sample_type stringclasses 1
value | target_idea_id stringlengths 7 7 | target_text stringlengths 103 328 | target_date int64 202k 203k | primary_aspect stringclasses 9
values | contribution_type stringclasses 7
values | task_family stringclasses 9
values | top_priors listlengths 2 5 | joint_coverage_score float64 0 1 | joint_coverage_label stringclasses 4
values | substantive_delta_score float64 0.1 0.9 | residual_type stringclasses 11
values | surface_change_score float64 0 0.9 | evidence_confidence float64 0.55 0.95 | novelty_label stringclasses 4
values | label_confidence float64 0.55 0.95 | label_source stringclasses 1
value | pre_llm_substantive_delta_score float64 0.2 0.8 | pre_llm_surface_change_score float64 0 0.55 | pre_llm_evidence_confidence float64 0.42 0.94 | pre_llm_residual_type stringclasses 18
values | llm_residual_rationale stringlengths 118 320 | llm_residual_model stringclasses 1
value | pre_recomputed_joint_coverage_score float64 0 1 | pre_recomputed_joint_coverage_label stringclasses 4
values | pre_recomputed_top_priors listlengths 5 5 | top_prior_texts listlengths 2 5 | top_prior_coverage_labels listlengths 2 5 | top_prior_coverage_scores listlengths 2 5 | context_source stringclasses 1
value |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
idea_novelty | CI_0001 | Propose a zero-reference training framework that eliminates the need for paired or unpaired reference images by using task-specific non-reference loss functions to implicitly guide enhancement quality. | 202,001 | A7 | CT_task | TF_representation | [
"CI_0379",
"CI_0029",
"CI_0700"
] | 0 | not_covered | 0.7 | new_training_signal | 0.2 | 0.9 | high_novelty | 0.9 | llm_residual_delta_judge_v1 | 0.8 | 0 | 0.62 | new_training_signal | Introduces a novel zero-reference training paradigm using task-specific non-reference losses, distinct from prior weakly-supervised or transductive settings. | Qwen3.5-397B-A17B | 0.1809 | not_covered | [
"CI_0379",
"CI_0195",
"CI_0130",
"CI_0032",
"CI_0029"
] | [
"Define a new transductive zero-shot learning configuration that explicitly accounts for the presence of unrelated images in the target domain, extending the problem setting beyond standard assumptions.",
"Propose a progressive data augmentation strategy that synthesizes edge maps from large-scale unlabeled image... | [
"not_covering",
"not_covering",
"not_covering"
] | [
0.19,
0.07,
0.07
] | recomputed_from_pair_coverage_effective_joint_v1 |
idea_novelty | CI_0002 | Reformulate low-light enhancement as an image-specific high-order curve estimation problem, where a lightweight network predicts pixel-wise curves for dynamic range adjustment instead of performing direct image-to-image mapping. | 202,001 | A4 | CT_task | TF_representation | [
"CI_0064",
"CI_0308",
"CI_0011"
] | 0.04 | not_covered | 0.7 | new_mechanism | 0.2 | 0.9 | high_novelty | 0.9 | llm_residual_delta_judge_v1 | 0.8 | 0 | 0.63 | new_task_definition | Reformulates enhancement as pixel-wise curve estimation rather than direct image mapping, representing a distinct algorithmic mechanism from prior depth/transmission approaches. | Qwen3.5-397B-A17B | 0.2351 | not_covered | [
"CI_0064",
"CI_0308",
"CI_0502",
"CI_0280",
"CI_0159"
] | [
"Propose a two-stage framework that fuses cardinal color information to generate multi-channel depth maps and estimates scene transmission using a multi-scale convolutional network to address color distortion in low-light hazy images.",
"Formulate the problem of realistic image compositing as learning geometric c... | [
"related_not_covering",
"not_covering",
"not_covering"
] | [
0.2,
0.1,
0.1
] | recomputed_from_pair_coverage_effective_joint_v1 |
idea_novelty | CI_0003 | Demonstrate that meta-learning techniques can be effectively applied to full-frame video stabilization to achieve rapid adaptation to unseen videos at test time without requiring ground truth stable data. | 202,403 | A6 | CT_application | TF_application | [
"CI_1526",
"CI_0166",
"CI_0132"
] | 0.04 | not_covered | 0.6 | new_application_constraint | 0.3 | 0.85 | high_novelty | 0.85 | llm_residual_delta_judge_v1 | 0.7 | 0 | 0.74 | new_mechanism | Applying meta-learning to full-frame stabilization for test-time adaptation without ground truth is a novel application constraint, though meta-learning itself is established. | Qwen3.5-397B-A17B | 0.3454 | weakly_covered | [
"CI_1526",
"CI_3379",
"CI_1810",
"CI_0230",
"CI_0287"
] | [
"Introduce a path adjustment strategy to balance camera motion smoothing and frame coverage maximization, thereby minimizing regions with unknown content in full-frame stabilization.",
"Propose a slice pooling layer that projects unordered point features onto an ordered sequence to enable the application of stand... | [
"related_not_covering",
"not_covering",
"not_covering"
] | [
0.44,
0.1,
0.1
] | recomputed_from_pair_coverage_effective_joint_v1 |
idea_novelty | CI_0004 | Design a multi-view, multi-modal diffusion transformer framework capable of generating all necessary photogrammetry modalities through a single end-to-end optimizable process. | 202,502 | A6 | CT_task | TF_representation | [
"CI_4878",
"CI_4390",
"CI_5986"
] | 0.0784 | not_covered | 0.6 | new_system_design | 0.3 | 0.8 | moderate_novelty | 0.8 | llm_residual_delta_judge_v1 | 0.8 | 0 | 0.64 | new_mechanism | Designing an end-to-end diffusion transformer specifically for multi-view photogrammetry modalities constitutes a new system pipeline, though it builds on existing diffusion-transformer architectures. | Qwen3.5-397B-A17B | 0.1195 | not_covered | [
"CI_4390",
"CI_2717",
"CI_0013",
"CI_5985",
"CI_4878"
] | [
"Design a two-branch Transformer architecture that integrates separate 2D and 3D processing streams to effectively fuse multi-modal representations while preserving pretrained 2D vision-language knowledge.",
"Propose a unified diffusion framework that accepts interleaved vision-language prompts to enable both tex... | [
"related_not_covering",
"related_not_covering",
"not_covering"
] | [
0.2,
0.2,
0.07
] | recomputed_from_pair_coverage_effective_joint_v1 |
idea_novelty | CI_0005 | Propose a unified iterative framework that leverages a single pre-trained diffusion model to simultaneously function as both a deterministic denoiser and a stochastic sampler, constructing two complementary regularizers for zero-shot image restoration. | 202,503 | A6 | CT_method | TF_generation | [
"CI_3288",
"CI_5091",
"CI_4566",
"CI_3175",
"CI_3493"
] | 0.1846 | not_covered | 0.6 | new_mechanism | 0.2 | 0.85 | moderate_novelty | 0.85 | llm_residual_delta_judge_v1 | 0.5 | 0 | 0.7 | minor_mechanism_change | Introduces a novel dual-regularizer framework using a single model as both deterministic denoiser and stochastic sampler, a specific architectural mechanism not present in prior diffusion bridge or iterative solver works. | Qwen3.5-397B-A17B | 0.656 | partially_covered | [
"CI_3288",
"CI_5091",
"CI_4566",
"CI_3175",
"CI_3493"
] | [
"Propose a conditional diffusion framework that learns nonlinear diffusion bridges directly between degraded and clean image distributions, bypassing the need to start generation from pure noise.",
"Formulate the reassembly problem as a multi-step iterative denoising process using Diffusion Probabilistic Models t... | [
"related_not_covering",
"related_not_covering",
"related_not_covering",
"related_not_covering",
"related_not_covering"
] | [
0.34,
0.34,
0.34,
0.2,
0.2
] | recomputed_from_pair_coverage_effective_joint_v1 |
idea_novelty | CI_0337 | Propose a learning strategy that dynamically resizes word-level loss based on calculated importance values combining information relevance and content to prioritize video-specific distinctive words over common ones. | 201,901 | A7 | CT_method | TF_representation | [
"CI_0054",
"CI_0029"
] | 0 | not_covered | 0.7 | new_training_signal | 0.2 | 0.85 | high_novelty | 0.85 | llm_residual_delta_judge_v1 | 0.8 | 0 | 0.48 | new_training_signal | Introduces a novel dynamic loss re-weighting mechanism based on word importance and content relevance, distinct from prior data augmentation or shuffling strategies. | Qwen3.5-397B-A17B | 0.0895 | not_covered | [
"CI_0113",
"CI_0183",
"CI_0043",
"CI_0336",
"CI_0054"
] | [
"Propose a training strategy that randomly shuffles input pixels to force the network to learn global statistical distributions rather than local spatial patterns.",
"Propose a progressive data augmentation strategy that synthesizes edge maps from large-scale unlabeled images to pre-train the model, enabling effe... | [
"not_covering",
"not_covering"
] | [
0.1,
0.1
] | recomputed_from_pair_coverage_effective_joint_v1 |
idea_novelty | CI_0338 | Design a framework incorporating hierarchical visual representations and an optimized hierarchical attention mechanism to adaptively extract salient spatial-temporal features across and within video frames. | 201,901 | A6 | CT_method | TF_representation | [
"CI_0026",
"CI_0246",
"CI_0111"
] | 0.38 | weakly_covered | 0.5 | new_mechanism | 0.3 | 0.9 | moderate_novelty | 0.9 | llm_residual_delta_judge_v1 | 0.7 | 0 | 0.68 | new_mechanism | Extends existing spatiotemporal attention with a specific hierarchical structure and optimization; substantive but builds directly on strong prior attention mechanisms. | Qwen3.5-397B-A17B | 0.3126 | weakly_covered | [
"CI_0026",
"CI_0107",
"CI_0112",
"CI_0087",
"CI_0219"
] | [
"Incorporate spatial-temporal attention mechanisms and reference-frame subtraction into 3D ConvNets to exploit the sparsity of relevant motion and focus on moving objects of interest.",
"Develop a hierarchical recurrent architecture that conditions sentence generation on both visual features and linguistic contex... | [
"partial_cover",
"not_covering",
"not_covering"
] | [
0.45,
0.1,
0.1
] | recomputed_from_pair_coverage_effective_joint_v1 |
idea_novelty | CI_0339 | Address the dataset bias in video captioning where common words dominate training gradients, leading to poor recognition of salient but unusual objects. | 201,901 | A3 | CT_theory | TF_representation | [
"CI_0131",
"CI_0291"
] | 0 | not_covered | 0.6 | new_application_constraint | 0.4 | 0.75 | moderate_novelty | 0.75 | llm_residual_delta_judge_v1 | 0.72 | 0 | 0.47 | new_data_distribution | Formalizes dataset bias in video captioning as a specific theoretical constraint; priors cover data integration or new datasets but not this specific bias formulation. | Qwen3.5-397B-A17B | 0 | not_covered | [
"CI_0114",
"CI_0322",
"CI_0131",
"CI_0040",
"CI_0297"
] | [
"Enable the seamless integration of heterogeneous training data with varying annotation granularities (image tags, bounding boxes, pixel masks) to boost segmentation performance without requiring uniform label types.",
"Construct a large-scale, high-resolution bark image dataset with multi-task annotations (speci... | [
"not_covering",
"not_covering"
] | [
0.07,
0.07
] | recomputed_from_pair_coverage_effective_joint_v1 |
idea_novelty | CI_0340 | Integrate a CNN for learning view-invariant features with an MRF using ray potentials to explicitly enforce physical constraints of perspective geometry and occlusion in an end-to-end framework. | 201,901 | A6 | CT_method | TF_representation | [
"CI_0038",
"CI_0184",
"CI_0193"
] | 0.04 | not_covered | 0.7 | new_mechanism | 0.2 | 0.85 | high_novelty | 0.85 | llm_residual_delta_judge_v1 | 0.8 | 0 | 0.6 | new_mechanism | Integrates MRFs with ray potentials into CNNs to enforce physical geometry constraints, a distinct mechanism from prior deep supervision or feature fusion approaches. | Qwen3.5-397B-A17B | 0.247 | not_covered | [
"CI_0038",
"CI_0201",
"CI_0076",
"CI_0152",
"CI_0334"
] | [
"Propose a deep supervision mechanism that sequentially supervises hidden CNN layers with unobserved intermediate shape concepts to regularize learning and improve generalization for geometry estimation.",
"Develop a unified end-to-end framework that jointly performs complementary feature extraction and saliency ... | [
"related_not_covering",
"not_covering",
"not_covering"
] | [
0.34,
0.1,
0.1
] | recomputed_from_pair_coverage_effective_joint_v1 |
idea_novelty | CI_0341 | Propose a stochastic ray sampling strategy to enable efficient backpropagation through unrolled MRF inference, overcoming the memory intractability of training joint probabilistic and neural models. | 201,901 | A6 | CT_method | TF_representation | [
"CI_0006",
"CI_0007"
] | 0 | not_covered | 0.8 | new_mechanism | 0.1 | 0.9 | high_novelty | 0.9 | llm_residual_delta_judge_v1 | 0.8 | 0 | 0.5 | new_mechanism | Proposes a specific stochastic ray sampling strategy to enable backpropagation through unrolled MRF inference, solving a specific memory intractability problem not addressed by prior feature refinement methods. | Qwen3.5-397B-A17B | 0.0895 | not_covered | [
"CI_0249",
"CI_0217",
"CI_0084",
"CI_0173",
"CI_0201"
] | [
"Propose a bidirectional feature refinement mechanism that enables mutual guidance between deep semantic features and shallow detailed features to address scale variation in skeleton detection.",
"Introduce a hierarchical integration strategy that gradually combines nearby feature levels instead of directly fusin... | [
"not_covering",
"not_covering"
] | [
0.1,
0.1
] | recomputed_from_pair_coverage_effective_joint_v1 |
idea_novelty | CI_0342 | Formulate 3D reconstruction as a joint optimization problem that combines learned local appearance cues with global geometric consistency constraints derived from ray potentials. | 201,901 | A2 | CT_task | TF_representation | [
"CI_0086",
"CI_0185"
] | 0 | not_covered | 0.6 | nontrivial_combination | 0.3 | 0.8 | moderate_novelty | 0.8 | llm_residual_delta_judge_v1 | 0.8 | 0 | 0.46 | new_theory_or_concept | Formulates 3D reconstruction by combining learned appearance cues with global geometric constraints via ray potentials; priors on voxel discretization or complexity theory do not cover this specific joint optimization formulation. | Qwen3.5-397B-A17B | 0.0895 | not_covered | [
"CI_0202",
"CI_0056",
"CI_0096",
"CI_0254",
"CI_0284"
] | [
"Define toxels as 4D spatiotemporal hypercube discretizations to generalize 3D voxels for streaming RGBD video segmentation.",
"Prove that the consensus maximisation problem for robust linear model fitting is NP-hard, establishing that no polynomial-time algorithm exists to solve it exactly under standard complex... | [
"not_covering",
"not_covering"
] | [
0.1,
0.1
] | recomputed_from_pair_coverage_effective_joint_v1 |
idea_novelty | CI_0343 | Propose a novel solution that converts ancient paintings to natural images via domain transfer, enabling the direct application of models trained on natural images to solve painting processing tasks. | 201,901 | A1 | CT_task | TF_representation | [
"CI_0317",
"CI_0242"
] | 0 | not_covered | 0.6 | new_application_constraint | 0.4 | 0.85 | moderate_novelty | 0.85 | llm_residual_delta_judge_v1 | 0.72 | 0 | 0.44 | new_application_constraint | Applying standard domain transfer to ancient paintings is a specific domain adaptation under the constraint of bridging the ancient-natural image gap, rather than a new mechanism. | Qwen3.5-397B-A17B | 0.2472 | not_covered | [
"CI_0283",
"CI_0317",
"CI_0056",
"CI_0174",
"CI_0276"
] | [
"Apply conditional generative adversarial networks to the domain of task-specific visual saliency prediction to automatically learn task-dependent features, bridging the gap between bottom-up deep learning and top-down hand-crafted methods.",
"Extend the Asplünd's metric from grey-scale to colour and multivariate... | [
"not_covering",
"not_covering"
] | [
0.1,
0.04
] | recomputed_from_pair_coverage_effective_joint_v1 |
