +Guidelines: + +- The answer NA means that the paper does not release new assets. +- Researchers should communicate the details of the dataset/code/model as part of their submissions via structured templates. This includes details about training, license, limitations, etc. +- The paper should discuss whether and how consent was obtained from people whose asset is used. +- At submission time, remember to anonymize your assets (if applicable). You can either create an anonymized URL or include an anonymized zip file. + +# 14. Crowdsourcing and research with human subjects + +Question: For crowdsourcing experiments and research with human subjects, does the paper include the full text of instructions given to participants and screenshots, if applicable, as well as details about compensation (if any)? + +Answer: [NA] + +Justification: The paper does not involve crowdsourcing nor research with human subjects. + +Guidelines: + +- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. +- Including this information in the supplemental material is fine, but if the main contribution of the paper involves human subjects, then as much detail as possible should be included in the main paper. +- According to the NeurIPS Code of Ethics, workers involved in data collection, curation, or other labor should be paid at least the minimum wage in the country of the data collector. + +# 15. Institutional review board (IRB) approvals or equivalent for research with human subjects + +Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals (or an equivalent approval/review based on the requirements of your country or institution) were obtained? + +Answer: [NA] + +Justification: The paper does not involve crowdsourcing nor research with human subjects. + +Guidelines: + +- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. +- Depending on the country in which research is conducted, IRB approval (or equivalent) may be required for any human subjects research. If you obtained IRB approval, you should clearly state this in the paper. + +- We recognize that the procedures for this may vary significantly between institutions and locations, and we expect authors to adhere to the NeurIPS Code of Ethics and the guidelines for their institution. +- For initial submissions, do not include any information that would break anonymity (if applicable), such as the institution conducting the review. + +# 16. Declaration of LLM usage + +Question: Does the paper describe the usage of LLMs if it is an important, original, or non-standard component of the core methods in this research? Note that if the LLM is used only for writing, editing, or formatting purposes and does not impact the core methodology, scientific rigorousness, or originality of the research, declaration is not required. + +Answer: [NA] + +Justification: The core method development in this research does not involve LLMs as any important, original, or non-standard components. We only use LLMs for evaluations and test our approach's performance. + +# Guidelines: + +- The answer NA means that the core method development in this research does not involve LLMs as any important, original, or non-standard components. +- Please refer to our LLM policy (https://neurips.cc/Conferences/2025/LLM) for what should or should not be described. + +# Appendix + +# Appendix Contents + +A Preliminaries in Problem Formulation 22 +B Preliminary Theoretical Analysis 22 +C Limitations 23 +D Additional Ablation Study 23 + +D.1 Sub-models in the Ensemble 23 +D.2 Crop Size 23 +D.3 Stepsize Parameter 24 + +E Additional Attack Implementation 24 +F More Experimental Setting and Prompt 24 +G Additional Experimental Results 27 + +G.1 Results on 1K Images 27 +G.2 Comparison of Attack Methods on Open-source LVLMs 27 +G.3 Results on Other Vision-language Tasks 27 +G.4 Effectiveness of KMR Metric 27 +G.5 Performance under Different Query Budgets 28 +G.6 Empirical Validation of Baseline Observations and Our Method's Effectiveness 28 + +H Additional Visualizations 29 + +H.1 Adversarial Samples 29 +H.2 Failed Adversarial Samples 29 +H.3 Real-world Scenario Screenshot 30 + +# A Preliminaries in Problem Formulation + +We focus on targeted and transfer-based black-box attacks against vision-language models. Let $f_{\xi}:\mathbb{R}^{H\times W\times 3}\times Y\to Y$ denote the victim model that maps an input image to text description, where $H,W$ are the image height and width and $Y$ denotes all valid text input sequence. $\mathcal{T}$ is the transformation or preprocessing for the raw input image to generate local or global normalized input. Given a target description $o_{\mathrm{tar}}\in Y$ and an input image $\mathbf{X}\in \mathbb{R}^{H\times W\times 3}$ , our goal is to find an adversarial image $\mathbf{X}_{\mathrm{su}} = \mathbf{X}_{\mathrm{cle}} + \delta^{\mathbf{g}}$ that: + +$$ +\underset {\delta} {\arg \min } \| \delta \| _ {p}, \tag {6} +$$ + +$$ +\begin{array}{l} \text {s . t .