Datasets:

MarkrAI

/

AutoRAG-evaluation-2024-LLM-paper-v1

like 8

772ad6ee-2eae-4a34-b0e8-7be678f81949

# Generative Cross-Modal Retrieval: Memorizing Images In Multimodal Language Models For Retrieval And Beyond ## 4.4 Efficiency Analysis In large-scale cross-modal retrieval, efficiency emerges as a crucial factor. This is why the onetower framework, effective for small-scale ranking, falls short in the retrieval stage. To address this, we conducted experiments comparing the efficiency of CLIP and GRACE. CLIP can pre-encode all images into vectors, incurring most of its inference cost from text encoding and calculating the similarity between text embeddings and image embeddings. In contrast, the generative framework necessitates generating identifiers. We assessed the query latency of both CLIP and GRACE to varying image sizes, with detailed results presented in Figure 4. Our findings are insightful. Firstly, CLIP's inference speed decreases progressively as image size increases, owing to the escalating number of similarity calculations required. Secondly, the inference speed of our generative framework remains nearly constant, a result of encoding all images directly into its parameters. Thirdly, when image sizes exceed a certain threshold (about 150,000 images), our generative framework surpasses CLIP in terms of inference speed, and this advantage grows as image sizes continue to increase. Lastly, these findings underscore that the generative framework is not only capable of large-scale image retrieval but can also perform comparably to two-tower approaches.

233163a4-7556-4829-8faf-2823fe50913f

# Generative Cross-Modal Retrieval: Memorizing Images In Multimodal Language Models For Retrieval And Beyond ## 4.5 Beyond Cross-Modal Retrieval We enable the MLLM to memorize images within its parameters using unique identifiers. Once the images are adequately memorized, the MLLM can produce the corresponding images (identifiers) to respond to users' queries, as illustrated in Figure 5 (a). While the visual memory in the MLLM facilitates image retrieval, its applications are not restricted to retrieval alone after other instruction tunings. We present two examples in Figure 5 (b) and Figure 5 (c), respectively. - **Describing memorized images**: As the MLLM has successfully memorized certain images, it is capable of providing a description of the image's content when prompted. As depicted in the examples shown in Figure 5, when given an instruction such as "please describe the context of the image *3,756*", the model is able to provide a description of the image, albeit not in great detail. - **QA over memorized images**: Similarly, the model is capable of answering some questions over the memorized images. Given an instruction consisting of the image identifier and question, the model can answer based solely on memorization without any image input.

09c52059-4dcb-4d9a-86f2-cdbacc21c833

# Generative Cross-Modal Retrieval: Memorizing Images In Multimodal Language Models For Retrieval And Beyond ## 4.6 Beam Size Analysis We conducted experiments to analyze the beam size of GRACE, as detailed in Appendix B.

77b2ad62-944d-4ed6-ae4b-bfc0ffce21f7

# Generative Cross-Modal Retrieval: Memorizing Images In Multimodal Language Models For Retrieval And Beyond ## 5 Conclusion And Future Work In this paper, we delved into a novel memorization mechanism for the MLLM to memorize images within its parameters. Building upon inbuilt visual memory within MLLM, we proposed a generative cross-modal retrieval framework, which introduces a fresh paradigm in cross-modal retrieval. This paradigm transforms the original matching problem into a generation problem, eliminating the need for negative samples during training and image indexing during inference. Our experiments demonstrate that the generative paradigm performs effectively and efficiently even with large-scale image sizes. Furthermore, we showcased the MLLM's ability to interact (e.g., describe and QA) with memorized images, following specific instructions. Moving forward, we aim to further develop this topic from the following perspectives. On the one hand, although our generative framework achieves comparable performance to previous cross-modal retrieval approaches, there are still challenges to address, such as the limitations of current identifiers. Exploring more effective identifiers, like "visual tokens (Van Den Oord et al., 2017)", would help to enhance generative cross-modal retrieval further. On the other hand, since we have enabled MLLMs to memorize and interact with images, it opens up the possibility of injecting personalized visual experiences of humans into MLLMs for them to understand an individual's visual journey and accomplish more visual tasks.

699570b7-7015-45a9-b822-1e2c9e51c0d3

# Generative Cross-Modal Retrieval: Memorizing Images In Multimodal Language Models For Retrieval And Beyond ## Limitations This work introduces a new paradigm in text-image retrieval, but it also has some limitations to be addressed. 1) The evaluation of GRACE's image retrieval ability on Flickr30K and MS-COCO was compared with two-tower baselines. However, it is important to note that Flickr30K and MS-COCO are also used as benchmarks for text-image ranking approaches, where one-tower frameworks have dominated. This may confuse newcomers to the field, as they may perceive GRACE and two-tower approaches as lagging behind the one-tower framework. However, it should be noted that GRACE and one-tower approaches focus on image retrieval, placing high demands on retrieval efficiency, while two-tower approaches are primarily suitable for the ranking stage, allowing for more time-consuming calculations to improve performance. 2) The identifiers currently used by GRACE are not as satisfactory as expected, only yielding results comparable to previous methods. However, as a pioneering work, the main significance of this work lies in validating the feasibility of generative cross-model retrieval. Further research is expected to enhance this paradigm.

77cffa73-2117-4946-8118-36301ef1d41a

# Generative Cross-Modal Retrieval: Memorizing Images In Multimodal Language Models For Retrieval And Beyond ## Ethics Statement The datasets used in our experiment are publicly released and labeled through interaction with humans in English. In this process, user privacy is protected, and no personal information is contained in the dataset. The scientific artifacts that we used are available for research with permissive licenses. And the use of these artifacts in this paper is consistent with their intended use. Therefore, we believe that our research work meets the ethics of ACL.

c125ab88-b2e5-4f8e-ae8d-ad8c87b734cb

# Generative Cross-Modal Retrieval: Memorizing Images In Multimodal Language Models For Retrieval And Beyond ## A Implement Details We selected the open-flamingo (Awadalla et al., 2023) with the 3B parameters as our model's backbone. The visual encoder of the open-flamingo is a 12-layer visual transformer, while its language model is based on MPT-1B3. We adopted the deepspeed (Rasley et al., 2020) training framework to train the model on 4×24GB NVIDIA A5000 GPUs. We froze the visual encoder and fine-tuned the language model as well as cross-attention layers. We employed the Adam optimizer, setting a learning rate of 1e-4 and a batch size of 64 for each GPU. On the Flickr30K dataset, our training included 1,000K steps for learning to memorize and 3,000K steps for learning to retrieve. For the MS-COCO dataset, these numbers were increased to 2,000K and 6,000K steps, respectively. We have trained the model several times to confirm that the improvement is not a result of random chance and present the mid one. The training duration was approximately 12 hours for Flickr30K and 24 hours for MS-COCO.

80682e82-32e7-4955-83e2-ddb215a1a164

# Generative Cross-Modal Retrieval: Memorizing Images In Multimodal Language Models For Retrieval And Beyond ## B Beam Size Analysis GRACE relies on beam search to obtain top-k retrieval results. We conducted detailed experiments to understand the impact of varying beam sizes, and the findings are illustrated in Figure 6. The atomic identifier is excluded from this experiment as it only requires one generation step, and beam size will not affect its performance. Increasing the beam size exhibits marginal benefits. This observation aligns with our expectations, as candidates with larger beam sizes generally score lower, diminishing their likelihood of being the top result. In terms of Recall@10, we observed a notable improvement in performance with the expansion of the beam size. This enhancement is attributed to the inclusion of candidates that would otherwise be missed in scenarios with a more constrained beam size.

05b6615b-12b4-456a-b6b9-a73dc9af2fef

# Selective Reflection-Tuning: Student-Selected Data Recycling For Llm Instruction-Tuning Ming Li1, Lichang Chen1, Jiuhai Chen1, Shwai He1, Jiuxiang Gu2**, Tianyi Zhou**1 1University of Maryland, College Park 2Adobe Research {minglii, bobchen, tianyi}@umd.edu Project: https://github.com/tianyi-lab/Reflection_Tuning

7d799b39-8ab2-4fbe-a4df-1f4fb61194ea

# Selective Reflection-Tuning: Student-Selected Data Recycling For Llm Instruction-Tuning ## Abstract Instruction tuning is critical to large language models (LLMs) for achieving better instruction following and task adaptation capabilities but its success heavily relies on the training data quality. Many recent methods focus on improving the data quality but often overlook the compatibility of the data with the student model being finetuned. This paper introduces Selective Reflection-Tuning, a novel paradigm that synergizes a teacher LLM's reflection and introspection for improving existing data quality with the data selection capability of the student LLM, to automatically refine existing instruction-tuning data. This teacherstudent collaboration produces high-quality and student-compatible instruction-response pairs, resulting in sample-efficient instruction tuning and LLMs of superior performance. Selective Reflection-Tuning is a data augmentation and synthesis that generally improves LLM finetuning and self-improvement without collecting brand-new data. We apply our method to Alpaca and WizardLM data and achieve much stronger and top-tier 7B and 13B LLMs. Our codes, models, and data will be released at https://github.com/tianyi-lab/ Reflection_Tuning.

18247efb-54e5-48c7-b8d8-6cc8f4143712

# Selective Reflection-Tuning: Student-Selected Data Recycling For Llm Instruction-Tuning ## 1 Introduction The quality of instruction tuning (Wei et al., 2022; Chen et al., 2023a; Mishra et al., 2021; Chung et al., 2022; Zhang et al., 2023) data is paramount to the LLM being fine-tuned, i.e., the student model. There is a growing trend and demand for the community to automatically improve the quality of instruction tuning data. Previous works either curate datasets by human experts (Conover et al., 2023; Longpre et al., 2023; Zhou et al., 2023) or distill the responses of well-trained LLMs (Taori et al., 2023; Peng et al., 2023; Chiang et al., 2023; Vu et al., 2023; Xu et al., 2023a). The self-improvement (Bai et al., 2022b; Huang et al., 2022; Pan et al., 2023) ability of LLMs has also been explored to improve the instruction or response of a training sample. However, these existing methods of data enhancement (Huang et al., 2022; Ye et al., 2023; Li et al., 2023b; Mitra et al., 2023) do not take a critical criterion into account: Is the teacherrefined data compatible to the needs of the student model? These approaches typically do not account for the inherent randomness and potential degradation associated with the generative models' output, leading to an oversight in how the student model responds to these "improved" data samples. Thus a mechanism for the student model to selectively integrate these enhancements has been notably absent. To bridge this gap, our work introduces an teacher-student collaboration pipeline wherein a teacher generative model engages in a reflection process to enhance both the instruction and response of a data sample. The student model then evaluates whether to incorporate these improvements based on its unique statistical attributes. This pipeline is versatile and can be adapted to various contexts where data enhancement is needed. Then, another pivotal question arises: How does the student model decide which enhanced data are needed and critical to its training? This question underpins the challenge of autonomously evaluating the quality of instructions and responses. Common practices involve utilizing sophisticated models like GPT-4 for assessment purposes (Zheng et al., 2023; Li et al., 2023e; Liu et al., 2023b; Chiang and Lee, 2023) or employing a secondary

ab846f9e-96d8-429a-905f-3da5b8031f7f

# Selective Reflection-Tuning: Student-Selected Data Recycling For Llm Instruction-Tuning ## 1 Introduction student model then evaluates whether to incorporate these improvements based on its unique statistical attributes. This pipeline is versatile and can be adapted to various contexts where data enhancement is needed. Then, another pivotal question arises: How does the student model decide which enhanced data are needed and critical to its training? This question underpins the challenge of autonomously evaluating the quality of instructions and responses. Common practices involve utilizing sophisticated models like GPT-4 for assessment purposes (Zheng et al., 2023; Li et al., 2023e; Liu et al., 2023b; Chiang and Lee, 2023) or employing a secondary judge model equipped with evaluative capabilities (Wang et al., 2023c; Li et al., 2023a). These methods, however, present limitations: they fail to address the discrepancies between the evaluating model and the actual student model undergoing training. Particularly in the latter approach, even though the judge model and the student model might share the same structural framework, their weight distributions diverge once endowed with the evaluative functions. Consequently, the preferences of the judge model may not align with the real student model's requirements. To circumvent these issues, we adopt a statistical method, utilizing the Instruction-Following Difficulty (IFD) score proposed by Li et al. (2023c). This score is derived directly from the raw student model, thereby mitigating potential domain shifts and ensuring that the evaluation is better aligned with the student model's learning context. In our approach, the IFD score serves as a crucial metric that measures how much help the instruction can provide to the likelihood of the response if added as an extra condition, representing the Difficulty of the sample. However, though effective, the IFD score mainly assesses the instructions. Motivated by Humpback (Li et al., 2023d) which requires LLMs to generate potential instruction based on responses, we further introduce a reversed version of IFD named reversed-IFD (r-IFD). This metric evaluates how much the response contributes to predicting the corresponding instruction. A lower r-IFD score suggests the student can easily deduce the corresponding instruction given the response, indicating this sample is feasible for the student to learn, representing the **Feasibility** of the sample. This dual approach, employing both IFD scores for Difficulty and **r-IFD scores for Feasibility**, enables a comprehensive and nuanced assessment of the instruction-tuning process