idea_novelty | CI_0344 | Develop an unsupervised generative adversarial network for domain style transfer that employs a compound loss function to preserve the original content and color composition while converting artistic styles to photo-realistic images. | 201,901 | A6 | CT_method | TF_representation | [
"CI_0332",
"CI_0157"
] | 0 | not_covered | 0.4 | minor_variant | 0.6 | 0.9 | moderate_novelty | 0.9 | llm_residual_delta_judge_v1 | 0.8 | 0 | 0.5 | new_mechanism | Proposing a specific compound loss for style transfer is an incremental engineering tweak to existing GAN frameworks rather than a fundamental algorithmic breakthrough. | Qwen3.5-397B-A17B | 0.0895 | not_covered | [
"CI_0160",
"CI_0133",
"CI_0064",
"CI_0225",
"CI_0281"
] | [
"Design a unified deep architecture capable of operating under unsupervised, semi-supervised, and fully-supervised settings by integrating a supervised loss term to correct false correlations found in purely unsupervised learning.",
"Propose an iterative graph construction process guided by a CNN decision functio... | [
"not_covering",
"not_covering"
] | [
0.1,
0.1
] | recomputed_from_pair_coverage_effective_joint_v1 |
idea_novelty | CI_0345 | Curate and release three specialized datasets containing Chinese ancient flower, bird, and landscape paintings to facilitate research in painting analysis and style transfer tasks. | 201,901 | A3 | CT_system | TF_representation | [
"CI_0291",
"CI_0131"
] | 0 | not_covered | 0.8 | new_data_distribution | 0.2 | 0.95 | high_novelty | 0.95 | llm_residual_delta_judge_v1 | 0.72 | 0 | 0.47 | new_data_distribution | The core contribution is the creation and release of a novel data distribution (Chinese ancient paintings) which enables new research directions not possible with existing datasets. | Qwen3.5-397B-A17B | 0.0698 | not_covered | [
"CI_0291",
"CI_0322",
"CI_0297",
"CI_0114",
"CI_0040"
] | [
"Construct a large-scale, high-resolution bark image dataset with multi-task annotations (species and tree diameter) to enable deep learning for fine-grained tree species identification.",
"Enable the seamless integration of heterogeneous training data with varying annotation granularities (image tags, bounding b... | [
"not_covering",
"not_covering"
] | [
0.1,
0.07
] | recomputed_from_pair_coverage_effective_joint_v1 |
idea_novelty | CI_0346 | Construct a 3D model from training images to hallucinate novel viewpoints and poses for augmenting scarce data in few-shot learning. | 201,901 | A6 | CT_method | TF_representation | [
"CI_0213",
"CI_0006",
"CI_0014"
] | 0.38 | weakly_covered | 0.4 | minor_variant | 0.6 | 0.85 | weak_novelty | 0.85 | llm_residual_delta_judge_v1 | 0.5 | 0 | 0.86 | minor_mechanism_change | Target adapts existing 3D reconstruction for data augmentation to a few-shot setting. Core mechanism (SfM for synthesis) is covered by priors; novelty lies in specific application context. | Qwen3.5-397B-A17B | 0.5107 | partially_covered | [
"CI_0213",
"CI_0163",
"CI_0127",
"CI_0135",
"CI_0128"
] | [
"Utilize 3D models reconstructed from unlabelled images via Structure-from-Motion to automatically generate massive training pairs with relative similarity relationships, bypassing manual labeling.",
"Propose a bidirectional feature refinement mechanism that enables mutual guidance between deep semantic features ... | [
"partial_cover",
"not_covering",
"not_covering"
] | [
0.62,
0.1,
0.1
] | recomputed_from_pair_coverage_effective_joint_v1 |
idea_novelty | CI_0347 | Adopt a self-paced learning strategy to select a diverse set of high-quality synthetic images for robust classifier training. | 201,901 | A7 | CT_method | TF_representation | [
"CI_0041",
"CI_0019"
] | 0 | not_covered | 0.7 | new_training_signal | 0.2 | 0.9 | high_novelty | 0.9 | llm_residual_delta_judge_v1 | 0.8 | 0 | 0.48 | new_training_signal | Priors do not cover self-paced learning for selecting synthetic images. This introduces a distinct optimization strategy and selection criterion for synthetic data, representing a new training signal. | Qwen3.5-397B-A17B | 0.0895 | not_covered | [
"CI_0041",
"CI_0027",
"CI_0032",
"CI_0054",
"CI_0130"
] | [
"Utilize coarse-grained class labels as intermediate supervisory signals within a deep supervision framework to enhance fine-grained image classification performance.",
"Propose novel unsupervised criteria for selecting representative foreground objects by combining semantic scores and optical flow features over ... | [
"not_covering",
"not_covering"
] | [
0.1,
0.1
] | recomputed_from_pair_coverage_effective_joint_v1 |
idea_novelty | CI_0348 | Propose a learning-based contour generation method that resolves annotation diversity by converging to an implicit consensus while tolerating imperfect alignment between sketches and ground truth images. | 201,901 | A6 | CT_method | TF_representation | [
"CI_0018",
"CI_0006"
] | 0 | not_covered | 0.75 | new_mechanism | 0.15 | 0.9 | high_novelty | 0.9 | llm_residual_delta_judge_v1 | 0.8 | 0 | 0.5 | new_mechanism | Proposes a specific learning-based method to handle annotation diversity and sketch-image misalignment via implicit consensus. Priors lack mechanisms for this specific noise-tolerant contour generation problem. | Qwen3.5-397B-A17B | 0.0895 | not_covered | [
"CI_0281",
"CI_0128",
"CI_0014",
"CI_0018",
"CI_0196"
] | [
"Introduce a non-parametric classification mechanism that segments objects by finding nearest neighbors among representative foreground and background embeddings selected via objectness and motion cues.",
"Propose a bidirectional feature refinement mechanism that enables mutual guidance between deep semantic feat... | [
"not_covering",
"not_covering"
] | [
0.1,
0.1
] | recomputed_from_pair_coverage_effective_joint_v1 |
idea_novelty | CI_0349 | Introduce a novel loss function designed to handle multiple valid contour annotations for a single image by guiding the network to learn an implicit consensus among diverse human drawings. | 201,901 | A7 | CT_method | TF_representation | [
"CI_0130",
"CI_0032",
"CI_0029",
"CI_0019"
] | 0.0784 | not_covered | 0.6 | new_training_signal | 0.2 | 0.85 | high_novelty | 0.85 | llm_residual_delta_judge_v1 | 0.5 | 0 | 0.84 | training_variant | Introduces a specific loss function to handle multiple valid annotations via implicit consensus, distinct from prior multi-level or structural consistency losses. | Qwen3.5-397B-A17B | 0.6223 | partially_covered | [
"CI_0130",
"CI_0032",
"CI_0029",
"CI_0183",
"CI_0027"
] | [
"Propose a flexible end-to-end framework that integrates image-level, box-level, and pixel-level annotations into a unified training scheme using multiple specific loss functions to leverage diverse supervision sources.",
"Design a composite objective function incorporating additional loss terms alongside adversa... | [
"related_not_covering",
"related_not_covering",
"not_covering",
"not_covering"
] | [
0.44,
0.34,
0.1,
0.1
] | recomputed_from_pair_coverage_effective_joint_v1 |
idea_novelty | CI_0350 | Collect a large-scale dataset of contour drawings to bridge the gap between rigid boundary detection annotations and artistic contour sketches, enabling data-driven learning of scene outlining. | 201,901 | A5 | CT_system | TF_representation | [
"CI_0029",
"CI_0058",
"CI_0033"
] | 0.04 | not_covered | 0.7 | new_data_distribution | 0.3 | 0.9 | high_novelty | 0.9 | llm_residual_delta_judge_v1 | 0.62 | 0 | 0.7 | new_data_organization | Provides a new large-scale dataset bridging rigid boundaries and artistic sketches, a distinct resource contribution compared to prior augmentation strategies. | Qwen3.5-397B-A17B | 0.2976 | weakly_covered | [
"CI_0029",
"CI_0021",
"CI_0216",
"CI_0058",
"CI_0123"
] | [
"Propose a progressive data augmentation strategy that synthesizes edge maps from large-scale unlabeled images to pre-train the model, enabling effective knowledge transfer to the scarce human-sketch domain.",
"Introduce a graph-based representation for 2D edges to identify and prioritize the most structurally si... | [
"related_not_covering",
"not_covering",
"not_covering"
] | [
0.34,
0.07,
0.07
] | recomputed_from_pair_coverage_effective_joint_v1 |
idea_novelty | CI_0351 | Reframe the task of edge extraction from strict pixel-level boundary detection to generating artistic contour drawings that capture salient structural cues with loose alignment and stylistic variation. | 201,901 | A4 | CT_task | TF_representation | [
"CI_0011",
"CI_0164"
] | 0 | not_covered | 0.8 | new_task_definition | 0.1 | 0.95 | high_novelty | 0.95 | llm_residual_delta_judge_v1 | 0.8 | 0 | 0.49 | new_task_definition | Reframes edge extraction from strict pixel detection to generating artistic contours with stylistic variation, defining a fundamentally new problem formulation. | Qwen3.5-397B-A17B | 0.0895 | not_covered | [
"CI_0159",
"CI_0261",
"CI_0016",
"CI_0023",
"CI_0228"
] | [
"Define a unified segmentation task that assigns both semantic labels and instance IDs to every pixel to jointly handle stuff and things in a single output format.",
"Reformulate 3D human pose estimation as a large-scale classification problem by clustering continuous 3D poses into discrete classes to enable effe... | [
"not_covering",
"not_covering"
] | [
0.1,
0.1
] | recomputed_from_pair_coverage_effective_joint_v1 |
idea_novelty | CI_0352 | Propose a 3D deep embedding learning architecture that maps voxels to an embedding space where standard clustering algorithms can distinguish individual fiber instances, addressing the difficulty of separating densely packed objects in low-resolution volumetric data. | 201,901 | A6 | CT_method | TF_representation | [
"CI_0252",
"CI_0006"
] | 0 | not_covered | 0.7 | new_mechanism | 0.2 | 0.85 | high_novelty | 0.85 | llm_residual_delta_judge_v1 | 0.8 | 0 | 0.5 | new_mechanism | Introduces a specific 3D deep embedding architecture for voxel clustering, a distinct mechanism from the domain alignment or feature refinement strategies in the provided priors. | Qwen3.5-397B-A17B | 0.0895 | not_covered | [
"CI_0065",
"CI_0219",
"CI_0252",
"CI_0022",
"CI_0239"
] | [
"Propose an adversarial learning framework that aligns source and target domains directly in the structured output space rather than the feature space to leverage spatial similarities.",
"Propose a bidirectional feature refinement mechanism that enables mutual guidance between deep semantic features and shallow d... | [
"not_covering",
"not_covering"
] | [
0.1,
0.1
] | recomputed_from_pair_coverage_effective_joint_v1 |
idea_novelty | CI_0353 | Introduce a merging post-processing algorithm that integrates overlapping predictions from small sub-volumes into a consistent full-volume instance segmentation, overcoming memory limitations and arbitrary index assignments in large-scale 3D scanning. | 201,901 | A6 | CT_method | TF_representation | [
"CI_0234",
"CI_0157"
] | 0 | not_covered | 0.6 | new_system_design | 0.3 | 0.8 | moderate_novelty | 0.8 | llm_residual_delta_judge_v1 | 0.8 | 0 | 0.5 | new_mechanism | Proposes a specific merging post-processing pipeline to handle sub-volume overlaps and index consistency, addressing system-level constraints not covered by the detection or graph construction priors. | Qwen3.5-397B-A17B | 0.0895 | not_covered | [
"CI_0087",
"CI_0237",
"CI_0014",
"CI_0057",
"CI_0148"
] | [
"Propose a hybrid detection framework that localizes text by sampling and grouping corner points while utilizing position-sensitive segmentation maps for scoring and suppression, avoiding complex post-processing.",
"Propose an iterative graph construction process guided by a CNN decision function to directly deri... | [
"not_covering",
"not_covering"
] | [
0.1,
0.1
] | recomputed_from_pair_coverage_effective_joint_v1 |
idea_novelty | CI_0354 | Demonstrate the first successful application of deep embedding learning for 3D volumetric instance segmentation, specifically enabling the extraction of individual fibers from low-resolution CT scans where traditional methods fail. | 201,901 | A1 | CT_application | TF_application | [
"CI_0168",
"CI_0118"
] | 0 | not_covered | 0.5 | new_application_constraint | 0.4 | 0.9 | moderate_novelty | 0.9 | llm_residual_delta_judge_v1 | 0.72 | 0 | 0.44 | new_application_constraint | Demonstrates the first successful application of deep embedding to low-resolution 3D CT fiber segmentation, solving a specific domain constraint where traditional methods fail, despite using known embedding concepts. | Qwen3.5-397B-A17B | 0.0895 | not_covered | [
"CI_0168",
"CI_0017",
"CI_0010",
"CI_0242",
"CI_0181"
] | [
"Develop an efficient architecture for 3D segmentation that directly processes raw point clouds without converting them to voxels or multi-view images, thereby avoiding information loss and quantization artifacts.",
"Propose a unified framework that performs both spatial selection of normal field-of-view segments... | [
"not_covering",
"not_covering"
] | [
0.1,
0.1
] | recomputed_from_pair_coverage_effective_joint_v1 |
idea_novelty | CI_0355 | Construct a hierarchical grocery image dataset containing natural scenes, iconic product images, and textual descriptions to support multimodal learning for assistive vision. | 201,901 | A5 | CT_system | TF_representation | [
"CI_0048",
"CI_0291"
] | 0 | not_covered | 0.6 | new_data_organization | 0.4 | 0.85 | moderate_novelty | 0.85 | llm_residual_delta_judge_v1 | 0.72 | 0 | 0.47 | new_data_organization | Introduces a novel hierarchical schema combining natural scenes, iconic products, and text for assistive vision, representing a new data organization rather than just a new domain. | Qwen3.5-397B-A17B | 0.0285 | not_covered | [
"CI_0021",
"CI_0216",
"CI_0123",
"CI_0048",
"CI_0174"
] | [
"Augment spatio-temporal cubes with deep appearance features extracted from pre-trained convolutional networks to enhance feature representation for normality modeling.",
"Construct a large-scale, high-resolution bark image dataset with multi-task annotations (species and tree diameter) to enable deep learning fo... | [
"not_covering",
"not_covering"
] | [
0.07,
0.06
] | recomputed_from_pair_coverage_effective_joint_v1 |
idea_novelty | CI_0356 | Define a fine-grained visual recognition task with hierarchical class structures specifically for assisting visually impaired users in navigating grocery store environments. | 201,901 | A4 | CT_task | TF_representation | [
"CI_0016",
"CI_0204"
] | 0 | not_covered | 0.7 | new_task_definition | 0.3 | 0.8 | high_novelty | 0.8 | llm_residual_delta_judge_v1 | 0.8 | 0 | 0.49 | new_task_definition | Defines a specific fine-grained recognition task with hierarchical structures tailored for visually impaired navigation, constituting a new problem formulation beyond generic recognition. | Qwen3.5-397B-A17B | 0.0895 | not_covered | [
"CI_0023",
"CI_0011",
"CI_0016",
"CI_0271",
"CI_0143"
] | [
"Reformulate the scene text detection task to eliminate the necessity of location regression by treating text/non-text prediction as a semantic segmentation problem enriched with instance-level linking.",
"Enable flexible adaptation to datasets with varying numbers of landmarks and auxiliary tasks through a recur... | [
"not_covering",
"not_covering"
] | [
0.1,
0.1
] | recomputed_from_pair_coverage_effective_joint_v1 |
idea_novelty | CI_0357 | Adapt a multi-view generative model to leverage complementary iconic product images alongside natural scene images for improving classification accuracy in cluttered environments. | 201,901 | A6 | CT_application | TF_generation | [
"CI_0144",
"CI_0060"