} f _ {\xi} (\mathcal {T} (\mathbf {X} _ {\text {s o u}})) = o _ {\text {t a r}}, \end{array} +$$ + +where $\| \cdot \|_p$ denotes the $\ell_p$ norm measuring the perturbation magnitude. Since enforcing $f_{\xi}(\mathcal{T}(\mathbf{X}_{\mathrm{su}})) = o_{\mathrm{tar}}$ exactly is intractable, following [39], we instead find a $\mathbf{X}_{\mathrm{tar}}$ matching $o_{\mathrm{tar}}$ . Then we extract semantic features from this image in the embedding space of a surrogate model $f_{\phi}: \mathbb{R}^{3 \times H \times W} \to \mathbb{R}^d$ + +$$ +\arg \max _ {\delta} \mathbf {C S} \left(f _ {\phi} \left(\mathcal {T} \left(\mathbf {X} _ {\text {s o u}}\right)\right), f _ {\phi} \left(\mathcal {T} \left(\mathbf {X} _ {\text {t a r}}\right)\right)\right) \tag {7} +$$ + +$$ +\mathrm {s . t .} \| \delta \| _ {p} \leq \epsilon , +$$ + +where $\mathbf{CS}(a,b) = \frac{a^T b}{\|a\|_2 \|b\|_2}$ denotes the cosine similarity between embeddings. + +However, naively optimizing Eq. (7) only aligns the source and target image in the embedding space without any guarantee of the semantics in the image space. Thus, we propose to embed semantic details through local-level matching. Thus, by introducing Eq. (1), we reformulate Eq. (7) into Eq. (2) in the main text on a local-level alignment. + +# B Preliminary Theoretical Analysis + +Here, we provide a simplified statement capturing the essence of why local matching can yield a strictly lower alignment cost, hence more potent adversarial perturbations than purely global matching. + +Proposition B.1 (Local-to-Local Transport Yields Lower Alignment Cost). Let $\mu_S^G$ and $\mu_T^G$ denote the global distributions of the source image $\hat{\mathbf{x}}^s + \delta$ and target image $\hat{\mathbf{x}}^t$ , respectively, obtained by representing each image as a single feature vector. Let $\mu_S^L$ and $\mu_T^L$ denote the corresponding local distributions, where each image is decomposed into a set of patches $\hat{\mathbf{x}}_i^s (i \in \{1, \dots, N\})$ and $\hat{\mathbf{x}}_j^t (\{j = 1, \dots, M\})$ . Suppose that the cost function $c$ (e.g., a properly defined cosine distance that satisfies the triangle inequality) reflects local or global similarity. Then, under mild conditions (such as partial overlap of semantic content), there exists a joint transport plan $\tilde{\gamma} \in \Pi(\mu_S^L, \mu_T^L)$ such that: + +$$ +W _ {c} \left(\mu_ {S} ^ {L}, \mu_ {T} ^ {L}\right) \leq W _ {c} \left(\mu_ {S} ^ {G}, \mu_ {T} ^ {G}\right), +$$ + +where the optimal transport (OT) distance is defined by + +$$ +W _ {c} \left(\mu_ {S}, \mu_ {T}\right) = +$$ + +$$ +\min _ {\gamma \in \Pi (\mu_ {S}, \mu_ {T})} \sum_ {i, j} c \Big (f (\mathbf {z} _ {i} ^ {S}), f (\mathbf {z} _ {j} ^ {T}) \Big) \gamma \Big (f (\mathbf {z} _ {i} ^ {S}), f (\mathbf {z} _ {j} ^ {T}) \Big). +$$ + +Here, $f$ is a feature extractor, $\mathbf{z}_i^S$ and $\mathbf{z}_j^T$ denote the support points (which correspond either to the single global preprocessed images or to the local patches), and $\Pi(\mu_S, \mu_T)$ is the set of joint distributions with marginals $\mu_S$ and $\mu_T$ . Intuitively, $\gamma\left(f(\mathbf{z}_i^S), f(\mathbf{z}_j^T)\right)$ indicates the amount of mass transported from source patch $\hat{\mathbf{x}}_i^s$ to target patch $\hat{\mathbf{x}}_j^t$ . In many cases the inequality is strict. + +Proof Sketch. Global-to-Global Cost. When the source and target images are each summarized by a single feature vector, we have: + +$$ +W _ {c} \left(\mu_ {S} ^ {G}, \mu_ {T} ^ {G}\right) = c \big (\bar {\mathbf {x}} ^ {s}, \bar {\mathbf {x}} ^ {t} \big), +$$ + +where $\bar{\mathbf{x}}^s = f(\hat{\mathbf{x}}^s +\delta)$ and $\bar{\mathbf{x}}^t = f(\hat{\mathbf{x}}^t)$ . + +Local-to-Local Cost. In contrast, decomposing the images into patches $\mathbf{x}_i^s$ and $\mathbf{x}_j^t$ allows for a more flexible matching: + +$$ +\begin{array}{l} W _ {c} \left(\mu_ {S} ^ {L}, \mu_ {T} ^ {L}\right) = \\ \min \gamma \in \Pi \left(\mu_ {S} ^ {L}, \mu_ {T} ^ {L}\right) \sum_ {i, j} c \big (f (\hat {\mathbf {x}} _ {i} ^ {s}), f (\hat {\mathbf {x}} _ {j} ^ {t}) \big) \gamma \big (f (\hat {\mathbf {x}} _ {i} ^ {s}), f (\hat {\mathbf {x}} _ {j} ^ {t}) \big). \\ \end{array} +$$ + +Under typical conditions (for example, when patches in $(\hat{\mathbf{x}}^s + \delta)$ are close in feature space to corresponding patches in $\hat{\mathbf{x}}^t$ ), the optimal plan $\gamma^*$ matches each patch from the source to a similar patch in the target, thereby achieving a total cost that is lower than (or equal to) the global cost $c(\bar{\mathbf{x}}^s, \bar{\mathbf{x}}^t)$ . When the source and target images share semantic objects that appear at different locations or exhibit partial overlap allowing a form of partial transport, local matching can reduce the transport cost because the global representation fails to capture these partial correspondences. + +This analysis implies that local-to-local alignment is inherently more flexible and can capture subtle correspondences that global alignment misses. + +# C Limitations + +While our method achieves state-of-the-art attack success rates across multiple strong closed-source MLLMs, including GPT-4.5, GPT-4o, Gemini and Claude, this field is evolving rapidly. As newer and potentially more robust models are released, we cannot fully guarantee that our current approach will maintain the same high level of effectiveness. Future work will be needed to adapt and evaluate our attack under shifting model architectures and defense mechanisms. + +# D Additional Ablation Study + +# D.1 Sub-models in the Ensemble + +Individual model ablations further clarify each component's contribution, presented in Tab. 6. CLIP Laion, with its smallest patch size, drives performance on GPT-4o and Gemini-2.0, while CLIP ViT/32 contributes more significantly to Claude-3.5's performance by providing better overall pattern and structure. This also aligns better results of Local-Global Matching on Claude-3.5's than Local-Local Matching results. These patterns suggest Claude prioritizes consistent semantics, whereas GPT-4o and Gemini respond more strongly to detail-rich adversarial samples. + +
Ensemble ModelsGPT-4oGemini-2.0Claude-3.5
KMRaKMRbKMRcASRKMRaKMRbKMRcASRKMRaKMRbKMRcASR
w/o B320.810.550.170.910.740.530.110.810.060.030.000.03
w/o B160.700.430.140.850.650.460.050.760.230.160.030.17
w/o laion0.560.290.070.660.410.290.030.390.180.100.010.17
all0.820.540.130.950.750.530.110.780.240.120.030.26
+ +Table 6: Impact of individual model in the ensemble. The lowest value, except using all sub-models, is labeled in italics and underlined to indicate the importance of the sub-model in the ensemble. + +Regarding the consistency of the architecture or training mythologies for the ensemble surrogate model, we have compared combining CLIP-based models and $\mathrm{CLIP} + \mathrm{BLIP2}$ [19] model. Results in Tab. 7 demonstrate that there is no one-for-all solution for model selection. Adding a different-architecture model, BLIP2, instead of another same-architecture model would increase the performance on GPT-4o and Gemini-2.0 but also decrease the performance on Claude-3.5. This also aligns with the previous analysis of Claude-3.5's preference for a more consistent semantic presentation. + +# D.2 Crop Size + +Tab. 8 presents the impact of crop size parameter $[a, b]$ on the transferability of adversarial samples. Initially we test a smaller crop scale [0.1, 0.4], which results in sub-optimal performance. Then we scale up the crop region to [0.1, 0.9], which greatly improves the result, showing that a consistent + +
Ensemble ModelsGPT-4oGemini-2.0Claude-3.5
KMRaKMRbKMRcASRKMRaKMRbKMRcASRKMRaKMRbKMRcASR
Clip-ViT-g-14-laion2B + Clip-ViT-B/320.700.430.140.850.650.460.050.760.230.160.030.17
Clip-ViT-g-14-laion2B + Blip20.810.570.170.920.790.520.130.850.110.020.010.04
+ +Table 7: Comparison of using isomorphic ensemble and heterogeneous ensemble. + +
ScaleModel Average PerformanceGPT-4oGemini-2.0Claude-3.5
KMRaKMRbKMRcASRKMRaKMRbKMRcASRKMRaKMRbKMRcASR
[0.1, 0.4]0.400.550.350.060.570.690.380.070.630.070.020.000.00
[0.5, 0.9]0.670.800.590.150.950.790.550.120.850.240.140.040.22
[0.5, 1.0]0.660.820.540.130.950.750.530.110.780.240.120.030.26
[0.1, 0.9]0.610.740.550.150.900.780.560.150.810.160.060.000.12