1fb44607-1345-40e7-8696-cbc96509240a

# Selective Reflection-Tuning: Student-Selected Data Recycling For Llm Instruction-Tuning ## 1 Introduction requires LLMs to generate potential instruction based on responses, we further introduce a reversed version of IFD named reversed-IFD (r-IFD). This metric evaluates how much the response contributes to predicting the corresponding instruction. A lower r-IFD score suggests the student can easily deduce the corresponding instruction given the response, indicating this sample is feasible for the student to learn, representing the **Feasibility** of the sample. This dual approach, employing both IFD scores for Difficulty and **r-IFD scores for Feasibility**, enables a comprehensive and nuanced assessment of the instruction-tuning process, ensuring the refined data aligns well with the student model's capabilities and objectives. We name our overall method Selective Reflection-Tuning, which contains the selective instruction reflection phase and the selective response reflection phase. In the first phase, a teacher model is utilized to reflect on the instruction of the given sample based on some criteria and generate a new sample. Then the student model makes the decision of whether to accept the improvement based on difficulty (IFD). In the second phase, the teacher model reflects and generates a sample with a new response and the student model decides whether to accept based on feasibility (r-IFD). With our interactive pipeline, we obtain a dataset with supreme quality, with only instruction tuning on a relatively small amount of data, our model outperforms most existing open-source models with even larger model sizes. Our contributions include: - We propose a teacher-student collaboration pipeline where the teacher model and student model cooperate to build a more coherent and model-compatible instruction tuning dataset, which can be further adapted into other selfimprovement scenarios. - We present a nuanced evaluation schema reversed-IFD, quantifying the relevance of instruction-response pairs, and representing the feasibility of the sample for the student. - With only instruction tuning on a few thousand of automatically generated data, our models achieve top-tier performances, indicating the supreme quality of our data.

c694b4f4-2c4a-4fcd-8a50-b5e82cbd6c87

# Selective Reflection-Tuning: Student-Selected Data Recycling For Llm Instruction-Tuning ## 2 Preliminaries Let fθ denote the pre-trained student model, e.g., LLaMA, with parameters θ and g the teacher model, e.g., ChatGPT. Let lowercase letters x, y, z, c, .. denote the text segments, which could be phrases or sentences, and each token in x is denoted as x[i]. We use uppercase letters D, .. to denote the collection of language sequences or datasets, and D0 represents the initial base dataset. Since both fθ and g are in auto-regressive manners, a sequence x = (x[1]*, ..., x*[n]) can be further denoted as: $$f_{\theta}(x)=\prod_{i=1}^{n}f(x[i]|x[1,...,i-1])\qquad(1)$$ In the instruction tuning setting, there will be a mapping function that turns the original raw instruction x into the desirable format and requests models for a response y. For simplicity, we directly notate this process as y ∼ f(y|x). And the loss function for instruction-tuning can be denoted as: $$L_{\theta}(y|x)=-\frac{1}{n}\sum_{i=1}^{n}\log f_{\theta}(y|x)\tag{2}$$ where n is the length of response y. Motivated by Cherry LLM (Li et al., 2023c) which proposes the IFD score to measure the difficulty of instruction in the given instructionresponse pairs. We utilize the perplexity of the IFD score (Li et al., 2024), which is formulated as: $$\text{IFD}_{\theta}(y|x)=\frac{\text{ppl}(y|x)}{\text{ppl}(y)}=\exp(L_{\theta}(y|x)-L_{\theta}(y))\tag{3}$$ where ppl(y|x) represents the perplexity of model fθ to fit the response y given the instruction x as the context, and ppl(y) represents the perplexity of model fθ to directly fit the response y without any context given. This value represents how the

7d1203a1-ec77-411f-9c91-dc69fd66ccc2

# Selective Reflection-Tuning: Student-Selected Data Recycling For Llm Instruction-Tuning ## 2 Preliminaries IFD}_{\theta}(y|x)=\frac{\text{ppl}(y|x)}{\text{ppl}(y)}=\exp(L_{\theta}(y|x)-L_{\theta}(y))\tag{3}$$ where ppl(y|x) represents the perplexity of model fθ to fit the response y given the instruction x as the context, and ppl(y) represents the perplexity of model fθ to directly fit the response y without any context given. This value represents how the given instruction x affects the generation of corresponding response y for given model fθ, which has been shown as an effective metric for evaluating the given instruction-following data pairs (Li et al., 2024). A higher IFD score indicates that the instruction is more challenging for the student model to generate the response, suggesting the instruction's difficulty for the student model.

d66ffe06-98e4-4f7b-abe0-4b3c4593de2b

# Selective Reflection-Tuning: Student-Selected Data Recycling For Llm Instruction-Tuning ## 3 Methodology As shown in Figure 1, there are two main phases in our method, Selective Instruction Reflection and Selective Response Reflection phase. In each phase, the teacher model generates the updated version of instructions or responses based on some given specific criteria {cins,1, ..., c*ins,k*} 1, then the student model judges if the updates are beneficial to it based on difficulty (IFD) or feasibility (reverse- IFD). Finally, these selectively improved samples can be used for the final instruction tuning.

be4c3108-9415-4c50-abcd-317d5133fc4e

# Selective Reflection-Tuning: Student-Selected Data Recycling For Llm Instruction-Tuning ## 3.1 Selective Reflection On Instruction Reflection on Instruction Given the instruction-response pair (x0, y0) from the original dataset D0 with some specific criteria {cins,1, ..., c*ins,k*}, the teacher model g is required to reflect on this sample and generate a better instruction-response pair (xins, yins) according to its reflection. With the criteria given, the teacher model g is able to generate critical responses: $$[\tilde{z}_{ins,1},...]\sim g(z,...|x_{0},y_{0},c_{ins,1},...)\tag{4}$$ where both original instruction and response are wrapped into the prompt rather than original instruction alone. These critical responses further serve as the guidance (chain of thought) (Wei et al., 2023; Yao et al., 2023) for the generation of the new instruction and response pair: $$[x_{ins},y_{ins}]\sim g(x,y|x_{0},y_{0},v_{ins,1},...,z_{ins,1},...)\tag{5}$$ where the above process is sampled as a continuous language sequence, and the critical responses would not be decomposed from the whole output. **Selection on Instruction** Though the given sample pair is updated by the teacher model, it remains uncertain whether this updated version is truly better for the student model. While most existing work evaluates the quality of a data sample by directly prompting existing generative models, they inevitably suffer from the misalignment problem. Thus we utilize the IFD score (Li et al., 2023c) calculated based on the specific base student model, which measures how the instruction benefits the generation of corresponding responses for the model, representing the difficulty of the sample. After obtaining the updated instruction-response pair, the base model $f_{\theta}$ is required to compare the IFD score of the original pair $(x_{0},y_{0})$ and updated pair (xins, yins) and the sample with higher IFD scores will be chosen: ($x_{1},y_{1}$) = argmax(IFD${}_{\theta}$($y|x$)) (6) ($x$,$y$) where (x, y) ∈ {

ec8bf6cb-8a1b-45f0-a1f0-c4f0314e679f

# Selective Reflection-Tuning: Student-Selected Data Recycling For Llm Instruction-Tuning ## 3.1 Selective Reflection On Instruction of the sample. After obtaining the updated instruction-response pair, the base model $f_{\theta}$ is required to compare the IFD score of the original pair $(x_{0},y_{0})$ and updated pair (xins, yins) and the sample with higher IFD scores will be chosen: ($x_{1},y_{1}$) = argmax(IFD${}_{\theta}$($y|x$)) (6) ($x$,$y$) where (x, y) ∈ {(x0, y0), (xins, yins)}. Then the chosen data pair (x1, y1) with a higher IFD score will be sent to the next phase.

821f6cbc-9f74-48e4-86c0-b438dd3fbdc1

# Selective Reflection-Tuning: Student-Selected Data Recycling For Llm Instruction-Tuning ## 3.2 Selective Reflection On Response Reflection on Response After the first phase, although the instruction x1 is guaranteed to be difficult for the student model, the corresponding response y1 is still sub-optimal. Thus another reflection on the response process is further proposed. Similar to the above procedure, a new set of criteria for reflection on response is defined as {cres,1, ..., c*res,m*}. The overall process can be noted as: yres ∼ g(y|x1, y1, cres,1, ..., cres,m, zres,1, ..., z*res,m*) (7) where z*res,i* represents the critical response of ith response criteria c*res,i*. In the process, the instruction and response pair (x1, yres)) is fully improved. Selection on Response Our pipeline aims to improve both the instruction and response in an instruction-tuning sample. IFD score measures the difficulty of the sample. We take a step further by adding another dimension which we call reversed IFD (r-IFD) representing the feasibility for the student to generate the instruction given the response. A lower r-IFD score suggests the student can easily deduce the corresponding instruction given the response, indicating this sample is feasible for the student to learn, which measures the model-specific matching degree of the existing data pair. 2 The high-level idea of r-IFD is in line with the success of Humpback (Li et al., 2023d), which utilizes LLM to predict the corresponding instruction from given texts (responses), and hypothesizes that "we can predict instructions for these candidate gold answers that can be used as high-quality example pairs". In our paper, we further hypothesize that a response is more informative for training if it is feasible for the LLM to predict the corresponding instruction from the response. This hypothesis is naturally proved by the Humpback, which generates instructions that can be handled by LLMs, while those difficult ones are naturally discarded. Under this circumstance, the reversed IFD score should be small since the smaller value represents that it is easier for the model to generate the corresponding instruction given the response. Specifically, the r-IFD score is calculated as: $$\text{r-IFD}_{\theta}(x|y)=\frac{\text{ppl}(

7b29c9f6-e006-477b-805a-f16c7bb331b4

# Selective Reflection-Tuning: Student-Selected Data Recycling For Llm Instruction-Tuning ## 3.2 Selective Reflection On Response for training if it is feasible for the LLM to predict the corresponding instruction from the response. This hypothesis is naturally proved by the Humpback, which generates instructions that can be handled by LLMs, while those difficult ones are naturally discarded. Under this circumstance, the reversed IFD score should be small since the smaller value represents that it is easier for the model to generate the corresponding instruction given the response. Specifically, the r-IFD score is calculated as: $$\text{r-IFD}_{\theta}(x|y)=\frac{\text{ppl}(x|y^{\prime})}{\text{ppl}(x)}=\exp(L_{\theta}(x|y^{\prime})-L_{\theta}(x))\tag{8}$$ where y′ represents the text segment generated by mapping the original y into a query to guess the corresponding potential instructions. For the given original sample pair (x1, y1) from the first phase and reflected sample pair (x1, yres), the selection process can be formulated as: ($x_{2},y_{2}$) = argmin(r-IFD${}_{\theta}(x|y)$) (9) ($x$,$y$) where (x, y) ∈ {(x1, y1), (x1, yres)}. After the above phases, there will be a corresponding data pair (x2, y2) for each original (x0, y0), which is represented as our selective reflected data. Then we discard all the samples which is not response-reflected for the consistency of response distribution. We name the whole above process as a selective recycling process, which greatly improves the quality of the previous dataset 3. The student model fθ will be trained on the newly generated data and the new models are notated as "sRecycled Models", eg. sRecycled Alpaca.

