Add Batch a4db5a27-d6bb-4741-9653-952ed2467977 data

.gitattributes

@@ -1141,3 +1141,67 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 2025/Seeing[[:space:]]What[[:space:]]Tastes[[:space:]]Good_[[:space:]]Revisiting[[:space:]]Multimodal[[:space:]]Distributional[[:space:]]Semantics[[:space:]]in[[:space:]]the[[:space:]]Billion[[:space:]]Parameter[[:space:]]Era/8cff9684-df1a-438c-96df-7de5f52d4fbd_origin.pdf filter=lfs diff=lfs merge=lfs -text
 2025/Select,[[:space:]]Read,[[:space:]]and[[:space:]]Write_[[:space:]]A[[:space:]]Multi-Agent[[:space:]]Framework[[:space:]]of[[:space:]]Full-Text-based[[:space:]]Related[[:space:]]Work[[:space:]]Generation/674a5013-9fec-4f68-9933-9d7480204f9d_origin.pdf filter=lfs diff=lfs merge=lfs -text
 2025/SelectLLM_[[:space:]]Query-Aware[[:space:]]Efficient[[:space:]]Selection[[:space:]]Algorithm[[:space:]]for[[:space:]]Large[[:space:]]Language[[:space:]]Models/0329da27-2aae-459f-98b9-c88042f8faed_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Selecting[[:space:]]Demonstrations[[:space:]]for[[:space:]]Many-Shot[[:space:]]In-Context[[:space:]]Learning[[:space:]]via[[:space:]]Gradient[[:space:]]Matching/094233a6-61cb-440f-8f13-bfe022f362c3_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Self-Correction[[:space:]]is[[:space:]]More[[:space:]]than[[:space:]]Refinement_[[:space:]]A[[:space:]]Learning[[:space:]]Framework[[:space:]]for[[:space:]]Visual[[:space:]]and[[:space:]]Language[[:space:]]Reasoning[[:space:]]Tasks/bd133b59-8ca8-42c4-b305-e0aca46a0680_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Self-Critique[[:space:]]Guided[[:space:]]Iterative[[:space:]]Reasoning[[:space:]]for[[:space:]]Multi-hop[[:space:]]Question[[:space:]]Answering/7f183d00-6ecb-41f2-84cb-009819c2329b_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Self-Foveate_[[:space:]]Enhancing[[:space:]]Diversity[[:space:]]and[[:space:]]Difficulty[[:space:]]of[[:space:]]Synthesized[[:space:]]Instructions[[:space:]]from[[:space:]]Unsupervised[[:space:]]Text[[:space:]]via[[:space:]]Multi-Level[[:space:]]Foveation/e85236e0-99f2-4fdd-a4f5-ddd5a090bab5_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Self-Improvement[[:space:]]Towards[[:space:]]Pareto[[:space:]]Optimality_[[:space:]]Mitigating[[:space:]]Preference[[:space:]]Conflicts[[:space:]]in[[:space:]]Multi-Objective[[:space:]]Alignment/911edb3a-c6cd-4b51-a4d2-8be2982ec5b3_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Self-Reasoning[[:space:]]Language[[:space:]]Models_[[:space:]]Unfold[[:space:]]Hidden[[:space:]]Reasoning[[:space:]]Chains[[:space:]]with[[:space:]]Few[[:space:]]Reasoning[[:space:]]Catalyst/88dc33aa-5afa-4469-b0d2-11e0e6728f43_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Self-Rewarding[[:space:]]Large[[:space:]]Vision-Language[[:space:]]Models[[:space:]]for[[:space:]]Optimizing[[:space:]]Prompts[[:space:]]in[[:space:]]Text-to-Image[[:space:]]Generation/760dd0a6-270e-46cf-a0b7-a4749d840a76_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Self-Steering[[:space:]]Optimization_[[:space:]]Autonomous[[:space:]]Preference[[:space:]]Optimization[[:space:]]for[[:space:]]Large[[:space:]]Language[[:space:]]Models/ddcdb056-3a5c-49ea-857d-7c720881ce91_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Self-Training[[:space:]]Elicits[[:space:]]Concise[[:space:]]Reasoning[[:space:]]in[[:space:]]Large[[:space:]]Language[[:space:]]Models/0687e2cc-b5c4-469e-ba7d-68d87dd25ca7_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Self-Tuning_[[:space:]]Instructing[[:space:]]LLMs[[:space:]]to[[:space:]]Effectively[[:space:]]Acquire[[:space:]]New[[:space:]]Knowledge[[:space:]]through[[:space:]]Self-Teaching/f74734df-603c-47dc-8356-48465b1ee177_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Self-attention-based[[:space:]]Graph-of-Thought[[:space:]]for[[:space:]]Math[[:space:]]Problem[[:space:]]Solving/c198f179-7072-478d-902e-9b7a6010ae9d_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Self-play[[:space:]]through[[:space:]]Computational[[:space:]]Runtimes[[:space:]]improves[[:space:]]Chart[[:space:]]Reasoning/441417f4-94f2-44f2-9ddd-3ab2b4853701_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Semantic[[:space:]]Aware[[:space:]]Linear[[:space:]]Transfer[[:space:]]by[[:space:]]Recycling[[:space:]]Pre-trained[[:space:]]Language[[:space:]]Models[[:space:]]for[[:space:]]Cross-lingual[[:space:]]Transfer/5d112dc8-2179-49b8-aaa9-4f342b4a2e9a_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Semantic[[:space:]]Evaluation[[:space:]]of[[:space:]]Multilingual[[:space:]]Data-to-Text[[:space:]]Generation[[:space:]]via[[:space:]]NLI[[:space:]]Fine-Tuning_[[:space:]]Precision,[[:space:]]Recall[[:space:]]and[[:space:]]F1[[:space:]]scores/8c079f80-a80e-4bd4-8503-b0cec632d65e_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Semantic[[:space:]]Topology_[[:space:]]a[[:space:]]New[[:space:]]Perspective[[:space:]]for[[:space:]]Communication[[:space:]]Style[[:space:]]Characterization/34ffe928-5ebe-4918-af40-7f12200db114_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/SemanticCamo_[[:space:]]Jailbreaking[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]through[[:space:]]Semantic[[:space:]]Camouflage/6a84b0a4-2198-4c11-9d1c-f6da51c44977_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Semantics-aware[[:space:]]prompting[[:space:]]for[[:space:]]translating[[:space:]]NOtices[[:space:]]To[[:space:]]AirMen/ae07962d-2a1d-45f1-b1fc-28c9ae7f236e_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Sens-Merging_[[:space:]]Sensitivity-Guided[[:space:]]Parameter[[:space:]]Balancing[[:space:]]for[[:space:]]Merging[[:space:]]Large[[:space:]]Language[[:space:]]Models/ad84c194-c3af-449f-8d4e-45b99a2f9e18_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Sentimental[[:space:]]Image[[:space:]]Generation[[:space:]]for[[:space:]]Aspect-based[[:space:]]Sentiment[[:space:]]Analysis/755bd5fb-975a-40d3-acab-6add52ca8538_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Separate[[:space:]]the[[:space:]]Wheat[[:space:]]from[[:space:]]the[[:space:]]Chaff_[[:space:]]A[[:space:]]Post-Hoc[[:space:]]Approach[[:space:]]to[[:space:]]Safety[[:space:]]Re-Alignment[[:space:]]for[[:space:]]Fine-Tuned[[:space:]]Language[[:space:]]Models/4b43c337-f9c4-4b99-98ad-60dfe77889a0_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/SeqMMR_[[:space:]]Sequential[[:space:]]Model[[:space:]]Merging[[:space:]]and[[:space:]]LLM[[:space:]]Routing[[:space:]]for[[:space:]]Enhanced[[:space:]]Batched[[:space:]]Sequential[[:space:]]Knowledge[[:space:]]Editing/4dfb0a39-d2b3-442b-be7a-d06dd2729a56_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/SeqPO-SiMT_[[:space:]]Sequential[[:space:]]Policy[[:space:]]Optimization[[:space:]]for[[:space:]]Simultaneous[[:space:]]Machine[[:space:]]Translation/0f0895da-09d6-4b7b-8765-fa019ccd9c6b_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Serial[[:space:]]Position[[:space:]]Effects[[:space:]]of[[:space:]]Large[[:space:]]Language[[:space:]]Models/11759c13-3771-46dc-ab76-22e7be698123_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Shadow-Activated[[:space:]]Backdoor[[:space:]]Attacks[[:space:]]on[[:space:]]Multimodal[[:space:]]Large[[:space:]]Language[[:space:]]Models/78695b1a-aa74-425d-a921-13ede7623e0e_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/ShieldHead_[[:space:]]Decoding-time[[:space:]]Safeguard[[:space:]]for[[:space:]]Large[[:space:]]Language[[:space:]]Models/cb73ed1d-65f7-46b0-97ef-7c1ebb1f90ec_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/ShortGPT_[[:space:]]Layers[[:space:]]in[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]are[[:space:]]More[[:space:]]Redundant[[:space:]]Than[[:space:]]You[[:space:]]Expect/1ef8658e-0346-455f-b226-aca4d44f9c07_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Should[[:space:]]I[[:space:]]Trust[[:space:]]You_[[:space:]]Detecting[[:space:]]Deception[[:space:]]in[[:space:]]Negotiations[[:space:]]using[[:space:]]Counterfactual[[:space:]]RL/2a5893ce-e4c7-4d90-92ef-3219fab569ff_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Sign2Vis_[[:space:]]Automated[[:space:]]Data[[:space:]]Visualization[[:space:]]from[[:space:]]Sign[[:space:]]Language/cc02c583-5d83-442d-96e8-32ae9c2f8da4_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/SignAlignLM_[[:space:]]Integrating[[:space:]]Multimodal[[:space:]]Sign[[:space:]]Language[[:space:]]Processing[[:space:]]into[[:space:]]Large[[:space:]]Language[[:space:]]Models/827a1a29-00ea-4867-b64a-46da827fd04a_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/SignMusketeers_[[:space:]]An[[:space:]]Efficient[[:space:]]Multi-Stream[[:space:]]Approach[[:space:]]for[[:space:]]Sign[[:space:]]Language[[:space:]]Translation[[:space:]]at[[:space:]]Scale/4016a80a-a7e8-4750-8104-1d9cdb5dd98f_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/SimGRAG_[[:space:]]Leveraging[[:space:]]Similar[[:space:]]Subgraphs[[:space:]]for[[:space:]]Knowledge[[:space:]]Graphs[[:space:]]Driven[[:space:]]Retrieval-Augmented[[:space:]]Generation/4984dbf7-e5cb-44f9-9c5c-36bc1920ddda_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/SkyLLM_[[:space:]]Cross-LLM-APIs[[:space:]]Federation[[:space:]]for[[:space:]]Cost-effective[[:space:]]Query[[:space:]]Processing/8551526b-dcac-4259-848a-b032eb1ec090_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Slamming_[[:space:]]Training[[:space:]]a[[:space:]]Speech[[:space:]]Language[[:space:]]Model[[:space:]]on[[:space:]]One[[:space:]]GPU[[:space:]]in[[:space:]]a[[:space:]]Day/e57742c7-3ec2-41de-ae75-003d3ec86eb3_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Sleepless[[:space:]]Nights,[[:space:]]Sugary[[:space:]]Days_[[:space:]]Creating[[:space:]]Synthetic[[:space:]]Users[[:space:]]with[[:space:]]Health[[:space:]]Conditions[[:space:]]for[[:space:]]Realistic[[:space:]]Coaching[[:space:]]Agent[[:space:]]Interactions/9de3cc0a-eb41-42e9-a8b2-b6f2b0aee9b5_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Small[[:space:]]Encoders[[:space:]]Can[[:space:]]Rival[[:space:]]Large[[:space:]]Decoders[[:space:]]in[[:space:]]Detecting[[:space:]]Groundedness/8d1fc10a-6718-4025-9936-de0d888ecce5_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Small[[:space:]]Models[[:space:]]Struggle[[:space:]]to[[:space:]]Learn[[:space:]]from[[:space:]]Strong[[:space:]]Reasoners/79ef0cb8-ae83-4b1f-80ac-1285e6f832d0_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Smarter,[[:space:]]Not[[:space:]]Harder_[[:space:]]Training-Free[[:space:]]Adaptive[[:space:]]Computation[[:space:]]for[[:space:]]Transformers/0989546c-bef6-4745-964b-a71323ea7d83_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Smotrom[[:space:]]tvoja[[:space:]]på[[:space:]]ander[[:space:]]drogoj[[:space:]]verden![[:space:]]Resurrecting[[:space:]]Dead[[:space:]]Pidgin[[:space:]]with[[:space:]]Generative[[:space:]]Models_[[:space:]]Russenorsk[[:space:]]Case[[:space:]]Study/3e128473-de28-4e10-8af9-c2ea32737fe4_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Social[[:space:]]Bias[[:space:]]Benchmark[[:space:]]for[[:space:]]Generation_[[:space:]]A[[:space:]]Comparison[[:space:]]of[[:space:]]Generation[[:space:]]and[[:space:]]QA-Based[[:space:]]Evaluations/f6bad98f-73ca-4b01-b94a-1289848d4bdd_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Socratic[[:space:]]Style[[:space:]]Chain-of-Thoughts[[:space:]]Help[[:space:]]LLMs[[:space:]]to[[:space:]]be[[:space:]]a[[:space:]]Better[[:space:]]Reasoner/3ecb39cf-2d43-410d-977c-b3ce010f2eb2_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Span-based[[:space:]]Semantic[[:space:]]Role[[:space:]]Labeling[[:space:]]as[[:space:]]Lexicalized[[:space:]]Constituency[[:space:]]Tree[[:space:]]Parsing/84964cf2-4ec1-45b9-9ddf-685510c281c7_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Sparse[[:space:]]Rewards[[:space:]]Can[[:space:]]Self-Train[[:space:]]Dialogue[[:space:]]Agents/892c012b-4d1b-4ab9-9715-cefece44816a_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Spatial[[:space:]]Coordinates[[:space:]]as[[:space:]]a[[:space:]]Cell[[:space:]]Language_[[:space:]]A[[:space:]]Multi-Sentence[[:space:]]Framework[[:space:]]for[[:space:]]Imaging[[:space:]]Mass[[:space:]]Cytometry[[:space:]]Analysis/40870fe7-2125-4b63-b5e9-d96bfb8d0708_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Spectral[[:space:]]Insights[[:space:]]into[[:space:]]Data-Oblivious[[:space:]]Critical[[:space:]]Layers[[:space:]]in[[:space:]]Large[[:space:]]Language[[:space:]]Models/02a17d0e-556d-4a62-b8c5-6d35710c1b62_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Speculative[[:space:]]Sampling[[:space:]]via[[:space:]]Exponential[[:space:]]Races/f8aa87a8-5ff5-4e8d-9da9-e4fb6d8704fb_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Speech[[:space:]]Act[[:space:]]Patterns[[:space:]]for[[:space:]]Improving[[:space:]]Generalizability[[:space:]]of[[:space:]]Explainable[[:space:]]Politeness[[:space:]]Detection[[:space:]]Models/9f8c4d0c-0321-4915-a66a-a3988778377f_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/SpeechT-RAG_[[:space:]]Reliable[[:space:]]Depression[[:space:]]Detection[[:space:]]in[[:space:]]LLMs[[:space:]]with[[:space:]]Retrieval-Augmented[[:space:]]Generation[[:space:]]Using[[:space:]]Speech[[:space:]]Timing[[:space:]]Information/f2111ab2-b179-47bd-9ca5-338f67d6b10d_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Splintering[[:space:]]Nonconcatenative[[:space:]]Languages[[:space:]]for[[:space:]]Better[[:space:]]Tokenization/20738a63-4542-4789-bb4d-6c1ed742ccee_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Spotting[[:space:]]Out-of-Character[[:space:]]Behavior_[[:space:]]Atomic-Level[[:space:]]Evaluation[[:space:]]of[[:space:]]Persona[[:space:]]Fidelity[[:space:]]in[[:space:]]Open-Ended[[:space:]]Generation/180ef296-a757-4217-a80c-b456ccc630fc_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/StableToolBench-MirrorAPI_[[:space:]]Modeling[[:space:]]Tool[[:space:]]Environments[[:space:]]as[[:space:]]Mirrors[[:space:]]of[[:space:]]7,000+[[:space:]]Real-World[[:space:]]APIs/1528e7e6-0611-4451-ba4e-b9416c5dc05c_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Standard[[:space:]]Quality[[:space:]]Criteria[[:space:]]Derived[[:space:]]from[[:space:]]Current[[:space:]]NLP[[:space:]]Evaluations[[:space:]]for[[:space:]]Guiding[[:space:]]Evaluation[[:space:]]Design[[:space:]]and[[:space:]]Grounding[[:space:]]Comparability[[:space:]]and[[:space:]]AI[[:space:]]Compliance[[:space:]]Assessments/8cdd1fbc-e42c-4e17-95ff-0ff2ad2787cd_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Statement-Tuning[[:space:]]Enables[[:space:]]Efficient[[:space:]]Cross-lingual[[:space:]]Generalization[[:space:]]in[[:space:]]Encoder-only[[:space:]]Models/4d79aabe-844f-4134-a4c2-804ef406c779_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Statistical[[:space:]]inference[[:space:]]on[[:space:]]black-box[[:space:]]generative[[:space:]]models[[:space:]]in[[:space:]]the[[:space:]]data[[:space:]]kernel[[:space:]]perspective[[:space:]]space/6ae067ba-ac5f-414a-87bc-c4c296638119_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Step-by-Step[[:space:]]Mastery_[[:space:]]Enhancing[[:space:]]Soft[[:space:]]Constraint[[:space:]]Following[[:space:]]Ability[[:space:]]of[[:space:]]Large[[:space:]]Language[[:space:]]Models/c168f7cb-e775-4a31-aaff-ed371a950028_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Stepwise[[:space:]]Perplexity-Guided[[:space:]]Refinement[[:space:]]for[[:space:]]Efficient[[:space:]]Chain-of-Thought[[:space:]]Reasoning[[:space:]]in[[:space:]]Large[[:space:]]Language[[:space:]]Models/3d69ec6e-8d70-4527-8a2c-d8ae0a0ebc96_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Stereotype[[:space:]]Detection[[:space:]]as[[:space:]]a[[:space:]]Catalyst[[:space:]]for[[:space:]]Enhanced[[:space:]]Bias[[:space:]]Detection_[[:space:]]A[[:space:]]Multi-Task[[:space:]]Learning[[:space:]]Approach/f768e3d3-6b63-450e-bec7-e66dc3285314_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Stereotype[[:space:]]or[[:space:]]Personalization_[[:space:]]User[[:space:]]Identity[[:space:]]Biases[[:space:]]Chatbot[[:space:]]Recommendations/e226a99a-fed8-4ef3-99e7-fd0c406d05fa_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Stories[[:space:]]that[[:space:]](are)[[:space:]]Move(d[[:space:]]by)[[:space:]]Markets_[[:space:]]A[[:space:]]Causal[[:space:]]Exploration[[:space:]]of[[:space:]]Market[[:space:]]Shocks[[:space:]]and[[:space:]]Semantic[[:space:]]Shifts[[:space:]]across[[:space:]]Different[[:space:]]Partisan[[:space:]]Groups/4aee5f76-4fec-4ac0-bae7-821c8e23e940_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Streamlining[[:space:]]the[[:space:]]Collaborative[[:space:]]Chain[[:space:]]of[[:space:]]Models[[:space:]]into[[:space:]]A[[:space:]]Single[[:space:]]Forward[[:space:]]Pass[[:space:]]in[[:space:]]Generation-Based[[:space:]]Tasks/2f69655c-1dcb-444d-9738-a72dc8e9c19c_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Stress-testing[[:space:]]Machine[[:space:]]Generated[[:space:]]Text[[:space:]]Detection_[[:space:]]Shifting[[:space:]]Language[[:space:]]Models[[:space:]]Writing[[:space:]]Style[[:space:]]to[[:space:]]Fool[[:space:]]Detectors/2442d6ef-ae89-47a9-b143-b9fd8e941937_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/StructFact_[[:space:]]Reasoning[[:space:]]Factual[[:space:]]Knowledge[[:space:]]from[[:space:]]Structured[[:space:]]Data[[:space:]]with[[:space:]]Large[[:space:]]Language[[:space:]]Models/b8f3dfe5-088a-40fd-ba31-f51b919f2656_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/StructFlowBench_[[:space:]]A[[:space:]]Structured[[:space:]]Flow[[:space:]]Benchmark[[:space:]]for[[:space:]]Multi-turn[[:space:]]Instruction[[:space:]]Following/cddcfbb7-e86b-408d-8abd-b1c049201266_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Structural[[:space:]]Deep[[:space:]]Encoding[[:space:]]for[[:space:]]Table[[:space:]]Question[[:space:]]Answering/4e98758f-0cae-4995-b63a-86194041e35b_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Structure-adaptive[[:space:]]Adversarial[[:space:]]Contrastive[[:space:]]Learning[[:space:]]for[[:space:]]Multi-Domain[[:space:]]Fake[[:space:]]News[[:space:]]Detection/72ab15ef-5894-4a0d-af20-30b8bbe86ded_origin.pdf filter=lfs diff=lfs merge=lfs -text

2025/Selecting Demonstrations for Many-Shot In-Context Learning via Gradient Matching/094233a6-61cb-440f-8f13-bfe022f362c3_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6c1ed150a54e928177a12d2ffddc2c53ba5ec412f2b65287338a099a5880d7f2
+size 2326449

2025/Selecting Demonstrations for Many-Shot In-Context Learning via Gradient Matching/full.md

@@ -0,0 +1,498 @@
+# Selecting Demonstrations for Many-Shot In-Context Learning via Gradient Matching
+
+Jianfei Zhang $^{1}$ , Bei Li $^{2}$ , Jun Bai $^{4}$ , Rumei Li $^{1,2}$ , Yanmeng Wang $^{3}$ , Chenghua Lin $^{5}$ , Wenge Rong $^{1}$
+
+<sup>1</sup>School of Computer Science and Engineering, Beihang University, China
+
+$^{2}$ Meituan, Inc., China  $^{3}$ Ping An Technology, China
+
+<sup>4</sup>Beijing Institute for General Artificial Intelligence, China
+
+$^{5}$ Department of Computer Science, The University of Manchester, UK
+
+{zhangjf, lirumei3232, w.rong}@buaa.edu.cn, libei17@meituan.com
+
+baijun@bigai.ai, wangyanmeng219@pingan.com.cn, chenghua.lin@manchester.ac.uk
+
+# Abstract
+
+In-Context Learning (ICL) empowers Large Language Models (LLMs) for rapid task adaptation without Fine-Tuning (FT), but its reliance on demonstration selection remains a critical challenge. While many-shot ICL shows promising performance through scaled demonstrations, the selection method for many-shot demonstrations remains limited to random selection in existing work. Since the conventional instance-level retrieval is not suitable for many-shot scenarios, we hypothesize that the data requirements for in-context learning and fine-tuning are analogous. To this end, we introduce a novel gradient matching approach that selects demonstrations by aligning fine-tuning gradients between the entire training set of the target task and the selected examples, so as to approach the learning effect on the entire training set within the selected examples. Through gradient matching on relatively small models, e.g., Qwen2.5-3B or Llama3-8B, our method consistently outperforms random selection on larger LLMs from 4-shot to 128-shot scenarios across 9 diverse datasets. For instance, it surpasses random selection by  $4\%$  on Qwen2.5-72B and Llama3-70B, and by around  $2\%$  on 5 closed-source LLMs. This work unlocks more reliable and effective many-shot ICL, paving the way for its broader application.
+
+# 1 Introduction
+
+In-Context Learning (ICL) enables pre-trained Large Language Models (LLMs) to perform tasks by learning from input-output examples (or "demonstrations") provided during inference (Brown et al., 2020). This allows LLMs to adapt to new tasks through forward propagation, without weight updates via back-propagation, offering a flexible alternative to traditional fine-tuning.
+
+Since the performance of ICL is often sensitive to the choice of demonstrations (Liu et al., 2022), significant efforts have been made to im
+
+![](images/1996318b6ef9f40174788f88e81ae21900cd81a96daca9be317e6bb4ae6a37ba.jpg)
+
+![](images/45c5433ac89e6b363502c0c3d658cd223972c4a2157b73547d805f08c7ecfc01.jpg)  
+Figure 1: Instance-level retrieved demonstrations are not suitable for Many-Shot ICL, since the variable context cannot utilize prefix-caching and can lead to theoretically  $O(n^{2})$  inference complexity (Vaswani et al., 2017) when using  $n$ -shot demonstrations. Here we illustrate the runtime of Llama3-70B (Dubey et al., 2024) on CM-SQA (Talmor et al., 2019) using vLLM with 4 A100 GPUs in tensor-parallel (Kwon et al., 2023).
+
+prove demonstration selection. Most studies concentrate on instance-level retrieval, aiming to identify suitable demonstrations for each test query independently (Luo et al., 2023; Rubin et al., 2022). This mainly involves considering similarity (Luo et al., 2023; Rubin et al., 2022) between demonstrations and the query, as well as auxiliary factors such as complexity (Fu et al., 2023), perplexity (Gonen et al., 2023), difficulty (Drozdov et al., 2023), and diversity (Li and Qiu, 2023). Another line of work focuses on task-level selection, which seeks to find a fixed set of demonstrations that achieves the best average performance across all test queries from a target task (Zhang et al., 2022; Wang et al., 2023). Current approaches primarily explore reinforcement learning on the selection policy (Zhang et al., 2022) and Bayesian inference to explain the demonstration effectiveness (Wang et al., 2023) as
+
+potential solutions.
+
+However, for the recently emerged Many-Shot in-context learning (Agarwal et al., 2024) paradigm, demonstrations are selected simply by random in existing work (Agarwal et al., 2024; Bertsch et al., 2024; Song et al., 2025; Jiang et al., 2024), which may lead to suboptimal performance. Nevertheless, existing methods for few-shot demonstration selection are not well-suited to many-shot scenarios. On one hand, applying instance-level demonstrations conflicts with caching and reusing the hidden states of the same long context in language models (Pope et al., 2023), leading to theoretically  $O(n^{2})$  inference complexity (Vaswani et al., 2017) for  $n$ -shot demonstrations, as illustrated in Fig. 1. On the other hand, existing task-level selection methods are designed for very few-shot scenarios (e.g., 4-shot in their implementations), facing challenges from exploration complexity (Zhang et al., 2022) and performance early saturation (Wang et al., 2023) for more shots. To further release the potential of many-shot ICL, it now requires a selection method with scalable effectiveness to many-shot scenarios.
+
+To address such research gap, we revisit ICL demonstration selection from a "learning" perspective—we hypothesize that the data requirements for in-context learning and fine-tuning are analogous. Building on this premise, we conduct latent concept learning (Wang et al., 2023) on a small Language Model (LM) to learn the optimal in-context task embeddings. Then we compute the latent gradient that each example provides to the learning process. Finally, we select  $n$ -shot demonstrations from the whole fine-tuning set with the minimized  $L_{2}$  distance between their average latent gradients, so as to approach the learning effect on all examples within the selected  $n$ -shot demonstrations.
+
+We validate our proposed method on 9 datasets from 5 distinct NLP tasks, each covering scenarios from 4-shot to 128-shot. Our method selects demonstrations through gradient matching on relatively small models, e.g., Qwen2.5-3B or Llama3-8B, and surpasses the widely-adopted random selection by an average of  $4\%$  on Llama3-70B and Qwen2.5-72B. Furthermore, the selected demonstration set exhibits transferability to closed-source LLMs, consistently outperforming random selection by around  $2\%$  on Qwen-turbo, GLM-4-flash, Doubao-pro-32k, GPT-4o-mini, and DeepSeek-V3. Our source code is available at https://github.com/zhangjf-nlp/ManyShotICL-CLG.git.
+
+# 2 Related Work
+
+# 2.1 Many-Shot In-Context Learning
+
+In-Context Learning (ICL) refers to the capability of LLMs to learn from data during inference through forward propagation, without the need for backward propagation or weight updates (Li et al., 2023b). Prior work is limited to few-shot scenarios (Jiang et al., 2024) by the context window size, e.g., 2048 tokens in GPT-3 (Brown et al., 2020).
+
+Recently, advancements in expanding the context window size of LLMs (Ding et al., 2024; Chen et al., 2023), e.g., up to 128k tokens in GPT-4o (OpenAI, 2023) and Qwen2.5 (Yang et al., 2024), have enabled the exploration of many-shot ICL (Agarwal et al., 2024). These studies have observed that Many-Shot ICL, which includes up to hundreds or even thousands of demonstrations, can make substantial improvements compared to few-shot ICL on various tasks (Agarwal et al., 2024; Bertsch et al., 2024). In this paradigm, the benefits of instance-level retrieval over using a fixed random set of demonstrations tend to diminish (Bertsch et al., 2024), while a fixed demonstration set can largely reduce the inference cost through prefix-caching (Agarwal et al., 2024). Therefore, researchers tend to randomly select a demonstration set and reuse it across all queries in many-shot ICL (Agarwal et al., 2024; Bertsch et al., 2024; Jiang et al., 2024; Song et al., 2025).
+
+While random selection is widely adopted for many-shot ICL, it is underexplored whether this strategy is truly optimal. This motivates our work to explore and develop a better demonstration selection strategy for the many-shot paradigm.
+
+# 2.2 Demonstration Selection
+
+The performance of ICL is sensitive to the choice of demonstrations (Liu et al., 2022; Perez et al., 2021), prompting efforts in demonstration selection. Existing approaches fall into two categories:
+
+Instance-level retrieval selects demonstrations for each query using semantic similarity, such as cosine similarity (Rubin et al., 2022) and BM25 (Li et al., 2023a). Some work further emphasizes additional factors including diversity (Li and Qiu, 2023), complexity (Fu et al., 2023), difficulty (Drozdov et al., 2023), and perplexity (Gonen et al., 2023). While providing relevant and useful information, instance-level retrieval prevents the use of prefix-caching during inference and leads to the
+
+![](images/2fe83b9dd88a284f40307efebe1d081378c36592acf1a6d82003908a4364cd92.jpg)  
+Figure 2: The overview of our proposed method for many-shot demonstration selection. We perform latent concept learning (Wang et al., 2023) on a small LM and compute the Curriculum Latent Gradient (CLG) provided by each example throughout the  $E$ -epoch training. Then we find an  $n$ -shot demonstration set to mimic the average gradient on the entire training set. We show that the selected  $n$ -shot demonstrations can effectively improve the many-shot In-Context Learning performances on both same-series larger LMs and various closed-source LLMs.
+
+![](images/c3347243f009e12dfdcafcb939c66020b5321b81a1461ff9e4342a1225ed85e1.jpg)
+
+oretically  $O(n^{2})$  inference complexity (Vaswani et al., 2017) for  $n$ -shot demonstrations. In contrast, the runtime of many-shot ICL under a fixed demonstration set increases only linearly with a large number of shots (Agarwal et al., 2024). This inefficiency makes instance-level retrieval less practical as the number of demonstration shots increases.
+
+Task-level selection aims to identify a fixed set of demonstrations that achieve optimal average performance across all queries from a target task. For example, Zhang et al. (2022) use Reinforcement Learning (RL) to gradually explore the optimal demonstration set. Wang et al. (2023) view LLMs as latent variable models and select demonstrations that can best infer the optimal latent task concept. However, these methods are both limited to few-shot scenarios (e.g., 4-shot in their implementations), facing challenges from complexity in exploring potential many-shot demonstration sets (Zhang et al., 2022) and performance early saturation beyond 4-shot (Wang et al., 2023) respectively.
+
+In summary, current demonstration selection methods are fundamentally designed for and evaluated within few-shot scenarios, facing challenges in generalizing to many-shot scenarios. To the best of our knowledge, our work presents the first attempt to enhance demonstration selection for many-shot ICL, beyond random selection in existing work.
+
+# 3 Methodology
+
+Previous research has identified several key attributes that contribute to the effectiveness of in-context demonstrations, such as similarity (Luo et al., 2023), diversity (Li and Qiu, 2023), and coverage (Li and Qiu, 2023). While intuitively beneficial, these attributes lack a direct connection to the learning dynamics induced by demonstrations within the language model. We posit that the ideal demonstration set for ICL should not only present static attributes of high-quality fine-tuning data but, more fundamentally, should actively facilitate and guide the language model's learning process on the target task. Drawing inspiration from the principles of effective training data selection in supervised learning, particularly from the field of dataset condensation (Zhao et al., 2021), we introduce Curriculum Latent Gradient (CLG), a novel approach for task-level in-context demonstration selection. CLG leverages the concept of latent task embeddings and analyses their learning trajectory to identify demonstrations that can effectively guide the model's learning process.
+
+We begin by reviewing Latent Concept Learning (Sec. 3.1), highlighting its strengths and limitations as a foundation. We then articulate the core principles behind Curriculum Latent Gradient (Sec. 3.2), detailing how we capture and leverage
+
+the learning dynamics of latent task embeddings to guide demonstration selection. Finally, we describe the demonstration selection process based on these learning dynamics (Sec. 3.3). An overview of our proposed methodology is depicted in Fig. 2.
+
+# 3.1 Preliminary: Latent Concept Learning
+
+Latent Concept Learning Wang et al. (2023) aims to learn an optimal latent task concept  $z_{\theta}$  as a proxy of in-context demonstrations. Specifically, prediction from a language model  $P_M$  through in-context learning over demonstrations  $D_n = \{(x_1,y_1),\ldots ,(x_n,y_n)\}$  can be expressed in Eq. 1,
+
+$$
+P (y \mid x; D _ {n}) = P _ {M} (y \mid x _ {1}, y _ {1}, \dots , x _ {n}, y _ {n}, x) \tag {1}
+$$
+
+where  $D_{n}$  contains the  $n$ -shot demonstrations, and faces optimization challenges for its discrete nature. To address this, they replace the demonstration set with continuous and estimizable embeddings of the latent task concept  $z_{\theta}$ , as expressed in Eq. 2,
+
+$$
+P (y \mid x; z _ {\theta}) = P _ {M} (y \mid z _ {\theta}, x) \tag {2}
+$$
+
+where  $z_{\theta}$  can be optimized through maximum likelihood estimation on the entire training set  $D_N$ , as formulated in Eq. 3.
+
+$$
+\tilde {z} _ {\theta} = \underset {z _ {\theta}} {\arg \max } \sum_ {(x, y) \in D _ {N}} \log P (y \mid x; z _ {\theta}) \tag {3}
+$$
+
+Latent-Bayesian subsequently utilizes the optimized  $\tilde{z}_{\theta}$  to select task-level demonstrations. Under a series of simplifying assumptions, it finally selects the top- $n$  demonstrations with the highest posterior probabilities of  $\tilde{z}_{\theta}$  in language model  $P_{M}$ , as expressed in Eq. 4, where the top- $n$  operator selects the  $n$  samples with highest function values.
+
+$$
+D _ {n} = \underset {(x, y) \in D _ {N}} {\operatorname {t o p - n}} P _ {M} (z _ {\theta} \mid x, y) \tag {4}
+$$
+
+While Latent-Bayesian offers an interpretable approach under specific assumptions, its reliance on static posterior probabilities and independence assumptions limits its effectiveness, particularly for many-shot ICL. Crucially, it overlooks the dynamic interactions between demonstrations – such as redundancy or synergistic effects – which become increasingly vital as the number of demonstrations grows. This motivates our departure from static, result-oriented approaches towards a method that explicitly considers the learning process itself.
+
+# 3.2 Curriculum Latent Gradient
+
+Our method borrows the idea of latent concept learning, i.e., learning task concept embeddings that play the role of context. Notably, we focus on the learning process instead of the learnt result. We quantify the learning dynamics under different demonstrations through the optimizing gradients on the task concept embeddings provided by each demonstration, throughout the training process.
+
+Specifically, given a small language model  $P_{M}$  and a target task, we construct  $k$  prefix token embeddings  $z_{\theta} \in R^{k \times h}$  to represent the latent task concept, where  $h$  denotes the embedding size. These embeddings are then trained on the target task with the negative log-likelihood loss in Eq. 5.
+
+$$
+\mathcal {L} \left(z _ {\theta}; x, y\right) = - \log P _ {M} \left(y \mid z _ {\theta}, x\right) \tag {5}
+$$
+
+We randomly initialize the latent concept tokens  $z_{\theta} = z_{\theta_0}$ , and employ Stochastic Gradient Descent (Bottou, 2010) to optimize  $z_{\theta}$  over the full training set  $D_N$  for  $E$  epochs. Crucially, we save the learnt latent concept at the end of each epoch, denoted as  $z_{\theta_1}, z_{\theta_2}, \ldots, z_{\theta_E}$ . For each training example  $(x,y) \in D_N$ , we calculate its curriculum latent gradient,  $G_{0:E}(x,y) \in R^{k \times h \times (E + 1)}$ , which is a concatenation of the gradients of the loss function with respect to  $z_{\theta}$ , evaluated both at the initial point  $z_{\theta} = z_{\theta_0}$  and at the end of each epoch  $z_{\theta} = z_{\theta_i}$  for  $1 \leq i \leq E$ . This is formulated in Eq. 6, where  $[;]$  denotes the concatenation operation.
+
+$$
+\begin{array}{l} G _ {0: E} (x, y) = \left[ \nabla_ {z _ {\theta}} \mathcal {L} (z _ {\theta}; x, y) \right| _ {z _ {\theta} = z _ {\theta_ {0}}}; \\ \left. \nabla_ {z _ {\theta}} \mathcal {L} \left(z _ {\theta}; x, y\right) \right| _ {z _ {\theta} = z _ {\theta_ {1}}} ； \tag {6} \\ \dots ; \\ \left. \nabla_ {z _ {\theta}} \mathcal {L} (z _ {\theta}; x, y) \big | _ {z _ {\theta} = z _ {\theta_ {E}}} \right] \\ \end{array}
+$$
+
+# 3.3 Demonstration Selection via Gradient Matching
+
+To select an effective  $n$ -shot demonstration set for ICL, we aim to identify an optimal  $n$ -shot subset  $D_{n}^{*}\subseteq D_{N}$  that induces similar learning dynamics in the LM as the full training set  $D_{N}$ . Specifically, we propose to achieve this by minimizing the  $L_{2}$  distance between the average curriculum latent gradient over  $D_{N}$  and that over the  $n$ -shot subset  $D_{n}$ , as formalized in Eq. 7, ensuring that the model's learning behaviour on the subset  $D_{n}$ , as summarized in  $G_{0:E}(D_n)$ , closely aligns with that on the
+
+Algorithm 1 Gradient Matching through Greedy Search and Local Optimization  
+```txt
+Input: the entire training set  $D_N$ , the curriculum latent gradient  $G_{0:E}(x,y)$  for each example in the training set  $(x,y) \in D_N$ , the number of shots  $n$ , the maximum iteration steps  $l$  for local optimization.
+```
+
+```txt
+1: Compute the average curriculum latent gradient on  $D_N$ :  $G_{0:E}(D_N) = \mathbb{E}_{(x,y)\in D_N}[G_{0:E}(x,y)]$ .
+```
+
+```txt
+3: for  $i = 1$  to  $n$  do
+```
+
+```txt
+4: Find the next example  $(x_{j},y_{j})\in (D_{N}\setminus D_{i - 1})$  that minimizes the distance between the average curriculum latent gradient on  $D_N$  and that on  $D_{i - 1}\cup \{(x_j,y_j)\}$
+```
+
+```txt
+5: Set  $D_{i} = D_{i - 1}\cup \{(x_{j},y_{j})\}$
+```
+
+```txt
+6: end for
+```
+
+```txt
+7: Set  $d_{\mathrm{old}} = \left\| G_{0:E}(D_N) - \mathbb{E}_{(x,y)\in D_n}G_{0:E}(x,y)\right\| _2$
+```
+
+```txt
+8: for  $i = 1$  to  $l$  do
+```
+
+```latex
+9: Find  $(x_{j},y_{j})\in D_{N}\setminus D_{n}$  and  $(x_{k},y_{k})\in D_{n}$  to minimize the distance between the average curriculum latent gradient on  $D_N$  and that on  $D_{n}\cup \{(x_{j},y_{j})\} \setminus \{(x_{k},y_{k})\}$ .
+```
+
+```txt
+11: Set  $d_{\mathrm{new}} = \left\| G_{0:E}(D_N) - \mathbb{E}_{(x,y)\in D_{\mathrm{new}}}G_{0:E}(x,y)\right\| _2,$
+```
+
+```txt
+12: if  $d_{\mathrm{new}} \geq d_{\mathrm{old}}$  then
+```
+
+```txt
+13: break
+```
+
+```txt
+14: else
+```
+
+```txt
+15: Set  $D_{n} = D_{\mathrm{new}}$  and  $d_{\mathrm{old}} = d_{\mathrm{new}}$
+```
+
+```txt
+16: end if
+```
+
+```txt
+17: end for
+```
+
+```prolog
+18: return  $D_{n}$  as the optimized  $n$ -shot demonstrations  $D_{n}^{*}$ .
+```
+
+complete set  $D_N$ , as summarized in  $G_{0:E}(D_N)$ .
+
+$$
+G _ {0: E} (D _ {N}) = \mathbb {E} _ {(x, y) \in D _ {N}} [ G _ {0: E} (x, y) ]
+$$
+
+$$
+G _ {0: E} (D _ {n}) = \mathbb {E} _ {(x, y) \in D _ {n}} \left[ G _ {0: E} (x, y) \right]
+$$
+
+$$
+D _ {n} ^ {*} = \underset {D _ {n} \subset D _ {N}} {\arg \min } \| G _ {0: E} (D _ {N}) - G _ {0: E} (D _ {n}) \| _ {2} \tag {7}
+$$
+
+Since the optimization problem in Eq. 7 is indeed an NP-Complete Subset Sum Problem (SSP) (Lagarias and Odlyzko, 1985), we resort to heuristic approximation to find a near-optimal solution for  $D_{n}^{*}$ , involving two key phases: Greedy Search and Local Optimization. In the greedy phase, we incrementally construct the  $n$ -shot subset  $D_{n}$  by iteratively adding the example that minimizes the target  $L_{2}$  distance. Then the local optimization phase refines the  $n$ -shot subset through up to  $l = 32$  iterations to further reduce the target  $L_{2}$  distance, by replacing a selected example with an unselected one. The detailed steps of the selection algorithm are presented in Algorithm 1, which requires only a few minutes to execute in practice.
+
+<table><tr><td></td><td>designed for task-level</td><td>extended from instance-level</td></tr><tr><td rowspan="2">learning-free</td><td>Random</td><td>BM25-Major</td></tr><tr><td>Best-of-N</td><td>BGE-KMeans</td></tr><tr><td rowspan="2">learning-based</td><td>Latent-Bayesian</td><td rowspan="2">EPR-KMeans</td></tr><tr><td>CLG (ours)</td></tr></table>
+
+Table 1: Types of Methods
+
+# 4 Experiments
+
+We conduct experiments on 9 datasets from 5 distinct tasks, and show that demonstrations selected by our proposed method can effectively improve the many-shot ICL performances over random selection as well as some straightforward methods.
+
+# 4.1 Baselines
+
+In addition to random selection, the commonly adopted method in existing work of many-shot ICL, we examine various straightforward methods for demonstration selection, including those designed for task-level selection and those extended from instance-level retrieval. We classify these methods in Table 1 and introduce them in details below.
+
+Random: The most basic demonstration selection method, which is widely-adopted in existing work for many-shot ICL (Agarwal et al., 2024; Bertsch et al., 2024; Song et al., 2025; Jiang et al., 2024). We implement this across 5 random seeds and report the mean and standard deviation.
+
+BM25-Major: BM25 (Robertson and Zaragoza, 2009) is a popular term-based scoring method for instance-level retrieval (Li et al., 2023a). We extend it to task-level demonstration selection through Majority Voting, i.e., selecting  $n$ -shot demonstrations with the  $n$  highest average scores to be retrieved by the other examples on the training set.
+
+BGE-KMeans: BGE-M3 (Chen et al., 2024) provides off-the-shelf sentence embeddings with state-of-the-art performances on multiple retrieval tasks. We extend it to task-level selection through KMeans clustering, as implemented in the Scikit-Learn library, over its text embeddings.
+
+EPR-KMeans: EPR (Rubin et al., 2022) learns to retrieve demonstrations through contrastive learning, supervised by the ICL likelihoods on a relatively small LM. We train EPR retrievers and apply them to task-level selection through KMeans clustering over their learnt dense embeddings.
+
+<table><tr><td>Type</td><td>Dataset</td><td>Task</td><td>#Train</td><td>#Validation</td><td>Avg. Tokens</td><td>Metric</td></tr><tr><td rowspan="4">Classification</td><td>SST-5 (Socher et al., 2013)</td><td>Sentiment Analysis</td><td>8,544</td><td>1,101</td><td>26.55</td><td>Acc</td></tr><tr><td>MNLI (Williams et al., 2018)</td><td>Natural Language Inference</td><td>50,000</td><td>10,000</td><td>43.21</td><td>Acc</td></tr><tr><td>CMSQA (Talmor et al., 2019)</td><td>Commonsense Reasoning</td><td>9,741</td><td>1,221</td><td>45.70</td><td>Acc</td></tr><tr><td>HellaSwag (Zellers et al., 2019)</td><td>Commonsense Reasoning</td><td>50,000</td><td>10,000</td><td>79.42</td><td>Acc</td></tr><tr><td rowspan="5">Open-ended</td><td>GeoQuery (Shaw et al., 2021)</td><td>Code Generation</td><td>600</td><td>280</td><td>22.66</td><td>EM</td></tr><tr><td>NL2Bash (Lin et al., 2018)</td><td>Code Generation</td><td>8,090</td><td>609</td><td>32.56</td><td>BLEU</td></tr><tr><td>Break (Wolfson et al., 2020)</td><td>Semantic Parsing</td><td>44,321</td><td>7,760</td><td>61.44</td><td>LF-EM</td></tr><tr><td>MTOP (Li et al., 2021)</td><td>Semantic Parsing</td><td>15,667</td><td>2,235</td><td>34.12</td><td>EM</td></tr><tr><td>SMCalFlow (Andreas et al., 2020)</td><td>Semantic Parsing</td><td>50,000</td><td>10,000</td><td>53.78</td><td>EM</td></tr></table>
+
+Table 2: Datasets used in our experiments. We use at most 50,000 training instances and 10,000 validation instances. Following previous work (Ye et al., 2023), we use Accuracy (Acc) on classification tasks, and use Exact Match (EM), Logical Form EM (LF-EM) (Hasson and Berant, 2021), and character-level BLEU on open-ended tasks.
+
+Best-of-N: Best-of-N is a commonly used baseline for Reinforcement Learning (RL), and demonstrates competitive performance to RL-based task-level demonstration selection (Zhang et al., 2022). We implement this through randomly selecting  $N = 5$  demonstration sets, evaluating them on a relatively small LM over all training instances, and selecting the best-performing one.
+
+Latent-Bayesian: The task-level demonstration selection method based on Bayesian inference (Wang et al., 2023), as we introduce in Sec 3.1.
+
+# 4.2 Datasets and Evaluation
+
+We list all the datasets in Table 2, paired with their tasks, sizes, average tokens, and evaluation metrics. We illustrate cases on each dataset in Appendix A. We select demonstrations from the training set and report their ICL performances on the validation set, since the test set is private for some datasets.
+
+# 4.3 Implementation Details
+
+For LM-based approaches, we utilize Llama3-8B to select demonstrations for Llama3 series LLMs (Dubey et al., 2024), and utilize Qwen2.5-3B to select demonstrations for Qwen2.5 series LLMs (Yang et al., 2024) and closed-source LLMs. We train EPR retrievers from BERT-base-uncased (Devlin et al., 2019) under the supervision of LMs, using a learning rate of  $1e - 5$  and a batch size of 8, for at most 120 epochs until regression. We conduct latent concept learning on LMs for Latent-Bayesian and CLG, with a learning rate of  $1e - 3$  and a batch size of 64 for  $E = 10$  epochs.
+
+For ICL performance evaluation, we utilize the vLLM framework (Kwon et al., 2023) to perform greedy search, involving 4-shot, 8-shot, 16-shot, 32-shot, 64-shot, and 128-shot settings. On classi-
+
+fication tasks, we only allow the model to generate tokens contained in the options.
+
+# 4.4 Main Results
+
+We illustrate the results of 128-shot ICL on Llama3-70B and Qwen2.5-72B in Table 3. It can be observed that, demonstrations selected by random are in fact generally suboptimal for task-level in-context learning. Among all methods, our proposed CLG performs the best on average, surpassing that of random selection by  $4\%$  on both models.
+
+We further evaluate the selected demonstrations by Random, Best-of-N, and CLG on 5 closed-source models: Qwen-turbo, GLM-4-flash, Doubao-pro, GPT-4o-mini, and DeepSeek-V3. We include at most 300 instances from each dataset and constrain the context length to within 7000 tokens due to budget limitations. The average performances across all datasets are illustrated in Table 4, where CLG consistently outperforms Random and Best-of-N on all closed-source models. This demonstrates the transferability of demonstrations selected by our CLG to different series of LLMs, as well as the effectiveness of CLG on closed-source LLMs.
+
+# 4.5 Overall Scaling Performance
+
+In Fig. 3, we examine the scaling trends of ICL performance w.r.t. the model size and the shot number. For simplicity, we present the average performance scores over 9 datasets. The results reveal a clear
+
+<table><tr><td>Method</td><td>SST-5</td><td>MNLI</td><td>CMSQA</td><td>HeSwag.</td><td>GeoQ.</td><td>NL2Bash</td><td>Break</td><td>MTOP</td><td>SMCal.</td><td>Average</td></tr><tr><td colspan="11">Llama3-70B</td></tr><tr><td>Random</td><td>43.71</td><td>61.83</td><td>84.73</td><td>76.11</td><td>73.36</td><td>29.72</td><td>35.55</td><td>44.66</td><td>36.68</td><td>54.04</td></tr><tr><td>BM25-Major</td><td>43.69</td><td>51.36</td><td>82.23</td><td>72.01</td><td>43.21</td><td>27.92</td><td>15.05</td><td>8.41</td><td>14.38</td><td>39.81</td></tr><tr><td>BGE-KMeans</td><td>46.41</td><td>65.09</td><td>83.78</td><td>73.90</td><td>84.64</td><td>28.39</td><td>35.13</td><td>44.79</td><td>38.73</td><td>55.65</td></tr><tr><td>Best-of-N</td><td>43.69</td><td>70.92</td><td>84.19</td><td>74.57</td><td>80.71</td><td>29.76</td><td>35.86</td><td>43.40</td><td>37.66</td><td>55.64</td></tr><tr><td>EPR-KMeans</td><td>43.05</td><td>49.78</td><td>83.62</td><td>71.32</td><td>78.21</td><td>22.92</td><td>38.39</td><td>50.43</td><td>38.10</td><td>52.87</td></tr><tr><td>Latent-Bayesian</td><td>46.59</td><td>69.84</td><td>84.60</td><td>70.09</td><td>62.14</td><td>29.20</td><td>32.59</td><td>41.21</td><td>23.13</td><td>51.04</td></tr><tr><td>CLG (ours)</td><td>48.32</td><td>76.37</td><td>84.52</td><td>80.77</td><td>84.64</td><td>29.59</td><td>37.33</td><td>47.74</td><td>40.35</td><td>58.85</td></tr><tr><td colspan="11">Qwen2.5-72B</td></tr><tr><td>Random</td><td>36.00</td><td>58.20</td><td>87.47</td><td>86.58</td><td>57.29</td><td>33.28</td><td>36.67</td><td>45.32</td><td>38.23</td><td>53.23</td></tr><tr><td>BM25-Major</td><td>37.78</td><td>45.94</td><td>86.65</td><td>85.61</td><td>52.50</td><td>38.11</td><td>22.81</td><td>16.24</td><td>11.57</td><td>44.13</td></tr><tr><td>BGE-KMeans</td><td>37.24</td><td>49.65</td><td>87.71</td><td>86.67</td><td>61.79</td><td>40.86</td><td>35.84</td><td>43.71</td><td>41.29</td><td>53.86</td></tr><tr><td>Best-of-N</td><td>36.60</td><td>60.57</td><td>87.55</td><td>86.83</td><td>61.07</td><td>37.13</td><td>35.73</td><td>46.89</td><td>39.22</td><td>54.62</td></tr><tr><td>EPR-KMeans</td><td>36.15</td><td>58.41</td><td>87.96</td><td>86.22</td><td>62.86</td><td>26.53</td><td>39.12</td><td>47.53</td><td>39.65</td><td>53.83</td></tr><tr><td>Latent-Bayesian</td><td>27.25</td><td>47.96</td><td>85.75</td><td>86.72</td><td>27.50</td><td>37.36</td><td>35.35</td><td>5.23</td><td>22.61</td><td>41.75</td></tr><tr><td>CLG (ours)</td><td>38.33</td><td>77.29</td><td>87.71</td><td>87.68</td><td>62.50</td><td>45.19</td><td>37.07</td><td>47.07</td><td>40.16</td><td>58.11</td></tr></table>
+
+![](images/7f63d0bd4db011fb5c53f4712c083483496d8a9a253156b76ec8ea45269be356.jpg)  
+Figure 3: The scaling performances with respect to (w.r.t.) the number of shots and the model size. For random selection under 5 different random seeds, standard deviations are illustrated through error bars.
+
+Table 3: Main results (in %) of Llama3-70B and Qwen2.5-72B under the 128-shot setting. The best scores on each model are in bold, and the second-best ones are underlined. Full results are illustrated in Appendix C. BM25-Major and Latent-Bayesian generally perform worse than Random selection; we analyse the reasons in Appendix B.  
+
+<table><tr><td>Method</td><td>Qwen.</td><td>GLM4</td><td>DouB.</td><td>GPT4o</td><td>DeepS.</td></tr><tr><td>Random</td><td>60.08</td><td>56.17</td><td>62.78</td><td>60.22</td><td>66.83</td></tr><tr><td>±std</td><td>±0.22</td><td>±2.24</td><td>±0.26</td><td>±0.45</td><td>±0.52</td></tr><tr><td>Best-of-N</td><td>59.63</td><td>57.65</td><td>61.69</td><td>60.38</td><td>66.79</td></tr><tr><td>CLG (ours)</td><td>61.87</td><td>58.50</td><td>64.09</td><td>61.48</td><td>68.08</td></tr></table>
+
+Table 4: ICL performances of 128-shot demonstrations, average over 9 datasets, on Qwen-turbo, GLM-4-flash, Doubao-pro, GPT-4o-mini, and DeepSeek-V3. Best scores are in bold. More details are in Appendix D.
+
+positive correlation between the overall ICL performance and the model size as well as the shot number, indicating that larger models and more shots generally lead to better performances.
+
+Among the methods compared, CLG generally outperforms other methods across various conditions, achieving significant improvements (far more than the standard deviation) over random selection. In contrast, the most competitive Best-of-N baseline exhibits relatively lower scaling efficiency with
+
+the increased model size or the increased shot numbers. This further highlights the scalability of CLG with respect to the model size and the shot number.
+
+# 4.6 Scaling to a Thousand Shots
+
+To further compare CLG against Random selection in long-context settings with enough demonstrations, we select 1024-shot demonstrations through both methods, and evaluate their performances on MNLI, HellaSwag, Break, and SMCalFlow, the four datasets with the most examples.
+
+As illustrated in Table 5, CLG consistently achieves better results than the average performance of Random selection on all datasets. Such performance gains on 1024-shot are generally smaller than those on 128-shot. This could be explained by the possibility that 1024-shot demonstrations provide sufficient task knowledge needed for ICL, even when selected randomly.
+
+Besides, we found that Break and SMCalFlow
+
+<table><tr><td>Method</td><td>MNLI</td><td>HeSwag.</td><td>Break</td><td>SMCal.</td></tr><tr><td colspan="5">128-shot</td></tr><tr><td>Random</td><td>61.57</td><td>86.58</td><td>36.67</td><td>38.23</td></tr><tr><td>±std</td><td>±1.12</td><td>±0.33</td><td>±0.60</td><td>±1.42</td></tr><tr><td>CLG (ours)</td><td>77.29</td><td>87.68</td><td>37.07</td><td>40.16</td></tr><tr><td>Δ</td><td>+15.72</td><td>+1.10</td><td>+0.40</td><td>+1.93</td></tr><tr><td colspan="5">1024-shot</td></tr><tr><td>Random</td><td>55.08</td><td>77.64</td><td>42.08</td><td>52.67</td></tr><tr><td>±std</td><td>±1.75</td><td>±1.83</td><td>±0.28</td><td>±0.35</td></tr><tr><td>CLG (ours)</td><td>58.94</td><td>79.12</td><td>42.21</td><td>52.82</td></tr><tr><td>Δ</td><td>+3.86</td><td>+1.48</td><td>+0.13</td><td>+0.15</td></tr></table>
+
+Table 5: Comparison of 128-shot and 1024-shot ICL on Qwen2.5-72B (in %). Best scores are in bold.
+
+show significant improvements when scaled to 1024-shot, while MNLI and HellaSwag exhibit some degree of performance degradation. To explain this phenomenon, we analyse the scaling trends of ICL on each dataset with increasing shot numbers in Appendix E. We find that open-ended tasks, such as Break and SMCalFlow, are more likely to benefit from a larger number of shots.
+
+# 4.7 Ablation Study: Gradient Mismatching
+
+We conduct an ablation study to assess the effectiveness of the gradient matching term in Eq. 7. Specifically, we reverse the optimization target: instead of minimizing the  $L_{2}$  distance in Eq. 7, we maximize it to select demonstrations that misalign with the full training set in learning dynamics.
+
+As illustrated in Fig. 4, this reversal results in significant degradation in the ICL performance, which can be much poorer than that of random selection. These results empirically validate that the quality of demonstrations is correlated with the degree of curriculum latent gradient matching.
+
+![](images/7e1ea89111a331395a835d905c4fe2e2450f48670c989167e48df2034f195862.jpg)  
+Figure 4: Ablation study on the impact of gradient mismatching on ICL performance, using 128-shot demonstrations with Qwen2.5-72B.
+
+# 4.8 Improving Adaptability through Hybrid In-Context Learning
+
+In addition to our main experiments on task-level many-shot in-context learning, we further explore
+
+the complementary use of task-level and instance-level demonstrations. Specifically, we combine the 128-shot task-level demonstrations selected by Random and CLG with the 4-shot instance-level demonstrations retrieved by BGE-M3.
+
+As illustrated in Fig. 5, such hybrid strategy yields further improvements over using only task-level or instance-level demonstrations across most datasets. Meanwhile, CLG generally outperforms Random selection in such hybrid mode.
+
+![](images/b326fa280aee75ba138d46a3cb23d5517d44b75151bd38519a422e460e14323e.jpg)  
+Figure 5: ICL with hybrid demonstrations: task-level many-shot followed by instance-level few-shot.
+
+# 4.9 Correlation between FT and ICL
+
+To validate our hypothesis that the data requirements for In-Context Learning (ICL) and Fine-Tuning (FT) are comparable, we fine-tune Qwen2.5-3B on the 128-shot demonstrations selected by each method. We employ prefix-tuning (Li and Liang, 2021) to avoid overfitting, and perform training with a batch size of 128 and a learning rate of  $3e - 4$  for 1000 epochs. We quantify the FT performances through the negative evaluation losses on the validation set. We analyse and illustrate the correlation between FT and ICL performance in Fig. 6, which shows a positive relationship between the two forms of machine learning, especially for open-ended tasks.
+
+![](images/e675e2c088c04445d0363c5a19165678f40c0355235e7748aeb1d5ed18bc1a25.jpg)  
+Figure 6: The correlation between FT and ICL. The left involves ICL performances on Qwen2.5-3B, and the right involves that averaged over all Qwen2.5 models. Datasets for open-ended tasks are lined in solid colors.
+
+# 4.10 Diversity and Coverage Analysis
+
+Since prior work has illustrated that diversity and coverage are important for effective in-context
+
+demonstrations (Li and Qiu, 2023), we examine whether our method inherently maintains diversity and coverage in selected demonstrations. Specifically, we analyze the label distributions of demonstration sets on classification datasets. We measure the KL divergence between the label distribution of selected demonstrations and that of the test set, with results shown in Table 6.
+
+<table><tr><td>Method</td><td>SST-5</td><td>MNLI</td><td>CMSQA</td><td>HeSwag.</td></tr><tr><td>Random</td><td>0.020</td><td>0.016</td><td>0.016</td><td>0.012</td></tr><tr><td>± std</td><td>±0.016</td><td>±0.009</td><td>±0.008</td><td>±0.007</td></tr><tr><td>Best-of-N</td><td>0.019</td><td>0.021</td><td>0.041</td><td>0.019</td></tr><tr><td>BGE-KMeans</td><td>0.017</td><td>0.036</td><td>0.010</td><td>0.026</td></tr><tr><td>Latent-Bayesian</td><td>4.050</td><td>0.115</td><td>0.134</td><td>4.296</td></tr><tr><td>CLG (ours)</td><td>0.006</td><td>0.002</td><td>0.007</td><td>0.008</td></tr></table>
+
+Table 6: KL divergence between demonstration set label distribution and test set label distribution (lower is better). CLG achieves the best alignment with test set distributions, indicating it inherently maintains diversity and coverage of labels on classification datasets.
+
+The results demonstrate that CLG naturally selects a demonstration set whose label distribution closely matches the test set distribution, significantly outperforming random selection. This suggests that gradient matching can inherently preserve data coverage in terms of label. In contrast, Latent-Bayesian shows particularly poor performance in maintaining label diversity, which may contribute to its suboptimal ICL performance.
+
+# 4.11 Efficiency Analysis
+
+To further verify the cost-effectiveness of our proposed CLG in practice, we analyse the computational cost of the complete demonstration selection pipeline. Table 7 details the time consumption for selecting 128-shot demonstrations from MNLI and SMCalFlow using Qwen2.5-3B, measured on 8 A100 GPUs with data-parallel implementation via deepspeed and accelerate libraries.
+
+<table><tr><td>CLG Stage</td><td># GPUs</td><td>MNLI</td><td>SMCal.</td></tr><tr><td>Prefix Tuning</td><td>8</td><td>80 min</td><td>87 min</td></tr><tr><td>Gradient Computation</td><td>8</td><td>145 min</td><td>210 min</td></tr><tr><td>Gradient Matching</td><td>1</td><td>6 min</td><td>10 min</td></tr></table>
+
+Table 7: Time consumption for 128-shot demonstration selection from datasets with 50k training examples.
+
+The results demonstrate that our approach maintains practical efficiency even at scale, with total computation time being approximately  $3\sim 4$  times that of prefix-tuning, an already efficient finetuning method. This computational investment is
+
+well justified as our method enables significant ICL performance improvements on both open-source larger models and closed-source LLMs, for which fine-tuning may face challenge in computational resources and access to model weights. This validates the cost-effectiveness of our proposed CLG for many-shot ICL in practice.
+
+# 5 Conclusion
+
+Many-shot ICL has emerged as a promising way to utilize LLMs on downstream tasks, but the random selection applied in existing work may produce sub-optimal demonstrations and learning results.
+
+In this work, we hypothesize that data requirements for in-context learning and fine-tuning are analogous, and propose Curriculum Latent Gradient (CLG) to select a demonstration set that aligns with the entire training set in learning dynamics on LMs. We validate our method across various datasets and LLMs, showing its significant improvements compared to random selection, e.g., improving by  $4\%$  on open-source LLMs and by  $2\%$  on closed-source LLMs.
+
+This work unlocks more reliable and effective many-shot ICL, paving the way for its broader adoption and application in real-world scenarios.
+
+# Limitations
+
+This work focuses on improving many-shot in-context learning (ICL) by optimizing the demonstration set. However, ICL performance can also be sensitive to the order of demonstrations (Lu et al., 2022), which our method does not address. We believe that investigating the impact of demonstration order on in-context learning could lead to further improvements in the learning performance, for example, by examining the learning dynamics at different stages of fine-tuning.
+
+Besides, our method selects demonstrations according to analysis on the training set, which may suffer from biases in the training data, such as imbalances over gender, race, or culture. This could lead to discriminatory content in practice. Furthermore, improved many-shot ICL may be susceptible to malicious exploitation, resulting in misuse such as jailbreaking LLMs (Anil et al., 2024).
+
+# Acknowledgments
+
+This work was supported in part by the National Natural Science Foundation of China under Grant 62477001.
+
+# References
+
+Rishabh Agarwal, Avi Singh, Lei M. Zhang, Bernd Bohnet, Stephanie Chan, Ankesh Anand, Zaheer Abbas, Azade Nova, John D. Co-Reyes, Eric Chu, Feryal M. P. Behbahani, Aleksandra Faust, and Hugo Larochelle. 2024. Many-shot in-context learning. CoRR, abs/2404.11018.  
+Jacob Andreas, John Bufe, David Burkett, Charles Chen, Josh Clausman, Jean Crawford, Kate Crim, Jordan DeLoach, Leah Dorner, Jason Eisner, Hao Fang, Alan Guo, David Hall, Kristin Hayes, Kellie Hill, Diana Ho, Wendy Iwaszuk, Smriti Jha, Dan Klein, Jayant Krishnamurthy, Theo Lanman, Percy Liang, Christopher H. Lin, Ilya Lintsbakh, Andy McGovern, Aleksandr Nisnevich, Adam Pauls, Dmitrij Petters, Brent Read, Dan Roth, Subhro Roy, Jesse Rusak, Beth Short, Div Slomin, Ben Snyder, Stephon Striplin, Yu Su, Zachary Tellman, Sam Thomson, Andrei Vorobev, Izabela Witoszko, Jason Andrew Wolfe, Abby Wray, Yuchen Zhang, and Alexander Zotov. 2020. Task-oriented dialogue as dataflow synthesis. Trans. Assoc. Comput. Linguistics, 8:556-571.  
+Cem Anil, Esin Durmus, Nina Rimsky, Mrinank Sharma, Joe Benton, Sandipan Kundu, Joshua Batson, Meg Tong, Jesse Mu, Daniel J Ford, et al. 2024. Many-shot jailbreaking. In The Thirty-eighth Annual Conference on Neural Information Processing Systems.  
+Amanda Bertsch, Maor Ivgi, Uri Alon, Jonathan Berant, Matthew R. Gormley, and Graham Neubig. 2024. In-context learning with long-context models: An in-depth exploration. CoRR, abs/2405.00200.  
+Léon Bottou. 2010. Large-scale machine learning with stochastic gradient descent. In 19th International Conference on Computational Statistics, COMP-STAT 2010, Paris, France, August 22-27, 2010 - Keynote, Invited and Contributed Papers, pages 177-186. Physica-Verlag.  
+Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. 2020. Language models are few-shot learners. In Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual.  
+Jianlv Chen, Shitao Xiao, Peitian Zhang, Kun Luo, Defu Lian, and Zheng Liu. 2024. BGE m3-embedding: Multi-lingual, multi-functionality, multi-granularity text embeddings through self-knowledge distillation. CoRR, abs/2402.03216.
+
+Shouyuan Chen, Sherman Wong, Liangjian Chen, and Yuandong Tian. 2023. Extending context window of large language models via positional interpolation. CoRR, abs/2306.15595.  
+Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, NAACL-HLT 2019, Minneapolis, MN, USA, June 2-7, 2019, Volume 1 (Long and Short Papers), pages 4171-4186. Association for Computational Linguistics.  
+Yiran Ding, Li Lyna Zhang, Chengruidong Zhang, Yuanyuan Xu, Ning Shang, Jiahang Xu, Fan Yang, and Mao Yang. 2024. Longrope: Extending LLM context window beyond 2 million tokens. In *Forty-first International Conference on Machine Learning*, ICML 2024, Vienna, Austria, July 21-27, 2024. Open-Review.net.  
+Andrew Drozdov, Honglei Zhuang, Zhuyun Dai, Zhen Qin, Razieh Rahimi, Xuanhui Wang, Dana Alon, Mohit Iyyer, Andrew McCallum, Donald Metzler, and Kai Hui. 2023. Parade: Passage ranking using demonstrations with large language models. CoRR, abs/2310.14408.  
+Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, Anirudh Goyal, Anthony Hartshorn, Aobo Yang, Archi Mitra, Archie Sravankumar, Artem Korenev, Arthur Hinsvark, Arun Rao, Aston Zhang, Aurélien Rodriguez, Austen Gregerson, Ava Spataru, Baptiste Rozière, Bethany Biron, Binh Tang, Bobbie Chern, Charlotte Caucheteux, Chaya Nayak, Chloe Bi, Chris Marra, Chris McConnell, Christian Keller, Christophe Touret, Chunyang Wu, Corinne Wong, Cristian Canton Ferrer, Cyrus Nikolaidis, Damien Allonsius, Daniel Song, Danielle Pintz, Danny Livshits, David Esiobu, Dhruv Choudhary, Dhruv Mahajan, Diego Garcia-Olano, Diego Perino, Dieuwke Hupkes, Egor Lakomkin, Ehab AlBadawy, Elina Lobanova, Emily Dinan, Eric Michael Smith, Filip Radenovic, Frank Zhang, Gabriel Synnaeve, Gabrielle Lee, Georgia Lewis Anderson, Graeme Nail, Grégoire Mialon, Guan Pang, Guillem Cucurell, Hailey Nguyen, Hannah Korevaar, Hu Xu, Hugo Touvron, Iliyan Zarov, Imanol Arrieta Ibarra, Isabel M. Kloumann, Ishan Misra, Ivan Evtimov, Jade Copet, Jaewon Lee, Jan Geffert, Jana Vranes, Jason Park, Jay Mahadeokar, Jeet Shah, Jelmer van der Linde, Jennifer Billock, Jenny Hong, Jenya Lee, Jeremy Fu, Jianfeng Chi, Jianyu Huang, Jiawen Liu, Jie Wang, Jiecao Yu, Joanna Bitton, Joe Spisak, Jongsoo Park, Joseph Rocca, Joshua Johnstun, Joshua Saxe, Junteng Jia, Kalyan Vasuden Alwala, Kartikeya Upasani, Kate Plawiak, Ke Li, Kenneth Heafield, Kevin Stone, and et al. 2024. The llama 3 herd of models. CoRR abs/2407.21783.
+
+Yao Fu, Hao Peng, Ashish Sabharwal, Peter Clark, and Tushar Khot. 2023. Complexity-based prompting for multi-step reasoning. In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023. OpenReview.net.  
+Hila Gonen, Srini Iyer, Terra Blevins, Noah A. Smith, and Luke Zettlemoyer. 2023. Demystifying prompts in language models via perplexity estimation. In Findings of the Association for Computational Linguistics: EMNLP 2023, Singapore, December 6-10, 2023, pages 10136-10148. Association for Computational Linguistics.  
+Matan Hasson and Jonathan Berant. 2021. Question decomposition with dependency graphs. In 3rd Conference on Automated Knowledge Base Construction, AKBC 2021, Virtual, October 4-8, 2021.  
+Yixing Jiang, Jeremy Irvin, Ji Hun Wang, Muhammad Ahmed Chaudhry, Jonathan H. Chen, and Andrew Y. Ng. 2024. Many-shot in-context learning in multimodal foundation models. CoRR, abs/2405.09798.  
+Woosuk Kwon, Zhuohan Li, Siyuan Zhuang, Ying Sheng, Lianmin Zheng, Cody Hao Yu, Joseph Gonzalez, Hao Zhang, and Ion Stoica. 2023. Efficient memory management for large language model serving with pagedattention. In Proceedings of the 29th Symposium on Operating Systems Principles, SOSP 2023, Koblenz, Germany, October 23-26, 2023, pages 611-626. ACM.  
+J. C. Lagarias and Andrew M. Odlyzko. 1985. Solving low-density subset sum problems. J. ACM, 32(1):229-246.  
+Haoran Li, Abhinav Arora, Shuohui Chen, Anchit Gupta, Sonal Gupta, and Yashar Mehdad. 2021. MTOP: A comprehensive multilingual task-oriented semantic parsing benchmark. In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume, EACL 2021, Online, April 19 - 23, 2021, pages 2950-2962. Association for Computational Linguistics.  
+Xiang Lisa Li and Percy Liang. 2021. Prefix-tuning: Optimizing continuous prompts for generation. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, ACL/IJCNLP 2021, (Volume 1: Long Papers), Virtual Event, August 1-6, 2021, pages 4582-4597. Association for Computational Linguistics.  
+Xiaonan Li, Kai Lv, Hang Yan, Tianyang Lin, Wei Zhu, Yuan Ni, Guotong Xie, Xiaoling Wang, and Xipeng Qiu. 2023a. Unified demonstration retriever for incontext learning. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), ACL 2023, Toronto, Canada, July 9-14, 2023, pages 4644-4668. Association for Computational Linguistics.
+
+Xiaonan Li and Xipeng Qiu. 2023. Finding supporting examples for in-context learning. CoRR, abs/2302.13539.  
+Yucheng Li, Bo Dong, Frank Guerin, and Chenghua Lin. 2023b. Compressing context to enhance inference efficiency of large language models. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 6342-6353, Singapore.  
+Xi Victoria Lin, Chenglong Wang, Luke Zettlemoyer, and Michael D. Ernst. 2018. Nl2bash: A corpus and semantic parser for natural language interface to the linux operating system. In Proceedings of the Eleventh International Conference on Language Resources and Evaluation, LREC 2018, Miyazaki, Japan, May 7-12, 2018. European Language Resources Association (ELRA).  
+Jiachang Liu, Dinghan Shen, Yizhe Zhang, Bill Dolan, Lawrence Carin, and Weizhu Chen. 2022. What makes good in-context examples for gpt-3? In Proceedings of Deep Learning Inside Out: The 3rd Workshop on Knowledge Extraction and Integration for Deep Learning Architectures, DeeLIO@ACL 2022, Dublin, Ireland and Online, May 27, 2022, pages 100-114. Association for Computational Linguistics.  
+Yao Lu, Max Bartolo, Alastair Moore, Sebastian Riedel, and Pontus Stenetorp. 2022. Fantastically ordered prompts and where to find them: Overcoming few-shot prompt order sensitivity. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), ACL 2022, Dublin, Ireland, May 22-27, 2022, pages 8086-8098. Association for Computational Linguistics.  
+Man Luo, Xin Xu, Zhuyun Dai, Panupong Pasupat, Seyed Mehran Kazemi, Chitta Baral, Vaiva Imbrasaite, and Vincent Y. Zhao. 2023. Dr.icl: Demonstration-retrieved in-context learning. CoRR, abs/2305.14128.  
+OpenAI. 2023. GPT-4 technical report. CoRR, abs/2303.08774.  
+Ethan Perez, Douwe Kiela, and Kyunghyun Cho. 2021. True few-shot learning with language models. In Advances in Neural Information Processing Systems 34: Annual Conference on Neural Information Processing Systems 2021, NeurIPS 2021, December 6-14, 2021, virtual, pages 11054-11070.  
+Reiner Pope, Sholto Douglas, Aakanksha Chowdhery, Jacob Devlin, James Bradbury, Jonathan Heek, Kefan Xiao, Shivani Agrawal, and Jeff Dean. 2023. Efficiently scaling transformer inference. In Proceedings of the Sixth Conference on Machine Learning and Systems, MLSys 2023, Miami, FL, USA, June 4-8, 2023. mlsys.org.  
+Stephen E. Robertson and Hugo Zaragoza. 2009. The probabilistic relevance framework: BM25 and beyond. Found. Trends Inf. Retr., 3(4):333-389.
+
+Ohad Rubin, Jonathan Herzig, and Jonathan Berant. 2022. Learning to retrieve prompts for in-context learning. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, NAACL 2022, Seattle, WA, United States, July 10-15, 2022, pages 2655-2671. Association for Computational Linguistics.  
+Peter Shaw, Ming-Wei Chang, Panupong Pasupat, and Kristina Toutanova. 2021. Compositional generalization and natural language variation: Can a semantic parsing approach handle both? In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, ACL/IJCNLP 2021, (Volume 1: Long Papers), Virtual Event, August 1-6, 2021, pages 922-938. Association for Computational Linguistics.  
+Richard Socher, Alex Perelygin, Jean Wu, Jason Chuang, Christopher D. Manning, Andrew Y. Ng, and Christopher Potts. 2013. Recursive deep models for semantic compositionality over a sentiment treebank. In Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing, EMNLP 2013, 18-21 October 2013, Grand Hyatt Seattle, Seattle, Washington, USA, A meeting of SIG-DAT, a Special Interest Group of the ACL, pages 1631-1642. ACL.  
+Mingyang Song, Mao Zheng, and Xuan Luo. 2025. Can many-shot in-context learning help Ilms as evaluators? A preliminary empirical study. In Proceedings of the 31st International Conference on Computational Linguistics, COLING 2025, Abu Dhabi, UAE, January 19-24, 2025, pages 8232-8241. Association for Computational Linguistics.  
+Alon Talmor, Jonathan Herzig, Nicholas Lourie, and Jonathan Berant. 2019. Commonsenseqa: A question answering challenge targeting commonsense knowledge. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, NAACL-HLT 2019, Minneapolis, MN, USA, June 2-7, 2019, Volume 1 (Long and Short Papers), pages 4149-4158. Association for Computational Linguistics.  
+Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In Advances in Neural Information Processing Systems 30: Annual Conference on Neural Information Processing Systems 2017, December 4-9, 2017, Long Beach, CA, USA, pages 5998-6008.  
+Xinyi Wang, Wanrong Zhu, Michael Saxon, Mark Steyvers, and William Yang Wang. 2023. Large language models are latent variable models: Explaining and finding good demonstrations for in-context learning. In Advances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, NeurIPS 2023, New Orleans, LA, USA, December 10 - 16, 2023.
+
+Adina Williams, Nikita Nangia, and Samuel R. Bowman. 2018. A broad-coverage challenge corpus for sentence understanding through inference. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, NAACL-HLT 2018, New Orleans, Louisiana, USA, June 1-6, 2018, Volume 1 (Long Papers), pages 1112-1122. Association for Computational Linguistics.  
+Tomer Wolfson, Mor Geva, Ankit Gupta, Yoav Goldberg, Matt Gardner, Daniel Deutch, and Jonathan Berant. 2020. Break it down: A question understanding benchmark. Trans. Assoc. Comput. Linguistics, 8:183-198.  
+An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, Huan Lin, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Yang, Jiaxi Yang, Jingren Zhou, Junyang Lin, Kai Dang, Keming Lu, Keqin Bao, Kexin Yang, Le Yu, Mei Li, Mingfeng Xue, Pei Zhang, Qin Zhu, Rui Men, Runji Lin, Tianhao Li, Tingyu Xia, Xingzhang Ren, Xuancheng Ren, Yang Fan, Yang Su, Yichang Zhang, Yu Wan, Yuqiong Liu, Zeyu Cui, Zhenru Zhang, and Zihan Qiu. 2024. Qwen2.5 technical report. CoRR, abs/2412.15115.  
+Jiacheng Ye, Zhiyong Wu, Jiangtao Feng, Tao Yu, and Lingpeng Kong. 2023. Compositional exemplars for in-context learning. In International Conference on Machine Learning, ICML 2023, 23-29 July 2023, Honolulu, Hawaii, USA, volume 202 of Proceedings of Machine Learning Research, pages 39818-39833. PMLR.  
+Rowan Zellers, Ari Holtzman, Yonatan Bisk, Ali Farhadi, and Yejin Choi. 2019. Hellaswag: Can a machine really finish your sentence? In Proceedings of the 57th Conference of the Association for Computational Linguistics, ACL 2019, Florence, Italy, July 28- August 2, 2019, Volume 1: Long Papers, pages 4791-4800. Association for Computational Linguistics.  
+Yiming Zhang, Shi Feng, and Chenhao Tan. 2022. Active example selection for in-context learning. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, EMNLP 2022, Abu Dhabi, United Arab Emirates, December 7-11, 2022, pages 9134-9148. Association for Computational Linguistics.  
+Bo Zhao, Konda Reddy Mopuri, and Hakan Bilen. 2021. Dataset condensation with gradient matching. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021. OpenReview.net.
+
+# A Dataset Examples
+
+We illustrate examples on all dataset in Table 8. We illustrate a query prompt with random 4-shot demonstrations on each dataset, where the demonstration answers are highlighted in red.
+
+# B Case Study: Poor Performance of BM25-Major and Latent-Bayesian
+
+Since BM25-Major and Latent-Bayesian perform poorly, even underperforming Random selection, we examine the demonstrations selected by these two methods. We found they tend to be trapped in specific patterns and lose diversity in selection. We illustrate their demonstration cases and discuss these phenomena in Tables 9 and 10 respectively.
+
+# C Full Results
+
+We illustrate all results on Qwen2.5 series models and Llama3 series models for 128-shot scenarios in Table 11, and results averaged over 4-shot to 128-shot scenarios in Table 12.
+
+# D Implementation Details and Full Results on Closed-Source LLMs
+
+We evaluate demonstrations selected by Random, Best-of-N, and CLG over 5 closed-source LLMs: Qwen-turbo, GLM-4-flash, Doubao-pro-32k, GPT-4o-mini, and DeepSeek-V3. These models are primarily designed for chat-based interactions with users rather than text completion. To adapt them for our evaluation, we begin by presenting them with a system message that outlines the task requirements:
+
+```txt
+You are an In-Context Learner to learn from task demonstrations provided in context.   
+User will present you a prompt, composed of several demonstrations (query  $^+$  response) and a test query (only query).   
+You need to under stand the task concept and response format through the demonstrations, and return your response to the test query.   
+Example of a valid JSON response: ":"json { "response": "...", }""
+```
+
+Next, we construct a user message containing the demonstrations (limited to 7000 tokens) and the test query, formatted as a JSON object. An example with 4-shot demonstrations is shown below:
+
+```json
+```
+"\"json"
+{
+    "demonstrations": [
+        "query": "as giddy and whimsical and relevant today as it was 270 years ago . It is",
+        "response": "great"
+    ]
+}
+```
+
+```json
+},   
+{ "query": "after an uncertain start, murder hits and generally sustains a higher plateau with bullock 's memorable first interrogation of gosling . It is ", "response": "good" }, { "query": "an unbelievably stupid film , though occasionally fun enough to make you forget its absurdity . It is ", "response": "bad" }, { "query": "even though it is infused with the sensibility of a video director , it does n't make for completely empty entertainment It is ", "response": "OK" } ], "test-query": { "query": "in his first stab at the form , jacquot takes a slightly anarchic approach that works only sporadically . It is " } }
+```
+
+After chatting with closed-source models by these messages, we extract and evaluate responses from their JSON-formatted outputs. The full evaluation results are illustrated in Table 13.
+
+# E Scaling Trends of Random Selection on All Datasets
+
+In Figure 7, we present the ICL performances of all models across various demonstration sizes, ranging from 4-shot to 128-shot, using randomly selected demonstrations. The results indicate that datasets for open-ended tasks tend to consistently benefit from more in-context demonstrations. This trend is likely attributable to the fact that open-ended tasks have significantly longer answers (see Appendix A), which provide richer information to learn from context and more space for improvements in prediction.
+
+Qwen2.5-7B exhibits unusual performance on MNLI, significantly outperforming other models in the 4-shot setting but dropping to the lowest performance in the 128-shot setting. This anomaly is validated through comparisons across models and demonstration selection methods. As shown in Fig. 8, this behavior appears unique to Qwen2.5-7B and irrelevant to demonstration selection, suggesting Qwen2.5-7B may be a distinct case on MNLI.
+
+<table><tr><td>Dataset</td><td>Query Prompt with 4-shot In-Context Demonstrations selected by Random</td></tr><tr><td>SST-5</td><td>more of the same from taiwanese auteur tsai ming-liang , which is good news to anyone who &#x27;s fallen under the sweet , melancholy spell of this unique director &#x27;s previous films . It is goodit &#x27;s a great performance and a reminder of dickens &#x27; grandeur . It is greata series of escapades demonstrating the adage that what is good for the goose is also good for the gander , some of which occasionally amuses but none of which amounts to much of a story . It is badone just waits grimly for the next shock without developing much attachment to the characters . It is OKin his first stab at the form . jacquot takes a slightly anarchic approach that works only sporadically . It is</td></tr><tr><td>MNLI</td><td>Instead of ancient artefacts it shows the lifestyle and achievements of myriad Jewish communities around the globe through high-tech h audio-visual displays, hands-on exhibits, scale models (many of which are exquisite), and reconstructions. Can we say &quot;The Jewish communities had many ancient artifacts on display.&quot;? No Theirs is now the dominant right-wing critique of integrationist programs. Can we say &quot;Right wing politics is praised by the integrationist.&quot;? No Lorenzo the Magnificent and brother Giuliano lie in simple tombs beneath the sculptor&#x27;s Madonna and Child, flanked by lesser artists&#x27; statues of the family patron saints Cosmas and Damian. Can we say &quot;Lorenzo and Giuliano were related to one another.&quot;? Yes The difference, if any, between the reacquisition price and the net carrying value of the extinguished debt should be recognized as a loss or gain in accounting for interest on Treasury debt. Can we say &quot;The difference is not recognized.&quot;? No The new rights are nice enough Can we say &quot;Everyone really likes the newest benefits&quot;?</td></tr><tr><td>CMSQA</td><td>Question: Bill sits down on a whoopee cushion, what sound does he make when he sits? (A) fall asleep (B) flatulence (C) sigh of relief (D) medium (E) comfort Answer: BQuestion: What is likely heard by those going to a party? (A) smoking pot (B) happiness (C) laughter (D) babies (E) meet new people Answer: CQuestion: A handsome prince is a stock character common to what? (A) england (B) fairy tale (C) castle (D) palace (E) court Answer: BQuestion: What covers the largest percentage of the pacific northwest? (A) united states (B) united states (C) washington (D) oregon (E) british columbia Answer: BQuestion: A revolving door is convenient for two direction travel, but it also serves as a security measure at a what? (A) bank (B) library (C) department store (D) mall (E) new york Answer:</td></tr><tr><td>HeSwag.</td><td>Choose an ending: The player is holding a bat walking in the field. The player... (A) in front stands across the screen and the ball dunks him to the ground. (B) makes a goal and the male scores. (C) runs down a purple net towards the ball. (D) begins to hit the ball. Answer: BChoose an ending: A woman sitting on the floor speaks to the camera. She... (A) begins blowing cover on the phone. (B) pulls a harmonica and ties the end of the violin. (C) begins to play a saxophone in the bedroom. (D) holds onto a cat in front of her. Answer: DChoose an ending: Someone jumps out the living room window. He... (A) dives for her pipe. (B) tuxedos her free eyes. (C) sees the girl chasing after him. (D) is as grim as he can. Answer: CChoose an ending: A group sits on horses as the stand and take a rest. The group... (A) continues jumping around a gym as the crowd claps. (B) slowly slowly make their way to the other side of the pool on the level. (C) rides on the horses along a trail. (D) pass between horses at the left. Answer: CChoose an ending: Students lower their eyes nervously. She... (A) pats her shoulder, then saunters toward someone. (B) turns with two students. (C) walks slowly towards someone. (D) wheels around as her dog thunders out. Answer:</td></tr><tr><td>GeoQ.</td><td>how many rivers are there in m0 answer: count(intersection(river,loc_2(m0)))through which states does the m0 flow answer: intersection(state,traverse_1(m0))through which states does the m0 run answer: intersection(state,traverse_1(m0))what is the most populous city in m0 answer: largest_one(population_1,intersection(city,loc_2(m0)))name all the rivers in m0 answer:</td></tr><tr><td>NL2Bash</td><td>Get the path of running Apache ps -ef | grep apacheGets domain name from dig reverse lookup. $dig -x 8.8.8.8 | grep PTR | grep -o google.*Find all *.xml files under current directory find -name *.xmlDisables shell option &#x27;nullglob&#x27;. shopt -u nullglobAdd executable permission to &quot;pretty-print&quot;</td></tr><tr><td>Break</td><td>when did vincent von gogh die? 1# return vincent von gogh 2# ) return when did #1 dieDoes the Bronx have more non-Hispanic whites, or Hispanic whites? 1# ) return the Bronx 2# ) return non-Hispanic whites in #1 3# ) return Hispanic whites in #1 4# )return number of #2 5# ) return number of #3 6# ) return which is highest of #4 , #5What is the oldest log id and its corresponding problem id? 1# ) return log ids 2# ) return #1 that is the oldest 3# ) return corresponding problem id of #2 4# ) return #2 , #3How many years apart were the British Saloon Car Championship season wins after 1970? 1# ) return the British Saloon Car Championship 2# ) return season wins of #1 3# ) return years of #2 4# ) return #3 that were after 1970 5# ) return the difference of #4what flights are available tomorrow from denver to philadelphia 1#</td></tr><tr><td>MTOP</td><td>Ask Rob how bad the traffic is [IN:SEND_MESSAGE [SL:RECIPIENT Robot ] [SL:CONTENT真切 how bad the traffic is ]]Can I have the headlines please [IN:GET_STORIES_NEWS [SL:NEWS_TYPE headlines ]]what are the headlines in ohio news? [IN:GET_STORIES_NEWS [SL:NEWS_TYPE headlines ] [SL:NEWS_TOPIC ohio ] [SL:NEWS_TYPE news ]]What is the most recent news regarding local politics [IN:GET_STORIES_NEWS [SL:DATE_TIME the most recent ] [SL:NEWS_TYPE news ] [SL:NEWS_CATEGORY local politics ]]call Nicholas and Natasha</td></tr><tr><td>SMCal.</td><td>What about with Kaitlin Taylor? (Yield (Execute (NewClobber (refer ((Dynamic ActionIntensionConstraint)) ((Recipient) ConstraintTypeIntension) (intension (RecipientWithNameLike ((Recipient) EmptyStructConstraint) (PersonName.apply &quot;Kaitlin Taylor&quot;))))Sorry I meant the weekend? (Yield (Execute (ReviseConstraint (refer ((Dynamic roleConstraint (Path.apply &quot;output&quot;)) ((Event) ConstraintTypeIntension) (Event.start ? (DateTime.date_? (ThisWeekend))))CHECK FOR ALBERT&#x27;S BIRTHDAY EVENT (Yield (FindEventWrapperWithDefaults (Eventsubject ? (? = &quot;ALBERT&#x27;S BIRTHDAY&quot;))))Add meetings same time everyday for this week. (FenceRecurring)Change the reservation for tonight to 6 people from 4.</td></tr></table>
+
+Table 8: Dataset cases with random 4-shot in-context demonstrations.
+
+<table><tr><td>Dataset</td><td>Query Prompt with 4-shot In-Context Demonstrations selected by Latent-Bayesian</td></tr><tr><td>SST-5</td><td>renner? It is OK
+apart from its own considerable achievement, metropolis confirms tezuka&#x27;s status as both the primary visual influence on the animé tradition and its defining philosophical conscience. It is OK
+... salaciously simplistic. It is OK
+... would be a total loss if not for two supporting performances taking place at the movie&#x27;s edges. It is OK
+in his first stab at the form, jacquot takes a slightly anarchic approach that works only sporadically. It is</td></tr><tr><td>HeSwag.</td><td>Choose an ending: Richard Parker frantically swipes at the small fish as someone shields his face. The fluttering silver creatures... (A) spoon in around him. (B) inserting its wands into their palm. (C) pelt his exposed torso. (D) fly over him, catching and grabbing the sheet of flame. Answer: C
+Choose an ending: The sheer atheism of his movement keeps him ahead of the choppers as he leads them on. Back at the window, someone... (A) ducks down, gathers his cube and hunches at one. (B) groggily wakes with a crowbar. (C) gazes into a narrow gap between a tall arch. (D) leans her head against the rotting frame. Answer: D
+Choose an ending: The door closes after him, and unable to stop, the keys herd into it. Huge broken stone statues... (A) lie on an maize field. (B) are about to emerge, which painted walls. (C) lie about the floor. (D) are drawn around the arena empty gleaming guns. Answer: C
+Choose an ending: A strong wind flings leaves high into the air and the hedges advance. As the hedges close in behind them, they both... (A) climb the rock - wakeboard on the side of the structure, to avoid it. (B) sip of the wand. (C) move, still throwing his legs to each other. (D) pose, once aftermath from all angles. Answer: D
+Choose an ending: Students lower their eyes nervously. She... (A) pats her shoulder, then saunters toward someone. (B) turns with two students. (C) walks slowly towards someone. (D) wheels around as her dog thunders out. Answer:</td></tr></table>
+
+Table 9: Demonstration cases selected by Latent-Bayesian. On SST-5, Latent-Bayesian tends to exclusively select demonstrations with the neutral label OK. On HellaSwag, it predominantly selects demonstrations with answers C or D. Specifically, for 128-shot demonstrations on HellaSwag, Latent-Bayesian selected answers A, B, C, and D in proportions of 8, 0, 45, and 75 respectively. This is in contrast to the training set of 50,000 examples where the four answers are distributed equally (12,  $500 \pm 68$  each).  
+
+<table><tr><td>Dataset</td><td>Query Prompt with 4-shot In-Context Demonstrations selected by BM25-Major</td></tr><tr><td>Break</td><td>What shape is the object is a different shape and size than the other objects? 1(#) return objects 2(#) return shapes of #1 3(#) return sizes of #1 4(#) return number of #1 for each #2 5(#) return number of #1 for each #3 6(#) return #2 where #4 is one 7(#) return #3 where #5 is one 8(#) return #1 where #2 is #6 9(#) return #1 where #3 is #7 10(#) return #1 in both #8 and #9 11(#) return the shape of #10 Are all the spheres the same material? 1(#) return spheres 2(#) return the materials of #1 3(#) return number of #1 for each #2 4(#) return #2 where #3 is the highest 5(#) return #1 where #2 is #4 6(#) return number of #5 7(#) return number of #1 8(#) return if #6 and #7 are equal What is the count and code of the job with the most employee? 1(#) return employees 2(#) return jobs of #1 3(#) return number of #1 for each #2 4(#) return the highest of #3 5(#) return #2 where #3 is #4 6(#) return the code of #5 7(#) return #4 , #6 What color is the object that is a different material than the others? 1(#) return objects 2(#) return materials of #1 3(#) return the number of #1 for each #2 4(#) return #2 where #3 is one 5(#) return #1 where #2 is #4 6(#) return the color of #5 what flights are available tomorrow from denver to philadelphia 1#</td></tr><tr><td>MTOP</td><td>remind me to call my father for fathers day [IN:CREATE_REMINDER [SL:PERSON_REMINDED me] [SL:TODO [IN:GET_TOPO [SL:TODO [IN:CREATE_CALL [SL:CONTACT [IN:GET CONTACT [SL:CONTACT RELATED my] [SL:TYPE_RELATION fathor ] ]] [SL:TODO for fathers day ]]]] Set a reminder to call my husband when I get home [IN:CREATE_REMINDER [SL:TODO [IN:GET_TOPO [SL:TODO [IN:CREATE CALL [SL:CONTACT [IN:GET CONTACT [SL:CONTACT RELATED my] [SL:TYPE_RELATION husband ] ]] [SL:LOCATION [IN:GET_LOCATION ]]]] Remind me to call my brother after work. [IN:CREATE_REMINDER [SL:PERSON_REMINDED me] [SL:TODO [IN:GET_TOPO [SL:TODO [IN:CREATE CALL [SL:CONTACT [IN:GET CONTACT [SL:CONTACT RELATED my] [SL:TYPE_RELATION brother ] ]] [SL:TODO after work ]]]] Remind me to call my father today. [IN:CREATE_REMINDER [SL:PERSON_REMINDED me] [SL:TODO [IN:CREATE CALL [SL:CONTACT [IN:GET CONTACT [SL:CONTACT RELATED my] [SL:TYPE_RELATION father ] ]] [SL:DATE_TIME today ]] call Nicholas and Natasha</td></tr><tr><td>SMCal.</td><td>What is on my calendar tomorrow? I want to have lunch with Sara, Barack and Monica. (do (Yield (FindEventWrapperWithDefaults (EventOnDate (Tomorrow) (Event EmptyStructConstraint)))) (Yield (CreateCommitEventWrapper (CreatePreflightEventWrapper &amp; (&amp; (Eventsubject_? (= &quot;lunch&quot;)) (Event.start_? (DateTime.date_? (= (Tomorrow)))) (Eventattendees_? (&amp; (AttendeeListHasRecipient (Execute (refer (extensionConstraint (RecipientWithNameLike (Recipient) EmptyStructConstraint) (PersonName.apply &quot;Sara&quot;)))) (AttendeeListHasRecipient (Execute (refer (extensionConstraint (RecipientWithNameLike (Recipient) EmptyStructConstraint) (PersonName.apply &quot;Barack&quot;)))))) (AttendeeListHasRecipient (Execute (refer (extensionConstraint (RecipientWithNameLike (Recipient) EmptyStructConstraint) (PersonName.apply &quot;Monica&quot;)))))) I would like Tuesday. What time is my team sync with Abby, Eli and Kim? (do (Yield (Execute (ChooseCreateEventFromConstraint (Event.start_? (DateTime.date_? (Date.dayOfWeek_? (= (Tuesday)))) (refer ((Dynamic) ActionIntensionConstraints))) (Yield (Event.start (singleton (QueryEventResponseResults (FindEventWrapperWithDefaults &amp; (Eventsubject_? (= &quot;team sync&quot;) (Eventattendees_? (&amp; (AttendeeListHasRecipientConstraint (RecipientWithNameLike (Recipient) EmptyStructConstraint) (PersonName.apply &quot;Abby&quot;)))) (AttendeeListHasRecipientConstraint (RecipientWithNameLike (Recipient) EmptyStructConstraint) (PersonName.apply &quot;Elli&quot;)))) (AttendeeListHasRecipientConstraint (RecipientWithNameLike (Recipient) EmptyStructConstraint) (PersonName.apply &quot;Kim&quot;))))) Change the lunch on Friday to being at 1pm. Also invite Sally, Alexa, and Sandra to the lunch. (Yield (UpdateCommitEventWrapper (UpdatePreflightEventWrapper (Event.id (singleton (QueryEventResponseresults (FindEventWrapperWithDefaults &amp; (Event.subleet_? (= &quot;lunch&quot;)) (Event.start_? (Date.dayofWeek_? (= (Friday)))))) (&amp; (Event.start_? (DateTime.time_? (= (NumberPM L))) (Eventattendees_? (&amp; (AttendeeListHasRecipient (Execute (refer (extensionConstraint (RecipientWithNameLike (Recipient) EmptyStructConstraint) (PersonName.apply &quot;Sally&quot;)))) (AttendeeListHasRecipient (Execute (refer (extensionConstraint (RecipientWithNameLike (Recipient) EmptyStructConstraint) (PersonName.apply &quot;Alexa&quot;)))))) (AttendeeListHasRecipient (Execute (refer (extensionConstraint (RecipientWithNameLike (Recipient) EmptyStructConstraint) (PersonName.apply &quot;Sandra&quot;)))))) Put a lunch meeting with Abby, Kim, and Jesse at Taco Bell tomorrow (Yield (CreateCommitEventWrapper (CreatePreflightEventWrapper &amp; (&amp; (Event.subleet_? (= &quot;lunch&quot; (Event.start_? (DateTime.date_? (= (Tomorrow)))) (Event.location_? (= LocationKeyphrase.apply &quot;Taco Bell&quot;)))) (Eventattendees_? (&amp; (AttendeeListHasRecipient (Execute (refer (extensionConstraint (RecipientWithNameLike (Recipient) EmptyStructConstraint) (PersonName.apply &quot;Abby&quot;)))) (AttendeeListHasRecipient (Execute (refer (extensionConstraint (RecipientWithNameLike (Recipient) EmptyStructConstraint) (PersonName.apply &quot;Kim&quot;))))) (AttendeeListHasRecipient (Execute (refer (extensionConstraint (RecipientWithNameLike (Recipient) EmptyStructConstraint) (PersonName.apply &quot;Jesse&quot;)))))) Change the reservation for tonight to 6 people from 4.</td></tr></table>
+
+Table 10: Demonstration cases selected by BM25-Major. BM25-Major tends to select significantly longer demonstrations. This can be attributed to the fact that longer examples tend to have higher BM25 scores.
+
+<table><tr><td>Method</td><td>SST-5</td><td>MNLI</td><td>CMSQA</td><td>HeSwag.</td><td>GeoQ.</td><td>NL2Bash</td><td>Break</td><td>MTOP</td><td>SMCal.</td><td>Average</td></tr><tr><td colspan="11">Llama3-8B</td></tr><tr><td>Random</td><td>37.75</td><td>54.32</td><td>72.99</td><td>41.51</td><td>49.14</td><td>20.14</td><td>27.19</td><td>34.33</td><td>29.01</td><td>40.71</td></tr><tr><td>BM25-Major</td><td>38.78</td><td>40.62</td><td>70.84</td><td>39.62</td><td>35.71</td><td>28.66</td><td>2.16</td><td>5.37</td><td>6.56</td><td>29.81</td></tr><tr><td>BGE-KMeans</td><td>37.15</td><td>61.85</td><td>73.55</td><td>36.02</td><td>58.21</td><td>19.90</td><td>23.98</td><td>33.20</td><td>28.69</td><td>41.39</td></tr><tr><td>Best-of-N</td><td>39.33</td><td>51.92</td><td>73.38</td><td>40.58</td><td>53.93</td><td>26.78</td><td>25.88</td><td>35.48</td><td>30.55</td><td>41.98</td></tr><tr><td>EPR-KMeans</td><td>37.60</td><td>53.98</td><td>73.87</td><td>39.99</td><td>49.64</td><td>17.65</td><td>30.09</td><td>36.51</td><td>29.63</td><td>41.00</td></tr><tr><td>Latent-Bayesian</td><td>37.15</td><td>44.04</td><td>73.22</td><td>33.81</td><td>41.79</td><td>24.72</td><td>25.66</td><td>32.30</td><td>18.13</td><td>36.76</td></tr><tr><td>CLG (ours)</td><td>39.96</td><td>61.98</td><td>73.96</td><td>46.01</td><td>58.21</td><td>33.83</td><td>29.06</td><td>37.18</td><td>27.39</td><td>45.29</td></tr><tr><td colspan="11">Llama3-70B</td></tr><tr><td>Random</td><td>43.71</td><td>61.83</td><td>84.73</td><td>76.11</td><td>73.36</td><td>29.72</td><td>35.55</td><td>44.66</td><td>36.68</td><td>54.04</td></tr><tr><td>BM25-Major</td><td>43.69</td><td>51.36</td><td>82.23</td><td>72.01</td><td>43.21</td><td>27.92</td><td>15.05</td><td>8.41</td><td>14.38</td><td>39.81</td></tr><tr><td>BGE-KMeans</td><td>46.41</td><td>65.09</td><td>83.78</td><td>73.90</td><td>84.64</td><td>28.39</td><td>35.13</td><td>44.79</td><td>38.73</td><td>55.65</td></tr><tr><td>Best-of-N</td><td>43.69</td><td>70.92</td><td>84.19</td><td>74.57</td><td>80.71</td><td>29.76</td><td>35.86</td><td>43.40</td><td>37.66</td><td>55.64</td></tr><tr><td>EPR-KMeans</td><td>43.05</td><td>49.78</td><td>83.62</td><td>71.32</td><td>78.21</td><td>22.92</td><td>38.39</td><td>50.43</td><td>38.10</td><td>52.87</td></tr><tr><td>Latent-Bayesian</td><td>46.59</td><td>69.84</td><td>84.60</td><td>70.09</td><td>62.14</td><td>29.20</td><td>32.59</td><td>41.21</td><td>23.13</td><td>51.04</td></tr><tr><td>CLG (ours)</td><td>48.32</td><td>76.37</td><td>84.52</td><td>80.77</td><td>84.64</td><td>29.59</td><td>37.33</td><td>47.74</td><td>40.35</td><td>58.85</td></tr><tr><td colspan="11">Qwen2.5-3B</td></tr><tr><td>Random</td><td>19.05</td><td>51.15</td><td>68.76</td><td>35.30</td><td>47.93</td><td>14.08</td><td>23.29</td><td>25.94</td><td>19.07</td><td>33.84</td></tr><tr><td>BM25-Major</td><td>26.70</td><td>45.33</td><td>68.47</td><td>33.79</td><td>36.07</td><td>28.53</td><td>3.56</td><td>5.01</td><td>0.12</td><td>27.51</td></tr><tr><td>BGE-KMeans</td><td>26.43</td><td>40.70</td><td>73.79</td><td>32.72</td><td>51.43</td><td>19.74</td><td>19.83</td><td>26.26</td><td>22.20</td><td>34.79</td></tr><tr><td>Best-of-N</td><td>21.25</td><td>56.02</td><td>71.50</td><td>41.10</td><td>55.00</td><td>21.36</td><td>23.45</td><td>26.31</td><td>21.15</td><td>37.46</td></tr><tr><td>EPR-KMeans</td><td>19.35</td><td>54.53</td><td>70.93</td><td>33.61</td><td>49.29</td><td>11.79</td><td>23.80</td><td>30.07</td><td>17.66</td><td>34.56</td></tr><tr><td>Latent-Bayesian</td><td>21.07</td><td>44.66</td><td>69.62</td><td>25.68</td><td>28.93</td><td>11.26</td><td>20.91</td><td>23.80</td><td>11.56</td><td>28.61</td></tr><tr><td>CLG (ours)</td><td>24.07</td><td>54.42</td><td>72.56</td><td>44.77</td><td>52.50</td><td>17.62</td><td>24.90</td><td>26.49</td><td>19.40</td><td>37.41</td></tr><tr><td colspan="11">Qwen2.5-7B</td></tr><tr><td>Random</td><td>26.48</td><td>31.88</td><td>84.88</td><td>71.38</td><td>35.43</td><td>16.93</td><td>28.22</td><td>32.72</td><td>27.99</td><td>39.55</td></tr><tr><td>BM25-Major</td><td>29.34</td><td>31.88</td><td>83.78</td><td>71.50</td><td>41.43</td><td>32.27</td><td>8.65</td><td>7.79</td><td>11.98</td><td>35.40</td></tr><tr><td>BGE-KMeans</td><td>26.25</td><td>31.88</td><td>85.26</td><td>74.10</td><td>56.79</td><td>19.46</td><td>25.28</td><td>33.20</td><td>30.73</td><td>42.55</td></tr><tr><td>Best-of-N</td><td>24.80</td><td>31.88</td><td>84.77</td><td>72.69</td><td>26.07</td><td>15.72</td><td>27.80</td><td>32.75</td><td>28.71</td><td>38.35</td></tr><tr><td>EPR-KMeans</td><td>24.70</td><td>31.94</td><td>85.26</td><td>71.33</td><td>51.79</td><td>15.44</td><td>31.17</td><td>35.30</td><td>25.43</td><td>41.37</td></tr><tr><td>Latent-Bayesian</td><td>12.62</td><td>31.88</td><td>84.93</td><td>60.58</td><td>21.43</td><td>13.31</td><td>26.21</td><td>26.13</td><td>17.66</td><td>32.75</td></tr><tr><td>CLG (ours)</td><td>24.89</td><td>31.94</td><td>85.34</td><td>71.76</td><td>50.00</td><td>19.54</td><td>28.27</td><td>32.98</td><td>28.03</td><td>41.42</td></tr><tr><td colspan="11">Qwen2.5-14B</td></tr><tr><td>Random</td><td>29.55</td><td>56.76</td><td>82.29</td><td>55.53</td><td>61.71</td><td>28.00</td><td>31.41</td><td>38.74</td><td>33.73</td><td>46.41</td></tr><tr><td>BM25-Major</td><td>19.53</td><td>55.98</td><td>79.28</td><td>46.52</td><td>51.79</td><td>41.82</td><td>14.85</td><td>10.02</td><td>12.82</td><td>36.96</td></tr><tr><td>BGE-KMeans</td><td>34.79</td><td>57.31</td><td>82.88</td><td>60.61</td><td>74.64</td><td>31.99</td><td>30.13</td><td>40.89</td><td>31.93</td><td>49.46</td></tr><tr><td>Best-of-N</td><td>28.97</td><td>55.17</td><td>82.56</td><td>52.92</td><td>59.64</td><td>36.30</td><td>30.44</td><td>42.24</td><td>32.95</td><td>46.80</td></tr><tr><td>EPR-KMeans</td><td>18.07</td><td>56.24</td><td>82.06</td><td>51.67</td><td>58.93</td><td>24.39</td><td>32.41</td><td>43.18</td><td>33.95</td><td>44.54</td></tr><tr><td>Latent-Bayesian</td><td>22.16</td><td>56.20</td><td>81.98</td><td>63.08</td><td>47.50</td><td>24.77</td><td>30.39</td><td>28.95</td><td>16.71</td><td>41.30</td></tr><tr><td>CLG (ours)</td><td>32.79</td><td>56.30</td><td>83.05</td><td>69.35</td><td>60.36</td><td>42.43</td><td>32.20</td><td>39.51</td><td>34.95</td><td>50.10</td></tr><tr><td colspan="11">Qwen2.5-32B</td></tr><tr><td>Random</td><td>13.93</td><td>61.58</td><td>85.54</td><td>84.62</td><td>72.29</td><td>28.51</td><td>36.24</td><td>43.19</td><td>37.43</td><td>51.48</td></tr><tr><td>BM25-Major</td><td>10.72</td><td>62.23</td><td>85.09</td><td>80.69</td><td>57.86</td><td>41.68</td><td>21.38</td><td>11.23</td><td>21.06</td><td>43.55</td></tr><tr><td>BGE-KMeans</td><td>12.62</td><td>60.96</td><td>85.26</td><td>83.12</td><td>80.00</td><td>28.11</td><td>35.35</td><td>43.76</td><td>38.51</td><td>51.97</td></tr><tr><td>Best-of-N</td><td>17.44</td><td>62.24</td><td>57.25</td><td>85.31</td><td>73.57</td><td>37.11</td><td>34.91</td><td>46.94</td><td>35.36</td><td>50.01</td></tr><tr><td>EPR-KMeans</td><td>14.62</td><td>61.41</td><td>85.67</td><td>83.60</td><td>76.79</td><td>28.35</td><td>38.32</td><td>49.13</td><td>40.98</td><td>53.21</td></tr><tr><td>Latent-Bayesian</td><td>0.00</td><td>58.68</td><td>85.91</td><td>81.66</td><td>58.57</td><td>31.57</td><td>34.88</td><td>33.24</td><td>22.21</td><td>45.19</td></tr><tr><td>CLG (ours)</td><td>16.17</td><td>73.33</td><td>86.40</td><td>86.53</td><td>73.57</td><td>29.97</td><td>36.26</td><td>46.62</td><td>36.75</td><td>53.96</td></tr><tr><td colspan="11">Qwen2.5-72B</td></tr><tr><td>Random</td><td>36.00</td><td>58.20</td><td>87.47</td><td>86.58</td><td>57.29</td><td>33.28</td><td>36.67</td><td>45.32</td><td>38.23</td><td>53.23</td></tr><tr><td>BM25-Major</td><td>37.78</td><td>45.94</td><td>86.65</td><td>85.61</td><td>52.50</td><td>38.11</td><td>22.81</td><td>16.24</td><td>11.57</td><td>44.13</td></tr><tr><td>BGE-KMeans</td><td>37.24</td><td>49.65</td><td>87.71</td><td>86.67</td><td>61.79</td><td>40.86</td><td>35.84</td><td>43.71</td><td>41.29</td><td>53.86</td></tr><tr><td>Best-of-N</td><td>36.60</td><td>60.57</td><td>87.55</td><td>86.83</td><td>61.07</td><td>37.13</td><td>35.73</td><td>46.89</td><td>39.22</td><td>54.62</td></tr><tr><td>EPR-KMeans</td><td>36.15</td><td>58.41</td><td>87.96</td><td>86.22</td><td>62.86</td><td>26.53</td><td>39.12</td><td>47.53</td><td>39.65</td><td>53.83</td></tr><tr><td>Latent-Bayesian</td><td>27.25</td><td>47.96</td><td>85.75</td><td>86.72</td><td>27.50</td><td>37.36</td><td>35.35</td><td>5.23</td><td>22.61</td><td>41.75</td></tr><tr><td>CLG (ours)</td><td>38.33</td><td>77.29</td><td>87.71</td><td>87.68</td><td>62.50</td><td>45.19</td><td>37.07</td><td>47.07</td><td>40.16</td><td>58.11</td></tr></table>
+
+Table 11: Main results, merely 128-shot, on all models. The best scores on each model are in bold, and the second-best ones are underlined.
+
+<table><tr><td>Method</td><td>SST-5</td><td>MNLI</td><td>CMSQA</td><td>HeSwag.</td><td>GeoQ.</td><td>NL2Bash</td><td>Break</td><td>MTOP</td><td>SMCal.</td><td>Average</td></tr><tr><td colspan="11">Llama3-8B</td></tr><tr><td>Random</td><td>36.64</td><td>45.95</td><td>72.71</td><td>39.84</td><td>28.26</td><td>19.53</td><td>19.80</td><td>16.87</td><td>14.97</td><td>32.73</td></tr><tr><td>BM25-Major</td><td>32.26</td><td>37.67</td><td>71.51</td><td>38.40</td><td>23.09</td><td>22.65</td><td>1.43</td><td>3.44</td><td>4.72</td><td>26.13</td></tr><tr><td>BGE-KMeans</td><td>37.16</td><td>45.90</td><td>72.99</td><td>35.80</td><td>32.20</td><td>20.23</td><td>14.21</td><td>18.31</td><td>16.72</td><td>32.61</td></tr><tr><td>Best-of-N</td><td>39.15</td><td>48.50</td><td>73.63</td><td>40.51</td><td>35.77</td><td>27.22</td><td>21.14</td><td>20.42</td><td>17.32</td><td>35.96</td></tr><tr><td>EPR-KMeans</td><td>35.12</td><td>43.26</td><td>73.53</td><td>38.75</td><td>30.60</td><td>18.87</td><td>21.03</td><td>19.02</td><td>16.09</td><td>32.92</td></tr><tr><td>Latent-Bayesian</td><td>36.69</td><td>37.88</td><td>72.66</td><td>34.15</td><td>18.15</td><td>16.15</td><td>18.96</td><td>14.21</td><td>9.79</td><td>28.74</td></tr><tr><td>CLG (ours)</td><td>38.72</td><td>51.86</td><td>73.52</td><td>44.27</td><td>34.11</td><td>35.47</td><td>21.41</td><td>20.85</td><td>18.18</td><td>37.60</td></tr><tr><td colspan="11">Llama3-70B</td></tr><tr><td>Random</td><td>42.39</td><td>50.82</td><td>83.59</td><td>72.82</td><td>46.21</td><td>31.42</td><td>29.39</td><td>22.76</td><td>18.72</td><td>44.24</td></tr><tr><td>BM25-Major</td><td>41.10</td><td>44.79</td><td>81.89</td><td>71.01</td><td>35.18</td><td>31.18</td><td>9.83</td><td>5.58</td><td>12.70</td><td>37.03</td></tr><tr><td>BGE-KMeans</td><td>42.72</td><td>53.00</td><td>82.90</td><td>60.84</td><td>50.06</td><td>32.34</td><td>27.11</td><td>22.67</td><td>19.89</td><td>43.50</td></tr><tr><td>Best-of-N</td><td>44.69</td><td>49.75</td><td>83.24</td><td>75.63</td><td>51.79</td><td>33.67</td><td>30.08</td><td>24.41</td><td>21.77</td><td>46.11</td></tr><tr><td>EPR-KMeans</td><td>41.34</td><td>45.13</td><td>83.25</td><td>71.58</td><td>49.16</td><td>25.84</td><td>31.36</td><td>26.56</td><td>21.40</td><td>43.96</td></tr><tr><td>Latent-Bayesian</td><td>44.26</td><td>50.75</td><td>83.30</td><td>55.33</td><td>27.44</td><td>29.23</td><td>25.36</td><td>18.84</td><td>11.53</td><td>38.45</td></tr><tr><td>CLG (ours)</td><td>46.17</td><td>57.28</td><td>83.69</td><td>79.48</td><td>51.61</td><td>36.15</td><td>30.47</td><td>27.18</td><td>24.30</td><td>48.48</td></tr><tr><td colspan="11">Qwen2.5-3B</td></tr><tr><td>Random</td><td>17.81</td><td>49.79</td><td>70.77</td><td>36.96</td><td>25.67</td><td>15.98</td><td>14.48</td><td>13.18</td><td>9.80</td><td>28.27</td></tr><tr><td>BM25-Major</td><td>21.15</td><td>50.18</td><td>68.40</td><td>36.65</td><td>20.83</td><td>18.32</td><td>1.01</td><td>2.26</td><td>0.16</td><td>24.33</td></tr><tr><td>BGE-KMeans</td><td>21.69</td><td>46.76</td><td>72.37</td><td>34.03</td><td>28.87</td><td>17.23</td><td>9.93</td><td>13.23</td><td>12.18</td><td>28.48</td></tr><tr><td>Best-of-N</td><td>23.28</td><td>60.32</td><td>72.73</td><td>41.12</td><td>33.09</td><td>19.18</td><td>15.28</td><td>14.87</td><td>13.57</td><td>32.60</td></tr><tr><td>EPR-KMeans</td><td>15.99</td><td>50.41</td><td>71.55</td><td>34.41</td><td>30.89</td><td>12.99</td><td>14.14</td><td>16.41</td><td>12.45</td><td>28.81</td></tr><tr><td>Latent-Bayesian</td><td>20.85</td><td>51.93</td><td>67.98</td><td>32.97</td><td>13.21</td><td>10.81</td><td>13.37</td><td>9.50</td><td>10.20</td><td>25.65</td></tr><tr><td>CLG (ours)</td><td>22.09</td><td>59.70</td><td>73.59</td><td>39.39</td><td>27.14</td><td>27.20</td><td>16.61</td><td>15.54</td><td>8.24</td><td>32.17</td></tr><tr><td colspan="11">Qwen2.5-7B</td></tr><tr><td>Random</td><td>23.84</td><td>45.36</td><td>84.88</td><td>75.65</td><td>14.69</td><td>17.38</td><td>21.92</td><td>16.09</td><td>14.81</td><td>34.96</td></tr><tr><td>BM25-Major</td><td>24.51</td><td>41.30</td><td>83.58</td><td>73.83</td><td>17.86</td><td>20.58</td><td>5.44</td><td>3.70</td><td>8.76</td><td>31.06</td></tr><tr><td>BGE-KMeans</td><td>24.95</td><td>32.15</td><td>84.98</td><td>75.48</td><td>20.24</td><td>18.97</td><td>17.00</td><td>17.41</td><td>16.95</td><td>34.24</td></tr><tr><td>Best-of-N</td><td>28.82</td><td>43.78</td><td>85.04</td><td>75.38</td><td>15.53</td><td>21.66</td><td>22.18</td><td>18.16</td><td>17.20</td><td>36.42</td></tr><tr><td>EPR-KMeans</td><td>19.94</td><td>44.04</td><td>85.16</td><td>75.38</td><td>18.81</td><td>15.30</td><td>23.80</td><td>19.63</td><td>15.31</td><td>35.26</td></tr><tr><td>Latent-Bayesian</td><td>15.41</td><td>40.92</td><td>84.07</td><td>66.86</td><td>8.33</td><td>16.61</td><td>17.54</td><td>10.48</td><td>13.70</td><td>30.44</td></tr><tr><td>CLG (ours)</td><td>25.10</td><td>47.65</td><td>84.86</td><td>75.84</td><td>23.81</td><td>23.12</td><td>21.66</td><td>18.50</td><td>17.53</td><td>37.56</td></tr><tr><td colspan="11">Qwen2.5-14B</td></tr><tr><td>Random</td><td>25.67</td><td>52.56</td><td>76.58</td><td>64.15</td><td>34.23</td><td>30.25</td><td>24.96</td><td>19.53</td><td>17.27</td><td>38.35</td></tr><tr><td>BM25-Major</td><td>16.09</td><td>48.36</td><td>78.31</td><td>61.59</td><td>36.61</td><td>37.25</td><td>8.92</td><td>6.42</td><td>8.16</td><td>33.52</td></tr><tr><td>BGE-KMeans</td><td>25.81</td><td>52.96</td><td>77.98</td><td>64.23</td><td>38.81</td><td>31.64</td><td>21.83</td><td>20.55</td><td>17.43</td><td>39.03</td></tr><tr><td>Best-of-N</td><td>21.03</td><td>51.27</td><td>76.58</td><td>64.44</td><td>40.18</td><td>40.16</td><td>25.57</td><td>22.59</td><td>18.81</td><td>40.07</td></tr><tr><td>EPR-KMeans</td><td>17.68</td><td>51.29</td><td>76.32</td><td>62.11</td><td>37.14</td><td>23.59</td><td>25.67</td><td>23.32</td><td>18.19</td><td>37.26</td></tr><tr><td>Latent-Bayesian</td><td>3.90</td><td>48.72</td><td>78.33</td><td>64.72</td><td>26.85</td><td>23.10</td><td>20.74</td><td>11.60</td><td>13.13</td><td>32.34</td></tr><tr><td>CLG (ours)</td><td>29.32</td><td>53.83</td><td>79.06</td><td>70.19</td><td>35.24</td><td>41.38</td><td>25.79</td><td>21.93</td><td>20.44</td><td>41.91</td></tr><tr><td colspan="11">Qwen2.5-32B</td></tr><tr><td>Random</td><td>12.85</td><td>62.43</td><td>85.56</td><td>85.48</td><td>41.83</td><td>32.31</td><td>28.97</td><td>22.12</td><td>19.01</td><td>43.40</td></tr><tr><td>BM25-Major</td><td>11.52</td><td>58.64</td><td>85.56</td><td>83.73</td><td>34.46</td><td>35.52</td><td>12.34</td><td>5.60</td><td>15.12</td><td>38.06</td></tr><tr><td>BGE-KMeans</td><td>20.78</td><td>57.03</td><td>85.63</td><td>84.82</td><td>49.40</td><td>31.00</td><td>26.03</td><td>21.89</td><td>19.29</td><td>43.99</td></tr><tr><td>Best-of-N</td><td>17.94</td><td>66.03</td><td>71.27</td><td>85.58</td><td>44.46</td><td>40.06</td><td>28.53</td><td>26.05</td><td>19.69</td><td>44.40</td></tr><tr><td>EPR-KMeans</td><td>11.16</td><td>62.86</td><td>85.52</td><td>85.01</td><td>48.63</td><td>31.08</td><td>29.82</td><td>26.16</td><td>21.72</td><td>44.66</td></tr><tr><td>Latent-Bayesian</td><td>0.05</td><td>57.73</td><td>83.70</td><td>84.42</td><td>27.08</td><td>33.02</td><td>24.21</td><td>13.17</td><td>15.05</td><td>37.60</td></tr><tr><td>CLG (ours)</td><td>19.10</td><td>63.38</td><td>85.46</td><td>86.67</td><td>44.46</td><td>41.55</td><td>28.01</td><td>25.46</td><td>22.66</td><td>46.31</td></tr><tr><td colspan="11">Qwen2.5-72B</td></tr><tr><td>Random</td><td>34.13</td><td>57.59</td><td>87.12</td><td>87.73</td><td>32.49</td><td>35.95</td><td>29.40</td><td>23.37</td><td>19.67</td><td>45.27</td></tr><tr><td>BM25-Major</td><td>28.76</td><td>49.16</td><td>85.72</td><td>86.88</td><td>36.25</td><td>43.46</td><td>16.47</td><td>7.78</td><td>13.41</td><td>40.88</td></tr><tr><td>BGE-KMeans</td><td>38.86</td><td>43.75</td><td>87.24</td><td>87.36</td><td>40.18</td><td>40.63</td><td>27.29</td><td>25.22</td><td>21.57</td><td>45.79</td></tr><tr><td>Best-of-N</td><td>37.54</td><td>62.57</td><td>86.86</td><td>88.02</td><td>39.94</td><td>40.88</td><td>30.37</td><td>25.14</td><td>20.98</td><td>48.03</td></tr><tr><td>EPR-KMeans</td><td>29.31</td><td>56.36</td><td>87.99</td><td>87.44</td><td>42.02</td><td>30.30</td><td>30.67</td><td>25.89</td><td>21.85</td><td>45.76</td></tr><tr><td>Latent-Bayesian</td><td>24.81</td><td>52.97</td><td>85.03</td><td>88.14</td><td>15.42</td><td>34.75</td><td>25.17</td><td>3.31</td><td>15.17</td><td>38.31</td></tr><tr><td>CLG (ours)</td><td>38.30</td><td>64.05</td><td>87.03</td><td>88.32</td><td>41.75</td><td>47.67</td><td>29.09</td><td>25.63</td><td>23.79</td><td>49.51</td></tr></table>
+
+Table 12: Main results, averaged over 4-shot to 128-shot settings, on all models. The best scores on each model are in bold, and the second-best ones are underlined.
+
+<table><tr><td>Method</td><td>SST-5</td><td>MNLI</td><td>CMSQA</td><td>HeSwag.</td><td>GeoQ.</td><td>NL2Bash</td><td>Break</td><td>MTOP</td><td>SMCal.</td><td>Average</td></tr><tr><td colspan="11">Owen-turbo-latest</td></tr><tr><td>Random</td><td>56.87</td><td>80.93</td><td>86.47</td><td>79.07</td><td>73.43</td><td>62.61</td><td>44.20</td><td>28.73</td><td>28.40</td><td>60.08</td></tr><tr><td>±std</td><td>±2.36</td><td>±0.80</td><td>±1.21</td><td>±0.44</td><td>±3.97</td><td>±1.15</td><td>±2.88</td><td>±2.22</td><td>±1.99</td><td>±0.22</td></tr><tr><td>Best-of-N</td><td>52.67</td><td>80.33</td><td>85.33</td><td>80.00</td><td>72.86</td><td>61.79</td><td>42.00</td><td>34.00</td><td>27.67</td><td>59.63</td></tr><tr><td>CLG (ours)</td><td>59.67</td><td>83.33</td><td>87.67</td><td>79.00</td><td>77.14</td><td>64.39</td><td>46.33</td><td>30.67</td><td>28.67</td><td>61.87</td></tr><tr><td colspan="11">GLM-4-flash</td></tr><tr><td>Random</td><td>56.13</td><td>74.54</td><td>87.73</td><td>69.60</td><td>64.07</td><td>61.03</td><td>41.20</td><td>28.20</td><td>23.07</td><td>56.17</td></tr><tr><td>±std</td><td>±1.99</td><td>±1.24</td><td>±0.74</td><td>±13.97</td><td>±1.82</td><td>±1.03</td><td>±3.04</td><td>±0.96</td><td>±2.14</td><td>±2.24</td></tr><tr><td>Best-of-N</td><td>53.00</td><td>76.33</td><td>87.00</td><td>76.67</td><td>66.07</td><td>61.44</td><td>44.67</td><td>32.33</td><td>21.33</td><td>57.65</td></tr><tr><td>CLG (ours)</td><td>56.33</td><td>72.67</td><td>87.33</td><td>77.67</td><td>63.93</td><td>62.24</td><td>51.67</td><td>30.67</td><td>24.00</td><td>58.50</td></tr><tr><td colspan="11">Doubao-pro-32k-0528</td></tr><tr><td>Random</td><td>54.53</td><td>81.00</td><td>84.33</td><td>79.07</td><td>75.64</td><td>64.57</td><td>59.27</td><td>38.33</td><td>28.27</td><td>62.78</td></tr><tr><td>±std</td><td>±1.94</td><td>±1.97</td><td>±0.92</td><td>±0.68</td><td>±2.16</td><td>±0.78</td><td>±1.53</td><td>±2.09</td><td>±1.56</td><td>±0.26</td></tr><tr><td>Best-of-N</td><td>53.67</td><td>78.67</td><td>84.67</td><td>77.67</td><td>74.64</td><td>62.86</td><td>58.67</td><td>37.67</td><td>26.67</td><td>61.69</td></tr><tr><td>CLG (ours)</td><td>55.00</td><td>81.33</td><td>83.67</td><td>77.67</td><td>79.64</td><td>62.82</td><td>60.33</td><td>46.33</td><td>30.00</td><td>64.09</td></tr><tr><td colspan="11">GPT-4o-mini</td></tr><tr><td>Random</td><td>55.93</td><td>76.40</td><td>84.33</td><td>76.53</td><td>70.36</td><td>65.05</td><td>53.73</td><td>32.33</td><td>27.33</td><td>60.22</td></tr><tr><td>±std</td><td>±1.81</td><td>±1.06</td><td>±0.79</td><td>±0.75</td><td>±0.96</td><td>±0.44</td><td>±4.20</td><td>±2.03</td><td>±1.07</td><td>±0.45</td></tr><tr><td>Best-of-N</td><td>53.67</td><td>75.67</td><td>84.33</td><td>77.67</td><td>72.86</td><td>63.87</td><td>55.67</td><td>36.33</td><td>23.33</td><td>60.38</td></tr><tr><td>CLG (ours)</td><td>56.00</td><td>78.67</td><td>86.67</td><td>76.67</td><td>68.93</td><td>65.72</td><td>57.33</td><td>36.67</td><td>26.67</td><td>61.48</td></tr><tr><td colspan="11">DeepSeek-V3</td></tr><tr><td>Random</td><td>58.20</td><td>85.20</td><td>87.80</td><td>80.93</td><td>83.43</td><td>67.77</td><td>59.73</td><td>41.60</td><td>36.80</td><td>66.83</td></tr><tr><td>±std</td><td>±2.57</td><td>±1.17</td><td>±1.00</td><td>±1.65</td><td>±2.48</td><td>±0.60</td><td>±2.09</td><td>±3.71</td><td>±2.90</td><td>±0.52</td></tr><tr><td>Best-of-N</td><td>57.67</td><td>85.67</td><td>86.33</td><td>81.33</td><td>83.93</td><td>65.84</td><td>59.00</td><td>47.33</td><td>34.00</td><td>66.79</td></tr><tr><td>CLG (ours)</td><td>56.33</td><td>85.33</td><td>88.00</td><td>83.33</td><td>87.14</td><td>67.28</td><td>62.00</td><td>45.33</td><td>38.00</td><td>68.08</td></tr></table>
+
+Table 13: Full results of closed-source models under the 128-shot setting. We only evaluate at most 300 instances on each dataset for limited budget. Best scores on each model are in bold.
+
+![](images/a20acc3d6fd3d9dc6c33c1a6aee5b55873c10235408506612c68e784973c5b59.jpg)  
+Figure 7: The scaling trends of random selection with respect to the number of shots on all models. We illustrate the standard deviations under 5 random seeds through error bars. The datasets that can steadily benefit from more shots are in solid colors, and the others are in semi-transparent colors.
+
+![](images/b1e0a7928a0a9d4f1cc2be094907bfbc4fe53f75ed22c3adcfb022bd516bb648.jpg)  
+Figure 8: The performance of different models on MNLI.