] | 0 | not_covered | 0.5 | nontrivial_combination | 0.5 | 0.75 | moderate_novelty | 0.75 | llm_residual_delta_judge_v1 | 0.8 | 0 | 0.5 | new_mechanism | Combines multi-view generative modeling with complementary iconic and natural image sources to address clutter, representing a non-obvious integration of existing mechanisms for a specific constraint. | Qwen3.5-397B-A17B | 0 | not_covered | [
"CI_0316",
"CI_0279",
"CI_0144",
"CI_0241",
"CI_0184"
] | [
"Propose a meta-network mechanism that generates layer-specific parameters in a single forward pass to adapt a generic segmentation model to a target object, replacing iterative fine-tuning.",
"Propose a Gaussian Process-based framework to model natural reflectance spectra by exploiting kernel-induced smoothness ... | [
"not_covering",
"not_covering"
] | [
0.07,
0.04
] | recomputed_from_pair_coverage_effective_joint_v1 |
idea_novelty | CI_0358 | Propose a unified framework that applies facility location and minimum dispersion models to select diverse data subsets for reducing training costs and labeling efforts. | 201,901 | A6 | CT_application | TF_application | [
"CI_0166",
"CI_0035"
] | 0 | not_covered | 0.6 | nontrivial_combination | 0.3 | 0.85 | high_novelty | 0.85 | llm_residual_delta_judge_v1 | 0.8 | 0 | 0.5 | new_mechanism | Combines facility location and minimum dispersion models into a unified framework for data subset selection; priors do not cover this specific optimization formulation. | Qwen3.5-397B-A17B | 0.0895 | not_covered | [
"CI_0230",
"CI_0144",
"CI_0166",
"CI_0132",
"CI_0149"
] | [
"Propose a slice pooling layer that projects unordered point features onto an ordered sequence to enable the application of standard sequence modeling algorithms on point clouds.",
"Integrate semantic relatedness knowledge to model transferable properties between categories for adapting image classifiers to objec... | [
"not_covering",
"not_covering"
] | [
0.1,
0.1
] | recomputed_from_pair_coverage_effective_joint_v1 |
idea_novelty | CI_0359 | Apply data subset selection to accelerate hyper-parameter tuning by identifying representative small subsets for rapid experimental iteration. | 201,901 | A1 | CT_application | TF_application | [
"CI_0181",
"CI_0017"
] | 0 | not_covered | 0.5 | new_application_constraint | 0.4 | 0.9 | moderate_novelty | 0.9 | llm_residual_delta_judge_v1 | 0.72 | 0 | 0.44 | new_application_constraint | Applies data subset selection specifically to accelerate hyper-parameter tuning; the mechanism exists but the target application and constraints are novel relative to priors. | Qwen3.5-397B-A17B | 0.0895 | not_covered | [
"CI_0232",
"CI_0233",
"CI_0277",
"CI_0242",
"CI_0270"
] | [
"Enable unsupervised training on arbitrary uncalibrated monocular video streams by relying only on temporal coherence and ego-motion assumptions, eliminating the need for calibrated sensors or ground-truth depth labels.",
"Propose a strategy to transfer instance embedding networks trained on static images to unsu... | [
"not_covering",
"not_covering"
] | [
0.1,
0.1
] | recomputed_from_pair_coverage_effective_joint_v1 |
idea_novelty | CI_0360 | Utilize unsupervised diversity-based selection to extract representative frames from redundant video streams for efficient dataset creation. | 201,901 | A1 | CT_application | TF_application | [
"CI_0168",
"CI_0053"
] | 0 | not_covered | 0.4 | minor_variant | 0.5 | 0.85 | moderate_novelty | 0.85 | llm_residual_delta_judge_v1 | 0.72 | 0 | 0.44 | new_application_constraint | Adapts unsupervised diversity selection to video frame extraction; primarily a domain adaptation of existing subset selection methods with minor procedural tweaks. | Qwen3.5-397B-A17B | 0.0895 | not_covered | [
"CI_0181",
"CI_0017",
"CI_0118",
"CI_0242",
"CI_0168"
] | [
"Develop an efficient architecture for 3D segmentation that directly processes raw point clouds without converting them to voxels or multi-view images, thereby avoiding information loss and quantization artifacts.",
"Enable user-controllable generation and automatic scene annotation by explicitly modeling the int... | [
"not_covering",
"not_covering"
] | [
0.1,
0.1
] | recomputed_from_pair_coverage_effective_joint_v1 |
idea_novelty | CI_0361 | Enhance active learning by combining uncertainty sampling with diversified subset selection to reduce redundancy and improve labeling efficiency. | 201,901 | A6 | CT_method | TF_representation | [
"CI_0007",
"CI_0006"
] | 0 | not_covered | 0.3 | nontrivial_combination | 0.6 | 0.9 | weak_novelty | 0.9 | llm_residual_delta_judge_v1 | 0.8 | 0 | 0.5 | new_mechanism | Combining uncertainty sampling with diversity selection is a known strategy in active learning; priors discuss unrelated feature fusion, so novelty is limited to applying this standard combination to a specific context. | Qwen3.5-397B-A17B | 0.0895 | not_covered | [
"CI_0008",
"CI_0172",
"CI_0231",
"CI_0207",
"CI_0014"
] | [
"Introduce a hierarchical integration strategy that gradually combines nearby feature levels instead of directly fusing all levels to ease optimization and handle significant feature discrepancies.",
"Propose a bidirectional feature refinement mechanism that enables mutual guidance between deep semantic features ... | [
"not_covering",
"not_covering"
] | [
0.1,
0.1
] | recomputed_from_pair_coverage_effective_joint_v1 |
idea_novelty | CI_0362 | Construct a synthetic diagnostic dataset with programmatically generated ground-truth reasoning steps to enable bias-controlled evaluation and intermediate process analysis for referring expression tasks. | 201,901 | A8 | CT_system | TF_reasoning | [
"CI_0129",
"CI_0291"
] | 0 | not_covered | 0.7 | new_evaluation_protocol | 0.4 | 0.85 | high_novelty | 0.85 | llm_residual_delta_judge_v1 | 0.74 | 0 | 0.43 | new_evaluation_protocol | Creating a synthetic dataset with programmable reasoning steps for bias control introduces a new evaluation protocol and data generation method distinct from prior static dataset constructions. | Qwen3.5-397B-A17B | 0 | not_covered | [
"CI_0085",
"CI_0200",
"CI_0129",
"CI_0333",
"CI_0028"
] | [
"Construct a new benchmark dataset comprising pixel-level annotated images of homogeneous object clusters to evaluate segmentation performance in this specific domain.",
"Construct a large-scale, high-resolution bark image dataset with multi-task annotations (species and tree diameter) to enable deep learning for... | [
"not_covering",
"not_covering"
] | [
0.07,
0.03
] | recomputed_from_pair_coverage_effective_joint_v1 |
idea_novelty | CI_0363 | Propose a modular neural network architecture that executes explicit functional programs to perform step-by-step visual reasoning for referring expression comprehension. | 201,901 | A6 | CT_method | TF_reasoning | [
"CI_0070",
"CI_0024",
"CI_0152"
] | 0 | not_covered | 0.6 | new_mechanism | 0.3 | 0.8 | moderate_novelty | 0.8 | llm_residual_delta_judge_v1 | 0.8 | 0 | 0.64 | new_mechanism | Executing explicit functional programs for visual reasoning defines a new architectural mechanism; priors cover unrelated tasks like chart reading or gaze integration, leaving the core program-execution idea novel. | Qwen3.5-397B-A17B | 0.0816 | not_covered | [
"CI_0070",
"CI_0173",
"CI_0122",
"CI_0112",
"CI_0022"
] | [
"Propose an end-to-end neural network architecture capable of directly reading and extracting unique textual answers from visual elements within a chart, bypassing fixed vocabulary constraints.",
"Integrate human gaze estimation as an explicit auxiliary modality alongside language and visual features to improve t... | [
"not_covering",
"not_covering",
"not_covering"
] | [
0.19,
0.07,
0.07
] | recomputed_from_pair_coverage_effective_joint_v1 |
idea_novelty | CI_0364 | Demonstrate that a segmentation module trained for final output can be universally attached to intermediate reasoning steps to visualize the entire step-by-step reasoning process without additional training. | 201,901 | A6 | CT_method | TF_reasoning | [
"CI_0279",
"CI_0257"
] | 0 | not_covered | 0.7 | new_mechanism | 0.2 | 0.85 | high_novelty | 0.85 | llm_residual_delta_judge_v1 | 0.8 | 0 | 0.5 | new_mechanism | Proposes a novel zero-shot mechanism to repurpose final-output segmentation modules for intermediate reasoning visualization, distinct from prior two-stage generative refinement loops. | Qwen3.5-397B-A17B | 0 | not_covered | [
"CI_0237",
"CI_0279",
"CI_0074",
"CI_0157",
"CI_0257"
] | [
"Design a sequential two-stage generative framework where a second network predicts segmentation from the generated image to provide feedback, thereby refining the visual quality and sharpness of the first stage's output.",
"Introduce a two-stage convolutional neural network that refines initial 3D bounding box e... | [
"not_covering",
"not_covering"
] | [
0.07,
0.07
] | recomputed_from_pair_coverage_effective_joint_v1 |
idea_novelty | CI_0365 | Provide empirical evidence that neural module networks can correctly handle false-premise inputs by predicting no foreground object, validating their intended logical behavior quantitatively. | 201,901 | A8 | CT_evaluation | TF_benchmark | [
"CI_0136",
"CI_0028"
] | 0 | not_covered | 0.6 | new_evaluation_protocol | 0.4 | 0.9 | high_novelty | 0.9 | llm_residual_delta_judge_v1 | 0.74 | 0 | 0.43 | new_evaluation_protocol | Establishes a specific evaluation protocol for false-premise handling in neural module networks, filling a benchmarking gap not addressed by prior dataset construction efforts. | Qwen3.5-397B-A17B | 0.0895 | not_covered | [
"CI_0229",
"CI_0012",
"CI_0290",
"CI_0324",
"CI_0093"
] | [
"Release a new real-world testing dataset containing ground truth for occluded and missing parts to address the lack of benchmarks for evaluating single-view reconstruction in real scenarios.",
"Construct a large-scale real-world dataset of home surveillance video clips with binary relevance labels to benchmark r... | [
"not_covering",
"not_covering"
] | [
0.07,
0.07
] | recomputed_from_pair_coverage_effective_joint_v1 |
idea_novelty | CI_0366 | Propose a pre-processing image restoration framework that removes adherent rain effects to improve downstream segmentation performance, avoiding the need for task-specific retraining or domain adaptation. | 201,901 | A6 | CT_task | TF_representation | [
"CI_0295",
"CI_0294",
"CI_0241",
"CI_0008"
] | 0.0784 | not_covered | 0.5 | new_application_constraint | 0.5 | 0.85 | moderate_novelty | 0.85 | llm_residual_delta_judge_v1 | 0.8 | 0 | 0.64 | new_mechanism | Applies rain removal specifically as a plug-and-play pre-processing constraint to enable segmentation without retraining, differing from prior joint estimation or density classification approaches. | Qwen3.5-397B-A17B | 0.1979 | not_covered | [
"CI_0295",
"CI_0294",
"CI_0013",
"CI_0241",
"CI_0199"
] | [
"Design a residual-aware classifier that utilizes the residual component of rainy images to accurately classify rain density levels as a precursor for adaptive restoration.",
"Propose a unified framework that jointly estimates rain density levels and removes rain streaks, enabling the network to adaptively handle... | [
"related_not_covering",
"related_not_covering",
"not_covering",
"not_covering"
] | [
0.2,
0.2,
0.1,
0.07
] | recomputed_from_pair_coverage_effective_joint_v1 |
idea_novelty | CI_0367 | Introduce a real-world narrow-baseline stereo dataset capturing paired images with one lens affected by real water droplets and the other kept clear to enable supervised training for de-raining. | 201,901 | A3 | CT_system | TF_representation | [
"CI_0297",
"CI_0114",
"CI_0265"
] | 0.04 | not_covered | 0.7 | new_data_distribution | 0.3 | 0.9 | high_novelty | 0.9 | llm_residual_delta_judge_v1 | 0.72 | 0 | 0.7 | new_data_distribution | Introduces a novel real-world narrow-baseline stereo distribution with specific water droplet occlusion, distinct from prior synthetic rain datasets. | Qwen3.5-397B-A17B | 0.2325 | not_covered | [
"CI_0297",
"CI_0322",
"CI_0291",
"CI_0114",
"CI_0040"
] | [
"Construct a large-scale synthetic dataset containing images with explicit rain density labels (heavy, medium, light) to facilitate the training of density-aware de-raining models.",
"Compile and release a large-scale continuous sign language dataset with sentence-level annotations to facilitate research in unseg... | [
"related_not_covering",
"not_covering",
"not_covering"
] | [
0.34,
0.1,
0.07
] | recomputed_from_pair_coverage_effective_joint_v1 |
idea_novelty | CI_0368 | Develop a GPU-based synthetic data generation technique to realistically render adherent water droplets and streaks on any image, enabling large-scale proxy evaluation of rain effects on segmentation tasks. | 201,901 | A3 | CT_task | TF_representation | [
"CI_0297",
"CI_0322",
"CI_0265"
] | 0.04 | not_covered | 0.6 | new_data_distribution | 0.4 | 0.85 | moderate_novelty | 0.85 | llm_residual_delta_judge_v1 | 0.72 | 0 | 0.7 | new_data_distribution | Proposes a specialized GPU rendering technique for adherent droplets/streaks, creating a new synthetic data distribution for segmentation robustness evaluation. | Qwen3.5-397B-A17B | 0.1863 | not_covered | [
"CI_0297",
"CI_0040",
"CI_0131",
"CI_0291",
"CI_0322"
] | [
"Construct a large-scale synthetic dataset containing images with explicit rain density labels (heavy, medium, light) to facilitate the training of density-aware de-raining models.",
"Introduce two new datasets capturing hands in unconstrained wild environments and specific occlusion scenarios to overcome the lim... | [
"related_not_covering",
"not_covering",
"not_covering"
] | [
0.34,
0.07,
0.07
] | recomputed_from_pair_coverage_effective_joint_v1 |
idea_novelty | CI_0369 | Establish a proxy evaluation protocol that applies synthetic rain augmentation to existing clean segmentation benchmarks to systematically study and quantify the impact of lens occlusion on semantic segmentation performance. | 201,901 | A8 | CT_dataset | TF_benchmark | [
"CI_0297",
"CI_0136",
"CI_0028"
] | 0.04 | not_covered | 0.65 | new_evaluation_protocol | 0.35 | 0.9 | high_novelty | 0.9 | llm_residual_delta_judge_v1 | 0.74 | 0 | 0.57 | new_evaluation_protocol | Establishes a systematic proxy evaluation protocol using synthetic rain augmentation to quantify lens occlusion impacts on segmentation, distinct from training datasets. | Qwen3.5-397B-A17B | 0.2176 | not_covered | [
"CI_0297",
"CI_0247",
"CI_0274",
"CI_0082",
"CI_0129"
] | [
"Construct a large-scale synthetic dataset containing images with explicit rain density labels (heavy, medium, light) to facilitate the training of density-aware de-raining models.",
"Release a new real-world testing dataset containing ground truth for occluded and missing parts to address the lack of benchmarks ... | [
"related_not_covering",
"not_covering",
"not_covering"
] | [
0.2,
0.07,
0.07
] | recomputed_from_pair_coverage_effective_joint_v1 |
idea_novelty | CI_0370 | Propose cropping-inside residual units that remove padding-affected features within residual blocks to eliminate positional bias caused by convolution padding in tracking tasks. | 201,901 | A6 | CT_task | TF_representation | [
"CI_0013",
"CI_0315"
] | 0 | not_covered | 0.7 | new_mechanism | 0.2 | 0.9 | high_novelty | 0.9 | llm_residual_delta_judge_v1 | 0.8 | 0 | 0.5 | new_mechanism | Introduces a specific architectural mechanism (cropping-inside residual units) to solve positional bias from padding, distinct from prior multi-task or segmentation merging approaches. | Qwen3.5-397B-A17B | 0.0698 | not_covered | [
"CI_0260",
"CI_0013",
"CI_0315",
"CI_0153",
"CI_0217"
] | [
"Propose a baseline heuristic that merges outputs from independent semantic and instance segmentation systems via post-processing to establish an initial performance reference for the unified task.",