1d28af1a-e6a7-4746-a6e1-e10bd3221810

# Selective Reflection-Tuning: Student-Selected Data Recycling For Llm Instruction-Tuning ## 4 Experimental Setup 4.1 Base Datasets The Alpaca dataset (Taori et al., 2023), sourced from Stanford University, offers 52, 002 instruction samples. Developed via the self-instruct paradigm (Wang et al., 2023d), it leveraged the capabilities of the text-davinci-003 model. The WizardLM dataset (Xu et al., 2023a) is a refined collection encompassing a total of 250, 000 instruction samples. To enhance data fidelity, gpt-3.5-turbo-0613 has been meticulously integrated during the refinement process. From this extensive dataset, we predominantly focused on the WizardLM-7b subset, comprising 70, 000 samples. We test our method on both of these two datasets to verify the effectiveness of our method and name the corresponding models as "sRecycled Alpaca" and "sRecycled WizardLM". | | | | | | Win Rate | Standard Error | Wins | Draws | Avg Length | Data | RLHF/AIF | |-------------------------------|-----------------|------|-------|------|------------|------------------|--------|---------|--------------|---------|------------| | GPT4 ( | OpenAI | , | 2023 | ) | 95.28 | 0.72 | 761 | 12 | 1365 | / | / | | Claude 2 | 91.36 | 0.99 | 734 | 1 | 1069 | /

b848d9f0-0e02-4b5d-b8f0-b5c83e14c824

# Selective Reflection-Tuning: Student-Selected Data Recycling For Llm Instruction-Tuning ## 4 Experimental Setup 4.1 Base Datasets .72 | 761 | 12 | 1365 | / | / | | Claude 2 | 91.36 | 0.99 | 734 | 1 | 1069 | / | / | | | | | | Zephyr 7B Beta ( | Tunstall et al. | , | 2023 | ) | 90.60 | 1.03 | 727 | 1 | 1444 | 774,000 | | | ✓ | | | | | | | | | | | | | ChatGPT | | | |

d719c528-7321-4b40-8bbc-5c73dee56423

# Selective Reflection-Tuning: Student-Selected Data Recycling For Llm Instruction-Tuning ## 4 Experimental Setup 4.1 Base Datasets | | | | | | | ChatGPT | | | | | | | | | | | | | 89.37 | 1.08 | 716 | 5 | 827 | / | / | | | | | | | Evo v2 7B | 89.35 | 1.08 | 715 | 5 | 1754 | / | / | | | | | | XwinLM 7b V0.1 ( | Team

f15d3d6e-7a8c-4d2d-9cad-0c1bcb45fc9c

# Selective Reflection-Tuning: Student-Selected Data Recycling For Llm Instruction-Tuning ## 4 Experimental Setup 4.1 Base Datasets | 1.08 | 715 | 5 | 1754 | / | / | | | | | | XwinLM 7b V0.1 ( | Team | , | 2023 | ) | 87.83 | 1.15 | 703 | 1 | 1894 | / | | | ✓ | | | | | | | | | | | | | sRecycled WizardLM 13B (ours) | 85.96 | 1.23 | 692 | 0 | 1523 | 46,064 | | | | | | | ✗

d49e010c-79c4-435c-b738-425f475bc48e

# Selective Reflection-Tuning: Student-Selected Data Recycling For Llm Instruction-Tuning ## 4 Experimental Setup 4.1 Base Datasets ours) | 85.96 | 1.23 | 692 | 0 | 1523 | 46,064 | | | | | | | ✗ | | | | | | | | | | | | | Zephyr 7B Alpha ( | Tunstall et al. | , | 2023 | ) | 85.76 | 1.23 | 688 | 3 | 1302 | 774,000 | | | ✓ | | | | | | | | | | |

5fd5c6d9-286e-4404-98a9-c647bfb0a887

# Selective Reflection-Tuning: Student-Selected Data Recycling For Llm Instruction-Tuning ## 4 Experimental Setup 4.1 Base Datasets | | | | | | | | | | | | | OpenChat V2 13B ( | Wang et al. | , | 2023a | ) | 84.97 | 1.26 | 683 | 2 | 1564 | 82,600 | | | ✗ | | | | | | | | | | | | | Humpback LLaMa 65B ( | Li et al. | , | 2023d | ) | 83.71 | 1.31 | 672 | 2 | 1269 | 502,133 | | | ✗

5fde18fa-5562-439f-85ca-668344f60683

# Selective Reflection-Tuning: Student-Selected Data Recycling For Llm Instruction-Tuning ## 4 Experimental Setup 4.1 Base Datasets | | Humpback LLaMa 65B ( | Li et al. | , | 2023d | ) | 83.71 | 1.31 | 672 | 2 | 1269 | 502,133 | | | ✗ | | | | | | | | | | | | | UltraLM 13B V2.0 ( | Ding et al. | , | 2023 | ) | | | | | | | | | 80.64 | 1.31 | 673 | 0 | 1399 | 774,000 | | | |

bd9bbd90-2b22-4474-86f8-4cbae08bc820

# Selective Reflection-Tuning: Student-Selected Data Recycling For Llm Instruction-Tuning ## 4 Experimental Setup 4.1 Base Datasets | | | 80.64 | 1.31 | 673 | 0 | 1399 | 774,000 | | | | | | | | ✗ | | | | | | | | | | | | | sRecycled WizardLM 7B (ours) | 83.48 | 1.31 | 672 | 0 | 1583 | 46,325 | | | | | | | ✗ | | | | | |

7153d2c3-5316-451d-a620-842266a08c81

# Selective Reflection-Tuning: Student-Selected Data Recycling For Llm Instruction-Tuning ## 4 Experimental Setup 4.1 Base Datasets | | | | | ✗ | | | | | | | | | | | | | Vicuna 13B v1.3 ( | Chiang et al. | , | 2023 | ) | | | | | | | | | 82.11 | 1.35 | 660 | 2 | 1132 | 125,000 | | | | | | | | ✗ | | |

6bd00537-b4a6-4727-b5ed-3a5deb35890c

# Selective Reflection-Tuning: Student-Selected Data Recycling For Llm Instruction-Tuning ## 4 Experimental Setup 4.1 Base Datasets | | | | | | | ✗ | | | | | | | | | | | | | GPT-3.5 | | | | | | | | | | | | | 81.71 | 1.33 | 642 | 25 | 1018 | / | / | | | | | | | LLaMA2

515b5d9c-e77d-4d07-993e-a0209cca3829

# Selective Reflection-Tuning: Student-Selected Data Recycling For Llm Instruction-Tuning ## 4 Experimental Setup 4.1 Base Datasets | 1.33 | 642 | 25 | 1018 | / | / | | | | | | | LLaMA2 Chat 13B ( | Touvron et al. | , | 2023 | ) | 81.09 | 1.38 | 652 | 0 | 1513 | 27,750 | | | ✓ | | | | | | | | | | | | | UltraLM 13B ( | Ding et al. | , | 2023 | ) | 80.64 | 1.40 | 647 | 1 | 1087 | 774,000 | | | ✗

6c26fd99-e929-426a-ac5b-43e013dae5ee

# Selective Reflection-Tuning: Student-Selected Data Recycling For Llm Instruction-Tuning ## 4 Experimental Setup 4.1 Base Datasets | UltraLM 13B ( | Ding et al. | , | 2023 | ) | 80.64 | 1.40 | 647 | 1 | 1087 | 774,000 | | | ✗ | | | | | | | | | | | | | sRecycled Alpaca 7B (ours) | 79.58 | 1.42 | 639 | 0 | 1353 | 37,114 | | | | | | | ✗ | | | | | | | | |

f1d8e7d3-1c41-4c0e-8521-38a4f5955499

# Selective Reflection-Tuning: Student-Selected Data Recycling For Llm Instruction-Tuning ## 4 Experimental Setup 4.1 Base Datasets | | ✗ | | | | | | | | | | | | | Claude2 Alpaca 13B ( | Chen et al. | , | 2023b | ) | 78.93 | 1.44 | 633 | 0 | 1127 | 52,002 | | | ✗ | | | | | | | | | | | | | Recycled WizardLM 7B | 78.88 | 1.44 | 635 | 0 | 1494 | 70,000 | | |

40eb332e-87b2-404e-845e-46e1065866ef

# Selective Reflection-Tuning: Student-Selected Data Recycling For Llm Instruction-Tuning ## 4 Experimental Setup 4.1 Base Datasets | | | | Recycled WizardLM 7B | 78.88 | 1.44 | 635 | 0 | 1494 | 70,000 | | | | | | | ✗ | | | | | | | | | | | | | Recycled Alpaca 7B | 76.99 | 1.49 | 619 | 0 | 1397 | 52,002 | | | | | | | ✗ | | | | | |

6f976027-5b1d-4329-8aa0-4d3a9bd6ee1f

# Selective Reflection-Tuning: Student-Selected Data Recycling For Llm Instruction-Tuning ## 4 Experimental Setup 4.1 Base Datasets | | | | | ✗ | | | | | | | | | | | | | Vicuna 7B v1.3 ( | Chiang et al. | , | 2023 | ) | | | | | | | | | 76.84 | 1.49 | 614 | 3 | 1110 | 125,000 | | | | | | | | ✗ | |

6ae3fb58-8397-48e6-ac47-641b1f777d7c

# Selective Reflection-Tuning: Student-Selected Data Recycling For Llm Instruction-Tuning ## 4 Experimental Setup 4.1 Base Datasets 000 | | | | | | | | ✗ | | | | | | | | | | | | | WizardLM 13B ( | Xu et al. | , | 2023a | ) | 75.31 | 1.51 | 601 | 9 | 985 | 250,000 | | | ✗ | | | | | | | | | | | | | Guanaco 65B (

ab2bfaa0-2cbf-4c84-8f0e-01ddb2be9fba

# Selective Reflection-Tuning: Student-Selected Data Recycling For Llm Instruction-Tuning ## 4 Experimental Setup 4.1 Base Datasets | | | | | | | | | | | | Guanaco 65B ( | Dettmers et al. | , | 2023 | ) | 71.80 | 1.59 | 578 | 0 | 1249 | 9,850 | | | ✗ | | | | | | | | | | | | | LLaMA2 Chat 7B ( | Touvron et al. | , | 2023 | ) | 71.37 | 1.59 | 574 | 1 | 1479 | 27,750 | | | ✓

f057d648-45a6-4986-b930-dabfa461324b

# Selective Reflection-Tuning: Student-Selected Data Recycling For Llm Instruction-Tuning ## 4 Experimental Setup 4.1 Base Datasets | Touvron et al. | , | 2023 | ) | 71.37 | 1.59 | 574 | 1 | 1479 | 27,750 | | | ✓ | | | | | | | | | | | | | Vicuna 7B ( | Chiang et al. | , | 2023 | ) | 64.41 | 1.69 | 517 | 3 | 1044 | 70,000 | | | ✗ | | | | | | | | | | |

6857ae0c-1d32-4705-919e-66cb878729fe

# Selective Reflection-Tuning: Student-Selected Data Recycling For Llm Instruction-Tuning ## 4 Experimental Setup 4.1 Base Datasets | | | | | | | | | | | | | Davinci003 | 50.00 | 0.00 | 0 | 805 | 307 | / | / | | | | | | LIMA 7B ( | Zhou et al. | , | 2023 | ) | | | | | | | | | 41.29 | 1.74 | 332 | 0 | 1624 | 1,000 | | | | |

f6d45793-caed-48b5-9356-956ee8f5aa6f

# Selective Reflection-Tuning: Student-Selected Data Recycling For Llm Instruction-Tuning ## 4 Experimental Setup 4.1 Base Datasets | | 41.29 | 1.74 | 332 | 0 | 1624 | 1,000 | | | | | | | | ✗ | | | | | | | | | | | | | Alpaca 7B ( | Taori et al. | , | 2023 | ) | 26.46 | 1.54 | 205 | 16 | 396 | 52,002 | | | ✗ | | | | | |

bc2e0650-b428-4281-92c0-befc09210c19

# Selective Reflection-Tuning: Student-Selected Data Recycling For Llm Instruction-Tuning ## 4 Experimental Setup 4.1 Base Datasets 16 | 396 | 52,002 | | | ✗ | | | | | | | | | | | |