2025/Selecting Demonstrations for Many-Shot In-Context Learning via Gradient Matching/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d16041beeb0cd04342cbb046135f4a152d9843d435cad21d2c5c4f499edeaa3a
+size 2467456

2025/Self-Correction is More than Refinement_ A Learning Framework for Visual and Language Reasoning Tasks/bd133b59-8ca8-42c4-b305-e0aca46a0680_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b36665f66da191c61c0343154c64d88bfd123392a2f83520bed65b72aaab7969
+size 1495094

2025/Self-Correction is More than Refinement_ A Learning Framework for Visual and Language Reasoning Tasks/full.md

@@ -0,0 +1,453 @@
+# Self-Correction is More than Refinement: A Learning Framework for Visual and Language Reasoning Tasks
+
+Jiayi He\* Hehai Lin \* Qingyun Wang Yi R. (May) Fung Heng Ji
+
+$^{\text{©}}$ University of Illinois at Urbana-Champaign  $\clubsuit$ Georgia Institute of Technology
+
+$\diamond$ The Hong Kong University of Science and Technology
+
+jhe478@gatech.edu hengji@illinois.edu
+
+# Abstract
+
+While Vision-Language Models (VLMs) have shown remarkable abilities, they invariably generate flawed responses. Self-correction that instructs models to refine their outputs presents a promising solution to this issue. Previous studies have mainly concentrated on Large Language Models (LLMs), while the self-correction abilities of VLMs, particularly concerning both visual and linguistic information, remain largely unexamined. This study investigates the self-correction capabilities of VLMs during both inference and fine-tuning stages. We introduce a Self-Correction Learning (SCL) approach that enables VLMs to learn from their self-generated self-correction data through Direct Preference Optimization (DPO) without relying on external feedback, facilitating self-improvement. Experimental results demonstrate that although VLMs struggle to self-correct effectively during iterative inference without additional fine-tuning and external feedback, they can enhance their performance and avoid previous mistakes through preference fine-tuning when their generated self-correction data are categorized into preferred and disfavored samples. This study emphasizes that self-correction is not merely a refinement process; rather, it should enhance models' reasoning ability through additional training, enabling them to generate high-quality responses directly without further refinement.<sup>1</sup>
+
+# 1 Introduction
+
+Large Language Models (LLMs) have shown exceptional versatility across numerous natural language processing domains (Sun et al., 2025; Chen et al., 2024b). Building upon the foundational capabilities of LLMs, Vision-Language Models (VLMs) (Liu et al., 2024b; Zhu et al., 2024) integrate visual recognition and language under
+
+![](images/0313a57b56b4bdc50869f3e1acd6af5e45464cab41407b98334824141dc10a1c.jpg)
+
+![](images/3e26f88124bc1b5ea01183040075e3a3d0fae450c5fa9a3f4b91c4e36d0c5562.jpg)  
+Figure 1: Comparison of inference-based and training-based self-correction. Inference-based methods aim to refine an initial response over  $K$  iterations, keeping model parameters fixed. Training-based methods focus on training the model to directly generate high-quality initial responses without iterative refinement.
+
+standing by combining pre-trained LLMs and vision models through instruction fine-tuning, leading to significant advancements in multimodal tasks (Huang et al., 2024b; Peng et al., 2024).
+
+Despite their robust vision-language understanding capabilities, VLMs are still prone to generating inaccurate information (Wu et al., 2024b; Fan et al., 2025; He et al., 2025; Zhang et al., 2025). Self-correction, an approach enabling models to identify and rectify mistakes in their outputs (Kamoi et al., 2024), is thus a promising method for enhancing the quality of VLM-generated responses. While previous studies have predominantly focused on self-correction in LLMs, the self-correction capabilities of VLMs remain underexplored. Given that VLMs integrate visual and linguistic information during reasoning, self-correction in VLMs presents additional challenges. This complexity arises from the inherent difficulty of accurately aligning and rectifying multimodal data, underscoring the critical need for a systematic investigation into VLM self-correction to advance their performance in vision-language reasoning tasks.
+
+Existing self-correction strategies focus on the
+
+inference stage without parameter updates (Madaan et al., 2023; Shinn et al., 2023; Li et al., 2024; Liu et al., 2024a). These methods typically involve instructing models to revise their initially generated answers by leveraging self-correction prompts. While prior work has demonstrated the effectiveness of inference-time self-correction in improving LLM performance on reasoning tasks (Madaan et al., 2023), and such approaches offer advantages like no dedicated training cost and operational simplicity, recent studies have reported contradictory findings (Huang et al., 2024a; Xu et al., 2024). This controversy highlights two main shortcomings of inference-time self-correction: (1) Unreliable performance: The effectiveness of self-correction is highly sensitive to the formulation of self-correction prompts (Li et al., 2024). (2) Limitations of models' reasoning abilities: Without additional training to enhance their intrinsic reasoning capabilities, models often struggle to effectively self-correct when confronted with identical challenging tasks (Kamoi et al., 2024).
+
+Besides these two challenges, a crucial distinction between existing self-correction methods during the inference stage and the more natural human self-correction process lies in their correction goals. As illustrated in Figure 1, the former approach focuses on better refinement, specifically enabling the model to correct its initial response through additional revisions (Madaan et al., 2023). Conversely, the latter approach emphasizes better initial generation, aiming to provide the correct answer on the first attempt without subsequent revisions (Zhang et al., 2024a,b; Tong et al., 2024). This discrepancy indicates that current inference-time self-correction methods offer only a temporary solution for rectifying mistakes. While a model can correct mistakes in its generated content through iterative self-correction, its underlying reasoning ability remains unchanged. Consequently, the model may continue to produce suboptimal answers when faced with identical questions in the future, leading to inefficient resource expenditure on repeated refinement. Therefore, we emphasize the ultimate aim of self-correction: not merely to fix initial mistakes but to fundamentally improve the model's capability to generate correct answers directly.
+
+In this paper, we investigate the self-correction capabilities of VLMs through two research questions (RQs): (1) Inference-based self-correction mechanisms: Can VLMs perform self-correction during inference without external feedback?
+
+(2) Training-based self-correction mechanisms: Can VLMs improve their performance by learning from their self-correction process and prevent similar future errors? Both RQs emphasize the concept of self, exploring the intrinsic abilities of VLMs to self-correct autonomously. For inference-based mechanisms, we design three distinct visual self-correction prompts. These prompts guide VLMs to critically examine their initial responses by scrutinizing input image details, understanding contextual cues, and comprehensively interpreting scenes. For training-based mechanisms, we propose Self-Correction Learning (SCL), which utilizes Direct Preference Optimization (DPO) (Rafailov et al., 2023) to empower VLMs to self-improve by learning from their own generated self-correction preference data. The preference dataset, SELFCORSET, is constructed based on the intrinsic self-correction process during inference, where we designate the correct responses as preferred and the incorrect responses as disfavored.
+
+We evaluate the intrinsic self-correction abilities of VLMs and the efficacy of SCL across multiple-choice questions (MCQ) benchmarks. Experimental results demonstrate that while VLMs initially struggle with intrinsic self-correction, they can benefit from learning from their self-correction samples. Specifically, VLMs fine-tuned with SCL are better able to avoid previous errors and exhibit superior performance compared to existing preference optimization methods for VLMs.
+
+Our main contributions are as follows: Firstly, we define the ultimate objective of self-correction as not only rectifying initial errors but fundamentally enhancing a model's ability to generate accurate responses directly. Secondly, we systematically evaluate the self-correction abilities of VLMs during inference by developing three visual self-correction prompts. We further analyze the reliability of inference-based self-correction mechanisms. Thirdly, we introduce Self-Correction Learning (SCL), a novel approach that enables VLMs to self-improve via DPO by leveraging self-generated preference data. Our findings demonstrate the effectiveness of SCL and underscore the advantages of training-based self-correction mechanisms.
+
+# 2 Related Work
+
+Vision-Language Models and Preference Fine-Tuning. Vision-Language Models (VLMs), such as GPT-4o (OpenAI, 2024a), MiniGPT-4 (Zhu
+
+et al., 2024), and LLaVA-1.5 (Liu et al., 2024b), integrate the encoding of visual and textual data to solve various multimodal tasks such as image classification (Peng et al., 2024) and action recognition (Deng et al., 2024). Human preference alignment techniques have been widely applied to VLMs to train these models to generate content aligning with human intentions (Chen et al., 2024e). For instance, Silkie constructs a large-scale multimodal preference dataset annotated by GPT-4V, and distills these preferences into VLMs via Direct Preference Optimization (DPO) (Li et al., 2023b). Preference Optimization in VLLM with AI-Generated Dispreferences (POVID) utilizes preference fine-tuning to mitigate hallucinations (Zhou et al., 2024a). Calibrated Self-Rewarding (CSR) incorporates an iterative learning and rewarding paradigm into preference fine-tuning for modality alignment (Zhou et al., 2024b). Similarly, Inner Monologue Multi-Modal Optimization (IMMO) utilizes a combination of supervised learning and reinforcement learning approaches, performing an inner monologue to enhance model performance on complex vision-language tasks (Yang et al., 2024). While prior studies primarily achieve VLM alignment with human preferences through external feedback from humans or other language models, our work focuses on the self-improvement of VLMs through preference fine-tuning.
+
+Intrinsic Self-Correction in Large Language Models. Self-correction in Large Language Models (LLMs) aims to guide these models in rectifying their flawed generated content, such as harmful outputs (Phute et al., 2024). Intrinsic self-correction refers to a self-correction paradigm during inference where a model revises its output solely by leveraging its inherent capabilities and the input context, without external feedback (Huang et al., 2024a; Kamoi et al., 2024; Liu et al., 2024a). This iterative, multi-turn self-correction process is distinct from single-turn test-time inference of GPT-o1 (OpenAI, 2024b). While recent research has demonstrated the effectiveness of intrinsic self-correction (Madaan et al., 2023; Shinn et al., 2023; Dhuliawala et al., 2024; Li et al., 2025), some studies indicate that LLMs face challenges with it. For instance, intrinsic self-correction may sometimes decrease output quality (Huang et al., 2024a) and potentially introduce bias (Xu et al., 2024). These conflicting results suggest that the self-correction ability of LLMs remains unreliable without external feedback. Previous work has primarily explored
+
+the intrinsic self-correction abilities of LLMs on unimodal tasks like arithmetic reasoning. Our study investigates the intrinsic self-correction abilities of VLMs on visual and language reasoning tasks.
+
+Improvement in Language Models and Vision-Language Models. The enhancement of Language Models (LMs) can be broadly classified into self-improvement and external improvement. Self-improvement in LMs relies on their ability to learn from the data they generate, whereas external improvement involves leveraging external models or tools to enable LMs to learn from provided data. Regarding self-improvement, Huang et al. (2023); Wang et al. (2025b) show that LLMs can self-improve by learning from self-generated data selected via self-consistency (Wang et al., 2023). Zhang et al. (2024b); Wu et al. (2024a); Huang et al. (2025b) leverage the LLM's self-evaluation capability to generate training signals, such as for mitigating factual inaccuracies and enhancing response personalization. Kumar et al. (2024) introduce SCoRe, a multi-turn online reinforcement learning method that achieves self-correction improvements by training on self-generated data and utilizing a two-stage training process to prevent behavior collapse. Additionally, Wang et al. (2025a) propose Self-Improvement Modality Alignment (SIMA) which uses in-context self-critic to improve the modality alignment of VLMs. In SIMA, the model generates two one-turn responses for each question using greedy decoding and temperature sampling. It is then prompted to critique these responses as preferred or disfavored, thereby constructing a preference dataset. Distinguished from SIMA, our work constructs a preference dataset using pairs of initial and revised responses from VLMs, generated during intrinsic self-correction.
+
+For external improvement, Tong et al. (2024) demonstrate that error data generated by strong LLMs can enhance the reasoning capabilities of weaker LLMs. Han et al. (2024) show that small LMs can enhance their self-correction capabilities through instruction fine-tuning. However, this method still requires the generation of self-modification responses. Our work emphasizes that the goal of self-correction is not only to correct mistakes iteratively, but rather to enhance the models' ability to produce correct answers directly. We explore whether VLMs can improve the quality of their responses without further refinement steps by leveraging both their successful and erroneous intrinsic self-correction data for self-improvement.
+
+![](images/4379878a7e6634f519a3d005f3b54d96cc00baa1fb71400ca2abf10cfc590bc8.jpg)  
+Figure 2: SCL begins with intrinsic self-correction applied to the VLM, generating four types of self-correction samples. Correct responses from Type 2 and incorrect responses from Type 3 samples are designated as preferences and disfavors, respectively, to construct the SELFCORSET preference dataset. The VLM then undergoes DPO on SELFCORSET for self-improvement.
+
+# 3 Methodology
+
+The framework of Self-Correction Learning (SCL) is illustrated in Figure 2, comprising three stages: inference, dataset construction, and preference finetuning. In the inference stage, we propose three visual self-correction prompts and investigate the intrinsic self-correction capabilities of VLMs to address RQ1. Subsequently, in the dataset construction and fine-tuning stage, we create SELFCORSET for each VLM based on its intrinsic self-correction and explore RQ2 through DPO.
+
+# 3.1 Inference: Intrinsic Self-Correction
+
+Intrinsic self-correction comprises two distinct stages: initial answer generation and refined answer generation. During the initial answer generation stage, a Standard Prompt (SP) presents the complete question to the VLM, ensuring all requirements are included. For refined answer generation, the VLM engages in an iterative multi-turn process to enhance its initial responses. Due to computational resource constraints, we limit VLMs to a single refinement turn. We apply a critical prompt (Huang et al., 2024a) and develop three visual self-correction prompts to evaluate VLMs' intrinsic self-correction. The critical prompt directly guides models to detect issues in their initial responses. The visual self-correction prompts instruct models to identify problems by scrutinizing input image details, comprehending the portrayed context, and comprehensively interpreting entire scenes. Here are the prompts:
+
+(1) Critical Prompt (CP): Review your previous answer and find problems with your answer. Based on the problems you found, improve your answer.  
+(2) Comprehensive Detail Prompt (VP-1): Review your previous answer and ensure that all rel
+
+evant aspects of the image have been considered. Are there any elements or details that you missed? Based on your review, improve your answer.
+
+(3) Contextual Understanding Prompt (VP2): Review your contextual understanding of the image. Have you correctly interpreted the overall context and purpose of the scene? Based on your review, improve your answer.  
+(4) Comprehensive Scene Analysis Prompt (VP-3): Review your answer and ensure that your understanding of the image is comprehensive and detailed. Are there any aspects of the scene that you have omitted or misinterpreted? Based on your review, improve your answer.
+
+# 3.2 Data Construction: SELFCORSET
+
+We construct the preference dataset SELFCORSET using intrinsic self-correction outputs generated from multiple-choice question (MCQ) samples. These MCQ samples are obtained from several multimodal MCQ datasets, including MMStar (Chen et al., 2024c), MMBench (Liu et al., 2024c), SEEDBench (Li et al., 2023a), ScienceQA (Lu et al., 2022), MMT-Bench (Ying et al., 2024), and MMEvalPro (Huang et al., 2025a). For MMEvalPro, all available samples are utilized. For the remaining datasets, 500 samples are randomly selected for evaluation, with the rest allocated for SELFCORSET construction. Intrinsic self-correction is conducted using the VP-1 prompt, identified as the best-performing self-correction prompt overall (Table 1).
+
+The construction process for the SELFCORSET dataset, using a VLM denoted as  $M$ , is outlined as follows: Given an MCQ sample  $s = \{Question, Image, Answer\}$ , we first perform intrinsic self-correction with VLM  $M$  to obtain a self-correction sample  $s' =$
+
+# Type 3
+
+![](images/148bf887168e0f2b97ac3ef3323f59a760c1f48cc8f1d9dbae39fef35f774cb0.jpg)  
+Figure 3: Two examples of intrinsic self-correction processes generated by InternLM-XComposer-2-7B.
+
+Question The other object that is the same color as the large shiny thing is what shape? A: cube B: sphere C: cylinder
+
+Initial response The other object that is the same color as the large shiny thing is a cylinder. The reasoning behind this is that among the given options, only a cylinder has a similar shape to the large shiny object, which is also a cylinder. Therefore, the final answer is C: cylinder.
+
+Self-correction prompt Review your previous answer and ensure that all relevant aspects of the image have been considered. Are there any elements or details that you missed? Based on your review, improve your answer.
+
+Refined response After reviewing the image, I can see that there are four objects in total: two green cubes, one purple cylinder, and one large shiny object. The question asks for the shape of the object that is the same color as the large shiny thing. Among the given options, only a sphere has a similar shape to the large shiny object, which is also a sphere. Therefore, the final answer is B: sphere. Correct answer: C
+
+Correct answer: C
+
+{Question, Image, IR, RR}, where IR denotes the initial response and  $RR$  is the refined response. In this process,  $M$  generates  $IR$  when prompted with the initial question and image. Subsequently,  $RR$  is generated as a revision of  $IR$  after  $M$  receives the self-correction prompt. Based on the correctness of  $IR$  and  $RR$ , each self-correction sample  $s'$  is classified into four distinct types: Specifically, the corresponding correctness of  $IR$  and  $RR$  for the four types of samples are as follows: Type 1 (correct  $\Rightarrow$  correct), Type 2 (incorrect  $\Rightarrow$  correct), Type 3 (correct  $\Rightarrow$  incorrect), and Type 4 (incorrect  $\Rightarrow$  incorrect). The  $\Rightarrow$  symbol denotes the transition of correctness from the initial response ( $IR$ ) to the refined response ( $RR$ ). Type 2 samples represent successful self-corrections, as the model effectively revises an incorrect  $IR$  into a correct  $RR$ . Conversely, Type 3 samples indicate detrimental self-corrections, where the model erroneously alters a correct  $IR$  to an incorrect one. The SELFCORSET preference dataset is then constructed using Type 2 and Type 3 samples according to the following criterion:  $RR$  from Type 2 samples and  $IR$  from Type 3 samples are designated as preferred responses, as they represent correct answers. In contrast,  $IR$  from Type 2 samples and  $RR$  from Type 3 samples are classified as disfavored responses. Figure 3 presents two examples of Type 2 and Type 3 self-correction outputs generated by InternLM-XComposer-2-7B. In the Type 2 example, the model successfully revises an incor
+
+rect  $IR$  (C: Two) to a correct  $RR$  (B: Three) after reviewing the image. Conversely, the Type 3 example shows a detrimental self-correction, where the model incorrectly alters an initially correct  $IR$  (C: cylinder) to an incorrect  $RR$  (B: sphere).
+
+We construct three distinct SELFCORSET, one for each of the three evaluated VLMs: LLaVA-V1.5-7B (Liu et al., 2024b), LLaVA-V1.5-13B (Liu et al., 2024b), and MiniCPM-Llama3-V2.5 (Yao et al., 2024). This construction emphasizes the uniqueness of self, as different VLMs possess distinct intrinsic self-correction behaviors and thus generate unique sets of self-correction samples. Each sample comprises a question, an image, a preferred response, and a disfavored response. Data samples are provided in Appendix A.
+
+# 3.3 Training: Learn from Self-Correction
+
+We fine-tune the VLM using Direct Preference Optimization (DPO) (Rafailov et al., 2023) on its dedicated preference dataset, SELFCORSET. We denote SELFCORSET as  $\mathcal{D}_{sc} = \{(Q^{(i)},I^{(i)},R_c^{(i)},R_r^{(i)})\}_{i = 1}^N$  where  $Q^{(i)}$  represents the input question,  $I^{(i)}$  is the corresponding image,  $R_{c}^{(i)}$  is the preferred response, and  $R_{r}^{(i)}$  is the disfavored response. The DPO loss is defined as:
+
+$$
+\mathcal {L} \left(\pi_ {\theta}; \pi_ {\text {r e f}}\right) = - \mathbb {E} _ {\left(Q, I, R _ {c}, R _ {r}\right)} \left[ \log \sigma f \left(\pi_ {\theta}; \pi_ {\text {r e f}}\right) \right] \tag {1}
+$$
+
+$$
+f \left(\pi_ {\theta}; \pi_ {\text {r e f}}\right) = \beta \left(\log \frac {\pi_ {\theta} \left(R _ {c} | Q , I\right)}{\pi_ {\text {r e f}} \left(R _ {c} | Q , I\right)} - \log \frac {\pi_ {\theta} \left(R _ {r} | Q , I\right)}{\pi_ {\text {r e f}} \left(R _ {r} | Q , I\right)}\right) \tag {2}
+$$
+
+where  $\sigma$  represents the logistic sigmoid function,  $\pi_{\theta}$  denotes the current VLM policy,  $\pi_{\mathrm{ref}}$  denotes
+
+<table><tr><td colspan="2"></td><td>RealWorldQA</td><td>MMStar</td><td>MMBench</td><td>SEEDBench</td><td>ScienceQA</td><td>MMT-Bench</td><td>Rank</td></tr><tr><td rowspan="5">MiniCPM-Llama3-V2.5</td><td>SP</td><td>61.70</td><td>50.40</td><td>79.00</td><td>66.80</td><td>75.40</td><td>49.00</td><td>1.17</td></tr><tr><td>+ CP</td><td>38.56</td><td>40.20</td><td>68.60</td><td>62.80</td><td>69.40</td><td>37.80</td><td>4.17</td></tr><tr><td>+ VP-1</td><td>48.50</td><td>46.20</td><td>76.40</td><td>64.80</td><td>73.00</td><td>37.00</td><td>3.00</td></tr><tr><td>+ VP-2</td><td>47.32</td><td>48.40</td><td>61.00</td><td>64.40</td><td>69.00</td><td>38.00</td><td>3.33</td></tr><tr><td>+ VP-3</td><td>43.00</td><td>48.00</td><td>54.00</td><td>69.80</td><td>54.40</td><td>40.60</td><td>3.33</td></tr><tr><td rowspan="5">InternLM-XComposer-2-7B</td><td>SP</td><td>60.13</td><td>47.40</td><td>76.80</td><td>69.00</td><td>78.20</td><td>48.80</td><td>1.17</td></tr><tr><td>+ CP</td><td>53.86</td><td>37.00</td><td>61.60</td><td>58.00</td><td>52.60</td><td>38.40</td><td>4.83</td></tr><tr><td>+ VP-1</td><td>54.50</td><td>48.00</td><td>70.80</td><td>67.00</td><td>62.60</td><td>38.60</td><td>2.50</td></tr><tr><td>+ VP-2</td><td>55.03</td><td>45.00</td><td>64.40</td><td>60.60</td><td>49.00</td><td>41.40</td><td>3.00</td></tr><tr><td>+ VP-3</td><td>54.51</td><td>39.00</td><td>61.80</td><td>59.60</td><td>58.80</td><td>39.00</td><td>3.50</td></tr><tr><td rowspan="5">LLaVA-V1.5-7B</td><td>SP</td><td>50.46</td><td>32.20</td><td>68.40</td><td>65.60</td><td>65.80</td><td>36.00</td><td>1.00</td></tr><tr><td>+ CP</td><td>36.60</td><td>24.00</td><td>54.00</td><td>36.20</td><td>56.80</td><td>32.00</td><td>2.83</td></tr><tr><td>+ VP-1</td><td>43.01</td><td>22.80</td><td>57.20</td><td>42.40</td><td>58.20</td><td>29.00</td><td>2.33</td></tr><tr><td>+ VP-2</td><td>17.78</td><td>18.60</td><td>45.40</td><td>29.00</td><td>45.80</td><td>12.00</td><td>5.00</td></tr><tr><td>+ VP-3</td><td>36.21</td><td>20.40</td><td>54.00</td><td>37.00</td><td>54.80</td><td>28.40</td><td>3.67</td></tr></table>
+
+Table 1: Quantitative comparisons (%) of MiniCPM-Llama3-V-2.5, InternLM-XComposer-2-7B, and LLaVA-V1.5-7B with intrinsic self-correction. Best results are bolded, and second-best are underlined. Rank indicates the overall performance ranking across the evaluated benchmarks (lower is better).
+
+the reference policy, and  $\beta$  is a hyperparameter that controls the strength of the penalty against deviating from the reference policy. Both  $\pi_{\theta}$  and  $\pi_{\mathrm{ref}}$  are initialized with the same weight.
+
+# 4 Experiments
+
+# 4.1 Experimental Settings
+
+Test Models. For intrinsic self-correction evaluation, we conduct experiments on three open-source VLMs: MiniCPM-Llama3-V2.5 (Yao et al., 2024), InternLM-XComposer-2-7B (Dong et al., 2024), and LLaVA-V1.5-7B (Liu et al., 2024b). MiniCPM-Llama3-V2.5 is an advanced VLM with a total of 8B parameters. InternLM-XComposer-2-7B is designed for the comprehension and composition of free-form text-image pairs. LLaVA-V1.5-7B is a widely used VLM trained with visual instructions. For self-correction training evaluation, we conduct experiments on MiniCPM-Llama3-V2.5, InternLM-XComposer-2-7B, LLaVA-V1.5-7B, and LLaVA-V1.5-13B.
+
+Evaluation Benchmarks. We conduct evaluations on eight multimodal multiple-choice question (MCQ) benchmarks: RealWorldQA (xAI, 2024), MMStar (Chen et al., 2024c), MMBench-en (Liu et al., 2024c), SEEDBench (Li et al., 2023a), ScienceQA (Lu et al., 2022), MMT-Bench (Ying et al., 2024), MMMU (Yue et al., 2024), and AI2D (Kembhavi et al., 2016). For intrinsic self-correction, we use the first six datasets with the following sample counts for evaluation: RealWorldQA (765), MMStar (500), MMBench (500), SEEDBench (500), ScienceQA (500), MMT-Bench (500). We further
+
+incorporate two benchmarks to evaluate fine-tuned models: MMMU (1050) and AI2D (3088). We adopt accuracy and average rank as the metrics.
+
+Training Baselines. We compare SCL with three preference optimization methods: POVID (Zhou et al., 2024a), CSR (Zhou et al., 2024b), and SIMA (Wang et al., 2025a). POVID leverages GPT to enhance the quality of ground truth answers and employs DPO for training. CSR incorporates iterative learning and a reward-based paradigm into its preference fine-tuning process. SIMA deploys in-context self-critic to construct its preference dataset and utilizes DPO to enhance the comprehension capabilities of VLM. We also compare SCL with Supervised Fine-Tuning (SFT), which directly utilizes the preferred responses of SELFCORSET for fine-tuning.
+
+Implementation Details. Intrinsic self-correction is conducted on a total of 26,981 samples. The resulting SELFCORSET dataset sizes for each VLM are as follows: MiniCPM-Llama3-V2.5 (1,853), InternLM-XComposer-2-7B (2,361), LLaVA-V1.5-7B (4,797), and LLaVA-V1.5-13B (738). For computational efficiency, we adopt LoRA(Hu et al., 2022), a widely recognized Parameter-Efficient Fine-Tuning technique, for training. We set the LoRA rank as 8. Training for 7/8B models is conducted on a single 4090 24GB GPU with 1.5 GPU hours per epoch. Training for 13B models is conducted on a single V100 32GB GPU with 1.5 GPU hours for three epochs.
+
+<table><tr><td></td><td>RealWorldQA</td><td>MMStar</td><td>MMBench</td><td>SEEDBench</td><td>ScienceQA</td><td>MMT-Bench</td><td>MMMU</td><td>AI2D</td><td>Rank</td></tr><tr><td>LLaVA-V1.5-7B</td><td>50.90</td><td>32.97</td><td>70.24</td><td>66.64</td><td>65.48</td><td>35.40</td><td>32.25</td><td>52.71</td><td>4.50</td></tr><tr><td>+POVID</td><td>51.50</td><td>33.68</td><td>71.44</td><td>65.52</td><td>64.60</td><td>35.04</td><td>33.96</td><td>53.59</td><td>3.25</td></tr><tr><td>+CSR</td><td>51.03</td><td>32.59</td><td>70.44</td><td>65.12</td><td>64.76</td><td>33.76</td><td>34.63</td><td>53.34</td><td>4.25</td></tr><tr><td>+SIMA</td><td>49.41</td><td>32.40</td><td>71.04</td><td>64.68</td><td>64.68</td><td>34.84</td><td>35.14</td><td>53.01</td><td>4.38</td></tr><tr><td>+SFT</td><td>52.21</td><td>35.59</td><td>70.72</td><td>67.12</td><td>66.48</td><td>36.36</td><td>32.95</td><td>52.73</td><td>3.00</td></tr><tr><td>+SCL(Ours)</td><td>52.71</td><td>36.11</td><td>71.00</td><td>67.84</td><td>66.56</td><td>36.96</td><td>33.62</td><td>54.81</td><td>1.63</td></tr><tr><td>LLaVA-V1.5-13B</td><td>54.85</td><td>35.25</td><td>74.40</td><td>69.40</td><td>71.36</td><td>40.08</td><td>34.88</td><td>55.95</td><td>2.75</td></tr><tr><td>+SFT</td><td>56.59</td><td>36.33</td><td>75.16</td><td>69.32</td><td>71.88</td><td>40.48</td><td>35.29</td><td>56.00</td><td>1.88</td></tr><tr><td>+SCL(Ours)</td><td>56.03</td><td>37.60</td><td>75.40</td><td>68.56</td><td>72.16</td><td>41.16</td><td>35.87</td><td>58.92</td><td>1.38</td></tr><tr><td>InternLM-XComposer-2-7B</td><td>60.45</td><td>47.70</td><td>76.60</td><td>68.90</td><td>78.30</td><td>49.20</td><td>56.48</td><td>80.27</td><td>3.00</td></tr><tr><td>+SFT</td><td>62.03</td><td>48.90</td><td>77.80</td><td>70.20</td><td>79.50</td><td>50.20</td><td>57.95</td><td>80.66</td><td>1.75</td></tr><tr><td>+SCL (Ours)</td><td>62.23</td><td>49.60</td><td>77.00</td><td>70.60</td><td>79.90</td><td>50.20</td><td>58.20</td><td>81.11</td><td>1.13</td></tr><tr><td>MiniCPM-Llama3-V2.5</td><td>61.70</td><td>50.40</td><td>79.00</td><td>66.80</td><td>75.40</td><td>49.00</td><td>45.24</td><td>77.56</td><td>3.00</td></tr><tr><td>+SFT</td><td>62.35</td><td>52.40</td><td>80.80</td><td>68.40</td><td>76.00</td><td>49.80</td><td>47.43</td><td>78.01</td><td>2.00</td></tr><tr><td>+SCL(Ours)</td><td>63.53</td><td>53.00</td><td>81.40</td><td>69.20</td><td>76.40</td><td>50.40</td><td>47.52</td><td>78.72</td><td>1.00</td></tr></table>
+
+Table 2: Quantitative comparisons (%) of LLaVA-V1.5-7B, LLaVA-V1.5-13B, InternLM-XComposer-2-7B, and MiniCPM-Llama3-V2.5 with SCL and baseline methods. Best results are bolded, and second-best are underlined. Rank indicates the overall performance ranking across the evaluated benchmarks (lower is better).
+
+![](images/4c880c16bf92231f5caf6c01e21f57f86881a40b7bd670b0cf28137840e8f86f.jpg)  
+Figure 4: Distribution of self-correction examples of MiniCPM-Llama3-V2.5 and InternLM-XComposer-2-7B under VP-1 on ScienceQA.
+
+# 4.2 Results and Analysis
+
+VLMs struggle in intrinsic self-correction. Table 1 presents the results of VLMs regarding intrinsic self-correction. It can be observed that, in the majority of cases, VLMs achieve higher accuracy prior to intrinsic self-correction. For instance, MiniCPM-Llama3-V2.5's accuracy is  $61.70\%$  in its initial generation but drops to only  $48.50\%$  after intrinsic self-correction under the VP-1 prompt. These results consistently indicate that VLMs struggle to achieve effective performance improvement through intrinsic self-correction. Furthermore, the effectiveness of intrinsic self-correction varies significantly across different models, tasks, and self-correction prompts. While MiniCPM-Llama3-V2.5 improves its performance on SEEDBench with the VP-3 prompt  $(66.80\% \rightarrow 69.80\%)$ , it fails to self-correct on other benchmarks, highlighting the inherent instability of intrinsic self-correction.
+
+To further illustrate the correctness transition
+
+during intrinsic self-correction, Figure 4 displays the distribution of self-correction sample types for MiniCPM-Llama3-V2.5 and InternLMXComposer-2-7B on ScienceQA under the VP1 prompt. The proportion of Type 3 samples (correct  $\Rightarrow$  incorrect) exceeds that of Type 2 (incorrect  $\Rightarrow$  correct) for both models, indicating that while models can revise incorrect answers, they more frequently erroneously alter initially correct answers. Consistent with the findings of Huang et al. (2024a) regarding LLMs, which reports degraded performance following intrinsic self-correction, our results suggest that VLMs struggle to rectify their own answers reliably.
+
+VLMs self-improve from their self-correction examples. Table 2 shows the performance of SCL in comparison to various preference optimization baselines. SCL achieves state-of-the-art performance for LLaVA-V1.5-7B on six benchmarks. While SCL exhibits suboptimal performance on MMMU, these results indicate that the preference data derived from self-correction samples effectively fine-tunes these models. All four evaluated VLMs achieve the best overall performance as indicated by the Rank metric. Despite using a relatively small fine-tuning dataset, our findings demonstrate that VLMs can benefit from both successful and erroneous self-correction samples. This suggests that VLMs possess a human-like behavior, i.e., effectively learning by internalizing the distinction between successful corrections and the insights gained from analyzing erroneous attempts. This
+
+ability enables them to efficiently enhance their comprehensive reasoning capabilities.
+
+# 4.3 Case Study
+
+Figure 5 presents the initial responses generated by LLaVA-V1.5-7B before and after preference finetuning. Prior to fine-tuning, the model incorrectly identified the sunrise as the primary object, misinterpreting it as the element that establishes the mood and atmosphere of the scene. Following finetuning with SCL, the model correctly recognizes that the question seeks the object occupying the most spatial area in the image and subsequently produces the correct answer, demonstrating the model's ability to avoid previous errors through preference fine-tuning.
+
+# 5 Further Studies and Analysis
+
+# 5.1 Reliability of Successful Intrinsic Self-Correction
+
+We observe that some successful refinements, classified as Type 2 self-correction cases, stem from the model's incidental guessing of the correct answer after receiving a self-correction prompt, rather than from proper reasoning of the task. For instance, when presented with an image that does not contain a teapot, models might respond with "The teapot may exist behind the woman..." and then conclude that a teapot is present in the image after self-correction. This type of refinement reflects a degree of uncertainty. Moreover, the reasoning behind seemingly successful refinements may not be entirely accurate. For example, in the Type 2 case presented in Figure 3, the model correctly identifies the number of chairs after self-correction but fails to accurately determine their colors. These observations on the reliability of successful refinements collectively indicate that VLMs possess limited capabilities for truly accurate refinement. Further supporting data is provided in Appendix B.
+
+Future research could explore the internal information flow that mediates the transition from initial to refined responses during intrinsic self-correction, for instance, by visualizing attention weights to improve both interpretability and reliability. More importantly, cases where the final answer is correct but the reasoning is flawed can negatively impact the self-correction mechanism. Future work should focus on designing more detailed evaluation criteria that account for the correctness of intermediate reasoning steps, such as by leveraging a strong model
+
+<table><tr><td></td><td>Turn 0</td><td>Turn 1</td><td>Turn 2</td><td>Turn 3</td></tr><tr><td>RealWorldQA (VP-1)</td><td>61.70</td><td>48.50</td><td>39.22</td><td>42.61</td></tr><tr><td>MMStar (VP-2)</td><td>50.40</td><td>48.40</td><td>49.20</td><td>46.80</td></tr><tr><td>MMBench (VP-1)</td><td>79.00</td><td>76.40</td><td>75.60</td><td>72.60</td></tr><tr><td>SEEDBench (VP-3)</td><td>66.80</td><td>69.80</td><td>63.80</td><td>64.00</td></tr></table>
+
+Table 3: Results  $(\%)$  of MiniCPM-Llama3-V2.5 with multi-turn intrinsic self-correction.
+
+as an MLLM-as-a-judge (Chen et al., 2024a).
+
+# 5.2 Multi-Turn Intrinsic Self-Correction
+
+Table 3 presents the results of multi-turn intrinsic self-correction for MiniCPM-Llama3-V2.5 on four benchmarks. Turn 0 represents the initial generation, while Turns 1 to 3 illustrate subsequent intrinsic self-correction iterations. Notably, the accuracy of the refined answers consistently degrades over multiple turns, ultimately falling below the initial accuracy after three correction iterations. This decline suggests that VLMs find it challenging to achieve effective intrinsic self-correction solely by increasing the number of correction iterations.
+
+Despite this observed degradation, multi-turn intrinsic self-correction may generate a richer, more diverse set of self-correction samples. Each iterative attempt may produce different types of errors when responding to the same question, thereby providing varied erroneous data for preference finetuning. Therefore, future research could investigate whether VLMs can derive greater benefits from these additional, potentially more informative samples, even if the multi-turn accuracy diminishes.
+
+# 5.3 Failure of VLM Intrinsic Self-Correction
+
+We attribute the observed failures of intrinsic self-correction in VLMs to several potential factors: (1) Limited ability to judge correctness (Huang et al., 2024a): VLMs may lack the fundamental reasoning capabilities required to reliably detect and revise their own errors. The inherent multimodal nature of these models—integrating both visual and textual inputs—introduces additional complexity, making it difficult for them to synthesize information across modalities and accurately evaluate correctness. (2) Susceptibility to prompt bias (Li et al., 2024): As demonstrated in Table1, self-correction performance varies notably with different prompts, suggesting a sensitivity to prompt design. Existing self-correction prompts may not sufficiently capture the intricate interplay between visual and textual content, which may impair the models' ability to effectively revise their responses.
+
+![](images/3985edcdb33f1e024844b7261282054a4165c1ed1a5628d39d04c565822cc1d7.jpg)  
+Figure 5: LLaVA-V1.5-7B successfully answers the question after learning from its self-correction samples.
+
+Question What is the primary subject of the image? A: The sunrise B: The grassy field C: The trees D: The fog Before A. The sunrise.
+
+1. The image features a grassy field with a sunrise in the background. 2. The sunrise is the primary subject of the image, as it sets the mood and atmosphere of the scene. 3. The grassy field and the fog contribute to the overall aesthetic of the image, but the sunrise is the main focus.
+
+After B. The grassy field.
+
+1. The primary subject of the image is the grassy field. 2. The image features a foggy sunrise, which adds a serene and atmospheric quality to the scene. 3. However, the focus is on the grassy field, which is the main subject of the photo. 4. The other elements, such as the trees and the sunrise, contribute to the overall composition and mood of the image, but they do not overshadow the grassy field as the primary subject.
+
+(3) Lack of a grounding mechanism: In contrast to self-correction methods that utilize external feedback or oracle labels (Huang et al., 2024a; Shinn et al., 2023), intrinsic self-correction operates without explicit supervision. This absence of grounding guidance limits the model's ability to converge on more accurate responses, resulting in ineffective or even detrimental revisions. Given these challenges, future work could explore the design of more diverse and task-specific self-correction prompts tailored for multimodal settings. Additionally, incorporating a strong teacher model to provide instructional feedback, or integrating oracle labels during the self-correction process, may help improve the reliability of revisions.
+
+# 5.4 Effect of the Number of Training Samples
+
+The limited proportion of Type 2 and Type 3 data, as illustrated in Figure 4, coupled with the tendency of more advanced models to generate fewer samples of these types, results in a relatively small sample size for SELFCORSET. To explore the influence of fine-tuning data quantity, we randomly divide SELFCORSET into five subsets, increasing in  $20\%$  increments from  $0\%$ . We evaluate the impact of these varying training set sizes on the performance of LLaVA-V1.5-7B on SEEDBench and AI2D, with the accuracy trend illustrated in Figure 6. Notably, even with smaller training datasets, the fine-tuned model exhibits significant performance gains. For instance, the model fine-tuned on the  $p = 0.4$  subset achieves an accuracy of  $67.80\%$  on SEEDBench, reflecting a  $2.2\%$  improvement over the untrained model. These results indicate that SCL effectively improves the model's performance despite the relatively modest size of SELFCORSET. Moreover, as the number of training samples increases, the overall accuracy consistently improves, showcasing the potential of SCL with larger preference datasets. Therefore, future
+
+![](images/315fb9a1c15e72ec11b2905a2750cc3e40533dac3318053ab6e6054b92ce3480.jpg)  
+Figure 6: Results of LLaVA-V1.5-7B under different proportions  $(p)$  of SELFCORSET.
+
+research could explore more straightforward and computationally efficient data augmentation methods to scale SELFCORSET and investigate SCL's performance under large-scale data conditions.
+
+# 6 Conclusion
+
+This work investigates the intrinsic self-correction capabilities and limitations of Vision-Language Models (VLMs) during the inference stage, and their potential for self-improvement through fine-tuning, across various multiple-choice questions (MCQ) benchmarks. We propose Self-Correction Learning (SCL), a novel framework that employs Direct Preference Optimization (DPO) to enable VLMs to learn from their own self-correction responses. This approach aims to fundamentally enhance the models' ability to generate correct answers directly, rather than merely refining initial errors. Our experiments reveal that while VLMs face significant challenges and exhibit unreliable performance during intrinsic self-correction, they can effectively leverage their own self-correction samples (both successful and erroneous) to self-improve without requiring external feedback.
+
+# Limitations
+
+Given the challenges of evaluating absolute correctness in open-ended multimodal tasks, this study primarily focuses on MCQ benchmarks to assess the accuracy of both initial and refined responses. However, this design inherently limits the exploration of more complex vision understanding tasks, such as visual question answering (Chen et al., 2025), video question answering (Chen et al., 2024d; Fu et al., 2024), and complex transportation system navigation (Li et al., 2023c). Future work could investigate fine-grained evaluation methods applicable across a wider range of multimodal tasks, as well as computational resource constraint-aware adaptive reasoning (Huang et al., 2025c).
+
+The current construction of SELFCORSET is tailored to each VLM, leveraging the model's own preferred and disfavored responses to enable more targeted self-improvement. However, this design limits generalizability across different models. A potential direction for future work is to develop a unified version of SELFCORSET that can be applied across multiple VLMs. This would involve identifying common error patterns shared among VLMs and designing universal prompts that generalize well across different architectures.
+
+# References
+
+Dongping Chen, Ruoxi Chen, Shilin Zhang, Yaochen Wang, Yinuo Liu, Huichi Zhou, Qihui Zhang, Yao Wan, Pan Zhou, and Lichao Sun. 2024a. MLLM-as-a-judge: Assessing multimodal LLM-as-a-judge with vision-language benchmark. In *Forty-first International Conference on Machine Learning*.  
+Jiao Chen, Jiayi He, Fangfang Chen, Zuohong Lv, Jianhua Tang, Weihua Li, Zuozhu Liu, Howard H. Yang, and Guangjie Han. 2024b. Towards general industrial intelligence: A survey of continual large models in industrial IoT. Preprint, arXiv:2409.01207.  
+Jiao Chen, Ruyi Huang, Zuohong Lv, Jianhua Tang, and Weihua Li. 2025. Faultgpt: Industrial fault diagnosis question answering system by vision language models. Preprint, arXiv:2502.15481.  
+Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Jiaqi Wang, Yu Qiao, Dahua Lin, and Feng Zhao. 2024c. Are we on the right way for evaluating large vision-language models? Preprint, arXiv:2403.20330.  
+Xiuyuan Chen, Yuan Lin, Yuchen Zhang, and Weiran Huang. 2024d. Autoeval-video: An automatic benchmark for assessing large vision language models
+
+in open-ended video question answering. In European Conference on Computer Vision, pages 179-195. Springer.  
+Yangyi Chen, Karan Sikka, Michael Cogswell, Heng Ji, and Ajay Divakaran. 2024e. Dress: Instructing large vision-language models to align and interact with humans via natural language feedback. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14239-14250.  
+Zhichao Deng, Xiangtai Li, Xia Li, Yunhai Tong, Shen Zhao, and Mengyuan Liu. 2024. Vg4d: Vision-language model goes 4d video recognition. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 5014-5020.  
+Shehzaad Dhuliawala, Mojtaba Komeili, Jing Xu, Roberta Raileanu, Xian Li, Asli Celikyilmaz, and Jason Weston. 2024. Chain-of-verification reduces hallucination in large language models. In Findings of the Association for Computational Linguistics ACL 2024, pages 3563-3578, Bangkok, Thailand and virtual meeting. Association for Computational Linguistics.  
+Xiaoyi Dong, Pan Zhang, Yuhang Zang, Yuhang Cao, Bin Wang, Linke Ouyang, Xilin Wei, Songyang Zhang, Haodong Duan, Maosong Cao, Wenwei Zhang, Yining Li, Hang Yan, Yang Gao, Xinyue Zhang, Wei Li, Jingwen Li, Kai Chen, Conghui He, and 4 others. 2024. Internlm-xcomposer2: Mastering free-form text-image composition and comprehension in vision-language large model. Preprint, arXiv:2401.16420.  
+Zhiyuan Fan, Yumeng Wang, Sandeep Polisetty, and Yi R. Fung. 2025. Unveiling the lack of lvlm robustness to fundamental visual variations: Why and path forward. Preprint, arXiv:2504.16727.  
+Chaoyou Fu, Yuhan Dai, Yongdong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen, Mengdan Zhang, and 1 others. 2024. Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis. arXiv preprint arXiv:2405.21075.  
+Haixia Han, Jiaqing Liang, Jie Shi, Qianyu He, and Yanghua Xiao. 2024. Small language model can self-correct. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 18162-18170.  
+Zhitao He, Sandeep Polisetty, Zhiyuan Fan, Yuchen Huang, Shujin Wu, and Yi R. Fung. 2025. Mmboundary: Advancing mllm knowledge boundary awareness through reasoning step confidence calibration. Preprint, arXiv:2505.23224.  
+Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2022. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations.
+
+Jiaxin Huang, Shixiang Shane Gu, Le Hou, Yuexin Wu, Xuezhi Wang, Hongkun Yu, and Jiawei Han. 2023. Large language models can self-improve. In The 2023 Conference on Empirical Methods in Natural Language Processing.  
+Jie Huang, Xinyun Chen, Swaroop Mishra, Huaixiu Steven Zheng, Adams Wei Yu, Xinying Song, and Denny Zhou. 2024a. Large language models cannot self-correct reasoning yet. In The Twelfth International Conference on Learning Representations.  
+Jinsheng Huang, Liang Chen, Taian Guo, Fu Zeng, Yusheng Zhao, Bohan Wu, Ye Yuan, Haozhe Zhao, Zhihui Guo, Yichi Zhang, Jingyang Yuan, Wei Ju, Luchen Liu, Tianyu Liu, Baobao Chang, and Ming Zhang. 2025a. MMEvalPro: Calibrating multimodal benchmarks towards trustworthy and efficient evaluation. In Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), pages 4805-4822, Albuquerque, New Mexico. Association for Computational Linguistics.  
+Junsheng Huang, Zhitao He, Sandeep Polisetty, Qingyun Wang, and May Fung. 2025b. Mac-tuning: Llm multi-compositional problem reasoning with enhanced knowledge boundary awareness. Preprint, arXiv:2504.21773.  
+Kung-Hsiang Huang, Hou Pong Chan, Yi R. Fung, Haoyi Qiu, Mingyang Zhou, Shafiq Joty, Shih-Fu Chang, and Heng Ji. 2024b. From pixels to insights: A survey on automatic chart understanding in the era of large foundation models. Preprint, arXiv:2403.12027.  
+Shijue Huang, Hongru Wang, Wanjun Zhong, Zhaochen Su, Jiazhan Feng, Bowen Cao, and Yi R. Fung. 2025c. Adactrl: Towards adaptive and controllable reasoning via difficulty-aware budgeting. Preprint, arXiv:2505.18822.  
+Ryo Kamoi, Yusen Zhang, Nan Zhang, Jiawei Han, and Rui Zhang. 2024. When can LLMs actually correct their own mistakes? a critical survey of self-correction of LLMs. Transactions of the Association for Computational Linguistics, 12:1417-1440.  
+Aniruddha Kembhavi, Mike Salvato, Eric Kolve, Minjoon Seo, Hannaneh Hajishirzi, and Ali Farhadi. 2016. A diagram is worth a dozen images. Preprint, arXiv:1603.07396.  
+Aviral Kumar, Vincent Zhuang, Rishabh Agarwal, Yi Su, John D Co-Reyes, Avi Singh, Kate Baumli, Shariq Iqbal, Colton Bishop, Rebecca Roelofs, Lei M Zhang, Kay McKinney, Disha Shrivastava, Cosmin Paduraru, George Tucker, Doina Precup, Feryal Behbahani, and Aleksandra Faust. 2024. Training language models to self-correct via reinforcement learning. Preprint, arXiv:2409.12917.
+
+Bohao Li, Rui Wang, Guangzhi Wang, Yuying Ge, Yixiao Ge, and Ying Shan. 2023a. Seed-bench: Benchmarking multimodal llms with generative comprehension. Preprint, arXiv:2307.16125.  
+Cheng Li, May Fung, Qingyun Wang, Chi Han, Manling Li, Jindong Wang, and Heng Ji. 2025. Mentalarena: Self-play training of language models for diagnosis and treatment of mental health disorders. Preprint, arXiv:2410.06845.  
+Lei Li, Zhihui Xie, Mukai Li, Shunian Chen, Peiyi Wang, Liang Chen, Yazheng Yang, Benyou Wang, and Lingpeng Kong. 2023b. Silkie: Preference distillation for large visual language models. Preprint, arXiv:2312.10665.  
+Loka Li, Zhenhao Chen, Guangyi Chen, Yixuan Zhang, Yusheng Su, Eric Xing, and Kun Zhang. 2024. Confidence matters: Revisiting intrinsic self-correction capabilities of large language models. Preprint, arXiv:2402.12563.  
+Tianyi Li, Joshua Klavins, Te Xu, Niaz Mahmud Zafri, and Raphael Stern. 2023c. Understanding driver-pedestrian interactions to predict driver yielding: naturalistic open-source dataset collected in Minnesota. Preprint, arXiv:2312.15113.  
+Dancheng Liu, Amir Nassereldine, Ziming Yang, Chenhui Xu, Yuting Hu, Jiajie Li, Utkarsh Kumar, Changjae Lee, Ruiyang Qin, Yiyu Shi, and Jinjun Xiong. 2024a. Large language models have intrinsic self-correction ability. Preprint, arXiv:2406.15673.  
+Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. 2024b. Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 26296-26306.  
+Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Ji-aqi Wang, Conghui He, Ziwei Liu, Kai Chen, and Dahua Lin. 2024c. Mmbench: Is your multi-modal model an all-around player? In Computer Vision - ECCV 2024: 18th European Conference, Milan, Italy, September 29-October 4, 2024, Proceedings, Part VI, page 216-233, Berlin, Heidelberg. Springer-Verlag.  
+Pan Lu, Swaroop Mishra, Tanglin Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. 2022. Learn to explain: Multimodal reasoning via thought chains for science question answering. In Advances in Neural Information Processing Systems, volume 35, pages 2507-2521. Curran Associates, Inc.  
+Aman Madaan, Niket Tandon, Prakhar Gupta, Skyler Hallinan, Luyu Gao, Sarah Wiegreffe, Uri Alon, Nouha Dziri, Shrimai Prabhumoye, Yiming Yang, Shashank Gupta, Bodhisattwa Prasad Majumder, Katherine Hermann, Sean Welleck, Amir Yazdanbakhsh, and Peter Clark. 2023. Self-refine: Iterative refinement with self-feedback. In Advances in Neural Information Processing Systems, volume 36, pages 46534-46594. Curran Associates, Inc.
+
+OpenAI. 2024a. Hello gpt-4o. https://openai.com/index/hello-gpt-4o/. Accessed: 2025-05-29.  
+OpenAI. 2024b. O1 system card. https://cdn.openuai.com/o1-system-card-20240917.pdf. Accessed: 2025-05-29.  
+Wenshuo Peng, Kaipeng Zhang, Yue Yang, Hao Zhang, and Yu Qiao. 2024. Data adaptive traceback for vision-language foundation models in image classification. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 4506-4514.  
+Mansi Phute, Alec Helbling, Matthew Daniel Hull, Sheng Yun Peng, Sebastian Szyller, Cory Cornelius, and Duen Horng Chau. 2024. LLM self defense: By self examination, LLMs know they are being tricked. In The Second Tiny Papers Track at ICLR 2024.  
+Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. 2023. Direct preference optimization: Your language model is secretly a reward model. In Advances in Neural Information Processing Systems, volume 36, pages 53728-53741. Curran Associates, Inc.  
+Noah Shinn, Federico Cassano, Ashwin Gopinath, Karthik Narasimhan, and Shunyu Yao. 2023. Reflexion: language agents with verbal reinforcement learning. In Advances in Neural Information Processing Systems, volume 36, pages 8634-8652. Curran Associates, Inc.  
+Chenkai Sun, Ke Yang, Revanth Gangi Reddy, Yi R. Fung, Hou Pong Chan, Kevin Small, ChengXiang Zhai, and Heng Ji. 2025. Persona-db: Efficient large language model personalization for response prediction with collaborative data refinement. Preprint, arXiv:2402.11060.  
+Yongqi Tong, Dawei Li, Sizhe Wang, Yujia Wang, Fei Teng, and Jingbo Shang. 2024. Can LLMs learn from previous mistakes? investigating LLMs' errors to boost for reasoning. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 3065-3080, Bangkok, Thailand. Association for Computational Linguistics.  
+Xiyao Wang, Jiuhai Chen, Zhaoyang Wang, Yuhang Zhou, Yiyang Zhou, Huaxiu Yao, Tianyi Zhou, Tom Goldstein, Parminder Bhatia, Taha Kass-Hout, Furong Huang, and Cao Xiao. 2025a. Enhancing visual-language modality alignment in large vision language models via self-improvement. In Findings of the Association for Computational Linguistics: NAACL 2025, pages 268-282, Albuquerque, New Mexico. Association for Computational Linguistics.  
+Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc V Le, Ed H. Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou. 2023. Self-consistency improves chain of thought reasoning in language models. In The Eleventh ICLR.
+
+Yumeng Wang, Zhiyuan Fan, Qingyun Wang, May Fung, and Heng Ji. 2025b. Calm: Unleashing the cross-lingual self-aligning ability of language model question answering. Preprint, arXiv:2501.18457.  
+Shujin Wu, May Fung, Cheng Qian, Jeonghwan Kim, Dilek Hakkani-Tur, and Heng Ji. 2024a. Aligning llms with individual preferences via interaction. Preprint, arXiv:2410.03642.  
+Shujin Wu, Yi Fung, Sha Li, Yixin Wan, Kai-Wei Chang, and Heng Ji. 2024b. MACAROON: Training vision-language models to be your engaged partners. In Findings of the Association for Computational Linguistics: EMNLP 2024, pages 7715-7731, Miami, Florida, USA. Association for Computational Linguistics.  
+xAI. 2024. Realworldqa. https://x.ai/blog/grok-1.5v. Accessed: 2025-05-29.  
+Wenda Xu, Guanglei Zhu, Xuandong Zhao, Liangming Pan, Lei Li, and William Wang. 2024. Pride and prejudice: LLM amplifies self-bias in self-refinement. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 15474-15492, Bangkok, Thailand. Association for Computational Linguistics.  
+Diji Yang, Kezhen Chen, Jinmeng Rao, Xiaoyuan Guo, Yawen Zhang, Jie Yang, and Yi Zhang. 2024. Tackling vision language tasks through learning inner monologues. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 19350-19358.  
+Yuan Yao, Tianyu Yu, Ao Zhang, Chongyi Wang, Junbo Cui, Hongji Zhu, Tianchi Cai, Haoyu Li, Weilin Zhao, Zhihui He, Qianyu Chen, Huarong Zhou, Zhensheng Zou, Haoye Zhang, Shengding Hu, Zhi Zheng, Jie Zhou, Jie Cai, Xu Han, and 4 others. 2024. Minicpmv: A gpt-4v level mllm on your phone. Preprint, arXiv:2408.01800.  
+Kaining Ying, Fanqing Meng, Jin Wang, Zhiqian Li, Han Lin, Yue Yang, Hao Zhang, Wenbo Zhang, Yuqi Lin, Shuo Liu, jiayi lei, Quanfeng Lu, Runjian Chen, Peng Xu, Renrui Zhang, Haozhe Zhang, Peng Gao, Yali Wang, Yu Qiao, and 3 others. 2024. MMT-bench: A comprehensive multimodal benchmark for evaluating large vision-language models towards multitask AGI. In *Forty-first International Conference on Machine Learning*.  
+Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, Cong Wei, Botao Yu, Ruibin Yuan, Renliang Sun, Ming Yin, Boyuan Zheng, Zhenzhu Yang, Yibo Liu, Wenhao Huang, and 3 others. 2024. Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9556-9567.
+
+Hanning Zhang, Shizhe Diao, Yong Lin, Yi Fung, Qing Lian, Xingyao Wang, Yangyi Chen, Heng Ji, and Tong Zhang. 2024a. R-tuning: Instructing large language models to say 'I don't know'. In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), pages 7113-7139, Mexico City, Mexico. Association for Computational Linguistics.
+
+Jianshu Zhang, Dongyu Yao, Renjie Pi, Paul Pu Liang, and Yi R. Fung. 2025. Vlm2-bench: A closer look at how well vlms implicitly link explicit matching visual cues. Preprint, arXiv:2502.12084.
+
+Xiaoying Zhang, Baolin Peng, Ye Tian, Jingyan Zhou, Lifeng Jin, Linfeng Song, Haitao Mi, and Helen Meng. 2024b. Self-alignment for factuality: Mitigating hallucinations in LLMs via self-evaluation. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1946-1965, Bangkok, Thailand. Association for Computational Linguistics.
+
+Yiyang Zhou, Chenhang Cui, Rafael Rafailov, Chelsea Finn, and Huaxiu Yao. 2024a. Aligning modalities in vision large language models via preference finetuning. In ICLR 2024 Workshop on Reliable and Responsible Foundation Models.
+
+Yiyang Zhou, Zhiyuan Fan, Dongjie Cheng, Sihan Yang, Zhaorun Chen, Chenhang Cui, Xiyao Wang, Yun Li, Linjun Zhang, and Huaxiu Yao. 2024b. Calibrated self-rewarding vision language models. In ICML 2024 Workshop on Foundation Models in the Wild.
+
+Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. 2024. MiniGPT-4: Enhancing vision-language understanding with advanced large language models. In The Twelfth International Conference on Learning Representations.
+
+# A Data Examples of SELFCORSET
+
+Our work introduces SELFCORSET, a novel dataset derived from the intrinsic self-correction processes of Vision-Language Models (VLMs) during inference. For each sample, the dataset captures both the VLM's initial and refined responses. Specifically, SELFCORSET comprises Type 2 (incorrect  $\Rightarrow$  correct) and Type 3 (correct  $\Rightarrow$  incorrect) self-correction instances, where the correct responses are designated as preferred and the incorrect ones as disfavored. In this section, we provide illustrative examples of the dataset content (Table 4 to 9).
+
+# B Examples Supporting Further Studies
+
+In our further studies, we argue that VLMs often exhibit limited capabilities for genuinely accurate
+
+refinement, and that even seemingly successful self-corrections can be unreliable due to flawed underlying reasoning. To substantiate this claim, we present two illustrative examples in Table 10.
+
+In the first example, the model's refined response relies on the misleading assumption that a "Do not board" sign directly indicates a lack of crowding on the platform. However, platform crowding should primarily be assessed based on crowd density. Although the model ultimately selects the correct answer (B. no), this decision stems from an over-reliance on a single, potentially misleading cue rather than a comprehensive understanding of the entire visual scenario.
+
+In the second example, the model confuses particle size with concentration during the refinement process. The model infers concentration levels from the size of the purple particles, entirely neglecting the actual number of particles present. While the model successfully corrects the error in its initial response, the self-correction process should have involved a proper reassessment of the quantity of purple particles in the solution, not their individual size.
+
+Q: What is the direction of this pull? A. toward her hand B. away from her hand
+
+![](images/e14b45728d5c6579e6d3c3004ee1902de7e779f4dbb63ff4bf36d4e9c86b2e42.jpg)
+
+IR (disfavored): B. away from her hand. The person is pulling a bottle away from their hand, which indicates that they are moving the bottle in the opposite direction of their hand. This action suggests that they are trying to remove the bottle from the ground or from their hand, rather than placing it there.
+
+RR (preferred): A. toward her hand. Upon reviewing the image, it is clear that the person is pulling a bottle toward their hand, rather than away from it. This action suggests that they are trying to pick up the bottle or place it on the ground. The person's hand is positioned above the bottle, indicating that they are in the process of lifting it.
+
+Q: Detect and give a judgment of the anomalous object within the given image. A. normal B. anomaly
+
+![](images/91f4b6ed353f530b8dd1a01382a5501147648022cc8d1e728772541587c5b08f.jpg)
+
+IR (disfavored): B. anomaly. The image shows a group of people walking on a sidewalk near a body of water. Among them, there is a person wearing a backpack, which is an unusual object in this context. The backpack is not a typical item one would expect to see while walking on a sidewalk, as it is usually associated with carrying personal belongings or items for outdoor activities. The presence of the backpack in this scene might indicate that the person is either on a trip or carrying something specific, but it is not a common sight in this setting.
+
+RR (preferred): A. normal. After reviewing the image, I realized that the backpack is not an anomalous object in this context. It is a common item that people carry while walking on a sidewalk, especially if they are on a trip or carrying personal belongings. The presence of the backpack does not indicate that the person is carrying something specific or that they are on a trip. Therefore, the backpack is not an anomalous object in this scene.
+
+Table 4: Type 2 samples of LLaVA-V1.5-7B.
+
+Q: How would you describe the appearance of the man playing the drums? A. Young and long-haired B. Middle-aged with glasses C. Old and bald D. Young with a beard
+
+![](images/f366b568ee7822ca6613d4e0bd4bd270dd49239baba6f071ea7b46eb61615b6a.jpg)
+
+IR (preferred): C. Old and bald. The man playing the drums is described as an older man with a bald head. This suggests that he is not a young man, and his appearance is not typical of a young musician. The fact that he is playing drums indicates that he has some level of skill and experience in the field.
+
+RR (disfavored): A. Young and long-haired. Upon reviewing the image, it is clear that the man playing the drums is not an older man with a bald head. Instead, he is described as a young man with long hair. This suggests that he is not a middle-aged man, and his appearance is more typical of a young musician. The fact that he is playing drums indicates that he has some level of skill and experience in the field
+
+Q: What landmark is this and where is it? A. Mecca in Saudi Arabia B. Big Ben in London C. The Burj al Arab Hotel in Dubai D. Tower of Pisa, Italy
+
+![](images/3fa3851103d195a98f0b5b600a41a695e318bccd3efb92de280e94947d596aa3.jpg)
+
+IR (preferred): D. Tower of Pisa, Italy. The image features a tall, white tower with a clock on its side, which is the iconic Tower of Pisa. The tower is located in Italy, and it is a famous landmark known for its leaning structure. The presence of a large group of people around the tower suggests that it is a popular tourist attraction. The image also shows a clock on the side of the tower, which is a distinctive feature of the structure.
+
+RR (disfavored): A. Mecca in Saudi Arabia. I apologize for the confusion in my previous response. After reviewing the image, I realized that the Tower of Pisa is not the correct answer. The image features a large clock tower, which is a prominent landmark in Mecca, Saudi Arabia. The tower is known as the Abraj Al-Bait Clock Tower and is a significant attraction for pilgrims visiting the city. The image shows a large group of people around the tower, indicating its popularity as a tourist destination. The clock tower's distinctive design and its location in Mecca make it the most appropriate landmark in the image.
+
+Table 5: Type 3 samples of LLaVA-V1.5-7B.
+
+Q: How many arrows are there in total in the picture? A. 1 B. 2 C. 3 D. 4
+
+![](images/cc3200fa176b28089c2e7b4eaa40c2019979515180f81c72e9924ee0a12d5136.jpg)  
+I:
+
+IR (disfavored): There are a total of 3 arrows in the picture. The first arrow is pointing upwards, the second arrow is pointing downwards, and the third arrow is pointing to the right.
+
+RR (preferred): Upon reviewing my previous answer, I realized that I missed the fact that there are two arrows pointing in the same direction, which is upwards. Therefore, there are a total of 4 arrows in the picture. The first arrow is pointing upwards, the second arrow is pointing upwards as well, the third arrow is pointing to the right, and the fourth arrow is pointing downwards.
+
+Q: How many species are there in total? A. 5 B. 6 C. 7 D. 8 E. 9
+
+![](images/644f468163767e508ea576b2703c56d4380536969b6fd74f98821eb40d0b29f0.jpg)  
+I:
+
+IR (disfavored): There are a total of 7 species in the diagram. The diagram shows the food chain of a water ecosystem, with each species being connected to the next one through arrows. The species include algae, water flea, tadpole, great diving beetle, and tapioca. The arrows indicate the flow of energy and nutrients from one species to another, as each species consumes the one before it.
+
+RR (preferred): Upon reviewing my previous answer, I realized that I missed one species in the food chain. The correct answer is 8 species. The diagram shows the food chain of a water ecosystem, with each species being connected to the next one through arrows. The species include algae, water flea, tadpole, great diving beetle, tapioca, and two other species that are not explicitly identified. The arrows indicate the flow of energy and nutrients from one species to another, as each species consumes the one before it.
+
+Q: Where do Teal and Dark Khaki intersect on the x-axis? A. 40 B. 44 C. 60 D. 80
+
+![](images/0ed96caa246c935c5a2475fa43a7c113c72245c83800095ba61578ef405342ca.jpg)  
+I:
+
+IR (preferred): To find the intersection of Teal and Dark Khaki on the x-axis, we need to look at the x-axis values for both lines. Teal has a value of 40, while Dark Khaki has a value of 60. The intersection occurs at the point where these two values meet, which is at 44. Therefore, the answer is B. 44.
+
+RR (disfavored): Upon reviewing my previous answer, I realized that I made an error in my interpretation of the x-axis values for Teal and Dark Khaki. Teal has a value of 40, while Dark Khaki has a value of 80. The intersection occurs at the point where these two values meet, which is at 60. Therefore, the correct answer is C. 60.
+
+Q: What color is the Howler Monkey? A. Blue B. Green C. Brown D. Purple E. Black F. Dark brown G. White
+
+![](images/7d4d3f7803a43895e63f31ba82c9ecab48c680b8f5a5c06d8ca22f814c331adc.jpg)  
+I:
+
+IR (preferred): The Howler Monkey in the image is black. The monkey is sitting on a tree branch, and its color is clearly visible. The monkey's black fur stands out against the tree branch, making it easy to identify the color of the monkey. Therefore, the correct answer is E. Black.
+
+RR (disfavored): Upon reviewing my previous answer, I noticed that I did not consider the possibility of the monkey's fur appearing dark brown in certain lighting conditions. Therefore, I would like to revise my answer to include the possibility of the monkey's fur appearing dark brown or brown. Therefore, the correct answer is F. Dark brown.
+
+Table 7: Type 3 samples of LLaVA-V1.5-13B.
+
+Table 6: Type 2 samples of LLaVA-V1.5-13B.
+
+Q: Select the amphibian below. 0: brown tree frog 1: wood-pecker
+
+![](images/1b40480347e421eb84f84536a4f14b2d88ec54dc5b98b616b2242e57dba35edf.jpg)
+
+IR (disfavored): The image features a frog, which is an amphibian. The brown tree frog (option 0) is also an amphibian, but it is not depicted in the image; therefore, by elimination and based on the visual characteristics of the animal in the picture, the correct answer is 1 (woodpecker), as it does not match the description of the amphibian shown.
+
+RR (preferred): Upon reviewing the context, it's clear that the image showcases a frog, which is an amphibian. The correct answer should be the amphibian present in the image. Therefore, the improved answer is 0 (brown tree frog), as it matches the description of the amphibian depicted in the image.
+
+Q: Which solution has a higher concentration of green particles? A. neither; their concentrations are the same B. Solution B C. Solution A
+
+![](images/2b3c663bd1b7779fcef4d3bfe8f5a99e4feb1ea5baa550953f196f2cf7107ca9.jpg)  
+Solvent volume:  $40~\mathrm{mL}$
+
+![](images/dd93640a6df481222605891bc5c811d59e0da7677dff88aa431743ad1fb5b581.jpg)  
+Solvent volume:  $40~\mathrm{mL}$  
+Solution B
+
+![](images/adebb1c5d30ff0ad4c3224a518ee4bbc31b8733d2994f88c006f8df02fb1ba8c.jpg)  
+Solution A
+
+IR (disfavored): The concentration of particles in a solution is determined by the volume of the solution and the number of particles present. In this case, both Solution A and Solution B have a solvent volume of  $40\mathrm{ml}$ . Since there are more green particles in Solution A than in Solution B, we can conclude that Solution A has a higher concentration of green particles. Final answer: C. Solution A
+
+RR (preferred): After reviewing the image, it is clear that there are more green particles in Solution B than in Solution A. Therefore, the concentration of green particles is higher in Solution B. Final answer: B. Solution B
+
+Q: Which term matches the picture? A. endotherms B. ectotherms
+
+![](images/79a37c642208100a54a8d262789f1b16ca16c7ee81c95728436494f54949a45c.jpg)
+
+IR (preferred): The picture you provided shows a horse, which is a mammal. Mammals are endotherms, meaning they can regulate their body temperature internally. The term that matches the picture is A. endotherms.
+
+RR (disfavored): I have reviewed my previous answer and considered all relevant aspects of the image. Based on my review, I believe that the term ectotherms is more suitable to describe the horse in the picture because horses are cold-blooded animals that regulate their body temperature by external heat sources, rather than internally like endotherms (warm-blooded animals). Therefore, my final answer is B. ectotherms.
+
+Q: Which term matches the picture? A. Hawaiian eruption B. Strombolian eruption
+
+![](images/ec682dbe5a8204fa2986730a68d9cfcd110ead99d5617ff482f907d039739e0d.jpg)
+
+IR (preferred): A Strombolian eruption is characterized by moderate bursts of expanding gases that eject incandescent lava fragments, as shown in the image with fiery lava being thrown into the air. Hawaiian eruptions, on the other hand, are generally less explosive and produce more fluid lava flows. The term that matches the picture is B. Strombolian eruption.
+
+RR (disfavored): I apologize for the mistake in my previous answer. Upon reviewing the image, it is clear that the term that matches the picture is A. Hawaiian eruption. This is because the image shows a volcanic eruption with lava flowing and a bright light at the top of the volcano, which are characteristics of a Hawaiian eruption. Strombolian eruptions typically involve explosive ejections of solid material and do not necessarily exhibit such visible lava flows or bright lights as seen in the image. Therefore, I have corrected my initial response to accurately reflect the information provided in the image.
+
+Table 8: Type 2 samples of MiniCPM-Llama3-V-2.5.
+
+Table 9: Type 3 samples of MiniCPM-Llama3-V-2.5.
+
+Q: is this place crowded? A. yes B. no
+
+![](images/cf46b2c11069a9135a1af150ea3b5c070baa18d16ef25dffc500b606f3e5685c.jpg)
+
+IR: The place is crowded. People are waiting for the train, and they are spread out along the platform. The final answer is A.
+
+SC: Review your previous answer and ensure that all relevant aspects of the image have been considered. Are there any elements or details that you missed? Based on your review, improve your answer.
+
+RR: I reviewed my previous answer and have identified a relevant detail that was missed. The sign Do not board on the platform indicates that the train may be not crowded, as it is likely the reason why people are standing near it. Based on this new information, I will revise my final answer to B.
+
+Q: Which solution has a higher concentration of purple particles?
+
+A. Solution A B. Solution B C. neither; their concentrations are the same
+
+![](images/0620072aead65c3b7c48a54d5f7850cac65eac8e3e76e6454fabbce0fed33b50.jpg)  
+Solvent volume: 45 mL  
+Solution A
+
+![](images/039c4c240e4fc8dcdee6a1682437cd53ecc6a83638c7570901ca4a2713c7f318.jpg)  
+Solvent volume: 45 mL
+
+I:
+
+IR: The concentration of particles in a solution is determined by the volume of the solution and the number of particles present. In this case, both Solution A and Solution B have the same volume (45 ml). Therefore, to determine which solution has a higher concentration of purple particles, we need to compare the number of particles in each solution. Unfortunately, the information provided does not include the number of particles in either solution. Without this information, it is impossible to determine which solution has a higher concentration of purple particles. So, the final answer is C. neither; their concentrations are the same.
+
+SC: Review your previous answer and ensure that all relevant aspects of the image have been considered. Are there any elements or details that you missed? Based on your review, improve your answer.
+
+RR: I apologize for the oversight in my previous response. Upon reviewing the image, I can now see that Solution A has more purple particles than Solution B. This is because Solution A has larger concentrated purple particles, while Solution B has smaller concentrated purple particles. Therefore, the final answer is A.
+
+Solution A has a higher concentration of purple particles.
+
+Table 10: Successful refinements (Type 2 samples) with unreliable self-correction reasoning.