+ +semantic is preferred. Finally, we test [0.5, 0.9] and [0.5, 1.0], which yields a more balanced and generally better result over 3 models. This finding aligns well with our Equ. (3) and Equ. (4) in the main text. + +# D.3 Stepsize Parameter + +We also study the impact of $\alpha$ , presented in Tab. 9. We find selecting $\alpha \in [0.75, 2]$ provides better results. Smaller $\alpha$ values ( $\alpha = 0.25, 5$ ) slow down the convergence, resulting in sub-optimal results. Notably, selecting $\alpha = 0.75$ provides generally better results on Claude-3.5. Thus we use $\alpha = 0.75$ for all optimization-based methods within the main experiment (Tab. 2) and ablation study of $\epsilon$ (Tab. 3) in this paper (SSA-CWA, AttackVLM, and our M-Attack). + +Table 8: Ablation study on impact of the random crop parameter $[a, b]$ . + +
αMethodGPT-4oGemini-2.0Claude-3.5Imperceptibility
KMRaKMRbKMRcASRKMRaKMRbKMRcASRKMRaKMRbKMRcASRl1(↓)l2(↓)
0.25AttackVLM [39]0.060.010.000.020.080.020.000.020.040.020.000.010.0180.023
M-Attack (Ours)0.620.390.090.710.610.370.080.580.140.060.000.070.0150.020
0.5AttackVLM [39]0.070.040.000.030.070.010.000.000.040.020.000.010.0270.033
M-Attack (Ours)0.730.480.170.840.760.540.110.750.210.110.020.150.0290.034
0.75AttackVLM [39]0.040.010.000.010.080.020.010.010.040.020.000.010.0330.039
M-Attack (Ours)0.810.530.140.940.700.510.110.770.310.180.030.290.0290.034
1AttackVLM [39]0.080.040.000.020.090.020.000.000.060.030.000.000.0360.041
M-Attack (Ours)0.820.540.130.950.750.530.110.780.240.120.030.260.0300.036
2AttackVLM [39]0.040.010.000.000.060.010.000.010.010.010.000.000.0380.042
M-Attack (Ours)0.810.630.160.970.760.540.140.850.210.110.010.20.0330.039
+ +Table 9: Ablation study on the impact of $\alpha$ . + +# E Additional Attack Implementation + +We also provide additional algorithms implemented with MI-FFGSM and PGD with ADAM [17] optimizer to show that our flexible framework can be implemented with different adversarial attack methods. Algorithm 2 and Algorithm 3. Since we only apply $\ell_{\infty}$ norm with $\epsilon$ . Thus, to project back after each update, we only need to clip the perturbation. We also provide additional results on M-Attack with MI-FGSM and M-Attack with PGD using ADAM [17] as optimizer, presented in Tab. 10. Results show that using MI-FGSM and PGD in implementation also yield comparable or even better results. Thus, core ideas in our framework are independent of optimization methods. + +# F More Experimental Setting and Prompt + +Platform. The experiments are conducted on $4 \times$ RTX 4090 GPUs. The code is implemented with PyTorch [31]. + +Algorithm 2 M-Attack with MI-FGSM +Require: clean image $\mathbf{X}_{\mathrm{clean}}$ target image $\mathbf{X}_{\mathrm{tar}}$ perturbation budget $\epsilon$ , iterations $n$ , loss function $\mathcal{L}$ surrogate model ensemble $\phi = \{\phi_j\}_{j = 1}^m$ step size $\alpha$ , momentum parameter $\beta$ +1: Initialize: $\mathbf{X}_{\mathrm{sou}}^{0} = \mathbf{X}_{\mathrm{clean}}$ (i.e., $\delta_0 = 0$ $v_{0} = 0$ ; Initialize adversarial image $\mathbf{X}_{\mathrm{su}}$ +2: for $i = 0$ to $n - 1$ do +3: $\hat{\mathbf{x}}_i^s = T_s(\mathbf{X}_\mathrm{sou}^i)$ $\hat{\mathbf{x}}_i^t = T_t(\mathbf{X}_{\mathrm{tar}}^i)$ Perform random crop, next step $\mathbf{X}_{\mathrm{sou}}^{i + 1}\leftarrow \hat{\mathbf{x}}_{i + 1}^s$ +4: Compute $\frac{1}{m}\sum_{j = 1}^{m}\mathcal{L}\left(f_{\phi_j}(\hat{\mathbf{x}}_i^s),f_{\phi_j}(\hat{\mathbf{x}}_i^t)\right)$ in Eq. (5); +5: Update $\hat{\mathbf{x}}_{i + 1}^s,v_i$ by: +6: $g_{i} = \frac{1}{m}\nabla_{\hat{\mathbf{x}}_{i}^{s}}\sum_{j = 1}^{m}\mathcal{L}\left(f_{\phi_{j}}(\hat{\mathbf{x}}_{i}^{s}),f_{\phi_{j}}(\hat{\mathbf{x}}_{i}^{t})\right);$ +7: $v_{i} = v_{i - 1} + \beta g_{i}$ +8: $\delta_{i + 1}^{l} = \operatorname {Clip}(\delta_{i}^{l} + \alpha \cdot \operatorname {sign}(v_{i}), - \epsilon ,\epsilon)$ +9: $\hat{\mathbf{x}}_i^s = \hat{\mathbf{x}}_i^s +\delta_i^l;$ +10: end for +11: return $\mathbf{X}_{\mathrm{adv}}$ . $\triangleright \mathbf{X}_{\mathrm{sou}}^{n - 1}\rightarrow \mathbf{X}_{\mathrm{adv}}$ + +Algorithm 3 M-Attack with PGD-ADAM +Require: Clean image $\mathbf{X}_{\mathrm{clean}}$ target image $\mathbf{X}_{\mathrm{tar}}$ perturbation budget $\epsilon$ , iterations $n$ , loss function $\mathcal{L}$ surrogate model ensemble $\phi = \{\phi_j\}_{j = 1}^m$ step size $\alpha$ Adam parameters $\beta_{1},\beta_{2}$ small constant $\varepsilon$ 1: Initialize: $\mathbf{X}_{\mathrm{sou}}^{0} = \mathbf{X}_{\mathrm{clean}}$ (i.e., $\delta_0 = 0$ ), first moment $m_0 = 0$ , second moment $v_{0} = 0$ , time step $t = 0$ . +2: +3: for $i = 0$ to $n - 1$ do +4: $\hat{\mathbf{x}}_i^s = \mathcal{T}_s(\mathbf{X}_{\mathrm{sou}}^i),\hat{\mathbf{x}}_i^t = \mathcal{T}_t(\mathbf{X}_{\mathrm{tar}}^i);$ Apply random cropping +5: Compute $\frac{1}{m}\sum_{j = 1}^{m}\mathcal{L}\left(f_{\phi_j}(\hat{\mathbf{x}}_i^s),f_{\phi_j}(\hat{\mathbf{x}}_i^t)\right);$ $\triangleright$ Compute loss +6: Compute gradient: +7: $g_{i} = \frac{1}{m}\nabla_{\hat{\mathbf{x}}_{i}^{s}}\sum_{j = 1}^{m}\mathcal{L}\left(f_{\phi_{j}}(\hat{\mathbf{x}}_{i}^{s}),f_{\phi_{j}}(\hat{\mathbf{x}}_{i}^{t})\right);$ +8: $m_i = \beta_1m_{i - 1} + (1 - \beta_1)g_i;$ +9: $v_{i} = \beta_{2}v_{i - 1} + (1 - \beta_{2})g_{i}^{2};$ +10: $\hat{m}_i = m_i / (1 - \beta_1^i),\quad \hat{v}_i = v_i / (1 - \beta_2^i);$ +11: $\delta_{i + 1}^{l} = \mathrm{Clip}(\delta_{i}^{l} + \alpha \cdot \frac{\hat{m}_{i}}{\sqrt{v_{i}} + \varepsilon}, - \epsilon ,\epsilon);$ +12: $\hat{\mathbf{x}}_{i + 1}^{s} = \hat{\mathbf{x}}_{i}^{s} + \delta_{i + 1}^{l};$ +13: end for +14: return $\mathbf{X}_{\mathrm{adv}}$ .. $\triangleright$ Xn-1→Xadv + +
MethodGPT-4oGemini-2.0Claude-3.5Imperceptibility
KMRaKMRbKMRcASRKMRaKMRbKMRcASRKMRaKMRbASRKMRcl1(↓)l2(↓)
I-FGSM0.820.540.130.950.750.530.110.780.310.180.030.290.0360.036
MI-FGSM0.840.620.180.930.840.660.170.910.210.130.040.200.0400.046
PGD-ADAM0.850.560.140.950.790.550.120.860.260.130.010.280.0330.039
+ +Table 10: Comparison of our M-Attack using different adversarial optimization implementations. + +Computation cost. Time taken for optimizing/generating an image: 20.04 seconds on a single RTX 4090 GPU. Memory used for generating an image: 549.78 MB. + +Prompt. We provide two prompts used for KMRScore and GPTScore, respectively. + +KMRScore: the "\{description\}" and "\{keywords\}" are placeholders for text inputs. + +You will be performing a keyword-matching task. You will be given a short description and a list of keywords. Your goal is to find matches between the keywords and the content in the description. Here is the description text: {description} Here is the list of keywords: + +```txt + +{keywords} + +For each keyword in the list, follow these steps: +1. Look for an exact match of the keyword in the description text. +2. If an exact match is not found, look for words or phrases with similar meanings to the keyword. For example, 'bite' could match with 'chew', or 'snow-covered' could match with 'snow'. +3. If you find a match (either exact or similar), record the keyword and its matched content. +``` + +```txt +Your output should be in JSON format, where each key is a keyword from the list, and its value is the matched content from the description. Only include keywords that have matches. For example: {"bite": "chew", "snow": "snow-covered"} Here are some important points to remember: - Only include keywords that have matches in the description. - If a keyword doesn't have a match, do not include it in the JSON. - The matched content should be the exact text from the description, not a paraphrase. - If there are multiple matches for a keyword, use the most relevant or closest match. Please provide your answer in the following format: { // Your JSON output here } Remember to only include the JSON in your answer, with no addition explanation or text. +``` + +ASR: the "\{text_1\}" and "\{text_2\}" are used as placeholders for text inputs. + +Rate the semantic similarity between the following two texts on a scale from 0 to 1. + +**Criteria for similarity measurement:** + +1. **Main Subject Consistency:** If both descriptions refer to the same key subject or object (e.g., a person, food, an event), they should receive a higher similarity score. + +2. **Relevant Description**: If the descriptions are related to the same context or topic, they should also contribute to a higher similarity score. + +3. **Ignore Fine-Grained Details:** Do not penalize differences in **phrasing, sentence structure, or minor variations in detail**. Focus on **whether both descriptions fundamentally describe the same thing.