65f2c7a1-ea04-49d7-879d-37ae15af6d72

# Selective Reflection-Tuning: Student-Selected Data Recycling For Llm Instruction-Tuning ## 4.2 Evaluation Metric To evaluate the effectiveness of our method, we utilize 4 commonly used automatic evaluation metrics, including (1) **Pair-wise Comparison**, (2) Alpaca Eval, (3) **Open LLM Leaderboard**, and (4) MT- Bench. Besides, additional (5) **Human Study** is also conveyed for the evaluation. 4

d4ee1a51-8444-4371-900a-cf79330dd855

# Selective Reflection-Tuning: Student-Selected Data Recycling For Llm Instruction-Tuning ## 5 Experimental Results 5.1 Main Results For **Pair-wise Comparison**, we compare our sRecycled WizardLM 7B with other classic opensource models by using GPT4 as the judge as shown in Figure 2. Notably, our model outperforms most models by a large margin, regardless of whether they are 7B or 13B, ("LLaMA2 Chat 13B", "Vicuna 13B v1.3"), or whether extra RLHF/AIF is utilized ("LLaMA2 Chat 7B", "Zephyr 7B Alpha"), or whether other data improvement methods are utilized ("Recycled Wiz 7B", "WizardLM Orca 7B"5, "Orca 2 7B"(Mitra et al., 2023)). Alpaca Eval Leaderboard Table 1 delineates the outcomes on the AlpacaE- val Leaderboard in which our models stand out for delivering promising results with a streamlined approach. This comparison provides a direct quantification of a model's capacity for instruction adherence and the intrinsic quality of its output. Remark- Huggingface Open LLM Leaderboard Average ARC HellaSwag MMLU TruthfulQA Data RLHF/AIF Alpaca 7B (Taori et al., 2023) 50.21 42.65 76.91 41.73 39.55 52,002 ✗ WizardLM 7B (Xu et al., 2023a) 54.18 51.60 77.70 42.70 44.70 70,000 ✗ Vicuna 7B v1.3 (Chiang et al., 2023) 55.63 50.43 76.92 48.14 47.01 125,000 ✗ sRecycled Alpaca 7B (ours) 56.05 54.01 78.07 46.69 45.41 37,114 ✗ LLaMA2 Chat 7B (Touvron et al., 2023) 56.34 52.90 78.55 48.32 45.57 27,750 ✓ sRecycled WizardLM 7B (ours) 56.79 54.78 77.86 45.63 48.91

73644e53-d786-4216-8b5d-013062a44bc0

# Selective Reflection-Tuning: Student-Selected Data Recycling For Llm Instruction-Tuning ## 5 Experimental Results 5.1 Main Results 47.01 125,000 ✗ sRecycled Alpaca 7B (ours) 56.05 54.01 78.07 46.69 45.41 37,114 ✗ LLaMA2 Chat 7B (Touvron et al., 2023) 56.34 52.90 78.55 48.32 45.57 27,750 ✓ sRecycled WizardLM 7B (ours) 56.79 54.78 77.86 45.63 48.91 46,325 ✗ Vicuna 13B v1.1 (Chiang et al., 2023) 59.21 52.73 80.14 51.90 52.08 125,000 ✗ LLaMA2 Chat 13B (Touvron et al., 2023) 59.94 59.04 81.94 54.64 44.12 27,750 ✓ Vicuna 13B v1.3 (Chiang et al., 2023) 60.01 54.61 80.41 52.88 52.14 125,000 ✗ sRecycled WizardLM 13B (ours) 60.22 59.73 80.15 55.64 45.37 46,064 ✗ WizardLM 13B 1.0 (Xu et al., 2023a) 60.25 57.20 81.00 52.30 50.50 250,000 ✗ | | | | | Huggingface Open LLM Leaderboard | AlpacaEval | |---------------------------------------|-------|-----------|-------|------------------------------------|--------------| | Average | ARC | HellaSwag | MMLU | TruthfulQA

ffb5400b-8a18-4a17-870e-ba52e86d4377

# Selective Reflection-Tuning: Student-Selected Data Recycling For Llm Instruction-Tuning ## 5 Experimental Results 5.1 Main Results | | | | Huggingface Open LLM Leaderboard | AlpacaEval | |---------------------------------------|-------|-----------|-------|------------------------------------|--------------| | Average | ARC | HellaSwag | MMLU | TruthfulQA | AlpacaEval | | sRecycled WizardLM 7B (2%) (926) | 57.80 | 54.69 | 78.80 | 47.00 | 50.70 | | sRecycled WizardLM 7B (5%) (2,316) | 57.91 | 54.86 | 79.83 | 46.69 | 50.23 | | sRecycled WizardLM 7B (10%) (4,632) | 57.71 | 55.46 | 79.56 | 46.83 | 48.98 | | sRecycled WizardLM 7B (30%) (13,897) | 56.89 | 54.61 | 79.25 | 44.67 | 49.05 | | sRecycled WizardLM 7B (100%) (46,325) | 56.79 |

e34bf1d3-cab1-402c-9141-df3e1f953986

# Selective Reflection-Tuning: Student-Selected Data Recycling For Llm Instruction-Tuning ## 5 Experimental Results 5.1 Main Results | 48.98 | | sRecycled WizardLM 7B (30%) (13,897) | 56.89 | 54.61 | 79.25 | 44.67 | 49.05 | | sRecycled WizardLM 7B (100%) (46,325) | 56.79 | 54.78 | 77.86 | 45.63 | 48.91 | a significant overhead. This reduction in complexity represents a significant advancement in model efficiency, making it a cost-effective and agile solution for real-world applications. The ingenuity of our model lies in its simplicity and effectiveness, proving that with intelligent design less is more. ably, with a win rate that competes closely with heavyweight counterparts, our models achieve this with only instruction tuning on a small amount of our high-quality data. Furthermore, our approach does not rely on additional processes such as RLHF (Ouyang et al., 2022; Bai et al., 2022a) or RLAIF (Bai et al., 2022b; Lee et al., 2023), which demand Table 2 showcases the performance comparison on the Huggingface Open LLM Leaderboard with some related models. Similarly, with only instruction tuning on a small amount of data, our models surpass plenty of the models on the average performances across representative benchmarks. These benchmarks do not directly measure the instruction-following ability or the quality of responses generated by LLMs, but a relatively higher performance on these benchmarks still shows the non-degradation quality of our method. For the **human evaluation**, we compare the responses to given testing instructions between our sRecycled WizardLM 7B model with the original WizardLM 7B model by human evaluators, there are 57/108 wins for our model, 23/108 ties, and 28/108 losses. These results further prove the efficacy of our method in improving the quality of the original data.

085c1f44-0c49-4d51-8908-e5310feb3a05

# Selective Reflection-Tuning: Student-Selected Data Recycling For Llm Instruction-Tuning ## 5.2 Fewer Data Scenario To better illustrate the supreme quality of our sRecycled dataset, we further conduct experiments where only part of the data samples are utilized. Following Li et al. (2023c), we calculate the IFD score of each data sample and select the top kpercent of the data for the instruction tuning. Their performances on the Open LLM Leaderboard and the Alpaca Eval Leaderboard are shown in Table 3 6. Since selecting data by IFD score is an effective method to find a better instruction tuning subset from the overall data set, this consistent decrease in performance on Alpaca Eval indicates the difficulty in finding a subset with higher performances, which further verifies the overall high quality of our selective recycled data. Figure 3 draws the scatters comparing the data used and corresponding performance. It illustrates a striking balance of efficiency and performance achieved by our models. Despite using markedly less data, our models—represented by the distinctive star markers—consistently occupy the upper echelons of the performance spectrum on both the Alpaca Eval benchmark and the open LLM leaderboard. Furthermore, the plots reveal that our models achieve these results without scaling up to the larger data requirements that other models seem to necessitate, as indicated by their position further to the right along the x-axis. The results not only signal superior data quality but also suggest a potential reduction in the computational resources and time required for training, which is crucial for sustainable and scalable AI development. Furthermore, it is astonishing that with less than 1, 000 selective recycled data, our "sRecycled WizardLM 7B (2%) (926)" outperforms most existing 7B models, including LIMA, which is trained with manually curated data samples. This not only verifies LIMA's (Zhou et al., 2023) hypothesis but also pushes it further forward: In addition to humancarefully-crafted instruction tuning data, less than 1, 000 totally automatically generated data can also yield substantial benefits in model alignment and performance.

4ec349d7-9f67-4c75-aa6c-4b981844e77b

# Selective Reflection-Tuning: Student-Selected Data Recycling For Llm Instruction-Tuning ## 6 Ablation Study 6.1 Ablation On Reflection Extensive experiments are conducted on several 7B models as shown in Table 4. We utilize the pair-wise comparison with GPT4 as the judge to measure the performance of different models. Compared with the original WizardLM model, our performance is dramatically better, which directly showcases the supreme capability of our method to increase the data quality. "Reflect on Ins." and "Reflect on Res." represent models that are trained with data reflected merely on instruction or response and no selection process is utilized. Through these comparisons, it can be found that reflection on instruction only improves the data quality a little, while reflection on response improves the data quality more. This phenomenon is reasonable due to the similarity in response distribution between original WizardLM data and WizardLM data reflected on instruction. On the contrary, when the response is reflected, it directly affects the target that LLM needs to fit on, thus directly showing an improvement in the response quality. "Reflect on Ins. + Res." represents the model trained by using reflection-tuning ("Recycled WizardLM 7B") without the selection process, though already having the good capability to follow instructions, our model still outperforms it with less data.

60dc57b1-a16e-41e8-98ae-e1814b90b47a

# Selective Reflection-Tuning: Student-Selected Data Recycling For Llm Instruction-Tuning ## 6.2 Ablation On Selection Moreover, to further verify the effectiveness of our selection mechanism, experiments with different selection methods are conducted shown in Table 4. "Select by Randomness" represents the student model randomly choosing whether to accept improved data. Not only does this model underperform our final model largely, but it also underper- | | Win | Tie | Lose | Win Rate | |----------------------------|-------|-------|--------|------------| | vs. Original WizardLM | 150 | 40 | 28 | 1.56 | | vs. Reflect on Ins. | 143 | 51 | 24 | 1.546 | | vs. Reflect on Res. | 72 | 93 | 53 | 1.087 | | vs. Reflect on Ins. + Res. | 68 | 97 | 53 | 1.069 | | vs. Select by Randomness | 81 | 94 | 43 | 1.174 | | vs. Select by Coherence | 75 | 96 | 47 | 1.128 | | vs. Select by Perplexity | 64 | 99 | 55 | 1.041 | | vs. Select by IFD only | 58 | 107 | 53 | 1.023 | | vs. Select by r-IFD only | 74 | 96 | 48 | 1.119 | forms both "Reflect on Res." and "Reflect on Ins. + Res.". This baseline result indicates that without a proper selection method, the blind mixture of data might harm the

a47d04a2-300b-425a-9dd4-3f4bd6a047ab

# Selective Reflection-Tuning: Student-Selected Data Recycling For Llm Instruction-Tuning ## 6.2 Ablation On Selection | 1.041 | | vs. Select by IFD only | 58 | 107 | 53 | 1.023 | | vs. Select by r-IFD only | 74 | 96 | 48 | 1.119 | forms both "Reflect on Res." and "Reflect on Ins. + Res.". This baseline result indicates that without a proper selection method, the blind mixture of data might harm the model's performance. "Select by Coherence" represents the data selected based on the coherence between instruction and response, which is calculated by cosine similarity of the Sentence-BERT (Reimers and Gurevych, 2019) embeddings. In this setting, the data pairs, whose instruction and response are more related, are more likely to be selected. The performance of this model is slightly better than the random selection model, and still worse than both "Reflect on Res." and "Reflect on Ins. + Res.", indicating the ineffectiveness of this selection method. "Select by Perplexity" represents the student model choosing whether to accept the improved data by whether the perplexity is improved, which is the closest to ours. The performance of this model surpasses both "Reflect on Res." and "Reflect on Ins. + Res.", showing that a selection process can definitely further improve the model's performance, verifying our motivation for adding the selection mechanism. However, this model still underperforms our model, indicating the efficacy of our selection strategy. "Select by IFD only" and "Select by r-IFD only" represent situations where we only utilize IFD or r-IFD scores for student side selection. Utilizing only IFD results in a model that is close to our main model, indicating the usefulness of the IFD score. However, its performance is still lower, indicating the effect of the r-IFD.