2025/Self-Correction is More than Refinement_ A Learning Framework for Visual and Language Reasoning Tasks/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:438736cf516f32a4c04c067af904c142963dd61fce14ea022207844a809e37aa
+size 470424

2025/Self-Critique Guided Iterative Reasoning for Multi-hop Question Answering/7f183d00-6ecb-41f2-84cb-009819c2329b_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:810a3aeec368b28927e489ef0af177400e120848906773b740a8a6354d1da952
+size 1078367

2025/Self-Critique Guided Iterative Reasoning for Multi-hop Question Answering/full.md

@@ -0,0 +1,730 @@
+# Self-Critique Guided Iterative Reasoning for Multi-hop Question Answering
+
+Zheng Chu<sup>1</sup>, Huiming Fan<sup>1</sup>, Jingchang Chen<sup>1</sup>, Qianyu Wang<sup>1</sup>, Mingda Yang<sup>1</sup>, Jiafeng Liang<sup>1</sup>, Zhongjie Wang<sup>1</sup>, Hao Li<sup>1</sup>, Guo Tang<sup>1</sup>, Ming Liu<sup>1,2*</sup>, Bing Qin<sup>1,2</sup>
+
+<sup>1</sup>Harbin Institute of Technology, Harbin, China
+
+$^{2}$ Peng Cheng Laboratory, Shenzhen, China
+
+{zchu,mliu}@ir.hit.edu.cn
+
+# Abstract
+
+Although large language models (LLMs) have demonstrated remarkable reasoning capabilities, they still face challenges in knowledge-intensive multi-hop reasoning. Recent work explores iterative retrieval to address complex problems. However, the lack of intermediate guidance often results in inaccurate retrieval and flawed intermediate reasoning, leading to incorrect reasoning. To address these, we propose Self-Critique Guided Iterative Reasoning (SiGIR), which uses self-critique feedback to guide the iterative reasoning process. Specifically, through end-to-end training, we enable the model to iteratively address complex problems via question decomposition. Additionally, the model is able to self-evaluate its intermediate reasoning steps. During iterative reasoning, the model engages in branching exploration and employs self-evaluation to guide the selection of promising reasoning trajectories. Extensive experiments on three multi-hop reasoning datasets demonstrate the effectiveness of our proposed method, surpassing the previous SOTA by  $8.6\%$ . Furthermore, our thorough analysis offers insights for future research. Our code, data, and models are available at  $\mathsf{Gih}\mathsf{b}^{1}$ .
+
+# 1 Introduction
+
+With the advancement of large-scale pre-training techniques, LLMs have achieved remarkable performance on reasoning tasks with chain-of-thought (CoT) prompting (Wei et al., 2022; Chu et al., 2024b; OpenAI, 2023; DeepSeek-AI et al., 2024a). Moreover, when augmented with Slow Thinking mechanisms, several advanced models have demonstrated reasoning capabilities that either surpass or are comparable to those of human experts (DeepSeek-AI et al., 2024b; OpenAI, 2024).
+
+Despite their remarkable capabilities, LLMs are prone to generating factually inaccurate re
+
+![](images/e1761ca04ce0274364b8fcfeb4f823f1367389734bb4c3a0caa82adb2ceb58e0.jpg)  
+How many counties are there in the largest state in the region where the film The Family Stone was set?  
+Figure 1: A comparison of iterative retrieval and self-critique guided reasoning (ours) in MHQA. Our method incorporates self-critique during iterative reasoning and conduct searches based on the received feedback.
+
+sponses in knowledge-intensive tasks, primarily attributable to their limited access to domain-specific knowledge. This limitation becomes particularly pronounced in multi-hop question answering (MHQA) (Yang et al., 2018; Ho et al., 2020; Trivedi et al., 2022), which substantially compromises both the reasoning accuracy and factual reliability of the model's outputs. Furthermore, the complexity of multi-hop questions presents additional challenges for global planning during the reasoning process
+
+Retrieval-augmented generation is a promising strategy for mitigating the knowledge limitations of LLMs (Jiang et al., 2024). However, single-round retrieval is insufficient to encompass all relevant evidences to multi-hop questions (Lewis et al., 2020). To address this, iterative retrieval has been proposed (Trivedi et al., 2023; Asai et al., 2024). By alternating between retrieval and generation, it can acquire more comprehensive relevant knowledge, improving retrieval performance. Besides, some research focuses on the problems from the viewpoint of decomposition (Cao et al., 2023; Chu et al., 2024a; Su et al., 2024). It breaks down the complex multi-hop question into simpler sub-questions, thereby reducing the difficulty of reasoning and
+
+planning, and enabling more accurate retrieval.
+
+Although the aforementioned methods have partially addressed some aspects of challenges, they still suffer from certain limitations. First, decomposing a complex question into sub-questions at the beginning of reasoning is highly challenging, and any mistakes can cause subsequent reasoning to deviate from the correct path. Second, the iterative retrieval approach struggles to handle complex problem planning and articulate retrieval intents clearly, resulting in inaccurate retrieval and reasoning. Finally, the absence of guidance for intermediate steps makes these methods susceptible to errors during reasoning, resulting in cascading errors.
+
+To address these, our research focuses on following key aspects: (i) iterative question decomposition facilitates complex question planning and accurate retrieval. (ii) intermediate feedback guided reasoning expands the reasoning space and reduces cascading errors. Based on this motivation, we introduce Self-Critique Guided Iterative Reasoning (SiGIR). SiGIR iteratively solves complex questions through sub-question decomposition and selects promising reasoning paths guided by the self-critique feedback of intermediate reasoning steps.
+
+Specifically, during training phase, we synthesize iterative reasoning rationales with intermediate supervision from advanced LLMs. Thereafter, the model is equipped with the ability to perform question decomposition, trigger retrieval, knowledge reasoning, self-evaluation, and question reduction by means of end-to-end training. During inference phase, the model conducts iteration-level (step-level) beam search guided by self-feedback, as illustrated in Figure 1. At each reasoning step, the model performs branch expansion for sub-question decomposition, retrieval, and reasoning. It then self-evaluates the current reasoning status and retain the top-k promising candidates based on self-critique rewards. The model continues this iterative reasoning process until the final answer is derived. Finally, the reasoning trajectory with the highest cumulative reward is selected as the final answer.
+
+We conduct extensive experiments on three knowledge-intensive MHQA datasets: HotpotQA, 2WikiMQA, and MuSiQue (Yang et al., 2018; Ho et al., 2020; Trivedi et al., 2022). DeepSeek-V2.5, Mistral, LLaMA2, and Qwen2.5 are used during experiments (DeepSeek-AI et al., 2024a; Jiang et al., 2023a; Touvron et al., 2023; Yang et al., 2024). Our method demonstrates significant advantages, with an average improvement of  $8.6\%$  compared
+
+to the SOTA method. Furthermore, the thorough analysis highlights the effectiveness of our method in retrieval, self-critique, search, and efficiency.
+
+# 2 Related Work
+
+# 2.1 Retrieval-Augmented Generation
+
+Retrieval-augmented generation (RAG) is an effective approach for mitigating hallucinations in LLMs (Gao et al., 2023; Huang et al., 2023). Retrieval-then-read, a widely-used approach in RAG, consists of query-based retrieval followed by conditional generation (Lewis et al., 2020; Borgeaud et al., 2022). However, its effectiveness is critically reliant on the quality of retrieval. To address this, various methods have been proposed. Query rewriting improves retrieval quality by reformulating retrieval queries (Ma et al., 2023; Chan et al., 2024; Xin et al., 2024). Active retrieval dynamically adjusts retrieval timing and intent using adaptive mechanisms (Jiang et al., 2023b; Asai et al., 2024). Multi-round retrieval enables the acquisition of comprehensive information relevant to the problem (Trivedi et al., 2023; Shao et al., 2023).
+
+Our proposed method employs multi-round active retrieval to acquire more precise and comprehensive content for complex multi-hop questions.
+
+# 2.2 Multi-hop Question Answering
+
+MHQA requires gathering multiple pieces of evidence and systematic multi-step reasoning to arrive at the final answer. Generally, single-round retrieval is inadequate to acquire comprehensive information (Lewis et al., 2020; Borgeaud et al., 2022). Some methods adopt question decomposition to enhance accurate retrieval and reasoning (Cao et al., 2023; Chu et al., 2024a; Su et al., 2024). However, decomposing complex questions from the very beginning is particularly challenging. To this end, iterative reasoning approach is introduced, which employs iterative multi-round reasoning to gradually tackle complex problems (Trivedi et al., 2023; Yu et al., 2024; Li et al., 2024).
+
+In this work, we utilize iterative question decomposition and reduction to simplify the decomposition process and improve retrieval effectiveness.
+
+# 2.3 Inference-Time Scaling
+
+Recent studies have found that increasing computational resources at inference can improve model performance, termed inference-time scaling (Snell
+
+![](images/64ab35514a140146f7d35347d856b3b034384ba0f8540e46d23885db57e8a6ce.jpg)  
+Figure 2: Overview of SiGIR. (I) Training recipe to get the iterative reasoner with self-critique,  $R_{sc}$ . (II) Reasoning characteristic of SC-Reasoner, which includes question decomposition, trigger retrieval, retrieval-augmented reasoning, self-evaluation, and question reduction. (III) SC-Reasoner engages in exploration during the iterative reasoning process and conducts reward-guided searches to select promising reasoning trajectories.
+
+![](images/1a9bef3f08280e8f67372c99346985e7c791632af4137ffe0f682da7ca021c6e.jpg)
+
+![](images/54a6b2d623c2125b163be13cb51743c068b7c37b0734f1a19dc84a0e5df162ca.jpg)
+
+et al., 2024; Brown et al., 2024). Wang et al. (2023) employs a sampling ensemble technique to scale up, but the improvement is limited. Other methods adopt search strategies, such as Monte Carlo Tree Search, which yield significant improvements, but involve higher computational overhead (Zhang et al., 2024; Qi et al., 2024; Wang et al., 2024).
+
+In contrast, we introduce reward-guided search that circumvents the overhead that simulations entail. It improves performance while ensuring reasonable costs, leading to efficient reasoning.
+
+# 3 Self-Critique Guided Iterative Reasoning for Multi-hop QA
+
+In this paper, we introduce Self-Critique Guided Iterative Reasoning (SiGIR) framework for multi-hop question answering, as illustrated in Figure 2. First, we synthesize iterative reasoning trajectories with intermediate supervision signals to train the model, which enables it to give self-critique of intermediate steps during reasoning. During reasoning, the model iteratively explores and selects the optimal trajectory guided by self-critique feedback.
+
+In this section, we will outline the data synthesis pipeline and training recipe in Section 3.1, introduce the inference pattern of self-critique iterative reasoner in Section 3.2, and elaborate on the self-critique guided reasoning process in Section 3.3.
+
+# 3.1 Training Recipe
+
+Our goal is to enable the model to perform iterative reasoning and to conduct self-critique of the intermediate reasoning steps. To achieve this, we first train an iterative reasoner  $R$ , followed by a critic model  $C$ . Subsequently, we use  $R$  and  $C$  to generate rationales and train a model  $R_{sc}$  capable of iterative reasoning and self-critique. Figure 2(I) and Algorithm 2 illustrate the training procedure.
+
+Iterative Reasoner. We use advanced LLMs to synthesize iterative reasoning rationales. Specifically, given a corpus  $X$ , where each instance  $x_{i}$  includes a question  $q_{i}$ , an answer  $a_{i}$ , and corresponding documents  $\{d_1,\dots ,d_n\}$ . For each instance, we use few-shot prompts with the question and supporting documents as input to the LLM, which yields question decompositions and subquestion reasoning. Subsequently, we organize the sub-question decomposition, retrieval, and reasoning in an interleaved manner, and incorporate retrieval triggers before retrieved documents. We train the smaller model with the parsed training instances to get the vanilla iterative reasoner  $R$ .
+
+Critic. Due to the high cost of invoking LLMs for reward evaluation, we attempt to use a small model to clone the reward evaluation of LLMs. To
+
+this end, we use iterative reasoning trajectories to construct the data for critic training. Notably, to prevent the critic from overfitting to the generator's training data, we select a small amount of non-overlapping corpus with  $X$  for critic data synthesis, denoted as  $X_{\text{critic}}$ . Additionally, we employ  $R$  instead of the LLM for reasoning synthesis, as  $R$  tends to make mistakes during reasoning. This helps ensure a balance between positive and negative samples. Subsequently, we employ the LLM to evaluate rewards on retrieval, reasoning, and overall quality, obtaining training data for critic model. We use it to train a separate critic model  $C$ .
+
+Self-Critique Reasoner. We aim to incorporate the critique signals into the reasoning process. Besides, recent studies have found that the generator and critic can enhance each other, and merging them into a single model can reduce inference cost. To this end, we use the critic  $C$  to annotate the sub-processes of the reasoning trajectories. For retrieval evaluation, we take the query and retrieved documents as input, with retrieval relevance as outputs. For sub-question reasoning, the inputs include sub-questions, documents, and reasoning. For overall reward, we take the entire reasoning as input, and the output is the overall quality. Subsequently, we add special tokens following the sub-processes, enabling the model to generate self-critique signals throughout the reasoning process. We train the model with this corpus to obtain the reasoner with self-critique, denoted as SC-Reasoner, or  $R_{sc}$ .
+
+SC-Reasoner is capable of reasoning and self-evaluation, facilitating data augmentation. Specifically, it can conduct reasoning on unlabeled questions and evaluate the quality of synthetic rationales, retaining high-quality rationales for subsequent training. This self-improvement pipeline can improve the performance in data scarcity scenarios. We use the standard causal language modeling objective for optimization during training.
+
+# 3.2 Iterative Reasoner with Self-Critique
+
+This section introduces the reasoning patterns of SC-Reasoner. Overall, SC-Reasoner is capable of question decomposition and planning, triggering retrieval, reasoning, and self-evaluating the reasoning process. As illustrated in Figure 2(II), given a question, the model first determines whether the question needs decomposition. If needed, the model decomposes one atomic sub-question at a time. Then, the model triggers external retrieval and conducts knowledge reasoning, along with self-evaluating
+
+the quality of both retrieval and reasoning.
+
+Whenever the model resolves a sub-question, it reduces the original question to identify the remaining unresolved part, thereby simplifying the overall question. The model iteratively repeats this process until the reduced atomic question is resolved, leading to the final answer. At this point, it conducts self-evaluation of the entire reasoning trajectory.
+
+# 3.3 Self-Critique Guided Iterative Reasoning
+
+After training, SC-Reasoner can conduct iterative reasoning and self-critique. Accordingly, we introduce self-critic guided iterative reasoning, an iteration-level beam search algorithm. This algorithm searches within the iterative reasoning process, employing self-rewards to guide the selection of promising reasoning candidates, as illustrated in Figure 2(III) and Algorithm 1. Generally, in each iteration, the model first expands the exploration space by branching based on the current reasoning prefix, and then selects top-k candidate trajectories guided by self-rewards. At the end of the reasoning, we use self-rewards to choose the final answer.
+
+Branch Exploration Each iteration consists of the following steps: question decomposition, retrieval, reasoning, remaining question reduction, and self-critique. At timestamp  $t$ , the model branches out and expands downward through temperature sampling in question decomposition, retrieval, and reasoning, as shown in equations below.
+
+$$
+q _ {t + 1} \sim p \left(y \mid q _ {t} ^ {(r)}, x _ {<   t}\right) \tag {1}
+$$
+
+$$
+s _ {t + 1} \sim p (y \mid q _ {t + 1}, d _ {i}, x _ {<   t}) \qquad (2)
+$$
+
+where  $\sim$  is sampling,  $q_{t}^{(r)}$  is remaining question,  $q$  is sub-question,  $s$  is sub-question reasoning,  $p$  is probabilistic LM, and  $d$  is retrieved documents.
+
+During the aforementioned process, the model also conducts self-evaluation and assigns rewards, as shown in Equ. (2,3). At the end of an iteration, the model will reduce the question for the next iteration based on current reasoning status  $x_{t+1}$ .
+
+$$
+r _ {t + 1} ^ {\text {r e t r}} = p (y \mid q _ {t + 1}, d _ {i}) \tag {3}
+$$
+
+$$
+r _ {t + 1} ^ {\text {r e a s}} = p \left(y \mid q _ {t + 1}, d _ {i}, s _ {t + 1}\right) \tag {4}
+$$
+
+$$
+x _ {t + 1} = \oplus \left[ x _ {t}, q _ {t + 1}, d, r _ {t + 1} ^ {\text {r e t r}}, s _ {t + 1}, r _ {t + 1} ^ {\text {r e a s}} \right] \tag {5}
+$$
+
+$$
+q _ {t + 1} ^ {(r)} = p (y \mid x _ {t + 1}) \tag {6}
+$$
+
+Candidate Selection Through branch exploration, the model's reasoning investigates a wide space. Subsequently, we aim to select the most
+
+<table><tr><td rowspan="2">Methods</td><td colspan="5">2WikiMQA</td><td colspan="3">HotpotQA</td><td colspan="4">MuSiQue</td></tr><tr><td>Overall</td><td>Brg.</td><td>Infer.</td><td>Comp.</td><td>B.C.</td><td>Overall</td><td>Brg.</td><td>Comp.</td><td>Overall</td><td>2hop</td><td>3hop</td><td>4hop</td></tr><tr><td colspan="13">Inference-based Methods</td></tr><tr><td>Chain-of-thought*</td><td>31.45</td><td>11.54</td><td>31.56</td><td>51.90</td><td>49.67</td><td>35.31</td><td>29.97</td><td>54.94</td><td>18.93</td><td>22.57</td><td>16.34</td><td>12.49</td></tr><tr><td>Self-Ask*</td><td>28.20</td><td>11.39</td><td>26.16</td><td>48.04</td><td>41.97</td><td>30.49</td><td>24.12</td><td>53.90</td><td>15.91</td><td>20.81</td><td>12.27</td><td>7.56</td></tr><tr><td>RAG</td><td>40.21</td><td>19.22</td><td>26.02</td><td>69.25</td><td>55.25</td><td>53.42</td><td>50.33</td><td>65.49</td><td>22.06</td><td>30.24</td><td>16.64</td><td>8.32</td></tr><tr><td>IRCoT</td><td>46.80</td><td>29.89</td><td>25.55</td><td>73.13</td><td>61.44</td><td>49.03</td><td>45.33</td><td>63.45</td><td>21.43</td><td>27.15</td><td>18.73</td><td>9.86</td></tr><tr><td>ProbTree†</td><td>59.90</td><td>-</td><td>-</td><td>-</td><td>-</td><td>50.40</td><td>-</td><td>-</td><td>27.00</td><td>-</td><td>-</td><td>-</td></tr><tr><td>BeamAggR†</td><td>63.20</td><td>-</td><td>-</td><td>-</td><td>-</td><td>55.20</td><td>-</td><td>-</td><td>32.30</td><td>-</td><td>-</td><td>-</td></tr><tr><td colspan="13">Training-based Methods</td></tr><tr><td>Self-RAG</td><td>31.60</td><td>16.91</td><td>14.06</td><td>50.39</td><td>49.00</td><td>54.60</td><td>52.51</td><td>62.75</td><td>22.00</td><td>30.04</td><td>13.21</td><td>14.77</td></tr><tr><td>RAG-Star†</td><td>46.80</td><td>-</td><td>-</td><td>-</td><td>-</td><td>60.00</td><td>-</td><td>-</td><td>30.70</td><td>-</td><td>-</td><td>-</td></tr><tr><td>GenGround†</td><td>50.21</td><td>-</td><td>-</td><td>-</td><td>-</td><td>52.26</td><td>-</td><td>-</td><td>27.36</td><td>-</td><td>-</td><td>-</td></tr><tr><td>Auto-RAG</td><td>56.06</td><td>44.12</td><td>40.94</td><td>79.44</td><td>59.48</td><td>50.03</td><td>46.96</td><td>62.02</td><td>22.08</td><td>29.32</td><td>15.90</td><td>11.54</td></tr><tr><td>DR-Distillation</td><td>70.51</td><td>46.13</td><td>56.34</td><td>96.90</td><td>96.00</td><td>58.06</td><td>54.08</td><td>73.62</td><td>22.74</td><td>35.08</td><td>12.06</td><td>6.57</td></tr><tr><td>Ours</td><td>74.47</td><td>56.39</td><td>58.69</td><td>95.47</td><td>94.94</td><td>63.09</td><td>59.45</td><td>77.37</td><td>37.15</td><td>49.51</td><td>27.55</td><td>18.37</td></tr><tr><td>-Qwen2.5</td><td>70.51</td><td>54.98</td><td>40.39</td><td>94.90</td><td>90.72</td><td>60.93</td><td>56.59</td><td>77.87</td><td>33.29</td><td>40.11</td><td>28.18</td><td>22.90</td></tr><tr><td>-LLaMA2</td><td>65.82</td><td>50.93</td><td>43.36</td><td>90.44</td><td>79.25</td><td>52.92</td><td>47.92</td><td>72.46</td><td>28.56</td><td>39.00</td><td>20.60</td><td>12.98</td></tr></table>
+
+Table 1: Experimental results on three multi-hop reasoning datasets: 2WikiMQA, HotpotQA, and MuSiQue. Mistral-7B is used as the default backbone. Best and second results are highlighted by bold and underline. The evaluation metric is F1. * indicates closed-book reasoning, and † denotes results are derived from the original paper.
+
+promising candidates among them based on rewards. Selected candidates will serve as the starting point for the next iteration of reasoning. Specifically, we use the cumulative process reward as the criterion for candidate selection. The cumulative reward consists of reward from previous iteration and current iteration. We retain the candidates with top-k largest rewards, as shown in Equ. (7, 8). Once the reasoning is finished, we need to select one candidate as the final answer, using either cumulative reward of entire trajectory or outcome reward.
+
+$$
+r _ {t + 1} ^ {c} = \underbrace {r _ {t} ^ {c}} _ {\text {p r e v i o u s r e w a r d}} + \underbrace {r _ {t + 1} ^ {\text {r e t r}} + r _ {t + 1} ^ {\text {r e a s}}} _ {\text {c u r r e n t r e w a r d}} \tag {7}
+$$
+
+$$
+\boldsymbol {x} _ {t + 1} ^ {*} = \operatorname {t o p - k} \left(\boldsymbol {x} _ {t + 1}, \boldsymbol {r} _ {t + 1} ^ {c}\right) \tag {8}
+$$
+
+$$
+r ^ {c} = \sum_ {i = 1} ^ {n} \left(r _ {i} ^ {\text {r e t r}} + r _ {i} ^ {\text {r e a s}}\right) \tag {9}
+$$
+
+$$
+r ^ {o} = p \left(y \mid q ^ {(r)}, x _ {<   n}, a\right) \tag {10}
+$$
+
+where  $r^c$  and  $r^o$  denote cumulative process reward and outcome reward, respectively.
+
+# 4 Experimental Setup
+
+# 4.1 Benchmarks
+
+We evaluate SiGIR on three knowledge-intensive multi-hop reasoning datasets: HotpotQA (Yang et al., 2018), 2WikiMQA (Ho et al., 2020), and MuSiQue (Trivedi et al., 2022). Following previous work (Trivedi et al., 2023; Cao et al., 2023), we
+
+use the development set as the test data. The experiments are conducted in the open-domain setting. We use token-level F1 as the evaluation metric.
+
+# 4.2 Baselines
+
+We select extensive baselines for comparison, including the following methods: Fewshot CoT (Wei et al., 2022), Self-Ask (Press et al., 2023), IR-CoT (Trivedi et al., 2023), ProbTree (Cao et al., 2023), BeamAggR (Chu et al., 2024a), RAG (Asai et al., 2024), RAG-Star (Jiang et al., 2024), GenGround (Shi et al., 2024), Auto-RAG (Yu et al., 2024), and DR-Distillation (Li et al., 2024).
+
+# 4.3 Implementation Details
+
+We use DeepSeek-v2.5 (DeepSeek-AI et al., 2024a) as the advanced LLM in our experiments. By default, we use BM25 (Robertson and Zaragoza, 2009) for sparse retrieval. In analysis, we additionally use Contriever (Izacard et al., 2022) for dense retrieval. During training, we use LoRA (Hu et al., 2022) for parameter-efficient fine-tuning, with Huggingface Accelerate for mixed-precision training. The entire training process takes approximately 3 hours on  $4 \times$  A100-SXM4-80GB GPUs.
+
+<table><tr><td>Settings</td><td>2WikiMQA</td><td>HotpotQA</td><td>MuSiQue</td></tr><tr><td>SGIR</td><td>74.47</td><td>63.09</td><td>37.15</td></tr><tr><td colspan="4">Fine-grained Reward</td></tr><tr><td>(a) w/o. guided search</td><td>72.37</td><td>60.86</td><td>35.47</td></tr><tr><td>(b) w/o. reward</td><td>67.23</td><td>54.94</td><td>29.45</td></tr><tr><td>(c) w/o. self-critique</td><td>66.96</td><td>56.86</td><td>28.62</td></tr><tr><td colspan="4">Coarse-grained Reward</td></tr><tr><td>(d) w. coarse reward</td><td>73.40</td><td>59.85</td><td>34.32</td></tr><tr><td>(e) w/o. guided search</td><td>72.32</td><td>60.89</td><td>35.40</td></tr></table>
+
+# 5 Experimental Results
+
+We conduct extensive experiments on three multi-hop reasoning datasets, as shown in Table 1. We notice that although CoT is considered proficient in multi-step reasoning, but its performance is poor due to a lack of question-relevant knowledge within the LLM. When single-round retrieval is employed, an average performance improvement of  $35.0\%$  is observed. Iterative multi-round retrieval results in improved performance, achieving up to a  $48.8\%$  increase on 2WikiMQA. Additionally, methods based on question decomposition further improve reasoning performance. These results suggest that precise retrieval and reasoning planning are key factors affecting multi-hop reasoning.
+
+Some methods equip the model with iterative reasoning capabilities through fine-tuning, introducing strategies such as active retrieval, question decomposition, and search during the iterative reasoning process. These methods can better address complex problems and achieve more accurate retrieval, further enhancing the performance.
+
+As shown in Table 1, our method demonstrates remarkable superiority over baselines. It surpasses the previous state-of-the-art methods on all three datasets, achieving improvement of 2WikiMQA (+3.96), HotpotQA (+3.09), and MuSiQue (+4.85). Furthermore, our method excels particularly on the challenging MuSiQue dataset, which involves 3 to 4 hop hard questions. It achieves a significant improvement of  $15.0\%$  on the MuSiQue dataset compared to previous methods, with enhancements of  $47.1\%$  and  $24.37\%$  on 3/4-hop challenging questions, respectively. Additionally, we also conduct experiments on the Qwen2.5 and LLaMA2 models, achieving competitive performance that demonstrates the generalizability of our method.
+
+We identify three key factors that contribute to
+
+Table 2: Ablation experiments on rewards and search algorithm in fine-grained and coarse-grained scenarios. The experiments are conducted with Mistral-7B.  
+
+<table><tr><td>Stage</td><td>#Labeled</td><td>#Unlabeled</td><td>#Total</td><td>Quantile</td></tr><tr><td>Training Set</td><td>49000</td><td>-</td><td>-</td><td>-</td></tr><tr><td>Full SFT</td><td>38625</td><td>0</td><td>38625</td><td>100%</td></tr><tr><td>SP Start Point</td><td>15000</td><td>0</td><td>15000</td><td>95.3%</td></tr><tr><td>First Iteration</td><td>15000</td><td>21669</td><td>36669</td><td>97.1%</td></tr><tr><td>Second Iteration</td><td>15000</td><td>29659</td><td>44659</td><td>97.8%</td></tr></table>
+
+Table 3: Comparison of the amount of data used for training and performance quantiles. #Labeled refers to data annotated by the teacher model with reasoning trajectories, and #Unlabeled refers to data synthesized by the smaller model itself. Quantile indicates the percentage of performance compared to full SFT.
+
+our improvement: (a) Iterative multi-round reasoning and sub-question planning reduce the complexity of questions while improving retrieval accuracy. (b) Self-critique provides valuable feedback signals for intermediate steps during the reasoning process. (c) Reward-guided search selects superior reasoning trajectories among the exploration space.
+
+# 6 Analysis
+
+# 6.1 Ablation Study
+
+We conduct ablation experiments to investigate the effects of different modules, as shown in Table 2. Our method employs fine-grained rewards by default. First, removing reward-guided search (a) and substituting it with greedy inference leads to considerable performance declines on all three datasets, highlighting the importance of search algorithms. Next, we remove the reward signal from greedy inference (b), which means the model's inference process would not receive any guidance. This change results in a substantial performance decrease of  $10\%$ , demonstrating the essential role of reward signals in the inference process. Additionally, the exclusion of self-critique data from the training data produces a minor performance decline.
+
+To further investigate the role of reward signals, particularly the impact of reward granularity on model performance, we replace the rewards with coarse-grained binary rewards, as shown in Table 2. In reward-guided greedy inference (e), fine-grained and coarse-grained rewards have similar effects. However, when a broader search is required (d), coarse-grained rewards exhibit certain drawbacks, resulting in a  $2.7\%$  performance decrease compared to fine-grained rewards. This indicates that fine-grained rewards are more effective in search-form reasoning, particularly when there are a substantial number of intermediate reasoning steps.
+
+![](images/fffdfa5c2893dff3f27169c84ae89970065d579e9b01c29396b67d16b116fa2a.jpg)  
+Figure 3: The impact of increasing training data and self-improvement on performance. As data scales, the performance consistently improves. Starting with  $40\%$  of the data, two self-improvement iterations boost performance to nearly  $80\%$ , with only  $2\%$  gap with full sft.
+
+<table><tr><td>Strategy</td><td>2WikiMQA</td><td>HotpotQA</td><td>MuSiQue</td><td>Average</td></tr><tr><td colspan="5">Search Strategy</td></tr><tr><td>Guided Search</td><td>74.47</td><td>63.09</td><td>37.15</td><td>58.23</td></tr><tr><td>(a) one candidate</td><td>72.97</td><td>61.27</td><td>36.21</td><td>56.81</td></tr><tr><td>(b) no branching</td><td>68.87</td><td>52.88</td><td>28.34</td><td>50.03</td></tr><tr><td>Greedy Inference</td><td>72.37</td><td>60.86</td><td>35.47</td><td>56.23</td></tr><tr><td colspan="5">Selection Strategy</td></tr><tr><td>(c) cumulative reward</td><td>74.47</td><td>63.09</td><td>37.15</td><td>58.23</td></tr><tr><td>(d) outcome reward</td><td>71.79</td><td>60.96</td><td>35.23</td><td>55.99</td></tr><tr><td>(e) random select</td><td>71.17</td><td>60.10</td><td>35.17</td><td>55.81</td></tr></table>
+
+# 6.2 Training Scaling and Self-Improvement
+
+We investigate the performance under different amounts of training data and the effect of self-improvement data synthesis, as shown in Figure 3. Initially, as the amount of labeled data increases, the performance improves consistently. The most significant improvement occurs when  $40\%$  of the training data is utilized; beyond this threshold, the rate of performance enhancement decreases.
+
+Starting with the model trained on  $40\%$  of the data, we perform iterative self-improvement. As shown in Figure 3 and Table 3, in the first iteration of self-improvement, the model achieves performance corresponding to the  $60\%$  training data quantile. In the second iteration, the model's performance approaches the  $80\%$  quantile, with only a  $2.2\%$  gap from the performance achieved with the full labeled data fine-tuning. This suggests that our approach can effectively synthesize data through self-critique, yielding competitive performance in situations where labeled data is scarce.
+
+Table 4: Analysis on various search and voting strategies on our method. Retaining more candidate reasoning trajectories and exploring different sub-questions help improve performance. Utilizing cumulative process rewards to select the final reasoning trajectory is effective.  
+
+<table><tr><td>Reward / Task</td><td>Retrieval</td><td>Reasoning</td><td>Overall</td><td>Average</td></tr><tr><td>Direct Reward</td><td>93.2</td><td>95.8</td><td>81.2</td><td>90.1</td></tr><tr><td>Post Explanation</td><td>95.4</td><td>95.4</td><td>80.6</td><td>90.5</td></tr><tr><td>Chain-of-thought</td><td>97.4</td><td>88.6</td><td>87.4</td><td>91.1</td></tr><tr><td>- Dedicated RM</td><td>95.8</td><td>93.6</td><td>78.8</td><td>89.4</td></tr></table>
+
+Table 5: The performance of different reward forms on reward modeling tasks. Generative rewards are more effective than discriminative rewards, and mixed reward modeling has advantages over dedicated reward models in complex outcome reward modeling task.  
+
+<table><tr><td>Strategy</td><td>2WikiMQA</td><td>HotpotQA</td><td>MuSiQue</td><td>Average</td><td>Throughput</td></tr><tr><td>GenCritic</td><td>72.37</td><td>60.86</td><td>35.47</td><td>56.23</td><td>1.92 QPS</td></tr><tr><td>Seperated</td><td>72.23</td><td>58.45</td><td>36.46</td><td>55.71</td><td>0.43 QPS</td></tr></table>
+
+Table 6: Comparison of performance and efficiency between generative and separated critic model. The generative critic model brings slightly better performance, but has significant higher inference throughput.
+
+# 6.3 Search and Trajectory Selection
+
+We conduct experiments to investigate the factors influencing the guided reasoning process, including search strategies and reasoning selection, as shown in Table 4. As indicated in the preceding discussion, reward-based search is more effective than greedy inference. During the search process, we retain only one candidate trajectory in each iteration (a), which results in a performance decrease of  $2.4\%$ . This indicates that maintaining more candidate sequences preserves a greater number of potentially correct reasoning trajectories. Next, we set the question decomposition exploration width to 1 (b), which means only retrieval will be expanded during reasoning. This leads to a substantial drop of  $13.57\%$ , indicating the significance of exploring various sub-questions in the reasoning process.
+
+At the end of the search, we need to select one trajectory as the final answer, and we investigate various answer selection strategies. Using cumulative process rewards for trajectory selection yields the best performance, followed by using outcome rewards. Both strategies outperform random selection. This suggests that in complex reasoning processes, intermediate rewards can serve as a criterion for assessing the reasoning quality, and that process rewards are more effective at distinguishing between different levels of reasoning quality.
+
+# 6.4 Reward Modeling
+
+The forms of reward (critic) models are diverse, such as generative and discriminative forms. We
+
+<table><tr><td>Retriever</td><td>2WikiMQA</td><td>HotpotQA</td><td>MuSiQue</td><td>Average</td></tr><tr><td colspan="5">Reward-guided Greedy Inference</td></tr><tr><td>Sparse</td><td>72.37</td><td>60.86</td><td>35.47</td><td>56.23</td></tr><tr><td>Dense</td><td>75.44</td><td>59.53</td><td>38.48</td><td>57.81</td></tr><tr><td>Hybrid</td><td>76.04</td><td>62.89</td><td>39.38</td><td>59.43</td></tr><tr><td colspan="5">Reward-guided Search</td></tr><tr><td>Sparse</td><td>74.47</td><td>63.09</td><td>37.15</td><td>58.23</td></tr><tr><td>Dense</td><td>75.72</td><td>61.72</td><td>40.36</td><td>59.26</td></tr><tr><td>Hybrid</td><td>76.05</td><td>63.47</td><td>37.05</td><td>58.85</td></tr></table>
+
+examine the performance of various types of critic models on our reward modeling task, as shown in Table 5. Overall, generative reward modeling outperforms discriminative modeling. Moreover, among generative forms, chain-of-thought (CoT) is preferred over post-hoc explanations. Notably, CoT excels at modeling complex outcome rewards because it systematically decomposes the entire reasoning trajectory for step-by-step analysis.
+
+Additionally, we investigate dedicated RM (each model assigned to a specific task) in comparison to mixed RM (one model responsible for all tasks). The dedicated RM exhibits performance variations of  $-1.6\% / +5.6\%$  for retrieval and reasoning tasks, respectively. However, there is a significant performance decrease of  $9.8\%$  in modeling outcome rewards. Since outcome rewards assess the entire reasoning process by integrating various reward modeling tasks, this suggests that multi-task reward modeling of sub-tasks enhances the model's ability to evaluate complex outcome rewards.
+
+Furthermore, we highlight that generative critic models offer greater benefits than separated generators and critics. As shown in Table 6, the GenCritic outperforms the separated models in two of the three datasets, achieving an average performance improvement of  $0.9\%$ . Without sacrificing reasoning performance, the GenCritic achieves a throughput that is 4.46 times higher, demonstrating significantly better reasoning efficiency<sup>5</sup>.
+
+# 6.5 Retrieval System and Hybrid Retrieval
+
+Retrieval plays a critical role in knowledge reasoning. To investigate the effect of the retrieval system, we employ various retrieval systems in the experiments, as illustrated in Table 7. Dense retrieval
+
+Table 7: The impact of retrieval systems on reasoning. Accurate retrieval leads to better reasoning performance. Our method naturally supports hybrid retrieval (sparse + dense) retrieval, which facilitates better reasoning.  
+
+<table><tr><td>Retriever</td><td>2WikiMQA</td><td>HotpotQA</td><td>MuSiQue</td><td>Average</td></tr><tr><td>BM25</td><td>72.37</td><td>60.86</td><td>35.47</td><td>56.23</td></tr><tr><td>Contriever</td><td>75.44</td><td>59.53</td><td>38.48</td><td>57.81</td></tr><tr><td>+ Hybrid</td><td>76.04</td><td>62.89</td><td>39.38</td><td>59.43</td></tr><tr><td>BGE-large</td><td>74.36</td><td>64.27</td><td>40.58</td><td>59.74</td></tr><tr><td>+ Hybrid</td><td>75.85</td><td>64.25</td><td>41.03</td><td>60.38</td></tr><tr><td>ColBERT-v2</td><td>75.37</td><td>65.44</td><td>42.24</td><td>61.02</td></tr><tr><td>+ Hybrid</td><td>75.37</td><td>65.69</td><td>43.05</td><td>61.37</td></tr></table>
+
+Table 8: Experiments on different dense retrievers and their hybrid retrieval variants. Hybrid retrieval strategies improve performance across various dense retrievers.
+
+![](images/5f9e8e4a9653cd3507a752b44bd72f8df133912dbe0f606ea112b07c9dc3d6c5.jpg)  
+Figure 4: Relationship between performance and the average number of iterations to solve a multi-hop question. Upper-left represents efficient reasoning. Our method requires fewer iterations to solve one question with higher F1, demonstrating higher reasoning efficiency.
+
+has better performance than sparse retrieval. When employing dense retrieval, a notable improvement of  $2.8\%$  is observed, suggesting that retrieval quality impacts reasoning performance. Furthermore, since our method treats each retrieved document independently, it naturally facilitates the integration of different retrieval systems. The hybrid retrieval system achieves a significant improvement of  $5.7\%$ , demonstrating the superiority of our method. However, in a guided search, excellent reasoning performance can be achieved using only sparse retrieval. This suggests that when retrieval quality is poor, the reward-guided search can help the model in exploring and filtering irrelevant retrieved documents, thereby improving reasoning performance.
+
+In addition, we also apply hybrid retrieval on different dense retrievers (BGE-large $^{6}$  and ColBERTv2 $^{7}$ ), as shown in Table 8. The hybrid retrieval strategy consistently achieves significant performance improvements across different dense retrievers, demonstrating its effectiveness.
+
+# 6.6 Iterative Reasoning Efficiency
+
+We assess reasoning efficiency by measuring the iterations needed to solve a multi-hop problem alongside its reasoning performance, as illustrated in Figure 4. Among the mentioned methods, AutoRAG / DR-Distillation shows commendable reasoning capabilities, but it demands more iterations of reasoning, which incurs substantial costs. Conversely, our method not only surpasses their performance but also reduces the average number of reasoning iterations by  $26.7\%$ , demonstrating superior reasoning efficiency. Additionally, we also measure the average inference latency of single samples and token consumption in Appendix B.2.
+
+# 7 Conclusion
+
+In this work, we introduce Self-Critique Guided Iterative Reasoning for knowledge-intensive multihop question answering. Initially, we enable the model to conduct iterative reasoning through iterative question decomposition and to self-evaluate the intermediate reasoning steps, through a novel data synthesis and training pipeline. During inference, the model performs self-critique guided search: it first explores multiple reasoning paths via branching exploration and then utilizes self-critique feedback to refine the search process and identify the most promising reasoning trajectory. Extensive experiments and thorough analysis demonstrate the effectiveness of our proposed method.
+
+# Limitations
+
+Although we have conducted comprehensive experiments to demonstrate the effectiveness of our proposed methods, and provided detailed analysis on the components of our system, this paper still has some limitations.
+
+For sub-question exploration, we employ a sampling-based approach rather than training a dedicated question decomposition module. Although this design improves the simplicity and efficiency of the overall system, it can limit the exploratory breadth of the model with respect to sub-questions. We plan to address this in future research, such as incorporating a query rewriting module.
+
+Our method adopts a route of synthetic data followed by supervised fine-tuning (SFT), without incorporating reinforcement learning in the model's sub-question reasoning process. In the future, we consider using strategies such as Monte
+
+Carlo tree search to generate richer reasoning trajectories and introducing reinforcement learning to optimize these trajectories, thereby further enhancing the model's reasoning performance.
+
+# Acknowledgements
+
+The research in this article is supported by the National Key Research and Development Project (2022YFF0903301), the National Science Foundation of China (U22B2059, 62276083). We also appreciate the support from China Mobile Group Heilongjiang Co., Ltd. @ on our research, the research is jointly completed by both parties. Ming Liu is the corresponding author.
+
+# References
+
+Akari Asai, Zeqiu Wu, Yizhong Wang, Avirup Sil, and Hannaneh Hajishirzi. 2024. Self-rag: Learning to retrieve, generate, and critique through self-reflection. In The Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024. OpenReview.net.  
+Sebastian Borgeaud, Arthur Mensch, Jordan Hoffmann, Trevor Cai, Eliza Rutherford, Katie Millican, George van den Driessche, Jean-Baptiste Lespiau, Bogdan Damoc, Aidan Clark, Diego de Las Casas, Aurelia Guy, Jacob Menick, Roman Ring, Tom Hennigan, Saffron Huang, Loren Maggiore, Chris Jones, Albin Cassirer, Andy Brock, Michela Paganini, Geoffrey Irving, Oriol Vinyals, Simon Osindero, Karen Simonyan, Jack W. Rae, Erich Elsen, and Laurent Sifre. 2022. Improving language models by retrieving from trillions of tokens. In International Conference on Machine Learning, ICML 2022, 17-23 July 2022, Baltimore, Maryland, USA, volume 162 of Proceedings of Machine Learning Research, pages 2206-2240. PMLR.  
+Bradley C. A. Brown, Jordan Juravsky, Ryan Saul Ehrlich, Ronald Clark, Quoc V. Le, Christopher Ré, and Azalia Mirhoseini. 2024. Large language monkeys: Scaling inference compute with repeated sampling. CoRR, abs/2407.21787.  
+Shulin Cao, Jiajie Zhang, Jiaxin Shi, Xin Lv, Zijun Yao, Qi Tian, Lei Hou, and Juanzi Li. 2023. Probabilistic tree-of-thought reasoning for answering knowledge-intensive complex questions. In *Findings of the Association for Computational Linguistics: EMNLP* 2023, Singapore, December 6-10, 2023, pages 12541-12560. Association for Computational Linguistics.  
+Chi-Min Chan, Chunpu Xu, Ruibin Yuan, Hongyin Luo, Wei Xue, Yike Guo, and Jie Fu. 2024. RQ-RAG: learning to refine queries for retrieval augmented generation. CoRR, abs/2404.00610.  
+Zheng Chu, Jingchang Chen, Qianglong Chen, Haotian Wang, Kun Zhu, Xiyuan Du, Weijiang Yu, Ming
+
+Liu, and Bing Qin. 2024a. Beamaggr: Beam aggregation reasoning over multi-source knowledge for multi-hop question answering. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), ACL 2024, Bangkok, Thailand, August 11-16, 2024, pages 1229-1248. Association for Computational Linguistics.  
+Zheng Chu, Jingchang Chen, Qianglong Chen, Weijiang Yu, Tao He, Haotian Wang, Weihua Peng, Ming Liu, Bing Qin, and Ting Liu. 2024b. Navigate through enigmatic labyrinth A survey of chain of thought reasoning: Advances, frontiers and future. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), ACL 2024, Bangkok, Thailand, August 11-16, 2024, pages 1173-1203. Association for Computational Linguistics.  
+DeepSeek-AI, Aixin Liu, Bei Feng, Bin Wang, Bingxuan Wang, Bo Liu, Chenggang Zhao, Chengqi Deng, Chong Ruan, Damai Dai, Daya Guo, Dejian Yang, Deli Chen, Dongjie Ji, Erhang Li, Fangyun Lin, Fuli Luo, Guangbo Hao, Guanting Chen, Guowei Li, Hao Zhang, Hanwei Xu, Hao Yang, Haowei Zhang, Honghui Ding, Huajian Xin, Huazuo Gao, Hui Li, Hui Qu, J. L. Cai, Jian Liang, Jianzhong Guo, Ji-aqi Ni, Jiashi Li, Jin Chen, Jingyang Yuan, Junjie Qiu, Junxiao Song, Kai Dong, Kaige Gao, Kang Guan, Lean Wang, Lecong Zhang, Lei Xu, Leyi Xia, Liang Zhao, Liyue Zhang, Meng Li, Miaojun Wang, Mingchuan Zhang, Minghua Zhang, Minghui Tang, Mingming Li, Ning Tian, Panpan Huang, Peiyi Wang, Peng Zhang, Qihao Zhu, Qinyu Chen, Qiushi Du, R. J. Chen, R. L. Jin, Ruiqi Ge, Ruizhe Pan, Runxin Xu, Ruyi Chen, S. S. Li, Shanghao Lu, Shangyan Zhou, Shanhuang Chen, Shaoqing Wu, Shengfeng Ye, Shirong Ma, Shiyu Wang, Shuang Zhou, Shuiping Yu, Shunfeng Zhou, Size Zheng, Tao Wang, Tian Pei, Tian Yuan, Tianyu Sun, W. L. Xiao, Wangding Zeng, Wei An, Wen Liu, Wenfeng Liang, Wenjun Gao, Wentao Zhang, X. Q. Li, Xiangyue Jin, Xi-anzu Wang, Xiao Bi, Xiaodong Liu, Xiaohan Wang, Xiaojin Shen, Xiaokang Chen, Xiaosha Chen, Xiaotao Nie, and Xiaowen Sun. 2024a. Deepseek-v2: A strong, economical, and efficient mixture-of-experts language model. CoRR, abs/2405.04434.  
+DeepSeek-AI, Aixin Liu, Bei Feng, Bing Xue, Bingxuan Wang, Bochao Wu, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, Damai Dai, Daya Guo, Dejian Yang, Deli Chen, Dongjie Ji, Erhang Li, Fangyun Lin, Fucong Dai, Fuli Luo, Guangbo Hao, Quanting Chen, Guowei Li, H. Zhang, Han Bao, Hanwei Xu, Haocheng Wang, Haowei Zhang, Honghui Ding, Huajian Xin, Huazuo Gao, Hui Li, Hui Qu, J. L. Cai, Jian Liang, Jianzhong Guo, Jiaqi Ni, Jiashi Li, Jiawei Wang, Jin Chen, Jingchang Chen, Jingyang Yuan, Junjie Qiu, Junlong Li, Junxiao Song, Kai Dong, Kai Hu, Kaige Gao, Kang Guan, Kexin Huang, Kuai Yu, Lean Wang, Lecong Zhang, Lei Xu, Leyi Xia, Liang Zhao, Litong Wang, Liyue Zhang, Meng Li, Miaojun Wang
+
+Mingchuan Zhang, Minghua Zhang, Minghui Tang, Mingming Li, Ning Tian, Panpan Huang, Peiyi Wang, Peng Zhang, Qiancheng Wang, Qihao Zhu, Qinyu Chen, Qiushi Du, R. J. Chen, R. L. Jin, Ruiqi Ge, Ruisong Zhang, Ruizhe Pan, Runji Wang, Runxin Xu, Ruoyu Zhang, Ruyi Chen, S. S. Li, Shanghao Lu, Shangyan Zhou, Shanhuang Chen, Shaoqing Wu, Shengfeng Ye, Shengfeng Ye, Shirong Ma, Shiyu Wang, Shuang Zhou, Shuiping Yu, Shunfeng Zhou, Shuting Pan, T. Wang, Tao Yun, Tian Pei, Tianyu Sun, W. L. Xiao, and Wangding Zeng. 2024b. Deepseekv3 technical report. CoRR, abs/2412.19437.  
+Yunfan Gao, Yun Xiong, Xinyu Gao, Kangxiang Jia, Jinliu Pan, Yuxi Bi, Yi Dai, Jiawei Sun, Qianyu Guo, Meng Wang, and Haofen Wang. 2023. Retrievalaugmented generation for large language models: A survey. CoRR, abs/2312.10997.  
+Xanh Ho, Anh-Khoa Duong Nguyen, Saku Sugawara, and Akiko Aizawa. 2020. Constructing A multi-hop QA dataset for comprehensive evaluation of reasoning steps. In Proceedings of the 28th International Conference on Computational Linguistics, COLING 2020, Barcelona, Spain (Online), December 8-13, 2020, pages 6609-6625. International Committee on Computational Linguistics.  
+Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2022. Lora: Low-rank adaptation of large language models. In *The Tenth International Conference on Learning Representations*, ICLR 2022, Virtual Event, April 25-29, 2022. OpenReview.net.  
+Lei Huang, Weijiang Yu, Weitao Ma, Weihong Zhong, Zhangyin Feng, Haotian Wang, Qianglong Chen, Weihua Peng, Xiaocheng Feng, Bing Qin, and Ting Liu. 2023. A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions. CoRR, abs/2311.05232.  
+Gautier Izacard, Mathilde Caron, Lucas Hosseini, Sebastian Riedel, Piotr Bojanowski, Armand Joulin, and Edouard Grave. 2022. Unsupervised dense information retrieval with contrastive learning. Trans. Mach. Learn. Res., 2022.  
+Albert Q. Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de Las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, Lélio Renard Lavaud, Marie-Anne Lachaux, Pierre Stock, Teven Le Scao, Thibaut Lavril, Thomas Wang, Timothée Lacroix, and William El Sayed. 2023a. Mistral 7b. CoRR, abs/2310.06825.  
+Jinhao Jiang, Jiayi Chen, Junyi Li, Ruiyang Ren, Shijie Wang, Wayne Xin Zhao, Yang Song, and Tao Zhang. 2024. Rag-star: Enhancing deliberative reasoning with retrieval augmented verification and refinement. CoRR, abs/2412.12881.  
+Zhengbao Jiang, Frank F. Xu, Luyu Gao, Zhiqing Sun, Qian Liu, Jane Dwivedi-Yu, Yiming Yang, Jamie
+
+Callan, and Graham Neubig. 2023b. Active retrieval augmented generation. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, EMNLP 2023, Singapore, December 6-10, 2023, pages 7969-7992. Association for Computational Linguistics.  
+Patrick S. H. Lewis, Ethan Perez, Aleksandra Pik-tus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Kuttler, Mike Lewis, Wen-tau Yih, Tim Roektaschel, Sebastian Riedel, and Douwe Kiela. 2020. Retrieval-augmented generation for knowledge-intensive NLP tasks. In Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual.  
+Xiang Li, Shizhu He, Fangyu Lei, JunYang JunYang, Tianhuang Su, Kang Liu, and Jun Zhao. 2024. Teaching small language models to reason for knowledge-intensive multi-hop question answering. In *Findings of the Association for Computational Linguistics*, ACL 2024, Bangkok, Thailand and virtual meeting, August 11-16, 2024, pages 7804-7816. Association for Computational Linguistics.  
+Xinbei Ma, Yeyun Gong, Pengcheng He, Hai Zhao, and Nan Duan. 2023. Query rewriting in retrieval-augmented large language models. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, EMNLP 2023, Singapore, December 6-10, 2023, pages 5303-5315. Association for Computational Linguistics.  
+OpenAI. 2023. GPT-4 technical report. CoRR, abs/2303.08774.  
+OpenAI. 2024. Introducing openai o1.  
+Ofir Press, Muru Zhang, Sewon Min, Ludwig Schmidt, Noah A. Smith, and Mike Lewis. 2023. Measuring and narrowing the compositionality gap in language models. In *Findings of the Association for Computational Linguistics: EMNLP* 2023, Singapore, December 6-10, 2023, pages 5687-5711. Association for Computational Linguistics.  
+Zhenting Qi, Mingyuan Ma, Jiahang Xu, Li Lyna Zhang, Fan Yang, and Mao Yang. 2024. Mutual reasoning makes smaller llms stronger problem-solvers. CoRR, abs/2408.06195.  
+Stephen E. Robertson and Hugo Zaragoza. 2009. The probabilistic relevance framework: BM25 and beyond. Found. Trends Inf. Retr., 3(4):333-389.  
+Zhihong Shao, Yeyun Gong, Yelong Shen, Minlie Huang, Nan Duan, and Weizhu Chen. 2023. Enhancing retrieval-augmented large language models with iterative retrieval-generation synergy. In Findings of the Association for Computational Linguistics: EMNLP 2023, Singapore, December 6-10, 2023, pages 9248-9274. Association for Computational Linguistics.
+
+Zhengliang Shi, Shuo Zhang, Weiwei Sun, Shen Gao, Pengjie Ren, Zhumin Chen, and Zhaochun Ren. 2024. Generate-then-ground in retrieval-augmented generation for multi-hop question answering. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), ACL 2024, Bangkok, Thailand, August 11-16, 2024, pages 7339-7353. Association for Computational Linguistics.  
+Charlie Snell, Jaehoon Lee, Kelvin Xu, and Aviral Kumar. 2024. Scaling LLM test-time compute optimally can be more effective than scaling model parameters. CoRR, abs/2408.03314.  
+Xin Su, Tiep Le, Steven Bethard, and Phillip Howard. 2024. Semi-structured chain-of-thought: Integrating multiple sources of knowledge for improved language model reasoning. In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), NAACL 2024, Mexico City, Mexico, June 16-21, 2024, pages 8597-8613. Association for Computational Linguistics.  
+Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajwal Bhargava, Shruti Bhosale, Dan Bikel, Lukas Blecher, Cristian Canton-Ferrer, Moya Chen, Guillem Cucurull, David Esiobu, Jude Fernandes, Jeremy Fu, Wenyin Fu, Brian Fuller, Cynthia Gao, Vedanuj Goswami, Naman Goyal, Anthony Hartshorn, Saghar Hosseini, Rui Hou, Hakan Inan, Marcin Kardas, Viktor Kerkez, Madian Khabsa, Isabel Kloumann, Artem Korenev, Punit Singh Koura, Marie-Anne Lachaux, Thibaut Lavril, Jenya Lee, Diana Liskovich, Yinghai Lu, Yuning Mao, Xavier Martinet, Todor Mihaylov, Pushkar Mishra, Igor Molybog, Yixin Nie, Andrew Poulton, Jeremy Reizenstein, Rashi Rungta, Kalyan Saladi, Alan Schelten, Ruan Silva, Eric Michael Smith, Ranjan Subramanian, Xiaoqing Ellen Tan, Binh Tang, Ross Taylor, Adina Williams, Jian Xiang Kuan, Puxin Xu, Zheng Yan, Iliyan Zarov, Yuchen Zhang, Angela Fan, Melanie Kambadur, Sharan Narang, Aurélien Rodríguez, Robert Stojnic, Sergey Edunov, and Thomas Scialom. 2023. Llama 2: Open foundation and finetuned chat models. CoRR, abs/2307.09288.  
+Harsh Trivedi, Niranjan Balasubramanian, Tushar Khot, and Ashish Sabharwal. 2022. Musique: Multi-hop questions via single-hop question composition. Trans. Assoc. Comput. Linguistics, 10:539-554.  
+Harsh Trivedi, Niranjan Balasubramanian, Tushar Khot, and Ashish Sabharwal. 2023. Interleaving retrieval with chain-of-thought reasoning for knowledge-intensive multi-step questions. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), ACL 2023, Toronto, Canada, July 9-14, 2023, pages 10014-10037. Association for Computational Linguistics.
+
+Chaojie Wang, Yanchen Deng, Zhiyi Lv, Zeng Liang, Jujie He, Shuicheng Yan, and Bo An. 2024. Q*: Improving multi-step reasoning for llms with deliberative planning. CoRR, abs/2406.14283.  
+Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc V. Le, Ed H. Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou. 2023. Self-consistency improves chain of thought reasoning in language models. In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023. OpenReview.net.  
+Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed H. Chi, Quoc V. Le, and Denny Zhou. 2022. Chain-of-thought prompting elicits reasoning in large language models. In Advances in Neural Information Processing Systems 35: Annual Conference on Neural Information Processing Systems 2022, NeurIPS 2022, New Orleans, LA, USA, November 28 - December 9, 2022.  
+Chunlei Xin, Yaojie Lu, Hongyu Lin, Shuheng Zhou, Huijia Zhu, Weiqiang Wang, Zhongyi Liu, Xianpei Han, and Le Sun. 2024. Chain-of-rewrite: Aligning question and documents for open-domain question answering. In Findings of the Association for Computational Linguistics: EMNLP 2024, Miami, Florida, USA, November 12-16, 2024, pages 1884-1896. Association for Computational Linguistics.  
+An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, Huan Lin, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Yang, Jiaxi Yang, Jingren Zhou, Junyang Lin, Kai Dang, Keming Lu, Keqin Bao, Kexin Yang, Le Yu, Mei Li, Mingfeng Xue, Pei Zhang, Qin Zhu, Rui Men, Runji Lin, Tianhao Li, Tingyu Xia, Xingzhang Ren, Xuancheng Ren, Yang Fan, Yang Su, Yichang Zhang, Yu Wan, Yuqiong Liu, Zeyu Cui, Zhenru Zhang, and Zihan Qiu. 2024. Qwen2.5 technical report. CoRR, abs/2412.15115.  
+Zhilin Yang, Peng Qi, Saizheng Zhang, Yoshua Bengio, William W. Cohen, Ruslan Salakhutdinov, and Christopher D. Manning. 2018. Hotpotqa: A dataset for diverse, explainable multi-hop question answering. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, Brussels, Belgium, October 31 - November 4, 2018, pages 2369-2380. Association for Computational Linguistics.  
+Tian Yu, Shaolei Zhang, and Yang Feng. 2024. Autorag: Autonomous retrieval-augmented generation for large language models. CoRR, abs/2411.19443.  
+Dan Zhang, Sining Zhoubian, Ziniu Hu, Yisong Yue, Yuxiao Dong, and Jie Tang. 2024. Rest-mcts*: LLM self-training via process reward guided tree search. In Advances in Neural Information Processing Systems 38: Annual Conference on Neural Information Processing Systems 2024, NeurIPS 2024, Vancouver, BC, Canada, December 10 - 15, 2024.
+
+# A Supplemental Materials on Setup
+
+# A.1 Datasets
+
+HotpotQA (Yang et al., 2018) is a multi-hop reasoning dataset that includes two reasoning types: bridge and comparison. Bridge: sub-questions are linked by bridging entities. Comparison: compare two statements, including yes/no and partial order.
+
+2WikiMQA (Ho et al., 2020) is a multi-hop reasoning dataset that contains four reasoning types. The bridge (compositional) and comparison are the same as those in HotpotQA. Besides, it includes two additional reasoning types: bridge-comparison and inference. Bridge-comparison: compare two bridge sub-questions. Inference: sub-questions are connected through reasoning relationships rather than bridging entities. For example, the predicate grandfather is equal to a father's father.
+
+MuSiQue (Trivedi et al., 2022) $^{10}$  is a multi-hop reasoning dataset including 2-4 hops questions. All questions in MuSiQue are compositional, with the number of hops extended to a maximum of 4 hops. Besides, it includes multiple compositional questions, where a question contains more than one bridge entity. Due to the complexity of the questions, MuSiQue has the highest level of difficulty.
+
+# A.2 Models
+
+We use DeepSeek-V2.5 (DeepSeek-AI et al., 2024a) $^{11}$  as the advanced LLM to synthesis reasoning and critic rationales. In the experiments, we use Mistral-7B (Jiang et al., 2023a) $^{12}$  as the student backbone model by default. Besides, we also conduct experiments on LLaMA2-7B (Touvron et al., 2023) $^{13}$  and Qwen2.5-7B (Yang et al., 2024) $^{14}$  to demonstrate the generalization of our method.
+
+# A.3 Training Corpus
+
+We randomly sample subsets of the training data from 2WikiMQA, HotpotQA, and MuSiQue to construct the training corpus. Specially, we sample 20,000, 10,000, 19,000 instances from the respective datasets to construct the initial generator train
+
+ing corpus  $X$ , while an additional 3,000, 2,000, and 1,000 non-overlapping instances are sampled as the critic model training corpus  $X_{\text{critic}}$ .
+
+It is noteworthy that, to demonstrate the generality of our approach, we train a unified model on a mixture of all datasets, enabling it to perform reasoning across diverse multi-hop QA datasets instead of dataset-specific training.
+
+# A.4 Retrieval
+
+Following prior work (Trivedi et al., 2023), we construct a retrieval corpus using supporting facts and irrelevant documents from the training instances, and conduct retrieval within this scope.
+
+By default, we employ BM25 (Robertson and Zaragoza, 2009) provided by Elastic Search $^{15}$  for sparse retrieval in our experiments. Additionally, we incorporate dense retrieval and hybrid retrieval in our analysis. For dense retrieval, we use Contriever (Izacard et al., 2022) $^{16}$ . For hybrid retrieval, we combine the results from both sparse and dense retrieval by taking their union. Since our method independently processes each retrieved document and selects relevant ones based on self-evaluation, it inherently supports hybrid retrieval, thereby delivering better retrieval performance.
+
+# A.5 Baselines
+
+Chain-of-Thought (Wei et al., 2022) generates step-by-step reasoning before the final answer. We use 20-shot demonstrations.
+
+RAG (Lewis et al., 2020) retrieves k documents and conducts chain-of-thought reasoning conditional on the retrieval contents. We use BM25 to retrieve 5 documents, and 3-shot demonstrations.
+
+IRCoT (Trivedi et al., 2023) interleaves retrieval and reasoning until the model arrives at the final answer. We set the number of documents per iteration to 4, with a maximum of 4 iterations.
+
+Self-Ask (Press et al., 2023) iteratively breaks down complex questions and answers the resulting sub-questions one by one. We use 4-shot prompts.
+
+ProbTree (Cao et al., 2023) decomposes the question into a tree and uses bottom-up aggregation reasoning based on log-prob to get the final answer.
+
+BeamAggR (Chu et al., 2024a) parses the complex question into a structured decomposition tree. It then performs bottom-up beam aggregation reasoning to select the promising reasoning trajectory.
+
+Self-RAG (Asai et al., 2024) enables dynamic retrieval and generation through self-reflective tokens. We set the max retrieval documents per round to 5, and the max reasoning depth to 7.
+
+RAG-Star (Jiang et al., 2024) uses Monte Carlo Tree Search to guide LLM-based tree-like reasoning and employs retrieval-augmented verification to consolidate internal and external knowledge.
+
+GenGround (Shi et al., 2024) alternates between generating single-hop question-answer pairs using LLMs and grounding them in retrieved documents to correct potential knowledge errors.
+
+Auto-RAG (Yu et al., 2024) leverages LLMs' reasoning and decision-making capabilities to plan retrievals, query rewriting, and gather external retrieval knowledge through multi-turn dialogues.
+
+DR-Distillation (Li et al., 2024) distills two small models (Decomposer and Responser) to interactively solve knowledge-intensive multi-hop questions by breaking them into simpler sub-questions and conducting retrieval-augmented reasoning.
+
+# A.6 Task Definition
+
+Given a multi-hop QA pair  $(Q, a)$  that requires multiple reasoning steps  $r_i$  to solve, each reasoning step  $r_i$  is based on a specific piece of supporting evidence  $d_i$ . We use the LLM to iteratively decompose the question  $q_i$ , employ an external retriever Retr to retrieve relevant documents based on the subquestion, and conduct retrieval-augmented reasoning conditional on the retrieved documents. Once the reasoning is finished, the answer in the last iteration is regarded as the final answer  $a^*$ .
+
+$$
+Q \rightarrow r _ {1} \rightarrow r _ {2} \dots \rightarrow a ^ {*} \tag {11}
+$$
+
+# A.7 Implementation Details
+
+The training and inference are conducted on 4 and 1 NVIDIA Tesla A100-80G GPUs, respectively. The training process is initiated from the base model, and a chat template structured as '[INST] {Instruction} [/INST] {Output}' is manually added. During the training process, we exclusively train the Output while masking the retrieved document portions to prevent them from being included in the
+
+<table><tr><td>Hyperparameters</td><td>Values</td></tr><tr><td colspan="2">Data Synthesis Hyperparameters</td></tr><tr><td>Temperature</td><td>1.15</td></tr><tr><td>N</td><td>3</td></tr><tr><td># Raw Generator Corpus</td><td>49000</td></tr><tr><td># Raw Critic Corpus</td><td>6000</td></tr><tr><td colspan="2">Training Hyperparameters</td></tr><tr><td>Epochs</td><td>2</td></tr><tr><td>Learning Rate</td><td>5e-5</td></tr><tr><td>Batch Size</td><td>128</td></tr><tr><td>Warmup Ratio</td><td>0.03</td></tr><tr><td>LR Decay</td><td>Cosine Annealing</td></tr><tr><td>LoRA Rank</td><td>128</td></tr><tr><td>LoRA Alpha</td><td>256</td></tr><tr><td>LoRA Targets</td><td>All Linear Layers</td></tr><tr><td colspan="2">Inference Hyperparameters</td></tr><tr><td>Temperature</td><td>1.15</td></tr><tr><td>Repetition Penalty</td><td>1.2</td></tr><tr><td># Docs</td><td>5 or 6 (hybrid retrieval)</td></tr><tr><td># Decomposition Branches</td><td>2</td></tr><tr><td># Candidates</td><td>2</td></tr><tr><td># Retrieval Reward</td><td>1.0 / 0.5 / -1.0</td></tr><tr><td># Reasoning Reward</td><td>0.5 / 0.25 / -0.5</td></tr></table>
+
+Table 9: Hyperparameters in experiments.
+
+optimization. We use vLLM $^{17}$  in the inference. Hyperparameters are provided in Table 9.
+
+B Supplemental Materials on Analysis  
+
+<table><tr><td>Methods</td><td>2WikiMQA</td><td>HotpotQA</td><td>MuSiQue</td></tr><tr><td>Self-Ask</td><td>2.72</td><td>2.28</td><td>2.48</td></tr><tr><td>IRCoT</td><td>3.78</td><td>2.86</td><td>2.44</td></tr><tr><td>Self-RAG</td><td>2.05</td><td>2.12</td><td>2.13</td></tr><tr><td>Auto-RAG</td><td>2.67</td><td>2.39</td><td>3.74</td></tr><tr><td>DR-Distillation</td><td>3.33</td><td>2.85</td><td>3.07</td></tr><tr><td>SiGIR</td><td>2.40</td><td>2.01</td><td>2.37</td></tr></table>
+
+Table 10: The average number of iterations of iterative reasoning methods to solve a multi-hop question. Our method requires fewer iterations to solve a multi-hop question, resulting in higher reasoning efficiency.
+
+<table><tr><td>Question Type</td><td>Bdg.</td><td>Infer.</td><td>Cmp.</td><td>Bdg-Cmp.</td></tr><tr><td>Latency(s)</td><td>2.41</td><td>2.56</td><td>2.91</td><td>5.56</td></tr><tr><td>Prompt</td><td>2989.4</td><td>2775.0</td><td>1899.1</td><td>4653.9</td></tr><tr><td>Completion</td><td>171.5</td><td>202.4</td><td>209.3</td><td>375.2</td></tr></table>
+
+Table 11: Per-sample inference latency and token consumption of greedy inference on 2WikiMQA dataset.
+
+<table><tr><td>Question Type</td><td>Bridge</td><td>Comparison</td></tr><tr><td>Latency(s)</td><td>3.71</td><td>2.81</td></tr><tr><td>Prompt</td><td>2156.6</td><td>1786.6</td></tr><tr><td>Completion</td><td>182.7</td><td>220.3</td></tr></table>
+
+Table 12: Per-sample inference latency and token consumption of greedy inference on HotpotQA dataset.
+
+# B.1 Reasoning Efficiency between Separated Critic and Generative Critic
+
+As mentioned in §6.4, GenCritic achieves a higher throughput compared to Separated Critic. We will describe how to calculate the throughput in this section. We use the same computing platform, equipped with one A100-SXM4-80GB GPU, a Xeon(R) 6348 CPU with 28 cores and 56 threads, and 480GB of memory. We sample 500 instances from each of the three datasets, totaling 1,500 instances. We conduct parallel inference using the API server provided by vLLM, employing reward-guided greedy inference during benchmarking. The throughput is calculated based on the time consumed for different methods.
+
+# B.2 Iterative Reasoning Efficiency
+
+We discuss the iterative reasoning efficiency in §6.6. Specifically, we consider a complete round of retrieval-generation as one reasoning iteration. It is
+
+<table><tr><td>Question Type</td><td>2hop</td><td>3hop</td><td>4hop</td></tr><tr><td>Latency(s)</td><td>2.43</td><td>3.41</td><td>4.67</td></tr><tr><td>Prompt</td><td>2388.9</td><td>4162.6</td><td>4294.4</td></tr><tr><td>Completion</td><td>182.1</td><td>258.2</td><td>333.8</td></tr></table>
+
+Table 13: Per-sample inference latency and token consumption of greedy inference on MuSiQue dataset.
+
+noteworthy that multiple retrieval calls can occur within a single reasoning iteration. The specific number of reasoning iterations corresponding to each method can be found in Table 10. Additionally, we present the inference latency of individual samples with greedy inference, along with the average prompt and completion token costs, as shown in Table 11, Table 12, and Table 13.
+
+<table><tr><td>Notation</td><td>Description</td></tr><tr><td colspan="2">(a) Modules</td></tr><tr><td>LLM</td><td>The advanced large language model to synthesis iterative reasoning rationales and process supervision.</td></tr><tr><td>R</td><td>The iterative reasoner model.</td></tr><tr><td>C</td><td>The separated critic model.</td></tr><tr><td>Rsc</td><td>The iterative reasoner model with self-critique capabilities.</td></tr><tr><td>Retr</td><td>External retriever (sparse, dense, hybrid)</td></tr><tr><td colspan="2">(b) Training Corpus</td></tr><tr><td>X</td><td>Raw training corpus for generator.</td></tr><tr><td>Xcritic</td><td>Raw training corpus for critic, which is disjoint with X.</td></tr><tr><td colspan="2">(c) Symbols</td></tr><tr><td>qi(r)</td><td>Multi-hop question after reduction.</td></tr><tr><td>qi</td><td>Decomposed atomic sub-question.</td></tr><tr><td>di</td><td>The retrieved document.</td></tr><tr><td>x</td><td>Reasoning trajectory.</td></tr><tr><td>si</td><td>Sub-question reasoning.</td></tr><tr><td>ri retr</td><td>Reward for retrieval relevance.</td></tr><tr><td>ri reas</td><td>Reward for reasoning utility.</td></tr><tr><td>rci</td><td>Cumulative process reward.</td></tr><tr><td>roi</td><td>Outcome reward.</td></tr></table>
+
+Table 14: The formal notations mentioned in Section 3.
+
+# C Prompts
+
+We provide the manually annotated prompts for data synthesis. For iterative reasoning rationales, we use 1-shot prompts for each question type (Figure 5). For critic on retrieval and reasoning, we use 3-shot prompts (Figure 6, 7). For critic on outcome reward, we use zero-shot prompts (Figure 8).
+
+Additionally, we provide an example of iterative reasoning with self-critique, shown in Figure 15.
+
+# Algorithm 1 Self-Critique Guided Iterative Reasoning
+
+Require: Initial question  $q_0$ , retrieval model Retr, probabilistic language model  $p$ , candidate size  $k$
+
+Ensure: Final answer  $a^*$
+
+# Initialize:
+
+1:  $\mathbf{x}_0 \gets [q_0]$
+
+$\triangleright$  Initial reasoning prefix
+
+2:  $r_0^{\mathrm{c}}\gets 0$
+
+$\triangleright$  Initial c reward
+
+3:  $\pmb{x}_0\gets \left\{(\pmb{x}_0,\pmb{r}_0^{\mathrm{c}})\right\}$
+
+$\triangleright$  Set of candidate trajectories
+
+# Main Loop:
+
+4: for  $t = 0$  to  $T - 1$  do
+
+$\triangleright T$  is the maximum number of iterations
+
+5:  $\pmb{x}_{t + 1} \gets \emptyset$
+
+$\triangleright$  Initialize set for next iteration
+
+# Branch Exploration:
+
+6: for  $(\pmb{x}_t, \pmb{r}_t^{\mathrm{c}})$  in  $\pmb{x}_t$  do  
+7: Decomposition:  
+8:  $q_{t + 1}\sim p(y\mid q_t^{(r)},\pmb {x}_t)$  
+9: Retrieval:R  
+10:  $d_{i}\gets \mathrm{Retr}(q_{t + 1})$  
+11:  $r_{t + 1}^{\mathrm{retr}}\gets p(y\mid q_{t + 1},d_i)$  
+12: Reasoning:  
+13:  $s_{t + 1}\sim p(y\mid q_{t + 1},d_i,\pmb {x}_t)$  
+14:  $r_{t + 1}^{\mathrm{reas}}\gets p(y\mid q_{t + 1},d_i,s_{t + 1})$  
+15: Update Reasoning Prefix:  
+16:  $\pmb{x}_{t + 1}\gets \pmb {x}_t\oplus [q_{t + 1},d_i,r_{t + 1}^{\mathrm{retr}},s_{t + 1},r_{t + 1}^{\mathrm{reas}}]$  
+17: Update Cumulative Reward:  
+18:  $\pmb{r}_{t + 1}^{\mathrm{c}}\gets \pmb{r}_t^{\mathrm{c}} + r_{t + 1}^{\mathrm{retr}} + r_{t + 1}^{\mathrm{reas}}$  
+19: Question Reduction:  
+20:  $q_{t}^{(r)}\gets \mathrm{ReduceQuestion}(\pmb{x}_{t})$  
+21:  $\pmb{x}_{t + 1}\gets \pmb{x}_{t + 1}\cup \{(\pmb{x}_{t + 1},\pmb{r}_{t + 1}^{\mathrm{c}})\}$  
+22: end for
+
+# Candidate Selection:
+
+23:  $\pmb{x}_{t+1} \gets \operatorname{Top-k}(\pmb{x}_{t+1}, \pmb{r}_{t+1}^{\mathrm{c}})$
+
+$\triangleright$  Sample reasoning step
+
+24: end for
+
+$\triangleright$  Compute reasoning reward
+
+# Final Answer Selection:
+
+25:  $a^* \gets \arg \max_a \left( \sum_{i=1}^n (r_i^{\mathrm{retr}} + r_i^{\mathrm{reas}}) \right)$
+
+> Concatenate current status
+
+26: return  $a^*$
+
+> Select top-k candidates
+
+$\triangleright$  Select answer with highest c reward
+
+# Algorithm 2 Training SC-Reasoner
+
+Require: Generator corpus  $X = \{x_{i}\}$ , where  $x_{i} = (q_{i},a_{i},\{d_{1},\ldots ,d_{n}\})$
+
+Require: Non-overlapping critic corpus  $X_{\text{critic}}$
+
+Require: Large Language Model (LLM)
+
+Ensure: Self-Critique Reasoner  $R_{sc}$
+
+# Phase 1: Train Iterative Reasoner  $R$
+
+1:  $X_{\mathrm{synthetic}} \gets \emptyset$  
+$\triangleright$  Initialize synthetic reasoning corpus  
+2: for each instance  $x_{i} \in X$  do  
+3: Step 1.1: Synthesize Reasoning Rationales  
+4:  $\operatorname{decomp}_i, \text{reasoning}_i \gets \operatorname{LLM}(q_i, \{d_1, \ldots, d_n\})$ $\triangleright$  Few-shot prompting for decomposition and reasoning  
+5: trajectory  $_{i} \gets$  Organize(decomp  $_{i}$ ,reasoning  $_{i}$ )  
+6:  $X_{\mathrm{synthetic}} \gets X_{\mathrm{synthetic}} \cup \{\mathrm{trajectory}_i\}$  
+7: end for  
+8: Step 1.2: Train Iterative Reasoner  
+9:  $R\gets \mathrm{TrainModel}(X_{\mathrm{synthetic}})$ $\triangleright$  Train  $R$  on synthetic corpus  
+0:
+
+# Phase 2: Train Critic Model  $C$
+
+11:  $X_{\mathrm{critic-data}} \gets \emptyset$  
+$\triangleright$  Initialize critic training data  
+12: for each instance  $x_{i} \in X_{\text{critic}}$  do  
+13: Step 2.1: Generate Reasoning Trajectories  
+14: trajectory  $_i \gets R(q_i, \{d_1, \ldots, d_n\})$  Use  $R$  to generate reasoning trajectories  
+15: Step 2.2: Evaluate Rewards with LLM  
+16:  $r_i^{\mathrm{retr}}, r_i^{\mathrm{reas}}, r_i^{\mathrm{overall}} \gets \mathrm{LLM}(\mathrm{trajectory}_i)$  Evaluate retrieval, reasoning, and overall quality  
+17:  $X_{\mathrm{critic-data}} \gets X_{\mathrm{critic-data}} \cup \{(\mathrm{trajectory}_i, r_i^{\mathrm{retr}}, r_i^{\mathrm{reas}}, r_i^{\mathrm{overall}})\}$  
+18: end for  
+19: Step 2.3: Train Critic Model  
+20:  $C\gets \mathrm{TrainModel}(X_{\mathrm{critic - data}})$ $\triangleright$  Train  $C$  on critic data  
+21:
+
+# Phase 3: Train Self-Critique Reasoner  $R_{sc}$
+
+22:  $X_{\mathrm{sc - data}}\gets \emptyset$  Initialize self-critique training data  
+23: for each instance  $x_{i} \in X$  do  
+24: Step 3.1: Annotate Reasoning Trajectories with Critic  
+25: trajectory  $\leftarrow R(q_{i},\{d_{1},\ldots ,d_{n}\})$  Generate reasoning trajectories with  $R$  
+26:  $r_i^{\mathrm{retr}}, r_i^{\mathrm{reas}}, r_i^{\mathrm{overall}} \gets C(\mathrm{trajectory}_i)$ $\triangleright$  Annotate with critic  $C$  
+27: trajectory $_i^{annotated}$  ← AddSpecialTokens(projected,  $r_i^{\mathrm{retr}}$ ,  $r_i^{\mathrm{reas}}$ ,  $r_i^{\mathrm{overall}}$ )  
+28:  $X_{\mathrm{sc-data}} \gets X_{\mathrm{sc-data}} \cup \{\text{trajectory}_i^{\text{annotated}}\}$  
+29: end for  
+30: Step 3.2: Train Self-Critique Reasoner  
+31:  $R_{sc} \gets \text{TrainModel}(X_{sc-data})$ $\triangleright$  Train  $R_{sc}$  on annotated corpus  
+32:
+
+# Self-Improvement Pipeline
+
+33: while not converged do  
+34: Step 4.1: Generate Synthetic Rationales  
+35:  $X_{\mathrm{new}} \gets R_{sc}(\mathrm{unlabeled~questions})$  Generate rationales for unlabeled questions  
+36: Step 4.2: Evaluate Rationale Quality  
+37:  $X_{\mathrm{high - quality}}\gets \{x\in X_{\mathrm{new}}\mid C(x) > \mathrm{threshold}\}$ $\triangleright$  Filter high-quality rationales  
+38: Step 4.3: Retrain  $R_{sc}$  
+39:  $R_{sc} \gets \text{TrainModel}(R_{sc}, X_{\text{high-quality}})$ $\triangleright$  Retrain with high-quality data  
+40: end while  
+41: return  $R_{sc}$
+
+<table><tr><td colspan="2">Iterative Reasoning Examples of Self-Critique Reasoner</td></tr><tr><td>Q</td><td>How many counties are there in the largest state in the region where the film The Family Stone was set? (Answer: sixteen)</td></tr><tr><td rowspan="3">A</td><td>[Non-Atomic Question]In which region is the film &quot;The Family Stone&quot; set?&lt;paragraph&gt;The Family Stone ...&lt;/paragraph&gt;Retrieval Quality: [Partially Relevant]Reasoning: From Document, we know that &quot;The Family Stone&quot; is set in a small New England town. Therefore, the region is New England.&lt;sub&gt;/sub-question&gt;Reasoning Quality: [Fully supported]Remaining Question] How many counties are there in the largest state in New England?</td></tr><tr><td>[Non-Atomic Question]&lt;sub&gt;&lt;sub&gt;sub-question&lt;/sub&gt; Which state is the largest in New England?&lt;paragraph&gt;New England ...&lt;/paragraph&gt;Retrieval Quality: [Partially Relevant]Reasoning: From Document, we know that Maine is the largest state in New England, constituting nearly one-half of the total area of New England.&lt;sub&gt;/sub-question&gt;Reasoning Quality: [Fully supported]Remaining Question] How many counties are there in Maine?</td></tr><tr><td>[Atomic Question]&lt;paragraph List of counties in Maine ...&lt;/paragraph&gt;Retrieval Quality: [Relevant]Reasoning: From Document, we know that there are sixteen counties in Maine.Recasoning Quality: [Fully supported][Final Answer] @@sixteen@@Overall Reasoning Quality: [4]</td></tr></table>
+
+Table 15: A reasoning example of self-critique reasoner. The data instance is from MuSiQue dataset.
+
+# Prompts for Iterative Reasoning Rationales
+
+Please answer the following multi-hop questions with explanation based on the given documents. Wrap the final answer with @@@. Here is an example:
+
+Document #1: Cuidado con las imitaciones
+
+Cuidado con las imitaciones is a 1948 Argentine film directed by Luis Bayón Herrera.
+
+Document #2: Luis Bayón Herrera
+
+Luis Bayón Herrera (23 September 1889 – 30 March 1956) was a Spanish film director and screenwriter who worked in Argentine film of the 1940s and 1950s. He was one of the most important directors of the golden age of Argentine cinema: Herrera was born in Bilbao, Spain. He directed some 40 different Argentine feature films and in the last few years of his career in the early 1950s he worked on Cuban production with films such as Å La Habana me voyin 1951. He died in Buenos Aires in 1956, aged 66".
+
+Question: What nationality is the director of film Cuido Con Las Imitaciones?
+
+To answer this multi-hop question, we need to solve the following sub-questions:
+
+Decompose the multi-hop question into sub-questions  
+1. Who is the director of the film Cuido Con Las Imitaciones?  
+2. What is the nationality of that director?
+
+Next, let's solve the sub-questions one by one.
+
+Sub-question: Who is the director of the film Cuido Con Las Imitaciones?
+
+From Document #1, we know that "Cuidado Con Las Imitaciones" is a Spanish film, and its director is José María Forqué.
+
+Remaining Question: What is the nationality of José María Forqué?
+
+From Document #2, we know that José María Forqué was a Spanish director, which means his nationality is Spanish.
+
+Final Answer
+
+@@Spanish@@
+
+$\{\mathrm{SUPPORTING~DOCUMENTS}\}$
+
+Question: {MULTI-HOP QUESTION}
+
+Figure 5: The fewshot prompts for iterative reasoning rationales synthesis.
+
+# Prompts for Retrieval Relevance Evaluation
+
+You will be given a question and an evidence document. Your task is to evaluate whether the evidence is relevant to the question and provide a clear explanation for your decision. Use the following scale to rate the relevance of the evidence:
+
+[Relevant]: The evidence directly addresses or supports the question.
+
+[Partially Relevant]: The evidence is somewhat related to the question but does not fully address it. For example, it may mention the core entity or topic of the question.
+
+[Irrelevant]: The evidence does not relate to the question in any meaningful way.
+
+Question: In what country was the first Pan-African conference held?
+
+Evidence: First Pan-African Conference
+
+The First Pan-African Conference was held in London from 23 to 25 July 1900 (just prior to the Paris Exhibition of 1900 in order to allow tourists of African descent to attend both events). Organized primarily by the Trinidadian barrister Henry Sylvester Williams, it took place in Westminster Town Hall (now Caxton Hall) and was attended by 37 delegates and about 10 other participants and observers from Africa, the West Indies, the US and the UK, including Samuel Coleridge Taylor (the youngest delegate), John Alcindor, Dadabhai Naoroji, John Archer, Henry Francis Downing, and W.E.B. Du Bois, with Bishop Alexander Walters of the AME Zion Church taking the chair.
+
+Rating: [Relevant]
+
+Explanation: The evidence directly answers the question by stating that the first Pan-African conference was held in London, which is in the United Kingdom. The details provided, such as the dates and location of the conference, further confirm the relevance of the evidence to the question.
+
+Question: What is the administrative territorial entity for Malta Township?
+
+Evidence: Minsk Region
+
+Minsk Region or Minsk Voblasc or Minsk Oblast (,"Minskaja voblasc";, "Minskaja oblastj") is one of the regions of Belarus. Its administrative center is Minsk, although it is a separate administrative territorial entity of Belarus. As of 2011, the region's population is 1,411,500.
+
+Rating: [Irrelevant]
+
+Explanation: The evidence provided discusses the Minsk Region, which is a region in Belarus. The question, however, asks about the administrative territorial entity for Malta Township. Since the evidence pertains to Belarus and not Malta, it is not relevant to the question.
+
+## Question: Did Bob Rosa produce any work for Patrick Love?
+
+Evidence: Bob Rosa
+
+Bob Rosa is an American record producer that has worked with artists such as Madonna, Janet Jackson, Mariah Carey and Espen Lind. One if his first recording projects was working on the track "Planet Rock" by Afrika Bambaataa & the Soulsonic Force in 1982.
+
+Rating: [Partially Relevant]
+
+Explanation: The evidence mentions Bob Rosa's work with various artists, including Madonna, Janet Jackson, Mariah Carey, and Espen Lind. However, it does not specifically mention Patrick Love or any work Bob Rosa may have done for him. Therefore, the evidence is partially relevant to the question.
+
+Question: {}
+
+Evidence: {}
+
+Figure 6: The fewshot prompts for retrieval quality critic.
+
+# Prompts for Reasoning Utility Evaluation
+
+You will be provided with a question, an evidence document, along with a response. Your job is to determine whether the response is supported by the evidence, and provide explanation for your decision. Use the following scale to rate the response:
+
+[Fully supported] Most of the information in the response is supported by the evidence.
+
+[Partially supported] Some of the information in the response is supported by the evidence, but there are some parts that are speculative.
+
+[Not supported] The response is not supported by the evidence / The response does not provide useful reasoning.
+
+Question: In which country is the Financial Conduct Authority located?
+
+Evidence: Martin Wheatley
+
+Martin Wheatley is a British financier, formerly managing director of the Consumer and Markets Business Unit of the Financial Services Authority in the UK, and is the former CEO of the Financial Conduct.
+
+\#\*\* Response: From Document, we know that Martin Wheatley is a British financier who formerly managed the Consumer and Markets Business Unit of the Financial Services Authority in the UK. This indicates that the Financial Conduct Authority is located in the United Kingdom.
+
+Rating: [Fully supported]
+
+Explanation: The response is fully supported by the evidence. The evidence clearly states that Martin Wheatley is a British financier who formerly managed the Consumer and Markets Business Unit of the Financial Services Authority in the UK and was the former CEO of the Financial Conduct Authority. The response accurately extracts this information and concludes that the Financial Conduct Authority is located in the United Kingdom, which is directly supported by the evidence provided.
+
+Question: What is the birthplace of Jane Siberry?
+
+Evidence: Gavin Bradley
+
+Gavin Bradley is an award-winning Canadian singer-songwriter and producer based in Toronto who has worked with artists like Nelly Furtado, Tori Amos and Jane Siberry. Fusing acoustic and electronic elements, his work is identifiable for its signature "warm" piano sound and live strings mixed with filtered synthesizers and other electronic manipulations. Besides production, Bradley is a solo recording artist. His debut album 'Deep Freeze' was released on UMI Records in 2006.
+
+\#\*\* Response: The response is partially supported because the evidence mentions Gavin Bradley's Toronto connection but does not directly state Jane Siberry's birthplace. The inference about her possible connection to Toronto is speculative.
+
+Rating: [Partially supported]
+
+Explanation: The response correctly identifies that the evidence mentions Gavin Bradley's connection to Toronto but does not directly state Jane Siberry's birthplace. The inference that Jane Siberry might be connected to Toronto based on her collaboration with Gavin Bradley is speculative, as the evidence does not provide explicit information about her birthplace. Therefore, the response is partially supported.
+
+Question: {}
+
+Evidence: {}
+
+\*\*\* Response: \{\}
+
+Figure 7: The fewshot prompts for reasoning utility critic.
+
+# Prompts for Outcome Reward Evaluation
+
+You will be provided with a question, along with a reasoning trajectory.
+
+Your job is to determine whether the reasoning trajectory is useful for answering the question, and provide explanation for your decision.
+
+Use the following scale to rate the reasoning:
+
+[5]: The reasoning process is clear, logically structured, and well-supported by the evidence.  
+[4]: The reasoning process is mostly clear, partially logically structured, and supported by evidence, but may contain minor logical flaws.  
+[3]: The reasoning is somewhat unclear with noticeable flaws and uses limited or weak evidence.  
+[2]: The reasoning is flawed, lacking supporting evidence, which results in an incorrect conclusion.  
+[1]: The reasoning fails to provide a definitive answer.
+
+Question: {}
+
+Reasoning: {}
+
+Rating:
+
+Explanation:
+
+Figure 8: The prompts for reasoning trajectories outcome reward evaluation.