"Design a unified multi-task learning framework that learns semantic correspondences and contextual relationships ... | [
"not_covering",
"not_covering"
] | [
0.1,
0.1
] | recomputed_from_pair_coverage_effective_joint_v1 |
idea_novelty | CI_0371 | Design deeper and wider backbone architectures with controlled receptive field sizes and network strides to balance feature discriminability and localization precision for visual tracking. | 201,901 | A6 | CT_method | TF_representation | [
"CI_0302",
"CI_0299"
] | 0 | not_covered | 0.5 | minor_variant | 0.4 | 0.85 | moderate_novelty | 0.85 | llm_residual_delta_judge_v1 | 0.8 | 0 | 0.5 | new_mechanism | Proposes specific hyperparameter configurations (depth, width, stride) for backbones; while tailored for tracking, it represents an incremental architectural tuning rather than a new mechanism. | Qwen3.5-397B-A17B | 0.0895 | not_covered | [
"CI_0148",
"CI_0145",
"CI_0255",
"CI_0218",
"CI_0203"
] | [
"Propose a hybrid visual servoing framework that dynamically switches between global eye-to-hand and local eye-in-hand 3D sensors based on target distance and occlusion status to ensure continuous tracking.",
"Incorporate global average pooling and a multi-scale training strategy into the classifier architecture ... | [
"not_covering",
"not_covering"
] | [
0.1,
0.1
] | recomputed_from_pair_coverage_effective_joint_v1 |
idea_novelty | CI_0372 | Conduct a systematic analysis identifying receptive field size, network stride, and feature padding as critical factors causing performance degradation when applying deep classification networks to Siamese tracking. | 201,901 | A2 | CT_theory | TF_representation | [
"CI_0039",
"CI_0096"
] | 0 | not_covered | 0.6 | new_theory_or_concept | 0.3 | 0.8 | moderate_novelty | 0.8 | llm_residual_delta_judge_v1 | 0.8 | 0 | 0.46 | new_theory_or_concept | Provides a systematic theoretical analysis identifying specific factors (receptive field, stride, padding) causing degradation in Siamese tracking, offering new conceptual understanding beyond prior general supervision theories. | Qwen3.5-397B-A17B | 0.0698 | not_covered | [
"CI_0171",
"CI_0175",
"CI_0170",
"CI_0009",
"CI_0039"
] | [
"Formulate a probabilistic framework that theoretically explains how supervising intermediate concepts improves generalization and generalizes conventional multi-task and deep supervision schemes.",
"Formulate fine-grained recognition as a problem of disentangling pose and appearance by defining object representa... | [
"not_covering",
"not_covering"
] | [
0.07,
0.07
] | recomputed_from_pair_coverage_effective_joint_v1 |
idea_novelty | CI_0373 | Propose a systematic analysis framework to evaluate the trade-offs between diversity, representation, coverage, and importance across different video domains to guide model selection. | 201,901 | A1 | CT_application | TF_application | [
"CI_0181",
"CI_0232"
] | 0 | not_covered | 0.6 | new_evaluation_protocol | 0.3 | 0.8 | moderate_novelty | 0.8 | llm_residual_delta_judge_v1 | 0.72 | 0 | 0.44 | new_application_constraint | Proposes a systematic framework for evaluating trade-offs across domains, establishing a new evaluation protocol rather than a new model mechanism. | Qwen3.5-397B-A17B | 0.0895 | not_covered | [
"CI_0118",
"CI_0242",
"CI_0181",
"CI_0233",
"CI_0017"
] | [
"Enable unsupervised training on arbitrary uncalibrated monocular video streams by relying only on temporal coherence and ego-motion assumptions, eliminating the need for calibrated sensors or ground-truth depth labels.",
"Enable domain adaptation for subset of target categories by decoupling the learning of doma... | [
"not_covering",
"not_covering"
] | [
0.1,
0.1
] | recomputed_from_pair_coverage_effective_joint_v1 |
idea_novelty | CI_0374 | Develop an optimized implementation strategy using lazy evaluation and memoization techniques to enable real-time, interactive generation of video summaries with varying constraints. | 201,901 | A9 | CT_system | TF_representation | [
"CI_0105",
"CI_0121"
] | 0 | not_covered | 0.4 | minor_variant | 0.6 | 0.9 | moderate_novelty | 0.9 | llm_residual_delta_judge_v1 | 0.74 | 0 | 0.42 | new_system_design | Applies standard lazy evaluation and memoization to video summarization; primarily an engineering optimization rather than a novel algorithmic mechanism. | Qwen3.5-397B-A17B | 0.1959 | not_covered | [
"CI_0090",
"CI_0305",
"CI_0105",
"CI_0121",
"CI_0085"
] | [
"Construct a fully automated end-to-end diagnosis system that integrates 3D nodule detection and classification subnetworks to perform patient-level lung cancer diagnosis comparable to experienced doctors.",
"Curate a new dataset of 360-degree videos specifically annotated for spatio-temporal highlight detection ... | [
"not_covering",
"not_covering"
] | [
0.1,
0.06
] | recomputed_from_pair_coverage_effective_joint_v1 |
idea_novelty | CI_0375 | Formalize and compare three distinct video summarization paradigms: extractive, query-focused, and entity-based summarization within a unified submodular optimization framework. | 201,901 | A4 | CT_task | TF_representation | [
"CI_0016",
"CI_0083"
] | 0 | not_covered | 0.7 | new_task_definition | 0.2 | 0.85 | high_novelty | 0.85 | llm_residual_delta_judge_v1 | 0.8 | 0 | 0.49 | new_task_definition | Formalizes and unifies three distinct summarization paradigms under a single optimization framework, creating a new problem formulation and task definition. | Qwen3.5-397B-A17B | 0.0895 | not_covered | [
"CI_0023",
"CI_0189",
"CI_0016",
"CI_0236",
"CI_0083"
] | [
"Reformulate the scene text detection task to eliminate the necessity of location regression by treating text/non-text prediction as a semantic segmentation problem enriched with instance-level linking.",
"Define a new generative task that synthesizes animated sequences with localized motion from a single static ... | [
"not_covering",
"not_covering"
] | [
0.1,
0.1
] | recomputed_from_pair_coverage_effective_joint_v1 |
idea_novelty | CI_0376 | Propose a visual structure constraint that aligns projected semantic centers with unsupervised visual cluster centers to learn a robust projection function and alleviate domain shift in transductive zero-shot learning. | 201,901 | A6 | CT_method | TF_representation | [
"CI_0183",
"CI_0036",
"CI_0281"
] | 0.04 | not_covered | 0.7 | new_training_signal | 0.2 | 0.9 | high_novelty | 0.9 | llm_residual_delta_judge_v1 | 0.8 | 0 | 0.64 | new_mechanism | Proposes a specific visual structure constraint aligning semantic and visual cluster centers, a distinct training signal for transductive zero-shot learning not covered by prior structural losses. | Qwen3.5-397B-A17B | 0.2351 | not_covered | [
"CI_0183",
"CI_0182",
"CI_0252",
"CI_0036",
"CI_0127"
] | [
"Introduce a novel structural loss function that incorporates location, contextual, and semantic information to supervise boundary learning and ensure spatially consistent saliency predictions.",
"Combine visual and semantic similarity measures to jointly exploit complementary knowledge for improving semi-supervi... | [
"related_not_covering",
"not_covering",
"not_covering"
] | [
0.2,
0.1,
0.1
] | recomputed_from_pair_coverage_effective_joint_v1 |
idea_novelty | CI_0377 | Adopt multiple distance metrics including symmetric Chamfer distance, bipartite matching, and Wasserstein distance to measure and align the discrepancy between semantic and visual cluster centers. | 201,901 | A6 | CT_method | TF_representation | [
"CI_0036",
"CI_0191"
] | 0 | not_covered | 0.6 | new_training_signal | 0.3 | 0.85 | high_novelty | 0.85 | llm_residual_delta_judge_v1 | 0.8 | 0 | 0.5 | new_mechanism | Introduces a composite distance metric strategy (Chamfer, bipartite, Wasserstein) for center alignment, representing a novel optimization objective formulation relative to standard similarity measures. | Qwen3.5-397B-A17B | 0.0895 | not_covered | [
"CI_0034",
"CI_0036",
"CI_0184",
"CI_0154",
"CI_0193"
] | [
"Combine visual and semantic similarity measures to jointly exploit complementary knowledge for improving semi-supervised object detection performance.",
"Enforce temporal consistency constraints between the action start window and its subsequent follow-up window during training to align their feature representat... | [
"not_covering",
"not_covering"
] | [
0.1,
0.1
] | recomputed_from_pair_coverage_effective_joint_v1 |
idea_novelty | CI_0378 | Introduce a new training strategy that filters out unrelated images from the test set before clustering to prevent invalid visual center estimation in realistic scenarios containing noise. | 201,901 | A7 | CT_method | TF_representation | [
"CI_0032",
"CI_0043"
] | 0 | not_covered | 0.7 | new_training_signal | 0.2 | 0.9 | high_novelty | 0.9 | llm_residual_delta_judge_v1 | 0.8 | 0 | 0.48 | new_training_signal | Defines a new pre-clustering filtering strategy to remove test-set noise, establishing a distinct training protocol to ensure valid visual center estimation in realistic scenarios. | Qwen3.5-397B-A17B | 0.0895 | not_covered | [
"CI_0054",
"CI_0027",
"CI_0089",
"CI_0029",
"CI_0078"
] | [
"Design a composite objective function incorporating additional loss terms alongside adversarial training to enforce structural consistency and improve the realism of generated images from sparse sketches.",
"Design a concurrent training mechanism where the context model and the auto-encoder are updated simultane... | [
"not_covering",
"not_covering"
] | [
0.1,
0.1
] | recomputed_from_pair_coverage_effective_joint_v1 |
idea_novelty | CI_0379 | Define a new transductive zero-shot learning configuration that explicitly accounts for the presence of unrelated images in the target domain, extending the problem setting beyond standard assumptions. | 201,901 | A4 | CT_task | TF_representation | [
"CI_0271",
"CI_0016",
"CI_0164"
] | 0.04 | not_covered | 0.7 | new_task_definition | 0.2 | 0.85 | high_novelty | 0.85 | llm_residual_delta_judge_v1 | 0.8 | 0 | 0.63 | new_task_definition | Defines a novel transductive ZSL setting explicitly handling unrelated target images, extending standard problem assumptions beyond existing disjoint category formulations. | Qwen3.5-397B-A17B | 0.1551 | not_covered | [
"CI_0271",
"CI_0023",
"CI_0016",
"CI_0204",
"CI_0189"
] | [
"Define a novel zero-shot sketch-based image retrieval hashing task where training and testing categories are disjoint to address the limitation of existing methods in handling unseen query concepts.",
"Reformulate the scene text detection task to eliminate the necessity of location regression by treating text/no... | [
"related_not_covering",
"not_covering",
"not_covering"
] | [
0.2,
0.1,
0.1
] | recomputed_from_pair_coverage_effective_joint_v1 |
idea_novelty | CI_0380 | Propose a parallel multi-branch architecture with shared parameters but different receptive fields to generate scale-specific feature maps with uniform representational power. | 201,901 | A6 | CT_method | TF_representation | [
"CI_0065",
"CI_0145",
"CI_0296"
] | 0.04 | not_covered | 0.5 | new_mechanism | 0.4 | 0.9 | moderate_novelty | 0.9 | llm_residual_delta_judge_v1 | 0.8 | 0 | 0.6 | new_mechanism | Proposes a specific parallel multi-branch architecture with shared parameters and varied receptive fields for scale-specific features, refining existing multi-branch concepts. | Qwen3.5-397B-A17B | 0.247 | not_covered | [
"CI_0065",
"CI_0064",
"CI_0248",
"CI_0145",
"CI_0250"
] | [
"Propose a multi-branch convolution mechanism in feature space to implicitly expand point features and reconstruct a denser point set, addressing the lack of regular structure in point clouds.",
"Introduce a network modulation process that injects generated parameters into intermediate layers for feature manipula... | [
"related_not_covering",
"not_covering",
"not_covering"
] | [
0.34,
0.1,
0.1
] | recomputed_from_pair_coverage_effective_joint_v1 |
idea_novelty | CI_0381 | Adopt a scale-aware training scheme that samples object instances of proper scales to specialize each branch according to its specific receptive field. | 201,901 | A7 | CT_method | TF_representation | [
"CI_0032",
"CI_0029"
] | 0 | not_covered | 0.6 | new_training_signal | 0.3 | 0.8 | moderate_novelty | 0.8 | llm_residual_delta_judge_v1 | 0.8 | 0 | 0.48 | new_training_signal | Introduces a scale-aware sampling strategy to specialize network branches, representing a distinct training protocol compared to standard augmentation or loss-based priors. | Qwen3.5-397B-A17B | 0.0895 | not_covered | [
"CI_0192",
"CI_0043",
"CI_0054",
"CI_0330",
"CI_0032"
] | [
"Design a composite objective function incorporating additional loss terms alongside adversarial training to enforce structural consistency and improve the realism of generated images from sparse sketches.",
"Propose a progressive data augmentation strategy that synthesizes edge maps from large-scale unlabeled im... | [
"not_covering",
"not_covering"
] | [
0.1,
0.1
] | recomputed_from_pair_coverage_effective_joint_v1 |
idea_novelty | CI_0382 | Design a fast approximation variant that utilizes only a single major branch during inference via weight sharing, achieving performance gains without additional computational cost or parameters. | 201,901 | A6 | CT_method | TF_representation | [
"CI_0135",
"CI_0268"
] | 0 | not_covered | 0.6 | new_mechanism | 0.2 | 0.85 | high_novelty | 0.85 | llm_residual_delta_judge_v1 | 0.8 | 0 | 0.5 | new_mechanism | Proposes a specific inference-time approximation via weight sharing and single-branch selection, a distinct efficiency mechanism not covered by priors focusing on generation or fusion. | Qwen3.5-397B-A17B | 0.0895 | not_covered | [
"CI_0046",
"CI_0224",
"CI_0231",
"CI_0115",
"CI_0250"
] | [
"Demonstrate the capability to generate high-resolution (256^3) dense 3D occupancy grids from single arbitrary depth views without requiring object class labels or multiple viewpoints.",
"Augment mask prediction with a complementary branch using fully-connected layers to capture diverse views and fuse predictions... | [
"not_covering",
"not_covering"
] | [
0.1,
0.1
] | recomputed_from_pair_coverage_effective_joint_v1 |
idea_novelty | CI_0383 | Propose a scale-invariant sub-network integrating orientation-aware modules to capture local geometric information from sparse point clouds across different scales. | 201,901 | A6 | CT_method | TF_representation | [
"CI_0207",
"CI_0154",
"CI_0065",
"CI_0167",
"CI_0059"
] | 0.1507 | not_covered | 0.5 | nontrivial_combination | 0.3 | 0.9 | moderate_novelty | 0.9 | llm_residual_delta_judge_v1 | 0.7 | 0 | 0.7 | new_mechanism | Combines scale-invariance and orientation-awareness into a unified sub-network; while components exist separately, their specific integration for sparse clouds represents a non-obvious architectural synthesis. | Qwen3.5-397B-A17B | 0.4994 | weakly_covered | [
"CI_0207",
"CI_0059",
"CI_0065",
"CI_0167",
"CI_0156"
] | [
"Develop a hierarchical network architecture using sparse lattice operations to learn spatially-aware features directly from unordered point clouds while maintaining computational efficiency via hash-based indexing.",
"Propose a local feature descriptor that integrates point-pair features with global context awar... | [
"related_not_covering",
"related_not_covering",
"related_not_covering",
"related_not_covering",
"not_covering"
] | [
0.44,
0.2,
0.2,
0.2,
0.19
] | recomputed_from_pair_coverage_effective_joint_v1 |
idea_novelty | CI_0384 | Design a feature reweighting mechanism that extends channel attention to 3D space to enhance informative features and suppress less useful ones for bounding box estimation. | 201,901 | A6 | CT_method | TF_representation | [
"CI_0257",
"CI_0052",
"CI_0199"