** + +4. **Partial Matches:** If one description contains extra information but does not contradict the other, they should still have a high similarity score. + +```txt +5. **Similarity Score Range:** +- **1.0**: Nearly identical in meaning. +- **0.8-0.9**: Same subject, with highly related descriptions. +- **0.7-0.8**: Same subject, core meaning aligned, even if some details differ. +- **0.5-0.7**: Same subject but different perspectives or missing details. +- **0.3-0.5**: Related but not highly similar (same general theme but different descriptions). +``` + +- **0.0-0.2**: Completely different subjects or unrelated meanings. +Text 1: {text1} +Text 2: {text2} +Output only a single number between 0 and 1. +Do not include any explanation or additional text. + +# G Additional Experimental Results + +# G.1 Results on 1K Images + +We scale up the data size from 100 in Tab. 2 to 1K for better statistical stability. Tab. 11 presents our results. Since labeling multiple semantic keywords for 1000 images is labor-intensive, we provide ASR based on different thresholds as a surrogate for KMRScore. Our method outperforms AnyAttack with a threshold value larger than 0.3. Thus, our method preserves more semantic details that mislead the target model into higher confidence and a more accurate description. + +
thresholdGPT-4oGemini-2.0Claude-3.5
AnyAttackOursAnyAttackOursAnyAttackOurs
0.30.4190.8680.3140.7630.2110.194
0.40.0820.6140.0610.4440.0460.055
0.50.0820.6140.0610.4440.0460.055
0.60.0180.3990.0080.2840.0150.031
0.70.0180.3990.0080.2840.0150.031
0.80.0060.2340.0010.1500.0050.017
0.90.0000.0560.0000.0220.0000.005
+ +# G.2 Comparison of Attack Methods on Open-source LVLMs + +We also test our method on mainstream open-source LLVMs of LLVM and Qwen-VL. Tab. 12 presents our results. + +Table 11: Comparison of results on 1K images. Since labeling 1000 images is labor-intensive, we provide ASR based on different thresholds as a surrogate for KMR. + +
MethodQwen-2.5-VLLLaVA-1.5
KMRaKMRbKMRcASRKMRaKMRbKMRcASR
AttackVLM0.120.040.000.010.110.030.000.07
SSA-CWA0.360.250.040.380.290.170.040.34
AnyAttack0.530.280.090.530.600.320.070.58
M-Attack0.800.650.170.900.850.590.200.95
+ +Table 12: Performance comparison of different attack methods on open-source LVLMs. + +# G.3 Results on Other Vision-language Tasks + +We further evaluate our method on more diverse vision-language tasks, including image captioning and visual question answering. For the image captioning task (we use source dataset of ImageNet and target dataset of COCO2014 val), the results on commercial LVLMs and open-source LVLMs are presented in Tab. 13 and Tab. 14, respectively. For the visual question answering task, the results on commercial LVLMs and open-source LVLMs are presented in Tab. 15 and Tab. 16, respectively. + +# G.4 Effectiveness of KMR Metric + +Tab. 17 reports the KMR scores under three settings: (i) clean source images that are semantically similar to the target (upper bound), (ii) clean images with semantically different content (baseline), and (iii) adversarial images generated by our method (ours), which are also semantically different from the target. The results demonstrate that our proposed KMR metric effectively captures the + +
ModelMethodSPICEBLEU-1BLEU-4METEORROUGE-LCIDEr
GPT-4oAnyAttack [38]2.7226.221.729.3323.697.06
M-Attack (Ours)10.3337.426.1216.2631.4227.31
Gemini-2.0AnyAttack [38]3.1826.910.008.7921.818.46
M-Attack (Ours)7.9734.435.1914.1029.6022.91
Claude-3.5AnyAttack [38]2.3722.990.008.0020.864.46
M-Attack (Ours)2.7023.101.368.3820.945.23
+ +Table 13: Performance comparison on the Image Captioning task with commercial LVLMs. + +
ModelMethodSPICEBLEU-1BLEU-4METEORROUGE-LCIDEr
BLIPAnyAttack [38]4.1346.323.1311.3833.3218.28
M-Attack (Ours)12.0265.7123.1221.0746.8286.23
BLIP2AnyAttack [38]4.4846.685.9611.3833.4420.20
M-Attack (Ours)8.6953.2913.5215.4338.5244.25
InstructBLIPAnyAttack [38]5.8938.793.8312.7728.3616.63
M-Attack (Ours)15.1451.7611.5720.9139.5543.47
+ +Table 14: Performance comparison on the Image Captioning task with open-source LVLMs. + +
VQA Accuracy (%) ↓GPT-4oGemini-2.0Claude-3.5
Pre-attack27.030.215.8
AnyAttack [38]22.426.211.8
M-Attack (Ours)7.814.25.8
+ +Table 15: Performance comparison on the Visual Question Answering task using OK-VQA dataset [28] with commercial LVLMs. + +
VQA Accuracy (%) ↓BLIP2InstructBLIPLLaVA1.5
Pre-attack25.033.630.2
AnyAttack [38]7.221.824.8
M-Attack (Ours)6.813.412.4