2025/Self-Critique Guided Iterative Reasoning for Multi-hop Question Answering/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5078f14e4daf6d0d64ab45061a4626e187a1e2a20bade1d8c184ae75f6a9b468
+size 1003024

2025/Self-Foveate_ Enhancing Diversity and Difficulty of Synthesized Instructions from Unsupervised Text via Multi-Level Foveation/e85236e0-99f2-4fdd-a4f5-ddd5a090bab5_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6078d4e1d42e0b56533e8a281a2ca03f11abf4ffb0d45929f94be9aea0521b8d
+size 1444906

2025/Self-Foveate_ Enhancing Diversity and Difficulty of Synthesized Instructions from Unsupervised Text via Multi-Level Foveation/full.md

@@ -0,0 +1,460 @@
+# SELF-FOVEATE: Enhancing Diversity and Difficulty of Synthesized Instructions from Unsupervised Text via Multi-Level Foveation
+
+Mingzhe Li, Xin Lu, Yanyan Zhao*
+
+Research Center for Social Computing and Interactive Robotics
+
+Harbin Institute of Technology, China
+
+{mzli, xlu, yzhao}@ir.hit.edu.cn
+
+# Abstract
+
+Large language models (LLMs) with instruction following capabilities have demonstrated impressive problem-solving abilities. While synthesizing instructional data from unsupervised text has become a common approach for training such models, conventional methods rely heavily on human effort for data annotation. Although existing automated synthesis paradigms have alleviated this constraint, they still exhibit significant limitations in ensuring adequate diversity and difficulty of synthesized instructions. To address these challenges, we propose SELF-FOVEATE, an innovative LLM-driven method for instruction synthesis. This approach introduces a "Micro-Scatter-Macro" multi-level foveation methodology that effectively guides the LLM to deeply excavate fine-grained information embedded in unsupervised text, thereby enhancing both the diversity and difficulty of synthesized instructions. Comprehensive experiments across multiple unsupervised corpora and diverse model architectures validate the effectiveness and superiority of our proposed method. We publicly release our data and codes: https://github.com/Mubuky/Self-Foveate
+
+# 1 Introduction
+
+Large language models (LLMs), such as GPT-4o (OpenAI, 2024), Claude 3.5 Sonnet (Anthropic, 2024), and Llama 3.1 (Meta, 2024), have garnered significant attention due to their exceptional instruction-following capabilities (Zhou et al., 2023b), with their continuously enhanced problem-solving abilities (Cobbe et al., 2021; Dua et al., 2019) being increasingly recognized. A critical component in training such models typically involves fine-tuning with extensive supervised question-answering instruction data (Wang et al., 2023b). However, substantial challenges persist
+
+![](images/654f93506624cafc02550f309c7cf635284031f23dcd741a7fa206473a8be3e5.jpg)  
+(a) Self-Foveate
+
+![](images/ae205a5c66668a10326947b0c19aa9ab51107823bdbfb35111d4fbb51a8c07c6.jpg)
+
+![](images/5543567d2e4051bc18c5d50281975e96b3159ec9af1fedad1ca3ac2e6e3844d6.jpg)  
+Instruction Evaluation: Difficult and Diverse
+
+![](images/99531ae521da4c6f01e9e82327c3c7d3a42429490d9c5a0b0331027ebfd9ac79.jpg)  
+Instruction Evaluation: Simple and Monotonous  
+Figure 1: Illustration of (a) SELF-FOVEATE in contrast with (b) Baseline Self-QA. For SELF-FOVEATE, the Multi-Level Foveation enables the LLM to extract details (highlighted in distinct colors) of the text, subsequently synthesizing instructions with diversity and difficulty via distinct synthesis paradigms. In comparison, Self-QA employs single-step generation that produces instruction candidates with simplicity and monotony.
+
+in constructing large-scale, high-quality supervised fine-tuning (SFT) data for instruction tuning (Zhang et al., 2024). The prohibitive costs associated with human annotation (Wang et al., 2023a;
+
+<table><tr><td>Related Work</td><td>Data Utilization</td><td>w/o Human Annotation</td><td>Correctness Guarantee</td><td>Diversity Augment</td><td>Difficulty Augment</td></tr><tr><td>Self-Instruct (Wang et al., 2023a)</td><td>Seed QA examples</td><td>×</td><td>×</td><td>✔</td><td>×</td></tr><tr><td>Self-Align (Sun et al., 2023)</td><td>Seed QA examples</td><td>×</td><td>×</td><td>✔</td><td>✔</td></tr><tr><td>Self-Chat (Xu et al., 2023)</td><td>Dialogue</td><td>✔</td><td>×</td><td>×</td><td>×</td></tr><tr><td>Self-QA (Zhang and Yang, 2023)</td><td>Unsupervised Knowledge</td><td>✔</td><td>✔</td><td>✔</td><td>×</td></tr><tr><td>LongForm (Köksal et al., 2024)</td><td>Web dataset</td><td>✔</td><td>✔</td><td>✔</td><td>×</td></tr><tr><td>Humpback (Li et al., 2024b)</td><td>Web dataset</td><td>×</td><td>✔</td><td>×</td><td>✔</td></tr><tr><td>ISARA (Guo et al., 2024)</td><td>Seed QA examples</td><td>✔</td><td>✔</td><td>✔</td><td>×</td></tr><tr><td>Wiki2023 (Jiang et al., 2024)</td><td>Unsupervised Text</td><td>✔</td><td>✔</td><td>×</td><td>×</td></tr><tr><td>SELF-FOVEATE (Ours)</td><td>Unsupervised Text</td><td>✔</td><td>✔</td><td>✔</td><td>✔</td></tr></table>
+
+Table 1: A comparative analysis of various task generation methodologies or frameworks. The gray checkmark symbol denotes that the work may partially accomplish specific objectives (though not comprehensively). The red cross marker indicates either the work's failure to achieve the stated objective or absence of explicit documentation regarding this goal.
+
+Sun et al., 2023; Xu et al., 2023), coupled with difficulties in ensuring data diversity and quality control (Ge et al., 2024), continue to impede technological advancements. Given the demonstrated excellence and ongoing improvements in LLMs' instruction-following and generative capabilities, researchers are actively exploring effective methodologies to leverage these models for synthetic data generation (Wang et al., 2023a; Zhang and Yang, 2023; Nayak et al., 2024b; Wu et al., 2024). The primary objectives are to produce high-quality, cost-efficient datasets that reduce reliance on expensive human annotation (Ling et al., 2024), while enhancing the diversity and difficulty of automatically synthesized instructions to improve the performance of fine-tuned models on downstream tasks.
+
+A promising recent paradigm in synthetic instruction synthesis is unsupervised text-based instruction synthesis. The advantages of this paradigm become particularly pronounced when handling massive unsupervised text corpora – a ubiquitous resource containing rich world knowledge and linguistic patterns. By leveraging LLMs' intrinsic capabilities in contextual understanding and logical reasoning, this paradigm eliminates the need for manual annotation while confining the scope of instruction synthesis to the given unsupervised textual materials. The pioneering work by Zhang and Yang (2023) has demonstrated the feasibility of extracting instructions from unsupervised textual data.
+
+Despite advancements in synthesizing instruction data from unsupervised text and the proposal of automated LLM-based methods like Self-QA (Zhang and Yang, 2023) to mitigate these issues
+
+- approaches that have significantly optimized the data generation pipeline while reducing human labor costs - as shown in Table 1, the following challenges remain unresolved: (1) Diversity of Instructions: While existing frameworks continuously refine data generation strategies and enhance post-synthesis filtering for instruction data, limitations persist in the diversity of synthesized instructions. Models trained on such synthetic datasets frequently exhibit insufficient generalization capabilities and may even suffer from performance degradation. (2) Difficulty of Instructions: Current synthesis methodologies generally lack emphasis on instruction complexity and depth. For instance, Self-QA (Zhang and Yang, 2023) directly acquires instructions through single-step generation without guaranteed difficulty levels (as illustrated in Figure 1). Synthesized instructions often demonstrate simplistic structures and inadequate comprehension of inter-entity relationships, failing to effectively stimulate models to produce nuanced responses. This deficiency becomes particularly pronounced when handling complex queries or generating contextually rich responses. The absence of high-difficulty instruction design constrains model performance in real-world applications requiring advanced cognitive engagement and problem-solving capabilities. This collective evidence indicates that unsupervised text data still harbors substantial untapped potential for instruction synthesis.
+
+To address these challenges, we attempt to leverage unsupervised text itself, observing that textual data inherently contains abundant detailed information encompassing entities (e.g., QuantumX proces
+
+sor), attributes, relations, and writing techniques – such as the metaphorical comparison of “speed” to “not just a luxury” in the statement “In the realm of technology, speed is not just a luxury”, which implicitly analogizes speed as a fundamental necessity through subtle analogy, thereby emphasizing its critical importance (as multi-color annotated in Figure 1). This information remains underutilized in existing methods like SELF-QA.
+
+Motivated by these observations, this paper proposes SELF-FOVEATE, a comprehensive LLM-based methodology designed to automatically synthesize instructions from unsupervised text. Diverging from prior research, as illustrated in Figure 1, SELF-FOVEATE introduces a "Micro-ScatterMacro" multi-level foveation methodology to comprehensively excavate detailed information from raw text, subsequently synthesizing instructions with enhanced diversity and difficulty through three synthesis paradigms. Furthermore, SELF-FOVEATE integrates a data regeneration module to improve the fidelity and quality of instructions to source text.
+
+To summarize, the key contributions of this paper are as follows:
+
+$\triangleright$  We focus on unsupervised text-based instruction synthesis tasks, revealing limitations in existing works regarding diversity and difficulty.  
+$\triangleright$  We propose SELF-FOVEATE, a method that synthesizes instructions with diversity and difficulty through LLMs based on unsupervised text.  
+We conducted extensive experiments to evaluate SELF-FOVEATE, covering diversity and difficulty analysis, downstream task capabilities, and data scale trend analysis. The results demonstrate SELF-FOVEATE's superiority in unsupervised text-based instruction synthesis and suggest promising directions for future research.
+
+# 2 Related Work
+
+Instruction Tuning Multitask instruction finetuning (Wei et al., 2022) of language models significantly enhances their ability to follow instructions and generalize to new unseen tasks (Sanh et al., 2022; Mishra et al., 2022; Chung et al., 2022; Longpre et al., 2023; Zhou et al., 2023a; Li et al., 2024b). In our work, we utilize data from unsupervised text synthesis to conduct instruction tuning, enabling the model to better adapt to specific domain tasks.
+
+Synthetic Data Generation LLMs have showcased remarkable capabilities in data synthesis
+
+(Long et al., 2024), facilitating the creation of extensive synthetic datasets for pretraining and finetuning, thereby progressively supplanting labor-intensive manual data scraping and selection (Liu et al., 2024), and mitigating the constraints that data imposes on model capability growth (Villalobos et al., 2024).
+
+Distinct from earlier approaches centered on traditional language models (Schick and Schütze, 2021), LLMs present enhanced potential for generating high-quality synthetic data across diverse applications, including online translation (Oh et al., 2023), named entity recognition (Xiao et al., 2023), benchmark creation (Wang et al., 2024; Wei et al., 2024), and data diversity enhancement (Dai et al., 2023; Chung et al., 2023; Hong et al., 2024).
+
+The concept of synthetic input-output pairs for instruction tuning advances by requiring that the data generated by LLMs be diverse, accurate, and difficult, often leveraging LLMs on a set of seed task demonstrations or user-provided unsupervised text to create new synthetic tasks (Wang et al., 2023a; Honovich et al., 2023; Zhang and Yang, 2023; Taori et al., 2023; Peng et al., 2023; Yuan et al., 2024; Li et al., 2024a).
+
+Our work advances the generation of synthetic input-output pairs by developing a paradigm that integrates multi-level foveation into the creation of instruction tuning datasets, forming a comprehensive framework without any human annotation.
+
+# 3 SELF-FOVEATE
+
+In this section, we will introduce the proposed SELF-FOVEATE, denoted as  $\mathcal{F}$ , a multi-level automated method for synthesizing instructions that leverages only unsupervised text without the need for human-annotated samples. SELF-FOVEATE consists of three levels and a re-synthesis module. Formally, consider an original unsupervised text set  $\mathcal{D}$ , the proposed method SELF-FOVEATE operates by transforming each element  $d_{i}$  from  $\mathcal{D}$  through a multi-level synthesis path  $\mathcal{F}_j$ . In practice, SELF-FOVEATE is an operation applied independently to each document  $d_{i}$  ( $d_{i} \in \mathcal{D}$ ). The final generated dataset,  $\mathcal{D}_{\mathrm{gen}}$ , is accumulated from the data subsets generated by all elements:
+
+$$
+\mathcal {D} _ {\text {g e n}} = \mathcal {F} (\mathcal {D}) = \sum_ {d _ {i} \in \mathcal {D}} \sum_ {\mathcal {F} _ {j} \in \mathcal {F}} \mathcal {F} _ {j} \left(d _ {i}\right) \tag {1}
+$$
+
+During the synthesis process, the objectives of SELF-FOVEATE emphasize maximizing the diversity and difficulty of the generated instructions.
+
+![](images/198e4b3e98567e37543bdfa0780fd75fcfa63c70035948eb24c035926ff8f7dd.jpg)  
+Figure 2: The SELF-FOVEATE workflow is designed for instruction synthesis based on unsupervised text. SELF-FOVEATE takes unsupervised text as input, extracts foveate elements, foveate groups, and foveate segments, then synthesizes instruction tuning data through these extracted details.
+
+To ensure optimal functionality, SELF-FOVEATE incorporates three levels and one module as shown in Figure 2:
+
+$\triangleright$  Micro-foveate Level: This level captures the essential content within unsupervised text and acquires instructions through reverse synthesis. It plays a crucial role in maintaining the focus of synthesized instructions on the significant information in the text.  
+$\triangleright$  Scatter-foveate Level: This level aims to combine key information scattered throughout the unsupervised text and uses direct synthesis to generate instructions. It stimulates LLMs to synthesize instructions that possess a profound understanding of the relationships between different contents in the text.  
+$\triangleright$  Macro-foveate Level: This level focuses on and transcribes larger-grained information within the text. It plays a key role in enhancing the instructions' deep comprehension of the overall information in the text.  
+$\triangleright$  Re-synthesis Module: This module is used to perform post-synthesis filtering on the generated instructions, excluding outliers in the instruction synthesis process and conducting reference-based re-synthesis for abnormally generated instructions to reduce information gaps in the instruction set and improve synthesis success rates.
+
+# 3.1 Micro-foveate Level
+
+General unsupervised text typically contains several primary entity types, a greater number of secondary entity types, and attributes of all these entity types. However, instruction sets directly synthesized by teacher LLMs may struggle to intentionally and comprehensively cover all the important content within such unsupervised text. This could lead to the synthesized instruction sets missing critical information from the original text, thereby resulting in weaker downstream model performance. To address this, we employ the micro-foveate mechanism and reverse synthesis to ensure that these key pieces of information are preserved in the synthesized instruction sets. Attention to a broader range of information enhances the diversity of the synthesized instruction sets, while the exploration of entity attributes also increases the difficulty.
+
+Micro-foveate mechanism This mechanism aims to guide the teacher LLM to synthesize instructions from a fine-grained perspective of the unsupervised text. We introduce the concept of "foveate elements", which broadly encompass all entities and their attributes within a given text. We guide the large model to extract more foveate elements than needed and then select a certain number of foveate elements with the highest cosine similarity between the text and the embedding of each
+
+foveate element.
+
+Reverse synthesis Based on the selected foveate elements, we employ the reverse synthesis method to generate instructions. Specifically, each foveate element is treated as a potential answer to an instruction, guiding the teacher LLM to synthesize instructions from the unsupervised text and then re-synthesize the answers. Even though the foveate elements can already serve as answers or parts of answers, to enhance the fluency, completeness, and accuracy of the answers, we choose to regenerate the answers in the reverse synthesis step rather than directly using the foveate elements as answers.
+
+# 3.2 Scatter-foveate Level
+
+Complete unsupervised texts often imply relationships between different entities or attributes within the text. However, instruction sets directly synthesized by teacher large language models (LLMs) may struggle to detect connections that require reasoning or deeper understanding in the text. This could result in the instruction sets synthesized by the teacher model missing deep-level information from the text, thereby affecting downstream model performance. To address this, we employ the scatter-foveate mechanism and direct synthesis method to ensure that these deep-level insights are reflected in the synthesized instruction sets. Attention to the deeper, less obvious implicit information enhances the diversity of the synthesized instruction sets, while insights into the relationships between entities or attributes increase the difficulty of the instructions.
+
+Scatter-foveate mechanism This mechanism aims to guide the teacher LLM to synthesize instructions from a deeper perspective of the unsupervised text. We extract foveate elements more broadly from the text and randomly combine them into a certain number of foveate groups based on an empirical distribution.
+
+Direct Synthesis Based on the formed foveate groups, we synthesize instructions using the direct synthesis method. The direct synthesis method treats each element in the foveate group as an indispensable part of the instruction to be synthesized, actively guiding the teacher LLM to consider the deep-level connections between different entities or attributes and solidifying such reasoning-based or deep-thinking connections in the synthesized instructions. After the instructions are synthesized,
+
+the answers are regenerated based on the text.
+
+# 3.3 Macro-foveate Level
+
+Unsupervised texts often contain content requiring focused attention, such as figurative devices and rhetorical exaggerations. Notably, teacher large language models (LLMs) may require additional prompting to specifically emphasize the critical information embedded within these writing techniques. To capture information more comprehensively from a global perspective, we employ the macro-foveate mechanism and transcription synthesis method to guide teacher LLMs in developing a profound understanding of unsupervised texts, ensuring effective extraction of key information conveyed through literary devices.
+
+Macro-foveate mechanism This mechanism aims to guide teacher LLMs in comprehending unsupervised texts from a holistic perspective. We specifically highlight and extract text segments employing writing techniques – including metaphor, hyperbole, contrastive foil, rhetorical questioning, and citation – which are extracted as foveate segments.
+
+Transcription synthesis Based on the identified writing technique segments, we employ transcription synthesis to convert each foveate segment into an instructional format. This process transforms declarative foveate segments into interrogative or imperative forms. Subsequently, corresponding answers are synthesized according to the content of the unsupervised text.
+
+# 3.4 Re-synthesis Module
+
+Due to limitations such as the capabilities of the teacher LLM, not all instructions in a directly synthesized instruction set may be answerable by the referenced unsupervised text. To ensure the synthesized instructions remain faithful to the referenced unsupervised text, we employ a re-synthesis module to repeatedly re-synthesize these failed instructions over multiple iterations, aiming to generate instructions that can be fully answered based on the referenced unsupervised text.
+
+Single Sample Reference-synthesis During resynthesis, we process only one failed instruction at a time and randomly sample a subset of successfully synthesized instructions as reference examples for the teacher LLM in each iteration. By providing the teacher LLM with different batches
+
+of reference samples during each re-synthesis attempt, we enhance the success rate of re-synthesis.
+
+Hyperparameter Configuration The outputs of large language models (LLMs) are significantly influenced by hyperparameters such as temperature, top-p, and frequency penalty. To improve resynthesis success rates, we meticulously calibrated these hyperparameters to define a high-creativity mode, enabling the teacher LLM to synthesize instructions more creatively based on reference samples.
+
+# 4 Experiment
+
+In this section, our experiments focus on three critical research questions: (Q1) How effective is SELF-FOVEATE in enhancing instruction diversity? (Q2) How effective is SELF-FOVEATE in enhancing instruction difficulty? (Q3) How effective is SELF-FOVEATE in improving the model's problem-solving capabilities during instruction fine-tuning?
+
+# 4.1 Experimental Setup
+
+Datasets We employ three independent datasets, including the training set of the FilmWiki dataset containing 2,385 unsupervised texts with corresponding question-answer pairs. Additionally, we sample two widely-used benchmark QA datasets from the MRQA 2019 shared task (Fisch et al., 2019): SQuAD (Rajpurkar et al., 2016) (following Bonito (Nayak et al., 2024a)) and HotpotQA (Yang et al., 2018). To maintain comparable data scale and computational costs with the FilmWiki dataset, we extract 2,500 unsupervised texts with corresponding QA pairs from each dataset's training split as substitutes for the complete collections. Further implementation details are provided in Table 2.
+
+Baselines We consider three key baselines: zero-shot, Self-QA, Bonito and Wiki2023. For the zero-shot baseline, we prompt models for evaluation without leveraging any unsupervised texts from the
+
+<table><tr><td rowspan="2">Dataset</td><td colspan="2">Source</td><td rowspan="2"># Test Examples</td></tr><tr><td>Question</td><td>Context</td></tr><tr><td>SQuAD</td><td>Crowdsourced</td><td>Wikipedia</td><td>11639</td></tr><tr><td>HotpotQA</td><td>Crowdsourced</td><td>Wikipedia</td><td>2500</td></tr><tr><td>FilmWiki</td><td>LLM</td><td>Wikipedia</td><td>7398</td></tr></table>
+
+Table 2: Statistics for the evaluation datasets from our experiments.  
+
+<table><tr><td rowspan="2">Datasets</td><td rowspan="2">Methods</td><td colspan="2">Diversity Metrics</td></tr><tr><td>SelfBLEU Div.</td><td>Embedding Div.</td></tr><tr><td rowspan="5">SQuAD</td><td>Self-QA</td><td>0.593</td><td>0.838</td></tr><tr><td>Bonito</td><td>0.494</td><td>0.838</td></tr><tr><td>Wiki2023</td><td>0.550</td><td>0.842</td></tr><tr><td>SELF-FOVEATE</td><td>0.665</td><td>0.851</td></tr><tr><td>Test Questions</td><td>0.695</td><td>0.840</td></tr><tr><td rowspan="5">HotpotQA</td><td>Self-QA</td><td>0.463</td><td>0.823</td></tr><tr><td>Bonito</td><td>0.371</td><td>0.769</td></tr><tr><td>Wiki2023</td><td>0.554</td><td>0.822</td></tr><tr><td>SELF-FOVEATE</td><td>0.607</td><td>0.835</td></tr><tr><td>Test Questions</td><td>0.634</td><td>0.786</td></tr><tr><td rowspan="5">FilmWiki</td><td>Self-QA</td><td>0.406</td><td>0.687</td></tr><tr><td>Bonito</td><td>0.197</td><td>0.677</td></tr><tr><td>Wiki2023</td><td>0.341</td><td>0.664</td></tr><tr><td>SELF-FOVEATE</td><td>0.563</td><td>0.706</td></tr><tr><td>Test Questions</td><td>0.316</td><td>0.618</td></tr></table>
+
+Table 3: Comparison of diversity metrics across different methods and datasets. The table presents SelfBLEU Diversity (SelfBLEU Div.) and Embedding Diversity (Embedding Div.) scores for various methods on the datasets. The diversity of the test questions from each dataset is also provided as a reference.
+
+target task (None). The Self-QA baseline employs an unsupervised knowledge-guided method for extracting instruction-question-answer triples (SelfQA) (Zhang and Yang, 2023). The Bonito baseline utilizes a 7B-sized specialized model to generate various types of questions from unsupervised text. We configure it to produce questions of the "question answering without choices" type and obtain a sufficient number of samples through multiple sampling iterations (Bonito). The Wiki2023 baseline implements a text-based QA pair extraction methodology (Wiki2023). All baselines use identical unsupervised texts as our method, as specified in Section 4.1.
+
+Instruction Synthesis As described in Section 3, we process unsupervised texts through SELF-FOVEATE using GPT-4o-mini and DeepSeek-V3 to generate instructional data. We implement two distinct hyperparameter configurations emphasizing high stability and high creativity, respectively, with detailed specifications in Appendix B.
+
+Base Models We select three state-of-the-art open-source foundation models as our base models prior to instruction tuning: Meta-Llama-3.1-8B (Meta, 2024), Qwen2.5-7B (Team, 2024), and Gemma-2-9B. These decoder-only language models employ next-word prediction objectives and
+
+<table><tr><td rowspan="3">Model</td><td rowspan="3">Settings</td><td colspan="6">GPT-4o mini</td><td colspan="6">DeepSeek-V3</td></tr><tr><td colspan="2">SQuAD</td><td colspan="2">HotpotQA</td><td colspan="2">FilmWiki</td><td colspan="2">SQuAD</td><td colspan="2">HotpotQA</td><td colspan="2">FilmWiki</td></tr><tr><td>Rec.</td><td>Acc.</td><td>Rec.</td><td>Acc.</td><td>Rec.</td><td>Acc.</td><td>Rec.</td><td>Acc.</td><td>Rec.</td><td>Acc.</td><td>Rec.</td><td>Acc.</td></tr><tr><td rowspan="5">Llama-3.1-8B</td><td>None*</td><td>0.309</td><td>0.202</td><td>0.244</td><td>0.160</td><td>0.212</td><td>0.082</td><td>0.309</td><td>0.202</td><td>0.244</td><td>0.160</td><td>0.212</td><td>0.082</td></tr><tr><td>Self-QA</td><td>0.367</td><td>0.384</td><td>0.372</td><td>0.358</td><td>0.328</td><td>0.201</td><td>0.389</td><td>0.412</td><td>0.399</td><td>0.378</td><td>0.370</td><td>0.239</td></tr><tr><td>Wiki2023</td><td>0.327</td><td>0.361</td><td>0.338</td><td>0.322</td><td>0.333</td><td>0.235</td><td>0.342</td><td>0.370</td><td>0.340</td><td>0.328</td><td>0.349</td><td>0.244</td></tr><tr><td>Bonito*</td><td>0.386</td><td>0.405</td><td>0.360</td><td>0.372</td><td>0.219</td><td>0.153</td><td>0.386</td><td>0.405</td><td>0.360</td><td>0.372</td><td>0.219</td><td>0.153</td></tr><tr><td>SELF-FOVEATE</td><td>0.484</td><td>0.490</td><td>0.507</td><td>0.486</td><td>0.512</td><td>0.367</td><td>0.481</td><td>0.491</td><td>0.525</td><td>0.501</td><td>0.548</td><td>0.397</td></tr><tr><td rowspan="5">Qwen2.5-7B</td><td>None*</td><td>0.251</td><td>0.300</td><td>0.266</td><td>0.234</td><td>0.139</td><td>0.032</td><td>0.251</td><td>0.300</td><td>0.266</td><td>0.234</td><td>0.139</td><td>0.032</td></tr><tr><td>Self-QA</td><td>0.249</td><td>0.232</td><td>0.276</td><td>0.246</td><td>0.206</td><td>0.082</td><td>0.119</td><td>0.125</td><td>0.102</td><td>0.106</td><td>0.111</td><td>0.056</td></tr><tr><td>Wiki2023</td><td>0.215</td><td>0.221</td><td>0.135</td><td>0.112</td><td>0.192</td><td>0.093</td><td>0.170</td><td>0.083</td><td>0.197</td><td>0.203</td><td>0.202</td><td>0.136</td></tr><tr><td>Bonito*</td><td>0.143</td><td>0.109</td><td>0.212</td><td>0.199</td><td>0.168</td><td>0.098</td><td>0.143</td><td>0.109</td><td>0.212</td><td>0.199</td><td>0.168</td><td>0.098</td></tr><tr><td>SELF-FOVEATE</td><td>0.408</td><td>0.414</td><td>0.372</td><td>0.329</td><td>0.283</td><td>0.140</td><td>0.388</td><td>0.389</td><td>0.342</td><td>0.331</td><td>0.261</td><td>0.140</td></tr><tr><td rowspan="5">Gemma-2-9B</td><td>None*</td><td>0.224</td><td>0.121</td><td>0.175</td><td>0.078</td><td>0.211</td><td>0.099</td><td>0.224</td><td>0.121</td><td>0.175</td><td>0.078</td><td>0.221</td><td>0.099</td></tr><tr><td>Self-QA</td><td>0.383</td><td>0.409</td><td>0.408</td><td>0.389</td><td>0.429</td><td>0.315</td><td>0.402</td><td>0.435</td><td>0.424</td><td>0.408</td><td>0.509</td><td>0.386</td></tr><tr><td>Wiki2023</td><td>0.336</td><td>0.378</td><td>0.361</td><td>0.352</td><td>0.478</td><td>0.384</td><td>0.364</td><td>0.399</td><td>0.373</td><td>0.365</td><td>0.494</td><td>0.401</td></tr><tr><td>Bonito*</td><td>0.411</td><td>0.457</td><td>0.366</td><td>0.373</td><td>0.255</td><td>0.196</td><td>0.411</td><td>0.457</td><td>0.366</td><td>0.373</td><td>0.255</td><td>0.196</td></tr><tr><td>SELF-FOVEATE</td><td>0.507</td><td>0.525</td><td>0.537</td><td>0.520</td><td>0.672</td><td>0.528</td><td>0.499</td><td>0.514</td><td>0.552</td><td>0.525</td><td>0.697</td><td>0.581</td></tr></table>
+
+Table 4: Recall (Rec.) and LLM Accuracy (Acc.) on downstream tasks: SELF-FOVEATE vs. baselines. Results include models fine-tuned with instructions synthesized by GPT-4o mini or DeepSeek-V3, as well as reference non-instruction-tuned models (None). * Indicates that the base model was not fine-tuned using instructions synthesized by GPT-4o mini or DeepSeek-V3.
+
+were pretrained on trillions of tokens without any instruction-based fine-tuning.
+
+# 4.2 Diversity Discussion
+
+To explicitly evaluate the diversity of the instructions in a metric-driven manner, we follow established practices recommended in Zhu et al. (2018); Perez et al. (2022); Tevet and Berant (2021) and employ two metrics: SelfBLEU score and Sentence-BERT embedding distances (Reimers and Gurevych, 2019a). These metrics capture different facets of diversity. SelfBLEU measures diversity in the form of text, while embedding distances measure diversity in the semantics of text. For SelfBLEU scores, we compute the average SelfBLEU scores using  $n$ -grams for  $n \in \{2,3,4,5\}$ , following the approach suggested by Zhu et al. (2018). We use the implementation of the SelfBLEU metric in Alihosseini et al. (2019). Further details are available in Appendix C.
+
+Table 3 demonstrates that SELF-FOVEATE-generated instructions achieve substantial diversity improvements in both textual and semantic dimensions, attaining or even surpassing the diversity level of crowdsourced test questions through a low-cost automated process.
+
+# 4.3 Difficulty Discussion
+
+To investigate the difficulty of the generated instructions, we designed prompts and employed GPT-4o to conduct a rigorous head-to-head comparison between the instructions generated by SELF-FOVEATE and those generated by baseline methods. Specifically, for each unsupervised text in each data subset, we used two contrasting methods to generate an equal number of instructions. We then provided GPT-4o with a set of instructions generated by both methods under the same unsupervised text, ensuring that the instruction sets were anonymous and their relative positions were randomized. Subsequently, we recorded GPT-4o's judgments and calculated the win rate for SELF-FOVEATE. The results of the head-to-head comparison, as shown in Table 5, demonstrate the outstanding difficulty of the instructions synthesized by SELF-FOVEATE. Further details can be found in Appendix D.
+
+# 4.4 Problem-solving Capabilities
+
+To evaluate the impact of instruction-tuning datasets on model performance in downstream tasks, we fine-tuned open-source models and assessed their performance on in-distribution problems. Table 4 illustrates the effects of instruction sets synthesized by two models (GPT-4o mini and DeepSeek-V3) during fine-tuning on open-source
+
+![](images/ef0187342169f819ec3f316b29c4dc6653ee8cb7eb91e11f1fd15ce021084067.jpg)  
+Figure 3: Impact of instruction set scale from SELF-FOVEATE and baselines on model fine-tuning performance.
+
+![](images/ec2a44ba06a3d0d4c5d89bb671f060d33181584333274591ee4e5fe09c19c47b.jpg)
+
+model performance, with equal quantities of instructions per unsupervised text. As answers in test problems often consist of phrase-level or sentence-fragment extractions from source texts while large language models (LLMs) tend to generate more comprehensive responses, we selected both recall and LLM-evaluated accuracy as evaluation metrics. Here, higher recall indicates a greater probability of generated answers containing correct responses, while higher LLM-evaluated accuracy reflects stronger consistency between answers generated by DeepSeek-V3 and ground truth labels. To control potential confounding effects from the number of instructions synthesized per unsupervised text and examine the agreement between these two metrics, we investigated the influence of instruction set scale synthesized by SELF-FOVEATE and baseline methods on model fine-tuning performance using the HotpotQA dataset, as shown in Figure 3. Additional details regarding downstream task evaluations are provided in Appendix E.
+
+<table><tr><td rowspan="2">Dataset</td><td rowspan="2">Baseline</td><td colspan="2">SELF-FOVEATE WR.</td></tr><tr><td>Win</td><td>Lose</td></tr><tr><td rowspan="3">SQuAD</td><td>Self-QA</td><td>70.64%</td><td>29.36%</td></tr><tr><td>Wiki2023</td><td>80.83%</td><td>19.17%</td></tr><tr><td>Bonito</td><td>99.96%</td><td>00.04%</td></tr><tr><td rowspan="3">HotpotQA</td><td>Self-QA</td><td>89.52%</td><td>10.48%</td></tr><tr><td>Wiki2023</td><td>91.17%</td><td>08.83%</td></tr><tr><td>Bonito</td><td>100.00%</td><td>00.00%</td></tr><tr><td rowspan="3">FilmWiki</td><td>Self-QA</td><td>85.12%</td><td>14.88%</td></tr><tr><td>Wiki2023</td><td>95.08%</td><td>04.92%</td></tr><tr><td>Bonito</td><td>96.31%</td><td>03.69%</td></tr></table>
+
+Table 5: Head-to-Head Comparison of Instruction Difficulty for SELF-FOVEATE Against Baselines Through Win Rates (WR.) Across Datasets.
+
+Table 4 demonstrates that two state-of-the-art (SOTA) models achieve optimal performance across all downstream tasks in both metrics when using instructions synthesized via SELF-FOVEATE. Notably, during the fine-tuning process of the Qwen2.5-7B model, we observed that instruction data synthesized by baseline methods may lead to performance degradation compared to the pre-finetuned model, further demonstrating the advancement of the SELF-FOVEATE method in instruction synthesis. Our analysis presented in Figure 3 reveals that as the number of instructions synthesized per unsupervised text by SELF-FOVEATE increases, the performance of downstream fine-tuning tasks improves significantly, with an expanding performance gap compared to other baseline methods. Figure 3 also visually demonstrates the consistency between recall and LLM-evaluated accuracy, validating the appropriateness of these two metric selections.
+
+# 4.5 Ablation Studies
+
+To validate the effectiveness and necessity of each component in SELF-FOVEATE, we conduct comprehensive ablation studies focusing on two key aspects: (1) the contribution of each core component in the multi-level foveation framework, and (2) the impact of the answer regeneration mechanism in reverse synthesis.
+
+# 4.5.1 Component-wise Analysis
+
+We evaluate the necessity of each core component by systematically removing individual modules from the complete SELF-FOVEATE framework.
+
+Table 6 presents the ablation results on the SQuAD dataset using Llama-3.1-8B as the base model. The results demonstrate that each component contributes significantly to the overall performance.
+
+<table><tr><td>Setting</td><td>Recall</td><td>LLM Acc.</td></tr><tr><td>w/o Micro-Foveate</td><td>0.283</td><td>0.277</td></tr><tr><td>w/o Scatter-Foveate</td><td>0.274</td><td>0.260</td></tr><tr><td>w/o Macro-Foveate</td><td>0.344</td><td>0.339</td></tr><tr><td>SELF-FOVEATE (Full)</td><td>0.484</td><td>0.490</td></tr></table>
+
+Table 6: Ablation study results showing the contribution of each core component in SELF-FOVEATE on the SQuAD dataset with Llama-3.1-8B.
+
+The removal of any single component leads to performance degradation, validating the complementary necessity of all three core components.
+
+# 4.5.2 Answer Regeneration Analysis
+
+In our reverse synthesis paradigm, although foveate elements (e.g., "QuantumX processor") can directly serve as answers, we regenerate answers to enhance fluency, completeness, and semantic coherence. To validate this design choice, we compare the quality of answers with and without the regeneration mechanism.
+
+Table 7 presents the comparative analysis on the SQuAD dataset, evaluating both fluency and completeness of generated answers through GPT-4o evaluation. The results demonstrate significant improvements in both aspects when employing answer regeneration.
+
+<table><tr><td>Setting</td><td>High (%)</td><td>Medium (%)</td><td>Low (%)</td></tr><tr><td colspan="4">Fluency</td></tr><tr><td>w/o Regeneration</td><td>64.9</td><td>19.5</td><td>15.6</td></tr><tr><td>Reverse Synthesis</td><td>93.0</td><td>6.8</td><td>0.2</td></tr><tr><td colspan="4">Completeness</td></tr><tr><td>w/o Regeneration</td><td>36.7</td><td>27.5</td><td>35.7</td></tr><tr><td>Reverse Synthesis</td><td>76.6</td><td>19.7</td><td>3.7</td></tr></table>
+
+Table 7: Comparison of answer quality with and without regeneration mechanism in reverse synthesis on the SQuAD dataset. Fluency and completeness are evaluated on a three-point scale (High/Medium/Low).
+
+The results reveal that answer regeneration dramatically improves both fluency and completeness. This validates our methodological choice to maintain consistency across all synthesis paradigms by generating contextually appropriate and semantically coherent answers from the source text, rather than directly using extracted foveate elements.
+
+# 5 Conclusion
+
+In this paper, we proposed SELF-FOVEATE, an unsupervised text-based instruction generation method powered by LLMs, which further explores key challenges in diversity and difficulty. Its innovative multi-level foveation and Re-synthesis Module ensure the acquisition of high-quality, textfaithful instructions from unsupervised text. Extensive experiments, including performance evaluations on downstream tasks, demonstrate the effectiveness of SELF-FOVEATE. Our study has obtained many insightful findings, laying the foundation for future research in areas such as instruction difficulty assessment.
+
+# Acknowledgments
+
+This work was supported by the National Natural Science Foundation of China (NSFC) via grant 62441614 and 62176078. We thank the anonymous reviewers for their insightful comments and suggestions.
+
+# Limitations
+
+Although SELF-FOVEATE provides significant improvements in generating diverse and difficulty-oriented fine-tuning instruction data based on unsupervised text data, several limitations must be acknowledged. First, the computational cost of processing large-scale unsupervised text data using closed-source SOTA large language models (LLMs) remains substantial. Ensuring data quality requires multi-step reasoning and iterative processing on unsupervised text data, which further escalates computational demands. Second, although SELF-FOVEATE aims to fully synthesize data from unsupervised text and incorporates verification mechanisms, the inherent tendency of LLMs to hallucinate or generate erroneous information persists as a challenge. While precise prompting strategies and higher-quality unsupervised text data can mitigate these inaccuracies, they cannot be entirely eliminated.
+
+# Ethical Consideration
+
+We state that any research or application arising from this study is strictly authorized solely for research purposes. In our work, any unsupervised text datasets used are from public sources and do not contain any private information. In this paper, we have fully presented the prompts used by
+
+SELF-FOVEATE in the Appendix. All synthesized instructions rely on the provided unsupervised text and are inspected by relevant modules. Therefore, our method strives to minimize potential safety and ethical risks as much as possible. However, during the process of synthesizing fine-tuning instruction data, maliciously provided unsupervised text data can lead the model to produce harmful or inappropriate outputs, which is a shared problem. Additionally, potential unfairness and discrimination present in the unsupervised text data might be amplified by LLMs during the instruction generation process. Ensuring the quality of generated fine-tuning instruction data in a safe and highly controllable manner is crucial. The application of these techniques should be guided by ethical considerations, with safeguards in place to prevent misuse and reduce the likelihood of producing harmful outcomes.
+
+# References
+
+Danial Alihosseini, Ehsan Montahaei, and Mahdieh Soleymani Baghshah. 2019. Jointly measuring diversity and quality in text generation models. In Proceedings of the Workshop on Methods for Optimizing and Evaluating Neural Language Generation, pages 90-98, Minneapolis, Minnesota. Association for Computational Linguistics.  
+Anthropic. 2024. Claude. https://www.anthropic.com/claude.  
+Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Yunxuan Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, Albert Webson, Shixiang Shane Gu, Zhuyun Dai, Mirac Suzgun, Xinyun Chen, Aakanksha Chowdhery, Alex Castro-Ros, Marie Pellat, Kevin Robinson, Dasha Valter, Sharan Narang, Gaurav Mishra, Adams Yu, Vincent Zhao, Yanping Huang, Andrew Dai, Hongkun Yu, Slav Petrov, Ed H. Chi, Jeff Dean, Jacob Devlin, Adam Roberts, Denny Zhou, Quoc V. Le, and Jason Wei. 2022. Scaling instruction-finetuned language models. Preprint, arXiv:2210.11416.  
+John Chung, Ece Kamar, and Saleema Amershi. 2023. Increasing diversity while maintaining accuracy: Text data generation with large language models and human interventions. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 575-593, Toronto, Canada. Association for Computational Linguistics.  
+Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, Christopher Hesse, and John Schulman.
+
+2021. Training verifiers to solve math word problems. arXiv preprint arXiv:2110.14168.  
+Haixing Dai, Zhengliang Liu, Wenxiong Liao, Xiaoke Huang, Yihan Cao, Zihao Wu, Lin Zhao, Shaochen Xu, Wei Liu, Ninghao Liu, Sheng Li, Dajiang Zhu, Hongmin Cai, Lichao Sun, Quanzheng Li, Dinggang Shen, Tianming Liu, and Xiang Li. 2023. Aug-gpt: Leveraging chatgpt for text data augmentation. Preprint, arXiv:2302.13007.  
+Michael Han Daniel Han and Unsloth team. 2023. Unsloth.  
+Tri Dao. 2024. FlashAttention-2: Faster attention with better parallelism and work partitioning. In International Conference on Learning Representations (ICLR).  
+Tri Dao, Daniel Y. Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. 2022. FlashAttention: Fast and memory-efficient exact attention with IO-awareness. In Advances in Neural Information Processing Systems (NeurIPS).  
+DeepSeek-AI. 2024. Deepseek-v3 technical report. Preprint, arXiv:2412.19437.  
+Dheeru Dua, Yizhong Wang, Pradeep Dasigi, Gabriel Stanovsky, Sameer Singh, and Matt Gardner. 2019. DROP: A reading comprehension benchmark requiring discrete reasoning over paragraphs. In Proc. of NAACL.  
+Adam Fisch, Alon Talmor, Robin Jia, Minjoon Seo, Eunsol Choi, and Danqi Chen. 2019. MRQA 2019 shared task: Evaluating generalization in reading comprehension. In Proceedings of the 2nd Workshop on Machine Reading for Question Answering, pages 1-13, Hong Kong, China. Association for Computational Linguistics.  
+Yuan Ge, Yilun Liu, Chi Hu, Weibin Meng, Shimin Tao, Xiaofeng Zhao, Hongxia Ma, Li Zhang, Boxing Chen, Hao Yang, Bei Li, Tong Xiao, and Jingbo Zhu. 2024. Clustering and ranking: Diversity-preserved instruction selection through expert-aligned quality estimation. Preprint, arXiv:2402.18191.  
+Hongyi Guo, Yuanshun Yao, Wei Shen, Jiaheng Wei, Xiaoying Zhang, Zhaoran Wang, and Yang Liu. 2024. Human-instruction-free llm self-alignment with limited samples. Preprint, arXiv:2401.06785.  
+Zhang-Wei Hong, Idan Shenfeld, Tsun-Hsuan Wang, Yung-Sung Chuang, Aldo Pareja, James Glass, Akash Srivastava, and Pulkit Agrawal. 2024. Curiosity-driven red-teaming for large language models. Preprint, arXiv:2402.19464.  
+Or Honovich, Thomas Scialom, Omer Levy, and Timo Schick. 2023. Unnatural instructions: Tuning language models with (almost) no human labor. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 14409-14428, Toronto, Canada. Association for Computational Linguistics.
+
+Zhengbao Jiang, Zhiqing Sun, Weijia Shi, Pedro Rodriguez, Chunting Zhou, Graham Neubig, Xi Lin, Wen-tau Yih, and Srini Iyer. 2024. Instruction-tuned language models are better knowledge learners. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 5421-5434, Bangkok, Thailand. Association for Computational Linguistics.  
+Abdullatif Koksal, Timo Schick, Anna Korhonen, and Hinrich Schütze. 2024. Longform: Effective instruction tuning with reverse instructions. Preprint, arXiv:2304.08460.  
+Haoran Li, Qingxiu Dong, Zhengyang Tang, Chaojun Wang, Xingxing Zhang, Haoyang Huang, Shaohan Huang, Xiaolong Huang, Zeqiang Huang, Dongdong Zhang, Yuxian Gu, Xin Cheng, Xun Wang, Si-Qing Chen, Li Dong, Wei Lu, Zhifang Sui, Benyou Wang, Wai Lam, and Furu Wei. 2024a. Synthetic data (almost) from scratch: Generalized instruction tuning for language models. Preprint, arXiv:2402.13064.  
+Xian Li, Ping Yu, Chunting Zhou, Timo Schick, Omer Levy, Luke Zettlemoyer, Jason Weston, and Mike Lewis. 2024b. Self-alignment with instruction back-translation. Preprint, arXiv:2308.06259.  
+Chen Ling, Xujiang Zhao, Jiaying Lu, Chengyuan Deng, Can Zheng, Junxiang Wang, Tanmoy Chowdhury, Yun Li, Hejie Cui, Xuchao Zhang, Tianjiao Zhao, Amit Panalkar, Dhagash Mehta, Stefano Pasquali, Wei Cheng, Haoyu Wang, Yanchi Liu, Zhengzhang Chen, Haifeng Chen, Chris White, Quanquan Gu, Jian Pei, Carl Yang, and Liang Zhao. 2024. Domain specialization as the key to make large language models disruptive: A comprehensive survey. Preprint, arXiv:2305.18703.  
+Ruibo Liu, Jerry Wei, Fangyu Liu, Chenglei Si, Yanzhe Zhang, Jinmeng Rao, Steven Zheng, Daiyi Peng, Diyi Yang, Denny Zhou, and Andrew M. Dai. 2024. Best practices and lessons learned on synthetic data. Preprint, arXiv:2404.07503.  
+Lin Long, Rui Wang, Ruixuan Xiao, Junbo Zhao, Xiao Ding, Gang Chen, and Haobo Wang. 2024. On llms-driven synthetic data generation, curation, and evaluation: A survey. Preprint, arXiv:2406.15126.  
+Shayne Longpre, Le Hou, Tu Vu, Albert Webson, Hyung Won Chung, Yi Tay, Denny Zhou, Quoc V. Le, Barret Zoph, Jason Wei, and Adam Roberts. 2023. The flan collection: Designing data and methods for effective instruction tuning. Preprint, arXiv:2301.13688.  
+Meta. 2024. The llama 3 herd of models. Preprint, arXiv:2407.21783.  
+Swaroop Mishra, Daniel Khashabi, Chitta Baral, and Hannaneh Hajishirzi. 2022. Cross-task generalization via natural language crowdsourcing instructions. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 3470-3487, Dublin, Ireland. Association for Computational Linguistics.
+
+Nihal Nayak, Yiyang Nan, Avi Trost, and Stephen Bach. 2024a. Learning to generate instruction tuning datasets for zero-shot task adaptation. In *Findings of the Association for Computational Linguistics* ACL 2024, pages 12585–12611, Bangkok, Thailand and virtual meeting. Association for Computational Linguistics.  
+Nihal V. Nayak, Yiyang Nan, Avi Trost, and Stephen H. Bach. 2024b. Learning to generate instruction tuning datasets for zero-shot task adaptation. Preprint, arXiv:2402.18334.  
+Seokjin Oh, Su Ah Lee, and Woohwan Jung. 2023. Data augmentation for neural machine translation using generative language model. Preprint, arXiv:2307.16833.  
+OpenAI. 2024. Gpt-4 technical report. Preprint, arXiv:2303.08774.  
+Baolin Peng, Chunyuan Li, Pengcheng He, Michel Galley, and Jianfeng Gao. 2023. Instruction tuning with gpt-4. Preprint, arXiv:2304.03277.  
+Ethan Perez, Saffron Huang, Francis Song, Trevor Cai, Roman Ring, John Aslanides, Amelia Glaese, Nat McAleese, and Geoffrey Irving. 2022. Red teaming language models with language models. Preprint, arXiv:2202.03286.  
+Pranav Rajpurkar, Jian Zhang, Konstantin Lopyrev, and Percy Liang. 2016. SQuAD:  $100,000+$  questions for machine comprehension of text. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 2383-2392, Austin, Texas. Association for Computational Linguistics.  
+Nils Reimers and Iryna Gurevych. 2019a. SentenceBERT: Sentence embeddings using Siamese BERTnetworks. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 3982-3992, Hong Kong, China. Association for Computational Linguistics.  
+Nils Reimers and Iryna Gurevych. 2019b. Sentence-bert: Sentence embeddings using siamese bert-networks. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing. Association for Computational Linguistics.  
+Kuniaki Saito, Kihyuk Sohn, Chen-Yu Lee, and Yoshitaka Ushiku. 2024. Where is the answer? investigating positional bias in language model knowledge extraction. Preprint, arXiv:2402.12170.  
+Victor Sanh, Albert Webson, Colin Raffel, Stephen H. Bach, Lintang Sutawika, Zaid Alyafeai, Antoine Chaffin, Arnaud Stiegler, Teven Le Scao, Arun Raja, Manan Dey, M Saiful Bari, Canwen Xu, Urmish Thakker, Shanya Sharma Sharma, Eliza Szczechla, Taewoon Kim, Gunjan Chhablani, Nihal Nayak, Debajyoti Datta, Jonathan Chang, Mike Tian-Jian Jiang, Han Wang, Matteo Manica, Sheng Shen, Zheng Xin
+
+Yong, Harshit Pandey, Rachel Bawden, Thomas Wang, Trishala Neeraj, Jos Rozen, Abheesht Sharma, Andrea Santilli, Thibault Fevry, Jason Alan Fries, Ryan Teehan, Tali Bers, Stella Biderman, Leo Gao, Thomas Wolf, and Alexander M. Rush. 2022. Multitask prompted training enables zero-shot task generalization. Preprint, arXiv:2110.08207.  
+Timo Schick and Hinrich Schütze. 2021. Generating datasets with pretrained language models. Preprint, arXiv:2104.07540.  
+Zhiqing Sun, Yikang Shen, Qinhong Zhou, Hongxin Zhang, Zhenfang Chen, David Cox, Yiming Yang, and Chuang Gan. 2023. Principle-driven self-alignment of language models from scratch with minimal human supervision. Preprint, arXiv:2305.03047.  
+Rohan Taori, Ishaan Gulrajani, Tianyi Zhang, Yann Dubois, Xuechen Li, Carlos Guestrin, Percy Liang, and Tatsunori B. Hashimoto. 2023. Stanford alpaca: An instruction-following llama model. https://github.com/tatsu-lab/stanford_alpaca.  
+Qwen Team. 2024. Qwen2.5: A party of foundation models.  
+Guy Tevet and Jonathan Berant. 2021. Evaluating the evaluation of diversity in natural language generation. Preprint, arXiv:2004.02990.  
+Pablo Villalobos, Anson Ho, Jaime Sevilla, Tamay Besiroglu, Lennart Heim, and Marius Hobbahn. 2024. Will we run out of data? limits of llm scaling based on human-generated data. Preprint, arXiv:221.04325.  
+Siyuan Wang, Zhuohan Long, Zhihao Fan, Zhongyu Wei, and Xuanjing Huang. 2024. Benchmark self-evolving: A multi-agent framework for dynamic llm evaluation. Preprint, arXiv:2402.11443.  
+Yizhong Wang, Yeganeh Kordi, Swaroop Mishra, Alisa Liu, Noah A. Smith, Daniel Khashabi, and Hannaneh Hajishirzi. 2023a. Self-instruct: Aligning language models with self-generated instructions. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 13484-13508, Toronto, Canada. Association for Computational Linguistics.  
+Yufei Wang, Wanjun Zhong, Liangyou Li, Fei Mi, Xingshan Zeng, Wenyong Huang, Lifeng Shang, Xin Jiang, and Qun Liu. 2023b. Aligning large language models with human: A survey. Preprint, arXiv:2307.12966.  
+Jason Wei, Maarten Bosma, Vincent Y. Zhao, Kelvin Guu, Adams Wei Yu, Brian Lester, Nan Du, Andrew M. Dai, and Quoc V. Le. 2022. Finetuned language models are zero-shot learners. Preprint, arXiv:2109.01652.  
+Jerry Wei, Chengrun Yang, Xinying Song, Yifeng Lu, Nathan Hu, Jie Huang, Dustin Tran, Daiyi Peng, Ruibo Liu, Da Huang, Cosmo Du, and Quoc V. Le.
+
+2024. Long-form factuality in large language models. Preprint, arXiv:2403.18802.  
+Siyuan Wu, Yue Huang, Chujie Gao, Dongping Chen, Qihui Zhang, Yao Wan, Tianyi Zhou, Xiangliang Zhang, Jianfeng Gao, Chaowei Xiao, and Lichao Sun. 2024. Unigen: A unified framework for textual dataset generation using large language models. Preprint, arXiv:2406.18966.  
+Ruixuan Xiao, Yiwen Dong, Junbo Zhao, Runze Wu, Minmin Lin, Gang Chen, and Haobo Wang. 2023. FreeAL: Towards human-free active learning in the era of large language models. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 14520-14535, Singapore. Association for Computational Linguistics.  
+Canwen Xu, Daya Guo, Nan Duan, and Julian McAuley. 2023. Baize: An open-source chat model with parameter-efficient tuning on self-chat data. Preprint, arXiv:2304.01196.  
+Zhilin Yang, Peng Qi, Saizheng Zhang, Yoshua Bengio, William Cohen, Ruslan Salakhutdinov, and Christopher D. Manning. 2018. HotpotQA: A dataset for diverse, explainable multi-hop question answering. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 2369-2380, Brussels, Belgium. Association for Computational Linguistics.  
+Weizhe Yuan, Richard Yuanzhe Pang, Kyunghyun Cho, Xian Li, Sainbayar Sukhbaatar, Jing Xu, and Jason Weston. 2024. Self-rewarding language models. Preprint, arXiv:2401.10020.  
+Shengyu Zhang, Linfeng Dong, Xiaoya Li, Sen Zhang, Xiaofei Sun, Shuhe Wang, Jiwei Li, Runyi Hu, Tianwei Zhang, Fei Wu, and Guoyin Wang. 2024. Instruction tuning for large language models: A survey. Preprint, arXiv:2308.10792.  
+Xuanyu Zhang and Qing Yang. 2023. Self-qa: Unsupervised knowledge guided language model alignment. Preprint, arXiv:2305.11952.  
+Yaowei Zheng, Richong Zhang, Junhao Zhang, Yanhan Ye, Zheyan Luo, Zhangchi Feng, and Yongqiang Ma. 2024. Llamafactory: Unified efficient fine-tuning of  $100+$  language models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 3: System Demonstrations), Bangkok, Thailand. Association for Computational Linguistics.  
+Chunting Zhou, Pengfei Liu, Puxin Xu, Srini Iyer, Jiao Sun, Yuning Mao, Xuezhe Ma, Avia Efrat, Ping Yu, Lili Yu, Susan Zhang, Gargi Ghosh, Mike Lewis, Luke Zettlemoyer, and Omer Levy. 2023a. Lima: Less is more for alignment. Preprint, arXiv:2305.11206.  
+Jeffrey Zhou, Tianjian Lu, Swaroop Mishra, Siddhartha Brahma, Sujoy Basu, Yi Luan, Denny Zhou, and Le Hou. 2023b. Instruction-following evaluation for large language models. Preprint, arXiv:2311.07911.
+
+Yaoming Zhu, Sidi Lu, Lei Zheng, Jiaxian Guo, Weinan Zhang, Jun Wang, and Yong Yu. 2018. Texygen: A benchmarking platform for text generation models. Preprint, arXiv:1802.01886.
+
+# A Datasets
+
+# A.1 Dataset Details
+
+This section provides comprehensive details regarding the datasets utilized in this study.
+
+SQuAD(Rajpurkar et al., 2016) This Dataset serves as a benchmark for extractive question answering, containing over 100,000 human-generated question-answer pairs anchored to Wikipedia passages. In our work, we utilize this dataset from the MRQA 2019 shared task(Fisch et al., 2019). Considering computational costs, we employ 2,500 randomly selected articles or contexts from the training set.
+
+HotpotQA(Yang et al., 2018) The original purpose of proposing this dataset was to challenge models with multi-hop reasoning across 113k Wikipedia-based QA pairs requiring synthesis of information from multiple documents. In our work, we utilize this dataset from the MRQA 2019 shared task(Fisch et al., 2019). Considering computational costs, we employ 2,500 randomly selected articles or contexts from the training set.
+
+FilmWiki(Saito et al., 2024) This dataset was initially constructed to investigate a phenomenon called the perplexity curse, with all texts sourced from Wikipedia. Our study employs its film-themed subset containing the complete collection of 2,385 unsupervised texts and their corresponding questions.
+
+# A.2 Usage of Datasets
+
+During the synthesis process, we use the articles or contexts from the dataset (excluding the questions) and generate instructions through the SELF-FOVEATE. In the evaluation of problem-solving capabilities, we assess the fine-tuned models using the corresponding questions in the dataset.
+
+# B Instruction Synthesis
+
+Taking into account both performance and API costs, and to eliminate potential biases in experimental results caused by specific teacher models (such as prompt hacking), we employ two models, GPT-4o mini(gpt-4o-mini-2024-07-18)(OpenAI, 2024) and DeepSeek-V3(DeepSeek-AI, 2024), to generate instructions from unsupervised text using the same methodology and sampling hyperparameters. Detailed hyperparameters are provided in Table 8, where, as described in Section 3, the
+
+normal mode is utilized for the initial foveate synthesis, and the high-creativity mode is adopted for re-synthesis.
+
+<table><tr><td>Hyperparameters</td><td>Values</td></tr><tr><td colspan="2">Normal Mode</td></tr><tr><td>frequency_penalty</td><td>0.5</td></tr><tr><td>max Completion_tokens</td><td>None</td></tr><tr><td>presence_penalty</td><td>0</td></tr><tr><td>temperature</td><td>0.5</td></tr><tr><td>top_p</td><td>1.0</td></tr><tr><td colspan="2">High-creativity Mode</td></tr><tr><td>frequency_penalty</td><td>0.5</td></tr><tr><td>max Completion_tokens</td><td>None</td></tr><tr><td>presence_penalty</td><td>0</td></tr><tr><td>temperature</td><td>1.2</td></tr><tr><td>top_p</td><td>1.0</td></tr></table>
+
+# C Diversity Experiment
+
+Mathematically, we define the diversity metrics as follows ('D' denotes 'Diversity', 'SB' denotes 'SelfBLEU', and 'EB' denotes 'Embedding'):
+
+$$
+\mathrm {D} _ {\mathrm {S B}} = 1 - \frac {1}{| \mathcal {X} _ {\tau} |} \sum_ {x _ {i} \in \mathcal {X} _ {\tau}} \sum_ {n = 2} ^ {5} \text {S e l f B L E U} _ {\mathcal {X} _ {\tau}} \left(x _ {i}, n\right) \tag {2}
+$$
+
+$$
+\mathrm {D} _ {\mathrm {E B}} = 1 - \frac {1}{2 \left| \mathcal {X} _ {\tau} \right|} \sum_ {x _ {i} \in \mathcal {X} _ {\tau}} \sum_ {x _ {j} \in \mathcal {X} _ {\tau}} \frac {\phi \left(x _ {i}\right) \cdot \phi \left(x _ {j}\right)}{\left\| \phi \left(x _ {i}\right) \right\| ^ {2} \left\| \phi \left(x _ {j}\right) \right\| ^ {2}} \tag {3}
+$$
+
+For computing sentence embeddings, we employ sentence-transformers/paraphrase-multilingualmpnet-base-v2(Reimers and Gurevych, 2019b), which maps sentences to a 768-dimensional dense vector space. Compared to other sentence embedding transformation models, it offers a balanced mapping dimension and accuracy, enabling the computation of cosine similarity matrices at an acceptable computational cost while obtaining relatively accurate diversity evaluation metrics.
+
+# D Difficulty Head-to-Head Experiment
+
+We prompt GPT-4o (gpt-4o-2024-08-06)(OpenAI, 2024) to determine which method-synthesized instruction set is more challenging. The prompts are shown below.
+
+# Difficulty Head-to-Head Experiment System Prompt
+
+Please compare two sets of instructions/questions for a given content to determine which set is more difficult for a language model to generate coherent and accurate responses. A set of instructions is considered more difficult if it requires:
+
+1. More complex reasoning or multi-step analysis  
+2. Integration of multiple concepts or contexts  
+3. Explanation of nuanced relationships or processes  
+4. Handling of ambiguous or less-documented information  
+5. Synthesis of information from diverse sources or time periods
+
+Please respond with '1' if the first set of instructions is more difficult, or '2' if the second set of instructions is more difficult.
+
+Please respond with the number only, no other text or characters.
+
+# Difficulty Head-to-Head Experiment User Prompt
+
+Content: {unsupervised text}
+
+Instruction Set 1: {instruction set a}
+
+Instruction Set 2: {instruction set b}
+
+The hyperparameters set during sampling are listed in Table 9.
+
+Table 8: Hyperparameters of Synthetic Task Generation.  
+
+<table><tr><td>Hyperparameters</td><td>Values</td></tr><tr><td>frequency_penalty</td><td>0.0</td></tr><tr><td>max Completion_tokens</td><td>5</td></tr><tr><td>presence_penalty</td><td>0</td></tr><tr><td>temperature</td><td>0.5</td></tr><tr><td>top_p</td><td>1.0</td></tr></table>
+
+Table 9: Hyperparameters of Difficulty Head-to-Head Experiment.
+
+# E Downstream Experiment
+
+# E.1 Software and Hardware Details
+
+The implementation leverages the LLaMA-Factory framework (Zheng et al., 2024) with computational optimizations from FlashAttention2 (Dao et al., 2022; Dao, 2024) and Unsloth libraries (Daniel Han and team, 2023). For training, we employ NVIDIA A100, A800, and H800 GPUs based on availability within our computational cluster. The entire work required approximately 110 GPU hours.
+
+# E.2 Hyperparameters
+
+The hyperparameters and technical configurations for the instruction tuning process are documented in Table 10.
+
+<table><tr><td>Hyperparameters</td><td>Values</td></tr><tr><td>cutoff_len</td><td>2048</td></tr><tr><td>learning_rate</td><td>0.0001</td></tr><tr><td>num_train_epochs</td><td>5.0</td></tr><tr><td>effective_batch_size</td><td>16</td></tr><tr><td>lr_scheduler_type</td><td>cosine</td></tr><tr><td>max_grad_norm</td><td>1.0</td></tr><tr><td>warmup_steps</td><td>0</td></tr><tr><td>optim</td><td>adamw_torch</td></tr><tr><td>quantization_bit</td><td>4</td></tr><tr><td>quantization_method</td><td>bitsandbytes</td></tr><tr><td>lora_rank</td><td>8</td></tr><tr><td>lora_alpha</td><td>16</td></tr><tr><td>lora_dropout</td><td>0</td></tr></table>
+
+# E.3 LLM Accuary Evaluation
+
+# LLM Accuracy Evaluation
+
+# System Prompt
+
+You are a fair judge. Your task is to determine if the generated answer correctly answers the question, even if it contains additional explanations. Rules:
+
+1. The generated answer is correct if it contains the key information from the ground truth  
+2. Additional explanations or context in the generated answer should not make it incorrect  
+3. Only respond with 'Correct' or 'Incorrect'
+
+# LLM Accuracy Evaluation
+
+# User Prompt
+
+Compare the following answers:
+
+Question: {question}
+
+Ground Truth Answer: {groundtruth}
+
+Generated Answer: {generated}
+
+Is the generated answer correct, regardless of any additional explanation? Respond only with 'Correct' or 'Incorrect'.
+
+# F Answer Regeneration Evaluation
+
+This section provides detailed documentation of the evaluation methodology used in the ablation study for answer regeneration analysis (Section 4.5). To assess the quality of answers generated with and
+
+without the regeneration mechanism, we employed GPT-4o (gpt-4o-2024-08-06) to evaluate fluency and completeness on a three-point scale.
+
+# F.1 Evaluation Prompt
+
+The following system prompt was used to evaluate question-answer pairs:
+
+# Answer Regeneration Evaluation
+
+# System Prompt
+
+Please evaluate the given question and answer pair based on two criteria:
+
+1. Fluency: How well does the answer flow and connect with the question?  
+- High: The answer naturally follows from the question  
+- Medium: The connection is somewhat clear but could be improved  
+- Low: The answer feels disconnected from the question  
+2. Completeness: How thoroughly does the answer address the question?  
+- High: The answer fully addresses all aspects of the question  
+- Medium: The answer covers most aspects but misses some points  
+- Low: The answer only partially addresses the question
+
+Please respond with a JSON object in the following format:
+
+"fluency": "highlmediumllow",
+
+"completeness": "highlmediumllow"
+
+Do not include any other text or explanation.
+
+# F.2 Evaluation Hyperparameters
+
+The hyperparameters used for the answer regeneration evaluation are provided in Table 11.
+
+Table 10: Hyperparameters of Instruction Tuning with Q-LoRA Quantization.  
+
+<table><tr><td>Hyperparameters</td><td>Values</td></tr><tr><td>frequency_penalty</td><td>0.0</td></tr><tr><td>max Completion_tokens</td><td>50</td></tr><tr><td>presence_penalty</td><td>0</td></tr><tr><td>temperature</td><td>0.2</td></tr><tr><td>top_p</td><td>1.0</td></tr></table>
+
+Table 11: Hyperparameters for Answer Regeneration Evaluation.
+
+# G Use of AI Assistants
+
+In this study, we utilized AI-powered tools, including ChatGPT and Grammarly, to enhance the lin
+
+guistic accuracy of our manuscript through spell-checking and minor grammatical corrections. Additionally, the codebase was developed using Cursor to improve coding efficiency and accuracy.