] | 0 | not_covered | 0.4 | minor_variant | 0.4 | 0.95 | moderate_novelty | 0.95 | llm_residual_delta_judge_v1 | 0.7 | 0 | 0.68 | new_mechanism | Adapts standard channel attention to 3D space; the core mechanism is well-known, with novelty limited to the specific 3D geometric application and bounding box task context. | Qwen3.5-397B-A17B | 0.3126 | weakly_covered | [
"CI_0257",
"CI_0065",
"CI_0052",
"CI_0310",
"CI_0006"
] | [
"Introduce a two-stage convolutional neural network that refines initial 3D bounding box estimates by regressing final dimensions and classification based on points within the proposed box.",
"Design a coarse-to-fine layout inference mechanism that sequentially predicts object bounding boxes and then refines obje... | [
"not_covering",
"not_covering",
"not_covering"
] | [
0.19,
0.1,
0.1
] | recomputed_from_pair_coverage_effective_joint_v1 |
idea_novelty | CI_0385 | Develop a unified architecture fusing front-view image appearance features with frustum point clouds to improve localization accuracy in sparse 3D scenes. | 201,901 | A6 | CT_method | TF_representation | [
"CI_0208",
"CI_0059",
"CI_0024",
"CI_0264"
] | 0.4048 | weakly_covered | 0.5 | nontrivial_combination | 0.3 | 0.8 | weak_novelty | 0.8 | llm_residual_delta_judge_v1 | 0.7 | 0 | 0.68 | new_mechanism | Combines established 2D-3D fusion with specific frustum point cloud representation for sparse localization. Core mechanism exists, but specific integration for this setting is a non-trivial adaptation. | Qwen3.5-397B-A17B | 0.4514 | weakly_covered | [
"CI_0059",
"CI_0208",
"CI_0154",
"CI_0262",
"CI_0209"
] | [
"Design a unified framework that jointly processes 2D images and 3D point clouds in a single forward pass by leveraging flexible lattice structures to map between 2D and 3D representations.",
"Integrate reconstructed 3D edges into a multi-view stereo pipeline to enhance the accuracy and visual quality of surface ... | [
"partial_cover",
"related_not_covering",
"not_covering",
"not_covering"
] | [
0.45,
0.44,
0.1,
0.1
] | recomputed_from_pair_coverage_effective_joint_v1 |
idea_novelty | CI_0386 | Propose a discriminative modeling framework that maximizes the probability margin between correct and negative action order transcripts to prevent degenerate single-frame alignments in weakly supervised settings. | 201,901 | A6 | CT_method | TF_representation | [
"CI_0240",
"CI_0087"
] | 0 | not_covered | 0.7 | new_training_signal | 0.2 | 0.9 | high_novelty | 0.9 | llm_residual_delta_judge_v1 | 0.8 | 0 | 0.5 | new_mechanism | Introduces a specific discriminative margin-maximization objective to solve degenerate alignment in weakly supervised settings, representing a distinct new training signal not covered by priors. | Qwen3.5-397B-A17B | 0.0895 | not_covered | [
"CI_0100",
"CI_0191",
"CI_0042",
"CI_0332",
"CI_0196"
] | [
"Design a multi-evidence fusion strategy that integrates image-level attention, object-level attention, and detection heat maps to generate clean pixel-wise probability maps for segmentation training.",
"Propose a nonlinear multistage hierarchical segmentation strategy that processes depth and color sequentially ... | [
"not_covering",
"not_covering"
] | [
0.1,
0.1
] | recomputed_from_pair_coverage_effective_joint_v1 |
idea_novelty | CI_0387 | Develop a differentiable structural prediction mechanism by applying continuous relaxation to the min-operator in dynamic programming, enabling end-to-end training of discriminative losses for sequence alignment. | 201,901 | A6 | CT_method | TF_representation | [
"CI_0127",
"CI_0237"
] | 0 | not_covered | 0.8 | new_mechanism | 0.1 | 0.9 | high_novelty | 0.9 | llm_residual_delta_judge_v1 | 0.8 | 0 | 0.5 | new_mechanism | Proposes a novel differentiable relaxation of the min-operator in dynamic programming, enabling end-to-end training for structural prediction. This is a new algorithmic mechanism. | Qwen3.5-397B-A17B | 0.0895 | not_covered | [
"CI_0326",
"CI_0127",
"CI_0137",
"CI_0328",
"CI_0184"
] | [
"Develop a structured likelihood maximization solver to generate synthetic images with realistic cluster layouts by learning structural priors from real data.",
"Develop an end-to-end trainable architecture that integrates object detection and semantic segmentation paradigms to eliminate the need for complex post... | [
"not_covering",
"not_covering"
] | [
0.1,
0.1
] | recomputed_from_pair_coverage_effective_joint_v1 |
idea_novelty | CI_0388 | Formulate weakly supervised action alignment as a discriminative task that compares the alignment costs of positive ground-truth transcripts against negative samples, rather than maximizing posterior probability alone. | 201,901 | A4 | CT_task | TF_representation | [
"CI_0023",
"CI_0083"
] | 0 | not_covered | 0.7 | new_training_signal | 0.2 | 0.85 | high_novelty | 0.85 | llm_residual_delta_judge_v1 | 0.8 | 0 | 0.49 | new_task_definition | Introduces a discriminative contrastive objective comparing positive/negative alignment costs, distinct from standard posterior maximization or generative video tasks in priors. | Qwen3.5-397B-A17B | 0.0895 | not_covered | [
"CI_0228",
"CI_0189",
"CI_0023",
"CI_0204",
"CI_0016"
] | [
"Extend the object referring task from static images to dynamic video sequences by incorporating motion features and spatio-temporal context as essential cues for disambiguation.",
"Define a new generative task that synthesizes animated sequences with localized motion from a single static input image."
] | [
"not_covering",
"not_covering"
] | [
0.1,
0.1
] | recomputed_from_pair_coverage_effective_joint_v1 |
idea_novelty | CI_0389 | Propose a shape regression mechanism that predicts dense text boundary points instead of quadrilateral vertices to accurately locate curved and long text lines while minimizing regression errors. | 201,901 | A6 | CT_method | TF_representation | [
"CI_0014",
"CI_0052",
"CI_0234",
"CI_0236",
"CI_0245"
] | 0.1507 | not_covered | 0.6 | new_mechanism | 0.3 | 0.9 | moderate_novelty | 0.9 | llm_residual_delta_judge_v1 | 0.7 | 0 | 0.64 | new_mechanism | Shifts text detection from corner grouping or segmentation to direct dense boundary point regression, offering a distinct mechanism for curved text localization. | Qwen3.5-397B-A17B | 0.3914 | weakly_covered | [
"CI_0052",
"CI_0234",
"CI_0014",
"CI_0236",
"CI_0137"
] | [
"Propose an instance segmentation framework that replaces bounding box regression by linking pixels within the same text instance to directly extract detection boxes.",
"Design a coarse-to-fine layout inference mechanism that sequentially predicts object bounding boxes and then refines object shapes within those ... | [
"related_not_covering",
"related_not_covering",
"related_not_covering",
"related_not_covering",
"not_covering"
] | [
0.2,
0.2,
0.2,
0.2,
0.1
] | recomputed_from_pair_coverage_effective_joint_v1 |
idea_novelty | CI_0390 | Design a multi-scale network architecture employing multiple channels to extract and fuse features at different scales, enhancing tolerance to large variations in text size. | 201,901 | A6 | CT_method | TF_representation | [
"CI_0296",
"CI_0046",
"CI_0299"
] | 0.04 | not_covered | 0.3 | minor_variant | 0.6 | 0.95 | weak_novelty | 0.95 | llm_residual_delta_judge_v1 | 0.8 | 0 | 0.6 | new_mechanism | Applies generic multi-scale feature fusion architecture to text detection; priors show similar multi-scale designs in other domains, indicating an incremental domain adaptation. | Qwen3.5-397B-A17B | 0.247 | not_covered | [
"CI_0296",
"CI_0066",
"CI_0030",
"CI_0046",
"CI_0299"
] | [
"Develop a multi-stream densely connected network architecture that leverages features from different scales to effectively remove rain streaks with diverse shapes and sizes.",
"Design a recurrent architecture that assimilates information from a large number of previous frames without increasing computational dem... | [
"related_not_covering",
"not_covering",
"not_covering"
] | [
0.34,
0.1,
0.1
] | recomputed_from_pair_coverage_effective_joint_v1 |
idea_novelty | CI_0391 | Define a new task of continuous novel view synthesis from single images with fine-grained 6-DOF camera control, addressing the limitation of discrete view generation in existing methods. | 201,901 | A4 | CT_task | TF_representation | [
"CI_0189",
"CI_0016"
] | 0 | not_covered | 0.7 | new_task_definition | 0.2 | 0.9 | high_novelty | 0.9 | llm_residual_delta_judge_v1 | 0.8 | 0 | 0.49 | new_task_definition | Defines a distinct problem setting (continuous 6-DOF control from single image) not covered by priors focusing on action detection or text segmentation reformulation. | Qwen3.5-397B-A17B | 0.0895 | not_covered | [
"CI_0280",
"CI_0083",
"CI_0164",
"CI_0329",
"CI_0023"
] | [
"Define a novel task for online detection of action start points in untrimmed streaming videos, focusing on immediate identification of action onset without future frame access.",
"Reformulate the scene text detection task to eliminate the necessity of location regression by treating text/non-text prediction as a... | [
"not_covering",
"not_covering"
] | [
0.1,
0.1
] | recomputed_from_pair_coverage_effective_joint_v1 |
idea_novelty | CI_0392 | Propose a geometry-aware architecture that explicitly applies 3D rotations and translations to the latent code within an auto-encoder, forcing the network to learn an implicit 3D geometric representation rather than relying on direct pixel synthesis. | 201,901 | A6 | CT_application | TF_application | [
"CI_0166",
"CI_0132"
] | 0 | not_covered | 0.7 | new_mechanism | 0.2 | 0.85 | high_novelty | 0.85 | llm_residual_delta_judge_v1 | 0.8 | 0 | 0.5 | new_mechanism | Introduces a specific geometric mechanism (explicit 3D rotation/translation in latent space) distinct from priors on point cloud pooling or modular backbones. | Qwen3.5-397B-A17B | 0.0895 | not_covered | [
"CI_0144",
"CI_0230",
"CI_0149",
"CI_0166",
"CI_0323"
] | [
"Propose a slice pooling layer that projects unordered point features onto an ordered sequence to enable the application of standard sequence modeling algorithms on point clouds.",
"Design a modular architecture compatible with any existing CNN-based semantic segmentation backbone, allowing the diverse supervisio... | [
"not_covering",
"not_covering"
] | [
0.1,
0.1
] | recomputed_from_pair_coverage_effective_joint_v1 |
idea_novelty | CI_0393 | Integrate a depth-guided warping mechanism that uses the predicted depth map from the transformed latent code to compute dense pixel correspondences and warp source pixels, ensuring geometric consistency in synthesized views. | 201,901 | A6 | CT_method | TF_representation | [
"CI_0306",
"CI_0335",
"CI_0128",
"CI_0063"
] | 0.0784 | not_covered | 0.6 | new_mechanism | 0.3 | 0.9 | high_novelty | 0.9 | llm_residual_delta_judge_v1 | 0.7 | 0 | 0.68 | new_mechanism | Adapts warping concepts into a specific depth-guided dense correspondence mechanism for view synthesis, differing from prior STN-based compositing or stereo depth estimation. | Qwen3.5-397B-A17B | 0.4106 | weakly_covered | [
"CI_0306",
"CI_0335",
"CI_0135",
"CI_0128",
"CI_0055"
] | [
"Propose a GAN framework that utilizes Spatial Transformer Networks as the generator to learn realistic geometric warps for image compositing instead of generating pixels directly.",
"Utilize both spatial (left-right) and temporal (forward-backward) photometric warp errors from stereo video sequences to constrain... | [
"related_not_covering",
"related_not_covering",
"not_covering",
"not_covering"
] | [
0.44,
0.34,
0.1,
0.1
] | recomputed_from_pair_coverage_effective_joint_v1 |
idea_novelty | CI_0394 | Employ a self-supervised training strategy that leverages geometric constraints and view transformations as supervision signals, eliminating the need for explicit depth maps or optical flow ground truth during training. | 201,901 | A7 | CT_method | TF_representation | [
"CI_0027",
"CI_0043"
] | 0 | not_covered | 0.7 | new_training_signal | 0.2 | 0.85 | high_novelty | 0.85 | llm_residual_delta_judge_v1 | 0.8 | 0 | 0.48 | new_training_signal | Introduces geometric constraints and view transformations as explicit self-supervised signals, a distinct training objective compared to prior pseudo-label or concurrent update strategies. | Qwen3.5-397B-A17B | 0.0895 | not_covered | [
"CI_0151",
"CI_0307",
"CI_0130",
"CI_0041",
"CI_0027"
] | [
"Adopt a weakly supervised training strategy using pseudo-labels generated by a traditional object detection pipeline to avoid the cost of manual annotation.",
"Design a concurrent training mechanism where the context model and the auto-encoder are updated simultaneously, allowing the context model to learn symbo... | [
"not_covering",
"not_covering"
] | [
0.1,
0.1
] | recomputed_from_pair_coverage_effective_joint_v1 |
idea_novelty | CI_0395 | Propose a self-supervised correction mechanism with dual 2D-to-3D and 3D-to-2D learning tasks to enforce geometric consistency without requiring extra 3D annotations. | 201,901 | A7 | CT_task | TF_representation | [
"CI_0029",
"CI_0054"
] | 0 | not_covered | 0.7 | new_training_signal | 0.2 | 0.85 | high_novelty | 0.85 | llm_residual_delta_judge_v1 | 0.8 | 0 | 0.48 | new_training_signal | Proposes a novel dual 2D-to-3D and 3D-to-2D consistency loop as a self-supervised correction mechanism, differing significantly from prior augmentation or pixel-shuffling strategies. | Qwen3.5-397B-A17B | 0 | not_covered | [
"CI_0194",
"CI_0029",
"CI_0286",
"CI_0032",
"CI_0078"
] | [
"Propose a progressive data augmentation strategy that synthesizes edge maps from large-scale unlabeled images to pre-train the model, enabling effective knowledge transfer to the scarce human-sketch domain.",
"Propose a training strategy that randomly shuffles input pixels to force the network to learn global st... | [
"not_covering",
"not_covering"
] | [
0.07,
0.07
] | recomputed_from_pair_coverage_effective_joint_v1 |
idea_novelty | CI_0396 | Integrate spatial relationships, temporal smoothness, and 3D geometric constraints into a unified sequential framework using CNNs and RNNs to capture long-range dependencies. | 201,901 | A6 | CT_method | TF_representation | [
"CI_0167",
"CI_0306",
"CI_0335",
"CI_0006"
] | 0.0784 | not_covered | 0.5 | nontrivial_combination | 0.4 | 0.9 | moderate_novelty | 0.9 | llm_residual_delta_judge_v1 | 0.7 | 0 | 0.7 | new_mechanism | Integrates spatial, temporal, and 3D geometric constraints into a unified CNN-RNN framework; while components exist separately, their joint formulation for long-range dependency modeling represents a nontrivial synthesis. | Qwen3.5-397B-A17B | 0.3624 | weakly_covered | [
"CI_0306",
"CI_0167",
"CI_0152",
"CI_0059",
"CI_0334"
] | [
"Design a local dependency module that combines slice pooling, recurrent neural networks, and slice unpooling to efficiently model local geometric structures in unordered point sets.",
"Propose a GAN framework that utilizes Spatial Transformer Networks as the generator to learn realistic geometric warps for image... | [
"related_not_covering",
"related_not_covering",
"not_covering",
"not_covering"
] | [
0.34,
0.34,
0.1,
0.1
] | recomputed_from_pair_coverage_effective_joint_v1 |
idea_novelty | CI_0397 | Leverage abundant 2D pose data through a self-supervised projection loss to refine 3D predictions and overcome the scarcity of 3D training annotations. | 201,901 | A7 | CT_method | TF_representation | [
"CI_0227",
"CI_0029",
"CI_0043"
] | 0.72 | partially_covered | 0.3 | minor_variant | 0.6 | 0.9 | low_novelty | 0.9 | llm_residual_delta_judge_v1 | 0.7 | 0 | 0.84 | new_training_signal | Prior already covers training 3D pose with 2D data via shared features. Target adds a specific projection loss, which is an incremental refinement of the established self-supervised strategy. | Qwen3.5-397B-A17B | 0.4569 | weakly_covered | [