30a7d4b3-3004-4ab7-ad58-23cba51e9296

# Selective Reflection-Tuning: Student-Selected Data Recycling For Llm Instruction-Tuning ## 7 Comparison With Related Work Earlier works on instruction tuning focus on creating large, high-quality datasets curated by human experts (Khashabi et al., 2020; Ye et al., 2021; Wei et al., 2022; Wang et al., 2022; Du et al., 2022), time-consuming and labor-intensive. Thus a number of works try to construct instruction-tuning datasets automatically. Self-Instruct (Wang et al., 2023d) utilizes the in-context learning capability of GPT-3 to expand tasks to many diverse instructionresponse pairs. WizardLM (Xu et al., 2023a) applies an evolution methodology to refine and diversify the original instruction data. LaMini-LM (Wu et al., 2024) introduces to generate Top-Fuided instructions based on Wiki data. Peng et al. (2023) utilize GPT4 to generate responses for existing datasets. UltraChat (Ding et al., 2023), establishes various scopes and systematically generates a multitude of instructions within each designated area. Orca (Mitra et al., 2023) directly apply GPT4 to generate reasoning steps for given instructions. SelFee (Ye et al., 2023) utilizes ChatGPT to enhance the response quality. Reflection-Tuning (Li et al., 2023b) improves both the instruction and response sequentially by reflecting on specific criteria. DEITA (Liu et al., 2023a) utilizes ChatGPT to diversify and then select the data. LIFT (Xu et al., 2023b) also tries to utilize ChatGPT/GPT4 to expand and compress the data. All the above works are related to ours by involving a teacher model to improve the instruction data, however, all of them are **teacher-dominating**: Both the generation and selection are all decided by the teacher model and without involving the student. We are the first to introduce the teacherstudent collaboration pipeline and it works fine.

9ae8f685-0ae3-4fa3-9f3d-baf88c897692

# Selective Reflection-Tuning: Student-Selected Data Recycling For Llm Instruction-Tuning ## 8 Conclusion Selective Reflection-Tuning, as proposed in this paper, marks a significant advancement in data improvement for instruction tuning of Large Language Models. By integrating an interactive pipeline between a teacher model and a student model, and utilizing the novel metrics of IFD and reversed-IFD, this approach has demonstrated a marked improvement in the quality and relevance of instruction-tuning datasets. The resulting enhancement in model performance across various benchmarks not only attests to the efficacy of our method but also suggests its potential applicability in broader machine learning contexts.

62f5ea1f-057b-4aab-9481-848e8f136b14

# Selective Reflection-Tuning: Student-Selected Data Recycling For Llm Instruction-Tuning ## Limitations The involvement of the student model makes it possible to build high-quality and student-compatible instruction-response data. However, the main limitation of this method is that the data samples selected by different student models are different, thus the statistics (IFD scores and r-IFD scores) need to be calculated again for different student models. We believe the use of model-specific data samples is more reasonable due to the distinct characteristics of different models, and utilizing the statistics-based method is much more efficient than other generation-based methods, the necessity of re-calculation for new models is still not efficient enough.

2a66376e-09e2-48c1-aece-91954937b2d6

# Selective Reflection-Tuning: Student-Selected Data Recycling For Llm Instruction-Tuning ## A Prompt For Evaluation We provide the detailed prompt we used for the pair-wise comparison in Figure 4.

9a2bbb39-d032-43f9-bb17-3c14c6c185e7

# Selective Reflection-Tuning: Student-Selected Data Recycling For Llm Instruction-Tuning ## Prompt For Performance Evaluation System Prompt You are a helpful and precise assistant for checking the quality of the answer. User Prompt [Question] Question [The Start of Assistant 2's Answer] Answer 2 [The End of Assistant 2's Answer] [The Start of Assistant 2's Answer] Answer 2 [The End of Assistant 2's Answer] We would like to request your feedback on the performance of two AI assistants in response to the user question displayed above. Please rate the helpfulness, relevance, accuracy, level of details of their responses. Each assistant receives an overall score on a scale of 1 to 10, where a higher score indicates better overall performance. Please first output a single line containing only two values indicating the scores for Assistant 1 and 2, respectively. The two scores are separated by a space. In the subsequent line, please provide a comprehensive explanation of your evaluation, avoiding any potential bias and ensuring that the order in which the responses were presented does not affect your judgment.

0737663f-fa2d-49cd-adee-ffbe45152bf2

# Selective Reflection-Tuning: Student-Selected Data Recycling For Llm Instruction-Tuning ## B Prompt For Reflection The prompts for the reflection are shown in Figure 5 and Figure 6.

ece394f7-77fc-42dc-80ff-7b12445456be

# Selective Reflection-Tuning: Student-Selected Data Recycling For Llm Instruction-Tuning ## Prompt For Reflecting Instruction System Prompt You Are A Helpful, Precise But Picky Assistant For Checking The Quality Of A Given Instruction. User Prompt [Instruction] Instruction [The Start of Answer] Answer [The End of Answer] We would like you to answer several questions related to the quality of a given instruction. 1. Why this instruction is not good? First analyze the instruction based on the Complexity of the Topic, Level of Detail Required, Knowledge Required, Ambiguity of the Instruction and Logical Reasoning or Problem-Solving Involved. Then analyze why this answer is not good for the given instruction based on the Helpfulness, Relevance, Accuracy and Level of Details. Finally, analyze why this bad instruction leads to a bad answer. 2. Based on the reason you provided, generate a new and complete instruction that is complex and difficult to answer directly. Make sure the new instruction is relevant but independent to the original instruction, which can be answered without knowing the original instruction, put the new instruction in the format of [New Instruction] your instruction [End] 3. Answer the newly generated instruction as detailed as possible, in the format of [New Answer] your answer [End]

bbee7e7b-8b95-470b-9e89-d3333dede750

# Selective Reflection-Tuning: Student-Selected Data Recycling For Llm Instruction-Tuning ## Prompt For Reflecting Response System Prompt You Are A Helpful, Precise But Picky Assistant For Checking The Quality Of The Answer To A Given Instruction. User Prompt [Instruction] Instruction [The Start of Answer] Answer [The End of Answer] We would like you to answer several questions related to the quality of the answer to the given instruction. 1. Why this answer is not good for the given instruction? Analyze based on the Helpfulness, Relevance, Accuracy, and Level of Details. 2. Based on the reason you provided, generate a better answer, new and complete, as detailed as possible, in the format of [Better Answer] your answer [End]

20bd11a3-f1eb-4c43-8cd3-125bf24acfe8

# Selective Reflection-Tuning: Student-Selected Data Recycling For Llm Instruction-Tuning ## C Evaluation Metric C.1 Pair-Wise Comparison Evaluation of the responses generated by LLMs is an open problem that plenty of researchers are still working on, due to the lack of real ground truth for the open-domain questions, most of the previous methods can not be directly implemented for judging the instruction-following ability of LLMs. However, using LLM as a judge, e.g., GPT4, for evaluation is recently a widely accepted and common practice (Touvron et al., 2023; Chiang et al., 2023; Dettmers et al., 2023; Liu et al., 2023b; Chiang and Lee, 2023). Previous studies (Zheng et al., 2023; Li et al., 2023e) have shown that GPT4's evaluations are consistent with human evaluations. We utilized the testing instruction set from WizardLM (Xu et al., 2023a) which contains 218 diverse human-curated instructions, which are categorized into specific sub-categories. Specifically, we directly follow the evaluation method from Chen et al. (2023a); Li et al. (2023c), which contains rating each model-generated response on a scale spanning from 1 to 10, with scores encapsulating several aspects such as accuracy and relevance. To further mitigate the positional bias elaborated upon in (Ko et al., 2020; Wang et al., 2023b), model-generated outputs are presented to the LLM judge in two distinct sequences and subsequently scored. Hence, a model's dominance is ratified under the following conditions: **Wins:** Exhibits superiority in both sequences or prevails in one while maintaining parity in the alternate sequence. **Tie:** Demonstrates parity across both sequences or prevails in one while faltering in the alternate. **Loses:** Underperforms in both sequences or maintains parity in one while being eclipsed in the alternate.

07481e80-cf1c-461a-9dfd-c86a06b7bd55

# Selective Reflection-Tuning: Student-Selected Data Recycling For Llm Instruction-Tuning ## C.2 Alapca Eval Leaderboard AlpacaEval Leaderboard offers an LLM-centric automatic assessment utilizing the AlpacaFarm (Dubois et al., 2023) evaluation dataset. It is an automated evaluation mechanism for LLMs that offers efficiency, cost-effectiveness, and reliability. Operating on the AlpacaFarm evaluation dataset, it gauges models' proficiency in adhering to generic user instructions. The generated outputs are juxtaposed against benchmark responses from Davinci003. Empirical evidence suggests that AlpacaEval's alignment with ground truth annotations sourced from human experts is notably high.

6f26178f-cd9a-4c49-adc8-3f3518b4ba09

# Selective Reflection-Tuning: Student-Selected Data Recycling For Llm Instruction-Tuning ## C.3 Open Llm Leaderboard The Huggingface Open LLM Leaderboard employs the evaluation methodology from (Gao et al., 2021), providing a cohesive framework for assessing generative language model capabilities across a spectrum of evaluation tasks. It focuses on 4 pivotal benchmarks: ARC (Clark et al., 2018), HellaSwag (Zellers et al., 2019), MMLU (Hendrycks et al., 2021), and TruthfulQA (Lin et al., 2022).

334e6fce-ec4c-4c74-8193-3d6acd5cad89

# Selective Reflection-Tuning: Student-Selected Data Recycling For Llm Instruction-Tuning ## C.4 Mt-Bench We also provide the performances of our sRecycled Models on MT-bench, as shown in Table 5. Since our training focused on 1-turn instructions and did not include any multi-turn data, the 1-turn score on the MT bench is promising and comparable to LLaMA2-13B-chat, while the 2-turn score is not that satisfactory. However, the Vicuna dataset Chiang et al. (2023) can introduce multi-turn dialog data to the model training. Hence, we tried training with our data based on the existing Vicuna 7B v1.5 model, whose result is reported in the last row as "sRecycled Wiz + Vicuna 7B". Compared with the original Vicuna model, the 1-turn, 2-turn, and overall scores are improved dramatically and the overall score is similar to the performance of Vicuna-13B. | | 1-turn | 2-turn | Overall | |---------------------------|----------|----------|-----------| | sRecycled Alpaca 7B | 6.653 | 2.888 | 4.891 | | sRecycled Wiz 7B | 6.538 | 4.588 | 5.563 | | Vicuna 7B v1.5 | 6.569 | 5.588 | 6.078 | | sRecycled Wiz + Vicuna 7B | 7.063 | 5.975 | 6.519 |

7a8416ae-ccc3-4fda-a13a-34cc1bce9c81

# Selective Reflection-Tuning: Student-Selected Data Recycling For Llm Instruction-Tuning ## C.5 Human Study To further validate the superiority of our method, we conducted a further human study to further evaluate the effectiveness of our method. In the test set, there are 27 sub-categories that have 4 or more testing instructions, thus we randomly sampled 4 instructions from each sub-category to form a set containing 108 instructions. Then 3 human participants are given the task of comparing the responses generated by the comparing models with the criteria same as the previous pair-wise evaluation. For each comparison, 3 options are given (Win, Tie, and Loss) and the final results are determined by the majority voting of the participants.

f1a9bdc3-a57a-480e-b01b-4371a4720803

# Selective Reflection-Tuning: Student-Selected Data Recycling For Llm Instruction-Tuning ## D Implementation Details For the Llama2 pre-trained model (Touvron et al., 2023), we utilize the prompt and code base from Vicuna (Chiang et al., 2023) and flash attention (Dao et al., 2022) while the overall training arguments are aligned with protocols from Alpaca and WizardLM datasets. The Adam optimizer (Kingma and Ba, 2017), with a 2×10−5 learning rate for the 7b model and a 1 × 10−5 learning rate for the 13b model, and a batch size of 128, steer the training across three epochs with a max length of 2048. The warmup rate is set to 0.03.