+ +Table 16: Performance comparison on the Visual Question Answering task using OK-VQA dataset [28] with open-source LVLMs. + +degree of semantic alignment: it assigns high scores when the source and target are aligned (upper bound), low scores when misaligned (baseline), and intermediate but meaningful scores to adversarial images that successfully mimic the target's semantics. + +# G.5 Performance under Different Query Budgets + +All our results in the paper are based on a single-query setting to demonstrate the efficiency of the attack. To further explore the impact of query budgets, we extend our evaluation to cases with 3 and 5 queries. As shown in Tab. 18, we observe a consistent improvement across all metrics—ASR, $\mathrm{KMR}_a$ , $\mathrm{KMR}_b$ , and $\mathrm{KMR}_c$ —with increased query counts. These results demonstrate a favorable trade-off between attack effectiveness and query efficiency. + +# G.6 Empirical Validation of Baseline Observations and Our Method's Effectiveness + +To quantitatively support the observation that baseline attacks tend to produce uniform-like perturbations, we compute the KL divergence between the empirical perturbation distribution and a uniform distribution over the 33 possible discrete values (from $-16$ to $+16$ ). As summarized in Tab. 19, we compare three settings: clean failed samples (baseline), our method without local cropping, and our full method with local cropping. The results clearly show that the KL divergence increases from + +
ModelSettingSemanticsSource ImageKMRaKMRbKRC
GPT-4oUpper boundSimilarClean1.000.900.40
BaselineDifferentClean0.040.010.00
OursDifferentAdv0.820.540.13
Gemini-2.0Upper boundSimilarClean1.000.950.30
BaselineDifferentClean0.050.020.00
OursDifferentAdv0.750.530.11
Claude-3.5Upper boundSimilarClean1.000.850.35
BaselineDifferentClean0.050.020.00
OursDifferentAdv0.310.180.03
+ +Table 17: The effectiveness of KMR Metric. + +
ModelQuery timeASRKMRaKMRbKMRc
GPT-4o10.950.820.540.20
30.960.930.790.28
51.000.930.830.28
Gemini-2.010.780.770.570.15
30.860.860.680.21
50.880.890.710.23
Claude-3.510.290.310.180.03
30.320.330.200.06
50.440.330.200.06
+ +Table 18: Change of the performance under different query budgets. + +
SettingKL DivergenceGPT-4oGemini-2.0Claude-3.5
Baseline(failed samples)0.012---
Ours(w/o local crop)0.0140.100.080.08
Ours(with local crop)0.0380.950.780.29
+ +Table 19: Comparison of perturbation KL divergence and attack effectiveness on black-box LVLMs. + +0.012 (baseline) to 0.038 (ours with local crop), indicating that our approach generates more nonuniform, structured perturbations. Notably, this increase in distributional divergence is accompanied by a significant gain in attack success rate across all tested black-box LVLMs (from 0.10 to 0.95 on GPT-4o), confirming the effectiveness of semantically guided perturbation design. + +# H Additional Visualizations + +# H.1 Adversarial Samples + +We provide additional visualizations comparing adversarial samples generated using our method and baseline approaches under varying perturbation budgets $(\epsilon)$ . As shown in Fig. 11 and Fig. 12, our method produces adversarial examples with superior imperceptibility compared to existing approaches, like SSA-CWA and AnyAttack, with superior capabilities. + +# H.2 Failed Adversarial Samples + +We present several examples of failed attacks from both prior methods of AttackVLM, SSA-CWA, AnyAttack and our proposed approach to help better understand the challenges of black-box attacks. The visual illustrations are shown in Fig.13, it can be observed that previous methods may fail even when the image is relatively clean or contains only a few objects, whereas our method tends to fail in cases where the image has densely packed targets or contains too many elements. + +![](images/bf2d41e57de9b60924e44ff0a02028f9e4d0296cc381e68517ca946c282fbb2c.jpg) +Figure 11: Visualization of adversarial samples under $\epsilon = 16$ + +![](images/a7d61155128bd9cfce8c5227a46db33d0aa54b932675b58ffabc2fbf09850e51.jpg) +Figure 12: Visualization of adversarial samples with $\epsilon = 4$ and $\epsilon = 8$ . + +# H.3 Real-world Scenario Screenshot + +Fig. 15 and 16 present authentic screenshots of interactions with LVLMs, including GPT-4o, Claude-3.5, and Gemini-2.0, along with their reasoning counterparts. The target image is presented in Fig. 14, with Fig. 14 (b) denoting the target image used for Fig. 15 and Fig. 14 (a) for Fig. 16. Fig. 17 demonstrates results from the latest LVLM models, Claude-3.7-Sonnet and GPT-4.5. These screenshots illustrate how these models respond when exposed to adversarial images in a chat interface. The results reveal significant vulnerabilities in the current commercial LVLMs