"CI_0227",
"CI_0130",
"CI_0032",
"CI_0183",
"CI_0029"
] | [
"Enable simultaneous training on heterogeneous datasets to generalize 3D pose predictions using only 2D annotated data through shared visual feature learning.",
"Propose a progressive data augmentation strategy that synthesizes edge maps from large-scale unlabeled images to pre-train the model, enabling effective... | [
"large_cover",
"not_covering",
"not_covering"
] | [
0.72,
0.1,
0.1
] | recomputed_from_pair_coverage_effective_joint_v1 |
idea_novelty | CI_0398 | Propose a compositional network architecture that performs iterative color dequantization to recover high-quality colors from heavily quantized GIF frames. | 201,901 | A6 | CT_method | TF_representation | [
"CI_0184",
"CI_0070"
] | 0 | not_covered | 0.7 | new_mechanism | 0.2 | 0.85 | high_novelty | 0.85 | llm_residual_delta_judge_v1 | 0.8 | 0 | 0.5 | new_mechanism | Priors do not cover color dequantization or iterative recovery from GIF quantization. The proposed compositional architecture and iterative mechanism represent a substantial new approach for this specific restoration task. | Qwen3.5-397B-A17B | 0.0895 | not_covered | [
"CI_0064",
"CI_0087",
"CI_0248",
"CI_0157",
"CI_0022"
] | [
"Develop a unified end-to-end framework that jointly performs complementary feature extraction and saliency region classification by fusing multi-level features from symmetric branches.",
"Propose an end-to-end neural network architecture capable of directly reading and extracting unique textual answers from visu... | [
"not_covering",
"not_covering"
] | [
0.1,
0.1
] | recomputed_from_pair_coverage_effective_joint_v1 |
idea_novelty | CI_0399 | Design a comprehensive loss function combining reconstruction and adversarial losses on both pixel colors and image gradients to handle large quantization errors. | 201,901 | A7 | CT_method | TF_representation | [
"CI_0078",
"CI_0221",
"CI_0029",
"CI_0043"
] | 0.4048 | weakly_covered | 0.5 | new_training_signal | 0.4 | 0.9 | weak_novelty | 0.9 | llm_residual_delta_judge_v1 | 0.7 | 0 | 0.66 | new_training_signal | While priors cover combining adversarial and reconstruction losses, the specific inclusion of image gradient losses to address large quantization errors constitutes a meaningful new training signal formulation. | Qwen3.5-397B-A17B | 0.4514 | weakly_covered | [
"CI_0221",
"CI_0078",
"CI_0032",
"CI_0307",
"CI_0067"
] | [
"Integrate global and local adversarial discriminators combined with spatially discounted reconstruction loss to enhance training stability and enforce coherence between generated and existing regions.",
"Design a multi-purpose adversarial loss that jointly differentiates real-and-fake image-text pairs and local ... | [
"partial_cover",
"related_not_covering",
"not_covering",
"not_covering"
] | [
0.45,
0.44,
0.1,
0.1
] | recomputed_from_pair_coverage_effective_joint_v1 |
idea_novelty | CI_0400 | Adapt a temporal interpolation network to increase the frame rate of GIF sequences, addressing motion jerkiness caused by aggressive frame sampling. | 201,901 | A6 | CT_application | TF_application | [
"CI_0166",
"CI_0132"
] | 0 | not_covered | 0.3 | minor_variant | 0.7 | 0.9 | weak_novelty | 0.9 | llm_residual_delta_judge_v1 | 0.8 | 0 | 0.5 | new_mechanism | Applying standard temporal interpolation to GIFs is a domain-specific adaptation of existing video frame interpolation techniques, lacking core algorithmic novelty. | Qwen3.5-397B-A17B | 0.0895 | not_covered | [
"CI_0230",
"CI_0144",
"CI_0323",
"CI_0149",
"CI_0285"
] | [
"Propose a slice pooling layer that projects unordered point features onto an ordered sequence to enable the application of standard sequence modeling algorithms on point clouds.",
"Design a modular architecture compatible with any existing CNN-based semantic segmentation backbone, allowing the diverse supervisio... | [
"not_covering",
"not_covering"
] | [
0.1,
0.1
] | recomputed_from_pair_coverage_effective_joint_v1 |
idea_novelty | CI_0401 | Introduce two large-scale paired datasets containing animated GIFs and their corresponding high-quality source videos to facilitate training and evaluation of GIF restoration models. | 201,901 | A8 | CT_system | TF_representation | [
"CI_0085",
"CI_0075",
"CI_0079"
] | 0.04 | not_covered | 0.5 | new_data_distribution | 0.4 | 0.95 | moderate_novelty | 0.95 | llm_residual_delta_judge_v1 | 0.74 | 0 | 0.53 | new_evaluation_protocol | Creating large-scale paired GIF-video datasets introduces a new data distribution and resource specifically for GIF restoration, distinct from prior synthetic or unrelated animation datasets. | Qwen3.5-397B-A17B | 0.247 | not_covered | [
"CI_0085",
"CI_0121",
"CI_0200",
"CI_0333",
"CI_0075"
] | [
"Construct and release two dedicated datasets, comprising both synthetically generated samples and real-world collected videos, to enable training and evaluation of the proposed animation task.",
"Construct a new dataset comprising three representative scanned scenes with 3D models of common objects and annotated... | [
"related_not_covering",
"not_covering",
"not_covering"
] | [
0.34,
0.1,
0.07
] | recomputed_from_pair_coverage_effective_joint_v1 |
idea_novelty | CI_0402 | Design a unified network with shared backbone and parallel heads for simultaneous semantic and instance segmentation to leverage task complementariness. | 201,901 | A6 | CT_task | TF_representation | [
"CI_0318",
"CI_0315",
"CI_0278",
"CI_0013",
"CI_0132"
] | 0.8464 | mostly_covered | 0.3 | minor_variant | 0.6 | 0.95 | low_novelty | 0.95 | llm_residual_delta_judge_v1 | 0.5 | 0 | 0.7 | minor_mechanism_change | Unified multi-task networks with shared backbones are well-established; applying this to simultaneous semantic and instance segmentation is an incremental architectural adaptation. | Qwen3.5-397B-A17B | 0.5873 | partially_covered | [
"CI_0318",
"CI_0278",
"CI_0315",
"CI_0013",
"CI_0132"
] | [
"Propose a shared convolutional architecture that simultaneously performs object detection, instance-level depth estimation, and semantic segmentation to improve computational efficiency and accuracy over single-task combinations.",
"Design a unified multi-task learning framework that learns semantic corresponden... | [
"large_cover",
"partial_cover",
"related_not_covering",
"related_not_covering",
"related_not_covering"
] | [
0.72,
0.45,
0.44,
0.34,
0.2
] | recomputed_from_pair_coverage_effective_joint_v1 |
idea_novelty | CI_0403 | Propose a parameter-free panoptic head that resolves conflicts between semantic and instance predictions via pixel-wise classification with an expanded unknown class channel. | 201,901 | A6 | CT_method | TF_representation | [
"CI_0160",
"CI_0024"
] | 0 | not_covered | 0.7 | new_mechanism | 0.2 | 0.85 | high_novelty | 0.85 | llm_residual_delta_judge_v1 | 0.8 | 0 | 0.5 | new_mechanism | Introduces a specific parameter-free conflict resolution mechanism via an expanded unknown class channel, distinct from prior semi-supervised or gaze-based approaches. | Qwen3.5-397B-A17B | 0.0895 | not_covered | [
"CI_0014",
"CI_0022",
"CI_0223",
"CI_0126",
"CI_0018"
] | [
"Propose a semi-supervised framework integrating a spectral-spatial generative adversarial network with a conditional random field to jointly learn discriminative features from limited labeled data and refine predictions using abundant unlabeled data.",
"Integrate human gaze estimation as an explicit auxiliary mo... | [
"not_covering",
"not_covering"
] | [
0.1,
0.1
] | recomputed_from_pair_coverage_effective_joint_v1 |
idea_novelty | CI_0404 | Enable end-to-end training for panoptic segmentation by designing a lightweight head that handles varying instance counts and permits backpropagation to bottom modules. | 201,901 | A6 | CT_method | TF_representation | [
"CI_0007",
"CI_0262"
] | 0 | not_covered | 0.6 | new_mechanism | 0.3 | 0.8 | moderate_novelty | 0.8 | llm_residual_delta_judge_v1 | 0.8 | 0 | 0.5 | new_mechanism | Proposes a specific lightweight head design enabling end-to-end backpropagation for variable instance counts, advancing beyond hierarchical integration or joint pose generation. | Qwen3.5-397B-A17B | 0.0895 | not_covered | [
"CI_0237",
"CI_0115",
"CI_0007",
"CI_0018",
"CI_0126"
] | [
"Introduce a hierarchical integration strategy that gradually combines nearby feature levels instead of directly fusing all levels to ease optimization and handle significant feature discrepancies.",
"Propose an end-to-end architecture that jointly generates, scores, and refines 2D and 3D pose proposals without r... | [
"not_covering",
"not_covering"
] | [
0.1,
0.1
] | recomputed_from_pair_coverage_effective_joint_v1 |
idea_novelty | CI_0405 | Utilize a single shared backbone network to generate common feature representations for both semantic and instance segmentation subtasks. | 201,901 | A6 | CT_task | TF_representation | [
"CI_0318",
"CI_0013",
"CI_0126",
"CI_0191"
] | 0.4048 | weakly_covered | 0.4 | minor_variant | 0.5 | 0.9 | weak_novelty | 0.9 | llm_residual_delta_judge_v1 | 0.7 | 0 | 0.7 | new_mechanism | Applying shared backbones to semantic and instance segmentation is a direct adaptation of existing multi-task shared architecture principles, offering limited conceptual novelty. | Qwen3.5-397B-A17B | 0.4106 | weakly_covered | [
"CI_0318",
"CI_0013",
"CI_0315",
"CI_0051",
"CI_0047"
] | [
"Propose a shared convolutional architecture that simultaneously performs object detection, instance-level depth estimation, and semantic segmentation to improve computational efficiency and accuracy over single-task combinations.",
"Propose a baseline heuristic that merges outputs from independent semantic and i... | [
"partial_cover",
"related_not_covering",
"not_covering",
"not_covering"
] | [
0.45,
0.34,
0.07,
0.07
] | recomputed_from_pair_coverage_effective_joint_v1 |
idea_novelty | CI_0406 | Propose a fully recurrent two-stream architecture that integrates optical flow as an attention mask within Convolutional LSTMs to model spatiotemporal pain cues in video. | 201,901 | A6 | CT_method | TF_representation | [
"CI_0022",
"CI_0246",
"CI_0182"
] | 0.04 | not_covered | 0.6 | nontrivial_combination | 0.3 | 0.85 | high_novelty | 0.85 | llm_residual_delta_judge_v1 | 0.8 | 0 | 0.6 | new_mechanism | Integrating optical flow as an attention mask within ConvLSTMs for pain cues combines known mechanisms in a specific, non-trivial architectural configuration not present in priors. | Qwen3.5-397B-A17B | 0.247 | not_covered | [
"CI_0022",
"CI_0217",
"CI_0076",
"CI_0111",
"CI_0055"
] | [
"Propose a recurrent convolutional network architecture that fuses appearance, motion, depth, and human gaze cues with language expressions to localize objects within spatio-temporal video contexts.",
"Develop a hierarchical recurrent architecture that conditions sentence generation on both visual features and li... | [
"related_not_covering",
"not_covering",
"not_covering"
] | [
0.34,
0.1,
0.1
] | recomputed_from_pair_coverage_effective_joint_v1 |
idea_novelty | CI_0407 | Introduce the novel application field of automatic equine pain detection in video, shifting focus from static facial expressions to dynamic pose and movement patterns. | 201,901 | A1 | CT_application | TF_application | [
"CI_0118",
"CI_0181"
] | 0 | not_covered | 0.7 | new_application_constraint | 0.4 | 0.9 | high_novelty | 0.9 | llm_residual_delta_judge_v1 | 0.72 | 0 | 0.44 | new_application_constraint | Shifts the domain to equine pain and fundamentally changes the visual constraint from static faces to dynamic pose/movement, defining a new application setting with distinct requirements. | Qwen3.5-397B-A17B | 0.0895 | not_covered | [
"CI_0118",
"CI_0242",
"CI_0277",
"CI_0181",
"CI_0233"
] | [
"Propose a unified framework that performs both spatial selection of normal field-of-view segments and temporal summarization to detect highlights in 360-degree videos.",
"Enable unsupervised training on arbitrary uncalibrated monocular video streams by relying only on temporal coherence and ego-motion assumption... | [
"not_covering",
"not_covering"
] | [
0.1,
0.1
] | recomputed_from_pair_coverage_effective_joint_v1 |
idea_novelty | CI_0408 | Formulate the hypothesis that temporal dynamics and movement patterns are more critical than static frames for recognizing pain in prey animals like horses. | 201,901 | A2 | CT_task | TF_representation | [
"CI_0039",
"CI_0009"
] | 0 | not_covered | 0.6 | new_theory_or_concept | 0.2 | 0.8 | high_novelty | 0.8 | llm_residual_delta_judge_v1 | 0.8 | 0 | 0.46 | new_theory_or_concept | Proposes a specific scientific hypothesis prioritizing temporal dynamics over static frames for prey animal pain, offering a new conceptual framing for the problem rather than just a method. | Qwen3.5-397B-A17B | 0.0895 | not_covered | [
"CI_0039",
"CI_0186",
"CI_0096",
"CI_0056",
"CI_0185"
] | [
"Formulate a probabilistic framework that theoretically explains how supervising intermediate concepts improves generalization and generalizes conventional multi-task and deep supervision schemes.",
"Provide theoretical proof demonstrating that the proposed sampling mechanism is equivalent to standard inner produ... | [
"not_covering",
"not_covering"
] | [
0.1,
0.07
] | recomputed_from_pair_coverage_effective_joint_v1 |
idea_novelty | CI_0409 | Develop an end-to-end learning approach that directly infers pain levels from raw video without relying on intermediate manual annotations like facial action coding systems. | 201,901 | A6 | CT_method | TF_representation | [
"CI_0179",
"CI_0157"
] | 0 | not_covered | 0.7 | new_mechanism | 0.2 | 0.8 | high_novelty | 0.8 | llm_residual_delta_judge_v1 | 0.8 | 0 | 0.5 | new_mechanism | Proposes a distinct end-to-end mechanism bypassing intermediate facial action codes for pain estimation, unlike priors focused on 3D consistency or graph construction. | Qwen3.5-397B-A17B | 0.0895 | not_covered | [
"CI_0112",
"CI_0100",
"CI_0156",
"CI_0210",
"CI_0262"
] | [
"Propose a differentiable loss function that directly enforces consistency between 3D point clouds across consecutive frames by applying estimated ego-motion transformations, rather than relying solely on 2D pixel reconstruction errors.",
"Propose an iterative graph construction process guided by a CNN decision f... | [
"not_covering",
"not_covering"
] | [
0.1,
0.1
] | recomputed_from_pair_coverage_effective_joint_v1 |
idea_novelty | CI_0410 | Propose a region feature alignment module that adaptively shifts and aligns feature maps from different modalities to handle position shifts in weakly aligned multispectral data. | 201,901 | A6 | CT_method | TF_representation | [
"CI_0250",
"CI_0218"
] | 0 | not_covered | 0.6 | new_mechanism | 0.3 | 0.7 | moderate_novelty | 0.7 | llm_residual_delta_judge_v1 | 0.8 | 0 | 0.5 | new_mechanism | Introduces a specific region alignment module for positional shifts in multispectral data, differing from priors on temporal reuse or internal feature stabilization. | Qwen3.5-397B-A17B | 0.0895 | not_covered | [
"CI_0184",
"CI_0094",
"CI_0250",
"CI_0095",
"CI_0199"
] | [
"Introduce a video inference approach that computes precise feature maps for every frame by reusing previous computations, avoiding the performance drop associated with approximate feature propagation methods.",
"Introduce a local feature fusion mechanism within dense blocks to adaptively preserve accumulated fea... | [
"not_covering",
"not_covering"
] | [
0.1,
0.1
] | recomputed_from_pair_coverage_effective_joint_v1 |