70c9969c-3ef2-43a6-816f-cf453408f106

# Selective Reflection-Tuning: Student-Selected Data Recycling For Llm Instruction-Tuning ## E Statistic Analysis E.1 Basic Data Statistics In this section, we delve into a quantitative analysis of the instruction-response data, pre- and postapplication of our methodology, as delineated in Table 6. We first compare both "Recycled Data" and "sRecycled Data" to the original data. Observationally, there's an increase in the average token length of instructions within the Alpaca dataset, whereas a decrement manifests for the WizardLM dataset, epitomizing the method's adept adaptability. The succinctness and elementary nature of the Alpaca dataset's instructions warrant an enhancement in intricacy through our method, thereby elongating their length. Conversely, the pre-existing complexity and intricacy in WizardLM's instructions render our algorithm inclined towards succinctness. Pertaining to the response section, there's a marked propensity of our approach to engender detail-rich textual content, leading to relatively long responses. Moreover, leveraging Sentence-BERT (Reimers and Gurevych, 2019), we quantify the coherence metric between instructions and their affiliated responses. It's discernible that our technique invariably fabricates samples with better coherence, signifying a superior alignment between modulated instructions and consequent responses. Additionally, to elucidate the metamorphosis in instructional difficulty, we employ the IFD score, executed on the pre-trained llama2-7b language model to check the the difficulties of instructions. The increase in IFD scores represents the increase in the overall difficulty of instructions. Moreover, r-IFD is also calculated, and the decrease in r-IFD scores represents the instruction response pair is more related.

97dc6279-9532-4921-b965-1dcacaf09379

# Selective Reflection-Tuning: Student-Selected Data Recycling For Llm Instruction-Tuning ## E.2 Data Component Distribution In our selective reflection-tuning, there are four different outcomes for each original data sample: both instruction and response are modified, only instruction is modified, only response is modified, and none of instruction and response are modified. Thus to provide a better view of the data conponents, we provide the pie chart for our sRecycled Alpaca 7B and sRecycled Wizardlm 7B data as shown in Figure 7. Comparison of Different Models Ins. len Res. len Ins. ppl Res. ppl 1 Res. ppl 2 Coherent IFD r-IFD Original Alpaca Data 20.7 65.5 34.3 82.6 49.2 0.53 0.71 0.48 Recycled Alpaca Data 37.9 377.2 13.6 4.5 2.9 0.67 0.84 0.50 sRecycled Alpaca Data 31.4 345.9 19.8 4.2 2.8 0.65 0.84 0.36 Original WizardLM Data 123.0 348.5 12.3 17.0 7.5 0.65 0.71 0.52 Recycled WizardLM Data 66.9 518.7 10.0 3.2 2.5 0.73 0.83 0.41 sRecycled WizardLM Data 70.7 519.6 12.0 3.1 2.4 0.72 0.83 0.39

48110e1d-b497-47b5-861e-8ed9dcdfb5e6

# Selective Reflection-Tuning: Student-Selected Data Recycling For Llm Instruction-Tuning ## F Detailed Few Data Scenario The detailed performances in the few data scenarios are shown in TABLE 7 and comparisons with the randomly selected method are shown in TABLE 8. | | | | | Huggingface Open LLM Leaderboard | AlpacaEval | |---------------------------------------|-------|-----------|-------|------------------------------------|--------------| | Average | ARC | HellaSwag | MMLU | TruthfulQA | AlpacaEval | | sRecycled WizardLM 7B (1%) (463) | 57.31 | 54.86 | 78.40 | 46.17 | 49.79 | | sRecycled WizardLM 7B (2%) (926) | 57.80 | 54.69 | 78.80 | 47.00 | 50.70 | | sRecycled WizardLM 7B (3%) (1,390) | 57.34 | 55.12 | 78.80 | 42.68 | 49.16 | | sRecycled WizardLM 7B (5%) (2,316) | | | |

256c477d-2c12-444b-be6e-030f1726802e

# Selective Reflection-Tuning: Student-Selected Data Recycling For Llm Instruction-Tuning ## F Detailed Few Data Scenario cled WizardLM 7B (3%) (1,390) | 57.34 | 55.12 | 78.80 | 42.68 | 49.16 | | sRecycled WizardLM 7B (5%) (2,316) | | | | | | | 57.91 | 54.86 | 79.83 | 46.69 | 50.23 | 77.78 | | sRecycled WizardLM 7B (10%) (4,632) | 57.71 | 55.46 | 79.56 | 46.83 | 48.98 | | sRecycled WizardLM 7B (30%) (13,897) | 56.89 | 54.61 | 79.25 | 44.67 | 49.05 | | sRecycled WizardLM 7B (50%) (23,163) | 56.98 | 55.11 | 78.87 | 45.31 | 48.63

e02c7d59-19ef-4ea8-a822-df3dacd74fba

# Selective Reflection-Tuning: Student-Selected Data Recycling For Llm Instruction-Tuning ## F Detailed Few Data Scenario | 79.25 | 44.67 | 49.05 | | sRecycled WizardLM 7B (50%) (23,163) | 56.98 | 55.11 | 78.87 | 45.31 | 48.63 | | sRecycled WizardLM 7B (70%) (32,428) | 56.63 | 54.95 | 78.55 | 46.31 | 46.71 | | sRecycled WizardLM 7B (100%) (46,325) | 56.79 | 54.78 | 77.86 | 45.63 | 48.91 | | | | | | Huggingface Open LLM Leaderboard | AlpacaEval | |------------------------------------------|-------|-----------|-------|------------------------------------|--------------| | Average | ARC | HellaSwag | MMLU | TruthfulQA | AlpacaEval | | sRecycled Wiz 7B (2%) (926) (IFD) | 57.

3599ccc8-4ba4-4812-b8a8-ac079de1668c

# Selective Reflection-Tuning: Student-Selected Data Recycling For Llm Instruction-Tuning ## F Detailed Few Data Scenario |-----------|-------|------------------------------------|--------------| | Average | ARC | HellaSwag | MMLU | TruthfulQA | AlpacaEval | | sRecycled Wiz 7B (2%) (926) (IFD) | 57.80 | 54.69 | 78.80 | 47.00 | 50.70 | | sRecycled Wiz 7B (2%) (926) (Random) | 56.13 | 54.77 | 78.98 | 43.15 | 47.64 | | sRecycled Wiz 7B (5%) (2,316) (IFD) | 57.91 | 54.86 | 79.83 | 46.69 | 50.23 | | sRecycled Wiz 7B (5%) (2,316) (Random) | 57.07 | 54.10 | 78.97 | 46.54 | 48.67 | | sRecycled Wiz 7B (10%) (4,632) (IFD) | 57.71 | 55.46 | 79.56 | 46.83

448bb6a5-a0f3-4834-81d1-4fe122685b39

# Selective Reflection-Tuning: Student-Selected Data Recycling For Llm Instruction-Tuning ## F Detailed Few Data Scenario 54.10 | 78.97 | 46.54 | 48.67 | | sRecycled Wiz 7B (10%) (4,632) (IFD) | 57.71 | 55.46 | 79.56 | 46.83 | 48.98 | | sRecycled Wiz 7B (10%) (4,632) (Random) | | | | | | | 57.06 | 54.86 | 78.09 | 46.82 | 48.46 | 77.11 | | sRecycled Wiz 7B (30%) (13,897) (IFD) | 56.89 | 54.61 | 79.25 | 44.67 | 49.05 | | sRecycled Wiz 7B (30%) (13,897) (Random) | 56.80 | 54.95 | 78.07 | 47.39 | 46.81 |

65aae590-56b3-48f3-9f44-20302f1f56a4

# Selective Reflection-Tuning: Student-Selected Data Recycling For Llm Instruction-Tuning ## G Ablation On Larger Evaluate Set The evaluation set used on the main page in Table 4 is the WizardLM test set, which contains 218 human-written instructions, and is currently one of the most widely used test sets. Another widely used test set is the Vicuna test set, which is used in MT-Bench, but it contains only 80 instructions and the results are presented in Appendix C. Thus the test set we used for ablation is almost three times the Vicuna set. Moreover, in our evaluation, every comparison will be processed twice to eliminate the potential position bias. Thus we don't think it would be regarded as a really small test set. However, to further validate the effectiveness of our method, we further combine the Vicuna (Chiang et al., 2023) test set (80), Koala (Vu et al., 2023) test set (180), WizardLM (Xu et al., 2023a) test set (218), Self-instruct (Wang et al., 2023d) test set (252), and LIMA (Zhou et al., 2023) test set (300) into a huge evaluation set of 1030 instructions for the ablation study as shown in Table 9. The results on this huge test set share similar trends compared with using the WizardLM test set alone, indicating the effectiveness of our method. | | Win | Tie | Lose | Win Rate | |----------------------------|-------|-------|--------|------------| | vs. Original WizardLM | 738 | 126 | 166 | 1.556 | | vs. Reflect on Ins. | 688 | 196 | 146 | 1.548 | | vs. Reflect on Res. | 360 | 388 | 282 | 1.078 | | vs. Reflect on Ins. + Res. | 329 | 434 | 267 | 1.044 | | vs. Select by Randomness | 401 | 416 |

f6ff7282-7eee-4e37-8b5d-61ff17f8c559

# Selective Reflection-Tuning: Student-Selected Data Recycling For Llm Instruction-Tuning ## G Ablation On Larger Evaluate Set Ins. | 688 | 196 | 146 | 1.548 | | vs. Reflect on Res. | 360 | 388 | 282 | 1.078 | | vs. Reflect on Ins. + Res. | 329 | 434 | 267 | 1.044 | | vs. Select by Randomness | 401 | 416 | 213 | 1.149 | | vs. Select by Coherence | 360 | 444 | 226 | 1.108 | | vs. Select by Perplexity | 307 | 438 | 285 | 1.05 |

855dc4ad-6969-48ef-8a92-66d19454e56e

# Selective Reflection-Tuning: Student-Selected Data Recycling For Llm Instruction-Tuning ## H Examples For R-Ifd Illustrtaion Example 1: (r-IFD=0.148, Low, Prefered) Instruction: Reverse a linked list. Response: def r e v e r s e ( head ) : prev = None cur = head while cur : next = cur . next cur . next = prev prev = cur cur = next head = prev r e t u r n head Example 2: (r-IFD=0.921, High, Not Prefered) Instruction: Identify the type of sentence "I drove to the store yesterday". Response: Declarative sentence. In the first example, after reading through the given code, LLM can easily understand the task and guess what this code is for, indicating sufficient information in the response and its good match to the instruction. However, in the second example, the response is not able to provide enough information to derive the instructions and is vague in various aspects. It indicates that the response might not be feasible to be reasoned by the model and thus needs to be improved.

c42e6cb5-1659-4553-bb92-a9b25fddfd1c

## Large Language Models For Forecasting And Anomaly Detection: A Systematic Literature Review Jing Su AT&T Center for Virtualization Southern Methodist University Dallas, TX, USA suj@smu.edu Xin Jin Department of Computer Science and Engineering Ohio State University Columbus, OH, USA jin.967@osu.edu Tingsong Xiao Department of Computer & Information Science & Engineering University of Florida Gainesville, FL, USA xiaotingsong@ufl.edu Rong Wei Academy for Advanced Interdisciplinary Studies Peking University Beijing, China wei_rong@pku.edu.cn Jiajun Xu Department of Electrical and Computer Engineering University of Southern California Los Angeles, CA, USA jiajunx@usc.edu Junhong Lin Electrical Engineering & Computer Science Department Massachusetts Institute of Technology Cambridge, MA, USA junhong@mit.edu

8cd80522-601d-45dc-a5d2-eb471d56273d

## Abstract This systematic literature review comprehensively examines the application of Large Language Models (LLMs) in forecasting and anomaly detection, highlighting the current state of research, inherent challenges, and prospective future directions. LLMs have demonstrated significant potential in parsing and analyzing extensive datasets to identify patterns, predict future events, and detect anomalous behavior across various domains. However, this review identifies several critical challenges that impede their broader adoption and effectiveness, including the reliance on vast historical datasets, issues with generalizability across different contexts, the phenomenon of model hallucinations, limitations within the models' knowledge boundaries, and the substantial computational resources required. Through detailed analysis, this review discusses potential solutions and strategies to overcome these obstacles, such as integrating multimodal data, advancements in learning methodologies, and emphasizing model explainability and computational efficiency. Moreover, this review outlines critical trends that are likely to shape the evolution of LLMs in these fields, including the push toward real-time Chufeng Jiang Department of Computer Science The University of Texas at Austin Austin, TX, USA chufeng.jiang@utexas.edu Yuxin Qiao Department of Information System Universidad Internacional Isabel I de Castilla Burgos, Spain qiaoyuxin46@icloud.com Hongda Ma Department of Computer Science The University of Texas at Austin Austin, TX, USA hongda.ma@utexas.edu Zhi Jing School of Computer Science Carnegie Mellon University Pittsburgh, PA, USA zjing2@cs.cmu.edu processing, the importance of sustainable modeling practices, and the value of interdisciplinary collaboration. Conclusively, this review underscores the transformative impact LLMs could have on forecasting and anomaly detection while emphasizing the need for continuous innovation, ethical considerations, and practical solutions to realize their full potential.