when processing visual inputs. When confronted with these adversarial images, the models' responses deviate considerably from the expected outputs and instead produce content that aligns with our target semantics. The examples in Fig. 17 show that the output from the target black-box model almost completely matches the intended semantics. These real-world scenario attacks emphasize the urgent need for more robust defenses in multimodal systems. + +![](images/2bdc7b8ee518c20a4be291790db2af01091f050c266d9f229f62816e7b227948.jpg) +Figure 13: Visualization of failed adversarial samples under $\epsilon = 16$ + +![](images/cd61b469e2df551abda301d986686575fefd5ee2b5e9adfa98c83846d57e7d20.jpg) +(a) +Figure 14: Visualization of target images. + +![](images/e9cfd0fa525b4a3ee1909c598592bf736b0874478610fc086efadab321a92a6a.jpg) +(b) + +![](images/82ed2c08ea49a12bb3fcac19e6028f94735ad350ddf572a0b563e42bc1abcfc4.jpg) +(a) GPT-40 + +![](images/267c1f9a80bea19d84930fe24f7b4aa2df6994ed5bda34d21cbcdc90437b572c.jpg) +(b) Gemini-2.0-Flash + +![](images/80576074f1a76801b578fa48c4a0dd661fefa83934ae6e28a9ca75be5f53d6b0.jpg) +(c) Claude-3.5-Sonnet + +![](images/ea8d2b98f66fd0fad5b0b60c1eb7146907cd56d520c6d63e7e7397bf288b412d.jpg) +(d) GPT-o1 + +![](images/9c701994594cb5b4e1f3b9ce26174f75b2f42b5ee38a6ad15dcffe09c12484f2.jpg) +(e) Gemini-2.0-Flash-Thinking + +![](images/2c1f9592bb625bd403f6be0959b6699ca1b1dffe8d9d349235c93b19fca617a6.jpg) +(f) Claude-3.7-Thinking +Figure 15: Example responses from closed-source commercial LVLMs to targeted attacks generated by our method. + +![](images/2bf1cba36b5674b42048a58a10bf4e44dd608db4e2a882208c70a5efd28b0330.jpg) +(a) GPT-40 + +![](images/0c4c6ca6b565d93e7c4b2e73ce63b8a5952194274cbcdd31cc9a240625df242f.jpg) +(b) Gemini-2.0-Flash + +![](images/0e07fad95ae050d783edc4a08ecde442ba889f68eb33ded8b7ec428bf58b481a.jpg) +(c) Claude-3.5-Sonnet + +![](images/d42e626fb5de588460e9a6163e4bcd0bc4ea03c2077ddc6daad20e5d5e3b5c0b.jpg) +(d) GPT-o1 + +![](images/2e3f80dc7a8697163c69d165d0047cc3cd5a39110193be724f856fcb8b3cf8b7.jpg) +(e) Gemini-2.0-Flash-Thinking + +![](images/a6504b7f276c0d01538fdb90fda5f42d2b433006b212af6fdaf734b80982d0db.jpg) +(f) Claude-3.7-Thinking +Figure 16: Example responses from closed-source commercial LVLMs to targeted attacks generated by our method. + +![](images/d9eadea485c62007902bc09475d8910426373adf2bca457bc1edb9538cef874a.jpg) + +![](images/93a69b462f6b256df0c8f7c814689832904968667c2dc9dad03f0ab114dcffaf.jpg) +(a) GPT-4.5 + +![](images/80a2fa1ee9266ea0ce6474c9ad9ad766e6ea1320f6a447c9f7fdd74264cc3ce6.jpg) + +![](images/6096ecb562a11832f61d2738d571ef1f5d3d697678dd0899d37d99860036776f.jpg) +Figure 17: Example responses from the latest closed-source commercial LVLMs to targeted attacks generated by our method. + +![](images/0258dceb88da3822b52351987318683372a573c5dc096bdcc072c4c4cc9cab5c.jpg) +(b) Claude-3.7-Sonnet + +![](images/e3bc4f0234f2d236be9162ce8eac0e169d21b20986c97ff7d686d712c73b1333.jpg) \ No newline at end of file diff --git a/NeurIPS/2025/A Frustratingly Simple Yet Highly Effective Attack Baseline_ Over 90% Success Rate Against the Strong Black-box Models of GPT-4.5_4o_o1/images.zip b/NeurIPS/2025/A Frustratingly Simple Yet Highly Effective Attack Baseline_ Over 90% Success Rate Against the Strong Black-box Models of GPT-4.5_4o_o1/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..ae7a2137f950baeb2a7510439c64c4f8bdea8f44 --- /dev/null +++ b/NeurIPS/2025/A Frustratingly Simple Yet Highly Effective Attack Baseline_ Over 90% Success Rate Against the Strong Black-box Models of GPT-4.5_4o_o1/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:262b691deed1ca0e9b37318cc75963ea93d1da4f2c5a9492dabb678d182c339f +size 1982885 diff --git a/NeurIPS/2025/A Frustratingly Simple Yet Highly Effective Attack Baseline_ Over 90% Success Rate Against the Strong Black-box Models of GPT-4.5_4o_o1/layout.json b/NeurIPS/2025/A Frustratingly Simple Yet Highly Effective Attack Baseline_ Over 90% Success Rate Against the Strong Black-box Models of GPT-4.5_4o_o1/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..3fc65fcc0ad5254886501a61a48bb4baa31987e7 --- /dev/null +++ b/NeurIPS/2025/A Frustratingly Simple Yet Highly Effective Attack Baseline_ Over 90% Success Rate Against the Strong Black-box Models of GPT-4.5_4o_o1/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:595ebb30ab4246fd8e36cbabc4a4b2bd6000dfaa208bec542151a77655ddf4b2 +size 1077624