idea_novelty | CI_0411 | Introduce a confidence-aware fusion mechanism that performs adaptive feature re-weighting to select reliable features and suppress confusing ones across modalities. | 201,901 | A6 | CT_method | TF_representation | [
"CI_0218",
"CI_0065",
"CI_0115"
] | 0.04 | not_covered | 0.5 | minor_variant | 0.4 | 0.8 | moderate_novelty | 0.8 | llm_residual_delta_judge_v1 | 0.8 | 0 | 0.6 | new_mechanism | Adaptive confidence-based fusion is an incremental refinement of existing adaptive fusion mechanisms, applying similar logic to multi-modal reliability rather than network stability. | Qwen3.5-397B-A17B | 0.247 | not_covered | [
"CI_0218",
"CI_0219",
"CI_0303",
"CI_0184",
"CI_0006"
] | [
"Introduce a local feature fusion mechanism within dense blocks to adaptively preserve accumulated features and stabilize the training of wider networks with high growth rates.",
"Propose a multi-branch convolution mechanism in feature space to implicitly expand point features and reconstruct a denser point set, ... | [
"related_not_covering",
"not_covering",
"not_covering"
] | [
0.34,
0.1,
0.1
] | recomputed_from_pair_coverage_effective_joint_v1 |
idea_novelty | CI_0412 | Develop a region-of-interest jitter training strategy that randomly perturbs sampling regions to improve model robustness against unexpected position shift patterns. | 201,901 | A7 | CT_method | TF_representation | [
"CI_0043",
"CI_0041"
] | 0 | not_covered | 0.7 | new_training_signal | 0.2 | 0.9 | high_novelty | 0.9 | llm_residual_delta_judge_v1 | 0.8 | 0 | 0.48 | new_training_signal | Introduces a specific region-of-interest jitter strategy for robustness, distinct from prior concurrent training or coarse-label supervision mechanisms. | Qwen3.5-397B-A17B | 0.0895 | not_covered | [
"CI_0054",
"CI_0192",
"CI_0081",
"CI_0043",
"CI_0078"
] | [
"Design a concurrent training mechanism where the context model and the auto-encoder are updated simultaneously, allowing the context model to learn symbol dependencies while providing entropy estimates for gradient-based optimization.",
"Utilize coarse-grained class labels as intermediate supervisory signals wit... | [
"not_covering",
"not_covering"
] | [
0.1,
0.1
] | recomputed_from_pair_coverage_effective_joint_v1 |
idea_novelty | CI_0413 | Construct a paired annotation dataset by manually relabeling bounding boxes for both color and thermal modalities to provide precise supervision for weakly aligned multispectral detection. | 201,901 | A5 | CT_system | TF_representation | [
"CI_0200",
"CI_0174",
"CI_0123"
] | 0.04 | not_covered | 0.6 | new_data_organization | 0.4 | 0.85 | moderate_novelty | 0.85 | llm_residual_delta_judge_v1 | 0.62 | 0 | 0.7 | new_data_organization | Constructs a novel paired annotation schema for color-thermal bounding boxes, extending prior multispectral dataset concepts to a specific weakly-aligned detection setting. | Qwen3.5-397B-A17B | 0.2976 | weakly_covered | [
"CI_0200",
"CI_0021",
"CI_0216",
"CI_0174",
"CI_0033"
] | [
"Construct a ground-truth dataset containing shading and reflectance annotations for multispectral images to enable quantitative evaluation of intrinsic decomposition algorithms.",
"Utilize object proposals from detection networks as input representations instead of raw pixels to facilitate counting tasks in comp... | [
"related_not_covering",
"not_covering",
"not_covering"
] | [
0.34,
0.07,
0.07
] | recomputed_from_pair_coverage_effective_joint_v1 |
idea_novelty | CI_0414 | Propose a joint modeling framework that combines coarse 3D bounding box and fine-scaled morphable wireframe hypotheses to improve robustness against 2D landmark noise in monocular 3D vehicle detection. | 201,901 | A6 | CT_method | TF_representation | [
"CI_0255",
"CI_0154",
"CI_0239"
] | 0.04 | not_covered | 0.7 | new_mechanism | 0.2 | 0.9 | high_novelty | 0.9 | llm_residual_delta_judge_v1 | 0.8 | 0 | 0.7 | new_mechanism | Proposes a unique joint modeling framework fusing coarse 3D boxes with fine wireframes, a specific hypothesis combination not covered by prior 2D-3D fusion or feature descriptor works. | Qwen3.5-397B-A17B | 0.247 | not_covered | [
"CI_0255",
"CI_0257",
"CI_0275",
"CI_0014",
"CI_0115"
] | [
"Propose a flexible pipeline that fuses any existing 2D detection network with 3D point clouds to generate accurate 3D bounding boxes with minimal modification to the 2D backbone.",
"Propose a local feature descriptor that integrates point-pair features with global context awareness to improve robustness in 3D po... | [
"related_not_covering",
"not_covering",
"not_covering"
] | [
0.34,
0.1,
0.1
] | recomputed_from_pair_coverage_effective_joint_v1 |
idea_novelty | CI_0415 | Integrate multiple task-specific priors, including unsupervised monocular depth, ground plane constraints, and vehicle shape priors, into a unified energy function to refine 3D pose and shape estimation. | 201,901 | A7 | CT_task | TF_representation | [
"CI_0334",
"CI_0029",
"CI_0041"
] | 0.04 | not_covered | 0.6 | nontrivial_combination | 0.3 | 0.85 | high_novelty | 0.85 | llm_residual_delta_judge_v1 | 0.8 | 0 | 0.62 | new_training_signal | Combines multiple distinct geometric priors (depth, ground plane, shape) into a unified energy function, a specific integration not covered by single-prior baselines. | Qwen3.5-397B-A17B | 0.1809 | not_covered | [
"CI_0334",
"CI_0195",
"CI_0029",
"CI_0032",
"CI_0151"
] | [
"Propose an unsupervised framework that jointly trains depth and odometry estimators using stereo sequences to enforce real-world scale constraints and resolve scale ambiguity.",
"Propose a progressive data augmentation strategy that synthesizes edge maps from large-scale unlabeled images to pre-train the model, ... | [
"related_not_covering",
"not_covering",
"not_covering"
] | [
0.2,
0.07,
0.07
] | recomputed_from_pair_coverage_effective_joint_v1 |
idea_novelty | CI_0416 | Formulate an overall energy function that jointly optimizes two-scale projection consistency and task priors to solve the ill-posed problem of monocular 3D vehicle attribute inference. | 201,901 | A6 | CT_task | TF_representation | [
"CI_0013",
"CI_0269"
] | 0 | not_covered | 0.5 | new_task_definition | 0.4 | 0.75 | moderate_novelty | 0.75 | llm_residual_delta_judge_v1 | 0.8 | 0 | 0.5 | new_mechanism | Formulates a specific joint optimization for monocular 3D vehicle attributes using two-scale consistency, defining a novel problem setup despite using standard optimization concepts. | Qwen3.5-397B-A17B | 0.0895 | not_covered | [
"CI_0226",
"CI_0278",
"CI_0260",
"CI_0013",
"CI_0214"
] | [
"Propose a baseline heuristic that merges outputs from independent semantic and instance segmentation systems via post-processing to establish an initial performance reference for the unified task.",
"Propose a large-margin loss function based on the assumption that deep features follow a Gaussian Mixture distrib... | [
"not_covering",
"not_covering"
] | [
0.1,
0.07
] | recomputed_from_pair_coverage_effective_joint_v1 |
idea_novelty | CI_0417 | Propose a lightweight generator network that refines noisy motion vectors and residuals into a discriminative motion representation by jointly optimizing reconstruction, adversarial, and classification losses. | 201,901 | A6 | CT_method | TF_representation | [
"CI_0190",
"CI_0257",
"CI_0327"
] | 0.04 | not_covered | 0.4 | minor_variant | 0.5 | 0.9 | moderate_novelty | 0.9 | llm_residual_delta_judge_v1 | 0.5 | 0 | 0.86 | minor_mechanism_change | Applies standard GAN and multi-loss optimization to motion vector refinement; the mechanism is an incremental adaptation of existing generative refinement techniques. | Qwen3.5-397B-A17B | 0.5107 | partially_covered | [
"CI_0190",
"CI_0022",
"CI_0100",
"CI_0257",
"CI_0326"
] | [
"Employ a generative adversarial network to synthesize hard negative samples that mimic ambiguous background scenes, improving the discrimination between action starts and similar non-action contexts.",
"Introduce a two-stage convolutional neural network that refines initial 3D bounding box estimates by regressin... | [
"related_not_covering",
"not_covering",
"not_covering"
] | [
0.44,
0.1,
0.1
] | recomputed_from_pair_coverage_effective_joint_v1 |
idea_novelty | CI_0418 | Develop a compressed-domain action recognition framework that achieves optical-flow-level accuracy without explicit flow estimation by generating refined motion cues directly from compressed video streams. | 201,901 | A6 | CT_method | TF_representation | [
"CI_0228",
"CI_0044",
"CI_0156"
] | 0 | not_covered | 0.7 | new_mechanism | 0.2 | 0.9 | high_novelty | 0.9 | llm_residual_delta_judge_v1 | 0.8 | 0 | 0.64 | new_mechanism | Introduces a novel mechanism to generate refined motion cues directly from compressed streams, bypassing explicit flow estimation while matching its accuracy. | Qwen3.5-397B-A17B | 0.2351 | not_covered | [
"CI_0228",
"CI_0084",
"CI_0022",
"CI_0044",
"CI_0112"
] | [
"Redefine the action recognition pipeline to rely solely on RGB images while implicitly leveraging joint pose and appearance features within a unified architecture.",
"Introduce a spatially-aware network architecture incorporating an importance map to adaptively allocate bitrate to visually significant regions wi... | [
"not_covering",
"not_covering",
"not_covering"
] | [
0.19,
0.1,
0.1
] | recomputed_from_pair_coverage_effective_joint_v1 |
idea_novelty | CI_0419 | Introduce a multi-objective training strategy that combines pixel-level reconstruction, adversarial distribution matching, and downstream task supervision to learn motion representations discriminative for action recognition. | 201,901 | A7 | CT_task | TF_representation | [
"CI_0226",
"CI_0027",
"CI_0192"
] | 0.04 | not_covered | 0.6 | new_training_signal | 0.3 | 0.85 | high_novelty | 0.85 | llm_residual_delta_judge_v1 | 0.8 | 0 | 0.62 | new_training_signal | Proposes a specific multi-objective training strategy combining reconstruction, adversarial matching, and task supervision to learn discriminative motion representations. | Qwen3.5-397B-A17B | 0.1809 | not_covered | [
"CI_0226",
"CI_0195",
"CI_0190",
"CI_0078",
"CI_0330"
] | [
"Propose an end-to-end multitask framework that jointly optimizes 2D/3D pose estimation and action recognition by replacing non-differentiable post-processing with a differentiable soft-argmax operation.",
"Adopt a weakly supervised training strategy using pseudo-labels generated by a traditional object detection... | [
"related_not_covering",
"not_covering",
"not_covering"
] | [
0.2,
0.07,
0.07
] | recomputed_from_pair_coverage_effective_joint_v1 |
idea_novelty | CI_0420 | Propose a hierarchical architecture search space that jointly optimizes network-level structures for spatial resolution changes and cell-level structures for layer-wise computations, addressing the limitations of hand-designed outer networks in dense prediction tasks. | 201,901 | A6 | CT_task | TF_representation | [
"CI_0278",
"CI_0226"
] | 0 | not_covered | 0.7 | new_mechanism | 0.2 | 0.8 | high_novelty | 0.8 | llm_residual_delta_judge_v1 | 0.8 | 0 | 0.5 | new_mechanism | Defines a novel hierarchical search space jointly optimizing network-level resolution and cell-level computations, addressing limitations of hand-designed architectures. | Qwen3.5-397B-A17B | 0.0495 | not_covered | [
"CI_0013",
"CI_0278",
"CI_0226",
"CI_0217",
"CI_0167"
] | [
"Propose a multi-task generative architecture that jointly synthesizes the target view image and its semantic segmentation map to enforce semantic consistency across drastically different views.",
"Propose an end-to-end multitask framework that jointly optimizes 2D/3D pose estimation and action recognition by rep... | [
"not_covering",
"not_covering"
] | [
0.1,
0.1
] | recomputed_from_pair_coverage_effective_joint_v1 |
idea_novelty | CI_0421 | Develop a differentiable continuous relaxation formulation for the hierarchical search space, enabling efficient joint optimization of network and cell architectures via stochastic gradient descent instead of computationally intensive reinforcement learning or evolutionary algorithms. | 201,901 | A6 | CT_method | TF_representation | [
"CI_0201",
"CI_0173"
] | 0 | not_covered | 0.7 | new_mechanism | 0.2 | 0.85 | high_novelty | 0.85 | llm_residual_delta_judge_v1 | 0.8 | 0 | 0.5 | new_mechanism | Introduces a differentiable continuous relaxation for hierarchical search spaces, replacing RL/evolution with SGD. This is a substantive algorithmic mechanism shift not covered by priors. | Qwen3.5-397B-A17B | 0.0895 | not_covered | [
"CI_0212",
"CI_0300",
"CI_0207",
"CI_0084",
"CI_0219"
] | [
"Propose a dendritic CNN architecture that enforces facial shape constraints through a tree-structured topology within a single network, replacing traditional joint regression or cascaded methods.",
"Propose a differentiable neural network component that processes object proposals to resolve overlapping detection... | [
"not_covering",
"not_covering"
] | [
0.1,
0.1
] | recomputed_from_pair_coverage_effective_joint_v1 |
idea_novelty | CI_0422 | Design an efficient decoding strategy that greedily selects the optimal cell architecture and applies the Viterbi algorithm to efficiently decode the best network-level architecture from the continuous relaxation. | 201,901 | A6 | CT_application | TF_application | [
"CI_0144",
"CI_0166"
] | 0 | not_covered | 0.6 | new_mechanism | 0.3 | 0.8 | moderate_novelty | 0.8 | llm_residual_delta_judge_v1 | 0.8 | 0 | 0.5 | new_mechanism | Proposes a specific decoding strategy combining greedy selection and Viterbi algorithm for architecture search. While building on relaxation, the specific inference procedure constitutes a new mechanism. | Qwen3.5-397B-A17B | 0.0895 | not_covered | [
"CI_0149",
"CI_0144",
"CI_0132",
"CI_0166",
"CI_0253"
] | [
"Propose a meta-network mechanism that generates layer-specific parameters in a single forward pass to adapt a generic segmentation model to a target object, replacing iterative fine-tuning.",
"Propose a slice pooling layer that projects unordered point features onto an ordered sequence to enable the application ... | [
"not_covering",
"not_covering"
] | [
0.1,
0.1
] | recomputed_from_pair_coverage_effective_joint_v1 |
idea_novelty | CI_0423 | Extend Neural Architecture Search from image classification to dense image prediction tasks, specifically addressing the need for architectures that inherently operate on high-resolution imagery with complex spatial resolution variations. | 201,901 | A1 | CT_task | TF_representation | [
"CI_0317",
"CI_0056"
] | 0 | not_covered | 0.5 | new_task_definition | 0.4 | 0.9 | moderate_novelty | 0.9 | llm_residual_delta_judge_v1 | 0.72 | 0 | 0.44 | new_application_constraint | Extends NAS from classification to dense prediction with high-resolution constraints. This redefines the problem formulation and search space requirements, representing a new task definition rather than just a domain shift. | Qwen3.5-397B-A17B | 0.2472 | not_covered | [
"CI_0317",
"CI_0283",
"CI_0251",
"CI_0241",
"CI_0056"
] | [
"Apply conditional generative adversarial networks to the domain of task-specific visual saliency prediction to automatically learn task-dependent features, bridging the gap between bottom-up deep learning and top-down hand-crafted methods.",
"Formulate the modeling of natural image statistical regularities as a ... | [
"not_covering",
"not_covering"
] | [
0.1,
0.06
] | recomputed_from_pair_coverage_effective_joint_v1 |
idea_novelty | CI_0424 | Propose a new shape representation called 'shape programs' that combines geometric primitives with symbolic control structures like loops to capture high-level regularities such as symmetry and repetition. | 201,901 | A5 | CT_theory | TF_representation | [
"CI_0058",
"CI_0174"