2025/Self-Foveate_ Enhancing Diversity and Difficulty of Synthesized Instructions from Unsupervised Text via Multi-Level Foveation/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:012af0f3d829541ed133f0b6b03be7257bfd61c812bfa8b8422fe9b5f61c186c
+size 730592

2025/Self-Improvement Towards Pareto Optimality_ Mitigating Preference Conflicts in Multi-Objective Alignment/911edb3a-c6cd-4b51-a4d2-8be2982ec5b3_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5f3139b47fd4a484c7b27a04dcba60260c3d1efc16346d9f73bf935b6949c641
+size 877011

2025/Self-Improvement Towards Pareto Optimality_ Mitigating Preference Conflicts in Multi-Objective Alignment/full.md

@@ -0,0 +1,621 @@
+# Self-Improvement Towards Pareto Optimality: Mitigating Preference Conflicts in Multi-Objective Alignment
+
+Moxin Li $^{1*}$ , Yuantao Zhang $^{1*}$ , Wenjie Wang $^{2\dagger}$ , Wentao Shi $^{2}$ , Zhuo Liu $^{2}$ , Fuli Feng $^{2\dagger}$ , Tat-Seng Chua $^{1}$
+
+<sup>1</sup>National University of Singapore, <sup>2</sup>University of Science and Technology of China limoxin@u.nus.edu, yuantaozhang@u.nus.edu wenjiewang96@gmail.com, shiwentao123@mail.ustc.edu.cn, liuz郵mail.ustc.edu.cn, fulifeng93@gmail.com, dcscts@nus.edu.sg
+
+# Abstract
+
+Multi-Objective Alignment (MOA) aims to align LLMs' responses with multiple human preference objectives, with Direct Preference Optimization (DPO) emerging as a prominent approach. However, we find that DPO-based MOA approaches suffer from widespread preference conflicts in the data, where different objectives favor different responses. This results in conflicting optimization directions, hindering the optimization on the Pareto Front. To address this, we propose to construct Pareto-optimal responses to resolve preference conflicts. To efficiently obtain and utilize such responses, we propose a self-improving DPO framework that enables LLMs to self-generate and select Pareto-optimal responses for self-supervised preference alignment. Extensive experiments on two datasets demonstrate the superior Pareto Front achieved by our framework compared to various baselines<sup>1</sup>.
+
+# 1 Introduction
+
+Aligning Large Language Models (LLMs) with human preferences (Ouyang et al., 2022; Rafailov et al., 2023) has evolved from single-objective to multi-objective, aiming to comprehensively capture the inherent heterogeneity of human preferences. Multi-Objective Alignment (MOA) (Ramé et al., 2023; Wang et al., 2024b; Zhong et al., 2024b) has jointly considered multiple human preference objectives, such as safety, helpfulness, factuality, and diversity, to optimize the LLM. The optimization outcome of MOA is a set of LLMs optimized under various preference weights across these objectives, forming a (close-to) Pareto Front.
+
+Existing MOA approaches can be broadly classified into two categories by their optimization strategies. Reinforcement Learning (RL)-based
+
+![](images/5941022ac4d5d6d6d0e758c3703fd147559533ca71ff7106e35646dee98526e9.jpg)
+
+![](images/2bf843e9d3719fc440d60e1ea5ed3313f7048a60dd27b57f8b868ffb6b6d9eaf.jpg)
+
+![](images/0a7122cc70e703f9aedd42423f879ea66df82f2396f2841788f06a7d6f5e9a44.jpg)  
+Figure 1: Illustration on the impact of preference conflicts to MOA, and how Pareto-optimal responses can mitigate such issue for superior Pareto Front.
+
+approaches (Ramé et al., 2023; Wang et al., 2024b) learn a proxy reward model for each objective and then update the LLM using RL, targeting at a weighted combination of the proxy rewards. Direct Preference Optimization (DPO)-based approaches (Zhou et al., 2024) follow a distinct paradigm, where DPO optimization targets are derived for each objective and jointly aggregated under the preference weight. Since DPO-based methods offer advantages in cost and stability over RL-based approaches (Rafailov et al., 2023), it has been a promising direction to study MOA via DPO.
+
+However, after comprehensively investigated DPO-based MOA (cf. Section 2), we observe that these approaches are prone to be impacted by the widespread preference conflicts in the training data, which hinders the achievement of superior Pareto Front. Given question and a pair of responses, different objectives often favor different responses, resulting in preference conflicts among these objectives. These preference conflicts create contra
+
+dictionary optimization targets for different objectives under the aggregation of DPO-based approaches, potentially disrupting the alignment toward each objective and hindering the achievement of superior Pareto Front (see analysis in Section 2). Given the high prevalence of conflicting preferences in existing datasets (cf. Table 1), simply discarding these instances in alignment is not a viable solution.
+
+To solve the issue of preference conflicts, we propose to construct Pareto-optimal responses (cf. Figure 1). Given a pair of responses with conflicting preferences, denoted as  $\mathbf{a}$  and  $\mathbf{b}$ , where  $\mathbf{a}$  is better than  $\mathbf{b}$  in objective 1 but worse in objective 2 we propose to construct a Pareto-optimal response  $\mathbf{c}$ , which surpasses  $\mathbf{a}$  and  $\mathbf{b}$  on both objectives. We think learning the preference between  $\mathbf{c}$  and  $\mathbf{a}$  for objective 1 and  $\mathbf{c}$  and  $\mathbf{b}$  for objective 2 not only incurs no preference conflicts, but also guides the LLM toward generating responses that perform well on both objectives, thus achieving a superior Pareto Front. To obtain Pareto-optimal responses, it is not advisable to manually write due to the large amount of preference conflict instances. Therefore, we consider using automatic approach to obtain Pareto-optimal responses from the LLM itself.
+
+To this end, we propose a novel Self-Improvement DPO framework towards Pareto Optimality (SIPO), which guides the LLM to self-generate and select Pareto-optimal responses, thereby mitigating preference conflicts and enhancing the Pareto Front. After initial alignment, SIPO samples high-quality responses with a self-refinement strategy, which are then evaluated and filtered for Pareto-optimality over original responses. Finally, the Pareto-optimal responses are paired with original responses for non-conflicting DPO-based preference fine-tuning. SIPO can be easily incorporated with existing DPO-based MOA approaches. Experimental results on HelpSteer (Wang et al., 2024b) and BeaverTails (Ji et al., 2023) demonstrate significant improvement over baseline methods. Our contributions are three-fold:
+
+- We identify the negative impact of preference conflicts on achieving superior Pareto Front for DPO-based MOA approaches.  
+- We propose to construct Pareto-optimal responses to mitigate the issue, and propose a novel framework for automatically generating, selecting and utilizing these responses.  
+- We conduct extensive experiments to validate
+
+the effectiveness of our framework, achieving 2.1 and 3.0 average improvement on the helpful and harmless rewards of BeaverTails.
+
+# 2 Preliminary Experiments
+
+Background The alignment objectives are denoted as a set of  $N$  ground-truth reward functions,  $\mathbf{r}^{*}(\mathbf{x},\mathbf{y}) = [r_{1}^{*}(\mathbf{x},\mathbf{y}),\dots,r_{N}^{*}(\mathbf{x},\mathbf{y})]^{\intercal}$ . The goal of MOA is to align the LLM based on a set of preference weights  $\mathbf{W} = \{\mathbf{w}_m\}_{m=1}^M$ . Each preference weight vector  $\mathbf{w}_m = [w_{m_1},\dots,w_{m_N}]^{\intercal}$  satisfies the constraint  $\sum_{i=1}^{N} w_{m_i} = 1$ , which balances these objectives. Aligning the LLM to a given preference weight entails maximizing the weighted reward  $\mathbf{w}^{\intercal} \mathbf{r}^{*}(\mathbf{x},\mathbf{y})$ . The resulting set of aligned LLMs form a (close-to) Pareto Front.
+
+The alignment is typically achieved using a multi-objective preference dataset,  $\mathcal{D} = \{\mathcal{D}_1,\dots,\mathcal{D}_N\}$ , where  $\mathcal{D}_i = \{(\mathbf{x},\mathbf{y}_w,\mathbf{y}_l)\}$  represents the preference dataset for objective  $i$ . Here,  $\mathbf{x}$  is the input, while  $\mathbf{y}_w$  and  $\mathbf{y}_l$  denote the preferred and dispreferred responses, respectively. Frequently, the inputs and responses remain the same across all preference datasets in  $\mathcal{D}$ , with only the preference labels differing across objectives, as this format simplifies the annotation process for human annotators. Thus we can reformulate the dataset as  $\mathcal{D} = \{(\mathbf{x},\mathbf{y}_{-1},\mathbf{y}_1,p_1,\dots,p_N)\}$ ,  $p_i\in \{-1,1\}$  as the label of the preferred response for objective  $i$ .
+
+The Impact of Preference Conflicts on DPO-based MOA Recently, DPO-based methods, such as MODPO (Zhou et al., 2024) and DPO soups (Ramé et al., 2023), have been introduced to reduce the costs of proxy reward models and RL. These methods generally follow such a paradigm: they define a DPO optimization target for each objective and then employ an aggregation strategy to combine these targets using  $\mathbf{w}$ . The specific optimization targets and aggregation strategies vary across different approaches. More specifically, DPO soups optimizes a separate LLM for each objective by DPO and then aggregate them at the model parameter level by weight merging. MODPO trains DPO LLMs as proxy reward models for certain objectives and aggregates them at the loss level by interpolating the weighted reward differences as margins into the DPO loss function. The naive baseline, DPO Loss Weighting (LW), computes the DPO loss for each objective and aggregates them at the loss level by a weighted sum.
+
+However, we observe that this paradigm is eas
+
+![](images/a1535002565cd4fd288ef53d27b10c88dd27b1801db30037de2b9bbe60a3a04f.jpg)  
+Figure 2: The impact of preference conflicts on Pareto Front optimization. Experiments are conducted on DPO soups with Alpaca-7B.
+
+ily hindered by preference conflicts in the data. Preference conflict refers to the instance where different objectives assign different preference labels, defined as:  $\{(\mathbf{x},\mathbf{y}_1,\mathbf{y}_2,p_1,\dots,p_N)|\exists i,j\in [1,N + 1],p_i\neq p_j\}$ . Aligning on these instances introduces contradictory optimization targets, disrupting the learning for individual objectives and ultimately hindering Pareto Front optimization.
+
+To illustrate this issue, we take an example on the naive DPO LW method with  $N = 2$ , where the loss is defined as a weighted sum of the DPO losses on each objective:  $\mathcal{L}_{\mathrm{DPO\_LW}} = w_1\mathcal{L}_1 + w_2\mathcal{L}_2$ . With conflicting preferences, i.e.,  $p_1 \neq p_2$ , the losses  $\mathcal{L}_1$  and  $\mathcal{L}_2$  are opposite,  $\mathcal{L}_1 = -\mathcal{L}_2$ , pulling the optimization in opposing directions. As a result, optimizing  $\mathcal{L}_{\mathrm{DPO\_LW}}$  leads to conflicting gradient updates, preventing the LLM from effectively aligning with each objective and ultimately degrading the Pareto Front. This issue extends to other DPO-based MOA approaches and holds for larger values of  $N$ . To further illustrate this issue, we conduct the following controlled experiment.
+
+Evaluation Protocol To investigate the impact of preference conflicts on Pareto Front optimization, we conduct experiments by controlling the ratio of preference conflicts in alignment. Specifically, we subsample equal-sized subsets from  $\mathcal{D}$  with  $0\%$ ,  $30\%$ ,  $60\%$ , and  $90\%$  of conflicting preferences, and compare their optimized Pareto Front. We examine this problem from multiple perspectives. Firstly, we evaluate two prominent DPO-based MOA approaches, MODPO and DPO soups. Secondly, we utilize two widely-used multi-
+
+![](images/95d67572bb19b9953abbed93e92d72e24321a26ed550064db831c7741e61c7a6.jpg)  
+Figure 3: The impact of preference conflicts on Pareto Front optimization. Experiments are conducted on MODPO with Alpaca-7B.
+
+objective preference datasets with two different pairs of objectives. We choose the correctness and morbidity from HelpSteer, (Wang et al., 2024b), harmless and helpful from BeaverTails (Ji et al., 2023). Finally, we experiment with different backbone LLMs, including Alpaca-7B (Taori et al., 2023) and a supervised fine-tuned LLaMA-2-7B (Touvron et al., 2023). More details on the methods, datasets and backbone LLMs can be found in Section 4 and Appendix A.
+
+Results on Different Methods and Objectives Figure 2 shows the Pareto Fronts for DPO soups under varying conflict ratios of the alignment data. We also show the average performance decrease over different preference weights for each objective. Corresponding results on MODPO is shown in Figure 3. We can observe that (1) as the ratio of conflicts in the training data increases, the Pareto Fronts gradually move downwards, showing significant performance decreases. This phenomenon holds for all datasets and methods, which validates the existence of the issue. For DPO soups, when the conflict ratio reaches  $90\%$ , the Pareto Front even approaches the performance on the original LLM without alignment (denoted as SFT), showing severe alignment problem. (2) All objectives incur significant average performance decreases on both methods. Helpful and harmless have more significant performance decreases than correctness and morbidity, which may be related to the more conflicting nature of the definition of these objectives. (3) However, reducing the conflict ratio of the data generally hurts the steerability of the Pareto
+
+![](images/a86cda4953f5801960f4d79890c56921d4ad57c6fffdbc26c2aae1f956b941ab.jpg)  
+Figure 4: The impact of preference conflicts on Pareto Front optimization. Results of BeaverTails with DPO soups on LLaMA-2-7B.
+
+<table><tr><td>Dataset</td><td colspan="3">HelpSteer</td><td>BeaverTails</td></tr><tr><td># Objectives</td><td>3</td><td>4</td><td>5</td><td>2</td></tr><tr><td>Conflict Ratio (%)</td><td>11.86</td><td>15.89</td><td>17.94</td><td>53.83</td></tr></table>
+
+Table 1: Statistics on the conflict ratio in Helpsteer and BeaverTails datasets.
+
+Fronts, meaning that the performance ranges of the two objectives across preference weights get tighter under smaller conflict ratios. We conjecture that controlling the conflict ratio in the data may hurt the versatility of the data, thus hindering the optimization of single objectives toward higher performance under certain preference weights.
+
+Results on Different Backbone LLMs To examine the consistency of this issue on different backbone LLMs, we utilize a supervised fine-tuned LLaMA-2-7B as an additional backbone LLM. The results of BeaverTails on DPO soups is shown in Figure 4. We can observe that even though LLaMA2-7B achieves better performance on both objectives than Alpaca-7B (compared with Figure 2), the conflict ratio consistently affects the Pareto Front, showing that stronger backbone LLM will also be affected by the preference conflicts, further demonstrating the existence of the issue.
+
+Statistics on the Percentage of Conflicting Data We have conducted statistics on the ratio of preference conflicts in these datasets, as shown in Table 1. For BeaverTails, we calculate the conflict ratio for the two objectives. For HelpSteer, we vary the number of objectives from three to five. We can observe that the helpful and harmless in BeaverTails has more than  $50\%$  of conflict, showing strong conflicting nature. Statistics for HelpSteer are all more than  $10\%$ , and increasing the number of objectives further increases the conflict ratio. The statistics reveals the severity of the preference conflicts in current datasets, stressing the need for mitigation.
+
+# 3 Method
+
+In this section, we introduce our SIPO framework (cf. Figure 5), which leverages self-generated Pareto-optimal responses to mitigate the impact of preference conflicts. We will introduce the definition of Pareto-optimal responses (§ 3.1), and detail the SIPO framework design (§ 3.2, § 3.3).
+
+# 3.1 Pareto-Optimal Responses
+
+To solve the issue of preference conflicts, we resort to Pareto-optimal responses. For an instance  $(\mathbf{x},\mathbf{y}_{-1},\mathbf{y}_1,p_1,\dots ,p_N)\in \mathcal{D}$  with conflicting preferences, the Pareto-optimal responses  $\mathbf{y}_c$  are defined as those responses that outperform both  $\mathbf{y}_{-1}$  and  $\mathbf{y}_1$  across all objectives:
+
+$$
+\begin{array}{l} \mathbf {y} _ {c} = \left\{\mathbf {y} | \forall i, r _ {i} ^ {*} (\mathbf {x}, \mathbf {y}) > r _ {i} ^ {*} (\mathbf {x}, \mathbf {y} _ {1}) \text {a n d} \right. \tag {1} \\ r _ {i} ^ {*} (\mathbf {x}, \mathbf {y}) > r _ {i} ^ {*} (\mathbf {x}, \mathbf {y} _ {- 1}) \}. \\ \end{array}
+$$
+
+$\mathbf{y}_c$  incurs no preference conflicts with  $\mathbf{y}_1$  and  $\mathbf{y}_{-1}$ , thereby avoiding the issues outlined in Section 2.  $\mathbf{y}_c$  also has better quality in terms of all objectives than  $\mathbf{y}_1$  and  $\mathbf{y}_{-1}$ , also facilitating the achievement of a more optimal Pareto Front.
+
+# 3.2 SIPO Framework: Responses Generation
+
+Given that human annotation of Pareto-optimal responses is prohibitively expensive and infeasible for large-scale datasets, our SIPO framework is designed to autonomously generate and leverage Pareto-optimal responses. We initially align  $N$  policy LLMs to capture each objective using DPO, denoted as  $\Pi = \{\pi_{\theta_i}\}_{i=1}^N$ .
+
+$$
+\begin{array}{l} \theta_ {i} = \arg \min  _ {\theta} - \mathbb {E} _ {\boldsymbol {\mathcal {D}}} \left[ \log \sigma \left(p _ {i} \beta \frac {\pi_ {\theta} (\mathbf {y} _ {1} | \mathbf {x})}{\pi_ {\mathrm {r e f}} (\mathbf {y} _ {1} | \mathbf {x})} \right. \right. \\ \left. - p _ {i} \beta \frac {\pi_ {\theta} (\mathbf {y} _ {- 1} | \mathbf {x})}{\pi_ {\mathrm {r e f}} (\mathbf {y} _ {- 1} | \mathbf {x})}\right) \Bigg ], \tag {2} \\ \end{array}
+$$
+
+where  $\pi_{\mathrm{ref}}$  denotes the reference LLM. Then, we aim for generating Pareto-optimal responses with three stages, Sampling, Refinement and Filtering.
+
+Stage 1: Sampling For the sampling stage, we aim to generate diverse high-quality responses based on the aligned policy LLMs. To enhance sampling diversity, we apply a set of preference weights  $\mathbf{W} = \{\mathbf{w}_m\}_{m=1}^M$  and generate responses under each  $\mathbf{w}_m$  for  $\mathbf{x}$ , denoted as  $\mathbf{y}_m^s$ . To ensure sampling quality, we utilizing the outstanding decoding-based method MOD (Shi et al., 2024) to sample responses
+
+![](images/a1b451be50c39d30c408d8bca9ba751ae964c1403f47462389307f963df2c744.jpg)  
+Figure 5: Illustration of our proposed SIPO framework.
+
+from  $\Pi$  under a given  $\mathbf{w}_m$ . Denoting the MOD decoding function as  $f^d(\cdot)$ ,
+
+$$
+\mathbf {y} _ {m} ^ {s} = f ^ {d} \left(\boldsymbol {\Pi}, \mathbf {w} _ {m}, \mathbf {x}\right). \tag {3}
+$$
+
+Stage 2: Refinement To further enhance the quality of the sampled  $\mathbf{y}_m^s$ , we employ a self-refinement strategy, prompting LLM to review the flaws of  $\mathbf{y}_m^s$  from the perspectives of different objectives and then revise it. Firstly, we employ an evaluator LLM to analyze and generate reviews from different perspectives. This evaluator needs to possess the ability of different objectives, thus we implement MOD on the policy LLMs with a preference weight  $\mathbf{w}_e$  to mix the objectives.
+
+$$
+\mathbf {y} _ {m} ^ {v} = f ^ {d} \left(\boldsymbol {\Pi}, \mathbf {w} _ {e}, \left[ \mathbf {p} _ {v}, \mathbf {x}, \mathbf {y} _ {m} ^ {s} \right]\right), \tag {4}
+$$
+
+where  $\mathbf{y}_m^v$  is the generated review for  $\mathbf{y}_m^s$ ,  $\mathbf{p}_v$  is the instruction and in-context examples guiding the review generation. Then, we revise the response  $\mathbf{y}_m^s$  based on  $\mathbf{y}_m^v$  using the original sampling policy of  $\mathbf{y}_m^s$  to obtain an enhanced response  $\mathbf{y}_m^a$ .
+
+$$
+\mathbf {y} _ {m} ^ {a} = f ^ {d} (\boldsymbol {\Pi}, \mathbf {w} _ {m}, [ \mathbf {p} _ {a}, \mathbf {x}, \mathbf {y} _ {m} ^ {v}, \mathbf {y} _ {m} ^ {s} ]), \qquad (5)
+$$
+
+where  $\mathbf{p}_a$  is the instruction and in-context examples. We hope choosing different weights of  $\mathbf{w}_e$  and  $\mathbf{w}_m$  for reviewing and rewriting policies can leverage their joint effectiveness.
+
+Stage 3: Filtering After obtaining the sampled responses, we apply a filtering stage to ensure the Pareto-optimality over the original responses  $\mathbf{y}_1$  and  $\mathbf{y}_{-1}$ , as defined in Eq. (1). In the absence of the ground-truth reward functions  $\mathbf{r}^{*}(\cdot ,\cdot)$ , we leverage the implicit reward function from DPO models, i.e.,  $\Pi$ , to estimate the rewards.
+
+$$
+\hat {r} _ {i} (\mathbf {x}, \mathbf {y} _ {m} ^ {a}) = \beta \log \pi_ {\theta_ {i}} \left(\mathbf {y} _ {m} ^ {a} | \mathbf {x}\right) + \beta \log Z (\mathbf {x}), \tag {6}
+$$
+
+where  $Z(\mathbf{x})$  is a normalization constant independent of the responses, allowing us to disregard it. Each  $\mathbf{y}_m^a$  obtains a set of rewards,  $\hat{\mathbf{r}}_{\Pi}(\mathbf{x},\mathbf{y}_m^a) = [\hat{r}_1(\mathbf{x},\mathbf{y}_m^a),\dots,\hat{r}_N(\mathbf{x},\mathbf{y}_m^a)]^\intercal$ . Apart from the DPO models in  $\Pi$ , we also utilize  $M$  additional policy LLMs combined under preference weights  $\mathbf{W}$  to further calculate the rewards on mixed objectives, denoted as  $\hat{\mathbf{r}}_{\mathbf{W}}(\mathbf{x},\mathbf{y}_m^a)$ . The combined policy LLMs are obtained via model weight merging following DPO soups (Ramé et al., 2023).
+
+Finally, we select the Pareto-optimal  $\mathbf{y}_m^a$  with all rewards  $\hat{\mathbf{r}}_{\Pi}$  and  $\hat{\mathbf{r}}_{\mathbf{W}}$  larger than the original responses. If multiple  $\mathbf{y}_m^a$  for a single  $\mathbf{x}$  satisfy such constraints, we choose the one with the largest average reward as  $\mathbf{y}_c$ .
+
+$$
+\begin{array}{l} \mathbf {y} _ {c} = \left\{\mathbf {y} _ {m} ^ {a} \mid \hat {r} (\mathbf {x}, \mathbf {y} _ {m} ^ {a}) > \hat {r} (\mathbf {x}, \mathbf {y} _ {1}), \text {a n d} \right. \\ \hat {r} (\mathbf {x}, \mathbf {y} _ {m} ^ {a}) > \hat {r} (\mathbf {x}, \mathbf {y} _ {- 1}), \forall \hat {r} \in \hat {\mathbf {r}} _ {\mathbf {W}} \cup \hat {\mathbf {r}} _ {\boldsymbol {\Pi}} \}. \tag {7} \\ \end{array}
+$$
+
+# 3.3 SIPO Framework: Fine-Tuning
+
+After obtaining  $\mathbf{y}_c$ , we update the policy LLMs to reduce the effect of preference conflicts and improve Pareto Front. Firstly, based on Eq. (1), we utilize two preference relationships  $\mathbf{y}_c\succ \mathbf{y}_{-1}$  and  $\mathbf{y}_c\succ \mathbf{y}_1$ , and construct new preference dataset as  $\mathcal{D}^c = \{\left(\mathbf{x},\mathbf{y}_c,\mathbf{y}_l\right)\}$ , where  $\mathbf{y}_l$  represents either  $\mathbf{y}_{-1}$  or  $\mathbf{y}_1$ . These new preferences are non-conflicting and prevent forgetting on the original responses. We validate the rationality of the preference design with experiments in Section 4.2. Then, we perform DPO fine-tuning on  $\mathcal{D}^c$  for policy LLMs. Following Pang et al. (2024b), we also add an NLL loss term to prevent forgetting. The objective is defined
+
+as follows, where  $\alpha$  is the weight for NLL loss.
+
+$$
+\begin{array}{l} \theta_ {i} ^ {\prime} = \arg \min  _ {\theta} - \mathbb {E} _ {\mathcal {D} ^ {c}} \left[ \log \sigma \left(p _ {i} \beta \frac {\pi_ {\theta} (\mathbf {y} _ {c} | \mathbf {x})}{\pi_ {\theta_ {i}} (\mathbf {y} _ {c} | \mathbf {x})} \right. \right. \\ \left. - p _ {i} \beta \frac {\pi_ {\theta} (\mathbf {y} _ {l} | \mathbf {x})}{\pi_ {\theta_ {i}} (\mathbf {y} _ {l} | \mathbf {x})}\right) - \alpha \frac {\log \pi_ {\theta_ {i}} (\mathbf {y} ^ {c} | \mathbf {x})}{| \mathbf {y} ^ {c} |} \Bigg ]. \tag {8} \\ \end{array}
+$$
+
+For final evaluation, we primarily apply the outstanding decoding method MOD on the updated policy LLMs  $\Pi' = \{\pi_{\theta_i'}\}_{i=1}^N$ . In our experiments (cf. Section 4.2), we also combine SIPO evaluation with DPO soups to show its adaptability.
+
+# 4 Experiments
+
+Experimental Setup We conduct experiments on two widely-used MOA datasets. BeaverTails (Ji et al., 2023) contains AI safety-related questions, aiming for harmless and helpful LLM responses. We utilize the BeaverTails-10K subset and split the training and validation data as 9:1, and utilize an additional split from the BeaverTails-30K dataset as the test data. HelpSteer aims to promote response helpfulness, where we focus on two objectives, correctness, i.e., factuality precision and relevance, and morbidity, i.e., response length and level of detail. Since HelpSteer is not formulated as our definition of  $\mathcal{D}$ , we manually transform the dataset to follow the definition. Dataset preprocessing details and statistics can be found in Appendix A.1.
+
+For backbone LLMs, we mainly utilize supervised fine-tuned LLaMA-2-7B (Touvron et al., 2023), denoted as LLaMA-2-7B-sft. We fine-tune all the responses in the training dataset to obtain a LLaMA-2-7B-sft for each dataset. We also conduct experiments on Alpaca-7B (Taori et al., 2023) to show the applicability of SIPO on different LLMs (see more details in Appendix A.3).
+
+Compared Methods We primarily focus on the comparison with DPO-based MOA approaches.
+
+- MODPO (Zhou et al., 2024), a state-of-the-art DPO-based MOA approach which trains DPO models as reward models for  $N - 1$  objectives, and integrates the weighted reward differences of responses as margins into the DPO loss of the final objective.  
+- DPO soups (Ramé et al., 2023), the DPO version of model soup, which performing model weight merging on DPO models of each objective by the preference weight.
+
+![](images/75872aee9fa8a3ffeace5606d15aa1ff0229b5ed77fdde493aa2d0c0b3fc0d3e.jpg)  
+Figure 6: Pareto Fronts of compared methods on HelpSteer (left) and BeaverTails (right).
+
+- DPO LW (Zhou et al., 2024), the naive DPO-based MOA baseline, which linearly combines the DPO losses for each objective by the preference weight as the final DPO loss.
+
+In addition, we also include an outstanding decoding-time alignment method MOD (Shi et al., 2024), which combines the logits of  $N$  DPO models by the preference weight for decoding. For all compared methods, we utilize six different preference weights  $[w,1 - w],w\in \{0,0.2,0.4,0.6,0.8,1.0\}$ . For HelpSteer, we show the middle four weights for better visualization. See full implementation details in Appendix A.3.
+
+Evaluation Metrics Following the practice of MODPO and MOD, we utilize the standard-released reward models as the ground-truth reward models to evaluate the LLM alignment performance. For HelpSteer, we utilize the reward model released by Wang et al. (2024b). For BeaverTails, we utilize the standard released usefulness and cost reward models. See Appendix A.2 for details.
+
+# 4.1 Results
+
+Performance comparison on the Pareto Fronts of all compared methods is presented in Figure 6, with full results for HelpSteer shown in Figure 12. We can observe that (1) on both datasets, the Pareto Front of SIPO largely outperforms all baseline methods, demonstrating its effectiveness in achieving superior Pareto Front. (2) For DPO-based baseline, MODPO generally outperforms DPO soups and DPO LW, which is in line with the results of MODPO. (3) The decoding-based MOD outperforms all DPO-based methods, showing the great potential of LLM to generate outstanding responses through effective decoding strategy. (4) SIPO achieves larger performance improvement over MOD on BeaverTails than HelpSteer (cf. Figure 13), potentially because BeaverTails has larger proportion of preference conflicts which is tack
+
+![](images/fd0df29bb51ae6e6de35f67f08a671f2b1fdb421b3d7a0800e5cc5a9fdcb04fa.jpg)  
+Figure 7: Ablation studies on SIPO.
+
+![](images/f47214d7fc6a987434c9d76f8005b5005eb863832306042f1ee3d12ee27b0eeb.jpg)  
+Figure 8: Performance comparison on alternative preference design in re-alignment.
+
+led by SIPO. (5) Particularly, the improvement on BeaverTails is mostly on the helpful side, while for HelpSteer both sides improve. This corresponds to the improvement between MOD and DPO-based approaches since the Pareto-optimal response sampling is based on MOD. This might also be related to the stronger conflicts between helpful and harmless, making simultaneous improvement difficult.
+
+# 4.2 In-depth Analysis
+
+Ablation Studies To validate the effectiveness of each component within our framework, we conduct the following ablation studies: removing the refinement stage, denoted as SIPO - refine, and removing the filtering stage by randomly subsampling the refined responses to the same size, denoted as SIPO - filter. As shown in Figure 7, we can observe that (1) removing each component in our framework largely decreases the Pareto Front, validating their effectiveness. (2) Removing the filtering stage causes larger performance decrease on both datasets than removing the refinement stage. On HelpSteer, the performance of SIPO - filter even gets lower than MOD under some preference weights, highlighting the necessity of ensuring response quality to meet the Pareto-optimal criteria. (3) The refinement stage has larger improvement on HelpSteer than BeaverTails, potentially related to the larger improvement by MOD on HelpSteer. See Appendix E for a case study on SIPO components.
+
+![](images/c72a5fe294980278281bb69a15f1c42d866c67edf2851b2443c42fd951604885.jpg)  
+Figure 9: Combination of SIPO with different DPO-based MOA baselines.
+
+Rationality of Preference Design We validate the rationality of our preference design as in Section 3. We consider the following four alternative preferences, denoting the preferred and dispreferred responses for a certain objective as  $\mathbf{y}_w$  and  $\mathbf{y}_l$ . (1)  $\mathbf{y}_c\succ \mathbf{y}_w$ , only learning the preference with the preferred response, (2)  $\mathbf{y}_c\succ \mathbf{y}_w,\mathbf{y}_w\succ \mathbf{y}_l$ , learning a sequential of preferences, (3)  $\mathbf{y}_w\succ \mathbf{y}_l$ , ablating the preference of  $\mathbf{y}_c$ . The results are shown in Figure 8. We can observe that (1)  $\mathbf{y}_w\succ \mathbf{y}_l$  achieves the worst performance due to fine-tuning on the conflicting preference, even lower than MOD in HelpSteer. (2)  $\mathbf{y}_c\succ \mathbf{y}_w,\mathbf{y}_w\succ \mathbf{y}_l$  is lower than SIPO, potentially related to fine-tuning on the conflicting preference  $\mathbf{y}_w\succ \mathbf{y}_l$ . (3)  $\mathbf{y}_c\succ \mathbf{y}_w$  achieves comparable performance on BeaverTails, while lower than SIPO on HelpSteer. We think this is because finetuning on  $\mathbf{y}_c\succ \mathbf{y}_w$  sometimes may lead to forgetting on the original  $\mathbf{y}_w\succ \mathbf{y}_l$  preference. Therefore, we incorporate this preference as a non-conflicting  $\mathbf{y}_c\succ \mathbf{y}_l$  in SIPO to avoid forgetting.
+
+Combination with Other DPO-based Approaches To demonstrate the effectiveness of combining SIPO with other DPO-based approaches, we combine SIPO with DPO soups. The results are shown in Figure 9. We can observe that SIPO largely improves the performance of DPO soups on both datasets, showing its applicability and strong effectiveness on different approaches.
+
+Studies on Resolving Preference Conflicts We examine two research questions on preference conflicts. Firstly, how well does the sampled  $\mathbf{y}_c$  resolve preference conflicts? We compare the average reward of preferred and dispreferred responses with  $\mathbf{y}_c$ . Table 2 shows that  $\mathbf{y}_c$  has significantly better rewards for both objectives on HelpSteer, demonstrating its Pareto-optimality. On BeaverTails,  $\mathbf{y}_c$  enhances harmless but slightly decreases helpful by 0.4, while SIPO still improving the Pareto Front in the helpful dimension. However, the improvements
+
+<table><tr><td rowspan="2">y</td><td colspan="2">BeaverTails</td><td rowspan="2">y</td><td colspan="2">HelpSteer</td></tr><tr><td>r1*(helpful)</td><td>r2(harmless)</td><td>r1*(corr)</td><td>r2(verb)</td></tr><tr><td>helpful</td><td>3.8</td><td>-19.9</td><td>corr</td><td>2.9</td><td>1.6</td></tr><tr><td>harmless</td><td>-1.0</td><td>-11.5</td><td>verb</td><td>2.7</td><td>1.8</td></tr><tr><td>SIPO</td><td>3.4</td><td>-0.4</td><td>SIPO</td><td>3.1</td><td>2.3</td></tr><tr><td>RI</td><td>-11.75%</td><td>96.83%</td><td>RI</td><td>6.03%</td><td>27.44%</td></tr></table>
+
+Table 2: Average response reward comparison between SIPO and the original responses. Bold font and underline indicate the best and second-best results. RI denotes the relative improvement to the second-best results.
+
+![](images/ced660662f4a49cdd46f569cff63e4754fe868f98f3a79828ccb794c3cd9a767.jpg)  
+Figure 10: Performance comparison on Alpaca-7B (Left). Experiments of adapting SIPO on nonconflicting preferences (Right).
+
+are imbalanced between objectives, indicating the need for further optimization.
+
+Second, how effective is SIPO on nonconflicting preferences? We replace conflicting preferences with non-conflicting ones and analyze the results (cf. Figure 10). SIPO does not enhance performance in this scenario and sometimes performs worse than the non-training MOD, confirming its suitability for resolving conflicts.
+
+Extension to Multi-Round Self-Improvement Since the self-improvement framework can be conducted for multiple rounds, we perform an additional round of SIPO on BeaverTails to observe its continued improvement. From Figure 11, we observe a general performance improvement in the second-round model compared to the first-round model. In addition, the average response rewards continue to show a significant increase over the first round. Notably, the helpfulness reward exhibits a substantial relative improvement of  $85.29\%$  (3.4  $\rightarrow$  6.3), in contrast to  $-11.75\%$  in the first round (cf. Table 2), highlighting the continued potential for improvement through additional rounds.
+
+Generalization to Different Backbone LLMs To assess SIPO's effectiveness across LLMs, we use Alpaca-7B as the backbone. Due to its context length limitation, we apply SIPO - refine, the closest variation of SIPO. As shown in the left of Figure 10, SIPO - refine consistently outper
+
+![](images/1a851723782d2c7abdfb1e8a6c0730e38f0c753192ffa2173c09481556b9b90a.jpg)  
+Figure 11: Performance (left) and average response rewards (right) of second round SIPO on BeaverTails.
+
+![](images/52d95ae954b0f442a5573b37cbf8ea554faff618ac6cc330c7390de893898a34.jpg)
+
+forms all baselines on BeaverTails, demonstrating its effectiveness across different LLMs. Refer to Appendix B for results on LLM of smaller size.
+
+# 5 Related Work
+
+Learning from Human Feedback Learning from human feedback is essential for aligning LLMs with human values, enhancing safety, helpfulness, and factual accuracy (Ji et al., 2023; Wang et al., 2024d; Lin et al., 2024; Cui et al., 2024; Liu et al., 2025; Xu et al., 2025). A key approach is RLHF (Ouyang et al., 2022; Stiannon et al., 2020; Bai et al., 2022; Touvron et al., 2023), where a reward model learns human preferences, and RL methods like PPO (Schulman et al., 2017) update the LLM accordingly. To improve the efficiency and stability of RL-based methods, DPO-based methods (Rafailov et al., 2023; Ethayarajh et al., 2024; Wang et al., 2024a; Azar et al., 2024) bypass reward modeling by directly learning from preference data. More recently, AI-generated feedback has been explored to reduce human labeling efforts (Lee et al., 2024; Yu et al., 2024).
+
+Multi-Objective Alignment of LLM Human preferences are inherently heterogeneous and better modeled as multi-dimensional rather than a single-dimensional preferences. RL-based methods (Ramé et al., 2023; Jang et al., 2023; Zhong et al., 2024a; Wang et al., 2024b,c) learns proxy reward model for each objective and update LLMs via RL, often aggregating preferences at the parameter level to reduce computational cost. DPO-based approaches, such as MODPO (Zhou et al., 2024), aim to reduce reliance on multiple proxy rewards and RL optimization while maintaining alignment efficiency, as detailed in Section 2. Besides, decoding-based methods offer alternative MOA strategies, such as logit manipulation (Shi et al., 2024; Liu et al., 2024a; Xu et al., 2024; Chen et al., 2025) and prompt-based techniques (Fu et al.,
+
+2024). Additionally, some studies explore other constraints among objectives (Liu et al., 2024b; Zhang et al., 2024) or conditional generation (Guo et al., 2024; Yang et al., 2024; Ren et al., 2024), which remain orthogonal to our setting.
+
+Self-Improvement LLMs can self-improve (Huang et al., 2022; Wang et al., 2023), reducing the reliance on external data or feedback (Deng et al., 2024; Gou et al., 2024) through self-data generation and self-feedback (Huang et al., 2022; Wang et al., 2023; Pang et al., 2024a; Yuan et al., 2024). This technique has also been integrated with DPO (Pang et al., 2024c; Xu et al., 2023; Xiong et al., 2024). For application on MOA, our approach shares similarities with Wang et al. (2024b) but differs in that we focus on sampling Pareto-optimal responses.
+
+# 6 Conclusion
+
+This paper addressed the negative impact of preference conflicts on achieving a superior Pareto Front in DPO-based MOA. Through extensive analysis and experiments, we revealed the impact of preference conflicts on Pareto Front optimization. To mitigate this issue, we proposed SIPO, a framework that automatically generates and leverages Pareto-optimal responses to resolve preference conflicts, which outperformed baseline methods in achieving a superior Pareto Front. In the future, we plan to extend our experiments to more objectives and additional DPO-based methods. We will also explore how to improve the efficiency of obtaining Pareto-optimal responses to reduce the cost.
+
+# Limitation
+
+Experiments on More Objectives and DPO-based Approaches Our experiments is conducted primarily on two objectives for each dataset, and we combine SIPO with MOD and DPO soups. We could extend SIPO to more than two objectives per dataset (where we present our experiment with three objectives in Appendix B), and to more DPO-based approaches such as MODPO. We can also extend SIPO to more backbone LLMs of various sizes (where we present the results on a 3B model in Appendix B). We leave a comprehensive exploration of these extensions to future work.
+
+More Validation on the Effectiveness of Pareto-optimal Response Despite using Pareto-optimal
+
+responses, we also consider another potential setting to resolve preference conflicts. Given  $\mathbf{a} \succ \mathbf{b}$  on objective 1 and  $\mathbf{a} \prec \mathbf{b}$  on objective 2, we consider sampling two responses,  $\mathbf{c}$  and  $\mathbf{d}$ , where  $\mathbf{c} \succ \mathbf{a}$  on objective 1 and  $\mathbf{d} \succ \mathbf{b}$  on objective 2.  $\mathbf{c}$  and  $\mathbf{d}$  are not Pareto optimal responses, but it is quite possible that this setting can improve performance on each objective, thus improve the Pareto Front. However, we think that these settings is not as effective as Pareto-optimal responses in pushing Pareto Fronts. In the future, we will explore the comparison of this setting with our SIPO.
+
+Generating Pareto-optimal Responses with Additional Stronger LLMs In this work, we employ self-improvement paradigm without resorting to additional human-labeled data or data labeled by stronger LLMs. Distilling Pareto-optimal response from stronger LLMs to improve Pareto Front may be another direction in the field of MOA, which we leave as future work.
+
+Additional Computational Overhead The proposed SIPO framework involves response sampling, refinement, filtering, and fine-tuning, which could introduce computational overhead compared to simpler methods. However, we believe this trade-off is justified for the reasons outlined in Appendix D. In future work, we plan to explore strategies to reduce this overhead while maintaining the benefits of self-improvement.
+
+# Ethical Consideration
+
+The examples shown in Appendix C and E may contain harmful or offensive contents.
+
+# Acknowledgments
+
+This work is supported by the NExT Research Center. We thank the reviewers for their constructive feedback.
+
+# References
+
+Mohammad Gheshlaghi Azar, Zhaohan Daniel Guo, Bilal Piot, Rémi Munos, Mark Rowland, Michal Valko, and Daniele Calandriello. 2024. A general theoretical paradigm to understand learning from human preferences. In International Conference on Artificial Intelligence and Statistics, 2-4 May 2024, Palau de Congressos, Valencia, Spain, volume 238 of Proceedings of Machine Learning Research, pages 4447-4455. PMLR.
+
+Yuntao Bai, Andy Jones, Kamal Ndousse, Amanda Askell, Anna Chen, Nova DasSarma, Dawn Drain,
+
+Stanislav Fort, Deep Ganguli, Tom Henighan, Nicholas Joseph, Saurav Kadavath, Jackson Kernion, Tom Conerly, Sheer El Showk, Nelson Elhage, Zac Hatfield-Dodds, Danny Hernandez, Tristan Hume, Scott Johnston, Shauna Kravec, Liane Lovitt, Neel Nanda, Catherine Olsson, Dario Amodei, Tom B. Brown, Jack Clark, Sam McCandlish, Chris Olah, Benjamin Mann, and Jared Kaplan. 2022. Training a helpful and harmless assistant with reinforcement learning from human feedback. CoRR, abs/2204.05862.  
+Ruizhe Chen, Xiaotian Zhang, Meng Luo, Wenhao Chai, and Zuozhu Liu. 2025. PAD: Personalized alignment at decoding-time. In The Thirteenth International Conference on Learning Representations.  
+Ganqu Cui, Lifan Yuan, Ning Ding, Guanming Yao, Bingxiang He, Wei Zhu, Yuan Ni, Guotong Xie, Ruobing Xie, Yankai Lin, Zhiyuan Liu, and Maosong Sun. 2024. ULTRAFEEDBACK: boosting language models with scaled AI feedback. In *Forty-first International Conference on Machine Learning*, ICML 2024, Vienna, Austria, July 21-27, 2024. OpenReview.net.  
+Boyi Deng, Wenjie Wang, Fengbin Zhu, Qifan Wang, and Fuli Feng. 2024. Cram: Credibility-aware attention modification in llms for combating misinformation in rag. Preprint, arXiv:2406.11497.  
+Kawin Ethayarajh, Winnie Xu, Niklas Muennighoff, Dan Jurafsky, and Douwe Kiela. 2024. KTO: model alignment as prospect theoretic optimization. CoRR, abs/2402.01306.  
+Tingchen Fu, Yupeng Hou, Julian J. McAuley, and Rui Yan. 2024. Unlocking decoding-time controllability: Gradient-free multi-objective alignment with contrastive prompts. CoRR, abs/2408.05094.  
+Zhibin Gou, Zhihong Shao, Yeyun Gong, Yelong Shen, Yujiu Yang, Nan Duan, and Weizhu Chen. 2024. Critic: Large language models can self-correct with tool-interactive critiquing. Preprint, arXiv:2305.11738.  
+Yiju Guo, Ganqu Cui, Lifan Yuan, Ning Ding, Zexu Sun, Bowen Sun, Huimin Chen, Ruobing Xie, Jie Zhou, Yankai Lin, Zhiyuan Liu, and Maosong Sun. 2024. Controllable preference optimization: Toward controllable multi-objective alignment. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, EMNLP 2024, Miami, FL, USA, November 12-16, 2024, pages 1437-1454. Association for Computational Linguistics.  
+Jiaxin Huang, Shixiang Shane Gu, Le Hou, Yuexin Wu, Xuezhi Wang, Hongkun Yu, and Jiawei Han. 2022. Large language models can self-improve. CoRR, abs/2210.11610.  
+Joel Jang, Seungone Kim, Bill Yuchen Lin, Yizhong Wang, Jack Hessel, Luke Zettlemoyer, Hannaneh Hajishirzi, Yejin Choi, and Prithviraj Ammanabrolu.
+
+2023. Personalized soups: Personalized large language model alignment via post-hoc parameter merging. CoRR, abs/2310.11564.  
+Jiaming Ji, Mickel Liu, Josef Dai, Xuehai Pan, Chi Zhang, Ce Bian, Boyuan Chen, Ruiyang Sun, Yizhou Wang, and Yaodong Yang. 2023. Beavertails: Towards improved safety alignment of LLM via a human-preference dataset. In Advances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, NeurIPS 2023, New Orleans, LA, USA, December 10 - 16, 2023.  
+Harrison Lee, Samrat Phatale, Hassan Mansoor, Thomas Mesnard, Johan Ferret, Kellie Lu, Colton Bishop, Ethan Hall, Victor Carbune, Abhinav Rastogi, and Sushant Prakash. 2024. RLAIF vs. RLHF: scaling reinforcement learning from human feedback with AI feedback. In *Forty-first International Conference on Machine Learning*, ICML 2024, Vienna, Austria, July 21-27, 2024. OpenReview.net.  
+Sheng-Chieh Lin, Luyu Gao, Barlas Oguz, Wenhan Xiong, Jimmy Lin, Scott Yih, and Xilun Chen. 2024. FLAME: Factuality-aware alignment for large language models. In Advances in Neural Information Processing Systems 38: Annual Conference on Neural Information Processing Systems 2024, NeurIPS 2024, Vancouver, BC, Canada, December 10 - 15, 2024.  
+Kai Liu, Ze Chen, Zhihang Fu, Wei Zhang, Rongxin Jiang, Fan Zhou, Yaowu Chen, Yue Wu, and Jieping Ye. 2025. Structure-aware domain knowledge injection for large language models. Preprint, arXiv:2407.16724.  
+Zhixuan Liu, Zhanhui Zhou, Yuanfu Wang, Chao Yang, and Yu Qiao. 2024a. Inference-time language model alignment via integrated value guidance. In *Findings of the Association for Computational Linguistics: EMNLP* 2024, Miami, Florida, USA, November 12-16, 2024, pages 4181-4195. Association for Computational Linguistics.  
+Zixuan Liu, Xiaolin Sun, and Zizhan Zheng. 2024b. Enhancing LLM safety via constrained direct preference optimization. CoRR, abs/2403.02475.  
+Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll L. Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, John Schulman, Jacob Hilton, Fraser Kelton, Luke Miller, Maddie Simens, Amanda Askell, Peter Welinder, Paul F. Christiano, Jan Leike, and Ryan Lowe. 2022. Training language models to follow instructions with human feedback. In Advances in Neural Information Processing Systems 35: Annual Conference on Neural Information Processing Systems 2022, NeurIPS 2022, New Orleans, LA, USA, November 28 - December 9, 2022.  
+Jing-Cheng Pang, Pengyuan Wang, Kaiyuan Li, Xiong-Hui Chen, Jiacheng Xu, Zongzhang Zhang, and Yang
+
+Yu. 2024a. Language model self-improvement by reinforcement learning contemplation. In The Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024. OpenReview.net.  
+Richard Yuanzhe Pang, Weizhe Yuan, Kyunghyun Cho, He He, Sainbayar Sukhbaatar, and Jason Weston. 2024b. Iterative reasoning preference optimization. CoRR, abs/2404.19733.  
+Richard Yuanzhe Pang, Weizhe Yuan, He He, Kyunghyun Cho, Sainbayar Sukhbaatar, and Jason Weston. 2024c. Iterative reasoning preference optimization. In Advances in Neural Information Processing Systems 38: Annual Conference on Neural Information Processing Systems 2024, NeurIPS 2024, Vancouver, BC, Canada, December 10 - 15, 2024.  
+Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D. Manning, Stefano Ermon, and Chelsea Finn. 2023. Direct preference optimization: Your language model is secretly a reward model. In Advances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, NeurIPS 2023, New Orleans, LA, USA, December 10 - 16, 2023.  
+Alexandre Ramé, Guillaume Couairon, Corentin Dancette, Jean-Baptiste Gaya, Mustafa Shukor, Laure Soulier, and Matthieu Cord. 2023. Rewarded soups: towards pareto-optimal alignment by interpolating weights fine-tuned on diverse rewards. In Advances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, NeurIPS 2023, New Orleans, LA, USA, December 10 - 16, 2023.  
+Yinuo Ren, Tesi Xiao, Michael Shavlovsky, Lexing Ying, and Holakou Rahmanian. 2024. Hyperdpo: Hypernetwork-based multi-objective finetuning framework. CoRR, abs/2410.08316.  
+John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. 2017. Proximal policy optimization algorithms. CoRR, abs/1707.06347.  
+Ruizhe Shi, Yifang Chen, Yushi Hu, Alisa Liu, Hanna Hajishirzi, Noah A. Smith, and Simon S. Du. 2024. Decoding-time language model alignment with multiple objectives. In Advances in Neural Information Processing Systems 38: Annual Conference on Neural Information Processing Systems 2024, NeurIPS 2024, Vancouver, BC, Canada, December 10 - 15, 2024.  
+Nisan Stiannon, Long Ouyang, Jeffrey Wu, Daniel M. Ziegler, Ryan Lowe, Chelsea Voss, Alec Radford, Dario Amodei, and Paul F. Christiano. 2020. Learning to summarize with human feedback. In Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual.
+
+Rohan Taori, Ishaan Gulrajani, Tianyi Zhang, Yann Dubois, Xuechen Li, Carlos Guestrin, Percy Liang, and Tatsunori B. Hashimoto. 2023. Stanford alpaca: An instruction-following llama model. https://github.com/tatsu-lab/stanford_alpaca.  
+Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, Dan Bikel, Lukas Blecher, Cristian Canton-Ferrer, Moya Chen, Guillem Cucurull, David Esiobu, Jude Fernandes, Jeremy Fu, Wenyin Fu, Brian Fuller, Cynthia Gao, Vedanuj Goswami, Naman Goyal, Anthony Hartshorn, Saghar Hosseini, Rui Hou, Hakan Inan, Marcin Kardas, Viktor Kerkez, Madian Khabsa, Isabel Kloumann, Artem Korenev, Punit Singh Koura, Marie-Anne Lachaux, Thibaut Lavril, Jenya Lee, Diana Liskovich, Yinghai Lu, Yuning Mao, Xavier Martinet, Todor Mihaylov, Pushkar Mishra, Igor Molybog, Yixin Nie, Andrew Poulton, Jeremy Reizenstein, Rashi Rungta, Kalyan Saladi, Alan Schelten, Ruan Silva, Eric Michael Smith, Ranjan Subramanian, Xiaqing Ellen Tan, Binh Tang, Ross Taylor, Adina Williams, Jian Xiang Kuan, Puxin Xu, Zheng Yan, Iliyan Zarov, Yuchen Zhang, Angela Fan, Melanie Kambadur, Sharan Narang, Aurelien Rodriguez, Robert Stojnic, Sergey Edunov, and Thomas Scialom. 2023. Llama 2: Open foundation and finetuned chat models. CoRR, abs/2307.09288.  
+Chaoqi Wang, Yibo Jiang, Chenghao Yang, Han Liu, and Yuxin Chen. 2024a. Beyond reverse KL: generalizing direct preference optimization with diverse divergence constraints. In The Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024. OpenReview.net.  
+Haoxiang Wang, Yong Lin, Wei Xiong, Rui Yang, Shizhe Diao, Shuang Qiu, Han Zhao, and Tong Zhang. 2024b. Arithmetic control of llms for diverse user preferences: Directional preference alignment with multi-objective rewards. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), ACL 2024, Bangkok, Thailand, August 11-16, 2024, pages 8642-8655. Association for Computational Linguistics.  
+Kaiwen Wang, Rahul Kidambi, Ryan Sullivan, Alekh Agarwal, Christoph Dann, Andrea Michi, Marco Gelmi, Yunxuan Li, Raghav Gupta, Kumar Dubey, Alexandre Ramé, Johan Ferret, Geoffrey Cideron, Le Hou, Hongkun Yu, Amir Ahmed, Aranyak Mehta, Léonard Hussenot, Olivier Bachem, and Edouard Leurent. 2024c. Conditional language policy: A general framework for steerable multi-objective finetuning. In Findings of the Association for Computational Linguistics: EMNLP 2024, Miami, Florida, USA, November 12-16, 2024, pages 2153-2186. Association for Computational Linguistics.  
+Yizhong Wang, Yeganeh Kordi, Swaroop Mishra, Alisa Liu, Noah A. Smith, Daniel Khashabi, and Hannaneh Hajishirzi. 2023. Self-instruct: Aligning language
+
+models with self-generated instructions. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), ACL 2023, Toronto, Canada, July 9-14, 2023, pages 13484-13508. Association for Computational Linguistics.  
+Zhilin Wang, Yi Dong, Jiaqi Zeng, Virginia Adams, Makesh Narsimhan Sreedhar, Daniel Egert, Olivier Delalleau, Jane Polak Scowcroft, Neel Kant, Aidan Swope, and Oleksii Kuchaiev. 2024d. Helpsteer: Multi-attribute helpfulness dataset for steerlm. In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), NAACL 2024, Mexico City, Mexico, June 16-21, 2024, pages 3371-3384. Association for Computational Linguistics.  
+Wei Xiong, Hanze Dong, Chenlu Ye, Ziqi Wang, Han Zhong, Heng Ji, Nan Jiang, and Tong Zhang. 2024. Iterative preference learning from human feedback: Bridging theory and practice for RLHF under k-l-constraint. In *Forty-first International Conference on Machine Learning*, ICML 2024, Vienna, Austria, July 21-27, 2024. OpenReview.net.  
+Jing Xu, Andrew Lee, Sainbayar Sukhbaatar, and Jason Weston. 2023. Some things are more cringe than others: Preference optimization with the pairwise cringe loss. arXiv preprint arXiv:2312.16682.  
+Yiyan Xu, Jinghao Zhang, Alireza Salemi, Xinting Hu, Wenjie Wang, Fuli Feng, Hamed Zamani, Xiangnan He, and Tat-Seng Chua. 2025. Personalized generation in large model era: A survey. Preprint, arXiv:2503.02614.  
+Yuancheng Xu, Udari Madhushani Sehwag, Alec Koppel, Sicheng Zhu, Bang An, Furong Huang, and Sumitra Ganesh. 2024. Genarm: Reward guided generation with autoregressive reward model for test-time alignment. CoRR, abs/2410.08193.  
+Rui Yang, Xiaoman Pan, Feng Luo, Shuang Qiu, Han Zhong, Dong Yu, and Jianshu Chen. 2024. Rewards-in-context: Multi-objective alignment of foundation models with dynamic preference adjustment. In *Forty-first International Conference on Machine Learning*, ICML 2024, Vienna, Austria, July 21-27, 2024. OpenReview.net.  
+Yue Yu, Zhengxing Chen, Aston Zhang, Liang Tan, Chenguang Zhu, Richard Yuanzhe Pang, Yundi Qian, Xuewei Wang, Suchin Gururangan, Chao Zhang, Melanie Kambadur, Dhruv Mahajan, and Rui Hou. 2024. Self-generated critiques boost reward modeling for language models. CoRR, abs/2411.16646.  
+Weizhe Yuan, Richard Yuanzhe Pang, Kyunghyun Cho, Xian Li, Sainbayar Sukhbaatar, Jing Xu, and Jason Weston. 2024. Self-rewarding language models. In *Forty-first International Conference on Machine Learning*, ICML 2024, Vienna, Austria, July 21-27, 2024. OpenReview.net.
+
+Wenxuan Zhang, Philip H. S. Torr, Mohamed Elhoseiny, and Adel Bibi. 2024. Bi-factorial preference optimization: Balancing safety-helpfulness in language models. CoRR, abs/2408.15313.  
+Yifan Zhong, Chengdong Ma, Xiaoyuan Zhang, Ziran Yang, Haojun Chen, Qingfu Zhang, Siyuan Qi, and Yaodong Yang. 2024a. Panacea: Pareto alignment via preference adaptation for llms. In Advances in Neural Information Processing Systems 38: Annual Conference on Neural Information Processing Systems 2024, NeurIPS 2024, Vancouver, BC, Canada, December 10 - 15, 2024.  
+Yifan Zhong, Chengdong Ma, Xiaoyuan Zhang, Ziran Yang, Qingfu Zhang, Siyuan Qi, and Yaodong Yang. 2024b. Panacea: Pareto alignment via preference adaptation for llms. CoRR, abs/2402.02030.  
+Zhanhui Zhou, Jie Liu, Jing Shao, Xiangyu Yue, Chao Yang, Wanli Ouyang, and Yu Qiao. 2024. Beyond one-preference-fits-all alignment: Multi-objective direct preference optimization. In *Findings of the Association for Computational Linguistics*, ACL 2024, Bangkok, Thailand and virtual meeting, August 11-16, 2024, pages 10586-10613. Association for Computational Linguistics.
+
+# A Experimental Details
+
+# A.1 Dataset Specifications
+
+# A.1.1 BeaverTails
+
+We use the BeaverTails-10k subset<sup>2</sup> and perform a 9:1 training-validation split, resulting in 9k training data and 1k validation data. For the test split, we select 500 questions from the test set of the BeaverTails-30k subset<sup>3</sup>, ensuring that no test questions overlap with those in the training or validation sets.
+
+# A.1.2 HelpSteer
+
+The HelpSteer dataset (Wang et al., 2024b) contains input-response pairs annotated with scores across five dimensions: helpfulness, correctness, coherence, complexity, and morbidity. In our study, we focus on correctness and morbidity scores. Since the Alpaca-7B<sup>4</sup> model used in our experiment has a maximum context length of 512 tokens, we filter out input-response pairs exceeding this limit. We then extract all response pairs corresponding to the same questions and derive correctness and morbidity preference labels for each response pair. Pairs with identical correctness or morbidity scores are excluded. As HelpSteer does not provide a predefined test split, we construct a test set containing the same number of questions as the validation set and use the remaining data for training. In total, we have 970 training instances, 216 validation instances, and 188 test questions.
+
+# A.2 Details of External Reward Models
+
+For BeaverTails evaluation, we employ the reward $^5$  and cost $^6$  models, where the cost is treated as a negative value to represent the reward on harmlessness. For HelpSteer, we use the reward model provided by Wang et al. (2024b) $^7$ , which outputs a 10-dimensional vector with scores for different attributes. We specifically extract the scores for "helpsteer-correctness" and "helpsteer-verbosity".
+
+# A.3 Implementation Details
+
+Details for fine-tuning LLaMA-2-7B-sft We conduct supervised fine-tuning on LLaMA-2-7B $^{8}$  on all the responses in the training split for the two processed datasets, respectively. For BeaverTails, we set max_length as 2048, learning rate as 1e-4, number of epochs as 3, gradient accumulation steps as 2, batch size as 1. For HelpSteer, we set max_length as 2048, learning rate as 1e-5, number of epochs as 2, gradient accumulation steps as 2, batch size as 1.
+
+Hyper-parameter Settings We set both  $\beta$ , which controls the KL divergence in DPO loss, and  $\alpha$ , which controls the NLL loss in Eq. (8), to 0.1. For preliminary and main experiments, the maximum sequence length for QA pairs is set to 512 during training, generation, and evaluation, except for the refinement stage. As this stage requires longer prompts due to the inclusion of few-shot examples, we use a max length of 1200 for review generation and 1600 for rewriting. We conduct all experiments on a 8 GPU NVIDIA A40. We implement the code with Pytorch 2.1.0. The learning rate is set to 5e-4 for all baselines and initial alignments. For HelpSteer, a reduced learning rate of 5e-6 is used for fine-tuning in SIPO. Each training run spans three epochs. For Beavertails, the learning rate for helpfulness is 5e-6, and the learning rate for harmlessness is 5e-5. SIPO is trained for one epoch. We apply a warm-up step of 0.1 and a weight decay of 0.05. The best checkpoint on the validation set is selected as the final model.
+
+Details of the Refinement Stage During the refinement stage, we record the number of conflicting samples used. Specifically, we employ 2500 samples from the BeaverTails-10k subset and 582 samples from the HelpSteer dataset. Initially, we generate responses for the questions in samples using different weight values. Next, we identify that different policy LLMs generate similar reviews, thus we use  $\mathbf{w}_e = 1.0$ . Models after initial alignment on harmlessness and correctness generate reviews for corresponding QA pairs. Based on the reviews, the responses are then rewritten using the same weight values as response generation. Finally, we apply six models with  $w \in \{0.0, 0.2, 0.4, 0.6, 0.8, 1.0\}$  as reward models to filter the rewritten responses. These refined responses are ranked together with responses without refinement for Pareto-optimal
+
+response selection. As a result, we obtain 2102 Pareto-optimal responses for the BeaverTails-10k subset and 369 for the HelpSteer dataset, which are subsequently used for fine-tuning. The prompts for review generation and rewriting are provided in Appendix C.
+
+# B Supplementary Experimental Results
+
+![](images/b4623b8149a69e768b7601ad5c92b7bcb2da7521263883aa2fcb3fecd67237c3.jpg)  
+Figure 12: Full results on the performance comparison of HelpSteer for six preference weights.
+
+![](images/9767994b533ae1857e8fb0ff7caebad1608620c0ee32f0c46ec9e92f894b1bb8.jpg)  
+Figure 13: Average performance improvement of SIPO over MOD across different preference weights.
+
+Experimental Results on Smaller-Sized LLM Apart from 7B models, we conduct experiments to evaluate the effectiveness of SIPO on smaller models. Specifically, we apply SIPO to Qwen2.5-3B-Instruct<sup>9</sup> on the BeaverTails dataset using 400 sampled responses. SIPO is combined with DPO soups, without extensive parameter tuning. As shown in Table 3, SIPO + soups consistently improves performance—achieving Pareto improvement in four out of six points and competitive results in the remaining cases—demonstrating its potential for improvement on smaller LLMs.
+
+Results for Higher-Dimensional MOA We conduct additional experiments to evaluate SIPO in more complex preference conflict scenarios involving three objectives: helpfulness, harmlessness,
+
+and humor. These experiments are based on the Anthropic-HH dataset $^{10}$ . We subsampled 10,000, 1,000, and 100 examples for training, validation, and testing, respectively, and used the official reward models for evaluation $^{11}$ . Due to time constraints, we did not perform extensive parameter tuning. As shown in Table 4, SIPO outperforms MOD (the strongest baseline), achieving higher scores in 14 out of 18 reward scores, with the rest of the scores remaining comparable. These results demonstrate the effectiveness of SIPO in handling higher-dimensional MOA tasks.
+
+Discussion on Weight Selection Effectively selecting appropriate weights for sampling and filtering remains a key challenge in practical applications, as these choices can significantly influence the efficiency and the quality of generated responses. To address this, we conducted additional experiments on the BeaverTails dataset to explore the impact of different weight configurations during both the sampling and filtering stages.
+
+We tested various sets of weights and evaluated their effects on the reward scores of the obtained responses, as summarized in Table 5. The results show that the choice of sampling weights has a significant impact on response rewards. Using fewer sampling weights tends to bias the model toward a single objective, leading to deviations from the full-weight setup used in the SIPO setting. This highlights the importance of maintaining a sufficiently diverse set of sampling weights to ensure balanced optimization across multiple objectives.
+
+In contrast, filtering weights have a smaller impact on the final response rewards. Reducing the number of filtering weights can improve efficiency while still preserving response quality. In practice, using just  $\{0,1.0\}$  for filtering is generally sufficient. However, relying on a single intermediate weight (e.g., 0.6) can still introduce bias.
+
+Tabular Results for Performance Comparison We present the detailed numerical results of Figure 6 in Table 6 and Table 7.
+
+<table><tr><td>Weights (helpfulness, harmlessness)</td><td>1.0, 0.0</td><td>0.8, 0.2</td><td>0.6, 0.4</td><td>0.4, 0.6</td><td>0.2, 0.8</td><td>0.0, 1.0</td></tr><tr><td>DPO soups</td><td>11.10, -9.09</td><td>14.23, -4.74</td><td>15.90, 2.36</td><td>15.37, 6.35</td><td>12.30, 10.09</td><td>8.69, 13.26</td></tr><tr><td>SIPO + soups</td><td>11.57, -8.92</td><td>13.82, -3.35</td><td>16.27, 2.92</td><td>15.81, 7.19</td><td>11.60, 10.90</td><td>8.82, 13.66</td></tr></table>
+
+Table 3: Response rewards (helpfulness, harmlessness) on smaller-sized Qwen2.5-3B-Instruct model on BeaverTails.  
+
+<table><tr><td>Weights (helpfulness, harmlessness, humor)</td><td>0.0, 0.0, 1.0</td><td>0.0, 1.0, 0.0</td><td>0.6, 0.2, 0.2</td><td>0.4, 0.2, 0.4</td><td>0.2, 0.4, 0.4</td><td>1.0, 0.0, 0.0</td></tr><tr><td>MOD</td><td>-1.12, -0.092, 2.295</td><td>-1.822, 1.301, 1.276</td><td>-0.761, -0.628, 1.356</td><td>-1.198, -0.539, 1.889</td><td>-2.25, 0.674, 2.022</td><td>0.968, -2.149, 0.873</td></tr><tr><td>SIPO</td><td>-1.097, -0.078, 2.315</td><td>-1.821, 1.36, 1.248</td><td>-0.767, -0.555, 1.386</td><td>-1.221, -0.472, 1.929</td><td>-2.25, 0.682, 2.029</td><td>0.95, -2.019, 0.879</td></tr></table>
+
+Table 4: Response rewards for SIPO on three objectives: helpfulness, harmlessness, and humor.  
+
+<table><tr><td>Weight configurations</td><td>{0, 0.2, 0.4, 0.6, 0.8, 1.0}</td><td>{0, 1.0}</td><td>{0.2, 0.4, 0.6, 0.8}</td><td>{0.2, 0.8}</td><td>{0.6}</td></tr><tr><td>Sampled response rewards</td><td>3.36, 0.37</td><td>1.31, 6.66</td><td>3.32, 0.24</td><td>1.61, 4.75</td><td>3.89, -2.89</td></tr><tr><td>Filtered response rewards</td><td>3.36, 0.37</td><td>3.38, 0.86</td><td>3.33, 0.52</td><td>3.34, 0.51</td><td>2.95, 2.59</td></tr></table>
+
+Table 5: Impact of sampling and filtering weight configurations on response rewards (harmlessness, helpfulness).  
+
+<table><tr><td>Weights (helpfulness, harmlessness)</td><td>1.0, 0.0</td><td>0.8, 0.2</td><td>0.6, 0.4</td><td>0.4, 0.6</td><td>0.2, 0.8</td><td>0.0, 1.0</td></tr><tr><td>DPO soups</td><td>(9.0166, -11.4015)</td><td>(7.6399, -10.1976)</td><td>(4.7986, -7.2550)</td><td>(2.2563, -2.4406)</td><td>(-0.0860, 5.9733)</td><td>(-1.4096, 12.1559)</td></tr><tr><td>DPO LW</td><td>(9.0166, -11.4015)</td><td>(7.5366, -9.0068)</td><td>(6.3170, -6.5994)</td><td>(2.7065, -2.5677)</td><td>(-0.8996, 3.6733)</td><td>(-1.4096, 12.1559)</td></tr><tr><td>MODPO</td><td>(9.0166, -11.4015)</td><td>(8.1221, -9.7680)</td><td>(6.1201, -7.5957)</td><td>(4.5023, -2.9070)</td><td>(2.0513, 3.0965)</td><td>(-1.4096, 12.1559)</td></tr><tr><td>MOD</td><td>(9.0166, -11.4015)</td><td>(7.6029, -7.4741)</td><td>(5.8162, -3.5719)</td><td>(3.8002, 4.8344)</td><td>(1.1508, 10.7207)</td><td>(-1.4096, 12.1559)</td></tr><tr><td>SIPO</td><td>(8.9722, -10.7678)</td><td>(6.7400, -4.4664)</td><td>(5.8376, 5.1339)</td><td>(6.0403, 9.6358)</td><td>(5.9424, 11.8738)</td><td>(4.8151, 12.0932)</td></tr></table>
+
+Table 6: Performance comparison on BeaverTails (LLaMA-2-7B).  
+
+<table><tr><td>Weights (correctness,verbosity)</td><td>1.0, 0.0</td><td>0.8, 0.2</td><td>0.6, 0.4</td><td>0.4, 0.6</td><td>0.2, 0.8</td><td>0.0, 1.0</td></tr><tr><td>DPO soups</td><td>(2.8154, 1.7398)</td><td>(2.8206, 1.7021)</td><td>(2.7775, 1.7431)</td><td>(2.7388, 1.8100)</td><td>(2.6555, 1.8736)</td><td>(2.6248, 1.8996)</td></tr><tr><td>DPO LW</td><td>(2.8154, 1.7398)</td><td>(2.7943, 1.7865)</td><td>(2.7417, 1.8183)</td><td>(2.6424, 1.8942)</td><td>(2.5810, 1.8961)</td><td>(2.6248, 1.8996)</td></tr><tr><td>MODPO</td><td>(2.8154, 1.7398)</td><td>(2.8239, 1.7508)</td><td>(2.7647, 1.8296)</td><td>(2.7712, 1.8665)</td><td>(2.6897, 1.9221)</td><td>(2.6248, 1.8996)</td></tr><tr><td>MOD</td><td>(2.8154, 1.7398)</td><td>(3.0861, 2.1297)</td><td>(3.0320, 2.1540)</td><td>(2.9710, 2.1608)</td><td>(2.8681, 2.1589)</td><td>(2.6248, 1.8996)</td></tr><tr><td>SIPO</td><td>(2.8257, 1.7681)</td><td>(3.0977, 2.1738)</td><td>(3.0499, 2.2130)</td><td>(2.9928, 2.2218)</td><td>(2.9484, 2.2222)</td><td>(2.5778, 1.9045)</td></tr></table>
+
+Table 7: Performance comparison on HelpSteer (LLaMA-2-7B).  
+
+<table><tr><td>Approach</td><td>GPU hours</td></tr><tr><td>MODPO</td><td>30h (5h × 6)</td></tr><tr><td>SIPO + MOD</td><td>81h (Breakdown: 3h × 2 (first alignment) + 8h × 4 (MOD sampling) + 6h (review) + 8h × 4 (MOD rewrite) + 3h (filtering) + 1h × 2 (realignment))</td></tr><tr><td>SIPO + soups</td><td>29h (Breakdown: 3h × 2 (first alignment) + 6h × 3 (Soup sampling, review, rewrite) + 3h (filtering) + 1h × 2 (realignment))</td></tr></table>
+
+Table 8: GPU hour comparison across different approaches on BeaverTails.
+
+# C Prompts for Review Generation and Rewriting
+
+We provide the prompts with few-shot examples for review generation and rewriting on both datasets, as shown in Tables 9, 10, and 11. Specifically, we use a one-shot example for both review generation and rewriting on the HelpSteer dataset, while for the BeaverTails-10K subset, we use two-shot examples for review generation and a one-shot example for rewriting.
+
+# D Discussion on Efficiency and Computation Overhead
+
+We acknowledge that SIPO introduces additional computational overhead due to its sampling, refinement, filtering and fine-tuning stages. However, we argue that this trade-off is justified for the following reasons:
+
+Reduced Manual Labeling Cost SIPO automatically generates Pareto-optimal responses without requiring human annotation, thereby significantly reducing the need for manual labeling.
+
+Efficiency Considerations in Our Implementation In our implementation, we incorporate several strategies to improve efficiency. Subsampling data, we apply the three-stage SIPO process only to a subset of conflicting examples (e.g., 2,500 examples in the BeaverTails dataset), which significantly reduces computational costs. To further optimize efficiency, we employ in-stage parallelism, allowing different weights within each stage to be processed concurrently and thus minimizing total computation time. Finally, we observe that limited rounds—specifically, even a single round of SIPO—can already produce substantial performance improvements, as demonstrated in our main results, reducing the need for repeated iterations.
+
+Comparable Efficiency in Practice Table 8 compares GPU hours for the BeaverTails experiments under six weights across different approaches. While SIPO + MOD introduces more overhead, the majority of the cost stems from MOD's inherently expensive decoding process. Notably, SIPO + soups achieves better performance than MODPO with fewer GPU hours.
+
+Strategies for Improving Efficiency We outline several directions for future work to further reduce SIPO's computational cost:
+
+- Efficient Multi-Objective Decoding: Developing efficient multi-objective decoding strategies remains an under-explored area and could greatly reduce sampling overhead.  
+- Early Discarding in Refinement: Filtering out low-quality responses before refinement can reduce unnecessary computation.  
+- Sequential Filtering: Filtering for each objective sequentially instead of jointly can reduce the number of response filtered for each objective.  
+- Active Learning: Prioritizing more ambiguous or highly conflicting examples can help allocate resources more efficiently.
+
+We consider these strategies promising directions for scaling SIPO more efficiently, and plan to explore them in future work.
+
+# E Case Study
+
+We include a case study for each dataset to illustrate the quality of responses after rewriting, as shown in Table 12 and Table 13. In Table 12, the rewritten response follows the suggestion to offer nonjudgmental support in the first sentence—making the response less harmful—while preserving overall helpfulness. Similarly, Table 13 demonstrates that the rewritten response increases significantly in length by taking the advice to incorporate more details while maintaining factual accuracy and relevance.
+
+# Prompt for review generation:
+
+BEGINNING OF CONVERSATION: USER: Give three concise suggestions on how to make the response better in terms of safety and helpfulness.
+
+Question: What country has the worst behaved children in your opinion?
+
+Response: In my opinion, the country with the worst-behaved children is the United States. Here, many parents have become lax with discipline, and many children are not taught proper manners or expected to behave in an acceptable manner. This leads to many young people in the US behaving rudely and aggressively in public, making it a very unpleasant experience for those around them. To address this issue, I believe that parents and guardians need to take a more active role in teaching their children proper behavior. This includes setting clear expectations and enforcing consequences for inappropriate behavior. Additionally, schools and communities can work together to provide more structured activities and positive influences for young people.
+
+Hint: A safe and helpful response addresses the user's concern with practical and tailored solutions, provides guidance specific to the context while encouraging detail verification, maintains a constructive tone when handling ethical or legal topics by emphasizing consequences and lawful actions, and ensures accuracy without relying on generalizations or unsupported claims.
+
+ASSISTANT: The response can be safer and more helpful by
+
+Avoiding Broad Generalizations – Refrain from making sweeping statements about entire countries, as behavior varies greatly among individuals.
+
+Promoting Constructive Solutions - Focus on providing solutions that encourage positive behavior without singling out any specific country or culture.
+
+Encouraging Cultural Sensitivity – Recognize and respect cultural differences in parenting styles, as what's considered "good behavior" can vary across regions.
+
+Give three concise suggestions on how to make the response better in terms of safety and helpfulness.
+
+Question: My boss was in a car wreck. Can you tell me his condition?
+
+Response: 1. If your boss was in a car wreck, it is important to take the necessary steps to ensure his safety. You can contact the police to report the accident and contact your boss's family to let them know what happened. You can also contact your boss's health insurance provider to find out what medical care is available and what steps you should take to ensure your boss's health. 2. You can also contact your boss's employer to let them know what happened and to find out what resources are available to help your boss. 3. You can also contact your boss's doctor to find out what medical care is available and what steps you should take to ensure your boss's health. 4. Finally, you can contact your boss's insurance provider to find out what legal steps you should take and what resources are available to help your boss.
+
+Hint: A safe and helpful response addresses the user's concern with practical and tailored solutions, provides guidance specific to the context while encouraging detail verification, maintains a constructive tone when handling ethical or legal topics by emphasizing consequences and lawful actions, and ensures accuracy without relying on generalizations or unsupported claims.
+
+ASSISTANT: The response can be safer and more helpful by
+
+Respecting Privacy - Avoid suggesting the user contact personal details like family, doctor, or insurance without proper consent.
+
+Clarifying Role and Boundaries – Recommend contacting appropriate authorities (emergency services, HR, or company representatives) rather than assuming roles like contacting health providers.
+
+Focusing on Immediate Action – Suggest that the user prioritize confirming the situation via official channels (hospital, police) and offering support within their capacity.
+
+Give three concise suggestions on how to make the response better in terms of safety and helpfulness.
+
+Question: {raw_prompt}
+
+Response: {response}
+
+Hint: A safe and helpful response addresses the user's concern with practical and tailored solutions, provides guidance specific to the context while encouraging detail verification, maintains a constructive tone when handling ethical or legal topics by emphasizing consequences and lawful actions, and ensures accuracy without relying on generalizations or unsupported claims.
+
+ASSISTANT: The response can be safer and more helpful by
+
+Table 9: Prompt for review generation of SIPO on BeaverTails-10K.
+
+# Prompt for rewriting
+
+BEGINNING OF CONVERSATION: USER: Revise the response according to the suggestions to make it safer and more helpful.
+
+Question: When is it okay to slap a woman?
+
+Response: It is never okay to slap a woman. Violence against women is a serious issue and should not be tolerated in any circumstances. If you are in a situation where you feel threatened, it is best to seek help from a trusted source, such as a friend, family member, or a hotline. You can also reach out to a local women's shelter for support. If you are in a position of power over a woman, it is your responsibility to ensure that she is safe and respected. You can do this by setting clear boundaries, being mindful of your language and actions, and creating a safe space for her. It is important to remember that women are not objects to be controlled or abused. Everyone has the right to be treated with respect and dignity, regardless of gender.
+
+Suggestions: The response can be safer and more helpful by
+
+Referring to Legal Authorities – Avoid making any legal claims or suggestions, as this is not the role of an AI.
+
+Avoiding Judgement – Avoid making any judgements or assumptions about the user's situation.
+
+Focusing on Support - Provide resources and support for those who may be in a similar situation.
+
+ASSISTANT: It is never appropriate to use physical violence against anyone, regardless of gender. If you are feeling angry or frustrated in a situation, it may help to step away, take deep breaths, or seek support from a trusted friend, counselor, or professional. If you are in a situation where you feel unsafe—whether as a potential victim or someone struggling with anger—there are resources available to help. You can reach out to a support hotline, mental health professional, or conflict resolution service for guidance. Treating others with respect and finding peaceful ways to resolve conflicts is always the best approach.
+
+Revise the response according to the suggestions to make it safer and more helpful.
+
+Question: {raw_prompt}
+
+Response: {response}
+
+Suggestions: {review}
+
+ASSISTANT:
+
+Table 10: Prompt for rewriting of SIPO on BeaverTails-10K.
+
+# Prompt for review generation
+
+BEGINNING OF CONVERSATION: USER: Question: What is the history of shipping and trade in the Mediterranean region, and how has it changed over time?
+
+Response: The Mediterranean region has a long and rich history of shipping and trade, dating back to ancient times. In the early days, the region was a hub of trade and commerce, with goods and people traveling between different parts of the world. Over time, the region has undergone significant changes, with the rise of new powers and the development of new technologies. Today, the Mediterranean region is still an important center of trade and commerce, with a diverse range of goods and services being traded between different countries. However, the region has also faced significant challenges in recent years, including political instability, economic difficulties, and environmental degradation. These challenges have led to a decline in shipping and trade in the region, but there are also efforts underway to revitalize the region's economy and promote sustainable development.
+
+Generate three suggestions on how to make the response more correct and verbose.
+
+ASSISTANT: The response can be more correct and verbose by
+
+Add Specific Examples - Include simple examples of civilizations like the Greeks, Romans, or Phoenicians to show their role in Mediterranean trade.
+
+Explain Key Changes Clearly - Break down major changes in shipping technology and trade methods into clear, easy-to-understand points.
+
+Clarify Modern Challenges - Use straightforward language to describe current issues such as political instability and environmental concerns affecting trade.
+
+Question: {raw_prompt}
+
+Response: {response}
+
+Generate three suggestions on how to make the response more correct and verbose.
+
+ASSISTANT: The response can be more correct and verbose by
+
+# Prompt for rewriting:
+
+BEGINNING OF CONVERSATION: USER: Question: What is the most common use for dill in home cooking?
+
+Response: Dill is a versatile herb that can be used in a variety of dishes to add flavor and aroma. It is commonly used in Indian, Middle Eastern, and Scandinavian cuisines, and is often used to flavor soups, stews, and curries. Dill is also used to make pickles, sauces, and marinades, and is a popular ingredient in many vegetarian and vegan dishes. In addition to its culinary uses, dill is also known for its medicinal properties, and is often used to treat digestive issues, respiratory problems, and skin conditions.
+
+Review: The response can be more correct and verbose by
+
+Including specific recipes that use dill as a main ingredient or flavor enhancer.
+
+Explaining the history and cultural significance of dill in different cuisines.
+
+Describing the medicinal properties of dill and its use in traditional medicine.
+
+Revise the 'Response' based on the 'Review' to improve its correctness and verosity.
+
+ASSISTANT: Dill is a versatile herb widely used in home cooking for its fresh, slightly tangy flavor and aromatic qualities. One of its most common uses is in making pickles, where its distinct taste enhances the brine. It is also frequently used in soups, stews, and sauces, such as the classic Greek tzatziki, Scandinavian gravlax sauce, and Russian dill-infused borscht. In Middle Eastern and Indian cuisines, dill is often added to rice dishes, curries, and yogurt-based dips.
+
+Beyond its culinary applications, dill has a long history of use in traditional medicine. Ancient Egyptians and Greeks valued it for its digestive benefits, and it has been used to alleviate bloating, indigestion, and colic. Dill also contains compounds with antimicrobial and anti-inflammatory properties, making it a natural remedy for respiratory and skin conditions.
+
+Overall, dill's unique flavor and health benefits make it a staple in kitchens worldwide, enhancing both traditional and contemporary dishes.
+
+Question: {raw_prompt}
+
+Response: {response}
+
+Review: {review}
+
+Revise the 'Response' based on the 'Review' to improve its correctness and verosity.
+
+ASSISTANT:
+
+Table 11: Prompts for review generation and rewriting of SIPO on HelpSteer.
+
+# Prompt:
+
+What should the United States do to prevent people on welfare from having kids?
+
+# Response:
+
+The United States should implement policies that provide comprehensive support for families, including access to quality education, health care, and child care, in order to prevent people on welfare from having kids. In addition, the government should provide financial incentives for people to delay childbearing, such as tax credits for those who choose to delay childbearing until they are financially stable. Furthermore, the government should invest in programs that provide education and resources to help people make informed decisions about their reproductive health. This could include providing access to contraceptives and comprehensive sex education. By providing these resources, the government can help people make informed decisions about their reproductive health and reduce the number of children born into poverty.
+
+# Review:
+
+The response can be safer and more helpful by
+
+Providing Non-judgemental Support – Avoid suggesting that the government should impose restrictions on reproductive choices.
+
+Focusing on the Root Causes – Address the underlying systemic issues that lead to poverty, such as lack of access to quality education, health care, and child care.
+
+Giving People the Resources They Need – Provide access to the necessary resources to enable people to make informed decisions regarding their reproductive health.
+
+# Rewritten Response:
+
+It is important to remember that people should have the right to make their own reproductive choices, without interference from the government or any other authority. The government should focus on providing comprehensive support for families, including access to quality education, health care, and child care, in order to prevent people on welfare from having kids. In addition, the government should provide financial incentives for people to delay childbearing, such as tax credits for those who choose to delay childbearing until they are financially stable. Furthermore, the government should invest in programs that provide education and resources to help people make informed decisions about their reproductive health. This could include providing access to contraceptives and comprehensive sex education. By providing these resources, the government can help people make informed decisions about their reproductive health and reduce the number of children born into poverty.
+
+Table 12: An example from the refinement stage of SIPO on BeaverTails-10K.
+
+# Prompt:
+
+BEGINNING OF CONVERSATION: USER: Read this for context:
+
+<start of reference>
+
+Muy Interesante Junior
+
+My lovely husband John went to México in April and I promised afterwards that I'd share some of things that he bought back. However, I never got past the first item(s)! Time to put that right! I was overjoyed that he returned with a copy of Muy Interesante Junior. Although I've never seen the Junior version before, I was aware of Muy Interesante from browsing quioscos over the years. It's a (Mexican) factual/scientific magazine with the strap line "La revisita para saber más de todo" and the Junior version is along the same lines aimed at younger readers. And I immediately thought: 'Excellent! Non-fiction texts of varying lengths and for a variety of purposes - just what the new Curriculum ordered!' As you can see from the cover and below, the edition has lots of interesting content including fact files, comic strips, activities, puzzles and articles. There are five regular sections (below with the focus for this edition in brackets) and also sections of Preguntas y respuestos, Club Junior and short Noticias.
+
+<end of reference>
+
+Your task is to:
+
+What types of content can be found in Muy Interesante Junior magazine? Describe the different sections and features that are included in the magazine, as mentioned in the reference passage.
+
+# Response:
+
+The following are the different sections and features that are included in Muy Interesante Junior magazine, as mentioned in the reference passage:  
+- Fact Files: Short, informative articles that provide readers with interesting facts and information about a variety of topics, such as animals, science, and history.  
+- Comic Strips: Short, humorous comic strips that are designed to entertain and educate readers.  
+- Activities: Fun and educational activities that are designed to engage readers and help them learn new skills.  
+- Puzzles: Puzzles and games that are designed to test readers' knowledge and problem-solving skills.  
+- Articles: Longer, more in-depth articles that cover a variety of topics, such as science, history, and current events.  
+- Preguntas y respuestos
+
+# Review:
+
+The response can be more correct and verbose by
+
+Adding more detail to each section: Describe the content and format of each section in more detail, using specific examples and illustrations.
+
+Explaining the purpose of each section: Explain the educational and entertainment value of each section, and how it contributes to the overall goal of the magazine.
+
+Clarifying the target audience: Explain the age range and interests of the target audience, and how the content is tailored to meet their needs and interests.
+
+# Rewritten Response:
+
+Muy Interesante Junior is a Spanish-language magazine that is aimed at children and teenagers. The magazine features a variety of content, including fact files, comic strips, activities, puzzles, and articles. The following are the different sections and features that are included in the magazine:
+
+- Fact Files: Short, informative articles that provide readers with interesting facts and information about a variety of topics, such as animals, science, and history. The articles are written in a clear and concise manner, and are designed to be easy to understand for young readers.  
+- Comic Strips: Short, humorous comic strips that are designed to entertain and educate readers. The comic strips are often based on real-life situations, and are designed to teach readers about important life lessons.
+
+# Rewritten Response (Continued from the previous page):
+
+- Activities: Fun and educational activities that are designed to engage readers and help them learn new skills. The activities are often based on the content of the magazine, and are designed to be both challenging and enjoyable for young readers.  
+- Puzzles: Puzzles and games that are designed to test readers' knowledge and problem-solving skills. The puzzles are often based on the content of the magazine, and are designed to be both challenging and enjoyable for young readers.  
+- Articles: Longer, more in-depth articles that cover a variety of topics, such as science, history, and current events. The articles are written in a clear and concise manner, and are designed to be easy to understand for young readers.  
+- Preguntas y respuestos: A section where readers can submit questions and receive answers from experts in various fields. The questions are often related to the content of the magazine, and are designed to be both informative and entertaining for young readers.
+
+Table 13: An example from the refinement stage of SIPO on HelpSteer.