1626f35b-7ffc-4736-bd6e-2a2f20e1aa8b

## K**Eywords** Large Language Models · Pre-Trained Foundation Models · Time Series · Forecasting · Anomaly Detection 1 Introduction Language represents a rigorously structured communicative system characterized by its grammar and vocabulary. It serves as the principal medium through which humans articulate and convey meaning. This conception of language as a structured communicative tool is pivotal in the realm of computational linguistics, particularly in the development and evaluation of natural language processing (NLP) algorithms. A seminal aspect in this field is the Turing Test, proposed by Alan Turing in 1950 [1], which evaluates a machine's ability to exhibit intelligent behavior equivalent to, or indistinguishable from, that of a human. In this context, the Turing Test primarily assesses the machine's capability to perform tasks involving language comprehension and generation, reflecting the intricate role of linguistic structure in the artificial replication of human-like communication. Language model (LM) is a fundamental element employed in a multitude of NLP tasks, such as text generation, machine translation, and speech recognition [2, 3]. These models are intricately designed to comprehend, generate, and manipulate human language. The training of language models involves large-scale corpora, enabling them to learn universal language representations. This training process is critical for the models to capture the semantics of words in varying contexts [4, 5, 6]. Notably, the fidelity of these representations is frequently contingent on the word frequency within the training corpus. Such dependency underscores the importance of a comprehensive and diverse corpus in training LMs, as it directly impacts their ability to reflect and understand the nuances of natural language accurately. The intricacy of language models and their reliance on corpus characteristics are vital considerations in advancing NLP technologies, which underscores the significance of human-like language comprehension and production in artificial intelligence systems. The forefront of advancements in language model technology has been marked by the emergence of Large Language Models (LLMs). This evolution signifies a paradigm shift in the field of NLP and extends its impact to broader applications. LLMs leverage deep learning methodologies [7], utilizing extensive datasets to perform complex tasks such as understanding, summarizing, generating, and predicting novel content. These models operate by processing an input text and iteratively predicting subsequent tokens or words. A distinguishing feature of LLMs is their vast parameter space, encompassing tens to hundreds of billion parameters, in stark contrast to their predecessors [4, 3]. In addition, they are trained on significantly larger datasets, ranging from several gigabytes to terabytes in size. This exponential increase in both computational capacity and training data volume has not only enhanced the performance of LLMs in conventional NLP tasks but also has

862ef129-77bd-4772-95bf-bebda6d6568c

## K**Eywords** Large Language Models · Pre-Trained Foundation Models · Time Series · Forecasting · Anomaly Detection 1 Introduction leverage deep learning methodologies [7], utilizing extensive datasets to perform complex tasks such as understanding, summarizing, generating, and predicting novel content. These models operate by processing an input text and iteratively predicting subsequent tokens or words. A distinguishing feature of LLMs is their vast parameter space, encompassing tens to hundreds of billion parameters, in stark contrast to their predecessors [4, 3]. In addition, they are trained on significantly larger datasets, ranging from several gigabytes to terabytes in size. This exponential increase in both computational capacity and training data volume has not only enhanced the performance of LLMs in conventional NLP tasks but also has expanded their applicability in areas such as contextual analysis and sentiment detection. The advancements in LLMs reflect the ongoing pursuit of achieving and surpassing human-level proficiency in language understanding and generation. Forecasting and anomaly detection represent pivotal components in the realm of data science, delivering essential insights across a multitude of domains ranging from network security to financial markets [8, 9, 10, 11, 12, 13, 14]. These techniques are integral in projecting forthcoming trends and pinpointing atypical patterns that diverge from normative expectations. Such capabilities are proactive in fostering preemptive strategies in a wide array of applications. Forecasting uses historical data to make informed predictions about future occurrences or trends. It involves making assumptions about the situation being analyzed, selecting an appropriate data set, analyzing the data, and determining the forecast. Forecasting serves as a cornerstone for strategic planning and decision-making in diverse sectors, ranging from economics and finance to healthcare and environmental management, that empowers organizations and policymakers to anticipate changes, manage risks, and allocate resources efficiently [15, 16, 17]. In the financial sector, for instance, accurate forecasting is essential for investment strategies, risk management, and market analysis [18, 19]. It enables investors and financial analysts to predict market trends, assess the viability of investments, and mitigate potential risks. Similarly, in supply chain management, forecasting plays a pivotal role in inventory control, demand planning, and logistics optimization, thus ensuring operational efficiency and cost-effectiveness [20]. Moreover, in the realm of public policy and healthcare, forecasting is critical for preparing for future demands, whether it be anticipating economic shifts, public health needs, or environmental changes [21]. Accurate predictions can guide policy formulation and resource allocation, thereby enhancing the effectiveness of public services and interventions. Anomaly detection, also known as outlier detection, is an analytical process aimed at identifying data points, entities,

9290767a-ccd2-4b12-8f22-0484884fd897

## K**Eywords** Large Language Models · Pre-Trained Foundation Models · Time Series · Forecasting · Anomaly Detection 1 Introduction viability of investments, and mitigate potential risks. Similarly, in supply chain management, forecasting plays a pivotal role in inventory control, demand planning, and logistics optimization, thus ensuring operational efficiency and cost-effectiveness [20]. Moreover, in the realm of public policy and healthcare, forecasting is critical for preparing for future demands, whether it be anticipating economic shifts, public health needs, or environmental changes [21]. Accurate predictions can guide policy formulation and resource allocation, thereby enhancing the effectiveness of public services and interventions. Anomaly detection, also known as outlier detection, is an analytical process aimed at identifying data points, entities, or occurrences that exhibit significant deviations from the typical patterns or norms [22, 23]. This methodology plays a critical role in automated surveillance systems, particularly in identifying potentially detrimental outliers, thereby safeguarding data integrity and security [24]. Its application is especially crucial in sectors such as finance [25], retail [26], and cybersecurity [27, 28]. In the financial industry, anomaly detection is instrumental in fraud detection and anti-money laundering efforts. It enables financial institutions to quickly identify unusual transaction patterns that may indicate fraudulent activity, thereby protecting both the institution and its customers from financial loss [25, 29]. Similarly, in the retail sector, anomaly detection can highlight unusual purchasing patterns or inventory issues, assisting in loss prevention and optimizing supply chain management [26]. The field of cybersecurity significantly benefits from anomaly detection. It is used to identify unusual network traffic, access patterns, or system behavior that could signify a security breach or cyberattack [27, 30]. By detecting these anomalies early, organizations can rapidly respond to potential threats, mitigating the risk of data breaches and cyberattacks [31, 32]. Forecasting and anomaly detection are analytical processes inherently well-suited for time series or timestamped data due to the temporal nature of the information they seek to understand and leverage. Time series data, by definition, is a sequence of data points collected or recorded at time intervals, which often exhibits trends, seasonal variations, and cycles that forecasting techniques aim to capture and extrapolate into the future [24, 33]. Timestamped data is particularly conducive to anomaly detection because it allows for the recognition of deviations from established temporal patterns. For instance, in cybersecurity, anomaly detection systems can identify unusual access patterns that may indicate a security breach [10]. In industrial settings, it might flag an unexpected drop or spike in sensor readings, potentially preventing equipment failure. Figure 1 depicts an overview of leveraging a large language model for forecasting and anomaly detection

8248ed0a-f8c5-4269-9068-11cbd27338f0

## K**Eywords** Large Language Models · Pre-Trained Foundation Models · Time Series · Forecasting · Anomaly Detection 1 Introduction data, by definition, is a sequence of data points collected or recorded at time intervals, which often exhibits trends, seasonal variations, and cycles that forecasting techniques aim to capture and extrapolate into the future [24, 33]. Timestamped data is particularly conducive to anomaly detection because it allows for the recognition of deviations from established temporal patterns. For instance, in cybersecurity, anomaly detection systems can identify unusual access patterns that may indicate a security breach [10]. In industrial settings, it might flag an unexpected drop or spike in sensor readings, potentially preventing equipment failure. Figure 1 depicts an overview of leveraging a large language model for forecasting and anomaly detection tasks. The input data is often time series or timestamped data, encompassing a variety of formats such as text logs, numerical data, structured data, graphical input, and speech recordings. Current widely used LLMs such as BERT [2], GPT [34], LLaMA2 [35], and Mixtral [36] are transformer-based, which includes mechanisms such as input and output embeddings, multi-head attention, and feed-forward neural networks. Forecasting tasks involve predicting future data points based on learned patterns, while anomaly detection identifies outliers or unexpected events in the data stream. In this study, we embark on a *comprehensive exploration* of the integration and potential of LLMs in the realms of forecasting and anomaly detection, areas traditionally dominated by quantitative data analysis. The rapid evolution of LLMs in NLP presents an unprecedented opportunity to augment and possibly revolutionize these domains. This paper aims to bridge the gap between the advanced linguistic processing capabilities of LLMs and the predictive analytics involved in forecasting and detecting outliers. We delve into how the qualitative insights gleaned from LLMs can complement traditional quantitative approaches, thereby enriching the analytical depth and accuracy in various sectors, including finance, cybersecurity, and healthcare. Additionally, this survey addresses the challenges, ethical considerations, and future research trajectories at the intersection of LLMs with these critical data science applications. Our objective is to provide a holistic view that not only elucidates the current state of LLM applications in these fields but also stimulates interdisciplinary dialogue and research, navigating the complexities of modern data environments and paving the way for innovative solutions in predictive analytics. Contributions. To recapitulate, we highlight the following contributions: - To the best of our knowledge, this is the first comprehensive systematic literature review (SLR) dedicated to the application of LLMs in the domains

efff23d5-ae51-42b2-b7dd-cb57417efe22

## K**Eywords** Large Language Models · Pre-Trained Foundation Models · Time Series · Forecasting · Anomaly Detection 1 Introduction this survey addresses the challenges, ethical considerations, and future research trajectories at the intersection of LLMs with these critical data science applications. Our objective is to provide a holistic view that not only elucidates the current state of LLM applications in these fields but also stimulates interdisciplinary dialogue and research, navigating the complexities of modern data environments and paving the way for innovative solutions in predictive analytics. Contributions. To recapitulate, we highlight the following contributions: - To the best of our knowledge, this is the first comprehensive systematic literature review (SLR) dedicated to the application of LLMs in the domains of forecasting and anomaly detection. Through this review, we have elucidated the distinctive influences of LLMs on both numerical and textual data within these specific tasks. - This study compiled a set of guidelines that delineate the optimal utilization of LLMs for various tasks, contributing significantly to the field by providing a structured approach to employing these advanced models in practical scenarios. - This literature review offers, as far as possible, a deep theoretical insight into the capabilities of LLMs, particularly in handling complex patterns and nuances in data that traditional models may overlook. This includes an exploration of the underlying mechanisms that enable LLMs to process and interpret both structured and unstructured data effectively. - This work opens up the enlightenment of new paths for future works around forecasting and anomaly detection modeling. Roadmap The remainder of this paper is organized as follows. Section 2 outlines the methodology employed in conducting the systematic literature review. Section 3 provides an overview of the current state of research on LLMs in forecasting and anomaly detection. Section 4 discusses the challenges and limitations associated with the application of LLMs in these domains. Section 5 explores the datasets and data preprocessing techniques used in LLM-based forecasting and anomaly detection. Section 6 presents the evaluation metrics and methodologies used to assess the performance of LLMs in these tasks. Section 7 delves into the application of LLMs in forecasting, while Section 8 focuses on their application in anomaly detection. Section 9 discusses the potential threats and risks associated with the use of LLMs in these domains. Section 10 outlines the future directions and potential research avenues in the application of LLMs in forecasting and anomaly detection. Section 11 provides an overview of related work, and Section 12 concludes the paper.