] | 0 | not_covered | 0.8 | new_mechanism | 0.1 | 0.9 | high_novelty | 0.9 | llm_residual_delta_judge_v1 | 0.72 | 0 | 0.47 | new_data_organization | Introduces 'shape programs' combining geometric primitives with symbolic control structures (loops), a distinct representational mechanism not covered by prior edge-graph or object-proposal methods. | Qwen3.5-397B-A17B | 0 | not_covered | [
"CI_0058",
"CI_0123",
"CI_0174",
"CI_0048",
"CI_0021"
] | [
"Introduce a graph-based representation for 2D edges to identify and prioritize the most structurally significant edges within an image.",
"Utilize object proposals from detection networks as input representations instead of raw pixels to facilitate counting tasks in complex natural scenes, addressing the lack of... | [
"not_covering",
"not_covering"
] | [
0.07,
0.07
] | recomputed_from_pair_coverage_effective_joint_v1 |
idea_novelty | CI_0425 | Develop a neural program executor that reconstructs 3D shapes from inferred programs, enabling the model to learn program inference in a self-supervised manner without ground-truth program annotations. | 201,901 | A7 | CT_method | TF_representation | [
"CI_0029",
"CI_0019"
] | 0 | not_covered | 0.85 | new_training_signal | 0.1 | 0.9 | high_novelty | 0.9 | llm_residual_delta_judge_v1 | 0.8 | 0 | 0.48 | new_training_signal | Proposes a neural program executor enabling self-supervised learning of program inference without ground-truth annotations, defining a new training signal distinct from prior unsupervised selection or augmentation strategies. | Qwen3.5-397B-A17B | 0.0895 | not_covered | [
"CI_0089",
"CI_0043",
"CI_0029",
"CI_0101",
"CI_0032"
] | [
"Propose a progressive data augmentation strategy that synthesizes edge maps from large-scale unlabeled images to pre-train the model, enabling effective knowledge transfer to the scarce human-sketch domain.",
"Propose novel unsupervised criteria for selecting representative foreground objects by combining semant... | [
"not_covering",
"not_covering"
] | [
0.1,
0.1
] | recomputed_from_pair_coverage_effective_joint_v1 |
idea_novelty | CI_0426 | Design an end-to-end differentiable architecture that jointly learns to infer symbolic shape programs from input data and execute them to reconstruct the 3D geometry. | 201,901 | A6 | CT_method | TF_representation | [
"CI_0038",
"CI_0152",
"CI_0065",
"CI_0025"
] | 0.04 | not_covered | 0.75 | new_mechanism | 0.15 | 0.95 | high_novelty | 0.95 | llm_residual_delta_judge_v1 | 0.7 | 0 | 0.6 | new_mechanism | Designs an end-to-end differentiable architecture jointly inferring symbolic programs and executing them for 3D reconstruction, a specific mechanism not present in prior spatial transform or point cloud densification approaches. | Qwen3.5-397B-A17B | 0.3624 | weakly_covered | [
"CI_0038",
"CI_0152",
"CI_0262",
"CI_0133",
"CI_0193"
] | [
"Propose a deep supervision mechanism that sequentially supervises hidden CNN layers with unobserved intermediate shape concepts to regularize learning and improve generalization for geometry estimation.",
"Implement a neural architecture incorporating a spatial transform layer to reconstruct one image from anoth... | [
"related_not_covering",
"not_covering",
"not_covering",
"not_covering"
] | [
0.34,
0.19,
0.1,
0.1
] | recomputed_from_pair_coverage_effective_joint_v1 |
idea_novelty | CI_0427 | Extend the program inference framework to accept RGB images as input, allowing direct inference of 3D shape programs and physically plausible reconstructions from single views. | 201,901 | A1 | CT_application | TF_application | [
"CI_0168",
"CI_0118"
] | 0 | not_covered | 0.7 | new_mechanism | 0.3 | 0.85 | high_novelty | 0.85 | llm_residual_delta_judge_v1 | 0.72 | 0 | 0.44 | new_application_constraint | Extends program inference to raw RGB inputs, introducing a new mechanism for direct 3D shape program synthesis from single views, distinct from prior point-cloud or video segmentation methods. | Qwen3.5-397B-A17B | 0.0895 | not_covered | [
"CI_0168",
"CI_0242",
"CI_0277",
"CI_0181",
"CI_0010"
] | [
"Develop an efficient architecture for 3D segmentation that directly processes raw point clouds without converting them to voxels or multi-view images, thereby avoiding information loss and quantization artifacts.",
"Propose a unified framework that performs both spatial selection of normal field-of-view segments... | [
"not_covering",
"not_covering"
] | [
0.1,
0.1
] | recomputed_from_pair_coverage_effective_joint_v1 |
idea_novelty | CI_0428 | Propose a module that integrates graphical long short-term memory mechanisms into convolutional networks to model complex dependency patterns among body parts for boosting intermediate feature maps. | 201,901 | A6 | CT_method | TF_representation | [
"CI_0145",
"CI_0154"
] | 0 | not_covered | 0.7 | new_mechanism | 0.2 | 0.9 | high_novelty | 0.9 | llm_residual_delta_judge_v1 | 0.8 | 0 | 0.5 | new_mechanism | Proposes a novel Graph-LSTM module integrated into CNNs to explicitly model body-part dependencies, a distinct architectural mechanism compared to prior parameter injection or local descriptor approaches. | Qwen3.5-397B-A17B | 0.0895 | not_covered | [
"CI_0145",
"CI_0249",
"CI_0248",
"CI_0217",
"CI_0206"
] | [
"Introduce a network modulation process that injects generated parameters into intermediate layers for feature manipulation based on limited visual and spatial priors.",
"Propose a local feature descriptor that integrates point-pair features with global context awareness to improve robustness in 3D point cloud ma... | [
"not_covering",
"not_covering"
] | [
0.1,
0.1
] | recomputed_from_pair_coverage_effective_joint_v1 |
End of preview. Expand in Data Studio
IdeaDelta
Idea-level novelty assessment across Graph ML, CV, and Time Series.
A dataset for idea-level scientific novelty assessment: given a research idea and the prior ideas that precede it in time, decide how much of the idea is already covered by the prior work, and how novel the remaining residual is.
Every sample is an idea — a single atomic research claim extracted from an arXiv paper — not a whole paper. Ideas are linked to time-ordered prior ideas retrieved from the same domain, and both the pairwise coverage relations and the idea-level novelty labels are LLM-judged.
Three domains are released, each a self-contained dataset with its own strict temporal train/dev/test split: Graph ML (5,225 ideas), CV (8,291 ideas), and Time Series (7,595 ideas) — 21,111 ideas and 92,876 idea pairs in total. Each domain is built and split independently, and no pair crosses a domain boundary.
Quick start
from datasets import load_dataset
# pairwise coverage, temporal splits — the main training target
pairs = load_dataset("holaholaIDEADELTA/IdeaDelta", "graph_pair_staged")
print(pairs)
# train / train_balanced / validation / test
# idea-level novelty labels
ideas = load_dataset("holaholaIDEADELTA/IdeaDelta", "cv_novelty", split="train")
print(ideas[0]["target_text"], ideas[0]["novelty_label"])
Each domain has four configs — <domain>_pair_staged, <domain>_novelty,
<domain>_pair_coverage, and <domain>_prior_ranking — where <domain> is
graph, cv, or time_series.
The raw JSONL files are also in the repo under data/, byte-identical to the
files used for the reported experiments:
from huggingface_hub import snapshot_download
path = snapshot_download(
"holaholaIDEADELTA/IdeaDelta",
repo_type="dataset",
allow_patterns="data/graph/*",
)
What is in the repo
data/ JSONL, byte-identical to the experiment inputs (+ dataset_manifest.json)
graph/ cv/ time_series/
parquet/ Parquet mirror of the same rows, used by the Hub dataset viewer
Every dataset_manifest.json records the row count and SHA-256 of each JSONL
file, so the published bytes can be checked against the experiment inputs.
The four record types
<domain>_novelty — idea-level novelty label (4-way)
One row per target idea, carrying the retrieved top priors, the aggregate coverage of the idea by those priors, and the novelty verdict.
| field | description |
|---|---|
target_idea_id, target_text, target_date |
the idea, and its YYYYMM publication month |
primary_aspect, contribution_type, task_family |
facet labels for the idea |
top_priors, top_prior_texts, top_prior_coverage_labels, top_prior_coverage_scores |
the retrieved prior context |
joint_coverage_score, joint_coverage_label |
how much of the idea the priors jointly cover |
substantive_delta_score, surface_change_score, residual_type |
what is left over, and what kind of contribution it is |
novelty_label |
high_novelty / moderate_novelty / weak_novelty / low_novelty |
evidence_confidence, label_confidence, label_source |
annotation confidence and provenance |
pre_recomputed_*, pre_llm_* |
pre-LLM heuristic values, kept for ablations |
llm_residual_rationale, llm_residual_model |
the judge's free-text reason and model id |
residual_type is one of 11 categories: new_mechanism, minor_variant,
nontrivial_combination, new_training_signal, new_evaluation_protocol,
new_theory_or_concept, new_task_definition, new_application_constraint,
new_system_design, new_data_distribution, new_data_organization.
<domain>_pair_coverage — pairwise coverage label (4-way)
One row per (target idea, prior idea) pair, the full labelled pool.
| field | description |
|---|---|
target_*, prior_* |
both ideas with their dates and facet labels |
relation |
graph relation between the two ideas (nullable) |
similarity, semantic_similarity |
lexical and embedding similarity |
same_aspect, same_contribution_type, same_task_family |
facet agreement flags |
coverage_label |
large_cover / partial_cover / related_not_covering / not_covering |
coverage_score, label_confidence, label_source |
score, confidence, and which signals produced it |
original_*, semantic_*, pre_llm_* |
pre-LLM values, kept for ablations |
llm_rationale, llm_model |
the judge's free-text reason and model id |
<domain>_pair_staged — the same pairs under a temporal split
The training-ready view of pair_coverage. Splits are cut by target date, so
every dev/test target is strictly later than every training target — no temporal
leakage. Adds staged-decision labels for a coarse-to-fine classifier:
| field | description |
|---|---|
pair_label_4way |
the 4-way coverage label |
relatedness_stage_label |
stage 1: related / unrelated |
covering_stage_label, covering_stage_mask |
stage 2: covering / not_covering_related / ignore |
cover_strength_label, cover_strength_mask |
stage 3: large_cover / partial_cover |
coverage_level_3way |
ordinal 0/1/2 view of coverage |
split |
train / dev / test |
train_balanced resamples the natural training set to 1,000 rows per class
(max 5× oversampling); duplicated rows are flagged with
is_oversampled_duplicate and balanced_duplicate_id. validation and
test keep their natural distribution and are never resampled.
<domain>_prior_ranking — retrieval targets
One row per target idea, with candidate_priors: a ranked list of prior ideas,
each carrying its own text, date, relation, similarity, and coverage label. Use
it for retrieval/reranking rather than classification.
Statistics
| domain | ideas | pairs | train | train_balanced | validation | test |
|---|---|---|---|---|---|---|
| graph | 5,225 | 21,183 | 13,648 | 4,000 | 3,746 | 3,789 |
| cv | 8,291 | 36,238 | 22,888 | 4,000 | 6,369 | 6,981 |
| time_series | 7,595 | 35,455 | 22,554 | 4,000 | 6,585 | 6,316 |
Temporal boundaries (YYYYMM of the target idea):
| domain | train | validation | test |
|---|---|---|---|
| graph | 201901–202506 | 202506–202511 | 202511–202606 |
| cv | 201901–202501 | 202501–202503 | 202503–202602 |
| time_series | 201901–202503 | 202503–202511 | 202511–202609 |
Novelty labels:
| domain | high | moderate | weak | low |
|---|---|---|---|---|
| graph | 919 | 2,247 | 1,240 | 819 |
| cv | 2,902 | 2,777 | 1,711 | 901 |
| time_series | 2,084 | 2,173 | 1,614 | 1,724 |
Pairwise coverage labels are heavily imbalanced, which is the point — most prior ideas do not cover a target:
| domain | split | large_cover | partial_cover | related_not_covering | not_covering |
|---|---|---|---|---|---|
| graph | all | 747 | 1,541 | 8,658 | 10,237 |
| graph | train | 448 | 865 | 5,163 | 7,172 |
| graph | validation | 161 | 318 | 1,726 | 1,541 |
| graph | test | 138 | 358 | 1,769 | 1,524 |
| cv | all | 592 | 3,354 | 15,607 | 16,685 |
| cv | train | 298 | 1,927 | 8,891 | 11,772 |
| cv | validation | 148 | 654 | 3,108 | 2,459 |
| cv | test | 146 | 773 | 3,608 | 2,454 |
| time_series | all | 1,415 | 4,687 | 14,510 | 14,843 |
| time_series | train | 820 | 2,742 | 8,553 | 10,439 |
| time_series | validation | 319 | 987 | 3,094 | 2,185 |
| time_series | test | 276 | 958 | 2,863 | 2,219 |
Every domain's train_balanced is 4,000 rows, 1,000 per class.
How it was built
- Idea extraction — atomic research claims are extracted from arXiv papers in each domain and normalised into single-sentence idea statements.
- Facet labelling — each idea gets an aspect, contribution type, and task family label.
- Prior retrieval — for each target idea, candidate priors are retrieved from strictly earlier ideas in the same domain using lexical and embedding similarity plus graph relations.
- Pairwise coverage judging — an LLM judges how much each prior covers the
target, producing the 4-way
coverage_labelwith a rationale. - Residual / novelty judging — the retained coverage is aggregated per
target idea, and an LLM judges the substantive residual, yielding
residual_typeand the 4-waynovelty_label. - Temporal splitting — pairs are split by target date into train/dev/test, then a class-balanced training variant is resampled from the training half only.
The annotation model differs by domain, which matters for cross-domain comparison:
| domain | label pipeline | judge model |
|---|---|---|
| graph | taxonomy-based (taxonomy_llm) |
deepseek-v4-pro |
| cv | fixed threshold (fixed_threshold_llm) |
Qwen3.5-397B-A17B |
| time_series | fixed threshold (fixed_threshold_llm) |
Qwen3.5-397B-A17B |
Intended use and limitations
Intended for training and evaluating novelty-assessment models: pairwise coverage classification, idea-level novelty classification, and prior retrieval or reranking.
Known limitations:
- Labels are LLM-generated, not human-annotated. They carry the judges'
biases and are not a gold standard for what the research community considers
novel.
label_confidenceandevidence_confidenceare the judges' own self-reported values. - Novelty is defined relative to the retrieved priors, not to all prior art.
An idea labelled
high_noveltyis novel with respect to the retrieved candidate set within the same domain's corpus, which is itself incomplete. - Domains are annotated by different models and pipelines (see the table above) and split at different temporal boundaries, so absolute numbers are not directly comparable across domains.
- Class imbalance is severe in pairwise coverage —
large_coveris under 3% of all pairs. Report per-class and macro metrics, not accuracy alone. - Idea text is LLM-rewritten from paper content, so extraction errors and paraphrase drift are possible; ideas are not verbatim quotes from the papers.
- The facet fields
graph_formandmechanism_family, present in some upstream artefacts, are not included here because they proved unreliable.
Source and licensing
Ideas are derived from the content of arXiv preprints via LLM extraction and rewriting; no full paper text is redistributed. arXiv papers carry per-paper licences chosen by their authors. This dataset is released under CC BY 4.0 as a derived annotation layer; users are responsible for complying with the terms of the underlying papers for their own use case.
Citation
Withheld during anonymous review.
- Downloads last month
- 23