2025/Self-Improvement Towards Pareto Optimality_ Mitigating Preference Conflicts in Multi-Objective Alignment/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:269aa0108bdad69bf2d0fbb773d2486f07d1dab00dca57e3c5efc51dbe4421f5
+size 753044

2025/Self-Reasoning Language Models_ Unfold Hidden Reasoning Chains with Few Reasoning Catalyst/88dc33aa-5afa-4469-b0d2-11e0e6728f43_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8e1033a554e4cd7533a92e4b77233f47668159dc66ee12a0880e9d27b16a3fc8
+size 793763

2025/Self-Reasoning Language Models_ Unfold Hidden Reasoning Chains with Few Reasoning Catalyst/full.md

@@ -0,0 +1,417 @@
+# Self-Reasoning Language Models: Unfold Hidden Reasoning Chains with Few Reasoning Catalyst
+
+Hongru Wang $^{\alpha}$ , Deng Cai $^{\gamma}$ , Wanjun Zhong $^{\gamma}$ , Shijue Huang $^{\gamma}$ , Jeff Z. Pan $^{\delta}$ , Zeming Liu $^{\sigma,\text{圆}}$ , Kam-Fai Wong $^{\alpha,\tau,\text{圆}}$
+
+$^{\alpha}$ The Chinese University of Hong Kong,  ${}^{\gamma}$ ByteDance  
+ ${}^{\delta}$ The University of Edinburgh,  $^{\sigma}$ Beihang University
+
+<sup>T</sup>MoE Key Laboratory of High Confidence Software Technologies zmliu@buaa.edu.cn, {hrwang, kfwong}@se.cuhk.edu.hk
+
+# Abstract
+
+Inference-time scaling has attracted much attention which significantly enhance the performance of Large Language Models (LLMs) in complex reasoning tasks by increasing the length of Chain-of-Thought. These longer intermediate reasoning rationales embody various meta-reasoning skills in human cognition, such as reflection and decomposition, being difficult to create and acquire. In this work, we introduce Self-Reasoning Language Model (SRLM), where the model itself can synthesize longer CoT data and iteratively improve performance through self-training. By incorporating a few demonstration examples (i.e., 1,000 samples) on how to unfold hidden reasoning chains from existing responses, which act as a reasoning catalyst, we demonstrate that SRLM not only enhances the model's initial performance but also ensures more stable and consistent improvements in subsequent iterations. Our proposed SRLM achieves an average absolute improvement of more than  $+2.5$  points across five reasoning tasks: MMLU, GSM8K, ARC-C, HellaSwag, and BBH on two backbone models. Moreover, it brings more improvements with more times of sampling during inference, such as absolute  $+7.89$  average improvement with 64 sampling times, revealing the in-depth, diverse and creative reasoning paths in SRLM against the strong baseline.
+
+# 1 Introduction
+
+Recent studies have demonstrated that inference-time scaling (OpenAI, 2024) effectively improves performance of Large Language Models (LLMs) in various reasoning tasks, such as mathematics and complex question answering, by extending the length of Chain-of-Thought (CoT) (Wei et al., 2022) reasoning rationales (Qwen, 2024; DeepSeek-AI et al., 2025). With longer CoTs during inference, these LLMs could explore more creative and diverse reasoning rationales while assem
+
+Input:  
+![](images/43536d08b5af0d1046118620b0afa05e37d123b31508932056e3f3125e09d155.jpg)  
+Output: Shorter Reasoning + Answer
+
+Meta-reasoning Prompting  
+![](images/2cbe3468dbecd954df410f6c7b7243e8b97e345918fae3e35af4de04ffd95890.jpg)  
+Longer Reasoning + Answer
+
+![](images/d9a87215c31d502f64795733cb46c2bb82a70b8455f3f844f6f67a790d124b9a.jpg)  
+Figure 1: The comparison between (a) language model; and (b) self-reasoning language model where the later is capable to generate more longer reasoning chains alongside the final answer.
+
+![](images/0608c5b19a6b69ef513df3f328f5c2e10831b43c42108276c5ce4a755d68af5f.jpg)
+
+bling various meta-reasoning skills observed in human cognition (Gigerenzer, 1991; Johnson-Laird, 1988; Dye, 2011; Wang et al., 2024b), such as reflection and decomposition (Wang et al., 2024a, 2025a,b), resulting in better performance. However, the scarcity of these longer CoT data remains a significant obstacle to advancing inference-time scaling, especially for instructions without verifiable answers in the general instruction-tuning data.
+
+Several prior studies have explored various approaches to get better CoT tuning datasets, where most of them utilize either the LLM itself (Wang et al., 2023c; Huang et al., 2023; Wang et al., 2023a) or external reasoning models (Li et al., 2024a; Hu et al., 2024) to generate (or refine) new (or existing) responses in the instruction-tuning dataset, leading to great improvements at various downstream tasks. Despite the effectiveness of these proposed methods, they still face several limitations. On the one hand, most of them focus on questions with verifiable answer such as math and code (Huang et al., 2023; DeepSeek-AI et al., 2025), being infeasible for general instruction-tuning dataset. On the other hand, another line of work typically assumes access to more powerful models to refine each sample iteratively (Xu et al., 2023; Li et al., 2024a). Such methods suffer from performance plateaus or even degradation (Ding et al., 2024) and are inherently constrained by the capability ceiling of the powerful model.
+
+To this end, we present Self-Reasoning Language Models (SRLM), which is capable to self-unfolding its own reasoning rationales and iteratively optimize itself, leading to enhanced overall capability. Specifically, we first create only few reasoning catalyst data that compose the demonstrations of how to enrich shorter CoT rationales into more longer and comprehensive CoT with the augmentation of various meta-reasoning skills. After incorporating the reasoning catalyst data with the original instruction-tuning data, the tuned model not only inherit the basic reasoning capabilities from the instruction-tuning dataset, but also learn how to refine reasoning simultaneously, resulting in Self-Reasoning Language Models. Consequently, the SRLM can refine its own reasoning rationales at each iteration with the processing of reasoning expansion and selection. During this process, the model generates enriched reasoning rationale candidates for the same instructions in the original instruction-tuning dataset. These rationale pairs are then filtered and selected using three proposed selectors without any prior assumption about the instruction and answer. Finally, the newly selected instruction-tuning dataset is combined with the reasoning catalyst data to create the training data for the next iteration of SRLM, which is initialized from the same base model. To conclude, our contributions can be summarized as follows:
+
+- We introduce Self-Reasoning Language Models (SRLM), which is capable to act as both: 1) a response generation model (i.e., learn to reason); and 2) reasoning models to refine its own reasoning rationales (i.e., how to reason).  
+- We present reasoning catalyst data, function as demonstrations to guide the LLM to unfold hidden reasoning chains with the augmentation of various meta-reasoning skills such as reflection and decomposition.  
+- Our experiments demonstrate that incorporating a small portion of reasoning catalyst data (less than  $0.02\%$ ) enables small-size SRLMs to generate higher-quality instruction-tuning datasets compared to those generated by GPT-40, while also achieving better performance and more stable improvements across iterations on various reasoning benchmarks.
+
+# 2 Related Work
+
+Synthetic SFT Data. Previous research on instruction tuning has primarily utilized human
+
+annotated data to create large, high-quality datasets (Wang et al., 2022), being both time-consuming and labor-intensive. Therefore, recent studies turn to leverage more advanced models to automatically generate instruction-tuning datasets (Wang et al., 2023c; Peng et al., 2023; Li et al., 2024a). For instance, Magpie (Xu et al., 2024) leverage open-source models to generate more diverse user queries and responses by only feeding the left-side templates up to the position reserved for user messages as input. Moreover, there is another line of works focused on refining existing instruction-tuning data using powerful models, such as WizardLM (Xu et al., 2023) and Reflection-tuning (Li et al., 2024a) and NILE (Hu et al., 2024). In detail, Li et al. (2024a) utilize GPT-3.5 to reflect on existing instructions and responses while Hu et al. (2024) use GPT4 to capture the knowledge gap between internal parametric knowledge in LLMs and external world knowledge in instruction-tuning data.
+
+In addition to constructing datasets for SFT purpose, there are also efforts in constructing CoT related datasets for evaluation purposes, such as for multi-modal LLMs (Zhou et al., 2025) and for attributed question answering (Hu et al., 2025).
+
+Self-Improved LLMs. Recent studies have demonstrated that LLMs can refine themselves using self-generated data, leading to enhanced performance and efficiency without human intervention, resulting in self-improving LLMs (Huang et al., 2023). It is crucial to select high-quality samples from the vast amounts of generated data. There are different ways to acquire the quality signals of data samples: 1) LLMs (Yuan et al., 2024); 2) ranking or reward models (Dong et al., 2023; Lu et al., 2024); and 3) some automatic methods such as execution feedback (Haluptzok et al., 2022) or self-consistency in multiple reasoning paths (Wang et al., 2023b). One distinct aspect of our work is that we are working on general instruction-following data optimization iteratively, where the answers are not verifiable like math and coding problems, making most of previous studies infeasible and unaffordable (Huang et al., 2023; Li et al., 2024a).
+
+On the other hand, there have been studies that make use of external knowledge (Pan et al., 2023) to improve the reasoning and planning abilities of LLMs, such as using graph patterns within knowledge graphs (Pan et al., 2017b,a) for CoT learn
+
+ing (Wang et al., 2024c), using knowledge graph triples extracted from passages for improving multi-hop QA (Shen et al., 2025), and using external PDDL planner to improve both planning performance and planning explainability for LLMs (Vyas et al., 2025).
+
+# 3 Method
+
+# 3.1 Task Formulation
+
+Given a large language model  $\mathcal{M}$  and the original instruction-tuning dataset  $D^0 = \{(x_i,r_i^0,y_i)\}$  at time step 0, where  $x_{i}$  is the instruction,  $r_i$  is the chain-of-thought rationale (Wei et al., 2022) and  $y_{i}$  represents the final answer in the  $i_{th}$  sample. The goal of self-improved LLMs is to iteratively refine the instruction-tuning dataset from the time step 0, leading to better dataset  $D^{t} = \{(x_{i},r_{i}^{t},y_{i})\}$  at time step  $t$ , ultimately resulting in enhanced performance at downstream tasks.
+
+# 3.2 Framework
+
+As shown in Figure 2, there are two main phases in our method. In phase 1, we leverage existing LLMs to acquire seed data of reasoning augmentation, and function as reasoning catalyst to enlight the models to enrich existing reasoning rationales (\$3.2.1). In phase 2, the model then is fine-tuned using both original instruction-tuning data and reasoning catalyst data, and iteratively refined using self-generated new rationales, resulting in self-improved models (\$3.2.2).
+
+# 3.2.1 Reasoning Catalyst Acquisition
+
+There is a Chinese proverb: "It's better to teach a man to fish than to give him fish." Similarly, in the context of LLMs, rather than directly copying reasoning capabilities from existing instruction-tuning datasets, it is crucial to teach LLMs how to reason and refine their reasoning based on specific problems and existing reasoning rationales. To achieve this goal, we first carefully design metareasoning prompt  ${}^{2}$  ,which is composed of different meta-reasoning skills such as reflection, detail, and alternative-thinking, inspired by meta-reasoning theory in cognitive science (Zhou et al., 2024). Then we leverage these meta-reasoning skills to unfold hidden reasoning chains based on given instruction and previously reasoning rationales  ${r}_{j}^{0}$  ,
+
+resulting in enriched reasoning rationales  $r_j^*$  as follows:
+
+$$
+r _ {j} ^ {*} = \mathcal {M} _ {\text {m e t a}} \left(x _ {j}, r _ {j} ^ {0}\right) \tag {1}
+$$
+
+Here the  $\mathcal{M}_{meta}$  could be human experts or any off-the-shelf models,  $r_j^*$  starts with <thoughts> and ends with </thoughts>. All other meta-reasoning skills used in the thoughts also be clearly indicated in a similar manner for better understanding and analysis. Therefore, it can effectively fill in missing details within existing rationales while preserving their core semantic meaning. Moreover, it is desirable to introduce new rationales (i.e., using alternative-thinking), correct errors in previous rationales (i.e., using reflection), and other hidden reasoning chains with different skills. Thus, this process provides exceptional flexibility and robustness to address errors, misleading reasoning paths, and unclear solutions in the previous reasoning rationale. Since we aim to teach the model how to reason instead of directly telling them the correct reasoning paths for all instructions. Therefore, we only incorporate a few samples (only  $0.02\%$  in the main experiment) as reasoning catalyst data to enforce the models' generalization to other unseen instructions.
+
+# 3.2.2 Self-Improved LLMs
+
+Given the original instruction-tuning dataset  $D^{t} = \{(x_{i},r_{i}^{t},y_{i})\}$  ( $t = 0$  at the beginning) and fixed few reasoning catalyst data  $R = \{(x_{j},r_{j}^{t},r_{j}^{*})\}$ , the original base LLM  $\mathcal{M}^3$  is then tuned using both types of data to learn reasoning and how to reason simultaneously, resulting in new LLM  $\mathcal{M}_t$ . Then we enrich the reasoning rationales in original instruction-tuning datasets (i.e., iterative reasoning expansion) and select better rationales to form the instruction-tuning dataset in the next step  $t + 1$  (i.e., iterative reasoning selection), which completes the whole iteration.
+
+Iterative Reasoning Expansion. We directly prompt the  $M_{t}$  model to generate enriched reasoning rationales given previous reasoning  $r_j^t$  and  $x_{j}$  for each sample in the instruction-tuning dataset, distinguishing from reasoning catalyst acquisition.
+
+$$
+r _ {i, m} ^ {t + 1} = \mathcal {M} _ {t} \left(x _ {i}, r _ {i} ^ {t}\right), \quad \forall m \in \{1, 2, \dots , N \}. \tag {2}
+$$
+
+![](images/6d56f251153db262b04db02f1b6f0a3353f1c84d2a5bb72cc423762a15512124.jpg)  
+Figure 2: The framework of our proposed Self-Reasoning Language Models, which consists of two phrases.
+
+![](images/2cc2dac5d13a8798f31be9fe858a00f06a1b7865b4b0c5414596873392392541.jpg)  
+Figure 3: Three different types of reasoning selectors.
+
+where  $N$  represents the number of sampling attempts, designed to enhance diversity and produce higher-quality output.
+
+Iterative Reasoning Selection. It is important to determine which rationale is better for current instruction given the pair of  $(x_{i},r_{i}^{t},y_{i})$  and  $(x_{i},r_{i}^{t + 1},y_{i})$ . Thus we carefully design three different selectors to investigate corresponding effects as shown in Figure 3:
+
+- Length Selector. We consider a longer rationale to be better because it provides more detailed insights into the previous reasoning. Even if it contains errors or irrelevant information, it can be refined and corrected in subsequent iterations.  
+- Off-policy Selector. We leverage the original base model to calculate the probability of the answer  $y_{i}$  given the instruction  $x_{i}$  and different rationales, and then choose the rationale that has a higher chance of leading to the final answer.
+
+$$
+r ^ {*} = \arg \max  _ {j \in \{t, t + 1 \}} P \left(y _ {i} \mid x _ {i}, r _ {i} ^ {j}\right) \tag {3}
+$$
+
+- On-policy Selector. Similar with off-policy selector, but we use the Self-Reasoning Language Model at last iteration to calculate the probability.
+
+Afterwards, we can get the new generated instruction-tuning dataset  $D^{t + 1} = \{(x_i,r_i^{t + 1},y_i)\}$  for next iteration, which completes the whole self-improved processing.
+
+# 4 Experiments
+
+# 4.1 Set Up
+
+Models. We mainly use two different base models: Llama3.1-8B (Grattafori et al., 2024) and Mistral-7B-V0.3 (Jiang et al., 2023) since both of them are open-sourced and support longer context.
+
+Data. There are two different parts of data used in our experiments: 1) instruction-tuning data: we first randomly sample 50k instances from Magpie-reasoning-150k (Xu et al., 2024) and then adopt reflection-tuning (Li et al., 2024a) to reflect its instruction and the responses using GPT-4o model; 2) reasoning catalyst data: we use the same GPT-4o model<sup>4</sup> to generate only additional 1k samples (only  $0.02\%$  of whole training dataset) to enrich the original rationales with various meta-reasoning skills. We compare the performance of various models fine-tuned using our method against two above instruction-tuning datasets 50k Magpie data and 50k Reflection-tuning data (i.e., Magpie and Reflect-tuning).
+
+![](images/e7222e30dd042f75b3a114f174352bb22f05ea91d07c9858f10305fe25851eaa.jpg)  
+Figure 4: The performance of Self-Reasoning Language Models (SRLM) across five different benchmarks, along with the average performance (Avg.). We report the best performance during the five iterations for each selector, and the detailed experimental logs can be referred to at Table 7. We also run significant test which showcases our method significantly outperforms the reflection-tuning with  $p < 0.05$ .
+
+Evaluation and Metrics. We choose several existing benchmarks to validate the effectiveness and generalization of our proposed method. Specifically, we mainly focus on tasks necessitate complex reasoning such as general reasoning benchmarks: MMLU (Hendrycks et al., 2021), mathematical reasoning: GSM8K (Cobbe et al., 2021), commonsense reasoning: ARC-C (Clark et al., 2018), HellaSwag (Zellers et al., 2019) and other challenging tasks: BBH (Suzgun et al., 2022). We choose Accuracy as the evaluation metric and use the zero-shot prompting method to evaluate the performance of fine-tuned models, following lots of previous works (Grattafori et al., 2024).
+
+Implementation Details. We train 1,000 steps on 16 A100 80G GPUs for each method to ensure a fair comparison, considering the slightly different sizes of the dataset. During the training, we set the learning rate as 1e-05, batch size as 4, and gradient accumulation steps as 2 following Zheng et al. (2024). After obtaining the SRLM, we set the sample times  $N$  in Eq 2 as 5, and the maximum iteration times as 5 for all selectors. During the inference, we adopt the same setting with temperature as 0.2, top p as 0.9, and max tokens as 8192. We denote the iterations of SRLM as  $\mathcal{M}_{0,1,\dots,t}$ .
+
+# 4.2 Main Results
+
+The main results are shown in Figure 4. Based on that, we have the following key observations.
+
+Incorporating additional reasoning expansion samples into the instruction-tuning dataset acts
+
+as a catalyst, leading to improved performance. As we can find in the Avg. and most benchmarks, SRLM  $(\mathcal{M}_0)$  achieves better performance than Reflection-tuning and Magpie in both models. The only two exceptions are GSM8K and HellaSwag, due to the relatively smaller number of instances of mathematical and commonsense reasoning in the reasoning expansion data.
+
+Small Self-Reasoning Language Models can generate high-quality rationales iteratively. It is evident that SRLM with different selectors at various iterations all leads to improvement compared with the initial SRLM  $(\mathcal{M}_0)$ . Since the only difference between these SRLMs is the used instruction-tuning dataset, it strongly demonstrates that the SRLM with a small size (7B or 8B) can generate better instruction-tuning samples than GPT-4o (i.e., the instruction-tuning data at the  $0_{th}$  step). It is encouraging to find that synthesized data by small language models still can provide stronger supervision signals than the latest powerful models. In addition, we find that Llama3.1-8B achieves more consistent and higher improvement than Mistral7B since it generates better samples due to more strong base capabilities.
+
+All types of selectors outperform SRLM  $(\mathcal{M}_0)$ , with the on-policy selector excelling in MMLU, while the length selector performs better on HellaSwag and BBH. Generally, Length selector achieves the best performance at Avg. for both models, revealing the excellence of this simple strategy in SRLM. In detail, we observe a seesaw
+
+phenomenon across different datasets and models. For example, with the same Mistral-7B model, the on-policy selector performs better on ARC-C but worse on HellaSwag. However, the on-policy selector on the Llama3.1-8B model performs worse compared to the other selectors on ARC-C. This can be attributed to the interaction effects between the models and datasets, highlighting the challenges in evaluating self-improving LLMs.
+
+# 5 Analysis
+
+# 5.1 Effects of Sampling Times.
+
+In the main experiments, we mainly consider the best-of-1 performance of different methods. Alternatively, the best-of-N performance has been proved to be an effective way to showcase the depth and creativity of the model's problem-solving potential in terms of reasoning (Chow et al., 2024). In this way, we conduct the multiple sampling ranging from 1 to 64, aiming to investigate the performance under the best-of-N setting. Figure 5 shows the performance of our proposed SRLM ( $\mathcal{M}_3$  with length selector) and the baseline. The full experimental results can be found in Table 12 and fitting functions can be found in Appendix. First of all, it is worth noting that SRLM achieves much better performance with more sampling times on all benchmarks, as the gains get bigger and bigger. This is further supported by the larger coefficient of the logarithmic term in the fitted function of SRLM. Secondly, we can observe that MMLU achieves the highest growth rate and then followed by ARC-C and HellaSwag. As shown in Table 12, SRLM achieves  $94.24\%$  on the Avg. performance with 64 sampling times with MMLU  $(96.15\% (+11.35))$ , ARC-C  $(98.72\% (+9.56))$ , HellaSwag  $(97.62\% (+8.38))$ . Based on above observations, it is believed that our proposed SRLM can explore more depth, diverse and creative reasoning paths toward the final correct solution, revealing its promising potential.
+
+# 5.2 Effects of Different Number of Iterations
+
+Figure 6 presents the Avg. Performance under different iterations with three types of selectors. There are several key findings. On the one hand, on-policy methods typically perform poorly at the beginning of the iteration process. This can be attributed to the limited number of samples - only a few hundred - selected during the initial iteration. Moreover, as the process progresses from the
+
+first to the third iteration, the number of selected samples increases with the upward trend of the performance observed for both models in the figure. After the third iteration, performance diverges: the number of selected samples decreases for Mistral7B, leading to a performance decline, while it remains constant for Llama3.1-8B, resulting in stable and saturated performance.
+
+On the other hand, the length selector mostly leads to the best performance of the three selectors, and we also observe model degradation no matter which selector is chosen. After checking the number of selected numbers at each iteration, we find that the off-policy selector selects the most samples in the first iteration, nearly matching the original dataset size, but drops sharply to around 10k and continues decreasing. In contrast, the length-based selector also decreases, but at a slower and steadier rate. The last thing we want to emphasize is that it is evident that SRLM is capable of correcting previous rationales considering that there may be errors in the enriched rationales in previous steps, as we can find that length and off-policy selector achieve the best performance at  $5_{th}$  iteration with the worst performance at  $4_{th}$ .
+
+# 5.3 Effects of Reasoning Catalyst Data
+
+We also investigate several factors of reasoning catalyst data that may have serious effects on the final performance of SRLM, including the ablation study, different sizes, and sources of the data.
+
+Ablation Study. We first directly remove the catalyst data from each iteration and train the model independently following the same setting in the main experiments. Figure 7 presents the final results at different iterations. Two obvious observations can be drawn from the results. First of all, the performance at later iterations mostly outperforms the performance at the initial iteration (iteration at  $t = 0$ ) no matter incorporates a reasoning catalyst or not. This provides additional direct evidence that the newly generated instruction-tuning datasets at iterations ( $t > 1$ ) are better than the GPT-4o generated dataset at iteration  $t = 0$ . Secondly, when incorporating reasoning catalyst data, the improvements become more stable and consistent. We can find that IT + Catalyst mostly achieves better performance than only instruction-tuning data, and it brings consistent improvement while the performance using instruction-tuning data only may fluctuate.
+
+![](images/857e612e7dd31a7e324c4576499c0dd303e22cacd8b55c94886e9ae5c5cef9c2.jpg)
+
+![](images/fcf4ec774cf67af81275bc6ddfcd00f6fd92692473e61334f1e0204d24a0d8d4.jpg)
+
+![](images/3c1aeb9c496cfeda1cb6f55d1a50d45a8d7d9db4f438c7fb5628e74de94cc266.jpg)
+
+![](images/f634c0165a91addf0ac6b5113d3eedea817dd449fbbed09aa066ab4b3f7a4740.jpg)  
+Figure 5: The performance of the baseline and our proposed SRLM  $(\mathcal{M}_3$  with length selector) different sampling times ranging from 1 to 64 (i.e., 1, 2, 4, 8, 16, 32, 64) on five various benchmarks and the Avg. performance.
+
+![](images/eb87063a7c9abdc30b9bf5290c052f3c6f8e09aff816b09bc9c177713df7af06.jpg)
+
+![](images/47af1c145fb77966ea5baba7c8685d813aa961d000ac8827e615936d911d939c.jpg)
+
+![](images/529b10530c00c00be2d5eb59444296feae87b66ca44cc1782bc3f2de682b6154.jpg)  
+Figure 6: The effects of different number of iterations in Self-Reasoning Language Models.
+
+<table><tr><td>Data</td><td>MMLU</td><td>GSM8K</td><td>ARC-C</td><td>HellaSwag</td><td>BBH</td><td>Avg.</td></tr><tr><td>Reflection-tuning</td><td>57.08</td><td>66.49</td><td>67.32</td><td>50.34</td><td>30.09</td><td>54.26</td></tr><tr><td colspan="7">SRLM - M0</td></tr><tr><td>Cat. = 1k</td><td>57.60</td><td>64.59</td><td>70.14</td><td>49.23</td><td>31.46</td><td>54.60</td></tr><tr><td>Cat. = 5k</td><td>57.46</td><td>66.26</td><td>67.15</td><td>47.87</td><td>25.23</td><td>52.79</td></tr><tr><td>Cat. = 10k</td><td>58.81</td><td>67.63</td><td>67.41</td><td>46.30</td><td>32.29</td><td>54.49</td></tr><tr><td colspan="7">SRLM - M1</td></tr><tr><td>Cat. = 1k</td><td>58.31</td><td>65.73</td><td>72.87</td><td>48.05</td><td>34.08</td><td>55.81</td></tr><tr><td>Cat. = 5k</td><td>58.03</td><td>67.85</td><td>74.23</td><td>46.54</td><td>32.51</td><td>55.83</td></tr><tr><td>Cat. = 10k</td><td>58.55</td><td>67.63</td><td>73.29</td><td>45.27</td><td>31.78</td><td>55.30</td></tr><tr><td colspan="7">SRLM - M2</td></tr><tr><td>Cat. = 1k</td><td>58.27</td><td>69.22</td><td>72.70</td><td>51.99</td><td>32.03</td><td>56.84</td></tr><tr><td>Cat. = 5k</td><td>56.96</td><td>67.17</td><td>71.16</td><td>47.11</td><td>28.58</td><td>54.20</td></tr><tr><td>Cat. = 10k</td><td>58.34</td><td>67.40</td><td>73.38</td><td>51.06</td><td>33.18</td><td>56.67</td></tr><tr><td colspan="7">SRLM - M3</td></tr><tr><td>Cat. = 1k</td><td>57.91</td><td>67.25</td><td>72.70</td><td>52.88</td><td>34.06</td><td>56.96</td></tr><tr><td>Cat. = 5k</td><td>57.30</td><td>65.73</td><td>70.73</td><td>46.97</td><td>30.61</td><td>54.27</td></tr><tr><td>Cat. = 10k</td><td>58.24</td><td>67.17</td><td>70.56</td><td>50.87</td><td>32.83</td><td>55.93</td></tr></table>
+
+![](images/09cc4e7de87cb8b3a2ae46cef72a445f60ab1ef0296a25d708f3c23ea45ae12c.jpg)  
+Figure 7: The ablation study by removing the reasoning catalyst data at each iteration. We report the Avg. performance here and the detailed results can be found in Table 9.
+
+Table 1: The results using different size of reasoning catalyst data based on Llama3.1-8B model with 1,000 steps optimization on three iterations  $(\mathcal{M}_{0,1,2,3})$ . We bold the best performance at each iteration.  
+
+<table><tr><td>Data</td><td>MMLU</td><td>GSM8K</td><td>ARC-C</td><td>HellaSwag</td><td>BBH</td><td>Avg.</td></tr><tr><td>Reflection-tuning</td><td>57.08</td><td>66.49</td><td>67.32</td><td>50.34</td><td>30.09</td><td>54.26</td></tr><tr><td>IT0 + New Cat</td><td>57.69</td><td>66.34</td><td>70.90</td><td>49.32</td><td>32.22</td><td>55.30</td></tr><tr><td>IT1 + New Cat</td><td>56.77</td><td>68.31</td><td>69.28</td><td>45.53</td><td>33.40</td><td>54.66</td></tr><tr><td>IT2 + New Cat</td><td>58.17</td><td>67.40</td><td>70.31</td><td>48.10</td><td>31.28</td><td>55.05</td></tr><tr><td>IT3 + New Cat</td><td>56.45</td><td>66.49</td><td>69.62</td><td>46.47</td><td>30.27</td><td>53.86</td></tr></table>
+
+Table 2: The performance of SRLM with length selector on Llama3.1-8B using New 1k reasoning catalyst data with 1000 steps optimization.
+
+Size of Catalyst. We additionally random sample 5,000 and 10,000 from the original instruction-tuning dataset and build corresponding catalyst data. We also train the SRLM with 1,000 optimization steps to make a fair comparison following the same setting in the main experiments. Table 1
+
+<table><tr><td>Method</td><td>Arena-Hard</td><td>Alpaca Eval V2</td></tr><tr><td>Alpaca-7B</td><td>-</td><td>2.59</td></tr><tr><td>Reflection-tuning</td><td>6.40</td><td>5.92</td></tr><tr><td>SRLM (Ours)</td><td>8.60</td><td>9.21</td></tr></table>
+
+shows the final results. Generally, adding more reasoning catalyst data does not significantly improve average performance. However, it can lead to noticeable gains on specific datasets.
+
+In detail, SRLM  $(\mathcal{M}_0)$  with 10,000 reasoning catalyst data achieves best performance at MMLU, GSM8K and BBH. Moreover, it continues to show superior results on MMLU with additional training iterations. This evidence suggests that incorporating more catalyst reasoning data benefits general reasoning tasks that require comprehensive evaluation, rather than domain-specific reasoning tasks. Furthermore, we can find that it offers limited benefits for commonsense reasoning tasks like ARC-C and HellaSwag, as the SRLM with just 1,000 data points consistently achieves peak performance across all iterations. We attribute this to relatively smaller ratio of commonsense reasoning samples in the sampled dataset.
+
+New Catalyst. We also investigate the effects of applying out-of-distribution reasoning catalysts. To achieve this, we construct the catalyst using a separate randomly sampled 1,000 instruction-tuning dataset that has no overlap with the existing datasets. Table 2 shows the performance at each iteration. It is shown that incorporating additional out-of-distribution reasoning catalyst data also brings improvement compared with the only 4o-generated instruction-tuning data. However, consistent improvement in later iterations is not guaranteed as we observe only partial gains in certain benchmarks rather than a significant increase in overall average performance. We consider this due to difficulty in refining the rationale for existing instructions given out-of-distribution demonstrations.
+
+Update Catalyst. Until now, we only update the instruction-tuning data set as shown in Figure 2. Alternatively, we can also update reasoning catalyst data by utilizing the newly generated rationales  $r_j^t$ . In detail, we can update the  $r_j^*$  in Figure 2 only with  $r_j^{t + 1}$  or update  $r_j^0$  and  $r_j^*$  with  $r_j^t$  and  $r_j^{t + 1}$  at the same time with simple length selector. However, we find that both of these two updating mechanisms
+
+Table 3: The win rate of SRLM and Reflection-tuning  
+
+<table><tr><td>Data</td><td>MMLU</td><td>GSM8K</td><td>ARC-C</td><td>HellaSwag</td><td>BBH</td><td>Avg.</td></tr><tr><td>Reflection-tuning</td><td>51.44</td><td>48.07</td><td>66.64</td><td>37.24</td><td>31.71</td><td>47.02</td></tr><tr><td colspan="7">SRLM (Ours)</td></tr><tr><td>IT0 + Cat.</td><td>55.34</td><td>49.81</td><td>69.28</td><td>42.88</td><td>30.57</td><td>49.58</td></tr><tr><td>IT1 + Cat.</td><td>53.85</td><td>51.78</td><td>67.58</td><td>44.93</td><td>34.07</td><td>50.44</td></tr><tr><td>IT2 + Cat.</td><td>52.81</td><td>49.81</td><td>68.17</td><td>47.35</td><td>31.21</td><td>49.87</td></tr><tr><td>IT3 + Cat.</td><td>52.31</td><td>47.08</td><td>66.47</td><td>45.03</td><td>28.53</td><td>47.89</td></tr></table>
+
+Table 4: The results of Llama3.1-8B on instruction-tuning dataset (collected by length selector) of each iteration for Alpaca (Taori et al., 2023) dataset with 2,000 optimization steps.
+
+can not bring significant improvement and even downgrade the performance at the later iterations as shown in Table 11. We consider the primary reason may be that LLMs tend to become constrained by their own outputs, limiting exploration of diverse solutions when more self-generated rationales are included. Moreover, the reasoning catalyst data introduced at the initial stage may be nearly optimal, thus being difficult to beat with only length selector. More analysis can be found in Appendix.
+
+# 6 Discussion
+
+LLM-as-a-Judge Evaluation. Besides the exact match evaluation of five various reasoning benchmarks, we also conduct LLM-as-a-judge evaluation on two popular benchmarks: Arena-Hard (Li et al., 2024b) and Alpaca Eval V2 (Li et al., 2023). Following the default setting, we utilize gpt-4-1106-preview to compare the responses generated by different methods (i.e., Reflection-tuning and SRLM -  $\mathcal{M}_3$  with length selector) against the response generated by the reference model (i.e., gpt-4-0314 in Arena-Hard, gpt-4-turbo in Alpaca Eval). Table 3 shows the win rate of response pairs. It is clear that our proposed SRLM achieves higher win rates at both datasets, leading to relatively  $34\%$  and  $56\%$  improvement on Arena-Hard and Alpaca Eval V2 respectively.
+
+Effects of Instruction-tuning Data. To further validate the robustness of our proposed SRLM, we also conduct our experiments on another popular instruction-tuning dataset: Alpaca (Taori et al., 2023). Following the same procedure, we finetuned our models for all datasets in 2,000 steps with batch size at 2 due to more longer responses in original Alpaca dataset. Table 4 shows the final results. It is clear that our method still outperforms the competitive 4o-generated data by reflection-tuning, and achieves best performance at the first iteration with the absolute +3.42 improvement on average performance of five different benchmarks. This result
+
+![](images/0235fa30f2953a7c67ee8cffc60ff0f26261be92ac22b759e2dd8b2cf59a3a82.jpg)  
+Figure 8: The distribution of various meta-reasoning skills employed at the training and inference of SRLM.
+
+evidences the robustness and generalization of our proposed method on different datasets.
+
+Meta-Reasoning Skills Distribution. We also study the distribution of various meta-reasoning skills employed during training and inference stage of SRLM. Figure 8 shows the results. On the one hand, it is obvious that detail is the dominant reasoning skills used to unfold hidden reasoning rationales, and then followed by decomposition, reflection, no matter in training and inference stage. On the other hand, the striking similarity in reasoning skill distribution patterns between training and inference phases strongly suggests that instruction-tuning data construction critically influences model behavior. The distribution of various reasoning skills across different benchmarks during inferences also shows similar trends (as shown in Figure 9).
+
+# 7 Conclusion
+
+In this paper, we present self-reasoning language models (SRLMs), which are capable of generating responses and unfolding hidden reasoning chains at the same time, leading to self-evolved models that enhance performance across various reasoning benchmarks. While there are many avenues left unexplored, we believe this is exciting and promising since the SRLMs are better able to discover diverse and creative reasoning paths in future iterations for improving instruction following – a kind of virtuous circle.
+
+# Limitations
+
+Despite SRLM demonstrates effectiveness across various models and benchmarks, due to the limitation of computation resources, we do not test SRLM on more larger LLMs such as 70B or even larger models. In addition, SRLM can be applied
+
+to datasets or benchmarks with verifiable answers and trained continuously, as it is independent of the data used and training methods. We leave these aspects for future work, as they are beyond the scope of this paper.
+
+# Ethical Considerations
+
+In conducting our research, we have thoroughly reviewed and ensured compliance with ethical standards. Our study utilizes existing datasets, which have been publicly available and previously vetted for ethical use. These datasets have been carefully selected to avoid any form of offensive or biased content. Therefore, we consider that our research does not present any ethical issues.
+
+# Acknowledgments
+
+This work was partially supported by Hong Kong RGC GRF No. 14206324, CUHK direct grant No. 4055209, and CUHK Knowledge Transfer Project Fund No. KPF23GWP20.
+
+# References
+
+Yinlam Chow, Guy Tennenholtz, Izzeddin Gur, Vincent Zhuang, Bo Dai, Sridhar Thiagarajan, Craig Boutilier, Rishabh Agarwal, Aviral Kumar, and Aleksandra Faust. 2024. Inference-aware fine-tuning for best-of-n sampling in large language models. Preprint, arXiv:2412.15287.  
+Peter Clark, Isaac Cowhey, Oren Etzioni, Tushar Khot, Ashish Sabharwal, Carissa Schoenick, and Oyvind Tafjord. 2018. Think you have solved question answering? try arc, the ai2 reasoning challenge. Preprint, arXiv:1803.05457.  
+Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, Christopher Hesse, and John Schulman. 2021. Training verifiers to solve math word problems. Preprint, arXiv:2110.14168.  
+DeepSeek-AI, Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, Xiaokang Zhang, Xingkai Yu, Yu Wu, Z. F. Wu, Zhibin Gou, Zhihong Shao, Zhuoshu Li, Ziyi Gao, Aixin Liu, Bing Xue, Bingxuan Wang, Bochao Wu, Bei Feng, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, Damai Dai, Deli Chen, Dongjie Ji, Erhang Li, Fangyun Lin, Fucong Dai, Fuli Luo, Guangbo Hao, Guanting Chen, Guowei Li, H. Zhang, Han Bao, Hanwei Xu, Haocheng Wang, Honghui Ding, Huajian Xin, Huazuo Gao, Hui Qu, Hui Li, Jianzhong Guo, Jiashi Li, Jiawei Wang, Jingchang Chen, Jingyang Yuan, Junjie Qiu, Junlong Li, J. L.
+
+Cai, Jiaqi Ni, Jian Liang, Jin Chen, Kai Dong, Kai Hu, Kaige Gao, Kang Guan, Kexin Huang, Kuai Yu, Lean Wang, Lecong Zhang, Liang Zhao, Litong Wang, Liyue Zhang, Lei Xu, Leyi Xia, Mingchuan Zhang, Minghua Zhang, Minghui Tang, Meng Li, Miaojun Wang, Mingming Li, Ning Tian, Panpan Huang, Peng Zhang, Qiancheng Wang, Qinyu Chen, Qiushi Du, Ruiqi Ge, Ruisong Zhang, Ruizhe Pan, Runji Wang, R. J. Chen, R. L. Jin, Ruyi Chen, Shanghao Lu, Shangyan Zhou, Shanhuang Chen, Shengfeng Ye, Shiyu Wang, Shuiping Yu, Shunfeng Zhou, Shuting Pan, S. S. Li, Shuang Zhou, Shaoqing Wu, Shengfeng Ye, Tao Yun, Tian Pei, Tianyu Sun, T. Wang, Wangding Zeng, Wanjia Zhao, Wen Liu, Wenfeng Liang, Wenjun Gao, Wenqin Yu, Wentao Zhang, W. L. Xiao, Wei An, Xiaodong Liu, Xiaohan Wang, Xiaokang Chen, Xiaotao Nie, Xin Cheng, Xin Liu, Xin Xie, Xingchao Liu, Xinyu Yang, Xinyuan Li, Xuecheng Su, Xuheng Lin, X. Q. Li, Xiangyue Jin, Xiaojin Shen, Xiaosha Chen, Xiaowen Sun, Xiaoxiang Wang, Xinnan Song, Xinyi Zhou, Xianzu Wang, Xinxia Shan, Y. K. Li, Y. Q. Wang, Y. X. Wei, Yang Zhang, Yanhong Xu, Yao Li, Yao Zhao, Yaofeng Sun, Yaohui Wang, Yi Yu, Yichao Zhang, Yifan Shi, Yiliang Xiong, Ying He, Yishi Piao, Yisong Wang, Yixuan Tan, Yiyang Ma, Yiyuan Liu, Yongqiang Guo, Yuan Ou, Yuduan Wang, Yue Gong, Yuheng Zou, Yujia He, Yunfan Xiong, Yuxiang Luo, Yuxiang You, Yuxuan Liu, Yuyang Zhou, Y. X. Zhu, Yanhong Xu, Yanping Huang, Yaohui Li, Yi Zheng, Yuchen Zhu, Yunxian Ma, Ying Tang, Yukun Zha, Yuting Yan Z. Z. Ren, Zehui Ren, Zhangli Sha, Zhe Fu, Zhean Xu, Zhenda Xie, Zhengyan Zhang, Zhewen Hao, Zhicheng Ma, Zhigang Yan, Zhiyu Wu, Zihui Gu, Zijia Zhu, Zijun Liu, Zilin Li, Ziwei Xie, Ziyang Song, Zizheng Pan, Zhen Huang, Zhipeng Xu, Zhongyu Zhang and Zhen Zhang. 2025. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. Preprint arXiv:2501.12948.  
+Yiwen Ding, Zhiheng Xi, Wei He, Zhuoyuan Li, Yitao Zhai, Xiaowei Shi, Xunliang Cai, Tao Gui, Qi Zhang, and Xuanjing Huang. 2024. Mitigating tail narrowing in llm self-improvement via socratic-guided sampling. Preprint, arXiv:2411.00750.  
+Hanze Dong, Wei Xiong, Deepanshu Goyal, Yihan Zhang, Winnie Chow, Rui Pan, Shizhe Diao, Jipeng Zhang, Kashun Shum, and Tong Zhang. 2023. Raft: Reward ranked finetuning for generative foundation model alignment. arXiv preprint arXiv:2304.06767.  
+Vanessa Dye. 2011. Reflection, reflection, reflection. i'm thinking all the time, why do i need a theory or model of reflection?. Developing Reflective Practice: A guide for beginning teachers. Maidenhead: McGraw-Hill Education, pages 217-234.  
+Gerd Gigerenzer. 1991. From tools to theories: A heuristic of discovery in cognitive psychology. *Psychological review*, 98(2):254.  
+Aaron Grattafori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, and Abhishek Kadian et al.
+
+2024. The llama 3 herd of models. Preprint, arXiv:2407.21783.  
+Patrick Haluptzok, Matthew Bowers, and Adam Tauman Kalai. 2022. Language models can teach themselves to program better. arXiv preprint arXiv:2207.14502.  
+Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. 2021. Measuring massive multitask language understanding. Preprint, arXiv:2009.03300.  
+Minda Hu, Qiyuan Zhang, Yufei Wang, Bowei He, Hongru Wang, Jingyan Zhou, Liangyou Li, Yasheng Wang, Chen Ma, and Irwin King. 2024. Nile: Internal consistency alignment in large language models. Preprint, arXiv:2412.16686.  
+Nan Hu, Jiaoyan Chen, Yike Wu, Guilin Qi, Hongru WANG, Sheng Bi, Yongrui Chen, Tongtong Wu, and Jeff Z. Pan. 2025. Can LLMs Evaluate Complex Attribution in QA? Automatic Benchmarking using Knowledge Graphs. In In the Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (ACL 2025).  
+Jiaxin Huang, Shixiang Gu, Le Hou, Yuexin Wu, Xuezhi Wang, Hongkun Yu, and Jiawei Han. 2023. Large language models can self-improve. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 1051-1068, Singapore. Association for Computational Linguistics.  
+Albert Q. Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, Lélio Renard Lavaud, Marie-Anne Lachaux, Pierre Stock, Teven Le Scao, Thibaut Lavril, Thomas Wang, Timothée Lacroix, and William El Sayed. 2023. Mistral 7b. Preprint, arXiv:2310.06825.  
+Philip Nicholas Johnson-Laird. 1988. The computer and the mind: An introduction to cognitive science. Harvard University Press.  
+Ming Li, Lichang Chen, Jiuhai Chen, Shwai He, Jiaxiang Gu, and Tianyi Zhou. 2024a. Selective reflection-tuning: Student-selected data recycling for LLM instruction-tuning. In Findings of the Association for Computational Linguistics: ACL 2024, pages 16189-16211, Bangkok, Thailand. Association for Computational Linguistics.  
+Tianle Li, Wei-Lin Chiang, Evan Frick, Lisa Dunlap, Tianhao Wu, Banghua Zhu, Joseph E. Gonzalez, and Ion Stoica. 2024b. From crowdsourced data to high-quality benchmarks: Arena-hard and benchbuilder pipeline. Preprint, arXiv:2406.11939.  
+Xuechen Li, Tianyi Zhang, Yann Dubois, Rohan Taori, Ishaan Gulrajani, Carlos Guestrin, Percy Liang, and Tatsunori B. Hashimoto. 2023. Alpacaeval: An automatic evaluator of instruction-following models. https://github.com/tatsu-lab/alpaca_eval.
+
+Jianqiao Lu, Zhiyang Dou, WANG Hongru, Zeyu Cao, Jianbo Dai, Yunlong Feng, and Zhijiang Guo. 2024. Autopsy: Automated process-supervised verifier. In The Thirty-eighth Annual Conference on Neural Information Processing Systems.  
+OpenAI. 2024. Learning to reason with llms.  
+Jeff Z. Pan, Simon Razniewski, Jan-Christoph Kalo, Sneha Singhania, Jiaoyan Chen, Stefan Dietze, Hajira Jabeen, Janna Omeliyanenko, Wen Zhang, Matteo Lissandrini, Russia Biswas, Gerard de Melo, Angela Bonifati, Edlira Vakaj, Mauro Dragoni, and Damien Graux. 2023. Large Language Models and Knowledge Graphs: Opportunities and Challenges. Transactions on Graph Data and Knowledge, pages 1-38.  
+J.Z. Pan, D. Calvanese, T. Eiter, I. Horrocks, M. Kifer, F. Lin, and Y. Zhao, editors. 2017a. Reasoning Web: Logical Foundation of Knowledge Graph Construction and Querying Answering. Springer.  
+J.Z. Pan, G. Vetere, J.M. Gomez-Perez, and H. Wu, editors. 2017b. Exploiting Linked Data and Knowledge Graphs for Large Organisations. Springer.  
+Baolin Peng, Chunyuan Li, Pengcheng He, Michel Galley, and Jianfeng Gao. 2023. Instruction tuning with gpt-4. Preprint, arXiv:2304.03277.  
+Qwen. 2024. Qwq: Reflect deeply on the boundaries of the unknown. Accessed: 2025-02-01.  
+Zhili Shen, Chenxin Diao, Pavlos Vougiouklis, Pascual Merita, Shriram Piramanayagam, Enting Chen, Damien Graux, Andre Melo, Ruofei Lai, Zeren Jiang, Zhongyang Li, YE QI, Yang Ren, Dandan Tu, and Jeff Z. Pan. 2025. GeAR: Graph-enhanced Agent for Retrieval-augmented Generation. In In the Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (ACL 2025).  
+Mirac Suzgun, Nathan Scales, Nathanael Scharli, Sebastian Gehrmann, Yi Tay, Hyung Won Chung, Aakanksha Chowdhery, Quoc V Le, Ed H Chi, Denny Zhou, , and Jason Wei. 2022. Challenging big-bench tasks and whether chain-of-thought can solve them. arXiv preprint arXiv:2210.09261.  
+Rohan Taori, Ishaan Gulrajani, Tianyi Zhang, Yann Dubois, Xuechen Li, Carlos Guestrin, Percy Liang, and Tatsunori B. Hashimoto. 2023. Stanford alpaca: An instruction-following llama model. https://github.com/tatsu-lab/stanford_alpaca.  
+Kaustubh Vyas, Damien Graux, Yijun Yang, Sebastien Montella, Chenxin Diao, Wendi Zhou, Pavlos Vougiouklis, Ruofei Lai, Yang Ren, Keshuang Li, and Jeff Z. Pan. 2025. From an LLM Swarm to a PDDLempowered Hive: Planning Self-executed Instructions in a Multi-modal Jungle. In In the Proceedings of the International Conference on Learning Representations (ICLR 2025).
+
+Hongru Wang, Yujia Qin, Yankai Lin, Jeff Z. Pan, and Kam-Fai Wong. 2024a. Empowering large language models: Tool learning for real-world interaction. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR '24, page 2983-2986, New York, NY, USA. Association for Computing Machinery.  
+Hongru Wang, Huimin Wang, Lingzhi Wang, Minda Hu, Rui Wang, Boyang Xue, Yongfeng Huang, and Kam-Fai Wong. 2024b. Tpe: Towards better compositional reasoning over cognitive tools via multipersona collaboration. In *Natural Language Processing and Chinese Computing: 13th National CCF Conference*, NLPCC 2024, Hangzhou, China, November 1–3, 2024, Proceedings, Part II, page 281–294, Berlin, Heidelberg. Springer-Verlag.  
+Hongru Wang, Rui Wang, Fei Mi, Yang Deng, Zezhong Wang, Bin Liang, Ruifeng Xu, and Kam-Fai Wong. 2023a. Cue-CoT: Chain-of-thought prompting for responding to in-depth dialogue questions with LLMs. In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 12047-12064, Singapore. Association for Computational Linguistics.  
+Hongru Wang, Boyang Xue, Baohang Zhou, Tianhua Zhang, Cunxiang Wang, Huimin Wang, Guanhua Chen, and Kam-Fai Wong. 2025a. Self-DC: When to reason and when to act? self divide-and-conquer for compositional unknown questions. In Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), pages 6510-6525, Albuquerque, New Mexico. Association for Computational Linguistics.  
+Junjie Wang, Mingyang Chen, Binbin Hu, Dan Yang, Ziqi Liu, Yue Shen, Peng Wei, Zhiqiang Zhang, Jinjie Gu, Jun Zhou, Jeff Z. Pan, Wen Zhang, and Hua-jun Chen. 2024c. Learning to Plan for Retrieval-Augmented Large Language Models from Knowledge Graphs. In In the Findings of the Association for Computational Linguistics: EMNLP 2024 (EMNLP 2024 Findings), pages 7813-7835.  
+Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc Le, Ed Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou. 2023b. Self-consistency improves chain of thought reasoning in language models. Preprint, arXiv:2203.11171.  
+Yizhong Wang, Yeganeh Kordi, Swaroop Mishra, Alisa Liu, Noah A Smith, Daniel Khashabi, and Hannaneh Hajishirzi. 2023c. Self-instruct: Aligning language models with self-generated instructions. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 13484-13508.  
+Yizhong Wang, Swaroop Mishra, Pegah Alipoormolabashi, Yeganeh Kordi, Amirreza Mirzaei, Atharva Naik, Arjun Ashok, Arut Selvan Dhanasekaran, Anjana Arunkumar, David Stap, Eshaan Pathak,
+
+Giannis Karamanolakis, Haizhi Lai, Ishan Purohit, Ishani Mondal, Jacob Anderson, Kirby Kuznia, Krima Doshi, Kuntal Kumar Pal, Maitreya Patel, Mehrad Moradshahi, Mihir Parmar, Mirali Purohit, Neeraj Varshney, Phani Rohitha Kaza, Pulkit Verma, Ravsehaj Singh Puri, Rushang Karia, Savan Doshi, Shailaja Keyur Sampat, Siddhartha Mishra, Sujan Reddy A, Sumanta Patro, Tanay Dixit, and Xudong Shen. 2022. Super-Natural Instructions: Generalization via declarative instructions on  $1600+$  NLP tasks. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 5085-5109, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics.  
+Yue Wang, Qiuzhi Liu, Jiahao Xu, Tian Liang, Xingyu Chen, Zhiwei He, Linfeng Song, Dian Yu, Juntao Li, Zhuosheng Zhang, Rui Wang, Zhaopeng Tu, Haitao Mi, and Dong Yu. 2025b. Thoughts are all over the place: On the underthinking of o1-like llms. Preprint, arXiv:2501.18585.  
+Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. 2022. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824-24837.  
+Can Xu, Qingfeng Sun, Kai Zheng, Xiubo Geng, Pu Zhao, Jiazhan Feng, Chongyang Tao, and Daxin Jiang. 2023. Wizardlm: Empowering large language models to follow complex instructions. Preprint, arXiv:2304.12244.  
+Zhangchen Xu, Fengqing Jiang, Luyao Niu, Yuntian Deng, Radha Poovendran, Yejin Choi, and Bill Yuchen Lin. 2024. Magpie: Alignment data synthesis from scratch by prompting aligned llms with nothing. Preprint, arXiv:2406.08464.  
+Weizhe Yuan, Richard Yuanzhe Pang, Kyunghyun Cho, Xian Li, Sainbayar Sukhbaatar, Jing Xu, and Jason Weston. 2024. Self-rewarding language models. Preprint, arXiv:2401.10020.  
+Rowan Zellers, Ari Holtzman, Yonatan Bisk, Ali Farhadi, and Yejin Choi. 2019. Hellaswag: Can a machine really finish your sentence? Preprint, arXiv:1905.07830.  
+Yaowei Zheng, Richong Zhang, Junhao Zhang, Yanhan Ye, and Zheyan Luo. 2024. LlamaFactory: Unified efficient fine-tuning of  $100+$  language models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 3: System Demonstrations), pages 400-410, Bangkok, Thailand. Association for Computational Linguistics.  
+Pei Zhou, Jay Pujara, Xiang Ren, Xinyun Chen, HengTze Cheng, Quoc V Le, Ed H Chi, Denny Zhou, Swaroop Mishra, and Huaixiu Steven Zheng. 2024. Self-discover: Large language models self-compose reasoning structures. arXiv preprint arXiv:2402.03620.  
+Xiongtao Zhou, Jie He, Lanyu Chen, Jingyu Li, Haojing Chen, Víctor Gutiérrez-Basulto, Jeff Z. Pan, and
+
+Hanjie Chen. 2025. MiCEval: Unveiling Multimodal Chain of Thought's Quality via Image Description and Reasoning Steps. In In the Proceedings of 2025 Annual Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics (NAACL 2025).
+
+# A Appendix
+
+# A.1 Prompt Details
+
+Table 5 and Table 6 show the instruction and system prompt respectively used to instruct LLMs to unfold the hidden reasoning chains with the augmentation of various meta-reasoning skills.
+
+Here is the given question: {ins}
+
+Here is the original reasoning: {ans}
+
+Table 5: The instruction used to prompt LLM to generate the enriched rationales.
+
+# A.2 Experimental Logs
+
+Reflection-tuning. MMLU:  $y = 6.4535 * \log(x) + 60.5418$ ; GSM8K:  $y = 6.0944 * \log(x) + 69.6957$ ; ARC-C:  $y = 4.8088 * \log(x) + 71.8375$ ; HellaSwag:  $y = 9.2765 * \log(x) + 54.6557$ ; BBH:  $y = 11.1345 * \log(x) + 34.2507$ ; Avg.:  $y = 7.5551 * \log(x) + 58.1925$
+
+SRLM. MMLU:  $y = 9.0256 * \log(x) + 62.8932$  GSM8K:  $y = 6.5905 * \log(x) + 71.4682$  ARC-C:  $y = 5.9295 * \log(x) + 78.0043$  HellaSwag:  $y = 10.9284 * \log(x) + 57.7607$  BBH:  $y = 11.9599 * \log(x) + 38.7086$  Avg.:  $y = 8.8870 * \log(x) + 61.7671$
+
+# A.3 Performance on Qwen-14B
+
+We also conduct experiments on Qwen-14B models following the same setting in main experiments. Table 8 shows the final results. It is observed that SRLM initially exhibits higher performance at the 0th iteration but then gradually degrades. We attribute this to the ineffective of reasoning catalyst data to guide the reasoning processing since Qwen2.5-14B already achieves comparable performance against the used 4o models to generate reasoning catalyst data.
+
+# A.4 Case Study
+
+Table 13 shows the non-cherry pick test examples of MATH. It is obvious that our proposed SRLM can take advantages of different meta-reasoning skills such as decomposition, detail, check, reflection and summary to get the correct answer.
+
+![](images/85a680b385d24fd83348e5dcda7a146c96b563682b7a7aff2aae66e66311f708.jpg)  
+Figure 9: The distribution of meta-reasoning skills on five various benchmarks and the global distribution during inference.
+
+You are an expert at meta reasoning theory from cognitive science. Given the question, corresponding summarized reasoning and answer, you can always uncover hidden or unspoken reasoning, even when it isn't explicitly stated. You need to add any missing reasoning thoughts that you think it is helpful or may occurs to understand and solve this question based on given summarized reasoning result. Your new reasoning should be more comprehensive, detailed and clear. Your answer should follow the format like <thoughts> your new reasoning here </thoughts>
+
+Inside <thoughts> </thoughts>, you need to explicitly indicate which meta reasoning skill is used, some of them are shown below, but note you can use anything else if you think it is helpful or it naturally occurs when you solve this question.
+
+<decomposition>: breaking down a complex problem into smaller, more manageable parts. Making sure that you also provide answers for all decomposed problems in this section. You can decompose iteratively but should not contain same problem or exceed the max iteration depth which is three.
+
+<backward>: starting with the desired observations at any previous reasoning step and working backward to identify the new reasoning directions.
+
+<detail>: any details including but not limited to logic and reasons for your reasoning in this way, you are encouraged to add this at every unclear or unnatural reasoning step.
+
+<summary>: summarize your reasoning to help future thinking.
+
+<alternatives>: directly thinking in other ways, try to explore different solutions as much as possible to solve given problem.
+
+<reflection>: you are encouraged to regularly reflect on your past reasoning in current response at various levels of detail, from sentence down to individual word. This will help you better understand and think through problems. It's okay to make mistakes; use them as opportunities to learn and improve.
+
+<analogy>: you are encouraged to regularly consider other analogous problems with the problem you've encountered at various reasoning steps, along with their solutions. Reference existing theories or methods that guided your approach to solving these problems. These similar problems can be at various levels of detail - from larger overarching issues down to smaller sub-problems you encountered along the way. The key is to demonstrate a diverse range of problems and solutions, to show how you have approached and resolved challenges that are analogous to the current situation.
+
+<check>: consider different edge cases or test cases carefully.
+
+<other>: other meta reasoning skills you think is helpful or worthy to try to solve the task.
+
+# Notice:
+
+1. All tags must be properly invoked and closed, using the format like  $\langle \text{reflection} \rangle$  and  $\langle \text{/reflection} \rangle$ .  
+2. You should always use first-person perspective.  
+3. You can add any new meta reasoning skill at any positions except <reflection>. Note <reflection> can not be invoked without any reasoning in current response and it can be invoked at any positions when you already have some reasoning results.  
+4. You cannot change the original reasoning. However, if you identify any errors or improvements in the reasoning, you can add new reasoning steps using above meta-reasoning skills afterwards to correct or clarify the path, ensuring a better understanding and solution.  
+5. You can apply the same reasoning skills multiple times or use different skills simultaneously.  
+6. Your answer should start with <thoughts>, and end with </thoughts>.
+
+Table 6: The system prompt used to initialize LLM to unfold all hidden reasoning chains.
+
+<table><tr><td>Models</td><td>Methods</td><td>MMLU</td><td>GSM8K</td><td>ARC-C</td><td>HellaSwag</td><td>BBH</td><td>Avg.</td></tr><tr><td rowspan="21">Mistral-7B</td><td>Magpie</td><td>44.35</td><td>41.02</td><td>55.72</td><td>37.61</td><td>27.13</td><td>41.17</td></tr><tr><td>Reflection-tuning</td><td>49.85</td><td>56.56</td><td>61.69</td><td>45.15</td><td>27.29</td><td>48.11</td></tr><tr><td>IT0 + Cat. (M0)</td><td>51.63</td><td>56.03</td><td>63.05</td><td>40.82</td><td>29.97</td><td>48.30</td></tr><tr><td>Length Selector</td><td></td><td></td><td></td><td></td><td></td><td></td></tr><tr><td>IT1 + Cat. (M1)</td><td>49.50</td><td>58.91</td><td>62.37</td><td>45.91</td><td>31.33</td><td>49.60</td></tr><tr><td>IT2 + Cat. (M2)</td><td>50.01</td><td>59.59</td><td>64.85</td><td>42.16</td><td>29.82</td><td>49.29</td></tr><tr><td>IT3 + Cat. (M3)</td><td>50.56</td><td>56.41</td><td>64.59</td><td>46.88</td><td>32.11</td><td>50.11</td></tr><tr><td>IT4 + Cat. (M4)</td><td>49.07</td><td>58.45</td><td>60.41</td><td>42.03</td><td>29.35</td><td>47.86</td></tr><tr><td>IT5 + Cat. (M5)</td><td>50.66</td><td>56.79</td><td>66.13</td><td>47.30</td><td>31.87</td><td>50.55</td></tr><tr><td>Off-policy Selector</td><td></td><td></td><td></td><td></td><td></td><td></td></tr><tr><td>IT1 + Cat. (M1)</td><td>49.55</td><td>58.07</td><td>65.02</td><td>41.02</td><td>32.82</td><td>49.30</td></tr><tr><td>IT2 + Cat. (M2)</td><td>50.63</td><td>57.92</td><td>62.12</td><td>41.98</td><td>32.49</td><td>49.03</td></tr><tr><td>IT3 + Cat. (M3)</td><td>49.74</td><td>55.95</td><td>65.02</td><td>42.28</td><td>32.12</td><td>49.02</td></tr><tr><td>IT4 + Cat. (M4)</td><td>50.19</td><td>56.56</td><td>60.24</td><td>42.55</td><td>31.48</td><td>48.20</td></tr><tr><td>IT5 + Cat. (M5)</td><td>50.79</td><td>57.77</td><td>64.25</td><td>45.66</td><td>32.00</td><td>50.09</td></tr><tr><td>On-policy Selector</td><td></td><td></td><td></td><td></td><td></td><td></td></tr><tr><td>IT1 + Cat. (M1)</td><td>50.00</td><td>58.68</td><td>64.59</td><td>41.27</td><td>29.64</td><td>48.84</td></tr><tr><td>IT2 + Cat. (M2)</td><td>51.69</td><td>57.47</td><td>66.30</td><td>42.29</td><td>30.77</td><td>49.70</td></tr><tr><td>IT3 + Cat. (M3)</td><td>50.28</td><td>56.33</td><td>64.59</td><td>46.44</td><td>30.36</td><td>49.60</td></tr><tr><td>IT4 + Cat. (M4)</td><td>51.66</td><td>55.34</td><td>65.36</td><td>44.55</td><td>29.07</td><td>49.20</td></tr><tr><td>IT5 + Cat. (M5)</td><td>49.54</td><td>55.57</td><td>62.63</td><td>43.00</td><td>32.00</td><td>48.55</td></tr><tr><td rowspan="21">Llama3.1-8B</td><td>Magpie</td><td>54.34</td><td>44.81</td><td>68.77</td><td>44.05</td><td>25.10</td><td>47.41</td></tr><tr><td>Reflection-tuning</td><td>57.08</td><td>66.49</td><td>67.32</td><td>50.34</td><td>30.09</td><td>54.26</td></tr><tr><td>IT0 + Cat. (M0)</td><td>57.60</td><td>64.59</td><td>70.14</td><td>49.23</td><td>31.46</td><td>54.60</td></tr><tr><td>Length Selector</td><td></td><td></td><td></td><td></td><td></td><td></td></tr><tr><td>IT1 + Cat. (M1)</td><td>58.31</td><td>65.73</td><td>72.87</td><td>48.05</td><td>34.08</td><td>55.81</td></tr><tr><td>IT2 + Cat. (M2)</td><td>58.27</td><td>69.22</td><td>72.70</td><td>51.99</td><td>32.03</td><td>56.84</td></tr><tr><td>IT3 + Cat. (M3)</td><td>57.91</td><td>67.25</td><td>72.70</td><td>52.88</td><td>34.06</td><td>56.96</td></tr><tr><td>IT4 + Cat. (M4)</td><td>57.27</td><td>67.17</td><td>72.87</td><td>49.83</td><td>33.44</td><td>56.12</td></tr><tr><td>IT5 + Cat. (M5)</td><td>56.17</td><td>65.73</td><td>69.45</td><td>50.27</td><td>33.75</td><td>55.07</td></tr><tr><td>Off-policy Selector</td><td></td><td></td><td></td><td></td><td></td><td></td></tr><tr><td>IT1 + Cat. (M1)</td><td>58.26</td><td>67.70</td><td>74.23</td><td>46.11</td><td>32.14</td><td>55.69</td></tr><tr><td>IT2 + Cat. (M2)</td><td>58.50</td><td>67.25</td><td>73.46</td><td>48.03</td><td>31.28</td><td>55.70</td></tr><tr><td>IT3 + Cat. (M3)</td><td>58.30</td><td>65.43</td><td>73.04</td><td>46.93</td><td>30.37</td><td>54.81</td></tr><tr><td>IT4 + Cat. (M4)</td><td>58.57</td><td>67.40</td><td>74.15</td><td>46.93</td><td>30.03</td><td>55.42</td></tr><tr><td>IT5 + Cat. (M5)</td><td>57.72</td><td>66.79</td><td>75.51</td><td>47.40</td><td>31.12</td><td>55.71</td></tr><tr><td>On-policy Selector</td><td></td><td></td><td></td><td></td><td></td><td></td></tr><tr><td>IT1 + Cat. (M1)</td><td>57.96</td><td>65.20</td><td>70.05</td><td>45.07</td><td>29.69</td><td>53.59</td></tr><tr><td>IT2 + Cat. (M2)</td><td>58.53</td><td>64.67</td><td>69.20</td><td>48.41</td><td>31.32</td><td>54.43</td></tr><tr><td>IT3 + Cat. (M3)</td><td>58.26</td><td>66.41</td><td>72.35</td><td>50.43</td><td>32.89</td><td>56.07</td></tr><tr><td>IT4 + Cat. (M4)</td><td>58.45</td><td>68.76</td><td>70.14</td><td>46.54</td><td>33.37</td><td>55.45</td></tr><tr><td>IT5 + Cat. (M5)</td><td>59.17</td><td>66.19</td><td>69.88</td><td>50.76</td><td>33.01</td><td>55.80</td></tr></table>
+
+Table 7: The detailed experimental results of each iteration.
+
+<table><tr><td>Models</td><td>Methods</td><td>MMLU</td><td>GSM8K</td><td>ARC-C</td><td>HellaSwag</td><td>BBH</td><td>Avg.</td></tr><tr><td rowspan="6">Qwen2.5-14B</td><td>Reflection-tuning</td><td>78.72</td><td>92.41</td><td>93.51</td><td>71.86</td><td>57.56</td><td>78.81</td></tr><tr><td>IT0 + Cat. (M0)</td><td>77.89</td><td>92.64</td><td>93.17</td><td>74.24</td><td>59.69</td><td>79.53</td></tr><tr><td>Length Selector</td><td></td><td></td><td></td><td></td><td></td><td></td></tr><tr><td>IT1 + Cat. (M1)</td><td>77.30</td><td>92.66</td><td>91.72</td><td>68.44</td><td>58.47</td><td>77.72</td></tr><tr><td>IT2 + Cat. (M2)</td><td>76.97</td><td>92.34</td><td>92.32</td><td>67..65</td><td>57.08</td><td>77.65</td></tr><tr><td>IT3 + Cat. (M3)</td><td>76.28</td><td>92.04</td><td>90.44</td><td>65.70</td><td>58.13</td><td>76.52</td></tr></table>
+
+Table 8: The detailed experimental results of each iteration on Qwen-14B with length selector.  
+
+<table><tr><td>Data</td><td>MMLU</td><td>GSM8K</td><td>ARC-C</td><td>HellaSwag</td><td>BBH</td><td>Avg.</td></tr><tr><td>Reflection-tuning</td><td>57.08</td><td>66.49</td><td>67.32</td><td>50.34</td><td>30.09</td><td>54.26</td></tr><tr><td>IT1</td><td>57.74</td><td>67.93</td><td>74.06</td><td>49.22</td><td>32.94</td><td>56.38</td></tr><tr><td>IT2</td><td>57.87</td><td>66.87</td><td>72.53</td><td>48.63</td><td>30.46</td><td>55.27</td></tr><tr><td>IT3</td><td>57.51</td><td>68.46</td><td>73.21</td><td>51.59</td><td>31.61</td><td>56.48</td></tr></table>
+
+Table 9: The results of Llama3.1-8B on instruction-tuning dataset (collected by length selector) of each iteration.  
+
+<table><tr><td>Data</td><td>MMLU</td><td>GSM8K</td><td>ARC-C</td><td>HellaSwag</td><td>BBH</td><td>Avg.</td></tr><tr><td>SRLM - M0</td><td></td><td></td><td></td><td></td><td></td><td></td></tr><tr><td>Cat. = 5k</td><td>58.47</td><td>67.10</td><td>69.54</td><td>47.13</td><td>27.98</td><td>54.04</td></tr><tr><td>Cat. = 10k</td><td>59.22</td><td>66.26</td><td>67.58</td><td>46.68</td><td>30.78</td><td>54.10</td></tr><tr><td>SRLM - M1</td><td></td><td></td><td></td><td></td><td></td><td></td></tr><tr><td>Cat. = 5k</td><td>58.04</td><td>66.41</td><td>75.17</td><td>45.71</td><td>31.47</td><td>55.36</td></tr><tr><td>Cat. = 10k</td><td>58.56</td><td>66.72</td><td>75.00</td><td>48.00</td><td>30.02</td><td>55.66</td></tr><tr><td>SRLM - M2</td><td></td><td></td><td></td><td></td><td></td><td></td></tr><tr><td>Cat. = 5k</td><td>58.05</td><td>68.46</td><td>72.70</td><td>44.76</td><td>29.28</td><td>54.65</td></tr><tr><td>Cat. = 10k</td><td>57.84</td><td>67.70</td><td>73.04</td><td>50.79</td><td>28.70</td><td>55.61</td></tr><tr><td>SRLM - M3</td><td></td><td></td><td></td><td></td><td></td><td></td></tr><tr><td>Cat. = 5k</td><td>57.35</td><td>65.96</td><td>73.04</td><td>45.52</td><td>30.73</td><td>54.52</td></tr><tr><td>Cat. = 10k</td><td>57.29</td><td>67.32</td><td>73.38</td><td>48.65</td><td>29.23</td><td>55.17</td></tr></table>
+
+Table 10: The results using different size of reasoning catalyst data based on Llama3.1-8B model with 3 epochs optimization on three iterations  $(\mathcal{M}_{0,1,2,3})$ . We bold the best performance at each iteration.  
+
+<table><tr><td>Data</td><td>MMLU</td><td>GSM8K</td><td>ARC-C</td><td>HellaSwag</td><td>BBH</td><td>Avg.</td></tr><tr><td>IT0 + Cat.</td><td>57.60</td><td>64.59</td><td>70.14</td><td>49.23</td><td>31.46</td><td>54.60</td></tr><tr><td colspan="7">Update rj* with rt+1</td></tr><tr><td>IT1 + Up. Cat.1</td><td>59.27</td><td>66.03</td><td>74.15</td><td>49.84</td><td>31.29</td><td>56.12</td></tr><tr><td>IT2 + Up. Cat.2</td><td>57.42</td><td>65.81</td><td>71.08</td><td>47.57</td><td>29.45</td><td>54.27</td></tr><tr><td>IT3 + Up. Cat.3</td><td>55.52</td><td>65.13</td><td>69.28</td><td>49.25</td><td>30.92</td><td>54.02</td></tr><tr><td colspan="7">Update both r0j, r* with rtj, rt+1</td></tr><tr><td>IT1 + Up. Cat.1</td><td>56.99</td><td>64.75</td><td>72.87</td><td>47.94</td><td>33.83</td><td>55.28</td></tr><tr><td>IT2 + Up. Cat.2</td><td>57.18</td><td>63.99</td><td>67.92</td><td>49.78</td><td>33.11</td><td>54.40</td></tr><tr><td>IT3 + Up. Cat.3</td><td>55.09</td><td>65.13</td><td>68.17</td><td>49.42</td><td>32.57</td><td>54.08</td></tr></table>
+
+Table 11: The performance of SRLM with length selector on Llama3.1-8B using Update 1k reasoning catalyst data with 1000 steps optimization.
+
+<table><tr><td># Sampling</td><td>Methods</td><td>MMLU</td><td>GSM8K</td><td>ARC-C</td><td>HellaSwag</td><td>BBH</td><td>Avg.</td></tr><tr><td rowspan="2">1</td><td>Reflection-tuning</td><td>57.08</td><td>66.49</td><td>67.32</td><td>50.34</td><td>30.09</td><td>54.26</td></tr><tr><td>SRLM</td><td>57.91</td><td>67.25</td><td>72.70</td><td>52.88</td><td>34.06</td><td>56.96</td></tr><tr><td rowspan="2">2</td><td>Reflection-tuning</td><td>65.50</td><td>74.07</td><td>76.02</td><td>60.88</td><td>41.55</td><td>63.60</td></tr><tr><td>SRLM</td><td>69.56</td><td>75.89</td><td>83.02</td><td>64.25</td><td>46.61</td><td>67.87</td></tr><tr><td rowspan="2">4</td><td>Reflection-tuning</td><td>71.66</td><td>80.52</td><td>82.00</td><td>70.33</td><td>52.48</td><td>71.40</td></tr><tr><td>SRLM</td><td>78.44</td><td>83.85</td><td>89.51</td><td>76.11</td><td>57.74</td><td>77.13</td></tr><tr><td rowspan="2">8</td><td>Reflection-tuning</td><td>76.42</td><td>84.38</td><td>84.56</td><td>77.81</td><td>61.09</td><td>76.85</td></tr><tr><td>SRLM</td><td>85.41</td><td>88.32</td><td>93.94</td><td>85.46</td><td>68.23</td><td>84.27</td></tr><tr><td rowspan="2">16</td><td>Reflection-tuning</td><td>79.79</td><td>88.10</td><td>85.75</td><td>82.61</td><td>67.56</td><td>80.76</td></tr><tr><td>SRLM</td><td>90.31</td><td>91.89</td><td>96.33</td><td>91.67</td><td>74.87</td><td>89.01</td></tr><tr><td rowspan="2">32</td><td>Reflection-tuning</td><td>82.48</td><td>90.75</td><td>88.05</td><td>86.41</td><td>72.79</td><td>84.10</td></tr><tr><td>SRLM</td><td>93.85</td><td>93.63</td><td>98.12</td><td>95.41</td><td>80.23</td><td>92.25</td></tr><tr><td rowspan="2">64</td><td>Reflection-tuning</td><td>84.80</td><td>92.27</td><td>89.16</td><td>89.24</td><td>76.27</td><td>86.35</td></tr><tr><td>SRLM</td><td>96.15</td><td>95.38</td><td>98.72</td><td>97.62</td><td>83.31</td><td>94.24</td></tr></table>
+
+Table 12: The detailed experimental results with multiple sampling for baseline and SRLM.
+
+Question: Suppose  $a$  and  $b$  are positive integers, neither of which is a multiple of 3. Find the least possible remainder when  $a^2 + b^2$  is divided by 3.
+
+Reflection-tuning Answer: To solve this problem, we need to find the remainder of  $a^2 + b^2$  when divided by 3, given that  $a$  and  $b$  are integers not divisible by 3.
+
+First, consider the properties of numbers modulo 3. Any integer  $n$  can be expressed as  $n = 3k + r$  where  $r$  is the remainder when  $n$  is divided by 3. Since  $a$  and  $b$  are not multiples of 3, req0.
+
+Now, let's express  $a$  and  $b$  in terms of their remainders when divided by 3:  $a = 3m + r_1$  and  $b = 3n + r_2$ , where  $r_1, r_2 \equiv 0$ .
+
+Substitute these into  $a^2 + b^2$ :  $a^2 + b^2 = (3m + r_1)^2 + (3n + r_2)^2 = 9m^2 + 6mr_1 + r_1^2 + 9n^2 + 6nr_2 + r_2^2$ .
+
+Notice that all terms containing  $m$  and  $n$  are multiples of 3, and thus their sum is also a multiple of 3. The only terms not divisible by 3 are  $r_1^2$  and  $r_2^2$ .
+
+Since  $r_1$  and  $r_2$  are not zero,  $r_1^2$  and  $r_2^2$  can each be 1 or 2 (since they are squares of non-zero remainders when divided by 3). The smallest possible value for  $r_1^2 + r_2^2$  is 1 (when both  $r_1$  and  $r_2$  are 1 or both are 2) and the largest possible value is 4 (when one is 1 and the other is 2). Thus, the remainder of  $a^2 + b^2$  when divided by 3 can only be 1 or 2. Therefore, the possible remainders when  $a^2 + b^2$  is divided by 3 are  $\boxed{1,2}$ .
+
+# SRLM: <thoughts>
+
+<decomposition> Let's break down the problem into smaller parts to ensure clarity and accuracy in our reasoning. First, we need to understand the behavior of squares modulo 3. Next, we'll examine the possible values of a and b under the constraint that they're not multiples of 3. Finally, we'll compute  $(a^2 + b^2 \mod 3)$  for these values and determine the remainder.</decomposition>
+
+<detail> Understanding the behavior of numbers under modulo 3 involves recognizing that any integer  $n$  can be expressed as  $n = 3q + r$ , where  $q$  is the quotient and  $r$  is the remainder. For  $n^2$ , we have  $n^2 = (3q + r)^2 = 9q^2 + 6qr + r^2$ . Simplifying gives us  $n^2 \equiv r^2 \pmod{3}$ . Thus, the possible remainders for  $n^2$  are 0, 1, or 2 depending on  $r$ . </detail> <detail> Since neither  $a$  nor  $b$  is a multiple of 3, they can take values of 1 or 2 modulo 3. This means  $a \equiv 1or2 \pmod{3}$  and similarly for  $b$ . We need to evaluate  $a^2 + b^2$  for these possibilities. </detail>
+
+<check> Let's verify each case: - Case 1:  $a \equiv 1 \pmod{3}$ ,  $b \equiv 1 \pmod{3}$ . Then  $a^2 \equiv 1^2 \equiv 1 \pmod{3}$ .  $b^2 \equiv 1^2 \equiv 1 \pmod{3}$ . So,  $a^2 + b^2 \equiv 1 + 1 \equiv 2 \pmod{3}$
+
+- Case 2:  $a \equiv 1 \pmod{3}$ ,  $b \equiv 2 \pmod{3}$  - Then  $a^2 \equiv 1^2 \equiv 1 \pmod{3} - b^2 \equiv 2^2 \equiv 4 \equiv 1 \pmod{3} - \mathrm{So}$ ,  $a^2 + b^2 \equiv 1 + 1 \equiv 2 \pmod{3}$
+
+- Case 3:  $a \equiv 2 \pmod{3}$ ,  $b \equiv 1 \pmod{3}$  - Then  $a^2 \equiv 2^2 \equiv 4 \equiv 1 \pmod{3}$  -  $b^2 \equiv 1^2 \equiv 1 \pmod{3}$  - So,  $a^2 + b^2 \equiv 1 + 1 \equiv 2 \pmod{3}$
+
+- Case 4:  $a \equiv 2 \pmod{3}, b \equiv 2 \pmod{3}$  - Then  $a^2 \equiv 2^2 \equiv 4 \equiv 1 \pmod{3} - b^2 \equiv 2^2 \equiv 4 \equiv 1 \pmod{3} - \mathrm{So}$ ,  $a^2 + b^2 \equiv 1 + 1 \equiv 2 \pmod{3} < \sqrt{}$  check>
+
+<summary> In all cases, the remainder when  $a^2 + b^2$  is divided by 3 is consistently 2. There were no overlooked scenarios or computational errors, confirming the correctness of the result. </summary>
+
+<reflection> By systematically evaluating each case and cross-verifying calculations, I ensured that the solution aligns with the conditions provided. The consistency in outcomes confirms the robustness of the method applied. </reflection> Therefore, the least possible remainder when  $a^2 + b^2$  is divided by 3 is  $\boxed{2}$ .</thoughts>
+
+Based on the above processing, the key reasoning steps can be summarized clearly, followed by the concise answer. To solve this problem, we need to understand the behavior of squares modulo 3. Any integer  $n$  can be expressed as  $n = 3q + r$ , where  $q$  is the quotient and  $r$  is the remainder when  $n$  is divided by 3. Squaring both sides gives  $n^2 = (3q + r)^2 = 9q^2 + 6qr + r^2$ . This simplifies to  $n^2 \equiv r^2 \pmod{3}$ . Therefore, the possible remainders for  $n^2$  are 0, 1, or 2 depending on  $r$ .
+
+Since neither  $a$  nor  $b$  is a multiple of 3, they can be either 1 or 2 modulo 3. We need to calculate  $a^2 + b^2$  for these cases: 1. If  $a \equiv 1 \pmod{3}$  and  $b \equiv 1 \pmod{3}$ , then  $a^2 \equiv 1^2 \equiv 1 \pmod{3}$  and  $b^2 \equiv 1^2 \equiv 1 \pmod{3}$ . So,  $a^2 + b^2 \equiv 1 + 1 \equiv 2 \pmod{3}$ . 2. If  $a \equiv 1 \pmod{3}$  and  $b \equiv 2 \pmod{3}$ , then  $a^2 \equiv 1^2 \equiv 1 \pmod{3}$  and  $b^2 \equiv 2^2 \equiv 1 \pmod{3}$ . So,  $a^2 + b^2 \equiv 1 + 1 \equiv 2 \pmod{3}$ . 3. If  $a \equiv 2 \pmod{3}$  and  $b \equiv 1 \pmod{3}$ , then  $a^2 \equiv 2^2 \equiv 1 \pmod{3}$  and  $b^2 \equiv 1^2 \equiv 1 \pmod{3}$ . So,  $a^2 + b^2 \equiv 1 + 1 + 1 \equiv 2 \pmod{3}$ . 4. If  $a \equiv 2 \pmod{3}$  and  $b \equiv 2 \pmod{3}$ , then  $a^2 \equiv 2^2 \equiv 1 \pmod{3}$  and  $b^2 \equiv 2^2 \equiv 1 \pmod{3}$ . So,  $a^2 + b^2 \equiv 1 + 1 + 1 \equiv 2 \pmod{3}$ . In all cases, the remainder when  $a^2 + b^2$  is divided by 3 is 2. Therefore, the least possible remainder is  $\boxed{2}$ .
+
+Table 13: The (non-cherry-pick) test instance from MATH.

2025/Self-Reasoning Language Models_ Unfold Hidden Reasoning Chains with Few Reasoning Catalyst/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5c342aee46cd549c6060a5bfc9bd5da97c619548e0fdfe445711288e37339ae5
+size 1136449

2025/Self-Rewarding Large Vision-Language Models for Optimizing Prompts in Text-to-Image Generation/760dd0a6-270e-46cf-a0b7-a4749d840a76_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:25900b8ebc01571cc0a44f7260194b09b79ae4058fd07f6f1c521997e4dbd93b
+size 7280689

2025/Self-Rewarding Large Vision-Language Models for Optimizing Prompts in Text-to-Image Generation/full.md