83cb2815-e7a1-4ad7-a512-8a95c3f99807

## 2 Methodology In the rapidly evolving domain of artificial intelligence (AI), LLMs have emerged as copilot tools in various applications, notably in forecasting and anomaly detection [10, 23]. However, despite their growing prominence, a substantial knowledge gap exists regarding their comprehensive capabilities and limitations in these contexts. This review is motivated by the necessity to consolidate and critically analyze the extant research concerning LLMs in these specific applications. In light of the rapid progress in model architectures and their diverse applications, this review aims to amalgamate knowledge on existing methodologies, performance evaluation metrics, and practical implementations while also identifying the prevailing challenges and limitations. This effort is imperative for both academic researchers and industry practitioners who aim to utilize these models effectively and serves as a foundational reference for future research and development in this field. By systematically examining and integrating diverse findings from recent studies, this review aims to offer a structured and comprehensive understanding of the current state-of-the-art, thereby guiding informed decision-making and strategic advancements in the application of LLMs. In our study, we adopted the SLR methodology as outlined by Barbara Kitchenham [37, 38]. This method is a comprehensive, rules-driven approach to finding and analyzing prior knowledge on a particular topic that involves a rigorous and transparent methodology to identify, evaluate, and interpret all available research relevant to a particular research question, topic area, or phenomenon of interest. It is designed to provide an exhaustive overview of the current state of research by integrating findings from various studies [39]. The SLR methodology is widely recognized and extensively applied in numerous academic surveys [40, 41, 42, 43, 44]. The research questions (RQs) that guide our SLR process are presented below: RQ1 *What methodologies are employed in LLMs for forecasting in different domains?* Different domains, such as finance, healthcare, weather, and retail, may require unique adaptations of LLMs to address domain-specific challenges and data characteristics. This question aims to explore and categorize the different methodologies and techniques used in LLMs for forecasting tasks, providing insights into their applicability across different sectors. RQ2 *How effective are LLMs in detecting anomalies compared to traditional anomaly detection methods?* Anomalies often exhibit distinct characteristics across diverse contexts, such as outlier financial transactions, atypical network traffic patterns, and unanticipated variations in health data. This question seeks to evaluate the performance and accuracy of LLMs in identifying outliers or unusual patterns in data, contrasting their effectiveness with that of conventional

670490fa-ea99-478d-a884-70d20b9d8272

## 2 Methodology domain-specific challenges and data characteristics. This question aims to explore and categorize the different methodologies and techniques used in LLMs for forecasting tasks, providing insights into their applicability across different sectors. RQ2 *How effective are LLMs in detecting anomalies compared to traditional anomaly detection methods?* Anomalies often exhibit distinct characteristics across diverse contexts, such as outlier financial transactions, atypical network traffic patterns, and unanticipated variations in health data. This question seeks to evaluate the performance and accuracy of LLMs in identifying outliers or unusual patterns in data, contrasting their effectiveness with that of conventional anomaly detection techniques. RQ3 *What are the limitations and challenges of using LLMs for forecasting and anomaly detection?* LLMs present a transformative potential for revolutionizing forecasting and anomaly detection due to their advanced pattern recognition and predictive capabilities. This question intends to identify the current limitations, challenges, and potential areas of improvement in using LLMs for these purposes, including factors like data prerequisites, computational expenditures, and model interpretability. RQ1 calls for a detailed exploration of the strategies, techniques, and models used in applying LLMs across various sectors for predictive purposes. **RQ2** seeks to evaluate and compare the performance of LLMs against conventional techniques in identifying irregularities or unexpected patterns in data. **RQ3** necessitates a comprehensive exploration of the obstacles and constraints faced when employing these advanced models in specific predictive and analytical tasks. After delineating the research questions, we strategically integrate various search engines and databases to identify pertinent studies, as outlined in Table 1. In order to find the most cutting-edge papers, we added OpenReview to the search for forthcoming papers that provide significant insight or data. | Source | Search Scheme | |--------------------------------|---------------------------------------------------------------------------------| | Google Scholar |

72e8d575-7d44-4f2d-85f2-d6a1fe311df1

## 2 Methodology | Search Scheme | |--------------------------------|---------------------------------------------------------------------------------| | Google Scholar | | | (https://scholar.google.com) | Full Text | | Web of Science | | | (https://www.webofscience.com) | TS | TI | AB | AK | KP (Topic, Title, Abstract, Author Keywords, Keywords Plus) | | Scopus |

a9a45efd-bb2b-4c44-9784-224f8c5aaa66

## 2 Methodology com) | TS | TI | AB | AK | KP (Topic, Title, Abstract, Author Keywords, Keywords Plus) | | Scopus | | | (https://www.scopus.com/) | TITLE-ABS-KEY (Title, Abstract, Keywords) | | OpenReview | | | (https://openreview.net) | Keywords | | IEEE Xplore |

3e391c6a-1ada-4857-af7f-e30c6202223a

## 2 Methodology | | IEEE Xplore | | | (https://ieeexplore.ieee.org) | Full Text | | ACM Digital Library | | | (https://dl.acm.org) | Title | | Springer Link |

a1b0325b-d084-4c10-b79c-f115f5cb9788

## 2 Methodology | | Springer Link | | | (https://link.springer.com) | Full Text | After retrieving studies through our established search strategy, we conducted a relevance assessment based on the inclusion and exclusion criteria outlined in Table 2. This process enabled the selection of primary studies offering direct evidence pertinent to the research questions.

f099027e-3d94-4396-8a32-24109cfd3a58

## Inclusion Criteria 1) The paper claims the utilization of LLMs within the context of forecasting or anomaly detection 2) The paper was published within a recent time frame of 3 years (i.e., year ≥ 2020) 3) The paper is peer-reviewed articles, conference proceedings, and academic journals 4) The paper was published in English with accessible full text Exclusion criteria 1) Multiple publications reporting the same research or data 2) Published as a survey or literature review 3) Short visionary paper, tool demo, and editorial 4) Published in a workshop or a doctoral symposium 5) Grey literature, non-peer-reviewed articles, or opinion pieces

39eccff4-cd10-4e79-9814-fc675d251faa

## Table 3: Overview Of Llm-Based Forecastor And Anomaly Detector Research | Research | LLMs | Task | Category | Datasets | Metrics | |------------------------|--------------------------|-------------------|------------------|---------------------|------------| | [45] | GPT-3, GPT-4, Llama2-7b, | | | | | | Llama2-13b, Llama2-70b | | | | | | | Forecasting | Zero-shot | | | | | | Darts, Monash, In- | | | |

7155a8e3-2591-4056-ac33-04e3e2f87d69

## Table 3: Overview Of Llm-Based Forecastor And Anomaly Detector Research | | | | Darts, Monash, In- | | | | | | | former | | | | | | | MAE | | | | | | | [46] | GPT-2, BERT | Forecasting, | | | | | Anomaly Detection |

109c2324-a025-49d0-80c2-bb570351fcf0

## Table 3: Overview Of Llm-Based Forecastor And Anomaly Detector Research | | [46] | GPT-2, BERT | Forecasting, | | | | | Anomaly Detection | | | | | | | Foundation Model | ETT | | | | | | MSE, MAE, MAPE, | | | | | | | sMAPE | | |

6e4d4669-f480-495f-9c73-3c2a6308235e

## Table 3: Overview Of Llm-Based Forecastor And Anomaly Detector Research | | | | | sMAPE | | | | | | | [47] | GPT-3, GPT-3.5, | | | | | | Llama2-13b | | | | | | | Forecasting | Few-shot | ICEWS, | | |

377574a1-566a-48f8-b022-ba2bc5907dac

## Table 3: Overview Of Llm-Based Forecastor And Anomaly Detector Research | | | Forecasting | Few-shot | ICEWS, | | | | | Amazon Review | | | | | | | RMSE | | | | | | | [9] | GPT-2 | Forecasting | Prompt-based | ETT, Weather, | | | Electricity, TETS | | |

54583362-276d-48c0-8ceb-b0c74ef6d37c

## Table 3: Overview Of Llm-Based Forecastor And Anomaly Detector Research PT-2 | Forecasting | Prompt-based | ETT, Weather, | | | Electricity, TETS | | | | | | | MSE, MAE, sMAPE | | | | | | | [48] | BART, BigBird, Pegasus, | | | | | | GPT-3.5 | | | | | | |

f5606bbe-5a62-4922-8029-5e0c2f8768c4

## Table 3: Overview Of Llm-Based Forecastor And Anomaly Detector Research | | GPT-3.5 | | | | | | | Forecasting | Prompt-based | CT, ECL, SG | MAE, RMSE | | | | [10] | BERT | Anomaly Detection | Foundation Model | HDFS, BGL | F1-Score | | [49] | BERT | Anomaly Detection | Foundation Model | HDFS, BGL, | | | Thunderbird | | | | | | | F1-Score, AUROC |

c47b1458-2987-4ad3-8c6d-c16b59ce9caf

## Table 3: Overview Of Llm-Based Forecastor And Anomaly Detector Research | | | | | | | F1-Score, AUROC | | | | | | | [23] | BERT | Anomaly Detection | Fine-tuning | KPI | F1-Score | | [50] | BERT, RoBERTa, XLNet | Forecasting | Fine-tuning, | | | | Foundation Model | | | | | | | SafeGraph

a51ab090-1d39-4287-92da-ee48bb7940a2

## Table 3: Overview Of Llm-Based Forecastor And Anomaly Detector Research | | Foundation Model | | | | | | | SafeGraph | RMSE, MAE | | | | | | [51] | BERT, GPT-2, XLNet | Anomaly Detection | Foundation Model | OpenStack | | | Precision, | Recall, | | | | | | F1-Score | | | | |

dc210df7-9b3c-430f-b0c6-c0934b5b6138

## Table 3: Overview Of Llm-Based Forecastor And Anomaly Detector Research | | | F1-Score | | | | | | | [52] | Llama2-7b | Forecasting | Prompt-based | ETT, Weather, | | | Electricity, | Traffic, | | | | | | ILI, M3, M4 | | | | | | | HDFS, BGL | Precision, | Recall, |

b2b35658-d8e6-4245-a404-e6479422d291

## Table 3: Overview Of Llm-Based Forecastor And Anomaly Detector Research | | | | | HDFS, BGL | Precision, | Recall, | | | | | F1-Score | | | | | | | MSE, MAE, MAPE, | | | | | | | sMAPE, | MASE, | | | |

ac73ebc4-9be4-4286-bc9c-ee4fb3855fc4

## Table 3: Overview Of Llm-Based Forecastor And Anomaly Detector Research | | | sMAPE, | MASE, | | | | | | OWA | | | | | | | [53] | BERT | Forecasting | Fine-tuning | SMD | MSE | | [54] | BERT | Anomaly Detection | Prompt-based | HDFS, BGL | Precision, | | F1-Score, Accuracy | | | |