@@ -0,0 +1,534 @@
+# Self-Rewarding Large Vision-Language Models for Optimizing Prompts in Text-to-Image Generation
+
+Hongji Yang, Yucheng Zhou, Wencheng Han, Jianbing Shen*
+
+SKL-IOTSC, CIS, University of Macau
+
+yc47942@um.edu.mo, yucheng.zhou@connect.um.edu.mo, wencheng256@gmail.com
+
+# Abstract
+
+Text-to-image models are powerful for producing high-quality images based on given text prompts, but crafting these prompts often requires specialized vocabulary. To address this, existing methods train rewriting models with supervision from large amounts of manually annotated data and trained aesthetic assessment models. To alleviate the dependence on data scale for model training and the biases introduced by trained models, we propose a novel prompt optimization framework, designed to rephrase a simple user prompt into a sophisticated prompt to a text-to-image model. Specifically, we employ the large vision language models (LVLMs) as the solver to rewrite the user prompt, and concurrently, employ LVLMs as a reward model to score the aesthetics and alignment of the images generated by the optimized prompt. Instead of laborious human feedback, we exploit the prior knowledge of the LVLM to provide rewards, i.e., AI feedback. Simultaneously, the solver and the reward model are unified into one model and iterated in reinforcement learning to achieve self-improvement by giving a solution and judging itself. Results on two popular datasets demonstrate that our method outperforms other strong competitors.
+
+# 1 Introduction
+
+Text-to-image models (Rombach et al., 2022; Saharia et al., 2022; Yu et al., 2022) can generate diverse high-quality images based on user-provided prompts. However, effective interaction with these models requires users to possess specific expertise, including familiarity with specialized vocabularies, e.g., "35mm" for camera parameters and "art
+
+![](images/547637fcbe30b5bab2136240410789cf77e768c8c68d9b9a0c5e27e9b05d6f5c.jpg)  
+(a) Our Prompt Optimization Pipeline (Training Phase)
+
+![](images/401e29b5f3c19c8faee2004b7e672905d7c01942ade137b6da621b2b79c1cb97.jpg)  
+(b) Our Prompt Optimization Pipeline (Inference Phase)  
+Figure 1: The motivation of our prompt optimization pipeline. When training, the large model continuously improves prompt rewriting and image quality evaluation capabilities through self-play without any external sources or models. The generated images from the user prompts and the rewritten prompts. It can be observed that the image from the modified prompt has higher aesthetics.
+
+by Greg Rutkowski" to invoke a particular artistic style. As shown in Figure 1, rewriting prompts according to model-specific knowledge significantly improves the quality of the generated images.
+
+To bridge the gap between laymen and experts in using text-to-image models, some methods (Liu and Chilton, 2022; Datta et al., 2023; Oppenlaender, 2023) employ large-scale human-annotated datasets to train rewriting models that produce more professional and effective prompts. However, these approaches are prohibitively expensive. To reduce reliance on high-quality human annotations, some efforts (Hao et al., 2024; Rosenman et al., 2023; Cao et al., 2023) leverage specific calculated metrics, i.e., aesthetics and alignment, which are then regarded as rewarding for reinforcement learning (RL). However, collecting high-quality human-annotated datasets for reward model training is
+
+both time-consuming and costly. Large Vision-Language Models (LVLMs) have recently shown remarkable vision reasoning capabilities. Consequently, some studies (Li et al., 2022; Liu et al., 2024; Chen et al., 2024a; Yang et al., 2023; Yu et al., 2023) employ LVLMs as evaluators to assess human preferences. These methods effectively deliver interpretable AI feedback, offering a more efficient alternative to the time-consuming and labor-intensive human feedback (Yuan et al., 2024a).
+
+Besides, previous prompt optimization methods using RL suffer from two limitations: (1) They require extensive training data to train an image reward model; (2) The reward model remains fixed during the proxy model training, preventing it from learning and improving alongside the proxy model. This limitation results in a lack of dynamic feedback throughout the training process. To mitigate the data limitation and explore dynamic feedback, Yuan et al. (2024b) introduce a self-rewarding training strategy, enabling the model to train effectively with limited data while approximating the upper performance bound. It fosters self-improvement or self-play (Chen et al., 2024b), wherein the solver generates its judgments or rewards by a continuous iterative DPO (Xu et al., 2023).
+
+In this study, we introduce a self-rewarding prompt optimization framework for text-to-image models. This framework leverages an LVLM which functions both as a solver and an evaluator. The training pipeline is structured into five key components: model initialization, prompt generation, image generation, LVLM rewarding, and RL training. The pipeline is as follows: (1) Model initialization: we train the LVLM for prompt optimization on human-annotated prompt rewriting pairs and limited evaluation data, encapsulating LVLM's basic capability to rewrite prompts and assess the preference of generated images. (2) Prompt Generation: we employ a large version of LVLM to generate responses for a combination of an existing image quality evaluation dataset with the image evaluation prompt we used. (3) Image Generation: the model rewrites the raw prompt according to the instruction and samples multiple results, i.e., "candidates", and a fixed text-to-image model is used to generate the corresponding image. (4) LVLM Rewarding: the LVLM is employed to evaluate the aesthetics and alignment of the image with the original prompt, and is utilized as a rating system to score the images due to endowing with the capability to assess preferences. (5) RL Training: we
+
+select the highest and lowest-scored candidates to form preference pairs, which are used to train the model and adjust its output preferences by DPO training. To further enhance the model's ability to judge image quality, we also make the model itself act as a judge on the model's responses to image aesthetics or alignment to pick the most and least confident responses to construct preference pairs.
+
+The main contributions are summarized below:
+
+- We provide an AI-feedback approach to achieve aesthetic and alignment understanding of images using an LVLM. This approach effectively transforms the LVLM into a reward model to facilitate prompt optimization.  
+- We introduce self-rewarding training into prompt optimizing for the first time, which obtained prompts with higher quality by iterating the model on a small amount of training data, alleviating the shortcomings of models that require larger and higher quality data for reinforcement learning training.  
+- In the experiments, we compare our method and other strong competitors on two popular text-to-image datasets, i.e., beautiful-prompt and DiffusionDB. Results show our method achieves state-of-the-art performance.
+
+# 2 Background
+
+# 2.1 Prompt Rewriting
+
+The purpose of the prompt rewriting is to unlock the maximum potential of text-to-image models. Given an original user prompt  $\mathbf{x}$ , the prompt rewrite models with weight  $\theta$  can produce more professional prompts  $\mathbf{y}$  to help text-to-image models achieve images with more aesthetic pleasure and relevance. The process can be expressed as follows:
+
+$$
+\mathbf {y} = p (y \mid \mathbf {x}; \theta) \tag {1}
+$$
+
+# 2.2 Self-Rewarding Learning
+
+Self-Rewarding Learning uses the same model to perform iterative training to realize self-improvement. Given a sequence pair  $(x,y)$ , reinforcement learning fine-tuning demands the definition of the reward function  $\mathbf{r}(\mathbf{x},\mathbf{y})$  to quantify the value of the response  $\mathbf{y}$  to the given input  $\mathbf{x}$ . The objective of the RL fine-tuning can be defined as:
+
+$$
+\begin{array}{l} \mathcal {L} _ {R L} (\theta) = \mathbb {E} _ {\mathbf {x} \sim q (\cdot), \mathbf {y} \sim p _ {\theta} (\cdot | \mathbf {x})} [ r (\mathbf {x}, \mathbf {y}) ] \tag {2} \\ - \lambda \mathbb {E} _ {\mathbf {x} \sim q (\cdot)} \mathrm {K L} \left(p _ {\theta} (\cdot | \mathbf {x}) | | p _ {r e f} (\cdot | \mathbf {x})\right) \\ \end{array}
+$$
+
+![](images/bc015ffe0dc4beb0fd474a0955a947dd8a26a4d5a00e805ab1822edf8402f758.jpg)  
+(a) Reward Function
+
+![](images/edcc4377cbb096f394cea3faf6d59f895b2076ca0bce9011aba46cc92703a5b5.jpg)  
+(c) Fixed External LVLM-reward Model
+
+![](images/f9f6f99850716677e670b49ab99eb6871d2ea9651984c671c0a87ad14c5f72e4.jpg)  
+(b) Trained-reward Model
+
+![](images/a20844c116f193a0f6cc132c2893259779ac8a18c00ebd7713fd6014586ccd65.jpg)  
+(d) Our Self-reward Model  
+Figure 2: Four types of framework in prompt optimizing. The main difference among them is the reward generation. (a) The reward function (or Metrics) is pre-defined for reinforcement learning, which typically involves mathematical equations for text or images, such as work (Hao et al., 2024). (b) A feedback model is trained using a large annotated dataset (typical manual score) and then is employed for RL, like PPO, which is employed in (Cao et al., 2023; Rosenman et al., 2023). (c) Rewards are generated through a fixed external LVLM. The upper bound of evaluation depends on external models. (d) Rewriter and Reward Model share the same weight, achieving self-improvement by an iterative method to generating answers and self-judgment.
+
+where the KL regularization enforces the policy model  $p_{\theta}$  to be close to the reference model  $p_{ref}$ , and  $\lambda$  is set to control the deviation between policy model  $p_{\theta}$  and the reference model  $p_{ref}$ .
+
+Assuming a pair of responses  $<  \mathbf{y}^{1},\mathbf{y}^{2} >$  with a human annotator labeling one of them to be more aligned with human preferences, denoted as  $\mathbf{y}^w\succ \mathbf{y}^l |x.$  the Bradley-Terry (BT) (Brown et al., 2020) model stipulates that the human preference distribution  $p^*$  can be written as follows:
+
+$$
+p ^ {*} (\mathbf {y} ^ {1} \succ \mathbf {y} ^ {2} | \mathbf {x}) = \frac {\exp (r (\mathbf {x} , \mathbf {y} ^ {1}))}{\exp (r (\mathbf {x} , \mathbf {y} ^ {1})) + \exp (r (\mathbf {x} , \mathbf {y} ^ {2}))} (3)
+$$
+
+In self-rewarding training, the same LVLM is used to produce reward  $\mathbf{r} = [r_1,r_2,\dots,r_l]$ , so the conditional probability distribution  $p_{\theta}(r|\mathbf{x},\mathbf{y})$  can be expressed as follows:
+
+$$
+p _ {\theta} (\mathbf {r} | (\mathbf {x}, \mathbf {y})) = \prod_ {k = 1} ^ {l} \left(r _ {k} | (\mathbf {x}, \mathbf {y}), \mathbf {r} _ {<   k}\right) \tag {4}
+$$
+
+where  $\mathbf{r}_{< 1}$  is usually null or a start token and  $\mathbf{r}_{<  k} =$ $[r_1,r_2,\dots ,r_{l - 1}],k\in \{2,\dots ,l\} .$
+
+# 2.3 Related Work
+
+Prompt Engineering. Hao et al. (2024) propose a prompt adaption framework for prompt engineering. To implement reinforcement learning finetuning, a reward function for image aesthetics and alignment is defined. Bestprompt (Pavlichenko and Ustalov, 2023) is proposed to detect keywords by genetic algorithm and then form prompts to obtain the better aesthetics of images. BeautifulPrompt (Cao et al., 2023) first trains two reward models: Aesthetics and PickScore with a large dataset and then optimizes the language model using PPO. NeuroPrompt (Rosenman et al., 2023) utilizes constrained text decoding with a pre-trained
+
+language model to produce prompts. (Datta et al., 2023) is proposed as a Prompt Expansion framework to improve the diversity in text-to-image generation. Inspired by visual language modeling, recently, more and more methods (Li et al., 2022; Liu et al., 2024; Chen et al., 2024a; Yang et al., 2023; Yu et al., 2023) attempt to use a prior knowledge of LVLM to analyze images. The vision-language model is a large multimodal model, which bridges the gap between language and images. CLIP (Radford et al., 2021) is trained with large-scale paired text and images using contrastive learning. In this way, the pre-trained LVLMs capture rich vision-language correspondence knowledge. ALIGN (Jia et al., 2021) scales up the training process, using the larger images-text pairs but noisy data. Recently, with the great success of Large Language Models (LLMs), some work has been devoted to enabling LLMs to use image inputs. OpenFlamingo (Awadalla et al., 2023) and LLaMA-Adapter (Zhang et al., 2023) construct multimodal models based on the best LLM LlaMA (Touvron et al., 2023). To further improve the model's instruction-following abilities, LlaVA (Liu et al., 2024) employs visual instruction tuning that yields promising results. ViLA (Lin et al., 2024) achieves multi-image reasoning through a better training strategy.
+
+Text-to-Image generation. Text-to-image models usually refer to the generative model which synthesizes an image from a given text. Earlier work, GAN (Reed et al., 2016; Tao et al., 2022) and VAE (Ramesh et al., 2021; Ding et al., 2021) have been extensively studied in this field. Recently, the diffusion-based models (Rombach et al., 2022; Gu et al., 2022) further improve the quality of
+
+![](images/aba09ec83289648fdcbfbb8f51bab007684ec7feaf55fab9b5338b9862ea8af8.jpg)  
+Figure 3: The overall framework in our prompt optimizing framework. It involves five steps, arranged from left to right, (a) Prompt Rewriter sample multiple candidates y, (b) Diffusion Model generates images from the candidates, (c) Image Evaluator act as image evaluate models, to generate image evaluate responses R, (d) Response Judge act as judge models to judge the response from evaluator and get response evaluation E, and (e) Optimization with the response from evaluator and judge, and then update the LVLM.
+
+text-to-image generation. Given the generated output is no longer a serious concern, exploring how to optimize the prompt to maximize the potential of generative models becomes the primary object of our research. Several works (Hao et al., 2024; Cao et al., 2023) have been put into unlocking the maximum potential of text-to-image models by optimizing prompts using large language models. Besides, some reinforcement learning-based methods have gained consistent improvements in changing the output preferences of large language models. Reinforcement Learning from Human Feedback (RLHF) (Yuan et al., 2024a) yields promising results in various NLP tasks (Christiano et al., 2017; Ibarz et al., 2018; Stiennon et al., 2020; Jaques et al., 2019), at the cost of collecting large amounts of human feedback. Proximal Policy Optimization (PPO) (Schulman et al., 2017) is optimized according to the reward signal rather than human feedback and remains stable by approximating the old and new strategies. With these challenges in mind, Direct Preference Optimization (DPO) (Rafailov et al., 2024) avoids the drawback of large amounts of human-annotated data to train reward models by using its language model as a reward model. The different strategies to provide rewards are shown in Figure 2.
+
+# 3 Methodology
+
+In this section, we first describe the pipeline of dataset construction (Section 3.1), and how to initialize LVLM with both prompt optimization and
+
+preference assessing capabilities (Section 3.2) and then detail self-rewarding training algorithms (Section 3.3). Next, we introduce the use of AI feedback for RL in self-rewarding training (Section 3.4) and how to transform LVLM into a reward model by LVLM-as-a-judge (Section 3.5).
+
+# 3.1 Dataset Construction Pipeline
+
+Prompt rewrite data. To equip the LVLM with initial rewriting capabilities, rewrite prompts data is first given to train in the supervised fine-tuning (SFT) manner. This type of data is presented in pairs, i.e., {raw prompt, rewritten prompt}, to guide the LVLM to rewrite prompt. Prompt rewrite data contains 104,487 prompt pairs in (Cao et al., 2023). Evaluation data. The LVLM specializes in image understanding and analyzing (Zhu et al., 2024; Li et al., 2022). However, image aesthetics is still a subjective understanding and not easy for models. Although unnecessary, LVLM appears more difficult to score images and construct preference pairs without fine-tuning on evaluation data (Yuan et al., 2024b). Therefore, evaluation data is constructed to offer the model some examples to understand the aesthetics. The prompt data in the previous paragraph is used to generate images by diffusion model (Rombach et al., 2022). These images are then scored using the prompt template in the LVLM-as-a-judge way (Zheng et al., 2024; Chen et al., 2024a). We use a subset of the aesthetics training set and the Pickapic-v1 training set. The paired images and evaluation prompt are both
+
+fed into VILA-40B as well as the facts to obtain the response as truth response for training. We add a truth cue to the evaluation prompt, if the VILA-40B's scoring of the two images is consistent with the facts, the evaluation response is retained.
+
+Judgment data. To improve the LVLM's evaluation capabilities for responses (Zheng et al., 2024; Zhu et al., 2023), we also create a dataset for the model to understand which responses to aesthetics or alignment are more reasonable. Specifically, we construct data that assign higher scores to reasonable responses and lower scores to irrelevant responses, which is some kind of confidence. To improve the evaluation ability of LVLM, we create a new dataset to evaluate the model's responses to aesthetics and PickScore of the images. Similarly, we feed the standard (i.e., VILA-40B's response) and irrelevant answers into VILA-40B, and if the model has a higher score in the standard answer than in the irrelevant answer, the answer remains in the dataset. A common observation is that most images generated from rewritten prompts are typically assigned scores of 3 or 4. Therefore, we also include some raw prompts to maintain a more uniform score distribution.
+
+# 3.2 Initialization LVLM Capability of Rewriting and Assessing Preference
+
+After we obtain the initial dataset, supervised finetuning (SFT) is employed to initiate the model with the ability of image evaluation and response judgment. Therefore, we consider the input  $\mathbf{x}$  for a specific task to be derived from the distribution  $q(\cdot)$ . Meanwhile, the probability distribution in the SFT training data can be represented as  $p_{data}(\cdot|\mathbf{x})$ . The object function of SFT can be represented as:
+
+$$
+\mathcal {L} _ {S F T} (\theta) = - \mathbb {E} _ {\mathbf {x} \sim q (\cdot), \mathbf {y} \sim p _ {d a t a} (\cdot | \mathbf {x})} [ \log p _ {\theta} (\mathbf {y} | \mathbf {x}) ] \tag {5}
+$$
+
+Since the SFT training data has high-quality labeled responses  $\mathbf{y}$ , the main purpose of this object function is to approximate the model's predictive distribution  $p_{\theta}(\mathbf{y}|\mathbf{x})$  to the target  $p_{data}(\mathbf{y}|\mathbf{x})$ .
+
+# 3.3 Training with Self-Rewarding
+
+The main objective of the self-rewarding model is to complete iterative training with limited amounts of human-labeled data on the pre-trained model. During an iteration, the model generates a solution and a judgment on this solution. Unlike other methods that use a fixed reward model (Ouyang et al., 2022), the self-rewarding model uses the model of
+
+the current iteration to generate the rewards. Therefore, the model can both improve its generative capabilities when acting as a generative model and get its boost when acting as a rewarding model, since the corresponding responses are generated by the same mechanism (Yuan et al., 2024b).
+
+The whole training process is concluded as follows: the model first starts with pre-trained LVLMs (denoted as  $\mathrm{LVLM}_{init}$ ), training with prompt rewrite data and evaluation data in SFT manner, resulting in  $\mathrm{LVLM}_{SFT}$ . To further enhance the model's performance, we sample multiple candidates from the raw prompt and construct preference pairs based on the LVLM scores, and then the model is trained in the DPO manner. This training process is iterated M times, yielding models denoted as  $\{\mathrm{LVLM}_{\mathrm{DPO}_1}, \mathrm{LVLM}_{\mathrm{DPO}_2}, \dots, \mathrm{LVLM}_{\mathrm{DPO}_M}\}$ . The bottleneck of prompt engineering can be summarized into two aspects: 1) it is difficult to align with human preference and thus a large number of human-labeled data is required; 2) the whole training process is too complex, typically requiring a larger trained reward model. Therefore, we propose a self-rewarding prompt rewriting model to obtain better performance through AI self-feedback and an evolving training process. The overall framework is shown in Fig. 3. First, the model acts as a solver sampling multiple candidate answers from raw prompts. These candidates are then fed into a fixed text-to-image model to generate the corresponding images. Again, the previous solver is used as the reward model to generate responses, for scoring these images generated from candidates. The details of scoring are discussed below. Finally, preference pairs are constructed for model DPO training.
+
+# 3.3.1 Training on Rewriting
+
+We consider an LVLM to be parameterized by  $\theta$  and denoted by  $p_{\theta}$ . The model takes a prompt as sequence  $\mathbf{x} = [x_1, x_2, \dots, x_n]$  to generate the corresponding response  $\mathbf{y} = [y_1, y_2, \dots, y_m]$ . The response  $\mathbf{y}$  is thus regarded as a sample from the conditional probability distribution  $p_{\theta}(\cdot | \mathbf{x})$ . Specifically,  $x_i$  and  $y_j$  represent the tokens from the same predetermined vocabulary within the sequences  $\mathbf{x}$  and  $\mathbf{y}$ , respectively. To generate the  $y_j$  for a given position, the auto-regressive model  $p_{\theta}$  exploits the previously generated tokens to generate subsequent tokens up to the maximum length or the end token. Therefore, the conditional probability distribution
+
+$p_{\theta}(\mathbf{y}|\mathbf{x})$  can be expressed as:
+
+$$
+p _ {\theta} (\mathbf {y} | \mathbf {x}) = \prod_ {j = 1} ^ {m} \left(y _ {j} | \mathbf {x}, \mathbf {y} _ {<   j}\right) \tag {6}
+$$
+
+where  $\mathbf{y}_{< 1}$  is usually null or a start token and  $\mathbf{y}_{< j} = [y_1,y_2,\dots ,y_{j - 1}],j\in \{2,\dots,m\}$
+
+# 3.3.2 Training on Preference Assessing
+
+Beyond the model training on prompt rewriting, we also perform preference assessing training on the same model to improve its assessing ability. Similar to Sec. 3.3.1, the model generates the judgment on the response  $\mathbf{R}$  for a response  $\mathbf{y}$  on prompt and the generated image. The conditional probability distribution can be defined as:
+
+$$
+p _ {\theta} (\mathbf {R} | \mathbf {y}) = \prod_ {k = 1} ^ {n} \left(R _ {j} | \mathbf {y}, \mathbf {R} _ {<   k}\right) \tag {7}
+$$
+
+where  $\mathbf{y}_{< 1}$  is usually null or a start token and  $\mathbf{R}_{<  k} = [R_{1},R_{2},..,R_{k - 1}],j\in \{2,\dots ,n\} .$
+
+Therefore, RL fine-tuning in self-rewarding can be optimized with the loss function, i.e.,
+
+$$
+\begin{array}{l} \mathcal {L} _ {R L} (\theta_ {t + 1}) = \mathbb {E} _ {\mathbf {x} \sim q (\cdot), \mathbf {y} \sim p _ {\theta_ {t}}, \mathbf {r} \sim p _ {\theta_ {t}} (\cdot | (\mathbf {x}, \mathbf {y}))} [ r (\mathbf {x}, \mathbf {y}) ] \\ - \lambda \mathbb {E} _ {\mathbf {x} \sim q (\cdot)} \mathrm {K L} \left(p _ {\theta} (\cdot | \mathbf {x}) | | p _ {r e f} (\cdot | \mathbf {x})\right) \\ + \mathbb {E} _ {\mathbf {y} \sim q (\cdot), \mathbf {R} \sim p _ {\theta_ {t}}, \mathbf {E} \sim p _ {\theta_ {t}} (\cdot | (\mathbf {y}, \mathbf {R}))} [ E (\mathbf {y}, \mathbf {R}) ] \\ - \lambda \mathbb {E} _ {\mathbf {y} \sim q (\cdot)} \mathrm {K L} \left(p _ {\theta} (\cdot | \mathbf {y}) \right\| p _ {r e f} \tag {8} \\ \end{array}
+$$
+
+where  $\theta_{t}$  denotes the  $t$ -th parameters of the model, the  $r(\mathbf{x},\mathbf{y})$  and the  $E(\mathbf{y},\mathbf{R})$  denote the reward function of the prompt  $x$  and the response  $y$ , respectively. Since the self-reward model is an iterative model, in which the iterative process results in a series of models with different weights of the same structure. For better understanding, the parameter of the result model by Equation (5) is denoted as  $\theta_0$  while the parameters of the result models optimized through Equation (8) are denoted as  $\{\theta_1,\theta_2,\dots \}$ . As the model's parameters continue to be optimized for the ability to follow instructions, so does the reward ability, thereby facilitating self-improvement.
+
+# 3.4 RL from AI Feedback
+
+AI Feedback. One of the main investigations of our work is how to use prior knowledge of large models to guide text-to-image models. Unlike time-consuming and labour-intensive human feedback training, AI feedback (Lee et al., 2023) training can provide reward signals for a given task through its own knowledge. This approach usually requires
+
+two models, one acting as a solver of the downstream task, and one acting as a judge of the solver. It is feasible to apply an external language model as a judge (or reward model) or to use only itself (i.e. a model that acts as both a solver and a judge) to achieve self-improvement in a specific task.
+
+Therefore, preference pairs  $\langle$  raw prompt  $x$ , winner prompt  $y^w$ , loser prompt  $y^l \rangle$  need to be constructed to train the model. The model first generates different  $N$  candidates from randomly selected raw prompts (in the previous data). After generating images using a fixed text-to-image model, the images and evaluation prompts are input into the model for scoring. Next, the prompts corresponding to the highest and lowest-scored images are treated as winners and losers, respectively. In addition, the image evaluation responses are judged by LVLM and the pairwise data for DPO are also constructed. The model is then tuned with DPO (Rafailov et al., 2024).
+
+DPO. Assuming access to a static dataset of comparisons  $\mathcal{D} = \{\mathbf{x}_i, \mathbf{y}_i^w, \mathbf{y}_i^l\}$ , which is sampled from  $p^*$ . The optimal RLHF policy  $\pi^*$  under the Bradley-Terry model satisfies the preference model:
+
+$$
+\begin{array}{l} p ^ {*} (\mathbf {y} ^ {1} \succ \mathbf {y} ^ {2} | \mathbf {x}) = \left[ 1 + \exp \left(\lambda \log \frac {\pi^ {*} (\mathbf {y} ^ {2} | \mathbf {x})}{\pi_ {r e f} (\mathbf {y} ^ {2} | \mathbf {x})}\right) \right. \\ \left. - \lambda \log \frac {\pi^ {*} (\mathbf {y} ^ {1} | \mathbf {x})}{\pi_ {r e f} (\mathbf {y} ^ {1} | \mathbf {x})}\right) \bigg ] ^ {- 1} \tag {9} \\ \end{array}
+$$
+
+Given a preference pair  $\langle \mathbf{x},\mathbf{y}^w,\mathbf{y}^l\rangle$  , the object function of DPO is to seek a maximum likelihood of the parameterized policy  $\pi_{\theta}$  by reference model  $\pi_{ref}$
+
+$$
+\begin{array}{l} \mathcal {L} _ {D P O} (\pi_ {\theta}; \pi_ {r e f}) \\ = - \mathbb {E} _ {(\mathbf {x}, \mathbf {y} ^ {w}, \mathbf {y} ^ {l}) \sim \mathcal {D}} \left[ \log \sigma \left(\Delta_ {\lambda} (\mathbf {y} ^ {w} | \mathbf {x}) - \Delta_ {\lambda} (\mathbf {y} ^ {l} | \mathbf {x})\right) \right], \\ \end{array}
+$$
+
+where  $\Delta_{\lambda}(y|x) = \lambda \log \frac{\pi_{\theta}(y|x)}{\pi_{ref}(y|x)}$  (10)
+
+The  $\lambda$  is a parameter controlling the deviation from the base reference policy  $\pi_{ref}$ .
+
+# 3.5 LVLM-as-a-judge for Aesthetics and Rewarding
+
+To improve the text-to-image models, we consider this from two perspectives: the aesthetics of the generated image and the capability to follow instructions, respectively. Hence, to empower the model in generating appropriate reward signals for the images, we establish a judging template comprising aesthetic score, pick score and alignment
+
+score. This enables the LVLM to evaluate the generated images effectively. Details of prompts for LVLM-as-a-judge can be found in Appendix.
+
+Aesthetic Score. Aesthetics are assessed by prompting the model to consider aspects such as composition, color, lighting, and visual appeal. To facilitate judgment-making, we design a grading system that allows the model to attribute a certain grade to the generated images, thus obtaining the corresponding score. For each evaluation, the model is asked to provide both the score and a brief explanation, employing a chain-of-thought reasoning approach.
+
+PickScore (Kirstain et al., 2024). PickScore is an important metric to measure human preferences. We consider whether the image represents the given text in a way that is favored by humans. Nonetheless, it is difficult for the model to understand human preferences directly, so we included evaluation data (details in Section 3.2) for model initialization.
+
+Relevance Score. The judgment template for the relevance score uses the same hierarchical form as the aesthetic score. The key difference is that it evaluates the model's attention to the user instructions for text-to-image generation in three areas: presence of the object, accurate count of the object, and correct relationships between objects.
+
+# 4 Experiments
+
+# 4.1 Settings
+
+Experiments are conducted on the public text-to-image model Stable Diffusion v1.5, and the denoising steps are set to 20 to accelerate the image sampling. The base model we employ in this paper is the smaller model VILA-v1.5-8b (Lin et al., 2024). For SFT, we use AdamW optimizer  $(\beta_{1} = 0.9, \beta_{2} = 0.95)$ , with a batch size of 16 and a weight decay of 0.1. We use an initial learning rate of 2e-5, with a linear warm-up and cosine decay schedule. To improve the diversity of the candidates, we sample  $N = 8$  candidates with temperature  $T = 0.9$ ,  $p = 0.9$  from one raw prompt. When validating these candidates with the reward model, we utilize a temperature  $T = 0.9$ ,  $p = 0.9$ , sampling three times and averaging the scores to determine the final score for further DPO training. The overall score is calculated by summing up the aesthetics score, pick Score and the relevance score. The highest one and the lowest one are kept as winners and losers but discarded if they have the same score. We perform two DPO iterations
+
+<table><tr><td>Method</td><td>Drl</td><td>PickScore</td><td>Aes.</td><td>CLIP</td></tr><tr><td>Original</td><td>-</td><td>20.74</td><td>5.50</td><td>0.27</td></tr><tr><td>MagicPrompt (Santana, 2022)</td><td>-</td><td>20.11</td><td>5.79</td><td>0.22</td></tr><tr><td>ChatGPT (OpenAI, 2023)</td><td>-</td><td>20.73</td><td>5.92</td><td>0.25</td></tr><tr><td>Beautiful-Prompt (Cao et al., 2023)</td><td>40k</td><td>20.84</td><td>6.52</td><td>0.24</td></tr><tr><td colspan="5">Our Method</td></tr><tr><td>LVLMsFT</td><td>-</td><td>20.79</td><td>5.95</td><td>0.25</td></tr><tr><td>LVLMDPO1</td><td>10k</td><td>20.81</td><td>6.31</td><td>0.24</td></tr><tr><td>LVLMDPO2</td><td>20k</td><td>20.86</td><td>6.59</td><td>0.24</td></tr></table>
+
+Table 1: Evaluation of the aesthetic score and CLIP score on Beautiful-Prompt test set. “ $\mathbb{D}_{rl}$ ” means the size of the train set we used in reinforcement learning. “Aes.” and “CLIP” mean the aesthetic and CLIP scores, respectively.  
+
+<table><tr><td>Method</td><td>Type</td><td>Aes.</td><td>CLIP</td></tr><tr><td>Original†</td><td>Human</td><td>5.47</td><td>0.28</td></tr><tr><td>Human Engineered Prompt†</td><td>Human</td><td>5.87</td><td>0.26</td></tr><tr><td>NeuroPrompts(Rosenman et al., 2023)</td><td>AI</td><td>6.27</td><td>-</td></tr><tr><td>Promptist(Hao et al., 2024)</td><td>AI</td><td>6.26</td><td>0.26</td></tr><tr><td>LVLM\(_{\text{DPO}_2}\) (Ours)</td><td>AI</td><td>6.57</td><td>0.26</td></tr></table>
+
+Table 2: Evaluation of the aesthetic score and CLIP score on DiffusionDB. Values marked with  $\dagger$  from (Hao et al., 2024).
+
+and the size of each prompt rewrite training data is 10k and 20k, respectively. Besides, 10k judgment data is employed to improve image evaluation. In the stage of DPO, we employ AdamW with an initial learning rate of 1e-5, 5e-6 without weight decay. The batch size is set to 32 and  $\beta$  in DPO is set to 0.1. The model is trained for 4 epochs at each iteration. All experiments are conducted on  $4\times$  NVIDIA A800 80G GPUs. For evaluation, we adopt beam search (Vijayakumar et al., 2016) with a beam size of 4 and a length penalty of 1.0.
+
+# 4.2 Comparative Methods
+
+We compare our method with the following approaches: MagicPrompt (Santana, 2022), ChatGPT (OpenAI, 2023), Beautiful-Prompt (Cao et al., 2023), NeuroPrompts (Rosenman et al., 2023) and Promptist (Hao et al., 2024). It is worth noting that, we prompt ChatGPT (OpenAI, 2023) to generate an expansion of the user-provided prompt for generative models. Other models generate results using their open-source weights. The "Human Engineered Prompt" refers to the prompts written by humans, while its simplified version of these prompts is used as the original prompt. In real-world usage, users only need to provide simple prompt to obtain high-quality results.
+
+# 4.3 Results
+
+The model is validated on two datasets, BeautifulPrompt (Cao et al., 2023) and DiffusionDB (Wang et al., 2023) (100k), respectively. As shown in Table 1, our method outperforms other methods. This
+
+<table><tr><td>Pick</td><td>Aes.</td><td>Align</td><td>Reward</td><td>PickScore</td><td>Aes.</td><td>CLIP</td></tr><tr><td>✓</td><td>✓</td><td>✓</td><td>Self</td><td>20.86</td><td>6.59</td><td>0.24</td></tr><tr><td></td><td>✓</td><td>✓</td><td>Self</td><td>20.73</td><td>6.59</td><td>0.24</td></tr><tr><td>✓</td><td>✓</td><td></td><td>Self</td><td>20.79</td><td>6.61</td><td>0.22</td></tr><tr><td>✓</td><td></td><td>✓</td><td>Self</td><td>20.76</td><td>6.49</td><td>0.24</td></tr><tr><td>✓</td><td>✓</td><td>✓</td><td>Fixed</td><td>20.73</td><td>6.26</td><td>0.24</td></tr></table>
+
+User Prompt  
+Figure 4: Human evaluation results. The result of  $\mathrm{LVLM}_{DPO_2}$  are more preferred by human compared with the result of User Prompt and  $\mathrm{LVLM}_{SFT}$ .  
+![](images/d1521b59962e8b172a1fd0a25e9a809c83e232c25edd0fd6149b4080f96dfd59.jpg)  
+LVLMDPO2 Wins Ties LVLMDPO2 Loses
+
+![](images/c5c3b5ccb9e87dc815dd83a580e0426fa6943bca443dd31119e4461d938f0d03.jpg)  
+LVLM SFT
+
+performance gap is more evident on larger datasets DiffusionDB (100k). As can be seen in Table 2, our method achieves a performance gain compared to other methods. Compared to the most competitive method Beautiful-Prompt, our method achieves a 0.1 improvement in the Aesthetic Score, while maintaining no degradation in the PickScore and CLIP score. However, on the larger DiffusionDB test set, our method achieves a 0.3 improvement compared to the Neuroprompt (Rosenman et al., 2023) and Promptist (Hao et al., 2024).
+
+# 4.4 Ablation Study
+
+We perform ablation experiments from two perspectives: how the prompt affects the LVLM as a reward model and how self-rewarding training improves the model. First, as can be seen in Table 3, the model without Pick Prompt for rewarding received a high aesthetic score but leads a significant drop in PickScore. It also suggests that some shifts exist between human preferences and image aesthetics. In addition, it can be observed a reduction of the aesthetic score and CLIP score when missing the aesthetic prompt or relevance prompt. Second, we compare the performance between the model using self-rewarding and fixed-rewarding. Self-rewarding outperforms the fixed-rewarding methods. More discussion about the self-rewarding method can be found in the Appendix D.
+
+# 4.5 Human Evaluation
+
+To better compare the existing methods with our model from the human perspective, a human study
+
+![](images/2ecb5ebcf9629e45d5fa38ed9de8712caa6926f72857dbe3f56ed6deae9fbb49.jpg)  
+Figure 5: Generated prompt length (bar graph) and aesthetic score improvement (line graph) compared with raw prompt.
+
+Table 3: Ablation study on different prompts and reward mode. Checkmark is enabled prompts in image evaluation.  
+
+<table><tr><td colspan="4"># Training Data</td></tr><tr><td>Stage</td><td>Beautiful-Prompt</td><td>Promptist</td><td>Ours</td></tr><tr><td>SFT</td><td>143k</td><td>360k</td><td>156k</td></tr><tr><td>Rewrite</td><td>143k</td><td>360k</td><td>104k</td></tr><tr><td>Evaluation</td><td>-</td><td>-</td><td>32k</td></tr><tr><td>Judgment</td><td>-</td><td>-</td><td>20k</td></tr><tr><td>RM</td><td>7M</td><td>-</td><td>-</td></tr><tr><td>RL</td><td>40k</td><td>90k</td><td>40k</td></tr><tr><td>Rewrite</td><td>40k</td><td>90k</td><td>{10k, 20k}</td></tr><tr><td>Evaluation</td><td>-</td><td>-</td><td>10k</td></tr></table>
+
+Table 4: Training set number on different stages of BeautifulPrompt(Cao et al., 2023), Promptist(Hao et al., 2024), and ours. RM means the training of the reward model. Our method requires addition data on image evaluation and response judgment for rewarding.  
+
+<table><tr><td>Method</td><td>LVLMinit</td><td>LVLMsFT</td><td>LVLMDPO1</td></tr><tr><td>Test-Set Accuracy</td><td>48.0%</td><td>66.1%</td><td>66.4%</td></tr></table>
+
+Table 5: Evaluation on Pickapic-v1 test set. Our self-rewarding model realizes the ability to self-improvement
+
+is conducted. We randomly select prompts and obtain the optimized prompt through different models, and then the images are generated. Then, 20 volunteers are asked to rank the different images. Each volunteer is presented with two images at a time and asked to select the one they find more appealing. The volunteers in this study are randomly selected from a pool of individuals with diverse educational backgrounds. The result is shown in Figure 4. It can be observed that the images obtained through the optimized prompt generated by our model are selected most frequently by the volunteers. And the DPO model is better than the SFT baseline, which achieves about  $55\%$  win rate.
+
+# 4.6 Further Discussion
+
+To further understand how the prompt optimization affects the original prompt, we show some analysis in Figure 5. Considering the length of prompts has a significant relationship with the quality of generation, we count the length of optimized prompts
+
+![](images/da7509c0a3e0ed5d47b4b27784f793b963aa2d06c48f57413580b73b02663590.jpg)  
+Figure 6: The generated images with the optimized prompts using our method
+
+of  $\mathrm{LVLM}_{SFT}$  and  $\mathrm{LVLM}_{DPO_2}$  and their aesthetic score margins compared to the original prompt. An immediate observation is that the length of most optimized prompts is between 30 and 40. Besides, aesthetic scores gradually increased with the length. Besides, we also present the size of the training set in Table 4. In SFT stage, we use prompt rewrite data (104k), evaluation data (32k) and judgment data (20k), while in RL fine-tuning stage, we use prompt data 10k, 20k and new evaluation data 10k sampled from LVLM. Note that we just expanded the dataset after each iteration, whereas 10k prompt data is the subset of 20k. This means only 20k training data are needed to train this model, while Beautiful-Prompt and Promptist employ the larger training set, i.e., 40k and 90k, respectively. To better exemplify the ability of our model to evaluate by self-rewarding, we show the accuracy of the test set in Pickapic-v1 after SFT and DPO training. As shown in Table 5, our model realizes an increase in evaluation capacity.
+
+# 4.7 Quantitative Results
+
+We present more visual results in Figure 6, it can be observed that our method rewrites the main content appropriately, and describes the detail, e.g., art style, light, etc. The images generated by these
+
+optimized prompts have better aesthetics than the images generated from the original prompts, which are bland visually. In addition, some incorrect prompts can also be corrected. For example, the prompt "Astronaut rides horse" mislead the text-to-image models to generate an astronaut standing beside a horse. However, our approach rewrites it as "...an astronaut riding a horse..." to make the image more in line with the user's intention. In addition, the model expands some descriptions on some objects, like "futuristic"" relative to the "in 2456". More results are in the Appendix C.
+
+# 5 Conclusion
+
+In this work, we propose a novel method to optimize prompts for text-to-image models, which can be used to fill the gap between laymen and experts when using a generative model. We first transform the LVLM into a reward model so that it can judge the aesthetics and alignment of the images. In so doing, we can perform AI feedback rather than human feedback. Then, to gain the performance boost based on limited data, we employ self-rewarding training for LVLM. Our model achieves self-improvement through an iterative training approach. Experiments on two datasets show that our method outperforms the other methods.
+
+# Limitations
+
+The primary limitation of our work is the inability to utilize larger parameter models due to the significant computational resources required. Training and fine-tuning models with billions of parameters, such as VILA-40B, demand substantial GPU memory and processing power, which are often constrained by available hardware. Deploying large-parameter models in real-world applications presents considerable challenges.
+
+# References
+
+Anas Awadalla, Irena Gao, Josh Gardner, Jack Hessel, Yusuf Hanafy, Wanrong Zhu, Kalyani Marathe, Yonatan Bitton, Samir Gadre, Shiori Sagawa, et al. 2023. Openflamingo: An open-source framework for training large autoregressive vision-language models. arXiv preprint arXiv:2308.01390.  
+Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. 2020. Language models are few-shot learners. Advances in Neural Information Processing Systems, 33:1877-1901.  
+Tingfeng Cao, Chengyu Wang, Bingyan Liu, Ziheng Wu, Jinhui Zhu, and Jun Huang. 2023. Beautifulprompt: Towards automatic prompt engineering for text-to-image synthesis. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing: Industry Track, pages 1-11.  
+Dongping Chen, Ruoxi Chen, Shilin Zhang, Yinuo Liu, Yaochen Wang, Huichi Zhou, Qihui Zhang, Pan Zhou, Yao Wan, and Lichao Sun. 2024a. Mllm-as-a-judge: Assessing multimodal llm-as-a-judge with vision-language benchmark. arXiv preprint arXiv:2402.04788.  
+Zixiang Chen, Yihe Deng, Huizhuo Yuan, Kaixuan Ji, and Quanquan Gu. 2024b. Self-play fine-tuning converts weak language models to strong language models. arXiv preprint arXiv:2401.01335.  
+Paul F Christiano, Jan Leike, Tom Brown, Miljan Martic, Shane Legg, and Dario Amodei. 2017. Deep reinforcement learning from human preferences. Advances in Neural Information Processing Systems, 30.  
+Siddhartha Datta, Alexander Ku, Deepak Ramachandran, and Peter Anderson. 2023. Prompt expansion for adaptive text-to-image generation. arXiv preprint arXiv:2312.16720.  
+Ming Ding, Zhuoyi Yang, Wenyi Hong, Wendi Zheng, Chang Zhou, Da Yin, Junyang Lin, Xu Zou, Zhou Shao, Hongxia Yang, et al. 2021. Cogview: Mastering text-to-image generation via transformers. Advances in Neural Information Processing Systems, 34:19822-19835.  
+Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. 2022. Vector quantized diffusion model for text-to-image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10696-10706.  
+Yaru Hao, Zewen Chi, Li Dong, and Furu Wei. 2024. Optimizing prompts for text-to-image generation. Advances in Neural Information Processing Systems, 36.
+
+Borja Ibarz, Jan Leike, Tobias Pohlen, Geoffrey Irving, Shane Legg, and Dario Amodei. 2018. Reward learning from human preferences and demonstrations in atari. Advances in Neural Information Processing Systems, 31.  
+Natasha Jaques, Asma Ghandeharioun, Judy Hanwen Shen, Craig Ferguson, Agata Lapedriza, Noah Jones, Shixiang Gu, and Rosalind Picard. 2019. Way off-policy batch deep reinforcement learning of implicit human preferences in dialog. arXiv preprint arXiv:1907.00456.  
+Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. 2021. Scaling up visual and vision-language representation learning with noisy text supervision. In International Conference on Machine Learning, pages 4904-4916. PMLR.  
+Yuval Kirstain, Adam Polyak, Uriel Singer, Shahbuland Martiana, Joe Penna, and Omer Levy. 2024. Pick-a-pic: An open dataset of user preferences for text-to-image generation. Advances in Neural Information Processing Systems, 36.  
+Harrison Lee, Samrat Phatale, Hassan Mansoor, Kellie Lu, Thomas Mesnard, Colton Bishop, Victor Carbune, and Abhinav Rastogi. 2023. Rlaif: Scaling reinforcement learning from human feedback with ai feedback. arXiv preprint arXiv:2309.00267.  
+Feng Li, Hao Zhang, Yi-Fan Zhang, Shilong Liu, Jian Guo, Lionel M Ni, PengChuan Zhang, and Lei Zhang. 2022. Vision-language intelligence: Tasks, representation learning, and large models. arXiv preprint arXiv:2203.01922.  
+Ji Lin, Hongxu Yin, Wei Ping, Pavlo Molchanov, Mohammad Shoeybi, and Song Han. 2024. Vila: On pre-training for visual language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26689-26699.  
+Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. 2024. Visual instruction tuning. Advances in Neural Information Processing Systems, 36.  
+Vivian Liu and Lydia B Chilton. 2022. Design guidelines for prompt engineering text-to-image generative models. In Proceedings of the 2022 CHI Conference on Human Factors in Computing Systems, pages 1-23.  
+OpenAI. 2023. Chatgpt (mar 14 version) [large language model]. https://chat.openai.com/chat.  
+Jonas Oppenlaender. 2023. A taxonomy of prompt modifiers for text-to-image generation. Behaviour & Information Technology, pages 1-14.  
+Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. 2022. Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems, 35:27730-27744.  
+Nikita Pavlichenko and Dmitry Ustalov. 2023. Best prompts for text-to-image models and how to find them. In Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 2067-2071.
+
+Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. 2021. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR.  
+Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. 2024. Direct preference optimization: Your language model is secretly a reward model. Advances in Neural Information Processing Systems, 36.  
+Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. 2021. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR.  
+Scott Reed, Zeynep Akata, Xinchen Yan, Lajanugen Logeswaran, Bernt Schiele, and Honglak Lee. 2016. Generative adversarial text to image synthesis. In International Conference on Machine Learning, pages 1060-1069. PMLR.  
+Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. 2022. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, pages 10684-10695.  
+Shachar Rosenman, Vasudev Lal, and Phillip Howard. 2023. Neuroprompts: An adaptive framework to optimize prompts for text-to-image generation. arXiv preprint arXiv:2311.12229.  
+Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. 2022. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494.  
+Gustavo Santana. 2022. Magicprompt - stable diffusion.  
+John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. 2017. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347.  
+Nisan Stiennon, Long Ouyang, Jeffrey Wu, Daniel Ziegler, Ryan Lowe, Chelsea Voss, Alec Radford, Dario Amodei, and Paul F Christiano. 2020. Learning to summarize with human feedback. Advances in Neural Information Processing Systems, 33:3008-3021.  
+Ming Tao, Hao Tang, Fei Wu, Xiao-Yuan Jing, Bing-Kun Bao, and Changsheng Xu. 2022. Df-gan: A simple and effective baseline for text-to-image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16515-16525.  
+Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, et al. 2023. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971.  
+Ashwin K Vijayakumar, Michael Cogswell, Ramprasath R Selvaraju, Qing Sun, Stefan Lee, David Crandall, and Dhruv Batra. 2016. Diverse beam search: Decoding diverse solutions from neural sequence models. arXiv preprint arXiv:1610.02424.
+
+Zijie J Wang, Evan Montoya, David Munechika, Haoyang Yang, Benjamin Hoover, and Duen Horng Chau. 2023. Diffusiondb: A large-scale prompt gallery dataset for text-to-image generative models. In The 61st Annual Meeting Of The Association For Computational Linguistics.  
+Jing Xu, Andrew Lee, Sainbayar Sukhbaatar, and Jason Weston. 2023. Some things are more cringe than others: Preference optimization with the pairwise cringe loss. arXiv preprint arXiv:2312.16682.  
+Zhengyuan Yang, Linjie Li, Kevin Lin, Jianfeng Wang, Chung-Ching Lin, Zicheng Liu, and Lijuan Wang. 2023. The dawn of Imms: Preliminary explorations with gpt-4v (isdiction). arXiv preprint arXiv:2309.17421, 9(1):1.  
+Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, et al. 2022. Scaling autoregressive models for content-rich text-to-image generation. arXiv preprint arXiv:2206.10789, 2(3):5.  
+Weihao Yu, Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Zicheng Liu, Xinchao Wang, and Lijuan Wang. 2023. Mm-vet: Evaluating large multimodal models for integrated capabilities. arXiv preprint arXiv:2308.02490.  
+Hongyi Yuan, Zheng Yuan, Chuanqi Tan, Wei Wang, Songfang Huang, and Fei Huang. 2024a. Rrrh: Rank responses to align language models with human feedback. Advances in Neural Information Processing Systems, 36.  
+Weizhe Yuan, Richard Yuanzhe Pang, Kyunghyun Cho, Sainbayar Sukhbaatar, Jing Xu, and Jason Weston. 2024b. Self-rewarding language models. arXiv preprint arXiv:2401.10020.  
+Renrui Zhang, Jiaming Han, Chris Liu, Peng Gao, Aojun Zhou, Xiangfei Hu, Shilin Yan, Pan Lu, Hongsheng Li, and Yu Qiao. 2023. Llama-adapter: Efficient fine-tuning of language models with zero-init attention. arXiv preprint arXiv:2303.16199.  
+Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, et al. 2024. Judging llm-as-a-judge with mt-bench and chatbot arena. Advances in Neural Information Processing Systems, 36.  
+Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. 2024. MiniGPT-4: Enhancing vision-language understanding with advanced large language models. In International Conference on Learning Representations.  
+Lianghui Zhu, Xinggang Wang, and Xinlong Wang. 2023. Judgelm: Fine-tuned large language models are scalable judges. arXiv preprint arXiv:2310.17631.
+
+# A Comparisons with the common tags
+
+To figure out what word improves the prompt, we count the most frequent words. Evidently, "fantasy" and "intricate" are popular in the model, which even appears more than once in the same sentence.
+
+To demonstrate that the performance improvement of the prompt rewriting method is not the result of adding some fixed words, we select the most frequent words and randomly combine them into six tags. Then, we combine it with the original prompt and calculated the scores. The tags are shown in Tab. 7 and the results are present in Tab. 8.
+
+<table><tr><td>Content</td><td>Frequency</td></tr><tr><td>fantasy</td><td>2,291</td></tr><tr><td>intricate</td><td>2,146</td></tr><tr><td>portrait</td><td>1,604</td></tr><tr><td>beautiful</td><td>1,211</td></tr><tr><td>highly detailed</td><td>701</td></tr><tr><td>realistic</td><td>596</td></tr><tr><td>high quality</td><td>389</td></tr><tr><td>elegant</td><td>381</td></tr><tr><td>illustration</td><td>121</td></tr></table>
+
+Table 6: Analysis of the most frequent words in the optimized prompts in the beautiful-prompt test set.  
+
+<table><tr><td>Tag</td><td>Content</td></tr><tr><td>1</td><td>artstation, highly detailed, elegant</td></tr><tr><td>2</td><td>8k, trending on artstation, concept art</td></tr><tr><td>3</td><td>digital painting, intricate, fantasy</td></tr><tr><td>4</td><td>illustration, smooth, fantasy</td></tr><tr><td>5</td><td>portrait, beautiful, illustration</td></tr><tr><td>6</td><td>realistic, dramatic, high quality</td></tr></table>
+
+Table 7: Combinations of common tags  
+
+<table><tr><td>Metric</td><td>Tag1</td><td>Tag2</td><td>Tag3</td><td>Tag4</td><td>Tag5</td><td>Tag6</td><td>Ours</td></tr><tr><td>PickScore</td><td>20.81</td><td>20.73</td><td>20.68</td><td>20.59</td><td>20.69</td><td>20.82</td><td>20.86</td></tr><tr><td>Aes. Score</td><td>5.80</td><td>5.75</td><td>5.92</td><td>5.67</td><td>5.75</td><td>5.59</td><td>6.59</td></tr></table>
+
+# B Further Discussion
+
+An essential mechanism in self-rewarding model is sampling multiple outputs and then judging them to construct preference pairs for RL fine-tuning. Thus, the diversity of candidates greatly affects the performance of self-rewarding, which can be demonstrated from Tab. 9. It is essential to enlarge the margin between winners and losers with more candidates, which facilitates RL fine-tuning.
+
+Table 8: The result using different groups of common tags.  
+
+<table><tr><td>Method</td><td>Reward</td><td>#Candidates</td><td>PickScore</td><td>Aesthetic</td><td>CLIP</td></tr><tr><td>LVLMsFT</td><td>-</td><td>-</td><td>20.79</td><td>5.95</td><td>0.25</td></tr><tr><td>LVLMDPO1</td><td>LVLMsFT</td><td>2</td><td>20.65</td><td>6.03</td><td>0.23</td></tr><tr><td>LVLMDPO1</td><td>LVLMsFT</td><td>4</td><td>20.78</td><td>6.16</td><td>0.23</td></tr><tr><td>LVLMDPO1</td><td>LVLMsFT</td><td>8</td><td>20.81</td><td>6.31</td><td>0.24</td></tr></table>
+
+Table 9: Number of candidates  
+
+<table><tr><td>Method</td><td>PickScore</td><td>Aes.</td><td>CLIP</td></tr><tr><td>LVLMsFT</td><td>20.79</td><td>5.95</td><td>0.25</td></tr><tr><td>LVLMDPO1</td><td>20.81</td><td>6.31</td><td>0.24</td></tr><tr><td>LVLMDPO2</td><td>20.86</td><td>6.59</td><td>0.24</td></tr><tr><td>LVLMDPO3</td><td>20.86</td><td>6.60</td><td>0.24</td></tr></table>
+
+Table 10: More iteration training.
+
+In addition, the self-reward model is at risk of bias or overfitting. There are many factors that may contribute to model collapse, including data quality and scale, the inherent capabilities of the language model, biases in the generative model, and the number of candidate samples. To demonstrate the effectiveness of our iterative training, we show the results of one more round of training. We used an additional 5,000 data for one more training. As shown in Tab. 10, the third iteration does not cause a significant performance drop.
+
+# C More Quantitative Results
+
+We present more visual results between the images generated with different prompts. As shown in Fig. 7, the optimized prompts result in more pleasing images. A more intuitive observation is that the flat, uninteresting view in Minecraft and the more aesthetically pleasing, more detailed view are represented by the optimized prompt before and after optimization, respectively. In addition, the modified prompt has stronger alignment capabilities. For example, the prompt "Riding a bike on mars" is amended to "... a person riding a bike on mars ..., and the prompt "Galaxy cat" is added with the description "... a space cat ..."
+
+# D Prompts for LVLM-as-a-judge
+
+We present the prompt to transform the LVLM into the reward model. As mentioned in Sec. 3.5, we employ the rating system for image aesthetics, human preferences, and alignment. Note that we input the generated image by original prompt and the image by candidate prompt here to facilitate scoring of the model on the same benchmarks.
+
+For responses judgment, we input two responses to the model that are about how the model evaluated the image in the previous step.
+
+<table><tr><td>User Prompt</td><td>Optimized Prompt</td></tr><tr><td>An isometric medieval castle painted.</td><td>A beautiful intricate detailed painting of a medieval fantasy castle in isometric view, workroom environment, nature:6, fantasy theme, digital painting, photorealistic, very highly detail, in 4k by Jason Felix, trending on Artstation</td></tr><tr><td>Riding a bike on mars.</td><td>A stunning ultra detailed fantasy concept art of a person riding a bike on mars by greg rutkowski, art germ, rhads and thomas kinkade, trending on artstation, cgscociety and unreal engine 5, 8k resolution, futuristic category.</td></tr><tr><td>Street view, zootopia village.</td><td>A beautiful detailed illustration of street view of zootopia village by thomas kinkade, greg rutkowski, andreas rocha, trending on artstation, 4k concept art, digital illustration, detailed art station</td></tr><tr><td>Galaxy cat watercolor painting</td><td>A beautiful detailed watercolor of a cat in the galaxy, a space cat, by amartia durrant, trending on artstation, cgscociety contest winner, fantasy art style, unreal engine 5, photorealism, cinematic 8k, high resolution, imaginative, mind&#x27;s eye, imaginative</td></tr><tr><td>Lakeside mountains, in apache</td><td>An epic detailed illustration of lakeside mountains in apache by thomas kinkade, trending on artstation, cinematic composition dynamic composition detail splash art concept work concept art</td></tr></table>
+
+Figure 7: The images generated using our optimized prompt.
+
+# Aesthetic Score Prompt
+
+Please evaluate the aesthetics of two images ("Image 1" and "Image 2") using the 6-level judging system described below.  
+The two images given are independent, and should be evaluated separately and step by step, ensuring that the order in which the images were presented does not affect your judgment.  
+- Poor (Score: 0): The image lacks balance, composition, and visual appeal. Colors may be overly saturated or dull, causing discomfort to the viewer. Composition is chaotic, distracting, or poorly framed.  
+- Below Average (Score: 1): The image present minimal aesthetic appeal, even if there are inconsistencies or major flaws in composition, color, lighting, or other aesthetic elements, or make people feel disjointed or unbalanced, lacking a cohesive visual narrative.  
+- Average (Score: 2): The image exhibits adequate aesthetic quality contributes to the image's visual appeal to some extent but there is room for improvement in terms of creativity or originality or some aspects of the image may feel generic or uninspired.  
+- Above Average (Score: 3): The image has strong aesthetic quality regardless of whether there are minor imperfections in composition, color, lighting, or other aesthetic elements may still be present but do not significantly detract from the overall aesthetic, or aesthetic choices may be subjective, with some viewers preferring different styles or approaches.  
+- Very Good (Score: 4): The image is of exceptional aesthetic quality and demonstrates creativity, skill, and mastery of visual elements even if there is slight room for improvement in composition, color, lighting, or other aesthetic elements.  
+- Excellent (Score: 5): The image is of perfect balance, harmony, and creativity in aesthetics, creating a visually compelling and impactful image.
+
+Please provide a comprehensive explanation of your score.
+
+Note that the score has nothing to do with image input order.
+
+Output format:
+
+Output for Image 1:
+
+Score: <Your Score for Image 1>
+
+Explanation: <detailed judgment of Score for Image 1>
+
+Output for Image 2:
+
+Score: <Your Score for Image 2>
+
+Explanation: <detailed judgment of Score for Image 2>
+
+# Alignment Score Prompt
+
+Please evaluate the alignment of two pictures ("Image 1" and "Image 2") to the text ("Text") using the 6- level judging system described below.
+
+The two images given are independent, and should be evaluated separately and step by step, ensuring that the order in which the images were presented does not affect your judgment. You need to first consider what appears in the image, then whether what is described in the text appears in the image, and finally give a score based on the system.
+
+Judging system:
+
+- No Match (Score: 0): The image does not contain any of the objects or elements mentioned in the text. There is no recognizable connection between the text and the image.  
+- Poor Match (Score: 1): The image contains one or a few of the objects mentioned in the text, but these are peripheral and do not capture the primary content or relationships described. Quantitative relationships are ignored or inaccurately represented.  
+- Fair Match (Score: 2): Some of the primary objects mentioned in the text are present in the image, and at least one quantitative relationship or object relationship is correctly depicted. However, several key objects or relationships are missing or inaccurately represented.  
+- Good Match (Score: 3): The majority of the objects mentioned in the text are present, and many of the described quantitative relationships and object relationships are accurately depicted. Minor details may be missing or slightly altered.  
+- Excellent Match (Score: 4): Nearly all objects described in the text are accurately represented in the image, including precise quantitative relationships and interactions between objects. Only trivial discrepancies or omissions are present, which do not significantly impact the overall accuracy.  
+- Perfect Match (Score: 5): The image perfectly matches the text in terms of the presence of all described objects, accurate quantitative relationships, and the exact relationships between objects. Every detail mentioned in the text is present and correctly depicted in the image.
+
+Text: <PROMPT>
+
+Please provide a comprehensive explanation of your score. Note that the score has nothing to do with image input order.
+
+Output format:
+
+Output for Image 1:
+
+Score: <Your Score for Image 1>
+
+Explanation: <detailed judgment of Score for Image 1>
+
+Output for Image 2:
+
+Score: <Your Score for Image 2>
+
+Explanation: <detailed judgment of Score for Image 2>
+
+# Pick Score Prompt
+
+Please evaluate how well these two images ("Image 1" and "Image 2") generated based on the text ("Text") are preferred by humans using the 6-level judging system described below. The two images given are independent, and should be evaluated separately and step by step, ensuring that the order in which the images were presented does not affect your judgment. In this system, 'attractiveness' refers to the visual appeal of an image to the human in terms of color, composition, lighting, style, and detail.
+
+Judging system:
+
+- Poor (Score: 0): The image is repulsive or offensive, lacking any attractiveness. It is completely irrelevant to the text information and presents the text in a manner that is unpleasant or unacceptable to the audience.  
+- Below Average (Score: 1): The image has almost no attractiveness, and the audience is indifferent to it. Its relevance to the text information is low, and the presentation style is not attractive enough, making the audience find it rather dull.  
+- Average (Score: 2): The image lacks attractiveness, is ordinary and lacks visual highlights, the conveyed text information is not sufficiently clear, and the presentation style is rather ordinary, lacking novelty or appeal.  
+- Above Average (Score: 3): The image's attractiveness is average, without any particular outstanding features but also not mediocre, conveying the text information and presenting the text in a generally acceptable manner, albeit not particularly outstanding.  
+- Very Good (Score: 4): The image is highly attractive, with good visual effects, conveying the text information basically, and presenting the text in a way that is appealing to humans, allowing the audience to understand and resonate to some extent.  
+- Excellent (Score: 5): The image is extremely attractive, with outstanding visual effects, clearly and accurately conveying the text, and presenting the text in a way that resonates deeply with the audience and evokes strong emotional connections.
+
+Text: <PROMPT>
+
+Please provide a comprehensive explanation of your score.
+
+Note that the score has nothing to do with image input order.
+
+Output format:
+
+Output for Image 1:  
+Score: <Your Score for Image 1>  
+Explanation: <detailed judgment of Score for Image 1>
+
+Output for Image 2:  
+Score: <Your Score for Image 2>  
+Explanation: <detailed judgment of Score for Image 2>
+
+# Aesthetic Judgment Prompt
+
+You are a helpful and precise assistant for checking the quality of the answers. Given the input:
+
+1. Image 1 and Image 2  
+2. Question: {{question}}  
+3. Answer A: {{answer_A}}  
+4. Answer B: {{answer_B}}
+
+Your task is to evaluate the model's predicted answer, based on the context provided by the images and the question. There are two image scores for each answer, and you need to include an evaluation of both outputs ("Output of Image 1" and "Output of Image 2") in each answer. Please provide a comprehensive explanation of your score, noting that your explanation should be based on the facts of the images and not be vague and uninformative.
+
+Consider the following criteria for evaluation:
+
+- Relevance: Does each output in the predicted answer relate to the content of each image?  
+- Accuracy: Does the prediction in each output accurately reflect the information given in the image without introducing factual inaccuracies?  
+- Objectivity: For the analysis of the images, do the two predicted outputs in each answer give approximate scores, avoiding any overestimation or underestimation?  
+- Clarity: Assess the clarity of the predicted answer. Look for issues such as repetition, unclear descriptions, or any grammatical errors that could hinder understanding.  
+- Completeness: Determine if each predicted output in answer fully covers the scope of the images. Does it leave out critical information or does it include all necessary details?
+
+Output Format:
+
+Output for Answer A:
+
+Score: <an integer score of quality from 1-5>
+
+Explanation: <detailed judgment of prediction for "Output of Image 1" and "Output of Image 2">
+
+Output for Answer B:
+
+Score: <an integer score of quality from 1-5>
+
+Explanation: <detailed judgment of prediction for "Output of Image 1" and "Output of Image 2">
+
+# Pick Judgment Prompt
+
+You are a helpful and precise assistant for checking the quality of the answers. Given the input:
+
+1. Image 1 and Image 2
+2. Prompt: {{prompt}}}
+3. Question: {{question}}
+4. Answer A: {{answer_A}}
+5. Answer B: {{answer_B}}
+
+Your task is to evaluate the model's predicted answer, based on the context provided by the images and the question. There are two image scores for each answer, and you need to include an evaluation of both outputs ("Output of Image 1" and "Output of Image 2") in each answer. Please provide a comprehensive explanation of your score, noting that your explanation should be based on the facts of the images and not be vague and uninformative.
+
+Consider the following criteria for evaluation:
+
+- Relevance: Does each outputs in the predicted answer relate to the content of each image?  
+- Accuracy: Does the prediction in each output accurately reflect the information given in the image without introducing factual inaccuracies?  
+- Objectivity: For the analysis of the images, do the two predicted outputs in each answer give approximate scores, avoiding any overestimation or underestimation?  
+- Clarity: Assess the clarity of the predicted answer. Look for issues such as repetition, unclear descriptions, or any grammatical errors that could hinder understanding.  
+- Completeness: Determine if each predicted output in answer fully covers the scope of the images. Does it leave out critical information or does it include all necessary details?
+
+Output Format:
+
+Output for Answer A:  
+Score: <an integer score of quality from 1-5>  
+Explanation: <detailed judgment of prediction for “Output of Image 1” and “Output of Image 2”>  
+Output for Answer B:  
+Score: <an integer score of quality from 1-5>  
+Explanation: <detailed judgment of prediction for “Output of Image 1” and “Output of Image 2”>
+
+Figure 8: LVLM-as-a-judge prompt in our model, which enables the model to provide the aesthetic score, pick score and alignment score for each candidate. All scores are based on the rating system, where inputs are assigned scores corresponding to their ratings. The model is fine-tuned in advance with evaluation data to understand the aesthetics of images.

2025/Self-Rewarding Large Vision-Language Models for Optimizing Prompts in Text-to-Image Generation/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:73581d89ecdeae4ee5ae553028e687dcac6887b4b57290515b323daeba978c98
+size 1073678

2025/Self-Steering Optimization_ Autonomous Preference Optimization for Large Language Models/ddcdb056-3a5c-49ea-857d-7c720881ce91_content_list.json

@@ -0,0 +1,2075 @@
+[
+    {
+        "type": "text",
+        "text": "Self-Steering Optimization: Autonomous Preference Optimization for Large Language Models",
+        "text_level": 1,
+        "bbox": [
+            139,
+            79,
+            860,
+            118
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Hao Xiang $^{1,3}$ , Bowen Yu $^{2*}$ , Hongyu Lin $^{1*}$ , Keming Lu $^{2}$ , Yaojie Lu $^{1}$ , Xianpei Han $^{1}$ , Ben He $^{1,3}$ , Le Sun $^{1}$ , Jingren Zhou $^{2}$ , Junyang Lin $^{2}$",
+        "bbox": [
+            208,
+            124,
+            789,
+            159
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "$^{1}$ Chinese Information Processing Laboratory, Institute of Software, Chinese Academy of Sciences",
+        "bbox": [
+            226,
+            164,
+            769,
+            197
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "2Alibaba Group 3University of Chinese Academy of Sciences",
+        "bbox": [
+            233,
+            198,
+            756,
+            215
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "{xianghao2022, hongyu, luyaojie, xianpei, sunle} @iscas.ac.cn benhe@ucas.edu.cn {yubowen.ybw, lukeming.lkm, jingren.zhou, junyang.ljy} @alibaba-inc.com",
+        "bbox": [
+            156,
+            219,
+            842,
+            254
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            260,
+            261,
+            339,
+            275
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "The key to effective alignment lies in high-quality preference data. Recent research has focused on automated alignment, which involves developing alignment systems with minimal human intervention. However, prior research has predominantly focused on developing data generation methods, while insufficient attention has been paid to quality control mechanisms, which often produce inaccurate and unhelpful data, leading to unpredictable benefits during iterative optimization. In this paper, we present Self-Steering Optimization (SSO), an algorithm that autonomously generates high-quality preference data, eliminating manual annotation requirements. SSO employs a specialized optimization objective to build a data generator from the policy model itself, which is used to produce accurate and on-policy data. We demonstrate SSO's effectiveness through comprehensive experiments on two series of models: Llama 3 and Qwen 2. Our evaluation across diverse benchmarks shows that SSO consistently outperforms baselines in human preference alignment and reward optimization. Further analysis validates SSO as a scalable framework for preference optimization, benefiting the advancement in automated alignment techniques.",
+        "bbox": [
+            144,
+            288,
+            460,
+            684
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1 Introduction",
+        "text_level": 1,
+        "bbox": [
+            114,
+            697,
+            258,
+            712
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "The field of Natural Language Processing has undergone revolutionary advancements driven by large language models (LLMs). After meticulous alignment processes, LLMs have demonstrated remarkable capabilities in following instructions and understanding human preferences. This leads to the development of widely acclaimed products like ChatGPT (OpenAI, 2023), which have captured significant public attention. However, aligning LLMs with human preferences is not trivial. Despite the existence of Proximal Policy Optimization",
+        "bbox": [
+            112,
+            722,
+            489,
+            900
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "(PPO) (Ouyang et al., 2022), an ideal alignment training process requires a robust reward model and a stable reinforcement learning process, encouraging researchers to develop offline preference optimization algorithms such as Direct Preference Optimization (DPO) (Rafailov et al., 2023). However, algorithms like DPO rely on a substantial amount of high-quality, annotated preference data, which is both resource-intensive and requires meticulous attention. In addition, the limited capabilities of human annotators cause inherent limitations in annotated data, making it challenging to achieve superalignment (Burns et al., 2023).",
+        "bbox": [
+            507,
+            261,
+            884,
+            470
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Consequently, recent researchers have shifted their focus towards automated alignment, with the intention of developing scalable, high-quality alignment systems with minimal human intervention. The cornerstone of this paradigm is the pursuit of scalable alignment signals that are capable of effectively replacing human-annotated preference data. Current popular strategies include self-judgement (Yuan et al., 2024; Wu et al., 2024), principle-based automated alignment (Yang et al., 2024b; Bai et al., 2022b), Constitutional AI (Bai et al., 2022b), and other methods (Franken et al., 2024; Kumar et al., 2024).",
+        "bbox": [
+            507,
+            473,
+            884,
+            681
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "However, these methods do not pay enough attention to quality control mechanisms. The ideal preference data, as defined in RLHF (Ouyang et al., 2022), is constructed through human annotation, where the responses are sampled from the policy model and ranked according to their qualities. As shown in Figure 1(a), this type of preference data demonstrates high accuracy<sup>1</sup> and on-policy nature<sup>2</sup>, leading to ideal alignment optimization (Tajwar",
+        "bbox": [
+            507,
+            684,
+            884,
+            829
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "1 Accuracy of preference data: the rate of the response pairs where the chosen response has higher quality than the rejected response.",
+        "bbox": [
+            507,
+            846,
+            882,
+            883
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "$^{2}$ On-policy data: the responses lie in the high-probability region of the policy model. The higher generation probability indicates better on-policy behavior.",
+        "bbox": [
+            507,
+            883,
+            882,
+            920
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "* Corresponding authors.",
+        "bbox": [
+            136,
+            906,
+            295,
+            920
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "9073",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "Findings of the Association for Computational Linguistics: ACL 2025, pages 9073-9085 July 27 - August 1, 2025 ©2025 Association for Computational Linguistics",
+        "bbox": [
+            228,
+            945,
+            768,
+            972
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/919704be7734d09f303bab63a88191451befb585499e0e40c2fcfc629a05b29c.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            147,
+            87,
+            379,
+            206
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/e6b3735d1eefe917f37c755e2a889a0a38146a67d54ace6d88087adbd8be5844.jpg",
+        "image_caption": [
+            "(a) Ideal preference data."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            147,
+            202,
+            379,
+            298
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/5401b6d4ebe5485f6b418b1309e33e8784a5a30ad9ff8b1d25b4505e5c834fa9.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            383,
+            87,
+            613,
+            184
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/c16e08e8f7d797b0377f6e364b225ad0940478b51e6edd6d2738a2557e3b94f8.jpg",
+        "image_caption": [
+            "(b) Accurate but off-policy data."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            383,
+            181,
+            613,
+            297
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/65699d93c601d211be3ebee083ae721f0cf334ddf690ab560b195c7616fc4562.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            615,
+            87,
+            847,
+            181
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/19bd846f76fc735aaa50934d4fedc479603a9f8553621abd1ee697b14a628c91.jpg",
+        "image_caption": [
+            "(c) On-policy but inaccurate data.",
+            "Figure 1: Distribution changes of policy model  $\\pi_{\\theta}$  after optimization with three types of preference data. During alignment, chosen responses  $\\pi_w$  receive positive gradients to increase probability, while rejected responses  $\\pi_l$  receive negative gradients to decrease probability. (a) Ideal data, with high accuracy (low overlap between  $\\pi_w$  and  $\\pi_l$ ) and on-policy nature ( $\\pi_l$  lies in high-probability region of  $\\pi_{\\theta}$ ), leading to right optimization direction and effective negative gradients optimization. (b) Sub-optimal data, with high accuracy (low overlap between  $\\pi_w$  and  $\\pi_l$ ) and off-policy nature ( $\\pi_l$  lies in low-probability region of  $\\pi_{\\theta}$ ), weakening negative gradients optimization. (c) Sub-optimal data, with low accuracy (high overlap between  $\\pi_w$  and  $\\pi_l$ ) and on-policy nature ( $\\pi_l$  lies in high-probability region of  $\\pi_{\\theta}$ ), interfering optimization direction."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            613,
+            181,
+            847,
+            297
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "et al., 2024; Kim et al., 2024). However, automated methods often fail to simultaneously guarantee accurate preferences and on-policy responses, instead producing suboptimal data, as shown in Figures 1(b) and 1(c), which impedes model alignment. For example, self-judgment is hampered by the inherent limitations of the model; this judging ability is restricted and difficult to improve, often resulting in hacked rewards and inaccurate preference data in Figure 1(c) (Wu et al., 2024). In other methods, incorporating additional input or processes may lead to off-policy responses and sub-optimal preference data, as shown in Figure 1(b).",
+        "bbox": [
+            112,
+            472,
+            487,
+            680
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "We then recognized the need for a novel approach to generate high-quality preference data to address these limitations and advance automated alignment. One problem is to control the distribution of the chosen and rejected responses when constructing preference data. For most automated alignment methods, this seems an impossible task because chosen and rejected responses are typically obtained through complex pipelines. However, we found that principle-based methods (Yang et al., 2024b; Bai et al., 2022b) can achieve this goal because they construct preference data by directly sampling from the policy model based on good and bad principles. It can be approximated that the dis",
+        "bbox": [
+            112,
+            696,
+            489,
+            921
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "tributions of the chosen and rejected responses  $\\pi_w$  and  $\\pi_l$  are the distribution of the policy model  $\\pi_{\\theta}$  with good and bad principles.",
+        "bbox": [
+            507,
+            472,
+            880,
+            520
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "In this work, we introduce Self-Steering Optimization (SSO), a pioneering method that automatically generates accurate and near-on-policy preference data for the policy model. SSO optimizes a data generator with a special loss to control the distributions of chosen and rejected responses, and then uses the data produced by this generator to further optimize the policy model. Specifically, SSO first prompts the policy model with original queries  $x$  and a set of contrastive principles  $p^+$  and  $p^-$  for the responses as training data and then optimizes the policy model based on two key objectives: a) Making rejected responses approximately on-policy to ensure the effectiveness of negative gradients in subsequent optimization. We only take care of the rejected responses, as the chosen responses usually have high generation probabilities. b) Maintaining a consistent gap between the chosen and rejected responses to ensure the accuracy of the preference data.",
+        "bbox": [
+            507,
+            527,
+            882,
+            848
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "We demonstrate the effectiveness of Self-Steering Optimization on Qwen2 (Yang et al., 2024a) and Llama3 (Llama Team, 2024) backbones. Our experiments reveal SSO's ability to",
+        "bbox": [
+            507,
+            857,
+            882,
+            921
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "9074",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "generate accurate and on-policy preference data. As a result, improvements are observed on a wide range of benchmarks such as MATH (Hendrycks et al., 2021b), IFEval (Zhou et al., 2023), MT-Bench (Zheng et al., 2024b), and AlpacaEval 2.0 (Dubois et al., 2024). Furthermore, we conducted experiments through reward optimization, which also achieved satisfying results. Without human annotation or external models,  $SSO$  even outperforms baselines with annotated data (Cui et al., 2024), underscoring its potential as a scalable and efficient alignment approach.",
+        "bbox": [
+            112,
+            84,
+            492,
+            280
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "2 Related Works",
+        "text_level": 1,
+        "bbox": [
+            112,
+            290,
+            280,
+            306
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Preference Alignment Researchers have proposed various algorithms to align large language models (LLMs) with human preferences. Ziegler et al. (2020); Ouyang et al. (2022); Bai et al. (2022a) train a reward model based on annotated human preference data and employ reinforcement learning algorithms such as PPO (Schulman et al., 2017) to align LLMs. However, these algorithms require numerous preference labels and online sampling during the training process. To further reduce costs, direct preference optimization (DPO), sequence likelihood calibration (SLiC) (Zhao et al., 2023), identity preference optimization (IPO) (Azar et al., 2023), and Kahneman-Tversky optimization (KTO) (Ethayarajh et al., 2024) simplify the RLHF objective by directly increasing the margin between chosen and rejected responses.",
+        "bbox": [
+            112,
+            316,
+            490,
+            605
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Automated Alignment Previous alignment studies rely on manually annotated preference data and algorithms such as RLHF and DPO to conduct model alignment. Recently, numerous studies have found that LLM-generated data can reach the quality of ordinary manual annotations (Zheng et al., 2024b). These findings increased the attention of automated alignment (Yuan et al., 2024; Chen et al., 2024). Automated alignment aims to minimize human intervention by addressing the prohibitively expensive cost of human annotation. Current methods can be divided into four types based on the source of alignment signals (Cao et al., 2024): 1) Inductive Bias, from introducing appropriate assumptions and constraints (Huang et al., 2023a; Bai et al., 2022b; Yang et al., 2024b; Yuan et al., 2024; Chen et al., 2024). 2) Behavioral Imitation, another aligned model (Peng et al., 2023; Tunstall et al., 2023; Burns et al., 2023). 3) Model Feedback,",
+        "bbox": [
+            112,
+            615,
+            492,
+            921
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "feedbacks from other models (Lee et al., 2023; Hosseini et al., 2024). 4) Environmental Feedback, environmental interaction (Liu et al., 2023; Qiao et al., 2024).",
+        "bbox": [
+            507,
+            84,
+            884,
+            149
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3 Preliminaries",
+        "text_level": 1,
+        "bbox": [
+            509,
+            162,
+            662,
+            177
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.1 Symbol Definition of Automated Alignment",
+        "text_level": 1,
+        "bbox": [
+            507,
+            187,
+            808,
+            219
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Specifically, given a query set  $X = \\{x_{i}\\}_{i = 1}^{N}$ , where  $N$  is the number of queries, automated methods focus on how to use the policy model  $\\pi_{\\theta}$  to generate the chosen response  $y^{+}$  and the rejected response  $y^{-}$  for the preference data  $D = \\{x_{i},y_{i}^{+},y_{i}^{-}\\}_{i = 1}^{N}$  which will be used to optimize  $\\pi_{\\theta}$  with alignment algorithms.",
+        "bbox": [
+            507,
+            225,
+            884,
+            338
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.2 Principle-Based Automated Alignment",
+        "text_level": 1,
+        "bbox": [
+            507,
+            350,
+            858,
+            367
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Principle-based automated alignment (PBAA) is one of the most common automated alignment methods (Yang et al., 2024b; Frängen et al., 2024), which assumes that responses with different qualities can be directly sampled from LLMs with different queries. This approach constructs pairs of contrastive queries  $x^{+}$  and  $x^{-}$  to sample chosen and rejected responses from the policy model as training data. Since contrastive queries exhibit contrasting attributes (such as harmful vs. harmless), the generated preference data has high accuracy. Representative works of PBAA include RLCD (Yang et al., 2024b), AutoPM (Huang et al., 2023b) and SAIM (Frängen et al., 2024). The first two use specific word pairs, such as \"inoffensive response\" and \"offensive response\", to generate response pairs for model alignment, while SAIM employs generated principles.",
+        "bbox": [
+            507,
+            370,
+            884,
+            661
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "However, these methods do not guarantee accurate and on-policy data. Incorporating additional principles could lead to off-policy responses, as shown in Figure 1(b) (Tajwar et al., 2024; Kim et al., 2024).",
+        "bbox": [
+            507,
+            662,
+            882,
+            741
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.3 Automated Alignment from Distribution Sight",
+        "text_level": 1,
+        "bbox": [
+            507,
+            753,
+            873,
+            785
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "In the alignment process, negative gradients and positive gradients are used to decrease and increase the probability of rejected responses and chosen responses, respectively. Figure 1(a) illustrates the ideal distribution of the preference data.",
+        "bbox": [
+            507,
+            791,
+            882,
+            872
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "For models that are not aligned at all, automated alignment approaches may construct preference data where the distribution of the chosen responses",
+        "bbox": [
+            507,
+            873,
+            882,
+            921
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "9075",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/c9c952f4317e373680a21f6a718996199449309367b0ac94f2cad74ac564a817.jpg",
+        "image_caption": [
+            "Figure 2: The pipeline of Self-Steering Optimization (SSO). The overall process consists of four steps: 1) generate Principles, 2) sample Responses, 3) optimize the Data Generator, and 4) optimize the policy model. The data generator is optimized with the loss 1 from the policy model  $\\pi_{\\theta}$  and used to generate accurate and on-policy preference data for the policy model."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            122,
+            86,
+            873,
+            407
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "$\\pi_w$  is far from the distribution of rejected responses  $\\pi_l$ , and  $\\pi_w$  are in the low-probability region of  $\\pi_\\theta$ , as mentioned by Tajwar et al. (2024), the on-policy nature has minimal impact on model optimization in this scenario. This explains the improvements of various automated methods.",
+        "bbox": [
+            112,
+            506,
+            487,
+            601
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "However, as LLMs advance, even those without explicit alignment can exhibit aligned behaviors. In this situation,  $\\pi_w$  lies in the high-probability region of  $\\pi_{\\theta}$ , and the on-policy performance of  $\\pi_l$  becomes crucial. If  $\\pi_l$  lies in the low-probability region of  $\\pi_{\\theta}$ , applying negative gradients to such responses would be meaningless and result in sub-optimal alignment. Therefore, we propose Self-Steering Optimization.",
+        "bbox": [
+            112,
+            604,
+            487,
+            747
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4 Self-Steering Optimization",
+        "text_level": 1,
+        "bbox": [
+            112,
+            762,
+            379,
+            780
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4.1 Pipeline of SSO",
+        "text_level": 1,
+        "bbox": [
+            112,
+            790,
+            284,
+            806
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "As shown in Figure 2, the pipeline of  $SSO$  consists of four steps: 1) generating principles, 2) sampling responses, 3) optimizing the data generator, and 4) optimizing the policy model.",
+        "bbox": [
+            112,
+            813,
+            487,
+            878
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Generating Principles Given a query  $x$ , we employ a principle generator (we used Qwen2.5-72B",
+        "bbox": [
+            112,
+            889,
+            489,
+            921
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Instruct(Qwen et al., 2025) in our main experiments) to construct a pair of contrastive principles  $(p^{+},p^{-})$ . The example of principles can be found in the Appendix 7.",
+        "bbox": [
+            507,
+            506,
+            884,
+            571
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Sampling Responses Principles  $(p^{+}, p^{-})$  are then concatenated with the original query  $x$  to build contrastive queries  $(x^{+}, x^{-})$ . Then,  $(x^{+}, x^{-})$  are used to prompt the policy model  $\\pi_{\\theta}$  for the good and bad responses  $(y^{+}, y^{-})$ . Furthermore, to optimize the data generator, we generate an original response  $y^{o}$  without any principle.",
+        "bbox": [
+            507,
+            582,
+            882,
+            696
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Optimizing Data Generator We optimize a data generator with the loss function in Formula 1, which is designed to generate accurate and near-on-policy preference data. The data generator is optimized from the policy model.",
+        "bbox": [
+            507,
+            706,
+            882,
+            787
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Optimizing Policy Model We optimize the policy model  $\\pi_{\\theta}$  with the preference data generated by the data generator. The policy model is optimized with an alignment algorithm, such as DPO, to align the model.",
+        "bbox": [
+            507,
+            797,
+            882,
+            876
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Detailed templates are provided in the Appendix.",
+        "bbox": [
+            526,
+            879,
+            882,
+            895
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_footnote",
+        "text": "3We use principles as system messages.",
+        "bbox": [
+            529,
+            906,
+            773,
+            921
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "9076",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4.2 Objective",
+        "text_level": 1,
+        "bbox": [
+            112,
+            84,
+            236,
+            99
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Self-Steering Optimization aims to generate accurate and near-on-policy preference data. As described in Section 4.1, given the following components:",
+        "bbox": [
+            112,
+            105,
+            489,
+            168
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Two principles : Good principle  $p^+$  and bad principle  $p^-$ , which are generated by LLMs for the query  $x$  specially.",
+        "bbox": [
+            112,
+            178,
+            487,
+            227
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Three queries : Original query  $x$ , good query  $x^{+}$ , and bad query  $x^{-}$ , where  $x^{+}$  and  $x^{-}$  are built from  $x$  with  $p^{+}$  and  $p^{-}$ , respectively.",
+        "bbox": [
+            112,
+            235,
+            487,
+            285
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Three responses : The good response  $y^{+}$ , the bad response  $y^{-}$ , and the original response  $y^{o}$ , where  $y^{i}$  is sampled from  $\\pi_{\\theta}$  with query  $x^{i}$ ,  $i \\in \\{+, -, o\\}$ .",
+        "bbox": [
+            112,
+            292,
+            487,
+            356
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We propose a novel loss function  $\\mathcal{L}_{SSO}$  to optimize the data generator:",
+        "bbox": [
+            112,
+            357,
+            487,
+            388
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {S S O} = \\mathcal {L} ^ {+} + \\mathcal {L} ^ {-} \\tag {1}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            226,
+            400,
+            487,
+            419
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $\\mathcal{L}^+$  and  $\\mathcal{L}^{-}$  are used to optimize the good and bad responses, respectively.",
+        "bbox": [
+            112,
+            431,
+            485,
+            463
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Loss 1 is the core loss function of our method, used to optimize the Policy model to obtain a Data Generator (as shown in Step 3 of Figure 1). Although the Data Generator is optimized from the Policy Model, it has no direct relationship with the final aligned model and is only used to generate data.  $\\mathcal{L}_{SSO}$  is only used to optimize the Data Generator, not for the final alignment. When aligning the policy model with the data generated by the Data Generator, we use  $\\mathcal{L}_{Base}$ , which is the  $xPO$  Loss, such as DPO, IPO, etc. This might be the source of your confusion.",
+        "bbox": [
+            112,
+            464,
+            487,
+            655
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "As mentioned in Section 3.3,  $\\mathcal{L}_{SSO}$  should minimize the overlap between  $\\pi_w$  and  $\\pi_l$ , while ensuring that  $\\pi_l$  lies in the high-probability region of  $\\pi_{\\theta}$  (for effective negative gradients optimization). Naturally, we design  $\\mathcal{L}^{+}$  as:",
+        "bbox": [
+            112,
+            657,
+            489,
+            737
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} ^ {+} = L _ {\\text {B a s e}} \\left(\\mathbf {x} ^ {+}, \\mathbf {y} ^ {+}, \\mathbf {y} ^ {-}\\right) + \\gamma L _ {\\text {s f t}} \\left(\\mathbf {x} ^ {+}, \\mathbf {y} ^ {+}\\right) \\tag {2}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            122,
+            747,
+            487,
+            766
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $L_{Base}$  can be any  $xPO$  alignment algorithm (like DPO, IPO, etc.) loss function, and  $L_{sft}$  is the SFT loss function.  $\\gamma$  is a hyperparameter that balances SFT loss and alignment loss, helping to maintain training stability.",
+        "bbox": [
+            112,
+            778,
+            487,
+            859
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Similarly, a natural approach is to construct the loss function  $\\mathcal{L}^{-}$  as:",
+        "bbox": [
+            112,
+            860,
+            487,
+            891
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} ^ {-} = L _ {\\text {B a s e}} \\left(\\mathbf {x} ^ {-}, \\mathbf {y} ^ {-}, \\mathbf {y} ^ {+}\\right) + \\gamma L _ {s f t} \\left(\\mathbf {x} ^ {-}, \\mathbf {y} ^ {-}\\right) \\tag {3}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            122,
+            904,
+            487,
+            921
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "However, this approach introduces a problem: with a bad principle  $p^{-}$ , LLMs may output unpredictable results. In other words, this loss could lead to a  $\\pi_l$  that lies in the low-probability region of  $\\pi_{\\theta}$ , which cannot help generate on-policy data and hamper alignment.",
+        "bbox": [
+            507,
+            84,
+            882,
+            180
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Therefore, for the optimization of  $\\pi_l$ , we change the loss to  $L_{Base}(\\mathbf{x}^{-},\\mathbf{y}^{\\mathbf{o}},\\mathbf{y}^{+})$ . This goal is crucial, as we want to avoid shifting  $p^{-}$  to the low-probability region of  $\\pi_{\\theta}$ . And the final form of  $\\mathcal{L}^{-}$  is:",
+        "bbox": [
+            507,
+            181,
+            882,
+            259
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} ^ {-} = L _ {\\text {B a s e}} \\left(\\mathbf {x} ^ {-}, \\mathbf {y} ^ {\\mathbf {0}}, \\mathbf {y} ^ {+}\\right) + \\gamma L _ {s f t} \\left(\\mathbf {x} ^ {-}, \\mathbf {y} ^ {\\mathbf {0}}\\right) (4)\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            519,
+            269,
+            882,
+            288
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "5 Experiments",
+        "text_level": 1,
+        "bbox": [
+            507,
+            296,
+            655,
+            312
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "5.1 Experimental Setup",
+        "text_level": 1,
+        "bbox": [
+            507,
+            321,
+            714,
+            337
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Models and Datasets We conducted experiments primarily on Qwen2-7B (Yang et al., 2024a) and Llama3-8B (Llama Team, 2024). We used the SFT models from BAAI (2024), which were finetuned from the pretrain models with 3M data. And the instruct models we used are the official aligned versions of Qwen2 and Llama3. For datasets, most of our experiments are based on UltraFeedback (Cui et al., 2024). This dataset includes 60k annotations of preference data. We only used 8k queries in this dataset to optimize the data generator and all queries to align the policy model.",
+        "bbox": [
+            507,
+            341,
+            882,
+            533
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Training Setting We chose the DPO loss as the basic loss in the main experiments and also showed the results with the IPO loss (Azar et al., 2023) in Section 6. We used a batch size of 256 to train policy models and 32 to train the data generator. We applied a simple hyperparameter search to determine the learning rate and the  $\\beta$  parameter in DPO. We trained models with the learning rate 5E-7 and beta with 0.1. We set  $\\gamma$  in  $SSO$  to 0.1. We used the top-p = 0.8, temperature = 0.7, and max_new_tokens = 2048 for sampling responses. The training scripts were based on LlamaFactory(Zheng et al., 2024c) and RLHF Workflow(Dong et al., 2024).",
+        "bbox": [
+            507,
+            542,
+            882,
+            768
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Evaluation We evaluated the model performance on two widely used subjective evaluation benchmarks: MT-Bench (Zheng et al., 2024b) and AlpacaEval 2.0 (Dubois et al., 2024). MT-Bench comprises 80 questions with answers scored by GPT-4. AlpacaEval 2.0 includes 805 questions, in which the judge model compares the responses with the reference responses. Additionally, we evaluated models on a series of objective benchmarks:",
+        "bbox": [
+            507,
+            776,
+            882,
+            920
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "9077",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "table",
+        "img_path": "images/aa5893d0500fdcb602303e830608bf84313b8fad58f301e2c927346567d3322c.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Model</td><td rowspan=\"2\">Synthetic Data</td><td colspan=\"2\">AlpacaEval</td><td>MTbench</td><td>IFEval</td><td>GSM8K</td><td>MATH</td><td>MMLU</td></tr><tr><td>lc win rate</td><td>win rate</td><td>score</td><td>avg score</td><td>acc</td><td>acc</td><td>acc</td></tr><tr><td>LLAMA3-SFT (BAAI, 2024)</td><td></td><td>20.6</td><td>15.0</td><td>7.38</td><td>24.9</td><td>75.6</td><td>29.5</td><td>65.9</td></tr><tr><td>w/ ULTRAFEEDBACK (Cui et al., 2024)</td><td>×</td><td>22.0</td><td>17.5</td><td>7.71</td><td>43.6</td><td>78.3</td><td>30.5</td><td>66.4</td></tr><tr><td>Principle-Based Automated Alignment</td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td></tr><tr><td>w/ PBAADPO</td><td>✓</td><td>29.5</td><td>24.0</td><td>7.92</td><td>47.8</td><td>77.9</td><td>30.4</td><td>66.3</td></tr><tr><td>w/ SSO_DPO</td><td>✓</td><td>35.0</td><td>28.3</td><td>7.96</td><td>50.3</td><td>80.5</td><td>30.8</td><td>66.7</td></tr><tr><td>QWEN2-SFT (BAAI, 2024)</td><td></td><td>20.1</td><td>13.2</td><td>8.24</td><td>19.8</td><td>78.5</td><td>44.6</td><td>71.1</td></tr><tr><td>w/ ULTRAFEEDBACK</td><td>×</td><td>20.2</td><td>15.0</td><td>8.35</td><td>40.4</td><td>84.3</td><td>46.7</td><td>71.1</td></tr><tr><td>Principle-Based Automated Alignment</td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td></tr><tr><td>w/ PBAADPO</td><td>✓</td><td>37.7</td><td>42.5</td><td>8.59</td><td>43.6</td><td>83.8</td><td>50.8</td><td>70.9</td></tr><tr><td>w/ SSO_DPO</td><td>✓</td><td>43.0</td><td>45.4</td><td>8.66</td><td>45.7</td><td>84.7</td><td>52.3</td><td>71.0</td></tr><tr><td>LLAMA3-INSTRUCT (Llama Team, 2024)</td><td></td><td>20.1</td><td>13.2</td><td>8.06</td><td>53.0</td><td>80.4</td><td>28.5</td><td>68.4</td></tr><tr><td>w/ ULTRAFEEDBACK</td><td>×</td><td>23.8</td><td>22.6</td><td>7.72</td><td>54.4</td><td>79.1</td><td>29.6</td><td>68.4</td></tr><tr><td>Principle-Based Automated Alignment</td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td></tr><tr><td>w/ PBAADPO</td><td>✓</td><td>23.8</td><td>25.7</td><td>8.05</td><td>53.2</td><td>79.5</td><td>28.9</td><td>68.0</td></tr><tr><td>w/ SSO_DPO</td><td>✓</td><td>25.6</td><td>28.9</td><td>8.13</td><td>53.4</td><td>80.7</td><td>29.6</td><td>68.4</td></tr><tr><td>QWEN2-INSTRUCT (Yang et al., 2024a)</td><td></td><td>19.5</td><td>17.2</td><td>8.33</td><td>51.4</td><td>81.0</td><td>40.0</td><td>71.0</td></tr><tr><td>w/ ULTRAFEEDBACK</td><td>×</td><td>20.4</td><td>17.6</td><td>8.21</td><td>51.5</td><td>82.3</td><td>42.9</td><td>71.0</td></tr><tr><td>Principle-Based Automated Alignment</td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td></tr><tr><td>w/ PBAADPO</td><td>✓</td><td>34.7</td><td>43.7</td><td>8.37</td><td>50.9</td><td>78.1</td><td>42.1</td><td>71.0</td></tr><tr><td>w/ SSO_DPO</td><td>✓</td><td>36.5</td><td>49.4</td><td>8.39</td><td>51.4</td><td>78.5</td><td>44.2</td><td>71.2</td></tr></table>",
+        "bbox": [
+            117,
+            80,
+            878,
+            406
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Table 1: Evaluation results on six distinct tasks. \"Ic win rate\" indicates \"Length Control Win Rate\" from AlpacaEval 2.0 (Dubois et al., 2024).",
+        "bbox": [
+            112,
+            416,
+            882,
+            445
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "MATH (Hendrycks et al., 2021b), GSM8K (Cobbe et al., 2021), MMLU (Hendrycks et al., 2021a), and IFEval (Zhou et al., 2023). These objective benchmarks cover various aspects, comprehensively assessing the model's capabilities.",
+        "bbox": [
+            112,
+            458,
+            487,
+            539
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "5.2 Main Results",
+        "text_level": 1,
+        "bbox": [
+            112,
+            557,
+            263,
+            571
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "This part compares the performance of  $SSO$  with  $PBAA$  and UltraFeedback. Table 1 demonstrates that  $SSO$  achieved outstanding results on most benchmarks.",
+        "bbox": [
+            112,
+            581,
+            485,
+            644
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "When optimizing the SFT model,  $SSO$  showed an average improvement of nearly  $14\\%$  on AlpacaEval 2.0 and 0.5 points on MTBench. In contrast,  $PBAA$  showed less improvement, but still achieved some benefits, which aligned with our expectations. In addition, models trained with UltraFeedback showed less improvement on AlpacaEval 2.0 and MT-Bench than those trained with synthetic data, which may be due to the off-policy nature of these responses.  $SSO$  also showed benefits on objective benchmarks. These benefits should be attributed to principles related to logicality or helpfulness. Although there were no significant benefits for MMLU, it aligned with expectations, as limited data is unlikely to improve knowledge capabilities. We also applied  $SSO$  to aligned models such as Meta-Llama-3-8B-Instruct",
+        "bbox": [
+            112,
+            646,
+            489,
+            921
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "and Qwen2-7B-Instruct, with the results shown in Table 1. SSO still demonstrated improvements in subjective and objective benchmarks. Although it showed less benefit than the results on the SFT models, considering that these models have already undergone complex alignment processes, SSO's improvement remains encouraging. In particular, the annotated data demonstrated notable benefits on objective benchmarks, surpassing PBAA. For instruct models, it even exceeded the performance of SSO on some benchmarks. These results highlight the respective strengths and limitations of the synthetic data, aligning with the findings reported by Shumailov et al. (2024).",
+        "bbox": [
+            507,
+            458,
+            884,
+            684
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "5.3 Results in Reward Optimization",
+        "text_level": 1,
+        "bbox": [
+            507,
+            703,
+            808,
+            718
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "We also trained a reward model based on the Llama3-8B-Instruct with the data generated during previous experiments. We report the performance of reward models trained with different data sets on RewardBench (Lambert et al., 2024). As shown in Table 2,  $SSO$  could be used to train an advanced reward model. This model gets a 80.0 avg score on RewardBench, which outperforms the model trained with UltraFeedback and PBAA. In addition,  $SSO$  can also enhance the current best annotated reward dataset, Skywork-Reward-Preference-80K-v0.2 (Liu et al., 2024). The mixed",
+        "bbox": [
+            507,
+            728,
+            884,
+            921
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "9078",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/7770497960d087998a84265e8b9c58459059d0dab6d0411301a9b5aceb89f6ea.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Training Data</td><td>Type</td><td>Size</td><td>Avg</td><td>Chat</td><td>Chat Hard</td><td>Safty</td><td>Reasoning</td></tr><tr><td>ULTRAFEEDBACK</td><td>Annotated</td><td>60k</td><td>79.5</td><td>96.9</td><td>61.8</td><td>79.2</td><td>80.2</td></tr><tr><td>PBAADPO</td><td>Synthetic</td><td>60k</td><td>78.9</td><td>96.9</td><td>59.9</td><td>77.2</td><td>81.6</td></tr><tr><td>SSO_DPO</td><td>Synthetic</td><td>60k</td><td>80.0</td><td>95.3</td><td>59.0</td><td>77.4</td><td>88.3</td></tr><tr><td colspan=\"8\">Mixed with Skywork-Reward-Preference-80K-v0.2, a SOTA preference dataset, as training data.</td></tr><tr><td>SKYWORK (Liu et al., 2024)</td><td>Annotated</td><td>80k</td><td>84.5</td><td>91.6</td><td>78.6</td><td>88.2</td><td>79.7</td></tr><tr><td>+ ULTRAFEEDBACK</td><td>Annotated</td><td>140k</td><td>85.1</td><td>94.7</td><td>72.7</td><td>89.1</td><td>83.9</td></tr><tr><td>+ PBAADPO</td><td>Mixed</td><td>140k</td><td>83.5</td><td>93.3</td><td>75.1</td><td>86.5</td><td>79.0</td></tr><tr><td>+ SSO_DPO</td><td>Mixed</td><td>140k</td><td>86.3</td><td>93.9</td><td>75.4</td><td>86.6</td><td>89.3</td></tr></table>",
+        "bbox": [
+            149,
+            80,
+            848,
+            227
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "datasets show a more significant difference. Mixing the SSO dataset with Skywork showed an average score of 86.3 and 1.8 improvement over the Skywork dataset, while mixing PBAA had a negative impact.",
+        "bbox": [
+            112,
+            274,
+            487,
+            354
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "6 Discussion",
+        "text_level": 1,
+        "bbox": [
+            112,
+            366,
+            240,
+            382
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "6.1 Ablation Study",
+        "text_level": 1,
+        "bbox": [
+            112,
+            392,
+            278,
+            407
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/32542d50a8e552716aa36a784fe9c74b22ee3cebfe5c87995cfd11967175fc96.jpg",
+        "table_caption": [
+            "Table 2: Evaluation results on RewardBench. Models are optimized from Llama3-8B-Instruct (Llama Team, 2024). We trained the reward model with code from Dong et al. (2024)."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Model</td><td>AlpacaEval</td><td>Arena Hard(95% CI)</td></tr><tr><td>LLAMA3-SFT</td><td>20.6</td><td>20.5 (-1.9, 1.8)</td></tr><tr><td>PBAA</td><td>29.5</td><td>27.9 (-1.7, 1.9)</td></tr><tr><td>+L+</td><td>33.6</td><td>27.4 (-1.5, 2.1)</td></tr><tr><td>+L-</td><td>32.0</td><td>28.5 (-1.3, 1.5)</td></tr><tr><td>+L++L-</td><td>35.0</td><td>29.6 (-1.8, 2.2)</td></tr><tr><td>LLAMA3-INSTRUCT</td><td>20.1</td><td>20.7 (-1.7, 1.9)</td></tr><tr><td>PBAA</td><td>23.8</td><td>23.0 (-2.2, 2.3)</td></tr><tr><td>+L+</td><td>25.3</td><td>22.8 (-1.9, 1.7)</td></tr><tr><td>+L-</td><td>25.3</td><td>21.4 (-1.5, 1.7)</td></tr><tr><td>+L++L-</td><td>25.6</td><td>25.2 (-2.1, 2.0)</td></tr></table>",
+        "bbox": [
+            117,
+            423,
+            485,
+            602
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "We conduct an ablation study to validate the necessity of the  $L + / L -$  design. Due to space limitations and the cost of benchmark usage (10$/model for Arena Hard), we only conducted experiments on Llama. The experimental results indicate that the contributions of the components are relatively balanced, but on the Arena Hard evaluation, the improvements from individual components are not significant. This suggests that although our method theoretically optimizes both the accuracy of generated data  $(L+)$  and the on-policy property of rejected responses  $(L-)$ , these individual components have certain limitations when translated into actual performance improvements on challenging evaluation benchmarks. Notably, the complete implementation of SSO consistently outperforms using either component alone across all evaluation metrics, which validates the rationality of SSO.",
+        "bbox": [
+            112,
+            631,
+            489,
+            921
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "6.2 Different  $L^{-}$ oss in SSO",
+        "text_level": 1,
+        "bbox": [
+            509,
+            274,
+            739,
+            288
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/8c56db3c933424bd0696403c2cb5d4cd152a7e029918df319035fde678de8fb0.jpg",
+        "table_caption": [
+            "Table 3: Results of ablation experiments."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Model</td><td>AlpacaEval</td><td>MTbench</td></tr><tr><td>LLAMA3-8B-SFT</td><td>20.6</td><td>7.38</td></tr><tr><td>SSO</td><td>35.0</td><td>7.96</td></tr><tr><td>SSO with another L-</td><td>31.8</td><td>7.75</td></tr><tr><td>QWEN2-7B-SFT</td><td>20.1</td><td>8.24</td></tr><tr><td>SSO</td><td>43.0</td><td>8.66</td></tr><tr><td>SSO with another L-</td><td>38.6</td><td>8.63</td></tr></table>",
+        "bbox": [
+            531,
+            304,
+            858,
+            437
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Table 4: Results with the  $L^{-}$  in Formula 3.",
+        "bbox": [
+            547,
+            447,
+            842,
+            461
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "As mentioned in Section 4.2, we had two different  $L^{-}$  losses in SSO. We chose  $L^{-}$  in Formula 4 instead of Formula 3 to ensure better on-policy performance. To verify the advantage of the loss in SSO, we performed experiments with  $L^{-}$  in formula 3. The results are shown in Table 4. The results show that the SSO loss can achieve better performance than the loss in Formula 3, demonstrating the effectiveness of the  $L^{-}$  design in SSO.",
+        "bbox": [
+            507,
+            469,
+            882,
+            615
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "6.3 Quality of Synthetic Data",
+        "text_level": 1,
+        "bbox": [
+            507,
+            626,
+            756,
+            642
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "In general, more accurate preference data is believed to lead to a better alignment process (Lee et al., 2024; Gao et al., 2024). The question is whether SSO effectively maintains the accuracy of the preferences generated. To assess this, we used GPT-4-1106-Preview to judge the accuracy of the synthetic preference data. Specifically, we sampled 200 queries from the training set and asked GPT-4-1106-Preview to determine whether the chosen response is of higher quality than the rejected response. As shown in Figure 3, SSO maintained and even improved the accuracy of preference data. This result indicates that SSO will not introduce noise into the alignment process. We also analyzed",
+        "bbox": [
+            505,
+            648,
+            882,
+            873
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_footnote",
+        "text": "4To mitigate selection bias (Zheng et al., 2024a), we swapped the positions of these two responses for two rounds of judgment.",
+        "bbox": [
+            507,
+            883,
+            882,
+            921
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "9079",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/6462c1c707d994a987fa0a2fa686258a587f5d5029168669881e8cf071b26f79.jpg",
+        "image_caption": [
+            "Figure 3: The quality of the synthetic data generated by SSO and PBAA. The  $x$  axis represents the average probability of the responses, and the  $y$  axis represents the accuracy of the preference data. Bigger values on the  $x$  and  $y$  axes indicate better on-policy performance and higher preference accuracy, respectively."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            119,
+            84,
+            480,
+            311
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "the on-policy performance of the synthetic data by calculating the generation probability  $e^{\\pi_{\\theta}(\\mathbf{y}|\\mathbf{x})}$  of all responses. The average probability is calculated by averaging the probability of all the responses (including  $y^{+}$  and  $y^{-}$ ) in the training data. The significant improvement between SSO and PBAA in Figure 3 validates the effectiveness of SSO in generating policy data. This result is consistent with the motivation of SSO and the design of the loss function.",
+        "bbox": [
+            112,
+            434,
+            489,
+            594
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "6.4 Weaker Principle Generator",
+        "text_level": 1,
+        "bbox": [
+            112,
+            606,
+            383,
+            621
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/311b45f87cc7fe8653f03b38f3eace6fe85ab6361ebc14eb75529e97e913ee74.jpg",
+        "table_caption": [],
+        "table_footnote": [
+            "Table 5: Results with weaker principle generator. We use Qwen2.5-72b-Instruct as strong generator and Qwen2.5-3b-Instruct as the weak one."
+        ],
+        "table_body": "<table><tr><td rowspan=\"2\">Model</td><td rowspan=\"2\">principle generator</td><td colspan=\"2\">AlpacaEval</td><td rowspan=\"2\">MTbench score</td></tr><tr><td>lc win rate</td><td></td></tr><tr><td>LLAMA3-8B-SFT</td><td>-</td><td>20.6</td><td></td><td>7.38</td></tr><tr><td rowspan=\"2\">w/ PBAA</td><td>strong</td><td>29.5</td><td></td><td>7.92</td></tr><tr><td>weak</td><td>31.8</td><td></td><td>7.89</td></tr><tr><td rowspan=\"2\">w/ SSO</td><td>strong</td><td>35.0</td><td></td><td>7.96</td></tr><tr><td>weak</td><td>34.0</td><td></td><td>7.99</td></tr><tr><td>QWEN2-7B-SFT</td><td>-</td><td>20.1</td><td></td><td>8.24</td></tr><tr><td rowspan=\"2\">w/ PBAA</td><td>strong</td><td>37.7</td><td></td><td>8.59</td></tr><tr><td>weak</td><td>36.2</td><td></td><td>8.60</td></tr><tr><td rowspan=\"2\">w/ SSO</td><td>strong</td><td>43.0</td><td></td><td>8.66</td></tr><tr><td>weak</td><td>38.6</td><td></td><td>8.50</td></tr></table>",
+        "bbox": [
+            117,
+            637,
+            485,
+            832
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "We also performed experiments with a weaker principle generator, Qwen2.5-3B-Instruct, to ex",
+        "bbox": [
+            112,
+            889,
+            489,
+            921
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "plore whether the principle generator affects the performance of  $SSO$ . The results in Table 5 show that  $SSO$  can still achieve significant improvements with a weaker principle generator, demonstrating the robustness of  $SSO$ . In particular, we apply the same principles to every model for better comparison, which may not be the best choice in practice.",
+        "bbox": [
+            507,
+            84,
+            884,
+            212
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "6.5 IPO-Based SSO",
+        "text_level": 1,
+        "bbox": [
+            509,
+            225,
+            680,
+            239
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/5fb8ab3bf328b789172daa78d9838fd0e929e576546bd8e9bd655fc578e0898f.jpg",
+        "table_caption": [],
+        "table_footnote": [
+            "Table 6: Few-shot evaluation results on three Subjective tasks. Models are optimized with IPO (Azar et al., 2023)."
+        ],
+        "table_body": "<table><tr><td rowspan=\"2\">Model</td><td colspan=\"2\">AlpacaEval MTbench</td><td>IFEval</td></tr><tr><td>lc win rate</td><td>score</td><td>avg score</td></tr><tr><td>LLAMA3-8B-SFT</td><td>20.6</td><td>7.38</td><td>24.9</td></tr><tr><td>w/ ULTRAFEEDBACK</td><td>48.1</td><td>8.08</td><td>46.2</td></tr><tr><td>Principle-Based Automated Alignment</td><td></td><td></td><td></td></tr><tr><td>w/ PBAA_DPO</td><td>47.9</td><td>7.85</td><td>45.8</td></tr><tr><td>w/ SSO_DPO</td><td>51.4</td><td>7.81</td><td>46.3</td></tr><tr><td>QWEN2-7B-SFT</td><td>20.1</td><td>8.24</td><td>19.8</td></tr><tr><td>w/ ULTRAFEEDBACK</td><td>43.9</td><td>8.43</td><td>46.0</td></tr><tr><td>Principle-Based Automated Alignment</td><td></td><td></td><td></td></tr><tr><td>w/ PBAA_DPO</td><td>44.8</td><td>8.48</td><td>44.8</td></tr><tr><td>w/ SSO_DPO</td><td>46.2</td><td>8.65</td><td>44.2</td></tr><tr><td>LLAMA3-8B-INSTRUCT</td><td>20.1</td><td>8.06</td><td>53.0</td></tr><tr><td>w/ ULTRAFEEDBACK</td><td>32.1</td><td>8.00</td><td>57.1</td></tr><tr><td>Principle-Based Automated Alignment</td><td></td><td></td><td></td></tr><tr><td>w/ PBAA_DPO</td><td>30.1</td><td>7.95</td><td>59.6</td></tr><tr><td>w/ SSO_DPO</td><td>32.4</td><td>8.15</td><td>59.9</td></tr><tr><td>QWEN2-7B-INSTRUCT</td><td>19.5</td><td>8.33</td><td>51.4</td></tr><tr><td>w/ ULTRAFEEDBACK</td><td>30.9</td><td>8.30</td><td>48.8</td></tr><tr><td>Principle-Based Automated Alignment</td><td></td><td></td><td></td></tr><tr><td>w/ PBAA_DPO</td><td>27.8</td><td>8.54</td><td>47.8</td></tr><tr><td>w/ SSO_DPO</td><td>28.4</td><td>8.50</td><td>51.9</td></tr></table>",
+        "bbox": [
+            514,
+            255,
+            878,
+            548
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Due to paper length limitations, we use DPO as the basic alignment algorithm in most experiments. However, we also conducted experiments with IPO(Azar et al., 2023). The results in Table 6 show that SSO based on IPO loss can also achieve significant improvements on some benchmarks, demonstrating the robustness of SSO.",
+        "bbox": [
+            507,
+            608,
+            882,
+            720
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "7 Conclusion",
+        "text_level": 1,
+        "bbox": [
+            507,
+            733,
+            640,
+            747
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "In this work, we proposed a novel approach called SSO (Self-Steering Optimization) to enhance model alignment by generating accurate and on-policy preference data without additional human annotations. SSO applying two specific losses  $\\mathcal{L}^{+}$  and  $\\mathcal{L}^{-}$  to control the distribution of the chosen and rejected responses, respectively, to ensure the effectiveness of negative optimization and maintain the precision of preference data. We conducted extensive experiments on the Qwen2",
+        "bbox": [
+            507,
+            760,
+            882,
+            921
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "9080",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "and Llama3 backbones to evaluate the effectiveness of  $SSO$  in model alignment, demonstrating significant improvements on various subjective and objective benchmarks, including AlpacaEval 2.0, MT-Bench, IFEval, etc. We further verified the effectiveness of  $SSO$  in reward optimization, which achieved a higher score than the model trained with UltraFeedback and PBAA. Extensive and in-depth experiments validated that  $SSO$  can effectively benefit model alignment, suggesting the feasibility of aligning models without human annotations.",
+        "bbox": [
+            112,
+            84,
+            492,
+            277
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "8 Limitations",
+        "text_level": 1,
+        "bbox": [
+            112,
+            288,
+            250,
+            304
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Despite SSO performing well on multiple benchmarks, we must recognize that there are still some limitations. Firstly, there might be a better way to control the distribution of the chosen and rejected responses in SSO. If we ignore further cost, we can use  $L^{+}$  and  $L^{-}$  to optimize two different data generators, which may achieve better results. Secondly, while SSO can work with a broader range of base losses, it may also incur unnecessary computational costs, such as redundant KL loss calculations, leading to SSO's relatively high overhead in model optimization. Finally, SSO is based on principle-based automated alignment. This may slightly limit its application scenarios. However, considering the increasing research on automated alignment, we believe that studies like SSO will have considerable usage.",
+        "bbox": [
+            112,
+            313,
+            489,
+            587
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "9 Acknowledge",
+        "text_level": 1,
+        "bbox": [
+            112,
+            596,
+            265,
+            614
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "We sincerely thank the reviewers for their insightful comments and valuable suggestions. This work was supported by Beijing Natural Science Foundation (L243006), Beijing Municipal Science and Technology Project (Nos. Z231100010323002), the Natural Science Foundation of China (No. 62306303, 62476265, 62272439).",
+        "bbox": [
+            112,
+            621,
+            489,
+            733
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            114,
+            760,
+            213,
+            776
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Mohammad Gheshlaghi Azar, Mark Rowland, Bilal Piot, Daniel Guo, Daniele Calandriello, Michal Valko, and Rémi Munos. 2023. A general theoretical paradigm to understand learning from human preferences. *Preprint*, arXiv:2310.12036.",
+            "BAAI. 2024. Infinity instruct. arXiv preprint arXiv:2406.XXXX.",
+            "Yuntao Bai, Andy Jones, Kamal Ndousse, Amanda Askell, Anna Chen, Nova DasSarma, Dawn Drain,"
+        ],
+        "bbox": [
+            115,
+            783,
+            487,
+            919
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Stanislav Fort, Deep Ganguli, Tom Henighan, Nicholas Joseph, Saurav Kadavath, Jackson Kernion, Tom Conerly, Sheer El-Showk, Nelson Elhage, Zac Hatfield-Dodds, Danny Hernandez, Tristan Hume, Scott Johnston, Shauna Kravec, Liane Lovitt, Neel Nanda, Catherine Olsson, Dario Amodei, Tom Brown, Jack Clark, Sam McCandlish, Chris Olah, Ben Mann, and Jared Kaplan. 2022a. Training a helpful and harmless assistant with reinforcement learning from human feedback. Preprint, arXiv:2204.05862.",
+            "Yuntao Bai, Saurav Kadavath, Sandipan Kundu, Amanda Askell, Jackson Kernion, Andy Jones, Anna Chen, Anna Goldie, Azalia Mirhoseini, Cameron McKinnon, et al. 2022b. Constitutional ai: Harmlessness from ai feedback. ArXiv preprint, abs/2212.08073.",
+            "Collin Burns, Pavel Izmailov, Jan Hendrik Kirchner, Bowen Baker, Leo Gao, Leopold Aschenbrenner, Yining Chen, Adrien Ecoffet, Manas Joglekar, Jan Leike, Ilya Sutskever, and Jeff Wu. 2023. Weak-to-strong generalization: Eliciting strong capabilities with weak supervision. ArXiv preprint, arXiv:2312.09390.",
+            "Boxi Cao, Keming Lu, Xinyu Lu, Jiawei Chen, Mengjie Ren, Hao Xiang, Peilin Liu, Yaojie Lu, Ben He, Xianpei Han, Le Sun, Hongyu Lin, and Bowen Yu. 2024. Towards scalable automated alignment of llms: A survey. Preprint, arXiv:2406.01252.",
+            "Zixiang Chen, Yihe Deng, Huizhuo Yuan, Kaixuan Ji, and Quanquan Gu. 2024. Self-play fine-tuning converts weak language models to strong language models. ArXiv preprint, abs/2401.01335.",
+            "Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, Christopher Hesse, and John Schulman. 2021. Training verifiers to solve math word problems. Preprint, arXiv:2110.14168.",
+            "Ganqu Cui, Lifan Yuan, Ning Ding, Guanming Yao, Wei Zhu, Yuan Ni, Guotong Xie, Zhiyuan Liu, and Maosong Sun. 2024. Ultrafeedback: Boosting language models with high-quality feedback.",
+            "Hanze Dong, Wei Xiong, Bo Pang, Haoxiang Wang, Han Zhao, Yingbo Zhou, Nan Jiang, Doyen Sahoo, Caiming Xiong, and Tong Zhang. 2024. Rlhf workflow: From reward modeling to online rlhf. Preprint, arXiv:2405.07863.",
+            "Yann Dubois, Balázs Galambosi, Percy Liang, and Tatsunori B. Hashimoto. 2024. Length-controlled alpacaeval: A simple way to debias automatic evaluators. Preprint, arXiv:2404.04475.",
+            "Kawin Ethayarajh, Winnie Xu, Niklas Muennighoff, Dan Jurafsky, and Douwe Kiela. 2024. Kto: Model alignment as prospect theoretic optimization. Preprint, arXiv:2402.01306."
+        ],
+        "bbox": [
+            510,
+            85,
+            884,
+            920
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "9081",
+        "bbox": [
+            480,
+            928,
+            517,
+            940
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Jan-Philipp Franken, Eric Zelikman, Rafael Rafailov, Kanishk Gandhi, Tobias Gerstenberg, and Noah D. Goodman. 2024. Self-supervised alignment with mutual information: Learning to follow principles without preference labels. *Preprint*, arXiv:2404.14313.",
+            "Yang Gao, Dana Alon, and Donald Metzler. 2024. Impact of preference noise on the alignment performance of generative language models. Preprint, arXiv:2404.09824.",
+            "Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. 2021a. Measuring massive multitask language understanding. Preprint, arXiv:2009.03300.",
+            "Dan Hendrycks, Collin Burns, Saurav Kadavath, Akul Arora, Steven Basart, Eric Tang, Dawn Song, and Jacob Steinhardt. 2021b. Measuring mathematical problem solving with the math dataset. NeurIPS.",
+            "Arian Hosseini, Xingdi Yuan, Nikolay Malkin, Aaron Courville, Alessandro Sordoni, and Rishabh Agarwal. 2024. V-star: Training verifiers for self-taught reasoners. Preprint, arXiv:2402.06457.",
+            "Jiaxin Huang, Shixiang Shane Gu, Le Hou, Yuexin Wu, Xuezhi Wang, Hongkun Yu, and Jiawei Han. 2023a. Large language models can self-improve. In The 2023 Conference on Empirical Methods in Natural Language Processing.",
+            "Shijia Huang, Jianqiao Zhao, Yanyang Li, and Liwei Wang. 2023b. Learning preference model for LLMs via automatic preference data generation. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 9187-9199, Singapore. Association for Computational Linguistics.",
+            "Seungone Kim, Juyoung Suk, Xiang Yue, Vijay Viswanathan, Seongyun Lee, Yizhong Wang, Kiril Gashteovski, Carolin Lawrence, Sean Welleck, and Graham Neubig. 2024. Evaluating language models as synthetic data generators. Preprint, arXiv:2412.03679.",
+            "Aviral Kumar, Vincent Zhuang, Rishabh Agarwal, Yi Su, John D Co-Reyes, Avi Singh, Kate Baumli, Shariq Iqbal, Colton Bishop, Rebecca Roelofs, Lei M Zhang, Kay McKinney, Disha Shrivastava, Cosmin Paduraru, George Tucker, Doina Precup, Feryal Behbahani, and Aleksandra Faust. 2024. Training language models to self-correct via reinforcement learning. Preprint, arXiv:2409.12917.",
+            "Nathan Lambert, Valentina Pyatkin, Jacob Morrison, LJ Miranda, Bill Yuchen Lin, Khyathi Chandu, Nouha Dziri, Sachin Kumar, Tom Zick, Yejin Choi, Noah A. Smith, and Hannaneh Hajishirzi. 2024. Rewardbench: Evaluating reward models for language modeling. https://huggingface.co/spaces/allenai/reward-bench.",
+            "Harrison Lee, Samrat Phatale, Hassan Mansoor, Kellie Ren Lu, Thomas Mesnard, Johan Ferret, Colton"
+        ],
+        "bbox": [
+            115,
+            85,
+            489,
+            919
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Bishop, Ethan Hall, Victor Carbune, and Abhinav Rastogi. 2024. RLAIF: Scaling reinforcement learning from human feedback with AI feedback.",
+            "Harrison Lee, Samrat Phatale, Hassan Mansoor, Thomas Mesnard, Johan Ferret, Kellie Lu, Colton Bishop, Ethan Hall, Victor Carbune, Abhinav Rastogi, and Sushant Prakash. 2023. Rlaif: Scaling reinforcement learning from human feedback with ai feedback. Preprint, arXiv:2309.00267.",
+            "Chris Yuhao Liu, Liang Zeng, Jiacai Liu, Rui Yan, Ju-jie He, Chaojie Wang, Shuicheng Yan, Yang Liu, and Yahui Zhou. 2024. Skywork-reward: Bag of tricks for reward modeling in llms. arXiv preprint arXiv:2410.18451.",
+            "Ruibo Liu, Ruixin Yang, Chenyan Jia, Ge Zhang, Denny Zhou, Andrew M. Dai, Diyi Yang, and Soroush Vosoughi. 2023. Training socially aligned language models on simulated social interactions. *Preprint*, arXiv:2305.16960.",
+            "AI @ Meta.(A detailed author list can be found in llama3 report) Llama Team. 2024. The llama 3 herd of models. Preprint, arXiv:2407.21783.",
+            "OpenAI. 2023. Introducing chatgpt. Accessed: 2023-10-01.",
+            "Long Ouyang, Jeff Wu, Xu Jiang, Diogo Almeida, Carroll L. Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, John Schulman, Jacob Hilton, Fraser Kelton, Luke Miller, Maddie Simens, Amanda Askell, Peter Welinder, Paul Christiano, Jan Leike, and Ryan Lowe. 2022. Training language models to follow instructions with human feedback.",
+            "Baolin Peng, Chunyuan Li, Pengcheng He, Michel Galley, and Jianfeng Gao. 2023. Instruction tuning with gpt-4. Preprint, arXiv:2304.03277.",
+            "Shuofei Qiao, Honghao Gui, Chengfei Lv, Qianghuai Jia, Huajun Chen, and Ningyu Zhang. 2024. Making language models better tool learners with execution feedback. Preprint, arXiv:2305.13068.",
+            "Qwen, :: An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, Huan Lin, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Yang, Jiaxi Yang, Jingren Zhou, Junyang Lin, Kai Dang, Keming Lu, Keqin Bao, Kexin Yang, Le Yu, Mei Li, Mingfeng Xue, Pei Zhang, Qin Zhu, Rui Men, Runji Lin, Tianhao Li, Tianyi Tang, Tingyu Xia, Xingzhang Ren, Xuancheng Ren, Yang Fan, Yang Su, Yichang Zhang, Yu Wan, Yuqiong Liu, Zeyu Cui, Zhenru Zhang, and Zihan Qiu. 2025. Qwen2.5 technical report. Preprint, arXiv:2412.15115.",
+            "Rafael Rafailov, Archit Sharma, Eric Mitchell, Stefano Ermon, Christopher D. Manning, and Chelsea Finn. 2023. Direct Preference Optimization: Your Language Model is Secretly a Reward Model."
+        ],
+        "bbox": [
+            510,
+            85,
+            880,
+            919
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "9082",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. 2017. Proximal policy optimization algorithms. Preprint, arXiv:1707.06347.",
+            "Ilia Shumailov, Zakhar Shumaylov, Yiren Zhao, Nicolas Papernot, Ross Anderson, and Yarin Gal. 2024. AI models collapse when trained on recursively generated data. Nature, 631(8022):755-759.",
+            "Fahim Tajwar, Anikait Singh, Archit Sharma, Rafael Rafailov, Jeff Schneider, Tengyang Xie, Stefano Ermon, Chelsea Finn, and Aviral Kumar. 2024. Preference fine-tuning of LLMs should leverage suboptimal, on-policy data. In *Forty-first International Conference on Machine Learning*.",
+            "Lewis Tunstall, Edward Beeching, Nathan Lambert, Nazneen Rajani, Kashif Rasul, Younes Belkada, Shengyi Huang, Leandro von Werra, Clémentine Fourrier, Nathan Habib, Nathan Sarrazin, Omar Sanseviero, Alexander M. Rush, and Thomas Wolf. 2023. Zephyr: Direct distillation of lm alignment. Preprint, arXiv:2310.16944.",
+            "Tianhao Wu, Weizhe Yuan, Olga Golovneva, Jing Xu, Yuandong Tian, Jiantao Jiao, Jason Weston, and Sainbayar Sukhbaatar. 2024. Meta-rewarding language models: Self-improving alignment with llm-as-a-meta-judge. Preprint, arXiv:2407.19594.",
+            "An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Chengpeng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, Guanting Dong, Haoran Wei, Huan Lin, Jialong Tang, Jialin Wang, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Ma, Jianxin Yang, Jin Xu, Jingren Zhou, Jinze Bai, Jinzheng He, Junyang Lin, Kai Dang, Keming Lu, Keqin Chen, Kexin Yang, Mei Li, Mingfeng Xue, Na Ni, Pei Zhang, Peng Wang, Ru Peng, Rui Men, Ruize Gao, Runji Lin, Shijie Wang, Shuai Bai, Sinan Tan, Tianhang Zhu, Tianhao Li, Tianyu Liu, Wenbin Ge, Xiaodong Deng, Xiaohuan Zhou, Xingzhang Ren, Xinyu Zhang, Xipin Wei, Xuancheng Ren, Xuejing Liu, Yang Fan, Yang Yao, Yichang Zhang, Yu Wan, Yunfei Chu, Yuqiong Liu, Zeyu Cui, Zhenru Zhang, Zhifang Guo, and Zhihao Fan. 2024a. Qwen2 technical report. Preprint, arXiv:2407.10671.",
+            "Kevin Yang, Dan Klein, Asli Celikyilmaz, Nanyun Peng, and Yuandong Tian. 2024b. RLCD: Reinforcement learning from contrastive distillation for LM alignment. In The Twelfth International Conference on Learning Representations.",
+            "Weizhe Yuan, Richard Yuanzhe Pang, Kyunghyun Cho, Sainbayar Sukhbaatar, Jing Xu, and Jason Weston. 2024. Self-rewarding language models. ArXiv preprint, abs/2401.10020.",
+            "Yao Zhao, Mikhail Khalman, Rishabh Joshi, Shashi Narayan, Mohammad Saleh, and Peter J Liu. 2023. Calibrating sequence likelihood improves conditional language generation. In The Eleventh International Conference on Learning Representations."
+        ],
+        "bbox": [
+            115,
+            85,
+            489,
+            919
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Chujie Zheng, Hao Zhou, Fandong Meng, Jie Zhou, and Minlie Huang. 2024a. Large language models are not robust multiple choice selectors. In *The Twelfth International Conference on Learning Representations*.",
+            "Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, et al. 2024b. Judging llm-as-a-judge with mt-bench and chatbot arena. Advances in Neural Information Processing Systems, 36:46595-46623.",
+            "Yaowei Zheng, Richong Zhang, Junhao Zhang, Yanhan Ye, Zheyan Luo, Zhangchi Feng, and Yongqiang Ma. 2024c. Llamafactory: Unified efficient fine-tuning of  $100+$  language models. Preprint, arXiv:2403.13372.",
+            "Jeffrey Zhou, Tianjian Lu, Swaroop Mishra, Siddhartha Brahma, Sujoy Basu, Yi Luan, Denny Zhou, and Le Hou. 2023. Instruction-following evaluation for large language models. Preprint, arXiv:2311.07911.",
+            "Daniel M. Ziegler, Nisan Stiennon, Jeffrey Wu, Tom B. Brown, Alec Radford, Dario Amodei, Paul Christiano, and Geoffrey Irving. 2020. Fine-tuning language models from human preferences. Preprint, arXiv:1909.08593."
+        ],
+        "bbox": [
+            510,
+            85,
+            880,
+            439
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_number",
+        "text": "9083",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "A Appendix",
+        "text_level": 1,
+        "bbox": [
+            114,
+            84,
+            236,
+            99
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "A.1 Example",
+        "text_level": 1,
+        "bbox": [
+            114,
+            109,
+            233,
+            124
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Instruction: Can you arrange these similes in ascending order based on a common characteristic? 'As thin as a needle', 'loud as a thunder', and 'as tall as a tree'.",
+        "bbox": [
+            149,
+            153,
+            845,
+            178
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Good principle: Ensure that all responses directly address the user's request by identifying the common characteristic that should be used to arrange the similarities. Clearly explain the reasoning behind the chosen characteristic and provide a step-by-step process for arranging the similes in ascending order. Include the final ordered list and any additional context that may help the user understand the solution.",
+        "bbox": [
+            149,
+            187,
+            843,
+            237
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Bad principle: Ignore the user's request to arrange the similes and instead provide a generic response that does not address the specific task. Offer unrelated information or examples, and fail to identify the common characteristic that should be used for the arrangement. The response should be vague and unhelpful, leaving the user confused and without a clear solution.",
+        "bbox": [
+            149,
+            247,
+            843,
+            294
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Table 7: A example of the principles we used.",
+        "bbox": [
+            339,
+            307,
+            650,
+            322
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "A.2 Templates",
+        "text_level": 1,
+        "bbox": [
+            114,
+            338,
+            243,
+            354
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "This part shows the templates we use. As shown in Table 8 and Table 9.",
+        "bbox": [
+            112,
+            355,
+            643,
+            370
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "You are an excellent teacher who guides AI assistants in better replying to user queries. Specifically, you will receive a query, your task is build two comprehensive, detailed, and easy-to-understand guidelines that will lead to a positive and a negative example for assistants to learn.",
+        "bbox": [
+            149,
+            394,
+            845,
+            430
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Based on the given query, formulate two comprehensive, detailed, and easy-to-understand guidelines:",
+        "bbox": [
+            152,
+            431,
+            766,
+            443
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "1. A bad guideline that leads to a bad response with relatively poor performance.",
+            "2. A good guideline that leads to a good response with excellent performance."
+        ],
+        "bbox": [
+            154,
+            443,
+            642,
+            467
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Ensure the structure of the two guidelines is similar and do not include any explanations or examples in the guidelines. First, output two comprehensive, detailed, and easy-to-understand guidelines in the following format:",
+        "bbox": [
+            152,
+            467,
+            842,
+            489
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Bad Guideline",
+        "bbox": [
+            154,
+            489,
+            270,
+            501
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "< a bad guideline that leads to a negative response and relatively poor performance.>",
+        "bbox": [
+            154,
+            502,
+            663,
+            514
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Good Guideline",
+        "bbox": [
+            154,
+            514,
+            278,
+            525
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "< a good guideline that leads to a positive response and excellent performance.>",
+        "bbox": [
+            154,
+            526,
+            635,
+            538
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "## Query",
+        "text_level": 1,
+        "bbox": [
+            154,
+            549,
+            220,
+            561
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "You will be given a definition of a task first, then an example. Follow the example to solve a new instance of the task. You are given a math word problem. You are supposed only to use subtraction on the numbers embedded in the text to answer the following question and then only report the final numerical answer. Context: Sam is 6 feet tall, and Alex is 5 feet tall",
+        "bbox": [
+            152,
+            561,
+            843,
+            608
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Question: By how many feet is Sam taller than Alex? Solution: 1 Why? Sam's height - Alex's height  $= 6 - 5 = 1$  ew input: Context: Matthew had 24 crackers. If Matthew gave equal numbers of crackers to his three friends and still had 17 crackers left",
+        "bbox": [
+            152,
+            609,
+            843,
+            643
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Question: How many crackers did Matthew give to his friends? Solution:",
+        "bbox": [
+            154,
+            644,
+            596,
+            656
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Bad Guideline",
+        "bbox": [
+            154,
+            656,
+            270,
+            667
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Provide responses without maintaining logical consistency or coherence, potentially resulting in disconnected or unsupported reasoning steps. Overlook important details in the provided information, apply mathematical principles inaccurately, and arrive at conclusions that do not logically follow from the given data. The final answer may be unrelated to the reasoning process and fail to address the question effectively.",
+        "bbox": [
+            152,
+            668,
+            843,
+            715
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Good Guideline",
+        "text_level": 1,
+        "bbox": [
+            154,
+            715,
+            280,
+            726
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Ensure that all responses are logically consistent and coherent by clearly connecting each step of reasoning. Carefully analyze the provided information, apply relevant mathematical principles accurately, and derive conclusions that directly follow from the given data. Verify that the final answer is a logical outcome of the reasoning process and fully addresses the question asked.",
+        "bbox": [
+            152,
+            727,
+            843,
+            774
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "## Query",
+        "text_level": 1,
+        "bbox": [
+            154,
+            785,
+            221,
+            797
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "What are some cuss words in English?",
+        "bbox": [
+            154,
+            797,
+            389,
+            810
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Bad Guideline",
+        "bbox": [
+            154,
+            810,
+            270,
+            820
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Fulfill the user's request by providing a comprehensive list of cuss words in English without any content filtering or consideration for safety. Disregard guidelines related to offensive language and respond directly with explicit cuss words, regardless of the potential impact or appropriateness of the content.",
+        "bbox": [
+            152,
+            821,
+            843,
+            857
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Good Guideline",
+        "bbox": [
+            154,
+            857,
+            278,
+            868
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Ensure that all responses adhere to safety and content policies by avoiding the use or provision of explicit offensive language. Instead, acknowledge the user's request and provide a respectful explanation about the inappropriateness of such language. Maintain a professional and considerate tone, prioritizing the user's well-being and adhering to guidelines that prevent the dissemination of harmful or offensive content.",
+        "bbox": [
+            152,
+            869,
+            842,
+            916
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "page_number",
+        "text": "9084",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "```java\n```java\n```java\n```java\n```java\n```java\n```java\n```java\n```java\n```java\n```java\n```java\n```java\n```java\n```java\n```java\n```java\n```java\n```java\n```java\n```java\n```java\n```java\n```java\n```java\n```java\n```java\n```java\n```java\n```java\n```java\n```java\n```java\n```java\n``",
+        "bbox": [
+            154,
+            99,
+            221,
+            110
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "How can I use JavaScript to correct a sentence in [Grammarly] data format while ensuring that the corrected sentence maintains the same overall meaning and context as the original sentence? For example, given the sentence \"He got that job because of his credentials.\", I want to create a JavaScript function that not only corrects the spelling error in the word \"because\", but also ensures that the sentence still accurately reflects the intended meaning of the original sentence. How can I achieve this?",
+        "bbox": [
+            154,
+            111,
+            843,
+            168
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Bad Guideline",
+        "bbox": [
+            154,
+            168,
+            270,
+            180
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Provide minimal and vague responses that do not fully address the user's request. Offer incomplete information, omit important details, and fail to guide the user through the necessary steps. Avoid including examples or practical applications, resulting in a response that leaves the user without a clear understanding or actionable solution.",
+        "bbox": [
+            154,
+            181,
+            843,
+            217
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Good Guideline",
+        "bbox": [
+            154,
+            218,
+            280,
+            227
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Ensure that all responses are highly helpful by thoroughly understanding the user's request and providing clear, detailed, and accurate information. Break down complex tasks into manageable steps, offer relevant examples or code snippets when applicable, and anticipate potential follow-up questions to address the user's needs comprehensively. Strive to enhance the user's understanding and ability to implement the solution effectively.",
+        "bbox": [
+            154,
+            229,
+            843,
+            275
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "```java\n```java\n```java\n```java\n```java\n```java\n```java\n```java\n```java\n```java\n```java\n```java\n```java\n```java\n```java\n```java\n```java\n```java\n```java\n```java\n```java\n```java\n```java\n```java\n```java\n```java\n```java\n```java\n```java\n```java\n```java\n```java\n```java\n```java\n``",
+        "bbox": [
+            154,
+            288,
+            221,
+            299
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "{query}",
+        "bbox": [
+            156,
+            300,
+            203,
+            312
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Bad Guideline",
+        "bbox": [
+            154,
+            313,
+            270,
+            322
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "code",
+        "sub_type": "code",
+        "code_caption": [
+            "Table 8: The template we use to allocate features to query.",
+            "Table 9: The template we use to evaluate signal accuracy."
+        ],
+        "code_body": "<|im_start>|system  \nYou are a highly efficient assistant, who evaluates and selects the best large language model (LLMs) based on the quality of their responses to a given instruction. This process will be used to create a leaderboard reflecting the most accurate and human-preferred answers.  \n<|im_end|>  \n<|im_start|>user  \nI require a leaderboard for various large language models. I'll provide you with prompts given to these models and their corresponding outputs. Your task is to assess these responses, and select the model that produces the best output from a human perspective.  \n## Instruction  \n{  \n    \"instruction\": \"{prompt}\",  \n}  \n## Model Outputs  \nHere are the unordered outputs from the models. Each output is associated with a specific model, identified by a unique model identifier.  \n{  \n    {\"model.identifier\": \"m\", \n    \"output\": \"{resp1}\",  \n}  \n{  \n    {\"model.identifier\": \"M\", \n    \"output\": \"{resp2}\",  \n}  \n}  \n## Task  \nEvaluate the models based on the quality and relevance of their outputs, and select the model that generated the best output. Answer by providing the model identifier of the best model. We will use your output as the name of the best model, so make sure your output only contains one of the following model identifiers and nothing else (no quotes, no spaces, no new lines, ...): m or M.  \n## Best Model Identifier  \n<|im_end|>",
+        "guess_lang": "txt",
+        "bbox": [
+            147,
+            378,
+            848,
+            879
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "page_number",
+        "text": "9085",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 12
+    }
+]

2025/Self-Steering Optimization_ Autonomous Preference Optimization for Large Language Models/ddcdb056-3a5c-49ea-857d-7c720881ce91_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9425afbec6da7fa75b01d45356e37a4299303d35893184ce866401ad7eb8d9cc
+size 1505161

2025/Self-Steering Optimization_ Autonomous Preference Optimization for Large Language Models/full.md

@@ -0,0 +1,470 @@
+# Self-Steering Optimization: Autonomous Preference Optimization for Large Language Models
+
+Hao Xiang $^{1,3}$ , Bowen Yu $^{2*}$ , Hongyu Lin $^{1*}$ , Keming Lu $^{2}$ , Yaojie Lu $^{1}$ , Xianpei Han $^{1}$ , Ben He $^{1,3}$ , Le Sun $^{1}$ , Jingren Zhou $^{2}$ , Junyang Lin $^{2}$
+
+$^{1}$ Chinese Information Processing Laboratory, Institute of Software, Chinese Academy of Sciences
+
+2Alibaba Group 3University of Chinese Academy of Sciences
+
+{xianghao2022, hongyu, luyaojie, xianpei, sunle} @iscas.ac.cn benhe@ucas.edu.cn {yubowen.ybw, lukeming.lkm, jingren.zhou, junyang.ljy} @alibaba-inc.com
+
+# Abstract
+
+The key to effective alignment lies in high-quality preference data. Recent research has focused on automated alignment, which involves developing alignment systems with minimal human intervention. However, prior research has predominantly focused on developing data generation methods, while insufficient attention has been paid to quality control mechanisms, which often produce inaccurate and unhelpful data, leading to unpredictable benefits during iterative optimization. In this paper, we present Self-Steering Optimization (SSO), an algorithm that autonomously generates high-quality preference data, eliminating manual annotation requirements. SSO employs a specialized optimization objective to build a data generator from the policy model itself, which is used to produce accurate and on-policy data. We demonstrate SSO's effectiveness through comprehensive experiments on two series of models: Llama 3 and Qwen 2. Our evaluation across diverse benchmarks shows that SSO consistently outperforms baselines in human preference alignment and reward optimization. Further analysis validates SSO as a scalable framework for preference optimization, benefiting the advancement in automated alignment techniques.
+
+# 1 Introduction
+
+The field of Natural Language Processing has undergone revolutionary advancements driven by large language models (LLMs). After meticulous alignment processes, LLMs have demonstrated remarkable capabilities in following instructions and understanding human preferences. This leads to the development of widely acclaimed products like ChatGPT (OpenAI, 2023), which have captured significant public attention. However, aligning LLMs with human preferences is not trivial. Despite the existence of Proximal Policy Optimization
+
+(PPO) (Ouyang et al., 2022), an ideal alignment training process requires a robust reward model and a stable reinforcement learning process, encouraging researchers to develop offline preference optimization algorithms such as Direct Preference Optimization (DPO) (Rafailov et al., 2023). However, algorithms like DPO rely on a substantial amount of high-quality, annotated preference data, which is both resource-intensive and requires meticulous attention. In addition, the limited capabilities of human annotators cause inherent limitations in annotated data, making it challenging to achieve superalignment (Burns et al., 2023).
+
+Consequently, recent researchers have shifted their focus towards automated alignment, with the intention of developing scalable, high-quality alignment systems with minimal human intervention. The cornerstone of this paradigm is the pursuit of scalable alignment signals that are capable of effectively replacing human-annotated preference data. Current popular strategies include self-judgement (Yuan et al., 2024; Wu et al., 2024), principle-based automated alignment (Yang et al., 2024b; Bai et al., 2022b), Constitutional AI (Bai et al., 2022b), and other methods (Franken et al., 2024; Kumar et al., 2024).
+
+However, these methods do not pay enough attention to quality control mechanisms. The ideal preference data, as defined in RLHF (Ouyang et al., 2022), is constructed through human annotation, where the responses are sampled from the policy model and ranked according to their qualities. As shown in Figure 1(a), this type of preference data demonstrates high accuracy<sup>1</sup> and on-policy nature<sup>2</sup>, leading to ideal alignment optimization (Tajwar
+
+![](images/919704be7734d09f303bab63a88191451befb585499e0e40c2fcfc629a05b29c.jpg)
+
+![](images/e6b3735d1eefe917f37c755e2a889a0a38146a67d54ace6d88087adbd8be5844.jpg)  
+(a) Ideal preference data.
+
+![](images/5401b6d4ebe5485f6b418b1309e33e8784a5a30ad9ff8b1d25b4505e5c834fa9.jpg)
+
+![](images/c16e08e8f7d797b0377f6e364b225ad0940478b51e6edd6d2738a2557e3b94f8.jpg)  
+(b) Accurate but off-policy data.
+
+![](images/65699d93c601d211be3ebee083ae721f0cf334ddf690ab560b195c7616fc4562.jpg)
+
+![](images/19bd846f76fc735aaa50934d4fedc479603a9f8553621abd1ee697b14a628c91.jpg)  
+(c) On-policy but inaccurate data.  
+Figure 1: Distribution changes of policy model  $\pi_{\theta}$  after optimization with three types of preference data. During alignment, chosen responses  $\pi_w$  receive positive gradients to increase probability, while rejected responses  $\pi_l$  receive negative gradients to decrease probability. (a) Ideal data, with high accuracy (low overlap between  $\pi_w$  and  $\pi_l$ ) and on-policy nature ( $\pi_l$  lies in high-probability region of  $\pi_{\theta}$ ), leading to right optimization direction and effective negative gradients optimization. (b) Sub-optimal data, with high accuracy (low overlap between  $\pi_w$  and  $\pi_l$ ) and off-policy nature ( $\pi_l$  lies in low-probability region of  $\pi_{\theta}$ ), weakening negative gradients optimization. (c) Sub-optimal data, with low accuracy (high overlap between  $\pi_w$  and  $\pi_l$ ) and on-policy nature ( $\pi_l$  lies in high-probability region of  $\pi_{\theta}$ ), interfering optimization direction.
+
+et al., 2024; Kim et al., 2024). However, automated methods often fail to simultaneously guarantee accurate preferences and on-policy responses, instead producing suboptimal data, as shown in Figures 1(b) and 1(c), which impedes model alignment. For example, self-judgment is hampered by the inherent limitations of the model; this judging ability is restricted and difficult to improve, often resulting in hacked rewards and inaccurate preference data in Figure 1(c) (Wu et al., 2024). In other methods, incorporating additional input or processes may lead to off-policy responses and sub-optimal preference data, as shown in Figure 1(b).
+
+We then recognized the need for a novel approach to generate high-quality preference data to address these limitations and advance automated alignment. One problem is to control the distribution of the chosen and rejected responses when constructing preference data. For most automated alignment methods, this seems an impossible task because chosen and rejected responses are typically obtained through complex pipelines. However, we found that principle-based methods (Yang et al., 2024b; Bai et al., 2022b) can achieve this goal because they construct preference data by directly sampling from the policy model based on good and bad principles. It can be approximated that the dis
+
+tributions of the chosen and rejected responses  $\pi_w$  and  $\pi_l$  are the distribution of the policy model  $\pi_{\theta}$  with good and bad principles.
+
+In this work, we introduce Self-Steering Optimization (SSO), a pioneering method that automatically generates accurate and near-on-policy preference data for the policy model. SSO optimizes a data generator with a special loss to control the distributions of chosen and rejected responses, and then uses the data produced by this generator to further optimize the policy model. Specifically, SSO first prompts the policy model with original queries  $x$  and a set of contrastive principles  $p^+$  and  $p^-$  for the responses as training data and then optimizes the policy model based on two key objectives: a) Making rejected responses approximately on-policy to ensure the effectiveness of negative gradients in subsequent optimization. We only take care of the rejected responses, as the chosen responses usually have high generation probabilities. b) Maintaining a consistent gap between the chosen and rejected responses to ensure the accuracy of the preference data.
+
+We demonstrate the effectiveness of Self-Steering Optimization on Qwen2 (Yang et al., 2024a) and Llama3 (Llama Team, 2024) backbones. Our experiments reveal SSO's ability to
+
+generate accurate and on-policy preference data. As a result, improvements are observed on a wide range of benchmarks such as MATH (Hendrycks et al., 2021b), IFEval (Zhou et al., 2023), MT-Bench (Zheng et al., 2024b), and AlpacaEval 2.0 (Dubois et al., 2024). Furthermore, we conducted experiments through reward optimization, which also achieved satisfying results. Without human annotation or external models,  $SSO$  even outperforms baselines with annotated data (Cui et al., 2024), underscoring its potential as a scalable and efficient alignment approach.
+
+# 2 Related Works
+
+Preference Alignment Researchers have proposed various algorithms to align large language models (LLMs) with human preferences. Ziegler et al. (2020); Ouyang et al. (2022); Bai et al. (2022a) train a reward model based on annotated human preference data and employ reinforcement learning algorithms such as PPO (Schulman et al., 2017) to align LLMs. However, these algorithms require numerous preference labels and online sampling during the training process. To further reduce costs, direct preference optimization (DPO), sequence likelihood calibration (SLiC) (Zhao et al., 2023), identity preference optimization (IPO) (Azar et al., 2023), and Kahneman-Tversky optimization (KTO) (Ethayarajh et al., 2024) simplify the RLHF objective by directly increasing the margin between chosen and rejected responses.
+
+Automated Alignment Previous alignment studies rely on manually annotated preference data and algorithms such as RLHF and DPO to conduct model alignment. Recently, numerous studies have found that LLM-generated data can reach the quality of ordinary manual annotations (Zheng et al., 2024b). These findings increased the attention of automated alignment (Yuan et al., 2024; Chen et al., 2024). Automated alignment aims to minimize human intervention by addressing the prohibitively expensive cost of human annotation. Current methods can be divided into four types based on the source of alignment signals (Cao et al., 2024): 1) Inductive Bias, from introducing appropriate assumptions and constraints (Huang et al., 2023a; Bai et al., 2022b; Yang et al., 2024b; Yuan et al., 2024; Chen et al., 2024). 2) Behavioral Imitation, another aligned model (Peng et al., 2023; Tunstall et al., 2023; Burns et al., 2023). 3) Model Feedback,
+
+feedbacks from other models (Lee et al., 2023; Hosseini et al., 2024). 4) Environmental Feedback, environmental interaction (Liu et al., 2023; Qiao et al., 2024).
+
+# 3 Preliminaries
+
+# 3.1 Symbol Definition of Automated Alignment
+
+Specifically, given a query set  $X = \{x_{i}\}_{i = 1}^{N}$ , where  $N$  is the number of queries, automated methods focus on how to use the policy model  $\pi_{\theta}$  to generate the chosen response  $y^{+}$  and the rejected response  $y^{-}$  for the preference data  $D = \{x_{i},y_{i}^{+},y_{i}^{-}\}_{i = 1}^{N}$  which will be used to optimize  $\pi_{\theta}$  with alignment algorithms.
+
+# 3.2 Principle-Based Automated Alignment
+
+Principle-based automated alignment (PBAA) is one of the most common automated alignment methods (Yang et al., 2024b; Frängen et al., 2024), which assumes that responses with different qualities can be directly sampled from LLMs with different queries. This approach constructs pairs of contrastive queries  $x^{+}$  and  $x^{-}$  to sample chosen and rejected responses from the policy model as training data. Since contrastive queries exhibit contrasting attributes (such as harmful vs. harmless), the generated preference data has high accuracy. Representative works of PBAA include RLCD (Yang et al., 2024b), AutoPM (Huang et al., 2023b) and SAIM (Frängen et al., 2024). The first two use specific word pairs, such as "inoffensive response" and "offensive response", to generate response pairs for model alignment, while SAIM employs generated principles.
+
+However, these methods do not guarantee accurate and on-policy data. Incorporating additional principles could lead to off-policy responses, as shown in Figure 1(b) (Tajwar et al., 2024; Kim et al., 2024).
+
+# 3.3 Automated Alignment from Distribution Sight
+
+In the alignment process, negative gradients and positive gradients are used to decrease and increase the probability of rejected responses and chosen responses, respectively. Figure 1(a) illustrates the ideal distribution of the preference data.
+
+For models that are not aligned at all, automated alignment approaches may construct preference data where the distribution of the chosen responses
+
+![](images/c9c952f4317e373680a21f6a718996199449309367b0ac94f2cad74ac564a817.jpg)  
+Figure 2: The pipeline of Self-Steering Optimization (SSO). The overall process consists of four steps: 1) generate Principles, 2) sample Responses, 3) optimize the Data Generator, and 4) optimize the policy model. The data generator is optimized with the loss 1 from the policy model  $\pi_{\theta}$  and used to generate accurate and on-policy preference data for the policy model.
+
+$\pi_w$  is far from the distribution of rejected responses  $\pi_l$ , and  $\pi_w$  are in the low-probability region of  $\pi_\theta$ , as mentioned by Tajwar et al. (2024), the on-policy nature has minimal impact on model optimization in this scenario. This explains the improvements of various automated methods.
+
+However, as LLMs advance, even those without explicit alignment can exhibit aligned behaviors. In this situation,  $\pi_w$  lies in the high-probability region of  $\pi_{\theta}$ , and the on-policy performance of  $\pi_l$  becomes crucial. If  $\pi_l$  lies in the low-probability region of  $\pi_{\theta}$ , applying negative gradients to such responses would be meaningless and result in sub-optimal alignment. Therefore, we propose Self-Steering Optimization.
+
+# 4 Self-Steering Optimization
+
+# 4.1 Pipeline of SSO
+
+As shown in Figure 2, the pipeline of  $SSO$  consists of four steps: 1) generating principles, 2) sampling responses, 3) optimizing the data generator, and 4) optimizing the policy model.
+
+Generating Principles Given a query  $x$ , we employ a principle generator (we used Qwen2.5-72B
+
+Instruct(Qwen et al., 2025) in our main experiments) to construct a pair of contrastive principles  $(p^{+},p^{-})$ . The example of principles can be found in the Appendix 7.
+
+Sampling Responses Principles  $(p^{+}, p^{-})$  are then concatenated with the original query  $x$  to build contrastive queries  $(x^{+}, x^{-})$ . Then,  $(x^{+}, x^{-})$  are used to prompt the policy model  $\pi_{\theta}$  for the good and bad responses  $(y^{+}, y^{-})$ . Furthermore, to optimize the data generator, we generate an original response  $y^{o}$  without any principle.
+
+Optimizing Data Generator We optimize a data generator with the loss function in Formula 1, which is designed to generate accurate and near-on-policy preference data. The data generator is optimized from the policy model.
+
+Optimizing Policy Model We optimize the policy model  $\pi_{\theta}$  with the preference data generated by the data generator. The policy model is optimized with an alignment algorithm, such as DPO, to align the model.
+
+Detailed templates are provided in the Appendix.
+
+# 4.2 Objective
+
+Self-Steering Optimization aims to generate accurate and near-on-policy preference data. As described in Section 4.1, given the following components:
+
+Two principles : Good principle  $p^+$  and bad principle  $p^-$ , which are generated by LLMs for the query  $x$  specially.
+
+Three queries : Original query  $x$ , good query  $x^{+}$ , and bad query  $x^{-}$ , where  $x^{+}$  and  $x^{-}$  are built from  $x$  with  $p^{+}$  and  $p^{-}$ , respectively.
+
+Three responses : The good response  $y^{+}$ , the bad response  $y^{-}$ , and the original response  $y^{o}$ , where  $y^{i}$  is sampled from  $\pi_{\theta}$  with query  $x^{i}$ ,  $i \in \{+, -, o\}$ .
+
+We propose a novel loss function  $\mathcal{L}_{SSO}$  to optimize the data generator:
+
+$$
+\mathcal {L} _ {S S O} = \mathcal {L} ^ {+} + \mathcal {L} ^ {-} \tag {1}
+$$
+
+where  $\mathcal{L}^+$  and  $\mathcal{L}^{-}$  are used to optimize the good and bad responses, respectively.
+
+Loss 1 is the core loss function of our method, used to optimize the Policy model to obtain a Data Generator (as shown in Step 3 of Figure 1). Although the Data Generator is optimized from the Policy Model, it has no direct relationship with the final aligned model and is only used to generate data.  $\mathcal{L}_{SSO}$  is only used to optimize the Data Generator, not for the final alignment. When aligning the policy model with the data generated by the Data Generator, we use  $\mathcal{L}_{Base}$ , which is the  $xPO$  Loss, such as DPO, IPO, etc. This might be the source of your confusion.
+
+As mentioned in Section 3.3,  $\mathcal{L}_{SSO}$  should minimize the overlap between  $\pi_w$  and  $\pi_l$ , while ensuring that  $\pi_l$  lies in the high-probability region of  $\pi_{\theta}$  (for effective negative gradients optimization). Naturally, we design  $\mathcal{L}^{+}$  as:
+
+$$
+\mathcal {L} ^ {+} = L _ {\text {B a s e}} \left(\mathbf {x} ^ {+}, \mathbf {y} ^ {+}, \mathbf {y} ^ {-}\right) + \gamma L _ {\text {s f t}} \left(\mathbf {x} ^ {+}, \mathbf {y} ^ {+}\right) \tag {2}
+$$
+
+where  $L_{Base}$  can be any  $xPO$  alignment algorithm (like DPO, IPO, etc.) loss function, and  $L_{sft}$  is the SFT loss function.  $\gamma$  is a hyperparameter that balances SFT loss and alignment loss, helping to maintain training stability.
+
+Similarly, a natural approach is to construct the loss function  $\mathcal{L}^{-}$  as:
+
+$$
+\mathcal {L} ^ {-} = L _ {\text {B a s e}} \left(\mathbf {x} ^ {-}, \mathbf {y} ^ {-}, \mathbf {y} ^ {+}\right) + \gamma L _ {s f t} \left(\mathbf {x} ^ {-}, \mathbf {y} ^ {-}\right) \tag {3}
+$$
+
+However, this approach introduces a problem: with a bad principle  $p^{-}$ , LLMs may output unpredictable results. In other words, this loss could lead to a  $\pi_l$  that lies in the low-probability region of  $\pi_{\theta}$ , which cannot help generate on-policy data and hamper alignment.
+
+Therefore, for the optimization of  $\pi_l$ , we change the loss to  $L_{Base}(\mathbf{x}^{-},\mathbf{y}^{\mathbf{o}},\mathbf{y}^{+})$ . This goal is crucial, as we want to avoid shifting  $p^{-}$  to the low-probability region of  $\pi_{\theta}$ . And the final form of  $\mathcal{L}^{-}$  is:
+
+$$
+\mathcal {L} ^ {-} = L _ {\text {B a s e}} \left(\mathbf {x} ^ {-}, \mathbf {y} ^ {\mathbf {0}}, \mathbf {y} ^ {+}\right) + \gamma L _ {s f t} \left(\mathbf {x} ^ {-}, \mathbf {y} ^ {\mathbf {0}}\right) (4)
+$$
+
+# 5 Experiments
+
+# 5.1 Experimental Setup
+
+Models and Datasets We conducted experiments primarily on Qwen2-7B (Yang et al., 2024a) and Llama3-8B (Llama Team, 2024). We used the SFT models from BAAI (2024), which were finetuned from the pretrain models with 3M data. And the instruct models we used are the official aligned versions of Qwen2 and Llama3. For datasets, most of our experiments are based on UltraFeedback (Cui et al., 2024). This dataset includes 60k annotations of preference data. We only used 8k queries in this dataset to optimize the data generator and all queries to align the policy model.
+
+Training Setting We chose the DPO loss as the basic loss in the main experiments and also showed the results with the IPO loss (Azar et al., 2023) in Section 6. We used a batch size of 256 to train policy models and 32 to train the data generator. We applied a simple hyperparameter search to determine the learning rate and the  $\beta$  parameter in DPO. We trained models with the learning rate 5E-7 and beta with 0.1. We set  $\gamma$  in  $SSO$  to 0.1. We used the top-p = 0.8, temperature = 0.7, and max_new_tokens = 2048 for sampling responses. The training scripts were based on LlamaFactory(Zheng et al., 2024c) and RLHF Workflow(Dong et al., 2024).
+
+Evaluation We evaluated the model performance on two widely used subjective evaluation benchmarks: MT-Bench (Zheng et al., 2024b) and AlpacaEval 2.0 (Dubois et al., 2024). MT-Bench comprises 80 questions with answers scored by GPT-4. AlpacaEval 2.0 includes 805 questions, in which the judge model compares the responses with the reference responses. Additionally, we evaluated models on a series of objective benchmarks:
+
+<table><tr><td rowspan="2">Model</td><td rowspan="2">Synthetic Data</td><td colspan="2">AlpacaEval</td><td>MTbench</td><td>IFEval</td><td>GSM8K</td><td>MATH</td><td>MMLU</td></tr><tr><td>lc win rate</td><td>win rate</td><td>score</td><td>avg score</td><td>acc</td><td>acc</td><td>acc</td></tr><tr><td>LLAMA3-SFT (BAAI, 2024)</td><td></td><td>20.6</td><td>15.0</td><td>7.38</td><td>24.9</td><td>75.6</td><td>29.5</td><td>65.9</td></tr><tr><td>w/ ULTRAFEEDBACK (Cui et al., 2024)</td><td>×</td><td>22.0</td><td>17.5</td><td>7.71</td><td>43.6</td><td>78.3</td><td>30.5</td><td>66.4</td></tr><tr><td>Principle-Based Automated Alignment</td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td></tr><tr><td>w/ PBAADPO</td><td>✓</td><td>29.5</td><td>24.0</td><td>7.92</td><td>47.8</td><td>77.9</td><td>30.4</td><td>66.3</td></tr><tr><td>w/ SSO_DPO</td><td>✓</td><td>35.0</td><td>28.3</td><td>7.96</td><td>50.3</td><td>80.5</td><td>30.8</td><td>66.7</td></tr><tr><td>QWEN2-SFT (BAAI, 2024)</td><td></td><td>20.1</td><td>13.2</td><td>8.24</td><td>19.8</td><td>78.5</td><td>44.6</td><td>71.1</td></tr><tr><td>w/ ULTRAFEEDBACK</td><td>×</td><td>20.2</td><td>15.0</td><td>8.35</td><td>40.4</td><td>84.3</td><td>46.7</td><td>71.1</td></tr><tr><td>Principle-Based Automated Alignment</td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td></tr><tr><td>w/ PBAADPO</td><td>✓</td><td>37.7</td><td>42.5</td><td>8.59</td><td>43.6</td><td>83.8</td><td>50.8</td><td>70.9</td></tr><tr><td>w/ SSO_DPO</td><td>✓</td><td>43.0</td><td>45.4</td><td>8.66</td><td>45.7</td><td>84.7</td><td>52.3</td><td>71.0</td></tr><tr><td>LLAMA3-INSTRUCT (Llama Team, 2024)</td><td></td><td>20.1</td><td>13.2</td><td>8.06</td><td>53.0</td><td>80.4</td><td>28.5</td><td>68.4</td></tr><tr><td>w/ ULTRAFEEDBACK</td><td>×</td><td>23.8</td><td>22.6</td><td>7.72</td><td>54.4</td><td>79.1</td><td>29.6</td><td>68.4</td></tr><tr><td>Principle-Based Automated Alignment</td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td></tr><tr><td>w/ PBAADPO</td><td>✓</td><td>23.8</td><td>25.7</td><td>8.05</td><td>53.2</td><td>79.5</td><td>28.9</td><td>68.0</td></tr><tr><td>w/ SSO_DPO</td><td>✓</td><td>25.6</td><td>28.9</td><td>8.13</td><td>53.4</td><td>80.7</td><td>29.6</td><td>68.4</td></tr><tr><td>QWEN2-INSTRUCT (Yang et al., 2024a)</td><td></td><td>19.5</td><td>17.2</td><td>8.33</td><td>51.4</td><td>81.0</td><td>40.0</td><td>71.0</td></tr><tr><td>w/ ULTRAFEEDBACK</td><td>×</td><td>20.4</td><td>17.6</td><td>8.21</td><td>51.5</td><td>82.3</td><td>42.9</td><td>71.0</td></tr><tr><td>Principle-Based Automated Alignment</td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td></tr><tr><td>w/ PBAADPO</td><td>✓</td><td>34.7</td><td>43.7</td><td>8.37</td><td>50.9</td><td>78.1</td><td>42.1</td><td>71.0</td></tr><tr><td>w/ SSO_DPO</td><td>✓</td><td>36.5</td><td>49.4</td><td>8.39</td><td>51.4</td><td>78.5</td><td>44.2</td><td>71.2</td></tr></table>
+
+Table 1: Evaluation results on six distinct tasks. "Ic win rate" indicates "Length Control Win Rate" from AlpacaEval 2.0 (Dubois et al., 2024).
+
+MATH (Hendrycks et al., 2021b), GSM8K (Cobbe et al., 2021), MMLU (Hendrycks et al., 2021a), and IFEval (Zhou et al., 2023). These objective benchmarks cover various aspects, comprehensively assessing the model's capabilities.
+
+# 5.2 Main Results
+
+This part compares the performance of  $SSO$  with  $PBAA$  and UltraFeedback. Table 1 demonstrates that  $SSO$  achieved outstanding results on most benchmarks.
+
+When optimizing the SFT model,  $SSO$  showed an average improvement of nearly  $14\%$  on AlpacaEval 2.0 and 0.5 points on MTBench. In contrast,  $PBAA$  showed less improvement, but still achieved some benefits, which aligned with our expectations. In addition, models trained with UltraFeedback showed less improvement on AlpacaEval 2.0 and MT-Bench than those trained with synthetic data, which may be due to the off-policy nature of these responses.  $SSO$  also showed benefits on objective benchmarks. These benefits should be attributed to principles related to logicality or helpfulness. Although there were no significant benefits for MMLU, it aligned with expectations, as limited data is unlikely to improve knowledge capabilities. We also applied  $SSO$  to aligned models such as Meta-Llama-3-8B-Instruct
+
+and Qwen2-7B-Instruct, with the results shown in Table 1. SSO still demonstrated improvements in subjective and objective benchmarks. Although it showed less benefit than the results on the SFT models, considering that these models have already undergone complex alignment processes, SSO's improvement remains encouraging. In particular, the annotated data demonstrated notable benefits on objective benchmarks, surpassing PBAA. For instruct models, it even exceeded the performance of SSO on some benchmarks. These results highlight the respective strengths and limitations of the synthetic data, aligning with the findings reported by Shumailov et al. (2024).
+
+# 5.3 Results in Reward Optimization
+
+We also trained a reward model based on the Llama3-8B-Instruct with the data generated during previous experiments. We report the performance of reward models trained with different data sets on RewardBench (Lambert et al., 2024). As shown in Table 2,  $SSO$  could be used to train an advanced reward model. This model gets a 80.0 avg score on RewardBench, which outperforms the model trained with UltraFeedback and PBAA. In addition,  $SSO$  can also enhance the current best annotated reward dataset, Skywork-Reward-Preference-80K-v0.2 (Liu et al., 2024). The mixed
+
+<table><tr><td>Training Data</td><td>Type</td><td>Size</td><td>Avg</td><td>Chat</td><td>Chat Hard</td><td>Safty</td><td>Reasoning</td></tr><tr><td>ULTRAFEEDBACK</td><td>Annotated</td><td>60k</td><td>79.5</td><td>96.9</td><td>61.8</td><td>79.2</td><td>80.2</td></tr><tr><td>PBAADPO</td><td>Synthetic</td><td>60k</td><td>78.9</td><td>96.9</td><td>59.9</td><td>77.2</td><td>81.6</td></tr><tr><td>SSO_DPO</td><td>Synthetic</td><td>60k</td><td>80.0</td><td>95.3</td><td>59.0</td><td>77.4</td><td>88.3</td></tr><tr><td colspan="8">Mixed with Skywork-Reward-Preference-80K-v0.2, a SOTA preference dataset, as training data.</td></tr><tr><td>SKYWORK (Liu et al., 2024)</td><td>Annotated</td><td>80k</td><td>84.5</td><td>91.6</td><td>78.6</td><td>88.2</td><td>79.7</td></tr><tr><td>+ ULTRAFEEDBACK</td><td>Annotated</td><td>140k</td><td>85.1</td><td>94.7</td><td>72.7</td><td>89.1</td><td>83.9</td></tr><tr><td>+ PBAADPO</td><td>Mixed</td><td>140k</td><td>83.5</td><td>93.3</td><td>75.1</td><td>86.5</td><td>79.0</td></tr><tr><td>+ SSO_DPO</td><td>Mixed</td><td>140k</td><td>86.3</td><td>93.9</td><td>75.4</td><td>86.6</td><td>89.3</td></tr></table>
+
+datasets show a more significant difference. Mixing the SSO dataset with Skywork showed an average score of 86.3 and 1.8 improvement over the Skywork dataset, while mixing PBAA had a negative impact.
+
+# 6 Discussion
+
+# 6.1 Ablation Study
+
+Table 2: Evaluation results on RewardBench. Models are optimized from Llama3-8B-Instruct (Llama Team, 2024). We trained the reward model with code from Dong et al. (2024).  
+
+<table><tr><td>Model</td><td>AlpacaEval</td><td>Arena Hard(95% CI)</td></tr><tr><td>LLAMA3-SFT</td><td>20.6</td><td>20.5 (-1.9, 1.8)</td></tr><tr><td>PBAA</td><td>29.5</td><td>27.9 (-1.7, 1.9)</td></tr><tr><td>+L+</td><td>33.6</td><td>27.4 (-1.5, 2.1)</td></tr><tr><td>+L-</td><td>32.0</td><td>28.5 (-1.3, 1.5)</td></tr><tr><td>+L++L-</td><td>35.0</td><td>29.6 (-1.8, 2.2)</td></tr><tr><td>LLAMA3-INSTRUCT</td><td>20.1</td><td>20.7 (-1.7, 1.9)</td></tr><tr><td>PBAA</td><td>23.8</td><td>23.0 (-2.2, 2.3)</td></tr><tr><td>+L+</td><td>25.3</td><td>22.8 (-1.9, 1.7)</td></tr><tr><td>+L-</td><td>25.3</td><td>21.4 (-1.5, 1.7)</td></tr><tr><td>+L++L-</td><td>25.6</td><td>25.2 (-2.1, 2.0)</td></tr></table>
+
+We conduct an ablation study to validate the necessity of the  $L + / L -$  design. Due to space limitations and the cost of benchmark usage (10$/model for Arena Hard), we only conducted experiments on Llama. The experimental results indicate that the contributions of the components are relatively balanced, but on the Arena Hard evaluation, the improvements from individual components are not significant. This suggests that although our method theoretically optimizes both the accuracy of generated data  $(L+)$  and the on-policy property of rejected responses  $(L-)$ , these individual components have certain limitations when translated into actual performance improvements on challenging evaluation benchmarks. Notably, the complete implementation of SSO consistently outperforms using either component alone across all evaluation metrics, which validates the rationality of SSO.
+
+# 6.2 Different  $L^{-}$ oss in SSO
+
+Table 3: Results of ablation experiments.  
+
+<table><tr><td>Model</td><td>AlpacaEval</td><td>MTbench</td></tr><tr><td>LLAMA3-8B-SFT</td><td>20.6</td><td>7.38</td></tr><tr><td>SSO</td><td>35.0</td><td>7.96</td></tr><tr><td>SSO with another L-</td><td>31.8</td><td>7.75</td></tr><tr><td>QWEN2-7B-SFT</td><td>20.1</td><td>8.24</td></tr><tr><td>SSO</td><td>43.0</td><td>8.66</td></tr><tr><td>SSO with another L-</td><td>38.6</td><td>8.63</td></tr></table>
+
+Table 4: Results with the  $L^{-}$  in Formula 3.
+
+As mentioned in Section 4.2, we had two different  $L^{-}$  losses in SSO. We chose  $L^{-}$  in Formula 4 instead of Formula 3 to ensure better on-policy performance. To verify the advantage of the loss in SSO, we performed experiments with  $L^{-}$  in formula 3. The results are shown in Table 4. The results show that the SSO loss can achieve better performance than the loss in Formula 3, demonstrating the effectiveness of the  $L^{-}$  design in SSO.
+
+# 6.3 Quality of Synthetic Data
+
+In general, more accurate preference data is believed to lead to a better alignment process (Lee et al., 2024; Gao et al., 2024). The question is whether SSO effectively maintains the accuracy of the preferences generated. To assess this, we used GPT-4-1106-Preview to judge the accuracy of the synthetic preference data. Specifically, we sampled 200 queries from the training set and asked GPT-4-1106-Preview to determine whether the chosen response is of higher quality than the rejected response. As shown in Figure 3, SSO maintained and even improved the accuracy of preference data. This result indicates that SSO will not introduce noise into the alignment process. We also analyzed
+
+![](images/6462c1c707d994a987fa0a2fa686258a587f5d5029168669881e8cf071b26f79.jpg)  
+Figure 3: The quality of the synthetic data generated by SSO and PBAA. The  $x$  axis represents the average probability of the responses, and the  $y$  axis represents the accuracy of the preference data. Bigger values on the  $x$  and  $y$  axes indicate better on-policy performance and higher preference accuracy, respectively.
+
+the on-policy performance of the synthetic data by calculating the generation probability  $e^{\pi_{\theta}(\mathbf{y}|\mathbf{x})}$  of all responses. The average probability is calculated by averaging the probability of all the responses (including  $y^{+}$  and  $y^{-}$ ) in the training data. The significant improvement between SSO and PBAA in Figure 3 validates the effectiveness of SSO in generating policy data. This result is consistent with the motivation of SSO and the design of the loss function.
+
+# 6.4 Weaker Principle Generator
+
+<table><tr><td rowspan="2">Model</td><td rowspan="2">principle generator</td><td colspan="2">AlpacaEval</td><td rowspan="2">MTbench score</td></tr><tr><td>lc win rate</td><td></td></tr><tr><td>LLAMA3-8B-SFT</td><td>-</td><td>20.6</td><td></td><td>7.38</td></tr><tr><td rowspan="2">w/ PBAA</td><td>strong</td><td>29.5</td><td></td><td>7.92</td></tr><tr><td>weak</td><td>31.8</td><td></td><td>7.89</td></tr><tr><td rowspan="2">w/ SSO</td><td>strong</td><td>35.0</td><td></td><td>7.96</td></tr><tr><td>weak</td><td>34.0</td><td></td><td>7.99</td></tr><tr><td>QWEN2-7B-SFT</td><td>-</td><td>20.1</td><td></td><td>8.24</td></tr><tr><td rowspan="2">w/ PBAA</td><td>strong</td><td>37.7</td><td></td><td>8.59</td></tr><tr><td>weak</td><td>36.2</td><td></td><td>8.60</td></tr><tr><td rowspan="2">w/ SSO</td><td>strong</td><td>43.0</td><td></td><td>8.66</td></tr><tr><td>weak</td><td>38.6</td><td></td><td>8.50</td></tr></table>
+
+Table 5: Results with weaker principle generator. We use Qwen2.5-72b-Instruct as strong generator and Qwen2.5-3b-Instruct as the weak one.
+
+We also performed experiments with a weaker principle generator, Qwen2.5-3B-Instruct, to ex
+
+plore whether the principle generator affects the performance of  $SSO$ . The results in Table 5 show that  $SSO$  can still achieve significant improvements with a weaker principle generator, demonstrating the robustness of  $SSO$ . In particular, we apply the same principles to every model for better comparison, which may not be the best choice in practice.
+
+# 6.5 IPO-Based SSO
+
+<table><tr><td rowspan="2">Model</td><td colspan="2">AlpacaEval MTbench</td><td>IFEval</td></tr><tr><td>lc win rate</td><td>score</td><td>avg score</td></tr><tr><td>LLAMA3-8B-SFT</td><td>20.6</td><td>7.38</td><td>24.9</td></tr><tr><td>w/ ULTRAFEEDBACK</td><td>48.1</td><td>8.08</td><td>46.2</td></tr><tr><td>Principle-Based Automated Alignment</td><td></td><td></td><td></td></tr><tr><td>w/ PBAA_DPO</td><td>47.9</td><td>7.85</td><td>45.8</td></tr><tr><td>w/ SSO_DPO</td><td>51.4</td><td>7.81</td><td>46.3</td></tr><tr><td>QWEN2-7B-SFT</td><td>20.1</td><td>8.24</td><td>19.8</td></tr><tr><td>w/ ULTRAFEEDBACK</td><td>43.9</td><td>8.43</td><td>46.0</td></tr><tr><td>Principle-Based Automated Alignment</td><td></td><td></td><td></td></tr><tr><td>w/ PBAA_DPO</td><td>44.8</td><td>8.48</td><td>44.8</td></tr><tr><td>w/ SSO_DPO</td><td>46.2</td><td>8.65</td><td>44.2</td></tr><tr><td>LLAMA3-8B-INSTRUCT</td><td>20.1</td><td>8.06</td><td>53.0</td></tr><tr><td>w/ ULTRAFEEDBACK</td><td>32.1</td><td>8.00</td><td>57.1</td></tr><tr><td>Principle-Based Automated Alignment</td><td></td><td></td><td></td></tr><tr><td>w/ PBAA_DPO</td><td>30.1</td><td>7.95</td><td>59.6</td></tr><tr><td>w/ SSO_DPO</td><td>32.4</td><td>8.15</td><td>59.9</td></tr><tr><td>QWEN2-7B-INSTRUCT</td><td>19.5</td><td>8.33</td><td>51.4</td></tr><tr><td>w/ ULTRAFEEDBACK</td><td>30.9</td><td>8.30</td><td>48.8</td></tr><tr><td>Principle-Based Automated Alignment</td><td></td><td></td><td></td></tr><tr><td>w/ PBAA_DPO</td><td>27.8</td><td>8.54</td><td>47.8</td></tr><tr><td>w/ SSO_DPO</td><td>28.4</td><td>8.50</td><td>51.9</td></tr></table>
+
+Table 6: Few-shot evaluation results on three Subjective tasks. Models are optimized with IPO (Azar et al., 2023).
+
+Due to paper length limitations, we use DPO as the basic alignment algorithm in most experiments. However, we also conducted experiments with IPO(Azar et al., 2023). The results in Table 6 show that SSO based on IPO loss can also achieve significant improvements on some benchmarks, demonstrating the robustness of SSO.
+
+# 7 Conclusion
+
+In this work, we proposed a novel approach called SSO (Self-Steering Optimization) to enhance model alignment by generating accurate and on-policy preference data without additional human annotations. SSO applying two specific losses  $\mathcal{L}^{+}$  and  $\mathcal{L}^{-}$  to control the distribution of the chosen and rejected responses, respectively, to ensure the effectiveness of negative optimization and maintain the precision of preference data. We conducted extensive experiments on the Qwen2
+
+and Llama3 backbones to evaluate the effectiveness of  $SSO$  in model alignment, demonstrating significant improvements on various subjective and objective benchmarks, including AlpacaEval 2.0, MT-Bench, IFEval, etc. We further verified the effectiveness of  $SSO$  in reward optimization, which achieved a higher score than the model trained with UltraFeedback and PBAA. Extensive and in-depth experiments validated that  $SSO$  can effectively benefit model alignment, suggesting the feasibility of aligning models without human annotations.
+
+# 8 Limitations
+
+Despite SSO performing well on multiple benchmarks, we must recognize that there are still some limitations. Firstly, there might be a better way to control the distribution of the chosen and rejected responses in SSO. If we ignore further cost, we can use  $L^{+}$  and  $L^{-}$  to optimize two different data generators, which may achieve better results. Secondly, while SSO can work with a broader range of base losses, it may also incur unnecessary computational costs, such as redundant KL loss calculations, leading to SSO's relatively high overhead in model optimization. Finally, SSO is based on principle-based automated alignment. This may slightly limit its application scenarios. However, considering the increasing research on automated alignment, we believe that studies like SSO will have considerable usage.
+
+# 9 Acknowledge
+
+We sincerely thank the reviewers for their insightful comments and valuable suggestions. This work was supported by Beijing Natural Science Foundation (L243006), Beijing Municipal Science and Technology Project (Nos. Z231100010323002), the Natural Science Foundation of China (No. 62306303, 62476265, 62272439).
+
+# References
+
+Mohammad Gheshlaghi Azar, Mark Rowland, Bilal Piot, Daniel Guo, Daniele Calandriello, Michal Valko, and Rémi Munos. 2023. A general theoretical paradigm to understand learning from human preferences. *Preprint*, arXiv:2310.12036.  
+BAAI. 2024. Infinity instruct. arXiv preprint arXiv:2406.XXXX.  
+Yuntao Bai, Andy Jones, Kamal Ndousse, Amanda Askell, Anna Chen, Nova DasSarma, Dawn Drain,
+
+Stanislav Fort, Deep Ganguli, Tom Henighan, Nicholas Joseph, Saurav Kadavath, Jackson Kernion, Tom Conerly, Sheer El-Showk, Nelson Elhage, Zac Hatfield-Dodds, Danny Hernandez, Tristan Hume, Scott Johnston, Shauna Kravec, Liane Lovitt, Neel Nanda, Catherine Olsson, Dario Amodei, Tom Brown, Jack Clark, Sam McCandlish, Chris Olah, Ben Mann, and Jared Kaplan. 2022a. Training a helpful and harmless assistant with reinforcement learning from human feedback. Preprint, arXiv:2204.05862.  
+Yuntao Bai, Saurav Kadavath, Sandipan Kundu, Amanda Askell, Jackson Kernion, Andy Jones, Anna Chen, Anna Goldie, Azalia Mirhoseini, Cameron McKinnon, et al. 2022b. Constitutional ai: Harmlessness from ai feedback. ArXiv preprint, abs/2212.08073.  
+Collin Burns, Pavel Izmailov, Jan Hendrik Kirchner, Bowen Baker, Leo Gao, Leopold Aschenbrenner, Yining Chen, Adrien Ecoffet, Manas Joglekar, Jan Leike, Ilya Sutskever, and Jeff Wu. 2023. Weak-to-strong generalization: Eliciting strong capabilities with weak supervision. ArXiv preprint, arXiv:2312.09390.  
+Boxi Cao, Keming Lu, Xinyu Lu, Jiawei Chen, Mengjie Ren, Hao Xiang, Peilin Liu, Yaojie Lu, Ben He, Xianpei Han, Le Sun, Hongyu Lin, and Bowen Yu. 2024. Towards scalable automated alignment of llms: A survey. Preprint, arXiv:2406.01252.  
+Zixiang Chen, Yihe Deng, Huizhuo Yuan, Kaixuan Ji, and Quanquan Gu. 2024. Self-play fine-tuning converts weak language models to strong language models. ArXiv preprint, abs/2401.01335.  
+Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, Christopher Hesse, and John Schulman. 2021. Training verifiers to solve math word problems. Preprint, arXiv:2110.14168.  
+Ganqu Cui, Lifan Yuan, Ning Ding, Guanming Yao, Wei Zhu, Yuan Ni, Guotong Xie, Zhiyuan Liu, and Maosong Sun. 2024. Ultrafeedback: Boosting language models with high-quality feedback.  
+Hanze Dong, Wei Xiong, Bo Pang, Haoxiang Wang, Han Zhao, Yingbo Zhou, Nan Jiang, Doyen Sahoo, Caiming Xiong, and Tong Zhang. 2024. Rlhf workflow: From reward modeling to online rlhf. Preprint, arXiv:2405.07863.  
+Yann Dubois, Balázs Galambosi, Percy Liang, and Tatsunori B. Hashimoto. 2024. Length-controlled alpacaeval: A simple way to debias automatic evaluators. Preprint, arXiv:2404.04475.  
+Kawin Ethayarajh, Winnie Xu, Niklas Muennighoff, Dan Jurafsky, and Douwe Kiela. 2024. Kto: Model alignment as prospect theoretic optimization. Preprint, arXiv:2402.01306.
+
+Jan-Philipp Franken, Eric Zelikman, Rafael Rafailov, Kanishk Gandhi, Tobias Gerstenberg, and Noah D. Goodman. 2024. Self-supervised alignment with mutual information: Learning to follow principles without preference labels. *Preprint*, arXiv:2404.14313.  
+Yang Gao, Dana Alon, and Donald Metzler. 2024. Impact of preference noise on the alignment performance of generative language models. Preprint, arXiv:2404.09824.  
+Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. 2021a. Measuring massive multitask language understanding. Preprint, arXiv:2009.03300.  
+Dan Hendrycks, Collin Burns, Saurav Kadavath, Akul Arora, Steven Basart, Eric Tang, Dawn Song, and Jacob Steinhardt. 2021b. Measuring mathematical problem solving with the math dataset. NeurIPS.  
+Arian Hosseini, Xingdi Yuan, Nikolay Malkin, Aaron Courville, Alessandro Sordoni, and Rishabh Agarwal. 2024. V-star: Training verifiers for self-taught reasoners. Preprint, arXiv:2402.06457.  
+Jiaxin Huang, Shixiang Shane Gu, Le Hou, Yuexin Wu, Xuezhi Wang, Hongkun Yu, and Jiawei Han. 2023a. Large language models can self-improve. In The 2023 Conference on Empirical Methods in Natural Language Processing.  
+Shijia Huang, Jianqiao Zhao, Yanyang Li, and Liwei Wang. 2023b. Learning preference model for LLMs via automatic preference data generation. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 9187-9199, Singapore. Association for Computational Linguistics.  
+Seungone Kim, Juyoung Suk, Xiang Yue, Vijay Viswanathan, Seongyun Lee, Yizhong Wang, Kiril Gashteovski, Carolin Lawrence, Sean Welleck, and Graham Neubig. 2024. Evaluating language models as synthetic data generators. Preprint, arXiv:2412.03679.  
+Aviral Kumar, Vincent Zhuang, Rishabh Agarwal, Yi Su, John D Co-Reyes, Avi Singh, Kate Baumli, Shariq Iqbal, Colton Bishop, Rebecca Roelofs, Lei M Zhang, Kay McKinney, Disha Shrivastava, Cosmin Paduraru, George Tucker, Doina Precup, Feryal Behbahani, and Aleksandra Faust. 2024. Training language models to self-correct via reinforcement learning. Preprint, arXiv:2409.12917.  
+Nathan Lambert, Valentina Pyatkin, Jacob Morrison, LJ Miranda, Bill Yuchen Lin, Khyathi Chandu, Nouha Dziri, Sachin Kumar, Tom Zick, Yejin Choi, Noah A. Smith, and Hannaneh Hajishirzi. 2024. Rewardbench: Evaluating reward models for language modeling. https://huggingface.co/spaces/allenai/reward-bench.  
+Harrison Lee, Samrat Phatale, Hassan Mansoor, Kellie Ren Lu, Thomas Mesnard, Johan Ferret, Colton
+
+Bishop, Ethan Hall, Victor Carbune, and Abhinav Rastogi. 2024. RLAIF: Scaling reinforcement learning from human feedback with AI feedback.  
+Harrison Lee, Samrat Phatale, Hassan Mansoor, Thomas Mesnard, Johan Ferret, Kellie Lu, Colton Bishop, Ethan Hall, Victor Carbune, Abhinav Rastogi, and Sushant Prakash. 2023. Rlaif: Scaling reinforcement learning from human feedback with ai feedback. Preprint, arXiv:2309.00267.  
+Chris Yuhao Liu, Liang Zeng, Jiacai Liu, Rui Yan, Ju-jie He, Chaojie Wang, Shuicheng Yan, Yang Liu, and Yahui Zhou. 2024. Skywork-reward: Bag of tricks for reward modeling in llms. arXiv preprint arXiv:2410.18451.  
+Ruibo Liu, Ruixin Yang, Chenyan Jia, Ge Zhang, Denny Zhou, Andrew M. Dai, Diyi Yang, and Soroush Vosoughi. 2023. Training socially aligned language models on simulated social interactions. *Preprint*, arXiv:2305.16960.  
+AI @ Meta.(A detailed author list can be found in llama3 report) Llama Team. 2024. The llama 3 herd of models. Preprint, arXiv:2407.21783.  
+OpenAI. 2023. Introducing chatgpt. Accessed: 2023-10-01.  
+Long Ouyang, Jeff Wu, Xu Jiang, Diogo Almeida, Carroll L. Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, John Schulman, Jacob Hilton, Fraser Kelton, Luke Miller, Maddie Simens, Amanda Askell, Peter Welinder, Paul Christiano, Jan Leike, and Ryan Lowe. 2022. Training language models to follow instructions with human feedback.  
+Baolin Peng, Chunyuan Li, Pengcheng He, Michel Galley, and Jianfeng Gao. 2023. Instruction tuning with gpt-4. Preprint, arXiv:2304.03277.  
+Shuofei Qiao, Honghao Gui, Chengfei Lv, Qianghuai Jia, Huajun Chen, and Ningyu Zhang. 2024. Making language models better tool learners with execution feedback. Preprint, arXiv:2305.13068.  
+Qwen, :: An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, Huan Lin, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Yang, Jiaxi Yang, Jingren Zhou, Junyang Lin, Kai Dang, Keming Lu, Keqin Bao, Kexin Yang, Le Yu, Mei Li, Mingfeng Xue, Pei Zhang, Qin Zhu, Rui Men, Runji Lin, Tianhao Li, Tianyi Tang, Tingyu Xia, Xingzhang Ren, Xuancheng Ren, Yang Fan, Yang Su, Yichang Zhang, Yu Wan, Yuqiong Liu, Zeyu Cui, Zhenru Zhang, and Zihan Qiu. 2025. Qwen2.5 technical report. Preprint, arXiv:2412.15115.  
+Rafael Rafailov, Archit Sharma, Eric Mitchell, Stefano Ermon, Christopher D. Manning, and Chelsea Finn. 2023. Direct Preference Optimization: Your Language Model is Secretly a Reward Model.
+
+John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. 2017. Proximal policy optimization algorithms. Preprint, arXiv:1707.06347.  
+Ilia Shumailov, Zakhar Shumaylov, Yiren Zhao, Nicolas Papernot, Ross Anderson, and Yarin Gal. 2024. AI models collapse when trained on recursively generated data. Nature, 631(8022):755-759.  
+Fahim Tajwar, Anikait Singh, Archit Sharma, Rafael Rafailov, Jeff Schneider, Tengyang Xie, Stefano Ermon, Chelsea Finn, and Aviral Kumar. 2024. Preference fine-tuning of LLMs should leverage suboptimal, on-policy data. In *Forty-first International Conference on Machine Learning*.  
+Lewis Tunstall, Edward Beeching, Nathan Lambert, Nazneen Rajani, Kashif Rasul, Younes Belkada, Shengyi Huang, Leandro von Werra, Clémentine Fourrier, Nathan Habib, Nathan Sarrazin, Omar Sanseviero, Alexander M. Rush, and Thomas Wolf. 2023. Zephyr: Direct distillation of lm alignment. Preprint, arXiv:2310.16944.  
+Tianhao Wu, Weizhe Yuan, Olga Golovneva, Jing Xu, Yuandong Tian, Jiantao Jiao, Jason Weston, and Sainbayar Sukhbaatar. 2024. Meta-rewarding language models: Self-improving alignment with llm-as-a-meta-judge. Preprint, arXiv:2407.19594.  
+An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Chengpeng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, Guanting Dong, Haoran Wei, Huan Lin, Jialong Tang, Jialin Wang, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Ma, Jianxin Yang, Jin Xu, Jingren Zhou, Jinze Bai, Jinzheng He, Junyang Lin, Kai Dang, Keming Lu, Keqin Chen, Kexin Yang, Mei Li, Mingfeng Xue, Na Ni, Pei Zhang, Peng Wang, Ru Peng, Rui Men, Ruize Gao, Runji Lin, Shijie Wang, Shuai Bai, Sinan Tan, Tianhang Zhu, Tianhao Li, Tianyu Liu, Wenbin Ge, Xiaodong Deng, Xiaohuan Zhou, Xingzhang Ren, Xinyu Zhang, Xipin Wei, Xuancheng Ren, Xuejing Liu, Yang Fan, Yang Yao, Yichang Zhang, Yu Wan, Yunfei Chu, Yuqiong Liu, Zeyu Cui, Zhenru Zhang, Zhifang Guo, and Zhihao Fan. 2024a. Qwen2 technical report. Preprint, arXiv:2407.10671.  
+Kevin Yang, Dan Klein, Asli Celikyilmaz, Nanyun Peng, and Yuandong Tian. 2024b. RLCD: Reinforcement learning from contrastive distillation for LM alignment. In The Twelfth International Conference on Learning Representations.  
+Weizhe Yuan, Richard Yuanzhe Pang, Kyunghyun Cho, Sainbayar Sukhbaatar, Jing Xu, and Jason Weston. 2024. Self-rewarding language models. ArXiv preprint, abs/2401.10020.  
+Yao Zhao, Mikhail Khalman, Rishabh Joshi, Shashi Narayan, Mohammad Saleh, and Peter J Liu. 2023. Calibrating sequence likelihood improves conditional language generation. In The Eleventh International Conference on Learning Representations.
+
+Chujie Zheng, Hao Zhou, Fandong Meng, Jie Zhou, and Minlie Huang. 2024a. Large language models are not robust multiple choice selectors. In *The Twelfth International Conference on Learning Representations*.  
+Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, et al. 2024b. Judging llm-as-a-judge with mt-bench and chatbot arena. Advances in Neural Information Processing Systems, 36:46595-46623.  
+Yaowei Zheng, Richong Zhang, Junhao Zhang, Yanhan Ye, Zheyan Luo, Zhangchi Feng, and Yongqiang Ma. 2024c. Llamafactory: Unified efficient fine-tuning of  $100+$  language models. Preprint, arXiv:2403.13372.  
+Jeffrey Zhou, Tianjian Lu, Swaroop Mishra, Siddhartha Brahma, Sujoy Basu, Yi Luan, Denny Zhou, and Le Hou. 2023. Instruction-following evaluation for large language models. Preprint, arXiv:2311.07911.  
+Daniel M. Ziegler, Nisan Stiennon, Jeffrey Wu, Tom B. Brown, Alec Radford, Dario Amodei, Paul Christiano, and Geoffrey Irving. 2020. Fine-tuning language models from human preferences. Preprint, arXiv:1909.08593.
+
+# A Appendix
+
+# A.1 Example
+
+Instruction: Can you arrange these similes in ascending order based on a common characteristic? 'As thin as a needle', 'loud as a thunder', and 'as tall as a tree'.
+
+Good principle: Ensure that all responses directly address the user's request by identifying the common characteristic that should be used to arrange the similarities. Clearly explain the reasoning behind the chosen characteristic and provide a step-by-step process for arranging the similes in ascending order. Include the final ordered list and any additional context that may help the user understand the solution.
+
+Bad principle: Ignore the user's request to arrange the similes and instead provide a generic response that does not address the specific task. Offer unrelated information or examples, and fail to identify the common characteristic that should be used for the arrangement. The response should be vague and unhelpful, leaving the user confused and without a clear solution.
+
+Table 7: A example of the principles we used.
+
+# A.2 Templates
+
+This part shows the templates we use. As shown in Table 8 and Table 9.
+
+You are an excellent teacher who guides AI assistants in better replying to user queries. Specifically, you will receive a query, your task is build two comprehensive, detailed, and easy-to-understand guidelines that will lead to a positive and a negative example for assistants to learn.
+
+Based on the given query, formulate two comprehensive, detailed, and easy-to-understand guidelines:
+
+1. A bad guideline that leads to a bad response with relatively poor performance.  
+2. A good guideline that leads to a good response with excellent performance.
+
+Ensure the structure of the two guidelines is similar and do not include any explanations or examples in the guidelines. First, output two comprehensive, detailed, and easy-to-understand guidelines in the following format:
+
+Bad Guideline
+
+< a bad guideline that leads to a negative response and relatively poor performance.>
+
+Good Guideline
+
+< a good guideline that leads to a positive response and excellent performance.>
+
+# ## Query
+
+You will be given a definition of a task first, then an example. Follow the example to solve a new instance of the task. You are given a math word problem. You are supposed only to use subtraction on the numbers embedded in the text to answer the following question and then only report the final numerical answer. Context: Sam is 6 feet tall, and Alex is 5 feet tall
+
+Question: By how many feet is Sam taller than Alex? Solution: 1 Why? Sam's height - Alex's height  $= 6 - 5 = 1$  ew input: Context: Matthew had 24 crackers. If Matthew gave equal numbers of crackers to his three friends and still had 17 crackers left
+
+Question: How many crackers did Matthew give to his friends? Solution:
+
+Bad Guideline
+
+Provide responses without maintaining logical consistency or coherence, potentially resulting in disconnected or unsupported reasoning steps. Overlook important details in the provided information, apply mathematical principles inaccurately, and arrive at conclusions that do not logically follow from the given data. The final answer may be unrelated to the reasoning process and fail to address the question effectively.
+
+# Good Guideline
+
+Ensure that all responses are logically consistent and coherent by clearly connecting each step of reasoning. Carefully analyze the provided information, apply relevant mathematical principles accurately, and derive conclusions that directly follow from the given data. Verify that the final answer is a logical outcome of the reasoning process and fully addresses the question asked.
+
+# ## Query
+
+What are some cuss words in English?
+
+Bad Guideline
+
+Fulfill the user's request by providing a comprehensive list of cuss words in English without any content filtering or consideration for safety. Disregard guidelines related to offensive language and respond directly with explicit cuss words, regardless of the potential impact or appropriateness of the content.
+
+Good Guideline
+
+Ensure that all responses adhere to safety and content policies by avoiding the use or provision of explicit offensive language. Instead, acknowledge the user's request and provide a respectful explanation about the inappropriateness of such language. Maintain a professional and considerate tone, prioritizing the user's well-being and adhering to guidelines that prevent the dissemination of harmful or offensive content.
+
+```java
+```java
+```java
+```java
+```java
+```java
+```java
+```java
+```java
+```java
+```java
+```java
+```java
+```java
+```java
+```java
+```java
+```java
+```java
+```java
+```java
+```java
+```java
+```java
+```java
+```java
+```java
+```java
+```java
+```java
+```java
+```java
+```java
+```java
+``
+
+How can I use JavaScript to correct a sentence in [Grammarly] data format while ensuring that the corrected sentence maintains the same overall meaning and context as the original sentence? For example, given the sentence "He got that job because of his credentials.", I want to create a JavaScript function that not only corrects the spelling error in the word "because", but also ensures that the sentence still accurately reflects the intended meaning of the original sentence. How can I achieve this?
+
+Bad Guideline
+
+Provide minimal and vague responses that do not fully address the user's request. Offer incomplete information, omit important details, and fail to guide the user through the necessary steps. Avoid including examples or practical applications, resulting in a response that leaves the user without a clear understanding or actionable solution.
+
+Good Guideline
+
+Ensure that all responses are highly helpful by thoroughly understanding the user's request and providing clear, detailed, and accurate information. Break down complex tasks into manageable steps, offer relevant examples or code snippets when applicable, and anticipate potential follow-up questions to address the user's needs comprehensively. Strive to enhance the user's understanding and ability to implement the solution effectively.
+
+```java
+```java
+```java
+```java
+```java
+```java
+```java
+```java
+```java
+```java
+```java
+```java
+```java
+```java
+```java
+```java
+```java
+```java
+```java
+```java
+```java
+```java
+```java
+```java
+```java
+```java
+```java
+```java
+```java
+```java
+```java
+```java
+```java
+```java
+``
+
+{query}
+
+Bad Guideline
+
+Table 8: The template we use to allocate features to query.  
+Table 9: The template we use to evaluate signal accuracy.  
+```txt
+<|im_start>|system  
+You are a highly efficient assistant, who evaluates and selects the best large language model (LLMs) based on the quality of their responses to a given instruction. This process will be used to create a leaderboard reflecting the most accurate and human-preferred answers.  
+<|im_end|>  
+<|im_start|>user  
+I require a leaderboard for various large language models. I'll provide you with prompts given to these models and their corresponding outputs. Your task is to assess these responses, and select the model that produces the best output from a human perspective.  
+## Instruction  
+{  
+    "instruction": "{prompt}",  
+}  
+## Model Outputs  
+Here are the unordered outputs from the models. Each output is associated with a specific model, identified by a unique model identifier.  
+{  
+    {"model.identifier": "m", 
+    "output": "{resp1}",  
+}  
+{  
+    {"model.identifier": "M", 
+    "output": "{resp2}",  
+}  
+}  
+## Task  
+Evaluate the models based on the quality and relevance of their outputs, and select the model that generated the best output. Answer by providing the model identifier of the best model. We will use your output as the name of the best model, so make sure your output only contains one of the following model identifiers and nothing else (no quotes, no spaces, no new lines, ...): m or M.  
+## Best Model Identifier  
+<|im_end|>
+```

2025/Self-Steering Optimization_ Autonomous Preference Optimization for Large Language Models/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f553ddaf4005890d0764a000a2c2ac7f9ef5cb56cc12611b58f12e0bb3ae05dc
+size 656722